diff --git a/UPDATING b/UPDATING index 8ed1564e6ba1..a538f0414635 100644 --- a/UPDATING +++ b/UPDATING @@ -1,2538 +1,2545 @@ Updating Information for FreeBSD stable/13 users. This file is maintained and copyrighted by M. Warner Losh . See end of file for further details. For commonly done items, please see the COMMON ITEMS: section later in the file. These instructions assume that you basically know what you are doing. If not, then please consult the FreeBSD handbook: https://www.freebsd.org/doc/en_US.ISO8859-1/books/handbook/makeworld.html Items affecting the ports and packages system can be found in /usr/ports/UPDATING. Please read that file before updating system packages and/or ports. +20220114: + The macros provided for the manipulation of CPU sets (e.g. CPU_AND) + have been modified to take 2 source arguments instead of only 1. + Externally maintained sources that use these macros will have to + be adapted. The FreeBSD version has been bumped to 1300524 to + reflect this change. + 20211218: Commit 18f5b477ee66 adds two arguments to VOP_ALLOCATE(). Normally changes to VOP calls are not MFC'd, but a discussion on freebsd-current@ seemed to agree that it was acceptable in this case, to fix the NFSv4.2 server. Any out of source tree file system with a custom VOP_ALLOCATE() call will need to be modified. Bump __FreeBSD_version to 1300523 since the NFS modules need to be rebuilt from sources. 20211124: Commit 6e8e261f0d4e changed the internal KAPI between the NFS modules. As such, all the modules must be rebuilt from sources. __FreeBSD_version has been bumped to 1300522. 20211119: Bump __FreeBSD_version to 1300521 after merging LinuxKPI and net80211 changes in order to support building various wireless drivers. This is to help other external consumers of LinuxKPI and net80211 to deal accordingly. 20211003: Commit a599f9f7620b deleted the variable called nfs_maxcopyrange from nfscommon.ko, since it no longer needs to be global. As such, the other nfs modules must be rebuilt from up to date sources. Bump __FreeBSD_version to 1300516 for this. 20210823: As of commit 622809b0868f OpenSSL no longer enables kernel TLS by default. Users can enable kernel TLS via the "KTLS" SSL option. This can be enabled globally by using a custom OpenSSL config file via OPENSSL_CONF or via an application-specific configuration option for applications which permit setting SSL options via SSL_CONF_cmd(3). 20210803: Commits 9fb6e613373c and 9ec7dbf46b0a both changed the internal KAPI between the NFS modules. Bump __FreeBSD_version to 1300514. All NFS modules must be rebuilt from sources. 20210730: Commit b69019c14cd8 removes pf's DIOCGETSTATESNV ioctl. As of be70c7a50d32 it is no longer used by userspace, but it does mean users may not be able to enumerate pf states if they update the kernel past c021ff986e2b without first updating userspace past be70c7a50d32. 20210718: Bump __FreeBSD_version to 1300512 after merging LinuxKPI, OFED, net80211, and driver changes in order to support building Intel wireless iwlwifi drivers. This is especially needed for drm-kmod which needs updates after this. 20210715: The 20210707 awk update brought in a change in behavior. This has been corrected as of d4d252c49976. Between these dates, if you installed a new awk binary, you may not be able to build a new kernel because the change in behavior affected the genoffset script used to build the kernel. If you did update, the fix is to update your sources past the above hash and do % cd usr.bin/awk % make clean all % sudo -E make install to enable building kernels again. 20210711: Commit 8a04edfdcbd2 changes the default minor version used for NFSv4 mounts to the highest minor version supported by the NFSv4 server. This default can be overridden by using the "minorversion" mount option. The consensus of a discussion on freebsd-current@ indicated that this would be acceptable to do and would not be considered a POLA violation. 20210710: awk has been updated to the latest one-true-awk version 20210215. This contains a number of minor bug fixes. 20210630: Commit ca179c4d74f2/632e3f2f3a66 changed the package in which the OpenSSL libraries and utilities are packaged. It is recommended for pkgbase user to do: pkg install -f FreeBSD-openssl before pkg upgrade otherwise some dependencies might not be met and pkg will stop working as libssl will not be present anymore on the system. 20210626: Commit 841006678745 changed the internal KAPI between the krpc and nfsd modules. As such, they must both be rebuilt from sources. __FreeBSD_version is bumped to 1300510. 20210510: Commit 272f39942254 changed the internal KAPI between the nscl.ko and nfscommon.ko modules, so they both need to be rebuilt from sources. 20210429: Commit 5a45802b3c8c changed the internal KAPI between the krpc and NFS. As such, the krpc, nfscommon and nfscl modules must all be rebuilt from sources. Without this patch, NFSv4.1/4.2 mounts should not be done with the nfscbd(8) daemon running, to avoid needing a working back channel for server->client RPCs. 20210202: Various LinuxKPI functionality was added which conflicts with DRM. Please update your drm-kmod port to after the __FreeBSD_verison 1300139 update. 20210122: The stable/13 branch has been created from main. 20210108: PC Card attachments for all devices have been removed. In the case of wi and cmx, the entire drivers were removed because they were only PC Card devices. FreeBSD_version 1300134 should be used for this since it was bumped so recently. 20210107: Transport-independent parts of HID support have been split off the USB code in to separate subsystem. Kernel configs which include one of ums, ukbd, uhid, atp, wsp, wmt, uaudio, ugold or ucycom drivers should be updated with adding of "device hid" line. 20210105: ncurses installation has been modified to only keep the widechar enabled version. Incremental build is broken for that change, so it requires a clean build. 20201223: The FreeBSD project has migrated from Subversion to Git. Temporary instructions can be found at https://github.com/bsdimp/freebsd-git-docs/blob/main/src-cvt.md and other documents in that repo. 20201216: The services database has been updated to cover more of the basic services expected in a modern system. The database is big enough that it will cause issues in mergemaster in Releases previous to 12.2 and 11.3, or in very old current systems from before r358154. 20201215: Obsolete in-tree GDB 6.1.1 has been removed. GDB (including kgdb) may be installed from ports or packages. 20201124: ping6 has been merged into ping. It can now be called as "ping -6". See ping(8) for details. 20201108: Default value of net.add_addr_allfibs has been changed to 0. If you have multi-fib configuration and rely on existence of all interface routes in every fib, you need to set the above sysctl to 1. 20201030: The internal pre-processor in the calendar(1) program has been extended to support more C pre-processor commands (e.g. #ifdef, #else, and #undef) and to detect unbalanced conditional statements. Error messages have been extended to include the filename and line number if processing stops to help fixing malformed data files. 20201026: All the data files for the calendar(1) program, except calendar.freebsd, have been moved to the deskutils/calendar-data port, much like the jewish calendar entries were moved to deskutils/hebcal years ago. After make delete-old-files, you need to install it to retain full functionality. calendar(1) will issue a reminder for files it can't find. 20200923: LINT files are no longer generated. We now include the relevant NOTES files. Note: This may cause conflicts with updating in some cases. find sys -name LINT\* -delete is suggested across this commit to remove the generated LINT files. If you have tried to update with generated files there, the svn command you want to un-auger the tree is cd sys/amd64/conf svn revert -R . and then do the above find from the top level. Substitute 'amd64' above with where the error message indicates a conflict. 20200824: OpenZFS support has been integrated. Do not upgrade root pools until the loader is updated to support zstd. Furthermore, we caution against 'zpool upgrade' for the next few weeks. The change should be transparent unless you want to use new features. Not all "NO_CLEAN" build scenarios work across these changes. Many scenarios have been tested and fixed, but rebuilding kernels without rebuilding world may fail. The ZFS cache file has moved from /boot to /etc to match the OpenZFS upstream default. A fallback to /boot has been added for mountroot. Pool auto import behavior at boot has been moved from the kernel module to an explicit "zpool import -a" in one of the rc scripts enabled by zfs_enable=YES. This means your non-root zpools won't auto import until you upgrade your /etc/rc.d files. 20200824: The resume code now notifies devd with the 'kernel' system rather than the old 'kern' subsystem to be consistent with other use. The old notification will be created as well, but will be removed prior to FreeBSD 14.0. 20200821: r362275 changed the internal API between the kernel RPC and the NFS modules. As such, all the modules must be recompiled from sources. 20200817: r364330 modified the internal API used between the NFS modules. As such, all the NFS modules must be re-compiled from sources. 20200816: Clang, llvm, lld, lldb, compiler-rt, libc++, libunwind and openmp have been upgraded to 11.0.0. Please see the 20141231 entry below for information about prerequisites and upgrading, if you are not already using clang 3.5.0 or higher. 20200810: r364092 modified the internal ABI used between the kernel NFS modules. As such, all of these modules need to be rebuilt from sources, so a version bump was done. 20200807: Makefile.inc has been updated to work around the issue documented in 20200729. It was a case where the optimization of using symbolic links to point to binaries created a situation where we'd run new binaries with old libraries starting midway through the installworld process. 20200729: r363679 has redefined some undefined behavior in regcomp(3); notably, extraneous escapes of most ordinary characters will no longer be accepted. An exp-run has identified all of the problems with this in ports, but other non-ports software may need extra escapes removed to continue to function. Because of this change, installworld may encounter the following error from rtld: Undefined symbol "regcomp@FBSD_1.6" -- It is imperative that you do not halt installworld. Instead, let it run to completion (whether successful or not) and run installworld once more. 20200627: A new implementation of bc and dc has been imported in r362681. This implementation corrects non-conformant behavior of the previous bc and adds GNU bc compatible options. It offers a number of extensions, is much faster on large values, and has support for message catalogs (a number of languages are already supported, contributions of further languages welcome). The option WITHOUT_GH_BC can be used to build the world with the previous versions of bc and dc. 20200625: r362639 changed the internal API used between the NFS kernel modules. As such, they all need to be rebuilt from sources. 20200613: r362158 changed the arguments for VFS_CHECKEXP(). As such, any out of tree file systems need to be modified and rebuilt. Also, any file systems that are modules must be rebuilt. 20200604: read(2) of a directory fd is now rejected by default. root may re-enable it for system root only on non-ZFS filesystems with the security.bsd.allow_read_dir sysctl(8) MIB if security.bsd.suser_enabled=1. It may be advised to setup aliases for grep to default to `-d skip` if commonly non-recursively grepping a list that includes directories and the potential for the resulting stderr output is not tolerable. Example aliases are now installed, commented out, in /root/.cshrc and /root/.shrc. 20200523: Clang, llvm, lld, lldb, compiler-rt, libc++, libunwind and openmp have been upgraded to 10.0.1. Please see the 20141231 entry below for information about prerequisites and upgrading, if you are not already using clang 3.5.0 or higher. 20200512: Support for obsolete compilers has been removed from the build system. Clang 6 and GCC 6.4 are the minimum supported versions. 20200424: closefrom(2) has been moved under COMPAT12, and replaced in libc with a stub that calls close_range(2). If using a custom kernel configuration, you may want to ensure that the COMPAT_FREEBSD12 option is included, as a slightly older -CURRENT userland and older FreeBSD userlands may not be functional without closefrom(2). 20200414: Upstream DTS from Linux 5.6 was merged and they now have the SID and THS (Secure ID controller and THermal Sensor) node present. The DTB overlays have now been removed from the tree for the H3/H5 and A64 SoCs and the aw_sid and aw_thermal driver have been updated to deal with upstream DTS. If you are using those overlays you need to remove them from loader.conf and update the DTBs on the FAT partition. 20200310: Clang, llvm, lld, lldb, compiler-rt, libc++, libunwind and openmp have been upgraded to 10.0.0. Please see the 20141231 entry below for information about prerequisites and upgrading, if you are not already using clang 3.5.0 or higher. 20200309: The amd(8) automount daemon has been removed from the source tree. As of FreeBSD 10.1 autofs(5) is the preferred tool for automounting. amd is still available in the sysutils/am-utils port. 20200301: Removed brooktree driver (bktr.4) from the tree. 20200229: The WITH_GPL_DTC option has been removed. The BSD-licenced device tree compiler in usr.bin/dtc is used on all architectures which use dtc, and the GPL dtc is available (if needed) from the sysutils/dtc port. 20200229: The WITHOUT_LLVM_LIBUNWIND option has been removed. LLVM's libunwind is used by all supported CPU architectures. 20200229: GCC 4.2.1 has been removed from the tree. The WITH_GCC, WITH_GCC_BOOTSTRAP, and WITH_GNUCXX options are no longer available. Users who wish to build FreeBSD with GCC must use the external toolchain ports or packages. 20200220: ncurses has been updated to a newer version (6.2-20200215). Given the ABI has changed, users will have to rebuild all the ports that are linked to ncurses. 20200217: The size of struct vnet and the magic cookie have changed. Users need to recompile libkvm and all modules using VIMAGE together with their new kernel. 20200212: Defining the long deprecated NO_CTF, NO_DEBUG_FILES, NO_INSTALLLIB, NO_MAN, NO_PROFILE, and NO_WARNS variables is now an error. Update your Makefiles and scripts to define MK_=no instead as required. One exception to this is that program or library Makefiles should define MAN to empty rather than setting MK_MAN=no. 20200108: Clang/LLVM is now the default compiler and LLD the default linker for riscv64. 20200107: make universe no longer uses GCC 4.2.1 on any architectures. Architectures not supported by in-tree Clang/LLVM require an external toolchain package. 20200104: GCC 4.2.1 is now not built by default, as part of the GCC 4.2.1 retirement plan. Specifically, the GCC, GCC_BOOTSTRAP, and GNUCXX options default to off for all supported CPU architectures. As a short-term transition aid they may be enabled via WITH_* options. GCC 4.2.1 is expected to be removed from the tree on 2020-03-31. 20200102: Support for armv5 has been disconnected and is being removed. The machine combination MACHINE=arm MACHINE_ARCH=arm is no longer valid. You must now use a MACHINE_ARCH of armv6 or armv7. The default MACHINE_ARCH for MACHINE=arm is now armv7. 20191226: Clang/LLVM is now the default compiler for all powerpc architectures. LLD is now the default linker for powerpc64. The change for powerpc64 also includes a change to the ELFv2 ABI, incompatible with the existing ABI. 20191226: Kernel-loadable random(4) modules are no longer unloadable. 20191222: Clang, llvm, lld, lldb, compiler-rt, libc++, libunwind and openmp have been upgraded to 9.0.1. Please see the 20141231 entry below for information about prerequisites and upgrading, if you are not already using clang 3.5.0 or higher. 20191212: r355677 has modified the internal interface used between the NFS modules in the kernel. As such, they must all be upgraded simultaneously. I will do a version bump for this. 20191205: The root certificates of the Mozilla CA Certificate Store have been imported into the base system and can be managed with the certctl(8) utility. If you have installed the security/ca_root_nss port or package with the ETCSYMLINK option (the default), be advised that there may be differences between those included in the port and those included in base due to differences in nss branch used as well as general update frequency. Note also that certctl(8) cannot manage certs in the format used by the security/ca_root_nss port. 20191120: The amd(8) automount daemon has been disabled by default, and will be removed in the future. As of FreeBSD 10.1 the autofs(5) is available for automounting. 20191107: The nctgpio and wbwd drivers have been moved to the superio bus. If you have one of these drivers in a kernel configuration, then you should add device superio to it. If you use one of these drivers as a module and you compile a custom set of modules, then you should add superio to the set. 20191021: KPIs for network drivers to access interface addresses have changed. Users need to recompile NIC driver modules together with kernel. 20191021: The net.link.tap.user_open sysctl no longer prevents user opening of already created /dev/tapNN devices. Access is still controlled by node permissions, just like tun devices. The net.link.tap.user_open sysctl is now used only to allow users to perform devfs cloning of tap devices, and the subsequent open may not succeed if the user is not in the appropriate group. This sysctl may be deprecated/removed completely in the future. 20191009: mips, powerpc, and sparc64 are no longer built as part of universe / tinderbox unless MAKE_OBSOLETE_GCC is defined. If not defined, mips, powerpc, and sparc64 builds will look for the xtoolchain binaries and if installed use them for universe builds. As llvm 9.0 becomes vetted for these architectures, they will be removed from the list. 20191009: Clang, llvm, lld, lldb, compiler-rt, libc++, libunwind and openmp have been upgraded to 9.0.0. Please see the 20141231 entry below for information about prerequisites and upgrading, if you are not already using clang 3.5.0 or higher. 20191003: The hpt27xx, hptmv, hptnr, and hptrr drivers have been removed from GENERIC. They are available as modules and can be loaded by adding to /boot/loader.conf hpt27xx_load="YES", hptmv_load="YES", hptnr_load="YES", or hptrr_load="YES", respectively. 20190913: ntpd no longer by default locks its pages in memory, allowing them to be paged out by the kernel. Use rlimit memlock to restore historic BSD behaviour. For example, add "rlimit memlock 32" to ntp.conf to lock up to 32 MB of ntpd address space in memory. 20190823: Several of ping6's options have been renamed for better consistency with ping. If you use any of -ARWXaghmrtwx, you must update your scripts. See ping6(8) for details. 20190727: The vfs.fusefs.sync_unmount and vfs.fusefs.init_backgrounded sysctls and the "-o sync_unmount" and "-o init_backgrounded" mount options have been removed from mount_fusefs(8). You can safely remove them from your scripts, because they had no effect. The vfs.fusefs.fix_broken_io, vfs.fusefs.sync_resize, vfs.fusefs.refresh_size, vfs.fusefs.mmap_enable, vfs.fusefs.reclaim_revoked, and vfs.fusefs.data_cache_invalidate sysctls have been removed. If you felt the need to set any of them to a non-default value, please tell asomers@FreeBSD.org why. 20190713: Default permissions on the /var/account/acct file (and copies of it rotated by periodic daily scripts) are changed from 0644 to 0640 because the file contains sensitive information that should not be world-readable. If the /var/account directory must be created by rc.d/accounting, the mode used is now 0750. Admins who use the accounting feature are encouraged to change the mode of an existing /var/account directory to 0750 or 0700. 20190620: Entropy collection and the /dev/random device are no longer optional components. The "device random" option has been removed. Implementations of distilling algorithms can still be made loadable with "options RANDOM_LOADABLE" (e.g., random_fortuna.ko). 20190612: Clang, llvm, lld, lldb, compiler-rt, libc++, libunwind and openmp have been upgraded to 8.0.1. Please see the 20141231 entry below for information about prerequisites and upgrading, if you are not already using clang 3.5.0 or higher. 20190608: A fix was applied to i386 kernel modules to avoid panics with dpcpu or vnet. Users need to recompile i386 kernel modules having pcpu or vnet sections or they will refuse to load. 20190513: User-wired pages now have their own counter, vm.stats.vm.v_user_wire_count. The vm.max_wired sysctl was renamed to vm.max_user_wired and changed from an unsigned int to an unsigned long. bhyve VMs wired with the -S are now subject to the user wiring limit; the vm.max_user_wired sysctl may need to be tuned to avoid running into the limit. 20190507: The IPSEC option has been removed from GENERIC. Users requiring ipsec(4) must now load the ipsec(4) kernel module. 20190507: The tap(4) driver has been folded into tun(4), and the module has been renamed to tuntap. You should update any kld_list="if_tap" or kld_list="if_tun" entries in /etc/rc.conf, if_tap_load="YES" or if_tun_load="YES" entries in /boot/loader.conf to load the if_tuntap module instead, and "device tap" or "device tun" entries in kernel config files to select the tuntap device instead. 20190418: The following knobs have been added related to tradeoffs between safe use of the random device and availability in the absence of entropy: kern.random.initial_seeding.bypass_before_seeding: tunable; set non-zero to bypass the random device prior to seeding, or zero to block random requests until the random device is initially seeded. For now, set to 1 (unsafe) by default to restore pre-r346250 boot availability properties. kern.random.initial_seeding.read_random_bypassed_before_seeding: read-only diagnostic sysctl that is set when bypass is enabled and read_random(9) is bypassed, to enable programmatic handling of this initial condition, if desired. kern.random.initial_seeding.arc4random_bypassed_before_seeding: Similar to the above, but for for arc4random(9) initial seeding. kern.random.initial_seeding.disable_bypass_warnings: tunable; set non-zero to disable warnings in dmesg when the same conditions are met as for the diagnostic sysctls above. Defaults to zero, i.e., produce warnings in dmesg when the conditions are met. 20190416: The loadable random module KPI has changed; the random_infra_init() routine now requires a 3rd function pointer for a bool (*)(void) method that returns true if the random device is seeded (and therefore unblocked). 20190404: r345895 reverts r320698. This implies that an nfsuserd(8) daemon built from head sources between r320757 (July 6, 2017) and r338192 (Aug. 22, 2018) will not work unless the "-use-udpsock" is added to the command line. nfsuserd daemons built from head sources that are post-r338192 are not affected and should continue to work. 20190320: The fuse(4) module has been renamed to fusefs(4) for consistency with other filesystems. You should update any kld_load="fuse" entries in /etc/rc.conf, fuse_load="YES" entries in /boot/loader.conf, and "options FUSE" entries in kernel config files. 20190304: Clang, llvm, lld, lldb, compiler-rt and libc++ have been upgraded to 8.0.0. Please see the 20141231 entry below for information about prerequisites and upgrading, if you are not already using clang 3.5.0 or higher. 20190226: geom_uzip(4) depends on the new module xz. If geom_uzip is statically compiled into your custom kernel, add 'device xz' statement to the kernel config. 20190219: drm and drm2 have been removed from the tree. Please see https://wiki.freebsd.org/Graphics for the latest information on migrating to the drm ports. 20190131: Iflib is no longer unconditionally compiled into the kernel. Drivers using iflib and statically compiled into the kernel, now require the 'device iflib' config option. For the same drivers loaded as modules on kernels not having 'device iflib', the iflib.ko module is loaded automatically. 20190125: The IEEE80211_AMPDU_AGE and AH_SUPPORT_AR5416 kernel configuration options no longer exist since r343219 and r343427 respectively; nothing uses them, so they should be just removed from custom kernel config files. 20181230: r342635 changes the way efibootmgr(8) works by requiring users to add the -b (bootnum) parameter for commands where the bootnum was previously specified with each option. For example 'efibootmgr -B 0001' is now 'efibootmgr -B -b 0001'. 20181220: r342286 modifies the NFSv4 server so that it obeys vfs.nfsd.nfs_privport in the same as it is applied to NFSv2 and 3. This implies that NFSv4 servers that have vfs.nfsd.nfs_privport set will only allow mounts from clients using a reserved port#. Since both the FreeBSD and Linux NFSv4 clients use reserved port#s by default, this should not affect most NFSv4 mounts. 20181219: The XLP config has been removed. We can't support 64-bit atomics in this kernel because it is running in 32-bit mode. XLP users must transition to running a 64-bit kernel (XLP64 or XLPN32). The mips GXEMUL support has been removed from FreeBSD. MALTA* + qemu is the preferred emulator today and we don't need two different ones. The old sibyte / swarm / Broadcom BCM1250 support has been removed from the mips port. 20181211: Clang, llvm, lld, lldb, compiler-rt and libc++ have been upgraded to 7.0.1. Please see the 20141231 entry below for information about prerequisites and upgrading, if you are not already using clang 3.5.0 or higher. 20181211: Remove the timed and netdate programs from the base tree. Setting the time with these daemons has been obsolete for over a decade. 20181126: On amd64, arm64 and armv7 (architectures that install LLVM's ld.lld linker as /usr/bin/ld) GNU ld is no longer installed as ld.bfd, as it produces broken binaries when ifuncs are in use. Users needing GNU ld should install the binutils port or package. 20181123: The BSD crtbegin and crtend code has been enabled by default. It has had extensive testing on amd64, arm64, and i386. It can be disabled by building a world with -DWITHOUT_BSD_CRTBEGIN. 20181115: The set of CTM commands (ctm, ctm_smail, ctm_rmail, ctm_dequeue) has been converted to a port (misc/ctm) and will be removed from FreeBSD-13. It is available as a package (ctm) for all supported FreeBSD versions. 20181110: The default newsyslog.conf(5) file has been changed to only include files in /etc/newsyslog.conf.d/ and /usr/local/etc/newsyslog.conf.d/ if the filenames end in '.conf' and do not begin with a '.'. You should check the configuration files in these two directories match this naming convention. You can verify which configuration files are being included using the command: $ newsyslog -Nrv 20181015: Ports for the DRM modules have been simplified. Now, amd64 users should just install the drm-kmod port. All others should install drm-legacy-kmod. Graphics hardware that's newer than about 2010 usually works with drm-kmod. For hardware older than 2013, however, some users will need to use drm-legacy-kmod if drm-kmod doesn't work for them. Hardware older than 2008 usually only works in drm-legacy-kmod. The graphics team can only commit to hardware made since 2013 due to the complexity of the market and difficulty to test all the older cards effectively. If you have hardware supported by drm-kmod, you are strongly encouraged to use that as you will get better support. Other than KPI chasing, drm-legacy-kmod will not be updated. As outlined elsewhere, the drm and drm2 modules will be eliminated from the src base soon (with a limited exception for arm). Please update to the package asap and report any issues to x11@freebsd.org. Generally, anybody using the drm*-kmod packages should add WITHOUT_DRM_MODULE=t and WITHOUT_DRM2_MODULE=t to avoid nasty cross-threading surprises, especially with automatic driver loading from X11 startup. These will become the defaults in 13-current shortly. 20181012: The ixlv(4) driver has been renamed to iavf(4). As a consequence, custom kernel and module loading configuration files must be updated accordingly. Moreover, interfaces previous presented as ixlvN to the system are now exposed as iavfN and network configuration files must be adjusted as necessary. 20181009: OpenSSL has been updated to version 1.1.1. This update included additional various API changes throughout the base system. It is important to rebuild third-party software after upgrading. The value of __FreeBSD_version has been bumped accordingly. 20181006: The legacy DRM modules and drivers have now been added to the loader's module blacklist, in favor of loading them with kld_list in rc.conf(5). The module blacklist may be overridden with the loader.conf(5) 'module_blacklist' variable, but loading them via rc.conf(5) is strongly encouraged. 20181002: The cam(4) based nda(4) driver will be used over nvd(4) by default on powerpc64. You may set 'options NVME_USE_NVD=1' in your kernel conf or loader tunable 'hw.nvme.use_nvd=1' if you wish to use the existing driver. Make sure to edit /boot/etc/kboot.conf and fstab to use the nda device name. 20180913: Reproducible build mode is now on by default, in preparation for FreeBSD 12.0. This eliminates build metadata such as the user, host, and time from the kernel (and uname), unless the working tree corresponds to a modified checkout from a version control system. The previous behavior can be obtained by setting the /etc/src.conf knob WITHOUT_REPRODUCIBLE_BUILD. 20180826: The Yarrow CSPRNG has been removed from the kernel as it has not been supported by its designers since at least 2003. Fortuna has been the default since FreeBSD-11. 20180822: devctl freeze/thaw have gone into the tree, the rc scripts have been updated to use them and devmatch has been changed. You should update kernel, userland and rc scripts all at the same time. 20180818: The default interpreter has been switched from 4th to Lua. LOADER_DEFAULT_INTERP, documented in build(7), will override the default interpreter. If you have custom FORTH code you will need to set LOADER_DEFAULT_INTERP=4th (valid values are 4th, lua or simp) in src.conf for the build. This will create default hard links between loader and loader_4th instead of loader and loader_lua, the new default. If you are using UEFI it will create the proper hard link to loader.efi. bhyve uses userboot.so. It remains 4th-only until some issues are solved regarding coexisting with multiple versions of FreeBSD are resolved. 20180815: ls(1) now respects the COLORTERM environment variable used in other systems and software to indicate that a colored terminal is both supported and desired. If ls(1) is suddenly emitting colors, they may be disabled again by either removing the unwanted COLORTERM from your environment, or using `ls --color=never`. The ls(1) specific CLICOLOR may not be observed in a future release. 20180808: The default pager for most commands has been changed to "less". To restore the old behavior, set PAGER="more" and MANPAGER="more -s" in your environment. 20180731: The jedec_ts(4) driver has been removed. A superset of its functionality is available in the jedec_dimm(4) driver, and the manpage for that driver includes migration instructions. If you have "device jedec_ts" in your kernel configuration file, it must be removed. 20180730: amd64/GENERIC now has EFI runtime services, EFIRT, enabled by default. This should have no effect if the kernel is booted via BIOS/legacy boot. EFIRT may be disabled via a loader tunable, efi.rt.disabled, if a system has a buggy firmware that prevents a successful boot due to use of runtime services. 20180727: Atmel AT91RM9200 and AT91SAM9, Cavium CNS 11xx and XScale support has been removed from the tree. These ports were obsolete and/or known to be broken for many years. 20180723: loader.efi has been augmented to participate more fully in the UEFI boot manager protocol. loader.efi will now look at the BootXXXX environment variable to determine if a specific kernel or root partition was specified. XXXX is derived from BootCurrent. efibootmgr(8) manages these standard UEFI variables. 20180720: zfsloader's functionality has now been folded into loader. zfsloader is no longer necessary once you've updated your boot blocks. For a transition period, we will install a hardlink for zfsloader to loader to allow a smooth transition until the boot blocks can be updated (hard link because old zfs boot blocks don't understand symlinks). 20180719: ARM64 now have efifb support, if you want to have serial console on your arm64 board when an screen is connected and the bootloader setup a frame buffer for us to use, just add : boot_serial=YES boot_multicons=YES in /boot/loader.conf For Raspberry Pi 3 (RPI) users, this is needed even if you don't have an screen connected as the firmware will setup a frame buffer are that u-boot will expose as an EFI frame buffer. 20180719: New uid:gid added, ntpd:ntpd (123:123). Be sure to run mergemaster or take steps to update /etc/passwd before doing installworld on existing systems. Do not skip the "mergemaster -Fp" step before installworld, as described in the update procedures near the bottom of this document. Also, rc.d/ntpd now starts ntpd(8) as user ntpd if the new mac_ntpd(4) policy is available, unless ntpd_flags or the ntp config file contain options that change file/dir locations. When such options (e.g., "statsdir" or "crypto") are used, ntpd can still be run as non-root by setting ntpd_user=ntpd in rc.conf, after taking steps to ensure that all required files/dirs are accessible by the ntpd user. 20180717: Big endian arm support has been removed. 20180711: The static environment setup in kernel configs is no longer mutually exclusive with the loader(8) environment by default. In order to restore the previous default behavior of disabling the loader(8) environment if a static environment is present, you must specify loader_env.disabled=1 in the static environment. 20180705: The ABI of syscalls used by management tools like sockstat and netstat has been broken to allow 32-bit binaries to work on 64-bit kernels without modification. These programs will need to match the kernel in order to function. External programs may require minor modifications to accommodate a change of type in structures from pointers to 64-bit virtual addresses. 20180702: On i386 and amd64 atomics are now inlined. Out of tree modules using atomics will need to be rebuilt. 20180701: The '%I' format in the kern.corefile sysctl limits the number of core files that a process can generate to the number stored in the debug.ncores sysctl. The '%I' format is replaced by the single digit index. Previously, if all indexes were taken the kernel would overwrite only a core file with the highest index in a filename. Currently the system will create a new core file if there is a free index or if all slots are taken it will overwrite the oldest one. 20180630: Clang, llvm, lld, lldb, compiler-rt and libc++ have been upgraded to 6.0.1. Please see the 20141231 entry below for information about prerequisites and upgrading, if you are not already using clang 3.5.0 or higher. 20180628: r335753 introduced a new quoting method. However, etc/devd/devmatch.conf needed to be changed to work with it. This change was made with r335763 and requires a mergemaster / etcupdate / etc to update the installed file. 20180612: r334930 changed the interface between the NFS modules, so they all need to be rebuilt. r335018 did a __FreeBSD_version bump for this. 20180530: As of r334391 lld is the default amd64 system linker; it is installed as /usr/bin/ld. Kernel build workarounds (see 20180510 entry) are no longer necessary. 20180530: The kernel / userland interface for devinfo changed, so you'll need a new kernel and userland as a pair for it to work (rebuilding lib/libdevinfo is all that's required). devinfo and devmatch will not work, but everything else will when there's a mismatch. 20180523: The on-disk format for hwpmc callchain records has changed to include threadid corresponding to a given record. This changes the field offsets and thus requires that libpmcstat be rebuilt before using a kernel later than r334108. 20180517: The vxge(4) driver has been removed. This driver was introduced into HEAD one week before the Exar left the Ethernet market and is not known to be used. If you have device vxge in your kernel config file it must be removed. 20180510: The amd64 kernel now requires a ld that supports ifunc to produce a working kernel, either lld or a newer binutils. lld is built by default on amd64, and the 'buildkernel' target uses it automatically. However, it is not the default linker, so building the kernel the traditional way requires LD=ld.lld on the command line (or LD=/usr/local/bin/ld for binutils port/package). lld will soon be default, and this requirement will go away. NOTE: As of r334391 lld is the default system linker on amd64, and no workaround is necessary. 20180508: The nxge(4) driver has been removed. This driver was for PCI-X 10g cards made by s2io/Neterion. The company was acquired by Exar and no longer sells or supports Ethernet products. If you have device nxge in your kernel config file it must be removed. 20180504: The tz database (tzdb) has been updated to 2018e. This version more correctly models time stamps in time zones with negative DST such as Europe/Dublin (from 1971 on), Europe/Prague (1946/7), and Africa/Windhoek (1994/2017). This does not affect the UT offsets, only time zone abbreviations and the tm_isdst flag. 20180502: The ixgb(4) driver has been removed. This driver was for an early and uncommon legacy PCI 10GbE for a single ASIC, Intel 82597EX. Intel quickly shifted to the long lived ixgbe family. If you have device ixgb in your kernel config file it must be removed. 20180501: The lmc(4) driver has been removed. This was a WAN interface card that was already reportedly rare in 2003, and had an ambiguous license. If you have device lmc in your kernel config file it must be removed. 20180413: Support for Arcnet networks has been removed. If you have device arcnet or device cm in your kernel config file they must be removed. 20180411: Support for FDDI networks has been removed. If you have device fddi or device fpa in your kernel config file they must be removed. 20180406: In addition to supporting RFC 3164 formatted messages, the syslogd(8) service is now capable of parsing RFC 5424 formatted log messages. The main benefit of using RFC 5424 is that clients may now send log messages with timestamps containing year numbers, microseconds and time zone offsets. Similarly, the syslog(3) C library function has been altered to send RFC 5424 formatted messages to the local system logging daemon. On systems using syslogd(8), this change should have no negative impact, as long as syslogd(8) and the C library are updated at the same time. On systems using a different system logging daemon, it may be necessary to make configuration adjustments, depending on the software used. When using syslog-ng, add the 'syslog-protocol' flag to local input sources to enable parsing of RFC 5424 formatted messages: source src { unix-dgram("/var/run/log" flags(syslog-protocol)); } When using rsyslog, disable the 'SysSock.UseSpecialParser' option of the 'imuxsock' module to let messages be processed by the regular RFC 3164/5424 parsing pipeline: module(load="imuxsock" SysSock.UseSpecialParser="off") Do note that these changes only affect communication between local applications and syslogd(8). The format that syslogd(8) uses to store messages on disk or forward messages to other systems remains unchanged. syslogd(8) still uses RFC 3164 for these purposes. Options to customize this behaviour will be added in the future. Utilities that process log files stored in /var/log are thus expected to continue to function as before. __FreeBSD_version has been incremented to 1200061 to denote this change. 20180328: Support for token ring networks has been removed. If you have "device token" in your kernel config you should remove it. No device drivers supported token ring. 20180323: makefs was modified to be able to tag ISO9660 El Torito boot catalog entries as EFI instead of overloading the i386 tag as done previously. The amd64 mkisoimages.sh script used to build amd64 ISO images for release was updated to use this. This may mean that makefs must be updated before "make cdrom" can be run in the release directory. This should be as simple as: $ cd $SRCDIR/usr.sbin/makefs $ make depend all install 20180212: FreeBSD boot loader enhanced with Lua scripting. It's purely opt-in for now by building WITH_LOADER_LUA and WITHOUT_FORTH in /etc/src.conf. Co-existence for the transition period will come shortly. Booting is a complex environment and test coverage for Lua-enabled loaders has been thin, so it would be prudent to assume it might not work and make provisions for backup boot methods. 20180211: devmatch functionality has been turned on in devd. It will automatically load drivers for unattached devices. This may cause unexpected drivers to be loaded. Please report any problems to current@ and imp@freebsd.org. 20180114: Clang, llvm, lld, lldb, compiler-rt and libc++ have been upgraded to 6.0.0. Please see the 20141231 entry below for information about prerequisites and upgrading, if you are not already using clang 3.5.0 or higher. 20180110: LLVM's lld linker is now used as the FreeBSD/amd64 bootstrap linker. This means it is used to link the kernel and userland libraries and executables, but is not yet installed as /usr/bin/ld by default. To revert to ld.bfd as the bootstrap linker, in /etc/src.conf set WITHOUT_LLD_BOOTSTRAP=yes 20180110: On i386, pmtimer has been removed. Its functionality has been folded into apm. It was a no-op on ACPI in current for a while now (but was still needed on i386 in FreeBSD 11 and earlier). Users may need to remove it from kernel config files. 20180104: The use of RSS hash from the network card aka flowid has been disabled by default for lagg(4) as it's currently incompatible with the lacp and loadbalance protocols. This can be re-enabled by setting the following in loader.conf: net.link.lagg.default_use_flowid="1" 20180102: The SW_WATCHDOG option is no longer necessary to enable the hardclock-based software watchdog if no hardware watchdog is configured. As before, SW_WATCHDOG will cause the software watchdog to be enabled even if a hardware watchdog is configured. 20171215: r326887 fixes the issue described in the 20171214 UPDATING entry. r326888 flips the switch back to building GELI support always. 20171214: r362593 broke ZFS + GELI support for reasons unknown. However, it also broke ZFS support generally, so GELI has been turned off by default as the lesser evil in r326857. If you boot off ZFS and/or GELI, it might not be a good time to update. 20171125: PowerPC users must update loader(8) by rebuilding world before installing a new kernel, as the protocol connecting them has changed. Without the update, loader metadata will not be passed successfully to the kernel and users will have to enter their root partition at the kernel mountroot prompt to continue booting. Newer versions of loader can boot old kernels without issue. 20171110: The LOADER_FIREWIRE_SUPPORT build variable as been renamed to WITH/OUT_LOADER_FIREWIRE. LOADER_{NO_,}GELI_SUPPORT has been renamed to WITH/OUT_LOADER_GELI. 20171106: The naive and non-compliant support of posix_fallocate(2) in ZFS has been removed as of r325320. The system call now returns EINVAL when used on a ZFS file. Although the new behavior complies with the standard, some consumers are not prepared to cope with it. One known victim is lld prior to r325420. 20171102: Building in a FreeBSD src checkout will automatically create object directories now rather than store files in the current directory if 'make obj' was not ran. Calling 'make obj' is no longer necessary. This feature can be disabled by setting WITHOUT_AUTO_OBJ=yes in /etc/src-env.conf (not /etc/src.conf), or passing the option in the environment. 20171101: The default MAKEOBJDIR has changed from /usr/obj/ for native builds, and /usr/obj// for cross-builds, to a unified /usr/obj//. This behavior can be changed to the old format by setting WITHOUT_UNIFIED_OBJDIR=yes in /etc/src-env.conf, the environment, or with -DWITHOUT_UNIFIED_OBJDIR when building. The UNIFIED_OBJDIR option is a transitional feature that will be removed for 12.0 release; please migrate to the new format for any tools by looking up the OBJDIR used by 'make -V .OBJDIR' means rather than hardcoding paths. 20171028: The native-xtools target no longer installs the files by default to the OBJDIR. Use the native-xtools-install target with a DESTDIR to install to ${DESTDIR}/${NXTP} where NXTP defaults to /nxb-bin. 20171021: As part of the boot loader infrastructure cleanup, LOADER_*_SUPPORT options are changing from controlling the build if defined / undefined to controlling the build with explicit 'yes' or 'no' values. They will shift to WITH/WITHOUT options to match other options in the system. 20171010: libstand has turned into a private library for sys/boot use only. It is no longer supported as a public interface outside of sys/boot. 20171005: The arm port has split armv6 into armv6 and armv7. armv7 is now a valid TARGET_ARCH/MACHINE_ARCH setting. If you have an armv7 system and are running a kernel from before r324363, you will need to add MACHINE_ARCH=armv7 to 'make buildworld' to do a native build. 20171003: When building multiple kernels using KERNCONF, non-existent KERNCONF files will produce an error and buildkernel will fail. Previously missing KERNCONF files silently failed giving no indication as to why, only to subsequently discover during installkernel that the desired kernel was never built in the first place. 20170912: The default serial number format for CTL LUNs has changed. This will affect users who use /dev/diskid/* device nodes, or whose FibreChannel or iSCSI clients care about their LUNs' serial numbers. Users who require serial number stability should hardcode serial numbers in /etc/ctl.conf . 20170912: For 32-bit arm compiled for hard-float support, soft-floating point binaries now always get their shared libraries from LD_SOFT_LIBRARY_PATH (in the past, this was only used if /usr/libsoft also existed). Only users with a hard-float ld.so, but soft-float everything else should be affected. 20170826: The geli password typed at boot is now hidden. To restore the previous behavior, see geli(8) for configuration options. 20170825: Move PMTUD blackhole counters to TCPSTATS and remove them from bare sysctl values. Minor nit, but requires a rebuild of both world/kernel to complete. 20170814: "make check" behavior (made in ^/head@r295380) has been changed to execute from a limited sandbox, as opposed to executing from ${TESTSDIR}. Behavioral changes: - The "beforecheck" and "aftercheck" targets are now specified. - ${CHECKDIR} (added in commit noted above) has been removed. - Legacy behavior can be enabled by setting WITHOUT_MAKE_CHECK_USE_SANDBOX in src.conf(5) or the environment. If the limited sandbox mode is enabled, "make check" will execute "make distribution", then install, execute the tests, and clean up the sandbox if successful. The "make distribution" and "make install" targets are typically run as root to set appropriate permissions and ownership at installation time. The end-user should set "WITH_INSTALL_AS_USER" in src.conf(5) or the environment if executing "make check" with limited sandbox mode using an unprivileged user. 20170808: Since the switch to GPT disk labels, fsck for UFS/FFS has been unable to automatically find alternate superblocks. As of r322297, the information needed to find alternate superblocks has been moved to the end of the area reserved for the boot block. Filesystems created with a newfs of this vintage or later will create the recovery information. If you have a filesystem created prior to this change and wish to have a recovery block created for your filesystem, you can do so by running fsck in foreground mode (i.e., do not use the -p or -y options). As it starts, fsck will ask ``SAVE DATA TO FIND ALTERNATE SUPERBLOCKS'' to which you should answer yes. 20170728: As of r321665, an NFSv4 server configuration that services Kerberos mounts or clients that do not support the uid/gid in owner/owner_group string capability, must explicitly enable the nfsuserd daemon by adding nfsuserd_enable="YES" to the machine's /etc/rc.conf file. 20170722: Clang, llvm, lldb, compiler-rt and libc++ have been upgraded to 5.0.0. Please see the 20141231 entry below for information about prerequisites and upgrading, if you are not already using clang 3.5.0 or higher. 20170701: WITHOUT_RCMDS is now the default. Set WITH_RCMDS if you need the r-commands (rlogin, rsh, etc.) to be built with the base system. 20170625: The FreeBSD/powerpc platform now uses a 64-bit type for time_t. This is a very major ABI incompatible change, so users of FreeBSD/powerpc must be careful when performing source upgrades. It is best to run 'make installworld' from an alternate root system, either a live CD/memory stick, or a temporary root partition. Additionally, all ports must be recompiled. powerpc64 is largely unaffected, except in the case of 32-bit compatibility. All 32-bit binaries will be affected. 20170623: Forward compatibility for the "ino64" project have been committed. This will allow most new binaries to run on older kernels in a limited fashion. This prevents many of the common foot-shooting actions in the upgrade as well as the limited ability to roll back the kernel across the ino64 upgrade. Complicated use cases may not work properly, though enough simpler ones work to allow recovery in most situations. 20170620: Switch back to the BSDL dtc (Device Tree Compiler). Set WITH_GPL_DTC if you require the GPL compiler. 20170618: The internal ABI used for communication between the NFS kernel modules was changed by r320085, so __FreeBSD_version was bumped to ensure all the NFS related modules are updated together. 20170617: The ABI of struct event was changed by extending the data member to 64bit and adding ext fields. For upgrade, same precautions as for the entry 20170523 "ino64" must be followed. 20170531: The GNU roff toolchain has been removed from base. To render manpages which are not supported by mandoc(1), man(1) can fallback on GNU roff from ports (and recommends to install it). To render roff(7) documents, consider using GNU roff from ports or the heirloom doctools roff toolchain from ports via pkg install groff or via pkg install heirloom-doctools. 20170524: The ath(4) and ath_hal(4) modules now build piecemeal to allow for smaller runtime footprint builds. This is useful for embedded systems which only require one chipset support. If you load it as a module, make sure this is in /boot/loader.conf: if_ath_load="YES" This will load the HAL, all chip/RF backends and if_ath_pci. If you have if_ath_pci in /boot/loader.conf, ensure it is after if_ath or it will not load any HAL chipset support. If you want to selectively load things (eg on ye cheape ARM/MIPS platforms where RAM is at a premium) you should: * load ath_hal * load the chip modules in question * load ath_rate, ath_dfs * load ath_main * load if_ath_pci and/or if_ath_ahb depending upon your particular bus bind type - this is where probe/attach is done. For further comments/feedback, poke adrian@ . 20170523: The "ino64" 64-bit inode project has been committed, which extends a number of types to 64 bits. Upgrading in place requires care and adherence to the documented upgrade procedure. If using a custom kernel configuration ensure that the COMPAT_FREEBSD11 option is included (as during the upgrade the system will be running the ino64 kernel with the existing world). For the safest in-place upgrade begin by removing previous build artifacts via "rm -rf /usr/obj/*". Then, carefully follow the full procedure documented below under the heading "To rebuild everything and install it on the current system." Specifically, a reboot is required after installing the new kernel before installing world. While an installworld normally works by accident from multiuser after rebooting the proper kernel, there are many cases where this will fail across this upgrade and installworld from single user is required. 20170424: The NATM framework including the en(4), fatm(4), hatm(4), and patm(4) devices has been removed. Consumers should plan a migration before the end-of-life date for FreeBSD 11. 20170420: GNU diff has been replaced by a BSD licensed diff. Some features of GNU diff has not been implemented, if those are needed a newer version of GNU diff is available via the diffutils package under the gdiff name. 20170413: As of r316810 for ipfilter, keep frags is no longer assumed when keep state is specified in a rule. r316810 aligns ipfilter with documentation in man pages separating keep frags from keep state. This allows keep state to be specified without forcing keep frags and allows keep frags to be specified independently of keep state. To maintain previous behaviour, also specify keep frags with keep state (as documented in ipf.conf.5). 20170407: arm64 builds now use the base system LLD 4.0.0 linker by default, instead of requiring that the aarch64-binutils port or package be installed. To continue using aarch64-binutils, set CROSS_BINUTILS_PREFIX=/usr/local/aarch64-freebsd/bin . 20170405: The UDP optimization in entry 20160818 that added the sysctl net.inet.udp.require_l2_bcast has been reverted. L2 broadcast packets will no longer be treated as L3 broadcast packets. 20170331: Binds and sends to the loopback addresses, IPv6 and IPv4, will now use any explicitly assigned loopback address available in the jail instead of using the first assigned address of the jail. 20170329: The ctl.ko module no longer implements the iSCSI target frontend: cfiscsi.ko does instead. If building cfiscsi.ko as a kernel module, the module can be loaded via one of the following methods: - `cfiscsi_load="YES"` in loader.conf(5). - Add `cfiscsi` to `$kld_list` in rc.conf(5). - ctladm(8)/ctld(8), when compiled with iSCSI support (`WITH_ISCSI=yes` in src.conf(5)) Please see cfiscsi(4) for more details. 20170316: The mmcsd.ko module now additionally depends on geom_flashmap.ko. Also, mmc.ko and mmcsd.ko need to be a matching pair built from the same source (previously, the dependency of mmcsd.ko on mmc.ko was missing, but mmcsd.ko now will refuse to load if it is incompatible with mmc.ko). 20170315: The syntax of ipfw(8) named states was changed to avoid ambiguity. If you have used named states in the firewall rules, you need to modify them after installworld and before rebooting. Now named states must be prefixed with colon. 20170311: The old drm (sys/dev/drm/) drivers for i915 and radeon have been removed as the userland we provide cannot use them. The KMS version (sys/dev/drm2) supports the same hardware. 20170302: Clang, llvm, lldb, compiler-rt and libc++ have been upgraded to 4.0.0. Please see the 20141231 entry below for information about prerequisites and upgrading, if you are not already using clang 3.5.0 or higher. 20170221: The code that provides support for ZFS .zfs/ directory functionality has been reimplemented. It's not possible now to create a snapshot by mkdir under .zfs/snapshot/. That should be the only user visible change. 20170216: EISA bus support has been removed. The WITH_EISA option is no longer valid. 20170215: MCA bus support has been removed. 20170127: The WITH_LLD_AS_LD / WITHOUT_LLD_AS_LD build knobs have been renamed WITH_LLD_IS_LD / WITHOUT_LLD_IS_LD, for consistency with CLANG_IS_CC. 20170112: The EM_MULTIQUEUE kernel configuration option is deprecated now that the em(4) driver conforms to iflib specifications. 20170109: The igb(4), em(4) and lem(4) ethernet drivers are now implemented via IFLIB. If you have a custom kernel configuration that excludes em(4) but you use igb(4), you need to re-add em(4) to your custom configuration. 20161217: Clang, llvm, lldb, compiler-rt and libc++ have been upgraded to 3.9.1. Please see the 20141231 entry below for information about prerequisites and upgrading, if you are not already using clang 3.5.0 or higher. 20161124: Clang, llvm, lldb, compiler-rt and libc++ have been upgraded to 3.9.0. Please see the 20141231 entry below for information about prerequisites and upgrading, if you are not already using clang 3.5.0 or higher. 20161119: The layout of the pmap structure has changed for powerpc to put the pmap statistics at the front for all CPU variations. libkvm(3) and all tools that link against it need to be recompiled. 20161030: isl(4) and cyapa(4) drivers now require a new driver, chromebook_platform(4), to work properly on Chromebook-class hardware. On other types of hardware the drivers may need to be configured using device hints. Please see the corresponding manual pages for details. 20161017: The urtwn(4) driver was merged into rtwn(4) and now consists of rtwn(4) main module + rtwn_usb(4) and rtwn_pci(4) bus-specific parts. Also, firmware for RTL8188CE was renamed due to possible name conflict (rtwnrtl8192cU(B) -> rtwnrtl8192cE(B)) 20161015: GNU rcs has been removed from base. It is available as packages: - rcs: Latest GPLv3 GNU rcs version. - rcs57: Copy of the latest version of GNU rcs (GPLv2) before it was removed from base. 20161008: Use of the cc_cdg, cc_chd, cc_hd, or cc_vegas congestion control modules now requires that the kernel configuration contain the TCP_HHOOK option. (This option is included in the GENERIC kernel.) 20161003: The WITHOUT_ELFCOPY_AS_OBJCOPY src.conf(5) knob has been retired. ELF Tool Chain's elfcopy is always installed as /usr/bin/objcopy. 20160924: Relocatable object files with the extension of .So have been renamed to use an extension of .pico instead. The purpose of this change is to avoid a name clash with shared libraries on case-insensitive file systems. On those file systems, foo.So is the same file as foo.so. 20160918: GNU rcs has been turned off by default. It can (temporarily) be built again by adding WITH_RCS knob in src.conf. Otherwise, GNU rcs is available from packages: - rcs: Latest GPLv3 GNU rcs version. - rcs57: Copy of the latest version of GNU rcs (GPLv2) from base. 20160918: The backup_uses_rcs functionality has been removed from rc.subr. 20160908: The queue(3) debugging macro, QUEUE_MACRO_DEBUG, has been split into two separate components, QUEUE_MACRO_DEBUG_TRACE and QUEUE_MACRO_DEBUG_TRASH. Define both for the original QUEUE_MACRO_DEBUG behavior. 20160824: r304787 changed some ioctl interfaces between the iSCSI userspace programs and the kernel. ctladm, ctld, iscsictl, and iscsid must be rebuilt to work with new kernels. __FreeBSD_version has been bumped to 1200005. 20160818: The UDP receive code has been updated to only treat incoming UDP packets that were addressed to an L2 broadcast address as L3 broadcast packets. It is not expected that this will affect any standards-conforming UDP application. The new behaviour can be disabled by setting the sysctl net.inet.udp.require_l2_bcast to 0. 20160818: Remove the openbsd_poll system call. __FreeBSD_version has been bumped because of this. 20160708: The stable/11 branch has been created from head@r302406. 20160622: The libc stub for the pipe(2) system call has been replaced with a wrapper that calls the pipe2(2) system call and the pipe(2) system call is now only implemented by the kernels that include "options COMPAT_FREEBSD10" in their config file (this is the default). Users should ensure that this option is enabled in their kernel or upgrade userspace to r302092 before upgrading their kernel. 20160527: CAM will now strip leading spaces from SCSI disks' serial numbers. This will affect users who create UFS filesystems on SCSI disks using those disk's diskid device nodes. For example, if /etc/fstab previously contained a line like "/dev/diskid/DISK-%20%20%20%20%20%20%20ABCDEFG0123456", you should change it to "/dev/diskid/DISK-ABCDEFG0123456". Users of geom transforms like gmirror may also be affected. ZFS users should generally be fine. 20160523: The bitstring(3) API has been updated with new functionality and improved performance. But it is binary-incompatible with the old API. Objects built with the new headers may not be linked against objects built with the old headers. 20160520: The brk and sbrk functions have been removed from libc on arm64. Binutils from ports has been updated to not link to these functions and should be updated to the latest version before installing a new libc. 20160517: The armv6 port now defaults to hard float ABI. Limited support for running both hardfloat and soft float on the same system is available using the libraries installed with -DWITH_LIBSOFT. This has only been tested as an upgrade path for installworld and packages may fail or need manual intervention to run. New packages will be needed. To update an existing self-hosted armv6hf system, you must add TARGET_ARCH=armv6 on the make command line for both the build and the install steps. 20160510: Kernel modules compiled outside of a kernel build now default to installing to /boot/modules instead of /boot/kernel. Many kernel modules built this way (such as those in ports) already overrode KMODDIR explicitly to install into /boot/modules. However, manually building and installing a module from /sys/modules will now install to /boot/modules instead of /boot/kernel. 20160414: The CAM I/O scheduler has been committed to the kernel. There should be no user visible impact. This does enable NCQ Trim on ada SSDs. While the list of known rogues that claim support for this but actually corrupt data is believed to be complete, be on the lookout for data corruption. The known rogue list is believed to be complete: o Crucial MX100, M550 drives with MU01 firmware. o Micron M510 and M550 drives with MU01 firmware. o Micron M500 prior to MU07 firmware o Samsung 830, 840, and 850 all firmwares o FCCT M500 all firmwares Crucial has firmware http://www.crucial.com/usa/en/support-ssd-firmware with working NCQ TRIM. For Micron branded drives, see your sales rep for updated firmware. Black listed drives will work correctly because these drives work correctly so long as no NCQ TRIMs are sent to them. Given this list is the same as found in Linux, it's believed there are no other rogues in the market place. All other models from the above vendors work. To be safe, if you are at all concerned, you can quirk each of your drives to prevent NCQ from being sent by setting: kern.cam.ada.X.quirks="0x2" in loader.conf. If the drive requires the 4k sector quirk, set the quirks entry to 0x3. 20160330: The FAST_DEPEND build option has been removed and its functionality is now the one true way. The old mkdep(1) style of 'make depend' has been removed. See 20160311 for further details. 20160317: Resource range types have grown from unsigned long to uintmax_t. All drivers, and anything using libdevinfo, need to be recompiled. 20160311: WITH_FAST_DEPEND is now enabled by default for in-tree and out-of-tree builds. It no longer runs mkdep(1) during 'make depend', and the 'make depend' stage can safely be skipped now as it is auto ran when building 'make all' and will generate all SRCS and DPSRCS before building anything else. Dependencies are gathered at compile time with -MF flags kept in separate .depend files per object file. Users should run 'make cleandepend' once if using -DNO_CLEAN to clean out older stale .depend files. 20160306: On amd64, clang 3.8.0 can now insert sections of type AMD64_UNWIND into kernel modules. Therefore, if you load any kernel modules at boot time, please install the boot loaders after you install the kernel, but before rebooting, e.g.: make buildworld make buildkernel KERNCONF=YOUR_KERNEL_HERE make installkernel KERNCONF=YOUR_KERNEL_HERE make -C sys/boot install Then follow the usual steps, described in the General Notes section, below. 20160305: Clang, llvm, lldb and compiler-rt have been upgraded to 3.8.0. Please see the 20141231 entry below for information about prerequisites and upgrading, if you are not already using clang 3.5.0 or higher. 20160301: The AIO subsystem is now a standard part of the kernel. The VFS_AIO kernel option and aio.ko kernel module have been removed. Due to stability concerns, asynchronous I/O requests are only permitted on sockets and raw disks by default. To enable asynchronous I/O requests on all file types, set the vfs.aio.enable_unsafe sysctl to a non-zero value. 20160226: The ELF object manipulation tool objcopy is now provided by the ELF Tool Chain project rather than by GNU binutils. It should be a drop-in replacement, with the addition of arm64 support. The (temporary) src.conf knob WITHOUT_ELFCOPY_AS_OBJCOPY knob may be set to obtain the GNU version if necessary. 20160129: Building ZFS pools on top of zvols is prohibited by default. That feature has never worked safely; it's always been prone to deadlocks. Using a zvol as the backing store for a VM guest's virtual disk will still work, even if the guest is using ZFS. Legacy behavior can be restored by setting vfs.zfs.vol.recursive=1. 20160119: The NONE and HPN patches has been removed from OpenSSH. They are still available in the security/openssh-portable port. 20160113: With the addition of ypldap(8), a new _ypldap user is now required during installworld. "mergemaster -p" can be used to add the user prior to installworld, as documented in the handbook. 20151216: The tftp loader (pxeboot) now uses the option root-path directive. As a consequence it no longer looks for a pxeboot.4th file on the tftp server. Instead it uses the regular /boot infrastructure as with the other loaders. 20151211: The code to start recording plug and play data into the modules has been committed. While the old tools will properly build a new kernel, a number of warnings about "unknown metadata record 4" will be produced for an older kldxref. To avoid such warnings, make sure to rebuild the kernel toolchain (or world). Make sure that you have r292078 or later when trying to build 292077 or later before rebuilding. 20151207: Debug data files are now built by default with 'make buildworld' and installed with 'make installworld'. This facilitates debugging but requires more disk space both during the build and for the installed world. Debug files may be disabled by setting WITHOUT_DEBUG_FILES=yes in src.conf(5). 20151130: r291527 changed the internal interface between the nfsd.ko and nfscommon.ko modules. As such, they must both be upgraded to-gether. __FreeBSD_version has been bumped because of this. 20151108: Add support for unicode collation strings leads to a change of order of files listed by ls(1) for example. To get back to the old behaviour, set LC_COLLATE environment variable to "C". Databases administrators will need to reindex their databases given collation results will be different. Due to a bug in install(1) it is recommended to remove the ancient locales before running make installworld. rm -rf /usr/share/locale/* 20151030: The OpenSSL has been upgraded to 1.0.2d. Any binaries requiring libcrypto.so.7 or libssl.so.7 must be recompiled. 20151020: Qlogic 24xx/25xx firmware images were updated from 5.5.0 to 7.3.0. Kernel modules isp_2400_multi and isp_2500_multi were removed and should be replaced with isp_2400 and isp_2500 modules respectively. 20151017: The build previously allowed using 'make -n' to not recurse into sub-directories while showing what commands would be executed, and 'make -n -n' to recursively show commands. Now 'make -n' will recurse and 'make -N' will not. 20151012: If you specify SENDMAIL_MC or SENDMAIL_CF in make.conf, mergemaster and etcupdate will now use this file. A custom sendmail.cf is now updated via this mechanism rather than via installworld. If you had excluded sendmail.cf in mergemaster.rc or etcupdate.conf, you may want to remove the exclusion or change it to "always install". /etc/mail/sendmail.cf is now managed the same way regardless of whether SENDMAIL_MC/SENDMAIL_CF is used. If you are not using SENDMAIL_MC/SENDMAIL_CF there should be no change in behavior. 20151011: Compatibility shims for legacy ATA device names have been removed. It includes ATA_STATIC_ID kernel option, kern.cam.ada.legacy_aliases and kern.geom.raid.legacy_aliases loader tunables, kern.devalias.* environment variables, /dev/ad* and /dev/ar* symbolic links. 20151006: Clang, llvm, lldb, compiler-rt and libc++ have been upgraded to 3.7.0. Please see the 20141231 entry below for information about prerequisites and upgrading, if you are not already using clang 3.5.0 or higher. 20150924: Kernel debug files have been moved to /usr/lib/debug/boot/kernel/, and renamed from .symbols to .debug. This reduces the size requirements on the boot partition or file system and provides consistency with userland debug files. When using the supported kernel installation method the /usr/lib/debug/boot/kernel directory will be renamed (to kernel.old) as is done with /boot/kernel. Developers wishing to maintain the historical behavior of installing debug files in /boot/kernel/ can set KERN_DEBUGDIR="" in src.conf(5). 20150827: The wireless drivers had undergone changes that remove the 'parent interface' from the ifconfig -l output. The rc.d network scripts used to check presence of a parent interface in the list, so old scripts would fail to start wireless networking. Thus, etcupdate(3) or mergemaster(8) run is required after kernel update, to update your rc.d scripts in /etc. 20150827: pf no longer supports 'scrub fragment crop' or 'scrub fragment drop-ovl' These configurations are now automatically interpreted as 'scrub fragment reassemble'. 20150817: Kernel-loadable modules for the random(4) device are back. To use them, the kernel must have device random options RANDOM_LOADABLE kldload(8) can then be used to load random_fortuna.ko or random_yarrow.ko. Please note that due to the indirect function calls that the loadable modules need to provide, the build-in variants will be slightly more efficient. The random(4) kernel option RANDOM_DUMMY has been retired due to unpopularity. It was not all that useful anyway. 20150813: The WITHOUT_ELFTOOLCHAIN_TOOLS src.conf(5) knob has been retired. Control over building the ELF Tool Chain tools is now provided by the WITHOUT_TOOLCHAIN knob. 20150810: The polarity of Pulse Per Second (PPS) capture events with the uart(4) driver has been corrected. Prior to this change the PPS "assert" event corresponded to the trailing edge of a positive PPS pulse and the "clear" event was the leading edge of the next pulse. As the width of a PPS pulse in a typical GPS receiver is on the order of 1 millisecond, most users will not notice any significant difference with this change. Anyone who has compensated for the historical polarity reversal by configuring a negative offset equal to the pulse width will need to remove that workaround. 20150809: The default group assigned to /dev/dri entries has been changed from 'wheel' to 'video' with the id of '44'. If you want to have access to the dri devices please add yourself to the video group with: # pw groupmod video -m $USER 20150806: The menu.rc and loader.rc files will now be replaced during upgrades. Please migrate local changes to menu.rc.local and loader.rc.local instead. 20150805: GNU Binutils versions of addr2line, c++filt, nm, readelf, size, strings and strip have been removed. The src.conf(5) knob WITHOUT_ELFTOOLCHAIN_TOOLS no longer provides the binutils tools. 20150728: As ZFS requires more kernel stack pages than is the default on some architectures e.g. i386, it now warns if KSTACK_PAGES is less than ZFS_MIN_KSTACK_PAGES (which is 4 at the time of writing). Please consider using 'options KSTACK_PAGES=X' where X is greater than or equal to ZFS_MIN_KSTACK_PAGES i.e. 4 in such configurations. 20150706: sendmail has been updated to 8.15.2. Starting with FreeBSD 11.0 and sendmail 8.15, sendmail uses uncompressed IPv6 addresses by default, i.e., they will not contain "::". For example, instead of ::1, it will be 0:0:0:0:0:0:0:1. This permits a zero subnet to have a more specific match, such as different map entries for IPv6:0:0 vs IPv6:0. This change requires that configuration data (including maps, files, classes, custom ruleset, etc.) must use the same format, so make certain such configuration data is upgrading. As a very simple check search for patterns like 'IPv6:[0-9a-fA-F:]*::' and 'IPv6::'. To return to the old behavior, set the m4 option confUSE_COMPRESSED_IPV6_ADDRESSES or the cf option UseCompressedIPv6Addresses. 20150630: The default kernel entropy-processing algorithm is now Fortuna, replacing Yarrow. Assuming you have 'device random' in your kernel config file, the configurations allow a kernel option to override this default. You may choose *ONE* of: options RANDOM_YARROW # Legacy /dev/random algorithm. options RANDOM_DUMMY # Blocking-only driver. If you have neither, you get Fortuna. For most people, read no further, Fortuna will give a /dev/random that works like it always used to, and the difference will be irrelevant. If you remove 'device random', you get *NO* kernel-processed entropy at all. This may be acceptable to folks building embedded systems, but has complications. Carry on reading, and it is assumed you know what you need. *PLEASE* read random(4) and random(9) if you are in the habit of tweaking kernel configs, and/or if you are a member of the embedded community, wanting specific and not-usual behaviour from your security subsystems. NOTE!! If you use RANDOM_DUMMY and/or have no 'device random', you will NOT have a functioning /dev/random, and many cryptographic features will not work, including SSH. You may also find strange behaviour from the random(3) set of library functions, in particular sranddev(3), srandomdev(3) and arc4random(3). The reason for this is that the KERN_ARND sysctl only returns entropy if it thinks it has some to share, and with RANDOM_DUMMY or no 'device random' this will never happen. 20150623: An additional fix for the issue described in the 20150614 sendmail entry below has been committed in revision 284717. 20150616: FreeBSD's old make (fmake) has been removed from the system. It is available as the devel/fmake port or via pkg install fmake. 20150615: The fix for the issue described in the 20150614 sendmail entry below has been committed in revision 284436. The work around described in that entry is no longer needed unless the default setting is overridden by a confDH_PARAMETERS configuration setting of '5' or pointing to a 512 bit DH parameter file. 20150614: ALLOW_DEPRECATED_ATF_TOOLS/ATFFILE support has been removed from atf.test.mk (included from bsd.test.mk). Please upgrade devel/atf and devel/kyua to version 0.20+ and adjust any calling code to work with Kyuafile and kyua. 20150614: The import of openssl to address the FreeBSD-SA-15:10.openssl security advisory includes a change which rejects handshakes with DH parameters below 768 bits. sendmail releases prior to 8.15.2 (not yet released), defaulted to a 512 bit DH parameter setting for client connections. To work around this interoperability, sendmail can be configured to use a 2048 bit DH parameter by: 1. Edit /etc/mail/`hostname`.mc 2. If a setting for confDH_PARAMETERS does not exist or exists and is set to a string beginning with '5', replace it with '2'. 3. If a setting for confDH_PARAMETERS exists and is set to a file path, create a new file with: openssl dhparam -out /path/to/file 2048 4. Rebuild the .cf file: cd /etc/mail/; make; make install 5. Restart sendmail: cd /etc/mail/; make restart A sendmail patch is coming, at which time this file will be updated. 20150604: Generation of legacy formatted entries have been disabled by default in pwd_mkdb(8), as all base system consumers of the legacy formatted entries were converted to use the new format by default when the new, machine independent format have been added and supported since FreeBSD 5.x. Please see the pwd_mkdb(8) manual page for further details. 20150525: Clang and llvm have been upgraded to 3.6.1 release. Please see the 20141231 entry below for information about prerequisites and upgrading, if you are not already using 3.5.0 or higher. 20150521: TI platform code switched to using vendor DTS files and this update may break existing systems running on Beaglebone, Beaglebone Black, and Pandaboard: - dtb files should be regenerated/reinstalled. Filenames are the same but content is different now - GPIO addressing was changed, now each GPIO bank (32 pins per bank) has its own /dev/gpiocX device, e.g. pin 121 on /dev/gpioc0 in old addressing scheme is now pin 25 on /dev/gpioc3. - Pandaboard: /etc/ttys should be updated, serial console device is now /dev/ttyu2, not /dev/ttyu0 20150501: soelim(1) from gnu/usr.bin/groff has been replaced by usr.bin/soelim. If you need the GNU extension from groff soelim(1), install groff from package: pkg install groff, or via ports: textproc/groff. 20150423: chmod, chflags, chown and chgrp now affect symlinks in -R mode as defined in symlink(7); previously symlinks were silently ignored. 20150415: The const qualifier has been removed from iconv(3) to comply with POSIX. The ports tree is aware of this from r384038 onwards. 20150416: Libraries specified by LIBADD in Makefiles must have a corresponding DPADD_ variable to ensure correct dependencies. This is now enforced in src.libnames.mk. 20150324: From legacy ata(4) driver was removed support for SATA controllers supported by more functional drivers ahci(4), siis(4) and mvs(4). Kernel modules ataahci and ataadaptec were removed completely, replaced by ahci and mvs modules respectively. 20150315: Clang, llvm and lldb have been upgraded to 3.6.0 release. Please see the 20141231 entry below for information about prerequisites and upgrading, if you are not already using 3.5.0 or higher. 20150307: The 32-bit PowerPC kernel has been changed to a position-independent executable. This can only be booted with a version of loader(8) newer than January 31, 2015, so make sure to update both world and kernel before rebooting. 20150217: If you are running a -CURRENT kernel since r273872 (Oct 30th, 2014), but before r278950, the RNG was not seeded properly. Immediately upgrade the kernel to r278950 or later and regenerate any keys (e.g. ssh keys or openssl keys) that were generated w/ a kernel from that range. This does not affect programs that directly used /dev/random or /dev/urandom. All userland uses of arc4random(3) are affected. 20150210: The autofs(4) ABI was changed in order to restore binary compatibility with 10.1-RELEASE. The automountd(8) daemon needs to be rebuilt to work with the new kernel. 20150131: The powerpc64 kernel has been changed to a position-independent executable. This can only be booted with a new version of loader(8), so make sure to update both world and kernel before rebooting. 20150118: Clang and llvm have been upgraded to 3.5.1 release. This is a bugfix only release, no new features have been added. Please see the 20141231 entry below for information about prerequisites and upgrading, if you are not already using 3.5.0. 20150107: ELF tools addr2line, elfcopy (strip), nm, size, and strings are now taken from the ELF Tool Chain project rather than GNU binutils. They should be drop-in replacements, with the addition of arm64 support. The WITHOUT_ELFTOOLCHAIN_TOOLS= knob may be used to obtain the binutils tools, if necessary. See 20150805 for updated information. 20150105: The default Unbound configuration now enables remote control using a local socket. Users who have already enabled the local_unbound service should regenerate their configuration by running "service local_unbound setup" as root. 20150102: The GNU texinfo and GNU info pages have been removed. To be able to view GNU info pages please install texinfo from ports. 20141231: Clang, llvm and lldb have been upgraded to 3.5.0 release. As of this release, a prerequisite for building clang, llvm and lldb is a C++11 capable compiler and C++11 standard library. This means that to be able to successfully build the cross-tools stage of buildworld, with clang as the bootstrap compiler, your system compiler or cross compiler should either be clang 3.3 or later, or gcc 4.8 or later, and your system C++ library should be libc++, or libdstdc++ from gcc 4.8 or later. On any standard FreeBSD 10.x or 11.x installation, where clang and libc++ are on by default (that is, on x86 or arm), this should work out of the box. On 9.x installations where clang is enabled by default, e.g. on x86 and powerpc, libc++ will not be enabled by default, so libc++ should be built (with clang) and installed first. If both clang and libc++ are missing, build clang first, then use it to build libc++. On 8.x and earlier installations, upgrade to 9.x first, and then follow the instructions for 9.x above. Sparc64 and mips users are unaffected, as they still use gcc 4.2.1 by default, and do not build clang. Many embedded systems are resource constrained, and will not be able to build clang in a reasonable time, or in some cases at all. In those cases, cross building bootable systems on amd64 is a workaround. This new version of clang introduces a number of new warnings, of which the following are most likely to appear: -Wabsolute-value This warns in two cases, for both C and C++: * When the code is trying to take the absolute value of an unsigned quantity, which is effectively a no-op, and almost never what was intended. The code should be fixed, if at all possible. If you are sure that the unsigned quantity can be safely cast to signed, without loss of information or undefined behavior, you can add an explicit cast, or disable the warning. * When the code is trying to take an absolute value, but the called abs() variant is for the wrong type, which can lead to truncation. If you want to disable the warning instead of fixing the code, please make sure that truncation will not occur, or it might lead to unwanted side-effects. -Wtautological-undefined-compare and -Wundefined-bool-conversion These warn when C++ code is trying to compare 'this' against NULL, while 'this' should never be NULL in well-defined C++ code. However, there is some legacy (pre C++11) code out there, which actively abuses this feature, which was less strictly defined in previous C++ versions. Squid and openjdk do this, for example. The warning can be turned off for C++98 and earlier, but compiling the code in C++11 mode might result in unexpected behavior; for example, the parts of the program that are unreachable could be optimized away. 20141222: The old NFS client and server (kernel options NFSCLIENT, NFSSERVER) kernel sources have been removed. The .h files remain, since some utilities include them. This will need to be fixed later. If "mount -t oldnfs ..." is attempted, it will fail. If the "-o" option on mountd(8), nfsd(8) or nfsstat(1) is used, the utilities will report errors. 20141121: The handling of LOCAL_LIB_DIRS has been altered to skip addition of directories to top level SUBDIR variable when their parent directory is included in LOCAL_DIRS. Users with build systems with such hierarchies and without SUBDIR entries in the parent directory Makefiles should add them or add the directories to LOCAL_DIRS. 20141109: faith(4) and faithd(8) have been removed from the base system. Faith has been obsolete for a very long time. 20141104: vt(4), the new console driver, is enabled by default. It brings support for Unicode and double-width characters, as well as support for UEFI and integration with the KMS kernel video drivers. You may need to update your console settings in /etc/rc.conf, most probably the keymap. During boot, /etc/rc.d/syscons will indicate what you need to do. vt(4) still has issues and lacks some features compared to syscons(4). See the wiki for up-to-date information: https://wiki.freebsd.org/Newcons If you want to keep using syscons(4), you can do so by adding the following line to /boot/loader.conf: kern.vty=sc 20141102: pjdfstest has been integrated into kyua as an opt-in test suite. Please see share/doc/pjdfstest/README for more details on how to execute it. 20141009: gperf has been removed from the base system for architectures that use clang. Ports that require gperf will obtain it from the devel/gperf port. 20140923: pjdfstest has been moved from tools/regression/pjdfstest to contrib/pjdfstest . 20140922: At svn r271982, The default linux compat kernel ABI has been adjusted to 2.6.18 in support of the linux-c6 compat ports infrastructure update. If you wish to continue using the linux-f10 compat ports, add compat.linux.osrelease=2.6.16 to your local sysctl.conf. Users are encouraged to update their linux-compat packages to linux-c6 during their next update cycle. 20140729: The ofwfb driver, used to provide a graphics console on PowerPC when using vt(4), no longer allows mmap() of all physical memory. This will prevent Xorg on PowerPC with some ATI graphics cards from initializing properly unless x11-servers/xorg-server is updated to 1.12.4_8 or newer. 20140723: The xdev targets have been converted to using TARGET and TARGET_ARCH instead of XDEV and XDEV_ARCH. 20140719: The default unbound configuration has been modified to address issues with reverse lookups on networks that use private address ranges. If you use the local_unbound service, run "service local_unbound setup" as root to regenerate your configuration, then "service local_unbound reload" to load the new configuration. 20140709: The GNU texinfo and GNU info pages are not built and installed anymore, WITH_INFO knob has been added to allow to built and install them again. UPDATE: see 20150102 entry on texinfo's removal 20140708: The GNU readline library is now an INTERNALLIB - that is, it is statically linked into consumers (GDB and variants) in the base system, and the shared library is no longer installed. The devel/readline port is available for third party software that requires readline. 20140702: The Itanium architecture (ia64) has been removed from the list of known architectures. This is the first step in the removal of the architecture. 20140701: Commit r268115 has added NFSv4.1 server support, merged from projects/nfsv4.1-server. Since this includes changes to the internal interfaces between the NFS related modules, a full build of the kernel and modules will be necessary. __FreeBSD_version has been bumped. 20140629: The WITHOUT_VT_SUPPORT kernel config knob has been renamed WITHOUT_VT. (The other _SUPPORT knobs have a consistent meaning which differs from the behaviour controlled by this knob.) 20140619: Maximal length of the serial number in CTL was increased from 16 to 64 chars, that breaks ABI. All CTL-related tools, such as ctladm and ctld, need to be rebuilt to work with a new kernel. 20140606: The libatf-c and libatf-c++ major versions were downgraded to 0 and 1 respectively to match the upstream numbers. They were out of sync because, when they were originally added to FreeBSD, the upstream versions were not respected. These libraries are private and not yet built by default, so renumbering them should be a non-issue. However, unclean source trees will yield broken test programs once the operator executes "make delete-old-libs" after a "make installworld". Additionally, the atf-sh binary was made private by moving it into /usr/libexec/. Already-built shell test programs will keep the path to the old binary so they will break after "make delete-old" is run. If you are using WITH_TESTS=yes (not the default), wipe the object tree and rebuild from scratch to prevent spurious test failures. This is only needed once: the misnumbered libraries and misplaced binaries have been added to OptionalObsoleteFiles.inc so they will be removed during a clean upgrade. 20140512: Clang and llvm have been upgraded to 3.4.1 release. 20140508: We bogusly installed src.opts.mk in /usr/share/mk. This file should be removed to avoid issues in the future (and has been added to ObsoleteFiles.inc). 20140505: /etc/src.conf now affects only builds of the FreeBSD src tree. In the past, it affected all builds that used the bsd.*.mk files. The old behavior was a bug, but people may have relied upon it. To get this behavior back, you can .include /etc/src.conf from /etc/make.conf (which is still global and isn't changed). This also changes the behavior of incremental builds inside the tree of individual directories. Set MAKESYSPATH to ".../share/mk" to do that. Although this has survived make universe and some upgrade scenarios, other upgrade scenarios may have broken. At least one form of temporary breakage was fixed with MAKESYSPATH settings for buildworld as well... In cases where MAKESYSPATH isn't working with this setting, you'll need to set it to the full path to your tree. One side effect of all this cleaning up is that bsd.compiler.mk is no longer implicitly included by bsd.own.mk. If you wish to use COMPILER_TYPE, you must now explicitly include bsd.compiler.mk as well. 20140430: The lindev device has been removed since /dev/full has been made a standard device. __FreeBSD_version has been bumped. 20140424: The knob WITHOUT_VI was added to the base system, which controls building ex(1), vi(1), etc. Older releases of FreeBSD required ex(1) in order to reorder files share/termcap and didn't build ex(1) as a build tool, so building/installing with WITH_VI is highly advised for build hosts for older releases. This issue has been fixed in stable/9 and stable/10 in r277022 and r276991, respectively. 20140418: The YES_HESIOD knob has been removed. It has been obsolete for a decade. Please move to using WITH_HESIOD instead or your builds will silently lack HESIOD. 20140405: The uart(4) driver has been changed with respect to its handling of the low-level console. Previously the uart(4) driver prevented any process from changing the baudrate or the CLOCAL and HUPCL control flags. By removing the restrictions, operators can make changes to the serial console port without having to reboot. However, when getty(8) is started on the serial device that is associated with the low-level console, a misconfigured terminal line in /etc/ttys will now have a real impact. Before upgrading the kernel, make sure that /etc/ttys has the serial console device configured as 3wire without baudrate to preserve the previous behaviour. E.g: ttyu0 "/usr/libexec/getty 3wire" vt100 on secure 20140306: Support for libwrap (TCP wrappers) in rpcbind was disabled by default to improve performance. To re-enable it, if needed, run rpcbind with command line option -W. 20140226: Switched back to the GPL dtc compiler due to updates in the upstream dts files not being supported by the BSDL dtc compiler. You will need to rebuild your kernel toolchain to pick up the new compiler. Core dumps may result while building dtb files during a kernel build if you fail to do so. Set WITHOUT_GPL_DTC if you require the BSDL compiler. 20140216: Clang and llvm have been upgraded to 3.4 release. 20140216: The nve(4) driver has been removed. Please use the nfe(4) driver for NVIDIA nForce MCP Ethernet adapters instead. 20140212: An ABI incompatibility crept into the libc++ 3.4 import in r261283. This could cause certain C++ applications using shared libraries built against the previous version of libc++ to crash. The incompatibility has now been fixed, but any C++ applications or shared libraries built between r261283 and r261801 should be recompiled. 20140204: OpenSSH will now ignore errors caused by kernel lacking of Capsicum capability mode support. Please note that enabling the feature in kernel is still highly recommended. 20140131: OpenSSH is now built with sandbox support, and will use sandbox as the default privilege separation method. This requires Capsicum capability mode support in kernel. 20140128: The libelf and libdwarf libraries have been updated to newer versions from upstream. Shared library version numbers for these two libraries were bumped. Any ports or binaries requiring these two libraries should be recompiled. __FreeBSD_version is bumped to 1100006. 20140110: If a Makefile in a tests/ directory was auto-generating a Kyuafile instead of providing an explicit one, this would prevent such Makefile from providing its own Kyuafile in the future during NO_CLEAN builds. This has been fixed in the Makefiles but manual intervention is needed to clean an objdir if you use NO_CLEAN: # find /usr/obj -name Kyuafile | xargs rm -f 20131213: The behavior of gss_pseudo_random() for the krb5 mechanism has changed, for applications requesting a longer random string than produced by the underlying enctype's pseudo-random() function. In particular, the random string produced from a session key of enctype aes256-cts-hmac-sha1-96 or aes256-cts-hmac-sha1-96 will be different at the 17th octet and later, after this change. The counter used in the PRF+ construction is now encoded as a big-endian integer in accordance with RFC 4402. __FreeBSD_version is bumped to 1100004. 20131108: The WITHOUT_ATF build knob has been removed and its functionality has been subsumed into the more generic WITHOUT_TESTS. If you were using the former to disable the build of the ATF libraries, you should change your settings to use the latter. 20131025: The default version of mtree is nmtree which is obtained from NetBSD. The output is generally the same, but may vary slightly. If you found you need identical output adding "-F freebsd9" to the command line should do the trick. For the time being, the old mtree is available as fmtree. 20131014: libbsdyml has been renamed to libyaml and moved to /usr/lib/private. This will break ports-mgmt/pkg. Rebuild the port, or upgrade to pkg 1.1.4_8 and verify bsdyml not linked in, before running "make delete-old-libs": # make -C /usr/ports/ports-mgmt/pkg build deinstall install clean or # pkg install pkg; ldd /usr/local/sbin/pkg | grep bsdyml 20131010: The stable/10 branch has been created in subversion from head revision r256279. COMMON ITEMS: General Notes ------------- Sometimes, obscure build problems are the result of environment poisoning. This can happen because the make utility reads its environment when searching for values for global variables. To run your build attempts in an "environmental clean room", prefix all make commands with 'env -i '. See the env(1) manual page for more details. Occasionally a build failure will occur with "make -j" due to a race condition. If this happens try building again without -j, and please report a bug if it happens consistently. When upgrading from one major version to another it is generally best to upgrade to the latest code in the currently installed branch first, then do an upgrade to the new branch. This is the best-tested upgrade path, and has the highest probability of being successful. Please try this approach if you encounter problems with a major version upgrade. Since the stable 4.x branch point, one has generally been able to upgrade from anywhere in the most recent stable branch to head / current (or even the last couple of stable branches). See the top of this file when there's an exception. The update process will emit an error on an attempt to perform a build or install from a FreeBSD version below the earliest supported version. When updating from an older version the update should be performed one major release at a time, including running `make delete-old` at each step. When upgrading a live system, having a root shell around before installing anything can help undo problems. Not having a root shell around can lead to problems if pam has changed too much from your starting point to allow continued authentication after the upgrade. This file should be read as a log of events. When a later event changes information of a prior event, the prior event should not be deleted. Instead, a pointer to the entry with the new information should be placed in the old entry. Readers of this file should also sanity check older entries before relying on them blindly. Authors of new entries should write them with this in mind. ZFS notes --------- When upgrading the boot ZFS pool to a new version, always follow these two steps: 1.) recompile and reinstall the ZFS boot loader and boot block (this is part of "make buildworld" and "make installworld") 2.) update the ZFS boot block on your boot drive The following example updates the ZFS boot block on the freebsd-boot partition of a GPT partitioned drive ada0: "gpart bootcode -p /boot/gptzfsboot -i $N ada0" The value $N will typically be 1 (if booting from BIOS) or 2 (if booting from EFI). Non-boot pools do not need these updates. To build a kernel ----------------- If you are updating from a prior version of FreeBSD (even one just a few days old), you should follow this procedure. It is the most failsafe as it uses a /usr/obj tree with a fresh mini-buildworld, make kernel-toolchain make -DALWAYS_CHECK_MAKE buildkernel KERNCONF=YOUR_KERNEL_HERE make -DALWAYS_CHECK_MAKE installkernel KERNCONF=YOUR_KERNEL_HERE To test a kernel once --------------------- If you just want to boot a kernel once (because you are not sure if it works, or if you want to boot a known bad kernel to provide debugging information) run make installkernel KERNCONF=YOUR_KERNEL_HERE KODIR=/boot/testkernel nextboot -k testkernel To rebuild everything and install it on the current system. ----------------------------------------------------------- # Note: sometimes if you are running current you gotta do more than # is listed here if you are upgrading from a really old current. make buildworld make buildkernel KERNCONF=YOUR_KERNEL_HERE make installkernel KERNCONF=YOUR_KERNEL_HERE [1] [3] mergemaster -Fp [5] make installworld mergemaster -Fi [4] make delete-old [6] To cross-install current onto a separate partition -------------------------------------------------- # In this approach we use a separate partition to hold # current's root, 'usr', and 'var' directories. A partition # holding "/", "/usr" and "/var" should be about 2GB in # size. make buildworld make buildkernel KERNCONF=YOUR_KERNEL_HERE make installworld DESTDIR=${CURRENT_ROOT} -DDB_FROM_SRC make distribution DESTDIR=${CURRENT_ROOT} # if newfs'd make installkernel KERNCONF=YOUR_KERNEL_HERE DESTDIR=${CURRENT_ROOT} cp /etc/fstab ${CURRENT_ROOT}/etc/fstab # if newfs'd To upgrade in-place from stable to current ---------------------------------------------- make buildworld [9] make buildkernel KERNCONF=YOUR_KERNEL_HERE [8] make installkernel KERNCONF=YOUR_KERNEL_HERE [1] [3] mergemaster -Fp [5] make installworld mergemaster -Fi [4] make delete-old [6] Make sure that you've read the UPDATING file to understand the tweaks to various things you need. At this point in the life cycle of current, things change often and you are on your own to cope. The defaults can also change, so please read ALL of the UPDATING entries. Also, if you are tracking -current, you must be subscribed to freebsd-current@freebsd.org. Make sure that before you update your sources that you have read and understood all the recent messages there. If in doubt, please track -stable which has much fewer pitfalls. [1] If you have third party modules, such as vmware, you should disable them at this point so they don't crash your system on reboot. Alternatively, you should rebuild all the modules you have in your system and install them as well. If you are running -current, you should seriously consider placing all sources to all the modules for your system (or symlinks to them) in /usr/local/sys/modules so this happens automatically. If all your modules come from ports, then adding the port origin directories to PORTS_MODULES instead is also automatic and effective, eg: PORTS_MODULES+=x11/nvidia-driver [3] From the bootblocks, boot -s, and then do fsck -p mount -u / mount -a sh /etc/rc.d/zfs start # mount zfs filesystem, if needed cd src # full path to source adjkerntz -i # if CMOS is wall time Also, when doing a major release upgrade, it is required that you boot into single user mode to do the installworld. [4] Note: This step is non-optional. Failure to do this step can result in a significant reduction in the functionality of the system. Attempting to do it by hand is not recommended and those that pursue this avenue should read this file carefully, as well as the archives of freebsd-current and freebsd-hackers mailing lists for potential gotchas. The -U option is also useful to consider. See mergemaster(8) for more information. [5] Usually this step is a no-op. However, from time to time you may need to do this if you get unknown user in the following step. It never hurts to do it all the time. You may need to install a new mergemaster (cd src/usr.sbin/mergemaster && make install) after the buildworld before this step if you last updated from current before 20130425 or from -stable before 20130430. [6] This only deletes old files and directories. Old libraries can be deleted by "make delete-old-libs", but you have to make sure that no program is using those libraries anymore. [8] The new kernel must be able to run existing binaries used by an installworld. When upgrading across major versions, the new kernel's configuration must include the correct COMPAT_FREEBSD option for existing binaries (e.g. COMPAT_FREEBSD11 to run 11.x binaries). Failure to do so may leave you with a system that is hard to boot to recover. A GENERIC kernel will include suitable compatibility options to run binaries from older branches. Note that the ability to run binaries from unsupported branches is not guaranteed. Make sure that you merge any new devices from GENERIC since the last time you updated your kernel config file. Options also change over time, so you may need to adjust your custom kernels for these as well. [9] If CPUTYPE is defined in your /etc/make.conf, make sure to use the "?=" instead of the "=" assignment operator, so that buildworld can override the CPUTYPE if it needs to. MAKEOBJDIRPREFIX must be defined in an environment variable, and not on the command line, or in /etc/make.conf. buildworld will warn if it is improperly defined. FORMAT: This file contains a list, in reverse chronological order, of major breakages in tracking -current. It is not guaranteed to be a complete list of such breakages, and only contains entries since September 23, 2011. If you need to see UPDATING entries from before that date, you will need to fetch an UPDATING file from an older FreeBSD release. Copyright information: Copyright 1998-2009 M. Warner Losh Redistribution, publication, translation and use, with or without modification, in full or in part, in any form or format of this document are permitted without further permission from the author. THIS DOCUMENT IS PROVIDED BY WARNER LOSH ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL WARNER LOSH BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. Contact Warner Losh if you have any questions about your use of this document. $FreeBSD$ diff --git a/contrib/ofed/libmlx5/mlx5.c b/contrib/ofed/libmlx5/mlx5.c index 1b860ed2ce76..8dbb268f55f7 100644 --- a/contrib/ofed/libmlx5/mlx5.c +++ b/contrib/ofed/libmlx5/mlx5.c @@ -1,1029 +1,1028 @@ /* * Copyright (c) 2012 Mellanox Technologies, Inc. All rights reserved. * * This software is available to you under a choice of one of two * licenses. You may choose to be licensed under the terms of the GNU * General Public License (GPL) Version 2, available from the file * COPYING in the main directory of this source tree, or the * OpenIB.org BSD license below: * * Redistribution and use in source and binary forms, with or * without modification, are permitted provided that the following * conditions are met: * * - Redistributions of source code must retain the above * copyright notice, this list of conditions and the following * disclaimer. * * - Redistributions in binary form must reproduce the above * copyright notice, this list of conditions and the following * disclaimer in the documentation and/or other materials * provided with the distribution. * * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF * MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND * NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS * BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN * ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN * CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ #define _GNU_SOURCE #include #include #include #include #include #include #include #include #include #include #include #include "mlx5.h" #include "mlx5-abi.h" #ifndef PCI_VENDOR_ID_MELLANOX #define PCI_VENDOR_ID_MELLANOX 0x15b3 #endif #ifndef CPU_OR #define CPU_OR(x, y, z) do {} while (0) #endif #ifndef CPU_EQUAL #define CPU_EQUAL(x, y) 1 #endif #define HCA(v, d) \ { .vendor = PCI_VENDOR_ID_##v, \ .device = d } static struct { unsigned vendor; unsigned device; } hca_table[] = { HCA(MELLANOX, 4113), /* MT4113 Connect-IB */ HCA(MELLANOX, 4114), /* Connect-IB Virtual Function */ HCA(MELLANOX, 4115), /* ConnectX-4 */ HCA(MELLANOX, 4116), /* ConnectX-4 Virtual Function */ HCA(MELLANOX, 4117), /* ConnectX-4LX */ HCA(MELLANOX, 4118), /* ConnectX-4LX Virtual Function */ HCA(MELLANOX, 4119), /* ConnectX-5, PCIe 3.0 */ HCA(MELLANOX, 4120), /* ConnectX-5 Virtual Function */ HCA(MELLANOX, 4121), /* ConnectX-5 Ex */ HCA(MELLANOX, 4122), /* ConnectX-5 Ex VF */ HCA(MELLANOX, 4123), /* ConnectX-6 */ HCA(MELLANOX, 4124), /* ConnectX-6 VF */ HCA(MELLANOX, 4125), /* ConnectX-6 DX */ HCA(MELLANOX, 4126), /* ConnectX family mlx5Gen Virtual Function */ HCA(MELLANOX, 41682), /* BlueField integrated ConnectX-5 network controller */ HCA(MELLANOX, 41683), /* BlueField integrated ConnectX-5 network controller VF */ }; uint32_t mlx5_debug_mask = 0; int mlx5_freeze_on_error_cqe; static struct ibv_context_ops mlx5_ctx_ops = { .query_device = mlx5_query_device, .query_port = mlx5_query_port, .alloc_pd = mlx5_alloc_pd, .dealloc_pd = mlx5_free_pd, .reg_mr = mlx5_reg_mr, .rereg_mr = mlx5_rereg_mr, .dereg_mr = mlx5_dereg_mr, .alloc_mw = mlx5_alloc_mw, .dealloc_mw = mlx5_dealloc_mw, .bind_mw = mlx5_bind_mw, .create_cq = mlx5_create_cq, .poll_cq = mlx5_poll_cq, .req_notify_cq = mlx5_arm_cq, .cq_event = mlx5_cq_event, .resize_cq = mlx5_resize_cq, .destroy_cq = mlx5_destroy_cq, .create_srq = mlx5_create_srq, .modify_srq = mlx5_modify_srq, .query_srq = mlx5_query_srq, .destroy_srq = mlx5_destroy_srq, .post_srq_recv = mlx5_post_srq_recv, .create_qp = mlx5_create_qp, .query_qp = mlx5_query_qp, .modify_qp = mlx5_modify_qp, .destroy_qp = mlx5_destroy_qp, .post_send = mlx5_post_send, .post_recv = mlx5_post_recv, .create_ah = mlx5_create_ah, .destroy_ah = mlx5_destroy_ah, .attach_mcast = mlx5_attach_mcast, .detach_mcast = mlx5_detach_mcast }; static int read_number_from_line(const char *line, int *value) { const char *ptr; ptr = strchr(line, ':'); if (!ptr) return 1; ++ptr; *value = atoi(ptr); return 0; } /** * The function looks for the first free user-index in all the * user-index tables. If all are used, returns -1, otherwise * a valid user-index. * In case the reference count of the table is zero, it means the * table is not in use and wasn't allocated yet, therefore the * mlx5_store_uidx allocates the table, and increment the reference * count on the table. */ static int32_t get_free_uidx(struct mlx5_context *ctx) { int32_t tind; int32_t i; for (tind = 0; tind < MLX5_UIDX_TABLE_SIZE; tind++) { if (ctx->uidx_table[tind].refcnt < MLX5_UIDX_TABLE_MASK) break; } if (tind == MLX5_UIDX_TABLE_SIZE) return -1; if (!ctx->uidx_table[tind].refcnt) return tind << MLX5_UIDX_TABLE_SHIFT; for (i = 0; i < MLX5_UIDX_TABLE_MASK + 1; i++) { if (!ctx->uidx_table[tind].table[i]) break; } return (tind << MLX5_UIDX_TABLE_SHIFT) | i; } int32_t mlx5_store_uidx(struct mlx5_context *ctx, void *rsc) { int32_t tind; int32_t ret = -1; int32_t uidx; pthread_mutex_lock(&ctx->uidx_table_mutex); uidx = get_free_uidx(ctx); if (uidx < 0) goto out; tind = uidx >> MLX5_UIDX_TABLE_SHIFT; if (!ctx->uidx_table[tind].refcnt) { ctx->uidx_table[tind].table = calloc(MLX5_UIDX_TABLE_MASK + 1, sizeof(struct mlx5_resource *)); if (!ctx->uidx_table[tind].table) goto out; } ++ctx->uidx_table[tind].refcnt; ctx->uidx_table[tind].table[uidx & MLX5_UIDX_TABLE_MASK] = rsc; ret = uidx; out: pthread_mutex_unlock(&ctx->uidx_table_mutex); return ret; } void mlx5_clear_uidx(struct mlx5_context *ctx, uint32_t uidx) { int tind = uidx >> MLX5_UIDX_TABLE_SHIFT; pthread_mutex_lock(&ctx->uidx_table_mutex); if (!--ctx->uidx_table[tind].refcnt) free(ctx->uidx_table[tind].table); else ctx->uidx_table[tind].table[uidx & MLX5_UIDX_TABLE_MASK] = NULL; pthread_mutex_unlock(&ctx->uidx_table_mutex); } static int mlx5_is_sandy_bridge(int *num_cores) { char line[128]; FILE *fd; int rc = 0; int cur_cpu_family = -1; int cur_cpu_model = -1; fd = fopen("/proc/cpuinfo", "r"); if (!fd) return 0; *num_cores = 0; while (fgets(line, 128, fd)) { int value; /* if this is information on new processor */ if (!strncmp(line, "processor", 9)) { ++*num_cores; cur_cpu_family = -1; cur_cpu_model = -1; } else if (!strncmp(line, "cpu family", 10)) { if ((cur_cpu_family < 0) && (!read_number_from_line(line, &value))) cur_cpu_family = value; } else if (!strncmp(line, "model", 5)) { if ((cur_cpu_model < 0) && (!read_number_from_line(line, &value))) cur_cpu_model = value; } /* if this is a Sandy Bridge CPU */ if ((cur_cpu_family == 6) && (cur_cpu_model == 0x2A || (cur_cpu_model == 0x2D) )) rc = 1; } fclose(fd); return rc; } /* man cpuset This format displays each 32-bit word in hexadecimal (using ASCII characters "0" - "9" and "a" - "f"); words are filled with leading zeros, if required. For masks longer than one word, a comma separator is used between words. Words are displayed in big-endian order, which has the most significant bit first. The hex digits within a word are also in big-endian order. The number of 32-bit words displayed is the minimum number needed to display all bits of the bitmask, based on the size of the bitmask. Examples of the Mask Format: 00000001 # just bit 0 set 40000000,00000000,00000000 # just bit 94 set 000000ff,00000000 # bits 32-39 set 00000000,000E3862 # 1,5,6,11-13,17-19 set A mask with bits 0, 1, 2, 4, 8, 16, 32, and 64 set displays as: 00000001,00000001,00010117 The first "1" is for bit 64, the second for bit 32, the third for bit 16, the fourth for bit 8, the fifth for bit 4, and the "7" is for bits 2, 1, and 0. */ static void mlx5_local_cpu_set(struct ibv_device *ibdev, cpuset_t *cpu_set) { char *p, buf[1024]; char *env_value; uint32_t word; int i, k; env_value = getenv("MLX5_LOCAL_CPUS"); if (env_value) strncpy(buf, env_value, sizeof(buf)); else { char fname[MAXPATHLEN]; snprintf(fname, MAXPATHLEN, "/sys/class/infiniband/%s", ibv_get_device_name(ibdev)); if (ibv_read_sysfs_file(fname, "device/local_cpus", buf, sizeof(buf))) { fprintf(stderr, PFX "Warning: can not get local cpu set: failed to open %s\n", fname); return; } } p = strrchr(buf, ','); if (!p) p = buf; i = 0; do { if (*p == ',') { *p = 0; p ++; } word = strtoul(p, NULL, 16); for (k = 0; word; ++k, word >>= 1) if (word & 1) CPU_SET(k+i, cpu_set); if (p == buf) break; p = strrchr(buf, ','); if (!p) p = buf; i += 32; } while (i < CPU_SETSIZE); } static int mlx5_enable_sandy_bridge_fix(struct ibv_device *ibdev) { cpuset_t my_cpus, dev_local_cpus, result_set; int stall_enable; int ret; int num_cores; if (!mlx5_is_sandy_bridge(&num_cores)) return 0; /* by default enable stall on sandy bridge arch */ stall_enable = 1; /* * check if app is bound to cpu set that is inside * of device local cpu set. Disable stalling if true */ /* use static cpu set - up to CPU_SETSIZE (1024) cpus/node */ CPU_ZERO(&my_cpus); CPU_ZERO(&dev_local_cpus); CPU_ZERO(&result_set); ret = cpuset_getaffinity(CPU_LEVEL_WHICH, CPU_WHICH_PID, -1, sizeof(my_cpus), &my_cpus); if (ret == -1) { if (errno == EINVAL) fprintf(stderr, PFX "Warning: my cpu set is too small\n"); else fprintf(stderr, PFX "Warning: failed to get my cpu set\n"); goto out; } /* get device local cpu set */ mlx5_local_cpu_set(ibdev, &dev_local_cpus); /* check if my cpu set is in dev cpu */ - CPU_OR(&result_set, &my_cpus); - CPU_OR(&result_set, &dev_local_cpus); + CPU_OR(&result_set, &my_cpus, &dev_local_cpus); stall_enable = CPU_EQUAL(&result_set, &dev_local_cpus) ? 0 : 1; out: return stall_enable; } static void mlx5_read_env(struct ibv_device *ibdev, struct mlx5_context *ctx) { char *env_value; env_value = getenv("MLX5_STALL_CQ_POLL"); if (env_value) /* check if cq stall is enforced by user */ ctx->stall_enable = (strcmp(env_value, "0")) ? 1 : 0; else /* autodetect if we need to do cq polling */ ctx->stall_enable = mlx5_enable_sandy_bridge_fix(ibdev); env_value = getenv("MLX5_STALL_NUM_LOOP"); if (env_value) mlx5_stall_num_loop = atoi(env_value); env_value = getenv("MLX5_STALL_CQ_POLL_MIN"); if (env_value) mlx5_stall_cq_poll_min = atoi(env_value); env_value = getenv("MLX5_STALL_CQ_POLL_MAX"); if (env_value) mlx5_stall_cq_poll_max = atoi(env_value); env_value = getenv("MLX5_STALL_CQ_INC_STEP"); if (env_value) mlx5_stall_cq_inc_step = atoi(env_value); env_value = getenv("MLX5_STALL_CQ_DEC_STEP"); if (env_value) mlx5_stall_cq_dec_step = atoi(env_value); ctx->stall_adaptive_enable = 0; ctx->stall_cycles = 0; if (mlx5_stall_num_loop < 0) { ctx->stall_adaptive_enable = 1; ctx->stall_cycles = mlx5_stall_cq_poll_min; } } static int get_total_uuars(int page_size) { int size = MLX5_DEF_TOT_UUARS; int uuars_in_page; char *env; env = getenv("MLX5_TOTAL_UUARS"); if (env) size = atoi(env); if (size < 1) return -EINVAL; uuars_in_page = page_size / MLX5_ADAPTER_PAGE_SIZE * MLX5_NUM_NON_FP_BFREGS_PER_UAR; size = max(uuars_in_page, size); size = align(size, MLX5_NUM_NON_FP_BFREGS_PER_UAR); if (size > MLX5_MAX_BFREGS) return -ENOMEM; return size; } static void open_debug_file(struct mlx5_context *ctx) { char *env; env = getenv("MLX5_DEBUG_FILE"); if (!env) { ctx->dbg_fp = stderr; return; } ctx->dbg_fp = fopen(env, "aw+"); if (!ctx->dbg_fp) { fprintf(stderr, "Failed opening debug file %s, using stderr\n", env); ctx->dbg_fp = stderr; return; } } static void close_debug_file(struct mlx5_context *ctx) { if (ctx->dbg_fp && ctx->dbg_fp != stderr) fclose(ctx->dbg_fp); } static void set_debug_mask(void) { char *env; env = getenv("MLX5_DEBUG_MASK"); if (env) mlx5_debug_mask = strtol(env, NULL, 0); } static void set_freeze_on_error(void) { char *env; env = getenv("MLX5_FREEZE_ON_ERROR_CQE"); if (env) mlx5_freeze_on_error_cqe = strtol(env, NULL, 0); } static int get_always_bf(void) { char *env; env = getenv("MLX5_POST_SEND_PREFER_BF"); if (!env) return 1; return strcmp(env, "0") ? 1 : 0; } static int get_shut_up_bf(void) { char *env; env = getenv("MLX5_SHUT_UP_BF"); if (!env) return 0; return strcmp(env, "0") ? 1 : 0; } static int get_num_low_lat_uuars(int tot_uuars) { char *env; int num = 4; env = getenv("MLX5_NUM_LOW_LAT_UUARS"); if (env) num = atoi(env); if (num < 0) return -EINVAL; num = max(num, tot_uuars - MLX5_MED_BFREGS_TSHOLD); return num; } /* The library allocates an array of uuar contexts. The one in index zero does * not to execersize odd/even policy so it can avoid a lock but it may not use * blue flame. The upper ones, low_lat_uuars can use blue flame with no lock * since they are assigned to one QP only. The rest can use blue flame but since * they are shared they need a lock */ static int need_uuar_lock(struct mlx5_context *ctx, int uuarn) { if (uuarn == 0 || mlx5_single_threaded) return 0; if (uuarn >= (ctx->tot_uuars - ctx->low_lat_uuars) * 2) return 0; return 1; } static int single_threaded_app(void) { char *env; env = getenv("MLX5_SINGLE_THREADED"); if (env) return strcmp(env, "1") ? 0 : 1; return 0; } static int mlx5_cmd_get_context(struct mlx5_context *context, struct mlx5_alloc_ucontext *req, size_t req_len, struct mlx5_alloc_ucontext_resp *resp, size_t resp_len) { if (!ibv_cmd_get_context(&context->ibv_ctx, &req->ibv_req, req_len, &resp->ibv_resp, resp_len)) return 0; /* The ibv_cmd_get_context fails in older kernels when passing * a request length that the kernel doesn't know. * To avoid breaking compatibility of new libmlx5 and older * kernels, when ibv_cmd_get_context fails with the full * request length, we try once again with the legacy length. * We repeat this process while reducing requested size based * on the feature input size. To avoid this in the future, we * will remove the check in kernel that requires fields unknown * to the kernel to be cleared. This will require that any new * feature that involves extending struct mlx5_alloc_ucontext * will be accompanied by an indication in the form of one or * more fields in struct mlx5_alloc_ucontext_resp. If the * response value can be interpreted as feature not supported * when the returned value is zero, this will suffice to * indicate to the library that the request was ignored by the * kernel, either because it is unaware or because it decided * to do so. If zero is a valid response, we will add a new * field that indicates whether the request was handled. */ if (!ibv_cmd_get_context(&context->ibv_ctx, &req->ibv_req, offsetof(struct mlx5_alloc_ucontext, lib_caps), &resp->ibv_resp, resp_len)) return 0; return ibv_cmd_get_context(&context->ibv_ctx, &req->ibv_req, offsetof(struct mlx5_alloc_ucontext, cqe_version), &resp->ibv_resp, resp_len); } static int mlx5_map_internal_clock(struct mlx5_device *mdev, struct ibv_context *ibv_ctx) { struct mlx5_context *context = to_mctx(ibv_ctx); void *hca_clock_page; off_t offset = 0; set_command(MLX5_MMAP_GET_CORE_CLOCK_CMD, &offset); hca_clock_page = mmap(NULL, mdev->page_size, PROT_READ, MAP_SHARED, ibv_ctx->cmd_fd, mdev->page_size * offset); if (hca_clock_page == MAP_FAILED) { fprintf(stderr, PFX "Warning: Timestamp available,\n" "but failed to mmap() hca core clock page.\n"); return -1; } context->hca_core_clock = hca_clock_page + (context->core_clock.offset & (mdev->page_size - 1)); return 0; } int mlx5dv_query_device(struct ibv_context *ctx_in, struct mlx5dv_context *attrs_out) { struct mlx5_context *mctx = to_mctx(ctx_in); uint64_t comp_mask_out = 0; attrs_out->version = 0; attrs_out->flags = 0; if (mctx->cqe_version == MLX5_CQE_VERSION_V1) attrs_out->flags |= MLX5DV_CONTEXT_FLAGS_CQE_V1; if (mctx->vendor_cap_flags & MLX5_VENDOR_CAP_FLAGS_MPW) attrs_out->flags |= MLX5DV_CONTEXT_FLAGS_MPW; if (attrs_out->comp_mask & MLX5DV_CONTEXT_MASK_CQE_COMPRESION) { attrs_out->cqe_comp_caps = mctx->cqe_comp_caps; comp_mask_out |= MLX5DV_CONTEXT_MASK_CQE_COMPRESION; } attrs_out->comp_mask = comp_mask_out; return 0; } static int mlx5dv_get_qp(struct ibv_qp *qp_in, struct mlx5dv_qp *qp_out) { struct mlx5_qp *mqp = to_mqp(qp_in); qp_out->comp_mask = 0; qp_out->dbrec = mqp->db; if (mqp->sq_buf_size) /* IBV_QPT_RAW_PACKET */ qp_out->sq.buf = (void *)((uintptr_t)mqp->sq_buf.buf); else qp_out->sq.buf = (void *)((uintptr_t)mqp->buf.buf + mqp->sq.offset); qp_out->sq.wqe_cnt = mqp->sq.wqe_cnt; qp_out->sq.stride = 1 << mqp->sq.wqe_shift; qp_out->rq.buf = (void *)((uintptr_t)mqp->buf.buf + mqp->rq.offset); qp_out->rq.wqe_cnt = mqp->rq.wqe_cnt; qp_out->rq.stride = 1 << mqp->rq.wqe_shift; qp_out->bf.reg = mqp->bf->reg; if (mqp->bf->uuarn > 0) qp_out->bf.size = mqp->bf->buf_size; else qp_out->bf.size = 0; return 0; } static int mlx5dv_get_cq(struct ibv_cq *cq_in, struct mlx5dv_cq *cq_out) { struct mlx5_cq *mcq = to_mcq(cq_in); struct mlx5_context *mctx = to_mctx(cq_in->context); cq_out->comp_mask = 0; cq_out->cqn = mcq->cqn; cq_out->cqe_cnt = mcq->ibv_cq.cqe + 1; cq_out->cqe_size = mcq->cqe_sz; cq_out->buf = mcq->active_buf->buf; cq_out->dbrec = mcq->dbrec; cq_out->uar = mctx->uar; mcq->flags |= MLX5_CQ_FLAGS_DV_OWNED; return 0; } static int mlx5dv_get_rwq(struct ibv_wq *wq_in, struct mlx5dv_rwq *rwq_out) { struct mlx5_rwq *mrwq = to_mrwq(wq_in); rwq_out->comp_mask = 0; rwq_out->buf = mrwq->pbuff; rwq_out->dbrec = mrwq->recv_db; rwq_out->wqe_cnt = mrwq->rq.wqe_cnt; rwq_out->stride = 1 << mrwq->rq.wqe_shift; return 0; } static int mlx5dv_get_srq(struct ibv_srq *srq_in, struct mlx5dv_srq *srq_out) { struct mlx5_srq *msrq; msrq = container_of(srq_in, struct mlx5_srq, vsrq.srq); srq_out->comp_mask = 0; srq_out->buf = msrq->buf.buf; srq_out->dbrec = msrq->db; srq_out->stride = 1 << msrq->wqe_shift; srq_out->head = msrq->head; srq_out->tail = msrq->tail; return 0; } int mlx5dv_init_obj(struct mlx5dv_obj *obj, uint64_t obj_type) { int ret = 0; if (obj_type & MLX5DV_OBJ_QP) ret = mlx5dv_get_qp(obj->qp.in, obj->qp.out); if (!ret && (obj_type & MLX5DV_OBJ_CQ)) ret = mlx5dv_get_cq(obj->cq.in, obj->cq.out); if (!ret && (obj_type & MLX5DV_OBJ_SRQ)) ret = mlx5dv_get_srq(obj->srq.in, obj->srq.out); if (!ret && (obj_type & MLX5DV_OBJ_RWQ)) ret = mlx5dv_get_rwq(obj->rwq.in, obj->rwq.out); return ret; } static void adjust_uar_info(struct mlx5_device *mdev, struct mlx5_context *context, struct mlx5_alloc_ucontext_resp resp) { if (!resp.log_uar_size && !resp.num_uars_per_page) { /* old kernel */ context->uar_size = mdev->page_size; context->num_uars_per_page = 1; return; } context->uar_size = 1 << resp.log_uar_size; context->num_uars_per_page = resp.num_uars_per_page; } static int mlx5_init_context(struct verbs_device *vdev, struct ibv_context *ctx, int cmd_fd) { struct mlx5_context *context; struct mlx5_alloc_ucontext req; struct mlx5_alloc_ucontext_resp resp; int i; int page_size; int tot_uuars; int low_lat_uuars; int gross_uuars; int j; off_t offset; struct mlx5_device *mdev; struct verbs_context *v_ctx; struct ibv_port_attr port_attr; struct ibv_device_attr_ex device_attr; int k; int bfi; int num_sys_page_map; mdev = to_mdev(&vdev->device); v_ctx = verbs_get_ctx(ctx); page_size = mdev->page_size; mlx5_single_threaded = single_threaded_app(); context = to_mctx(ctx); context->ibv_ctx.cmd_fd = cmd_fd; open_debug_file(context); set_debug_mask(); set_freeze_on_error(); if (gethostname(context->hostname, sizeof(context->hostname))) strcpy(context->hostname, "host_unknown"); tot_uuars = get_total_uuars(page_size); if (tot_uuars < 0) { errno = -tot_uuars; goto err_free; } low_lat_uuars = get_num_low_lat_uuars(tot_uuars); if (low_lat_uuars < 0) { errno = -low_lat_uuars; goto err_free; } if (low_lat_uuars > tot_uuars - 1) { errno = ENOMEM; goto err_free; } memset(&req, 0, sizeof(req)); memset(&resp, 0, sizeof(resp)); req.total_num_uuars = tot_uuars; req.num_low_latency_uuars = low_lat_uuars; req.cqe_version = MLX5_CQE_VERSION_V1; req.lib_caps |= MLX5_LIB_CAP_4K_UAR; if (mlx5_cmd_get_context(context, &req, sizeof(req), &resp, sizeof(resp))) goto err_free; context->max_num_qps = resp.qp_tab_size; context->bf_reg_size = resp.bf_reg_size; context->tot_uuars = resp.tot_uuars; context->low_lat_uuars = low_lat_uuars; context->cache_line_size = resp.cache_line_size; context->max_sq_desc_sz = resp.max_sq_desc_sz; context->max_rq_desc_sz = resp.max_rq_desc_sz; context->max_send_wqebb = resp.max_send_wqebb; context->num_ports = resp.num_ports; context->max_recv_wr = resp.max_recv_wr; context->max_srq_recv_wr = resp.max_srq_recv_wr; context->cqe_version = resp.cqe_version; if (context->cqe_version) { if (context->cqe_version == MLX5_CQE_VERSION_V1) mlx5_ctx_ops.poll_cq = mlx5_poll_cq_v1; else goto err_free; } adjust_uar_info(mdev, context, resp); gross_uuars = context->tot_uuars / MLX5_NUM_NON_FP_BFREGS_PER_UAR * NUM_BFREGS_PER_UAR; context->bfs = calloc(gross_uuars, sizeof(*context->bfs)); if (!context->bfs) { errno = ENOMEM; goto err_free; } context->cmds_supp_uhw = resp.cmds_supp_uhw; context->vendor_cap_flags = 0; pthread_mutex_init(&context->qp_table_mutex, NULL); pthread_mutex_init(&context->srq_table_mutex, NULL); pthread_mutex_init(&context->uidx_table_mutex, NULL); for (i = 0; i < MLX5_QP_TABLE_SIZE; ++i) context->qp_table[i].refcnt = 0; for (i = 0; i < MLX5_QP_TABLE_SIZE; ++i) context->uidx_table[i].refcnt = 0; context->db_list = NULL; pthread_mutex_init(&context->db_list_mutex, NULL); num_sys_page_map = context->tot_uuars / (context->num_uars_per_page * MLX5_NUM_NON_FP_BFREGS_PER_UAR); for (i = 0; i < num_sys_page_map; ++i) { offset = 0; set_command(MLX5_MMAP_GET_REGULAR_PAGES_CMD, &offset); set_index(i, &offset); context->uar[i] = mmap(NULL, page_size, PROT_WRITE, MAP_SHARED, cmd_fd, page_size * offset); if (context->uar[i] == MAP_FAILED) { context->uar[i] = NULL; goto err_free_bf; } } for (i = 0; i < num_sys_page_map; i++) { for (j = 0; j < context->num_uars_per_page; j++) { for (k = 0; k < NUM_BFREGS_PER_UAR; k++) { bfi = (i * context->num_uars_per_page + j) * NUM_BFREGS_PER_UAR + k; context->bfs[bfi].reg = context->uar[i] + MLX5_ADAPTER_PAGE_SIZE * j + MLX5_BF_OFFSET + k * context->bf_reg_size; context->bfs[bfi].need_lock = need_uuar_lock(context, bfi); mlx5_spinlock_init(&context->bfs[bfi].lock); context->bfs[bfi].offset = 0; if (bfi) context->bfs[bfi].buf_size = context->bf_reg_size / 2; context->bfs[bfi].uuarn = bfi; } } } context->hca_core_clock = NULL; if (resp.response_length + sizeof(resp.ibv_resp) >= offsetof(struct mlx5_alloc_ucontext_resp, hca_core_clock_offset) + sizeof(resp.hca_core_clock_offset) && resp.comp_mask & MLX5_IB_ALLOC_UCONTEXT_RESP_MASK_CORE_CLOCK_OFFSET) { context->core_clock.offset = resp.hca_core_clock_offset; mlx5_map_internal_clock(mdev, ctx); } mlx5_spinlock_init(&context->lock32); context->prefer_bf = get_always_bf(); context->shut_up_bf = get_shut_up_bf(); mlx5_read_env(&vdev->device, context); mlx5_spinlock_init(&context->hugetlb_lock); TAILQ_INIT(&context->hugetlb_list); context->ibv_ctx.ops = mlx5_ctx_ops; verbs_set_ctx_op(v_ctx, create_qp_ex, mlx5_create_qp_ex); verbs_set_ctx_op(v_ctx, open_xrcd, mlx5_open_xrcd); verbs_set_ctx_op(v_ctx, close_xrcd, mlx5_close_xrcd); verbs_set_ctx_op(v_ctx, create_srq_ex, mlx5_create_srq_ex); verbs_set_ctx_op(v_ctx, get_srq_num, mlx5_get_srq_num); verbs_set_ctx_op(v_ctx, query_device_ex, mlx5_query_device_ex); verbs_set_ctx_op(v_ctx, query_rt_values, mlx5_query_rt_values); verbs_set_ctx_op(v_ctx, ibv_create_flow, ibv_cmd_create_flow); verbs_set_ctx_op(v_ctx, ibv_destroy_flow, ibv_cmd_destroy_flow); verbs_set_ctx_op(v_ctx, create_cq_ex, mlx5_create_cq_ex); verbs_set_ctx_op(v_ctx, create_wq, mlx5_create_wq); verbs_set_ctx_op(v_ctx, modify_wq, mlx5_modify_wq); verbs_set_ctx_op(v_ctx, destroy_wq, mlx5_destroy_wq); verbs_set_ctx_op(v_ctx, create_rwq_ind_table, mlx5_create_rwq_ind_table); verbs_set_ctx_op(v_ctx, destroy_rwq_ind_table, mlx5_destroy_rwq_ind_table); memset(&device_attr, 0, sizeof(device_attr)); if (!mlx5_query_device_ex(ctx, NULL, &device_attr, sizeof(struct ibv_device_attr_ex))) { context->cached_device_cap_flags = device_attr.orig_attr.device_cap_flags; context->atomic_cap = device_attr.orig_attr.atomic_cap; context->cached_tso_caps = device_attr.tso_caps; } for (j = 0; j < min(MLX5_MAX_PORTS_NUM, context->num_ports); ++j) { memset(&port_attr, 0, sizeof(port_attr)); if (!mlx5_query_port(ctx, j + 1, &port_attr)) context->cached_link_layer[j] = port_attr.link_layer; } return 0; err_free_bf: free(context->bfs); err_free: for (i = 0; i < MLX5_MAX_UARS; ++i) { if (context->uar[i]) munmap(context->uar[i], page_size); } close_debug_file(context); return errno; } static void mlx5_cleanup_context(struct verbs_device *device, struct ibv_context *ibctx) { struct mlx5_context *context = to_mctx(ibctx); int page_size = to_mdev(ibctx->device)->page_size; int i; free(context->bfs); for (i = 0; i < MLX5_MAX_UARS; ++i) { if (context->uar[i]) munmap(context->uar[i], page_size); } if (context->hca_core_clock) munmap(context->hca_core_clock - context->core_clock.offset, page_size); close_debug_file(context); } static struct verbs_device_ops mlx5_dev_ops = { .init_context = mlx5_init_context, .uninit_context = mlx5_cleanup_context, }; static struct verbs_device *mlx5_driver_init(const char *uverbs_sys_path, int abi_version) { char value[8]; struct mlx5_device *dev; unsigned vendor, device; int i; if (ibv_read_sysfs_file(uverbs_sys_path, "device/vendor", value, sizeof value) < 0) return NULL; sscanf(value, "%i", &vendor); if (ibv_read_sysfs_file(uverbs_sys_path, "device/device", value, sizeof value) < 0) return NULL; sscanf(value, "%i", &device); for (i = 0; i < sizeof hca_table / sizeof hca_table[0]; ++i) if (vendor == hca_table[i].vendor && device == hca_table[i].device) goto found; return NULL; found: if (abi_version < MLX5_UVERBS_MIN_ABI_VERSION || abi_version > MLX5_UVERBS_MAX_ABI_VERSION) { fprintf(stderr, PFX "Fatal: ABI version %d of %s is not supported " "(min supported %d, max supported %d)\n", abi_version, uverbs_sys_path, MLX5_UVERBS_MIN_ABI_VERSION, MLX5_UVERBS_MAX_ABI_VERSION); return NULL; } dev = calloc(1, sizeof *dev); if (!dev) { fprintf(stderr, PFX "Fatal: couldn't allocate device for %s\n", uverbs_sys_path); return NULL; } dev->page_size = sysconf(_SC_PAGESIZE); dev->driver_abi_ver = abi_version; dev->verbs_dev.ops = &mlx5_dev_ops; dev->verbs_dev.sz = sizeof(*dev); dev->verbs_dev.size_of_context = sizeof(struct mlx5_context) - sizeof(struct ibv_context); return &dev->verbs_dev; } static __attribute__((constructor)) void mlx5_register_driver(void) { verbs_register_driver("mlx5", mlx5_driver_init); } diff --git a/lib/libc/gen/Makefile.inc b/lib/libc/gen/Makefile.inc index 6a9e284b5860..fd8a0feeb748 100644 --- a/lib/libc/gen/Makefile.inc +++ b/lib/libc/gen/Makefile.inc @@ -1,578 +1,580 @@ # @(#)Makefile.inc 8.6 (Berkeley) 5/4/95 # $FreeBSD$ # machine-independent gen sources .PATH: ${LIBC_SRCTOP}/${LIBC_ARCH}/gen ${LIBC_SRCTOP}/gen ${SRCTOP}/etc CONFS+= group master.passwd shells CONFSMODE_master.passwd= 600 CONFSPACKAGE= runtime SRCS+= __getosreldate.c \ __pthread_mutex_init_calloc_cb_stub.c \ __xuname.c \ _once_stub.c \ _pthread_stubs.c \ _rand48.c \ _spinlock_stub.c \ _thread_init.c \ alarm.c \ arc4random.c \ arc4random-compat.c \ arc4random_uniform.c \ assert.c \ auxv.c \ basename.c \ basename_compat.c \ cap_sandboxed.c \ check_utility_compat.c \ clock.c \ clock_getcpuclockid.c \ closedir.c \ confstr.c \ + cpuset_alloc.c \ + cpuset_free.c \ crypt.c \ ctermid.c \ daemon.c \ devname.c \ devname-compat11.c \ dirfd.c \ dirname.c \ dirname_compat.c \ disklabel.c \ dlfcn.c \ drand48.c \ dup3.c \ elf_utils.c \ erand48.c \ err.c \ errlst.c \ errno.c \ eventfd.c \ exec.c \ exect.c \ fdevname.c \ feature_present.c \ fmtcheck.c \ fmtmsg.c \ fnmatch.c \ fpclassify.c \ frexp.c \ fstab.c \ ftok.c \ fts.c \ fts-compat.c \ fts-compat11.c \ ftw.c \ ftw-compat11.c \ getbootfile.c \ getbsize.c \ getcap.c \ getcwd.c \ getdomainname.c \ getentropy.c \ getgrent.c \ getgrouplist.c \ gethostname.c \ getloadavg.c \ getlogin.c \ getmntinfo.c \ getmntinfo-compat11.c \ getnetgrent.c \ getosreldate.c \ getpagesize.c \ getpagesizes.c \ getpeereid.c \ getprogname.c \ getpwent.c \ getttyent.c \ getusershell.c \ getutxent.c \ getvfsbyname.c \ glob.c \ glob-compat11.c \ initgroups.c \ isatty.c \ isinf.c \ isnan.c \ jrand48.c \ lcong48.c \ libc_dlopen.c \ lockf.c \ lrand48.c \ memalign.c \ mrand48.c \ nftw.c \ nftw-compat11.c \ nice.c \ nlist.c \ nrand48.c \ opendir.c \ pause.c \ pmadvise.c \ popen.c \ posix_spawn.c \ psignal.c \ pututxline.c \ pw_scan.c \ raise.c \ readdir.c \ readdir-compat11.c \ readpassphrase.c \ recvmmsg.c \ rewinddir.c \ scandir.c \ scandir_b.c \ scandir-compat11.c \ sched_getaffinity.c \ sched_setaffinity.c \ seed48.c \ seekdir.c \ semctl.c \ sendmmsg.c \ setdomainname.c \ sethostname.c \ setjmperr.c \ setmode.c \ setproctitle.c \ setprogname.c \ siginterrupt.c \ siglist.c \ signal.c \ sigsetops.c \ sleep.c \ srand48.c \ statvfs.c \ stringlist.c \ strtofflags.c \ sysconf.c \ sysctl.c \ sysctlbyname.c \ sysctlnametomib.c \ syslog.c \ telldir.c \ termios.c \ time.c \ times.c \ timespec_get.c \ timezone.c \ tls.c \ ttyname.c \ ttyslot.c \ ualarm.c \ ulimit.c \ uname.c \ unvis-compat.c \ usleep.c \ utime.c \ utxdb.c \ valloc.c \ wait.c \ wait3.c \ waitpid.c \ waitid.c \ wordexp.c CFLAGS.arc4random.c= -I${SRCTOP}/sys -I${SRCTOP}/sys/crypto/chacha20 CFLAGS.dlfcn.c= ${RTLD_HDRS} CFLAGS.tls.c= ${RTLD_HDRS} .PATH: ${SRCTOP}/contrib/libc-pwcache SRCS+= pwcache.c pwcache.h .PATH: ${SRCTOP}/contrib/libc-vis CFLAGS+= -I${SRCTOP}/contrib/libc-vis SRCS+= unvis.c vis.c MISRCS+=modf.c CANCELPOINTS_SRCS=sem.c sem_new.c .for src in ${CANCELPOINTS_SRCS} SRCS+=cancelpoints_${src} CLEANFILES+=cancelpoints_${src} cancelpoints_${src}: ${LIBC_SRCTOP}/gen/${src} .NOMETA ln -sf ${.ALLSRC} ${.TARGET} .endfor SYM_MAPS+=${LIBC_SRCTOP}/gen/Symbol.map # machine-dependent gen sources .sinclude "${LIBC_SRCTOP}/${LIBC_ARCH}/gen/Makefile.inc" MAN+= alarm.3 \ arc4random.3 \ auxv.3 \ basename.3 \ cap_rights_get.3 \ cap_sandboxed.3 \ check_utility_compat.3 \ clock.3 \ clock_getcpuclockid.3 \ confstr.3 \ ctermid.3 \ daemon.3 \ devname.3 \ directory.3 \ dirname.3 \ dl_iterate_phdr.3 \ dladdr.3 \ dlinfo.3 \ dllockinit.3 \ dlopen.3 \ dup3.3 \ err.3 \ exec.3 \ feature_present.3 \ fmtcheck.3 \ fmtmsg.3 \ fnmatch.3 \ fpclassify.3 \ frexp.3 \ ftok.3 \ fts.3 \ ftw.3 \ getbootfile.3 \ getbsize.3 \ getcap.3 \ getcontext.3 \ getcwd.3 \ getdiskbyname.3 \ getdomainname.3 \ getentropy.3 \ getfsent.3 \ getgrent.3 \ getgrouplist.3 \ gethostname.3 \ getloadavg.3 \ getmntinfo.3 \ getnetgrent.3 \ getosreldate.3 \ getpagesize.3 \ getpagesizes.3 \ getpass.3 \ getpeereid.3 \ getprogname.3 \ getpwent.3 \ getttyent.3 \ getusershell.3 \ getutxent.3 \ getvfsbyname.3 \ glob.3 \ initgroups.3 \ isgreater.3 \ ldexp.3 \ lockf.3 \ makecontext.3 \ modf.3 \ nice.3 \ nlist.3 \ pause.3 \ popen.3 \ posix_spawn.3 \ posix_spawn_file_actions_addopen.3 \ posix_spawn_file_actions_init.3 \ posix_spawnattr_getflags.3 \ posix_spawnattr_getpgroup.3 \ posix_spawnattr_getschedparam.3 \ posix_spawnattr_getschedpolicy.3 \ posix_spawnattr_init.3 \ posix_spawnattr_getsigdefault.3 \ posix_spawnattr_getsigmask.3 \ psignal.3 \ pwcache.3 \ raise.3 \ rand48.3 \ readpassphrase.3 \ rfork_thread.3 \ scandir.3 \ sem_destroy.3 \ sem_getvalue.3 \ sem_init.3 \ sem_open.3 \ sem_post.3 \ sem_timedwait.3 \ sem_wait.3 \ setjmp.3 \ setmode.3 \ setproctitle.3 \ siginterrupt.3 \ signal.3 \ sigsetops.3 \ sleep.3 \ statvfs.3 \ stringlist.3 \ strtofflags.3 \ sysconf.3 \ sysctl.3 \ syslog.3 \ tcgetpgrp.3 \ tcgetsid.3 \ tcgetwinsize.3 \ tcsendbreak.3 \ tcsetattr.3 \ tcsetpgrp.3 \ tcsetsid.3 \ time.3 \ times.3 \ timespec_get.3 \ timezone.3 \ ttyname.3 \ tzset.3 \ ualarm.3 \ ucontext.3 \ ulimit.3 \ uname.3 \ unvis.3 \ usleep.3 \ utime.3 \ valloc.3 \ vis.3 \ wordexp.3 MLINKS+=arc4random.3 arc4random_buf.3 \ arc4random.3 arc4random_uniform.3 MLINKS+=auxv.3 elf_aux_info.3 MLINKS+=ctermid.3 ctermid_r.3 MLINKS+=daemon.3 daemonfd.3 MLINKS+=devname.3 devname_r.3 MLINKS+=devname.3 fdevname.3 MLINKS+=devname.3 fdevname_r.3 MLINKS+=directory.3 closedir.3 \ directory.3 dirfd.3 \ directory.3 fdclosedir.3 \ directory.3 fdopendir.3 \ directory.3 opendir.3 \ directory.3 readdir.3 \ directory.3 readdir_r.3 \ directory.3 rewinddir.3 \ directory.3 seekdir.3 \ directory.3 telldir.3 MLINKS+=dlopen.3 fdlopen.3 \ dlopen.3 dlclose.3 \ dlopen.3 dlerror.3 \ dlopen.3 dlfunc.3 \ dlopen.3 dlsym.3 \ dlopen.3 dlvsym.3 MLINKS+=err.3 err_set_exit.3 \ err.3 err_set_file.3 \ err.3 errc.3 \ err.3 errx.3 \ err.3 verr.3 \ err.3 verrc.3 \ err.3 verrx.3 \ err.3 vwarn.3 \ err.3 vwarnc.3 \ err.3 vwarnx.3 \ err.3 warnc.3 \ err.3 warn.3 \ err.3 warnx.3 MLINKS+=exec.3 execl.3 \ exec.3 execle.3 \ exec.3 execlp.3 \ exec.3 exect.3 \ exec.3 execv.3 \ exec.3 execvP.3 \ exec.3 execvp.3 MLINKS+=fpclassify.3 finite.3 \ fpclassify.3 finitef.3 \ fpclassify.3 isfinite.3 \ fpclassify.3 isinf.3 \ fpclassify.3 isnan.3 \ fpclassify.3 isnormal.3 MLINKS+=frexp.3 frexpf.3 \ frexp.3 frexpl.3 MLINKS+=fts.3 fts_children.3 \ fts.3 fts_close.3 \ fts.3 fts_open.3 \ fts.3 fts_read.3 \ fts.3 fts_set.3 \ fts.3 fts_set_clientptr.3 \ fts.3 fts_get_clientptr.3 \ fts.3 fts_get_stream.3 MLINKS+=ftw.3 nftw.3 MLINKS+=getcap.3 cgetcap.3 \ getcap.3 cgetclose.3 \ getcap.3 cgetent.3 \ getcap.3 cgetfirst.3 \ getcap.3 cgetmatch.3 \ getcap.3 cgetnext.3 \ getcap.3 cgetnum.3 \ getcap.3 cgetset.3 \ getcap.3 cgetstr.3 \ getcap.3 cgetustr.3 MLINKS+=getcwd.3 getwd.3 MLINKS+=getcontext.3 getcontextx.3 MLINKS+=getcontext.3 setcontext.3 MLINKS+=getdomainname.3 setdomainname.3 MLINKS+=getfsent.3 endfsent.3 \ getfsent.3 getfsfile.3 \ getfsent.3 getfsspec.3 \ getfsent.3 getfstype.3 \ getfsent.3 setfsent.3 \ getfsent.3 setfstab.3 \ getfsent.3 getfstab.3 MLINKS+=getgrent.3 endgrent.3 \ getgrent.3 getgrgid.3 \ getgrent.3 getgrnam.3 \ getgrent.3 setgrent.3 \ getgrent.3 setgroupent.3 \ getgrent.3 getgrent_r.3 \ getgrent.3 getgrnam_r.3 \ getgrent.3 getgrgid_r.3 MLINKS+=gethostname.3 sethostname.3 MLINKS+=getnetgrent.3 endnetgrent.3 \ getnetgrent.3 getnetgrent_r.3 \ getnetgrent.3 innetgr.3 \ getnetgrent.3 setnetgrent.3 MLINKS+=getprogname.3 setprogname.3 MLINKS+=getpwent.3 endpwent.3 \ getpwent.3 getpwnam.3 \ getpwent.3 getpwuid.3 \ getpwent.3 setpassent.3 \ getpwent.3 setpwent.3 \ getpwent.3 setpwfile.3 \ getpwent.3 getpwent_r.3 \ getpwent.3 getpwnam_r.3 \ getpwent.3 getpwuid_r.3 MLINKS+=getttyent.3 endttyent.3 \ getttyent.3 getttynam.3 \ getttyent.3 isdialuptty.3 \ getttyent.3 isnettty.3 \ getttyent.3 setttyent.3 MLINKS+=getusershell.3 endusershell.3 \ getusershell.3 setusershell.3 MLINKS+=getutxent.3 endutxent.3 \ getutxent.3 getutxid.3 \ getutxent.3 getutxline.3 \ getutxent.3 getutxuser.3 \ getutxent.3 pututxline.3 \ getutxent.3 setutxdb.3 \ getutxent.3 setutxent.3 \ getutxent.3 utmpx.3 MLINKS+=glob.3 globfree.3 MLINKS+=isgreater.3 isgreaterequal.3 \ isgreater.3 isless.3 \ isgreater.3 islessequal.3 \ isgreater.3 islessgreater.3 \ isgreater.3 isunordered.3 MLINKS+=ldexp.3 ldexpf.3 \ ldexp.3 ldexpl.3 MLINKS+=makecontext.3 swapcontext.3 MLINKS+=modf.3 modff.3 \ modf.3 modfl.3 MLINKS+=popen.3 pclose.3 MLINKS+=posix_spawn.3 posix_spawnp.3 \ posix_spawn_file_actions_addopen.3 posix_spawn_file_actions_addclose.3 \ posix_spawn_file_actions_addopen.3 posix_spawn_file_actions_addclosefrom_np.3 \ posix_spawn_file_actions_addopen.3 posix_spawn_file_actions_adddup2.3 \ posix_spawn_file_actions_addopen.3 posix_spawn_file_actions_addchdir_np.3 \ posix_spawn_file_actions_addopen.3 posix_spawn_file_actions_addfchdir_np.3 \ posix_spawn_file_actions_init.3 posix_spawn_file_actions_destroy.3 \ posix_spawnattr_getflags.3 posix_spawnattr_setflags.3 \ posix_spawnattr_getpgroup.3 posix_spawnattr_setpgroup.3 \ posix_spawnattr_getschedparam.3 posix_spawnattr_setschedparam.3 \ posix_spawnattr_getschedpolicy.3 posix_spawnattr_setschedpolicy.3 \ posix_spawnattr_getsigdefault.3 posix_spawnattr_setsigdefault.3 \ posix_spawnattr_getsigmask.3 posix_spawnattr_setsigmask.3 \ posix_spawnattr_init.3 posix_spawnattr_destroy.3 MLINKS+=psignal.3 strsignal.3 \ psignal.3 sys_siglist.3 \ psignal.3 sys_signame.3 MLINKS+=pwcache.3 group_from_gid.3 \ pwcache.3 user_from_uid.3 MLINKS+=rand48.3 _rand48.3 \ rand48.3 drand48.3 \ rand48.3 erand48.3 \ rand48.3 jrand48.3 \ rand48.3 lcong48.3 \ rand48.3 lrand48.3 \ rand48.3 mrand48.3 \ rand48.3 nrand48.3 \ rand48.3 seed48.3 \ rand48.3 srand48.3 MLINKS+=recv.2 recvmmsg.2 MLINKS+=scandir.3 alphasort.3 MLINKS+=sem_open.3 sem_close.3 \ sem_open.3 sem_unlink.3 MLINKS+=sem_wait.3 sem_trywait.3 MLINKS+=sem_timedwait.3 sem_clockwait_np.3 MLINKS+=send.2 sendmmsg.2 MLINKS+=setjmp.3 _longjmp.3 \ setjmp.3 _setjmp.3 \ setjmp.3 longjmp.3 \ setjmp.3 longjmperr.3 \ setjmp.3 longjmperror.3 \ setjmp.3 siglongjmp.3 \ setjmp.3 sigsetjmp.3 MLINKS+=setmode.3 getmode.3 MLINKS+=setproctitle.3 setproctitle_fast.3 MLINKS+=sigsetops.3 sigaddset.3 \ sigsetops.3 sigandset.3 \ sigsetops.3 sigdelset.3 \ sigsetops.3 sigemptyset.3 \ sigsetops.3 sigfillset.3 \ sigsetops.3 sigisemptyset.3 \ sigsetops.3 sigismember.3 \ sigsetops.3 sigorset.3 MLINKS+=statvfs.3 fstatvfs.3 MLINKS+=stringlist.3 sl_add.3 \ stringlist.3 sl_find.3 \ stringlist.3 sl_free.3 \ stringlist.3 sl_init.3 MLINKS+=strtofflags.3 fflagstostr.3 MLINKS+=sysctl.3 sysctlbyname.3 \ sysctl.3 sysctlnametomib.3 MLINKS+=syslog.3 closelog.3 \ syslog.3 openlog.3 \ syslog.3 setlogmask.3 \ syslog.3 vsyslog.3 MLINKS+=tcgetwinsize.3 tcsetwinsize.3 MLINKS+=tcsendbreak.3 tcdrain.3 \ tcsendbreak.3 tcflow.3 \ tcsendbreak.3 tcflush.3 MLINKS+=tcsetattr.3 cfgetispeed.3 \ tcsetattr.3 cfgetospeed.3 \ tcsetattr.3 cfmakeraw.3 \ tcsetattr.3 cfmakesane.3 \ tcsetattr.3 cfsetispeed.3 \ tcsetattr.3 cfsetospeed.3 \ tcsetattr.3 cfsetspeed.3 \ tcsetattr.3 tcgetattr.3 MLINKS+=ttyname.3 isatty.3 \ ttyname.3 ttyname_r.3 MLINKS+=tzset.3 tzsetwall.3 MLINKS+=unvis.3 strunvis.3 \ unvis.3 strunvisx.3 MLINKS+=vis.3 nvis.3 \ vis.3 snvis.3 \ vis.3 strenvisx.3 \ vis.3 strnunvis.3 \ vis.3 strnunvisx.3 \ vis.3 strnvis.3 \ vis.3 strnvisx.3 \ vis.3 strsenvisx.3 \ vis.3 strsnvis.3 \ vis.3 strsnvisx.3 \ vis.3 strsvis.3 \ vis.3 strsvisx.3 \ vis.3 strvis.3 \ vis.3 strvisx.3 \ vis.3 svis.3 MLINKS+=wordexp.3 wordfree.3 .include afterinstallconfig: install-passwd install-passwd: .PHONY .if ${MK_TCSH} == "no" sed -i "" -e 's;/bin/csh;/bin/sh;' ${DESTDIR}/etc/master.passwd .endif ${PWD_MKDB_CMD} -i -p -d ${DESTDIR}/etc ${DESTDIR}/etc/master.passwd .if defined(NO_ROOT) && defined(METALOG) ( \ echo "./etc/pwd.db type=file mode=0644 uname=root gname=wheel"; \ echo "./etc/spwd.db type=file mode=0600 uname=root gname=wheel"; \ echo "./etc/passwd type=file mode=0644 uname=root gname=wheel"; \ ) | cat -l >> ${METALOG} .endif diff --git a/lib/libc/gen/Symbol.map b/lib/libc/gen/Symbol.map index 5554c32ab4f2..012c9ff11f04 100644 --- a/lib/libc/gen/Symbol.map +++ b/lib/libc/gen/Symbol.map @@ -1,572 +1,574 @@ /* * $FreeBSD$ */ FBSD_1.0 { __xuname; pthread_atfork; pthread_attr_destroy; pthread_attr_get_np; pthread_attr_getdetachstate; pthread_attr_getguardsize; pthread_attr_getinheritsched; pthread_attr_getschedparam; pthread_attr_getschedpolicy; pthread_attr_getscope; pthread_attr_getstackaddr; pthread_attr_getstacksize; pthread_attr_init; pthread_attr_setdetachstate; pthread_attr_setguardsize; pthread_attr_setinheritsched; pthread_attr_setschedparam; pthread_attr_setschedpolicy; pthread_attr_setscope; pthread_attr_setstackaddr; pthread_attr_setstacksize; pthread_cancel; pthread_cleanup_pop; pthread_cleanup_push; pthread_cond_broadcast; pthread_cond_destroy; pthread_cond_init; pthread_cond_signal; pthread_cond_timedwait; pthread_cond_wait; pthread_detach; pthread_equal; pthread_exit; pthread_getspecific; pthread_join; pthread_key_create; pthread_key_delete; pthread_kill; pthread_main_np; pthread_mutex_destroy; pthread_mutex_init; pthread_mutex_lock; pthread_mutex_trylock; pthread_mutex_unlock; pthread_mutexattr_destroy; pthread_mutexattr_init; pthread_mutexattr_settype; pthread_once; pthread_rwlock_destroy; pthread_rwlock_init; pthread_rwlock_rdlock; pthread_rwlock_tryrdlock; pthread_rwlock_trywrlock; pthread_rwlock_unlock; pthread_rwlock_wrlock; pthread_self; pthread_setcancelstate; pthread_setcanceltype; pthread_setspecific; pthread_sigmask; pthread_testcancel; alarm; arc4random; __assert; check_utility_compat; clock; closedir; confstr; ctermid; ctermid_r; daemon; getdiskbyname; dladdr; dlclose; dlerror; dlfunc; dllockinit; dlopen; dlsym; dlvsym; dlinfo; dl_iterate_phdr; drand48; erand48; err_set_file; err_set_exit; err; verr; errc; verrc; errx; verrx; warn; vwarn; warnc; vwarnc; warnx; vwarnx; sys_errlist; sys_nerr; errno; exect; execl; execle; execlp; execv; execvp; execvP; fmtcheck; fmtmsg; fnmatch; __fpclassifyf; __fpclassifyd; __fpclassifyl; frexp; setfstab; getfstab; getfsent; getfsspec; getfsfile; setfsent; endfsent; ftok; getbootfile; getbsize; cgetset; cgetcap; cgetent; cgetmatch; cgetfirst; cgetclose; cgetnext; cgetstr; cgetustr; cgetnum; getcwd; getdomainname; setgrent; setgroupent; endgrent; getgrent_r; getgrnam_r; getgrgid_r; getgrnam; getgrgid; getgrent; /* * Why are __gr_parse_entry() and __gr_match_entry() not static in * gen/getgrent.c? */ getgrouplist; gethostname; getloadavg; getlogin; setnetgrent; getnetgrent; endnetgrent; innetgr; getosreldate; getpagesize; getpeereid; _getprogname; getprogname; setpwent; setpassent; endpwent; getpwent_r; getpwnam_r; getpwuid_r; getpwnam; getpwuid; getpwent; getttynam; getttyent; setttyent; endttyent; isdialuptty; isnettty; getusershell; endusershell; setusershell; getvfsbyname; __isnan; isnan; __isnanf; isnanf; __isinf; isinf; __isinff; __isinfl; isatty; initgroups; jrand48; lcong48; ldexp; lockf; lrand48; modf; mrand48; nice; nlist; nrand48; opendir; pause; posix_madvise; popen; pclose; psignal; raise; readpassphrase; getpass; rewinddir; seed48; seekdir; user_from_uid; group_from_gid; setdomainname; sethostname; longjmperror; getmode; setmode; setproctitle; setprogname; siginterrupt; sys_signame; sys_siglist; sys_nsig; signal; sigaddset; sigdelset; sigemptyset; sigfillset; sigismember; sleep; srand48; fstatvfs; statvfs; sl_init; sl_add; sl_free; sl_find; fflagstostr; strtofflags; sysconf; sysctl; sysctlbyname; sysctlnametomib; syslog; vsyslog; openlog; closelog; setlogmask; ttyname_r; ttyname; timezone; times; time; telldir; tcgetattr; tcsetattr; tcsetpgrp; tcgetpgrp; cfgetospeed; cfgetispeed; cfsetospeed; cfsetispeed; cfsetspeed; cfmakeraw; tcsendbreak; _init_tls; __tls_get_addr; tcdrain; tcflush; tcflow; ualarm; ulimit; uname; strunvis; strunvisx; usleep; utime; valloc; vis; strvis; strvisx; wait; wait3; waitpid; wordexp; wordfree; }; FBSD_1.1 { arc4random_buf; arc4random_uniform; fdevname; fdevname_r; fdopendir; feature_present; posix_spawn; posix_spawn_file_actions_addclose; posix_spawn_file_actions_adddup2; posix_spawn_file_actions_addopen; posix_spawn_file_actions_destroy; posix_spawn_file_actions_init; posix_spawnattr_destroy; posix_spawnattr_getflags; posix_spawnattr_getpgroup; posix_spawnattr_getschedparam; posix_spawnattr_getschedpolicy; posix_spawnattr_getsigdefault; posix_spawnattr_getsigmask; posix_spawnattr_init; posix_spawnattr_setflags; posix_spawnattr_setpgroup; posix_spawnattr_setschedparam; posix_spawnattr_setschedpolicy; posix_spawnattr_setsigdefault; posix_spawnattr_setsigmask; posix_spawnp; semctl; tcgetsid; tcsetsid; __pthread_cleanup_pop_imp; __pthread_cleanup_push_imp; }; FBSD_1.2 { cfmakesane; endutxent; getpagesizes; getutxent; getutxid; getutxline; getutxuser; pthread_getthreadid_np; pututxline; sem_close; sem_destroy; sem_getvalue; sem_init; sem_open; sem_post; sem_timedwait; sem_trywait; sem_unlink; sem_wait; setutxdb; setutxent; }; FBSD_1.3 { clock_getcpuclockid; dirfd; dup3; fdclosedir; fdlopen; __FreeBSD_libc_enter_restricted_mode; getcontextx; gid_from_group; nvis; pwcache_userdb; pwcache_groupdb; snvis; strenvisx; strnunvis; strnunvisx; strnvis; strnvisx; strsenvisx; strsnvis; strsnvisx; strsvis; strsvisx; svis; uid_from_user; unvis; waitid; }; FBSD_1.4 { getnetgrent_r; pthread_mutex_consistent; pthread_mutexattr_getrobust; pthread_mutexattr_setrobust; stravis; }; FBSD_1.5 { alphasort; basename; daemonfd; devname; devname_r; dirname; elf_aux_info; fts_children; fts_close; fts_get_clientptr; fts_get_stream; fts_open; fts_read; fts_set; fts_set_clientptr; ftw; getentropy; getmntinfo; glob; globfree; nftw; readdir; readdir_r; scandir; sem_clockwait_np; setproctitle_fast; timespec_get; }; FBSD_1.6 { eventfd; eventfd_read; eventfd_write; getlogin_r; memalign; scandir_b; sigandset; sigisemptyset; sigorset; tcgetwinsize; tcsetwinsize; }; FBSD_1.7 { posix_spawn_file_actions_addchdir_np; posix_spawn_file_actions_addclosefrom_np; posix_spawn_file_actions_addfchdir_np; sched_getaffinity; sched_setaffinity; sched_getcpu; + __cpuset_alloc; + __cpuset_free; }; FBSDprivate_1.0 { /* needed by thread libraries */ __thr_jtable; _pthread_atfork; _pthread_attr_destroy; _pthread_attr_getdetachstate; _pthread_attr_getguardsize; _pthread_attr_getinheritsched; _pthread_attr_getschedparam; _pthread_attr_getschedpolicy; _pthread_attr_getscope; _pthread_attr_getstackaddr; _pthread_attr_getstacksize; _pthread_attr_init; _pthread_attr_setdetachstate; _pthread_attr_setguardsize; _pthread_attr_setinheritsched; _pthread_attr_setschedparam; _pthread_attr_setschedpolicy; _pthread_attr_setscope; _pthread_attr_setstackaddr; _pthread_attr_setstacksize; _pthread_cancel; _pthread_cancel_enter; _pthread_cancel_leave; _pthread_cleanup_pop; _pthread_cleanup_push; _pthread_cond_broadcast; _pthread_cond_destroy; _pthread_cond_init; _pthread_cond_signal; _pthread_cond_timedwait; _pthread_cond_wait; _pthread_detach; _pthread_equal; _pthread_exit; _pthread_getspecific; _pthread_join; _pthread_key_create; _pthread_key_delete; _pthread_kill; _pthread_main_np; _pthread_mutex_destroy; _pthread_mutex_init_calloc_cb; _pthread_mutex_init; _pthread_mutex_lock; _pthread_mutex_trylock; _pthread_mutex_unlock; _pthread_mutexattr_destroy; _pthread_mutexattr_init; _pthread_mutexattr_settype; _pthread_once; _pthread_rwlock_destroy; _pthread_rwlock_init; _pthread_rwlock_rdlock; _pthread_rwlock_tryrdlock; _pthread_rwlock_trywrlock; _pthread_rwlock_unlock; _pthread_rwlock_wrlock; _pthread_self; _pthread_setcancelstate; _pthread_setcanceltype; _pthread_setspecific; _pthread_sigmask; _pthread_testcancel; _spinlock; _spinunlock; _rtld_addr_phdr; _rtld_atfork_pre; _rtld_atfork_post; _rtld_error; /* for private use */ _rtld_get_stack_prot; _rtld_is_dlopened; _rtld_thread_init; /* for private use */ __elf_phdr_match_addr; _err; _warn; __fmtcheck; /* __pw_match_entry; */ /* __pw_parse_entry; */ __fdnlist; /* used by libkvm */ /* __elf_is_okay__; */ /* __elf_fdnlist; */ __opendir2; __pause; _pause; __raise; _raise; __sleep; _sleep; _rtld_allocate_tls; _rtld_free_tls; #if defined(i386) ___libc_tls_get_addr; /* x86 only */ #endif __libc_tls_get_addr; __tcdrain; _tcdrain; __usleep; _usleep; __wait; _wait; __waitpid; _waitpid; _libc_sem_init_compat; _libc_sem_destroy_compat; _libc_sem_open_compat; _libc_sem_close_compat; _libc_sem_unlink_compat; _libc_sem_wait_compat; _libc_sem_trywait_compat; _libc_sem_timedwait_compat; _libc_sem_post_compat; _libc_sem_getvalue_compat; __libc_tcdrain; __elf_aux_vector; __pthread_distribute_static_tls; __pthread_map_stacks_exec; __fillcontextx; __fillcontextx2; __getcontextx_size; }; diff --git a/lib/libc/gen/sched_getcpu_gen.c b/lib/libc/gen/cpuset_alloc.c similarity index 82% copy from lib/libc/gen/sched_getcpu_gen.c copy to lib/libc/gen/cpuset_alloc.c index 704075c0ae53..ac307f07f263 100644 --- a/lib/libc/gen/sched_getcpu_gen.c +++ b/lib/libc/gen/cpuset_alloc.c @@ -1,37 +1,33 @@ /*- - * Copyright (c) 2021 The FreeBSD Foundation - * - * This software were developed by Konstantin Belousov - * under sponsorship from the FreeBSD Foundation. + * Copyright (c) 2021 Stefan Esser * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ -#define _WITH_CPU_SET_T +#include #include -#include "libc_private.h" -int -sched_getcpu(void) +cpuset_t * +__cpuset_alloc(size_t ncpus) { - return (__sys_sched_getcpu()); + return (malloc(CPU_ALLOC_SIZE(ncpus))); } diff --git a/lib/libc/gen/sched_getcpu_gen.c b/lib/libc/gen/cpuset_free.c similarity index 82% copy from lib/libc/gen/sched_getcpu_gen.c copy to lib/libc/gen/cpuset_free.c index 704075c0ae53..4cd993853241 100644 --- a/lib/libc/gen/sched_getcpu_gen.c +++ b/lib/libc/gen/cpuset_free.c @@ -1,37 +1,33 @@ /*- - * Copyright (c) 2021 The FreeBSD Foundation - * - * This software were developed by Konstantin Belousov - * under sponsorship from the FreeBSD Foundation. + * Copyright (c) 2021 Stefan Esser * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ -#define _WITH_CPU_SET_T +#include #include -#include "libc_private.h" -int -sched_getcpu(void) +void +__cpuset_free(cpuset_t *ptr) { - return (__sys_sched_getcpu()); + free(ptr); } diff --git a/lib/libc/gen/sched_getaffinity.c b/lib/libc/gen/sched_getaffinity.c index 7fca53e39b4b..2ae8c5b763a3 100644 --- a/lib/libc/gen/sched_getaffinity.c +++ b/lib/libc/gen/sched_getaffinity.c @@ -1,37 +1,36 @@ /*- * Copyright (c) 2021 The FreeBSD Foundation * * This software were developed by Konstantin Belousov * under sponsorship from the FreeBSD Foundation. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ -#define _WITH_CPU_SET_T #include int sched_getaffinity(pid_t pid, size_t cpusetsz, cpuset_t *cpuset) { return (cpuset_getaffinity(CPU_LEVEL_WHICH, CPU_WHICH_PID, pid == 0 ? -1 : pid, cpusetsz, cpuset)); } diff --git a/lib/libc/gen/sched_getcpu_gen.c b/lib/libc/gen/sched_getcpu_gen.c index 704075c0ae53..ef1bebc2915a 100644 --- a/lib/libc/gen/sched_getcpu_gen.c +++ b/lib/libc/gen/sched_getcpu_gen.c @@ -1,37 +1,36 @@ /*- * Copyright (c) 2021 The FreeBSD Foundation * * This software were developed by Konstantin Belousov * under sponsorship from the FreeBSD Foundation. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ -#define _WITH_CPU_SET_T #include #include "libc_private.h" int sched_getcpu(void) { return (__sys_sched_getcpu()); } diff --git a/lib/libc/gen/sched_setaffinity.c b/lib/libc/gen/sched_setaffinity.c index 1c083b4b108a..52fc6d4bc811 100644 --- a/lib/libc/gen/sched_setaffinity.c +++ b/lib/libc/gen/sched_setaffinity.c @@ -1,37 +1,36 @@ /*- * Copyright (c) 2021 The FreeBSD Foundation * * This software were developed by Konstantin Belousov * under sponsorship from the FreeBSD Foundation. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ -#define _WITH_CPU_SET_T #include int sched_setaffinity(pid_t pid, size_t cpusetsz, const cpuset_t *cpuset) { return (cpuset_setaffinity(CPU_LEVEL_WHICH, CPU_WHICH_PID, pid == 0 ? -1 : pid, cpusetsz, cpuset)); } diff --git a/lib/libc/x86/sys/sched_getcpu_x86.c b/lib/libc/x86/sys/sched_getcpu_x86.c index 5e3bd5311734..13ba18ef5e36 100644 --- a/lib/libc/x86/sys/sched_getcpu_x86.c +++ b/lib/libc/x86/sys/sched_getcpu_x86.c @@ -1,82 +1,81 @@ /*- * Copyright (c) 2021 The FreeBSD Foundation * * This software were developed by Konstantin Belousov * under sponsorship from the FreeBSD Foundation. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include #include #include #include #include #include -#define _WITH_CPU_SET_T #include #include "libc_private.h" static int sched_getcpu_sys(void) { return (__sys_sched_getcpu()); } static int sched_getcpu_rdpid(void) { register_t res; __asm("rdpid %0" : "=r" (res)); return ((int)res); } static int sched_getcpu_rdtscp(void) { int res; __asm("rdtscp" : "=c" (res) : : "eax", "edx"); return (res); } DEFINE_UIFUNC(, int, sched_getcpu, (void)) { u_int amd_feature, cpu_exthigh, p[4]; if ((cpu_stdext_feature2 & CPUID_STDEXT2_RDPID) != 0) return (sched_getcpu_rdpid); amd_feature = 0; if (cpu_feature != 0) { do_cpuid(0x80000000, p); cpu_exthigh = p[0]; if (cpu_exthigh >= 0x80000001) { do_cpuid(0x80000001, p); amd_feature = p[3]; } } return ((amd_feature & AMDID_RDTSCP) == 0 ? sched_getcpu_sys : sched_getcpu_rdtscp); } diff --git a/share/man/man9/cpuset.9 b/share/man/man9/cpuset.9 index 93cd6fb7f115..92b759b438a3 100644 --- a/share/man/man9/cpuset.9 +++ b/share/man/man9/cpuset.9 @@ -1,352 +1,354 @@ .\" Copyright (c) 2015 Conrad Meyer .\" All rights reserved. .\" .\" Redistribution and use in source and binary forms, with or without .\" modification, are permitted provided that the following conditions .\" are met: .\" 1. Redistributions of source code must retain the above copyright .\" notice, this list of conditions and the following disclaimer. .\" 2. Redistributions in binary form must reproduce the above copyright .\" notice, this list of conditions and the following disclaimer in the .\" documentation and/or other materials provided with the distribution. .\" .\" THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' .\" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED .\" TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR .\" PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE .\" LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR .\" CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF .\" SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS .\" INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN .\" CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) .\" ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE .\" POSSIBILITY OF SUCH DAMAGE. .\" .\" $FreeBSD$ .\" .Dd December 12, 2019 .Dt CPUSET 9 .Os .Sh NAME .Nm cpuset(9) \(em .Nm CPUSET_T_INITIALIZER , .Nm CPUSET_FSET , .Nm CPU_CLR , .Nm CPU_COPY , .Nm CPU_ISSET , .Nm CPU_SET , .Nm CPU_ZERO , .Nm CPU_FILL , .Nm CPU_SETOF , .Nm CPU_EMPTY , .Nm CPU_ISFULLSET , .Nm CPU_FFS , .Nm CPU_COUNT , .Nm CPU_SUBSET , .Nm CPU_OVERLAP , .Nm CPU_CMP , .Nm CPU_OR , .Nm CPU_AND , .Nm CPU_ANDNOT , +.Nm CPU_XOR , .Nm CPU_CLR_ATOMIC , .Nm CPU_SET_ATOMIC , .Nm CPU_SET_ATOMIC_ACQ , .Nm CPU_AND_ATOMIC , .Nm CPU_OR_ATOMIC , .Nm CPU_COPY_STORE_REL .Nd cpuset manipulation macros .Sh SYNOPSIS .In sys/_cpuset.h .In sys/cpuset.h .\" .Fn CPUSET_T_INITIALIZER "ARRAY_CONTENTS" .Vt CPUSET_FSET .\" .Fn CPU_CLR "size_t cpu_idx" "cpuset_t *cpuset" .Fn CPU_COPY "cpuset_t *from" "cpuset_t *to" .Ft bool .Fn CPU_ISSET "size_t cpu_idx" "cpuset_t *cpuset" .Fn CPU_SET "size_t cpu_idx" "cpuset_t *cpuset" .Fn CPU_ZERO "cpuset_t *cpuset" .Fn CPU_FILL "cpuset_t *cpuset" .Fn CPU_SETOF "size_t cpu_idx" "cpuset_t *cpuset" .Ft bool .Fn CPU_EMPTY "cpuset_t *cpuset" .Ft bool .Fn CPU_ISFULLSET "cpuset_t *cpuset" .Ft int .Fn CPU_FFS "cpuset_t *cpuset" .Ft int .Fn CPU_COUNT "cpuset_t *cpuset" .\" .Ft bool .Fn CPU_SUBSET "cpuset_t *haystack" "cpuset_t *needle" .Ft bool .Fn CPU_OVERLAP "cpuset_t *cpuset1" "cpuset_t *cpuset2" .Ft bool .Fn CPU_CMP "cpuset_t *cpuset1" "cpuset_t *cpuset2" -.Fn CPU_OR "cpuset_t *dst" "cpuset_t *src" -.Fn CPU_AND "cpuset_t *dst" "cpuset_t *src" -.Fn CPU_ANDNOT "cpuset_t *dst" "cpuset_t *src" +.Fn CPU_OR "cpuset_t *dst" "cpuset_t *src1" "cpuset_t *src2" +.Fn CPU_AND "cpuset_t *dst" "cpuset_t *src1" "cpuset_t *src2" +.Fn CPU_ANDNOT "cpuset_t *dst" "cpuset_t *src1" "cpuset_t *src2" +.Fn CPU_XOR "cpuset_t *dst" "cpuset_t *src1" "cpuset_t *src2" .\" .Fn CPU_CLR_ATOMIC "size_t cpu_idx" "cpuset_t *cpuset" .Fn CPU_SET_ATOMIC "size_t cpu_idx" "cpuset_t *cpuset" .Fn CPU_SET_ATOMIC_ACQ "size_t cpu_idx" "cpuset_t *cpuset" .\" .Fn CPU_AND_ATOMIC "cpuset_t *dst" "cpuset_t *src" .Fn CPU_OR_ATOMIC "cpuset_t *dst" "cpuset_t *src" .Fn CPU_COPY_STORE_REL "cpuset_t *from" "cpuset_t *to" .Sh DESCRIPTION The .Nm family of macros provide a flexible and efficient CPU set implementation, backed by the .Xr bitset 9 macros. Each CPU is represented by a single bit. The maximum number of CPUs representable by .Vt cpuset_t is .Va MAXCPU . Individual CPUs in cpusets are referenced with indices zero through .Fa MAXCPU - 1 . .Pp The .Fn CPUSET_T_INITIALIZER macro allows one to initialize a .Vt cpuset_t with a compile time literal value. .Pp The .Fn CPUSET_FSET macro defines a compile time literal, usable by .Fn CPUSET_T_INITIALIZER , representing a full cpuset (all CPUs present). For examples of .Fn CPUSET_T_INITIALIZER and .Fn CPUSET_FSET usage, see the .Sx CPUSET_T_INITIALIZER EXAMPLE section. .Pp The .Fn CPU_CLR macro removes CPU .Fa cpu_idx from the cpuset pointed to by .Fa cpuset . The .Fn CPU_CLR_ATOMIC macro is identical, but the bit representing the CPU is cleared with atomic machine instructions. .Pp The .Fn CPU_COPY macro copies the contents of the cpuset .Fa from to the cpuset .Fa to . .Fn CPU_COPY_STORE_REL is similar, but copies component machine words from .Fa from and writes them to .Fa to with atomic store with release semantics. (That is, if .Fa to is composed of multiple machine words, .Fn CPU_COPY_STORE_REL performs multiple individually atomic operations.) .Pp The .Fn CPU_SET macro adds CPU .Fa cpu_idx to the cpuset pointed to by .Fa cpuset , if it is not already present. The .Fn CPU_SET_ATOMIC macro is identical, but the bit representing the CPU is set with atomic machine instructions. The .Fn CPU_SET_ATOMIC_ACQ macro sets the bit representing the CPU with atomic acquire semantics. .Pp The .Fn CPU_ZERO macro removes all CPUs from .Fa cpuset . .Pp The .Fn CPU_FILL macro adds all CPUs to .Fa cpuset . .Pp The .Fn CPU_SETOF macro removes all CPUs in .Fa cpuset before adding only CPU .Fa cpu_idx . .Pp The .Fn CPU_EMPTY macro returns .Dv true if .Fa cpuset is empty. .Pp The .Fn CPU_ISFULLSET macro returns .Dv true if .Fa cpuset is full (the set of all CPUs). .Pp The .Fn CPU_FFS macro returns the 1-index of the first (lowest) CPU in .Fa cpuset , or zero if .Fa cpuset is empty. Like with .Xr ffs 3 , to use the non-zero result of .Fn CPU_FFS as a .Fa cpu_idx index parameter to any other .Nm macro, you must subtract one from the result. .Pp The .Fn CPU_COUNT macro returns the total number of CPUs in .Fa cpuset . .Pp The .Fn CPU_SUBSET macro returns .Dv true if .Fa needle is a subset of .Fa haystack . .Pp The .Fn CPU_OVERLAP macro returns .Dv true if .Fa cpuset1 and .Fa cpuset2 have any common CPUs. (That is, if .Fa cpuset1 AND .Fa cpuset2 is not the empty set.) .Pp The .Fn CPU_CMP macro returns .Dv true if .Fa cpuset1 is NOT equal to .Fa cpuset2 . .Pp The .Fn CPU_OR macro adds CPUs present in .Fa src to .Fa dst . (It is the .Nm equivalent of the scalar: .Fa dst |= .Fa src . ) .Fn CPU_OR_ATOMIC is similar, but sets the bits representing CPUs in the component machine words in .Fa dst with atomic machine instructions. (That is, if .Fa dst is composed of multiple machine words, .Fn CPU_OR_ATOMIC performs multiple individually atomic operations.) .Pp The .Fn CPU_AND macro removes CPUs absent from .Fa src from .Fa dst . (It is the .Nm equivalent of the scalar: .Fa dst &= .Fa src . ) .Fn CPU_AND_ATOMIC is similar, with the same atomic semantics as .Fn CPU_OR_ATOMIC . .Pp The .Fn CPU_ANDNOT macro removes CPUs in .Fa src from .Fa dst . (It is the .Nm equivalent of the scalar: .Fa dst &= .Fa ~ src . ) .Sh CPUSET_T_INITIALIZER EXAMPLE .Bd -literal cpuset_t myset; /* Initialize myset to filled (all CPUs) */ myset = CPUSET_T_INITIALIZER(CPUSET_FSET); /* Initialize myset to only the lowest CPU */ myset = CPUSET_T_INITIALIZER(0x1); .Ed .Sh SEE ALSO .Xr cpuset 1 , .Xr cpuset 2 , .Xr bitset 9 .Sh HISTORY .In sys/cpuset.h first appeared in .Fx 7.1 , released in January 2009, and in .Fx 8.0 , released in November 2009. .Pp This manual page first appeared in .Fx 11.0 . .Sh AUTHORS .An -nosplit The .Nm macros were written by .An Jeff Roberson Aq Mt jeff@FreeBSD.org . This manual page was written by .An Conrad Meyer Aq Mt cem@FreeBSD.org . .Sh CAVEATS Unlike every other reference to individual set members, which are zero-indexed, .Fn CPU_FFS returns a one-indexed result (or zero if the cpuset is empty). diff --git a/sys/amd64/amd64/pmap.c b/sys/amd64/amd64/pmap.c index 0672c957ff75..9f405baaa007 100644 --- a/sys/amd64/amd64/pmap.c +++ b/sys/amd64/amd64/pmap.c @@ -1,11893 +1,11893 @@ /*- * SPDX-License-Identifier: BSD-4-Clause * * Copyright (c) 1991 Regents of the University of California. * All rights reserved. * Copyright (c) 1994 John S. Dyson * All rights reserved. * Copyright (c) 1994 David Greenman * All rights reserved. * Copyright (c) 2003 Peter Wemm * All rights reserved. * Copyright (c) 2005-2010 Alan L. Cox * All rights reserved. * * This code is derived from software contributed to Berkeley by * the Systems Programming Group of the University of Utah Computer * Science Department and William Jolitz of UUNET Technologies Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. All advertising materials mentioning features or use of this software * must display the following acknowledgement: * This product includes software developed by the University of * California, Berkeley and its contributors. * 4. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91 */ /*- * Copyright (c) 2003 Networks Associates Technology, Inc. * Copyright (c) 2014-2020 The FreeBSD Foundation * All rights reserved. * * This software was developed for the FreeBSD Project by Jake Burkholder, * Safeport Network Services, and Network Associates Laboratories, the * Security Research Division of Network Associates, Inc. under * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the DARPA * CHATS research program. * * Portions of this software were developed by * Konstantin Belousov under sponsorship from * the FreeBSD Foundation. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #define AMD64_NPT_AWARE #include __FBSDID("$FreeBSD$"); /* * Manages physical address maps. * * Since the information managed by this module is * also stored by the logical address mapping module, * this module may throw away valid virtual-to-physical * mappings at almost any time. However, invalidations * of virtual-to-physical mappings must be done as * requested. * * In order to cope with hardware architectures which * make virtual-to-physical map invalidates expensive, * this module may delay invalidate or reduced protection * operations until such time as they are actually * necessary. This module is given full information as * to which processors are currently using which maps, * and to when physical maps must be made correct. */ #include "opt_ddb.h" #include "opt_pmap.h" #include "opt_vm.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef DDB #include #include #endif #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef SMP #include #endif #include #include #ifdef NUMA #define PMAP_MEMDOM MAXMEMDOM #else #define PMAP_MEMDOM 1 #endif static __inline boolean_t pmap_type_guest(pmap_t pmap) { return ((pmap->pm_type == PT_EPT) || (pmap->pm_type == PT_RVI)); } static __inline boolean_t pmap_emulate_ad_bits(pmap_t pmap) { return ((pmap->pm_flags & PMAP_EMULATE_AD_BITS) != 0); } static __inline pt_entry_t pmap_valid_bit(pmap_t pmap) { pt_entry_t mask; switch (pmap->pm_type) { case PT_X86: case PT_RVI: mask = X86_PG_V; break; case PT_EPT: if (pmap_emulate_ad_bits(pmap)) mask = EPT_PG_EMUL_V; else mask = EPT_PG_READ; break; default: panic("pmap_valid_bit: invalid pm_type %d", pmap->pm_type); } return (mask); } static __inline pt_entry_t pmap_rw_bit(pmap_t pmap) { pt_entry_t mask; switch (pmap->pm_type) { case PT_X86: case PT_RVI: mask = X86_PG_RW; break; case PT_EPT: if (pmap_emulate_ad_bits(pmap)) mask = EPT_PG_EMUL_RW; else mask = EPT_PG_WRITE; break; default: panic("pmap_rw_bit: invalid pm_type %d", pmap->pm_type); } return (mask); } static pt_entry_t pg_g; static __inline pt_entry_t pmap_global_bit(pmap_t pmap) { pt_entry_t mask; switch (pmap->pm_type) { case PT_X86: mask = pg_g; break; case PT_RVI: case PT_EPT: mask = 0; break; default: panic("pmap_global_bit: invalid pm_type %d", pmap->pm_type); } return (mask); } static __inline pt_entry_t pmap_accessed_bit(pmap_t pmap) { pt_entry_t mask; switch (pmap->pm_type) { case PT_X86: case PT_RVI: mask = X86_PG_A; break; case PT_EPT: if (pmap_emulate_ad_bits(pmap)) mask = EPT_PG_READ; else mask = EPT_PG_A; break; default: panic("pmap_accessed_bit: invalid pm_type %d", pmap->pm_type); } return (mask); } static __inline pt_entry_t pmap_modified_bit(pmap_t pmap) { pt_entry_t mask; switch (pmap->pm_type) { case PT_X86: case PT_RVI: mask = X86_PG_M; break; case PT_EPT: if (pmap_emulate_ad_bits(pmap)) mask = EPT_PG_WRITE; else mask = EPT_PG_M; break; default: panic("pmap_modified_bit: invalid pm_type %d", pmap->pm_type); } return (mask); } static __inline pt_entry_t pmap_pku_mask_bit(pmap_t pmap) { return (pmap->pm_type == PT_X86 ? X86_PG_PKU_MASK : 0); } #if !defined(DIAGNOSTIC) #ifdef __GNUC_GNU_INLINE__ #define PMAP_INLINE __attribute__((__gnu_inline__)) inline #else #define PMAP_INLINE extern inline #endif #else #define PMAP_INLINE #endif #ifdef PV_STATS #define PV_STAT(x) do { x ; } while (0) #else #define PV_STAT(x) do { } while (0) #endif #undef pa_index #ifdef NUMA #define pa_index(pa) ({ \ KASSERT((pa) <= vm_phys_segs[vm_phys_nsegs - 1].end, \ ("address %lx beyond the last segment", (pa))); \ (pa) >> PDRSHIFT; \ }) #define pa_to_pmdp(pa) (&pv_table[pa_index(pa)]) #define pa_to_pvh(pa) (&(pa_to_pmdp(pa)->pv_page)) #define PHYS_TO_PV_LIST_LOCK(pa) ({ \ struct rwlock *_lock; \ if (__predict_false((pa) > pmap_last_pa)) \ _lock = &pv_dummy_large.pv_lock; \ else \ _lock = &(pa_to_pmdp(pa)->pv_lock); \ _lock; \ }) #else #define pa_index(pa) ((pa) >> PDRSHIFT) #define pa_to_pvh(pa) (&pv_table[pa_index(pa)]) #define NPV_LIST_LOCKS MAXCPU #define PHYS_TO_PV_LIST_LOCK(pa) \ (&pv_list_locks[pa_index(pa) % NPV_LIST_LOCKS]) #endif #define CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, pa) do { \ struct rwlock **_lockp = (lockp); \ struct rwlock *_new_lock; \ \ _new_lock = PHYS_TO_PV_LIST_LOCK(pa); \ if (_new_lock != *_lockp) { \ if (*_lockp != NULL) \ rw_wunlock(*_lockp); \ *_lockp = _new_lock; \ rw_wlock(*_lockp); \ } \ } while (0) #define CHANGE_PV_LIST_LOCK_TO_VM_PAGE(lockp, m) \ CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, VM_PAGE_TO_PHYS(m)) #define RELEASE_PV_LIST_LOCK(lockp) do { \ struct rwlock **_lockp = (lockp); \ \ if (*_lockp != NULL) { \ rw_wunlock(*_lockp); \ *_lockp = NULL; \ } \ } while (0) #define VM_PAGE_TO_PV_LIST_LOCK(m) \ PHYS_TO_PV_LIST_LOCK(VM_PAGE_TO_PHYS(m)) struct pmap kernel_pmap_store; vm_offset_t virtual_avail; /* VA of first avail page (after kernel bss) */ vm_offset_t virtual_end; /* VA of last avail page (end of kernel AS) */ int nkpt; SYSCTL_INT(_machdep, OID_AUTO, nkpt, CTLFLAG_RD, &nkpt, 0, "Number of kernel page table pages allocated on bootup"); static int ndmpdp; vm_paddr_t dmaplimit; vm_offset_t kernel_vm_end = VM_MIN_KERNEL_ADDRESS; pt_entry_t pg_nx; static SYSCTL_NODE(_vm, OID_AUTO, pmap, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "VM/pmap parameters"); static int pg_ps_enabled = 1; SYSCTL_INT(_vm_pmap, OID_AUTO, pg_ps_enabled, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pg_ps_enabled, 0, "Are large page mappings enabled?"); int __read_frequently la57 = 0; SYSCTL_INT(_vm_pmap, OID_AUTO, la57, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &la57, 0, "5-level paging for host is enabled"); static bool pmap_is_la57(pmap_t pmap) { if (pmap->pm_type == PT_X86) return (la57); return (false); /* XXXKIB handle EPT */ } #define PAT_INDEX_SIZE 8 static int pat_index[PAT_INDEX_SIZE]; /* cache mode to PAT index conversion */ static u_int64_t KPTphys; /* phys addr of kernel level 1 */ static u_int64_t KPDphys; /* phys addr of kernel level 2 */ static u_int64_t KPDPphys; /* phys addr of kernel level 3 */ u_int64_t KPML4phys; /* phys addr of kernel level 4 */ u_int64_t KPML5phys; /* phys addr of kernel level 5, if supported */ #ifdef KASAN static uint64_t KASANPDPphys; #endif static pml4_entry_t *kernel_pml4; static u_int64_t DMPDphys; /* phys addr of direct mapped level 2 */ static u_int64_t DMPDPphys; /* phys addr of direct mapped level 3 */ static int ndmpdpphys; /* number of DMPDPphys pages */ vm_paddr_t kernphys; /* phys addr of start of bootstrap data */ vm_paddr_t KERNend; /* and the end */ /* * pmap_mapdev support pre initialization (i.e. console) */ #define PMAP_PREINIT_MAPPING_COUNT 8 static struct pmap_preinit_mapping { vm_paddr_t pa; vm_offset_t va; vm_size_t sz; int mode; } pmap_preinit_mapping[PMAP_PREINIT_MAPPING_COUNT]; static int pmap_initialized; /* * Data for the pv entry allocation mechanism. * Updates to pv_invl_gen are protected by the pv list lock but reads are not. */ #ifdef NUMA static __inline int pc_to_domain(struct pv_chunk *pc) { return (vm_phys_domain(DMAP_TO_PHYS((vm_offset_t)pc))); } #else static __inline int pc_to_domain(struct pv_chunk *pc __unused) { return (0); } #endif struct pv_chunks_list { struct mtx pvc_lock; TAILQ_HEAD(pch, pv_chunk) pvc_list; int active_reclaims; } __aligned(CACHE_LINE_SIZE); struct pv_chunks_list __exclusive_cache_line pv_chunks[PMAP_MEMDOM]; #ifdef NUMA struct pmap_large_md_page { struct rwlock pv_lock; struct md_page pv_page; u_long pv_invl_gen; }; __exclusive_cache_line static struct pmap_large_md_page pv_dummy_large; #define pv_dummy pv_dummy_large.pv_page __read_mostly static struct pmap_large_md_page *pv_table; __read_mostly vm_paddr_t pmap_last_pa; #else static struct rwlock __exclusive_cache_line pv_list_locks[NPV_LIST_LOCKS]; static u_long pv_invl_gen[NPV_LIST_LOCKS]; static struct md_page *pv_table; static struct md_page pv_dummy; #endif /* * All those kernel PT submaps that BSD is so fond of */ pt_entry_t *CMAP1 = NULL; caddr_t CADDR1 = 0; static vm_offset_t qframe = 0; static struct mtx qframe_mtx; static int pmap_flags = PMAP_PDE_SUPERPAGE; /* flags for x86 pmaps */ static vmem_t *large_vmem; static u_int lm_ents; #define PMAP_ADDRESS_IN_LARGEMAP(va) ((va) >= LARGEMAP_MIN_ADDRESS && \ (va) < LARGEMAP_MIN_ADDRESS + NBPML4 * (u_long)lm_ents) int pmap_pcid_enabled = 1; SYSCTL_INT(_vm_pmap, OID_AUTO, pcid_enabled, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pmap_pcid_enabled, 0, "Is TLB Context ID enabled ?"); int invpcid_works = 0; SYSCTL_INT(_vm_pmap, OID_AUTO, invpcid_works, CTLFLAG_RD, &invpcid_works, 0, "Is the invpcid instruction available ?"); int __read_frequently pti = 0; SYSCTL_INT(_vm_pmap, OID_AUTO, pti, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &pti, 0, "Page Table Isolation enabled"); static vm_object_t pti_obj; static pml4_entry_t *pti_pml4; static vm_pindex_t pti_pg_idx; static bool pti_finalized; struct pmap_pkru_range { struct rs_el pkru_rs_el; u_int pkru_keyidx; int pkru_flags; }; static uma_zone_t pmap_pkru_ranges_zone; static bool pmap_pkru_same(pmap_t pmap, vm_offset_t sva, vm_offset_t eva); static pt_entry_t pmap_pkru_get(pmap_t pmap, vm_offset_t va); static void pmap_pkru_on_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva); static void *pkru_dup_range(void *ctx, void *data); static void pkru_free_range(void *ctx, void *node); static int pmap_pkru_copy(pmap_t dst_pmap, pmap_t src_pmap); static int pmap_pkru_deassign(pmap_t pmap, vm_offset_t sva, vm_offset_t eva); static void pmap_pkru_deassign_all(pmap_t pmap); static COUNTER_U64_DEFINE_EARLY(pcid_save_cnt); SYSCTL_COUNTER_U64(_vm_pmap, OID_AUTO, pcid_save_cnt, CTLFLAG_RD, &pcid_save_cnt, "Count of saved TLB context on switch"); static LIST_HEAD(, pmap_invl_gen) pmap_invl_gen_tracker = LIST_HEAD_INITIALIZER(&pmap_invl_gen_tracker); static struct mtx invl_gen_mtx; /* Fake lock object to satisfy turnstiles interface. */ static struct lock_object invl_gen_ts = { .lo_name = "invlts", }; static struct pmap_invl_gen pmap_invl_gen_head = { .gen = 1, .next = NULL, }; static u_long pmap_invl_gen = 1; static int pmap_invl_waiters; static struct callout pmap_invl_callout; static bool pmap_invl_callout_inited; #define PMAP_ASSERT_NOT_IN_DI() \ KASSERT(pmap_not_in_di(), ("DI already started")) static bool pmap_di_locked(void) { int tun; if ((cpu_feature2 & CPUID2_CX16) == 0) return (true); tun = 0; TUNABLE_INT_FETCH("vm.pmap.di_locked", &tun); return (tun != 0); } static int sysctl_pmap_di_locked(SYSCTL_HANDLER_ARGS) { int locked; locked = pmap_di_locked(); return (sysctl_handle_int(oidp, &locked, 0, req)); } SYSCTL_PROC(_vm_pmap, OID_AUTO, di_locked, CTLTYPE_INT | CTLFLAG_RDTUN | CTLFLAG_MPSAFE, 0, 0, sysctl_pmap_di_locked, "", "Locked delayed invalidation"); static bool pmap_not_in_di_l(void); static bool pmap_not_in_di_u(void); DEFINE_IFUNC(, bool, pmap_not_in_di, (void)) { return (pmap_di_locked() ? pmap_not_in_di_l : pmap_not_in_di_u); } static bool pmap_not_in_di_l(void) { struct pmap_invl_gen *invl_gen; invl_gen = &curthread->td_md.md_invl_gen; return (invl_gen->gen == 0); } static void pmap_thread_init_invl_gen_l(struct thread *td) { struct pmap_invl_gen *invl_gen; invl_gen = &td->td_md.md_invl_gen; invl_gen->gen = 0; } static void pmap_delayed_invl_wait_block(u_long *m_gen, u_long *invl_gen) { struct turnstile *ts; ts = turnstile_trywait(&invl_gen_ts); if (*m_gen > atomic_load_long(invl_gen)) turnstile_wait(ts, NULL, TS_SHARED_QUEUE); else turnstile_cancel(ts); } static void pmap_delayed_invl_finish_unblock(u_long new_gen) { struct turnstile *ts; turnstile_chain_lock(&invl_gen_ts); ts = turnstile_lookup(&invl_gen_ts); if (new_gen != 0) pmap_invl_gen = new_gen; if (ts != NULL) { turnstile_broadcast(ts, TS_SHARED_QUEUE); turnstile_unpend(ts); } turnstile_chain_unlock(&invl_gen_ts); } /* * Start a new Delayed Invalidation (DI) block of code, executed by * the current thread. Within a DI block, the current thread may * destroy both the page table and PV list entries for a mapping and * then release the corresponding PV list lock before ensuring that * the mapping is flushed from the TLBs of any processors with the * pmap active. */ static void pmap_delayed_invl_start_l(void) { struct pmap_invl_gen *invl_gen; u_long currgen; invl_gen = &curthread->td_md.md_invl_gen; PMAP_ASSERT_NOT_IN_DI(); mtx_lock(&invl_gen_mtx); if (LIST_EMPTY(&pmap_invl_gen_tracker)) currgen = pmap_invl_gen; else currgen = LIST_FIRST(&pmap_invl_gen_tracker)->gen; invl_gen->gen = currgen + 1; LIST_INSERT_HEAD(&pmap_invl_gen_tracker, invl_gen, link); mtx_unlock(&invl_gen_mtx); } /* * Finish the DI block, previously started by the current thread. All * required TLB flushes for the pages marked by * pmap_delayed_invl_page() must be finished before this function is * called. * * This function works by bumping the global DI generation number to * the generation number of the current thread's DI, unless there is a * pending DI that started earlier. In the latter case, bumping the * global DI generation number would incorrectly signal that the * earlier DI had finished. Instead, this function bumps the earlier * DI's generation number to match the generation number of the * current thread's DI. */ static void pmap_delayed_invl_finish_l(void) { struct pmap_invl_gen *invl_gen, *next; invl_gen = &curthread->td_md.md_invl_gen; KASSERT(invl_gen->gen != 0, ("missed invl_start")); mtx_lock(&invl_gen_mtx); next = LIST_NEXT(invl_gen, link); if (next == NULL) pmap_delayed_invl_finish_unblock(invl_gen->gen); else next->gen = invl_gen->gen; LIST_REMOVE(invl_gen, link); mtx_unlock(&invl_gen_mtx); invl_gen->gen = 0; } static bool pmap_not_in_di_u(void) { struct pmap_invl_gen *invl_gen; invl_gen = &curthread->td_md.md_invl_gen; return (((uintptr_t)invl_gen->next & PMAP_INVL_GEN_NEXT_INVALID) != 0); } static void pmap_thread_init_invl_gen_u(struct thread *td) { struct pmap_invl_gen *invl_gen; invl_gen = &td->td_md.md_invl_gen; invl_gen->gen = 0; invl_gen->next = (void *)PMAP_INVL_GEN_NEXT_INVALID; } static bool pmap_di_load_invl(struct pmap_invl_gen *ptr, struct pmap_invl_gen *out) { uint64_t new_high, new_low, old_high, old_low; char res; old_low = new_low = 0; old_high = new_high = (uintptr_t)0; __asm volatile("lock;cmpxchg16b\t%1" : "=@cce" (res), "+m" (*ptr), "+a" (old_low), "+d" (old_high) : "b"(new_low), "c" (new_high) : "memory", "cc"); if (res == 0) { if ((old_high & PMAP_INVL_GEN_NEXT_INVALID) != 0) return (false); out->gen = old_low; out->next = (void *)old_high; } else { out->gen = new_low; out->next = (void *)new_high; } return (true); } static bool pmap_di_store_invl(struct pmap_invl_gen *ptr, struct pmap_invl_gen *old_val, struct pmap_invl_gen *new_val) { uint64_t new_high, new_low, old_high, old_low; char res; new_low = new_val->gen; new_high = (uintptr_t)new_val->next; old_low = old_val->gen; old_high = (uintptr_t)old_val->next; __asm volatile("lock;cmpxchg16b\t%1" : "=@cce" (res), "+m" (*ptr), "+a" (old_low), "+d" (old_high) : "b"(new_low), "c" (new_high) : "memory", "cc"); return (res); } static COUNTER_U64_DEFINE_EARLY(pv_page_count); SYSCTL_COUNTER_U64(_vm_pmap, OID_AUTO, pv_page_count, CTLFLAG_RD, &pv_page_count, "Current number of allocated pv pages"); static COUNTER_U64_DEFINE_EARLY(user_pt_page_count); SYSCTL_COUNTER_U64(_vm_pmap, OID_AUTO, user_pt_page_count, CTLFLAG_RD, &user_pt_page_count, "Current number of allocated page table pages for userspace"); static COUNTER_U64_DEFINE_EARLY(kernel_pt_page_count); SYSCTL_COUNTER_U64(_vm_pmap, OID_AUTO, kernel_pt_page_count, CTLFLAG_RD, &kernel_pt_page_count, "Current number of allocated page table pages for the kernel"); #ifdef PV_STATS static COUNTER_U64_DEFINE_EARLY(invl_start_restart); SYSCTL_COUNTER_U64(_vm_pmap, OID_AUTO, invl_start_restart, CTLFLAG_RD, &invl_start_restart, "Number of delayed TLB invalidation request restarts"); static COUNTER_U64_DEFINE_EARLY(invl_finish_restart); SYSCTL_COUNTER_U64(_vm_pmap, OID_AUTO, invl_finish_restart, CTLFLAG_RD, &invl_finish_restart, "Number of delayed TLB invalidation completion restarts"); static int invl_max_qlen; SYSCTL_INT(_vm_pmap, OID_AUTO, invl_max_qlen, CTLFLAG_RD, &invl_max_qlen, 0, "Maximum delayed TLB invalidation request queue length"); #endif #define di_delay locks_delay static void pmap_delayed_invl_start_u(void) { struct pmap_invl_gen *invl_gen, *p, prev, new_prev; struct thread *td; struct lock_delay_arg lda; uintptr_t prevl; u_char pri; #ifdef PV_STATS int i, ii; #endif td = curthread; invl_gen = &td->td_md.md_invl_gen; PMAP_ASSERT_NOT_IN_DI(); lock_delay_arg_init(&lda, &di_delay); invl_gen->saved_pri = 0; pri = td->td_base_pri; if (pri > PVM) { thread_lock(td); pri = td->td_base_pri; if (pri > PVM) { invl_gen->saved_pri = pri; sched_prio(td, PVM); } thread_unlock(td); } again: PV_STAT(i = 0); for (p = &pmap_invl_gen_head;; p = prev.next) { PV_STAT(i++); prevl = (uintptr_t)atomic_load_ptr(&p->next); if ((prevl & PMAP_INVL_GEN_NEXT_INVALID) != 0) { PV_STAT(counter_u64_add(invl_start_restart, 1)); lock_delay(&lda); goto again; } if (prevl == 0) break; prev.next = (void *)prevl; } #ifdef PV_STATS if ((ii = invl_max_qlen) < i) atomic_cmpset_int(&invl_max_qlen, ii, i); #endif if (!pmap_di_load_invl(p, &prev) || prev.next != NULL) { PV_STAT(counter_u64_add(invl_start_restart, 1)); lock_delay(&lda); goto again; } new_prev.gen = prev.gen; new_prev.next = invl_gen; invl_gen->gen = prev.gen + 1; /* Formal fence between store to invl->gen and updating *p. */ atomic_thread_fence_rel(); /* * After inserting an invl_gen element with invalid bit set, * this thread blocks any other thread trying to enter the * delayed invalidation block. Do not allow to remove us from * the CPU, because it causes starvation for other threads. */ critical_enter(); /* * ABA for *p is not possible there, since p->gen can only * increase. So if the *p thread finished its di, then * started a new one and got inserted into the list at the * same place, its gen will appear greater than the previously * read gen. */ if (!pmap_di_store_invl(p, &prev, &new_prev)) { critical_exit(); PV_STAT(counter_u64_add(invl_start_restart, 1)); lock_delay(&lda); goto again; } /* * There we clear PMAP_INVL_GEN_NEXT_INVALID in * invl_gen->next, allowing other threads to iterate past us. * pmap_di_store_invl() provides fence between the generation * write and the update of next. */ invl_gen->next = NULL; critical_exit(); } static bool pmap_delayed_invl_finish_u_crit(struct pmap_invl_gen *invl_gen, struct pmap_invl_gen *p) { struct pmap_invl_gen prev, new_prev; u_long mygen; /* * Load invl_gen->gen after setting invl_gen->next * PMAP_INVL_GEN_NEXT_INVALID. This prevents larger * generations to propagate to our invl_gen->gen. Lock prefix * in atomic_set_ptr() worked as seq_cst fence. */ mygen = atomic_load_long(&invl_gen->gen); if (!pmap_di_load_invl(p, &prev) || prev.next != invl_gen) return (false); KASSERT(prev.gen < mygen, ("invalid di gen sequence %lu %lu", prev.gen, mygen)); new_prev.gen = mygen; new_prev.next = (void *)((uintptr_t)invl_gen->next & ~PMAP_INVL_GEN_NEXT_INVALID); /* Formal fence between load of prev and storing update to it. */ atomic_thread_fence_rel(); return (pmap_di_store_invl(p, &prev, &new_prev)); } static void pmap_delayed_invl_finish_u(void) { struct pmap_invl_gen *invl_gen, *p; struct thread *td; struct lock_delay_arg lda; uintptr_t prevl; td = curthread; invl_gen = &td->td_md.md_invl_gen; KASSERT(invl_gen->gen != 0, ("missed invl_start: gen 0")); KASSERT(((uintptr_t)invl_gen->next & PMAP_INVL_GEN_NEXT_INVALID) == 0, ("missed invl_start: INVALID")); lock_delay_arg_init(&lda, &di_delay); again: for (p = &pmap_invl_gen_head; p != NULL; p = (void *)prevl) { prevl = (uintptr_t)atomic_load_ptr(&p->next); if ((prevl & PMAP_INVL_GEN_NEXT_INVALID) != 0) { PV_STAT(counter_u64_add(invl_finish_restart, 1)); lock_delay(&lda); goto again; } if ((void *)prevl == invl_gen) break; } /* * It is legitimate to not find ourself on the list if a * thread before us finished its DI and started it again. */ if (__predict_false(p == NULL)) { PV_STAT(counter_u64_add(invl_finish_restart, 1)); lock_delay(&lda); goto again; } critical_enter(); atomic_set_ptr((uintptr_t *)&invl_gen->next, PMAP_INVL_GEN_NEXT_INVALID); if (!pmap_delayed_invl_finish_u_crit(invl_gen, p)) { atomic_clear_ptr((uintptr_t *)&invl_gen->next, PMAP_INVL_GEN_NEXT_INVALID); critical_exit(); PV_STAT(counter_u64_add(invl_finish_restart, 1)); lock_delay(&lda); goto again; } critical_exit(); if (atomic_load_int(&pmap_invl_waiters) > 0) pmap_delayed_invl_finish_unblock(0); if (invl_gen->saved_pri != 0) { thread_lock(td); sched_prio(td, invl_gen->saved_pri); thread_unlock(td); } } #ifdef DDB DB_SHOW_COMMAND(di_queue, pmap_di_queue) { struct pmap_invl_gen *p, *pn; struct thread *td; uintptr_t nextl; bool first; for (p = &pmap_invl_gen_head, first = true; p != NULL; p = pn, first = false) { nextl = (uintptr_t)atomic_load_ptr(&p->next); pn = (void *)(nextl & ~PMAP_INVL_GEN_NEXT_INVALID); td = first ? NULL : __containerof(p, struct thread, td_md.md_invl_gen); db_printf("gen %lu inv %d td %p tid %d\n", p->gen, (nextl & PMAP_INVL_GEN_NEXT_INVALID) != 0, td, td != NULL ? td->td_tid : -1); } } #endif #ifdef PV_STATS static COUNTER_U64_DEFINE_EARLY(invl_wait); SYSCTL_COUNTER_U64(_vm_pmap, OID_AUTO, invl_wait, CTLFLAG_RD, &invl_wait, "Number of times DI invalidation blocked pmap_remove_all/write"); static COUNTER_U64_DEFINE_EARLY(invl_wait_slow); SYSCTL_COUNTER_U64(_vm_pmap, OID_AUTO, invl_wait_slow, CTLFLAG_RD, &invl_wait_slow, "Number of slow invalidation waits for lockless DI"); #endif #ifdef NUMA static u_long * pmap_delayed_invl_genp(vm_page_t m) { vm_paddr_t pa; u_long *gen; pa = VM_PAGE_TO_PHYS(m); if (__predict_false((pa) > pmap_last_pa)) gen = &pv_dummy_large.pv_invl_gen; else gen = &(pa_to_pmdp(pa)->pv_invl_gen); return (gen); } #else static u_long * pmap_delayed_invl_genp(vm_page_t m) { return (&pv_invl_gen[pa_index(VM_PAGE_TO_PHYS(m)) % NPV_LIST_LOCKS]); } #endif static void pmap_delayed_invl_callout_func(void *arg __unused) { if (atomic_load_int(&pmap_invl_waiters) == 0) return; pmap_delayed_invl_finish_unblock(0); } static void pmap_delayed_invl_callout_init(void *arg __unused) { if (pmap_di_locked()) return; callout_init(&pmap_invl_callout, 1); pmap_invl_callout_inited = true; } SYSINIT(pmap_di_callout, SI_SUB_CPU + 1, SI_ORDER_ANY, pmap_delayed_invl_callout_init, NULL); /* * Ensure that all currently executing DI blocks, that need to flush * TLB for the given page m, actually flushed the TLB at the time the * function returned. If the page m has an empty PV list and we call * pmap_delayed_invl_wait(), upon its return we know that no CPU has a * valid mapping for the page m in either its page table or TLB. * * This function works by blocking until the global DI generation * number catches up with the generation number associated with the * given page m and its PV list. Since this function's callers * typically own an object lock and sometimes own a page lock, it * cannot sleep. Instead, it blocks on a turnstile to relinquish the * processor. */ static void pmap_delayed_invl_wait_l(vm_page_t m) { u_long *m_gen; #ifdef PV_STATS bool accounted = false; #endif m_gen = pmap_delayed_invl_genp(m); while (*m_gen > pmap_invl_gen) { #ifdef PV_STATS if (!accounted) { counter_u64_add(invl_wait, 1); accounted = true; } #endif pmap_delayed_invl_wait_block(m_gen, &pmap_invl_gen); } } static void pmap_delayed_invl_wait_u(vm_page_t m) { u_long *m_gen; struct lock_delay_arg lda; bool fast; fast = true; m_gen = pmap_delayed_invl_genp(m); lock_delay_arg_init(&lda, &di_delay); while (*m_gen > atomic_load_long(&pmap_invl_gen_head.gen)) { if (fast || !pmap_invl_callout_inited) { PV_STAT(counter_u64_add(invl_wait, 1)); lock_delay(&lda); fast = false; } else { /* * The page's invalidation generation number * is still below the current thread's number. * Prepare to block so that we do not waste * CPU cycles or worse, suffer livelock. * * Since it is impossible to block without * racing with pmap_delayed_invl_finish_u(), * prepare for the race by incrementing * pmap_invl_waiters and arming a 1-tick * callout which will unblock us if we lose * the race. */ atomic_add_int(&pmap_invl_waiters, 1); /* * Re-check the current thread's invalidation * generation after incrementing * pmap_invl_waiters, so that there is no race * with pmap_delayed_invl_finish_u() setting * the page generation and checking * pmap_invl_waiters. The only race allowed * is for a missed unblock, which is handled * by the callout. */ if (*m_gen > atomic_load_long(&pmap_invl_gen_head.gen)) { callout_reset(&pmap_invl_callout, 1, pmap_delayed_invl_callout_func, NULL); PV_STAT(counter_u64_add(invl_wait_slow, 1)); pmap_delayed_invl_wait_block(m_gen, &pmap_invl_gen_head.gen); } atomic_add_int(&pmap_invl_waiters, -1); } } } DEFINE_IFUNC(, void, pmap_thread_init_invl_gen, (struct thread *)) { return (pmap_di_locked() ? pmap_thread_init_invl_gen_l : pmap_thread_init_invl_gen_u); } DEFINE_IFUNC(static, void, pmap_delayed_invl_start, (void)) { return (pmap_di_locked() ? pmap_delayed_invl_start_l : pmap_delayed_invl_start_u); } DEFINE_IFUNC(static, void, pmap_delayed_invl_finish, (void)) { return (pmap_di_locked() ? pmap_delayed_invl_finish_l : pmap_delayed_invl_finish_u); } DEFINE_IFUNC(static, void, pmap_delayed_invl_wait, (vm_page_t)) { return (pmap_di_locked() ? pmap_delayed_invl_wait_l : pmap_delayed_invl_wait_u); } /* * Mark the page m's PV list as participating in the current thread's * DI block. Any threads concurrently using m's PV list to remove or * restrict all mappings to m will wait for the current thread's DI * block to complete before proceeding. * * The function works by setting the DI generation number for m's PV * list to at least the DI generation number of the current thread. * This forces a caller of pmap_delayed_invl_wait() to block until * current thread calls pmap_delayed_invl_finish(). */ static void pmap_delayed_invl_page(vm_page_t m) { u_long gen, *m_gen; rw_assert(VM_PAGE_TO_PV_LIST_LOCK(m), RA_WLOCKED); gen = curthread->td_md.md_invl_gen.gen; if (gen == 0) return; m_gen = pmap_delayed_invl_genp(m); if (*m_gen < gen) *m_gen = gen; } /* * Crashdump maps. */ static caddr_t crashdumpmap; /* * Internal flags for pmap_enter()'s helper functions. */ #define PMAP_ENTER_NORECLAIM 0x1000000 /* Don't reclaim PV entries. */ #define PMAP_ENTER_NOREPLACE 0x2000000 /* Don't replace mappings. */ /* * Internal flags for pmap_mapdev_internal() and * pmap_change_props_locked(). */ #define MAPDEV_FLUSHCACHE 0x00000001 /* Flush cache after mapping. */ #define MAPDEV_SETATTR 0x00000002 /* Modify existing attrs. */ #define MAPDEV_ASSERTVALID 0x00000004 /* Assert mapping validity. */ TAILQ_HEAD(pv_chunklist, pv_chunk); static void free_pv_chunk(struct pv_chunk *pc); static void free_pv_chunk_batch(struct pv_chunklist *batch); static void free_pv_entry(pmap_t pmap, pv_entry_t pv); static pv_entry_t get_pv_entry(pmap_t pmap, struct rwlock **lockp); static int popcnt_pc_map_pq(uint64_t *map); static vm_page_t reclaim_pv_chunk(pmap_t locked_pmap, struct rwlock **lockp); static void reserve_pv_entries(pmap_t pmap, int needed, struct rwlock **lockp); static void pmap_pv_demote_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa, struct rwlock **lockp); static bool pmap_pv_insert_pde(pmap_t pmap, vm_offset_t va, pd_entry_t pde, u_int flags, struct rwlock **lockp); #if VM_NRESERVLEVEL > 0 static void pmap_pv_promote_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa, struct rwlock **lockp); #endif static void pmap_pvh_free(struct md_page *pvh, pmap_t pmap, vm_offset_t va); static pv_entry_t pmap_pvh_remove(struct md_page *pvh, pmap_t pmap, vm_offset_t va); static void pmap_abort_ptp(pmap_t pmap, vm_offset_t va, vm_page_t mpte); static int pmap_change_props_locked(vm_offset_t va, vm_size_t size, vm_prot_t prot, int mode, int flags); static boolean_t pmap_demote_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va); static boolean_t pmap_demote_pde_locked(pmap_t pmap, pd_entry_t *pde, vm_offset_t va, struct rwlock **lockp); static boolean_t pmap_demote_pdpe(pmap_t pmap, pdp_entry_t *pdpe, vm_offset_t va); static bool pmap_enter_2mpage(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, struct rwlock **lockp); static int pmap_enter_pde(pmap_t pmap, vm_offset_t va, pd_entry_t newpde, u_int flags, vm_page_t m, struct rwlock **lockp); static vm_page_t pmap_enter_quick_locked(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, vm_page_t mpte, struct rwlock **lockp); static void pmap_fill_ptp(pt_entry_t *firstpte, pt_entry_t newpte); static int pmap_insert_pt_page(pmap_t pmap, vm_page_t mpte, bool promoted); static void pmap_invalidate_cache_range_selfsnoop(vm_offset_t sva, vm_offset_t eva); static void pmap_invalidate_cache_range_all(vm_offset_t sva, vm_offset_t eva); static void pmap_invalidate_pde_page(pmap_t pmap, vm_offset_t va, pd_entry_t pde); static void pmap_kenter_attr(vm_offset_t va, vm_paddr_t pa, int mode); static vm_page_t pmap_large_map_getptp_unlocked(void); static vm_paddr_t pmap_large_map_kextract(vm_offset_t va); #if VM_NRESERVLEVEL > 0 static void pmap_promote_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va, struct rwlock **lockp); #endif static boolean_t pmap_protect_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t sva, vm_prot_t prot); static void pmap_pte_props(pt_entry_t *pte, u_long bits, u_long mask); static void pmap_pti_add_kva_locked(vm_offset_t sva, vm_offset_t eva, bool exec); static pdp_entry_t *pmap_pti_pdpe(vm_offset_t va); static pd_entry_t *pmap_pti_pde(vm_offset_t va); static void pmap_pti_wire_pte(void *pte); static int pmap_remove_pde(pmap_t pmap, pd_entry_t *pdq, vm_offset_t sva, struct spglist *free, struct rwlock **lockp); static int pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t sva, pd_entry_t ptepde, struct spglist *free, struct rwlock **lockp); static vm_page_t pmap_remove_pt_page(pmap_t pmap, vm_offset_t va); static void pmap_remove_page(pmap_t pmap, vm_offset_t va, pd_entry_t *pde, struct spglist *free); static bool pmap_remove_ptes(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, pd_entry_t *pde, struct spglist *free, struct rwlock **lockp); static boolean_t pmap_try_insert_pv_entry(pmap_t pmap, vm_offset_t va, vm_page_t m, struct rwlock **lockp); static void pmap_update_pde(pmap_t pmap, vm_offset_t va, pd_entry_t *pde, pd_entry_t newpde); static void pmap_update_pde_invalidate(pmap_t, vm_offset_t va, pd_entry_t pde); static pd_entry_t *pmap_alloc_pde(pmap_t pmap, vm_offset_t va, vm_page_t *pdpgp, struct rwlock **lockp); static vm_page_t pmap_allocpte_alloc(pmap_t pmap, vm_pindex_t ptepindex, struct rwlock **lockp, vm_offset_t va); static vm_page_t pmap_allocpte_nosleep(pmap_t pmap, vm_pindex_t ptepindex, struct rwlock **lockp, vm_offset_t va); static vm_page_t pmap_allocpte(pmap_t pmap, vm_offset_t va, struct rwlock **lockp); static void _pmap_unwire_ptp(pmap_t pmap, vm_offset_t va, vm_page_t m, struct spglist *free); static int pmap_unuse_pt(pmap_t, vm_offset_t, pd_entry_t, struct spglist *); static vm_page_t pmap_alloc_pt_page(pmap_t, vm_pindex_t, int); static void pmap_free_pt_page(pmap_t, vm_page_t, bool); /********************/ /* Inline functions */ /********************/ /* * Return a non-clipped indexes for a given VA, which are page table * pages indexes at the corresponding level. */ static __inline vm_pindex_t pmap_pde_pindex(vm_offset_t va) { return (va >> PDRSHIFT); } static __inline vm_pindex_t pmap_pdpe_pindex(vm_offset_t va) { return (NUPDE + (va >> PDPSHIFT)); } static __inline vm_pindex_t pmap_pml4e_pindex(vm_offset_t va) { return (NUPDE + NUPDPE + (va >> PML4SHIFT)); } static __inline vm_pindex_t pmap_pml5e_pindex(vm_offset_t va) { return (NUPDE + NUPDPE + NUPML4E + (va >> PML5SHIFT)); } static __inline pml4_entry_t * pmap_pml5e(pmap_t pmap, vm_offset_t va) { MPASS(pmap_is_la57(pmap)); return (&pmap->pm_pmltop[pmap_pml5e_index(va)]); } static __inline pml4_entry_t * pmap_pml5e_u(pmap_t pmap, vm_offset_t va) { MPASS(pmap_is_la57(pmap)); return (&pmap->pm_pmltopu[pmap_pml5e_index(va)]); } static __inline pml4_entry_t * pmap_pml5e_to_pml4e(pml5_entry_t *pml5e, vm_offset_t va) { pml4_entry_t *pml4e; /* XXX MPASS(pmap_is_la57(pmap); */ pml4e = (pml4_entry_t *)PHYS_TO_DMAP(*pml5e & PG_FRAME); return (&pml4e[pmap_pml4e_index(va)]); } /* Return a pointer to the PML4 slot that corresponds to a VA */ static __inline pml4_entry_t * pmap_pml4e(pmap_t pmap, vm_offset_t va) { pml5_entry_t *pml5e; pml4_entry_t *pml4e; pt_entry_t PG_V; if (pmap_is_la57(pmap)) { pml5e = pmap_pml5e(pmap, va); PG_V = pmap_valid_bit(pmap); if ((*pml5e & PG_V) == 0) return (NULL); pml4e = (pml4_entry_t *)PHYS_TO_DMAP(*pml5e & PG_FRAME); } else { pml4e = pmap->pm_pmltop; } return (&pml4e[pmap_pml4e_index(va)]); } static __inline pml4_entry_t * pmap_pml4e_u(pmap_t pmap, vm_offset_t va) { MPASS(!pmap_is_la57(pmap)); return (&pmap->pm_pmltopu[pmap_pml4e_index(va)]); } /* Return a pointer to the PDP slot that corresponds to a VA */ static __inline pdp_entry_t * pmap_pml4e_to_pdpe(pml4_entry_t *pml4e, vm_offset_t va) { pdp_entry_t *pdpe; pdpe = (pdp_entry_t *)PHYS_TO_DMAP(*pml4e & PG_FRAME); return (&pdpe[pmap_pdpe_index(va)]); } /* Return a pointer to the PDP slot that corresponds to a VA */ static __inline pdp_entry_t * pmap_pdpe(pmap_t pmap, vm_offset_t va) { pml4_entry_t *pml4e; pt_entry_t PG_V; PG_V = pmap_valid_bit(pmap); pml4e = pmap_pml4e(pmap, va); if (pml4e == NULL || (*pml4e & PG_V) == 0) return (NULL); return (pmap_pml4e_to_pdpe(pml4e, va)); } /* Return a pointer to the PD slot that corresponds to a VA */ static __inline pd_entry_t * pmap_pdpe_to_pde(pdp_entry_t *pdpe, vm_offset_t va) { pd_entry_t *pde; KASSERT((*pdpe & PG_PS) == 0, ("%s: pdpe %#lx is a leaf", __func__, *pdpe)); pde = (pd_entry_t *)PHYS_TO_DMAP(*pdpe & PG_FRAME); return (&pde[pmap_pde_index(va)]); } /* Return a pointer to the PD slot that corresponds to a VA */ static __inline pd_entry_t * pmap_pde(pmap_t pmap, vm_offset_t va) { pdp_entry_t *pdpe; pt_entry_t PG_V; PG_V = pmap_valid_bit(pmap); pdpe = pmap_pdpe(pmap, va); if (pdpe == NULL || (*pdpe & PG_V) == 0) return (NULL); KASSERT((*pdpe & PG_PS) == 0, ("pmap_pde for 1G page, pmap %p va %#lx", pmap, va)); return (pmap_pdpe_to_pde(pdpe, va)); } /* Return a pointer to the PT slot that corresponds to a VA */ static __inline pt_entry_t * pmap_pde_to_pte(pd_entry_t *pde, vm_offset_t va) { pt_entry_t *pte; KASSERT((*pde & PG_PS) == 0, ("%s: pde %#lx is a leaf", __func__, *pde)); pte = (pt_entry_t *)PHYS_TO_DMAP(*pde & PG_FRAME); return (&pte[pmap_pte_index(va)]); } /* Return a pointer to the PT slot that corresponds to a VA */ static __inline pt_entry_t * pmap_pte(pmap_t pmap, vm_offset_t va) { pd_entry_t *pde; pt_entry_t PG_V; PG_V = pmap_valid_bit(pmap); pde = pmap_pde(pmap, va); if (pde == NULL || (*pde & PG_V) == 0) return (NULL); if ((*pde & PG_PS) != 0) /* compat with i386 pmap_pte() */ return ((pt_entry_t *)pde); return (pmap_pde_to_pte(pde, va)); } static __inline void pmap_resident_count_adj(pmap_t pmap, int count) { PMAP_LOCK_ASSERT(pmap, MA_OWNED); KASSERT(pmap->pm_stats.resident_count + count >= 0, ("pmap %p resident count underflow %ld %d", pmap, pmap->pm_stats.resident_count, count)); pmap->pm_stats.resident_count += count; } static __inline void pmap_pt_page_count_pinit(pmap_t pmap, int count) { KASSERT(pmap->pm_stats.resident_count + count >= 0, ("pmap %p resident count underflow %ld %d", pmap, pmap->pm_stats.resident_count, count)); pmap->pm_stats.resident_count += count; } static __inline void pmap_pt_page_count_adj(pmap_t pmap, int count) { if (pmap == kernel_pmap) counter_u64_add(kernel_pt_page_count, count); else { if (pmap != NULL) pmap_resident_count_adj(pmap, count); counter_u64_add(user_pt_page_count, count); } } PMAP_INLINE pt_entry_t * vtopte(vm_offset_t va) { u_int64_t mask; KASSERT(va >= VM_MAXUSER_ADDRESS, ("vtopte on a uva/gpa 0x%0lx", va)); if (la57) { mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT + NPDPEPGSHIFT + NPML4EPGSHIFT + NPML5EPGSHIFT)) - 1); return (P5Tmap + ((va >> PAGE_SHIFT) & mask)); } else { mask = ((1ul << (NPTEPGSHIFT + NPDEPGSHIFT + NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1); return (P4Tmap + ((va >> PAGE_SHIFT) & mask)); } } static __inline pd_entry_t * vtopde(vm_offset_t va) { u_int64_t mask; KASSERT(va >= VM_MAXUSER_ADDRESS, ("vtopde on a uva/gpa 0x%0lx", va)); if (la57) { mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT + NPML4EPGSHIFT + NPML5EPGSHIFT)) - 1); return (P5Dmap + ((va >> PDRSHIFT) & mask)); } else { mask = ((1ul << (NPDEPGSHIFT + NPDPEPGSHIFT + NPML4EPGSHIFT)) - 1); return (P4Dmap + ((va >> PDRSHIFT) & mask)); } } static u_int64_t allocpages(vm_paddr_t *firstaddr, int n) { u_int64_t ret; ret = *firstaddr; bzero((void *)ret, n * PAGE_SIZE); *firstaddr += n * PAGE_SIZE; return (ret); } CTASSERT(powerof2(NDMPML4E)); /* number of kernel PDP slots */ #define NKPDPE(ptpgs) howmany(ptpgs, NPDEPG) static void nkpt_init(vm_paddr_t addr) { int pt_pages; #ifdef NKPT pt_pages = NKPT; #else pt_pages = howmany(addr - kernphys, NBPDR) + 1; /* +1 for 2M hole @0 */ pt_pages += NKPDPE(pt_pages); /* * Add some slop beyond the bare minimum required for bootstrapping * the kernel. * * This is quite important when allocating KVA for kernel modules. * The modules are required to be linked in the negative 2GB of * the address space. If we run out of KVA in this region then * pmap_growkernel() will need to allocate page table pages to map * the entire 512GB of KVA space which is an unnecessary tax on * physical memory. * * Secondly, device memory mapped as part of setting up the low- * level console(s) is taken from KVA, starting at virtual_avail. * This is because cninit() is called after pmap_bootstrap() but * before vm_init() and pmap_init(). 20MB for a frame buffer is * not uncommon. */ pt_pages += 32; /* 64MB additional slop. */ #endif nkpt = pt_pages; } /* * Returns the proper write/execute permission for a physical page that is * part of the initial boot allocations. * * If the page has kernel text, it is marked as read-only. If the page has * kernel read-only data, it is marked as read-only/not-executable. If the * page has only read-write data, it is marked as read-write/not-executable. * If the page is below/above the kernel range, it is marked as read-write. * * This function operates on 2M pages, since we map the kernel space that * way. */ static inline pt_entry_t bootaddr_rwx(vm_paddr_t pa) { /* * The kernel is loaded at a 2MB-aligned address, and memory below that * need not be executable. The .bss section is padded to a 2MB * boundary, so memory following the kernel need not be executable * either. Preloaded kernel modules have their mapping permissions * fixed up by the linker. */ if (pa < trunc_2mpage(kernphys + btext - KERNSTART) || pa >= trunc_2mpage(kernphys + _end - KERNSTART)) return (X86_PG_RW | pg_nx); /* * The linker should ensure that the read-only and read-write * portions don't share the same 2M page, so this shouldn't * impact read-only data. However, in any case, any page with * read-write data needs to be read-write. */ if (pa >= trunc_2mpage(kernphys + brwsection - KERNSTART)) return (X86_PG_RW | pg_nx); /* * Mark any 2M page containing kernel text as read-only. Mark * other pages with read-only data as read-only and not executable. * (It is likely a small portion of the read-only data section will * be marked as read-only, but executable. This should be acceptable * since the read-only protection will keep the data from changing.) * Note that fixups to the .text section will still work until we * set CR0.WP. */ if (pa < round_2mpage(kernphys + etext - KERNSTART)) return (0); return (pg_nx); } static void create_pagetables(vm_paddr_t *firstaddr) { pd_entry_t *pd_p; pdp_entry_t *pdp_p; pml4_entry_t *p4_p; uint64_t DMPDkernphys; vm_paddr_t pax; #ifdef KASAN pt_entry_t *pt_p; uint64_t KASANPDphys, KASANPTphys, KASANphys; vm_offset_t kasankernbase; int kasankpdpi, kasankpdi, nkasanpte; #endif int i, j, ndm1g, nkpdpe, nkdmpde; /* Allocate page table pages for the direct map */ ndmpdp = howmany(ptoa(Maxmem), NBPDP); if (ndmpdp < 4) /* Minimum 4GB of dirmap */ ndmpdp = 4; ndmpdpphys = howmany(ndmpdp, NPDPEPG); if (ndmpdpphys > NDMPML4E) { /* * Each NDMPML4E allows 512 GB, so limit to that, * and then readjust ndmpdp and ndmpdpphys. */ printf("NDMPML4E limits system to %d GB\n", NDMPML4E * 512); Maxmem = atop(NDMPML4E * NBPML4); ndmpdpphys = NDMPML4E; ndmpdp = NDMPML4E * NPDEPG; } DMPDPphys = allocpages(firstaddr, ndmpdpphys); ndm1g = 0; if ((amd_feature & AMDID_PAGE1GB) != 0) { /* * Calculate the number of 1G pages that will fully fit in * Maxmem. */ ndm1g = ptoa(Maxmem) >> PDPSHIFT; /* * Allocate 2M pages for the kernel. These will be used in * place of the one or more 1G pages from ndm1g that maps * kernel memory into DMAP. */ nkdmpde = howmany((vm_offset_t)brwsection - KERNSTART + kernphys - rounddown2(kernphys, NBPDP), NBPDP); DMPDkernphys = allocpages(firstaddr, nkdmpde); } if (ndm1g < ndmpdp) DMPDphys = allocpages(firstaddr, ndmpdp - ndm1g); dmaplimit = (vm_paddr_t)ndmpdp << PDPSHIFT; /* Allocate pages */ KPML4phys = allocpages(firstaddr, 1); KPDPphys = allocpages(firstaddr, NKPML4E); #ifdef KASAN KASANPDPphys = allocpages(firstaddr, NKASANPML4E); KASANPDphys = allocpages(firstaddr, 1); #endif /* * Allocate the initial number of kernel page table pages required to * bootstrap. We defer this until after all memory-size dependent * allocations are done (e.g. direct map), so that we don't have to * build in too much slop in our estimate. * * Note that when NKPML4E > 1, we have an empty page underneath * all but the KPML4I'th one, so we need NKPML4E-1 extra (zeroed) * pages. (pmap_enter requires a PD page to exist for each KPML4E.) */ nkpt_init(*firstaddr); nkpdpe = NKPDPE(nkpt); KPTphys = allocpages(firstaddr, nkpt); KPDphys = allocpages(firstaddr, nkpdpe); #ifdef KASAN nkasanpte = howmany(nkpt, KASAN_SHADOW_SCALE); KASANPTphys = allocpages(firstaddr, nkasanpte); KASANphys = allocpages(firstaddr, nkasanpte * NPTEPG); #endif /* * Connect the zero-filled PT pages to their PD entries. This * implicitly maps the PT pages at their correct locations within * the PTmap. */ pd_p = (pd_entry_t *)KPDphys; for (i = 0; i < nkpt; i++) pd_p[i] = (KPTphys + ptoa(i)) | X86_PG_RW | X86_PG_V; /* * Map from start of the kernel in physical memory (staging * area) to the end of loader preallocated memory using 2MB * pages. This replaces some of the PD entries created above. * For compatibility, identity map 2M at the start. */ pd_p[0] = X86_PG_V | PG_PS | pg_g | X86_PG_M | X86_PG_A | X86_PG_RW | pg_nx; for (i = 1, pax = kernphys; pax < KERNend; i++, pax += NBPDR) { /* Preset PG_M and PG_A because demotion expects it. */ pd_p[i] = pax | X86_PG_V | PG_PS | pg_g | X86_PG_M | X86_PG_A | bootaddr_rwx(pax); } /* * Because we map the physical blocks in 2M pages, adjust firstaddr * to record the physical blocks we've actually mapped into kernel * virtual address space. */ if (*firstaddr < round_2mpage(KERNend)) *firstaddr = round_2mpage(KERNend); /* And connect up the PD to the PDP (leaving room for L4 pages) */ pdp_p = (pdp_entry_t *)(KPDPphys + ptoa(KPML4I - KPML4BASE)); for (i = 0; i < nkpdpe; i++) pdp_p[i + KPDPI] = (KPDphys + ptoa(i)) | X86_PG_RW | X86_PG_V; #ifdef KASAN kasankernbase = kasan_md_addr_to_shad(KERNBASE); kasankpdpi = pmap_pdpe_index(kasankernbase); kasankpdi = pmap_pde_index(kasankernbase); pdp_p = (pdp_entry_t *)KASANPDPphys; pdp_p[kasankpdpi] = (KASANPDphys | X86_PG_RW | X86_PG_V | pg_nx); pd_p = (pd_entry_t *)KASANPDphys; for (i = 0; i < nkasanpte; i++) pd_p[i + kasankpdi] = (KASANPTphys + ptoa(i)) | X86_PG_RW | X86_PG_V | pg_nx; pt_p = (pt_entry_t *)KASANPTphys; for (i = 0; i < nkasanpte * NPTEPG; i++) pt_p[i] = (KASANphys + ptoa(i)) | X86_PG_RW | X86_PG_V | X86_PG_M | X86_PG_A | pg_nx; #endif /* * Now, set up the direct map region using 2MB and/or 1GB pages. If * the end of physical memory is not aligned to a 1GB page boundary, * then the residual physical memory is mapped with 2MB pages. Later, * if pmap_mapdev{_attr}() uses the direct map for non-write-back * memory, pmap_change_attr() will demote any 2MB or 1GB page mappings * that are partially used. */ pd_p = (pd_entry_t *)DMPDphys; for (i = NPDEPG * ndm1g, j = 0; i < NPDEPG * ndmpdp; i++, j++) { pd_p[j] = (vm_paddr_t)i << PDRSHIFT; /* Preset PG_M and PG_A because demotion expects it. */ pd_p[j] |= X86_PG_RW | X86_PG_V | PG_PS | pg_g | X86_PG_M | X86_PG_A | pg_nx; } pdp_p = (pdp_entry_t *)DMPDPphys; for (i = 0; i < ndm1g; i++) { pdp_p[i] = (vm_paddr_t)i << PDPSHIFT; /* Preset PG_M and PG_A because demotion expects it. */ pdp_p[i] |= X86_PG_RW | X86_PG_V | PG_PS | pg_g | X86_PG_M | X86_PG_A | pg_nx; } for (j = 0; i < ndmpdp; i++, j++) { pdp_p[i] = DMPDphys + ptoa(j); pdp_p[i] |= X86_PG_RW | X86_PG_V | pg_nx; } /* * Instead of using a 1G page for the memory containing the kernel, * use 2M pages with read-only and no-execute permissions. (If using 1G * pages, this will partially overwrite the PDPEs above.) */ if (ndm1g > 0) { pd_p = (pd_entry_t *)DMPDkernphys; for (i = 0, pax = rounddown2(kernphys, NBPDP); i < NPDEPG * nkdmpde; i++, pax += NBPDR) { pd_p[i] = pax | X86_PG_V | PG_PS | pg_g | X86_PG_M | X86_PG_A | pg_nx | bootaddr_rwx(pax); } j = rounddown2(kernphys, NBPDP) >> PDPSHIFT; for (i = 0; i < nkdmpde; i++) { pdp_p[i + j] = (DMPDkernphys + ptoa(i)) | X86_PG_RW | X86_PG_V | pg_nx; } } /* And recursively map PML4 to itself in order to get PTmap */ p4_p = (pml4_entry_t *)KPML4phys; p4_p[PML4PML4I] = KPML4phys; p4_p[PML4PML4I] |= X86_PG_RW | X86_PG_V | pg_nx; #ifdef KASAN /* Connect the KASAN shadow map slots up to the PML4. */ for (i = 0; i < NKASANPML4E; i++) { p4_p[KASANPML4I + i] = KASANPDPphys + ptoa(i); p4_p[KASANPML4I + i] |= X86_PG_RW | X86_PG_V | pg_nx; } #endif /* Connect the Direct Map slots up to the PML4. */ for (i = 0; i < ndmpdpphys; i++) { p4_p[DMPML4I + i] = DMPDPphys + ptoa(i); p4_p[DMPML4I + i] |= X86_PG_RW | X86_PG_V | pg_nx; } /* Connect the KVA slots up to the PML4 */ for (i = 0; i < NKPML4E; i++) { p4_p[KPML4BASE + i] = KPDPphys + ptoa(i); p4_p[KPML4BASE + i] |= X86_PG_RW | X86_PG_V; } kernel_pml4 = (pml4_entry_t *)PHYS_TO_DMAP(KPML4phys); } /* * Bootstrap the system enough to run with virtual memory. * * On amd64 this is called after mapping has already been enabled * and just syncs the pmap module with what has already been done. * [We can't call it easily with mapping off since the kernel is not * mapped with PA == VA, hence we would have to relocate every address * from the linked base (virtual) address "KERNBASE" to the actual * (physical) address starting relative to 0] */ void pmap_bootstrap(vm_paddr_t *firstaddr) { vm_offset_t va; pt_entry_t *pte, *pcpu_pte; struct region_descriptor r_gdt; uint64_t cr4, pcpu_phys; u_long res; int i; KERNend = *firstaddr; res = atop(KERNend - (vm_paddr_t)kernphys); if (!pti) pg_g = X86_PG_G; /* * Create an initial set of page tables to run the kernel in. */ create_pagetables(firstaddr); pcpu_phys = allocpages(firstaddr, MAXCPU); /* * Add a physical memory segment (vm_phys_seg) corresponding to the * preallocated kernel page table pages so that vm_page structures * representing these pages will be created. The vm_page structures * are required for promotion of the corresponding kernel virtual * addresses to superpage mappings. */ vm_phys_early_add_seg(KPTphys, KPTphys + ptoa(nkpt)); /* * Account for the virtual addresses mapped by create_pagetables(). */ virtual_avail = (vm_offset_t)KERNSTART + round_2mpage(KERNend - (vm_paddr_t)kernphys); virtual_end = VM_MAX_KERNEL_ADDRESS; /* * Enable PG_G global pages, then switch to the kernel page * table from the bootstrap page table. After the switch, it * is possible to enable SMEP and SMAP since PG_U bits are * correct now. */ cr4 = rcr4(); cr4 |= CR4_PGE; load_cr4(cr4); load_cr3(KPML4phys); if (cpu_stdext_feature & CPUID_STDEXT_SMEP) cr4 |= CR4_SMEP; if (cpu_stdext_feature & CPUID_STDEXT_SMAP) cr4 |= CR4_SMAP; load_cr4(cr4); /* * Initialize the kernel pmap (which is statically allocated). * Count bootstrap data as being resident in case any of this data is * later unmapped (using pmap_remove()) and freed. */ PMAP_LOCK_INIT(kernel_pmap); kernel_pmap->pm_pmltop = kernel_pml4; kernel_pmap->pm_cr3 = KPML4phys; kernel_pmap->pm_ucr3 = PMAP_NO_CR3; TAILQ_INIT(&kernel_pmap->pm_pvchunk); kernel_pmap->pm_stats.resident_count = res; kernel_pmap->pm_flags = pmap_flags; /* * The kernel pmap is always active on all CPUs. Once CPUs are * enumerated, the mask will be set equal to all_cpus. */ CPU_FILL(&kernel_pmap->pm_active); /* * Initialize the TLB invalidations generation number lock. */ mtx_init(&invl_gen_mtx, "invlgn", NULL, MTX_DEF); /* * Reserve some special page table entries/VA space for temporary * mapping of pages. */ #define SYSMAP(c, p, v, n) \ v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n); va = virtual_avail; pte = vtopte(va); /* * Crashdump maps. The first page is reused as CMAP1 for the * memory test. */ SYSMAP(caddr_t, CMAP1, crashdumpmap, MAXDUMPPGS) CADDR1 = crashdumpmap; SYSMAP(struct pcpu *, pcpu_pte, __pcpu, MAXCPU); virtual_avail = va; for (i = 0; i < MAXCPU; i++) { pcpu_pte[i] = (pcpu_phys + ptoa(i)) | X86_PG_V | X86_PG_RW | pg_g | pg_nx | X86_PG_M | X86_PG_A; } /* * Re-initialize PCPU area for BSP after switching. * Make hardware use gdt and common_tss from the new PCPU. */ STAILQ_INIT(&cpuhead); wrmsr(MSR_GSBASE, (uint64_t)&__pcpu[0]); pcpu_init(&__pcpu[0], 0, sizeof(struct pcpu)); amd64_bsp_pcpu_init1(&__pcpu[0]); amd64_bsp_ist_init(&__pcpu[0]); __pcpu[0].pc_common_tss.tss_iobase = sizeof(struct amd64tss) + IOPERM_BITMAP_SIZE; memcpy(__pcpu[0].pc_gdt, temp_bsp_pcpu.pc_gdt, NGDT * sizeof(struct user_segment_descriptor)); gdt_segs[GPROC0_SEL].ssd_base = (uintptr_t)&__pcpu[0].pc_common_tss; ssdtosyssd(&gdt_segs[GPROC0_SEL], (struct system_segment_descriptor *)&__pcpu[0].pc_gdt[GPROC0_SEL]); r_gdt.rd_limit = NGDT * sizeof(struct user_segment_descriptor) - 1; r_gdt.rd_base = (long)__pcpu[0].pc_gdt; lgdt(&r_gdt); wrmsr(MSR_GSBASE, (uint64_t)&__pcpu[0]); ltr(GSEL(GPROC0_SEL, SEL_KPL)); __pcpu[0].pc_dynamic = temp_bsp_pcpu.pc_dynamic; __pcpu[0].pc_acpi_id = temp_bsp_pcpu.pc_acpi_id; /* * Initialize the PAT MSR. * pmap_init_pat() clears and sets CR4_PGE, which, as a * side-effect, invalidates stale PG_G TLB entries that might * have been created in our pre-boot environment. */ pmap_init_pat(); /* Initialize TLB Context Id. */ if (pmap_pcid_enabled) { for (i = 0; i < MAXCPU; i++) { kernel_pmap->pm_pcids[i].pm_pcid = PMAP_PCID_KERN; kernel_pmap->pm_pcids[i].pm_gen = 1; } /* * PMAP_PCID_KERN + 1 is used for initialization of * proc0 pmap. The pmap' pcid state might be used by * EFIRT entry before first context switch, so it * needs to be valid. */ PCPU_SET(pcid_next, PMAP_PCID_KERN + 2); PCPU_SET(pcid_gen, 1); /* * pcpu area for APs is zeroed during AP startup. * pc_pcid_next and pc_pcid_gen are initialized by AP * during pcpu setup. */ load_cr4(rcr4() | CR4_PCIDE); } } /* * Setup the PAT MSR. */ void pmap_init_pat(void) { uint64_t pat_msr; u_long cr0, cr4; int i; /* Bail if this CPU doesn't implement PAT. */ if ((cpu_feature & CPUID_PAT) == 0) panic("no PAT??"); /* Set default PAT index table. */ for (i = 0; i < PAT_INDEX_SIZE; i++) pat_index[i] = -1; pat_index[PAT_WRITE_BACK] = 0; pat_index[PAT_WRITE_THROUGH] = 1; pat_index[PAT_UNCACHEABLE] = 3; pat_index[PAT_WRITE_COMBINING] = 6; pat_index[PAT_WRITE_PROTECTED] = 5; pat_index[PAT_UNCACHED] = 2; /* * Initialize default PAT entries. * Leave the indices 0-3 at the default of WB, WT, UC-, and UC. * Program 5 and 6 as WP and WC. * * Leave 4 and 7 as WB and UC. Note that a recursive page table * mapping for a 2M page uses a PAT value with the bit 3 set due * to its overload with PG_PS. */ pat_msr = PAT_VALUE(0, PAT_WRITE_BACK) | PAT_VALUE(1, PAT_WRITE_THROUGH) | PAT_VALUE(2, PAT_UNCACHED) | PAT_VALUE(3, PAT_UNCACHEABLE) | PAT_VALUE(4, PAT_WRITE_BACK) | PAT_VALUE(5, PAT_WRITE_PROTECTED) | PAT_VALUE(6, PAT_WRITE_COMBINING) | PAT_VALUE(7, PAT_UNCACHEABLE); /* Disable PGE. */ cr4 = rcr4(); load_cr4(cr4 & ~CR4_PGE); /* Disable caches (CD = 1, NW = 0). */ cr0 = rcr0(); load_cr0((cr0 & ~CR0_NW) | CR0_CD); /* Flushes caches and TLBs. */ wbinvd(); invltlb(); /* Update PAT and index table. */ wrmsr(MSR_PAT, pat_msr); /* Flush caches and TLBs again. */ wbinvd(); invltlb(); /* Restore caches and PGE. */ load_cr0(cr0); load_cr4(cr4); } vm_page_t pmap_page_alloc_below_4g(bool zeroed) { return (vm_page_alloc_noobj_contig((zeroed ? VM_ALLOC_ZERO : 0), 1, 0, (1ULL << 32), PAGE_SIZE, 0, VM_MEMATTR_DEFAULT)); } extern const char la57_trampoline[], la57_trampoline_gdt_desc[], la57_trampoline_gdt[], la57_trampoline_end[]; static void pmap_bootstrap_la57(void *arg __unused) { char *v_code; pml5_entry_t *v_pml5; pml4_entry_t *v_pml4; pdp_entry_t *v_pdp; pd_entry_t *v_pd; pt_entry_t *v_pt; vm_page_t m_code, m_pml4, m_pdp, m_pd, m_pt, m_pml5; void (*la57_tramp)(uint64_t pml5); struct region_descriptor r_gdt; if ((cpu_stdext_feature2 & CPUID_STDEXT2_LA57) == 0) return; TUNABLE_INT_FETCH("vm.pmap.la57", &la57); if (!la57) return; r_gdt.rd_limit = NGDT * sizeof(struct user_segment_descriptor) - 1; r_gdt.rd_base = (long)__pcpu[0].pc_gdt; m_code = pmap_page_alloc_below_4g(true); v_code = (char *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m_code)); m_pml5 = pmap_page_alloc_below_4g(true); KPML5phys = VM_PAGE_TO_PHYS(m_pml5); v_pml5 = (pml5_entry_t *)PHYS_TO_DMAP(KPML5phys); m_pml4 = pmap_page_alloc_below_4g(true); v_pml4 = (pdp_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m_pml4)); m_pdp = pmap_page_alloc_below_4g(true); v_pdp = (pdp_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m_pdp)); m_pd = pmap_page_alloc_below_4g(true); v_pd = (pdp_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m_pd)); m_pt = pmap_page_alloc_below_4g(true); v_pt = (pt_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m_pt)); /* * Map m_code 1:1, it appears below 4G in KVA due to physical * address being below 4G. Since kernel KVA is in upper half, * the pml4e should be zero and free for temporary use. */ kernel_pmap->pm_pmltop[pmap_pml4e_index(VM_PAGE_TO_PHYS(m_code))] = VM_PAGE_TO_PHYS(m_pdp) | X86_PG_V | X86_PG_RW | X86_PG_A | X86_PG_M; v_pdp[pmap_pdpe_index(VM_PAGE_TO_PHYS(m_code))] = VM_PAGE_TO_PHYS(m_pd) | X86_PG_V | X86_PG_RW | X86_PG_A | X86_PG_M; v_pd[pmap_pde_index(VM_PAGE_TO_PHYS(m_code))] = VM_PAGE_TO_PHYS(m_pt) | X86_PG_V | X86_PG_RW | X86_PG_A | X86_PG_M; v_pt[pmap_pte_index(VM_PAGE_TO_PHYS(m_code))] = VM_PAGE_TO_PHYS(m_code) | X86_PG_V | X86_PG_RW | X86_PG_A | X86_PG_M; /* * Add pml5 entry at top of KVA pointing to existing pml4 table, * entering all existing kernel mappings into level 5 table. */ v_pml5[pmap_pml5e_index(UPT_MAX_ADDRESS)] = KPML4phys | X86_PG_V | X86_PG_RW | X86_PG_A | X86_PG_M | pg_g; /* * Add pml5 entry for 1:1 trampoline mapping after LA57 is turned on. */ v_pml5[pmap_pml5e_index(VM_PAGE_TO_PHYS(m_code))] = VM_PAGE_TO_PHYS(m_pml4) | X86_PG_V | X86_PG_RW | X86_PG_A | X86_PG_M; v_pml4[pmap_pml4e_index(VM_PAGE_TO_PHYS(m_code))] = VM_PAGE_TO_PHYS(m_pdp) | X86_PG_V | X86_PG_RW | X86_PG_A | X86_PG_M; /* * Copy and call the 48->57 trampoline, hope we return there, alive. */ bcopy(la57_trampoline, v_code, la57_trampoline_end - la57_trampoline); *(u_long *)(v_code + 2 + (la57_trampoline_gdt_desc - la57_trampoline)) = la57_trampoline_gdt - la57_trampoline + VM_PAGE_TO_PHYS(m_code); la57_tramp = (void (*)(uint64_t))VM_PAGE_TO_PHYS(m_code); invlpg((vm_offset_t)la57_tramp); la57_tramp(KPML5phys); /* * gdt was necessary reset, switch back to our gdt. */ lgdt(&r_gdt); wrmsr(MSR_GSBASE, (uint64_t)&__pcpu[0]); load_ds(_udatasel); load_es(_udatasel); load_fs(_ufssel); ssdtosyssd(&gdt_segs[GPROC0_SEL], (struct system_segment_descriptor *)&__pcpu[0].pc_gdt[GPROC0_SEL]); ltr(GSEL(GPROC0_SEL, SEL_KPL)); /* * Now unmap the trampoline, and free the pages. * Clear pml5 entry used for 1:1 trampoline mapping. */ pte_clear(&v_pml5[pmap_pml5e_index(VM_PAGE_TO_PHYS(m_code))]); invlpg((vm_offset_t)v_code); vm_page_free(m_code); vm_page_free(m_pdp); vm_page_free(m_pd); vm_page_free(m_pt); /* * Recursively map PML5 to itself in order to get PTmap and * PDmap. */ v_pml5[PML5PML5I] = KPML5phys | X86_PG_RW | X86_PG_V | pg_nx; kernel_pmap->pm_cr3 = KPML5phys; kernel_pmap->pm_pmltop = v_pml5; pmap_pt_page_count_adj(kernel_pmap, 1); } SYSINIT(la57, SI_SUB_KMEM, SI_ORDER_ANY, pmap_bootstrap_la57, NULL); /* * Initialize a vm_page's machine-dependent fields. */ void pmap_page_init(vm_page_t m) { TAILQ_INIT(&m->md.pv_list); m->md.pat_mode = PAT_WRITE_BACK; } static int pmap_allow_2m_x_ept; SYSCTL_INT(_vm_pmap, OID_AUTO, allow_2m_x_ept, CTLFLAG_RWTUN | CTLFLAG_NOFETCH, &pmap_allow_2m_x_ept, 0, "Allow executable superpage mappings in EPT"); void pmap_allow_2m_x_ept_recalculate(void) { /* * SKL002, SKL012S. Since the EPT format is only used by * Intel CPUs, the vendor check is merely a formality. */ if (!(cpu_vendor_id != CPU_VENDOR_INTEL || (cpu_ia32_arch_caps & IA32_ARCH_CAP_IF_PSCHANGE_MC_NO) != 0 || (CPUID_TO_FAMILY(cpu_id) == 0x6 && (CPUID_TO_MODEL(cpu_id) == 0x26 || /* Atoms */ CPUID_TO_MODEL(cpu_id) == 0x27 || CPUID_TO_MODEL(cpu_id) == 0x35 || CPUID_TO_MODEL(cpu_id) == 0x36 || CPUID_TO_MODEL(cpu_id) == 0x37 || CPUID_TO_MODEL(cpu_id) == 0x86 || CPUID_TO_MODEL(cpu_id) == 0x1c || CPUID_TO_MODEL(cpu_id) == 0x4a || CPUID_TO_MODEL(cpu_id) == 0x4c || CPUID_TO_MODEL(cpu_id) == 0x4d || CPUID_TO_MODEL(cpu_id) == 0x5a || CPUID_TO_MODEL(cpu_id) == 0x5c || CPUID_TO_MODEL(cpu_id) == 0x5d || CPUID_TO_MODEL(cpu_id) == 0x5f || CPUID_TO_MODEL(cpu_id) == 0x6e || CPUID_TO_MODEL(cpu_id) == 0x7a || CPUID_TO_MODEL(cpu_id) == 0x57 || /* Knights */ CPUID_TO_MODEL(cpu_id) == 0x85)))) pmap_allow_2m_x_ept = 1; TUNABLE_INT_FETCH("hw.allow_2m_x_ept", &pmap_allow_2m_x_ept); } static bool pmap_allow_2m_x_page(pmap_t pmap, bool executable) { return (pmap->pm_type != PT_EPT || !executable || !pmap_allow_2m_x_ept); } #ifdef NUMA static void pmap_init_pv_table(void) { struct pmap_large_md_page *pvd; vm_size_t s; long start, end, highest, pv_npg; int domain, i, j, pages; /* * We strongly depend on the size being a power of two, so the assert * is overzealous. However, should the struct be resized to a * different power of two, the code below needs to be revisited. */ CTASSERT((sizeof(*pvd) == 64)); /* * Calculate the size of the array. */ pmap_last_pa = vm_phys_segs[vm_phys_nsegs - 1].end; pv_npg = howmany(pmap_last_pa, NBPDR); s = (vm_size_t)pv_npg * sizeof(struct pmap_large_md_page); s = round_page(s); pv_table = (struct pmap_large_md_page *)kva_alloc(s); if (pv_table == NULL) panic("%s: kva_alloc failed\n", __func__); /* * Iterate physical segments to allocate space for respective pages. */ highest = -1; s = 0; for (i = 0; i < vm_phys_nsegs; i++) { end = vm_phys_segs[i].end / NBPDR; domain = vm_phys_segs[i].domain; if (highest >= end) continue; start = highest + 1; pvd = &pv_table[start]; pages = end - start + 1; s = round_page(pages * sizeof(*pvd)); highest = start + (s / sizeof(*pvd)) - 1; for (j = 0; j < s; j += PAGE_SIZE) { vm_page_t m = vm_page_alloc_noobj_domain(domain, 0); if (m == NULL) panic("failed to allocate PV table page"); pmap_qenter((vm_offset_t)pvd + j, &m, 1); } for (j = 0; j < s / sizeof(*pvd); j++) { rw_init_flags(&pvd->pv_lock, "pmap pv list", RW_NEW); TAILQ_INIT(&pvd->pv_page.pv_list); pvd->pv_page.pv_gen = 0; pvd->pv_page.pat_mode = 0; pvd->pv_invl_gen = 0; pvd++; } } pvd = &pv_dummy_large; rw_init_flags(&pvd->pv_lock, "pmap pv list dummy", RW_NEW); TAILQ_INIT(&pvd->pv_page.pv_list); pvd->pv_page.pv_gen = 0; pvd->pv_page.pat_mode = 0; pvd->pv_invl_gen = 0; } #else static void pmap_init_pv_table(void) { vm_size_t s; long i, pv_npg; /* * Initialize the pool of pv list locks. */ for (i = 0; i < NPV_LIST_LOCKS; i++) rw_init(&pv_list_locks[i], "pmap pv list"); /* * Calculate the size of the pv head table for superpages. */ pv_npg = howmany(vm_phys_segs[vm_phys_nsegs - 1].end, NBPDR); /* * Allocate memory for the pv head table for superpages. */ s = (vm_size_t)pv_npg * sizeof(struct md_page); s = round_page(s); pv_table = (struct md_page *)kmem_malloc(s, M_WAITOK | M_ZERO); for (i = 0; i < pv_npg; i++) TAILQ_INIT(&pv_table[i].pv_list); TAILQ_INIT(&pv_dummy.pv_list); } #endif /* * Initialize the pmap module. * Called by vm_init, to initialize any structures that the pmap * system needs to map virtual memory. */ void pmap_init(void) { struct pmap_preinit_mapping *ppim; vm_page_t m, mpte; int error, i, ret, skz63; /* L1TF, reserve page @0 unconditionally */ vm_page_blacklist_add(0, bootverbose); /* Detect bare-metal Skylake Server and Skylake-X. */ if (vm_guest == VM_GUEST_NO && cpu_vendor_id == CPU_VENDOR_INTEL && CPUID_TO_FAMILY(cpu_id) == 0x6 && CPUID_TO_MODEL(cpu_id) == 0x55) { /* * Skylake-X errata SKZ63. Processor May Hang When * Executing Code In an HLE Transaction Region between * 40000000H and 403FFFFFH. * * Mark the pages in the range as preallocated. It * seems to be impossible to distinguish between * Skylake Server and Skylake X. */ skz63 = 1; TUNABLE_INT_FETCH("hw.skz63_enable", &skz63); if (skz63 != 0) { if (bootverbose) printf("SKZ63: skipping 4M RAM starting " "at physical 1G\n"); for (i = 0; i < atop(0x400000); i++) { ret = vm_page_blacklist_add(0x40000000 + ptoa(i), FALSE); if (!ret && bootverbose) printf("page at %#lx already used\n", 0x40000000 + ptoa(i)); } } } /* IFU */ pmap_allow_2m_x_ept_recalculate(); /* * Initialize the vm page array entries for the kernel pmap's * page table pages. */ PMAP_LOCK(kernel_pmap); for (i = 0; i < nkpt; i++) { mpte = PHYS_TO_VM_PAGE(KPTphys + (i << PAGE_SHIFT)); KASSERT(mpte >= vm_page_array && mpte < &vm_page_array[vm_page_array_size], ("pmap_init: page table page is out of range")); mpte->pindex = pmap_pde_pindex(KERNBASE) + i; mpte->phys_addr = KPTphys + (i << PAGE_SHIFT); mpte->ref_count = 1; /* * Collect the page table pages that were replaced by a 2MB * page in create_pagetables(). They are zero filled. */ if ((i == 0 || kernphys + ((vm_paddr_t)(i - 1) << PDRSHIFT) < KERNend) && pmap_insert_pt_page(kernel_pmap, mpte, false)) panic("pmap_init: pmap_insert_pt_page failed"); } PMAP_UNLOCK(kernel_pmap); vm_wire_add(nkpt); /* * If the kernel is running on a virtual machine, then it must assume * that MCA is enabled by the hypervisor. Moreover, the kernel must * be prepared for the hypervisor changing the vendor and family that * are reported by CPUID. Consequently, the workaround for AMD Family * 10h Erratum 383 is enabled if the processor's feature set does not * include at least one feature that is only supported by older Intel * or newer AMD processors. */ if (vm_guest != VM_GUEST_NO && (cpu_feature & CPUID_SS) == 0 && (cpu_feature2 & (CPUID2_SSSE3 | CPUID2_SSE41 | CPUID2_AESNI | CPUID2_AVX | CPUID2_XSAVE)) == 0 && (amd_feature2 & (AMDID2_XOP | AMDID2_FMA4)) == 0) workaround_erratum383 = 1; /* * Are large page mappings enabled? */ TUNABLE_INT_FETCH("vm.pmap.pg_ps_enabled", &pg_ps_enabled); if (pg_ps_enabled) { KASSERT(MAXPAGESIZES > 1 && pagesizes[1] == 0, ("pmap_init: can't assign to pagesizes[1]")); pagesizes[1] = NBPDR; if ((amd_feature & AMDID_PAGE1GB) != 0) { KASSERT(MAXPAGESIZES > 2 && pagesizes[2] == 0, ("pmap_init: can't assign to pagesizes[2]")); pagesizes[2] = NBPDP; } } /* * Initialize pv chunk lists. */ for (i = 0; i < PMAP_MEMDOM; i++) { mtx_init(&pv_chunks[i].pvc_lock, "pmap pv chunk list", NULL, MTX_DEF); TAILQ_INIT(&pv_chunks[i].pvc_list); } pmap_init_pv_table(); pmap_initialized = 1; for (i = 0; i < PMAP_PREINIT_MAPPING_COUNT; i++) { ppim = pmap_preinit_mapping + i; if (ppim->va == 0) continue; /* Make the direct map consistent */ if (ppim->pa < dmaplimit && ppim->pa + ppim->sz <= dmaplimit) { (void)pmap_change_attr(PHYS_TO_DMAP(ppim->pa), ppim->sz, ppim->mode); } if (!bootverbose) continue; printf("PPIM %u: PA=%#lx, VA=%#lx, size=%#lx, mode=%#x\n", i, ppim->pa, ppim->va, ppim->sz, ppim->mode); } mtx_init(&qframe_mtx, "qfrmlk", NULL, MTX_SPIN); error = vmem_alloc(kernel_arena, PAGE_SIZE, M_BESTFIT | M_WAITOK, (vmem_addr_t *)&qframe); if (error != 0) panic("qframe allocation failed"); lm_ents = 8; TUNABLE_INT_FETCH("vm.pmap.large_map_pml4_entries", &lm_ents); if (lm_ents > LMEPML4I - LMSPML4I + 1) lm_ents = LMEPML4I - LMSPML4I + 1; if (bootverbose) printf("pmap: large map %u PML4 slots (%lu GB)\n", lm_ents, (u_long)lm_ents * (NBPML4 / 1024 / 1024 / 1024)); if (lm_ents != 0) { large_vmem = vmem_create("large", LARGEMAP_MIN_ADDRESS, (vmem_size_t)lm_ents * NBPML4, PAGE_SIZE, 0, M_WAITOK); if (large_vmem == NULL) { printf("pmap: cannot create large map\n"); lm_ents = 0; } for (i = 0; i < lm_ents; i++) { m = pmap_large_map_getptp_unlocked(); /* XXXKIB la57 */ kernel_pml4[LMSPML4I + i] = X86_PG_V | X86_PG_RW | X86_PG_A | X86_PG_M | pg_nx | VM_PAGE_TO_PHYS(m); } } } SYSCTL_UINT(_vm_pmap, OID_AUTO, large_map_pml4_entries, CTLFLAG_RDTUN | CTLFLAG_NOFETCH, &lm_ents, 0, "Maximum number of PML4 entries for use by large map (tunable). " "Each entry corresponds to 512GB of address space."); static SYSCTL_NODE(_vm_pmap, OID_AUTO, pde, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "2MB page mapping counters"); static COUNTER_U64_DEFINE_EARLY(pmap_pde_demotions); SYSCTL_COUNTER_U64(_vm_pmap_pde, OID_AUTO, demotions, CTLFLAG_RD, &pmap_pde_demotions, "2MB page demotions"); static COUNTER_U64_DEFINE_EARLY(pmap_pde_mappings); SYSCTL_COUNTER_U64(_vm_pmap_pde, OID_AUTO, mappings, CTLFLAG_RD, &pmap_pde_mappings, "2MB page mappings"); static COUNTER_U64_DEFINE_EARLY(pmap_pde_p_failures); SYSCTL_COUNTER_U64(_vm_pmap_pde, OID_AUTO, p_failures, CTLFLAG_RD, &pmap_pde_p_failures, "2MB page promotion failures"); static COUNTER_U64_DEFINE_EARLY(pmap_pde_promotions); SYSCTL_COUNTER_U64(_vm_pmap_pde, OID_AUTO, promotions, CTLFLAG_RD, &pmap_pde_promotions, "2MB page promotions"); static SYSCTL_NODE(_vm_pmap, OID_AUTO, pdpe, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "1GB page mapping counters"); static COUNTER_U64_DEFINE_EARLY(pmap_pdpe_demotions); SYSCTL_COUNTER_U64(_vm_pmap_pdpe, OID_AUTO, demotions, CTLFLAG_RD, &pmap_pdpe_demotions, "1GB page demotions"); /*************************************************** * Low level helper routines..... ***************************************************/ static pt_entry_t pmap_swap_pat(pmap_t pmap, pt_entry_t entry) { int x86_pat_bits = X86_PG_PTE_PAT | X86_PG_PDE_PAT; switch (pmap->pm_type) { case PT_X86: case PT_RVI: /* Verify that both PAT bits are not set at the same time */ KASSERT((entry & x86_pat_bits) != x86_pat_bits, ("Invalid PAT bits in entry %#lx", entry)); /* Swap the PAT bits if one of them is set */ if ((entry & x86_pat_bits) != 0) entry ^= x86_pat_bits; break; case PT_EPT: /* * Nothing to do - the memory attributes are represented * the same way for regular pages and superpages. */ break; default: panic("pmap_switch_pat_bits: bad pm_type %d", pmap->pm_type); } return (entry); } boolean_t pmap_is_valid_memattr(pmap_t pmap __unused, vm_memattr_t mode) { return (mode >= 0 && mode < PAT_INDEX_SIZE && pat_index[(int)mode] >= 0); } /* * Determine the appropriate bits to set in a PTE or PDE for a specified * caching mode. */ int pmap_cache_bits(pmap_t pmap, int mode, boolean_t is_pde) { int cache_bits, pat_flag, pat_idx; if (!pmap_is_valid_memattr(pmap, mode)) panic("Unknown caching mode %d\n", mode); switch (pmap->pm_type) { case PT_X86: case PT_RVI: /* The PAT bit is different for PTE's and PDE's. */ pat_flag = is_pde ? X86_PG_PDE_PAT : X86_PG_PTE_PAT; /* Map the caching mode to a PAT index. */ pat_idx = pat_index[mode]; /* Map the 3-bit index value into the PAT, PCD, and PWT bits. */ cache_bits = 0; if (pat_idx & 0x4) cache_bits |= pat_flag; if (pat_idx & 0x2) cache_bits |= PG_NC_PCD; if (pat_idx & 0x1) cache_bits |= PG_NC_PWT; break; case PT_EPT: cache_bits = EPT_PG_IGNORE_PAT | EPT_PG_MEMORY_TYPE(mode); break; default: panic("unsupported pmap type %d", pmap->pm_type); } return (cache_bits); } static int pmap_cache_mask(pmap_t pmap, boolean_t is_pde) { int mask; switch (pmap->pm_type) { case PT_X86: case PT_RVI: mask = is_pde ? X86_PG_PDE_CACHE : X86_PG_PTE_CACHE; break; case PT_EPT: mask = EPT_PG_IGNORE_PAT | EPT_PG_MEMORY_TYPE(0x7); break; default: panic("pmap_cache_mask: invalid pm_type %d", pmap->pm_type); } return (mask); } static int pmap_pat_index(pmap_t pmap, pt_entry_t pte, bool is_pde) { int pat_flag, pat_idx; pat_idx = 0; switch (pmap->pm_type) { case PT_X86: case PT_RVI: /* The PAT bit is different for PTE's and PDE's. */ pat_flag = is_pde ? X86_PG_PDE_PAT : X86_PG_PTE_PAT; if ((pte & pat_flag) != 0) pat_idx |= 0x4; if ((pte & PG_NC_PCD) != 0) pat_idx |= 0x2; if ((pte & PG_NC_PWT) != 0) pat_idx |= 0x1; break; case PT_EPT: if ((pte & EPT_PG_IGNORE_PAT) != 0) panic("EPT PTE %#lx has no PAT memory type", pte); pat_idx = (pte & EPT_PG_MEMORY_TYPE(0x7)) >> 3; break; } /* See pmap_init_pat(). */ if (pat_idx == 4) pat_idx = 0; if (pat_idx == 7) pat_idx = 3; return (pat_idx); } bool pmap_ps_enabled(pmap_t pmap) { return (pg_ps_enabled && (pmap->pm_flags & PMAP_PDE_SUPERPAGE) != 0); } static void pmap_update_pde_store(pmap_t pmap, pd_entry_t *pde, pd_entry_t newpde) { switch (pmap->pm_type) { case PT_X86: break; case PT_RVI: case PT_EPT: /* * XXX * This is a little bogus since the generation number is * supposed to be bumped up when a region of the address * space is invalidated in the page tables. * * In this case the old PDE entry is valid but yet we want * to make sure that any mappings using the old entry are * invalidated in the TLB. * * The reason this works as expected is because we rendezvous * "all" host cpus and force any vcpu context to exit as a * side-effect. */ atomic_add_long(&pmap->pm_eptgen, 1); break; default: panic("pmap_update_pde_store: bad pm_type %d", pmap->pm_type); } pde_store(pde, newpde); } /* * After changing the page size for the specified virtual address in the page * table, flush the corresponding entries from the processor's TLB. Only the * calling processor's TLB is affected. * * The calling thread must be pinned to a processor. */ static void pmap_update_pde_invalidate(pmap_t pmap, vm_offset_t va, pd_entry_t newpde) { pt_entry_t PG_G; if (pmap_type_guest(pmap)) return; KASSERT(pmap->pm_type == PT_X86, ("pmap_update_pde_invalidate: invalid type %d", pmap->pm_type)); PG_G = pmap_global_bit(pmap); if ((newpde & PG_PS) == 0) /* Demotion: flush a specific 2MB page mapping. */ invlpg(va); else if ((newpde & PG_G) == 0) /* * Promotion: flush every 4KB page mapping from the TLB * because there are too many to flush individually. */ invltlb(); else { /* * Promotion: flush every 4KB page mapping from the TLB, * including any global (PG_G) mappings. */ invltlb_glob(); } } /* * The amd64 pmap uses different approaches to TLB invalidation * depending on the kernel configuration, available hardware features, * and known hardware errata. The kernel configuration option that * has the greatest operational impact on TLB invalidation is PTI, * which is enabled automatically on affected Intel CPUs. The most * impactful hardware features are first PCID, and then INVPCID * instruction presence. PCID usage is quite different for PTI * vs. non-PTI. * * * Kernel Page Table Isolation (PTI or KPTI) is used to mitigate * the Meltdown bug in some Intel CPUs. Under PTI, each user address * space is served by two page tables, user and kernel. The user * page table only maps user space and a kernel trampoline. The * kernel trampoline includes the entirety of the kernel text but * only the kernel data that is needed to switch from user to kernel * mode. The kernel page table maps the user and kernel address * spaces in their entirety. It is identical to the per-process * page table used in non-PTI mode. * * User page tables are only used when the CPU is in user mode. * Consequently, some TLB invalidations can be postponed until the * switch from kernel to user mode. In contrast, the user * space part of the kernel page table is used for copyout(9), so * TLB invalidations on this page table cannot be similarly postponed. * * The existence of a user mode page table for the given pmap is * indicated by a pm_ucr3 value that differs from PMAP_NO_CR3, in * which case pm_ucr3 contains the %cr3 register value for the user * mode page table's root. * * * The pm_active bitmask indicates which CPUs currently have the * pmap active. A CPU's bit is set on context switch to the pmap, and * cleared on switching off this CPU. For the kernel page table, * the pm_active field is immutable and contains all CPUs. The * kernel page table is always logically active on every processor, * but not necessarily in use by the hardware, e.g., in PTI mode. * * When requesting invalidation of virtual addresses with * pmap_invalidate_XXX() functions, the pmap sends shootdown IPIs to * all CPUs recorded as active in pm_active. Updates to and reads * from pm_active are not synchronized, and so they may race with * each other. Shootdown handlers are prepared to handle the race. * * * PCID is an optional feature of the long mode x86 MMU where TLB * entries are tagged with the 'Process ID' of the address space * they belong to. This feature provides a limited namespace for * process identifiers, 12 bits, supporting 4095 simultaneous IDs * total. * * Allocation of a PCID to a pmap is done by an algorithm described * in section 15.12, "Other TLB Consistency Algorithms", of * Vahalia's book "Unix Internals". A PCID cannot be allocated for * the whole lifetime of a pmap in pmap_pinit() due to the limited * namespace. Instead, a per-CPU, per-pmap PCID is assigned when * the CPU is about to start caching TLB entries from a pmap, * i.e., on the context switch that activates the pmap on the CPU. * * The PCID allocator maintains a per-CPU, per-pmap generation * count, pm_gen, which is incremented each time a new PCID is * allocated. On TLB invalidation, the generation counters for the * pmap are zeroed, which signals the context switch code that the * previously allocated PCID is no longer valid. Effectively, * zeroing any of these counters triggers a TLB shootdown for the * given CPU/address space, due to the allocation of a new PCID. * * Zeroing can be performed remotely. Consequently, if a pmap is * inactive on a CPU, then a TLB shootdown for that pmap and CPU can * be initiated by an ordinary memory access to reset the target * CPU's generation count within the pmap. The CPU initiating the * TLB shootdown does not need to send an IPI to the target CPU. * * * PTI + PCID. The available PCIDs are divided into two sets: PCIDs * for complete (kernel) page tables, and PCIDs for user mode page * tables. A user PCID value is obtained from the kernel PCID value * by setting the highest bit, 11, to 1 (0x800 == PMAP_PCID_USER_PT). * * User space page tables are activated on return to user mode, by * loading pm_ucr3 into %cr3. If the PCPU(ucr3_load_mask) requests * clearing bit 63 of the loaded ucr3, this effectively causes * complete invalidation of the user mode TLB entries for the * current pmap. In which case, local invalidations of individual * pages in the user page table are skipped. * * * Local invalidation, all modes. If the requested invalidation is * for a specific address or the total invalidation of a currently * active pmap, then the TLB is flushed using INVLPG for a kernel * page table, and INVPCID(INVPCID_CTXGLOB)/invltlb_glob() for a * user space page table(s). * * If the INVPCID instruction is available, it is used to flush entries * from the kernel page table. * * * mode: PTI disabled, PCID present. The kernel reserves PCID 0 for its * address space, all other 4095 PCIDs are used for user mode spaces * as described above. A context switch allocates a new PCID if * the recorded PCID is zero or the recorded generation does not match * the CPU's generation, effectively flushing the TLB for this address space. * Total remote invalidation is performed by zeroing pm_gen for all CPUs. * local user page: INVLPG * local kernel page: INVLPG * local user total: INVPCID(CTX) * local kernel total: INVPCID(CTXGLOB) or invltlb_glob() * remote user page, inactive pmap: zero pm_gen * remote user page, active pmap: zero pm_gen + IPI:INVLPG * (Both actions are required to handle the aforementioned pm_active races.) * remote kernel page: IPI:INVLPG * remote user total, inactive pmap: zero pm_gen * remote user total, active pmap: zero pm_gen + IPI:(INVPCID(CTX) or * reload %cr3) * (See note above about pm_active races.) * remote kernel total: IPI:(INVPCID(CTXGLOB) or invltlb_glob()) * * PTI enabled, PCID present. * local user page: INVLPG for kpt, INVPCID(ADDR) or (INVLPG for ucr3) * for upt * local kernel page: INVLPG * local user total: INVPCID(CTX) or reload %cr3 for kpt, clear PCID_SAVE * on loading UCR3 into %cr3 for upt * local kernel total: INVPCID(CTXGLOB) or invltlb_glob() * remote user page, inactive pmap: zero pm_gen * remote user page, active pmap: zero pm_gen + IPI:(INVLPG for kpt, * INVPCID(ADDR) for upt) * remote kernel page: IPI:INVLPG * remote user total, inactive pmap: zero pm_gen * remote user total, active pmap: zero pm_gen + IPI:(INVPCID(CTX) for kpt, * clear PCID_SAVE on loading UCR3 into $cr3 for upt) * remote kernel total: IPI:(INVPCID(CTXGLOB) or invltlb_glob()) * * No PCID. * local user page: INVLPG * local kernel page: INVLPG * local user total: reload %cr3 * local kernel total: invltlb_glob() * remote user page, inactive pmap: - * remote user page, active pmap: IPI:INVLPG * remote kernel page: IPI:INVLPG * remote user total, inactive pmap: - * remote user total, active pmap: IPI:(reload %cr3) * remote kernel total: IPI:invltlb_glob() * Since on return to user mode, the reload of %cr3 with ucr3 causes * TLB invalidation, no specific action is required for user page table. * * EPT. EPT pmaps do not map KVA, all mappings are userspace. * XXX TODO */ #ifdef SMP /* * Interrupt the cpus that are executing in the guest context. * This will force the vcpu to exit and the cached EPT mappings * will be invalidated by the host before the next vmresume. */ static __inline void pmap_invalidate_ept(pmap_t pmap) { smr_seq_t goal; int ipinum; sched_pin(); KASSERT(!CPU_ISSET(curcpu, &pmap->pm_active), ("pmap_invalidate_ept: absurd pm_active")); /* * The TLB mappings associated with a vcpu context are not * flushed each time a different vcpu is chosen to execute. * * This is in contrast with a process's vtop mappings that * are flushed from the TLB on each context switch. * * Therefore we need to do more than just a TLB shootdown on * the active cpus in 'pmap->pm_active'. To do this we keep * track of the number of invalidations performed on this pmap. * * Each vcpu keeps a cache of this counter and compares it * just before a vmresume. If the counter is out-of-date an * invept will be done to flush stale mappings from the TLB. * * To ensure that all vCPU threads have observed the new counter * value before returning, we use SMR. Ordering is important here: * the VMM enters an SMR read section before loading the counter * and after updating the pm_active bit set. Thus, pm_active is * a superset of active readers, and any reader that has observed * the goal has observed the new counter value. */ atomic_add_long(&pmap->pm_eptgen, 1); goal = smr_advance(pmap->pm_eptsmr); /* * Force the vcpu to exit and trap back into the hypervisor. */ ipinum = pmap->pm_flags & PMAP_NESTED_IPIMASK; ipi_selected(pmap->pm_active, ipinum); sched_unpin(); /* * Ensure that all active vCPUs will observe the new generation counter * value before executing any more guest instructions. */ smr_wait(pmap->pm_eptsmr, goal); } static inline void pmap_invalidate_preipi_pcid(pmap_t pmap) { u_int cpuid, i; sched_pin(); cpuid = PCPU_GET(cpuid); if (pmap != PCPU_GET(curpmap)) cpuid = 0xffffffff; /* An impossible value */ CPU_FOREACH(i) { if (cpuid != i) pmap->pm_pcids[i].pm_gen = 0; } /* * The fence is between stores to pm_gen and the read of the * pm_active mask. We need to ensure that it is impossible * for us to miss the bit update in pm_active and * simultaneously observe a non-zero pm_gen in * pmap_activate_sw(), otherwise TLB update is missed. * Without the fence, IA32 allows such an outcome. Note that * pm_active is updated by a locked operation, which provides * the reciprocal fence. */ atomic_thread_fence_seq_cst(); } static void pmap_invalidate_preipi_nopcid(pmap_t pmap __unused) { sched_pin(); } DEFINE_IFUNC(static, void, pmap_invalidate_preipi, (pmap_t)) { return (pmap_pcid_enabled ? pmap_invalidate_preipi_pcid : pmap_invalidate_preipi_nopcid); } static inline void pmap_invalidate_page_pcid_cb(pmap_t pmap, vm_offset_t va, const bool invpcid_works1) { struct invpcid_descr d; uint64_t kcr3, ucr3; uint32_t pcid; u_int cpuid; /* * Because pm_pcid is recalculated on a context switch, we * must ensure there is no preemption, not just pinning. * Otherwise, we might use a stale value below. */ CRITICAL_ASSERT(curthread); /* * No need to do anything with user page tables invalidation * if there is no user page table, or invalidation is deferred * until the return to userspace. ucr3_load_mask is stable * because we have preemption disabled. */ if (pmap->pm_ucr3 == PMAP_NO_CR3 || PCPU_GET(ucr3_load_mask) != PMAP_UCR3_NOMASK) return; cpuid = PCPU_GET(cpuid); pcid = pmap->pm_pcids[cpuid].pm_pcid; if (invpcid_works1) { d.pcid = pcid | PMAP_PCID_USER_PT; d.pad = 0; d.addr = va; invpcid(&d, INVPCID_ADDR); } else { kcr3 = pmap->pm_cr3 | pcid | CR3_PCID_SAVE; ucr3 = pmap->pm_ucr3 | pcid | PMAP_PCID_USER_PT | CR3_PCID_SAVE; pmap_pti_pcid_invlpg(ucr3, kcr3, va); } } static void pmap_invalidate_page_pcid_invpcid_cb(pmap_t pmap, vm_offset_t va) { pmap_invalidate_page_pcid_cb(pmap, va, true); } static void pmap_invalidate_page_pcid_noinvpcid_cb(pmap_t pmap, vm_offset_t va) { pmap_invalidate_page_pcid_cb(pmap, va, false); } static void pmap_invalidate_page_nopcid_cb(pmap_t pmap __unused, vm_offset_t va __unused) { } DEFINE_IFUNC(static, void, pmap_invalidate_page_cb, (pmap_t, vm_offset_t)) { if (pmap_pcid_enabled) return (invpcid_works ? pmap_invalidate_page_pcid_invpcid_cb : pmap_invalidate_page_pcid_noinvpcid_cb); return (pmap_invalidate_page_nopcid_cb); } static void pmap_invalidate_page_curcpu_cb(pmap_t pmap, vm_offset_t va, vm_offset_t addr2 __unused) { if (pmap == kernel_pmap) { invlpg(va); } else if (pmap == PCPU_GET(curpmap)) { invlpg(va); pmap_invalidate_page_cb(pmap, va); } } void pmap_invalidate_page(pmap_t pmap, vm_offset_t va) { if (pmap_type_guest(pmap)) { pmap_invalidate_ept(pmap); return; } KASSERT(pmap->pm_type == PT_X86, ("pmap_invalidate_page: invalid type %d", pmap->pm_type)); pmap_invalidate_preipi(pmap); smp_masked_invlpg(va, pmap, pmap_invalidate_page_curcpu_cb); } /* 4k PTEs -- Chosen to exceed the total size of Broadwell L2 TLB */ #define PMAP_INVLPG_THRESHOLD (4 * 1024 * PAGE_SIZE) static void pmap_invalidate_range_pcid_cb(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, const bool invpcid_works1) { struct invpcid_descr d; uint64_t kcr3, ucr3; uint32_t pcid; u_int cpuid; CRITICAL_ASSERT(curthread); if (pmap != PCPU_GET(curpmap) || pmap->pm_ucr3 == PMAP_NO_CR3 || PCPU_GET(ucr3_load_mask) != PMAP_UCR3_NOMASK) return; cpuid = PCPU_GET(cpuid); pcid = pmap->pm_pcids[cpuid].pm_pcid; if (invpcid_works1) { d.pcid = pcid | PMAP_PCID_USER_PT; d.pad = 0; for (d.addr = sva; d.addr < eva; d.addr += PAGE_SIZE) invpcid(&d, INVPCID_ADDR); } else { kcr3 = pmap->pm_cr3 | pcid | CR3_PCID_SAVE; ucr3 = pmap->pm_ucr3 | pcid | PMAP_PCID_USER_PT | CR3_PCID_SAVE; pmap_pti_pcid_invlrng(ucr3, kcr3, sva, eva); } } static void pmap_invalidate_range_pcid_invpcid_cb(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { pmap_invalidate_range_pcid_cb(pmap, sva, eva, true); } static void pmap_invalidate_range_pcid_noinvpcid_cb(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { pmap_invalidate_range_pcid_cb(pmap, sva, eva, false); } static void pmap_invalidate_range_nopcid_cb(pmap_t pmap __unused, vm_offset_t sva __unused, vm_offset_t eva __unused) { } DEFINE_IFUNC(static, void, pmap_invalidate_range_cb, (pmap_t, vm_offset_t, vm_offset_t)) { if (pmap_pcid_enabled) return (invpcid_works ? pmap_invalidate_range_pcid_invpcid_cb : pmap_invalidate_range_pcid_noinvpcid_cb); return (pmap_invalidate_range_nopcid_cb); } static void pmap_invalidate_range_curcpu_cb(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { vm_offset_t addr; if (pmap == kernel_pmap) { for (addr = sva; addr < eva; addr += PAGE_SIZE) invlpg(addr); } else if (pmap == PCPU_GET(curpmap)) { for (addr = sva; addr < eva; addr += PAGE_SIZE) invlpg(addr); pmap_invalidate_range_cb(pmap, sva, eva); } } void pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { if (eva - sva >= PMAP_INVLPG_THRESHOLD) { pmap_invalidate_all(pmap); return; } if (pmap_type_guest(pmap)) { pmap_invalidate_ept(pmap); return; } KASSERT(pmap->pm_type == PT_X86, ("pmap_invalidate_range: invalid type %d", pmap->pm_type)); pmap_invalidate_preipi(pmap); smp_masked_invlpg_range(sva, eva, pmap, pmap_invalidate_range_curcpu_cb); } static inline void pmap_invalidate_all_pcid_cb(pmap_t pmap, bool invpcid_works1) { struct invpcid_descr d; uint64_t kcr3; uint32_t pcid; u_int cpuid; if (pmap == kernel_pmap) { if (invpcid_works1) { bzero(&d, sizeof(d)); invpcid(&d, INVPCID_CTXGLOB); } else { invltlb_glob(); } } else if (pmap == PCPU_GET(curpmap)) { CRITICAL_ASSERT(curthread); cpuid = PCPU_GET(cpuid); pcid = pmap->pm_pcids[cpuid].pm_pcid; if (invpcid_works1) { d.pcid = pcid; d.pad = 0; d.addr = 0; invpcid(&d, INVPCID_CTX); } else { kcr3 = pmap->pm_cr3 | pcid; load_cr3(kcr3); } if (pmap->pm_ucr3 != PMAP_NO_CR3) PCPU_SET(ucr3_load_mask, ~CR3_PCID_SAVE); } } static void pmap_invalidate_all_pcid_invpcid_cb(pmap_t pmap) { pmap_invalidate_all_pcid_cb(pmap, true); } static void pmap_invalidate_all_pcid_noinvpcid_cb(pmap_t pmap) { pmap_invalidate_all_pcid_cb(pmap, false); } static void pmap_invalidate_all_nopcid_cb(pmap_t pmap) { if (pmap == kernel_pmap) invltlb_glob(); else if (pmap == PCPU_GET(curpmap)) invltlb(); } DEFINE_IFUNC(static, void, pmap_invalidate_all_cb, (pmap_t)) { if (pmap_pcid_enabled) return (invpcid_works ? pmap_invalidate_all_pcid_invpcid_cb : pmap_invalidate_all_pcid_noinvpcid_cb); return (pmap_invalidate_all_nopcid_cb); } static void pmap_invalidate_all_curcpu_cb(pmap_t pmap, vm_offset_t addr1 __unused, vm_offset_t addr2 __unused) { pmap_invalidate_all_cb(pmap); } void pmap_invalidate_all(pmap_t pmap) { if (pmap_type_guest(pmap)) { pmap_invalidate_ept(pmap); return; } KASSERT(pmap->pm_type == PT_X86, ("pmap_invalidate_all: invalid type %d", pmap->pm_type)); pmap_invalidate_preipi(pmap); smp_masked_invltlb(pmap, pmap_invalidate_all_curcpu_cb); } static void pmap_invalidate_cache_curcpu_cb(pmap_t pmap __unused, vm_offset_t va __unused, vm_offset_t addr2 __unused) { wbinvd(); } void pmap_invalidate_cache(void) { sched_pin(); smp_cache_flush(pmap_invalidate_cache_curcpu_cb); } struct pde_action { cpuset_t invalidate; /* processors that invalidate their TLB */ pmap_t pmap; vm_offset_t va; pd_entry_t *pde; pd_entry_t newpde; u_int store; /* processor that updates the PDE */ }; static void pmap_update_pde_action(void *arg) { struct pde_action *act = arg; if (act->store == PCPU_GET(cpuid)) pmap_update_pde_store(act->pmap, act->pde, act->newpde); } static void pmap_update_pde_teardown(void *arg) { struct pde_action *act = arg; if (CPU_ISSET(PCPU_GET(cpuid), &act->invalidate)) pmap_update_pde_invalidate(act->pmap, act->va, act->newpde); } /* * Change the page size for the specified virtual address in a way that * prevents any possibility of the TLB ever having two entries that map the * same virtual address using different page sizes. This is the recommended * workaround for Erratum 383 on AMD Family 10h processors. It prevents a * machine check exception for a TLB state that is improperly diagnosed as a * hardware error. */ static void pmap_update_pde(pmap_t pmap, vm_offset_t va, pd_entry_t *pde, pd_entry_t newpde) { struct pde_action act; cpuset_t active, other_cpus; u_int cpuid; sched_pin(); cpuid = PCPU_GET(cpuid); other_cpus = all_cpus; CPU_CLR(cpuid, &other_cpus); if (pmap == kernel_pmap || pmap_type_guest(pmap)) active = all_cpus; else { active = pmap->pm_active; } if (CPU_OVERLAP(&active, &other_cpus)) { act.store = cpuid; act.invalidate = active; act.va = va; act.pmap = pmap; act.pde = pde; act.newpde = newpde; CPU_SET(cpuid, &active); smp_rendezvous_cpus(active, smp_no_rendezvous_barrier, pmap_update_pde_action, pmap_update_pde_teardown, &act); } else { pmap_update_pde_store(pmap, pde, newpde); if (CPU_ISSET(cpuid, &active)) pmap_update_pde_invalidate(pmap, va, newpde); } sched_unpin(); } #else /* !SMP */ /* * Normal, non-SMP, invalidation functions. */ void pmap_invalidate_page(pmap_t pmap, vm_offset_t va) { struct invpcid_descr d; uint64_t kcr3, ucr3; uint32_t pcid; if (pmap->pm_type == PT_RVI || pmap->pm_type == PT_EPT) { pmap->pm_eptgen++; return; } KASSERT(pmap->pm_type == PT_X86, ("pmap_invalidate_range: unknown type %d", pmap->pm_type)); if (pmap == kernel_pmap || pmap == PCPU_GET(curpmap)) { invlpg(va); if (pmap == PCPU_GET(curpmap) && pmap_pcid_enabled && pmap->pm_ucr3 != PMAP_NO_CR3) { critical_enter(); pcid = pmap->pm_pcids[0].pm_pcid; if (invpcid_works) { d.pcid = pcid | PMAP_PCID_USER_PT; d.pad = 0; d.addr = va; invpcid(&d, INVPCID_ADDR); } else { kcr3 = pmap->pm_cr3 | pcid | CR3_PCID_SAVE; ucr3 = pmap->pm_ucr3 | pcid | PMAP_PCID_USER_PT | CR3_PCID_SAVE; pmap_pti_pcid_invlpg(ucr3, kcr3, va); } critical_exit(); } } else if (pmap_pcid_enabled) pmap->pm_pcids[0].pm_gen = 0; } void pmap_invalidate_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { struct invpcid_descr d; vm_offset_t addr; uint64_t kcr3, ucr3; if (pmap->pm_type == PT_RVI || pmap->pm_type == PT_EPT) { pmap->pm_eptgen++; return; } KASSERT(pmap->pm_type == PT_X86, ("pmap_invalidate_range: unknown type %d", pmap->pm_type)); if (pmap == kernel_pmap || pmap == PCPU_GET(curpmap)) { for (addr = sva; addr < eva; addr += PAGE_SIZE) invlpg(addr); if (pmap == PCPU_GET(curpmap) && pmap_pcid_enabled && pmap->pm_ucr3 != PMAP_NO_CR3) { critical_enter(); if (invpcid_works) { d.pcid = pmap->pm_pcids[0].pm_pcid | PMAP_PCID_USER_PT; d.pad = 0; d.addr = sva; for (; d.addr < eva; d.addr += PAGE_SIZE) invpcid(&d, INVPCID_ADDR); } else { kcr3 = pmap->pm_cr3 | pmap->pm_pcids[0]. pm_pcid | CR3_PCID_SAVE; ucr3 = pmap->pm_ucr3 | pmap->pm_pcids[0]. pm_pcid | PMAP_PCID_USER_PT | CR3_PCID_SAVE; pmap_pti_pcid_invlrng(ucr3, kcr3, sva, eva); } critical_exit(); } } else if (pmap_pcid_enabled) { pmap->pm_pcids[0].pm_gen = 0; } } void pmap_invalidate_all(pmap_t pmap) { struct invpcid_descr d; uint64_t kcr3, ucr3; if (pmap->pm_type == PT_RVI || pmap->pm_type == PT_EPT) { pmap->pm_eptgen++; return; } KASSERT(pmap->pm_type == PT_X86, ("pmap_invalidate_all: unknown type %d", pmap->pm_type)); if (pmap == kernel_pmap) { if (pmap_pcid_enabled && invpcid_works) { bzero(&d, sizeof(d)); invpcid(&d, INVPCID_CTXGLOB); } else { invltlb_glob(); } } else if (pmap == PCPU_GET(curpmap)) { if (pmap_pcid_enabled) { critical_enter(); if (invpcid_works) { d.pcid = pmap->pm_pcids[0].pm_pcid; d.pad = 0; d.addr = 0; invpcid(&d, INVPCID_CTX); if (pmap->pm_ucr3 != PMAP_NO_CR3) { d.pcid |= PMAP_PCID_USER_PT; invpcid(&d, INVPCID_CTX); } } else { kcr3 = pmap->pm_cr3 | pmap->pm_pcids[0].pm_pcid; if (pmap->pm_ucr3 != PMAP_NO_CR3) { ucr3 = pmap->pm_ucr3 | pmap->pm_pcids[ 0].pm_pcid | PMAP_PCID_USER_PT; pmap_pti_pcid_invalidate(ucr3, kcr3); } else load_cr3(kcr3); } critical_exit(); } else { invltlb(); } } else if (pmap_pcid_enabled) { pmap->pm_pcids[0].pm_gen = 0; } } PMAP_INLINE void pmap_invalidate_cache(void) { wbinvd(); } static void pmap_update_pde(pmap_t pmap, vm_offset_t va, pd_entry_t *pde, pd_entry_t newpde) { pmap_update_pde_store(pmap, pde, newpde); if (pmap == kernel_pmap || pmap == PCPU_GET(curpmap)) pmap_update_pde_invalidate(pmap, va, newpde); else pmap->pm_pcids[0].pm_gen = 0; } #endif /* !SMP */ static void pmap_invalidate_pde_page(pmap_t pmap, vm_offset_t va, pd_entry_t pde) { /* * When the PDE has PG_PROMOTED set, the 2MB page mapping was created * by a promotion that did not invalidate the 512 4KB page mappings * that might exist in the TLB. Consequently, at this point, the TLB * may hold both 4KB and 2MB page mappings for the address range [va, * va + NBPDR). Therefore, the entire range must be invalidated here. * In contrast, when PG_PROMOTED is clear, the TLB will not hold any * 4KB page mappings for the address range [va, va + NBPDR), and so a * single INVLPG suffices to invalidate the 2MB page mapping from the * TLB. */ if ((pde & PG_PROMOTED) != 0) pmap_invalidate_range(pmap, va, va + NBPDR - 1); else pmap_invalidate_page(pmap, va); } DEFINE_IFUNC(, void, pmap_invalidate_cache_range, (vm_offset_t sva, vm_offset_t eva)) { if ((cpu_feature & CPUID_SS) != 0) return (pmap_invalidate_cache_range_selfsnoop); if ((cpu_feature & CPUID_CLFSH) != 0) return (pmap_force_invalidate_cache_range); return (pmap_invalidate_cache_range_all); } #define PMAP_CLFLUSH_THRESHOLD (2 * 1024 * 1024) static void pmap_invalidate_cache_range_check_align(vm_offset_t sva, vm_offset_t eva) { KASSERT((sva & PAGE_MASK) == 0, ("pmap_invalidate_cache_range: sva not page-aligned")); KASSERT((eva & PAGE_MASK) == 0, ("pmap_invalidate_cache_range: eva not page-aligned")); } static void pmap_invalidate_cache_range_selfsnoop(vm_offset_t sva, vm_offset_t eva) { pmap_invalidate_cache_range_check_align(sva, eva); } void pmap_force_invalidate_cache_range(vm_offset_t sva, vm_offset_t eva) { sva &= ~(vm_offset_t)(cpu_clflush_line_size - 1); /* * XXX: Some CPUs fault, hang, or trash the local APIC * registers if we use CLFLUSH on the local APIC range. The * local APIC is always uncached, so we don't need to flush * for that range anyway. */ if (pmap_kextract(sva) == lapic_paddr) return; if ((cpu_stdext_feature & CPUID_STDEXT_CLFLUSHOPT) != 0) { /* * Do per-cache line flush. Use a locked * instruction to insure that previous stores are * included in the write-back. The processor * propagates flush to other processors in the cache * coherence domain. */ atomic_thread_fence_seq_cst(); for (; sva < eva; sva += cpu_clflush_line_size) clflushopt(sva); atomic_thread_fence_seq_cst(); } else { /* * Writes are ordered by CLFLUSH on Intel CPUs. */ if (cpu_vendor_id != CPU_VENDOR_INTEL) mfence(); for (; sva < eva; sva += cpu_clflush_line_size) clflush(sva); if (cpu_vendor_id != CPU_VENDOR_INTEL) mfence(); } } static void pmap_invalidate_cache_range_all(vm_offset_t sva, vm_offset_t eva) { pmap_invalidate_cache_range_check_align(sva, eva); pmap_invalidate_cache(); } /* * Remove the specified set of pages from the data and instruction caches. * * In contrast to pmap_invalidate_cache_range(), this function does not * rely on the CPU's self-snoop feature, because it is intended for use * when moving pages into a different cache domain. */ void pmap_invalidate_cache_pages(vm_page_t *pages, int count) { vm_offset_t daddr, eva; int i; bool useclflushopt; useclflushopt = (cpu_stdext_feature & CPUID_STDEXT_CLFLUSHOPT) != 0; if (count >= PMAP_CLFLUSH_THRESHOLD / PAGE_SIZE || ((cpu_feature & CPUID_CLFSH) == 0 && !useclflushopt)) pmap_invalidate_cache(); else { if (useclflushopt) atomic_thread_fence_seq_cst(); else if (cpu_vendor_id != CPU_VENDOR_INTEL) mfence(); for (i = 0; i < count; i++) { daddr = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pages[i])); eva = daddr + PAGE_SIZE; for (; daddr < eva; daddr += cpu_clflush_line_size) { if (useclflushopt) clflushopt(daddr); else clflush(daddr); } } if (useclflushopt) atomic_thread_fence_seq_cst(); else if (cpu_vendor_id != CPU_VENDOR_INTEL) mfence(); } } void pmap_flush_cache_range(vm_offset_t sva, vm_offset_t eva) { pmap_invalidate_cache_range_check_align(sva, eva); if ((cpu_stdext_feature & CPUID_STDEXT_CLWB) == 0) { pmap_force_invalidate_cache_range(sva, eva); return; } /* See comment in pmap_force_invalidate_cache_range(). */ if (pmap_kextract(sva) == lapic_paddr) return; atomic_thread_fence_seq_cst(); for (; sva < eva; sva += cpu_clflush_line_size) clwb(sva); atomic_thread_fence_seq_cst(); } void pmap_flush_cache_phys_range(vm_paddr_t spa, vm_paddr_t epa, vm_memattr_t mattr) { pt_entry_t *pte; vm_offset_t vaddr; int error, pte_bits; KASSERT((spa & PAGE_MASK) == 0, ("pmap_flush_cache_phys_range: spa not page-aligned")); KASSERT((epa & PAGE_MASK) == 0, ("pmap_flush_cache_phys_range: epa not page-aligned")); if (spa < dmaplimit) { pmap_flush_cache_range(PHYS_TO_DMAP(spa), PHYS_TO_DMAP(MIN( dmaplimit, epa))); if (dmaplimit >= epa) return; spa = dmaplimit; } pte_bits = pmap_cache_bits(kernel_pmap, mattr, 0) | X86_PG_RW | X86_PG_V; error = vmem_alloc(kernel_arena, PAGE_SIZE, M_BESTFIT | M_WAITOK, &vaddr); KASSERT(error == 0, ("vmem_alloc failed: %d", error)); pte = vtopte(vaddr); for (; spa < epa; spa += PAGE_SIZE) { sched_pin(); pte_store(pte, spa | pte_bits); invlpg(vaddr); /* XXXKIB atomic inside flush_cache_range are excessive */ pmap_flush_cache_range(vaddr, vaddr + PAGE_SIZE); sched_unpin(); } vmem_free(kernel_arena, vaddr, PAGE_SIZE); } /* * Routine: pmap_extract * Function: * Extract the physical page address associated * with the given map/virtual_address pair. */ vm_paddr_t pmap_extract(pmap_t pmap, vm_offset_t va) { pdp_entry_t *pdpe; pd_entry_t *pde; pt_entry_t *pte, PG_V; vm_paddr_t pa; pa = 0; PG_V = pmap_valid_bit(pmap); PMAP_LOCK(pmap); pdpe = pmap_pdpe(pmap, va); if (pdpe != NULL && (*pdpe & PG_V) != 0) { if ((*pdpe & PG_PS) != 0) pa = (*pdpe & PG_PS_FRAME) | (va & PDPMASK); else { pde = pmap_pdpe_to_pde(pdpe, va); if ((*pde & PG_V) != 0) { if ((*pde & PG_PS) != 0) { pa = (*pde & PG_PS_FRAME) | (va & PDRMASK); } else { pte = pmap_pde_to_pte(pde, va); pa = (*pte & PG_FRAME) | (va & PAGE_MASK); } } } } PMAP_UNLOCK(pmap); return (pa); } /* * Routine: pmap_extract_and_hold * Function: * Atomically extract and hold the physical page * with the given pmap and virtual address pair * if that mapping permits the given protection. */ vm_page_t pmap_extract_and_hold(pmap_t pmap, vm_offset_t va, vm_prot_t prot) { pdp_entry_t pdpe, *pdpep; pd_entry_t pde, *pdep; pt_entry_t pte, PG_RW, PG_V; vm_page_t m; m = NULL; PG_RW = pmap_rw_bit(pmap); PG_V = pmap_valid_bit(pmap); PMAP_LOCK(pmap); pdpep = pmap_pdpe(pmap, va); if (pdpep == NULL || ((pdpe = *pdpep) & PG_V) == 0) goto out; if ((pdpe & PG_PS) != 0) { if ((pdpe & PG_RW) == 0 && (prot & VM_PROT_WRITE) != 0) goto out; m = PHYS_TO_VM_PAGE((pdpe & PG_PS_FRAME) | (va & PDPMASK)); goto check_page; } pdep = pmap_pdpe_to_pde(pdpep, va); if (pdep == NULL || ((pde = *pdep) & PG_V) == 0) goto out; if ((pde & PG_PS) != 0) { if ((pde & PG_RW) == 0 && (prot & VM_PROT_WRITE) != 0) goto out; m = PHYS_TO_VM_PAGE((pde & PG_PS_FRAME) | (va & PDRMASK)); goto check_page; } pte = *pmap_pde_to_pte(pdep, va); if ((pte & PG_V) == 0 || ((pte & PG_RW) == 0 && (prot & VM_PROT_WRITE) != 0)) goto out; m = PHYS_TO_VM_PAGE(pte & PG_FRAME); check_page: if (m != NULL && !vm_page_wire_mapped(m)) m = NULL; out: PMAP_UNLOCK(pmap); return (m); } vm_paddr_t pmap_kextract(vm_offset_t va) { pd_entry_t pde; vm_paddr_t pa; if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) { pa = DMAP_TO_PHYS(va); } else if (PMAP_ADDRESS_IN_LARGEMAP(va)) { pa = pmap_large_map_kextract(va); } else { pde = *vtopde(va); if (pde & PG_PS) { pa = (pde & PG_PS_FRAME) | (va & PDRMASK); } else { /* * Beware of a concurrent promotion that changes the * PDE at this point! For example, vtopte() must not * be used to access the PTE because it would use the * new PDE. It is, however, safe to use the old PDE * because the page table page is preserved by the * promotion. */ pa = *pmap_pde_to_pte(&pde, va); pa = (pa & PG_FRAME) | (va & PAGE_MASK); } } return (pa); } /*************************************************** * Low level mapping routines..... ***************************************************/ /* * Add a wired page to the kva. * Note: not SMP coherent. */ PMAP_INLINE void pmap_kenter(vm_offset_t va, vm_paddr_t pa) { pt_entry_t *pte; pte = vtopte(va); pte_store(pte, pa | X86_PG_RW | X86_PG_V | pg_g | pg_nx); } static __inline void pmap_kenter_attr(vm_offset_t va, vm_paddr_t pa, int mode) { pt_entry_t *pte; int cache_bits; pte = vtopte(va); cache_bits = pmap_cache_bits(kernel_pmap, mode, 0); pte_store(pte, pa | X86_PG_RW | X86_PG_V | pg_g | pg_nx | cache_bits); } /* * Remove a page from the kernel pagetables. * Note: not SMP coherent. */ PMAP_INLINE void pmap_kremove(vm_offset_t va) { pt_entry_t *pte; pte = vtopte(va); pte_clear(pte); } /* * Used to map a range of physical addresses into kernel * virtual address space. * * The value passed in '*virt' is a suggested virtual address for * the mapping. Architectures which can support a direct-mapped * physical to virtual region can return the appropriate address * within that region, leaving '*virt' unchanged. Other * architectures should map the pages starting at '*virt' and * update '*virt' with the first usable address after the mapped * region. */ vm_offset_t pmap_map(vm_offset_t *virt, vm_paddr_t start, vm_paddr_t end, int prot) { return PHYS_TO_DMAP(start); } /* * Add a list of wired pages to the kva * this routine is only used for temporary * kernel mappings that do not need to have * page modification or references recorded. * Note that old mappings are simply written * over. The page *must* be wired. * Note: SMP coherent. Uses a ranged shootdown IPI. */ void pmap_qenter(vm_offset_t sva, vm_page_t *ma, int count) { pt_entry_t *endpte, oldpte, pa, *pte; vm_page_t m; int cache_bits; oldpte = 0; pte = vtopte(sva); endpte = pte + count; while (pte < endpte) { m = *ma++; cache_bits = pmap_cache_bits(kernel_pmap, m->md.pat_mode, 0); pa = VM_PAGE_TO_PHYS(m) | cache_bits; if ((*pte & (PG_FRAME | X86_PG_PTE_CACHE)) != pa) { oldpte |= *pte; pte_store(pte, pa | pg_g | pg_nx | X86_PG_RW | X86_PG_V); } pte++; } if (__predict_false((oldpte & X86_PG_V) != 0)) pmap_invalidate_range(kernel_pmap, sva, sva + count * PAGE_SIZE); } /* * This routine tears out page mappings from the * kernel -- it is meant only for temporary mappings. * Note: SMP coherent. Uses a ranged shootdown IPI. */ void pmap_qremove(vm_offset_t sva, int count) { vm_offset_t va; va = sva; while (count-- > 0) { KASSERT(va >= VM_MIN_KERNEL_ADDRESS, ("usermode va %lx", va)); pmap_kremove(va); va += PAGE_SIZE; } pmap_invalidate_range(kernel_pmap, sva, va); } /*************************************************** * Page table page management routines..... ***************************************************/ /* * Schedule the specified unused page table page to be freed. Specifically, * add the page to the specified list of pages that will be released to the * physical memory manager after the TLB has been updated. */ static __inline void pmap_add_delayed_free_list(vm_page_t m, struct spglist *free, boolean_t set_PG_ZERO) { if (set_PG_ZERO) m->flags |= PG_ZERO; else m->flags &= ~PG_ZERO; SLIST_INSERT_HEAD(free, m, plinks.s.ss); } /* * Inserts the specified page table page into the specified pmap's collection * of idle page table pages. Each of a pmap's page table pages is responsible * for mapping a distinct range of virtual addresses. The pmap's collection is * ordered by this virtual address range. * * If "promoted" is false, then the page table page "mpte" must be zero filled. */ static __inline int pmap_insert_pt_page(pmap_t pmap, vm_page_t mpte, bool promoted) { PMAP_LOCK_ASSERT(pmap, MA_OWNED); mpte->valid = promoted ? VM_PAGE_BITS_ALL : 0; return (vm_radix_insert(&pmap->pm_root, mpte)); } /* * Removes the page table page mapping the specified virtual address from the * specified pmap's collection of idle page table pages, and returns it. * Otherwise, returns NULL if there is no page table page corresponding to the * specified virtual address. */ static __inline vm_page_t pmap_remove_pt_page(pmap_t pmap, vm_offset_t va) { PMAP_LOCK_ASSERT(pmap, MA_OWNED); return (vm_radix_remove(&pmap->pm_root, pmap_pde_pindex(va))); } /* * Decrements a page table page's reference count, which is used to record the * number of valid page table entries within the page. If the reference count * drops to zero, then the page table page is unmapped. Returns TRUE if the * page table page was unmapped and FALSE otherwise. */ static inline boolean_t pmap_unwire_ptp(pmap_t pmap, vm_offset_t va, vm_page_t m, struct spglist *free) { --m->ref_count; if (m->ref_count == 0) { _pmap_unwire_ptp(pmap, va, m, free); return (TRUE); } else return (FALSE); } static void _pmap_unwire_ptp(pmap_t pmap, vm_offset_t va, vm_page_t m, struct spglist *free) { pml5_entry_t *pml5; pml4_entry_t *pml4; pdp_entry_t *pdp; pd_entry_t *pd; vm_page_t pdpg, pdppg, pml4pg; PMAP_LOCK_ASSERT(pmap, MA_OWNED); /* * unmap the page table page */ if (m->pindex >= NUPDE + NUPDPE + NUPML4E) { /* PML4 page */ MPASS(pmap_is_la57(pmap)); pml5 = pmap_pml5e(pmap, va); *pml5 = 0; if (pmap->pm_pmltopu != NULL && va <= VM_MAXUSER_ADDRESS) { pml5 = pmap_pml5e_u(pmap, va); *pml5 = 0; } } else if (m->pindex >= NUPDE + NUPDPE) { /* PDP page */ pml4 = pmap_pml4e(pmap, va); *pml4 = 0; if (!pmap_is_la57(pmap) && pmap->pm_pmltopu != NULL && va <= VM_MAXUSER_ADDRESS) { pml4 = pmap_pml4e_u(pmap, va); *pml4 = 0; } } else if (m->pindex >= NUPDE) { /* PD page */ pdp = pmap_pdpe(pmap, va); *pdp = 0; } else { /* PTE page */ pd = pmap_pde(pmap, va); *pd = 0; } if (m->pindex < NUPDE) { /* We just released a PT, unhold the matching PD */ pdpg = PHYS_TO_VM_PAGE(*pmap_pdpe(pmap, va) & PG_FRAME); pmap_unwire_ptp(pmap, va, pdpg, free); } else if (m->pindex < NUPDE + NUPDPE) { /* We just released a PD, unhold the matching PDP */ pdppg = PHYS_TO_VM_PAGE(*pmap_pml4e(pmap, va) & PG_FRAME); pmap_unwire_ptp(pmap, va, pdppg, free); } else if (m->pindex < NUPDE + NUPDPE + NUPML4E && pmap_is_la57(pmap)) { /* We just released a PDP, unhold the matching PML4 */ pml4pg = PHYS_TO_VM_PAGE(*pmap_pml5e(pmap, va) & PG_FRAME); pmap_unwire_ptp(pmap, va, pml4pg, free); } pmap_pt_page_count_adj(pmap, -1); /* * Put page on a list so that it is released after * *ALL* TLB shootdown is done */ pmap_add_delayed_free_list(m, free, TRUE); } /* * After removing a page table entry, this routine is used to * conditionally free the page, and manage the reference count. */ static int pmap_unuse_pt(pmap_t pmap, vm_offset_t va, pd_entry_t ptepde, struct spglist *free) { vm_page_t mpte; if (va >= VM_MAXUSER_ADDRESS) return (0); KASSERT(ptepde != 0, ("pmap_unuse_pt: ptepde != 0")); mpte = PHYS_TO_VM_PAGE(ptepde & PG_FRAME); return (pmap_unwire_ptp(pmap, va, mpte, free)); } /* * Release a page table page reference after a failed attempt to create a * mapping. */ static void pmap_abort_ptp(pmap_t pmap, vm_offset_t va, vm_page_t mpte) { struct spglist free; SLIST_INIT(&free); if (pmap_unwire_ptp(pmap, va, mpte, &free)) { /* * Although "va" was never mapped, paging-structure caches * could nonetheless have entries that refer to the freed * page table pages. Invalidate those entries. */ pmap_invalidate_page(pmap, va); vm_page_free_pages_toq(&free, true); } } void pmap_pinit0(pmap_t pmap) { struct proc *p; struct thread *td; int i; PMAP_LOCK_INIT(pmap); pmap->pm_pmltop = kernel_pmap->pm_pmltop; pmap->pm_pmltopu = NULL; pmap->pm_cr3 = kernel_pmap->pm_cr3; /* hack to keep pmap_pti_pcid_invalidate() alive */ pmap->pm_ucr3 = PMAP_NO_CR3; vm_radix_init(&pmap->pm_root); CPU_ZERO(&pmap->pm_active); TAILQ_INIT(&pmap->pm_pvchunk); bzero(&pmap->pm_stats, sizeof pmap->pm_stats); pmap->pm_flags = pmap_flags; CPU_FOREACH(i) { pmap->pm_pcids[i].pm_pcid = PMAP_PCID_KERN + 1; pmap->pm_pcids[i].pm_gen = 1; } pmap_activate_boot(pmap); td = curthread; if (pti) { p = td->td_proc; PROC_LOCK(p); p->p_md.md_flags |= P_MD_KPTI; PROC_UNLOCK(p); } pmap_thread_init_invl_gen(td); if ((cpu_stdext_feature2 & CPUID_STDEXT2_PKU) != 0) { pmap_pkru_ranges_zone = uma_zcreate("pkru ranges", sizeof(struct pmap_pkru_range), NULL, NULL, NULL, NULL, UMA_ALIGN_PTR, 0); } } void pmap_pinit_pml4(vm_page_t pml4pg) { pml4_entry_t *pm_pml4; int i; pm_pml4 = (pml4_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pml4pg)); /* Wire in kernel global address entries. */ for (i = 0; i < NKPML4E; i++) { pm_pml4[KPML4BASE + i] = (KPDPphys + ptoa(i)) | X86_PG_RW | X86_PG_V; } #ifdef KASAN for (i = 0; i < NKASANPML4E; i++) { pm_pml4[KASANPML4I + i] = (KASANPDPphys + ptoa(i)) | X86_PG_RW | X86_PG_V | pg_nx; } #endif for (i = 0; i < ndmpdpphys; i++) { pm_pml4[DMPML4I + i] = (DMPDPphys + ptoa(i)) | X86_PG_RW | X86_PG_V; } /* install self-referential address mapping entry(s) */ pm_pml4[PML4PML4I] = VM_PAGE_TO_PHYS(pml4pg) | X86_PG_V | X86_PG_RW | X86_PG_A | X86_PG_M; /* install large map entries if configured */ for (i = 0; i < lm_ents; i++) pm_pml4[LMSPML4I + i] = kernel_pmap->pm_pmltop[LMSPML4I + i]; } void pmap_pinit_pml5(vm_page_t pml5pg) { pml5_entry_t *pm_pml5; pm_pml5 = (pml5_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pml5pg)); /* * Add pml5 entry at top of KVA pointing to existing pml4 table, * entering all existing kernel mappings into level 5 table. */ pm_pml5[pmap_pml5e_index(UPT_MAX_ADDRESS)] = KPML4phys | X86_PG_V | X86_PG_RW | X86_PG_A | X86_PG_M | pg_g | pmap_cache_bits(kernel_pmap, VM_MEMATTR_DEFAULT, FALSE); /* * Install self-referential address mapping entry. */ pm_pml5[PML5PML5I] = VM_PAGE_TO_PHYS(pml5pg) | X86_PG_RW | X86_PG_V | X86_PG_M | X86_PG_A | pmap_cache_bits(kernel_pmap, VM_MEMATTR_DEFAULT, FALSE); } static void pmap_pinit_pml4_pti(vm_page_t pml4pgu) { pml4_entry_t *pm_pml4u; int i; pm_pml4u = (pml4_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pml4pgu)); for (i = 0; i < NPML4EPG; i++) pm_pml4u[i] = pti_pml4[i]; } static void pmap_pinit_pml5_pti(vm_page_t pml5pgu) { pml5_entry_t *pm_pml5u; pm_pml5u = (pml5_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pml5pgu)); pagezero(pm_pml5u); /* * Add pml5 entry at top of KVA pointing to existing pml4 pti * table, entering all kernel mappings needed for usermode * into level 5 table. */ pm_pml5u[pmap_pml5e_index(UPT_MAX_ADDRESS)] = pmap_kextract((vm_offset_t)pti_pml4) | X86_PG_V | X86_PG_RW | X86_PG_A | X86_PG_M | pg_g | pmap_cache_bits(kernel_pmap, VM_MEMATTR_DEFAULT, FALSE); } /* Allocate a page table page and do related bookkeeping */ static vm_page_t pmap_alloc_pt_page(pmap_t pmap, vm_pindex_t pindex, int flags) { vm_page_t m; m = vm_page_alloc_noobj(flags); if (__predict_false(m == NULL)) return (NULL); m->pindex = pindex; pmap_pt_page_count_adj(pmap, 1); return (m); } static void pmap_free_pt_page(pmap_t pmap, vm_page_t m, bool zerofilled) { /* * This function assumes the page will need to be unwired, * even though the counterpart allocation in pmap_alloc_pt_page() * doesn't enforce VM_ALLOC_WIRED. However, all current uses * of pmap_free_pt_page() require unwiring. The case in which * a PT page doesn't require unwiring because its ref_count has * naturally reached 0 is handled through _pmap_unwire_ptp(). */ vm_page_unwire_noq(m); if (zerofilled) vm_page_free_zero(m); else vm_page_free(m); pmap_pt_page_count_adj(pmap, -1); } /* * Initialize a preallocated and zeroed pmap structure, * such as one in a vmspace structure. */ int pmap_pinit_type(pmap_t pmap, enum pmap_type pm_type, int flags) { vm_page_t pmltop_pg, pmltop_pgu; vm_paddr_t pmltop_phys; int i; bzero(&pmap->pm_stats, sizeof pmap->pm_stats); /* * Allocate the page directory page. Pass NULL instead of a * pointer to the pmap here to avoid calling * pmap_resident_count_adj() through pmap_pt_page_count_adj(), * since that requires pmap lock. Instead do the accounting * manually. * * Note that final call to pmap_remove() optimization that * checks for zero resident_count is basically disabled by * accounting for top-level page. But the optimization was * not effective since we started using non-managed mapping of * the shared page. */ pmltop_pg = pmap_alloc_pt_page(NULL, 0, VM_ALLOC_WIRED | VM_ALLOC_ZERO | VM_ALLOC_WAITOK); pmap_pt_page_count_pinit(pmap, 1); pmltop_phys = VM_PAGE_TO_PHYS(pmltop_pg); pmap->pm_pmltop = (pml5_entry_t *)PHYS_TO_DMAP(pmltop_phys); CPU_FOREACH(i) { pmap->pm_pcids[i].pm_pcid = PMAP_PCID_NONE; pmap->pm_pcids[i].pm_gen = 0; } pmap->pm_cr3 = PMAP_NO_CR3; /* initialize to an invalid value */ pmap->pm_ucr3 = PMAP_NO_CR3; pmap->pm_pmltopu = NULL; pmap->pm_type = pm_type; /* * Do not install the host kernel mappings in the nested page * tables. These mappings are meaningless in the guest physical * address space. * Install minimal kernel mappings in PTI case. */ switch (pm_type) { case PT_X86: pmap->pm_cr3 = pmltop_phys; if (pmap_is_la57(pmap)) pmap_pinit_pml5(pmltop_pg); else pmap_pinit_pml4(pmltop_pg); if ((curproc->p_md.md_flags & P_MD_KPTI) != 0) { /* * As with pmltop_pg, pass NULL instead of a * pointer to the pmap to ensure that the PTI * page counted explicitly. */ pmltop_pgu = pmap_alloc_pt_page(NULL, 0, VM_ALLOC_WIRED | VM_ALLOC_WAITOK); pmap_pt_page_count_pinit(pmap, 1); pmap->pm_pmltopu = (pml4_entry_t *)PHYS_TO_DMAP( VM_PAGE_TO_PHYS(pmltop_pgu)); if (pmap_is_la57(pmap)) pmap_pinit_pml5_pti(pmltop_pgu); else pmap_pinit_pml4_pti(pmltop_pgu); pmap->pm_ucr3 = VM_PAGE_TO_PHYS(pmltop_pgu); } if ((cpu_stdext_feature2 & CPUID_STDEXT2_PKU) != 0) { rangeset_init(&pmap->pm_pkru, pkru_dup_range, pkru_free_range, pmap, M_NOWAIT); } break; case PT_EPT: case PT_RVI: pmap->pm_eptsmr = smr_create("pmap", 0, 0); break; } vm_radix_init(&pmap->pm_root); CPU_ZERO(&pmap->pm_active); TAILQ_INIT(&pmap->pm_pvchunk); pmap->pm_flags = flags; pmap->pm_eptgen = 0; return (1); } int pmap_pinit(pmap_t pmap) { return (pmap_pinit_type(pmap, PT_X86, pmap_flags)); } static void pmap_allocpte_free_unref(pmap_t pmap, vm_offset_t va, pt_entry_t *pte) { vm_page_t mpg; struct spglist free; mpg = PHYS_TO_VM_PAGE(*pte & PG_FRAME); if (mpg->ref_count != 0) return; SLIST_INIT(&free); _pmap_unwire_ptp(pmap, va, mpg, &free); pmap_invalidate_page(pmap, va); vm_page_free_pages_toq(&free, true); } static pml4_entry_t * pmap_allocpte_getpml4(pmap_t pmap, struct rwlock **lockp, vm_offset_t va, bool addref) { vm_pindex_t pml5index; pml5_entry_t *pml5; pml4_entry_t *pml4; vm_page_t pml4pg; pt_entry_t PG_V; bool allocated; if (!pmap_is_la57(pmap)) return (&pmap->pm_pmltop[pmap_pml4e_index(va)]); PG_V = pmap_valid_bit(pmap); pml5index = pmap_pml5e_index(va); pml5 = &pmap->pm_pmltop[pml5index]; if ((*pml5 & PG_V) == 0) { if (pmap_allocpte_nosleep(pmap, pmap_pml5e_pindex(va), lockp, va) == NULL) return (NULL); allocated = true; } else { allocated = false; } pml4 = (pml4_entry_t *)PHYS_TO_DMAP(*pml5 & PG_FRAME); pml4 = &pml4[pmap_pml4e_index(va)]; if ((*pml4 & PG_V) == 0) { pml4pg = PHYS_TO_VM_PAGE(*pml5 & PG_FRAME); if (allocated && !addref) pml4pg->ref_count--; else if (!allocated && addref) pml4pg->ref_count++; } return (pml4); } static pdp_entry_t * pmap_allocpte_getpdp(pmap_t pmap, struct rwlock **lockp, vm_offset_t va, bool addref) { vm_page_t pdppg; pml4_entry_t *pml4; pdp_entry_t *pdp; pt_entry_t PG_V; bool allocated; PG_V = pmap_valid_bit(pmap); pml4 = pmap_allocpte_getpml4(pmap, lockp, va, false); if (pml4 == NULL) return (NULL); if ((*pml4 & PG_V) == 0) { /* Have to allocate a new pdp, recurse */ if (pmap_allocpte_nosleep(pmap, pmap_pml4e_pindex(va), lockp, va) == NULL) { if (pmap_is_la57(pmap)) pmap_allocpte_free_unref(pmap, va, pmap_pml5e(pmap, va)); return (NULL); } allocated = true; } else { allocated = false; } pdp = (pdp_entry_t *)PHYS_TO_DMAP(*pml4 & PG_FRAME); pdp = &pdp[pmap_pdpe_index(va)]; if ((*pdp & PG_V) == 0) { pdppg = PHYS_TO_VM_PAGE(*pml4 & PG_FRAME); if (allocated && !addref) pdppg->ref_count--; else if (!allocated && addref) pdppg->ref_count++; } return (pdp); } /* * The ptepindexes, i.e. page indices, of the page table pages encountered * while translating virtual address va are defined as follows: * - for the page table page (last level), * ptepindex = pmap_pde_pindex(va) = va >> PDRSHIFT, * in other words, it is just the index of the PDE that maps the page * table page. * - for the page directory page, * ptepindex = NUPDE (number of userland PD entries) + * (pmap_pde_index(va) >> NPDEPGSHIFT) * i.e. index of PDPE is put after the last index of PDE, * - for the page directory pointer page, * ptepindex = NUPDE + NUPDPE + (pmap_pde_index(va) >> (NPDEPGSHIFT + * NPML4EPGSHIFT), * i.e. index of pml4e is put after the last index of PDPE, * - for the PML4 page (if LA57 mode is enabled), * ptepindex = NUPDE + NUPDPE + NUPML4E + (pmap_pde_index(va) >> * (NPDEPGSHIFT + NPML4EPGSHIFT + NPML5EPGSHIFT), * i.e. index of pml5e is put after the last index of PML4E. * * Define an order on the paging entries, where all entries of the * same height are put together, then heights are put from deepest to * root. Then ptexpindex is the sequential number of the * corresponding paging entry in this order. * * The values of NUPDE, NUPDPE, and NUPML4E are determined by the size of * LA57 paging structures even in LA48 paging mode. Moreover, the * ptepindexes are calculated as if the paging structures were 5-level * regardless of the actual mode of operation. * * The root page at PML4/PML5 does not participate in this indexing scheme, * since it is statically allocated by pmap_pinit() and not by pmap_allocpte(). */ static vm_page_t pmap_allocpte_nosleep(pmap_t pmap, vm_pindex_t ptepindex, struct rwlock **lockp, vm_offset_t va) { vm_pindex_t pml5index, pml4index; pml5_entry_t *pml5, *pml5u; pml4_entry_t *pml4, *pml4u; pdp_entry_t *pdp; pd_entry_t *pd; vm_page_t m, pdpg; pt_entry_t PG_A, PG_M, PG_RW, PG_V; PMAP_LOCK_ASSERT(pmap, MA_OWNED); PG_A = pmap_accessed_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_V = pmap_valid_bit(pmap); PG_RW = pmap_rw_bit(pmap); /* * Allocate a page table page. */ m = pmap_alloc_pt_page(pmap, ptepindex, VM_ALLOC_WIRED | VM_ALLOC_ZERO); if (m == NULL) return (NULL); /* * Map the pagetable page into the process address space, if * it isn't already there. */ if (ptepindex >= NUPDE + NUPDPE + NUPML4E) { MPASS(pmap_is_la57(pmap)); pml5index = pmap_pml5e_index(va); pml5 = &pmap->pm_pmltop[pml5index]; KASSERT((*pml5 & PG_V) == 0, ("pmap %p va %#lx pml5 %#lx", pmap, va, *pml5)); *pml5 = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M; if (pmap->pm_pmltopu != NULL && pml5index < NUPML5E) { if (pmap->pm_ucr3 != PMAP_NO_CR3) *pml5 |= pg_nx; pml5u = &pmap->pm_pmltopu[pml5index]; *pml5u = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M; } } else if (ptepindex >= NUPDE + NUPDPE) { pml4index = pmap_pml4e_index(va); /* Wire up a new PDPE page */ pml4 = pmap_allocpte_getpml4(pmap, lockp, va, true); if (pml4 == NULL) { pmap_free_pt_page(pmap, m, true); return (NULL); } KASSERT((*pml4 & PG_V) == 0, ("pmap %p va %#lx pml4 %#lx", pmap, va, *pml4)); *pml4 = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M; if (!pmap_is_la57(pmap) && pmap->pm_pmltopu != NULL && pml4index < NUPML4E) { /* * PTI: Make all user-space mappings in the * kernel-mode page table no-execute so that * we detect any programming errors that leave * the kernel-mode page table active on return * to user space. */ if (pmap->pm_ucr3 != PMAP_NO_CR3) *pml4 |= pg_nx; pml4u = &pmap->pm_pmltopu[pml4index]; *pml4u = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M; } } else if (ptepindex >= NUPDE) { /* Wire up a new PDE page */ pdp = pmap_allocpte_getpdp(pmap, lockp, va, true); if (pdp == NULL) { pmap_free_pt_page(pmap, m, true); return (NULL); } KASSERT((*pdp & PG_V) == 0, ("pmap %p va %#lx pdp %#lx", pmap, va, *pdp)); *pdp = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M; } else { /* Wire up a new PTE page */ pdp = pmap_allocpte_getpdp(pmap, lockp, va, false); if (pdp == NULL) { pmap_free_pt_page(pmap, m, true); return (NULL); } if ((*pdp & PG_V) == 0) { /* Have to allocate a new pd, recurse */ if (pmap_allocpte_nosleep(pmap, pmap_pdpe_pindex(va), lockp, va) == NULL) { pmap_allocpte_free_unref(pmap, va, pmap_pml4e(pmap, va)); pmap_free_pt_page(pmap, m, true); return (NULL); } } else { /* Add reference to the pd page */ pdpg = PHYS_TO_VM_PAGE(*pdp & PG_FRAME); pdpg->ref_count++; } pd = (pd_entry_t *)PHYS_TO_DMAP(*pdp & PG_FRAME); /* Now we know where the page directory page is */ pd = &pd[pmap_pde_index(va)]; KASSERT((*pd & PG_V) == 0, ("pmap %p va %#lx pd %#lx", pmap, va, *pd)); *pd = VM_PAGE_TO_PHYS(m) | PG_U | PG_RW | PG_V | PG_A | PG_M; } return (m); } /* * This routine is called if the desired page table page does not exist. * * If page table page allocation fails, this routine may sleep before * returning NULL. It sleeps only if a lock pointer was given. Sleep * occurs right before returning to the caller. This way, we never * drop pmap lock to sleep while a page table page has ref_count == 0, * which prevents the page from being freed under us. */ static vm_page_t pmap_allocpte_alloc(pmap_t pmap, vm_pindex_t ptepindex, struct rwlock **lockp, vm_offset_t va) { vm_page_t m; m = pmap_allocpte_nosleep(pmap, ptepindex, lockp, va); if (m == NULL && lockp != NULL) { RELEASE_PV_LIST_LOCK(lockp); PMAP_UNLOCK(pmap); PMAP_ASSERT_NOT_IN_DI(); vm_wait(NULL); PMAP_LOCK(pmap); } return (m); } static pd_entry_t * pmap_alloc_pde(pmap_t pmap, vm_offset_t va, vm_page_t *pdpgp, struct rwlock **lockp) { pdp_entry_t *pdpe, PG_V; pd_entry_t *pde; vm_page_t pdpg; vm_pindex_t pdpindex; PG_V = pmap_valid_bit(pmap); retry: pdpe = pmap_pdpe(pmap, va); if (pdpe != NULL && (*pdpe & PG_V) != 0) { pde = pmap_pdpe_to_pde(pdpe, va); if (va < VM_MAXUSER_ADDRESS) { /* Add a reference to the pd page. */ pdpg = PHYS_TO_VM_PAGE(*pdpe & PG_FRAME); pdpg->ref_count++; } else pdpg = NULL; } else if (va < VM_MAXUSER_ADDRESS) { /* Allocate a pd page. */ pdpindex = pmap_pde_pindex(va) >> NPDPEPGSHIFT; pdpg = pmap_allocpte_alloc(pmap, NUPDE + pdpindex, lockp, va); if (pdpg == NULL) { if (lockp != NULL) goto retry; else return (NULL); } pde = (pd_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pdpg)); pde = &pde[pmap_pde_index(va)]; } else panic("pmap_alloc_pde: missing page table page for va %#lx", va); *pdpgp = pdpg; return (pde); } static vm_page_t pmap_allocpte(pmap_t pmap, vm_offset_t va, struct rwlock **lockp) { vm_pindex_t ptepindex; pd_entry_t *pd, PG_V; vm_page_t m; PG_V = pmap_valid_bit(pmap); /* * Calculate pagetable page index */ ptepindex = pmap_pde_pindex(va); retry: /* * Get the page directory entry */ pd = pmap_pde(pmap, va); /* * This supports switching from a 2MB page to a * normal 4K page. */ if (pd != NULL && (*pd & (PG_PS | PG_V)) == (PG_PS | PG_V)) { if (!pmap_demote_pde_locked(pmap, pd, va, lockp)) { /* * Invalidation of the 2MB page mapping may have caused * the deallocation of the underlying PD page. */ pd = NULL; } } /* * If the page table page is mapped, we just increment the * hold count, and activate it. */ if (pd != NULL && (*pd & PG_V) != 0) { m = PHYS_TO_VM_PAGE(*pd & PG_FRAME); m->ref_count++; } else { /* * Here if the pte page isn't mapped, or if it has been * deallocated. */ m = pmap_allocpte_alloc(pmap, ptepindex, lockp, va); if (m == NULL && lockp != NULL) goto retry; } return (m); } /*************************************************** * Pmap allocation/deallocation routines. ***************************************************/ /* * Release any resources held by the given physical map. * Called when a pmap initialized by pmap_pinit is being released. * Should only be called if the map contains no valid mappings. */ void pmap_release(pmap_t pmap) { vm_page_t m; int i; KASSERT(vm_radix_is_empty(&pmap->pm_root), ("pmap_release: pmap %p has reserved page table page(s)", pmap)); KASSERT(CPU_EMPTY(&pmap->pm_active), ("releasing active pmap %p", pmap)); m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)pmap->pm_pmltop)); if (pmap_is_la57(pmap)) { pmap->pm_pmltop[pmap_pml5e_index(UPT_MAX_ADDRESS)] = 0; pmap->pm_pmltop[PML5PML5I] = 0; } else { for (i = 0; i < NKPML4E; i++) /* KVA */ pmap->pm_pmltop[KPML4BASE + i] = 0; #ifdef KASAN for (i = 0; i < NKASANPML4E; i++) /* KASAN shadow map */ pmap->pm_pmltop[KASANPML4I + i] = 0; #endif for (i = 0; i < ndmpdpphys; i++)/* Direct Map */ pmap->pm_pmltop[DMPML4I + i] = 0; pmap->pm_pmltop[PML4PML4I] = 0; /* Recursive Mapping */ for (i = 0; i < lm_ents; i++) /* Large Map */ pmap->pm_pmltop[LMSPML4I + i] = 0; } pmap_free_pt_page(NULL, m, true); pmap_pt_page_count_pinit(pmap, -1); if (pmap->pm_pmltopu != NULL) { m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)pmap-> pm_pmltopu)); pmap_free_pt_page(NULL, m, false); pmap_pt_page_count_pinit(pmap, -1); } if (pmap->pm_type == PT_X86 && (cpu_stdext_feature2 & CPUID_STDEXT2_PKU) != 0) rangeset_fini(&pmap->pm_pkru); KASSERT(pmap->pm_stats.resident_count == 0, ("pmap_release: pmap %p resident count %ld != 0", pmap, pmap->pm_stats.resident_count)); } static int kvm_size(SYSCTL_HANDLER_ARGS) { unsigned long ksize = VM_MAX_KERNEL_ADDRESS - VM_MIN_KERNEL_ADDRESS; return sysctl_handle_long(oidp, &ksize, 0, req); } SYSCTL_PROC(_vm, OID_AUTO, kvm_size, CTLTYPE_LONG | CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 0, kvm_size, "LU", "Size of KVM"); static int kvm_free(SYSCTL_HANDLER_ARGS) { unsigned long kfree = VM_MAX_KERNEL_ADDRESS - kernel_vm_end; return sysctl_handle_long(oidp, &kfree, 0, req); } SYSCTL_PROC(_vm, OID_AUTO, kvm_free, CTLTYPE_LONG | CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 0, kvm_free, "LU", "Amount of KVM free"); /* * Allocate physical memory for the vm_page array and map it into KVA, * attempting to back the vm_pages with domain-local memory. */ void pmap_page_array_startup(long pages) { pdp_entry_t *pdpe; pd_entry_t *pde, newpdir; vm_offset_t va, start, end; vm_paddr_t pa; long pfn; int domain, i; vm_page_array_size = pages; start = VM_MIN_KERNEL_ADDRESS; end = start + pages * sizeof(struct vm_page); for (va = start; va < end; va += NBPDR) { pfn = first_page + (va - start) / sizeof(struct vm_page); domain = vm_phys_domain(ptoa(pfn)); pdpe = pmap_pdpe(kernel_pmap, va); if ((*pdpe & X86_PG_V) == 0) { pa = vm_phys_early_alloc(domain, PAGE_SIZE); dump_add_page(pa); pagezero((void *)PHYS_TO_DMAP(pa)); *pdpe = (pdp_entry_t)(pa | X86_PG_V | X86_PG_RW | X86_PG_A | X86_PG_M); } pde = pmap_pdpe_to_pde(pdpe, va); if ((*pde & X86_PG_V) != 0) panic("Unexpected pde"); pa = vm_phys_early_alloc(domain, NBPDR); for (i = 0; i < NPDEPG; i++) dump_add_page(pa + i * PAGE_SIZE); newpdir = (pd_entry_t)(pa | X86_PG_V | X86_PG_RW | X86_PG_A | X86_PG_M | PG_PS | pg_g | pg_nx); pde_store(pde, newpdir); } vm_page_array = (vm_page_t)start; } /* * grow the number of kernel page table entries, if needed */ void pmap_growkernel(vm_offset_t addr) { vm_paddr_t paddr; vm_page_t nkpg; pd_entry_t *pde, newpdir; pdp_entry_t *pdpe; mtx_assert(&kernel_map->system_mtx, MA_OWNED); /* * Return if "addr" is within the range of kernel page table pages * that were preallocated during pmap bootstrap. Moreover, leave * "kernel_vm_end" and the kernel page table as they were. * * The correctness of this action is based on the following * argument: vm_map_insert() allocates contiguous ranges of the * kernel virtual address space. It calls this function if a range * ends after "kernel_vm_end". If the kernel is mapped between * "kernel_vm_end" and "addr", then the range cannot begin at * "kernel_vm_end". In fact, its beginning address cannot be less * than the kernel. Thus, there is no immediate need to allocate * any new kernel page table pages between "kernel_vm_end" and * "KERNBASE". */ if (KERNBASE < addr && addr <= KERNBASE + nkpt * NBPDR) return; addr = roundup2(addr, NBPDR); if (addr - 1 >= vm_map_max(kernel_map)) addr = vm_map_max(kernel_map); if (kernel_vm_end < addr) kasan_shadow_map(kernel_vm_end, addr - kernel_vm_end); while (kernel_vm_end < addr) { pdpe = pmap_pdpe(kernel_pmap, kernel_vm_end); if ((*pdpe & X86_PG_V) == 0) { /* We need a new PDP entry */ nkpg = pmap_alloc_pt_page(kernel_pmap, kernel_vm_end >> PDPSHIFT, VM_ALLOC_WIRED | VM_ALLOC_INTERRUPT | VM_ALLOC_ZERO); if (nkpg == NULL) panic("pmap_growkernel: no memory to grow kernel"); paddr = VM_PAGE_TO_PHYS(nkpg); *pdpe = (pdp_entry_t)(paddr | X86_PG_V | X86_PG_RW | X86_PG_A | X86_PG_M); continue; /* try again */ } pde = pmap_pdpe_to_pde(pdpe, kernel_vm_end); if ((*pde & X86_PG_V) != 0) { kernel_vm_end = (kernel_vm_end + NBPDR) & ~PDRMASK; if (kernel_vm_end - 1 >= vm_map_max(kernel_map)) { kernel_vm_end = vm_map_max(kernel_map); break; } continue; } nkpg = pmap_alloc_pt_page(kernel_pmap, pmap_pde_pindex(kernel_vm_end), VM_ALLOC_WIRED | VM_ALLOC_INTERRUPT | VM_ALLOC_ZERO); if (nkpg == NULL) panic("pmap_growkernel: no memory to grow kernel"); paddr = VM_PAGE_TO_PHYS(nkpg); newpdir = paddr | X86_PG_V | X86_PG_RW | X86_PG_A | X86_PG_M; pde_store(pde, newpdir); kernel_vm_end = (kernel_vm_end + NBPDR) & ~PDRMASK; if (kernel_vm_end - 1 >= vm_map_max(kernel_map)) { kernel_vm_end = vm_map_max(kernel_map); break; } } } /*************************************************** * page management routines. ***************************************************/ CTASSERT(sizeof(struct pv_chunk) == PAGE_SIZE); CTASSERT(_NPCM == 3); CTASSERT(_NPCPV == 168); static __inline struct pv_chunk * pv_to_chunk(pv_entry_t pv) { return ((struct pv_chunk *)((uintptr_t)pv & ~(uintptr_t)PAGE_MASK)); } #define PV_PMAP(pv) (pv_to_chunk(pv)->pc_pmap) #define PC_FREE0 0xfffffffffffffffful #define PC_FREE1 0xfffffffffffffffful #define PC_FREE2 0x000000fffffffffful static const uint64_t pc_freemask[_NPCM] = { PC_FREE0, PC_FREE1, PC_FREE2 }; #ifdef PV_STATS static COUNTER_U64_DEFINE_EARLY(pc_chunk_count); SYSCTL_COUNTER_U64(_vm_pmap, OID_AUTO, pc_chunk_count, CTLFLAG_RD, &pc_chunk_count, "Current number of pv entry cnunks"); static COUNTER_U64_DEFINE_EARLY(pc_chunk_allocs); SYSCTL_COUNTER_U64(_vm_pmap, OID_AUTO, pc_chunk_allocs, CTLFLAG_RD, &pc_chunk_allocs, "Total number of pv entry chunks allocated"); static COUNTER_U64_DEFINE_EARLY(pc_chunk_frees); SYSCTL_COUNTER_U64(_vm_pmap, OID_AUTO, pc_chunk_frees, CTLFLAG_RD, &pc_chunk_frees, "Total number of pv entry chunks freed"); static COUNTER_U64_DEFINE_EARLY(pc_chunk_tryfail); SYSCTL_COUNTER_U64(_vm_pmap, OID_AUTO, pc_chunk_tryfail, CTLFLAG_RD, &pc_chunk_tryfail, "Number of failed attempts to get a pv entry chunk page"); static COUNTER_U64_DEFINE_EARLY(pv_entry_frees); SYSCTL_COUNTER_U64(_vm_pmap, OID_AUTO, pv_entry_frees, CTLFLAG_RD, &pv_entry_frees, "Total number of pv entries freed"); static COUNTER_U64_DEFINE_EARLY(pv_entry_allocs); SYSCTL_COUNTER_U64(_vm_pmap, OID_AUTO, pv_entry_allocs, CTLFLAG_RD, &pv_entry_allocs, "Total number of pv entries allocated"); static COUNTER_U64_DEFINE_EARLY(pv_entry_count); SYSCTL_COUNTER_U64(_vm_pmap, OID_AUTO, pv_entry_count, CTLFLAG_RD, &pv_entry_count, "Current number of pv entries"); static COUNTER_U64_DEFINE_EARLY(pv_entry_spare); SYSCTL_COUNTER_U64(_vm_pmap, OID_AUTO, pv_entry_spare, CTLFLAG_RD, &pv_entry_spare, "Current number of spare pv entries"); #endif static void reclaim_pv_chunk_leave_pmap(pmap_t pmap, pmap_t locked_pmap, bool start_di) { if (pmap == NULL) return; pmap_invalidate_all(pmap); if (pmap != locked_pmap) PMAP_UNLOCK(pmap); if (start_di) pmap_delayed_invl_finish(); } /* * We are in a serious low memory condition. Resort to * drastic measures to free some pages so we can allocate * another pv entry chunk. * * Returns NULL if PV entries were reclaimed from the specified pmap. * * We do not, however, unmap 2mpages because subsequent accesses will * allocate per-page pv entries until repromotion occurs, thereby * exacerbating the shortage of free pv entries. */ static vm_page_t reclaim_pv_chunk_domain(pmap_t locked_pmap, struct rwlock **lockp, int domain) { struct pv_chunks_list *pvc; struct pv_chunk *pc, *pc_marker, *pc_marker_end; struct pv_chunk_header pc_marker_b, pc_marker_end_b; struct md_page *pvh; pd_entry_t *pde; pmap_t next_pmap, pmap; pt_entry_t *pte, tpte; pt_entry_t PG_G, PG_A, PG_M, PG_RW; pv_entry_t pv; vm_offset_t va; vm_page_t m, m_pc; struct spglist free; uint64_t inuse; int bit, field, freed; bool start_di, restart; PMAP_LOCK_ASSERT(locked_pmap, MA_OWNED); KASSERT(lockp != NULL, ("reclaim_pv_chunk: lockp is NULL")); pmap = NULL; m_pc = NULL; PG_G = PG_A = PG_M = PG_RW = 0; SLIST_INIT(&free); bzero(&pc_marker_b, sizeof(pc_marker_b)); bzero(&pc_marker_end_b, sizeof(pc_marker_end_b)); pc_marker = (struct pv_chunk *)&pc_marker_b; pc_marker_end = (struct pv_chunk *)&pc_marker_end_b; /* * A delayed invalidation block should already be active if * pmap_advise() or pmap_remove() called this function by way * of pmap_demote_pde_locked(). */ start_di = pmap_not_in_di(); pvc = &pv_chunks[domain]; mtx_lock(&pvc->pvc_lock); pvc->active_reclaims++; TAILQ_INSERT_HEAD(&pvc->pvc_list, pc_marker, pc_lru); TAILQ_INSERT_TAIL(&pvc->pvc_list, pc_marker_end, pc_lru); while ((pc = TAILQ_NEXT(pc_marker, pc_lru)) != pc_marker_end && SLIST_EMPTY(&free)) { next_pmap = pc->pc_pmap; if (next_pmap == NULL) { /* * The next chunk is a marker. However, it is * not our marker, so active_reclaims must be * > 1. Consequently, the next_chunk code * will not rotate the pv_chunks list. */ goto next_chunk; } mtx_unlock(&pvc->pvc_lock); /* * A pv_chunk can only be removed from the pc_lru list * when both pc_chunks_mutex is owned and the * corresponding pmap is locked. */ if (pmap != next_pmap) { restart = false; reclaim_pv_chunk_leave_pmap(pmap, locked_pmap, start_di); pmap = next_pmap; /* Avoid deadlock and lock recursion. */ if (pmap > locked_pmap) { RELEASE_PV_LIST_LOCK(lockp); PMAP_LOCK(pmap); if (start_di) pmap_delayed_invl_start(); mtx_lock(&pvc->pvc_lock); restart = true; } else if (pmap != locked_pmap) { if (PMAP_TRYLOCK(pmap)) { if (start_di) pmap_delayed_invl_start(); mtx_lock(&pvc->pvc_lock); restart = true; } else { pmap = NULL; /* pmap is not locked */ mtx_lock(&pvc->pvc_lock); pc = TAILQ_NEXT(pc_marker, pc_lru); if (pc == NULL || pc->pc_pmap != next_pmap) continue; goto next_chunk; } } else if (start_di) pmap_delayed_invl_start(); PG_G = pmap_global_bit(pmap); PG_A = pmap_accessed_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_RW = pmap_rw_bit(pmap); if (restart) continue; } /* * Destroy every non-wired, 4 KB page mapping in the chunk. */ freed = 0; for (field = 0; field < _NPCM; field++) { for (inuse = ~pc->pc_map[field] & pc_freemask[field]; inuse != 0; inuse &= ~(1UL << bit)) { bit = bsfq(inuse); pv = &pc->pc_pventry[field * 64 + bit]; va = pv->pv_va; pde = pmap_pde(pmap, va); if ((*pde & PG_PS) != 0) continue; pte = pmap_pde_to_pte(pde, va); if ((*pte & PG_W) != 0) continue; tpte = pte_load_clear(pte); if ((tpte & PG_G) != 0) pmap_invalidate_page(pmap, va); m = PHYS_TO_VM_PAGE(tpte & PG_FRAME); if ((tpte & (PG_M | PG_RW)) == (PG_M | PG_RW)) vm_page_dirty(m); if ((tpte & PG_A) != 0) vm_page_aflag_set(m, PGA_REFERENCED); CHANGE_PV_LIST_LOCK_TO_VM_PAGE(lockp, m); TAILQ_REMOVE(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; if (TAILQ_EMPTY(&m->md.pv_list) && (m->flags & PG_FICTITIOUS) == 0) { pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); if (TAILQ_EMPTY(&pvh->pv_list)) { vm_page_aflag_clear(m, PGA_WRITEABLE); } } pmap_delayed_invl_page(m); pc->pc_map[field] |= 1UL << bit; pmap_unuse_pt(pmap, va, *pde, &free); freed++; } } if (freed == 0) { mtx_lock(&pvc->pvc_lock); goto next_chunk; } /* Every freed mapping is for a 4 KB page. */ pmap_resident_count_adj(pmap, -freed); PV_STAT(counter_u64_add(pv_entry_frees, freed)); PV_STAT(counter_u64_add(pv_entry_spare, freed)); PV_STAT(counter_u64_add(pv_entry_count, -freed)); TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); if (pc->pc_map[0] == PC_FREE0 && pc->pc_map[1] == PC_FREE1 && pc->pc_map[2] == PC_FREE2) { PV_STAT(counter_u64_add(pv_entry_spare, -_NPCPV)); PV_STAT(counter_u64_add(pc_chunk_count, -1)); PV_STAT(counter_u64_add(pc_chunk_frees, 1)); /* Entire chunk is free; return it. */ m_pc = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)pc)); dump_drop_page(m_pc->phys_addr); mtx_lock(&pvc->pvc_lock); TAILQ_REMOVE(&pvc->pvc_list, pc, pc_lru); break; } TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list); mtx_lock(&pvc->pvc_lock); /* One freed pv entry in locked_pmap is sufficient. */ if (pmap == locked_pmap) break; next_chunk: TAILQ_REMOVE(&pvc->pvc_list, pc_marker, pc_lru); TAILQ_INSERT_AFTER(&pvc->pvc_list, pc, pc_marker, pc_lru); if (pvc->active_reclaims == 1 && pmap != NULL) { /* * Rotate the pv chunks list so that we do not * scan the same pv chunks that could not be * freed (because they contained a wired * and/or superpage mapping) on every * invocation of reclaim_pv_chunk(). */ while ((pc = TAILQ_FIRST(&pvc->pvc_list)) != pc_marker) { MPASS(pc->pc_pmap != NULL); TAILQ_REMOVE(&pvc->pvc_list, pc, pc_lru); TAILQ_INSERT_TAIL(&pvc->pvc_list, pc, pc_lru); } } } TAILQ_REMOVE(&pvc->pvc_list, pc_marker, pc_lru); TAILQ_REMOVE(&pvc->pvc_list, pc_marker_end, pc_lru); pvc->active_reclaims--; mtx_unlock(&pvc->pvc_lock); reclaim_pv_chunk_leave_pmap(pmap, locked_pmap, start_di); if (m_pc == NULL && !SLIST_EMPTY(&free)) { m_pc = SLIST_FIRST(&free); SLIST_REMOVE_HEAD(&free, plinks.s.ss); /* Recycle a freed page table page. */ m_pc->ref_count = 1; } vm_page_free_pages_toq(&free, true); return (m_pc); } static vm_page_t reclaim_pv_chunk(pmap_t locked_pmap, struct rwlock **lockp) { vm_page_t m; int i, domain; domain = PCPU_GET(domain); for (i = 0; i < vm_ndomains; i++) { m = reclaim_pv_chunk_domain(locked_pmap, lockp, domain); if (m != NULL) break; domain = (domain + 1) % vm_ndomains; } return (m); } /* * free the pv_entry back to the free list */ static void free_pv_entry(pmap_t pmap, pv_entry_t pv) { struct pv_chunk *pc; int idx, field, bit; PMAP_LOCK_ASSERT(pmap, MA_OWNED); PV_STAT(counter_u64_add(pv_entry_frees, 1)); PV_STAT(counter_u64_add(pv_entry_spare, 1)); PV_STAT(counter_u64_add(pv_entry_count, -1)); pc = pv_to_chunk(pv); idx = pv - &pc->pc_pventry[0]; field = idx / 64; bit = idx % 64; pc->pc_map[field] |= 1ul << bit; if (pc->pc_map[0] != PC_FREE0 || pc->pc_map[1] != PC_FREE1 || pc->pc_map[2] != PC_FREE2) { /* 98% of the time, pc is already at the head of the list. */ if (__predict_false(pc != TAILQ_FIRST(&pmap->pm_pvchunk))) { TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list); } return; } TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); free_pv_chunk(pc); } static void free_pv_chunk_dequeued(struct pv_chunk *pc) { vm_page_t m; PV_STAT(counter_u64_add(pv_entry_spare, -_NPCPV)); PV_STAT(counter_u64_add(pc_chunk_count, -1)); PV_STAT(counter_u64_add(pc_chunk_frees, 1)); counter_u64_add(pv_page_count, -1); /* entire chunk is free, return it */ m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)pc)); dump_drop_page(m->phys_addr); vm_page_unwire_noq(m); vm_page_free(m); } static void free_pv_chunk(struct pv_chunk *pc) { struct pv_chunks_list *pvc; pvc = &pv_chunks[pc_to_domain(pc)]; mtx_lock(&pvc->pvc_lock); TAILQ_REMOVE(&pvc->pvc_list, pc, pc_lru); mtx_unlock(&pvc->pvc_lock); free_pv_chunk_dequeued(pc); } static void free_pv_chunk_batch(struct pv_chunklist *batch) { struct pv_chunks_list *pvc; struct pv_chunk *pc, *npc; int i; for (i = 0; i < vm_ndomains; i++) { if (TAILQ_EMPTY(&batch[i])) continue; pvc = &pv_chunks[i]; mtx_lock(&pvc->pvc_lock); TAILQ_FOREACH(pc, &batch[i], pc_list) { TAILQ_REMOVE(&pvc->pvc_list, pc, pc_lru); } mtx_unlock(&pvc->pvc_lock); } for (i = 0; i < vm_ndomains; i++) { TAILQ_FOREACH_SAFE(pc, &batch[i], pc_list, npc) { free_pv_chunk_dequeued(pc); } } } /* * Returns a new PV entry, allocating a new PV chunk from the system when * needed. If this PV chunk allocation fails and a PV list lock pointer was * given, a PV chunk is reclaimed from an arbitrary pmap. Otherwise, NULL is * returned. * * The given PV list lock may be released. */ static pv_entry_t get_pv_entry(pmap_t pmap, struct rwlock **lockp) { struct pv_chunks_list *pvc; int bit, field; pv_entry_t pv; struct pv_chunk *pc; vm_page_t m; PMAP_LOCK_ASSERT(pmap, MA_OWNED); PV_STAT(counter_u64_add(pv_entry_allocs, 1)); retry: pc = TAILQ_FIRST(&pmap->pm_pvchunk); if (pc != NULL) { for (field = 0; field < _NPCM; field++) { if (pc->pc_map[field]) { bit = bsfq(pc->pc_map[field]); break; } } if (field < _NPCM) { pv = &pc->pc_pventry[field * 64 + bit]; pc->pc_map[field] &= ~(1ul << bit); /* If this was the last item, move it to tail */ if (pc->pc_map[0] == 0 && pc->pc_map[1] == 0 && pc->pc_map[2] == 0) { TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); TAILQ_INSERT_TAIL(&pmap->pm_pvchunk, pc, pc_list); } PV_STAT(counter_u64_add(pv_entry_count, 1)); PV_STAT(counter_u64_add(pv_entry_spare, -1)); return (pv); } } /* No free items, allocate another chunk */ m = vm_page_alloc_noobj(VM_ALLOC_WIRED); if (m == NULL) { if (lockp == NULL) { PV_STAT(counter_u64_add(pc_chunk_tryfail, 1)); return (NULL); } m = reclaim_pv_chunk(pmap, lockp); if (m == NULL) goto retry; } else counter_u64_add(pv_page_count, 1); PV_STAT(counter_u64_add(pc_chunk_count, 1)); PV_STAT(counter_u64_add(pc_chunk_allocs, 1)); dump_add_page(m->phys_addr); pc = (void *)PHYS_TO_DMAP(m->phys_addr); pc->pc_pmap = pmap; pc->pc_map[0] = PC_FREE0 & ~1ul; /* preallocated bit 0 */ pc->pc_map[1] = PC_FREE1; pc->pc_map[2] = PC_FREE2; pvc = &pv_chunks[vm_page_domain(m)]; mtx_lock(&pvc->pvc_lock); TAILQ_INSERT_TAIL(&pvc->pvc_list, pc, pc_lru); mtx_unlock(&pvc->pvc_lock); pv = &pc->pc_pventry[0]; TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list); PV_STAT(counter_u64_add(pv_entry_count, 1)); PV_STAT(counter_u64_add(pv_entry_spare, _NPCPV - 1)); return (pv); } /* * Returns the number of one bits within the given PV chunk map. * * The erratas for Intel processors state that "POPCNT Instruction May * Take Longer to Execute Than Expected". It is believed that the * issue is the spurious dependency on the destination register. * Provide a hint to the register rename logic that the destination * value is overwritten, by clearing it, as suggested in the * optimization manual. It should be cheap for unaffected processors * as well. * * Reference numbers for erratas are * 4th Gen Core: HSD146 * 5th Gen Core: BDM85 * 6th Gen Core: SKL029 */ static int popcnt_pc_map_pq(uint64_t *map) { u_long result, tmp; __asm __volatile("xorl %k0,%k0;popcntq %2,%0;" "xorl %k1,%k1;popcntq %3,%1;addl %k1,%k0;" "xorl %k1,%k1;popcntq %4,%1;addl %k1,%k0" : "=&r" (result), "=&r" (tmp) : "m" (map[0]), "m" (map[1]), "m" (map[2])); return (result); } /* * Ensure that the number of spare PV entries in the specified pmap meets or * exceeds the given count, "needed". * * The given PV list lock may be released. */ static void reserve_pv_entries(pmap_t pmap, int needed, struct rwlock **lockp) { struct pv_chunks_list *pvc; struct pch new_tail[PMAP_MEMDOM]; struct pv_chunk *pc; vm_page_t m; int avail, free, i; bool reclaimed; PMAP_LOCK_ASSERT(pmap, MA_OWNED); KASSERT(lockp != NULL, ("reserve_pv_entries: lockp is NULL")); /* * Newly allocated PV chunks must be stored in a private list until * the required number of PV chunks have been allocated. Otherwise, * reclaim_pv_chunk() could recycle one of these chunks. In * contrast, these chunks must be added to the pmap upon allocation. */ for (i = 0; i < PMAP_MEMDOM; i++) TAILQ_INIT(&new_tail[i]); retry: avail = 0; TAILQ_FOREACH(pc, &pmap->pm_pvchunk, pc_list) { #ifndef __POPCNT__ if ((cpu_feature2 & CPUID2_POPCNT) == 0) bit_count((bitstr_t *)pc->pc_map, 0, sizeof(pc->pc_map) * NBBY, &free); else #endif free = popcnt_pc_map_pq(pc->pc_map); if (free == 0) break; avail += free; if (avail >= needed) break; } for (reclaimed = false; avail < needed; avail += _NPCPV) { m = vm_page_alloc_noobj(VM_ALLOC_WIRED); if (m == NULL) { m = reclaim_pv_chunk(pmap, lockp); if (m == NULL) goto retry; reclaimed = true; } else counter_u64_add(pv_page_count, 1); PV_STAT(counter_u64_add(pc_chunk_count, 1)); PV_STAT(counter_u64_add(pc_chunk_allocs, 1)); dump_add_page(m->phys_addr); pc = (void *)PHYS_TO_DMAP(m->phys_addr); pc->pc_pmap = pmap; pc->pc_map[0] = PC_FREE0; pc->pc_map[1] = PC_FREE1; pc->pc_map[2] = PC_FREE2; TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list); TAILQ_INSERT_TAIL(&new_tail[vm_page_domain(m)], pc, pc_lru); PV_STAT(counter_u64_add(pv_entry_spare, _NPCPV)); /* * The reclaim might have freed a chunk from the current pmap. * If that chunk contained available entries, we need to * re-count the number of available entries. */ if (reclaimed) goto retry; } for (i = 0; i < vm_ndomains; i++) { if (TAILQ_EMPTY(&new_tail[i])) continue; pvc = &pv_chunks[i]; mtx_lock(&pvc->pvc_lock); TAILQ_CONCAT(&pvc->pvc_list, &new_tail[i], pc_lru); mtx_unlock(&pvc->pvc_lock); } } /* * First find and then remove the pv entry for the specified pmap and virtual * address from the specified pv list. Returns the pv entry if found and NULL * otherwise. This operation can be performed on pv lists for either 4KB or * 2MB page mappings. */ static __inline pv_entry_t pmap_pvh_remove(struct md_page *pvh, pmap_t pmap, vm_offset_t va) { pv_entry_t pv; TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) { if (pmap == PV_PMAP(pv) && va == pv->pv_va) { TAILQ_REMOVE(&pvh->pv_list, pv, pv_next); pvh->pv_gen++; break; } } return (pv); } /* * After demotion from a 2MB page mapping to 512 4KB page mappings, * destroy the pv entry for the 2MB page mapping and reinstantiate the pv * entries for each of the 4KB page mappings. */ static void pmap_pv_demote_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa, struct rwlock **lockp) { struct md_page *pvh; struct pv_chunk *pc; pv_entry_t pv; vm_offset_t va_last; vm_page_t m; int bit, field; PMAP_LOCK_ASSERT(pmap, MA_OWNED); KASSERT((pa & PDRMASK) == 0, ("pmap_pv_demote_pde: pa is not 2mpage aligned")); CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, pa); /* * Transfer the 2mpage's pv entry for this mapping to the first * page's pv list. Once this transfer begins, the pv list lock * must not be released until the last pv entry is reinstantiated. */ pvh = pa_to_pvh(pa); va = trunc_2mpage(va); pv = pmap_pvh_remove(pvh, pmap, va); KASSERT(pv != NULL, ("pmap_pv_demote_pde: pv not found")); m = PHYS_TO_VM_PAGE(pa); TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; /* Instantiate the remaining NPTEPG - 1 pv entries. */ PV_STAT(counter_u64_add(pv_entry_allocs, NPTEPG - 1)); va_last = va + NBPDR - PAGE_SIZE; for (;;) { pc = TAILQ_FIRST(&pmap->pm_pvchunk); KASSERT(pc->pc_map[0] != 0 || pc->pc_map[1] != 0 || pc->pc_map[2] != 0, ("pmap_pv_demote_pde: missing spare")); for (field = 0; field < _NPCM; field++) { while (pc->pc_map[field]) { bit = bsfq(pc->pc_map[field]); pc->pc_map[field] &= ~(1ul << bit); pv = &pc->pc_pventry[field * 64 + bit]; va += PAGE_SIZE; pv->pv_va = va; m++; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_pv_demote_pde: page %p is not managed", m)); TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; if (va == va_last) goto out; } } TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); TAILQ_INSERT_TAIL(&pmap->pm_pvchunk, pc, pc_list); } out: if (pc->pc_map[0] == 0 && pc->pc_map[1] == 0 && pc->pc_map[2] == 0) { TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); TAILQ_INSERT_TAIL(&pmap->pm_pvchunk, pc, pc_list); } PV_STAT(counter_u64_add(pv_entry_count, NPTEPG - 1)); PV_STAT(counter_u64_add(pv_entry_spare, -(NPTEPG - 1))); } #if VM_NRESERVLEVEL > 0 /* * After promotion from 512 4KB page mappings to a single 2MB page mapping, * replace the many pv entries for the 4KB page mappings by a single pv entry * for the 2MB page mapping. */ static void pmap_pv_promote_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa, struct rwlock **lockp) { struct md_page *pvh; pv_entry_t pv; vm_offset_t va_last; vm_page_t m; KASSERT((pa & PDRMASK) == 0, ("pmap_pv_promote_pde: pa is not 2mpage aligned")); CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, pa); /* * Transfer the first page's pv entry for this mapping to the 2mpage's * pv list. Aside from avoiding the cost of a call to get_pv_entry(), * a transfer avoids the possibility that get_pv_entry() calls * reclaim_pv_chunk() and that reclaim_pv_chunk() removes one of the * mappings that is being promoted. */ m = PHYS_TO_VM_PAGE(pa); va = trunc_2mpage(va); pv = pmap_pvh_remove(&m->md, pmap, va); KASSERT(pv != NULL, ("pmap_pv_promote_pde: pv not found")); pvh = pa_to_pvh(pa); TAILQ_INSERT_TAIL(&pvh->pv_list, pv, pv_next); pvh->pv_gen++; /* Free the remaining NPTEPG - 1 pv entries. */ va_last = va + NBPDR - PAGE_SIZE; do { m++; va += PAGE_SIZE; pmap_pvh_free(&m->md, pmap, va); } while (va < va_last); } #endif /* VM_NRESERVLEVEL > 0 */ /* * First find and then destroy the pv entry for the specified pmap and virtual * address. This operation can be performed on pv lists for either 4KB or 2MB * page mappings. */ static void pmap_pvh_free(struct md_page *pvh, pmap_t pmap, vm_offset_t va) { pv_entry_t pv; pv = pmap_pvh_remove(pvh, pmap, va); KASSERT(pv != NULL, ("pmap_pvh_free: pv not found")); free_pv_entry(pmap, pv); } /* * Conditionally create the PV entry for a 4KB page mapping if the required * memory can be allocated without resorting to reclamation. */ static boolean_t pmap_try_insert_pv_entry(pmap_t pmap, vm_offset_t va, vm_page_t m, struct rwlock **lockp) { pv_entry_t pv; PMAP_LOCK_ASSERT(pmap, MA_OWNED); /* Pass NULL instead of the lock pointer to disable reclamation. */ if ((pv = get_pv_entry(pmap, NULL)) != NULL) { pv->pv_va = va; CHANGE_PV_LIST_LOCK_TO_VM_PAGE(lockp, m); TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; return (TRUE); } else return (FALSE); } /* * Create the PV entry for a 2MB page mapping. Always returns true unless the * flag PMAP_ENTER_NORECLAIM is specified. If that flag is specified, returns * false if the PV entry cannot be allocated without resorting to reclamation. */ static bool pmap_pv_insert_pde(pmap_t pmap, vm_offset_t va, pd_entry_t pde, u_int flags, struct rwlock **lockp) { struct md_page *pvh; pv_entry_t pv; vm_paddr_t pa; PMAP_LOCK_ASSERT(pmap, MA_OWNED); /* Pass NULL instead of the lock pointer to disable reclamation. */ if ((pv = get_pv_entry(pmap, (flags & PMAP_ENTER_NORECLAIM) != 0 ? NULL : lockp)) == NULL) return (false); pv->pv_va = va; pa = pde & PG_PS_FRAME; CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, pa); pvh = pa_to_pvh(pa); TAILQ_INSERT_TAIL(&pvh->pv_list, pv, pv_next); pvh->pv_gen++; return (true); } /* * Fills a page table page with mappings to consecutive physical pages. */ static void pmap_fill_ptp(pt_entry_t *firstpte, pt_entry_t newpte) { pt_entry_t *pte; for (pte = firstpte; pte < firstpte + NPTEPG; pte++) { *pte = newpte; newpte += PAGE_SIZE; } } /* * Tries to demote a 2MB page mapping. If demotion fails, the 2MB page * mapping is invalidated. */ static boolean_t pmap_demote_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va) { struct rwlock *lock; boolean_t rv; lock = NULL; rv = pmap_demote_pde_locked(pmap, pde, va, &lock); if (lock != NULL) rw_wunlock(lock); return (rv); } static void pmap_demote_pde_check(pt_entry_t *firstpte __unused, pt_entry_t newpte __unused) { #ifdef INVARIANTS #ifdef DIAGNOSTIC pt_entry_t *xpte, *ypte; for (xpte = firstpte; xpte < firstpte + NPTEPG; xpte++, newpte += PAGE_SIZE) { if ((*xpte & PG_FRAME) != (newpte & PG_FRAME)) { printf("pmap_demote_pde: xpte %zd and newpte map " "different pages: found %#lx, expected %#lx\n", xpte - firstpte, *xpte, newpte); printf("page table dump\n"); for (ypte = firstpte; ypte < firstpte + NPTEPG; ypte++) printf("%zd %#lx\n", ypte - firstpte, *ypte); panic("firstpte"); } } #else KASSERT((*firstpte & PG_FRAME) == (newpte & PG_FRAME), ("pmap_demote_pde: firstpte and newpte map different physical" " addresses")); #endif #endif } static void pmap_demote_pde_abort(pmap_t pmap, vm_offset_t va, pd_entry_t *pde, pd_entry_t oldpde, struct rwlock **lockp) { struct spglist free; vm_offset_t sva; SLIST_INIT(&free); sva = trunc_2mpage(va); pmap_remove_pde(pmap, pde, sva, &free, lockp); if ((oldpde & pmap_global_bit(pmap)) == 0) pmap_invalidate_pde_page(pmap, sva, oldpde); vm_page_free_pages_toq(&free, true); CTR2(KTR_PMAP, "pmap_demote_pde: failure for va %#lx in pmap %p", va, pmap); } static boolean_t pmap_demote_pde_locked(pmap_t pmap, pd_entry_t *pde, vm_offset_t va, struct rwlock **lockp) { pd_entry_t newpde, oldpde; pt_entry_t *firstpte, newpte; pt_entry_t PG_A, PG_G, PG_M, PG_PKU_MASK, PG_RW, PG_V; vm_paddr_t mptepa; vm_page_t mpte; int PG_PTE_CACHE; bool in_kernel; PG_A = pmap_accessed_bit(pmap); PG_G = pmap_global_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_RW = pmap_rw_bit(pmap); PG_V = pmap_valid_bit(pmap); PG_PTE_CACHE = pmap_cache_mask(pmap, 0); PG_PKU_MASK = pmap_pku_mask_bit(pmap); PMAP_LOCK_ASSERT(pmap, MA_OWNED); in_kernel = va >= VM_MAXUSER_ADDRESS; oldpde = *pde; KASSERT((oldpde & (PG_PS | PG_V)) == (PG_PS | PG_V), ("pmap_demote_pde: oldpde is missing PG_PS and/or PG_V")); /* * Invalidate the 2MB page mapping and return "failure" if the * mapping was never accessed. */ if ((oldpde & PG_A) == 0) { KASSERT((oldpde & PG_W) == 0, ("pmap_demote_pde: a wired mapping is missing PG_A")); pmap_demote_pde_abort(pmap, va, pde, oldpde, lockp); return (FALSE); } mpte = pmap_remove_pt_page(pmap, va); if (mpte == NULL) { KASSERT((oldpde & PG_W) == 0, ("pmap_demote_pde: page table page for a wired mapping" " is missing")); /* * If the page table page is missing and the mapping * is for a kernel address, the mapping must belong to * the direct map. Page table pages are preallocated * for every other part of the kernel address space, * so the direct map region is the only part of the * kernel address space that must be handled here. */ KASSERT(!in_kernel || (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS), ("pmap_demote_pde: No saved mpte for va %#lx", va)); /* * If the 2MB page mapping belongs to the direct map * region of the kernel's address space, then the page * allocation request specifies the highest possible * priority (VM_ALLOC_INTERRUPT). Otherwise, the * priority is normal. */ mpte = pmap_alloc_pt_page(pmap, pmap_pde_pindex(va), (in_kernel ? VM_ALLOC_INTERRUPT : 0) | VM_ALLOC_WIRED); /* * If the allocation of the new page table page fails, * invalidate the 2MB page mapping and return "failure". */ if (mpte == NULL) { pmap_demote_pde_abort(pmap, va, pde, oldpde, lockp); return (FALSE); } if (!in_kernel) mpte->ref_count = NPTEPG; } mptepa = VM_PAGE_TO_PHYS(mpte); firstpte = (pt_entry_t *)PHYS_TO_DMAP(mptepa); newpde = mptepa | PG_M | PG_A | (oldpde & PG_U) | PG_RW | PG_V; KASSERT((oldpde & (PG_M | PG_RW)) != PG_RW, ("pmap_demote_pde: oldpde is missing PG_M")); newpte = oldpde & ~PG_PS; newpte = pmap_swap_pat(pmap, newpte); /* * If the page table page is not leftover from an earlier promotion, * initialize it. */ if (mpte->valid == 0) pmap_fill_ptp(firstpte, newpte); pmap_demote_pde_check(firstpte, newpte); /* * If the mapping has changed attributes, update the page table * entries. */ if ((*firstpte & PG_PTE_PROMOTE) != (newpte & PG_PTE_PROMOTE)) pmap_fill_ptp(firstpte, newpte); /* * The spare PV entries must be reserved prior to demoting the * mapping, that is, prior to changing the PDE. Otherwise, the state * of the PDE and the PV lists will be inconsistent, which can result * in reclaim_pv_chunk() attempting to remove a PV entry from the * wrong PV list and pmap_pv_demote_pde() failing to find the expected * PV entry for the 2MB page mapping that is being demoted. */ if ((oldpde & PG_MANAGED) != 0) reserve_pv_entries(pmap, NPTEPG - 1, lockp); /* * Demote the mapping. This pmap is locked. The old PDE has * PG_A set. If the old PDE has PG_RW set, it also has PG_M * set. Thus, there is no danger of a race with another * processor changing the setting of PG_A and/or PG_M between * the read above and the store below. */ if (workaround_erratum383) pmap_update_pde(pmap, va, pde, newpde); else pde_store(pde, newpde); /* * Invalidate a stale recursive mapping of the page table page. */ if (in_kernel) pmap_invalidate_page(pmap, (vm_offset_t)vtopte(va)); /* * Demote the PV entry. */ if ((oldpde & PG_MANAGED) != 0) pmap_pv_demote_pde(pmap, va, oldpde & PG_PS_FRAME, lockp); counter_u64_add(pmap_pde_demotions, 1); CTR2(KTR_PMAP, "pmap_demote_pde: success for va %#lx in pmap %p", va, pmap); return (TRUE); } /* * pmap_remove_kernel_pde: Remove a kernel superpage mapping. */ static void pmap_remove_kernel_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va) { pd_entry_t newpde; vm_paddr_t mptepa; vm_page_t mpte; KASSERT(pmap == kernel_pmap, ("pmap %p is not kernel_pmap", pmap)); PMAP_LOCK_ASSERT(pmap, MA_OWNED); mpte = pmap_remove_pt_page(pmap, va); if (mpte == NULL) panic("pmap_remove_kernel_pde: Missing pt page."); mptepa = VM_PAGE_TO_PHYS(mpte); newpde = mptepa | X86_PG_M | X86_PG_A | X86_PG_RW | X86_PG_V; /* * If this page table page was unmapped by a promotion, then it * contains valid mappings. Zero it to invalidate those mappings. */ if (mpte->valid != 0) pagezero((void *)PHYS_TO_DMAP(mptepa)); /* * Demote the mapping. */ if (workaround_erratum383) pmap_update_pde(pmap, va, pde, newpde); else pde_store(pde, newpde); /* * Invalidate a stale recursive mapping of the page table page. */ pmap_invalidate_page(pmap, (vm_offset_t)vtopte(va)); } /* * pmap_remove_pde: do the things to unmap a superpage in a process */ static int pmap_remove_pde(pmap_t pmap, pd_entry_t *pdq, vm_offset_t sva, struct spglist *free, struct rwlock **lockp) { struct md_page *pvh; pd_entry_t oldpde; vm_offset_t eva, va; vm_page_t m, mpte; pt_entry_t PG_G, PG_A, PG_M, PG_RW; PG_G = pmap_global_bit(pmap); PG_A = pmap_accessed_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_RW = pmap_rw_bit(pmap); PMAP_LOCK_ASSERT(pmap, MA_OWNED); KASSERT((sva & PDRMASK) == 0, ("pmap_remove_pde: sva is not 2mpage aligned")); oldpde = pte_load_clear(pdq); if (oldpde & PG_W) pmap->pm_stats.wired_count -= NBPDR / PAGE_SIZE; if ((oldpde & PG_G) != 0) pmap_invalidate_pde_page(kernel_pmap, sva, oldpde); pmap_resident_count_adj(pmap, -NBPDR / PAGE_SIZE); if (oldpde & PG_MANAGED) { CHANGE_PV_LIST_LOCK_TO_PHYS(lockp, oldpde & PG_PS_FRAME); pvh = pa_to_pvh(oldpde & PG_PS_FRAME); pmap_pvh_free(pvh, pmap, sva); eva = sva + NBPDR; for (va = sva, m = PHYS_TO_VM_PAGE(oldpde & PG_PS_FRAME); va < eva; va += PAGE_SIZE, m++) { if ((oldpde & (PG_M | PG_RW)) == (PG_M | PG_RW)) vm_page_dirty(m); if (oldpde & PG_A) vm_page_aflag_set(m, PGA_REFERENCED); if (TAILQ_EMPTY(&m->md.pv_list) && TAILQ_EMPTY(&pvh->pv_list)) vm_page_aflag_clear(m, PGA_WRITEABLE); pmap_delayed_invl_page(m); } } if (pmap == kernel_pmap) { pmap_remove_kernel_pde(pmap, pdq, sva); } else { mpte = pmap_remove_pt_page(pmap, sva); if (mpte != NULL) { KASSERT(mpte->valid == VM_PAGE_BITS_ALL, ("pmap_remove_pde: pte page not promoted")); pmap_resident_count_adj(pmap, -1); KASSERT(mpte->ref_count == NPTEPG, ("pmap_remove_pde: pte page ref count error")); mpte->ref_count = 0; pmap_add_delayed_free_list(mpte, free, FALSE); } } return (pmap_unuse_pt(pmap, sva, *pmap_pdpe(pmap, sva), free)); } /* * pmap_remove_pte: do the things to unmap a page in a process */ static int pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t va, pd_entry_t ptepde, struct spglist *free, struct rwlock **lockp) { struct md_page *pvh; pt_entry_t oldpte, PG_A, PG_M, PG_RW; vm_page_t m; PG_A = pmap_accessed_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_RW = pmap_rw_bit(pmap); PMAP_LOCK_ASSERT(pmap, MA_OWNED); oldpte = pte_load_clear(ptq); if (oldpte & PG_W) pmap->pm_stats.wired_count -= 1; pmap_resident_count_adj(pmap, -1); if (oldpte & PG_MANAGED) { m = PHYS_TO_VM_PAGE(oldpte & PG_FRAME); if ((oldpte & (PG_M | PG_RW)) == (PG_M | PG_RW)) vm_page_dirty(m); if (oldpte & PG_A) vm_page_aflag_set(m, PGA_REFERENCED); CHANGE_PV_LIST_LOCK_TO_VM_PAGE(lockp, m); pmap_pvh_free(&m->md, pmap, va); if (TAILQ_EMPTY(&m->md.pv_list) && (m->flags & PG_FICTITIOUS) == 0) { pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); if (TAILQ_EMPTY(&pvh->pv_list)) vm_page_aflag_clear(m, PGA_WRITEABLE); } pmap_delayed_invl_page(m); } return (pmap_unuse_pt(pmap, va, ptepde, free)); } /* * Remove a single page from a process address space */ static void pmap_remove_page(pmap_t pmap, vm_offset_t va, pd_entry_t *pde, struct spglist *free) { struct rwlock *lock; pt_entry_t *pte, PG_V; PG_V = pmap_valid_bit(pmap); PMAP_LOCK_ASSERT(pmap, MA_OWNED); if ((*pde & PG_V) == 0) return; pte = pmap_pde_to_pte(pde, va); if ((*pte & PG_V) == 0) return; lock = NULL; pmap_remove_pte(pmap, pte, va, *pde, free, &lock); if (lock != NULL) rw_wunlock(lock); pmap_invalidate_page(pmap, va); } /* * Removes the specified range of addresses from the page table page. */ static bool pmap_remove_ptes(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, pd_entry_t *pde, struct spglist *free, struct rwlock **lockp) { pt_entry_t PG_G, *pte; vm_offset_t va; bool anyvalid; PMAP_LOCK_ASSERT(pmap, MA_OWNED); PG_G = pmap_global_bit(pmap); anyvalid = false; va = eva; for (pte = pmap_pde_to_pte(pde, sva); sva != eva; pte++, sva += PAGE_SIZE) { if (*pte == 0) { if (va != eva) { pmap_invalidate_range(pmap, va, sva); va = eva; } continue; } if ((*pte & PG_G) == 0) anyvalid = true; else if (va == eva) va = sva; if (pmap_remove_pte(pmap, pte, sva, *pde, free, lockp)) { sva += PAGE_SIZE; break; } } if (va != eva) pmap_invalidate_range(pmap, va, sva); return (anyvalid); } /* * Remove the given range of addresses from the specified map. * * It is assumed that the start and end are properly * rounded to the page size. */ void pmap_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { struct rwlock *lock; vm_page_t mt; vm_offset_t va_next; pml5_entry_t *pml5e; pml4_entry_t *pml4e; pdp_entry_t *pdpe; pd_entry_t ptpaddr, *pde; pt_entry_t PG_G, PG_V; struct spglist free; int anyvalid; PG_G = pmap_global_bit(pmap); PG_V = pmap_valid_bit(pmap); /* * If there are no resident pages besides the top level page * table page(s), there is nothing to do. Kernel pmap always * accounts whole preloaded area as resident, which makes its * resident count > 2. * Perform an unsynchronized read. This is, however, safe. */ if (pmap->pm_stats.resident_count <= 1 + (pmap->pm_pmltopu != NULL ? 1 : 0)) return; anyvalid = 0; SLIST_INIT(&free); pmap_delayed_invl_start(); PMAP_LOCK(pmap); pmap_pkru_on_remove(pmap, sva, eva); /* * special handling of removing one page. a very * common operation and easy to short circuit some * code. */ if (sva + PAGE_SIZE == eva) { pde = pmap_pde(pmap, sva); if (pde && (*pde & PG_PS) == 0) { pmap_remove_page(pmap, sva, pde, &free); goto out; } } lock = NULL; for (; sva < eva; sva = va_next) { if (pmap->pm_stats.resident_count == 0) break; if (pmap_is_la57(pmap)) { pml5e = pmap_pml5e(pmap, sva); if ((*pml5e & PG_V) == 0) { va_next = (sva + NBPML5) & ~PML5MASK; if (va_next < sva) va_next = eva; continue; } pml4e = pmap_pml5e_to_pml4e(pml5e, sva); } else { pml4e = pmap_pml4e(pmap, sva); } if ((*pml4e & PG_V) == 0) { va_next = (sva + NBPML4) & ~PML4MASK; if (va_next < sva) va_next = eva; continue; } va_next = (sva + NBPDP) & ~PDPMASK; if (va_next < sva) va_next = eva; pdpe = pmap_pml4e_to_pdpe(pml4e, sva); if ((*pdpe & PG_V) == 0) continue; if ((*pdpe & PG_PS) != 0) { KASSERT(va_next <= eva, ("partial update of non-transparent 1G mapping " "pdpe %#lx sva %#lx eva %#lx va_next %#lx", *pdpe, sva, eva, va_next)); MPASS(pmap != kernel_pmap); /* XXXKIB */ MPASS((*pdpe & (PG_MANAGED | PG_G)) == 0); anyvalid = 1; *pdpe = 0; pmap_resident_count_adj(pmap, -NBPDP / PAGE_SIZE); mt = PHYS_TO_VM_PAGE(*pmap_pml4e(pmap, sva) & PG_FRAME); pmap_unwire_ptp(pmap, sva, mt, &free); continue; } /* * Calculate index for next page table. */ va_next = (sva + NBPDR) & ~PDRMASK; if (va_next < sva) va_next = eva; pde = pmap_pdpe_to_pde(pdpe, sva); ptpaddr = *pde; /* * Weed out invalid mappings. */ if (ptpaddr == 0) continue; /* * Check for large page. */ if ((ptpaddr & PG_PS) != 0) { /* * Are we removing the entire large page? If not, * demote the mapping and fall through. */ if (sva + NBPDR == va_next && eva >= va_next) { /* * The TLB entry for a PG_G mapping is * invalidated by pmap_remove_pde(). */ if ((ptpaddr & PG_G) == 0) anyvalid = 1; pmap_remove_pde(pmap, pde, sva, &free, &lock); continue; } else if (!pmap_demote_pde_locked(pmap, pde, sva, &lock)) { /* The large page mapping was destroyed. */ continue; } else ptpaddr = *pde; } /* * Limit our scan to either the end of the va represented * by the current page table page, or to the end of the * range being removed. */ if (va_next > eva) va_next = eva; if (pmap_remove_ptes(pmap, sva, va_next, pde, &free, &lock)) anyvalid = 1; } if (lock != NULL) rw_wunlock(lock); out: if (anyvalid) pmap_invalidate_all(pmap); PMAP_UNLOCK(pmap); pmap_delayed_invl_finish(); vm_page_free_pages_toq(&free, true); } /* * Routine: pmap_remove_all * Function: * Removes this physical page from * all physical maps in which it resides. * Reflects back modify bits to the pager. * * Notes: * Original versions of this routine were very * inefficient because they iteratively called * pmap_remove (slow...) */ void pmap_remove_all(vm_page_t m) { struct md_page *pvh; pv_entry_t pv; pmap_t pmap; struct rwlock *lock; pt_entry_t *pte, tpte, PG_A, PG_M, PG_RW; pd_entry_t *pde; vm_offset_t va; struct spglist free; int pvh_gen, md_gen; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_remove_all: page %p is not managed", m)); SLIST_INIT(&free); lock = VM_PAGE_TO_PV_LIST_LOCK(m); pvh = (m->flags & PG_FICTITIOUS) != 0 ? &pv_dummy : pa_to_pvh(VM_PAGE_TO_PHYS(m)); rw_wlock(lock); retry: while ((pv = TAILQ_FIRST(&pvh->pv_list)) != NULL) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { pvh_gen = pvh->pv_gen; rw_wunlock(lock); PMAP_LOCK(pmap); rw_wlock(lock); if (pvh_gen != pvh->pv_gen) { PMAP_UNLOCK(pmap); goto retry; } } va = pv->pv_va; pde = pmap_pde(pmap, va); (void)pmap_demote_pde_locked(pmap, pde, va, &lock); PMAP_UNLOCK(pmap); } while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { pvh_gen = pvh->pv_gen; md_gen = m->md.pv_gen; rw_wunlock(lock); PMAP_LOCK(pmap); rw_wlock(lock); if (pvh_gen != pvh->pv_gen || md_gen != m->md.pv_gen) { PMAP_UNLOCK(pmap); goto retry; } } PG_A = pmap_accessed_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_RW = pmap_rw_bit(pmap); pmap_resident_count_adj(pmap, -1); pde = pmap_pde(pmap, pv->pv_va); KASSERT((*pde & PG_PS) == 0, ("pmap_remove_all: found" " a 2mpage in page %p's pv list", m)); pte = pmap_pde_to_pte(pde, pv->pv_va); tpte = pte_load_clear(pte); if (tpte & PG_W) pmap->pm_stats.wired_count--; if (tpte & PG_A) vm_page_aflag_set(m, PGA_REFERENCED); /* * Update the vm_page_t clean and reference bits. */ if ((tpte & (PG_M | PG_RW)) == (PG_M | PG_RW)) vm_page_dirty(m); pmap_unuse_pt(pmap, pv->pv_va, *pde, &free); pmap_invalidate_page(pmap, pv->pv_va); TAILQ_REMOVE(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; free_pv_entry(pmap, pv); PMAP_UNLOCK(pmap); } vm_page_aflag_clear(m, PGA_WRITEABLE); rw_wunlock(lock); pmap_delayed_invl_wait(m); vm_page_free_pages_toq(&free, true); } /* * pmap_protect_pde: do the things to protect a 2mpage in a process */ static boolean_t pmap_protect_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t sva, vm_prot_t prot) { pd_entry_t newpde, oldpde; vm_page_t m, mt; boolean_t anychanged; pt_entry_t PG_G, PG_M, PG_RW; PG_G = pmap_global_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_RW = pmap_rw_bit(pmap); PMAP_LOCK_ASSERT(pmap, MA_OWNED); KASSERT((sva & PDRMASK) == 0, ("pmap_protect_pde: sva is not 2mpage aligned")); anychanged = FALSE; retry: oldpde = newpde = *pde; if ((prot & VM_PROT_WRITE) == 0) { if ((oldpde & (PG_MANAGED | PG_M | PG_RW)) == (PG_MANAGED | PG_M | PG_RW)) { m = PHYS_TO_VM_PAGE(oldpde & PG_PS_FRAME); for (mt = m; mt < &m[NBPDR / PAGE_SIZE]; mt++) vm_page_dirty(mt); } newpde &= ~(PG_RW | PG_M); } if ((prot & VM_PROT_EXECUTE) == 0) newpde |= pg_nx; if (newpde != oldpde) { /* * As an optimization to future operations on this PDE, clear * PG_PROMOTED. The impending invalidation will remove any * lingering 4KB page mappings from the TLB. */ if (!atomic_cmpset_long(pde, oldpde, newpde & ~PG_PROMOTED)) goto retry; if ((oldpde & PG_G) != 0) pmap_invalidate_pde_page(kernel_pmap, sva, oldpde); else anychanged = TRUE; } return (anychanged); } /* * Set the physical protection on the * specified range of this map as requested. */ void pmap_protect(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot) { vm_page_t m; vm_offset_t va_next; pml4_entry_t *pml4e; pdp_entry_t *pdpe; pd_entry_t ptpaddr, *pde; pt_entry_t *pte, PG_G, PG_M, PG_RW, PG_V; pt_entry_t obits, pbits; boolean_t anychanged; KASSERT((prot & ~VM_PROT_ALL) == 0, ("invalid prot %x", prot)); if (prot == VM_PROT_NONE) { pmap_remove(pmap, sva, eva); return; } if ((prot & (VM_PROT_WRITE|VM_PROT_EXECUTE)) == (VM_PROT_WRITE|VM_PROT_EXECUTE)) return; PG_G = pmap_global_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_V = pmap_valid_bit(pmap); PG_RW = pmap_rw_bit(pmap); anychanged = FALSE; /* * Although this function delays and batches the invalidation * of stale TLB entries, it does not need to call * pmap_delayed_invl_start() and * pmap_delayed_invl_finish(), because it does not * ordinarily destroy mappings. Stale TLB entries from * protection-only changes need only be invalidated before the * pmap lock is released, because protection-only changes do * not destroy PV entries. Even operations that iterate over * a physical page's PV list of mappings, like * pmap_remove_write(), acquire the pmap lock for each * mapping. Consequently, for protection-only changes, the * pmap lock suffices to synchronize both page table and TLB * updates. * * This function only destroys a mapping if pmap_demote_pde() * fails. In that case, stale TLB entries are immediately * invalidated. */ PMAP_LOCK(pmap); for (; sva < eva; sva = va_next) { pml4e = pmap_pml4e(pmap, sva); if (pml4e == NULL || (*pml4e & PG_V) == 0) { va_next = (sva + NBPML4) & ~PML4MASK; if (va_next < sva) va_next = eva; continue; } va_next = (sva + NBPDP) & ~PDPMASK; if (va_next < sva) va_next = eva; pdpe = pmap_pml4e_to_pdpe(pml4e, sva); if ((*pdpe & PG_V) == 0) continue; if ((*pdpe & PG_PS) != 0) { KASSERT(va_next <= eva, ("partial update of non-transparent 1G mapping " "pdpe %#lx sva %#lx eva %#lx va_next %#lx", *pdpe, sva, eva, va_next)); retry_pdpe: obits = pbits = *pdpe; MPASS((pbits & (PG_MANAGED | PG_G)) == 0); MPASS(pmap != kernel_pmap); /* XXXKIB */ if ((prot & VM_PROT_WRITE) == 0) pbits &= ~(PG_RW | PG_M); if ((prot & VM_PROT_EXECUTE) == 0) pbits |= pg_nx; if (pbits != obits) { if (!atomic_cmpset_long(pdpe, obits, pbits)) /* PG_PS cannot be cleared under us, */ goto retry_pdpe; anychanged = TRUE; } continue; } va_next = (sva + NBPDR) & ~PDRMASK; if (va_next < sva) va_next = eva; pde = pmap_pdpe_to_pde(pdpe, sva); ptpaddr = *pde; /* * Weed out invalid mappings. */ if (ptpaddr == 0) continue; /* * Check for large page. */ if ((ptpaddr & PG_PS) != 0) { /* * Are we protecting the entire large page? If not, * demote the mapping and fall through. */ if (sva + NBPDR == va_next && eva >= va_next) { /* * The TLB entry for a PG_G mapping is * invalidated by pmap_protect_pde(). */ if (pmap_protect_pde(pmap, pde, sva, prot)) anychanged = TRUE; continue; } else if (!pmap_demote_pde(pmap, pde, sva)) { /* * The large page mapping was destroyed. */ continue; } } if (va_next > eva) va_next = eva; for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++, sva += PAGE_SIZE) { retry: obits = pbits = *pte; if ((pbits & PG_V) == 0) continue; if ((prot & VM_PROT_WRITE) == 0) { if ((pbits & (PG_MANAGED | PG_M | PG_RW)) == (PG_MANAGED | PG_M | PG_RW)) { m = PHYS_TO_VM_PAGE(pbits & PG_FRAME); vm_page_dirty(m); } pbits &= ~(PG_RW | PG_M); } if ((prot & VM_PROT_EXECUTE) == 0) pbits |= pg_nx; if (pbits != obits) { if (!atomic_cmpset_long(pte, obits, pbits)) goto retry; if (obits & PG_G) pmap_invalidate_page(pmap, sva); else anychanged = TRUE; } } } if (anychanged) pmap_invalidate_all(pmap); PMAP_UNLOCK(pmap); } #if VM_NRESERVLEVEL > 0 static bool pmap_pde_ept_executable(pmap_t pmap, pd_entry_t pde) { if (pmap->pm_type != PT_EPT) return (false); return ((pde & EPT_PG_EXECUTE) != 0); } /* * Tries to promote the 512, contiguous 4KB page mappings that are within a * single page table page (PTP) to a single 2MB page mapping. For promotion * to occur, two conditions must be met: (1) the 4KB page mappings must map * aligned, contiguous physical memory and (2) the 4KB page mappings must have * identical characteristics. */ static void pmap_promote_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va, struct rwlock **lockp) { pd_entry_t newpde; pt_entry_t *firstpte, oldpte, pa, *pte; pt_entry_t PG_G, PG_A, PG_M, PG_RW, PG_V, PG_PKU_MASK; vm_page_t mpte; int PG_PTE_CACHE; PG_A = pmap_accessed_bit(pmap); PG_G = pmap_global_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_V = pmap_valid_bit(pmap); PG_RW = pmap_rw_bit(pmap); PG_PKU_MASK = pmap_pku_mask_bit(pmap); PG_PTE_CACHE = pmap_cache_mask(pmap, 0); PMAP_LOCK_ASSERT(pmap, MA_OWNED); /* * Examine the first PTE in the specified PTP. Abort if this PTE is * either invalid, unused, or does not map the first 4KB physical page * within a 2MB page. */ firstpte = (pt_entry_t *)PHYS_TO_DMAP(*pde & PG_FRAME); newpde = *firstpte; if ((newpde & ((PG_FRAME & PDRMASK) | PG_A | PG_V)) != (PG_A | PG_V) || !pmap_allow_2m_x_page(pmap, pmap_pde_ept_executable(pmap, newpde))) { counter_u64_add(pmap_pde_p_failures, 1); CTR2(KTR_PMAP, "pmap_promote_pde: failure for va %#lx" " in pmap %p", va, pmap); return; } setpde: if ((newpde & (PG_M | PG_RW)) == PG_RW) { /* * When PG_M is already clear, PG_RW can be cleared without * a TLB invalidation. */ if (!atomic_fcmpset_long(firstpte, &newpde, newpde & ~PG_RW)) goto setpde; newpde &= ~PG_RW; } /* * Examine each of the other PTEs in the specified PTP. Abort if this * PTE maps an unexpected 4KB physical page or does not have identical * characteristics to the first PTE. */ pa = (newpde & (PG_PS_FRAME | PG_A | PG_V)) + NBPDR - PAGE_SIZE; for (pte = firstpte + NPTEPG - 1; pte > firstpte; pte--) { oldpte = *pte; if ((oldpte & (PG_FRAME | PG_A | PG_V)) != pa) { counter_u64_add(pmap_pde_p_failures, 1); CTR2(KTR_PMAP, "pmap_promote_pde: failure for va %#lx" " in pmap %p", va, pmap); return; } setpte: if ((oldpte & (PG_M | PG_RW)) == PG_RW) { /* * When PG_M is already clear, PG_RW can be cleared * without a TLB invalidation. */ if (!atomic_fcmpset_long(pte, &oldpte, oldpte & ~PG_RW)) goto setpte; oldpte &= ~PG_RW; CTR2(KTR_PMAP, "pmap_promote_pde: protect for va %#lx" " in pmap %p", (oldpte & PG_FRAME & PDRMASK) | (va & ~PDRMASK), pmap); } if ((oldpte & PG_PTE_PROMOTE) != (newpde & PG_PTE_PROMOTE)) { counter_u64_add(pmap_pde_p_failures, 1); CTR2(KTR_PMAP, "pmap_promote_pde: failure for va %#lx" " in pmap %p", va, pmap); return; } pa -= PAGE_SIZE; } /* * Save the page table page in its current state until the PDE * mapping the superpage is demoted by pmap_demote_pde() or * destroyed by pmap_remove_pde(). */ mpte = PHYS_TO_VM_PAGE(*pde & PG_FRAME); KASSERT(mpte >= vm_page_array && mpte < &vm_page_array[vm_page_array_size], ("pmap_promote_pde: page table page is out of range")); KASSERT(mpte->pindex == pmap_pde_pindex(va), ("pmap_promote_pde: page table page's pindex is wrong " "mpte %p pidx %#lx va %#lx va pde pidx %#lx", mpte, mpte->pindex, va, pmap_pde_pindex(va))); if (pmap_insert_pt_page(pmap, mpte, true)) { counter_u64_add(pmap_pde_p_failures, 1); CTR2(KTR_PMAP, "pmap_promote_pde: failure for va %#lx in pmap %p", va, pmap); return; } /* * Promote the pv entries. */ if ((newpde & PG_MANAGED) != 0) pmap_pv_promote_pde(pmap, va, newpde & PG_PS_FRAME, lockp); /* * Propagate the PAT index to its proper position. */ newpde = pmap_swap_pat(pmap, newpde); /* * Map the superpage. */ if (workaround_erratum383) pmap_update_pde(pmap, va, pde, PG_PS | newpde); else pde_store(pde, PG_PROMOTED | PG_PS | newpde); counter_u64_add(pmap_pde_promotions, 1); CTR2(KTR_PMAP, "pmap_promote_pde: success for va %#lx" " in pmap %p", va, pmap); } #endif /* VM_NRESERVLEVEL > 0 */ static int pmap_enter_largepage(pmap_t pmap, vm_offset_t va, pt_entry_t newpte, int flags, int psind) { vm_page_t mp; pt_entry_t origpte, *pml4e, *pdpe, *pde, pten, PG_V; PMAP_LOCK_ASSERT(pmap, MA_OWNED); KASSERT(psind > 0 && psind < MAXPAGESIZES && pagesizes[psind] != 0, ("psind %d unexpected", psind)); KASSERT(((newpte & PG_FRAME) & (pagesizes[psind] - 1)) == 0, ("unaligned phys address %#lx newpte %#lx psind %d", newpte & PG_FRAME, newpte, psind)); KASSERT((va & (pagesizes[psind] - 1)) == 0, ("unaligned va %#lx psind %d", va, psind)); KASSERT(va < VM_MAXUSER_ADDRESS, ("kernel mode non-transparent superpage")); /* XXXKIB */ KASSERT(va + pagesizes[psind] < VM_MAXUSER_ADDRESS, ("overflowing user map va %#lx psind %d", va, psind)); /* XXXKIB */ PG_V = pmap_valid_bit(pmap); restart: if (!pmap_pkru_same(pmap, va, va + pagesizes[psind])) return (KERN_PROTECTION_FAILURE); pten = newpte; if (va < VM_MAXUSER_ADDRESS && pmap->pm_type == PT_X86) pten |= pmap_pkru_get(pmap, va); if (psind == 2) { /* 1G */ pml4e = pmap_pml4e(pmap, va); if (pml4e == NULL || (*pml4e & PG_V) == 0) { mp = pmap_allocpte_alloc(pmap, pmap_pml4e_pindex(va), NULL, va); if (mp == NULL) goto allocf; pdpe = (pdp_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(mp)); pdpe = &pdpe[pmap_pdpe_index(va)]; origpte = *pdpe; MPASS(origpte == 0); } else { pdpe = pmap_pml4e_to_pdpe(pml4e, va); KASSERT(pdpe != NULL, ("va %#lx lost pdpe", va)); origpte = *pdpe; if ((origpte & PG_V) == 0) { mp = PHYS_TO_VM_PAGE(*pml4e & PG_FRAME); mp->ref_count++; } } *pdpe = pten; } else /* (psind == 1) */ { /* 2M */ pde = pmap_pde(pmap, va); if (pde == NULL) { mp = pmap_allocpte_alloc(pmap, pmap_pdpe_pindex(va), NULL, va); if (mp == NULL) goto allocf; pde = (pd_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(mp)); pde = &pde[pmap_pde_index(va)]; origpte = *pde; MPASS(origpte == 0); } else { origpte = *pde; if ((origpte & PG_V) == 0) { pdpe = pmap_pdpe(pmap, va); MPASS(pdpe != NULL && (*pdpe & PG_V) != 0); mp = PHYS_TO_VM_PAGE(*pdpe & PG_FRAME); mp->ref_count++; } } *pde = pten; } KASSERT((origpte & PG_V) == 0 || ((origpte & PG_PS) != 0 && (origpte & PG_PS_FRAME) == (pten & PG_PS_FRAME)), ("va %#lx changing %s phys page origpte %#lx pten %#lx", va, psind == 2 ? "1G" : "2M", origpte, pten)); if ((pten & PG_W) != 0 && (origpte & PG_W) == 0) pmap->pm_stats.wired_count += pagesizes[psind] / PAGE_SIZE; else if ((pten & PG_W) == 0 && (origpte & PG_W) != 0) pmap->pm_stats.wired_count -= pagesizes[psind] / PAGE_SIZE; if ((origpte & PG_V) == 0) pmap_resident_count_adj(pmap, pagesizes[psind] / PAGE_SIZE); return (KERN_SUCCESS); allocf: if ((flags & PMAP_ENTER_NOSLEEP) != 0) return (KERN_RESOURCE_SHORTAGE); PMAP_UNLOCK(pmap); vm_wait(NULL); PMAP_LOCK(pmap); goto restart; } /* * Insert the given physical page (p) at * the specified virtual address (v) in the * target physical map with the protection requested. * * If specified, the page will be wired down, meaning * that the related pte can not be reclaimed. * * NB: This is the only routine which MAY NOT lazy-evaluate * or lose information. That is, this routine must actually * insert this page into the given map NOW. * * When destroying both a page table and PV entry, this function * performs the TLB invalidation before releasing the PV list * lock, so we do not need pmap_delayed_invl_page() calls here. */ int pmap_enter(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, u_int flags, int8_t psind) { struct rwlock *lock; pd_entry_t *pde; pt_entry_t *pte, PG_G, PG_A, PG_M, PG_RW, PG_V; pt_entry_t newpte, origpte; pv_entry_t pv; vm_paddr_t opa, pa; vm_page_t mpte, om; int rv; boolean_t nosleep; PG_A = pmap_accessed_bit(pmap); PG_G = pmap_global_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_V = pmap_valid_bit(pmap); PG_RW = pmap_rw_bit(pmap); va = trunc_page(va); KASSERT(va <= VM_MAX_KERNEL_ADDRESS, ("pmap_enter: toobig")); KASSERT(va < UPT_MIN_ADDRESS || va >= UPT_MAX_ADDRESS, ("pmap_enter: invalid to pmap_enter page table pages (va: 0x%lx)", va)); KASSERT((m->oflags & VPO_UNMANAGED) != 0 || va < kmi.clean_sva || va >= kmi.clean_eva, ("pmap_enter: managed mapping within the clean submap")); if ((m->oflags & VPO_UNMANAGED) == 0) VM_PAGE_OBJECT_BUSY_ASSERT(m); KASSERT((flags & PMAP_ENTER_RESERVED) == 0, ("pmap_enter: flags %u has reserved bits set", flags)); pa = VM_PAGE_TO_PHYS(m); newpte = (pt_entry_t)(pa | PG_A | PG_V); if ((flags & VM_PROT_WRITE) != 0) newpte |= PG_M; if ((prot & VM_PROT_WRITE) != 0) newpte |= PG_RW; KASSERT((newpte & (PG_M | PG_RW)) != PG_M, ("pmap_enter: flags includes VM_PROT_WRITE but prot doesn't")); if ((prot & VM_PROT_EXECUTE) == 0) newpte |= pg_nx; if ((flags & PMAP_ENTER_WIRED) != 0) newpte |= PG_W; if (va < VM_MAXUSER_ADDRESS) newpte |= PG_U; if (pmap == kernel_pmap) newpte |= PG_G; newpte |= pmap_cache_bits(pmap, m->md.pat_mode, psind > 0); /* * Set modified bit gratuitously for writeable mappings if * the page is unmanaged. We do not want to take a fault * to do the dirty bit accounting for these mappings. */ if ((m->oflags & VPO_UNMANAGED) != 0) { if ((newpte & PG_RW) != 0) newpte |= PG_M; } else newpte |= PG_MANAGED; lock = NULL; PMAP_LOCK(pmap); if ((flags & PMAP_ENTER_LARGEPAGE) != 0) { KASSERT((m->oflags & VPO_UNMANAGED) != 0, ("managed largepage va %#lx flags %#x", va, flags)); rv = pmap_enter_largepage(pmap, va, newpte | PG_PS, flags, psind); goto out; } if (psind == 1) { /* Assert the required virtual and physical alignment. */ KASSERT((va & PDRMASK) == 0, ("pmap_enter: va unaligned")); KASSERT(m->psind > 0, ("pmap_enter: m->psind < psind")); rv = pmap_enter_pde(pmap, va, newpte | PG_PS, flags, m, &lock); goto out; } mpte = NULL; /* * In the case that a page table page is not * resident, we are creating it here. */ retry: pde = pmap_pde(pmap, va); if (pde != NULL && (*pde & PG_V) != 0 && ((*pde & PG_PS) == 0 || pmap_demote_pde_locked(pmap, pde, va, &lock))) { pte = pmap_pde_to_pte(pde, va); if (va < VM_MAXUSER_ADDRESS && mpte == NULL) { mpte = PHYS_TO_VM_PAGE(*pde & PG_FRAME); mpte->ref_count++; } } else if (va < VM_MAXUSER_ADDRESS) { /* * Here if the pte page isn't mapped, or if it has been * deallocated. */ nosleep = (flags & PMAP_ENTER_NOSLEEP) != 0; mpte = pmap_allocpte_alloc(pmap, pmap_pde_pindex(va), nosleep ? NULL : &lock, va); if (mpte == NULL && nosleep) { rv = KERN_RESOURCE_SHORTAGE; goto out; } goto retry; } else panic("pmap_enter: invalid page directory va=%#lx", va); origpte = *pte; pv = NULL; if (va < VM_MAXUSER_ADDRESS && pmap->pm_type == PT_X86) newpte |= pmap_pkru_get(pmap, va); /* * Is the specified virtual address already mapped? */ if ((origpte & PG_V) != 0) { /* * Wiring change, just update stats. We don't worry about * wiring PT pages as they remain resident as long as there * are valid mappings in them. Hence, if a user page is wired, * the PT page will be also. */ if ((newpte & PG_W) != 0 && (origpte & PG_W) == 0) pmap->pm_stats.wired_count++; else if ((newpte & PG_W) == 0 && (origpte & PG_W) != 0) pmap->pm_stats.wired_count--; /* * Remove the extra PT page reference. */ if (mpte != NULL) { mpte->ref_count--; KASSERT(mpte->ref_count > 0, ("pmap_enter: missing reference to page table page," " va: 0x%lx", va)); } /* * Has the physical page changed? */ opa = origpte & PG_FRAME; if (opa == pa) { /* * No, might be a protection or wiring change. */ if ((origpte & PG_MANAGED) != 0 && (newpte & PG_RW) != 0) vm_page_aflag_set(m, PGA_WRITEABLE); if (((origpte ^ newpte) & ~(PG_M | PG_A)) == 0) goto unchanged; goto validate; } /* * The physical page has changed. Temporarily invalidate * the mapping. This ensures that all threads sharing the * pmap keep a consistent view of the mapping, which is * necessary for the correct handling of COW faults. It * also permits reuse of the old mapping's PV entry, * avoiding an allocation. * * For consistency, handle unmanaged mappings the same way. */ origpte = pte_load_clear(pte); KASSERT((origpte & PG_FRAME) == opa, ("pmap_enter: unexpected pa update for %#lx", va)); if ((origpte & PG_MANAGED) != 0) { om = PHYS_TO_VM_PAGE(opa); /* * The pmap lock is sufficient to synchronize with * concurrent calls to pmap_page_test_mappings() and * pmap_ts_referenced(). */ if ((origpte & (PG_M | PG_RW)) == (PG_M | PG_RW)) vm_page_dirty(om); if ((origpte & PG_A) != 0) { pmap_invalidate_page(pmap, va); vm_page_aflag_set(om, PGA_REFERENCED); } CHANGE_PV_LIST_LOCK_TO_PHYS(&lock, opa); pv = pmap_pvh_remove(&om->md, pmap, va); KASSERT(pv != NULL, ("pmap_enter: no PV entry for %#lx", va)); if ((newpte & PG_MANAGED) == 0) free_pv_entry(pmap, pv); if ((om->a.flags & PGA_WRITEABLE) != 0 && TAILQ_EMPTY(&om->md.pv_list) && ((om->flags & PG_FICTITIOUS) != 0 || TAILQ_EMPTY(&pa_to_pvh(opa)->pv_list))) vm_page_aflag_clear(om, PGA_WRITEABLE); } else { /* * Since this mapping is unmanaged, assume that PG_A * is set. */ pmap_invalidate_page(pmap, va); } origpte = 0; } else { /* * Increment the counters. */ if ((newpte & PG_W) != 0) pmap->pm_stats.wired_count++; pmap_resident_count_adj(pmap, 1); } /* * Enter on the PV list if part of our managed memory. */ if ((newpte & PG_MANAGED) != 0) { if (pv == NULL) { pv = get_pv_entry(pmap, &lock); pv->pv_va = va; } CHANGE_PV_LIST_LOCK_TO_PHYS(&lock, pa); TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; if ((newpte & PG_RW) != 0) vm_page_aflag_set(m, PGA_WRITEABLE); } /* * Update the PTE. */ if ((origpte & PG_V) != 0) { validate: origpte = pte_load_store(pte, newpte); KASSERT((origpte & PG_FRAME) == pa, ("pmap_enter: unexpected pa update for %#lx", va)); if ((newpte & PG_M) == 0 && (origpte & (PG_M | PG_RW)) == (PG_M | PG_RW)) { if ((origpte & PG_MANAGED) != 0) vm_page_dirty(m); /* * Although the PTE may still have PG_RW set, TLB * invalidation may nonetheless be required because * the PTE no longer has PG_M set. */ } else if ((origpte & PG_NX) != 0 || (newpte & PG_NX) == 0) { /* * This PTE change does not require TLB invalidation. */ goto unchanged; } if ((origpte & PG_A) != 0) pmap_invalidate_page(pmap, va); } else pte_store(pte, newpte); unchanged: #if VM_NRESERVLEVEL > 0 /* * If both the page table page and the reservation are fully * populated, then attempt promotion. */ if ((mpte == NULL || mpte->ref_count == NPTEPG) && pmap_ps_enabled(pmap) && (m->flags & PG_FICTITIOUS) == 0 && vm_reserv_level_iffullpop(m) == 0) pmap_promote_pde(pmap, pde, va, &lock); #endif rv = KERN_SUCCESS; out: if (lock != NULL) rw_wunlock(lock); PMAP_UNLOCK(pmap); return (rv); } /* * Tries to create a read- and/or execute-only 2MB page mapping. Returns true * if successful. Returns false if (1) a page table page cannot be allocated * without sleeping, (2) a mapping already exists at the specified virtual * address, or (3) a PV entry cannot be allocated without reclaiming another * PV entry. */ static bool pmap_enter_2mpage(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, struct rwlock **lockp) { pd_entry_t newpde; pt_entry_t PG_V; PMAP_LOCK_ASSERT(pmap, MA_OWNED); PG_V = pmap_valid_bit(pmap); newpde = VM_PAGE_TO_PHYS(m) | pmap_cache_bits(pmap, m->md.pat_mode, 1) | PG_PS | PG_V; if ((m->oflags & VPO_UNMANAGED) == 0) newpde |= PG_MANAGED; if ((prot & VM_PROT_EXECUTE) == 0) newpde |= pg_nx; if (va < VM_MAXUSER_ADDRESS) newpde |= PG_U; return (pmap_enter_pde(pmap, va, newpde, PMAP_ENTER_NOSLEEP | PMAP_ENTER_NOREPLACE | PMAP_ENTER_NORECLAIM, NULL, lockp) == KERN_SUCCESS); } /* * Returns true if every page table entry in the specified page table page is * zero. */ static bool pmap_every_pte_zero(vm_paddr_t pa) { pt_entry_t *pt_end, *pte; KASSERT((pa & PAGE_MASK) == 0, ("pa is misaligned")); pte = (pt_entry_t *)PHYS_TO_DMAP(pa); for (pt_end = pte + NPTEPG; pte < pt_end; pte++) { if (*pte != 0) return (false); } return (true); } /* * Tries to create the specified 2MB page mapping. Returns KERN_SUCCESS if * the mapping was created, and either KERN_FAILURE or KERN_RESOURCE_SHORTAGE * otherwise. Returns KERN_FAILURE if PMAP_ENTER_NOREPLACE was specified and * a mapping already exists at the specified virtual address. Returns * KERN_RESOURCE_SHORTAGE if PMAP_ENTER_NOSLEEP was specified and a page table * page allocation failed. Returns KERN_RESOURCE_SHORTAGE if * PMAP_ENTER_NORECLAIM was specified and a PV entry allocation failed. * * The parameter "m" is only used when creating a managed, writeable mapping. */ static int pmap_enter_pde(pmap_t pmap, vm_offset_t va, pd_entry_t newpde, u_int flags, vm_page_t m, struct rwlock **lockp) { struct spglist free; pd_entry_t oldpde, *pde; pt_entry_t PG_G, PG_RW, PG_V; vm_page_t mt, pdpg; KASSERT(pmap == kernel_pmap || (newpde & PG_W) == 0, ("pmap_enter_pde: cannot create wired user mapping")); PG_G = pmap_global_bit(pmap); PG_RW = pmap_rw_bit(pmap); KASSERT((newpde & (pmap_modified_bit(pmap) | PG_RW)) != PG_RW, ("pmap_enter_pde: newpde is missing PG_M")); PG_V = pmap_valid_bit(pmap); PMAP_LOCK_ASSERT(pmap, MA_OWNED); if (!pmap_allow_2m_x_page(pmap, pmap_pde_ept_executable(pmap, newpde))) { CTR2(KTR_PMAP, "pmap_enter_pde: 2m x blocked for va %#lx" " in pmap %p", va, pmap); return (KERN_FAILURE); } if ((pde = pmap_alloc_pde(pmap, va, &pdpg, (flags & PMAP_ENTER_NOSLEEP) != 0 ? NULL : lockp)) == NULL) { CTR2(KTR_PMAP, "pmap_enter_pde: failure for va %#lx" " in pmap %p", va, pmap); return (KERN_RESOURCE_SHORTAGE); } /* * If pkru is not same for the whole pde range, return failure * and let vm_fault() cope. Check after pde allocation, since * it could sleep. */ if (!pmap_pkru_same(pmap, va, va + NBPDR)) { pmap_abort_ptp(pmap, va, pdpg); return (KERN_PROTECTION_FAILURE); } if (va < VM_MAXUSER_ADDRESS && pmap->pm_type == PT_X86) { newpde &= ~X86_PG_PKU_MASK; newpde |= pmap_pkru_get(pmap, va); } /* * If there are existing mappings, either abort or remove them. */ oldpde = *pde; if ((oldpde & PG_V) != 0) { KASSERT(pdpg == NULL || pdpg->ref_count > 1, ("pmap_enter_pde: pdpg's reference count is too low")); if ((flags & PMAP_ENTER_NOREPLACE) != 0 && (va < VM_MAXUSER_ADDRESS || (oldpde & PG_PS) != 0 || !pmap_every_pte_zero(oldpde & PG_FRAME))) { if (pdpg != NULL) pdpg->ref_count--; CTR2(KTR_PMAP, "pmap_enter_pde: failure for va %#lx" " in pmap %p", va, pmap); return (KERN_FAILURE); } /* Break the existing mapping(s). */ SLIST_INIT(&free); if ((oldpde & PG_PS) != 0) { /* * The reference to the PD page that was acquired by * pmap_alloc_pde() ensures that it won't be freed. * However, if the PDE resulted from a promotion, then * a reserved PT page could be freed. */ (void)pmap_remove_pde(pmap, pde, va, &free, lockp); if ((oldpde & PG_G) == 0) pmap_invalidate_pde_page(pmap, va, oldpde); } else { pmap_delayed_invl_start(); if (pmap_remove_ptes(pmap, va, va + NBPDR, pde, &free, lockp)) pmap_invalidate_all(pmap); pmap_delayed_invl_finish(); } if (va < VM_MAXUSER_ADDRESS) { vm_page_free_pages_toq(&free, true); KASSERT(*pde == 0, ("pmap_enter_pde: non-zero pde %p", pde)); } else { KASSERT(SLIST_EMPTY(&free), ("pmap_enter_pde: freed kernel page table page")); /* * Both pmap_remove_pde() and pmap_remove_ptes() will * leave the kernel page table page zero filled. */ mt = PHYS_TO_VM_PAGE(*pde & PG_FRAME); if (pmap_insert_pt_page(pmap, mt, false)) panic("pmap_enter_pde: trie insert failed"); } } if ((newpde & PG_MANAGED) != 0) { /* * Abort this mapping if its PV entry could not be created. */ if (!pmap_pv_insert_pde(pmap, va, newpde, flags, lockp)) { if (pdpg != NULL) pmap_abort_ptp(pmap, va, pdpg); CTR2(KTR_PMAP, "pmap_enter_pde: failure for va %#lx" " in pmap %p", va, pmap); return (KERN_RESOURCE_SHORTAGE); } if ((newpde & PG_RW) != 0) { for (mt = m; mt < &m[NBPDR / PAGE_SIZE]; mt++) vm_page_aflag_set(mt, PGA_WRITEABLE); } } /* * Increment counters. */ if ((newpde & PG_W) != 0) pmap->pm_stats.wired_count += NBPDR / PAGE_SIZE; pmap_resident_count_adj(pmap, NBPDR / PAGE_SIZE); /* * Map the superpage. (This is not a promoted mapping; there will not * be any lingering 4KB page mappings in the TLB.) */ pde_store(pde, newpde); counter_u64_add(pmap_pde_mappings, 1); CTR2(KTR_PMAP, "pmap_enter_pde: success for va %#lx in pmap %p", va, pmap); return (KERN_SUCCESS); } /* * Maps a sequence of resident pages belonging to the same object. * The sequence begins with the given page m_start. This page is * mapped at the given virtual address start. Each subsequent page is * mapped at a virtual address that is offset from start by the same * amount as the page is offset from m_start within the object. The * last page in the sequence is the page with the largest offset from * m_start that can be mapped at a virtual address less than the given * virtual address end. Not every virtual page between start and end * is mapped; only those for which a resident page exists with the * corresponding offset from m_start are mapped. */ void pmap_enter_object(pmap_t pmap, vm_offset_t start, vm_offset_t end, vm_page_t m_start, vm_prot_t prot) { struct rwlock *lock; vm_offset_t va; vm_page_t m, mpte; vm_pindex_t diff, psize; VM_OBJECT_ASSERT_LOCKED(m_start->object); psize = atop(end - start); mpte = NULL; m = m_start; lock = NULL; PMAP_LOCK(pmap); while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) { va = start + ptoa(diff); if ((va & PDRMASK) == 0 && va + NBPDR <= end && m->psind == 1 && pmap_ps_enabled(pmap) && pmap_enter_2mpage(pmap, va, m, prot, &lock)) m = &m[NBPDR / PAGE_SIZE - 1]; else mpte = pmap_enter_quick_locked(pmap, va, m, prot, mpte, &lock); m = TAILQ_NEXT(m, listq); } if (lock != NULL) rw_wunlock(lock); PMAP_UNLOCK(pmap); } /* * this code makes some *MAJOR* assumptions: * 1. Current pmap & pmap exists. * 2. Not wired. * 3. Read access. * 4. No page table pages. * but is *MUCH* faster than pmap_enter... */ void pmap_enter_quick(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot) { struct rwlock *lock; lock = NULL; PMAP_LOCK(pmap); (void)pmap_enter_quick_locked(pmap, va, m, prot, NULL, &lock); if (lock != NULL) rw_wunlock(lock); PMAP_UNLOCK(pmap); } static vm_page_t pmap_enter_quick_locked(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, vm_page_t mpte, struct rwlock **lockp) { pt_entry_t newpte, *pte, PG_V; KASSERT(va < kmi.clean_sva || va >= kmi.clean_eva || (m->oflags & VPO_UNMANAGED) != 0, ("pmap_enter_quick_locked: managed mapping within the clean submap")); PG_V = pmap_valid_bit(pmap); PMAP_LOCK_ASSERT(pmap, MA_OWNED); /* * In the case that a page table page is not * resident, we are creating it here. */ if (va < VM_MAXUSER_ADDRESS) { vm_pindex_t ptepindex; pd_entry_t *ptepa; /* * Calculate pagetable page index */ ptepindex = pmap_pde_pindex(va); if (mpte && (mpte->pindex == ptepindex)) { mpte->ref_count++; } else { /* * Get the page directory entry */ ptepa = pmap_pde(pmap, va); /* * If the page table page is mapped, we just increment * the hold count, and activate it. Otherwise, we * attempt to allocate a page table page. If this * attempt fails, we don't retry. Instead, we give up. */ if (ptepa && (*ptepa & PG_V) != 0) { if (*ptepa & PG_PS) return (NULL); mpte = PHYS_TO_VM_PAGE(*ptepa & PG_FRAME); mpte->ref_count++; } else { /* * Pass NULL instead of the PV list lock * pointer, because we don't intend to sleep. */ mpte = pmap_allocpte_alloc(pmap, ptepindex, NULL, va); if (mpte == NULL) return (mpte); } } pte = (pt_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(mpte)); pte = &pte[pmap_pte_index(va)]; } else { mpte = NULL; pte = vtopte(va); } if (*pte) { if (mpte != NULL) mpte->ref_count--; return (NULL); } /* * Enter on the PV list if part of our managed memory. */ if ((m->oflags & VPO_UNMANAGED) == 0 && !pmap_try_insert_pv_entry(pmap, va, m, lockp)) { if (mpte != NULL) pmap_abort_ptp(pmap, va, mpte); return (NULL); } /* * Increment counters */ pmap_resident_count_adj(pmap, 1); newpte = VM_PAGE_TO_PHYS(m) | PG_V | pmap_cache_bits(pmap, m->md.pat_mode, 0); if ((m->oflags & VPO_UNMANAGED) == 0) newpte |= PG_MANAGED; if ((prot & VM_PROT_EXECUTE) == 0) newpte |= pg_nx; if (va < VM_MAXUSER_ADDRESS) newpte |= PG_U | pmap_pkru_get(pmap, va); pte_store(pte, newpte); return (mpte); } /* * Make a temporary mapping for a physical address. This is only intended * to be used for panic dumps. */ void * pmap_kenter_temporary(vm_paddr_t pa, int i) { vm_offset_t va; va = (vm_offset_t)crashdumpmap + (i * PAGE_SIZE); pmap_kenter(va, pa); invlpg(va); return ((void *)crashdumpmap); } /* * This code maps large physical mmap regions into the * processor address space. Note that some shortcuts * are taken, but the code works. */ void pmap_object_init_pt(pmap_t pmap, vm_offset_t addr, vm_object_t object, vm_pindex_t pindex, vm_size_t size) { pd_entry_t *pde; pt_entry_t PG_A, PG_M, PG_RW, PG_V; vm_paddr_t pa, ptepa; vm_page_t p, pdpg; int pat_mode; PG_A = pmap_accessed_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_V = pmap_valid_bit(pmap); PG_RW = pmap_rw_bit(pmap); VM_OBJECT_ASSERT_WLOCKED(object); KASSERT(object->type == OBJT_DEVICE || object->type == OBJT_SG, ("pmap_object_init_pt: non-device object")); if ((addr & (NBPDR - 1)) == 0 && (size & (NBPDR - 1)) == 0) { if (!pmap_ps_enabled(pmap)) return; if (!vm_object_populate(object, pindex, pindex + atop(size))) return; p = vm_page_lookup(object, pindex); KASSERT(p->valid == VM_PAGE_BITS_ALL, ("pmap_object_init_pt: invalid page %p", p)); pat_mode = p->md.pat_mode; /* * Abort the mapping if the first page is not physically * aligned to a 2MB page boundary. */ ptepa = VM_PAGE_TO_PHYS(p); if (ptepa & (NBPDR - 1)) return; /* * Skip the first page. Abort the mapping if the rest of * the pages are not physically contiguous or have differing * memory attributes. */ p = TAILQ_NEXT(p, listq); for (pa = ptepa + PAGE_SIZE; pa < ptepa + size; pa += PAGE_SIZE) { KASSERT(p->valid == VM_PAGE_BITS_ALL, ("pmap_object_init_pt: invalid page %p", p)); if (pa != VM_PAGE_TO_PHYS(p) || pat_mode != p->md.pat_mode) return; p = TAILQ_NEXT(p, listq); } /* * Map using 2MB pages. Since "ptepa" is 2M aligned and * "size" is a multiple of 2M, adding the PAT setting to "pa" * will not affect the termination of this loop. */ PMAP_LOCK(pmap); for (pa = ptepa | pmap_cache_bits(pmap, pat_mode, 1); pa < ptepa + size; pa += NBPDR) { pde = pmap_alloc_pde(pmap, addr, &pdpg, NULL); if (pde == NULL) { /* * The creation of mappings below is only an * optimization. If a page directory page * cannot be allocated without blocking, * continue on to the next mapping rather than * blocking. */ addr += NBPDR; continue; } if ((*pde & PG_V) == 0) { pde_store(pde, pa | PG_PS | PG_M | PG_A | PG_U | PG_RW | PG_V); pmap_resident_count_adj(pmap, NBPDR / PAGE_SIZE); counter_u64_add(pmap_pde_mappings, 1); } else { /* Continue on if the PDE is already valid. */ pdpg->ref_count--; KASSERT(pdpg->ref_count > 0, ("pmap_object_init_pt: missing reference " "to page directory page, va: 0x%lx", addr)); } addr += NBPDR; } PMAP_UNLOCK(pmap); } } /* * Clear the wired attribute from the mappings for the specified range of * addresses in the given pmap. Every valid mapping within that range * must have the wired attribute set. In contrast, invalid mappings * cannot have the wired attribute set, so they are ignored. * * The wired attribute of the page table entry is not a hardware * feature, so there is no need to invalidate any TLB entries. * Since pmap_demote_pde() for the wired entry must never fail, * pmap_delayed_invl_start()/finish() calls around the * function are not needed. */ void pmap_unwire(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { vm_offset_t va_next; pml4_entry_t *pml4e; pdp_entry_t *pdpe; pd_entry_t *pde; pt_entry_t *pte, PG_V, PG_G; PG_V = pmap_valid_bit(pmap); PG_G = pmap_global_bit(pmap); PMAP_LOCK(pmap); for (; sva < eva; sva = va_next) { pml4e = pmap_pml4e(pmap, sva); if (pml4e == NULL || (*pml4e & PG_V) == 0) { va_next = (sva + NBPML4) & ~PML4MASK; if (va_next < sva) va_next = eva; continue; } va_next = (sva + NBPDP) & ~PDPMASK; if (va_next < sva) va_next = eva; pdpe = pmap_pml4e_to_pdpe(pml4e, sva); if ((*pdpe & PG_V) == 0) continue; if ((*pdpe & PG_PS) != 0) { KASSERT(va_next <= eva, ("partial update of non-transparent 1G mapping " "pdpe %#lx sva %#lx eva %#lx va_next %#lx", *pdpe, sva, eva, va_next)); MPASS(pmap != kernel_pmap); /* XXXKIB */ MPASS((*pdpe & (PG_MANAGED | PG_G)) == 0); atomic_clear_long(pdpe, PG_W); pmap->pm_stats.wired_count -= NBPDP / PAGE_SIZE; continue; } va_next = (sva + NBPDR) & ~PDRMASK; if (va_next < sva) va_next = eva; pde = pmap_pdpe_to_pde(pdpe, sva); if ((*pde & PG_V) == 0) continue; if ((*pde & PG_PS) != 0) { if ((*pde & PG_W) == 0) panic("pmap_unwire: pde %#jx is missing PG_W", (uintmax_t)*pde); /* * Are we unwiring the entire large page? If not, * demote the mapping and fall through. */ if (sva + NBPDR == va_next && eva >= va_next) { atomic_clear_long(pde, PG_W); pmap->pm_stats.wired_count -= NBPDR / PAGE_SIZE; continue; } else if (!pmap_demote_pde(pmap, pde, sva)) panic("pmap_unwire: demotion failed"); } if (va_next > eva) va_next = eva; for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++, sva += PAGE_SIZE) { if ((*pte & PG_V) == 0) continue; if ((*pte & PG_W) == 0) panic("pmap_unwire: pte %#jx is missing PG_W", (uintmax_t)*pte); /* * PG_W must be cleared atomically. Although the pmap * lock synchronizes access to PG_W, another processor * could be setting PG_M and/or PG_A concurrently. */ atomic_clear_long(pte, PG_W); pmap->pm_stats.wired_count--; } } PMAP_UNLOCK(pmap); } /* * Copy the range specified by src_addr/len * from the source map to the range dst_addr/len * in the destination map. * * This routine is only advisory and need not do anything. */ void pmap_copy(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, vm_size_t len, vm_offset_t src_addr) { struct rwlock *lock; pml4_entry_t *pml4e; pdp_entry_t *pdpe; pd_entry_t *pde, srcptepaddr; pt_entry_t *dst_pte, PG_A, PG_M, PG_V, ptetemp, *src_pte; vm_offset_t addr, end_addr, va_next; vm_page_t dst_pdpg, dstmpte, srcmpte; if (dst_addr != src_addr) return; if (dst_pmap->pm_type != src_pmap->pm_type) return; /* * EPT page table entries that require emulation of A/D bits are * sensitive to clearing the PG_A bit (aka EPT_PG_READ). Although * we clear PG_M (aka EPT_PG_WRITE) concomitantly, the PG_U bit * (aka EPT_PG_EXECUTE) could still be set. Since some EPT * implementations flag an EPT misconfiguration for exec-only * mappings we skip this function entirely for emulated pmaps. */ if (pmap_emulate_ad_bits(dst_pmap)) return; end_addr = src_addr + len; lock = NULL; if (dst_pmap < src_pmap) { PMAP_LOCK(dst_pmap); PMAP_LOCK(src_pmap); } else { PMAP_LOCK(src_pmap); PMAP_LOCK(dst_pmap); } PG_A = pmap_accessed_bit(dst_pmap); PG_M = pmap_modified_bit(dst_pmap); PG_V = pmap_valid_bit(dst_pmap); for (addr = src_addr; addr < end_addr; addr = va_next) { KASSERT(addr < UPT_MIN_ADDRESS, ("pmap_copy: invalid to pmap_copy page tables")); pml4e = pmap_pml4e(src_pmap, addr); if (pml4e == NULL || (*pml4e & PG_V) == 0) { va_next = (addr + NBPML4) & ~PML4MASK; if (va_next < addr) va_next = end_addr; continue; } va_next = (addr + NBPDP) & ~PDPMASK; if (va_next < addr) va_next = end_addr; pdpe = pmap_pml4e_to_pdpe(pml4e, addr); if ((*pdpe & PG_V) == 0) continue; if ((*pdpe & PG_PS) != 0) { KASSERT(va_next <= end_addr, ("partial update of non-transparent 1G mapping " "pdpe %#lx sva %#lx eva %#lx va_next %#lx", *pdpe, addr, end_addr, va_next)); MPASS((addr & PDPMASK) == 0); MPASS((*pdpe & PG_MANAGED) == 0); srcptepaddr = *pdpe; pdpe = pmap_pdpe(dst_pmap, addr); if (pdpe == NULL) { if (pmap_allocpte_alloc(dst_pmap, pmap_pml4e_pindex(addr), NULL, addr) == NULL) break; pdpe = pmap_pdpe(dst_pmap, addr); } else { pml4e = pmap_pml4e(dst_pmap, addr); dst_pdpg = PHYS_TO_VM_PAGE(*pml4e & PG_FRAME); dst_pdpg->ref_count++; } KASSERT(*pdpe == 0, ("1G mapping present in dst pmap " "pdpe %#lx sva %#lx eva %#lx va_next %#lx", *pdpe, addr, end_addr, va_next)); *pdpe = srcptepaddr & ~PG_W; pmap_resident_count_adj(dst_pmap, NBPDP / PAGE_SIZE); continue; } va_next = (addr + NBPDR) & ~PDRMASK; if (va_next < addr) va_next = end_addr; pde = pmap_pdpe_to_pde(pdpe, addr); srcptepaddr = *pde; if (srcptepaddr == 0) continue; if (srcptepaddr & PG_PS) { /* * We can only virtual copy whole superpages. */ if ((addr & PDRMASK) != 0 || addr + NBPDR > end_addr) continue; pde = pmap_alloc_pde(dst_pmap, addr, &dst_pdpg, NULL); if (pde == NULL) break; if (*pde == 0 && ((srcptepaddr & PG_MANAGED) == 0 || pmap_pv_insert_pde(dst_pmap, addr, srcptepaddr, PMAP_ENTER_NORECLAIM, &lock))) { /* * We leave the dirty bit unchanged because * managed read/write superpage mappings are * required to be dirty. However, managed * superpage mappings are not required to * have their accessed bit set, so we clear * it because we don't know if this mapping * will be used. */ srcptepaddr &= ~PG_W; if ((srcptepaddr & PG_MANAGED) != 0) srcptepaddr &= ~PG_A; *pde = srcptepaddr; pmap_resident_count_adj(dst_pmap, NBPDR / PAGE_SIZE); counter_u64_add(pmap_pde_mappings, 1); } else pmap_abort_ptp(dst_pmap, addr, dst_pdpg); continue; } srcptepaddr &= PG_FRAME; srcmpte = PHYS_TO_VM_PAGE(srcptepaddr); KASSERT(srcmpte->ref_count > 0, ("pmap_copy: source page table page is unused")); if (va_next > end_addr) va_next = end_addr; src_pte = (pt_entry_t *)PHYS_TO_DMAP(srcptepaddr); src_pte = &src_pte[pmap_pte_index(addr)]; dstmpte = NULL; for (; addr < va_next; addr += PAGE_SIZE, src_pte++) { ptetemp = *src_pte; /* * We only virtual copy managed pages. */ if ((ptetemp & PG_MANAGED) == 0) continue; if (dstmpte != NULL) { KASSERT(dstmpte->pindex == pmap_pde_pindex(addr), ("dstmpte pindex/addr mismatch")); dstmpte->ref_count++; } else if ((dstmpte = pmap_allocpte(dst_pmap, addr, NULL)) == NULL) goto out; dst_pte = (pt_entry_t *) PHYS_TO_DMAP(VM_PAGE_TO_PHYS(dstmpte)); dst_pte = &dst_pte[pmap_pte_index(addr)]; if (*dst_pte == 0 && pmap_try_insert_pv_entry(dst_pmap, addr, PHYS_TO_VM_PAGE(ptetemp & PG_FRAME), &lock)) { /* * Clear the wired, modified, and accessed * (referenced) bits during the copy. */ *dst_pte = ptetemp & ~(PG_W | PG_M | PG_A); pmap_resident_count_adj(dst_pmap, 1); } else { pmap_abort_ptp(dst_pmap, addr, dstmpte); goto out; } /* Have we copied all of the valid mappings? */ if (dstmpte->ref_count >= srcmpte->ref_count) break; } } out: if (lock != NULL) rw_wunlock(lock); PMAP_UNLOCK(src_pmap); PMAP_UNLOCK(dst_pmap); } int pmap_vmspace_copy(pmap_t dst_pmap, pmap_t src_pmap) { int error; if (dst_pmap->pm_type != src_pmap->pm_type || dst_pmap->pm_type != PT_X86 || (cpu_stdext_feature2 & CPUID_STDEXT2_PKU) == 0) return (0); for (;;) { if (dst_pmap < src_pmap) { PMAP_LOCK(dst_pmap); PMAP_LOCK(src_pmap); } else { PMAP_LOCK(src_pmap); PMAP_LOCK(dst_pmap); } error = pmap_pkru_copy(dst_pmap, src_pmap); /* Clean up partial copy on failure due to no memory. */ if (error == ENOMEM) pmap_pkru_deassign_all(dst_pmap); PMAP_UNLOCK(src_pmap); PMAP_UNLOCK(dst_pmap); if (error != ENOMEM) break; vm_wait(NULL); } return (error); } /* * Zero the specified hardware page. */ void pmap_zero_page(vm_page_t m) { vm_offset_t va = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)); pagezero((void *)va); } /* * Zero an an area within a single hardware page. off and size must not * cover an area beyond a single hardware page. */ void pmap_zero_page_area(vm_page_t m, int off, int size) { vm_offset_t va = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)); if (off == 0 && size == PAGE_SIZE) pagezero((void *)va); else bzero((char *)va + off, size); } /* * Copy 1 specified hardware page to another. */ void pmap_copy_page(vm_page_t msrc, vm_page_t mdst) { vm_offset_t src = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(msrc)); vm_offset_t dst = PHYS_TO_DMAP(VM_PAGE_TO_PHYS(mdst)); pagecopy((void *)src, (void *)dst); } int unmapped_buf_allowed = 1; void pmap_copy_pages(vm_page_t ma[], vm_offset_t a_offset, vm_page_t mb[], vm_offset_t b_offset, int xfersize) { void *a_cp, *b_cp; vm_page_t pages[2]; vm_offset_t vaddr[2], a_pg_offset, b_pg_offset; int cnt; boolean_t mapped; while (xfersize > 0) { a_pg_offset = a_offset & PAGE_MASK; pages[0] = ma[a_offset >> PAGE_SHIFT]; b_pg_offset = b_offset & PAGE_MASK; pages[1] = mb[b_offset >> PAGE_SHIFT]; cnt = min(xfersize, PAGE_SIZE - a_pg_offset); cnt = min(cnt, PAGE_SIZE - b_pg_offset); mapped = pmap_map_io_transient(pages, vaddr, 2, FALSE); a_cp = (char *)vaddr[0] + a_pg_offset; b_cp = (char *)vaddr[1] + b_pg_offset; bcopy(a_cp, b_cp, cnt); if (__predict_false(mapped)) pmap_unmap_io_transient(pages, vaddr, 2, FALSE); a_offset += cnt; b_offset += cnt; xfersize -= cnt; } } /* * Returns true if the pmap's pv is one of the first * 16 pvs linked to from this page. This count may * be changed upwards or downwards in the future; it * is only necessary that true be returned for a small * subset of pmaps for proper page aging. */ boolean_t pmap_page_exists_quick(pmap_t pmap, vm_page_t m) { struct md_page *pvh; struct rwlock *lock; pv_entry_t pv; int loops = 0; boolean_t rv; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_page_exists_quick: page %p is not managed", m)); rv = FALSE; lock = VM_PAGE_TO_PV_LIST_LOCK(m); rw_rlock(lock); TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) { if (PV_PMAP(pv) == pmap) { rv = TRUE; break; } loops++; if (loops >= 16) break; } if (!rv && loops < 16 && (m->flags & PG_FICTITIOUS) == 0) { pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) { if (PV_PMAP(pv) == pmap) { rv = TRUE; break; } loops++; if (loops >= 16) break; } } rw_runlock(lock); return (rv); } /* * pmap_page_wired_mappings: * * Return the number of managed mappings to the given physical page * that are wired. */ int pmap_page_wired_mappings(vm_page_t m) { struct rwlock *lock; struct md_page *pvh; pmap_t pmap; pt_entry_t *pte; pv_entry_t pv; int count, md_gen, pvh_gen; if ((m->oflags & VPO_UNMANAGED) != 0) return (0); lock = VM_PAGE_TO_PV_LIST_LOCK(m); rw_rlock(lock); restart: count = 0; TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { md_gen = m->md.pv_gen; rw_runlock(lock); PMAP_LOCK(pmap); rw_rlock(lock); if (md_gen != m->md.pv_gen) { PMAP_UNLOCK(pmap); goto restart; } } pte = pmap_pte(pmap, pv->pv_va); if ((*pte & PG_W) != 0) count++; PMAP_UNLOCK(pmap); } if ((m->flags & PG_FICTITIOUS) == 0) { pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { md_gen = m->md.pv_gen; pvh_gen = pvh->pv_gen; rw_runlock(lock); PMAP_LOCK(pmap); rw_rlock(lock); if (md_gen != m->md.pv_gen || pvh_gen != pvh->pv_gen) { PMAP_UNLOCK(pmap); goto restart; } } pte = pmap_pde(pmap, pv->pv_va); if ((*pte & PG_W) != 0) count++; PMAP_UNLOCK(pmap); } } rw_runlock(lock); return (count); } /* * Returns TRUE if the given page is mapped individually or as part of * a 2mpage. Otherwise, returns FALSE. */ boolean_t pmap_page_is_mapped(vm_page_t m) { struct rwlock *lock; boolean_t rv; if ((m->oflags & VPO_UNMANAGED) != 0) return (FALSE); lock = VM_PAGE_TO_PV_LIST_LOCK(m); rw_rlock(lock); rv = !TAILQ_EMPTY(&m->md.pv_list) || ((m->flags & PG_FICTITIOUS) == 0 && !TAILQ_EMPTY(&pa_to_pvh(VM_PAGE_TO_PHYS(m))->pv_list)); rw_runlock(lock); return (rv); } /* * Destroy all managed, non-wired mappings in the given user-space * pmap. This pmap cannot be active on any processor besides the * caller. * * This function cannot be applied to the kernel pmap. Moreover, it * is not intended for general use. It is only to be used during * process termination. Consequently, it can be implemented in ways * that make it faster than pmap_remove(). First, it can more quickly * destroy mappings by iterating over the pmap's collection of PV * entries, rather than searching the page table. Second, it doesn't * have to test and clear the page table entries atomically, because * no processor is currently accessing the user address space. In * particular, a page table entry's dirty bit won't change state once * this function starts. * * Although this function destroys all of the pmap's managed, * non-wired mappings, it can delay and batch the invalidation of TLB * entries without calling pmap_delayed_invl_start() and * pmap_delayed_invl_finish(). Because the pmap is not active on * any other processor, none of these TLB entries will ever be used * before their eventual invalidation. Consequently, there is no need * for either pmap_remove_all() or pmap_remove_write() to wait for * that eventual TLB invalidation. */ void pmap_remove_pages(pmap_t pmap) { pd_entry_t ptepde; pt_entry_t *pte, tpte; pt_entry_t PG_M, PG_RW, PG_V; struct spglist free; struct pv_chunklist free_chunks[PMAP_MEMDOM]; vm_page_t m, mpte, mt; pv_entry_t pv; struct md_page *pvh; struct pv_chunk *pc, *npc; struct rwlock *lock; int64_t bit; uint64_t inuse, bitmask; int allfree, field, freed, i, idx; boolean_t superpage; vm_paddr_t pa; /* * Assert that the given pmap is only active on the current * CPU. Unfortunately, we cannot block another CPU from * activating the pmap while this function is executing. */ KASSERT(pmap == PCPU_GET(curpmap), ("non-current pmap %p", pmap)); #ifdef INVARIANTS { cpuset_t other_cpus; other_cpus = all_cpus; critical_enter(); CPU_CLR(PCPU_GET(cpuid), &other_cpus); - CPU_AND(&other_cpus, &pmap->pm_active); + CPU_AND(&other_cpus, &other_cpus, &pmap->pm_active); critical_exit(); KASSERT(CPU_EMPTY(&other_cpus), ("pmap active %p", pmap)); } #endif lock = NULL; PG_M = pmap_modified_bit(pmap); PG_V = pmap_valid_bit(pmap); PG_RW = pmap_rw_bit(pmap); for (i = 0; i < PMAP_MEMDOM; i++) TAILQ_INIT(&free_chunks[i]); SLIST_INIT(&free); PMAP_LOCK(pmap); TAILQ_FOREACH_SAFE(pc, &pmap->pm_pvchunk, pc_list, npc) { allfree = 1; freed = 0; for (field = 0; field < _NPCM; field++) { inuse = ~pc->pc_map[field] & pc_freemask[field]; while (inuse != 0) { bit = bsfq(inuse); bitmask = 1UL << bit; idx = field * 64 + bit; pv = &pc->pc_pventry[idx]; inuse &= ~bitmask; pte = pmap_pdpe(pmap, pv->pv_va); ptepde = *pte; pte = pmap_pdpe_to_pde(pte, pv->pv_va); tpte = *pte; if ((tpte & (PG_PS | PG_V)) == PG_V) { superpage = FALSE; ptepde = tpte; pte = (pt_entry_t *)PHYS_TO_DMAP(tpte & PG_FRAME); pte = &pte[pmap_pte_index(pv->pv_va)]; tpte = *pte; } else { /* * Keep track whether 'tpte' is a * superpage explicitly instead of * relying on PG_PS being set. * * This is because PG_PS is numerically * identical to PG_PTE_PAT and thus a * regular page could be mistaken for * a superpage. */ superpage = TRUE; } if ((tpte & PG_V) == 0) { panic("bad pte va %lx pte %lx", pv->pv_va, tpte); } /* * We cannot remove wired pages from a process' mapping at this time */ if (tpte & PG_W) { allfree = 0; continue; } /* Mark free */ pc->pc_map[field] |= bitmask; /* * Because this pmap is not active on other * processors, the dirty bit cannot have * changed state since we last loaded pte. */ pte_clear(pte); if (superpage) pa = tpte & PG_PS_FRAME; else pa = tpte & PG_FRAME; m = PHYS_TO_VM_PAGE(pa); KASSERT(m->phys_addr == pa, ("vm_page_t %p phys_addr mismatch %016jx %016jx", m, (uintmax_t)m->phys_addr, (uintmax_t)tpte)); KASSERT((m->flags & PG_FICTITIOUS) != 0 || m < &vm_page_array[vm_page_array_size], ("pmap_remove_pages: bad tpte %#jx", (uintmax_t)tpte)); /* * Update the vm_page_t clean/reference bits. */ if ((tpte & (PG_M | PG_RW)) == (PG_M | PG_RW)) { if (superpage) { for (mt = m; mt < &m[NBPDR / PAGE_SIZE]; mt++) vm_page_dirty(mt); } else vm_page_dirty(m); } CHANGE_PV_LIST_LOCK_TO_VM_PAGE(&lock, m); if (superpage) { pmap_resident_count_adj(pmap, -NBPDR / PAGE_SIZE); pvh = pa_to_pvh(tpte & PG_PS_FRAME); TAILQ_REMOVE(&pvh->pv_list, pv, pv_next); pvh->pv_gen++; if (TAILQ_EMPTY(&pvh->pv_list)) { for (mt = m; mt < &m[NBPDR / PAGE_SIZE]; mt++) if ((mt->a.flags & PGA_WRITEABLE) != 0 && TAILQ_EMPTY(&mt->md.pv_list)) vm_page_aflag_clear(mt, PGA_WRITEABLE); } mpte = pmap_remove_pt_page(pmap, pv->pv_va); if (mpte != NULL) { KASSERT(mpte->valid == VM_PAGE_BITS_ALL, ("pmap_remove_pages: pte page not promoted")); pmap_resident_count_adj(pmap, -1); KASSERT(mpte->ref_count == NPTEPG, ("pmap_remove_pages: pte page reference count error")); mpte->ref_count = 0; pmap_add_delayed_free_list(mpte, &free, FALSE); } } else { pmap_resident_count_adj(pmap, -1); TAILQ_REMOVE(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; if ((m->a.flags & PGA_WRITEABLE) != 0 && TAILQ_EMPTY(&m->md.pv_list) && (m->flags & PG_FICTITIOUS) == 0) { pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); if (TAILQ_EMPTY(&pvh->pv_list)) vm_page_aflag_clear(m, PGA_WRITEABLE); } } pmap_unuse_pt(pmap, pv->pv_va, ptepde, &free); freed++; } } PV_STAT(counter_u64_add(pv_entry_frees, freed)); PV_STAT(counter_u64_add(pv_entry_spare, freed)); PV_STAT(counter_u64_add(pv_entry_count, -freed)); if (allfree) { TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); TAILQ_INSERT_TAIL(&free_chunks[pc_to_domain(pc)], pc, pc_list); } } if (lock != NULL) rw_wunlock(lock); pmap_invalidate_all(pmap); pmap_pkru_deassign_all(pmap); free_pv_chunk_batch((struct pv_chunklist *)&free_chunks); PMAP_UNLOCK(pmap); vm_page_free_pages_toq(&free, true); } static boolean_t pmap_page_test_mappings(vm_page_t m, boolean_t accessed, boolean_t modified) { struct rwlock *lock; pv_entry_t pv; struct md_page *pvh; pt_entry_t *pte, mask; pt_entry_t PG_A, PG_M, PG_RW, PG_V; pmap_t pmap; int md_gen, pvh_gen; boolean_t rv; rv = FALSE; lock = VM_PAGE_TO_PV_LIST_LOCK(m); rw_rlock(lock); restart: TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { md_gen = m->md.pv_gen; rw_runlock(lock); PMAP_LOCK(pmap); rw_rlock(lock); if (md_gen != m->md.pv_gen) { PMAP_UNLOCK(pmap); goto restart; } } pte = pmap_pte(pmap, pv->pv_va); mask = 0; if (modified) { PG_M = pmap_modified_bit(pmap); PG_RW = pmap_rw_bit(pmap); mask |= PG_RW | PG_M; } if (accessed) { PG_A = pmap_accessed_bit(pmap); PG_V = pmap_valid_bit(pmap); mask |= PG_V | PG_A; } rv = (*pte & mask) == mask; PMAP_UNLOCK(pmap); if (rv) goto out; } if ((m->flags & PG_FICTITIOUS) == 0) { pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { md_gen = m->md.pv_gen; pvh_gen = pvh->pv_gen; rw_runlock(lock); PMAP_LOCK(pmap); rw_rlock(lock); if (md_gen != m->md.pv_gen || pvh_gen != pvh->pv_gen) { PMAP_UNLOCK(pmap); goto restart; } } pte = pmap_pde(pmap, pv->pv_va); mask = 0; if (modified) { PG_M = pmap_modified_bit(pmap); PG_RW = pmap_rw_bit(pmap); mask |= PG_RW | PG_M; } if (accessed) { PG_A = pmap_accessed_bit(pmap); PG_V = pmap_valid_bit(pmap); mask |= PG_V | PG_A; } rv = (*pte & mask) == mask; PMAP_UNLOCK(pmap); if (rv) goto out; } } out: rw_runlock(lock); return (rv); } /* * pmap_is_modified: * * Return whether or not the specified physical page was modified * in any physical maps. */ boolean_t pmap_is_modified(vm_page_t m) { KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_is_modified: page %p is not managed", m)); /* * If the page is not busied then this check is racy. */ if (!pmap_page_is_write_mapped(m)) return (FALSE); return (pmap_page_test_mappings(m, FALSE, TRUE)); } /* * pmap_is_prefaultable: * * Return whether or not the specified virtual address is eligible * for prefault. */ boolean_t pmap_is_prefaultable(pmap_t pmap, vm_offset_t addr) { pd_entry_t *pde; pt_entry_t *pte, PG_V; boolean_t rv; PG_V = pmap_valid_bit(pmap); rv = FALSE; PMAP_LOCK(pmap); pde = pmap_pde(pmap, addr); if (pde != NULL && (*pde & (PG_PS | PG_V)) == PG_V) { pte = pmap_pde_to_pte(pde, addr); rv = (*pte & PG_V) == 0; } PMAP_UNLOCK(pmap); return (rv); } /* * pmap_is_referenced: * * Return whether or not the specified physical page was referenced * in any physical maps. */ boolean_t pmap_is_referenced(vm_page_t m) { KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_is_referenced: page %p is not managed", m)); return (pmap_page_test_mappings(m, TRUE, FALSE)); } /* * Clear the write and modified bits in each of the given page's mappings. */ void pmap_remove_write(vm_page_t m) { struct md_page *pvh; pmap_t pmap; struct rwlock *lock; pv_entry_t next_pv, pv; pd_entry_t *pde; pt_entry_t oldpte, *pte, PG_M, PG_RW; vm_offset_t va; int pvh_gen, md_gen; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_remove_write: page %p is not managed", m)); vm_page_assert_busied(m); if (!pmap_page_is_write_mapped(m)) return; lock = VM_PAGE_TO_PV_LIST_LOCK(m); pvh = (m->flags & PG_FICTITIOUS) != 0 ? &pv_dummy : pa_to_pvh(VM_PAGE_TO_PHYS(m)); rw_wlock(lock); retry: TAILQ_FOREACH_SAFE(pv, &pvh->pv_list, pv_next, next_pv) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { pvh_gen = pvh->pv_gen; rw_wunlock(lock); PMAP_LOCK(pmap); rw_wlock(lock); if (pvh_gen != pvh->pv_gen) { PMAP_UNLOCK(pmap); goto retry; } } PG_RW = pmap_rw_bit(pmap); va = pv->pv_va; pde = pmap_pde(pmap, va); if ((*pde & PG_RW) != 0) (void)pmap_demote_pde_locked(pmap, pde, va, &lock); KASSERT(lock == VM_PAGE_TO_PV_LIST_LOCK(m), ("inconsistent pv lock %p %p for page %p", lock, VM_PAGE_TO_PV_LIST_LOCK(m), m)); PMAP_UNLOCK(pmap); } TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { pvh_gen = pvh->pv_gen; md_gen = m->md.pv_gen; rw_wunlock(lock); PMAP_LOCK(pmap); rw_wlock(lock); if (pvh_gen != pvh->pv_gen || md_gen != m->md.pv_gen) { PMAP_UNLOCK(pmap); goto retry; } } PG_M = pmap_modified_bit(pmap); PG_RW = pmap_rw_bit(pmap); pde = pmap_pde(pmap, pv->pv_va); KASSERT((*pde & PG_PS) == 0, ("pmap_remove_write: found a 2mpage in page %p's pv list", m)); pte = pmap_pde_to_pte(pde, pv->pv_va); oldpte = *pte; if (oldpte & PG_RW) { while (!atomic_fcmpset_long(pte, &oldpte, oldpte & ~(PG_RW | PG_M))) cpu_spinwait(); if ((oldpte & PG_M) != 0) vm_page_dirty(m); pmap_invalidate_page(pmap, pv->pv_va); } PMAP_UNLOCK(pmap); } rw_wunlock(lock); vm_page_aflag_clear(m, PGA_WRITEABLE); pmap_delayed_invl_wait(m); } static __inline boolean_t safe_to_clear_referenced(pmap_t pmap, pt_entry_t pte) { if (!pmap_emulate_ad_bits(pmap)) return (TRUE); KASSERT(pmap->pm_type == PT_EPT, ("invalid pm_type %d", pmap->pm_type)); /* * XWR = 010 or 110 will cause an unconditional EPT misconfiguration * so we don't let the referenced (aka EPT_PG_READ) bit to be cleared * if the EPT_PG_WRITE bit is set. */ if ((pte & EPT_PG_WRITE) != 0) return (FALSE); /* * XWR = 100 is allowed only if the PMAP_SUPPORTS_EXEC_ONLY is set. */ if ((pte & EPT_PG_EXECUTE) == 0 || ((pmap->pm_flags & PMAP_SUPPORTS_EXEC_ONLY) != 0)) return (TRUE); else return (FALSE); } /* * pmap_ts_referenced: * * Return a count of reference bits for a page, clearing those bits. * It is not necessary for every reference bit to be cleared, but it * is necessary that 0 only be returned when there are truly no * reference bits set. * * As an optimization, update the page's dirty field if a modified bit is * found while counting reference bits. This opportunistic update can be * performed at low cost and can eliminate the need for some future calls * to pmap_is_modified(). However, since this function stops after * finding PMAP_TS_REFERENCED_MAX reference bits, it may not detect some * dirty pages. Those dirty pages will only be detected by a future call * to pmap_is_modified(). * * A DI block is not needed within this function, because * invalidations are performed before the PV list lock is * released. */ int pmap_ts_referenced(vm_page_t m) { struct md_page *pvh; pv_entry_t pv, pvf; pmap_t pmap; struct rwlock *lock; pd_entry_t oldpde, *pde; pt_entry_t *pte, PG_A, PG_M, PG_RW; vm_offset_t va; vm_paddr_t pa; int cleared, md_gen, not_cleared, pvh_gen; struct spglist free; boolean_t demoted; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_ts_referenced: page %p is not managed", m)); SLIST_INIT(&free); cleared = 0; pa = VM_PAGE_TO_PHYS(m); lock = PHYS_TO_PV_LIST_LOCK(pa); pvh = (m->flags & PG_FICTITIOUS) != 0 ? &pv_dummy : pa_to_pvh(pa); rw_wlock(lock); retry: not_cleared = 0; if ((pvf = TAILQ_FIRST(&pvh->pv_list)) == NULL) goto small_mappings; pv = pvf; do { if (pvf == NULL) pvf = pv; pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { pvh_gen = pvh->pv_gen; rw_wunlock(lock); PMAP_LOCK(pmap); rw_wlock(lock); if (pvh_gen != pvh->pv_gen) { PMAP_UNLOCK(pmap); goto retry; } } PG_A = pmap_accessed_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_RW = pmap_rw_bit(pmap); va = pv->pv_va; pde = pmap_pde(pmap, pv->pv_va); oldpde = *pde; if ((oldpde & (PG_M | PG_RW)) == (PG_M | PG_RW)) { /* * Although "oldpde" is mapping a 2MB page, because * this function is called at a 4KB page granularity, * we only update the 4KB page under test. */ vm_page_dirty(m); } if ((oldpde & PG_A) != 0) { /* * Since this reference bit is shared by 512 4KB * pages, it should not be cleared every time it is * tested. Apply a simple "hash" function on the * physical page number, the virtual superpage number, * and the pmap address to select one 4KB page out of * the 512 on which testing the reference bit will * result in clearing that reference bit. This * function is designed to avoid the selection of the * same 4KB page for every 2MB page mapping. * * On demotion, a mapping that hasn't been referenced * is simply destroyed. To avoid the possibility of a * subsequent page fault on a demoted wired mapping, * always leave its reference bit set. Moreover, * since the superpage is wired, the current state of * its reference bit won't affect page replacement. */ if ((((pa >> PAGE_SHIFT) ^ (pv->pv_va >> PDRSHIFT) ^ (uintptr_t)pmap) & (NPTEPG - 1)) == 0 && (oldpde & PG_W) == 0) { if (safe_to_clear_referenced(pmap, oldpde)) { atomic_clear_long(pde, PG_A); pmap_invalidate_page(pmap, pv->pv_va); demoted = FALSE; } else if (pmap_demote_pde_locked(pmap, pde, pv->pv_va, &lock)) { /* * Remove the mapping to a single page * so that a subsequent access may * repromote. Since the underlying * page table page is fully populated, * this removal never frees a page * table page. */ demoted = TRUE; va += VM_PAGE_TO_PHYS(m) - (oldpde & PG_PS_FRAME); pte = pmap_pde_to_pte(pde, va); pmap_remove_pte(pmap, pte, va, *pde, NULL, &lock); pmap_invalidate_page(pmap, va); } else demoted = TRUE; if (demoted) { /* * The superpage mapping was removed * entirely and therefore 'pv' is no * longer valid. */ if (pvf == pv) pvf = NULL; pv = NULL; } cleared++; KASSERT(lock == VM_PAGE_TO_PV_LIST_LOCK(m), ("inconsistent pv lock %p %p for page %p", lock, VM_PAGE_TO_PV_LIST_LOCK(m), m)); } else not_cleared++; } PMAP_UNLOCK(pmap); /* Rotate the PV list if it has more than one entry. */ if (pv != NULL && TAILQ_NEXT(pv, pv_next) != NULL) { TAILQ_REMOVE(&pvh->pv_list, pv, pv_next); TAILQ_INSERT_TAIL(&pvh->pv_list, pv, pv_next); pvh->pv_gen++; } if (cleared + not_cleared >= PMAP_TS_REFERENCED_MAX) goto out; } while ((pv = TAILQ_FIRST(&pvh->pv_list)) != pvf); small_mappings: if ((pvf = TAILQ_FIRST(&m->md.pv_list)) == NULL) goto out; pv = pvf; do { if (pvf == NULL) pvf = pv; pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { pvh_gen = pvh->pv_gen; md_gen = m->md.pv_gen; rw_wunlock(lock); PMAP_LOCK(pmap); rw_wlock(lock); if (pvh_gen != pvh->pv_gen || md_gen != m->md.pv_gen) { PMAP_UNLOCK(pmap); goto retry; } } PG_A = pmap_accessed_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_RW = pmap_rw_bit(pmap); pde = pmap_pde(pmap, pv->pv_va); KASSERT((*pde & PG_PS) == 0, ("pmap_ts_referenced: found a 2mpage in page %p's pv list", m)); pte = pmap_pde_to_pte(pde, pv->pv_va); if ((*pte & (PG_M | PG_RW)) == (PG_M | PG_RW)) vm_page_dirty(m); if ((*pte & PG_A) != 0) { if (safe_to_clear_referenced(pmap, *pte)) { atomic_clear_long(pte, PG_A); pmap_invalidate_page(pmap, pv->pv_va); cleared++; } else if ((*pte & PG_W) == 0) { /* * Wired pages cannot be paged out so * doing accessed bit emulation for * them is wasted effort. We do the * hard work for unwired pages only. */ pmap_remove_pte(pmap, pte, pv->pv_va, *pde, &free, &lock); pmap_invalidate_page(pmap, pv->pv_va); cleared++; if (pvf == pv) pvf = NULL; pv = NULL; KASSERT(lock == VM_PAGE_TO_PV_LIST_LOCK(m), ("inconsistent pv lock %p %p for page %p", lock, VM_PAGE_TO_PV_LIST_LOCK(m), m)); } else not_cleared++; } PMAP_UNLOCK(pmap); /* Rotate the PV list if it has more than one entry. */ if (pv != NULL && TAILQ_NEXT(pv, pv_next) != NULL) { TAILQ_REMOVE(&m->md.pv_list, pv, pv_next); TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next); m->md.pv_gen++; } } while ((pv = TAILQ_FIRST(&m->md.pv_list)) != pvf && cleared + not_cleared < PMAP_TS_REFERENCED_MAX); out: rw_wunlock(lock); vm_page_free_pages_toq(&free, true); return (cleared + not_cleared); } /* * Apply the given advice to the specified range of addresses within the * given pmap. Depending on the advice, clear the referenced and/or * modified flags in each mapping and set the mapped page's dirty field. */ void pmap_advise(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, int advice) { struct rwlock *lock; pml4_entry_t *pml4e; pdp_entry_t *pdpe; pd_entry_t oldpde, *pde; pt_entry_t *pte, PG_A, PG_G, PG_M, PG_RW, PG_V; vm_offset_t va, va_next; vm_page_t m; bool anychanged; if (advice != MADV_DONTNEED && advice != MADV_FREE) return; /* * A/D bit emulation requires an alternate code path when clearing * the modified and accessed bits below. Since this function is * advisory in nature we skip it entirely for pmaps that require * A/D bit emulation. */ if (pmap_emulate_ad_bits(pmap)) return; PG_A = pmap_accessed_bit(pmap); PG_G = pmap_global_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_V = pmap_valid_bit(pmap); PG_RW = pmap_rw_bit(pmap); anychanged = false; pmap_delayed_invl_start(); PMAP_LOCK(pmap); for (; sva < eva; sva = va_next) { pml4e = pmap_pml4e(pmap, sva); if (pml4e == NULL || (*pml4e & PG_V) == 0) { va_next = (sva + NBPML4) & ~PML4MASK; if (va_next < sva) va_next = eva; continue; } va_next = (sva + NBPDP) & ~PDPMASK; if (va_next < sva) va_next = eva; pdpe = pmap_pml4e_to_pdpe(pml4e, sva); if ((*pdpe & PG_V) == 0) continue; if ((*pdpe & PG_PS) != 0) { KASSERT(va_next <= eva, ("partial update of non-transparent 1G mapping " "pdpe %#lx sva %#lx eva %#lx va_next %#lx", *pdpe, sva, eva, va_next)); continue; } va_next = (sva + NBPDR) & ~PDRMASK; if (va_next < sva) va_next = eva; pde = pmap_pdpe_to_pde(pdpe, sva); oldpde = *pde; if ((oldpde & PG_V) == 0) continue; else if ((oldpde & PG_PS) != 0) { if ((oldpde & PG_MANAGED) == 0) continue; lock = NULL; if (!pmap_demote_pde_locked(pmap, pde, sva, &lock)) { if (lock != NULL) rw_wunlock(lock); /* * The large page mapping was destroyed. */ continue; } /* * Unless the page mappings are wired, remove the * mapping to a single page so that a subsequent * access may repromote. Choosing the last page * within the address range [sva, min(va_next, eva)) * generally results in more repromotions. Since the * underlying page table page is fully populated, this * removal never frees a page table page. */ if ((oldpde & PG_W) == 0) { va = eva; if (va > va_next) va = va_next; va -= PAGE_SIZE; KASSERT(va >= sva, ("pmap_advise: no address gap")); pte = pmap_pde_to_pte(pde, va); KASSERT((*pte & PG_V) != 0, ("pmap_advise: invalid PTE")); pmap_remove_pte(pmap, pte, va, *pde, NULL, &lock); anychanged = true; } if (lock != NULL) rw_wunlock(lock); } if (va_next > eva) va_next = eva; va = va_next; for (pte = pmap_pde_to_pte(pde, sva); sva != va_next; pte++, sva += PAGE_SIZE) { if ((*pte & (PG_MANAGED | PG_V)) != (PG_MANAGED | PG_V)) goto maybe_invlrng; else if ((*pte & (PG_M | PG_RW)) == (PG_M | PG_RW)) { if (advice == MADV_DONTNEED) { /* * Future calls to pmap_is_modified() * can be avoided by making the page * dirty now. */ m = PHYS_TO_VM_PAGE(*pte & PG_FRAME); vm_page_dirty(m); } atomic_clear_long(pte, PG_M | PG_A); } else if ((*pte & PG_A) != 0) atomic_clear_long(pte, PG_A); else goto maybe_invlrng; if ((*pte & PG_G) != 0) { if (va == va_next) va = sva; } else anychanged = true; continue; maybe_invlrng: if (va != va_next) { pmap_invalidate_range(pmap, va, sva); va = va_next; } } if (va != va_next) pmap_invalidate_range(pmap, va, sva); } if (anychanged) pmap_invalidate_all(pmap); PMAP_UNLOCK(pmap); pmap_delayed_invl_finish(); } /* * Clear the modify bits on the specified physical page. */ void pmap_clear_modify(vm_page_t m) { struct md_page *pvh; pmap_t pmap; pv_entry_t next_pv, pv; pd_entry_t oldpde, *pde; pt_entry_t *pte, PG_M, PG_RW; struct rwlock *lock; vm_offset_t va; int md_gen, pvh_gen; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_clear_modify: page %p is not managed", m)); vm_page_assert_busied(m); if (!pmap_page_is_write_mapped(m)) return; pvh = (m->flags & PG_FICTITIOUS) != 0 ? &pv_dummy : pa_to_pvh(VM_PAGE_TO_PHYS(m)); lock = VM_PAGE_TO_PV_LIST_LOCK(m); rw_wlock(lock); restart: TAILQ_FOREACH_SAFE(pv, &pvh->pv_list, pv_next, next_pv) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { pvh_gen = pvh->pv_gen; rw_wunlock(lock); PMAP_LOCK(pmap); rw_wlock(lock); if (pvh_gen != pvh->pv_gen) { PMAP_UNLOCK(pmap); goto restart; } } PG_M = pmap_modified_bit(pmap); PG_RW = pmap_rw_bit(pmap); va = pv->pv_va; pde = pmap_pde(pmap, va); oldpde = *pde; /* If oldpde has PG_RW set, then it also has PG_M set. */ if ((oldpde & PG_RW) != 0 && pmap_demote_pde_locked(pmap, pde, va, &lock) && (oldpde & PG_W) == 0) { /* * Write protect the mapping to a single page so that * a subsequent write access may repromote. */ va += VM_PAGE_TO_PHYS(m) - (oldpde & PG_PS_FRAME); pte = pmap_pde_to_pte(pde, va); atomic_clear_long(pte, PG_M | PG_RW); vm_page_dirty(m); pmap_invalidate_page(pmap, va); } PMAP_UNLOCK(pmap); } TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) { pmap = PV_PMAP(pv); if (!PMAP_TRYLOCK(pmap)) { md_gen = m->md.pv_gen; pvh_gen = pvh->pv_gen; rw_wunlock(lock); PMAP_LOCK(pmap); rw_wlock(lock); if (pvh_gen != pvh->pv_gen || md_gen != m->md.pv_gen) { PMAP_UNLOCK(pmap); goto restart; } } PG_M = pmap_modified_bit(pmap); PG_RW = pmap_rw_bit(pmap); pde = pmap_pde(pmap, pv->pv_va); KASSERT((*pde & PG_PS) == 0, ("pmap_clear_modify: found" " a 2mpage in page %p's pv list", m)); pte = pmap_pde_to_pte(pde, pv->pv_va); if ((*pte & (PG_M | PG_RW)) == (PG_M | PG_RW)) { atomic_clear_long(pte, PG_M); pmap_invalidate_page(pmap, pv->pv_va); } PMAP_UNLOCK(pmap); } rw_wunlock(lock); } /* * Miscellaneous support routines follow */ /* Adjust the properties for a leaf page table entry. */ static __inline void pmap_pte_props(pt_entry_t *pte, u_long bits, u_long mask) { u_long opte, npte; opte = *(u_long *)pte; do { npte = opte & ~mask; npte |= bits; } while (npte != opte && !atomic_fcmpset_long((u_long *)pte, &opte, npte)); } /* * Map a set of physical memory pages into the kernel virtual * address space. Return a pointer to where it is mapped. This * routine is intended to be used for mapping device memory, * NOT real memory. */ static void * pmap_mapdev_internal(vm_paddr_t pa, vm_size_t size, int mode, int flags) { struct pmap_preinit_mapping *ppim; vm_offset_t va, offset; vm_size_t tmpsize; int i; offset = pa & PAGE_MASK; size = round_page(offset + size); pa = trunc_page(pa); if (!pmap_initialized) { va = 0; for (i = 0; i < PMAP_PREINIT_MAPPING_COUNT; i++) { ppim = pmap_preinit_mapping + i; if (ppim->va == 0) { ppim->pa = pa; ppim->sz = size; ppim->mode = mode; ppim->va = virtual_avail; virtual_avail += size; va = ppim->va; break; } } if (va == 0) panic("%s: too many preinit mappings", __func__); } else { /* * If we have a preinit mapping, re-use it. */ for (i = 0; i < PMAP_PREINIT_MAPPING_COUNT; i++) { ppim = pmap_preinit_mapping + i; if (ppim->pa == pa && ppim->sz == size && (ppim->mode == mode || (flags & MAPDEV_SETATTR) == 0)) return ((void *)(ppim->va + offset)); } /* * If the specified range of physical addresses fits within * the direct map window, use the direct map. */ if (pa < dmaplimit && pa + size <= dmaplimit) { va = PHYS_TO_DMAP(pa); if ((flags & MAPDEV_SETATTR) != 0) { PMAP_LOCK(kernel_pmap); i = pmap_change_props_locked(va, size, PROT_NONE, mode, flags); PMAP_UNLOCK(kernel_pmap); } else i = 0; if (!i) return ((void *)(va + offset)); } va = kva_alloc(size); if (va == 0) panic("%s: Couldn't allocate KVA", __func__); } for (tmpsize = 0; tmpsize < size; tmpsize += PAGE_SIZE) pmap_kenter_attr(va + tmpsize, pa + tmpsize, mode); pmap_invalidate_range(kernel_pmap, va, va + tmpsize); if ((flags & MAPDEV_FLUSHCACHE) != 0) pmap_invalidate_cache_range(va, va + tmpsize); return ((void *)(va + offset)); } void * pmap_mapdev_attr(vm_paddr_t pa, vm_size_t size, int mode) { return (pmap_mapdev_internal(pa, size, mode, MAPDEV_FLUSHCACHE | MAPDEV_SETATTR)); } void * pmap_mapdev(vm_paddr_t pa, vm_size_t size) { return (pmap_mapdev_attr(pa, size, PAT_UNCACHEABLE)); } void * pmap_mapdev_pciecfg(vm_paddr_t pa, vm_size_t size) { return (pmap_mapdev_internal(pa, size, PAT_UNCACHEABLE, MAPDEV_SETATTR)); } void * pmap_mapbios(vm_paddr_t pa, vm_size_t size) { return (pmap_mapdev_internal(pa, size, PAT_WRITE_BACK, MAPDEV_FLUSHCACHE)); } void pmap_unmapdev(vm_offset_t va, vm_size_t size) { struct pmap_preinit_mapping *ppim; vm_offset_t offset; int i; /* If we gave a direct map region in pmap_mapdev, do nothing */ if (va >= DMAP_MIN_ADDRESS && va < DMAP_MAX_ADDRESS) return; offset = va & PAGE_MASK; size = round_page(offset + size); va = trunc_page(va); for (i = 0; i < PMAP_PREINIT_MAPPING_COUNT; i++) { ppim = pmap_preinit_mapping + i; if (ppim->va == va && ppim->sz == size) { if (pmap_initialized) return; ppim->pa = 0; ppim->va = 0; ppim->sz = 0; ppim->mode = 0; if (va + size == virtual_avail) virtual_avail = va; return; } } if (pmap_initialized) { pmap_qremove(va, atop(size)); kva_free(va, size); } } /* * Tries to demote a 1GB page mapping. */ static boolean_t pmap_demote_pdpe(pmap_t pmap, pdp_entry_t *pdpe, vm_offset_t va) { pdp_entry_t newpdpe, oldpdpe; pd_entry_t *firstpde, newpde, *pde; pt_entry_t PG_A, PG_M, PG_RW, PG_V; vm_paddr_t pdpgpa; vm_page_t pdpg; PG_A = pmap_accessed_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_V = pmap_valid_bit(pmap); PG_RW = pmap_rw_bit(pmap); PMAP_LOCK_ASSERT(pmap, MA_OWNED); oldpdpe = *pdpe; KASSERT((oldpdpe & (PG_PS | PG_V)) == (PG_PS | PG_V), ("pmap_demote_pdpe: oldpdpe is missing PG_PS and/or PG_V")); pdpg = pmap_alloc_pt_page(pmap, va >> PDPSHIFT, VM_ALLOC_WIRED | VM_ALLOC_INTERRUPT); if (pdpg == NULL) { CTR2(KTR_PMAP, "pmap_demote_pdpe: failure for va %#lx" " in pmap %p", va, pmap); return (FALSE); } pdpgpa = VM_PAGE_TO_PHYS(pdpg); firstpde = (pd_entry_t *)PHYS_TO_DMAP(pdpgpa); newpdpe = pdpgpa | PG_M | PG_A | (oldpdpe & PG_U) | PG_RW | PG_V; KASSERT((oldpdpe & PG_A) != 0, ("pmap_demote_pdpe: oldpdpe is missing PG_A")); KASSERT((oldpdpe & (PG_M | PG_RW)) != PG_RW, ("pmap_demote_pdpe: oldpdpe is missing PG_M")); newpde = oldpdpe; /* * Initialize the page directory page. */ for (pde = firstpde; pde < firstpde + NPDEPG; pde++) { *pde = newpde; newpde += NBPDR; } /* * Demote the mapping. */ *pdpe = newpdpe; /* * Invalidate a stale recursive mapping of the page directory page. */ pmap_invalidate_page(pmap, (vm_offset_t)vtopde(va)); counter_u64_add(pmap_pdpe_demotions, 1); CTR2(KTR_PMAP, "pmap_demote_pdpe: success for va %#lx" " in pmap %p", va, pmap); return (TRUE); } /* * Sets the memory attribute for the specified page. */ void pmap_page_set_memattr(vm_page_t m, vm_memattr_t ma) { m->md.pat_mode = ma; /* * If "m" is a normal page, update its direct mapping. This update * can be relied upon to perform any cache operations that are * required for data coherence. */ if ((m->flags & PG_FICTITIOUS) == 0 && pmap_change_attr(PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)), PAGE_SIZE, m->md.pat_mode)) panic("memory attribute change on the direct map failed"); } void pmap_page_set_memattr_noflush(vm_page_t m, vm_memattr_t ma) { int error; m->md.pat_mode = ma; if ((m->flags & PG_FICTITIOUS) != 0) return; PMAP_LOCK(kernel_pmap); error = pmap_change_props_locked(PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)), PAGE_SIZE, PROT_NONE, m->md.pat_mode, 0); PMAP_UNLOCK(kernel_pmap); if (error != 0) panic("memory attribute change on the direct map failed"); } /* * Changes the specified virtual address range's memory type to that given by * the parameter "mode". The specified virtual address range must be * completely contained within either the direct map or the kernel map. If * the virtual address range is contained within the kernel map, then the * memory type for each of the corresponding ranges of the direct map is also * changed. (The corresponding ranges of the direct map are those ranges that * map the same physical pages as the specified virtual address range.) These * changes to the direct map are necessary because Intel describes the * behavior of their processors as "undefined" if two or more mappings to the * same physical page have different memory types. * * Returns zero if the change completed successfully, and either EINVAL or * ENOMEM if the change failed. Specifically, EINVAL is returned if some part * of the virtual address range was not mapped, and ENOMEM is returned if * there was insufficient memory available to complete the change. In the * latter case, the memory type may have been changed on some part of the * virtual address range or the direct map. */ int pmap_change_attr(vm_offset_t va, vm_size_t size, int mode) { int error; PMAP_LOCK(kernel_pmap); error = pmap_change_props_locked(va, size, PROT_NONE, mode, MAPDEV_FLUSHCACHE); PMAP_UNLOCK(kernel_pmap); return (error); } /* * Changes the specified virtual address range's protections to those * specified by "prot". Like pmap_change_attr(), protections for aliases * in the direct map are updated as well. Protections on aliasing mappings may * be a subset of the requested protections; for example, mappings in the direct * map are never executable. */ int pmap_change_prot(vm_offset_t va, vm_size_t size, vm_prot_t prot) { int error; /* Only supported within the kernel map. */ if (va < VM_MIN_KERNEL_ADDRESS) return (EINVAL); PMAP_LOCK(kernel_pmap); error = pmap_change_props_locked(va, size, prot, -1, MAPDEV_ASSERTVALID); PMAP_UNLOCK(kernel_pmap); return (error); } static int pmap_change_props_locked(vm_offset_t va, vm_size_t size, vm_prot_t prot, int mode, int flags) { vm_offset_t base, offset, tmpva; vm_paddr_t pa_start, pa_end, pa_end1; pdp_entry_t *pdpe; pd_entry_t *pde, pde_bits, pde_mask; pt_entry_t *pte, pte_bits, pte_mask; int error; bool changed; PMAP_LOCK_ASSERT(kernel_pmap, MA_OWNED); base = trunc_page(va); offset = va & PAGE_MASK; size = round_page(offset + size); /* * Only supported on kernel virtual addresses, including the direct * map but excluding the recursive map. */ if (base < DMAP_MIN_ADDRESS) return (EINVAL); /* * Construct our flag sets and masks. "bits" is the subset of * "mask" that will be set in each modified PTE. * * Mappings in the direct map are never allowed to be executable. */ pde_bits = pte_bits = 0; pde_mask = pte_mask = 0; if (mode != -1) { pde_bits |= pmap_cache_bits(kernel_pmap, mode, true); pde_mask |= X86_PG_PDE_CACHE; pte_bits |= pmap_cache_bits(kernel_pmap, mode, false); pte_mask |= X86_PG_PTE_CACHE; } if (prot != VM_PROT_NONE) { if ((prot & VM_PROT_WRITE) != 0) { pde_bits |= X86_PG_RW; pte_bits |= X86_PG_RW; } if ((prot & VM_PROT_EXECUTE) == 0 || va < VM_MIN_KERNEL_ADDRESS) { pde_bits |= pg_nx; pte_bits |= pg_nx; } pde_mask |= X86_PG_RW | pg_nx; pte_mask |= X86_PG_RW | pg_nx; } /* * Pages that aren't mapped aren't supported. Also break down 2MB pages * into 4KB pages if required. */ for (tmpva = base; tmpva < base + size; ) { pdpe = pmap_pdpe(kernel_pmap, tmpva); if (pdpe == NULL || *pdpe == 0) { KASSERT((flags & MAPDEV_ASSERTVALID) == 0, ("%s: addr %#lx is not mapped", __func__, tmpva)); return (EINVAL); } if (*pdpe & PG_PS) { /* * If the current 1GB page already has the required * properties, then we need not demote this page. Just * increment tmpva to the next 1GB page frame. */ if ((*pdpe & pde_mask) == pde_bits) { tmpva = trunc_1gpage(tmpva) + NBPDP; continue; } /* * If the current offset aligns with a 1GB page frame * and there is at least 1GB left within the range, then * we need not break down this page into 2MB pages. */ if ((tmpva & PDPMASK) == 0 && tmpva + PDPMASK < base + size) { tmpva += NBPDP; continue; } if (!pmap_demote_pdpe(kernel_pmap, pdpe, tmpva)) return (ENOMEM); } pde = pmap_pdpe_to_pde(pdpe, tmpva); if (*pde == 0) { KASSERT((flags & MAPDEV_ASSERTVALID) == 0, ("%s: addr %#lx is not mapped", __func__, tmpva)); return (EINVAL); } if (*pde & PG_PS) { /* * If the current 2MB page already has the required * properties, then we need not demote this page. Just * increment tmpva to the next 2MB page frame. */ if ((*pde & pde_mask) == pde_bits) { tmpva = trunc_2mpage(tmpva) + NBPDR; continue; } /* * If the current offset aligns with a 2MB page frame * and there is at least 2MB left within the range, then * we need not break down this page into 4KB pages. */ if ((tmpva & PDRMASK) == 0 && tmpva + PDRMASK < base + size) { tmpva += NBPDR; continue; } if (!pmap_demote_pde(kernel_pmap, pde, tmpva)) return (ENOMEM); } pte = pmap_pde_to_pte(pde, tmpva); if (*pte == 0) { KASSERT((flags & MAPDEV_ASSERTVALID) == 0, ("%s: addr %#lx is not mapped", __func__, tmpva)); return (EINVAL); } tmpva += PAGE_SIZE; } error = 0; /* * Ok, all the pages exist, so run through them updating their * properties if required. */ changed = false; pa_start = pa_end = 0; for (tmpva = base; tmpva < base + size; ) { pdpe = pmap_pdpe(kernel_pmap, tmpva); if (*pdpe & PG_PS) { if ((*pdpe & pde_mask) != pde_bits) { pmap_pte_props(pdpe, pde_bits, pde_mask); changed = true; } if (tmpva >= VM_MIN_KERNEL_ADDRESS && (*pdpe & PG_PS_FRAME) < dmaplimit) { if (pa_start == pa_end) { /* Start physical address run. */ pa_start = *pdpe & PG_PS_FRAME; pa_end = pa_start + NBPDP; } else if (pa_end == (*pdpe & PG_PS_FRAME)) pa_end += NBPDP; else { /* Run ended, update direct map. */ error = pmap_change_props_locked( PHYS_TO_DMAP(pa_start), pa_end - pa_start, prot, mode, flags); if (error != 0) break; /* Start physical address run. */ pa_start = *pdpe & PG_PS_FRAME; pa_end = pa_start + NBPDP; } } tmpva = trunc_1gpage(tmpva) + NBPDP; continue; } pde = pmap_pdpe_to_pde(pdpe, tmpva); if (*pde & PG_PS) { if ((*pde & pde_mask) != pde_bits) { pmap_pte_props(pde, pde_bits, pde_mask); changed = true; } if (tmpva >= VM_MIN_KERNEL_ADDRESS && (*pde & PG_PS_FRAME) < dmaplimit) { if (pa_start == pa_end) { /* Start physical address run. */ pa_start = *pde & PG_PS_FRAME; pa_end = pa_start + NBPDR; } else if (pa_end == (*pde & PG_PS_FRAME)) pa_end += NBPDR; else { /* Run ended, update direct map. */ error = pmap_change_props_locked( PHYS_TO_DMAP(pa_start), pa_end - pa_start, prot, mode, flags); if (error != 0) break; /* Start physical address run. */ pa_start = *pde & PG_PS_FRAME; pa_end = pa_start + NBPDR; } } tmpva = trunc_2mpage(tmpva) + NBPDR; } else { pte = pmap_pde_to_pte(pde, tmpva); if ((*pte & pte_mask) != pte_bits) { pmap_pte_props(pte, pte_bits, pte_mask); changed = true; } if (tmpva >= VM_MIN_KERNEL_ADDRESS && (*pte & PG_FRAME) < dmaplimit) { if (pa_start == pa_end) { /* Start physical address run. */ pa_start = *pte & PG_FRAME; pa_end = pa_start + PAGE_SIZE; } else if (pa_end == (*pte & PG_FRAME)) pa_end += PAGE_SIZE; else { /* Run ended, update direct map. */ error = pmap_change_props_locked( PHYS_TO_DMAP(pa_start), pa_end - pa_start, prot, mode, flags); if (error != 0) break; /* Start physical address run. */ pa_start = *pte & PG_FRAME; pa_end = pa_start + PAGE_SIZE; } } tmpva += PAGE_SIZE; } } if (error == 0 && pa_start != pa_end && pa_start < dmaplimit) { pa_end1 = MIN(pa_end, dmaplimit); if (pa_start != pa_end1) error = pmap_change_props_locked(PHYS_TO_DMAP(pa_start), pa_end1 - pa_start, prot, mode, flags); } /* * Flush CPU caches if required to make sure any data isn't cached that * shouldn't be, etc. */ if (changed) { pmap_invalidate_range(kernel_pmap, base, tmpva); if ((flags & MAPDEV_FLUSHCACHE) != 0) pmap_invalidate_cache_range(base, tmpva); } return (error); } /* * Demotes any mapping within the direct map region that covers more than the * specified range of physical addresses. This range's size must be a power * of two and its starting address must be a multiple of its size. Since the * demotion does not change any attributes of the mapping, a TLB invalidation * is not mandatory. The caller may, however, request a TLB invalidation. */ void pmap_demote_DMAP(vm_paddr_t base, vm_size_t len, boolean_t invalidate) { pdp_entry_t *pdpe; pd_entry_t *pde; vm_offset_t va; boolean_t changed; if (len == 0) return; KASSERT(powerof2(len), ("pmap_demote_DMAP: len is not a power of 2")); KASSERT((base & (len - 1)) == 0, ("pmap_demote_DMAP: base is not a multiple of len")); if (len < NBPDP && base < dmaplimit) { va = PHYS_TO_DMAP(base); changed = FALSE; PMAP_LOCK(kernel_pmap); pdpe = pmap_pdpe(kernel_pmap, va); if ((*pdpe & X86_PG_V) == 0) panic("pmap_demote_DMAP: invalid PDPE"); if ((*pdpe & PG_PS) != 0) { if (!pmap_demote_pdpe(kernel_pmap, pdpe, va)) panic("pmap_demote_DMAP: PDPE failed"); changed = TRUE; } if (len < NBPDR) { pde = pmap_pdpe_to_pde(pdpe, va); if ((*pde & X86_PG_V) == 0) panic("pmap_demote_DMAP: invalid PDE"); if ((*pde & PG_PS) != 0) { if (!pmap_demote_pde(kernel_pmap, pde, va)) panic("pmap_demote_DMAP: PDE failed"); changed = TRUE; } } if (changed && invalidate) pmap_invalidate_page(kernel_pmap, va); PMAP_UNLOCK(kernel_pmap); } } /* * Perform the pmap work for mincore(2). If the page is not both referenced and * modified by this pmap, returns its physical address so that the caller can * find other mappings. */ int pmap_mincore(pmap_t pmap, vm_offset_t addr, vm_paddr_t *pap) { pdp_entry_t *pdpe; pd_entry_t *pdep; pt_entry_t pte, PG_A, PG_M, PG_RW, PG_V; vm_paddr_t pa; int val; PG_A = pmap_accessed_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_V = pmap_valid_bit(pmap); PG_RW = pmap_rw_bit(pmap); PMAP_LOCK(pmap); pte = 0; pa = 0; val = 0; pdpe = pmap_pdpe(pmap, addr); if (pdpe == NULL) goto out; if ((*pdpe & PG_V) != 0) { if ((*pdpe & PG_PS) != 0) { pte = *pdpe; pa = ((pte & PG_PS_PDP_FRAME) | (addr & PDPMASK)) & PG_FRAME; val = MINCORE_PSIND(2); } else { pdep = pmap_pde(pmap, addr); if (pdep != NULL && (*pdep & PG_V) != 0) { if ((*pdep & PG_PS) != 0) { pte = *pdep; /* Compute the physical address of the 4KB page. */ pa = ((pte & PG_PS_FRAME) | (addr & PDRMASK)) & PG_FRAME; val = MINCORE_PSIND(1); } else { pte = *pmap_pde_to_pte(pdep, addr); pa = pte & PG_FRAME; val = 0; } } } } if ((pte & PG_V) != 0) { val |= MINCORE_INCORE; if ((pte & (PG_M | PG_RW)) == (PG_M | PG_RW)) val |= MINCORE_MODIFIED | MINCORE_MODIFIED_OTHER; if ((pte & PG_A) != 0) val |= MINCORE_REFERENCED | MINCORE_REFERENCED_OTHER; } if ((val & (MINCORE_MODIFIED_OTHER | MINCORE_REFERENCED_OTHER)) != (MINCORE_MODIFIED_OTHER | MINCORE_REFERENCED_OTHER) && (pte & (PG_MANAGED | PG_V)) == (PG_MANAGED | PG_V)) { *pap = pa; } out: PMAP_UNLOCK(pmap); return (val); } static uint64_t pmap_pcid_alloc(pmap_t pmap, u_int cpuid) { uint32_t gen, new_gen, pcid_next; CRITICAL_ASSERT(curthread); gen = PCPU_GET(pcid_gen); if (pmap->pm_pcids[cpuid].pm_pcid == PMAP_PCID_KERN) return (pti ? 0 : CR3_PCID_SAVE); if (pmap->pm_pcids[cpuid].pm_gen == gen) return (CR3_PCID_SAVE); pcid_next = PCPU_GET(pcid_next); KASSERT((!pti && pcid_next <= PMAP_PCID_OVERMAX) || (pti && pcid_next <= PMAP_PCID_OVERMAX_KERN), ("cpu %d pcid_next %#x", cpuid, pcid_next)); if ((!pti && pcid_next == PMAP_PCID_OVERMAX) || (pti && pcid_next == PMAP_PCID_OVERMAX_KERN)) { new_gen = gen + 1; if (new_gen == 0) new_gen = 1; PCPU_SET(pcid_gen, new_gen); pcid_next = PMAP_PCID_KERN + 1; } else { new_gen = gen; } pmap->pm_pcids[cpuid].pm_pcid = pcid_next; pmap->pm_pcids[cpuid].pm_gen = new_gen; PCPU_SET(pcid_next, pcid_next + 1); return (0); } static uint64_t pmap_pcid_alloc_checked(pmap_t pmap, u_int cpuid) { uint64_t cached; cached = pmap_pcid_alloc(pmap, cpuid); KASSERT(pmap->pm_pcids[cpuid].pm_pcid < PMAP_PCID_OVERMAX, ("pmap %p cpu %d pcid %#x", pmap, cpuid, pmap->pm_pcids[cpuid].pm_pcid)); KASSERT(pmap->pm_pcids[cpuid].pm_pcid != PMAP_PCID_KERN || pmap == kernel_pmap, ("non-kernel pmap pmap %p cpu %d pcid %#x", pmap, cpuid, pmap->pm_pcids[cpuid].pm_pcid)); return (cached); } static void pmap_activate_sw_pti_post(struct thread *td, pmap_t pmap) { PCPU_GET(tssp)->tss_rsp0 = pmap->pm_ucr3 != PMAP_NO_CR3 ? PCPU_GET(pti_rsp0) : (uintptr_t)td->td_md.md_stack_base; } static void pmap_activate_sw_pcid_pti(struct thread *td, pmap_t pmap, u_int cpuid) { pmap_t old_pmap; uint64_t cached, cr3, kcr3, ucr3; KASSERT((read_rflags() & PSL_I) == 0, ("PCID needs interrupts disabled in pmap_activate_sw()")); /* See the comment in pmap_invalidate_page_pcid(). */ if (PCPU_GET(ucr3_load_mask) != PMAP_UCR3_NOMASK) { PCPU_SET(ucr3_load_mask, PMAP_UCR3_NOMASK); old_pmap = PCPU_GET(curpmap); MPASS(old_pmap->pm_ucr3 != PMAP_NO_CR3); old_pmap->pm_pcids[cpuid].pm_gen = 0; } cached = pmap_pcid_alloc_checked(pmap, cpuid); cr3 = rcr3(); if ((cr3 & ~CR3_PCID_MASK) != pmap->pm_cr3) load_cr3(pmap->pm_cr3 | pmap->pm_pcids[cpuid].pm_pcid); PCPU_SET(curpmap, pmap); kcr3 = pmap->pm_cr3 | pmap->pm_pcids[cpuid].pm_pcid; ucr3 = pmap->pm_ucr3 | pmap->pm_pcids[cpuid].pm_pcid | PMAP_PCID_USER_PT; if (!cached && pmap->pm_ucr3 != PMAP_NO_CR3) PCPU_SET(ucr3_load_mask, ~CR3_PCID_SAVE); PCPU_SET(kcr3, kcr3 | CR3_PCID_SAVE); PCPU_SET(ucr3, ucr3 | CR3_PCID_SAVE); if (cached) counter_u64_add(pcid_save_cnt, 1); pmap_activate_sw_pti_post(td, pmap); } static void pmap_activate_sw_pcid_nopti(struct thread *td __unused, pmap_t pmap, u_int cpuid) { uint64_t cached, cr3; KASSERT((read_rflags() & PSL_I) == 0, ("PCID needs interrupts disabled in pmap_activate_sw()")); cached = pmap_pcid_alloc_checked(pmap, cpuid); cr3 = rcr3(); if (!cached || (cr3 & ~CR3_PCID_MASK) != pmap->pm_cr3) load_cr3(pmap->pm_cr3 | pmap->pm_pcids[cpuid].pm_pcid | cached); PCPU_SET(curpmap, pmap); if (cached) counter_u64_add(pcid_save_cnt, 1); } static void pmap_activate_sw_nopcid_nopti(struct thread *td __unused, pmap_t pmap, u_int cpuid __unused) { load_cr3(pmap->pm_cr3); PCPU_SET(curpmap, pmap); } static void pmap_activate_sw_nopcid_pti(struct thread *td, pmap_t pmap, u_int cpuid __unused) { pmap_activate_sw_nopcid_nopti(td, pmap, cpuid); PCPU_SET(kcr3, pmap->pm_cr3); PCPU_SET(ucr3, pmap->pm_ucr3); pmap_activate_sw_pti_post(td, pmap); } DEFINE_IFUNC(static, void, pmap_activate_sw_mode, (struct thread *, pmap_t, u_int)) { if (pmap_pcid_enabled && pti) return (pmap_activate_sw_pcid_pti); else if (pmap_pcid_enabled && !pti) return (pmap_activate_sw_pcid_nopti); else if (!pmap_pcid_enabled && pti) return (pmap_activate_sw_nopcid_pti); else /* if (!pmap_pcid_enabled && !pti) */ return (pmap_activate_sw_nopcid_nopti); } void pmap_activate_sw(struct thread *td) { pmap_t oldpmap, pmap; u_int cpuid; oldpmap = PCPU_GET(curpmap); pmap = vmspace_pmap(td->td_proc->p_vmspace); if (oldpmap == pmap) { if (cpu_vendor_id != CPU_VENDOR_INTEL) mfence(); return; } cpuid = PCPU_GET(cpuid); #ifdef SMP CPU_SET_ATOMIC(cpuid, &pmap->pm_active); #else CPU_SET(cpuid, &pmap->pm_active); #endif pmap_activate_sw_mode(td, pmap, cpuid); #ifdef SMP CPU_CLR_ATOMIC(cpuid, &oldpmap->pm_active); #else CPU_CLR(cpuid, &oldpmap->pm_active); #endif } void pmap_activate(struct thread *td) { /* * invltlb_{invpcid,}_pcid_handler() is used to handle an * invalidate_all IPI, which checks for curpmap == * smp_tlb_pmap. The below sequence of operations has a * window where %CR3 is loaded with the new pmap's PML4 * address, but the curpmap value has not yet been updated. * This causes the invltlb IPI handler, which is called * between the updates, to execute as a NOP, which leaves * stale TLB entries. * * Note that the most common use of pmap_activate_sw(), from * a context switch, is immune to this race, because * interrupts are disabled (while the thread lock is owned), * so the IPI is delayed until after curpmap is updated. Protect * other callers in a similar way, by disabling interrupts * around the %cr3 register reload and curpmap assignment. */ spinlock_enter(); pmap_activate_sw(td); spinlock_exit(); } void pmap_activate_boot(pmap_t pmap) { uint64_t kcr3; u_int cpuid; /* * kernel_pmap must be never deactivated, and we ensure that * by never activating it at all. */ MPASS(pmap != kernel_pmap); cpuid = PCPU_GET(cpuid); #ifdef SMP CPU_SET_ATOMIC(cpuid, &pmap->pm_active); #else CPU_SET(cpuid, &pmap->pm_active); #endif PCPU_SET(curpmap, pmap); if (pti) { kcr3 = pmap->pm_cr3; if (pmap_pcid_enabled) kcr3 |= pmap->pm_pcids[cpuid].pm_pcid | CR3_PCID_SAVE; } else { kcr3 = PMAP_NO_CR3; } PCPU_SET(kcr3, kcr3); PCPU_SET(ucr3, PMAP_NO_CR3); } void pmap_sync_icache(pmap_t pm, vm_offset_t va, vm_size_t sz) { } /* * Increase the starting virtual address of the given mapping if a * different alignment might result in more superpage mappings. */ void pmap_align_superpage(vm_object_t object, vm_ooffset_t offset, vm_offset_t *addr, vm_size_t size) { vm_offset_t superpage_offset; if (size < NBPDR) return; if (object != NULL && (object->flags & OBJ_COLORED) != 0) offset += ptoa(object->pg_color); superpage_offset = offset & PDRMASK; if (size - ((NBPDR - superpage_offset) & PDRMASK) < NBPDR || (*addr & PDRMASK) == superpage_offset) return; if ((*addr & PDRMASK) < superpage_offset) *addr = (*addr & ~PDRMASK) + superpage_offset; else *addr = ((*addr + PDRMASK) & ~PDRMASK) + superpage_offset; } #ifdef INVARIANTS static unsigned long num_dirty_emulations; SYSCTL_ULONG(_vm_pmap, OID_AUTO, num_dirty_emulations, CTLFLAG_RW, &num_dirty_emulations, 0, NULL); static unsigned long num_accessed_emulations; SYSCTL_ULONG(_vm_pmap, OID_AUTO, num_accessed_emulations, CTLFLAG_RW, &num_accessed_emulations, 0, NULL); static unsigned long num_superpage_accessed_emulations; SYSCTL_ULONG(_vm_pmap, OID_AUTO, num_superpage_accessed_emulations, CTLFLAG_RW, &num_superpage_accessed_emulations, 0, NULL); static unsigned long ad_emulation_superpage_promotions; SYSCTL_ULONG(_vm_pmap, OID_AUTO, ad_emulation_superpage_promotions, CTLFLAG_RW, &ad_emulation_superpage_promotions, 0, NULL); #endif /* INVARIANTS */ int pmap_emulate_accessed_dirty(pmap_t pmap, vm_offset_t va, int ftype) { int rv; struct rwlock *lock; #if VM_NRESERVLEVEL > 0 vm_page_t m, mpte; #endif pd_entry_t *pde; pt_entry_t *pte, PG_A, PG_M, PG_RW, PG_V; KASSERT(ftype == VM_PROT_READ || ftype == VM_PROT_WRITE, ("pmap_emulate_accessed_dirty: invalid fault type %d", ftype)); if (!pmap_emulate_ad_bits(pmap)) return (-1); PG_A = pmap_accessed_bit(pmap); PG_M = pmap_modified_bit(pmap); PG_V = pmap_valid_bit(pmap); PG_RW = pmap_rw_bit(pmap); rv = -1; lock = NULL; PMAP_LOCK(pmap); pde = pmap_pde(pmap, va); if (pde == NULL || (*pde & PG_V) == 0) goto done; if ((*pde & PG_PS) != 0) { if (ftype == VM_PROT_READ) { #ifdef INVARIANTS atomic_add_long(&num_superpage_accessed_emulations, 1); #endif *pde |= PG_A; rv = 0; } goto done; } pte = pmap_pde_to_pte(pde, va); if ((*pte & PG_V) == 0) goto done; if (ftype == VM_PROT_WRITE) { if ((*pte & PG_RW) == 0) goto done; /* * Set the modified and accessed bits simultaneously. * * Intel EPT PTEs that do software emulation of A/D bits map * PG_A and PG_M to EPT_PG_READ and EPT_PG_WRITE respectively. * An EPT misconfiguration is triggered if the PTE is writable * but not readable (WR=10). This is avoided by setting PG_A * and PG_M simultaneously. */ *pte |= PG_M | PG_A; } else { *pte |= PG_A; } #if VM_NRESERVLEVEL > 0 /* try to promote the mapping */ if (va < VM_MAXUSER_ADDRESS) mpte = PHYS_TO_VM_PAGE(*pde & PG_FRAME); else mpte = NULL; m = PHYS_TO_VM_PAGE(*pte & PG_FRAME); if ((mpte == NULL || mpte->ref_count == NPTEPG) && pmap_ps_enabled(pmap) && (m->flags & PG_FICTITIOUS) == 0 && vm_reserv_level_iffullpop(m) == 0) { pmap_promote_pde(pmap, pde, va, &lock); #ifdef INVARIANTS atomic_add_long(&ad_emulation_superpage_promotions, 1); #endif } #endif #ifdef INVARIANTS if (ftype == VM_PROT_WRITE) atomic_add_long(&num_dirty_emulations, 1); else atomic_add_long(&num_accessed_emulations, 1); #endif rv = 0; /* success */ done: if (lock != NULL) rw_wunlock(lock); PMAP_UNLOCK(pmap); return (rv); } void pmap_get_mapping(pmap_t pmap, vm_offset_t va, uint64_t *ptr, int *num) { pml4_entry_t *pml4; pdp_entry_t *pdp; pd_entry_t *pde; pt_entry_t *pte, PG_V; int idx; idx = 0; PG_V = pmap_valid_bit(pmap); PMAP_LOCK(pmap); pml4 = pmap_pml4e(pmap, va); if (pml4 == NULL) goto done; ptr[idx++] = *pml4; if ((*pml4 & PG_V) == 0) goto done; pdp = pmap_pml4e_to_pdpe(pml4, va); ptr[idx++] = *pdp; if ((*pdp & PG_V) == 0 || (*pdp & PG_PS) != 0) goto done; pde = pmap_pdpe_to_pde(pdp, va); ptr[idx++] = *pde; if ((*pde & PG_V) == 0 || (*pde & PG_PS) != 0) goto done; pte = pmap_pde_to_pte(pde, va); ptr[idx++] = *pte; done: PMAP_UNLOCK(pmap); *num = idx; } /** * Get the kernel virtual address of a set of physical pages. If there are * physical addresses not covered by the DMAP perform a transient mapping * that will be removed when calling pmap_unmap_io_transient. * * \param page The pages the caller wishes to obtain the virtual * address on the kernel memory map. * \param vaddr On return contains the kernel virtual memory address * of the pages passed in the page parameter. * \param count Number of pages passed in. * \param can_fault TRUE if the thread using the mapped pages can take * page faults, FALSE otherwise. * * \returns TRUE if the caller must call pmap_unmap_io_transient when * finished or FALSE otherwise. * */ boolean_t pmap_map_io_transient(vm_page_t page[], vm_offset_t vaddr[], int count, boolean_t can_fault) { vm_paddr_t paddr; boolean_t needs_mapping; pt_entry_t *pte; int cache_bits, error __unused, i; /* * Allocate any KVA space that we need, this is done in a separate * loop to prevent calling vmem_alloc while pinned. */ needs_mapping = FALSE; for (i = 0; i < count; i++) { paddr = VM_PAGE_TO_PHYS(page[i]); if (__predict_false(paddr >= dmaplimit)) { error = vmem_alloc(kernel_arena, PAGE_SIZE, M_BESTFIT | M_WAITOK, &vaddr[i]); KASSERT(error == 0, ("vmem_alloc failed: %d", error)); needs_mapping = TRUE; } else { vaddr[i] = PHYS_TO_DMAP(paddr); } } /* Exit early if everything is covered by the DMAP */ if (!needs_mapping) return (FALSE); /* * NB: The sequence of updating a page table followed by accesses * to the corresponding pages used in the !DMAP case is subject to * the situation described in the "AMD64 Architecture Programmer's * Manual Volume 2: System Programming" rev. 3.23, "7.3.1 Special * Coherency Considerations". Therefore, issuing the INVLPG right * after modifying the PTE bits is crucial. */ if (!can_fault) sched_pin(); for (i = 0; i < count; i++) { paddr = VM_PAGE_TO_PHYS(page[i]); if (paddr >= dmaplimit) { if (can_fault) { /* * Slow path, since we can get page faults * while mappings are active don't pin the * thread to the CPU and instead add a global * mapping visible to all CPUs. */ pmap_qenter(vaddr[i], &page[i], 1); } else { pte = vtopte(vaddr[i]); cache_bits = pmap_cache_bits(kernel_pmap, page[i]->md.pat_mode, 0); pte_store(pte, paddr | X86_PG_RW | X86_PG_V | cache_bits); invlpg(vaddr[i]); } } } return (needs_mapping); } void pmap_unmap_io_transient(vm_page_t page[], vm_offset_t vaddr[], int count, boolean_t can_fault) { vm_paddr_t paddr; int i; if (!can_fault) sched_unpin(); for (i = 0; i < count; i++) { paddr = VM_PAGE_TO_PHYS(page[i]); if (paddr >= dmaplimit) { if (can_fault) pmap_qremove(vaddr[i], 1); vmem_free(kernel_arena, vaddr[i], PAGE_SIZE); } } } vm_offset_t pmap_quick_enter_page(vm_page_t m) { vm_paddr_t paddr; paddr = VM_PAGE_TO_PHYS(m); if (paddr < dmaplimit) return (PHYS_TO_DMAP(paddr)); mtx_lock_spin(&qframe_mtx); KASSERT(*vtopte(qframe) == 0, ("qframe busy")); pte_store(vtopte(qframe), paddr | X86_PG_RW | X86_PG_V | X86_PG_A | X86_PG_M | pmap_cache_bits(kernel_pmap, m->md.pat_mode, 0)); return (qframe); } void pmap_quick_remove_page(vm_offset_t addr) { if (addr != qframe) return; pte_store(vtopte(qframe), 0); invlpg(qframe); mtx_unlock_spin(&qframe_mtx); } /* * Pdp pages from the large map are managed differently from either * kernel or user page table pages. They are permanently allocated at * initialization time, and their reference count is permanently set to * zero. The pml4 entries pointing to those pages are copied into * each allocated pmap. * * In contrast, pd and pt pages are managed like user page table * pages. They are dynamically allocated, and their reference count * represents the number of valid entries within the page. */ static vm_page_t pmap_large_map_getptp_unlocked(void) { return (pmap_alloc_pt_page(kernel_pmap, 0, VM_ALLOC_ZERO)); } static vm_page_t pmap_large_map_getptp(void) { vm_page_t m; PMAP_LOCK_ASSERT(kernel_pmap, MA_OWNED); m = pmap_large_map_getptp_unlocked(); if (m == NULL) { PMAP_UNLOCK(kernel_pmap); vm_wait(NULL); PMAP_LOCK(kernel_pmap); /* Callers retry. */ } return (m); } static pdp_entry_t * pmap_large_map_pdpe(vm_offset_t va) { vm_pindex_t pml4_idx; vm_paddr_t mphys; pml4_idx = pmap_pml4e_index(va); KASSERT(LMSPML4I <= pml4_idx && pml4_idx < LMSPML4I + lm_ents, ("pmap_large_map_pdpe: va %#jx out of range idx %#jx LMSPML4I " "%#jx lm_ents %d", (uintmax_t)va, (uintmax_t)pml4_idx, LMSPML4I, lm_ents)); KASSERT((kernel_pml4[pml4_idx] & X86_PG_V) != 0, ("pmap_large_map_pdpe: invalid pml4 for va %#jx idx %#jx " "LMSPML4I %#jx lm_ents %d", (uintmax_t)va, (uintmax_t)pml4_idx, LMSPML4I, lm_ents)); mphys = kernel_pml4[pml4_idx] & PG_FRAME; return ((pdp_entry_t *)PHYS_TO_DMAP(mphys) + pmap_pdpe_index(va)); } static pd_entry_t * pmap_large_map_pde(vm_offset_t va) { pdp_entry_t *pdpe; vm_page_t m; vm_paddr_t mphys; retry: pdpe = pmap_large_map_pdpe(va); if (*pdpe == 0) { m = pmap_large_map_getptp(); if (m == NULL) goto retry; mphys = VM_PAGE_TO_PHYS(m); *pdpe = mphys | X86_PG_A | X86_PG_RW | X86_PG_V | pg_nx; } else { MPASS((*pdpe & X86_PG_PS) == 0); mphys = *pdpe & PG_FRAME; } return ((pd_entry_t *)PHYS_TO_DMAP(mphys) + pmap_pde_index(va)); } static pt_entry_t * pmap_large_map_pte(vm_offset_t va) { pd_entry_t *pde; vm_page_t m; vm_paddr_t mphys; retry: pde = pmap_large_map_pde(va); if (*pde == 0) { m = pmap_large_map_getptp(); if (m == NULL) goto retry; mphys = VM_PAGE_TO_PHYS(m); *pde = mphys | X86_PG_A | X86_PG_RW | X86_PG_V | pg_nx; PHYS_TO_VM_PAGE(DMAP_TO_PHYS((uintptr_t)pde))->ref_count++; } else { MPASS((*pde & X86_PG_PS) == 0); mphys = *pde & PG_FRAME; } return ((pt_entry_t *)PHYS_TO_DMAP(mphys) + pmap_pte_index(va)); } static vm_paddr_t pmap_large_map_kextract(vm_offset_t va) { pdp_entry_t *pdpe, pdp; pd_entry_t *pde, pd; pt_entry_t *pte, pt; KASSERT(PMAP_ADDRESS_IN_LARGEMAP(va), ("not largemap range %#lx", (u_long)va)); pdpe = pmap_large_map_pdpe(va); pdp = *pdpe; KASSERT((pdp & X86_PG_V) != 0, ("invalid pdp va %#lx pdpe %#lx pdp %#lx", va, (u_long)pdpe, pdp)); if ((pdp & X86_PG_PS) != 0) { KASSERT((amd_feature & AMDID_PAGE1GB) != 0, ("no 1G pages, va %#lx pdpe %#lx pdp %#lx", va, (u_long)pdpe, pdp)); return ((pdp & PG_PS_PDP_FRAME) | (va & PDPMASK)); } pde = pmap_pdpe_to_pde(pdpe, va); pd = *pde; KASSERT((pd & X86_PG_V) != 0, ("invalid pd va %#lx pde %#lx pd %#lx", va, (u_long)pde, pd)); if ((pd & X86_PG_PS) != 0) return ((pd & PG_PS_FRAME) | (va & PDRMASK)); pte = pmap_pde_to_pte(pde, va); pt = *pte; KASSERT((pt & X86_PG_V) != 0, ("invalid pte va %#lx pte %#lx pt %#lx", va, (u_long)pte, pt)); return ((pt & PG_FRAME) | (va & PAGE_MASK)); } static int pmap_large_map_getva(vm_size_t len, vm_offset_t align, vm_offset_t phase, vmem_addr_t *vmem_res) { /* * Large mappings are all but static. Consequently, there * is no point in waiting for an earlier allocation to be * freed. */ return (vmem_xalloc(large_vmem, len, align, phase, 0, VMEM_ADDR_MIN, VMEM_ADDR_MAX, M_NOWAIT | M_BESTFIT, vmem_res)); } int pmap_large_map(vm_paddr_t spa, vm_size_t len, void **addr, vm_memattr_t mattr) { pdp_entry_t *pdpe; pd_entry_t *pde; pt_entry_t *pte; vm_offset_t va, inc; vmem_addr_t vmem_res; vm_paddr_t pa; int error; if (len == 0 || spa + len < spa) return (EINVAL); /* See if DMAP can serve. */ if (spa + len <= dmaplimit) { va = PHYS_TO_DMAP(spa); *addr = (void *)va; return (pmap_change_attr(va, len, mattr)); } /* * No, allocate KVA. Fit the address with best possible * alignment for superpages. Fall back to worse align if * failed. */ error = ENOMEM; if ((amd_feature & AMDID_PAGE1GB) != 0 && rounddown2(spa + len, NBPDP) >= roundup2(spa, NBPDP) + NBPDP) error = pmap_large_map_getva(len, NBPDP, spa & PDPMASK, &vmem_res); if (error != 0 && rounddown2(spa + len, NBPDR) >= roundup2(spa, NBPDR) + NBPDR) error = pmap_large_map_getva(len, NBPDR, spa & PDRMASK, &vmem_res); if (error != 0) error = pmap_large_map_getva(len, PAGE_SIZE, 0, &vmem_res); if (error != 0) return (error); /* * Fill pagetable. PG_M is not pre-set, we scan modified bits * in the pagetable to minimize flushing. No need to * invalidate TLB, since we only update invalid entries. */ PMAP_LOCK(kernel_pmap); for (pa = spa, va = vmem_res; len > 0; pa += inc, va += inc, len -= inc) { if ((amd_feature & AMDID_PAGE1GB) != 0 && len >= NBPDP && (pa & PDPMASK) == 0 && (va & PDPMASK) == 0) { pdpe = pmap_large_map_pdpe(va); MPASS(*pdpe == 0); *pdpe = pa | pg_g | X86_PG_PS | X86_PG_RW | X86_PG_V | X86_PG_A | pg_nx | pmap_cache_bits(kernel_pmap, mattr, TRUE); inc = NBPDP; } else if (len >= NBPDR && (pa & PDRMASK) == 0 && (va & PDRMASK) == 0) { pde = pmap_large_map_pde(va); MPASS(*pde == 0); *pde = pa | pg_g | X86_PG_PS | X86_PG_RW | X86_PG_V | X86_PG_A | pg_nx | pmap_cache_bits(kernel_pmap, mattr, TRUE); PHYS_TO_VM_PAGE(DMAP_TO_PHYS((uintptr_t)pde))-> ref_count++; inc = NBPDR; } else { pte = pmap_large_map_pte(va); MPASS(*pte == 0); *pte = pa | pg_g | X86_PG_RW | X86_PG_V | X86_PG_A | pg_nx | pmap_cache_bits(kernel_pmap, mattr, FALSE); PHYS_TO_VM_PAGE(DMAP_TO_PHYS((uintptr_t)pte))-> ref_count++; inc = PAGE_SIZE; } } PMAP_UNLOCK(kernel_pmap); MPASS(len == 0); *addr = (void *)vmem_res; return (0); } void pmap_large_unmap(void *svaa, vm_size_t len) { vm_offset_t sva, va; vm_size_t inc; pdp_entry_t *pdpe, pdp; pd_entry_t *pde, pd; pt_entry_t *pte; vm_page_t m; struct spglist spgf; sva = (vm_offset_t)svaa; if (len == 0 || sva + len < sva || (sva >= DMAP_MIN_ADDRESS && sva + len <= DMAP_MIN_ADDRESS + dmaplimit)) return; SLIST_INIT(&spgf); KASSERT(PMAP_ADDRESS_IN_LARGEMAP(sva) && PMAP_ADDRESS_IN_LARGEMAP(sva + len - 1), ("not largemap range %#lx %#lx", (u_long)svaa, (u_long)svaa + len)); PMAP_LOCK(kernel_pmap); for (va = sva; va < sva + len; va += inc) { pdpe = pmap_large_map_pdpe(va); pdp = *pdpe; KASSERT((pdp & X86_PG_V) != 0, ("invalid pdp va %#lx pdpe %#lx pdp %#lx", va, (u_long)pdpe, pdp)); if ((pdp & X86_PG_PS) != 0) { KASSERT((amd_feature & AMDID_PAGE1GB) != 0, ("no 1G pages, va %#lx pdpe %#lx pdp %#lx", va, (u_long)pdpe, pdp)); KASSERT((va & PDPMASK) == 0, ("PDPMASK bit set, va %#lx pdpe %#lx pdp %#lx", va, (u_long)pdpe, pdp)); KASSERT(va + NBPDP <= sva + len, ("unmap covers partial 1GB page, sva %#lx va %#lx " "pdpe %#lx pdp %#lx len %#lx", sva, va, (u_long)pdpe, pdp, len)); *pdpe = 0; inc = NBPDP; continue; } pde = pmap_pdpe_to_pde(pdpe, va); pd = *pde; KASSERT((pd & X86_PG_V) != 0, ("invalid pd va %#lx pde %#lx pd %#lx", va, (u_long)pde, pd)); if ((pd & X86_PG_PS) != 0) { KASSERT((va & PDRMASK) == 0, ("PDRMASK bit set, va %#lx pde %#lx pd %#lx", va, (u_long)pde, pd)); KASSERT(va + NBPDR <= sva + len, ("unmap covers partial 2MB page, sva %#lx va %#lx " "pde %#lx pd %#lx len %#lx", sva, va, (u_long)pde, pd, len)); pde_store(pde, 0); inc = NBPDR; m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)pde)); m->ref_count--; if (m->ref_count == 0) { *pdpe = 0; SLIST_INSERT_HEAD(&spgf, m, plinks.s.ss); } continue; } pte = pmap_pde_to_pte(pde, va); KASSERT((*pte & X86_PG_V) != 0, ("invalid pte va %#lx pte %#lx pt %#lx", va, (u_long)pte, *pte)); pte_clear(pte); inc = PAGE_SIZE; m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)pte)); m->ref_count--; if (m->ref_count == 0) { *pde = 0; SLIST_INSERT_HEAD(&spgf, m, plinks.s.ss); m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((vm_offset_t)pde)); m->ref_count--; if (m->ref_count == 0) { *pdpe = 0; SLIST_INSERT_HEAD(&spgf, m, plinks.s.ss); } } } pmap_invalidate_range(kernel_pmap, sva, sva + len); PMAP_UNLOCK(kernel_pmap); vm_page_free_pages_toq(&spgf, false); vmem_free(large_vmem, sva, len); } static void pmap_large_map_wb_fence_mfence(void) { mfence(); } static void pmap_large_map_wb_fence_atomic(void) { atomic_thread_fence_seq_cst(); } static void pmap_large_map_wb_fence_nop(void) { } DEFINE_IFUNC(static, void, pmap_large_map_wb_fence, (void)) { if (cpu_vendor_id != CPU_VENDOR_INTEL) return (pmap_large_map_wb_fence_mfence); else if ((cpu_stdext_feature & (CPUID_STDEXT_CLWB | CPUID_STDEXT_CLFLUSHOPT)) == 0) return (pmap_large_map_wb_fence_atomic); else /* clflush is strongly enough ordered */ return (pmap_large_map_wb_fence_nop); } static void pmap_large_map_flush_range_clwb(vm_offset_t va, vm_size_t len) { for (; len > 0; len -= cpu_clflush_line_size, va += cpu_clflush_line_size) clwb(va); } static void pmap_large_map_flush_range_clflushopt(vm_offset_t va, vm_size_t len) { for (; len > 0; len -= cpu_clflush_line_size, va += cpu_clflush_line_size) clflushopt(va); } static void pmap_large_map_flush_range_clflush(vm_offset_t va, vm_size_t len) { for (; len > 0; len -= cpu_clflush_line_size, va += cpu_clflush_line_size) clflush(va); } static void pmap_large_map_flush_range_nop(vm_offset_t sva __unused, vm_size_t len __unused) { } DEFINE_IFUNC(static, void, pmap_large_map_flush_range, (vm_offset_t, vm_size_t)) { if ((cpu_stdext_feature & CPUID_STDEXT_CLWB) != 0) return (pmap_large_map_flush_range_clwb); else if ((cpu_stdext_feature & CPUID_STDEXT_CLFLUSHOPT) != 0) return (pmap_large_map_flush_range_clflushopt); else if ((cpu_feature & CPUID_CLFSH) != 0) return (pmap_large_map_flush_range_clflush); else return (pmap_large_map_flush_range_nop); } static void pmap_large_map_wb_large(vm_offset_t sva, vm_offset_t eva) { volatile u_long *pe; u_long p; vm_offset_t va; vm_size_t inc; bool seen_other; for (va = sva; va < eva; va += inc) { inc = 0; if ((amd_feature & AMDID_PAGE1GB) != 0) { pe = (volatile u_long *)pmap_large_map_pdpe(va); p = *pe; if ((p & X86_PG_PS) != 0) inc = NBPDP; } if (inc == 0) { pe = (volatile u_long *)pmap_large_map_pde(va); p = *pe; if ((p & X86_PG_PS) != 0) inc = NBPDR; } if (inc == 0) { pe = (volatile u_long *)pmap_large_map_pte(va); p = *pe; inc = PAGE_SIZE; } seen_other = false; for (;;) { if ((p & X86_PG_AVAIL1) != 0) { /* * Spin-wait for the end of a parallel * write-back. */ cpu_spinwait(); p = *pe; /* * If we saw other write-back * occuring, we cannot rely on PG_M to * indicate state of the cache. The * PG_M bit is cleared before the * flush to avoid ignoring new writes, * and writes which are relevant for * us might happen after. */ seen_other = true; continue; } if ((p & X86_PG_M) != 0 || seen_other) { if (!atomic_fcmpset_long(pe, &p, (p & ~X86_PG_M) | X86_PG_AVAIL1)) /* * If we saw PG_M without * PG_AVAIL1, and then on the * next attempt we do not * observe either PG_M or * PG_AVAIL1, the other * write-back started after us * and finished before us. We * can rely on it doing our * work. */ continue; pmap_large_map_flush_range(va, inc); atomic_clear_long(pe, X86_PG_AVAIL1); } break; } maybe_yield(); } } /* * Write-back cache lines for the given address range. * * Must be called only on the range or sub-range returned from * pmap_large_map(). Must not be called on the coalesced ranges. * * Does nothing on CPUs without CLWB, CLFLUSHOPT, or CLFLUSH * instructions support. */ void pmap_large_map_wb(void *svap, vm_size_t len) { vm_offset_t eva, sva; sva = (vm_offset_t)svap; eva = sva + len; pmap_large_map_wb_fence(); if (sva >= DMAP_MIN_ADDRESS && eva <= DMAP_MIN_ADDRESS + dmaplimit) { pmap_large_map_flush_range(sva, len); } else { KASSERT(sva >= LARGEMAP_MIN_ADDRESS && eva <= LARGEMAP_MIN_ADDRESS + lm_ents * NBPML4, ("pmap_large_map_wb: not largemap %#lx %#lx", sva, len)); pmap_large_map_wb_large(sva, eva); } pmap_large_map_wb_fence(); } static vm_page_t pmap_pti_alloc_page(void) { vm_page_t m; VM_OBJECT_ASSERT_WLOCKED(pti_obj); m = vm_page_grab(pti_obj, pti_pg_idx++, VM_ALLOC_NOBUSY | VM_ALLOC_WIRED | VM_ALLOC_ZERO); return (m); } static bool pmap_pti_free_page(vm_page_t m) { KASSERT(m->ref_count > 0, ("page %p not referenced", m)); if (!vm_page_unwire_noq(m)) return (false); vm_page_free_zero(m); return (true); } static void pmap_pti_init(void) { vm_page_t pml4_pg; pdp_entry_t *pdpe; vm_offset_t va; int i; if (!pti) return; pti_obj = vm_pager_allocate(OBJT_PHYS, NULL, 0, VM_PROT_ALL, 0, NULL); VM_OBJECT_WLOCK(pti_obj); pml4_pg = pmap_pti_alloc_page(); pti_pml4 = (pml4_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pml4_pg)); for (va = VM_MIN_KERNEL_ADDRESS; va <= VM_MAX_KERNEL_ADDRESS && va >= VM_MIN_KERNEL_ADDRESS && va > NBPML4; va += NBPML4) { pdpe = pmap_pti_pdpe(va); pmap_pti_wire_pte(pdpe); } pmap_pti_add_kva_locked((vm_offset_t)&__pcpu[0], (vm_offset_t)&__pcpu[0] + sizeof(__pcpu[0]) * MAXCPU, false); pmap_pti_add_kva_locked((vm_offset_t)idt, (vm_offset_t)idt + sizeof(struct gate_descriptor) * NIDT, false); CPU_FOREACH(i) { /* Doublefault stack IST 1 */ va = __pcpu[i].pc_common_tss.tss_ist1 + sizeof(struct nmi_pcpu); pmap_pti_add_kva_locked(va - DBLFAULT_STACK_SIZE, va, false); /* NMI stack IST 2 */ va = __pcpu[i].pc_common_tss.tss_ist2 + sizeof(struct nmi_pcpu); pmap_pti_add_kva_locked(va - NMI_STACK_SIZE, va, false); /* MC# stack IST 3 */ va = __pcpu[i].pc_common_tss.tss_ist3 + sizeof(struct nmi_pcpu); pmap_pti_add_kva_locked(va - MCE_STACK_SIZE, va, false); /* DB# stack IST 4 */ va = __pcpu[i].pc_common_tss.tss_ist4 + sizeof(struct nmi_pcpu); pmap_pti_add_kva_locked(va - DBG_STACK_SIZE, va, false); } pmap_pti_add_kva_locked((vm_offset_t)KERNSTART, (vm_offset_t)etext, true); pti_finalized = true; VM_OBJECT_WUNLOCK(pti_obj); } static void pmap_cpu_init(void *arg __unused) { CPU_COPY(&all_cpus, &kernel_pmap->pm_active); pmap_pti_init(); } SYSINIT(pmap_cpu, SI_SUB_CPU + 1, SI_ORDER_ANY, pmap_cpu_init, NULL); static pdp_entry_t * pmap_pti_pdpe(vm_offset_t va) { pml4_entry_t *pml4e; pdp_entry_t *pdpe; vm_page_t m; vm_pindex_t pml4_idx; vm_paddr_t mphys; VM_OBJECT_ASSERT_WLOCKED(pti_obj); pml4_idx = pmap_pml4e_index(va); pml4e = &pti_pml4[pml4_idx]; m = NULL; if (*pml4e == 0) { if (pti_finalized) panic("pml4 alloc after finalization\n"); m = pmap_pti_alloc_page(); if (*pml4e != 0) { pmap_pti_free_page(m); mphys = *pml4e & ~PAGE_MASK; } else { mphys = VM_PAGE_TO_PHYS(m); *pml4e = mphys | X86_PG_RW | X86_PG_V; } } else { mphys = *pml4e & ~PAGE_MASK; } pdpe = (pdp_entry_t *)PHYS_TO_DMAP(mphys) + pmap_pdpe_index(va); return (pdpe); } static void pmap_pti_wire_pte(void *pte) { vm_page_t m; VM_OBJECT_ASSERT_WLOCKED(pti_obj); m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((uintptr_t)pte)); m->ref_count++; } static void pmap_pti_unwire_pde(void *pde, bool only_ref) { vm_page_t m; VM_OBJECT_ASSERT_WLOCKED(pti_obj); m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((uintptr_t)pde)); MPASS(m->ref_count > 0); MPASS(only_ref || m->ref_count > 1); pmap_pti_free_page(m); } static void pmap_pti_unwire_pte(void *pte, vm_offset_t va) { vm_page_t m; pd_entry_t *pde; VM_OBJECT_ASSERT_WLOCKED(pti_obj); m = PHYS_TO_VM_PAGE(DMAP_TO_PHYS((uintptr_t)pte)); MPASS(m->ref_count > 0); if (pmap_pti_free_page(m)) { pde = pmap_pti_pde(va); MPASS((*pde & (X86_PG_PS | X86_PG_V)) == X86_PG_V); *pde = 0; pmap_pti_unwire_pde(pde, false); } } static pd_entry_t * pmap_pti_pde(vm_offset_t va) { pdp_entry_t *pdpe; pd_entry_t *pde; vm_page_t m; vm_pindex_t pd_idx; vm_paddr_t mphys; VM_OBJECT_ASSERT_WLOCKED(pti_obj); pdpe = pmap_pti_pdpe(va); if (*pdpe == 0) { m = pmap_pti_alloc_page(); if (*pdpe != 0) { pmap_pti_free_page(m); MPASS((*pdpe & X86_PG_PS) == 0); mphys = *pdpe & ~PAGE_MASK; } else { mphys = VM_PAGE_TO_PHYS(m); *pdpe = mphys | X86_PG_RW | X86_PG_V; } } else { MPASS((*pdpe & X86_PG_PS) == 0); mphys = *pdpe & ~PAGE_MASK; } pde = (pd_entry_t *)PHYS_TO_DMAP(mphys); pd_idx = pmap_pde_index(va); pde += pd_idx; return (pde); } static pt_entry_t * pmap_pti_pte(vm_offset_t va, bool *unwire_pde) { pd_entry_t *pde; pt_entry_t *pte; vm_page_t m; vm_paddr_t mphys; VM_OBJECT_ASSERT_WLOCKED(pti_obj); pde = pmap_pti_pde(va); if (unwire_pde != NULL) { *unwire_pde = true; pmap_pti_wire_pte(pde); } if (*pde == 0) { m = pmap_pti_alloc_page(); if (*pde != 0) { pmap_pti_free_page(m); MPASS((*pde & X86_PG_PS) == 0); mphys = *pde & ~(PAGE_MASK | pg_nx); } else { mphys = VM_PAGE_TO_PHYS(m); *pde = mphys | X86_PG_RW | X86_PG_V; if (unwire_pde != NULL) *unwire_pde = false; } } else { MPASS((*pde & X86_PG_PS) == 0); mphys = *pde & ~(PAGE_MASK | pg_nx); } pte = (pt_entry_t *)PHYS_TO_DMAP(mphys); pte += pmap_pte_index(va); return (pte); } static void pmap_pti_add_kva_locked(vm_offset_t sva, vm_offset_t eva, bool exec) { vm_paddr_t pa; pd_entry_t *pde; pt_entry_t *pte, ptev; bool unwire_pde; VM_OBJECT_ASSERT_WLOCKED(pti_obj); sva = trunc_page(sva); MPASS(sva > VM_MAXUSER_ADDRESS); eva = round_page(eva); MPASS(sva < eva); for (; sva < eva; sva += PAGE_SIZE) { pte = pmap_pti_pte(sva, &unwire_pde); pa = pmap_kextract(sva); ptev = pa | X86_PG_RW | X86_PG_V | X86_PG_A | X86_PG_G | (exec ? 0 : pg_nx) | pmap_cache_bits(kernel_pmap, VM_MEMATTR_DEFAULT, FALSE); if (*pte == 0) { pte_store(pte, ptev); pmap_pti_wire_pte(pte); } else { KASSERT(!pti_finalized, ("pti overlap after fin %#lx %#lx %#lx", sva, *pte, ptev)); KASSERT(*pte == ptev, ("pti non-identical pte after fin %#lx %#lx %#lx", sva, *pte, ptev)); } if (unwire_pde) { pde = pmap_pti_pde(sva); pmap_pti_unwire_pde(pde, true); } } } void pmap_pti_add_kva(vm_offset_t sva, vm_offset_t eva, bool exec) { if (!pti) return; VM_OBJECT_WLOCK(pti_obj); pmap_pti_add_kva_locked(sva, eva, exec); VM_OBJECT_WUNLOCK(pti_obj); } void pmap_pti_remove_kva(vm_offset_t sva, vm_offset_t eva) { pt_entry_t *pte; vm_offset_t va; if (!pti) return; sva = rounddown2(sva, PAGE_SIZE); MPASS(sva > VM_MAXUSER_ADDRESS); eva = roundup2(eva, PAGE_SIZE); MPASS(sva < eva); VM_OBJECT_WLOCK(pti_obj); for (va = sva; va < eva; va += PAGE_SIZE) { pte = pmap_pti_pte(va, NULL); KASSERT((*pte & X86_PG_V) != 0, ("invalid pte va %#lx pte %#lx pt %#lx", va, (u_long)pte, *pte)); pte_clear(pte); pmap_pti_unwire_pte(pte, va); } pmap_invalidate_range(kernel_pmap, sva, eva); VM_OBJECT_WUNLOCK(pti_obj); } static void * pkru_dup_range(void *ctx __unused, void *data) { struct pmap_pkru_range *node, *new_node; new_node = uma_zalloc(pmap_pkru_ranges_zone, M_NOWAIT); if (new_node == NULL) return (NULL); node = data; memcpy(new_node, node, sizeof(*node)); return (new_node); } static void pkru_free_range(void *ctx __unused, void *node) { uma_zfree(pmap_pkru_ranges_zone, node); } static int pmap_pkru_assign(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, u_int keyidx, int flags) { struct pmap_pkru_range *ppr; int error; PMAP_LOCK_ASSERT(pmap, MA_OWNED); MPASS(pmap->pm_type == PT_X86); MPASS((cpu_stdext_feature2 & CPUID_STDEXT2_PKU) != 0); if ((flags & AMD64_PKRU_EXCL) != 0 && !rangeset_check_empty(&pmap->pm_pkru, sva, eva)) return (EBUSY); ppr = uma_zalloc(pmap_pkru_ranges_zone, M_NOWAIT); if (ppr == NULL) return (ENOMEM); ppr->pkru_keyidx = keyidx; ppr->pkru_flags = flags & AMD64_PKRU_PERSIST; error = rangeset_insert(&pmap->pm_pkru, sva, eva, ppr); if (error != 0) uma_zfree(pmap_pkru_ranges_zone, ppr); return (error); } static int pmap_pkru_deassign(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { PMAP_LOCK_ASSERT(pmap, MA_OWNED); MPASS(pmap->pm_type == PT_X86); MPASS((cpu_stdext_feature2 & CPUID_STDEXT2_PKU) != 0); return (rangeset_remove(&pmap->pm_pkru, sva, eva)); } static void pmap_pkru_deassign_all(pmap_t pmap) { PMAP_LOCK_ASSERT(pmap, MA_OWNED); if (pmap->pm_type == PT_X86 && (cpu_stdext_feature2 & CPUID_STDEXT2_PKU) != 0) rangeset_remove_all(&pmap->pm_pkru); } static bool pmap_pkru_same(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { struct pmap_pkru_range *ppr, *prev_ppr; vm_offset_t va; PMAP_LOCK_ASSERT(pmap, MA_OWNED); if (pmap->pm_type != PT_X86 || (cpu_stdext_feature2 & CPUID_STDEXT2_PKU) == 0 || sva >= VM_MAXUSER_ADDRESS) return (true); MPASS(eva <= VM_MAXUSER_ADDRESS); for (va = sva; va < eva; prev_ppr = ppr) { ppr = rangeset_lookup(&pmap->pm_pkru, va); if (va == sva) prev_ppr = ppr; else if ((ppr == NULL) ^ (prev_ppr == NULL)) return (false); if (ppr == NULL) { va += PAGE_SIZE; continue; } if (prev_ppr->pkru_keyidx != ppr->pkru_keyidx) return (false); va = ppr->pkru_rs_el.re_end; } return (true); } static pt_entry_t pmap_pkru_get(pmap_t pmap, vm_offset_t va) { struct pmap_pkru_range *ppr; PMAP_LOCK_ASSERT(pmap, MA_OWNED); if (pmap->pm_type != PT_X86 || (cpu_stdext_feature2 & CPUID_STDEXT2_PKU) == 0 || va >= VM_MAXUSER_ADDRESS) return (0); ppr = rangeset_lookup(&pmap->pm_pkru, va); if (ppr != NULL) return (X86_PG_PKU(ppr->pkru_keyidx)); return (0); } static bool pred_pkru_on_remove(void *ctx __unused, void *r) { struct pmap_pkru_range *ppr; ppr = r; return ((ppr->pkru_flags & AMD64_PKRU_PERSIST) == 0); } static void pmap_pkru_on_remove(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { PMAP_LOCK_ASSERT(pmap, MA_OWNED); if (pmap->pm_type == PT_X86 && (cpu_stdext_feature2 & CPUID_STDEXT2_PKU) != 0) { rangeset_remove_pred(&pmap->pm_pkru, sva, eva, pred_pkru_on_remove); } } static int pmap_pkru_copy(pmap_t dst_pmap, pmap_t src_pmap) { PMAP_LOCK_ASSERT(dst_pmap, MA_OWNED); PMAP_LOCK_ASSERT(src_pmap, MA_OWNED); MPASS(dst_pmap->pm_type == PT_X86); MPASS(src_pmap->pm_type == PT_X86); MPASS((cpu_stdext_feature2 & CPUID_STDEXT2_PKU) != 0); if (src_pmap->pm_pkru.rs_data_ctx == NULL) return (0); return (rangeset_copy(&dst_pmap->pm_pkru, &src_pmap->pm_pkru)); } static void pmap_pkru_update_range(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, u_int keyidx) { pml4_entry_t *pml4e; pdp_entry_t *pdpe; pd_entry_t newpde, ptpaddr, *pde; pt_entry_t newpte, *ptep, pte; vm_offset_t va, va_next; bool changed; PMAP_LOCK_ASSERT(pmap, MA_OWNED); MPASS(pmap->pm_type == PT_X86); MPASS(keyidx <= PMAP_MAX_PKRU_IDX); for (changed = false, va = sva; va < eva; va = va_next) { pml4e = pmap_pml4e(pmap, va); if (pml4e == NULL || (*pml4e & X86_PG_V) == 0) { va_next = (va + NBPML4) & ~PML4MASK; if (va_next < va) va_next = eva; continue; } pdpe = pmap_pml4e_to_pdpe(pml4e, va); if ((*pdpe & X86_PG_V) == 0) { va_next = (va + NBPDP) & ~PDPMASK; if (va_next < va) va_next = eva; continue; } va_next = (va + NBPDR) & ~PDRMASK; if (va_next < va) va_next = eva; pde = pmap_pdpe_to_pde(pdpe, va); ptpaddr = *pde; if (ptpaddr == 0) continue; MPASS((ptpaddr & X86_PG_V) != 0); if ((ptpaddr & PG_PS) != 0) { if (va + NBPDR == va_next && eva >= va_next) { newpde = (ptpaddr & ~X86_PG_PKU_MASK) | X86_PG_PKU(keyidx); if (newpde != ptpaddr) { *pde = newpde; changed = true; } continue; } else if (!pmap_demote_pde(pmap, pde, va)) { continue; } } if (va_next > eva) va_next = eva; for (ptep = pmap_pde_to_pte(pde, va); va != va_next; ptep++, va += PAGE_SIZE) { pte = *ptep; if ((pte & X86_PG_V) == 0) continue; newpte = (pte & ~X86_PG_PKU_MASK) | X86_PG_PKU(keyidx); if (newpte != pte) { *ptep = newpte; changed = true; } } } if (changed) pmap_invalidate_range(pmap, sva, eva); } static int pmap_pkru_check_uargs(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, u_int keyidx, int flags) { if (pmap->pm_type != PT_X86 || keyidx > PMAP_MAX_PKRU_IDX || (flags & ~(AMD64_PKRU_PERSIST | AMD64_PKRU_EXCL)) != 0) return (EINVAL); if (eva <= sva || eva > VM_MAXUSER_ADDRESS) return (EFAULT); if ((cpu_stdext_feature2 & CPUID_STDEXT2_PKU) == 0) return (ENOTSUP); return (0); } int pmap_pkru_set(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, u_int keyidx, int flags) { int error; sva = trunc_page(sva); eva = round_page(eva); error = pmap_pkru_check_uargs(pmap, sva, eva, keyidx, flags); if (error != 0) return (error); for (;;) { PMAP_LOCK(pmap); error = pmap_pkru_assign(pmap, sva, eva, keyidx, flags); if (error == 0) pmap_pkru_update_range(pmap, sva, eva, keyidx); PMAP_UNLOCK(pmap); if (error != ENOMEM) break; vm_wait(NULL); } return (error); } int pmap_pkru_clear(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { int error; sva = trunc_page(sva); eva = round_page(eva); error = pmap_pkru_check_uargs(pmap, sva, eva, 0, 0); if (error != 0) return (error); for (;;) { PMAP_LOCK(pmap); error = pmap_pkru_deassign(pmap, sva, eva); if (error == 0) pmap_pkru_update_range(pmap, sva, eva, 0); PMAP_UNLOCK(pmap); if (error != ENOMEM) break; vm_wait(NULL); } return (error); } #ifdef KASAN static vm_page_t pmap_kasan_enter_alloc_4k(void) { vm_page_t m; m = vm_page_alloc_noobj(VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED | VM_ALLOC_ZERO); if (m == NULL) panic("%s: no memory to grow shadow map", __func__); return (m); } static vm_page_t pmap_kasan_enter_alloc_2m(void) { return (vm_page_alloc_noobj_contig(VM_ALLOC_WIRED | VM_ALLOC_ZERO, NPTEPG, 0, ~0ul, NBPDR, 0, VM_MEMATTR_DEFAULT)); } /* * Grow the shadow map by at least one 4KB page at the specified address. Use * 2MB pages when possible. */ void pmap_kasan_enter(vm_offset_t va) { pdp_entry_t *pdpe; pd_entry_t *pde; pt_entry_t *pte; vm_page_t m; mtx_assert(&kernel_map->system_mtx, MA_OWNED); pdpe = pmap_pdpe(kernel_pmap, va); if ((*pdpe & X86_PG_V) == 0) { m = pmap_kasan_enter_alloc_4k(); *pdpe = (pdp_entry_t)(VM_PAGE_TO_PHYS(m) | X86_PG_RW | X86_PG_V | pg_nx); } pde = pmap_pdpe_to_pde(pdpe, va); if ((*pde & X86_PG_V) == 0) { m = pmap_kasan_enter_alloc_2m(); if (m != NULL) { *pde = (pd_entry_t)(VM_PAGE_TO_PHYS(m) | X86_PG_RW | X86_PG_PS | X86_PG_V | X86_PG_A | X86_PG_M | pg_nx); } else { m = pmap_kasan_enter_alloc_4k(); *pde = (pd_entry_t)(VM_PAGE_TO_PHYS(m) | X86_PG_RW | X86_PG_V | pg_nx); } } if ((*pde & X86_PG_PS) != 0) return; pte = pmap_pde_to_pte(pde, va); if ((*pte & X86_PG_V) != 0) return; m = pmap_kasan_enter_alloc_4k(); *pte = (pt_entry_t)(VM_PAGE_TO_PHYS(m) | X86_PG_RW | X86_PG_V | X86_PG_M | X86_PG_A | pg_nx); } #endif #ifdef KMSAN static vm_page_t pmap_kmsan_enter_alloc_4k(void) { vm_page_t m; m = vm_page_alloc_noobj(VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED | VM_ALLOC_ZERO); if (m == NULL) panic("%s: no memory to grow shadow map", __func__); return (m); } static vm_page_t pmap_kmsan_enter_alloc_2m(void) { return (vm_page_alloc_noobj_contig(VM_ALLOC_ZERO | VM_ALLOC_WIRED, NPTEPG, 0, ~0ul, NBPDR, 0, VM_MEMATTR_DEFAULT)); } /* * Grow the shadow or origin maps by at least one 4KB page at the specified * address. Use 2MB pages when possible. */ void pmap_kmsan_enter(vm_offset_t va) { pdp_entry_t *pdpe; pd_entry_t *pde; pt_entry_t *pte; vm_page_t m; mtx_assert(&kernel_map->system_mtx, MA_OWNED); pdpe = pmap_pdpe(kernel_pmap, va); if ((*pdpe & X86_PG_V) == 0) { m = pmap_kmsan_enter_alloc_4k(); *pdpe = (pdp_entry_t)(VM_PAGE_TO_PHYS(m) | X86_PG_RW | X86_PG_V | pg_nx); } pde = pmap_pdpe_to_pde(pdpe, va); if ((*pde & X86_PG_V) == 0) { m = pmap_kmsan_enter_alloc_2m(); if (m != NULL) { *pde = (pd_entry_t)(VM_PAGE_TO_PHYS(m) | X86_PG_RW | X86_PG_PS | X86_PG_V | X86_PG_A | X86_PG_M | pg_nx); } else { m = pmap_kmsan_enter_alloc_4k(); *pde = (pd_entry_t)(VM_PAGE_TO_PHYS(m) | X86_PG_RW | X86_PG_V | pg_nx); } } if ((*pde & X86_PG_PS) != 0) return; pte = pmap_pde_to_pte(pde, va); if ((*pte & X86_PG_V) != 0) return; m = pmap_kmsan_enter_alloc_4k(); *pte = (pt_entry_t)(VM_PAGE_TO_PHYS(m) | X86_PG_RW | X86_PG_V | X86_PG_M | X86_PG_A | pg_nx); } #endif /* * Track a range of the kernel's virtual address space that is contiguous * in various mapping attributes. */ struct pmap_kernel_map_range { vm_offset_t sva; pt_entry_t attrs; int ptes; int pdes; int pdpes; }; static void sysctl_kmaps_dump(struct sbuf *sb, struct pmap_kernel_map_range *range, vm_offset_t eva) { const char *mode; int i, pat_idx; if (eva <= range->sva) return; pat_idx = pmap_pat_index(kernel_pmap, range->attrs, true); for (i = 0; i < PAT_INDEX_SIZE; i++) if (pat_index[i] == pat_idx) break; switch (i) { case PAT_WRITE_BACK: mode = "WB"; break; case PAT_WRITE_THROUGH: mode = "WT"; break; case PAT_UNCACHEABLE: mode = "UC"; break; case PAT_UNCACHED: mode = "U-"; break; case PAT_WRITE_PROTECTED: mode = "WP"; break; case PAT_WRITE_COMBINING: mode = "WC"; break; default: printf("%s: unknown PAT mode %#x for range 0x%016lx-0x%016lx\n", __func__, pat_idx, range->sva, eva); mode = "??"; break; } sbuf_printf(sb, "0x%016lx-0x%016lx r%c%c%c%c %s %d %d %d\n", range->sva, eva, (range->attrs & X86_PG_RW) != 0 ? 'w' : '-', (range->attrs & pg_nx) != 0 ? '-' : 'x', (range->attrs & X86_PG_U) != 0 ? 'u' : 's', (range->attrs & X86_PG_G) != 0 ? 'g' : '-', mode, range->pdpes, range->pdes, range->ptes); /* Reset to sentinel value. */ range->sva = la57 ? KV5ADDR(NPML5EPG - 1, NPML4EPG - 1, NPDPEPG - 1, NPDEPG - 1, NPTEPG - 1) : KV4ADDR(NPML4EPG - 1, NPDPEPG - 1, NPDEPG - 1, NPTEPG - 1); } /* * Determine whether the attributes specified by a page table entry match those * being tracked by the current range. This is not quite as simple as a direct * flag comparison since some PAT modes have multiple representations. */ static bool sysctl_kmaps_match(struct pmap_kernel_map_range *range, pt_entry_t attrs) { pt_entry_t diff, mask; mask = X86_PG_G | X86_PG_RW | X86_PG_U | X86_PG_PDE_CACHE | pg_nx; diff = (range->attrs ^ attrs) & mask; if (diff == 0) return (true); if ((diff & ~X86_PG_PDE_PAT) == 0 && pmap_pat_index(kernel_pmap, range->attrs, true) == pmap_pat_index(kernel_pmap, attrs, true)) return (true); return (false); } static void sysctl_kmaps_reinit(struct pmap_kernel_map_range *range, vm_offset_t va, pt_entry_t attrs) { memset(range, 0, sizeof(*range)); range->sva = va; range->attrs = attrs; } /* * Given a leaf PTE, derive the mapping's attributes. If they do not match * those of the current run, dump the address range and its attributes, and * begin a new run. */ static void sysctl_kmaps_check(struct sbuf *sb, struct pmap_kernel_map_range *range, vm_offset_t va, pml4_entry_t pml4e, pdp_entry_t pdpe, pd_entry_t pde, pt_entry_t pte) { pt_entry_t attrs; attrs = pml4e & (X86_PG_RW | X86_PG_U | pg_nx); attrs |= pdpe & pg_nx; attrs &= pg_nx | (pdpe & (X86_PG_RW | X86_PG_U)); if ((pdpe & PG_PS) != 0) { attrs |= pdpe & (X86_PG_G | X86_PG_PDE_CACHE); } else if (pde != 0) { attrs |= pde & pg_nx; attrs &= pg_nx | (pde & (X86_PG_RW | X86_PG_U)); } if ((pde & PG_PS) != 0) { attrs |= pde & (X86_PG_G | X86_PG_PDE_CACHE); } else if (pte != 0) { attrs |= pte & pg_nx; attrs &= pg_nx | (pte & (X86_PG_RW | X86_PG_U)); attrs |= pte & (X86_PG_G | X86_PG_PTE_CACHE); /* Canonicalize by always using the PDE PAT bit. */ if ((attrs & X86_PG_PTE_PAT) != 0) attrs ^= X86_PG_PDE_PAT | X86_PG_PTE_PAT; } if (range->sva > va || !sysctl_kmaps_match(range, attrs)) { sysctl_kmaps_dump(sb, range, va); sysctl_kmaps_reinit(range, va, attrs); } } static int sysctl_kmaps(SYSCTL_HANDLER_ARGS) { struct pmap_kernel_map_range range; struct sbuf sbuf, *sb; pml4_entry_t pml4e; pdp_entry_t *pdp, pdpe; pd_entry_t *pd, pde; pt_entry_t *pt, pte; vm_offset_t sva; vm_paddr_t pa; int error, i, j, k, l; error = sysctl_wire_old_buffer(req, 0); if (error != 0) return (error); sb = &sbuf; sbuf_new_for_sysctl(sb, NULL, PAGE_SIZE, req); /* Sentinel value. */ range.sva = la57 ? KV5ADDR(NPML5EPG - 1, NPML4EPG - 1, NPDPEPG - 1, NPDEPG - 1, NPTEPG - 1) : KV4ADDR(NPML4EPG - 1, NPDPEPG - 1, NPDEPG - 1, NPTEPG - 1); /* * Iterate over the kernel page tables without holding the kernel pmap * lock. Outside of the large map, kernel page table pages are never * freed, so at worst we will observe inconsistencies in the output. * Within the large map, ensure that PDP and PD page addresses are * valid before descending. */ for (sva = 0, i = pmap_pml4e_index(sva); i < NPML4EPG; i++) { switch (i) { case PML4PML4I: sbuf_printf(sb, "\nRecursive map:\n"); break; case DMPML4I: sbuf_printf(sb, "\nDirect map:\n"); break; #ifdef KASAN case KASANPML4I: sbuf_printf(sb, "\nKASAN shadow map:\n"); break; #endif #ifdef KMSAN case KMSANSHADPML4I: sbuf_printf(sb, "\nKMSAN shadow map:\n"); break; case KMSANORIGPML4I: sbuf_printf(sb, "\nKMSAN origin map:\n"); break; #endif case KPML4BASE: sbuf_printf(sb, "\nKernel map:\n"); break; case LMSPML4I: sbuf_printf(sb, "\nLarge map:\n"); break; } /* Convert to canonical form. */ if (sva == 1ul << 47) sva |= -1ul << 48; restart: pml4e = kernel_pml4[i]; if ((pml4e & X86_PG_V) == 0) { sva = rounddown2(sva, NBPML4); sysctl_kmaps_dump(sb, &range, sva); sva += NBPML4; continue; } pa = pml4e & PG_FRAME; pdp = (pdp_entry_t *)PHYS_TO_DMAP(pa); for (j = pmap_pdpe_index(sva); j < NPDPEPG; j++) { pdpe = pdp[j]; if ((pdpe & X86_PG_V) == 0) { sva = rounddown2(sva, NBPDP); sysctl_kmaps_dump(sb, &range, sva); sva += NBPDP; continue; } pa = pdpe & PG_FRAME; if ((pdpe & PG_PS) != 0) { sva = rounddown2(sva, NBPDP); sysctl_kmaps_check(sb, &range, sva, pml4e, pdpe, 0, 0); range.pdpes++; sva += NBPDP; continue; } if (PMAP_ADDRESS_IN_LARGEMAP(sva) && vm_phys_paddr_to_vm_page(pa) == NULL) { /* * Page table pages for the large map may be * freed. Validate the next-level address * before descending. */ goto restart; } pd = (pd_entry_t *)PHYS_TO_DMAP(pa); for (k = pmap_pde_index(sva); k < NPDEPG; k++) { pde = pd[k]; if ((pde & X86_PG_V) == 0) { sva = rounddown2(sva, NBPDR); sysctl_kmaps_dump(sb, &range, sva); sva += NBPDR; continue; } pa = pde & PG_FRAME; if ((pde & PG_PS) != 0) { sva = rounddown2(sva, NBPDR); sysctl_kmaps_check(sb, &range, sva, pml4e, pdpe, pde, 0); range.pdes++; sva += NBPDR; continue; } if (PMAP_ADDRESS_IN_LARGEMAP(sva) && vm_phys_paddr_to_vm_page(pa) == NULL) { /* * Page table pages for the large map * may be freed. Validate the * next-level address before descending. */ goto restart; } pt = (pt_entry_t *)PHYS_TO_DMAP(pa); for (l = pmap_pte_index(sva); l < NPTEPG; l++, sva += PAGE_SIZE) { pte = pt[l]; if ((pte & X86_PG_V) == 0) { sysctl_kmaps_dump(sb, &range, sva); continue; } sysctl_kmaps_check(sb, &range, sva, pml4e, pdpe, pde, pte); range.ptes++; } } } } error = sbuf_finish(sb); sbuf_delete(sb); return (error); } SYSCTL_OID(_vm_pmap, OID_AUTO, kernel_maps, CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE | CTLFLAG_SKIP, NULL, 0, sysctl_kmaps, "A", "Dump kernel address layout"); #ifdef DDB DB_SHOW_COMMAND(pte, pmap_print_pte) { pmap_t pmap; pml5_entry_t *pml5; pml4_entry_t *pml4; pdp_entry_t *pdp; pd_entry_t *pde; pt_entry_t *pte, PG_V; vm_offset_t va; if (!have_addr) { db_printf("show pte addr\n"); return; } va = (vm_offset_t)addr; if (kdb_thread != NULL) pmap = vmspace_pmap(kdb_thread->td_proc->p_vmspace); else pmap = PCPU_GET(curpmap); PG_V = pmap_valid_bit(pmap); db_printf("VA 0x%016lx", va); if (pmap_is_la57(pmap)) { pml5 = pmap_pml5e(pmap, va); db_printf(" pml5e 0x%016lx", *pml5); if ((*pml5 & PG_V) == 0) { db_printf("\n"); return; } pml4 = pmap_pml5e_to_pml4e(pml5, va); } else { pml4 = pmap_pml4e(pmap, va); } db_printf(" pml4e 0x%016lx", *pml4); if ((*pml4 & PG_V) == 0) { db_printf("\n"); return; } pdp = pmap_pml4e_to_pdpe(pml4, va); db_printf(" pdpe 0x%016lx", *pdp); if ((*pdp & PG_V) == 0 || (*pdp & PG_PS) != 0) { db_printf("\n"); return; } pde = pmap_pdpe_to_pde(pdp, va); db_printf(" pde 0x%016lx", *pde); if ((*pde & PG_V) == 0 || (*pde & PG_PS) != 0) { db_printf("\n"); return; } pte = pmap_pde_to_pte(pde, va); db_printf(" pte 0x%016lx\n", *pte); } DB_SHOW_COMMAND(phys2dmap, pmap_phys2dmap) { vm_paddr_t a; if (have_addr) { a = (vm_paddr_t)addr; db_printf("0x%jx\n", (uintmax_t)PHYS_TO_DMAP(a)); } else { db_printf("show phys2dmap addr\n"); } } static void ptpages_show_page(int level, int idx, vm_page_t pg) { db_printf("l %d i %d pg %p phys %#lx ref %x\n", level, idx, pg, VM_PAGE_TO_PHYS(pg), pg->ref_count); } static void ptpages_show_complain(int level, int idx, uint64_t pte) { db_printf("l %d i %d pte %#lx\n", level, idx, pte); } static void ptpages_show_pml4(vm_page_t pg4, int num_entries, uint64_t PG_V) { vm_page_t pg3, pg2, pg1; pml4_entry_t *pml4; pdp_entry_t *pdp; pd_entry_t *pd; int i4, i3, i2; pml4 = (pml4_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pg4)); for (i4 = 0; i4 < num_entries; i4++) { if ((pml4[i4] & PG_V) == 0) continue; pg3 = PHYS_TO_VM_PAGE(pml4[i4] & PG_FRAME); if (pg3 == NULL) { ptpages_show_complain(3, i4, pml4[i4]); continue; } ptpages_show_page(3, i4, pg3); pdp = (pdp_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pg3)); for (i3 = 0; i3 < NPDPEPG; i3++) { if ((pdp[i3] & PG_V) == 0) continue; pg2 = PHYS_TO_VM_PAGE(pdp[i3] & PG_FRAME); if (pg3 == NULL) { ptpages_show_complain(2, i3, pdp[i3]); continue; } ptpages_show_page(2, i3, pg2); pd = (pd_entry_t *)PHYS_TO_DMAP(VM_PAGE_TO_PHYS(pg2)); for (i2 = 0; i2 < NPDEPG; i2++) { if ((pd[i2] & PG_V) == 0) continue; pg1 = PHYS_TO_VM_PAGE(pd[i2] & PG_FRAME); if (pg1 == NULL) { ptpages_show_complain(1, i2, pd[i2]); continue; } ptpages_show_page(1, i2, pg1); } } } } DB_SHOW_COMMAND(ptpages, pmap_ptpages) { pmap_t pmap; vm_page_t pg; pml5_entry_t *pml5; uint64_t PG_V; int i5; if (have_addr) pmap = (pmap_t)addr; else pmap = PCPU_GET(curpmap); PG_V = pmap_valid_bit(pmap); if (pmap_is_la57(pmap)) { pml5 = pmap->pm_pmltop; for (i5 = 0; i5 < NUPML5E; i5++) { if ((pml5[i5] & PG_V) == 0) continue; pg = PHYS_TO_VM_PAGE(pml5[i5] & PG_FRAME); if (pg == NULL) { ptpages_show_complain(4, i5, pml5[i5]); continue; } ptpages_show_page(4, i5, pg); ptpages_show_pml4(pg, NPML4EPG, PG_V); } } else { ptpages_show_pml4(PHYS_TO_VM_PAGE(DMAP_TO_PHYS( (vm_offset_t)pmap->pm_pmltop)), NUP4ML4E, PG_V); } } #endif diff --git a/sys/amd64/vmm/vmm.c b/sys/amd64/vmm/vmm.c index 80fcde8d80fa..f0674784903f 100644 --- a/sys/amd64/vmm/vmm.c +++ b/sys/amd64/vmm/vmm.c @@ -1,2953 +1,2953 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2011 NetApp, Inc. * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY NETAPP, INC ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL NETAPP, INC OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * $FreeBSD$ */ #include __FBSDID("$FreeBSD$"); #include "opt_bhyve_snapshot.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include "vmm_ioport.h" #include "vmm_ktr.h" #include "vmm_host.h" #include "vmm_mem.h" #include "vmm_util.h" #include "vatpic.h" #include "vatpit.h" #include "vhpet.h" #include "vioapic.h" #include "vlapic.h" #include "vpmtmr.h" #include "vrtc.h" #include "vmm_stat.h" #include "vmm_lapic.h" #include "io/ppt.h" #include "io/iommu.h" struct vlapic; /* * Initialization: * (a) allocated when vcpu is created * (i) initialized when vcpu is created and when it is reinitialized * (o) initialized the first time the vcpu is created * (x) initialized before use */ struct vcpu { struct mtx mtx; /* (o) protects 'state' and 'hostcpu' */ enum vcpu_state state; /* (o) vcpu state */ int hostcpu; /* (o) vcpu's host cpu */ int reqidle; /* (i) request vcpu to idle */ struct vlapic *vlapic; /* (i) APIC device model */ enum x2apic_state x2apic_state; /* (i) APIC mode */ uint64_t exitintinfo; /* (i) events pending at VM exit */ int nmi_pending; /* (i) NMI pending */ int extint_pending; /* (i) INTR pending */ int exception_pending; /* (i) exception pending */ int exc_vector; /* (x) exception collateral */ int exc_errcode_valid; uint32_t exc_errcode; struct savefpu *guestfpu; /* (a,i) guest fpu state */ uint64_t guest_xcr0; /* (i) guest %xcr0 register */ void *stats; /* (a,i) statistics */ struct vm_exit exitinfo; /* (x) exit reason and collateral */ uint64_t nextrip; /* (x) next instruction to execute */ uint64_t tsc_offset; /* (o) TSC offsetting */ }; #define vcpu_lock_initialized(v) mtx_initialized(&((v)->mtx)) #define vcpu_lock_init(v) mtx_init(&((v)->mtx), "vcpu lock", 0, MTX_SPIN) #define vcpu_lock(v) mtx_lock_spin(&((v)->mtx)) #define vcpu_unlock(v) mtx_unlock_spin(&((v)->mtx)) #define vcpu_assert_locked(v) mtx_assert(&((v)->mtx), MA_OWNED) struct mem_seg { size_t len; bool sysmem; struct vm_object *object; }; #define VM_MAX_MEMSEGS 3 struct mem_map { vm_paddr_t gpa; size_t len; vm_ooffset_t segoff; int segid; int prot; int flags; }; #define VM_MAX_MEMMAPS 8 /* * Initialization: * (o) initialized the first time the VM is created * (i) initialized when VM is created and when it is reinitialized * (x) initialized before use */ struct vm { void *cookie; /* (i) cpu-specific data */ void *iommu; /* (x) iommu-specific data */ struct vhpet *vhpet; /* (i) virtual HPET */ struct vioapic *vioapic; /* (i) virtual ioapic */ struct vatpic *vatpic; /* (i) virtual atpic */ struct vatpit *vatpit; /* (i) virtual atpit */ struct vpmtmr *vpmtmr; /* (i) virtual ACPI PM timer */ struct vrtc *vrtc; /* (o) virtual RTC */ volatile cpuset_t active_cpus; /* (i) active vcpus */ volatile cpuset_t debug_cpus; /* (i) vcpus stopped for debug */ int suspend; /* (i) stop VM execution */ volatile cpuset_t suspended_cpus; /* (i) suspended vcpus */ volatile cpuset_t halted_cpus; /* (x) cpus in a hard halt */ cpuset_t rendezvous_req_cpus; /* (x) rendezvous requested */ cpuset_t rendezvous_done_cpus; /* (x) rendezvous finished */ void *rendezvous_arg; /* (x) rendezvous func/arg */ vm_rendezvous_func_t rendezvous_func; struct mtx rendezvous_mtx; /* (o) rendezvous lock */ struct mem_map mem_maps[VM_MAX_MEMMAPS]; /* (i) guest address space */ struct mem_seg mem_segs[VM_MAX_MEMSEGS]; /* (o) guest memory regions */ struct vmspace *vmspace; /* (o) guest's address space */ char name[VM_MAX_NAMELEN+1]; /* (o) virtual machine name */ struct vcpu vcpu[VM_MAXCPU]; /* (i) guest vcpus */ /* The following describe the vm cpu topology */ uint16_t sockets; /* (o) num of sockets */ uint16_t cores; /* (o) num of cores/socket */ uint16_t threads; /* (o) num of threads/core */ uint16_t maxcpus; /* (o) max pluggable cpus */ }; static int vmm_initialized; static void vmmops_panic(void); static void vmmops_panic(void) { panic("vmm_ops func called when !vmm_is_intel() && !vmm_is_svm()"); } #define DEFINE_VMMOPS_IFUNC(ret_type, opname, args) \ DEFINE_IFUNC(static, ret_type, vmmops_##opname, args) \ { \ if (vmm_is_intel()) \ return (vmm_ops_intel.opname); \ else if (vmm_is_svm()) \ return (vmm_ops_amd.opname); \ else \ return ((ret_type (*)args)vmmops_panic); \ } DEFINE_VMMOPS_IFUNC(int, modinit, (int ipinum)) DEFINE_VMMOPS_IFUNC(int, modcleanup, (void)) DEFINE_VMMOPS_IFUNC(void, modresume, (void)) DEFINE_VMMOPS_IFUNC(void *, init, (struct vm *vm, struct pmap *pmap)) DEFINE_VMMOPS_IFUNC(int, run, (void *vmi, int vcpu, register_t rip, struct pmap *pmap, struct vm_eventinfo *info)) DEFINE_VMMOPS_IFUNC(void, cleanup, (void *vmi)) DEFINE_VMMOPS_IFUNC(int, getreg, (void *vmi, int vcpu, int num, uint64_t *retval)) DEFINE_VMMOPS_IFUNC(int, setreg, (void *vmi, int vcpu, int num, uint64_t val)) DEFINE_VMMOPS_IFUNC(int, getdesc, (void *vmi, int vcpu, int num, struct seg_desc *desc)) DEFINE_VMMOPS_IFUNC(int, setdesc, (void *vmi, int vcpu, int num, struct seg_desc *desc)) DEFINE_VMMOPS_IFUNC(int, getcap, (void *vmi, int vcpu, int num, int *retval)) DEFINE_VMMOPS_IFUNC(int, setcap, (void *vmi, int vcpu, int num, int val)) DEFINE_VMMOPS_IFUNC(struct vmspace *, vmspace_alloc, (vm_offset_t min, vm_offset_t max)) DEFINE_VMMOPS_IFUNC(void, vmspace_free, (struct vmspace *vmspace)) DEFINE_VMMOPS_IFUNC(struct vlapic *, vlapic_init, (void *vmi, int vcpu)) DEFINE_VMMOPS_IFUNC(void, vlapic_cleanup, (void *vmi, struct vlapic *vlapic)) #ifdef BHYVE_SNAPSHOT DEFINE_VMMOPS_IFUNC(int, snapshot, (void *vmi, struct vm_snapshot_meta *meta)) DEFINE_VMMOPS_IFUNC(int, vmcx_snapshot, (void *vmi, struct vm_snapshot_meta *meta, int vcpu)) DEFINE_VMMOPS_IFUNC(int, restore_tsc, (void *vmi, int vcpuid, uint64_t now)) #endif #define fpu_start_emulating() load_cr0(rcr0() | CR0_TS) #define fpu_stop_emulating() clts() SDT_PROVIDER_DEFINE(vmm); static MALLOC_DEFINE(M_VM, "vm", "vm"); /* statistics */ static VMM_STAT(VCPU_TOTAL_RUNTIME, "vcpu total runtime"); SYSCTL_NODE(_hw, OID_AUTO, vmm, CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, NULL); /* * Halt the guest if all vcpus are executing a HLT instruction with * interrupts disabled. */ static int halt_detection_enabled = 1; SYSCTL_INT(_hw_vmm, OID_AUTO, halt_detection, CTLFLAG_RDTUN, &halt_detection_enabled, 0, "Halt VM if all vcpus execute HLT with interrupts disabled"); static int vmm_ipinum; SYSCTL_INT(_hw_vmm, OID_AUTO, ipinum, CTLFLAG_RD, &vmm_ipinum, 0, "IPI vector used for vcpu notifications"); static int trace_guest_exceptions; SYSCTL_INT(_hw_vmm, OID_AUTO, trace_guest_exceptions, CTLFLAG_RDTUN, &trace_guest_exceptions, 0, "Trap into hypervisor on all guest exceptions and reflect them back"); static void vm_free_memmap(struct vm *vm, int ident); static bool sysmem_mapping(struct vm *vm, struct mem_map *mm); static void vcpu_notify_event_locked(struct vcpu *vcpu, bool lapic_intr); #ifdef KTR static const char * vcpu_state2str(enum vcpu_state state) { switch (state) { case VCPU_IDLE: return ("idle"); case VCPU_FROZEN: return ("frozen"); case VCPU_RUNNING: return ("running"); case VCPU_SLEEPING: return ("sleeping"); default: return ("unknown"); } } #endif static void vcpu_cleanup(struct vm *vm, int i, bool destroy) { struct vcpu *vcpu = &vm->vcpu[i]; vmmops_vlapic_cleanup(vm->cookie, vcpu->vlapic); if (destroy) { vmm_stat_free(vcpu->stats); fpu_save_area_free(vcpu->guestfpu); } } static void vcpu_init(struct vm *vm, int vcpu_id, bool create) { struct vcpu *vcpu; KASSERT(vcpu_id >= 0 && vcpu_id < vm->maxcpus, ("vcpu_init: invalid vcpu %d", vcpu_id)); vcpu = &vm->vcpu[vcpu_id]; if (create) { KASSERT(!vcpu_lock_initialized(vcpu), ("vcpu %d already " "initialized", vcpu_id)); vcpu_lock_init(vcpu); vcpu->state = VCPU_IDLE; vcpu->hostcpu = NOCPU; vcpu->guestfpu = fpu_save_area_alloc(); vcpu->stats = vmm_stat_alloc(); vcpu->tsc_offset = 0; } vcpu->vlapic = vmmops_vlapic_init(vm->cookie, vcpu_id); vm_set_x2apic_state(vm, vcpu_id, X2APIC_DISABLED); vcpu->reqidle = 0; vcpu->exitintinfo = 0; vcpu->nmi_pending = 0; vcpu->extint_pending = 0; vcpu->exception_pending = 0; vcpu->guest_xcr0 = XFEATURE_ENABLED_X87; fpu_save_area_reset(vcpu->guestfpu); vmm_stat_init(vcpu->stats); } int vcpu_trace_exceptions(struct vm *vm, int vcpuid) { return (trace_guest_exceptions); } struct vm_exit * vm_exitinfo(struct vm *vm, int cpuid) { struct vcpu *vcpu; if (cpuid < 0 || cpuid >= vm->maxcpus) panic("vm_exitinfo: invalid cpuid %d", cpuid); vcpu = &vm->vcpu[cpuid]; return (&vcpu->exitinfo); } static int vmm_init(void) { int error; if (!vmm_is_hw_supported()) return (ENXIO); vmm_host_state_init(); vmm_ipinum = lapic_ipi_alloc(pti ? &IDTVEC(justreturn1_pti) : &IDTVEC(justreturn)); if (vmm_ipinum < 0) vmm_ipinum = IPI_AST; error = vmm_mem_init(); if (error) return (error); vmm_resume_p = vmmops_modresume; return (vmmops_modinit(vmm_ipinum)); } static int vmm_handler(module_t mod, int what, void *arg) { int error; switch (what) { case MOD_LOAD: if (vmm_is_hw_supported()) { vmmdev_init(); error = vmm_init(); if (error == 0) vmm_initialized = 1; } else { error = ENXIO; } break; case MOD_UNLOAD: if (vmm_is_hw_supported()) { error = vmmdev_cleanup(); if (error == 0) { vmm_resume_p = NULL; iommu_cleanup(); if (vmm_ipinum != IPI_AST) lapic_ipi_free(vmm_ipinum); error = vmmops_modcleanup(); /* * Something bad happened - prevent new * VMs from being created */ if (error) vmm_initialized = 0; } } else { error = 0; } break; default: error = 0; break; } return (error); } static moduledata_t vmm_kmod = { "vmm", vmm_handler, NULL }; /* * vmm initialization has the following dependencies: * * - VT-x initialization requires smp_rendezvous() and therefore must happen * after SMP is fully functional (after SI_SUB_SMP). */ DECLARE_MODULE(vmm, vmm_kmod, SI_SUB_SMP + 1, SI_ORDER_ANY); MODULE_VERSION(vmm, 1); static void vm_init(struct vm *vm, bool create) { int i; vm->cookie = vmmops_init(vm, vmspace_pmap(vm->vmspace)); vm->iommu = NULL; vm->vioapic = vioapic_init(vm); vm->vhpet = vhpet_init(vm); vm->vatpic = vatpic_init(vm); vm->vatpit = vatpit_init(vm); vm->vpmtmr = vpmtmr_init(vm); if (create) vm->vrtc = vrtc_init(vm); CPU_ZERO(&vm->active_cpus); CPU_ZERO(&vm->debug_cpus); vm->suspend = 0; CPU_ZERO(&vm->suspended_cpus); for (i = 0; i < vm->maxcpus; i++) vcpu_init(vm, i, create); } /* * The default CPU topology is a single thread per package. */ u_int cores_per_package = 1; u_int threads_per_core = 1; int vm_create(const char *name, struct vm **retvm) { struct vm *vm; struct vmspace *vmspace; /* * If vmm.ko could not be successfully initialized then don't attempt * to create the virtual machine. */ if (!vmm_initialized) return (ENXIO); if (name == NULL || strnlen(name, VM_MAX_NAMELEN + 1) == VM_MAX_NAMELEN + 1) return (EINVAL); vmspace = vmmops_vmspace_alloc(0, VM_MAXUSER_ADDRESS_LA48); if (vmspace == NULL) return (ENOMEM); vm = malloc(sizeof(struct vm), M_VM, M_WAITOK | M_ZERO); strcpy(vm->name, name); vm->vmspace = vmspace; mtx_init(&vm->rendezvous_mtx, "vm rendezvous lock", 0, MTX_DEF); vm->sockets = 1; vm->cores = cores_per_package; /* XXX backwards compatibility */ vm->threads = threads_per_core; /* XXX backwards compatibility */ vm->maxcpus = VM_MAXCPU; /* XXX temp to keep code working */ vm_init(vm, true); *retvm = vm; return (0); } void vm_get_topology(struct vm *vm, uint16_t *sockets, uint16_t *cores, uint16_t *threads, uint16_t *maxcpus) { *sockets = vm->sockets; *cores = vm->cores; *threads = vm->threads; *maxcpus = vm->maxcpus; } uint16_t vm_get_maxcpus(struct vm *vm) { return (vm->maxcpus); } int vm_set_topology(struct vm *vm, uint16_t sockets, uint16_t cores, uint16_t threads, uint16_t maxcpus) { if (maxcpus != 0) return (EINVAL); /* XXX remove when supported */ if ((sockets * cores * threads) > vm->maxcpus) return (EINVAL); /* XXX need to check sockets * cores * threads == vCPU, how? */ vm->sockets = sockets; vm->cores = cores; vm->threads = threads; vm->maxcpus = VM_MAXCPU; /* XXX temp to keep code working */ return(0); } static void vm_cleanup(struct vm *vm, bool destroy) { struct mem_map *mm; int i; ppt_unassign_all(vm); if (vm->iommu != NULL) iommu_destroy_domain(vm->iommu); if (destroy) vrtc_cleanup(vm->vrtc); else vrtc_reset(vm->vrtc); vpmtmr_cleanup(vm->vpmtmr); vatpit_cleanup(vm->vatpit); vhpet_cleanup(vm->vhpet); vatpic_cleanup(vm->vatpic); vioapic_cleanup(vm->vioapic); for (i = 0; i < vm->maxcpus; i++) vcpu_cleanup(vm, i, destroy); vmmops_cleanup(vm->cookie); /* * System memory is removed from the guest address space only when * the VM is destroyed. This is because the mapping remains the same * across VM reset. * * Device memory can be relocated by the guest (e.g. using PCI BARs) * so those mappings are removed on a VM reset. */ for (i = 0; i < VM_MAX_MEMMAPS; i++) { mm = &vm->mem_maps[i]; if (destroy || !sysmem_mapping(vm, mm)) vm_free_memmap(vm, i); } if (destroy) { for (i = 0; i < VM_MAX_MEMSEGS; i++) vm_free_memseg(vm, i); vmmops_vmspace_free(vm->vmspace); vm->vmspace = NULL; } } void vm_destroy(struct vm *vm) { vm_cleanup(vm, true); free(vm, M_VM); } int vm_reinit(struct vm *vm) { int error; /* * A virtual machine can be reset only if all vcpus are suspended. */ if (CPU_CMP(&vm->suspended_cpus, &vm->active_cpus) == 0) { vm_cleanup(vm, false); vm_init(vm, false); error = 0; } else { error = EBUSY; } return (error); } const char * vm_name(struct vm *vm) { return (vm->name); } int vm_map_mmio(struct vm *vm, vm_paddr_t gpa, size_t len, vm_paddr_t hpa) { vm_object_t obj; if ((obj = vmm_mmio_alloc(vm->vmspace, gpa, len, hpa)) == NULL) return (ENOMEM); else return (0); } int vm_unmap_mmio(struct vm *vm, vm_paddr_t gpa, size_t len) { vmm_mmio_free(vm->vmspace, gpa, len); return (0); } /* * Return 'true' if 'gpa' is allocated in the guest address space. * * This function is called in the context of a running vcpu which acts as * an implicit lock on 'vm->mem_maps[]'. */ bool vm_mem_allocated(struct vm *vm, int vcpuid, vm_paddr_t gpa) { struct mem_map *mm; int i; #ifdef INVARIANTS int hostcpu, state; state = vcpu_get_state(vm, vcpuid, &hostcpu); KASSERT(state == VCPU_RUNNING && hostcpu == curcpu, ("%s: invalid vcpu state %d/%d", __func__, state, hostcpu)); #endif for (i = 0; i < VM_MAX_MEMMAPS; i++) { mm = &vm->mem_maps[i]; if (mm->len != 0 && gpa >= mm->gpa && gpa < mm->gpa + mm->len) return (true); /* 'gpa' is sysmem or devmem */ } if (ppt_is_mmio(vm, gpa)) return (true); /* 'gpa' is pci passthru mmio */ return (false); } int vm_alloc_memseg(struct vm *vm, int ident, size_t len, bool sysmem) { struct mem_seg *seg; vm_object_t obj; if (ident < 0 || ident >= VM_MAX_MEMSEGS) return (EINVAL); if (len == 0 || (len & PAGE_MASK)) return (EINVAL); seg = &vm->mem_segs[ident]; if (seg->object != NULL) { if (seg->len == len && seg->sysmem == sysmem) return (EEXIST); else return (EINVAL); } obj = vm_object_allocate(OBJT_DEFAULT, len >> PAGE_SHIFT); if (obj == NULL) return (ENOMEM); seg->len = len; seg->object = obj; seg->sysmem = sysmem; return (0); } int vm_get_memseg(struct vm *vm, int ident, size_t *len, bool *sysmem, vm_object_t *objptr) { struct mem_seg *seg; if (ident < 0 || ident >= VM_MAX_MEMSEGS) return (EINVAL); seg = &vm->mem_segs[ident]; if (len) *len = seg->len; if (sysmem) *sysmem = seg->sysmem; if (objptr) *objptr = seg->object; return (0); } void vm_free_memseg(struct vm *vm, int ident) { struct mem_seg *seg; KASSERT(ident >= 0 && ident < VM_MAX_MEMSEGS, ("%s: invalid memseg ident %d", __func__, ident)); seg = &vm->mem_segs[ident]; if (seg->object != NULL) { vm_object_deallocate(seg->object); bzero(seg, sizeof(struct mem_seg)); } } int vm_mmap_memseg(struct vm *vm, vm_paddr_t gpa, int segid, vm_ooffset_t first, size_t len, int prot, int flags) { struct mem_seg *seg; struct mem_map *m, *map; vm_ooffset_t last; int i, error; if (prot == 0 || (prot & ~(VM_PROT_ALL)) != 0) return (EINVAL); if (flags & ~VM_MEMMAP_F_WIRED) return (EINVAL); if (segid < 0 || segid >= VM_MAX_MEMSEGS) return (EINVAL); seg = &vm->mem_segs[segid]; if (seg->object == NULL) return (EINVAL); last = first + len; if (first < 0 || first >= last || last > seg->len) return (EINVAL); if ((gpa | first | last) & PAGE_MASK) return (EINVAL); map = NULL; for (i = 0; i < VM_MAX_MEMMAPS; i++) { m = &vm->mem_maps[i]; if (m->len == 0) { map = m; break; } } if (map == NULL) return (ENOSPC); error = vm_map_find(&vm->vmspace->vm_map, seg->object, first, &gpa, len, 0, VMFS_NO_SPACE, prot, prot, 0); if (error != KERN_SUCCESS) return (EFAULT); vm_object_reference(seg->object); if (flags & VM_MEMMAP_F_WIRED) { error = vm_map_wire(&vm->vmspace->vm_map, gpa, gpa + len, VM_MAP_WIRE_USER | VM_MAP_WIRE_NOHOLES); if (error != KERN_SUCCESS) { vm_map_remove(&vm->vmspace->vm_map, gpa, gpa + len); return (error == KERN_RESOURCE_SHORTAGE ? ENOMEM : EFAULT); } } map->gpa = gpa; map->len = len; map->segoff = first; map->segid = segid; map->prot = prot; map->flags = flags; return (0); } int vm_munmap_memseg(struct vm *vm, vm_paddr_t gpa, size_t len) { struct mem_map *m; int i; for (i = 0; i < VM_MAX_MEMMAPS; i++) { m = &vm->mem_maps[i]; if (m->gpa == gpa && m->len == len && (m->flags & VM_MEMMAP_F_IOMMU) == 0) { vm_free_memmap(vm, i); return (0); } } return (EINVAL); } int vm_mmap_getnext(struct vm *vm, vm_paddr_t *gpa, int *segid, vm_ooffset_t *segoff, size_t *len, int *prot, int *flags) { struct mem_map *mm, *mmnext; int i; mmnext = NULL; for (i = 0; i < VM_MAX_MEMMAPS; i++) { mm = &vm->mem_maps[i]; if (mm->len == 0 || mm->gpa < *gpa) continue; if (mmnext == NULL || mm->gpa < mmnext->gpa) mmnext = mm; } if (mmnext != NULL) { *gpa = mmnext->gpa; if (segid) *segid = mmnext->segid; if (segoff) *segoff = mmnext->segoff; if (len) *len = mmnext->len; if (prot) *prot = mmnext->prot; if (flags) *flags = mmnext->flags; return (0); } else { return (ENOENT); } } static void vm_free_memmap(struct vm *vm, int ident) { struct mem_map *mm; int error; mm = &vm->mem_maps[ident]; if (mm->len) { error = vm_map_remove(&vm->vmspace->vm_map, mm->gpa, mm->gpa + mm->len); KASSERT(error == KERN_SUCCESS, ("%s: vm_map_remove error %d", __func__, error)); bzero(mm, sizeof(struct mem_map)); } } static __inline bool sysmem_mapping(struct vm *vm, struct mem_map *mm) { if (mm->len != 0 && vm->mem_segs[mm->segid].sysmem) return (true); else return (false); } vm_paddr_t vmm_sysmem_maxaddr(struct vm *vm) { struct mem_map *mm; vm_paddr_t maxaddr; int i; maxaddr = 0; for (i = 0; i < VM_MAX_MEMMAPS; i++) { mm = &vm->mem_maps[i]; if (sysmem_mapping(vm, mm)) { if (maxaddr < mm->gpa + mm->len) maxaddr = mm->gpa + mm->len; } } return (maxaddr); } static void vm_iommu_modify(struct vm *vm, bool map) { int i, sz; vm_paddr_t gpa, hpa; struct mem_map *mm; void *vp, *cookie, *host_domain; sz = PAGE_SIZE; host_domain = iommu_host_domain(); for (i = 0; i < VM_MAX_MEMMAPS; i++) { mm = &vm->mem_maps[i]; if (!sysmem_mapping(vm, mm)) continue; if (map) { KASSERT((mm->flags & VM_MEMMAP_F_IOMMU) == 0, ("iommu map found invalid memmap %#lx/%#lx/%#x", mm->gpa, mm->len, mm->flags)); if ((mm->flags & VM_MEMMAP_F_WIRED) == 0) continue; mm->flags |= VM_MEMMAP_F_IOMMU; } else { if ((mm->flags & VM_MEMMAP_F_IOMMU) == 0) continue; mm->flags &= ~VM_MEMMAP_F_IOMMU; KASSERT((mm->flags & VM_MEMMAP_F_WIRED) != 0, ("iommu unmap found invalid memmap %#lx/%#lx/%#x", mm->gpa, mm->len, mm->flags)); } gpa = mm->gpa; while (gpa < mm->gpa + mm->len) { vp = vm_gpa_hold(vm, -1, gpa, PAGE_SIZE, VM_PROT_WRITE, &cookie); KASSERT(vp != NULL, ("vm(%s) could not map gpa %#lx", vm_name(vm), gpa)); vm_gpa_release(cookie); hpa = DMAP_TO_PHYS((uintptr_t)vp); if (map) { iommu_create_mapping(vm->iommu, gpa, hpa, sz); iommu_remove_mapping(host_domain, hpa, sz); } else { iommu_remove_mapping(vm->iommu, gpa, sz); iommu_create_mapping(host_domain, hpa, hpa, sz); } gpa += PAGE_SIZE; } } /* * Invalidate the cached translations associated with the domain * from which pages were removed. */ if (map) iommu_invalidate_tlb(host_domain); else iommu_invalidate_tlb(vm->iommu); } #define vm_iommu_unmap(vm) vm_iommu_modify((vm), false) #define vm_iommu_map(vm) vm_iommu_modify((vm), true) int vm_unassign_pptdev(struct vm *vm, int bus, int slot, int func) { int error; error = ppt_unassign_device(vm, bus, slot, func); if (error) return (error); if (ppt_assigned_devices(vm) == 0) vm_iommu_unmap(vm); return (0); } int vm_assign_pptdev(struct vm *vm, int bus, int slot, int func) { int error; vm_paddr_t maxaddr; /* Set up the IOMMU to do the 'gpa' to 'hpa' translation */ if (ppt_assigned_devices(vm) == 0) { KASSERT(vm->iommu == NULL, ("vm_assign_pptdev: iommu must be NULL")); maxaddr = vmm_sysmem_maxaddr(vm); vm->iommu = iommu_create_domain(maxaddr); if (vm->iommu == NULL) return (ENXIO); vm_iommu_map(vm); } error = ppt_assign_device(vm, bus, slot, func); return (error); } void * vm_gpa_hold(struct vm *vm, int vcpuid, vm_paddr_t gpa, size_t len, int reqprot, void **cookie) { int i, count, pageoff; struct mem_map *mm; vm_page_t m; #ifdef INVARIANTS /* * All vcpus are frozen by ioctls that modify the memory map * (e.g. VM_MMAP_MEMSEG). Therefore 'vm->memmap[]' stability is * guaranteed if at least one vcpu is in the VCPU_FROZEN state. */ int state; KASSERT(vcpuid >= -1 && vcpuid < vm->maxcpus, ("%s: invalid vcpuid %d", __func__, vcpuid)); for (i = 0; i < vm->maxcpus; i++) { if (vcpuid != -1 && vcpuid != i) continue; state = vcpu_get_state(vm, i, NULL); KASSERT(state == VCPU_FROZEN, ("%s: invalid vcpu state %d", __func__, state)); } #endif pageoff = gpa & PAGE_MASK; if (len > PAGE_SIZE - pageoff) panic("vm_gpa_hold: invalid gpa/len: 0x%016lx/%lu", gpa, len); count = 0; for (i = 0; i < VM_MAX_MEMMAPS; i++) { mm = &vm->mem_maps[i]; if (gpa >= mm->gpa && gpa < mm->gpa + mm->len) { count = vm_fault_quick_hold_pages(&vm->vmspace->vm_map, trunc_page(gpa), PAGE_SIZE, reqprot, &m, 1); break; } } if (count == 1) { *cookie = m; return ((void *)(PHYS_TO_DMAP(VM_PAGE_TO_PHYS(m)) + pageoff)); } else { *cookie = NULL; return (NULL); } } void vm_gpa_release(void *cookie) { vm_page_t m = cookie; vm_page_unwire(m, PQ_ACTIVE); } int vm_get_register(struct vm *vm, int vcpu, int reg, uint64_t *retval) { if (vcpu < 0 || vcpu >= vm->maxcpus) return (EINVAL); if (reg >= VM_REG_LAST) return (EINVAL); return (vmmops_getreg(vm->cookie, vcpu, reg, retval)); } int vm_set_register(struct vm *vm, int vcpuid, int reg, uint64_t val) { struct vcpu *vcpu; int error; if (vcpuid < 0 || vcpuid >= vm->maxcpus) return (EINVAL); if (reg >= VM_REG_LAST) return (EINVAL); error = vmmops_setreg(vm->cookie, vcpuid, reg, val); if (error || reg != VM_REG_GUEST_RIP) return (error); /* Set 'nextrip' to match the value of %rip */ VCPU_CTR1(vm, vcpuid, "Setting nextrip to %#lx", val); vcpu = &vm->vcpu[vcpuid]; vcpu->nextrip = val; return (0); } static bool is_descriptor_table(int reg) { switch (reg) { case VM_REG_GUEST_IDTR: case VM_REG_GUEST_GDTR: return (true); default: return (false); } } static bool is_segment_register(int reg) { switch (reg) { case VM_REG_GUEST_ES: case VM_REG_GUEST_CS: case VM_REG_GUEST_SS: case VM_REG_GUEST_DS: case VM_REG_GUEST_FS: case VM_REG_GUEST_GS: case VM_REG_GUEST_TR: case VM_REG_GUEST_LDTR: return (true); default: return (false); } } int vm_get_seg_desc(struct vm *vm, int vcpu, int reg, struct seg_desc *desc) { if (vcpu < 0 || vcpu >= vm->maxcpus) return (EINVAL); if (!is_segment_register(reg) && !is_descriptor_table(reg)) return (EINVAL); return (vmmops_getdesc(vm->cookie, vcpu, reg, desc)); } int vm_set_seg_desc(struct vm *vm, int vcpu, int reg, struct seg_desc *desc) { if (vcpu < 0 || vcpu >= vm->maxcpus) return (EINVAL); if (!is_segment_register(reg) && !is_descriptor_table(reg)) return (EINVAL); return (vmmops_setdesc(vm->cookie, vcpu, reg, desc)); } static void restore_guest_fpustate(struct vcpu *vcpu) { /* flush host state to the pcb */ fpuexit(curthread); /* restore guest FPU state */ fpu_stop_emulating(); fpurestore(vcpu->guestfpu); /* restore guest XCR0 if XSAVE is enabled in the host */ if (rcr4() & CR4_XSAVE) load_xcr(0, vcpu->guest_xcr0); /* * The FPU is now "dirty" with the guest's state so turn on emulation * to trap any access to the FPU by the host. */ fpu_start_emulating(); } static void save_guest_fpustate(struct vcpu *vcpu) { if ((rcr0() & CR0_TS) == 0) panic("fpu emulation not enabled in host!"); /* save guest XCR0 and restore host XCR0 */ if (rcr4() & CR4_XSAVE) { vcpu->guest_xcr0 = rxcr(0); load_xcr(0, vmm_get_host_xcr0()); } /* save guest FPU state */ fpu_stop_emulating(); fpusave(vcpu->guestfpu); fpu_start_emulating(); } static VMM_STAT(VCPU_IDLE_TICKS, "number of ticks vcpu was idle"); static int vcpu_set_state_locked(struct vm *vm, int vcpuid, enum vcpu_state newstate, bool from_idle) { struct vcpu *vcpu; int error; vcpu = &vm->vcpu[vcpuid]; vcpu_assert_locked(vcpu); /* * State transitions from the vmmdev_ioctl() must always begin from * the VCPU_IDLE state. This guarantees that there is only a single * ioctl() operating on a vcpu at any point. */ if (from_idle) { while (vcpu->state != VCPU_IDLE) { vcpu->reqidle = 1; vcpu_notify_event_locked(vcpu, false); VCPU_CTR1(vm, vcpuid, "vcpu state change from %s to " "idle requested", vcpu_state2str(vcpu->state)); msleep_spin(&vcpu->state, &vcpu->mtx, "vmstat", hz); } } else { KASSERT(vcpu->state != VCPU_IDLE, ("invalid transition from " "vcpu idle state")); } if (vcpu->state == VCPU_RUNNING) { KASSERT(vcpu->hostcpu == curcpu, ("curcpu %d and hostcpu %d " "mismatch for running vcpu", curcpu, vcpu->hostcpu)); } else { KASSERT(vcpu->hostcpu == NOCPU, ("Invalid hostcpu %d for a " "vcpu that is not running", vcpu->hostcpu)); } /* * The following state transitions are allowed: * IDLE -> FROZEN -> IDLE * FROZEN -> RUNNING -> FROZEN * FROZEN -> SLEEPING -> FROZEN */ switch (vcpu->state) { case VCPU_IDLE: case VCPU_RUNNING: case VCPU_SLEEPING: error = (newstate != VCPU_FROZEN); break; case VCPU_FROZEN: error = (newstate == VCPU_FROZEN); break; default: error = 1; break; } if (error) return (EBUSY); VCPU_CTR2(vm, vcpuid, "vcpu state changed from %s to %s", vcpu_state2str(vcpu->state), vcpu_state2str(newstate)); vcpu->state = newstate; if (newstate == VCPU_RUNNING) vcpu->hostcpu = curcpu; else vcpu->hostcpu = NOCPU; if (newstate == VCPU_IDLE) wakeup(&vcpu->state); return (0); } static void vcpu_require_state(struct vm *vm, int vcpuid, enum vcpu_state newstate) { int error; if ((error = vcpu_set_state(vm, vcpuid, newstate, false)) != 0) panic("Error %d setting state to %d\n", error, newstate); } static void vcpu_require_state_locked(struct vm *vm, int vcpuid, enum vcpu_state newstate) { int error; if ((error = vcpu_set_state_locked(vm, vcpuid, newstate, false)) != 0) panic("Error %d setting state to %d", error, newstate); } #define RENDEZVOUS_CTR0(vm, vcpuid, fmt) \ do { \ if (vcpuid >= 0) \ VCPU_CTR0(vm, vcpuid, fmt); \ else \ VM_CTR0(vm, fmt); \ } while (0) static int vm_handle_rendezvous(struct vm *vm, int vcpuid) { struct thread *td; int error; KASSERT(vcpuid == -1 || (vcpuid >= 0 && vcpuid < vm->maxcpus), ("vm_handle_rendezvous: invalid vcpuid %d", vcpuid)); error = 0; td = curthread; mtx_lock(&vm->rendezvous_mtx); while (vm->rendezvous_func != NULL) { /* 'rendezvous_req_cpus' must be a subset of 'active_cpus' */ - CPU_AND(&vm->rendezvous_req_cpus, &vm->active_cpus); + CPU_AND(&vm->rendezvous_req_cpus, &vm->rendezvous_req_cpus, &vm->active_cpus); if (vcpuid != -1 && CPU_ISSET(vcpuid, &vm->rendezvous_req_cpus) && !CPU_ISSET(vcpuid, &vm->rendezvous_done_cpus)) { VCPU_CTR0(vm, vcpuid, "Calling rendezvous func"); (*vm->rendezvous_func)(vm, vcpuid, vm->rendezvous_arg); CPU_SET(vcpuid, &vm->rendezvous_done_cpus); } if (CPU_CMP(&vm->rendezvous_req_cpus, &vm->rendezvous_done_cpus) == 0) { VCPU_CTR0(vm, vcpuid, "Rendezvous completed"); vm->rendezvous_func = NULL; wakeup(&vm->rendezvous_func); break; } RENDEZVOUS_CTR0(vm, vcpuid, "Wait for rendezvous completion"); mtx_sleep(&vm->rendezvous_func, &vm->rendezvous_mtx, 0, "vmrndv", hz); if ((td->td_flags & TDF_NEEDSUSPCHK) != 0) { mtx_unlock(&vm->rendezvous_mtx); error = thread_check_susp(td, true); if (error != 0) return (error); mtx_lock(&vm->rendezvous_mtx); } } mtx_unlock(&vm->rendezvous_mtx); return (0); } /* * Emulate a guest 'hlt' by sleeping until the vcpu is ready to run. */ static int vm_handle_hlt(struct vm *vm, int vcpuid, bool intr_disabled, bool *retu) { struct vcpu *vcpu; const char *wmesg; struct thread *td; int error, t, vcpu_halted, vm_halted; KASSERT(!CPU_ISSET(vcpuid, &vm->halted_cpus), ("vcpu already halted")); vcpu = &vm->vcpu[vcpuid]; vcpu_halted = 0; vm_halted = 0; error = 0; td = curthread; vcpu_lock(vcpu); while (1) { /* * Do a final check for pending NMI or interrupts before * really putting this thread to sleep. Also check for * software events that would cause this vcpu to wakeup. * * These interrupts/events could have happened after the * vcpu returned from vmmops_run() and before it acquired the * vcpu lock above. */ if (vm->rendezvous_func != NULL || vm->suspend || vcpu->reqidle) break; if (vm_nmi_pending(vm, vcpuid)) break; if (!intr_disabled) { if (vm_extint_pending(vm, vcpuid) || vlapic_pending_intr(vcpu->vlapic, NULL)) { break; } } /* Don't go to sleep if the vcpu thread needs to yield */ if (vcpu_should_yield(vm, vcpuid)) break; if (vcpu_debugged(vm, vcpuid)) break; /* * Some Linux guests implement "halt" by having all vcpus * execute HLT with interrupts disabled. 'halted_cpus' keeps * track of the vcpus that have entered this state. When all * vcpus enter the halted state the virtual machine is halted. */ if (intr_disabled) { wmesg = "vmhalt"; VCPU_CTR0(vm, vcpuid, "Halted"); if (!vcpu_halted && halt_detection_enabled) { vcpu_halted = 1; CPU_SET_ATOMIC(vcpuid, &vm->halted_cpus); } if (CPU_CMP(&vm->halted_cpus, &vm->active_cpus) == 0) { vm_halted = 1; break; } } else { wmesg = "vmidle"; } t = ticks; vcpu_require_state_locked(vm, vcpuid, VCPU_SLEEPING); /* * XXX msleep_spin() cannot be interrupted by signals so * wake up periodically to check pending signals. */ msleep_spin(vcpu, &vcpu->mtx, wmesg, hz); vcpu_require_state_locked(vm, vcpuid, VCPU_FROZEN); vmm_stat_incr(vm, vcpuid, VCPU_IDLE_TICKS, ticks - t); if ((td->td_flags & TDF_NEEDSUSPCHK) != 0) { vcpu_unlock(vcpu); error = thread_check_susp(td, false); if (error != 0) return (error); vcpu_lock(vcpu); } } if (vcpu_halted) CPU_CLR_ATOMIC(vcpuid, &vm->halted_cpus); vcpu_unlock(vcpu); if (vm_halted) vm_suspend(vm, VM_SUSPEND_HALT); return (0); } static int vm_handle_paging(struct vm *vm, int vcpuid, bool *retu) { int rv, ftype; struct vm_map *map; struct vcpu *vcpu; struct vm_exit *vme; vcpu = &vm->vcpu[vcpuid]; vme = &vcpu->exitinfo; KASSERT(vme->inst_length == 0, ("%s: invalid inst_length %d", __func__, vme->inst_length)); ftype = vme->u.paging.fault_type; KASSERT(ftype == VM_PROT_READ || ftype == VM_PROT_WRITE || ftype == VM_PROT_EXECUTE, ("vm_handle_paging: invalid fault_type %d", ftype)); if (ftype == VM_PROT_READ || ftype == VM_PROT_WRITE) { rv = pmap_emulate_accessed_dirty(vmspace_pmap(vm->vmspace), vme->u.paging.gpa, ftype); if (rv == 0) { VCPU_CTR2(vm, vcpuid, "%s bit emulation for gpa %#lx", ftype == VM_PROT_READ ? "accessed" : "dirty", vme->u.paging.gpa); goto done; } } map = &vm->vmspace->vm_map; rv = vm_fault(map, vme->u.paging.gpa, ftype, VM_FAULT_NORMAL, NULL); VCPU_CTR3(vm, vcpuid, "vm_handle_paging rv = %d, gpa = %#lx, " "ftype = %d", rv, vme->u.paging.gpa, ftype); if (rv != KERN_SUCCESS) return (EFAULT); done: return (0); } static int vm_handle_inst_emul(struct vm *vm, int vcpuid, bool *retu) { struct vie *vie; struct vcpu *vcpu; struct vm_exit *vme; uint64_t gla, gpa, cs_base; struct vm_guest_paging *paging; mem_region_read_t mread; mem_region_write_t mwrite; enum vm_cpu_mode cpu_mode; int cs_d, error, fault; vcpu = &vm->vcpu[vcpuid]; vme = &vcpu->exitinfo; KASSERT(vme->inst_length == 0, ("%s: invalid inst_length %d", __func__, vme->inst_length)); gla = vme->u.inst_emul.gla; gpa = vme->u.inst_emul.gpa; cs_base = vme->u.inst_emul.cs_base; cs_d = vme->u.inst_emul.cs_d; vie = &vme->u.inst_emul.vie; paging = &vme->u.inst_emul.paging; cpu_mode = paging->cpu_mode; VCPU_CTR1(vm, vcpuid, "inst_emul fault accessing gpa %#lx", gpa); /* Fetch, decode and emulate the faulting instruction */ if (vie->num_valid == 0) { error = vmm_fetch_instruction(vm, vcpuid, paging, vme->rip + cs_base, VIE_INST_SIZE, vie, &fault); } else { /* * The instruction bytes have already been copied into 'vie' */ error = fault = 0; } if (error || fault) return (error); if (vmm_decode_instruction(vm, vcpuid, gla, cpu_mode, cs_d, vie) != 0) { VCPU_CTR1(vm, vcpuid, "Error decoding instruction at %#lx", vme->rip + cs_base); *retu = true; /* dump instruction bytes in userspace */ return (0); } /* * Update 'nextrip' based on the length of the emulated instruction. */ vme->inst_length = vie->num_processed; vcpu->nextrip += vie->num_processed; VCPU_CTR1(vm, vcpuid, "nextrip updated to %#lx after instruction " "decoding", vcpu->nextrip); /* return to userland unless this is an in-kernel emulated device */ if (gpa >= DEFAULT_APIC_BASE && gpa < DEFAULT_APIC_BASE + PAGE_SIZE) { mread = lapic_mmio_read; mwrite = lapic_mmio_write; } else if (gpa >= VIOAPIC_BASE && gpa < VIOAPIC_BASE + VIOAPIC_SIZE) { mread = vioapic_mmio_read; mwrite = vioapic_mmio_write; } else if (gpa >= VHPET_BASE && gpa < VHPET_BASE + VHPET_SIZE) { mread = vhpet_mmio_read; mwrite = vhpet_mmio_write; } else { *retu = true; return (0); } error = vmm_emulate_instruction(vm, vcpuid, gpa, vie, paging, mread, mwrite, retu); return (error); } static int vm_handle_suspend(struct vm *vm, int vcpuid, bool *retu) { int error, i; struct vcpu *vcpu; struct thread *td; error = 0; vcpu = &vm->vcpu[vcpuid]; td = curthread; CPU_SET_ATOMIC(vcpuid, &vm->suspended_cpus); /* * Wait until all 'active_cpus' have suspended themselves. * * Since a VM may be suspended at any time including when one or * more vcpus are doing a rendezvous we need to call the rendezvous * handler while we are waiting to prevent a deadlock. */ vcpu_lock(vcpu); while (error == 0) { if (CPU_CMP(&vm->suspended_cpus, &vm->active_cpus) == 0) { VCPU_CTR0(vm, vcpuid, "All vcpus suspended"); break; } if (vm->rendezvous_func == NULL) { VCPU_CTR0(vm, vcpuid, "Sleeping during suspend"); vcpu_require_state_locked(vm, vcpuid, VCPU_SLEEPING); msleep_spin(vcpu, &vcpu->mtx, "vmsusp", hz); vcpu_require_state_locked(vm, vcpuid, VCPU_FROZEN); if ((td->td_flags & TDF_NEEDSUSPCHK) != 0) { vcpu_unlock(vcpu); error = thread_check_susp(td, false); vcpu_lock(vcpu); } } else { VCPU_CTR0(vm, vcpuid, "Rendezvous during suspend"); vcpu_unlock(vcpu); error = vm_handle_rendezvous(vm, vcpuid); vcpu_lock(vcpu); } } vcpu_unlock(vcpu); /* * Wakeup the other sleeping vcpus and return to userspace. */ for (i = 0; i < vm->maxcpus; i++) { if (CPU_ISSET(i, &vm->suspended_cpus)) { vcpu_notify_event(vm, i, false); } } *retu = true; return (error); } static int vm_handle_reqidle(struct vm *vm, int vcpuid, bool *retu) { struct vcpu *vcpu = &vm->vcpu[vcpuid]; vcpu_lock(vcpu); KASSERT(vcpu->reqidle, ("invalid vcpu reqidle %d", vcpu->reqidle)); vcpu->reqidle = 0; vcpu_unlock(vcpu); *retu = true; return (0); } int vm_suspend(struct vm *vm, enum vm_suspend_how how) { int i; if (how <= VM_SUSPEND_NONE || how >= VM_SUSPEND_LAST) return (EINVAL); if (atomic_cmpset_int(&vm->suspend, 0, how) == 0) { VM_CTR2(vm, "virtual machine already suspended %d/%d", vm->suspend, how); return (EALREADY); } VM_CTR1(vm, "virtual machine successfully suspended %d", how); /* * Notify all active vcpus that they are now suspended. */ for (i = 0; i < vm->maxcpus; i++) { if (CPU_ISSET(i, &vm->active_cpus)) vcpu_notify_event(vm, i, false); } return (0); } void vm_exit_suspended(struct vm *vm, int vcpuid, uint64_t rip) { struct vm_exit *vmexit; KASSERT(vm->suspend > VM_SUSPEND_NONE && vm->suspend < VM_SUSPEND_LAST, ("vm_exit_suspended: invalid suspend type %d", vm->suspend)); vmexit = vm_exitinfo(vm, vcpuid); vmexit->rip = rip; vmexit->inst_length = 0; vmexit->exitcode = VM_EXITCODE_SUSPENDED; vmexit->u.suspended.how = vm->suspend; } void vm_exit_debug(struct vm *vm, int vcpuid, uint64_t rip) { struct vm_exit *vmexit; vmexit = vm_exitinfo(vm, vcpuid); vmexit->rip = rip; vmexit->inst_length = 0; vmexit->exitcode = VM_EXITCODE_DEBUG; } void vm_exit_rendezvous(struct vm *vm, int vcpuid, uint64_t rip) { struct vm_exit *vmexit; KASSERT(vm->rendezvous_func != NULL, ("rendezvous not in progress")); vmexit = vm_exitinfo(vm, vcpuid); vmexit->rip = rip; vmexit->inst_length = 0; vmexit->exitcode = VM_EXITCODE_RENDEZVOUS; vmm_stat_incr(vm, vcpuid, VMEXIT_RENDEZVOUS, 1); } void vm_exit_reqidle(struct vm *vm, int vcpuid, uint64_t rip) { struct vm_exit *vmexit; vmexit = vm_exitinfo(vm, vcpuid); vmexit->rip = rip; vmexit->inst_length = 0; vmexit->exitcode = VM_EXITCODE_REQIDLE; vmm_stat_incr(vm, vcpuid, VMEXIT_REQIDLE, 1); } void vm_exit_astpending(struct vm *vm, int vcpuid, uint64_t rip) { struct vm_exit *vmexit; vmexit = vm_exitinfo(vm, vcpuid); vmexit->rip = rip; vmexit->inst_length = 0; vmexit->exitcode = VM_EXITCODE_BOGUS; vmm_stat_incr(vm, vcpuid, VMEXIT_ASTPENDING, 1); } int vm_run(struct vm *vm, struct vm_run *vmrun) { struct vm_eventinfo evinfo; int error, vcpuid; struct vcpu *vcpu; struct pcb *pcb; uint64_t tscval; struct vm_exit *vme; bool retu, intr_disabled; pmap_t pmap; vcpuid = vmrun->cpuid; if (vcpuid < 0 || vcpuid >= vm->maxcpus) return (EINVAL); if (!CPU_ISSET(vcpuid, &vm->active_cpus)) return (EINVAL); if (CPU_ISSET(vcpuid, &vm->suspended_cpus)) return (EINVAL); pmap = vmspace_pmap(vm->vmspace); vcpu = &vm->vcpu[vcpuid]; vme = &vcpu->exitinfo; evinfo.rptr = &vm->rendezvous_func; evinfo.sptr = &vm->suspend; evinfo.iptr = &vcpu->reqidle; restart: critical_enter(); KASSERT(!CPU_ISSET(curcpu, &pmap->pm_active), ("vm_run: absurd pm_active")); tscval = rdtsc(); pcb = PCPU_GET(curpcb); set_pcb_flags(pcb, PCB_FULL_IRET); restore_guest_fpustate(vcpu); vcpu_require_state(vm, vcpuid, VCPU_RUNNING); error = vmmops_run(vm->cookie, vcpuid, vcpu->nextrip, pmap, &evinfo); vcpu_require_state(vm, vcpuid, VCPU_FROZEN); save_guest_fpustate(vcpu); vmm_stat_incr(vm, vcpuid, VCPU_TOTAL_RUNTIME, rdtsc() - tscval); critical_exit(); if (error == 0) { retu = false; vcpu->nextrip = vme->rip + vme->inst_length; switch (vme->exitcode) { case VM_EXITCODE_REQIDLE: error = vm_handle_reqidle(vm, vcpuid, &retu); break; case VM_EXITCODE_SUSPENDED: error = vm_handle_suspend(vm, vcpuid, &retu); break; case VM_EXITCODE_IOAPIC_EOI: vioapic_process_eoi(vm, vcpuid, vme->u.ioapic_eoi.vector); break; case VM_EXITCODE_RENDEZVOUS: error = vm_handle_rendezvous(vm, vcpuid); break; case VM_EXITCODE_HLT: intr_disabled = ((vme->u.hlt.rflags & PSL_I) == 0); error = vm_handle_hlt(vm, vcpuid, intr_disabled, &retu); break; case VM_EXITCODE_PAGING: error = vm_handle_paging(vm, vcpuid, &retu); break; case VM_EXITCODE_INST_EMUL: error = vm_handle_inst_emul(vm, vcpuid, &retu); break; case VM_EXITCODE_INOUT: case VM_EXITCODE_INOUT_STR: error = vm_handle_inout(vm, vcpuid, vme, &retu); break; case VM_EXITCODE_MONITOR: case VM_EXITCODE_MWAIT: case VM_EXITCODE_VMINSN: vm_inject_ud(vm, vcpuid); break; default: retu = true; /* handled in userland */ break; } } if (error == 0 && retu == false) goto restart; VCPU_CTR2(vm, vcpuid, "retu %d/%d", error, vme->exitcode); /* copy the exit information */ bcopy(vme, &vmrun->vm_exit, sizeof(struct vm_exit)); return (error); } int vm_restart_instruction(void *arg, int vcpuid) { struct vm *vm; struct vcpu *vcpu; enum vcpu_state state; uint64_t rip; int error; vm = arg; if (vcpuid < 0 || vcpuid >= vm->maxcpus) return (EINVAL); vcpu = &vm->vcpu[vcpuid]; state = vcpu_get_state(vm, vcpuid, NULL); if (state == VCPU_RUNNING) { /* * When a vcpu is "running" the next instruction is determined * by adding 'rip' and 'inst_length' in the vcpu's 'exitinfo'. * Thus setting 'inst_length' to zero will cause the current * instruction to be restarted. */ vcpu->exitinfo.inst_length = 0; VCPU_CTR1(vm, vcpuid, "restarting instruction at %#lx by " "setting inst_length to zero", vcpu->exitinfo.rip); } else if (state == VCPU_FROZEN) { /* * When a vcpu is "frozen" it is outside the critical section * around vmmops_run() and 'nextrip' points to the next * instruction. Thus instruction restart is achieved by setting * 'nextrip' to the vcpu's %rip. */ error = vm_get_register(vm, vcpuid, VM_REG_GUEST_RIP, &rip); KASSERT(!error, ("%s: error %d getting rip", __func__, error)); VCPU_CTR2(vm, vcpuid, "restarting instruction by updating " "nextrip from %#lx to %#lx", vcpu->nextrip, rip); vcpu->nextrip = rip; } else { panic("%s: invalid state %d", __func__, state); } return (0); } int vm_exit_intinfo(struct vm *vm, int vcpuid, uint64_t info) { struct vcpu *vcpu; int type, vector; if (vcpuid < 0 || vcpuid >= vm->maxcpus) return (EINVAL); vcpu = &vm->vcpu[vcpuid]; if (info & VM_INTINFO_VALID) { type = info & VM_INTINFO_TYPE; vector = info & 0xff; if (type == VM_INTINFO_NMI && vector != IDT_NMI) return (EINVAL); if (type == VM_INTINFO_HWEXCEPTION && vector >= 32) return (EINVAL); if (info & VM_INTINFO_RSVD) return (EINVAL); } else { info = 0; } VCPU_CTR2(vm, vcpuid, "%s: info1(%#lx)", __func__, info); vcpu->exitintinfo = info; return (0); } enum exc_class { EXC_BENIGN, EXC_CONTRIBUTORY, EXC_PAGEFAULT }; #define IDT_VE 20 /* Virtualization Exception (Intel specific) */ static enum exc_class exception_class(uint64_t info) { int type, vector; KASSERT(info & VM_INTINFO_VALID, ("intinfo must be valid: %#lx", info)); type = info & VM_INTINFO_TYPE; vector = info & 0xff; /* Table 6-4, "Interrupt and Exception Classes", Intel SDM, Vol 3 */ switch (type) { case VM_INTINFO_HWINTR: case VM_INTINFO_SWINTR: case VM_INTINFO_NMI: return (EXC_BENIGN); default: /* * Hardware exception. * * SVM and VT-x use identical type values to represent NMI, * hardware interrupt and software interrupt. * * SVM uses type '3' for all exceptions. VT-x uses type '3' * for exceptions except #BP and #OF. #BP and #OF use a type * value of '5' or '6'. Therefore we don't check for explicit * values of 'type' to classify 'intinfo' into a hardware * exception. */ break; } switch (vector) { case IDT_PF: case IDT_VE: return (EXC_PAGEFAULT); case IDT_DE: case IDT_TS: case IDT_NP: case IDT_SS: case IDT_GP: return (EXC_CONTRIBUTORY); default: return (EXC_BENIGN); } } static int nested_fault(struct vm *vm, int vcpuid, uint64_t info1, uint64_t info2, uint64_t *retinfo) { enum exc_class exc1, exc2; int type1, vector1; KASSERT(info1 & VM_INTINFO_VALID, ("info1 %#lx is not valid", info1)); KASSERT(info2 & VM_INTINFO_VALID, ("info2 %#lx is not valid", info2)); /* * If an exception occurs while attempting to call the double-fault * handler the processor enters shutdown mode (aka triple fault). */ type1 = info1 & VM_INTINFO_TYPE; vector1 = info1 & 0xff; if (type1 == VM_INTINFO_HWEXCEPTION && vector1 == IDT_DF) { VCPU_CTR2(vm, vcpuid, "triple fault: info1(%#lx), info2(%#lx)", info1, info2); vm_suspend(vm, VM_SUSPEND_TRIPLEFAULT); *retinfo = 0; return (0); } /* * Table 6-5 "Conditions for Generating a Double Fault", Intel SDM, Vol3 */ exc1 = exception_class(info1); exc2 = exception_class(info2); if ((exc1 == EXC_CONTRIBUTORY && exc2 == EXC_CONTRIBUTORY) || (exc1 == EXC_PAGEFAULT && exc2 != EXC_BENIGN)) { /* Convert nested fault into a double fault. */ *retinfo = IDT_DF; *retinfo |= VM_INTINFO_VALID | VM_INTINFO_HWEXCEPTION; *retinfo |= VM_INTINFO_DEL_ERRCODE; } else { /* Handle exceptions serially */ *retinfo = info2; } return (1); } static uint64_t vcpu_exception_intinfo(struct vcpu *vcpu) { uint64_t info = 0; if (vcpu->exception_pending) { info = vcpu->exc_vector & 0xff; info |= VM_INTINFO_VALID | VM_INTINFO_HWEXCEPTION; if (vcpu->exc_errcode_valid) { info |= VM_INTINFO_DEL_ERRCODE; info |= (uint64_t)vcpu->exc_errcode << 32; } } return (info); } int vm_entry_intinfo(struct vm *vm, int vcpuid, uint64_t *retinfo) { struct vcpu *vcpu; uint64_t info1, info2; int valid; KASSERT(vcpuid >= 0 && vcpuid < vm->maxcpus, ("invalid vcpu %d", vcpuid)); vcpu = &vm->vcpu[vcpuid]; info1 = vcpu->exitintinfo; vcpu->exitintinfo = 0; info2 = 0; if (vcpu->exception_pending) { info2 = vcpu_exception_intinfo(vcpu); vcpu->exception_pending = 0; VCPU_CTR2(vm, vcpuid, "Exception %d delivered: %#lx", vcpu->exc_vector, info2); } if ((info1 & VM_INTINFO_VALID) && (info2 & VM_INTINFO_VALID)) { valid = nested_fault(vm, vcpuid, info1, info2, retinfo); } else if (info1 & VM_INTINFO_VALID) { *retinfo = info1; valid = 1; } else if (info2 & VM_INTINFO_VALID) { *retinfo = info2; valid = 1; } else { valid = 0; } if (valid) { VCPU_CTR4(vm, vcpuid, "%s: info1(%#lx), info2(%#lx), " "retinfo(%#lx)", __func__, info1, info2, *retinfo); } return (valid); } int vm_get_intinfo(struct vm *vm, int vcpuid, uint64_t *info1, uint64_t *info2) { struct vcpu *vcpu; if (vcpuid < 0 || vcpuid >= vm->maxcpus) return (EINVAL); vcpu = &vm->vcpu[vcpuid]; *info1 = vcpu->exitintinfo; *info2 = vcpu_exception_intinfo(vcpu); return (0); } int vm_inject_exception(struct vm *vm, int vcpuid, int vector, int errcode_valid, uint32_t errcode, int restart_instruction) { struct vcpu *vcpu; uint64_t regval; int error; if (vcpuid < 0 || vcpuid >= vm->maxcpus) return (EINVAL); if (vector < 0 || vector >= 32) return (EINVAL); /* * A double fault exception should never be injected directly into * the guest. It is a derived exception that results from specific * combinations of nested faults. */ if (vector == IDT_DF) return (EINVAL); vcpu = &vm->vcpu[vcpuid]; if (vcpu->exception_pending) { VCPU_CTR2(vm, vcpuid, "Unable to inject exception %d due to " "pending exception %d", vector, vcpu->exc_vector); return (EBUSY); } if (errcode_valid) { /* * Exceptions don't deliver an error code in real mode. */ error = vm_get_register(vm, vcpuid, VM_REG_GUEST_CR0, ®val); KASSERT(!error, ("%s: error %d getting CR0", __func__, error)); if (!(regval & CR0_PE)) errcode_valid = 0; } /* * From section 26.6.1 "Interruptibility State" in Intel SDM: * * Event blocking by "STI" or "MOV SS" is cleared after guest executes * one instruction or incurs an exception. */ error = vm_set_register(vm, vcpuid, VM_REG_GUEST_INTR_SHADOW, 0); KASSERT(error == 0, ("%s: error %d clearing interrupt shadow", __func__, error)); if (restart_instruction) vm_restart_instruction(vm, vcpuid); vcpu->exception_pending = 1; vcpu->exc_vector = vector; vcpu->exc_errcode = errcode; vcpu->exc_errcode_valid = errcode_valid; VCPU_CTR1(vm, vcpuid, "Exception %d pending", vector); return (0); } void vm_inject_fault(void *vmarg, int vcpuid, int vector, int errcode_valid, int errcode) { struct vm *vm; int error, restart_instruction; vm = vmarg; restart_instruction = 1; error = vm_inject_exception(vm, vcpuid, vector, errcode_valid, errcode, restart_instruction); KASSERT(error == 0, ("vm_inject_exception error %d", error)); } void vm_inject_pf(void *vmarg, int vcpuid, int error_code, uint64_t cr2) { struct vm *vm; int error; vm = vmarg; VCPU_CTR2(vm, vcpuid, "Injecting page fault: error_code %#x, cr2 %#lx", error_code, cr2); error = vm_set_register(vm, vcpuid, VM_REG_GUEST_CR2, cr2); KASSERT(error == 0, ("vm_set_register(cr2) error %d", error)); vm_inject_fault(vm, vcpuid, IDT_PF, 1, error_code); } static VMM_STAT(VCPU_NMI_COUNT, "number of NMIs delivered to vcpu"); int vm_inject_nmi(struct vm *vm, int vcpuid) { struct vcpu *vcpu; if (vcpuid < 0 || vcpuid >= vm->maxcpus) return (EINVAL); vcpu = &vm->vcpu[vcpuid]; vcpu->nmi_pending = 1; vcpu_notify_event(vm, vcpuid, false); return (0); } int vm_nmi_pending(struct vm *vm, int vcpuid) { struct vcpu *vcpu; if (vcpuid < 0 || vcpuid >= vm->maxcpus) panic("vm_nmi_pending: invalid vcpuid %d", vcpuid); vcpu = &vm->vcpu[vcpuid]; return (vcpu->nmi_pending); } void vm_nmi_clear(struct vm *vm, int vcpuid) { struct vcpu *vcpu; if (vcpuid < 0 || vcpuid >= vm->maxcpus) panic("vm_nmi_pending: invalid vcpuid %d", vcpuid); vcpu = &vm->vcpu[vcpuid]; if (vcpu->nmi_pending == 0) panic("vm_nmi_clear: inconsistent nmi_pending state"); vcpu->nmi_pending = 0; vmm_stat_incr(vm, vcpuid, VCPU_NMI_COUNT, 1); } static VMM_STAT(VCPU_EXTINT_COUNT, "number of ExtINTs delivered to vcpu"); int vm_inject_extint(struct vm *vm, int vcpuid) { struct vcpu *vcpu; if (vcpuid < 0 || vcpuid >= vm->maxcpus) return (EINVAL); vcpu = &vm->vcpu[vcpuid]; vcpu->extint_pending = 1; vcpu_notify_event(vm, vcpuid, false); return (0); } int vm_extint_pending(struct vm *vm, int vcpuid) { struct vcpu *vcpu; if (vcpuid < 0 || vcpuid >= vm->maxcpus) panic("vm_extint_pending: invalid vcpuid %d", vcpuid); vcpu = &vm->vcpu[vcpuid]; return (vcpu->extint_pending); } void vm_extint_clear(struct vm *vm, int vcpuid) { struct vcpu *vcpu; if (vcpuid < 0 || vcpuid >= vm->maxcpus) panic("vm_extint_pending: invalid vcpuid %d", vcpuid); vcpu = &vm->vcpu[vcpuid]; if (vcpu->extint_pending == 0) panic("vm_extint_clear: inconsistent extint_pending state"); vcpu->extint_pending = 0; vmm_stat_incr(vm, vcpuid, VCPU_EXTINT_COUNT, 1); } int vm_get_capability(struct vm *vm, int vcpu, int type, int *retval) { if (vcpu < 0 || vcpu >= vm->maxcpus) return (EINVAL); if (type < 0 || type >= VM_CAP_MAX) return (EINVAL); return (vmmops_getcap(vm->cookie, vcpu, type, retval)); } int vm_set_capability(struct vm *vm, int vcpu, int type, int val) { if (vcpu < 0 || vcpu >= vm->maxcpus) return (EINVAL); if (type < 0 || type >= VM_CAP_MAX) return (EINVAL); return (vmmops_setcap(vm->cookie, vcpu, type, val)); } struct vlapic * vm_lapic(struct vm *vm, int cpu) { return (vm->vcpu[cpu].vlapic); } struct vioapic * vm_ioapic(struct vm *vm) { return (vm->vioapic); } struct vhpet * vm_hpet(struct vm *vm) { return (vm->vhpet); } bool vmm_is_pptdev(int bus, int slot, int func) { int b, f, i, n, s; char *val, *cp, *cp2; bool found; /* * XXX * The length of an environment variable is limited to 128 bytes which * puts an upper limit on the number of passthru devices that may be * specified using a single environment variable. * * Work around this by scanning multiple environment variable * names instead of a single one - yuck! */ const char *names[] = { "pptdevs", "pptdevs2", "pptdevs3", NULL }; /* set pptdevs="1/2/3 4/5/6 7/8/9 10/11/12" */ found = false; for (i = 0; names[i] != NULL && !found; i++) { cp = val = kern_getenv(names[i]); while (cp != NULL && *cp != '\0') { if ((cp2 = strchr(cp, ' ')) != NULL) *cp2 = '\0'; n = sscanf(cp, "%d/%d/%d", &b, &s, &f); if (n == 3 && bus == b && slot == s && func == f) { found = true; break; } if (cp2 != NULL) *cp2++ = ' '; cp = cp2; } freeenv(val); } return (found); } void * vm_iommu_domain(struct vm *vm) { return (vm->iommu); } int vcpu_set_state(struct vm *vm, int vcpuid, enum vcpu_state newstate, bool from_idle) { int error; struct vcpu *vcpu; if (vcpuid < 0 || vcpuid >= vm->maxcpus) panic("vm_set_run_state: invalid vcpuid %d", vcpuid); vcpu = &vm->vcpu[vcpuid]; vcpu_lock(vcpu); error = vcpu_set_state_locked(vm, vcpuid, newstate, from_idle); vcpu_unlock(vcpu); return (error); } enum vcpu_state vcpu_get_state(struct vm *vm, int vcpuid, int *hostcpu) { struct vcpu *vcpu; enum vcpu_state state; if (vcpuid < 0 || vcpuid >= vm->maxcpus) panic("vm_get_run_state: invalid vcpuid %d", vcpuid); vcpu = &vm->vcpu[vcpuid]; vcpu_lock(vcpu); state = vcpu->state; if (hostcpu != NULL) *hostcpu = vcpu->hostcpu; vcpu_unlock(vcpu); return (state); } int vm_activate_cpu(struct vm *vm, int vcpuid) { if (vcpuid < 0 || vcpuid >= vm->maxcpus) return (EINVAL); if (CPU_ISSET(vcpuid, &vm->active_cpus)) return (EBUSY); VCPU_CTR0(vm, vcpuid, "activated"); CPU_SET_ATOMIC(vcpuid, &vm->active_cpus); return (0); } int vm_suspend_cpu(struct vm *vm, int vcpuid) { int i; if (vcpuid < -1 || vcpuid >= vm->maxcpus) return (EINVAL); if (vcpuid == -1) { vm->debug_cpus = vm->active_cpus; for (i = 0; i < vm->maxcpus; i++) { if (CPU_ISSET(i, &vm->active_cpus)) vcpu_notify_event(vm, i, false); } } else { if (!CPU_ISSET(vcpuid, &vm->active_cpus)) return (EINVAL); CPU_SET_ATOMIC(vcpuid, &vm->debug_cpus); vcpu_notify_event(vm, vcpuid, false); } return (0); } int vm_resume_cpu(struct vm *vm, int vcpuid) { if (vcpuid < -1 || vcpuid >= vm->maxcpus) return (EINVAL); if (vcpuid == -1) { CPU_ZERO(&vm->debug_cpus); } else { if (!CPU_ISSET(vcpuid, &vm->debug_cpus)) return (EINVAL); CPU_CLR_ATOMIC(vcpuid, &vm->debug_cpus); } return (0); } int vcpu_debugged(struct vm *vm, int vcpuid) { return (CPU_ISSET(vcpuid, &vm->debug_cpus)); } cpuset_t vm_active_cpus(struct vm *vm) { return (vm->active_cpus); } cpuset_t vm_debug_cpus(struct vm *vm) { return (vm->debug_cpus); } cpuset_t vm_suspended_cpus(struct vm *vm) { return (vm->suspended_cpus); } void * vcpu_stats(struct vm *vm, int vcpuid) { return (vm->vcpu[vcpuid].stats); } int vm_get_x2apic_state(struct vm *vm, int vcpuid, enum x2apic_state *state) { if (vcpuid < 0 || vcpuid >= vm->maxcpus) return (EINVAL); *state = vm->vcpu[vcpuid].x2apic_state; return (0); } int vm_set_x2apic_state(struct vm *vm, int vcpuid, enum x2apic_state state) { if (vcpuid < 0 || vcpuid >= vm->maxcpus) return (EINVAL); if (state >= X2APIC_STATE_LAST) return (EINVAL); vm->vcpu[vcpuid].x2apic_state = state; vlapic_set_x2apic_state(vm, vcpuid, state); return (0); } /* * This function is called to ensure that a vcpu "sees" a pending event * as soon as possible: * - If the vcpu thread is sleeping then it is woken up. * - If the vcpu is running on a different host_cpu then an IPI will be directed * to the host_cpu to cause the vcpu to trap into the hypervisor. */ static void vcpu_notify_event_locked(struct vcpu *vcpu, bool lapic_intr) { int hostcpu; hostcpu = vcpu->hostcpu; if (vcpu->state == VCPU_RUNNING) { KASSERT(hostcpu != NOCPU, ("vcpu running on invalid hostcpu")); if (hostcpu != curcpu) { if (lapic_intr) { vlapic_post_intr(vcpu->vlapic, hostcpu, vmm_ipinum); } else { ipi_cpu(hostcpu, vmm_ipinum); } } else { /* * If the 'vcpu' is running on 'curcpu' then it must * be sending a notification to itself (e.g. SELF_IPI). * The pending event will be picked up when the vcpu * transitions back to guest context. */ } } else { KASSERT(hostcpu == NOCPU, ("vcpu state %d not consistent " "with hostcpu %d", vcpu->state, hostcpu)); if (vcpu->state == VCPU_SLEEPING) wakeup_one(vcpu); } } void vcpu_notify_event(struct vm *vm, int vcpuid, bool lapic_intr) { struct vcpu *vcpu = &vm->vcpu[vcpuid]; vcpu_lock(vcpu); vcpu_notify_event_locked(vcpu, lapic_intr); vcpu_unlock(vcpu); } struct vmspace * vm_get_vmspace(struct vm *vm) { return (vm->vmspace); } int vm_apicid2vcpuid(struct vm *vm, int apicid) { /* * XXX apic id is assumed to be numerically identical to vcpu id */ return (apicid); } int vm_smp_rendezvous(struct vm *vm, int vcpuid, cpuset_t dest, vm_rendezvous_func_t func, void *arg) { int error, i; /* * Enforce that this function is called without any locks */ WITNESS_WARN(WARN_PANIC, NULL, "vm_smp_rendezvous"); KASSERT(vcpuid == -1 || (vcpuid >= 0 && vcpuid < vm->maxcpus), ("vm_smp_rendezvous: invalid vcpuid %d", vcpuid)); restart: mtx_lock(&vm->rendezvous_mtx); if (vm->rendezvous_func != NULL) { /* * If a rendezvous is already in progress then we need to * call the rendezvous handler in case this 'vcpuid' is one * of the targets of the rendezvous. */ RENDEZVOUS_CTR0(vm, vcpuid, "Rendezvous already in progress"); mtx_unlock(&vm->rendezvous_mtx); error = vm_handle_rendezvous(vm, vcpuid); if (error != 0) return (error); goto restart; } KASSERT(vm->rendezvous_func == NULL, ("vm_smp_rendezvous: previous " "rendezvous is still in progress")); RENDEZVOUS_CTR0(vm, vcpuid, "Initiating rendezvous"); vm->rendezvous_req_cpus = dest; CPU_ZERO(&vm->rendezvous_done_cpus); vm->rendezvous_arg = arg; vm->rendezvous_func = func; mtx_unlock(&vm->rendezvous_mtx); /* * Wake up any sleeping vcpus and trigger a VM-exit in any running * vcpus so they handle the rendezvous as soon as possible. */ for (i = 0; i < vm->maxcpus; i++) { if (CPU_ISSET(i, &dest)) vcpu_notify_event(vm, i, false); } return (vm_handle_rendezvous(vm, vcpuid)); } struct vatpic * vm_atpic(struct vm *vm) { return (vm->vatpic); } struct vatpit * vm_atpit(struct vm *vm) { return (vm->vatpit); } struct vpmtmr * vm_pmtmr(struct vm *vm) { return (vm->vpmtmr); } struct vrtc * vm_rtc(struct vm *vm) { return (vm->vrtc); } enum vm_reg_name vm_segment_name(int seg) { static enum vm_reg_name seg_names[] = { VM_REG_GUEST_ES, VM_REG_GUEST_CS, VM_REG_GUEST_SS, VM_REG_GUEST_DS, VM_REG_GUEST_FS, VM_REG_GUEST_GS }; KASSERT(seg >= 0 && seg < nitems(seg_names), ("%s: invalid segment encoding %d", __func__, seg)); return (seg_names[seg]); } void vm_copy_teardown(struct vm *vm, int vcpuid, struct vm_copyinfo *copyinfo, int num_copyinfo) { int idx; for (idx = 0; idx < num_copyinfo; idx++) { if (copyinfo[idx].cookie != NULL) vm_gpa_release(copyinfo[idx].cookie); } bzero(copyinfo, num_copyinfo * sizeof(struct vm_copyinfo)); } int vm_copy_setup(struct vm *vm, int vcpuid, struct vm_guest_paging *paging, uint64_t gla, size_t len, int prot, struct vm_copyinfo *copyinfo, int num_copyinfo, int *fault) { int error, idx, nused; size_t n, off, remaining; void *hva, *cookie; uint64_t gpa; bzero(copyinfo, sizeof(struct vm_copyinfo) * num_copyinfo); nused = 0; remaining = len; while (remaining > 0) { KASSERT(nused < num_copyinfo, ("insufficient vm_copyinfo")); error = vm_gla2gpa(vm, vcpuid, paging, gla, prot, &gpa, fault); if (error || *fault) return (error); off = gpa & PAGE_MASK; n = min(remaining, PAGE_SIZE - off); copyinfo[nused].gpa = gpa; copyinfo[nused].len = n; remaining -= n; gla += n; nused++; } for (idx = 0; idx < nused; idx++) { hva = vm_gpa_hold(vm, vcpuid, copyinfo[idx].gpa, copyinfo[idx].len, prot, &cookie); if (hva == NULL) break; copyinfo[idx].hva = hva; copyinfo[idx].cookie = cookie; } if (idx != nused) { vm_copy_teardown(vm, vcpuid, copyinfo, num_copyinfo); return (EFAULT); } else { *fault = 0; return (0); } } void vm_copyin(struct vm *vm, int vcpuid, struct vm_copyinfo *copyinfo, void *kaddr, size_t len) { char *dst; int idx; dst = kaddr; idx = 0; while (len > 0) { bcopy(copyinfo[idx].hva, dst, copyinfo[idx].len); len -= copyinfo[idx].len; dst += copyinfo[idx].len; idx++; } } void vm_copyout(struct vm *vm, int vcpuid, const void *kaddr, struct vm_copyinfo *copyinfo, size_t len) { const char *src; int idx; src = kaddr; idx = 0; while (len > 0) { bcopy(src, copyinfo[idx].hva, copyinfo[idx].len); len -= copyinfo[idx].len; src += copyinfo[idx].len; idx++; } } /* * Return the amount of in-use and wired memory for the VM. Since * these are global stats, only return the values with for vCPU 0 */ VMM_STAT_DECLARE(VMM_MEM_RESIDENT); VMM_STAT_DECLARE(VMM_MEM_WIRED); static void vm_get_rescnt(struct vm *vm, int vcpu, struct vmm_stat_type *stat) { if (vcpu == 0) { vmm_stat_set(vm, vcpu, VMM_MEM_RESIDENT, PAGE_SIZE * vmspace_resident_count(vm->vmspace)); } } static void vm_get_wiredcnt(struct vm *vm, int vcpu, struct vmm_stat_type *stat) { if (vcpu == 0) { vmm_stat_set(vm, vcpu, VMM_MEM_WIRED, PAGE_SIZE * pmap_wired_count(vmspace_pmap(vm->vmspace))); } } VMM_STAT_FUNC(VMM_MEM_RESIDENT, "Resident memory", vm_get_rescnt); VMM_STAT_FUNC(VMM_MEM_WIRED, "Wired memory", vm_get_wiredcnt); #ifdef BHYVE_SNAPSHOT static int vm_snapshot_vcpus(struct vm *vm, struct vm_snapshot_meta *meta) { int ret; int i; struct vcpu *vcpu; for (i = 0; i < VM_MAXCPU; i++) { vcpu = &vm->vcpu[i]; SNAPSHOT_VAR_OR_LEAVE(vcpu->x2apic_state, meta, ret, done); SNAPSHOT_VAR_OR_LEAVE(vcpu->exitintinfo, meta, ret, done); SNAPSHOT_VAR_OR_LEAVE(vcpu->exc_vector, meta, ret, done); SNAPSHOT_VAR_OR_LEAVE(vcpu->exc_errcode_valid, meta, ret, done); SNAPSHOT_VAR_OR_LEAVE(vcpu->exc_errcode, meta, ret, done); SNAPSHOT_VAR_OR_LEAVE(vcpu->guest_xcr0, meta, ret, done); SNAPSHOT_VAR_OR_LEAVE(vcpu->exitinfo, meta, ret, done); SNAPSHOT_VAR_OR_LEAVE(vcpu->nextrip, meta, ret, done); /* XXX we're cheating here, since the value of tsc_offset as * saved here is actually the value of the guest's TSC value. * * It will be turned turned back into an actual offset when the * TSC restore function is called */ SNAPSHOT_VAR_OR_LEAVE(vcpu->tsc_offset, meta, ret, done); } done: return (ret); } static int vm_snapshot_vm(struct vm *vm, struct vm_snapshot_meta *meta) { int ret; int i; uint64_t now; ret = 0; now = rdtsc(); if (meta->op == VM_SNAPSHOT_SAVE) { /* XXX make tsc_offset take the value TSC proper as seen by the * guest */ for (i = 0; i < VM_MAXCPU; i++) vm->vcpu[i].tsc_offset += now; } ret = vm_snapshot_vcpus(vm, meta); if (ret != 0) { printf("%s: failed to copy vm data to user buffer", __func__); goto done; } if (meta->op == VM_SNAPSHOT_SAVE) { /* XXX turn tsc_offset back into an offset; actual value is only * required for restore; using it otherwise would be wrong */ for (i = 0; i < VM_MAXCPU; i++) vm->vcpu[i].tsc_offset -= now; } done: return (ret); } static int vm_snapshot_vmcx(struct vm *vm, struct vm_snapshot_meta *meta) { int i, error; error = 0; for (i = 0; i < VM_MAXCPU; i++) { error = vmmops_vmcx_snapshot(vm->cookie, meta, i); if (error != 0) { printf("%s: failed to snapshot vmcs/vmcb data for " "vCPU: %d; error: %d\n", __func__, i, error); goto done; } } done: return (error); } /* * Save kernel-side structures to user-space for snapshotting. */ int vm_snapshot_req(struct vm *vm, struct vm_snapshot_meta *meta) { int ret = 0; switch (meta->dev_req) { case STRUCT_VMX: ret = vmmops_snapshot(vm->cookie, meta); break; case STRUCT_VMCX: ret = vm_snapshot_vmcx(vm, meta); break; case STRUCT_VM: ret = vm_snapshot_vm(vm, meta); break; case STRUCT_VIOAPIC: ret = vioapic_snapshot(vm_ioapic(vm), meta); break; case STRUCT_VLAPIC: ret = vlapic_snapshot(vm, meta); break; case STRUCT_VHPET: ret = vhpet_snapshot(vm_hpet(vm), meta); break; case STRUCT_VATPIC: ret = vatpic_snapshot(vm_atpic(vm), meta); break; case STRUCT_VATPIT: ret = vatpit_snapshot(vm_atpit(vm), meta); break; case STRUCT_VPMTMR: ret = vpmtmr_snapshot(vm_pmtmr(vm), meta); break; case STRUCT_VRTC: ret = vrtc_snapshot(vm_rtc(vm), meta); break; default: printf("%s: failed to find the requested type %#x\n", __func__, meta->dev_req); ret = (EINVAL); } return (ret); } int vm_set_tsc_offset(struct vm *vm, int vcpuid, uint64_t offset) { struct vcpu *vcpu; if (vcpuid < 0 || vcpuid >= VM_MAXCPU) return (EINVAL); vcpu = &vm->vcpu[vcpuid]; vcpu->tsc_offset = offset; return (0); } int vm_restore_time(struct vm *vm) { int error, i; uint64_t now; struct vcpu *vcpu; now = rdtsc(); error = vhpet_restore_time(vm_hpet(vm)); if (error) return (error); for (i = 0; i < nitems(vm->vcpu); i++) { vcpu = &vm->vcpu[i]; error = vmmops_restore_tsc(vm->cookie, i, vcpu->tsc_offset - now); if (error) return (error); } return (0); } #endif diff --git a/sys/dev/acpica/acpi.c b/sys/dev/acpica/acpi.c index c5c0a38e3001..c1b79fb15727 100644 --- a/sys/dev/acpica/acpi.c +++ b/sys/dev/acpica/acpi.c @@ -1,4375 +1,4375 @@ /*- * Copyright (c) 2000 Takanori Watanabe * Copyright (c) 2000 Mitsuru IWASAKI * Copyright (c) 2000, 2001 Michael Smith * Copyright (c) 2000 BSDi * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include "opt_acpi.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #if defined(__i386__) || defined(__amd64__) #include #include #endif #include #include #include #include #include #include #include #include #include #include #include #include static MALLOC_DEFINE(M_ACPIDEV, "acpidev", "ACPI devices"); /* Hooks for the ACPI CA debugging infrastructure */ #define _COMPONENT ACPI_BUS ACPI_MODULE_NAME("ACPI") static d_open_t acpiopen; static d_close_t acpiclose; static d_ioctl_t acpiioctl; static struct cdevsw acpi_cdevsw = { .d_version = D_VERSION, .d_open = acpiopen, .d_close = acpiclose, .d_ioctl = acpiioctl, .d_name = "acpi", }; struct acpi_interface { ACPI_STRING *data; int num; }; static char *sysres_ids[] = { "PNP0C01", "PNP0C02", NULL }; static char *pcilink_ids[] = { "PNP0C0F", NULL }; /* Global mutex for locking access to the ACPI subsystem. */ struct mtx acpi_mutex; struct callout acpi_sleep_timer; /* Bitmap of device quirks. */ int acpi_quirks; /* Supported sleep states. */ static BOOLEAN acpi_sleep_states[ACPI_S_STATE_COUNT]; static void acpi_lookup(void *arg, const char *name, device_t *dev); static int acpi_modevent(struct module *mod, int event, void *junk); static int acpi_probe(device_t dev); static int acpi_attach(device_t dev); static int acpi_suspend(device_t dev); static int acpi_resume(device_t dev); static int acpi_shutdown(device_t dev); static device_t acpi_add_child(device_t bus, u_int order, const char *name, int unit); static int acpi_print_child(device_t bus, device_t child); static void acpi_probe_nomatch(device_t bus, device_t child); static void acpi_driver_added(device_t dev, driver_t *driver); static void acpi_child_deleted(device_t dev, device_t child); static int acpi_read_ivar(device_t dev, device_t child, int index, uintptr_t *result); static int acpi_write_ivar(device_t dev, device_t child, int index, uintptr_t value); static struct resource_list *acpi_get_rlist(device_t dev, device_t child); static void acpi_reserve_resources(device_t dev); static int acpi_sysres_alloc(device_t dev); static int acpi_set_resource(device_t dev, device_t child, int type, int rid, rman_res_t start, rman_res_t count); static struct resource *acpi_alloc_resource(device_t bus, device_t child, int type, int *rid, rman_res_t start, rman_res_t end, rman_res_t count, u_int flags); static int acpi_adjust_resource(device_t bus, device_t child, int type, struct resource *r, rman_res_t start, rman_res_t end); static int acpi_release_resource(device_t bus, device_t child, int type, int rid, struct resource *r); static void acpi_delete_resource(device_t bus, device_t child, int type, int rid); static uint32_t acpi_isa_get_logicalid(device_t dev); static int acpi_isa_get_compatid(device_t dev, uint32_t *cids, int count); static int acpi_device_id_probe(device_t bus, device_t dev, char **ids, char **match); static ACPI_STATUS acpi_device_eval_obj(device_t bus, device_t dev, ACPI_STRING pathname, ACPI_OBJECT_LIST *parameters, ACPI_BUFFER *ret); static ACPI_STATUS acpi_device_scan_cb(ACPI_HANDLE h, UINT32 level, void *context, void **retval); static ACPI_STATUS acpi_device_scan_children(device_t bus, device_t dev, int max_depth, acpi_scan_cb_t user_fn, void *arg); static int acpi_isa_pnp_probe(device_t bus, device_t child, struct isa_pnp_id *ids); static void acpi_platform_osc(device_t dev); static void acpi_probe_children(device_t bus); static void acpi_probe_order(ACPI_HANDLE handle, int *order); static ACPI_STATUS acpi_probe_child(ACPI_HANDLE handle, UINT32 level, void *context, void **status); static void acpi_sleep_enable(void *arg); static ACPI_STATUS acpi_sleep_disable(struct acpi_softc *sc); static ACPI_STATUS acpi_EnterSleepState(struct acpi_softc *sc, int state); static void acpi_shutdown_final(void *arg, int howto); static void acpi_enable_fixed_events(struct acpi_softc *sc); static void acpi_resync_clock(struct acpi_softc *sc); static int acpi_wake_sleep_prep(ACPI_HANDLE handle, int sstate); static int acpi_wake_run_prep(ACPI_HANDLE handle, int sstate); static int acpi_wake_prep_walk(int sstate); static int acpi_wake_sysctl_walk(device_t dev); static int acpi_wake_set_sysctl(SYSCTL_HANDLER_ARGS); static void acpi_system_eventhandler_sleep(void *arg, int state); static void acpi_system_eventhandler_wakeup(void *arg, int state); static int acpi_sname2sstate(const char *sname); static const char *acpi_sstate2sname(int sstate); static int acpi_supported_sleep_state_sysctl(SYSCTL_HANDLER_ARGS); static int acpi_sleep_state_sysctl(SYSCTL_HANDLER_ARGS); static int acpi_debug_objects_sysctl(SYSCTL_HANDLER_ARGS); static int acpi_pm_func(u_long cmd, void *arg, ...); static int acpi_child_location_str_method(device_t acdev, device_t child, char *buf, size_t buflen); static int acpi_child_pnpinfo_str_method(device_t acdev, device_t child, char *buf, size_t buflen); static void acpi_enable_pcie(void); static void acpi_hint_device_unit(device_t acdev, device_t child, const char *name, int *unitp); static void acpi_reset_interfaces(device_t dev); static device_method_t acpi_methods[] = { /* Device interface */ DEVMETHOD(device_probe, acpi_probe), DEVMETHOD(device_attach, acpi_attach), DEVMETHOD(device_shutdown, acpi_shutdown), DEVMETHOD(device_detach, bus_generic_detach), DEVMETHOD(device_suspend, acpi_suspend), DEVMETHOD(device_resume, acpi_resume), /* Bus interface */ DEVMETHOD(bus_add_child, acpi_add_child), DEVMETHOD(bus_print_child, acpi_print_child), DEVMETHOD(bus_probe_nomatch, acpi_probe_nomatch), DEVMETHOD(bus_driver_added, acpi_driver_added), DEVMETHOD(bus_child_deleted, acpi_child_deleted), DEVMETHOD(bus_read_ivar, acpi_read_ivar), DEVMETHOD(bus_write_ivar, acpi_write_ivar), DEVMETHOD(bus_get_resource_list, acpi_get_rlist), DEVMETHOD(bus_set_resource, acpi_set_resource), DEVMETHOD(bus_get_resource, bus_generic_rl_get_resource), DEVMETHOD(bus_alloc_resource, acpi_alloc_resource), DEVMETHOD(bus_adjust_resource, acpi_adjust_resource), DEVMETHOD(bus_release_resource, acpi_release_resource), DEVMETHOD(bus_delete_resource, acpi_delete_resource), DEVMETHOD(bus_child_pnpinfo_str, acpi_child_pnpinfo_str_method), DEVMETHOD(bus_child_location_str, acpi_child_location_str_method), DEVMETHOD(bus_activate_resource, bus_generic_activate_resource), DEVMETHOD(bus_deactivate_resource, bus_generic_deactivate_resource), DEVMETHOD(bus_setup_intr, bus_generic_setup_intr), DEVMETHOD(bus_teardown_intr, bus_generic_teardown_intr), DEVMETHOD(bus_hint_device_unit, acpi_hint_device_unit), DEVMETHOD(bus_get_cpus, acpi_get_cpus), DEVMETHOD(bus_get_domain, acpi_get_domain), /* ACPI bus */ DEVMETHOD(acpi_id_probe, acpi_device_id_probe), DEVMETHOD(acpi_evaluate_object, acpi_device_eval_obj), DEVMETHOD(acpi_pwr_for_sleep, acpi_device_pwr_for_sleep), DEVMETHOD(acpi_scan_children, acpi_device_scan_children), /* ISA emulation */ DEVMETHOD(isa_pnp_probe, acpi_isa_pnp_probe), DEVMETHOD_END }; static driver_t acpi_driver = { "acpi", acpi_methods, sizeof(struct acpi_softc), }; static devclass_t acpi_devclass; EARLY_DRIVER_MODULE(acpi, nexus, acpi_driver, acpi_devclass, acpi_modevent, 0, BUS_PASS_BUS + BUS_PASS_ORDER_MIDDLE); MODULE_VERSION(acpi, 1); ACPI_SERIAL_DECL(acpi, "ACPI root bus"); /* Local pools for managing system resources for ACPI child devices. */ static struct rman acpi_rman_io, acpi_rman_mem; #define ACPI_MINIMUM_AWAKETIME 5 /* Holds the description of the acpi0 device. */ static char acpi_desc[ACPI_OEM_ID_SIZE + ACPI_OEM_TABLE_ID_SIZE + 2]; SYSCTL_NODE(_debug, OID_AUTO, acpi, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "ACPI debugging"); static char acpi_ca_version[12]; SYSCTL_STRING(_debug_acpi, OID_AUTO, acpi_ca_version, CTLFLAG_RD, acpi_ca_version, 0, "Version of Intel ACPI-CA"); /* * Allow overriding _OSI methods. */ static char acpi_install_interface[256]; TUNABLE_STR("hw.acpi.install_interface", acpi_install_interface, sizeof(acpi_install_interface)); static char acpi_remove_interface[256]; TUNABLE_STR("hw.acpi.remove_interface", acpi_remove_interface, sizeof(acpi_remove_interface)); /* Allow users to dump Debug objects without ACPI debugger. */ static int acpi_debug_objects; TUNABLE_INT("debug.acpi.enable_debug_objects", &acpi_debug_objects); SYSCTL_PROC(_debug_acpi, OID_AUTO, enable_debug_objects, CTLFLAG_RW | CTLTYPE_INT | CTLFLAG_MPSAFE, NULL, 0, acpi_debug_objects_sysctl, "I", "Enable Debug objects"); /* Allow the interpreter to ignore common mistakes in BIOS. */ static int acpi_interpreter_slack = 1; TUNABLE_INT("debug.acpi.interpreter_slack", &acpi_interpreter_slack); SYSCTL_INT(_debug_acpi, OID_AUTO, interpreter_slack, CTLFLAG_RDTUN, &acpi_interpreter_slack, 1, "Turn on interpreter slack mode."); /* Ignore register widths set by FADT and use default widths instead. */ static int acpi_ignore_reg_width = 1; TUNABLE_INT("debug.acpi.default_register_width", &acpi_ignore_reg_width); SYSCTL_INT(_debug_acpi, OID_AUTO, default_register_width, CTLFLAG_RDTUN, &acpi_ignore_reg_width, 1, "Ignore register widths set by FADT"); /* Allow users to override quirks. */ TUNABLE_INT("debug.acpi.quirks", &acpi_quirks); int acpi_susp_bounce; SYSCTL_INT(_debug_acpi, OID_AUTO, suspend_bounce, CTLFLAG_RW, &acpi_susp_bounce, 0, "Don't actually suspend, just test devices."); /* * ACPI can only be loaded as a module by the loader; activating it after * system bootstrap time is not useful, and can be fatal to the system. * It also cannot be unloaded, since the entire system bus hierarchy hangs * off it. */ static int acpi_modevent(struct module *mod, int event, void *junk) { switch (event) { case MOD_LOAD: if (!cold) { printf("The ACPI driver cannot be loaded after boot.\n"); return (EPERM); } break; case MOD_UNLOAD: if (!cold && power_pm_get_type() == POWER_PM_TYPE_ACPI) return (EBUSY); break; default: break; } return (0); } /* * Perform early initialization. */ ACPI_STATUS acpi_Startup(void) { static int started = 0; ACPI_STATUS status; int val; ACPI_FUNCTION_TRACE((char *)(uintptr_t)__func__); /* Only run the startup code once. The MADT driver also calls this. */ if (started) return_VALUE (AE_OK); started = 1; /* * Initialize the ACPICA subsystem. */ if (ACPI_FAILURE(status = AcpiInitializeSubsystem())) { printf("ACPI: Could not initialize Subsystem: %s\n", AcpiFormatException(status)); return_VALUE (status); } /* * Pre-allocate space for RSDT/XSDT and DSDT tables and allow resizing * if more tables exist. */ if (ACPI_FAILURE(status = AcpiInitializeTables(NULL, 2, TRUE))) { printf("ACPI: Table initialisation failed: %s\n", AcpiFormatException(status)); return_VALUE (status); } /* Set up any quirks we have for this system. */ if (acpi_quirks == ACPI_Q_OK) acpi_table_quirks(&acpi_quirks); /* If the user manually set the disabled hint to 0, force-enable ACPI. */ if (resource_int_value("acpi", 0, "disabled", &val) == 0 && val == 0) acpi_quirks &= ~ACPI_Q_BROKEN; if (acpi_quirks & ACPI_Q_BROKEN) { printf("ACPI disabled by blacklist. Contact your BIOS vendor.\n"); status = AE_SUPPORT; } return_VALUE (status); } /* * Detect ACPI and perform early initialisation. */ int acpi_identify(void) { ACPI_TABLE_RSDP *rsdp; ACPI_TABLE_HEADER *rsdt; ACPI_PHYSICAL_ADDRESS paddr; struct sbuf sb; ACPI_FUNCTION_TRACE((char *)(uintptr_t)__func__); if (!cold) return (ENXIO); /* Check that we haven't been disabled with a hint. */ if (resource_disabled("acpi", 0)) return (ENXIO); /* Check for other PM systems. */ if (power_pm_get_type() != POWER_PM_TYPE_NONE && power_pm_get_type() != POWER_PM_TYPE_ACPI) { printf("ACPI identify failed, other PM system enabled.\n"); return (ENXIO); } /* Initialize root tables. */ if (ACPI_FAILURE(acpi_Startup())) { printf("ACPI: Try disabling either ACPI or apic support.\n"); return (ENXIO); } if ((paddr = AcpiOsGetRootPointer()) == 0 || (rsdp = AcpiOsMapMemory(paddr, sizeof(ACPI_TABLE_RSDP))) == NULL) return (ENXIO); if (rsdp->Revision > 1 && rsdp->XsdtPhysicalAddress != 0) paddr = (ACPI_PHYSICAL_ADDRESS)rsdp->XsdtPhysicalAddress; else paddr = (ACPI_PHYSICAL_ADDRESS)rsdp->RsdtPhysicalAddress; AcpiOsUnmapMemory(rsdp, sizeof(ACPI_TABLE_RSDP)); if ((rsdt = AcpiOsMapMemory(paddr, sizeof(ACPI_TABLE_HEADER))) == NULL) return (ENXIO); sbuf_new(&sb, acpi_desc, sizeof(acpi_desc), SBUF_FIXEDLEN); sbuf_bcat(&sb, rsdt->OemId, ACPI_OEM_ID_SIZE); sbuf_trim(&sb); sbuf_putc(&sb, ' '); sbuf_bcat(&sb, rsdt->OemTableId, ACPI_OEM_TABLE_ID_SIZE); sbuf_trim(&sb); sbuf_finish(&sb); sbuf_delete(&sb); AcpiOsUnmapMemory(rsdt, sizeof(ACPI_TABLE_HEADER)); snprintf(acpi_ca_version, sizeof(acpi_ca_version), "%x", ACPI_CA_VERSION); return (0); } /* * Fetch some descriptive data from ACPI to put in our attach message. */ static int acpi_probe(device_t dev) { ACPI_FUNCTION_TRACE((char *)(uintptr_t)__func__); device_set_desc(dev, acpi_desc); return_VALUE (BUS_PROBE_NOWILDCARD); } static int acpi_attach(device_t dev) { struct acpi_softc *sc; ACPI_STATUS status; int error, state; UINT32 flags; UINT8 TypeA, TypeB; char *env; ACPI_FUNCTION_TRACE((char *)(uintptr_t)__func__); sc = device_get_softc(dev); sc->acpi_dev = dev; callout_init(&sc->susp_force_to, 1); error = ENXIO; /* Initialize resource manager. */ acpi_rman_io.rm_type = RMAN_ARRAY; acpi_rman_io.rm_start = 0; acpi_rman_io.rm_end = 0xffff; acpi_rman_io.rm_descr = "ACPI I/O ports"; if (rman_init(&acpi_rman_io) != 0) panic("acpi rman_init IO ports failed"); acpi_rman_mem.rm_type = RMAN_ARRAY; acpi_rman_mem.rm_descr = "ACPI I/O memory addresses"; if (rman_init(&acpi_rman_mem) != 0) panic("acpi rman_init memory failed"); /* Initialise the ACPI mutex */ mtx_init(&acpi_mutex, "ACPI global lock", NULL, MTX_DEF); /* * Set the globals from our tunables. This is needed because ACPI-CA * uses UINT8 for some values and we have no tunable_byte. */ AcpiGbl_EnableInterpreterSlack = acpi_interpreter_slack ? TRUE : FALSE; AcpiGbl_EnableAmlDebugObject = acpi_debug_objects ? TRUE : FALSE; AcpiGbl_UseDefaultRegisterWidths = acpi_ignore_reg_width ? TRUE : FALSE; #ifndef ACPI_DEBUG /* * Disable all debugging layers and levels. */ AcpiDbgLayer = 0; AcpiDbgLevel = 0; #endif /* Override OS interfaces if the user requested. */ acpi_reset_interfaces(dev); /* Load ACPI name space. */ status = AcpiLoadTables(); if (ACPI_FAILURE(status)) { device_printf(dev, "Could not load Namespace: %s\n", AcpiFormatException(status)); goto out; } /* Handle MCFG table if present. */ acpi_enable_pcie(); /* * Note that some systems (specifically, those with namespace evaluation * issues that require the avoidance of parts of the namespace) must * avoid running _INI and _STA on everything, as well as dodging the final * object init pass. * * For these devices, we set ACPI_NO_DEVICE_INIT and ACPI_NO_OBJECT_INIT). * * XXX We should arrange for the object init pass after we have attached * all our child devices, but on many systems it works here. */ flags = 0; if (testenv("debug.acpi.avoid")) flags = ACPI_NO_DEVICE_INIT | ACPI_NO_OBJECT_INIT; /* Bring the hardware and basic handlers online. */ if (ACPI_FAILURE(status = AcpiEnableSubsystem(flags))) { device_printf(dev, "Could not enable ACPI: %s\n", AcpiFormatException(status)); goto out; } /* * Call the ECDT probe function to provide EC functionality before * the namespace has been evaluated. * * XXX This happens before the sysresource devices have been probed and * attached so its resources come from nexus0. In practice, this isn't * a problem but should be addressed eventually. */ acpi_ec_ecdt_probe(dev); /* Bring device objects and regions online. */ if (ACPI_FAILURE(status = AcpiInitializeObjects(flags))) { device_printf(dev, "Could not initialize ACPI objects: %s\n", AcpiFormatException(status)); goto out; } /* * Setup our sysctl tree. * * XXX: This doesn't check to make sure that none of these fail. */ sysctl_ctx_init(&sc->acpi_sysctl_ctx); sc->acpi_sysctl_tree = SYSCTL_ADD_NODE(&sc->acpi_sysctl_ctx, SYSCTL_STATIC_CHILDREN(_hw), OID_AUTO, device_get_name(dev), CTLFLAG_RD | CTLFLAG_MPSAFE, 0, ""); SYSCTL_ADD_PROC(&sc->acpi_sysctl_ctx, SYSCTL_CHILDREN(sc->acpi_sysctl_tree), OID_AUTO, "supported_sleep_state", CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 0, acpi_supported_sleep_state_sysctl, "A", "List supported ACPI sleep states."); SYSCTL_ADD_PROC(&sc->acpi_sysctl_ctx, SYSCTL_CHILDREN(sc->acpi_sysctl_tree), OID_AUTO, "power_button_state", CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE, &sc->acpi_power_button_sx, 0, acpi_sleep_state_sysctl, "A", "Power button ACPI sleep state."); SYSCTL_ADD_PROC(&sc->acpi_sysctl_ctx, SYSCTL_CHILDREN(sc->acpi_sysctl_tree), OID_AUTO, "sleep_button_state", CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE, &sc->acpi_sleep_button_sx, 0, acpi_sleep_state_sysctl, "A", "Sleep button ACPI sleep state."); SYSCTL_ADD_PROC(&sc->acpi_sysctl_ctx, SYSCTL_CHILDREN(sc->acpi_sysctl_tree), OID_AUTO, "lid_switch_state", CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE, &sc->acpi_lid_switch_sx, 0, acpi_sleep_state_sysctl, "A", "Lid ACPI sleep state. Set to S3 if you want to suspend your laptop when close the Lid."); SYSCTL_ADD_PROC(&sc->acpi_sysctl_ctx, SYSCTL_CHILDREN(sc->acpi_sysctl_tree), OID_AUTO, "standby_state", CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE, &sc->acpi_standby_sx, 0, acpi_sleep_state_sysctl, "A", ""); SYSCTL_ADD_PROC(&sc->acpi_sysctl_ctx, SYSCTL_CHILDREN(sc->acpi_sysctl_tree), OID_AUTO, "suspend_state", CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE, &sc->acpi_suspend_sx, 0, acpi_sleep_state_sysctl, "A", ""); SYSCTL_ADD_INT(&sc->acpi_sysctl_ctx, SYSCTL_CHILDREN(sc->acpi_sysctl_tree), OID_AUTO, "sleep_delay", CTLFLAG_RW, &sc->acpi_sleep_delay, 0, "sleep delay in seconds"); SYSCTL_ADD_INT(&sc->acpi_sysctl_ctx, SYSCTL_CHILDREN(sc->acpi_sysctl_tree), OID_AUTO, "s4bios", CTLFLAG_RW, &sc->acpi_s4bios, 0, "S4BIOS mode"); SYSCTL_ADD_INT(&sc->acpi_sysctl_ctx, SYSCTL_CHILDREN(sc->acpi_sysctl_tree), OID_AUTO, "verbose", CTLFLAG_RW, &sc->acpi_verbose, 0, "verbose mode"); SYSCTL_ADD_INT(&sc->acpi_sysctl_ctx, SYSCTL_CHILDREN(sc->acpi_sysctl_tree), OID_AUTO, "disable_on_reboot", CTLFLAG_RW, &sc->acpi_do_disable, 0, "Disable ACPI when rebooting/halting system"); SYSCTL_ADD_INT(&sc->acpi_sysctl_ctx, SYSCTL_CHILDREN(sc->acpi_sysctl_tree), OID_AUTO, "handle_reboot", CTLFLAG_RW, &sc->acpi_handle_reboot, 0, "Use ACPI Reset Register to reboot"); /* * Default to 1 second before sleeping to give some machines time to * stabilize. */ sc->acpi_sleep_delay = 1; if (bootverbose) sc->acpi_verbose = 1; if ((env = kern_getenv("hw.acpi.verbose")) != NULL) { if (strcmp(env, "0") != 0) sc->acpi_verbose = 1; freeenv(env); } /* Only enable reboot by default if the FADT says it is available. */ if (AcpiGbl_FADT.Flags & ACPI_FADT_RESET_REGISTER) sc->acpi_handle_reboot = 1; #if !ACPI_REDUCED_HARDWARE /* Only enable S4BIOS by default if the FACS says it is available. */ if (AcpiGbl_FACS != NULL && AcpiGbl_FACS->Flags & ACPI_FACS_S4_BIOS_PRESENT) sc->acpi_s4bios = 1; #endif /* Probe all supported sleep states. */ acpi_sleep_states[ACPI_STATE_S0] = TRUE; for (state = ACPI_STATE_S1; state < ACPI_S_STATE_COUNT; state++) if (ACPI_SUCCESS(AcpiEvaluateObject(ACPI_ROOT_OBJECT, __DECONST(char *, AcpiGbl_SleepStateNames[state]), NULL, NULL)) && ACPI_SUCCESS(AcpiGetSleepTypeData(state, &TypeA, &TypeB))) acpi_sleep_states[state] = TRUE; /* * Dispatch the default sleep state to devices. The lid switch is set * to UNKNOWN by default to avoid surprising users. */ sc->acpi_power_button_sx = acpi_sleep_states[ACPI_STATE_S5] ? ACPI_STATE_S5 : ACPI_STATE_UNKNOWN; sc->acpi_lid_switch_sx = ACPI_STATE_UNKNOWN; sc->acpi_standby_sx = acpi_sleep_states[ACPI_STATE_S1] ? ACPI_STATE_S1 : ACPI_STATE_UNKNOWN; sc->acpi_suspend_sx = acpi_sleep_states[ACPI_STATE_S3] ? ACPI_STATE_S3 : ACPI_STATE_UNKNOWN; /* Pick the first valid sleep state for the sleep button default. */ sc->acpi_sleep_button_sx = ACPI_STATE_UNKNOWN; for (state = ACPI_STATE_S1; state <= ACPI_STATE_S4; state++) if (acpi_sleep_states[state]) { sc->acpi_sleep_button_sx = state; break; } acpi_enable_fixed_events(sc); /* * Scan the namespace and attach/initialise children. */ /* Register our shutdown handler. */ EVENTHANDLER_REGISTER(shutdown_final, acpi_shutdown_final, sc, SHUTDOWN_PRI_LAST); /* * Register our acpi event handlers. * XXX should be configurable eg. via userland policy manager. */ EVENTHANDLER_REGISTER(acpi_sleep_event, acpi_system_eventhandler_sleep, sc, ACPI_EVENT_PRI_LAST); EVENTHANDLER_REGISTER(acpi_wakeup_event, acpi_system_eventhandler_wakeup, sc, ACPI_EVENT_PRI_LAST); /* Flag our initial states. */ sc->acpi_enabled = TRUE; sc->acpi_sstate = ACPI_STATE_S0; sc->acpi_sleep_disabled = TRUE; /* Create the control device */ sc->acpi_dev_t = make_dev(&acpi_cdevsw, 0, UID_ROOT, GID_OPERATOR, 0664, "acpi"); sc->acpi_dev_t->si_drv1 = sc; if ((error = acpi_machdep_init(dev))) goto out; /* Register ACPI again to pass the correct argument of pm_func. */ power_pm_register(POWER_PM_TYPE_ACPI, acpi_pm_func, sc); acpi_platform_osc(dev); if (!acpi_disabled("bus")) { EVENTHANDLER_REGISTER(dev_lookup, acpi_lookup, NULL, 1000); acpi_probe_children(dev); } /* Update all GPEs and enable runtime GPEs. */ status = AcpiUpdateAllGpes(); if (ACPI_FAILURE(status)) device_printf(dev, "Could not update all GPEs: %s\n", AcpiFormatException(status)); /* Allow sleep request after a while. */ callout_init_mtx(&acpi_sleep_timer, &acpi_mutex, 0); callout_reset(&acpi_sleep_timer, hz * ACPI_MINIMUM_AWAKETIME, acpi_sleep_enable, sc); error = 0; out: return_VALUE (error); } static void acpi_set_power_children(device_t dev, int state) { device_t child; device_t *devlist; int dstate, i, numdevs; if (device_get_children(dev, &devlist, &numdevs) != 0) return; /* * Retrieve and set D-state for the sleep state if _SxD is present. * Skip children who aren't attached since they are handled separately. */ for (i = 0; i < numdevs; i++) { child = devlist[i]; dstate = state; if (device_is_attached(child) && acpi_device_pwr_for_sleep(dev, child, &dstate) == 0) acpi_set_powerstate(child, dstate); } free(devlist, M_TEMP); } static int acpi_suspend(device_t dev) { int error; GIANT_REQUIRED; error = bus_generic_suspend(dev); if (error == 0) acpi_set_power_children(dev, ACPI_STATE_D3); return (error); } static int acpi_resume(device_t dev) { GIANT_REQUIRED; acpi_set_power_children(dev, ACPI_STATE_D0); return (bus_generic_resume(dev)); } static int acpi_shutdown(device_t dev) { GIANT_REQUIRED; /* Allow children to shutdown first. */ bus_generic_shutdown(dev); /* * Enable any GPEs that are able to power-on the system (i.e., RTC). * Also, disable any that are not valid for this state (most). */ acpi_wake_prep_walk(ACPI_STATE_S5); return (0); } /* * Handle a new device being added */ static device_t acpi_add_child(device_t bus, u_int order, const char *name, int unit) { struct acpi_device *ad; device_t child; if ((ad = malloc(sizeof(*ad), M_ACPIDEV, M_NOWAIT | M_ZERO)) == NULL) return (NULL); resource_list_init(&ad->ad_rl); child = device_add_child_ordered(bus, order, name, unit); if (child != NULL) device_set_ivars(child, ad); else free(ad, M_ACPIDEV); return (child); } static int acpi_print_child(device_t bus, device_t child) { struct acpi_device *adev = device_get_ivars(child); struct resource_list *rl = &adev->ad_rl; int retval = 0; retval += bus_print_child_header(bus, child); retval += resource_list_print_type(rl, "port", SYS_RES_IOPORT, "%#jx"); retval += resource_list_print_type(rl, "iomem", SYS_RES_MEMORY, "%#jx"); retval += resource_list_print_type(rl, "irq", SYS_RES_IRQ, "%jd"); retval += resource_list_print_type(rl, "drq", SYS_RES_DRQ, "%jd"); if (device_get_flags(child)) retval += printf(" flags %#x", device_get_flags(child)); retval += bus_print_child_domain(bus, child); retval += bus_print_child_footer(bus, child); return (retval); } /* * If this device is an ACPI child but no one claimed it, attempt * to power it off. We'll power it back up when a driver is added. * * XXX Disabled for now since many necessary devices (like fdc and * ATA) don't claim the devices we created for them but still expect * them to be powered up. */ static void acpi_probe_nomatch(device_t bus, device_t child) { #ifdef ACPI_ENABLE_POWERDOWN_NODRIVER acpi_set_powerstate(child, ACPI_STATE_D3); #endif } /* * If a new driver has a chance to probe a child, first power it up. * * XXX Disabled for now (see acpi_probe_nomatch for details). */ static void acpi_driver_added(device_t dev, driver_t *driver) { device_t child, *devlist; int i, numdevs; DEVICE_IDENTIFY(driver, dev); if (device_get_children(dev, &devlist, &numdevs)) return; for (i = 0; i < numdevs; i++) { child = devlist[i]; if (device_get_state(child) == DS_NOTPRESENT) { #ifdef ACPI_ENABLE_POWERDOWN_NODRIVER acpi_set_powerstate(child, ACPI_STATE_D0); if (device_probe_and_attach(child) != 0) acpi_set_powerstate(child, ACPI_STATE_D3); #else device_probe_and_attach(child); #endif } } free(devlist, M_TEMP); } /* Location hint for devctl(8) */ static int acpi_child_location_str_method(device_t cbdev, device_t child, char *buf, size_t buflen) { struct acpi_device *dinfo = device_get_ivars(child); char buf2[32]; int pxm; if (dinfo->ad_handle) { snprintf(buf, buflen, "handle=%s", acpi_name(dinfo->ad_handle)); if (ACPI_SUCCESS(acpi_GetInteger(dinfo->ad_handle, "_PXM", &pxm))) { snprintf(buf2, 32, " _PXM=%d", pxm); strlcat(buf, buf2, buflen); } } else { snprintf(buf, buflen, ""); } return (0); } /* PnP information for devctl(8) */ int acpi_pnpinfo_str(ACPI_HANDLE handle, char *buf, size_t buflen) { ACPI_DEVICE_INFO *adinfo; if (ACPI_FAILURE(AcpiGetObjectInfo(handle, &adinfo))) { snprintf(buf, buflen, "unknown"); return (0); } snprintf(buf, buflen, "_HID=%s _UID=%lu _CID=%s", (adinfo->Valid & ACPI_VALID_HID) ? adinfo->HardwareId.String : "none", (adinfo->Valid & ACPI_VALID_UID) ? strtoul(adinfo->UniqueId.String, NULL, 10) : 0UL, ((adinfo->Valid & ACPI_VALID_CID) && adinfo->CompatibleIdList.Count > 0) ? adinfo->CompatibleIdList.Ids[0].String : "none"); AcpiOsFree(adinfo); return (0); } static int acpi_child_pnpinfo_str_method(device_t cbdev, device_t child, char *buf, size_t buflen) { struct acpi_device *dinfo = device_get_ivars(child); return (acpi_pnpinfo_str(dinfo->ad_handle, buf, buflen)); } /* * Handle device deletion. */ static void acpi_child_deleted(device_t dev, device_t child) { struct acpi_device *dinfo = device_get_ivars(child); if (acpi_get_device(dinfo->ad_handle) == child) AcpiDetachData(dinfo->ad_handle, acpi_fake_objhandler); } /* * Handle per-device ivars */ static int acpi_read_ivar(device_t dev, device_t child, int index, uintptr_t *result) { struct acpi_device *ad; if ((ad = device_get_ivars(child)) == NULL) { device_printf(child, "device has no ivars\n"); return (ENOENT); } /* ACPI and ISA compatibility ivars */ switch(index) { case ACPI_IVAR_HANDLE: *(ACPI_HANDLE *)result = ad->ad_handle; break; case ACPI_IVAR_PRIVATE: *(void **)result = ad->ad_private; break; case ACPI_IVAR_FLAGS: *(int *)result = ad->ad_flags; break; case ISA_IVAR_VENDORID: case ISA_IVAR_SERIAL: case ISA_IVAR_COMPATID: *(int *)result = -1; break; case ISA_IVAR_LOGICALID: *(int *)result = acpi_isa_get_logicalid(child); break; case PCI_IVAR_CLASS: *(uint8_t*)result = (ad->ad_cls_class >> 16) & 0xff; break; case PCI_IVAR_SUBCLASS: *(uint8_t*)result = (ad->ad_cls_class >> 8) & 0xff; break; case PCI_IVAR_PROGIF: *(uint8_t*)result = (ad->ad_cls_class >> 0) & 0xff; break; default: return (ENOENT); } return (0); } static int acpi_write_ivar(device_t dev, device_t child, int index, uintptr_t value) { struct acpi_device *ad; if ((ad = device_get_ivars(child)) == NULL) { device_printf(child, "device has no ivars\n"); return (ENOENT); } switch(index) { case ACPI_IVAR_HANDLE: ad->ad_handle = (ACPI_HANDLE)value; break; case ACPI_IVAR_PRIVATE: ad->ad_private = (void *)value; break; case ACPI_IVAR_FLAGS: ad->ad_flags = (int)value; break; default: panic("bad ivar write request (%d)", index); return (ENOENT); } return (0); } /* * Handle child resource allocation/removal */ static struct resource_list * acpi_get_rlist(device_t dev, device_t child) { struct acpi_device *ad; ad = device_get_ivars(child); return (&ad->ad_rl); } static int acpi_match_resource_hint(device_t dev, int type, long value) { struct acpi_device *ad = device_get_ivars(dev); struct resource_list *rl = &ad->ad_rl; struct resource_list_entry *rle; STAILQ_FOREACH(rle, rl, link) { if (rle->type != type) continue; if (rle->start <= value && rle->end >= value) return (1); } return (0); } /* * Wire device unit numbers based on resource matches in hints. */ static void acpi_hint_device_unit(device_t acdev, device_t child, const char *name, int *unitp) { const char *s; long value; int line, matches, unit; /* * Iterate over all the hints for the devices with the specified * name to see if one's resources are a subset of this device. */ line = 0; while (resource_find_dev(&line, name, &unit, "at", NULL) == 0) { /* Must have an "at" for acpi or isa. */ resource_string_value(name, unit, "at", &s); if (!(strcmp(s, "acpi0") == 0 || strcmp(s, "acpi") == 0 || strcmp(s, "isa0") == 0 || strcmp(s, "isa") == 0)) continue; /* * Check for matching resources. We must have at least one match. * Since I/O and memory resources cannot be shared, if we get a * match on either of those, ignore any mismatches in IRQs or DRQs. * * XXX: We may want to revisit this to be more lenient and wire * as long as it gets one match. */ matches = 0; if (resource_long_value(name, unit, "port", &value) == 0) { /* * Floppy drive controllers are notorious for having a * wide variety of resources not all of which include the * first port that is specified by the hint (typically * 0x3f0) (see the comment above fdc_isa_alloc_resources() * in fdc_isa.c). However, they do all seem to include * port + 2 (e.g. 0x3f2) so for a floppy device, look for * 'value + 2' in the port resources instead of the hint * value. */ if (strcmp(name, "fdc") == 0) value += 2; if (acpi_match_resource_hint(child, SYS_RES_IOPORT, value)) matches++; else continue; } if (resource_long_value(name, unit, "maddr", &value) == 0) { if (acpi_match_resource_hint(child, SYS_RES_MEMORY, value)) matches++; else continue; } if (matches > 0) goto matched; if (resource_long_value(name, unit, "irq", &value) == 0) { if (acpi_match_resource_hint(child, SYS_RES_IRQ, value)) matches++; else continue; } if (resource_long_value(name, unit, "drq", &value) == 0) { if (acpi_match_resource_hint(child, SYS_RES_DRQ, value)) matches++; else continue; } matched: if (matches > 0) { /* We have a winner! */ *unitp = unit; break; } } } /* * Fetch the NUMA domain for a device by mapping the value returned by * _PXM to a NUMA domain. If the device does not have a _PXM method, * -2 is returned. If any other error occurs, -1 is returned. */ static int acpi_parse_pxm(device_t dev) { #ifdef NUMA #if defined(__i386__) || defined(__amd64__) || defined(__aarch64__) ACPI_HANDLE handle; ACPI_STATUS status; int pxm; handle = acpi_get_handle(dev); if (handle == NULL) return (-2); status = acpi_GetInteger(handle, "_PXM", &pxm); if (ACPI_SUCCESS(status)) return (acpi_map_pxm_to_vm_domainid(pxm)); if (status == AE_NOT_FOUND) return (-2); #endif #endif return (-1); } int acpi_get_cpus(device_t dev, device_t child, enum cpu_sets op, size_t setsize, cpuset_t *cpuset) { int d, error; d = acpi_parse_pxm(child); if (d < 0) return (bus_generic_get_cpus(dev, child, op, setsize, cpuset)); switch (op) { case LOCAL_CPUS: if (setsize != sizeof(cpuset_t)) return (EINVAL); *cpuset = cpuset_domain[d]; return (0); case INTR_CPUS: error = bus_generic_get_cpus(dev, child, op, setsize, cpuset); if (error != 0) return (error); if (setsize != sizeof(cpuset_t)) return (EINVAL); - CPU_AND(cpuset, &cpuset_domain[d]); + CPU_AND(cpuset, cpuset, &cpuset_domain[d]); return (0); default: return (bus_generic_get_cpus(dev, child, op, setsize, cpuset)); } } /* * Fetch the NUMA domain for the given device 'dev'. * * If a device has a _PXM method, map that to a NUMA domain. * Otherwise, pass the request up to the parent. * If there's no matching domain or the domain cannot be * determined, return ENOENT. */ int acpi_get_domain(device_t dev, device_t child, int *domain) { int d; d = acpi_parse_pxm(child); if (d >= 0) { *domain = d; return (0); } if (d == -1) return (ENOENT); /* No _PXM node; go up a level */ return (bus_generic_get_domain(dev, child, domain)); } /* * Pre-allocate/manage all memory and IO resources. Since rman can't handle * duplicates, we merge any in the sysresource attach routine. */ static int acpi_sysres_alloc(device_t dev) { struct resource *res; struct resource_list *rl; struct resource_list_entry *rle; struct rman *rm; device_t *children; int child_count, i; /* * Probe/attach any sysresource devices. This would be unnecessary if we * had multi-pass probe/attach. */ if (device_get_children(dev, &children, &child_count) != 0) return (ENXIO); for (i = 0; i < child_count; i++) { if (ACPI_ID_PROBE(dev, children[i], sysres_ids, NULL) <= 0) device_probe_and_attach(children[i]); } free(children, M_TEMP); rl = BUS_GET_RESOURCE_LIST(device_get_parent(dev), dev); STAILQ_FOREACH(rle, rl, link) { if (rle->res != NULL) { device_printf(dev, "duplicate resource for %jx\n", rle->start); continue; } /* Only memory and IO resources are valid here. */ switch (rle->type) { case SYS_RES_IOPORT: rm = &acpi_rman_io; break; case SYS_RES_MEMORY: rm = &acpi_rman_mem; break; default: continue; } /* Pre-allocate resource and add to our rman pool. */ res = BUS_ALLOC_RESOURCE(device_get_parent(dev), dev, rle->type, &rle->rid, rle->start, rle->start + rle->count - 1, rle->count, 0); if (res != NULL) { rman_manage_region(rm, rman_get_start(res), rman_get_end(res)); rle->res = res; } else if (bootverbose) device_printf(dev, "reservation of %jx, %jx (%d) failed\n", rle->start, rle->count, rle->type); } return (0); } /* * Reserve declared resources for devices found during attach once system * resources have been allocated. */ static void acpi_reserve_resources(device_t dev) { struct resource_list_entry *rle; struct resource_list *rl; struct acpi_device *ad; struct acpi_softc *sc; device_t *children; int child_count, i; sc = device_get_softc(dev); if (device_get_children(dev, &children, &child_count) != 0) return; for (i = 0; i < child_count; i++) { ad = device_get_ivars(children[i]); rl = &ad->ad_rl; /* Don't reserve system resources. */ if (ACPI_ID_PROBE(dev, children[i], sysres_ids, NULL) <= 0) continue; STAILQ_FOREACH(rle, rl, link) { /* * Don't reserve IRQ resources. There are many sticky things * to get right otherwise (e.g. IRQs for psm, atkbd, and HPET * when using legacy routing). */ if (rle->type == SYS_RES_IRQ) continue; /* * Don't reserve the resource if it is already allocated. * The acpi_ec(4) driver can allocate its resources early * if ECDT is present. */ if (rle->res != NULL) continue; /* * Try to reserve the resource from our parent. If this * fails because the resource is a system resource, just * let it be. The resource range is already reserved so * that other devices will not use it. If the driver * needs to allocate the resource, then * acpi_alloc_resource() will sub-alloc from the system * resource. */ resource_list_reserve(rl, dev, children[i], rle->type, &rle->rid, rle->start, rle->end, rle->count, 0); } } free(children, M_TEMP); sc->acpi_resources_reserved = 1; } static int acpi_set_resource(device_t dev, device_t child, int type, int rid, rman_res_t start, rman_res_t count) { struct acpi_softc *sc = device_get_softc(dev); struct acpi_device *ad = device_get_ivars(child); struct resource_list *rl = &ad->ad_rl; ACPI_DEVICE_INFO *devinfo; rman_res_t end; int allow; /* Ignore IRQ resources for PCI link devices. */ if (type == SYS_RES_IRQ && ACPI_ID_PROBE(dev, child, pcilink_ids, NULL) <= 0) return (0); /* * Ignore most resources for PCI root bridges. Some BIOSes * incorrectly enumerate the memory ranges they decode as plain * memory resources instead of as ResourceProducer ranges. Other * BIOSes incorrectly list system resource entries for I/O ranges * under the PCI bridge. Do allow the one known-correct case on * x86 of a PCI bridge claiming the I/O ports used for PCI config * access. */ if (type == SYS_RES_MEMORY || type == SYS_RES_IOPORT) { if (ACPI_SUCCESS(AcpiGetObjectInfo(ad->ad_handle, &devinfo))) { if ((devinfo->Flags & ACPI_PCI_ROOT_BRIDGE) != 0) { #if defined(__i386__) || defined(__amd64__) allow = (type == SYS_RES_IOPORT && start == CONF1_ADDR_PORT); #else allow = 0; #endif if (!allow) { AcpiOsFree(devinfo); return (0); } } AcpiOsFree(devinfo); } } #ifdef INTRNG /* map with default for now */ if (type == SYS_RES_IRQ) start = (rman_res_t)acpi_map_intr(child, (u_int)start, acpi_get_handle(child)); #endif /* If the resource is already allocated, fail. */ if (resource_list_busy(rl, type, rid)) return (EBUSY); /* If the resource is already reserved, release it. */ if (resource_list_reserved(rl, type, rid)) resource_list_unreserve(rl, dev, child, type, rid); /* Add the resource. */ end = (start + count - 1); resource_list_add(rl, type, rid, start, end, count); /* Don't reserve resources until the system resources are allocated. */ if (!sc->acpi_resources_reserved) return (0); /* Don't reserve system resources. */ if (ACPI_ID_PROBE(dev, child, sysres_ids, NULL) <= 0) return (0); /* * Don't reserve IRQ resources. There are many sticky things to * get right otherwise (e.g. IRQs for psm, atkbd, and HPET when * using legacy routing). */ if (type == SYS_RES_IRQ) return (0); /* * Don't reserve resources for CPU devices. Some of these * resources need to be allocated as shareable, but reservations * are always non-shareable. */ if (device_get_devclass(child) == devclass_find("cpu")) return (0); /* * Reserve the resource. * * XXX: Ignores failure for now. Failure here is probably a * BIOS/firmware bug? */ resource_list_reserve(rl, dev, child, type, &rid, start, end, count, 0); return (0); } static struct resource * acpi_alloc_resource(device_t bus, device_t child, int type, int *rid, rman_res_t start, rman_res_t end, rman_res_t count, u_int flags) { #ifndef INTRNG ACPI_RESOURCE ares; #endif struct acpi_device *ad; struct resource_list_entry *rle; struct resource_list *rl; struct resource *res; int isdefault = RMAN_IS_DEFAULT_RANGE(start, end); /* * First attempt at allocating the resource. For direct children, * use resource_list_alloc() to handle reserved resources. For * other devices, pass the request up to our parent. */ if (bus == device_get_parent(child)) { ad = device_get_ivars(child); rl = &ad->ad_rl; /* * Simulate the behavior of the ISA bus for direct children * devices. That is, if a non-default range is specified for * a resource that doesn't exist, use bus_set_resource() to * add the resource before allocating it. Note that these * resources will not be reserved. */ if (!isdefault && resource_list_find(rl, type, *rid) == NULL) resource_list_add(rl, type, *rid, start, end, count); res = resource_list_alloc(rl, bus, child, type, rid, start, end, count, flags); #ifndef INTRNG if (res != NULL && type == SYS_RES_IRQ) { /* * Since bus_config_intr() takes immediate effect, we cannot * configure the interrupt associated with a device when we * parse the resources but have to defer it until a driver * actually allocates the interrupt via bus_alloc_resource(). * * XXX: Should we handle the lookup failing? */ if (ACPI_SUCCESS(acpi_lookup_irq_resource(child, *rid, res, &ares))) acpi_config_intr(child, &ares); } #endif /* * If this is an allocation of the "default" range for a given * RID, fetch the exact bounds for this resource from the * resource list entry to try to allocate the range from the * system resource regions. */ if (res == NULL && isdefault) { rle = resource_list_find(rl, type, *rid); if (rle != NULL) { start = rle->start; end = rle->end; count = rle->count; } } } else res = BUS_ALLOC_RESOURCE(device_get_parent(bus), child, type, rid, start, end, count, flags); /* * If the first attempt failed and this is an allocation of a * specific range, try to satisfy the request via a suballocation * from our system resource regions. */ if (res == NULL && start + count - 1 == end) res = acpi_alloc_sysres(child, type, rid, start, end, count, flags); return (res); } /* * Attempt to allocate a specific resource range from the system * resource ranges. Note that we only handle memory and I/O port * system resources. */ struct resource * acpi_alloc_sysres(device_t child, int type, int *rid, rman_res_t start, rman_res_t end, rman_res_t count, u_int flags) { struct rman *rm; struct resource *res; switch (type) { case SYS_RES_IOPORT: rm = &acpi_rman_io; break; case SYS_RES_MEMORY: rm = &acpi_rman_mem; break; default: return (NULL); } KASSERT(start + count - 1 == end, ("wildcard resource range")); res = rman_reserve_resource(rm, start, end, count, flags & ~RF_ACTIVE, child); if (res == NULL) return (NULL); rman_set_rid(res, *rid); /* If requested, activate the resource using the parent's method. */ if (flags & RF_ACTIVE) if (bus_activate_resource(child, type, *rid, res) != 0) { rman_release_resource(res); return (NULL); } return (res); } static int acpi_is_resource_managed(int type, struct resource *r) { /* We only handle memory and IO resources through rman. */ switch (type) { case SYS_RES_IOPORT: return (rman_is_region_manager(r, &acpi_rman_io)); case SYS_RES_MEMORY: return (rman_is_region_manager(r, &acpi_rman_mem)); } return (0); } static int acpi_adjust_resource(device_t bus, device_t child, int type, struct resource *r, rman_res_t start, rman_res_t end) { if (acpi_is_resource_managed(type, r)) return (rman_adjust_resource(r, start, end)); return (bus_generic_adjust_resource(bus, child, type, r, start, end)); } static int acpi_release_resource(device_t bus, device_t child, int type, int rid, struct resource *r) { int ret; /* * If this resource belongs to one of our internal managers, * deactivate it and release it to the local pool. */ if (acpi_is_resource_managed(type, r)) { if (rman_get_flags(r) & RF_ACTIVE) { ret = bus_deactivate_resource(child, type, rid, r); if (ret != 0) return (ret); } return (rman_release_resource(r)); } return (bus_generic_rl_release_resource(bus, child, type, rid, r)); } static void acpi_delete_resource(device_t bus, device_t child, int type, int rid) { struct resource_list *rl; rl = acpi_get_rlist(bus, child); if (resource_list_busy(rl, type, rid)) { device_printf(bus, "delete_resource: Resource still owned by child" " (type=%d, rid=%d)\n", type, rid); return; } resource_list_unreserve(rl, bus, child, type, rid); resource_list_delete(rl, type, rid); } /* Allocate an IO port or memory resource, given its GAS. */ int acpi_bus_alloc_gas(device_t dev, int *type, int *rid, ACPI_GENERIC_ADDRESS *gas, struct resource **res, u_int flags) { int error, res_type; error = ENOMEM; if (type == NULL || rid == NULL || gas == NULL || res == NULL) return (EINVAL); /* We only support memory and IO spaces. */ switch (gas->SpaceId) { case ACPI_ADR_SPACE_SYSTEM_MEMORY: res_type = SYS_RES_MEMORY; break; case ACPI_ADR_SPACE_SYSTEM_IO: res_type = SYS_RES_IOPORT; break; default: return (EOPNOTSUPP); } /* * If the register width is less than 8, assume the BIOS author means * it is a bit field and just allocate a byte. */ if (gas->BitWidth && gas->BitWidth < 8) gas->BitWidth = 8; /* Validate the address after we're sure we support the space. */ if (gas->Address == 0 || gas->BitWidth == 0) return (EINVAL); bus_set_resource(dev, res_type, *rid, gas->Address, gas->BitWidth / 8); *res = bus_alloc_resource_any(dev, res_type, rid, RF_ACTIVE | flags); if (*res != NULL) { *type = res_type; error = 0; } else bus_delete_resource(dev, res_type, *rid); return (error); } /* Probe _HID and _CID for compatible ISA PNP ids. */ static uint32_t acpi_isa_get_logicalid(device_t dev) { ACPI_DEVICE_INFO *devinfo; ACPI_HANDLE h; uint32_t pnpid; ACPI_FUNCTION_TRACE((char *)(uintptr_t)__func__); /* Fetch and validate the HID. */ if ((h = acpi_get_handle(dev)) == NULL || ACPI_FAILURE(AcpiGetObjectInfo(h, &devinfo))) return_VALUE (0); pnpid = (devinfo->Valid & ACPI_VALID_HID) != 0 && devinfo->HardwareId.Length >= ACPI_EISAID_STRING_SIZE ? PNP_EISAID(devinfo->HardwareId.String) : 0; AcpiOsFree(devinfo); return_VALUE (pnpid); } static int acpi_isa_get_compatid(device_t dev, uint32_t *cids, int count) { ACPI_DEVICE_INFO *devinfo; ACPI_PNP_DEVICE_ID *ids; ACPI_HANDLE h; uint32_t *pnpid; int i, valid; ACPI_FUNCTION_TRACE((char *)(uintptr_t)__func__); pnpid = cids; /* Fetch and validate the CID */ if ((h = acpi_get_handle(dev)) == NULL || ACPI_FAILURE(AcpiGetObjectInfo(h, &devinfo))) return_VALUE (0); if ((devinfo->Valid & ACPI_VALID_CID) == 0) { AcpiOsFree(devinfo); return_VALUE (0); } if (devinfo->CompatibleIdList.Count < count) count = devinfo->CompatibleIdList.Count; ids = devinfo->CompatibleIdList.Ids; for (i = 0, valid = 0; i < count; i++) if (ids[i].Length >= ACPI_EISAID_STRING_SIZE && strncmp(ids[i].String, "PNP", 3) == 0) { *pnpid++ = PNP_EISAID(ids[i].String); valid++; } AcpiOsFree(devinfo); return_VALUE (valid); } static int acpi_device_id_probe(device_t bus, device_t dev, char **ids, char **match) { ACPI_HANDLE h; ACPI_OBJECT_TYPE t; int rv; int i; h = acpi_get_handle(dev); if (ids == NULL || h == NULL) return (ENXIO); t = acpi_get_type(dev); if (t != ACPI_TYPE_DEVICE && t != ACPI_TYPE_PROCESSOR) return (ENXIO); /* Try to match one of the array of IDs with a HID or CID. */ for (i = 0; ids[i] != NULL; i++) { rv = acpi_MatchHid(h, ids[i]); if (rv == ACPI_MATCHHID_NOMATCH) continue; if (match != NULL) { *match = ids[i]; } return ((rv == ACPI_MATCHHID_HID)? BUS_PROBE_DEFAULT : BUS_PROBE_LOW_PRIORITY); } return (ENXIO); } static ACPI_STATUS acpi_device_eval_obj(device_t bus, device_t dev, ACPI_STRING pathname, ACPI_OBJECT_LIST *parameters, ACPI_BUFFER *ret) { ACPI_HANDLE h; if (dev == NULL) h = ACPI_ROOT_OBJECT; else if ((h = acpi_get_handle(dev)) == NULL) return (AE_BAD_PARAMETER); return (AcpiEvaluateObject(h, pathname, parameters, ret)); } int acpi_device_pwr_for_sleep(device_t bus, device_t dev, int *dstate) { struct acpi_softc *sc; ACPI_HANDLE handle; ACPI_STATUS status; char sxd[8]; handle = acpi_get_handle(dev); /* * XXX If we find these devices, don't try to power them down. * The serial and IRDA ports on my T23 hang the system when * set to D3 and it appears that such legacy devices may * need special handling in their drivers. */ if (dstate == NULL || handle == NULL || acpi_MatchHid(handle, "PNP0500") || acpi_MatchHid(handle, "PNP0501") || acpi_MatchHid(handle, "PNP0502") || acpi_MatchHid(handle, "PNP0510") || acpi_MatchHid(handle, "PNP0511")) return (ENXIO); /* * Override next state with the value from _SxD, if present. * Note illegal _S0D is evaluated because some systems expect this. */ sc = device_get_softc(bus); snprintf(sxd, sizeof(sxd), "_S%dD", sc->acpi_sstate); status = acpi_GetInteger(handle, sxd, dstate); if (ACPI_FAILURE(status) && status != AE_NOT_FOUND) { device_printf(dev, "failed to get %s on %s: %s\n", sxd, acpi_name(handle), AcpiFormatException(status)); return (ENXIO); } return (0); } /* Callback arg for our implementation of walking the namespace. */ struct acpi_device_scan_ctx { acpi_scan_cb_t user_fn; void *arg; ACPI_HANDLE parent; }; static ACPI_STATUS acpi_device_scan_cb(ACPI_HANDLE h, UINT32 level, void *arg, void **retval) { struct acpi_device_scan_ctx *ctx; device_t dev, old_dev; ACPI_STATUS status; ACPI_OBJECT_TYPE type; /* * Skip this device if we think we'll have trouble with it or it is * the parent where the scan began. */ ctx = (struct acpi_device_scan_ctx *)arg; if (acpi_avoid(h) || h == ctx->parent) return (AE_OK); /* If this is not a valid device type (e.g., a method), skip it. */ if (ACPI_FAILURE(AcpiGetType(h, &type))) return (AE_OK); if (type != ACPI_TYPE_DEVICE && type != ACPI_TYPE_PROCESSOR && type != ACPI_TYPE_THERMAL && type != ACPI_TYPE_POWER) return (AE_OK); /* * Call the user function with the current device. If it is unchanged * afterwards, return. Otherwise, we update the handle to the new dev. */ old_dev = acpi_get_device(h); dev = old_dev; status = ctx->user_fn(h, &dev, level, ctx->arg); if (ACPI_FAILURE(status) || old_dev == dev) return (status); /* Remove the old child and its connection to the handle. */ if (old_dev != NULL) device_delete_child(device_get_parent(old_dev), old_dev); /* Recreate the handle association if the user created a device. */ if (dev != NULL) AcpiAttachData(h, acpi_fake_objhandler, dev); return (AE_OK); } static ACPI_STATUS acpi_device_scan_children(device_t bus, device_t dev, int max_depth, acpi_scan_cb_t user_fn, void *arg) { ACPI_HANDLE h; struct acpi_device_scan_ctx ctx; if (acpi_disabled("children")) return (AE_OK); if (dev == NULL) h = ACPI_ROOT_OBJECT; else if ((h = acpi_get_handle(dev)) == NULL) return (AE_BAD_PARAMETER); ctx.user_fn = user_fn; ctx.arg = arg; ctx.parent = h; return (AcpiWalkNamespace(ACPI_TYPE_ANY, h, max_depth, acpi_device_scan_cb, NULL, &ctx, NULL)); } /* * Even though ACPI devices are not PCI, we use the PCI approach for setting * device power states since it's close enough to ACPI. */ int acpi_set_powerstate(device_t child, int state) { ACPI_HANDLE h; ACPI_STATUS status; h = acpi_get_handle(child); if (state < ACPI_STATE_D0 || state > ACPI_D_STATES_MAX) return (EINVAL); if (h == NULL) return (0); /* Ignore errors if the power methods aren't present. */ status = acpi_pwr_switch_consumer(h, state); if (ACPI_SUCCESS(status)) { if (bootverbose) device_printf(child, "set ACPI power state D%d on %s\n", state, acpi_name(h)); } else if (status != AE_NOT_FOUND) device_printf(child, "failed to set ACPI power state D%d on %s: %s\n", state, acpi_name(h), AcpiFormatException(status)); return (0); } static int acpi_isa_pnp_probe(device_t bus, device_t child, struct isa_pnp_id *ids) { int result, cid_count, i; uint32_t lid, cids[8]; ACPI_FUNCTION_TRACE((char *)(uintptr_t)__func__); /* * ISA-style drivers attached to ACPI may persist and * probe manually if we return ENOENT. We never want * that to happen, so don't ever return it. */ result = ENXIO; /* Scan the supplied IDs for a match */ lid = acpi_isa_get_logicalid(child); cid_count = acpi_isa_get_compatid(child, cids, 8); while (ids && ids->ip_id) { if (lid == ids->ip_id) { result = 0; goto out; } for (i = 0; i < cid_count; i++) { if (cids[i] == ids->ip_id) { result = 0; goto out; } } ids++; } out: if (result == 0 && ids->ip_desc) device_set_desc(child, ids->ip_desc); return_VALUE (result); } /* * Look for a MCFG table. If it is present, use the settings for * domain (segment) 0 to setup PCI config space access via the memory * map. * * On non-x86 architectures (arm64 for now), this will be done from the * PCI host bridge driver. */ static void acpi_enable_pcie(void) { #if defined(__i386__) || defined(__amd64__) ACPI_TABLE_HEADER *hdr; ACPI_MCFG_ALLOCATION *alloc, *end; ACPI_STATUS status; status = AcpiGetTable(ACPI_SIG_MCFG, 1, &hdr); if (ACPI_FAILURE(status)) return; end = (ACPI_MCFG_ALLOCATION *)((char *)hdr + hdr->Length); alloc = (ACPI_MCFG_ALLOCATION *)((ACPI_TABLE_MCFG *)hdr + 1); while (alloc < end) { if (alloc->PciSegment == 0) { pcie_cfgregopen(alloc->Address, alloc->StartBusNumber, alloc->EndBusNumber); return; } alloc++; } #endif } static void acpi_platform_osc(device_t dev) { ACPI_HANDLE sb_handle; ACPI_STATUS status; uint32_t cap_set[2]; /* 0811B06E-4A27-44F9-8D60-3CBBC22E7B48 */ static uint8_t acpi_platform_uuid[ACPI_UUID_LENGTH] = { 0x6e, 0xb0, 0x11, 0x08, 0x27, 0x4a, 0xf9, 0x44, 0x8d, 0x60, 0x3c, 0xbb, 0xc2, 0x2e, 0x7b, 0x48 }; if (ACPI_FAILURE(AcpiGetHandle(ACPI_ROOT_OBJECT, "\\_SB_", &sb_handle))) return; cap_set[1] = 0x10; /* APEI Support */ status = acpi_EvaluateOSC(sb_handle, acpi_platform_uuid, 1, nitems(cap_set), cap_set, cap_set, false); if (ACPI_FAILURE(status)) { if (status == AE_NOT_FOUND) return; device_printf(dev, "_OSC failed: %s\n", AcpiFormatException(status)); return; } } /* * Scan all of the ACPI namespace and attach child devices. * * We should only expect to find devices in the \_PR, \_TZ, \_SI, and * \_SB scopes, and \_PR and \_TZ became obsolete in the ACPI 2.0 spec. * However, in violation of the spec, some systems place their PCI link * devices in \, so we have to walk the whole namespace. We check the * type of namespace nodes, so this should be ok. */ static void acpi_probe_children(device_t bus) { ACPI_FUNCTION_TRACE((char *)(uintptr_t)__func__); /* * Scan the namespace and insert placeholders for all the devices that * we find. We also probe/attach any early devices. * * Note that we use AcpiWalkNamespace rather than AcpiGetDevices because * we want to create nodes for all devices, not just those that are * currently present. (This assumes that we don't want to create/remove * devices as they appear, which might be smarter.) */ ACPI_DEBUG_PRINT((ACPI_DB_OBJECTS, "namespace scan\n")); AcpiWalkNamespace(ACPI_TYPE_ANY, ACPI_ROOT_OBJECT, 100, acpi_probe_child, NULL, bus, NULL); /* Pre-allocate resources for our rman from any sysresource devices. */ acpi_sysres_alloc(bus); /* Reserve resources already allocated to children. */ acpi_reserve_resources(bus); /* Create any static children by calling device identify methods. */ ACPI_DEBUG_PRINT((ACPI_DB_OBJECTS, "device identify routines\n")); bus_generic_probe(bus); /* Probe/attach all children, created statically and from the namespace. */ ACPI_DEBUG_PRINT((ACPI_DB_OBJECTS, "acpi bus_generic_attach\n")); bus_generic_attach(bus); /* Attach wake sysctls. */ acpi_wake_sysctl_walk(bus); ACPI_DEBUG_PRINT((ACPI_DB_OBJECTS, "done attaching children\n")); return_VOID; } /* * Determine the probe order for a given device. */ static void acpi_probe_order(ACPI_HANDLE handle, int *order) { ACPI_OBJECT_TYPE type; /* * 0. CPUs * 1. I/O port and memory system resource holders * 2. Clocks and timers (to handle early accesses) * 3. Embedded controllers (to handle early accesses) * 4. PCI Link Devices */ AcpiGetType(handle, &type); if (type == ACPI_TYPE_PROCESSOR) *order = 0; else if (acpi_MatchHid(handle, "PNP0C01") || acpi_MatchHid(handle, "PNP0C02")) *order = 1; else if (acpi_MatchHid(handle, "PNP0100") || acpi_MatchHid(handle, "PNP0103") || acpi_MatchHid(handle, "PNP0B00")) *order = 2; else if (acpi_MatchHid(handle, "PNP0C09")) *order = 3; else if (acpi_MatchHid(handle, "PNP0C0F")) *order = 4; } /* * Evaluate a child device and determine whether we might attach a device to * it. */ static ACPI_STATUS acpi_probe_child(ACPI_HANDLE handle, UINT32 level, void *context, void **status) { ACPI_DEVICE_INFO *devinfo; struct acpi_device *ad; struct acpi_prw_data prw; ACPI_OBJECT_TYPE type; ACPI_HANDLE h; device_t bus, child; char *handle_str; int order; ACPI_FUNCTION_TRACE((char *)(uintptr_t)__func__); if (acpi_disabled("children")) return_ACPI_STATUS (AE_OK); /* Skip this device if we think we'll have trouble with it. */ if (acpi_avoid(handle)) return_ACPI_STATUS (AE_OK); bus = (device_t)context; if (ACPI_SUCCESS(AcpiGetType(handle, &type))) { handle_str = acpi_name(handle); switch (type) { case ACPI_TYPE_DEVICE: /* * Since we scan from \, be sure to skip system scope objects. * \_SB_ and \_TZ_ are defined in ACPICA as devices to work around * BIOS bugs. For example, \_SB_ is to allow \_SB_._INI to be run * during the initialization and \_TZ_ is to support Notify() on it. */ if (strcmp(handle_str, "\\_SB_") == 0 || strcmp(handle_str, "\\_TZ_") == 0) break; if (acpi_parse_prw(handle, &prw) == 0) AcpiSetupGpeForWake(handle, prw.gpe_handle, prw.gpe_bit); /* * Ignore devices that do not have a _HID or _CID. They should * be discovered by other buses (e.g. the PCI bus driver). */ if (!acpi_has_hid(handle)) break; /* FALLTHROUGH */ case ACPI_TYPE_PROCESSOR: case ACPI_TYPE_THERMAL: case ACPI_TYPE_POWER: /* * Create a placeholder device for this node. Sort the * placeholder so that the probe/attach passes will run * breadth-first. Orders less than ACPI_DEV_BASE_ORDER * are reserved for special objects (i.e., system * resources). */ ACPI_DEBUG_PRINT((ACPI_DB_OBJECTS, "scanning '%s'\n", handle_str)); order = level * 10 + ACPI_DEV_BASE_ORDER; acpi_probe_order(handle, &order); child = BUS_ADD_CHILD(bus, order, NULL, -1); if (child == NULL) break; /* Associate the handle with the device_t and vice versa. */ acpi_set_handle(child, handle); AcpiAttachData(handle, acpi_fake_objhandler, child); /* * Check that the device is present. If it's not present, * leave it disabled (so that we have a device_t attached to * the handle, but we don't probe it). * * XXX PCI link devices sometimes report "present" but not * "functional" (i.e. if disabled). Go ahead and probe them * anyway since we may enable them later. */ if (type == ACPI_TYPE_DEVICE && !acpi_DeviceIsPresent(child)) { /* Never disable PCI link devices. */ if (acpi_MatchHid(handle, "PNP0C0F")) break; /* * Docking stations should remain enabled since the system * may be undocked at boot. */ if (ACPI_SUCCESS(AcpiGetHandle(handle, "_DCK", &h))) break; device_disable(child); break; } /* * Get the device's resource settings and attach them. * Note that if the device has _PRS but no _CRS, we need * to decide when it's appropriate to try to configure the * device. Ignore the return value here; it's OK for the * device not to have any resources. */ acpi_parse_resources(child, handle, &acpi_res_parse_set, NULL); ad = device_get_ivars(child); ad->ad_cls_class = 0xffffff; if (ACPI_SUCCESS(AcpiGetObjectInfo(handle, &devinfo))) { if ((devinfo->Valid & ACPI_VALID_CLS) != 0 && devinfo->ClassCode.Length >= ACPI_PCICLS_STRING_SIZE) { ad->ad_cls_class = strtoul(devinfo->ClassCode.String, NULL, 16); } AcpiOsFree(devinfo); } break; } } return_ACPI_STATUS (AE_OK); } /* * AcpiAttachData() requires an object handler but never uses it. This is a * placeholder object handler so we can store a device_t in an ACPI_HANDLE. */ void acpi_fake_objhandler(ACPI_HANDLE h, void *data) { } static void acpi_shutdown_final(void *arg, int howto) { struct acpi_softc *sc = (struct acpi_softc *)arg; register_t intr; ACPI_STATUS status; /* * XXX Shutdown code should only run on the BSP (cpuid 0). * Some chipsets do not power off the system correctly if called from * an AP. */ if ((howto & RB_POWEROFF) != 0) { status = AcpiEnterSleepStatePrep(ACPI_STATE_S5); if (ACPI_FAILURE(status)) { device_printf(sc->acpi_dev, "AcpiEnterSleepStatePrep failed - %s\n", AcpiFormatException(status)); return; } device_printf(sc->acpi_dev, "Powering system off\n"); intr = intr_disable(); status = AcpiEnterSleepState(ACPI_STATE_S5); if (ACPI_FAILURE(status)) { intr_restore(intr); device_printf(sc->acpi_dev, "power-off failed - %s\n", AcpiFormatException(status)); } else { DELAY(1000000); intr_restore(intr); device_printf(sc->acpi_dev, "power-off failed - timeout\n"); } } else if ((howto & RB_HALT) == 0 && sc->acpi_handle_reboot) { /* Reboot using the reset register. */ status = AcpiReset(); if (ACPI_SUCCESS(status)) { DELAY(1000000); device_printf(sc->acpi_dev, "reset failed - timeout\n"); } else if (status != AE_NOT_EXIST) device_printf(sc->acpi_dev, "reset failed - %s\n", AcpiFormatException(status)); } else if (sc->acpi_do_disable && !KERNEL_PANICKED()) { /* * Only disable ACPI if the user requested. On some systems, writing * the disable value to SMI_CMD hangs the system. */ device_printf(sc->acpi_dev, "Shutting down\n"); AcpiTerminate(); } } static void acpi_enable_fixed_events(struct acpi_softc *sc) { static int first_time = 1; /* Enable and clear fixed events and install handlers. */ if ((AcpiGbl_FADT.Flags & ACPI_FADT_POWER_BUTTON) == 0) { AcpiClearEvent(ACPI_EVENT_POWER_BUTTON); AcpiInstallFixedEventHandler(ACPI_EVENT_POWER_BUTTON, acpi_event_power_button_sleep, sc); if (first_time) device_printf(sc->acpi_dev, "Power Button (fixed)\n"); } if ((AcpiGbl_FADT.Flags & ACPI_FADT_SLEEP_BUTTON) == 0) { AcpiClearEvent(ACPI_EVENT_SLEEP_BUTTON); AcpiInstallFixedEventHandler(ACPI_EVENT_SLEEP_BUTTON, acpi_event_sleep_button_sleep, sc); if (first_time) device_printf(sc->acpi_dev, "Sleep Button (fixed)\n"); } first_time = 0; } /* * Returns true if the device is actually present and should * be attached to. This requires the present, enabled, UI-visible * and diagnostics-passed bits to be set. */ BOOLEAN acpi_DeviceIsPresent(device_t dev) { ACPI_HANDLE h; UINT32 s; ACPI_STATUS status; h = acpi_get_handle(dev); if (h == NULL) return (FALSE); #ifdef ACPI_EARLY_EPYC_WAR /* * Certain Treadripper boards always returns 0 for FreeBSD because it * only returns non-zero for the OS string "Windows 2015". Otherwise it * will return zero. Force them to always be treated as present. * Beata versions were worse: they always returned 0. */ if (acpi_MatchHid(h, "AMDI0020") || acpi_MatchHid(h, "AMDI0010")) return (TRUE); #endif status = acpi_GetInteger(h, "_STA", &s); /* * If no _STA method or if it failed, then assume that * the device is present. */ if (ACPI_FAILURE(status)) return (TRUE); return (ACPI_DEVICE_PRESENT(s) ? TRUE : FALSE); } /* * Returns true if the battery is actually present and inserted. */ BOOLEAN acpi_BatteryIsPresent(device_t dev) { ACPI_HANDLE h; UINT32 s; ACPI_STATUS status; h = acpi_get_handle(dev); if (h == NULL) return (FALSE); status = acpi_GetInteger(h, "_STA", &s); /* * If no _STA method or if it failed, then assume that * the device is present. */ if (ACPI_FAILURE(status)) return (TRUE); return (ACPI_BATTERY_PRESENT(s) ? TRUE : FALSE); } /* * Returns true if a device has at least one valid device ID. */ BOOLEAN acpi_has_hid(ACPI_HANDLE h) { ACPI_DEVICE_INFO *devinfo; BOOLEAN ret; if (h == NULL || ACPI_FAILURE(AcpiGetObjectInfo(h, &devinfo))) return (FALSE); ret = FALSE; if ((devinfo->Valid & ACPI_VALID_HID) != 0) ret = TRUE; else if ((devinfo->Valid & ACPI_VALID_CID) != 0) if (devinfo->CompatibleIdList.Count > 0) ret = TRUE; AcpiOsFree(devinfo); return (ret); } /* * Match a HID string against a handle * returns ACPI_MATCHHID_HID if _HID match * ACPI_MATCHHID_CID if _CID match and not _HID match. * ACPI_MATCHHID_NOMATCH=0 if no match. */ int acpi_MatchHid(ACPI_HANDLE h, const char *hid) { ACPI_DEVICE_INFO *devinfo; BOOLEAN ret; int i; if (hid == NULL || h == NULL || ACPI_FAILURE(AcpiGetObjectInfo(h, &devinfo))) return (ACPI_MATCHHID_NOMATCH); ret = ACPI_MATCHHID_NOMATCH; if ((devinfo->Valid & ACPI_VALID_HID) != 0 && strcmp(hid, devinfo->HardwareId.String) == 0) ret = ACPI_MATCHHID_HID; else if ((devinfo->Valid & ACPI_VALID_CID) != 0) for (i = 0; i < devinfo->CompatibleIdList.Count; i++) { if (strcmp(hid, devinfo->CompatibleIdList.Ids[i].String) == 0) { ret = ACPI_MATCHHID_CID; break; } } AcpiOsFree(devinfo); return (ret); } /* * Return the handle of a named object within our scope, ie. that of (parent) * or one if its parents. */ ACPI_STATUS acpi_GetHandleInScope(ACPI_HANDLE parent, char *path, ACPI_HANDLE *result) { ACPI_HANDLE r; ACPI_STATUS status; /* Walk back up the tree to the root */ for (;;) { status = AcpiGetHandle(parent, path, &r); if (ACPI_SUCCESS(status)) { *result = r; return (AE_OK); } /* XXX Return error here? */ if (status != AE_NOT_FOUND) return (AE_OK); if (ACPI_FAILURE(AcpiGetParent(parent, &r))) return (AE_NOT_FOUND); parent = r; } } /* * Allocate a buffer with a preset data size. */ ACPI_BUFFER * acpi_AllocBuffer(int size) { ACPI_BUFFER *buf; if ((buf = malloc(size + sizeof(*buf), M_ACPIDEV, M_NOWAIT)) == NULL) return (NULL); buf->Length = size; buf->Pointer = (void *)(buf + 1); return (buf); } ACPI_STATUS acpi_SetInteger(ACPI_HANDLE handle, char *path, UINT32 number) { ACPI_OBJECT arg1; ACPI_OBJECT_LIST args; arg1.Type = ACPI_TYPE_INTEGER; arg1.Integer.Value = number; args.Count = 1; args.Pointer = &arg1; return (AcpiEvaluateObject(handle, path, &args, NULL)); } /* * Evaluate a path that should return an integer. */ ACPI_STATUS acpi_GetInteger(ACPI_HANDLE handle, char *path, UINT32 *number) { ACPI_STATUS status; ACPI_BUFFER buf; ACPI_OBJECT param; if (handle == NULL) handle = ACPI_ROOT_OBJECT; /* * Assume that what we've been pointed at is an Integer object, or * a method that will return an Integer. */ buf.Pointer = ¶m; buf.Length = sizeof(param); status = AcpiEvaluateObject(handle, path, NULL, &buf); if (ACPI_SUCCESS(status)) { if (param.Type == ACPI_TYPE_INTEGER) *number = param.Integer.Value; else status = AE_TYPE; } /* * In some applications, a method that's expected to return an Integer * may instead return a Buffer (probably to simplify some internal * arithmetic). We'll try to fetch whatever it is, and if it's a Buffer, * convert it into an Integer as best we can. * * This is a hack. */ if (status == AE_BUFFER_OVERFLOW) { if ((buf.Pointer = AcpiOsAllocate(buf.Length)) == NULL) { status = AE_NO_MEMORY; } else { status = AcpiEvaluateObject(handle, path, NULL, &buf); if (ACPI_SUCCESS(status)) status = acpi_ConvertBufferToInteger(&buf, number); AcpiOsFree(buf.Pointer); } } return (status); } ACPI_STATUS acpi_ConvertBufferToInteger(ACPI_BUFFER *bufp, UINT32 *number) { ACPI_OBJECT *p; UINT8 *val; int i; p = (ACPI_OBJECT *)bufp->Pointer; if (p->Type == ACPI_TYPE_INTEGER) { *number = p->Integer.Value; return (AE_OK); } if (p->Type != ACPI_TYPE_BUFFER) return (AE_TYPE); if (p->Buffer.Length > sizeof(int)) return (AE_BAD_DATA); *number = 0; val = p->Buffer.Pointer; for (i = 0; i < p->Buffer.Length; i++) *number += val[i] << (i * 8); return (AE_OK); } /* * Iterate over the elements of an a package object, calling the supplied * function for each element. * * XXX possible enhancement might be to abort traversal on error. */ ACPI_STATUS acpi_ForeachPackageObject(ACPI_OBJECT *pkg, void (*func)(ACPI_OBJECT *comp, void *arg), void *arg) { ACPI_OBJECT *comp; int i; if (pkg == NULL || pkg->Type != ACPI_TYPE_PACKAGE) return (AE_BAD_PARAMETER); /* Iterate over components */ i = 0; comp = pkg->Package.Elements; for (; i < pkg->Package.Count; i++, comp++) func(comp, arg); return (AE_OK); } /* * Find the (index)th resource object in a set. */ ACPI_STATUS acpi_FindIndexedResource(ACPI_BUFFER *buf, int index, ACPI_RESOURCE **resp) { ACPI_RESOURCE *rp; int i; rp = (ACPI_RESOURCE *)buf->Pointer; i = index; while (i-- > 0) { /* Range check */ if (rp > (ACPI_RESOURCE *)((u_int8_t *)buf->Pointer + buf->Length)) return (AE_BAD_PARAMETER); /* Check for terminator */ if (rp->Type == ACPI_RESOURCE_TYPE_END_TAG || rp->Length == 0) return (AE_NOT_FOUND); rp = ACPI_NEXT_RESOURCE(rp); } if (resp != NULL) *resp = rp; return (AE_OK); } /* * Append an ACPI_RESOURCE to an ACPI_BUFFER. * * Given a pointer to an ACPI_RESOURCE structure, expand the ACPI_BUFFER * provided to contain it. If the ACPI_BUFFER is empty, allocate a sensible * backing block. If the ACPI_RESOURCE is NULL, return an empty set of * resources. */ #define ACPI_INITIAL_RESOURCE_BUFFER_SIZE 512 ACPI_STATUS acpi_AppendBufferResource(ACPI_BUFFER *buf, ACPI_RESOURCE *res) { ACPI_RESOURCE *rp; void *newp; /* Initialise the buffer if necessary. */ if (buf->Pointer == NULL) { buf->Length = ACPI_INITIAL_RESOURCE_BUFFER_SIZE; if ((buf->Pointer = AcpiOsAllocate(buf->Length)) == NULL) return (AE_NO_MEMORY); rp = (ACPI_RESOURCE *)buf->Pointer; rp->Type = ACPI_RESOURCE_TYPE_END_TAG; rp->Length = ACPI_RS_SIZE_MIN; } if (res == NULL) return (AE_OK); /* * Scan the current buffer looking for the terminator. * This will either find the terminator or hit the end * of the buffer and return an error. */ rp = (ACPI_RESOURCE *)buf->Pointer; for (;;) { /* Range check, don't go outside the buffer */ if (rp >= (ACPI_RESOURCE *)((u_int8_t *)buf->Pointer + buf->Length)) return (AE_BAD_PARAMETER); if (rp->Type == ACPI_RESOURCE_TYPE_END_TAG || rp->Length == 0) break; rp = ACPI_NEXT_RESOURCE(rp); } /* * Check the size of the buffer and expand if required. * * Required size is: * size of existing resources before terminator + * size of new resource and header + * size of terminator. * * Note that this loop should really only run once, unless * for some reason we are stuffing a *really* huge resource. */ while ((((u_int8_t *)rp - (u_int8_t *)buf->Pointer) + res->Length + ACPI_RS_SIZE_NO_DATA + ACPI_RS_SIZE_MIN) >= buf->Length) { if ((newp = AcpiOsAllocate(buf->Length * 2)) == NULL) return (AE_NO_MEMORY); bcopy(buf->Pointer, newp, buf->Length); rp = (ACPI_RESOURCE *)((u_int8_t *)newp + ((u_int8_t *)rp - (u_int8_t *)buf->Pointer)); AcpiOsFree(buf->Pointer); buf->Pointer = newp; buf->Length += buf->Length; } /* Insert the new resource. */ bcopy(res, rp, res->Length + ACPI_RS_SIZE_NO_DATA); /* And add the terminator. */ rp = ACPI_NEXT_RESOURCE(rp); rp->Type = ACPI_RESOURCE_TYPE_END_TAG; rp->Length = ACPI_RS_SIZE_MIN; return (AE_OK); } UINT64 acpi_DSMQuery(ACPI_HANDLE h, const uint8_t *uuid, int revision) { /* * ACPI spec 9.1.1 defines this. * * "Arg2: Function Index Represents a specific function whose meaning is * specific to the UUID and Revision ID. Function indices should start * with 1. Function number zero is a query function (see the special * return code defined below)." */ ACPI_BUFFER buf; ACPI_OBJECT *obj; UINT64 ret = 0; int i; if (!ACPI_SUCCESS(acpi_EvaluateDSM(h, uuid, revision, 0, NULL, &buf))) { ACPI_INFO(("Failed to enumerate DSM functions\n")); return (0); } obj = (ACPI_OBJECT *)buf.Pointer; KASSERT(obj, ("Object not allowed to be NULL\n")); /* * From ACPI 6.2 spec 9.1.1: * If Function Index = 0, a Buffer containing a function index bitfield. * Otherwise, the return value and type depends on the UUID and revision * ID (see below). */ switch (obj->Type) { case ACPI_TYPE_BUFFER: for (i = 0; i < MIN(obj->Buffer.Length, sizeof(ret)); i++) ret |= (((uint64_t)obj->Buffer.Pointer[i]) << (i * 8)); break; case ACPI_TYPE_INTEGER: ACPI_BIOS_WARNING((AE_INFO, "Possibly buggy BIOS with ACPI_TYPE_INTEGER for function enumeration\n")); ret = obj->Integer.Value; break; default: ACPI_WARNING((AE_INFO, "Unexpected return type %u\n", obj->Type)); }; AcpiOsFree(obj); return ret; } /* * DSM may return multiple types depending on the function. It is therefore * unsafe to use the typed evaluation. It is highly recommended that the caller * check the type of the returned object. */ ACPI_STATUS acpi_EvaluateDSM(ACPI_HANDLE handle, const uint8_t *uuid, int revision, UINT64 function, ACPI_OBJECT *package, ACPI_BUFFER *out_buf) { return (acpi_EvaluateDSMTyped(handle, uuid, revision, function, package, out_buf, ACPI_TYPE_ANY)); } ACPI_STATUS acpi_EvaluateDSMTyped(ACPI_HANDLE handle, const uint8_t *uuid, int revision, UINT64 function, ACPI_OBJECT *package, ACPI_BUFFER *out_buf, ACPI_OBJECT_TYPE type) { ACPI_OBJECT arg[4]; ACPI_OBJECT_LIST arglist; ACPI_BUFFER buf; ACPI_STATUS status; if (out_buf == NULL) return (AE_NO_MEMORY); arg[0].Type = ACPI_TYPE_BUFFER; arg[0].Buffer.Length = ACPI_UUID_LENGTH; arg[0].Buffer.Pointer = __DECONST(uint8_t *, uuid); arg[1].Type = ACPI_TYPE_INTEGER; arg[1].Integer.Value = revision; arg[2].Type = ACPI_TYPE_INTEGER; arg[2].Integer.Value = function; if (package) { arg[3] = *package; } else { arg[3].Type = ACPI_TYPE_PACKAGE; arg[3].Package.Count = 0; arg[3].Package.Elements = NULL; } arglist.Pointer = arg; arglist.Count = 4; buf.Pointer = NULL; buf.Length = ACPI_ALLOCATE_BUFFER; status = AcpiEvaluateObjectTyped(handle, "_DSM", &arglist, &buf, type); if (ACPI_FAILURE(status)) return (status); KASSERT(ACPI_SUCCESS(status), ("Unexpected status")); *out_buf = buf; return (status); } ACPI_STATUS acpi_EvaluateOSC(ACPI_HANDLE handle, uint8_t *uuid, int revision, int count, uint32_t *caps_in, uint32_t *caps_out, bool query) { ACPI_OBJECT arg[4], *ret; ACPI_OBJECT_LIST arglist; ACPI_BUFFER buf; ACPI_STATUS status; arglist.Pointer = arg; arglist.Count = 4; arg[0].Type = ACPI_TYPE_BUFFER; arg[0].Buffer.Length = ACPI_UUID_LENGTH; arg[0].Buffer.Pointer = uuid; arg[1].Type = ACPI_TYPE_INTEGER; arg[1].Integer.Value = revision; arg[2].Type = ACPI_TYPE_INTEGER; arg[2].Integer.Value = count; arg[3].Type = ACPI_TYPE_BUFFER; arg[3].Buffer.Length = count * sizeof(*caps_in); arg[3].Buffer.Pointer = (uint8_t *)caps_in; caps_in[0] = query ? 1 : 0; buf.Pointer = NULL; buf.Length = ACPI_ALLOCATE_BUFFER; status = AcpiEvaluateObjectTyped(handle, "_OSC", &arglist, &buf, ACPI_TYPE_BUFFER); if (ACPI_FAILURE(status)) return (status); if (caps_out != NULL) { ret = buf.Pointer; if (ret->Buffer.Length != count * sizeof(*caps_out)) { AcpiOsFree(buf.Pointer); return (AE_BUFFER_OVERFLOW); } bcopy(ret->Buffer.Pointer, caps_out, ret->Buffer.Length); } AcpiOsFree(buf.Pointer); return (status); } /* * Set interrupt model. */ ACPI_STATUS acpi_SetIntrModel(int model) { return (acpi_SetInteger(ACPI_ROOT_OBJECT, "_PIC", model)); } /* * Walk subtables of a table and call a callback routine for each * subtable. The caller should provide the first subtable and a * pointer to the end of the table. This can be used to walk tables * such as MADT and SRAT that use subtable entries. */ void acpi_walk_subtables(void *first, void *end, acpi_subtable_handler *handler, void *arg) { ACPI_SUBTABLE_HEADER *entry; for (entry = first; (void *)entry < end; ) { /* Avoid an infinite loop if we hit a bogus entry. */ if (entry->Length < sizeof(ACPI_SUBTABLE_HEADER)) return; handler(entry, arg); entry = ACPI_ADD_PTR(ACPI_SUBTABLE_HEADER, entry, entry->Length); } } /* * DEPRECATED. This interface has serious deficiencies and will be * removed. * * Immediately enter the sleep state. In the old model, acpiconf(8) ran * rc.suspend and rc.resume so we don't have to notify devd(8) to do this. */ ACPI_STATUS acpi_SetSleepState(struct acpi_softc *sc, int state) { static int once; if (!once) { device_printf(sc->acpi_dev, "warning: acpi_SetSleepState() deprecated, need to update your software\n"); once = 1; } return (acpi_EnterSleepState(sc, state)); } #if defined(__amd64__) || defined(__i386__) static void acpi_sleep_force_task(void *context) { struct acpi_softc *sc = (struct acpi_softc *)context; if (ACPI_FAILURE(acpi_EnterSleepState(sc, sc->acpi_next_sstate))) device_printf(sc->acpi_dev, "force sleep state S%d failed\n", sc->acpi_next_sstate); } static void acpi_sleep_force(void *arg) { struct acpi_softc *sc = (struct acpi_softc *)arg; device_printf(sc->acpi_dev, "suspend request timed out, forcing sleep now\n"); /* * XXX Suspending from callout causes freezes in DEVICE_SUSPEND(). * Suspend from acpi_task thread instead. */ if (ACPI_FAILURE(AcpiOsExecute(OSL_NOTIFY_HANDLER, acpi_sleep_force_task, sc))) device_printf(sc->acpi_dev, "AcpiOsExecute() for sleeping failed\n"); } #endif /* * Request that the system enter the given suspend state. All /dev/apm * devices and devd(8) will be notified. Userland then has a chance to * save state and acknowledge the request. The system sleeps once all * acks are in. */ int acpi_ReqSleepState(struct acpi_softc *sc, int state) { #if defined(__amd64__) || defined(__i386__) struct apm_clone_data *clone; ACPI_STATUS status; if (state < ACPI_STATE_S1 || state > ACPI_S_STATES_MAX) return (EINVAL); if (!acpi_sleep_states[state]) return (EOPNOTSUPP); /* * If a reboot/shutdown/suspend request is already in progress or * suspend is blocked due to an upcoming shutdown, just return. */ if (rebooting || sc->acpi_next_sstate != 0 || suspend_blocked) { return (0); } /* Wait until sleep is enabled. */ while (sc->acpi_sleep_disabled) { AcpiOsSleep(1000); } ACPI_LOCK(acpi); sc->acpi_next_sstate = state; /* S5 (soft-off) should be entered directly with no waiting. */ if (state == ACPI_STATE_S5) { ACPI_UNLOCK(acpi); status = acpi_EnterSleepState(sc, state); return (ACPI_SUCCESS(status) ? 0 : ENXIO); } /* Record the pending state and notify all apm devices. */ STAILQ_FOREACH(clone, &sc->apm_cdevs, entries) { clone->notify_status = APM_EV_NONE; if ((clone->flags & ACPI_EVF_DEVD) == 0) { selwakeuppri(&clone->sel_read, PZERO); KNOTE_LOCKED(&clone->sel_read.si_note, 0); } } /* If devd(8) is not running, immediately enter the sleep state. */ if (!devctl_process_running()) { ACPI_UNLOCK(acpi); status = acpi_EnterSleepState(sc, state); return (ACPI_SUCCESS(status) ? 0 : ENXIO); } /* * Set a timeout to fire if userland doesn't ack the suspend request * in time. This way we still eventually go to sleep if we were * overheating or running low on battery, even if userland is hung. * We cancel this timeout once all userland acks are in or the * suspend request is aborted. */ callout_reset(&sc->susp_force_to, 10 * hz, acpi_sleep_force, sc); ACPI_UNLOCK(acpi); /* Now notify devd(8) also. */ acpi_UserNotify("Suspend", ACPI_ROOT_OBJECT, state); return (0); #else /* This platform does not support acpi suspend/resume. */ return (EOPNOTSUPP); #endif } /* * Acknowledge (or reject) a pending sleep state. The caller has * prepared for suspend and is now ready for it to proceed. If the * error argument is non-zero, it indicates suspend should be cancelled * and gives an errno value describing why. Once all votes are in, * we suspend the system. */ int acpi_AckSleepState(struct apm_clone_data *clone, int error) { #if defined(__amd64__) || defined(__i386__) struct acpi_softc *sc; int ret, sleeping; /* If no pending sleep state, return an error. */ ACPI_LOCK(acpi); sc = clone->acpi_sc; if (sc->acpi_next_sstate == 0) { ACPI_UNLOCK(acpi); return (ENXIO); } /* Caller wants to abort suspend process. */ if (error) { sc->acpi_next_sstate = 0; callout_stop(&sc->susp_force_to); device_printf(sc->acpi_dev, "listener on %s cancelled the pending suspend\n", devtoname(clone->cdev)); ACPI_UNLOCK(acpi); return (0); } /* * Mark this device as acking the suspend request. Then, walk through * all devices, seeing if they agree yet. We only count devices that * are writable since read-only devices couldn't ack the request. */ sleeping = TRUE; clone->notify_status = APM_EV_ACKED; STAILQ_FOREACH(clone, &sc->apm_cdevs, entries) { if ((clone->flags & ACPI_EVF_WRITE) != 0 && clone->notify_status != APM_EV_ACKED) { sleeping = FALSE; break; } } /* If all devices have voted "yes", we will suspend now. */ if (sleeping) callout_stop(&sc->susp_force_to); ACPI_UNLOCK(acpi); ret = 0; if (sleeping) { if (ACPI_FAILURE(acpi_EnterSleepState(sc, sc->acpi_next_sstate))) ret = ENODEV; } return (ret); #else /* This platform does not support acpi suspend/resume. */ return (EOPNOTSUPP); #endif } static void acpi_sleep_enable(void *arg) { struct acpi_softc *sc = (struct acpi_softc *)arg; ACPI_LOCK_ASSERT(acpi); /* Reschedule if the system is not fully up and running. */ if (!AcpiGbl_SystemAwakeAndRunning) { callout_schedule(&acpi_sleep_timer, hz * ACPI_MINIMUM_AWAKETIME); return; } sc->acpi_sleep_disabled = FALSE; } static ACPI_STATUS acpi_sleep_disable(struct acpi_softc *sc) { ACPI_STATUS status; /* Fail if the system is not fully up and running. */ if (!AcpiGbl_SystemAwakeAndRunning) return (AE_ERROR); ACPI_LOCK(acpi); status = sc->acpi_sleep_disabled ? AE_ERROR : AE_OK; sc->acpi_sleep_disabled = TRUE; ACPI_UNLOCK(acpi); return (status); } enum acpi_sleep_state { ACPI_SS_NONE, ACPI_SS_GPE_SET, ACPI_SS_DEV_SUSPEND, ACPI_SS_SLP_PREP, ACPI_SS_SLEPT, }; /* * Enter the desired system sleep state. * * Currently we support S1-S5 but S4 is only S4BIOS */ static ACPI_STATUS acpi_EnterSleepState(struct acpi_softc *sc, int state) { register_t intr; ACPI_STATUS status; ACPI_EVENT_STATUS power_button_status; enum acpi_sleep_state slp_state; int sleep_result; ACPI_FUNCTION_TRACE_U32((char *)(uintptr_t)__func__, state); if (state < ACPI_STATE_S1 || state > ACPI_S_STATES_MAX) return_ACPI_STATUS (AE_BAD_PARAMETER); if (!acpi_sleep_states[state]) { device_printf(sc->acpi_dev, "Sleep state S%d not supported by BIOS\n", state); return (AE_SUPPORT); } /* Re-entry once we're suspending is not allowed. */ status = acpi_sleep_disable(sc); if (ACPI_FAILURE(status)) { device_printf(sc->acpi_dev, "suspend request ignored (not ready yet)\n"); return (status); } if (state == ACPI_STATE_S5) { /* * Shut down cleanly and power off. This will call us back through the * shutdown handlers. */ shutdown_nice(RB_POWEROFF); return_ACPI_STATUS (AE_OK); } EVENTHANDLER_INVOKE(power_suspend_early); stop_all_proc(); suspend_all_fs(); EVENTHANDLER_INVOKE(power_suspend); #ifdef EARLY_AP_STARTUP MPASS(mp_ncpus == 1 || smp_started); thread_lock(curthread); sched_bind(curthread, 0); thread_unlock(curthread); #else if (smp_started) { thread_lock(curthread); sched_bind(curthread, 0); thread_unlock(curthread); } #endif /* * Be sure to hold Giant across DEVICE_SUSPEND/RESUME since non-MPSAFE * drivers need this. */ mtx_lock(&Giant); slp_state = ACPI_SS_NONE; sc->acpi_sstate = state; /* Enable any GPEs as appropriate and requested by the user. */ acpi_wake_prep_walk(state); slp_state = ACPI_SS_GPE_SET; /* * Inform all devices that we are going to sleep. If at least one * device fails, DEVICE_SUSPEND() automatically resumes the tree. * * XXX Note that a better two-pass approach with a 'veto' pass * followed by a "real thing" pass would be better, but the current * bus interface does not provide for this. */ if (DEVICE_SUSPEND(root_bus) != 0) { device_printf(sc->acpi_dev, "device_suspend failed\n"); goto backout; } slp_state = ACPI_SS_DEV_SUSPEND; status = AcpiEnterSleepStatePrep(state); if (ACPI_FAILURE(status)) { device_printf(sc->acpi_dev, "AcpiEnterSleepStatePrep failed - %s\n", AcpiFormatException(status)); goto backout; } slp_state = ACPI_SS_SLP_PREP; if (sc->acpi_sleep_delay > 0) DELAY(sc->acpi_sleep_delay * 1000000); suspendclock(); intr = intr_disable(); if (state != ACPI_STATE_S1) { sleep_result = acpi_sleep_machdep(sc, state); acpi_wakeup_machdep(sc, state, sleep_result, 0); /* * XXX According to ACPI specification SCI_EN bit should be restored * by ACPI platform (BIOS, firmware) to its pre-sleep state. * Unfortunately some BIOSes fail to do that and that leads to * unexpected and serious consequences during wake up like a system * getting stuck in SMI handlers. * This hack is picked up from Linux, which claims that it follows * Windows behavior. */ if (sleep_result == 1 && state != ACPI_STATE_S4) AcpiWriteBitRegister(ACPI_BITREG_SCI_ENABLE, ACPI_ENABLE_EVENT); if (sleep_result == 1 && state == ACPI_STATE_S3) { /* * Prevent mis-interpretation of the wakeup by power button * as a request for power off. * Ideally we should post an appropriate wakeup event, * perhaps using acpi_event_power_button_wake or alike. * * Clearing of power button status after wakeup is mandated * by ACPI specification in section "Fixed Power Button". * * XXX As of ACPICA 20121114 AcpiGetEventStatus provides * status as 0/1 corressponding to inactive/active despite * its type being ACPI_EVENT_STATUS. In other words, * we should not test for ACPI_EVENT_FLAG_SET for time being. */ if (ACPI_SUCCESS(AcpiGetEventStatus(ACPI_EVENT_POWER_BUTTON, &power_button_status)) && power_button_status != 0) { AcpiClearEvent(ACPI_EVENT_POWER_BUTTON); device_printf(sc->acpi_dev, "cleared fixed power button status\n"); } } intr_restore(intr); /* call acpi_wakeup_machdep() again with interrupt enabled */ acpi_wakeup_machdep(sc, state, sleep_result, 1); AcpiLeaveSleepStatePrep(state); if (sleep_result == -1) goto backout; /* Re-enable ACPI hardware on wakeup from sleep state 4. */ if (state == ACPI_STATE_S4) AcpiEnable(); } else { status = AcpiEnterSleepState(state); intr_restore(intr); AcpiLeaveSleepStatePrep(state); if (ACPI_FAILURE(status)) { device_printf(sc->acpi_dev, "AcpiEnterSleepState failed - %s\n", AcpiFormatException(status)); goto backout; } } slp_state = ACPI_SS_SLEPT; /* * Back out state according to how far along we got in the suspend * process. This handles both the error and success cases. */ backout: if (slp_state >= ACPI_SS_SLP_PREP) resumeclock(); if (slp_state >= ACPI_SS_GPE_SET) { acpi_wake_prep_walk(state); sc->acpi_sstate = ACPI_STATE_S0; } if (slp_state >= ACPI_SS_DEV_SUSPEND) DEVICE_RESUME(root_bus); if (slp_state >= ACPI_SS_SLP_PREP) AcpiLeaveSleepState(state); if (slp_state >= ACPI_SS_SLEPT) { #if defined(__i386__) || defined(__amd64__) /* NB: we are still using ACPI timecounter at this point. */ resume_TSC(); #endif acpi_resync_clock(sc); acpi_enable_fixed_events(sc); } sc->acpi_next_sstate = 0; mtx_unlock(&Giant); #ifdef EARLY_AP_STARTUP thread_lock(curthread); sched_unbind(curthread); thread_unlock(curthread); #else if (smp_started) { thread_lock(curthread); sched_unbind(curthread); thread_unlock(curthread); } #endif resume_all_fs(); resume_all_proc(); EVENTHANDLER_INVOKE(power_resume); /* Allow another sleep request after a while. */ callout_schedule(&acpi_sleep_timer, hz * ACPI_MINIMUM_AWAKETIME); /* Run /etc/rc.resume after we are back. */ if (devctl_process_running()) acpi_UserNotify("Resume", ACPI_ROOT_OBJECT, state); return_ACPI_STATUS (status); } static void acpi_resync_clock(struct acpi_softc *sc) { /* * Warm up timecounter again and reset system clock. */ (void)timecounter->tc_get_timecount(timecounter); inittodr(time_second + sc->acpi_sleep_delay); } /* Enable or disable the device's wake GPE. */ int acpi_wake_set_enable(device_t dev, int enable) { struct acpi_prw_data prw; ACPI_STATUS status; int flags; /* Make sure the device supports waking the system and get the GPE. */ if (acpi_parse_prw(acpi_get_handle(dev), &prw) != 0) return (ENXIO); flags = acpi_get_flags(dev); if (enable) { status = AcpiSetGpeWakeMask(prw.gpe_handle, prw.gpe_bit, ACPI_GPE_ENABLE); if (ACPI_FAILURE(status)) { device_printf(dev, "enable wake failed\n"); return (ENXIO); } acpi_set_flags(dev, flags | ACPI_FLAG_WAKE_ENABLED); } else { status = AcpiSetGpeWakeMask(prw.gpe_handle, prw.gpe_bit, ACPI_GPE_DISABLE); if (ACPI_FAILURE(status)) { device_printf(dev, "disable wake failed\n"); return (ENXIO); } acpi_set_flags(dev, flags & ~ACPI_FLAG_WAKE_ENABLED); } return (0); } static int acpi_wake_sleep_prep(ACPI_HANDLE handle, int sstate) { struct acpi_prw_data prw; device_t dev; /* Check that this is a wake-capable device and get its GPE. */ if (acpi_parse_prw(handle, &prw) != 0) return (ENXIO); dev = acpi_get_device(handle); /* * The destination sleep state must be less than (i.e., higher power) * or equal to the value specified by _PRW. If this GPE cannot be * enabled for the next sleep state, then disable it. If it can and * the user requested it be enabled, turn on any required power resources * and set _PSW. */ if (sstate > prw.lowest_wake) { AcpiSetGpeWakeMask(prw.gpe_handle, prw.gpe_bit, ACPI_GPE_DISABLE); if (bootverbose) device_printf(dev, "wake_prep disabled wake for %s (S%d)\n", acpi_name(handle), sstate); } else if (dev && (acpi_get_flags(dev) & ACPI_FLAG_WAKE_ENABLED) != 0) { acpi_pwr_wake_enable(handle, 1); acpi_SetInteger(handle, "_PSW", 1); if (bootverbose) device_printf(dev, "wake_prep enabled for %s (S%d)\n", acpi_name(handle), sstate); } return (0); } static int acpi_wake_run_prep(ACPI_HANDLE handle, int sstate) { struct acpi_prw_data prw; device_t dev; /* * Check that this is a wake-capable device and get its GPE. Return * now if the user didn't enable this device for wake. */ if (acpi_parse_prw(handle, &prw) != 0) return (ENXIO); dev = acpi_get_device(handle); if (dev == NULL || (acpi_get_flags(dev) & ACPI_FLAG_WAKE_ENABLED) == 0) return (0); /* * If this GPE couldn't be enabled for the previous sleep state, it was * disabled before going to sleep so re-enable it. If it was enabled, * clear _PSW and turn off any power resources it used. */ if (sstate > prw.lowest_wake) { AcpiSetGpeWakeMask(prw.gpe_handle, prw.gpe_bit, ACPI_GPE_ENABLE); if (bootverbose) device_printf(dev, "run_prep re-enabled %s\n", acpi_name(handle)); } else { acpi_SetInteger(handle, "_PSW", 0); acpi_pwr_wake_enable(handle, 0); if (bootverbose) device_printf(dev, "run_prep cleaned up for %s\n", acpi_name(handle)); } return (0); } static ACPI_STATUS acpi_wake_prep(ACPI_HANDLE handle, UINT32 level, void *context, void **status) { int sstate; /* If suspending, run the sleep prep function, otherwise wake. */ sstate = *(int *)context; if (AcpiGbl_SystemAwakeAndRunning) acpi_wake_sleep_prep(handle, sstate); else acpi_wake_run_prep(handle, sstate); return (AE_OK); } /* Walk the tree rooted at acpi0 to prep devices for suspend/resume. */ static int acpi_wake_prep_walk(int sstate) { ACPI_HANDLE sb_handle; if (ACPI_SUCCESS(AcpiGetHandle(ACPI_ROOT_OBJECT, "\\_SB_", &sb_handle))) AcpiWalkNamespace(ACPI_TYPE_DEVICE, sb_handle, 100, acpi_wake_prep, NULL, &sstate, NULL); return (0); } /* Walk the tree rooted at acpi0 to attach per-device wake sysctls. */ static int acpi_wake_sysctl_walk(device_t dev) { int error, i, numdevs; device_t *devlist; device_t child; ACPI_STATUS status; error = device_get_children(dev, &devlist, &numdevs); if (error != 0 || numdevs == 0) { if (numdevs == 0) free(devlist, M_TEMP); return (error); } for (i = 0; i < numdevs; i++) { child = devlist[i]; acpi_wake_sysctl_walk(child); if (!device_is_attached(child)) continue; status = AcpiEvaluateObject(acpi_get_handle(child), "_PRW", NULL, NULL); if (ACPI_SUCCESS(status)) { SYSCTL_ADD_PROC(device_get_sysctl_ctx(child), SYSCTL_CHILDREN(device_get_sysctl_tree(child)), OID_AUTO, "wake", CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, child, 0, acpi_wake_set_sysctl, "I", "Device set to wake the system"); } } free(devlist, M_TEMP); return (0); } /* Enable or disable wake from userland. */ static int acpi_wake_set_sysctl(SYSCTL_HANDLER_ARGS) { int enable, error; device_t dev; dev = (device_t)arg1; enable = (acpi_get_flags(dev) & ACPI_FLAG_WAKE_ENABLED) ? 1 : 0; error = sysctl_handle_int(oidp, &enable, 0, req); if (error != 0 || req->newptr == NULL) return (error); if (enable != 0 && enable != 1) return (EINVAL); return (acpi_wake_set_enable(dev, enable)); } /* Parse a device's _PRW into a structure. */ int acpi_parse_prw(ACPI_HANDLE h, struct acpi_prw_data *prw) { ACPI_STATUS status; ACPI_BUFFER prw_buffer; ACPI_OBJECT *res, *res2; int error, i, power_count; if (h == NULL || prw == NULL) return (EINVAL); /* * The _PRW object (7.2.9) is only required for devices that have the * ability to wake the system from a sleeping state. */ error = EINVAL; prw_buffer.Pointer = NULL; prw_buffer.Length = ACPI_ALLOCATE_BUFFER; status = AcpiEvaluateObject(h, "_PRW", NULL, &prw_buffer); if (ACPI_FAILURE(status)) return (ENOENT); res = (ACPI_OBJECT *)prw_buffer.Pointer; if (res == NULL) return (ENOENT); if (!ACPI_PKG_VALID(res, 2)) goto out; /* * Element 1 of the _PRW object: * The lowest power system sleeping state that can be entered while still * providing wake functionality. The sleeping state being entered must * be less than (i.e., higher power) or equal to this value. */ if (acpi_PkgInt32(res, 1, &prw->lowest_wake) != 0) goto out; /* * Element 0 of the _PRW object: */ switch (res->Package.Elements[0].Type) { case ACPI_TYPE_INTEGER: /* * If the data type of this package element is numeric, then this * _PRW package element is the bit index in the GPEx_EN, in the * GPE blocks described in the FADT, of the enable bit that is * enabled for the wake event. */ prw->gpe_handle = NULL; prw->gpe_bit = res->Package.Elements[0].Integer.Value; error = 0; break; case ACPI_TYPE_PACKAGE: /* * If the data type of this package element is a package, then this * _PRW package element is itself a package containing two * elements. The first is an object reference to the GPE Block * device that contains the GPE that will be triggered by the wake * event. The second element is numeric and it contains the bit * index in the GPEx_EN, in the GPE Block referenced by the * first element in the package, of the enable bit that is enabled for * the wake event. * * For example, if this field is a package then it is of the form: * Package() {\_SB.PCI0.ISA.GPE, 2} */ res2 = &res->Package.Elements[0]; if (!ACPI_PKG_VALID(res2, 2)) goto out; prw->gpe_handle = acpi_GetReference(NULL, &res2->Package.Elements[0]); if (prw->gpe_handle == NULL) goto out; if (acpi_PkgInt32(res2, 1, &prw->gpe_bit) != 0) goto out; error = 0; break; default: goto out; } /* Elements 2 to N of the _PRW object are power resources. */ power_count = res->Package.Count - 2; if (power_count > ACPI_PRW_MAX_POWERRES) { printf("ACPI device %s has too many power resources\n", acpi_name(h)); power_count = 0; } prw->power_res_count = power_count; for (i = 0; i < power_count; i++) prw->power_res[i] = res->Package.Elements[i]; out: if (prw_buffer.Pointer != NULL) AcpiOsFree(prw_buffer.Pointer); return (error); } /* * ACPI Event Handlers */ /* System Event Handlers (registered by EVENTHANDLER_REGISTER) */ static void acpi_system_eventhandler_sleep(void *arg, int state) { struct acpi_softc *sc = (struct acpi_softc *)arg; int ret; ACPI_FUNCTION_TRACE_U32((char *)(uintptr_t)__func__, state); /* Check if button action is disabled or unknown. */ if (state == ACPI_STATE_UNKNOWN) return; /* Request that the system prepare to enter the given suspend state. */ ret = acpi_ReqSleepState(sc, state); if (ret != 0) device_printf(sc->acpi_dev, "request to enter state S%d failed (err %d)\n", state, ret); return_VOID; } static void acpi_system_eventhandler_wakeup(void *arg, int state) { ACPI_FUNCTION_TRACE_U32((char *)(uintptr_t)__func__, state); /* Currently, nothing to do for wakeup. */ return_VOID; } /* * ACPICA Event Handlers (FixedEvent, also called from button notify handler) */ static void acpi_invoke_sleep_eventhandler(void *context) { EVENTHANDLER_INVOKE(acpi_sleep_event, *(int *)context); } static void acpi_invoke_wake_eventhandler(void *context) { EVENTHANDLER_INVOKE(acpi_wakeup_event, *(int *)context); } UINT32 acpi_event_power_button_sleep(void *context) { struct acpi_softc *sc = (struct acpi_softc *)context; ACPI_FUNCTION_TRACE((char *)(uintptr_t)__func__); if (ACPI_FAILURE(AcpiOsExecute(OSL_NOTIFY_HANDLER, acpi_invoke_sleep_eventhandler, &sc->acpi_power_button_sx))) return_VALUE (ACPI_INTERRUPT_NOT_HANDLED); return_VALUE (ACPI_INTERRUPT_HANDLED); } UINT32 acpi_event_power_button_wake(void *context) { struct acpi_softc *sc = (struct acpi_softc *)context; ACPI_FUNCTION_TRACE((char *)(uintptr_t)__func__); if (ACPI_FAILURE(AcpiOsExecute(OSL_NOTIFY_HANDLER, acpi_invoke_wake_eventhandler, &sc->acpi_power_button_sx))) return_VALUE (ACPI_INTERRUPT_NOT_HANDLED); return_VALUE (ACPI_INTERRUPT_HANDLED); } UINT32 acpi_event_sleep_button_sleep(void *context) { struct acpi_softc *sc = (struct acpi_softc *)context; ACPI_FUNCTION_TRACE((char *)(uintptr_t)__func__); if (ACPI_FAILURE(AcpiOsExecute(OSL_NOTIFY_HANDLER, acpi_invoke_sleep_eventhandler, &sc->acpi_sleep_button_sx))) return_VALUE (ACPI_INTERRUPT_NOT_HANDLED); return_VALUE (ACPI_INTERRUPT_HANDLED); } UINT32 acpi_event_sleep_button_wake(void *context) { struct acpi_softc *sc = (struct acpi_softc *)context; ACPI_FUNCTION_TRACE((char *)(uintptr_t)__func__); if (ACPI_FAILURE(AcpiOsExecute(OSL_NOTIFY_HANDLER, acpi_invoke_wake_eventhandler, &sc->acpi_sleep_button_sx))) return_VALUE (ACPI_INTERRUPT_NOT_HANDLED); return_VALUE (ACPI_INTERRUPT_HANDLED); } /* * XXX This static buffer is suboptimal. There is no locking so only * use this for single-threaded callers. */ char * acpi_name(ACPI_HANDLE handle) { ACPI_BUFFER buf; static char data[256]; buf.Length = sizeof(data); buf.Pointer = data; if (handle && ACPI_SUCCESS(AcpiGetName(handle, ACPI_FULL_PATHNAME, &buf))) return (data); return ("(unknown)"); } /* * Debugging/bug-avoidance. Avoid trying to fetch info on various * parts of the namespace. */ int acpi_avoid(ACPI_HANDLE handle) { char *cp, *env, *np; int len; np = acpi_name(handle); if (*np == '\\') np++; if ((env = kern_getenv("debug.acpi.avoid")) == NULL) return (0); /* Scan the avoid list checking for a match */ cp = env; for (;;) { while (*cp != 0 && isspace(*cp)) cp++; if (*cp == 0) break; len = 0; while (cp[len] != 0 && !isspace(cp[len])) len++; if (!strncmp(cp, np, len)) { freeenv(env); return(1); } cp += len; } freeenv(env); return (0); } /* * Debugging/bug-avoidance. Disable ACPI subsystem components. */ int acpi_disabled(char *subsys) { char *cp, *env; int len; if ((env = kern_getenv("debug.acpi.disabled")) == NULL) return (0); if (strcmp(env, "all") == 0) { freeenv(env); return (1); } /* Scan the disable list, checking for a match. */ cp = env; for (;;) { while (*cp != '\0' && isspace(*cp)) cp++; if (*cp == '\0') break; len = 0; while (cp[len] != '\0' && !isspace(cp[len])) len++; if (strncmp(cp, subsys, len) == 0) { freeenv(env); return (1); } cp += len; } freeenv(env); return (0); } static void acpi_lookup(void *arg, const char *name, device_t *dev) { ACPI_HANDLE handle; if (*dev != NULL) return; /* * Allow any handle name that is specified as an absolute path and * starts with '\'. We could restrict this to \_SB and friends, * but see acpi_probe_children() for notes on why we scan the entire * namespace for devices. * * XXX: The pathname argument to AcpiGetHandle() should be fixed to * be const. */ if (name[0] != '\\') return; if (ACPI_FAILURE(AcpiGetHandle(ACPI_ROOT_OBJECT, __DECONST(char *, name), &handle))) return; *dev = acpi_get_device(handle); } /* * Control interface. * * We multiplex ioctls for all participating ACPI devices here. Individual * drivers wanting to be accessible via /dev/acpi should use the * register/deregister interface to make their handlers visible. */ struct acpi_ioctl_hook { TAILQ_ENTRY(acpi_ioctl_hook) link; u_long cmd; acpi_ioctl_fn fn; void *arg; }; static TAILQ_HEAD(,acpi_ioctl_hook) acpi_ioctl_hooks; static int acpi_ioctl_hooks_initted; int acpi_register_ioctl(u_long cmd, acpi_ioctl_fn fn, void *arg) { struct acpi_ioctl_hook *hp; if ((hp = malloc(sizeof(*hp), M_ACPIDEV, M_NOWAIT)) == NULL) return (ENOMEM); hp->cmd = cmd; hp->fn = fn; hp->arg = arg; ACPI_LOCK(acpi); if (acpi_ioctl_hooks_initted == 0) { TAILQ_INIT(&acpi_ioctl_hooks); acpi_ioctl_hooks_initted = 1; } TAILQ_INSERT_TAIL(&acpi_ioctl_hooks, hp, link); ACPI_UNLOCK(acpi); return (0); } void acpi_deregister_ioctl(u_long cmd, acpi_ioctl_fn fn) { struct acpi_ioctl_hook *hp; ACPI_LOCK(acpi); TAILQ_FOREACH(hp, &acpi_ioctl_hooks, link) if (hp->cmd == cmd && hp->fn == fn) break; if (hp != NULL) { TAILQ_REMOVE(&acpi_ioctl_hooks, hp, link); free(hp, M_ACPIDEV); } ACPI_UNLOCK(acpi); } static int acpiopen(struct cdev *dev, int flag, int fmt, struct thread *td) { return (0); } static int acpiclose(struct cdev *dev, int flag, int fmt, struct thread *td) { return (0); } static int acpiioctl(struct cdev *dev, u_long cmd, caddr_t addr, int flag, struct thread *td) { struct acpi_softc *sc; struct acpi_ioctl_hook *hp; int error, state; error = 0; hp = NULL; sc = dev->si_drv1; /* * Scan the list of registered ioctls, looking for handlers. */ ACPI_LOCK(acpi); if (acpi_ioctl_hooks_initted) TAILQ_FOREACH(hp, &acpi_ioctl_hooks, link) { if (hp->cmd == cmd) break; } ACPI_UNLOCK(acpi); if (hp) return (hp->fn(cmd, addr, hp->arg)); /* * Core ioctls are not permitted for non-writable user. * Currently, other ioctls just fetch information. * Not changing system behavior. */ if ((flag & FWRITE) == 0) return (EPERM); /* Core system ioctls. */ switch (cmd) { case ACPIIO_REQSLPSTATE: state = *(int *)addr; if (state != ACPI_STATE_S5) return (acpi_ReqSleepState(sc, state)); device_printf(sc->acpi_dev, "power off via acpi ioctl not supported\n"); error = EOPNOTSUPP; break; case ACPIIO_ACKSLPSTATE: error = *(int *)addr; error = acpi_AckSleepState(sc->acpi_clone, error); break; case ACPIIO_SETSLPSTATE: /* DEPRECATED */ state = *(int *)addr; if (state < ACPI_STATE_S0 || state > ACPI_S_STATES_MAX) return (EINVAL); if (!acpi_sleep_states[state]) return (EOPNOTSUPP); if (ACPI_FAILURE(acpi_SetSleepState(sc, state))) error = ENXIO; break; default: error = ENXIO; break; } return (error); } static int acpi_sname2sstate(const char *sname) { int sstate; if (toupper(sname[0]) == 'S') { sstate = sname[1] - '0'; if (sstate >= ACPI_STATE_S0 && sstate <= ACPI_STATE_S5 && sname[2] == '\0') return (sstate); } else if (strcasecmp(sname, "NONE") == 0) return (ACPI_STATE_UNKNOWN); return (-1); } static const char * acpi_sstate2sname(int sstate) { static const char *snames[] = { "S0", "S1", "S2", "S3", "S4", "S5" }; if (sstate >= ACPI_STATE_S0 && sstate <= ACPI_STATE_S5) return (snames[sstate]); else if (sstate == ACPI_STATE_UNKNOWN) return ("NONE"); return (NULL); } static int acpi_supported_sleep_state_sysctl(SYSCTL_HANDLER_ARGS) { int error; struct sbuf sb; UINT8 state; sbuf_new(&sb, NULL, 32, SBUF_AUTOEXTEND); for (state = ACPI_STATE_S1; state < ACPI_S_STATE_COUNT; state++) if (acpi_sleep_states[state]) sbuf_printf(&sb, "%s ", acpi_sstate2sname(state)); sbuf_trim(&sb); sbuf_finish(&sb); error = sysctl_handle_string(oidp, sbuf_data(&sb), sbuf_len(&sb), req); sbuf_delete(&sb); return (error); } static int acpi_sleep_state_sysctl(SYSCTL_HANDLER_ARGS) { char sleep_state[10]; int error, new_state, old_state; old_state = *(int *)oidp->oid_arg1; strlcpy(sleep_state, acpi_sstate2sname(old_state), sizeof(sleep_state)); error = sysctl_handle_string(oidp, sleep_state, sizeof(sleep_state), req); if (error == 0 && req->newptr != NULL) { new_state = acpi_sname2sstate(sleep_state); if (new_state < ACPI_STATE_S1) return (EINVAL); if (new_state < ACPI_S_STATE_COUNT && !acpi_sleep_states[new_state]) return (EOPNOTSUPP); if (new_state != old_state) *(int *)oidp->oid_arg1 = new_state; } return (error); } /* Inform devctl(4) when we receive a Notify. */ void acpi_UserNotify(const char *subsystem, ACPI_HANDLE h, uint8_t notify) { char notify_buf[16]; ACPI_BUFFER handle_buf; ACPI_STATUS status; if (subsystem == NULL) return; handle_buf.Pointer = NULL; handle_buf.Length = ACPI_ALLOCATE_BUFFER; status = AcpiNsHandleToPathname(h, &handle_buf, FALSE); if (ACPI_FAILURE(status)) return; snprintf(notify_buf, sizeof(notify_buf), "notify=0x%02x", notify); devctl_notify("ACPI", subsystem, handle_buf.Pointer, notify_buf); AcpiOsFree(handle_buf.Pointer); } #ifdef ACPI_DEBUG /* * Support for parsing debug options from the kernel environment. * * Bits may be set in the AcpiDbgLayer and AcpiDbgLevel debug registers * by specifying the names of the bits in the debug.acpi.layer and * debug.acpi.level environment variables. Bits may be unset by * prefixing the bit name with !. */ struct debugtag { char *name; UINT32 value; }; static struct debugtag dbg_layer[] = { {"ACPI_UTILITIES", ACPI_UTILITIES}, {"ACPI_HARDWARE", ACPI_HARDWARE}, {"ACPI_EVENTS", ACPI_EVENTS}, {"ACPI_TABLES", ACPI_TABLES}, {"ACPI_NAMESPACE", ACPI_NAMESPACE}, {"ACPI_PARSER", ACPI_PARSER}, {"ACPI_DISPATCHER", ACPI_DISPATCHER}, {"ACPI_EXECUTER", ACPI_EXECUTER}, {"ACPI_RESOURCES", ACPI_RESOURCES}, {"ACPI_CA_DEBUGGER", ACPI_CA_DEBUGGER}, {"ACPI_OS_SERVICES", ACPI_OS_SERVICES}, {"ACPI_CA_DISASSEMBLER", ACPI_CA_DISASSEMBLER}, {"ACPI_ALL_COMPONENTS", ACPI_ALL_COMPONENTS}, {"ACPI_AC_ADAPTER", ACPI_AC_ADAPTER}, {"ACPI_BATTERY", ACPI_BATTERY}, {"ACPI_BUS", ACPI_BUS}, {"ACPI_BUTTON", ACPI_BUTTON}, {"ACPI_EC", ACPI_EC}, {"ACPI_FAN", ACPI_FAN}, {"ACPI_POWERRES", ACPI_POWERRES}, {"ACPI_PROCESSOR", ACPI_PROCESSOR}, {"ACPI_THERMAL", ACPI_THERMAL}, {"ACPI_TIMER", ACPI_TIMER}, {"ACPI_ALL_DRIVERS", ACPI_ALL_DRIVERS}, {NULL, 0} }; static struct debugtag dbg_level[] = { {"ACPI_LV_INIT", ACPI_LV_INIT}, {"ACPI_LV_DEBUG_OBJECT", ACPI_LV_DEBUG_OBJECT}, {"ACPI_LV_INFO", ACPI_LV_INFO}, {"ACPI_LV_REPAIR", ACPI_LV_REPAIR}, {"ACPI_LV_ALL_EXCEPTIONS", ACPI_LV_ALL_EXCEPTIONS}, /* Trace verbosity level 1 [Standard Trace Level] */ {"ACPI_LV_INIT_NAMES", ACPI_LV_INIT_NAMES}, {"ACPI_LV_PARSE", ACPI_LV_PARSE}, {"ACPI_LV_LOAD", ACPI_LV_LOAD}, {"ACPI_LV_DISPATCH", ACPI_LV_DISPATCH}, {"ACPI_LV_EXEC", ACPI_LV_EXEC}, {"ACPI_LV_NAMES", ACPI_LV_NAMES}, {"ACPI_LV_OPREGION", ACPI_LV_OPREGION}, {"ACPI_LV_BFIELD", ACPI_LV_BFIELD}, {"ACPI_LV_TABLES", ACPI_LV_TABLES}, {"ACPI_LV_VALUES", ACPI_LV_VALUES}, {"ACPI_LV_OBJECTS", ACPI_LV_OBJECTS}, {"ACPI_LV_RESOURCES", ACPI_LV_RESOURCES}, {"ACPI_LV_USER_REQUESTS", ACPI_LV_USER_REQUESTS}, {"ACPI_LV_PACKAGE", ACPI_LV_PACKAGE}, {"ACPI_LV_VERBOSITY1", ACPI_LV_VERBOSITY1}, /* Trace verbosity level 2 [Function tracing and memory allocation] */ {"ACPI_LV_ALLOCATIONS", ACPI_LV_ALLOCATIONS}, {"ACPI_LV_FUNCTIONS", ACPI_LV_FUNCTIONS}, {"ACPI_LV_OPTIMIZATIONS", ACPI_LV_OPTIMIZATIONS}, {"ACPI_LV_VERBOSITY2", ACPI_LV_VERBOSITY2}, {"ACPI_LV_ALL", ACPI_LV_ALL}, /* Trace verbosity level 3 [Threading, I/O, and Interrupts] */ {"ACPI_LV_MUTEX", ACPI_LV_MUTEX}, {"ACPI_LV_THREADS", ACPI_LV_THREADS}, {"ACPI_LV_IO", ACPI_LV_IO}, {"ACPI_LV_INTERRUPTS", ACPI_LV_INTERRUPTS}, {"ACPI_LV_VERBOSITY3", ACPI_LV_VERBOSITY3}, /* Exceptionally verbose output -- also used in the global "DebugLevel" */ {"ACPI_LV_AML_DISASSEMBLE", ACPI_LV_AML_DISASSEMBLE}, {"ACPI_LV_VERBOSE_INFO", ACPI_LV_VERBOSE_INFO}, {"ACPI_LV_FULL_TABLES", ACPI_LV_FULL_TABLES}, {"ACPI_LV_EVENTS", ACPI_LV_EVENTS}, {"ACPI_LV_VERBOSE", ACPI_LV_VERBOSE}, {NULL, 0} }; static void acpi_parse_debug(char *cp, struct debugtag *tag, UINT32 *flag) { char *ep; int i, l; int set; while (*cp) { if (isspace(*cp)) { cp++; continue; } ep = cp; while (*ep && !isspace(*ep)) ep++; if (*cp == '!') { set = 0; cp++; if (cp == ep) continue; } else { set = 1; } l = ep - cp; for (i = 0; tag[i].name != NULL; i++) { if (!strncmp(cp, tag[i].name, l)) { if (set) *flag |= tag[i].value; else *flag &= ~tag[i].value; } } cp = ep; } } static void acpi_set_debugging(void *junk) { char *layer, *level; if (cold) { AcpiDbgLayer = 0; AcpiDbgLevel = 0; } layer = kern_getenv("debug.acpi.layer"); level = kern_getenv("debug.acpi.level"); if (layer == NULL && level == NULL) return; printf("ACPI set debug"); if (layer != NULL) { if (strcmp("NONE", layer) != 0) printf(" layer '%s'", layer); acpi_parse_debug(layer, &dbg_layer[0], &AcpiDbgLayer); freeenv(layer); } if (level != NULL) { if (strcmp("NONE", level) != 0) printf(" level '%s'", level); acpi_parse_debug(level, &dbg_level[0], &AcpiDbgLevel); freeenv(level); } printf("\n"); } SYSINIT(acpi_debugging, SI_SUB_TUNABLES, SI_ORDER_ANY, acpi_set_debugging, NULL); static int acpi_debug_sysctl(SYSCTL_HANDLER_ARGS) { int error, *dbg; struct debugtag *tag; struct sbuf sb; char temp[128]; if (sbuf_new(&sb, NULL, 128, SBUF_AUTOEXTEND) == NULL) return (ENOMEM); if (strcmp(oidp->oid_arg1, "debug.acpi.layer") == 0) { tag = &dbg_layer[0]; dbg = &AcpiDbgLayer; } else { tag = &dbg_level[0]; dbg = &AcpiDbgLevel; } /* Get old values if this is a get request. */ ACPI_SERIAL_BEGIN(acpi); if (*dbg == 0) { sbuf_cpy(&sb, "NONE"); } else if (req->newptr == NULL) { for (; tag->name != NULL; tag++) { if ((*dbg & tag->value) == tag->value) sbuf_printf(&sb, "%s ", tag->name); } } sbuf_trim(&sb); sbuf_finish(&sb); strlcpy(temp, sbuf_data(&sb), sizeof(temp)); sbuf_delete(&sb); error = sysctl_handle_string(oidp, temp, sizeof(temp), req); /* Check for error or no change */ if (error == 0 && req->newptr != NULL) { *dbg = 0; kern_setenv((char *)oidp->oid_arg1, temp); acpi_set_debugging(NULL); } ACPI_SERIAL_END(acpi); return (error); } SYSCTL_PROC(_debug_acpi, OID_AUTO, layer, CTLFLAG_RW | CTLTYPE_STRING | CTLFLAG_MPSAFE, "debug.acpi.layer", 0, acpi_debug_sysctl, "A", ""); SYSCTL_PROC(_debug_acpi, OID_AUTO, level, CTLFLAG_RW | CTLTYPE_STRING | CTLFLAG_MPSAFE, "debug.acpi.level", 0, acpi_debug_sysctl, "A", ""); #endif /* ACPI_DEBUG */ static int acpi_debug_objects_sysctl(SYSCTL_HANDLER_ARGS) { int error; int old; old = acpi_debug_objects; error = sysctl_handle_int(oidp, &acpi_debug_objects, 0, req); if (error != 0 || req->newptr == NULL) return (error); if (old == acpi_debug_objects || (old && acpi_debug_objects)) return (0); ACPI_SERIAL_BEGIN(acpi); AcpiGbl_EnableAmlDebugObject = acpi_debug_objects ? TRUE : FALSE; ACPI_SERIAL_END(acpi); return (0); } static int acpi_parse_interfaces(char *str, struct acpi_interface *iface) { char *p; size_t len; int i, j; p = str; while (isspace(*p) || *p == ',') p++; len = strlen(p); if (len == 0) return (0); p = strdup(p, M_TEMP); for (i = 0; i < len; i++) if (p[i] == ',') p[i] = '\0'; i = j = 0; while (i < len) if (isspace(p[i]) || p[i] == '\0') i++; else { i += strlen(p + i) + 1; j++; } if (j == 0) { free(p, M_TEMP); return (0); } iface->data = malloc(sizeof(*iface->data) * j, M_TEMP, M_WAITOK); iface->num = j; i = j = 0; while (i < len) if (isspace(p[i]) || p[i] == '\0') i++; else { iface->data[j] = p + i; i += strlen(p + i) + 1; j++; } return (j); } static void acpi_free_interfaces(struct acpi_interface *iface) { free(iface->data[0], M_TEMP); free(iface->data, M_TEMP); } static void acpi_reset_interfaces(device_t dev) { struct acpi_interface list; ACPI_STATUS status; int i; if (acpi_parse_interfaces(acpi_install_interface, &list) > 0) { for (i = 0; i < list.num; i++) { status = AcpiInstallInterface(list.data[i]); if (ACPI_FAILURE(status)) device_printf(dev, "failed to install _OSI(\"%s\"): %s\n", list.data[i], AcpiFormatException(status)); else if (bootverbose) device_printf(dev, "installed _OSI(\"%s\")\n", list.data[i]); } acpi_free_interfaces(&list); } if (acpi_parse_interfaces(acpi_remove_interface, &list) > 0) { for (i = 0; i < list.num; i++) { status = AcpiRemoveInterface(list.data[i]); if (ACPI_FAILURE(status)) device_printf(dev, "failed to remove _OSI(\"%s\"): %s\n", list.data[i], AcpiFormatException(status)); else if (bootverbose) device_printf(dev, "removed _OSI(\"%s\")\n", list.data[i]); } acpi_free_interfaces(&list); } } static int acpi_pm_func(u_long cmd, void *arg, ...) { int state, acpi_state; int error; struct acpi_softc *sc; va_list ap; error = 0; switch (cmd) { case POWER_CMD_SUSPEND: sc = (struct acpi_softc *)arg; if (sc == NULL) { error = EINVAL; goto out; } va_start(ap, arg); state = va_arg(ap, int); va_end(ap); switch (state) { case POWER_SLEEP_STATE_STANDBY: acpi_state = sc->acpi_standby_sx; break; case POWER_SLEEP_STATE_SUSPEND: acpi_state = sc->acpi_suspend_sx; break; case POWER_SLEEP_STATE_HIBERNATE: acpi_state = ACPI_STATE_S4; break; default: error = EINVAL; goto out; } if (ACPI_FAILURE(acpi_EnterSleepState(sc, acpi_state))) error = ENXIO; break; default: error = EINVAL; goto out; } out: return (error); } static void acpi_pm_register(void *arg) { if (!cold || resource_disabled("acpi", 0)) return; power_pm_register(POWER_PM_TYPE_ACPI, acpi_pm_func, NULL); } SYSINIT(power, SI_SUB_KLD, SI_ORDER_ANY, acpi_pm_register, NULL); diff --git a/sys/i386/i386/pmap.c b/sys/i386/i386/pmap.c index af5b5ac73ce3..fa9ee33db70e 100644 --- a/sys/i386/i386/pmap.c +++ b/sys/i386/i386/pmap.c @@ -1,6395 +1,6395 @@ /*- * SPDX-License-Identifier: BSD-4-Clause * * Copyright (c) 1991 Regents of the University of California. * All rights reserved. * Copyright (c) 1994 John S. Dyson * All rights reserved. * Copyright (c) 1994 David Greenman * All rights reserved. * Copyright (c) 2005-2010 Alan L. Cox * All rights reserved. * * This code is derived from software contributed to Berkeley by * the Systems Programming Group of the University of Utah Computer * Science Department and William Jolitz of UUNET Technologies Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. All advertising materials mentioning features or use of this software * must display the following acknowledgement: * This product includes software developed by the University of * California, Berkeley and its contributors. * 4. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)pmap.c 7.7 (Berkeley) 5/12/91 */ /*- * Copyright (c) 2003 Networks Associates Technology, Inc. * All rights reserved. * Copyright (c) 2018 The FreeBSD Foundation * All rights reserved. * * This software was developed for the FreeBSD Project by Jake Burkholder, * Safeport Network Services, and Network Associates Laboratories, the * Security Research Division of Network Associates, Inc. under * DARPA/SPAWAR contract N66001-01-C-8035 ("CBOSS"), as part of the DARPA * CHATS research program. * * Portions of this software were developed by * Konstantin Belousov under sponsorship from * the FreeBSD Foundation. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); /* * Manages physical address maps. * * Since the information managed by this module is * also stored by the logical address mapping module, * this module may throw away valid virtual-to-physical * mappings at almost any time. However, invalidations * of virtual-to-physical mappings must be done as * requested. * * In order to cope with hardware architectures which * make virtual-to-physical map invalidates expensive, * this module may delay invalidate or reduced protection * operations until such time as they are actually * necessary. This module is given full information as * to which processors are currently using which maps, * and to when physical maps must be made correct. */ #include "opt_apic.h" #include "opt_cpu.h" #include "opt_pmap.h" #include "opt_smp.h" #include "opt_vm.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef DEV_APIC #include #include #include #endif #include #include #include #include #include #include #include #ifdef SMP #include #endif #include #if !defined(DIAGNOSTIC) #ifdef __GNUC_GNU_INLINE__ #define PMAP_INLINE __attribute__((__gnu_inline__)) inline #else #define PMAP_INLINE extern inline #endif #else #define PMAP_INLINE #endif #ifdef PV_STATS #define PV_STAT(x) do { x ; } while (0) #else #define PV_STAT(x) do { } while (0) #endif #define pa_index(pa) ((pa) >> PDRSHIFT) #define pa_to_pvh(pa) (&pv_table[pa_index(pa)]) /* * PTmap is recursive pagemap at top of virtual address space. * Within PTmap, the page directory can be found (third indirection). */ #define PTmap ((pt_entry_t *)(PTDPTDI << PDRSHIFT)) #define PTD ((pd_entry_t *)((PTDPTDI << PDRSHIFT) + (PTDPTDI * PAGE_SIZE))) #define PTDpde ((pd_entry_t *)((PTDPTDI << PDRSHIFT) + (PTDPTDI * PAGE_SIZE) + \ (PTDPTDI * PDESIZE))) /* * Translate a virtual address to the kernel virtual address of its page table * entry (PTE). This can be used recursively. If the address of a PTE as * previously returned by this macro is itself given as the argument, then the * address of the page directory entry (PDE) that maps the PTE will be * returned. * * This macro may be used before pmap_bootstrap() is called. */ #define vtopte(va) (PTmap + i386_btop(va)) /* * Get PDEs and PTEs for user/kernel address space */ #define pmap_pde(m, v) (&((m)->pm_pdir[(vm_offset_t)(v) >> PDRSHIFT])) #define pdir_pde(m, v) (m[(vm_offset_t)(v) >> PDRSHIFT]) #define pmap_pde_v(pte) ((*(int *)pte & PG_V) != 0) #define pmap_pte_w(pte) ((*(int *)pte & PG_W) != 0) #define pmap_pte_m(pte) ((*(int *)pte & PG_M) != 0) #define pmap_pte_u(pte) ((*(int *)pte & PG_A) != 0) #define pmap_pte_v(pte) ((*(int *)pte & PG_V) != 0) #define pmap_pte_set_w(pte, v) ((v) ? atomic_set_int((u_int *)(pte), PG_W) : \ atomic_clear_int((u_int *)(pte), PG_W)) #define pmap_pte_set_prot(pte, v) ((*(int *)pte &= ~PG_PROT), (*(int *)pte |= (v))) static int pgeflag = 0; /* PG_G or-in */ static int pseflag = 0; /* PG_PS or-in */ static int nkpt = NKPT; #ifdef PMAP_PAE_COMP pt_entry_t pg_nx; static uma_zone_t pdptzone; #else #define pg_nx 0 #endif _Static_assert(VM_MAXUSER_ADDRESS == VADDR(TRPTDI, 0), "VM_MAXUSER_ADDRESS"); _Static_assert(VM_MAX_KERNEL_ADDRESS <= VADDR(PTDPTDI, 0), "VM_MAX_KERNEL_ADDRESS"); _Static_assert(PMAP_MAP_LOW == VADDR(LOWPTDI, 0), "PMAP_MAP_LOW"); _Static_assert(KERNLOAD == (KERNPTDI << PDRSHIFT), "KERNLOAD"); extern int pat_works; extern int pg_ps_enabled; extern int elf32_nxstack; #define PAT_INDEX_SIZE 8 static int pat_index[PAT_INDEX_SIZE]; /* cache mode to PAT index conversion */ /* * pmap_mapdev support pre initialization (i.e. console) */ #define PMAP_PREINIT_MAPPING_COUNT 8 static struct pmap_preinit_mapping { vm_paddr_t pa; vm_offset_t va; vm_size_t sz; int mode; } pmap_preinit_mapping[PMAP_PREINIT_MAPPING_COUNT]; static int pmap_initialized; static struct rwlock_padalign pvh_global_lock; /* * Data for the pv entry allocation mechanism */ static TAILQ_HEAD(pch, pv_chunk) pv_chunks = TAILQ_HEAD_INITIALIZER(pv_chunks); extern int pv_entry_max, pv_entry_count; static int pv_entry_high_water = 0; static struct md_page *pv_table; extern int shpgperproc; static struct pv_chunk *pv_chunkbase; /* KVA block for pv_chunks */ static int pv_maxchunks; /* How many chunks we have KVA for */ static vm_offset_t pv_vafree; /* freelist stored in the PTE */ /* * All those kernel PT submaps that BSD is so fond of */ static pt_entry_t *CMAP3; static pd_entry_t *KPTD; static caddr_t CADDR3; /* * Crashdump maps. */ static caddr_t crashdumpmap; static pt_entry_t *PMAP1 = NULL, *PMAP2, *PMAP3; static pt_entry_t *PADDR1 = NULL, *PADDR2, *PADDR3; #ifdef SMP static int PMAP1cpu, PMAP3cpu; extern int PMAP1changedcpu; #endif extern int PMAP1changed; extern int PMAP1unchanged; static struct mtx PMAP2mutex; /* * Internal flags for pmap_enter()'s helper functions. */ #define PMAP_ENTER_NORECLAIM 0x1000000 /* Don't reclaim PV entries. */ #define PMAP_ENTER_NOREPLACE 0x2000000 /* Don't replace mappings. */ static void free_pv_chunk(struct pv_chunk *pc); static void free_pv_entry(pmap_t pmap, pv_entry_t pv); static pv_entry_t get_pv_entry(pmap_t pmap, boolean_t try); static void pmap_pv_demote_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa); static bool pmap_pv_insert_pde(pmap_t pmap, vm_offset_t va, pd_entry_t pde, u_int flags); #if VM_NRESERVLEVEL > 0 static void pmap_pv_promote_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa); #endif static void pmap_pvh_free(struct md_page *pvh, pmap_t pmap, vm_offset_t va); static pv_entry_t pmap_pvh_remove(struct md_page *pvh, pmap_t pmap, vm_offset_t va); static int pmap_pvh_wired_mappings(struct md_page *pvh, int count); static void pmap_abort_ptp(pmap_t pmap, vm_offset_t va, vm_page_t mpte); static boolean_t pmap_demote_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va); static bool pmap_enter_4mpage(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot); static int pmap_enter_pde(pmap_t pmap, vm_offset_t va, pd_entry_t newpde, u_int flags, vm_page_t m); static vm_page_t pmap_enter_quick_locked(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, vm_page_t mpte); static int pmap_insert_pt_page(pmap_t pmap, vm_page_t mpte, bool promoted); static void pmap_invalidate_pde_page(pmap_t pmap, vm_offset_t va, pd_entry_t pde); static void pmap_fill_ptp(pt_entry_t *firstpte, pt_entry_t newpte); static boolean_t pmap_is_modified_pvh(struct md_page *pvh); static boolean_t pmap_is_referenced_pvh(struct md_page *pvh); static void pmap_kenter_attr(vm_offset_t va, vm_paddr_t pa, int mode); static void pmap_kenter_pde(vm_offset_t va, pd_entry_t newpde); static void pmap_pde_attr(pd_entry_t *pde, int cache_bits); #if VM_NRESERVLEVEL > 0 static void pmap_promote_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va); #endif static boolean_t pmap_protect_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t sva, vm_prot_t prot); static void pmap_pte_attr(pt_entry_t *pte, int cache_bits); static void pmap_remove_pde(pmap_t pmap, pd_entry_t *pdq, vm_offset_t sva, struct spglist *free); static int pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t sva, struct spglist *free); static vm_page_t pmap_remove_pt_page(pmap_t pmap, vm_offset_t va); static void pmap_remove_page(pmap_t pmap, vm_offset_t va, struct spglist *free); static bool pmap_remove_ptes(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, struct spglist *free); static void pmap_remove_entry(pmap_t pmap, vm_page_t m, vm_offset_t va); static void pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t m); static boolean_t pmap_try_insert_pv_entry(pmap_t pmap, vm_offset_t va, vm_page_t m); static void pmap_update_pde(pmap_t pmap, vm_offset_t va, pd_entry_t *pde, pd_entry_t newpde); static void pmap_update_pde_invalidate(vm_offset_t va, pd_entry_t newpde); static vm_page_t pmap_allocpte(pmap_t pmap, vm_offset_t va, u_int flags); static vm_page_t _pmap_allocpte(pmap_t pmap, u_int ptepindex, u_int flags); static void _pmap_unwire_ptp(pmap_t pmap, vm_page_t m, struct spglist *free); static pt_entry_t *pmap_pte_quick(pmap_t pmap, vm_offset_t va); static void pmap_pte_release(pt_entry_t *pte); static int pmap_unuse_pt(pmap_t, vm_offset_t, struct spglist *); #ifdef PMAP_PAE_COMP static void *pmap_pdpt_allocf(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags, int wait); #endif static void pmap_init_trm(void); static void pmap_invalidate_all_int(pmap_t pmap); static __inline void pagezero(void *page); CTASSERT(1 << PDESHIFT == sizeof(pd_entry_t)); CTASSERT(1 << PTESHIFT == sizeof(pt_entry_t)); extern char _end[]; extern u_long physfree; /* phys addr of next free page */ extern u_long vm86phystk;/* PA of vm86/bios stack */ extern u_long vm86paddr;/* address of vm86 region */ extern int vm86pa; /* phys addr of vm86 region */ extern u_long KERNend; /* phys addr end of kernel (just after bss) */ #ifdef PMAP_PAE_COMP pd_entry_t *IdlePTD_pae; /* phys addr of kernel PTD */ pdpt_entry_t *IdlePDPT; /* phys addr of kernel PDPT */ pt_entry_t *KPTmap_pae; /* address of kernel page tables */ #define IdlePTD IdlePTD_pae #define KPTmap KPTmap_pae #else pd_entry_t *IdlePTD_nopae; pt_entry_t *KPTmap_nopae; #define IdlePTD IdlePTD_nopae #define KPTmap KPTmap_nopae #endif extern u_long KPTphys; /* phys addr of kernel page tables */ extern u_long tramp_idleptd; static u_long allocpages(u_int cnt, u_long *physfree) { u_long res; res = *physfree; *physfree += PAGE_SIZE * cnt; bzero((void *)res, PAGE_SIZE * cnt); return (res); } static void pmap_cold_map(u_long pa, u_long va, u_long cnt) { pt_entry_t *pt; for (pt = (pt_entry_t *)KPTphys + atop(va); cnt > 0; cnt--, pt++, va += PAGE_SIZE, pa += PAGE_SIZE) *pt = pa | PG_V | PG_RW | PG_A | PG_M; } static void pmap_cold_mapident(u_long pa, u_long cnt) { pmap_cold_map(pa, pa, cnt); } _Static_assert(LOWPTDI * 2 * NBPDR == KERNBASE, "Broken double-map of zero PTD"); static void __CONCAT(PMTYPE, remap_lower)(bool enable) { int i; for (i = 0; i < LOWPTDI; i++) IdlePTD[i] = enable ? IdlePTD[LOWPTDI + i] : 0; load_cr3(rcr3()); /* invalidate TLB */ } /* * Called from locore.s before paging is enabled. Sets up the first * kernel page table. Since kernel is mapped with PA == VA, this code * does not require relocations. */ void __CONCAT(PMTYPE, cold)(void) { pt_entry_t *pt; u_long a; u_int cr3, ncr4; physfree = (u_long)&_end; if (bootinfo.bi_esymtab != 0) physfree = bootinfo.bi_esymtab; if (bootinfo.bi_kernend != 0) physfree = bootinfo.bi_kernend; physfree = roundup2(physfree, NBPDR); KERNend = physfree; /* Allocate Kernel Page Tables */ KPTphys = allocpages(NKPT, &physfree); KPTmap = (pt_entry_t *)KPTphys; /* Allocate Page Table Directory */ #ifdef PMAP_PAE_COMP /* XXX only need 32 bytes (easier for now) */ IdlePDPT = (pdpt_entry_t *)allocpages(1, &physfree); #endif IdlePTD = (pd_entry_t *)allocpages(NPGPTD, &physfree); /* * Allocate KSTACK. Leave a guard page between IdlePTD and * proc0kstack, to control stack overflow for thread0 and * prevent corruption of the page table. We leak the guard * physical memory due to 1:1 mappings. */ allocpages(1, &physfree); proc0kstack = allocpages(TD0_KSTACK_PAGES, &physfree); /* vm86/bios stack */ vm86phystk = allocpages(1, &physfree); /* pgtable + ext + IOPAGES */ vm86paddr = vm86pa = allocpages(3, &physfree); /* Install page tables into PTD. Page table page 1 is wasted. */ for (a = 0; a < NKPT; a++) IdlePTD[a] = (KPTphys + ptoa(a)) | PG_V | PG_RW | PG_A | PG_M; #ifdef PMAP_PAE_COMP /* PAE install PTD pointers into PDPT */ for (a = 0; a < NPGPTD; a++) IdlePDPT[a] = ((u_int)IdlePTD + ptoa(a)) | PG_V; #endif /* * Install recursive mapping for kernel page tables into * itself. */ for (a = 0; a < NPGPTD; a++) IdlePTD[PTDPTDI + a] = ((u_int)IdlePTD + ptoa(a)) | PG_V | PG_RW; /* * Initialize page table pages mapping physical address zero * through the (physical) end of the kernel. Many of these * pages must be reserved, and we reserve them all and map * them linearly for convenience. We do this even if we've * enabled PSE above; we'll just switch the corresponding * kernel PDEs before we turn on paging. * * This and all other page table entries allow read and write * access for various reasons. Kernel mappings never have any * access restrictions. */ pmap_cold_mapident(0, atop(NBPDR) * LOWPTDI); pmap_cold_map(0, NBPDR * LOWPTDI, atop(NBPDR) * LOWPTDI); pmap_cold_mapident(KERNBASE, atop(KERNend - KERNBASE)); /* Map page table directory */ #ifdef PMAP_PAE_COMP pmap_cold_mapident((u_long)IdlePDPT, 1); #endif pmap_cold_mapident((u_long)IdlePTD, NPGPTD); /* Map early KPTmap. It is really pmap_cold_mapident. */ pmap_cold_map(KPTphys, (u_long)KPTmap, NKPT); /* Map proc0kstack */ pmap_cold_mapident(proc0kstack, TD0_KSTACK_PAGES); /* ISA hole already mapped */ pmap_cold_mapident(vm86phystk, 1); pmap_cold_mapident(vm86pa, 3); /* Map page 0 into the vm86 page table */ *(pt_entry_t *)vm86pa = 0 | PG_RW | PG_U | PG_A | PG_M | PG_V; /* ...likewise for the ISA hole for vm86 */ for (pt = (pt_entry_t *)vm86pa + atop(ISA_HOLE_START), a = 0; a < atop(ISA_HOLE_LENGTH); a++, pt++) *pt = (ISA_HOLE_START + ptoa(a)) | PG_RW | PG_U | PG_A | PG_M | PG_V; /* Enable PSE, PGE, VME, and PAE if configured. */ ncr4 = 0; if ((cpu_feature & CPUID_PSE) != 0) { ncr4 |= CR4_PSE; pseflag = PG_PS; /* * Superpage mapping of the kernel text. Existing 4k * page table pages are wasted. */ for (a = KERNBASE; a < KERNend; a += NBPDR) IdlePTD[a >> PDRSHIFT] = a | PG_PS | PG_A | PG_M | PG_RW | PG_V; } if ((cpu_feature & CPUID_PGE) != 0) { ncr4 |= CR4_PGE; pgeflag = PG_G; } ncr4 |= (cpu_feature & CPUID_VME) != 0 ? CR4_VME : 0; #ifdef PMAP_PAE_COMP ncr4 |= CR4_PAE; #endif if (ncr4 != 0) load_cr4(rcr4() | ncr4); /* Now enable paging */ #ifdef PMAP_PAE_COMP cr3 = (u_int)IdlePDPT; if ((cpu_feature & CPUID_PAT) == 0) wbinvd(); #else cr3 = (u_int)IdlePTD; #endif tramp_idleptd = cr3; load_cr3(cr3); load_cr0(rcr0() | CR0_PG); /* * Now running relocated at KERNBASE where the system is * linked to run. */ /* * Remove the lowest part of the double mapping of low memory * to get some null pointer checks. */ __CONCAT(PMTYPE, remap_lower)(false); kernel_vm_end = /* 0 + */ NKPT * NBPDR; #ifdef PMAP_PAE_COMP i386_pmap_VM_NFREEORDER = VM_NFREEORDER_PAE; i386_pmap_VM_LEVEL_0_ORDER = VM_LEVEL_0_ORDER_PAE; i386_pmap_PDRSHIFT = PDRSHIFT_PAE; #else i386_pmap_VM_NFREEORDER = VM_NFREEORDER_NOPAE; i386_pmap_VM_LEVEL_0_ORDER = VM_LEVEL_0_ORDER_NOPAE; i386_pmap_PDRSHIFT = PDRSHIFT_NOPAE; #endif } static void __CONCAT(PMTYPE, set_nx)(void) { #ifdef PMAP_PAE_COMP if ((amd_feature & AMDID_NX) == 0) return; pg_nx = PG_NX; elf32_nxstack = 1; /* EFER.EFER_NXE is set in initializecpu(). */ #endif } /* * Bootstrap the system enough to run with virtual memory. * * On the i386 this is called after pmap_cold() created initial * kernel page table and enabled paging, and just syncs the pmap * module with what has already been done. */ static void __CONCAT(PMTYPE, bootstrap)(vm_paddr_t firstaddr) { vm_offset_t va; pt_entry_t *pte, *unused __unused; struct pcpu *pc; u_long res; int i; res = atop(firstaddr - (vm_paddr_t)KERNLOAD); /* * Add a physical memory segment (vm_phys_seg) corresponding to the * preallocated kernel page table pages so that vm_page structures * representing these pages will be created. The vm_page structures * are required for promotion of the corresponding kernel virtual * addresses to superpage mappings. */ vm_phys_early_add_seg(KPTphys, KPTphys + ptoa(nkpt)); /* * Initialize the first available kernel virtual address. * However, using "firstaddr" may waste a few pages of the * kernel virtual address space, because pmap_cold() may not * have mapped every physical page that it allocated. * Preferably, pmap_cold() would provide a first unused * virtual address in addition to "firstaddr". */ virtual_avail = (vm_offset_t)firstaddr; virtual_end = VM_MAX_KERNEL_ADDRESS; /* * Initialize the kernel pmap (which is statically allocated). * Count bootstrap data as being resident in case any of this data is * later unmapped (using pmap_remove()) and freed. */ PMAP_LOCK_INIT(kernel_pmap); kernel_pmap->pm_pdir = IdlePTD; #ifdef PMAP_PAE_COMP kernel_pmap->pm_pdpt = IdlePDPT; #endif CPU_FILL(&kernel_pmap->pm_active); /* don't allow deactivation */ kernel_pmap->pm_stats.resident_count = res; TAILQ_INIT(&kernel_pmap->pm_pvchunk); /* * Initialize the global pv list lock. */ rw_init(&pvh_global_lock, "pmap pv global"); /* * Reserve some special page table entries/VA space for temporary * mapping of pages. */ #define SYSMAP(c, p, v, n) \ v = (c)va; va += ((n)*PAGE_SIZE); p = pte; pte += (n); va = virtual_avail; pte = vtopte(va); /* * Initialize temporary map objects on the current CPU for use * during early boot. * CMAP1/CMAP2 are used for zeroing and copying pages. * CMAP3 is used for the boot-time memory test. */ pc = get_pcpu(); mtx_init(&pc->pc_cmap_lock, "SYSMAPS", NULL, MTX_DEF); SYSMAP(caddr_t, pc->pc_cmap_pte1, pc->pc_cmap_addr1, 1) SYSMAP(caddr_t, pc->pc_cmap_pte2, pc->pc_cmap_addr2, 1) SYSMAP(vm_offset_t, pte, pc->pc_qmap_addr, 1) SYSMAP(caddr_t, CMAP3, CADDR3, 1); /* * Crashdump maps. */ SYSMAP(caddr_t, unused, crashdumpmap, MAXDUMPPGS) /* * ptvmmap is used for reading arbitrary physical pages via /dev/mem. */ SYSMAP(caddr_t, unused, ptvmmap, 1) /* * msgbufp is used to map the system message buffer. */ SYSMAP(struct msgbuf *, unused, msgbufp, atop(round_page(msgbufsize))) /* * KPTmap is used by pmap_kextract(). * * KPTmap is first initialized by pmap_cold(). However, that initial * KPTmap can only support NKPT page table pages. Here, a larger * KPTmap is created that can support KVA_PAGES page table pages. */ SYSMAP(pt_entry_t *, KPTD, KPTmap, KVA_PAGES) for (i = 0; i < NKPT; i++) KPTD[i] = (KPTphys + ptoa(i)) | PG_RW | PG_V; /* * PADDR1 and PADDR2 are used by pmap_pte_quick() and pmap_pte(), * respectively. */ SYSMAP(pt_entry_t *, PMAP1, PADDR1, 1) SYSMAP(pt_entry_t *, PMAP2, PADDR2, 1) SYSMAP(pt_entry_t *, PMAP3, PADDR3, 1) mtx_init(&PMAP2mutex, "PMAP2", NULL, MTX_DEF); virtual_avail = va; /* * Initialize the PAT MSR if present. * pmap_init_pat() clears and sets CR4_PGE, which, as a * side-effect, invalidates stale PG_G TLB entries that might * have been created in our pre-boot environment. We assume * that PAT support implies PGE and in reverse, PGE presence * comes with PAT. Both features were added for Pentium Pro. */ pmap_init_pat(); } static void pmap_init_reserved_pages(void) { struct pcpu *pc; vm_offset_t pages; int i; #ifdef PMAP_PAE_COMP if (!pae_mode) return; #else if (pae_mode) return; #endif CPU_FOREACH(i) { pc = pcpu_find(i); mtx_init(&pc->pc_copyout_mlock, "cpmlk", NULL, MTX_DEF | MTX_NEW); pc->pc_copyout_maddr = kva_alloc(ptoa(2)); if (pc->pc_copyout_maddr == 0) panic("unable to allocate non-sleepable copyout KVA"); sx_init(&pc->pc_copyout_slock, "cpslk"); pc->pc_copyout_saddr = kva_alloc(ptoa(2)); if (pc->pc_copyout_saddr == 0) panic("unable to allocate sleepable copyout KVA"); pc->pc_pmap_eh_va = kva_alloc(ptoa(1)); if (pc->pc_pmap_eh_va == 0) panic("unable to allocate pmap_extract_and_hold KVA"); pc->pc_pmap_eh_ptep = (char *)vtopte(pc->pc_pmap_eh_va); /* * Skip if the mappings have already been initialized, * i.e. this is the BSP. */ if (pc->pc_cmap_addr1 != 0) continue; mtx_init(&pc->pc_cmap_lock, "SYSMAPS", NULL, MTX_DEF); pages = kva_alloc(PAGE_SIZE * 3); if (pages == 0) panic("unable to allocate CMAP KVA"); pc->pc_cmap_pte1 = vtopte(pages); pc->pc_cmap_pte2 = vtopte(pages + PAGE_SIZE); pc->pc_cmap_addr1 = (caddr_t)pages; pc->pc_cmap_addr2 = (caddr_t)(pages + PAGE_SIZE); pc->pc_qmap_addr = pages + ptoa(2); } } SYSINIT(rpages_init, SI_SUB_CPU, SI_ORDER_ANY, pmap_init_reserved_pages, NULL); /* * Setup the PAT MSR. */ static void __CONCAT(PMTYPE, init_pat)(void) { int pat_table[PAT_INDEX_SIZE]; uint64_t pat_msr; u_long cr0, cr4; int i; /* Set default PAT index table. */ for (i = 0; i < PAT_INDEX_SIZE; i++) pat_table[i] = -1; pat_table[PAT_WRITE_BACK] = 0; pat_table[PAT_WRITE_THROUGH] = 1; pat_table[PAT_UNCACHEABLE] = 3; pat_table[PAT_WRITE_COMBINING] = 3; pat_table[PAT_WRITE_PROTECTED] = 3; pat_table[PAT_UNCACHED] = 3; /* * Bail if this CPU doesn't implement PAT. * We assume that PAT support implies PGE. */ if ((cpu_feature & CPUID_PAT) == 0) { for (i = 0; i < PAT_INDEX_SIZE; i++) pat_index[i] = pat_table[i]; pat_works = 0; return; } /* * Due to some Intel errata, we can only safely use the lower 4 * PAT entries. * * Intel Pentium III Processor Specification Update * Errata E.27 (Upper Four PAT Entries Not Usable With Mode B * or Mode C Paging) * * Intel Pentium IV Processor Specification Update * Errata N46 (PAT Index MSB May Be Calculated Incorrectly) */ if (cpu_vendor_id == CPU_VENDOR_INTEL && !(CPUID_TO_FAMILY(cpu_id) == 6 && CPUID_TO_MODEL(cpu_id) >= 0xe)) pat_works = 0; /* Initialize default PAT entries. */ pat_msr = PAT_VALUE(0, PAT_WRITE_BACK) | PAT_VALUE(1, PAT_WRITE_THROUGH) | PAT_VALUE(2, PAT_UNCACHED) | PAT_VALUE(3, PAT_UNCACHEABLE) | PAT_VALUE(4, PAT_WRITE_BACK) | PAT_VALUE(5, PAT_WRITE_THROUGH) | PAT_VALUE(6, PAT_UNCACHED) | PAT_VALUE(7, PAT_UNCACHEABLE); if (pat_works) { /* * Leave the indices 0-3 at the default of WB, WT, UC-, and UC. * Program 5 and 6 as WP and WC. * Leave 4 and 7 as WB and UC. */ pat_msr &= ~(PAT_MASK(5) | PAT_MASK(6)); pat_msr |= PAT_VALUE(5, PAT_WRITE_PROTECTED) | PAT_VALUE(6, PAT_WRITE_COMBINING); pat_table[PAT_UNCACHED] = 2; pat_table[PAT_WRITE_PROTECTED] = 5; pat_table[PAT_WRITE_COMBINING] = 6; } else { /* * Just replace PAT Index 2 with WC instead of UC-. */ pat_msr &= ~PAT_MASK(2); pat_msr |= PAT_VALUE(2, PAT_WRITE_COMBINING); pat_table[PAT_WRITE_COMBINING] = 2; } /* Disable PGE. */ cr4 = rcr4(); load_cr4(cr4 & ~CR4_PGE); /* Disable caches (CD = 1, NW = 0). */ cr0 = rcr0(); load_cr0((cr0 & ~CR0_NW) | CR0_CD); /* Flushes caches and TLBs. */ wbinvd(); invltlb(); /* Update PAT and index table. */ wrmsr(MSR_PAT, pat_msr); for (i = 0; i < PAT_INDEX_SIZE; i++) pat_index[i] = pat_table[i]; /* Flush caches and TLBs again. */ wbinvd(); invltlb(); /* Restore caches and PGE. */ load_cr0(cr0); load_cr4(cr4); } #ifdef PMAP_PAE_COMP static void * pmap_pdpt_allocf(uma_zone_t zone, vm_size_t bytes, int domain, uint8_t *flags, int wait) { /* Inform UMA that this allocator uses kernel_map/object. */ *flags = UMA_SLAB_KERNEL; return ((void *)kmem_alloc_contig_domainset(DOMAINSET_FIXED(domain), bytes, wait, 0x0ULL, 0xffffffffULL, 1, 0, VM_MEMATTR_DEFAULT)); } #endif /* * Abuse the pte nodes for unmapped kva to thread a kva freelist through. * Requirements: * - Must deal with pages in order to ensure that none of the PG_* bits * are ever set, PG_V in particular. * - Assumes we can write to ptes without pte_store() atomic ops, even * on PAE systems. This should be ok. * - Assumes nothing will ever test these addresses for 0 to indicate * no mapping instead of correctly checking PG_V. * - Assumes a vm_offset_t will fit in a pte (true for i386). * Because PG_V is never set, there can be no mappings to invalidate. */ static vm_offset_t pmap_ptelist_alloc(vm_offset_t *head) { pt_entry_t *pte; vm_offset_t va; va = *head; if (va == 0) panic("pmap_ptelist_alloc: exhausted ptelist KVA"); pte = vtopte(va); *head = *pte; if (*head & PG_V) panic("pmap_ptelist_alloc: va with PG_V set!"); *pte = 0; return (va); } static void pmap_ptelist_free(vm_offset_t *head, vm_offset_t va) { pt_entry_t *pte; if (va & PG_V) panic("pmap_ptelist_free: freeing va with PG_V set!"); pte = vtopte(va); *pte = *head; /* virtual! PG_V is 0 though */ *head = va; } static void pmap_ptelist_init(vm_offset_t *head, void *base, int npages) { int i; vm_offset_t va; *head = 0; for (i = npages - 1; i >= 0; i--) { va = (vm_offset_t)base + i * PAGE_SIZE; pmap_ptelist_free(head, va); } } /* * Initialize the pmap module. * Called by vm_init, to initialize any structures that the pmap * system needs to map virtual memory. */ static void __CONCAT(PMTYPE, init)(void) { struct pmap_preinit_mapping *ppim; vm_page_t mpte; vm_size_t s; int i, pv_npg; /* * Initialize the vm page array entries for the kernel pmap's * page table pages. */ PMAP_LOCK(kernel_pmap); for (i = 0; i < NKPT; i++) { mpte = PHYS_TO_VM_PAGE(KPTphys + ptoa(i)); KASSERT(mpte >= vm_page_array && mpte < &vm_page_array[vm_page_array_size], ("pmap_init: page table page is out of range")); mpte->pindex = i + KPTDI; mpte->phys_addr = KPTphys + ptoa(i); mpte->ref_count = 1; /* * Collect the page table pages that were replaced by a 2/4MB * page. They are filled with equivalent 4KB page mappings. */ if (pseflag != 0 && KERNBASE <= i << PDRSHIFT && i << PDRSHIFT < KERNend && pmap_insert_pt_page(kernel_pmap, mpte, true)) panic("pmap_init: pmap_insert_pt_page failed"); } PMAP_UNLOCK(kernel_pmap); vm_wire_add(NKPT); /* * Initialize the address space (zone) for the pv entries. Set a * high water mark so that the system can recover from excessive * numbers of pv entries. */ TUNABLE_INT_FETCH("vm.pmap.shpgperproc", &shpgperproc); pv_entry_max = shpgperproc * maxproc + vm_cnt.v_page_count; TUNABLE_INT_FETCH("vm.pmap.pv_entries", &pv_entry_max); pv_entry_max = roundup(pv_entry_max, _NPCPV); pv_entry_high_water = 9 * (pv_entry_max / 10); /* * If the kernel is running on a virtual machine, then it must assume * that MCA is enabled by the hypervisor. Moreover, the kernel must * be prepared for the hypervisor changing the vendor and family that * are reported by CPUID. Consequently, the workaround for AMD Family * 10h Erratum 383 is enabled if the processor's feature set does not * include at least one feature that is only supported by older Intel * or newer AMD processors. */ if (vm_guest != VM_GUEST_NO && (cpu_feature & CPUID_SS) == 0 && (cpu_feature2 & (CPUID2_SSSE3 | CPUID2_SSE41 | CPUID2_AESNI | CPUID2_AVX | CPUID2_XSAVE)) == 0 && (amd_feature2 & (AMDID2_XOP | AMDID2_FMA4)) == 0) workaround_erratum383 = 1; /* * Are large page mappings supported and enabled? */ TUNABLE_INT_FETCH("vm.pmap.pg_ps_enabled", &pg_ps_enabled); if (pseflag == 0) pg_ps_enabled = 0; else if (pg_ps_enabled) { KASSERT(MAXPAGESIZES > 1 && pagesizes[1] == 0, ("pmap_init: can't assign to pagesizes[1]")); pagesizes[1] = NBPDR; } /* * Calculate the size of the pv head table for superpages. * Handle the possibility that "vm_phys_segs[...].end" is zero. */ pv_npg = trunc_4mpage(vm_phys_segs[vm_phys_nsegs - 1].end - PAGE_SIZE) / NBPDR + 1; /* * Allocate memory for the pv head table for superpages. */ s = (vm_size_t)(pv_npg * sizeof(struct md_page)); s = round_page(s); pv_table = (struct md_page *)kmem_malloc(s, M_WAITOK | M_ZERO); for (i = 0; i < pv_npg; i++) TAILQ_INIT(&pv_table[i].pv_list); pv_maxchunks = MAX(pv_entry_max / _NPCPV, maxproc); pv_chunkbase = (struct pv_chunk *)kva_alloc(PAGE_SIZE * pv_maxchunks); if (pv_chunkbase == NULL) panic("pmap_init: not enough kvm for pv chunks"); pmap_ptelist_init(&pv_vafree, pv_chunkbase, pv_maxchunks); #ifdef PMAP_PAE_COMP pdptzone = uma_zcreate("PDPT", NPGPTD * sizeof(pdpt_entry_t), NULL, NULL, NULL, NULL, (NPGPTD * sizeof(pdpt_entry_t)) - 1, UMA_ZONE_CONTIG | UMA_ZONE_VM | UMA_ZONE_NOFREE); uma_zone_set_allocf(pdptzone, pmap_pdpt_allocf); #endif pmap_initialized = 1; pmap_init_trm(); if (!bootverbose) return; for (i = 0; i < PMAP_PREINIT_MAPPING_COUNT; i++) { ppim = pmap_preinit_mapping + i; if (ppim->va == 0) continue; printf("PPIM %u: PA=%#jx, VA=%#x, size=%#x, mode=%#x\n", i, (uintmax_t)ppim->pa, ppim->va, ppim->sz, ppim->mode); } } extern u_long pmap_pde_demotions; extern u_long pmap_pde_mappings; extern u_long pmap_pde_p_failures; extern u_long pmap_pde_promotions; /*************************************************** * Low level helper routines..... ***************************************************/ static boolean_t __CONCAT(PMTYPE, is_valid_memattr)(pmap_t pmap __unused, vm_memattr_t mode) { return (mode >= 0 && mode < PAT_INDEX_SIZE && pat_index[(int)mode] >= 0); } /* * Determine the appropriate bits to set in a PTE or PDE for a specified * caching mode. */ static int __CONCAT(PMTYPE, cache_bits)(pmap_t pmap, int mode, boolean_t is_pde) { int cache_bits, pat_flag, pat_idx; if (!pmap_is_valid_memattr(pmap, mode)) panic("Unknown caching mode %d\n", mode); /* The PAT bit is different for PTE's and PDE's. */ pat_flag = is_pde ? PG_PDE_PAT : PG_PTE_PAT; /* Map the caching mode to a PAT index. */ pat_idx = pat_index[mode]; /* Map the 3-bit index value into the PAT, PCD, and PWT bits. */ cache_bits = 0; if (pat_idx & 0x4) cache_bits |= pat_flag; if (pat_idx & 0x2) cache_bits |= PG_NC_PCD; if (pat_idx & 0x1) cache_bits |= PG_NC_PWT; return (cache_bits); } static int pmap_pat_index(pmap_t pmap, pt_entry_t pte, bool is_pde) { int pat_flag, pat_idx; if ((cpu_feature & CPUID_PAT) == 0) return (0); pat_idx = 0; /* The PAT bit is different for PTE's and PDE's. */ pat_flag = is_pde ? PG_PDE_PAT : PG_PTE_PAT; if ((pte & pat_flag) != 0) pat_idx |= 0x4; if ((pte & PG_NC_PCD) != 0) pat_idx |= 0x2; if ((pte & PG_NC_PWT) != 0) pat_idx |= 0x1; /* See pmap_init_pat(). */ if (pat_works) { if (pat_idx == 4) pat_idx = 0; if (pat_idx == 7) pat_idx = 3; } else { /* XXXKIB */ } return (pat_idx); } static bool __CONCAT(PMTYPE, ps_enabled)(pmap_t pmap __unused) { return (pg_ps_enabled); } /* * The caller is responsible for maintaining TLB consistency. */ static void pmap_kenter_pde(vm_offset_t va, pd_entry_t newpde) { pd_entry_t *pde; pde = pmap_pde(kernel_pmap, va); pde_store(pde, newpde); } /* * After changing the page size for the specified virtual address in the page * table, flush the corresponding entries from the processor's TLB. Only the * calling processor's TLB is affected. * * The calling thread must be pinned to a processor. */ static void pmap_update_pde_invalidate(vm_offset_t va, pd_entry_t newpde) { if ((newpde & PG_PS) == 0) /* Demotion: flush a specific 2MB page mapping. */ invlpg(va); else /* if ((newpde & PG_G) == 0) */ /* * Promotion: flush every 4KB page mapping from the TLB * because there are too many to flush individually. */ invltlb(); } #ifdef SMP static void pmap_curcpu_cb_dummy(pmap_t pmap __unused, vm_offset_t addr1 __unused, vm_offset_t addr2 __unused) { } /* * For SMP, these functions have to use the IPI mechanism for coherence. * * N.B.: Before calling any of the following TLB invalidation functions, * the calling processor must ensure that all stores updating a non- * kernel page table are globally performed. Otherwise, another * processor could cache an old, pre-update entry without being * invalidated. This can happen one of two ways: (1) The pmap becomes * active on another processor after its pm_active field is checked by * one of the following functions but before a store updating the page * table is globally performed. (2) The pmap becomes active on another * processor before its pm_active field is checked but due to * speculative loads one of the following functions stills reads the * pmap as inactive on the other processor. * * The kernel page table is exempt because its pm_active field is * immutable. The kernel page table is always active on every * processor. */ static void pmap_invalidate_page_int(pmap_t pmap, vm_offset_t va) { cpuset_t *mask, other_cpus; u_int cpuid; sched_pin(); if (pmap == kernel_pmap) { invlpg(va); mask = &all_cpus; } else if (!CPU_CMP(&pmap->pm_active, &all_cpus)) { mask = &all_cpus; } else { cpuid = PCPU_GET(cpuid); other_cpus = all_cpus; CPU_CLR(cpuid, &other_cpus); - CPU_AND(&other_cpus, &pmap->pm_active); + CPU_AND(&other_cpus, &other_cpus, &pmap->pm_active); mask = &other_cpus; } smp_masked_invlpg(*mask, va, pmap, pmap_curcpu_cb_dummy); sched_unpin(); } /* 4k PTEs -- Chosen to exceed the total size of Broadwell L2 TLB */ #define PMAP_INVLPG_THRESHOLD (4 * 1024 * PAGE_SIZE) static void pmap_invalidate_range_int(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { cpuset_t *mask, other_cpus; vm_offset_t addr; u_int cpuid; if (eva - sva >= PMAP_INVLPG_THRESHOLD) { pmap_invalidate_all_int(pmap); return; } sched_pin(); if (pmap == kernel_pmap) { for (addr = sva; addr < eva; addr += PAGE_SIZE) invlpg(addr); mask = &all_cpus; } else if (!CPU_CMP(&pmap->pm_active, &all_cpus)) { mask = &all_cpus; } else { cpuid = PCPU_GET(cpuid); other_cpus = all_cpus; CPU_CLR(cpuid, &other_cpus); - CPU_AND(&other_cpus, &pmap->pm_active); + CPU_AND(&other_cpus, &other_cpus, &pmap->pm_active); mask = &other_cpus; } smp_masked_invlpg_range(*mask, sva, eva, pmap, pmap_curcpu_cb_dummy); sched_unpin(); } static void pmap_invalidate_all_int(pmap_t pmap) { cpuset_t *mask, other_cpus; u_int cpuid; sched_pin(); if (pmap == kernel_pmap) { invltlb(); mask = &all_cpus; } else if (!CPU_CMP(&pmap->pm_active, &all_cpus)) { mask = &all_cpus; } else { cpuid = PCPU_GET(cpuid); other_cpus = all_cpus; CPU_CLR(cpuid, &other_cpus); - CPU_AND(&other_cpus, &pmap->pm_active); + CPU_AND(&other_cpus, &other_cpus, &pmap->pm_active); mask = &other_cpus; } smp_masked_invltlb(*mask, pmap, pmap_curcpu_cb_dummy); sched_unpin(); } static void pmap_invalidate_cache_curcpu_cb(pmap_t pmap __unused, vm_offset_t addr1 __unused, vm_offset_t addr2 __unused) { wbinvd(); } static void __CONCAT(PMTYPE, invalidate_cache)(void) { smp_cache_flush(pmap_invalidate_cache_curcpu_cb); } struct pde_action { cpuset_t invalidate; /* processors that invalidate their TLB */ vm_offset_t va; pd_entry_t *pde; pd_entry_t newpde; u_int store; /* processor that updates the PDE */ }; static void pmap_update_pde_kernel(void *arg) { struct pde_action *act = arg; pd_entry_t *pde; if (act->store == PCPU_GET(cpuid)) { pde = pmap_pde(kernel_pmap, act->va); pde_store(pde, act->newpde); } } static void pmap_update_pde_user(void *arg) { struct pde_action *act = arg; if (act->store == PCPU_GET(cpuid)) pde_store(act->pde, act->newpde); } static void pmap_update_pde_teardown(void *arg) { struct pde_action *act = arg; if (CPU_ISSET(PCPU_GET(cpuid), &act->invalidate)) pmap_update_pde_invalidate(act->va, act->newpde); } /* * Change the page size for the specified virtual address in a way that * prevents any possibility of the TLB ever having two entries that map the * same virtual address using different page sizes. This is the recommended * workaround for Erratum 383 on AMD Family 10h processors. It prevents a * machine check exception for a TLB state that is improperly diagnosed as a * hardware error. */ static void pmap_update_pde(pmap_t pmap, vm_offset_t va, pd_entry_t *pde, pd_entry_t newpde) { struct pde_action act; cpuset_t active, other_cpus; u_int cpuid; sched_pin(); cpuid = PCPU_GET(cpuid); other_cpus = all_cpus; CPU_CLR(cpuid, &other_cpus); if (pmap == kernel_pmap) active = all_cpus; else active = pmap->pm_active; if (CPU_OVERLAP(&active, &other_cpus)) { act.store = cpuid; act.invalidate = active; act.va = va; act.pde = pde; act.newpde = newpde; CPU_SET(cpuid, &active); smp_rendezvous_cpus(active, smp_no_rendezvous_barrier, pmap == kernel_pmap ? pmap_update_pde_kernel : pmap_update_pde_user, pmap_update_pde_teardown, &act); } else { if (pmap == kernel_pmap) pmap_kenter_pde(va, newpde); else pde_store(pde, newpde); if (CPU_ISSET(cpuid, &active)) pmap_update_pde_invalidate(va, newpde); } sched_unpin(); } #else /* !SMP */ /* * Normal, non-SMP, 486+ invalidation functions. * We inline these within pmap.c for speed. */ static void pmap_invalidate_page_int(pmap_t pmap, vm_offset_t va) { if (pmap == kernel_pmap) invlpg(va); } static void pmap_invalidate_range_int(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { vm_offset_t addr; if (pmap == kernel_pmap) for (addr = sva; addr < eva; addr += PAGE_SIZE) invlpg(addr); } static void pmap_invalidate_all_int(pmap_t pmap) { if (pmap == kernel_pmap) invltlb(); } static void __CONCAT(PMTYPE, invalidate_cache)(void) { wbinvd(); } static void pmap_update_pde(pmap_t pmap, vm_offset_t va, pd_entry_t *pde, pd_entry_t newpde) { if (pmap == kernel_pmap) pmap_kenter_pde(va, newpde); else pde_store(pde, newpde); if (pmap == kernel_pmap || !CPU_EMPTY(&pmap->pm_active)) pmap_update_pde_invalidate(va, newpde); } #endif /* !SMP */ static void __CONCAT(PMTYPE, invalidate_page)(pmap_t pmap, vm_offset_t va) { pmap_invalidate_page_int(pmap, va); } static void __CONCAT(PMTYPE, invalidate_range)(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { pmap_invalidate_range_int(pmap, sva, eva); } static void __CONCAT(PMTYPE, invalidate_all)(pmap_t pmap) { pmap_invalidate_all_int(pmap); } static void pmap_invalidate_pde_page(pmap_t pmap, vm_offset_t va, pd_entry_t pde) { /* * When the PDE has PG_PROMOTED set, the 2- or 4MB page mapping was * created by a promotion that did not invalidate the 512 or 1024 4KB * page mappings that might exist in the TLB. Consequently, at this * point, the TLB may hold both 4KB and 2- or 4MB page mappings for * the address range [va, va + NBPDR). Therefore, the entire range * must be invalidated here. In contrast, when PG_PROMOTED is clear, * the TLB will not hold any 4KB page mappings for the address range * [va, va + NBPDR), and so a single INVLPG suffices to invalidate the * 2- or 4MB page mapping from the TLB. */ if ((pde & PG_PROMOTED) != 0) pmap_invalidate_range_int(pmap, va, va + NBPDR - 1); else pmap_invalidate_page_int(pmap, va); } /* * Are we current address space or kernel? */ static __inline int pmap_is_current(pmap_t pmap) { return (pmap == kernel_pmap); } /* * If the given pmap is not the current or kernel pmap, the returned pte must * be released by passing it to pmap_pte_release(). */ static pt_entry_t * __CONCAT(PMTYPE, pte)(pmap_t pmap, vm_offset_t va) { pd_entry_t newpf; pd_entry_t *pde; pde = pmap_pde(pmap, va); if (*pde & PG_PS) return (pde); if (*pde != 0) { /* are we current address space or kernel? */ if (pmap_is_current(pmap)) return (vtopte(va)); mtx_lock(&PMAP2mutex); newpf = *pde & PG_FRAME; if ((*PMAP2 & PG_FRAME) != newpf) { *PMAP2 = newpf | PG_RW | PG_V | PG_A | PG_M; pmap_invalidate_page_int(kernel_pmap, (vm_offset_t)PADDR2); } return (PADDR2 + (i386_btop(va) & (NPTEPG - 1))); } return (NULL); } /* * Releases a pte that was obtained from pmap_pte(). Be prepared for the pte * being NULL. */ static __inline void pmap_pte_release(pt_entry_t *pte) { if ((pt_entry_t *)((vm_offset_t)pte & ~PAGE_MASK) == PADDR2) mtx_unlock(&PMAP2mutex); } /* * NB: The sequence of updating a page table followed by accesses to the * corresponding pages is subject to the situation described in the "AMD64 * Architecture Programmer's Manual Volume 2: System Programming" rev. 3.23, * "7.3.1 Special Coherency Considerations". Therefore, issuing the INVLPG * right after modifying the PTE bits is crucial. */ static __inline void invlcaddr(void *caddr) { invlpg((u_int)caddr); } /* * Super fast pmap_pte routine best used when scanning * the pv lists. This eliminates many coarse-grained * invltlb calls. Note that many of the pv list * scans are across different pmaps. It is very wasteful * to do an entire invltlb for checking a single mapping. * * If the given pmap is not the current pmap, pvh_global_lock * must be held and curthread pinned to a CPU. */ static pt_entry_t * pmap_pte_quick(pmap_t pmap, vm_offset_t va) { pd_entry_t newpf; pd_entry_t *pde; pde = pmap_pde(pmap, va); if (*pde & PG_PS) return (pde); if (*pde != 0) { /* are we current address space or kernel? */ if (pmap_is_current(pmap)) return (vtopte(va)); rw_assert(&pvh_global_lock, RA_WLOCKED); KASSERT(curthread->td_pinned > 0, ("curthread not pinned")); newpf = *pde & PG_FRAME; if ((*PMAP1 & PG_FRAME) != newpf) { *PMAP1 = newpf | PG_RW | PG_V | PG_A | PG_M; #ifdef SMP PMAP1cpu = PCPU_GET(cpuid); #endif invlcaddr(PADDR1); PMAP1changed++; } else #ifdef SMP if (PMAP1cpu != PCPU_GET(cpuid)) { PMAP1cpu = PCPU_GET(cpuid); invlcaddr(PADDR1); PMAP1changedcpu++; } else #endif PMAP1unchanged++; return (PADDR1 + (i386_btop(va) & (NPTEPG - 1))); } return (0); } static pt_entry_t * pmap_pte_quick3(pmap_t pmap, vm_offset_t va) { pd_entry_t newpf; pd_entry_t *pde; pde = pmap_pde(pmap, va); if (*pde & PG_PS) return (pde); if (*pde != 0) { rw_assert(&pvh_global_lock, RA_WLOCKED); KASSERT(curthread->td_pinned > 0, ("curthread not pinned")); newpf = *pde & PG_FRAME; if ((*PMAP3 & PG_FRAME) != newpf) { *PMAP3 = newpf | PG_RW | PG_V | PG_A | PG_M; #ifdef SMP PMAP3cpu = PCPU_GET(cpuid); #endif invlcaddr(PADDR3); PMAP1changed++; } else #ifdef SMP if (PMAP3cpu != PCPU_GET(cpuid)) { PMAP3cpu = PCPU_GET(cpuid); invlcaddr(PADDR3); PMAP1changedcpu++; } else #endif PMAP1unchanged++; return (PADDR3 + (i386_btop(va) & (NPTEPG - 1))); } return (0); } static pt_entry_t pmap_pte_ufast(pmap_t pmap, vm_offset_t va, pd_entry_t pde) { pt_entry_t *eh_ptep, pte, *ptep; PMAP_LOCK_ASSERT(pmap, MA_OWNED); pde &= PG_FRAME; critical_enter(); eh_ptep = (pt_entry_t *)PCPU_GET(pmap_eh_ptep); if ((*eh_ptep & PG_FRAME) != pde) { *eh_ptep = pde | PG_RW | PG_V | PG_A | PG_M; invlcaddr((void *)PCPU_GET(pmap_eh_va)); } ptep = (pt_entry_t *)PCPU_GET(pmap_eh_va) + (i386_btop(va) & (NPTEPG - 1)); pte = *ptep; critical_exit(); return (pte); } /* * Extract from the kernel page table the physical address that is mapped by * the given virtual address "va". * * This function may be used before pmap_bootstrap() is called. */ static vm_paddr_t __CONCAT(PMTYPE, kextract)(vm_offset_t va) { vm_paddr_t pa; if ((pa = pte_load(&PTD[va >> PDRSHIFT])) & PG_PS) { pa = (pa & PG_PS_FRAME) | (va & PDRMASK); } else { /* * Beware of a concurrent promotion that changes the PDE at * this point! For example, vtopte() must not be used to * access the PTE because it would use the new PDE. It is, * however, safe to use the old PDE because the page table * page is preserved by the promotion. */ pa = KPTmap[i386_btop(va)]; pa = (pa & PG_FRAME) | (va & PAGE_MASK); } return (pa); } /* * Routine: pmap_extract * Function: * Extract the physical page address associated * with the given map/virtual_address pair. */ static vm_paddr_t __CONCAT(PMTYPE, extract)(pmap_t pmap, vm_offset_t va) { vm_paddr_t rtval; pt_entry_t pte; pd_entry_t pde; rtval = 0; PMAP_LOCK(pmap); pde = pmap->pm_pdir[va >> PDRSHIFT]; if (pde != 0) { if ((pde & PG_PS) != 0) rtval = (pde & PG_PS_FRAME) | (va & PDRMASK); else { pte = pmap_pte_ufast(pmap, va, pde); rtval = (pte & PG_FRAME) | (va & PAGE_MASK); } } PMAP_UNLOCK(pmap); return (rtval); } /* * Routine: pmap_extract_and_hold * Function: * Atomically extract and hold the physical page * with the given pmap and virtual address pair * if that mapping permits the given protection. */ static vm_page_t __CONCAT(PMTYPE, extract_and_hold)(pmap_t pmap, vm_offset_t va, vm_prot_t prot) { pd_entry_t pde; pt_entry_t pte; vm_page_t m; m = NULL; PMAP_LOCK(pmap); pde = *pmap_pde(pmap, va); if (pde != 0) { if (pde & PG_PS) { if ((pde & PG_RW) || (prot & VM_PROT_WRITE) == 0) m = PHYS_TO_VM_PAGE((pde & PG_PS_FRAME) | (va & PDRMASK)); } else { pte = pmap_pte_ufast(pmap, va, pde); if (pte != 0 && ((pte & PG_RW) || (prot & VM_PROT_WRITE) == 0)) m = PHYS_TO_VM_PAGE(pte & PG_FRAME); } if (m != NULL && !vm_page_wire_mapped(m)) m = NULL; } PMAP_UNLOCK(pmap); return (m); } /*************************************************** * Low level mapping routines..... ***************************************************/ /* * Add a wired page to the kva. * Note: not SMP coherent. * * This function may be used before pmap_bootstrap() is called. */ static void __CONCAT(PMTYPE, kenter)(vm_offset_t va, vm_paddr_t pa) { pt_entry_t *pte; pte = vtopte(va); pte_store(pte, pa | PG_RW | PG_V); } static __inline void pmap_kenter_attr(vm_offset_t va, vm_paddr_t pa, int mode) { pt_entry_t *pte; pte = vtopte(va); pte_store(pte, pa | PG_RW | PG_V | pmap_cache_bits(kernel_pmap, mode, 0)); } /* * Remove a page from the kernel pagetables. * Note: not SMP coherent. * * This function may be used before pmap_bootstrap() is called. */ static void __CONCAT(PMTYPE, kremove)(vm_offset_t va) { pt_entry_t *pte; pte = vtopte(va); pte_clear(pte); } /* * Used to map a range of physical addresses into kernel * virtual address space. * * The value passed in '*virt' is a suggested virtual address for * the mapping. Architectures which can support a direct-mapped * physical to virtual region can return the appropriate address * within that region, leaving '*virt' unchanged. Other * architectures should map the pages starting at '*virt' and * update '*virt' with the first usable address after the mapped * region. */ static vm_offset_t __CONCAT(PMTYPE, map)(vm_offset_t *virt, vm_paddr_t start, vm_paddr_t end, int prot) { vm_offset_t va, sva; vm_paddr_t superpage_offset; pd_entry_t newpde; va = *virt; /* * Does the physical address range's size and alignment permit at * least one superpage mapping to be created? */ superpage_offset = start & PDRMASK; if ((end - start) - ((NBPDR - superpage_offset) & PDRMASK) >= NBPDR) { /* * Increase the starting virtual address so that its alignment * does not preclude the use of superpage mappings. */ if ((va & PDRMASK) < superpage_offset) va = (va & ~PDRMASK) + superpage_offset; else if ((va & PDRMASK) > superpage_offset) va = ((va + PDRMASK) & ~PDRMASK) + superpage_offset; } sva = va; while (start < end) { if ((start & PDRMASK) == 0 && end - start >= NBPDR && pseflag != 0) { KASSERT((va & PDRMASK) == 0, ("pmap_map: misaligned va %#x", va)); newpde = start | PG_PS | PG_RW | PG_V; pmap_kenter_pde(va, newpde); va += NBPDR; start += NBPDR; } else { pmap_kenter(va, start); va += PAGE_SIZE; start += PAGE_SIZE; } } pmap_invalidate_range_int(kernel_pmap, sva, va); *virt = va; return (sva); } /* * Add a list of wired pages to the kva * this routine is only used for temporary * kernel mappings that do not need to have * page modification or references recorded. * Note that old mappings are simply written * over. The page *must* be wired. * Note: SMP coherent. Uses a ranged shootdown IPI. */ static void __CONCAT(PMTYPE, qenter)(vm_offset_t sva, vm_page_t *ma, int count) { pt_entry_t *endpte, oldpte, pa, *pte; vm_page_t m; oldpte = 0; pte = vtopte(sva); endpte = pte + count; while (pte < endpte) { m = *ma++; pa = VM_PAGE_TO_PHYS(m) | pmap_cache_bits(kernel_pmap, m->md.pat_mode, 0); if ((*pte & (PG_FRAME | PG_PTE_CACHE)) != pa) { oldpte |= *pte; pte_store(pte, pa | pg_nx | PG_RW | PG_V); } pte++; } if (__predict_false((oldpte & PG_V) != 0)) pmap_invalidate_range_int(kernel_pmap, sva, sva + count * PAGE_SIZE); } /* * This routine tears out page mappings from the * kernel -- it is meant only for temporary mappings. * Note: SMP coherent. Uses a ranged shootdown IPI. */ static void __CONCAT(PMTYPE, qremove)(vm_offset_t sva, int count) { vm_offset_t va; va = sva; while (count-- > 0) { pmap_kremove(va); va += PAGE_SIZE; } pmap_invalidate_range_int(kernel_pmap, sva, va); } /*************************************************** * Page table page management routines..... ***************************************************/ /* * Schedule the specified unused page table page to be freed. Specifically, * add the page to the specified list of pages that will be released to the * physical memory manager after the TLB has been updated. */ static __inline void pmap_add_delayed_free_list(vm_page_t m, struct spglist *free, boolean_t set_PG_ZERO) { if (set_PG_ZERO) m->flags |= PG_ZERO; else m->flags &= ~PG_ZERO; SLIST_INSERT_HEAD(free, m, plinks.s.ss); } /* * Inserts the specified page table page into the specified pmap's collection * of idle page table pages. Each of a pmap's page table pages is responsible * for mapping a distinct range of virtual addresses. The pmap's collection is * ordered by this virtual address range. * * If "promoted" is false, then the page table page "mpte" must be zero filled. */ static __inline int pmap_insert_pt_page(pmap_t pmap, vm_page_t mpte, bool promoted) { PMAP_LOCK_ASSERT(pmap, MA_OWNED); mpte->valid = promoted ? VM_PAGE_BITS_ALL : 0; return (vm_radix_insert(&pmap->pm_root, mpte)); } /* * Removes the page table page mapping the specified virtual address from the * specified pmap's collection of idle page table pages, and returns it. * Otherwise, returns NULL if there is no page table page corresponding to the * specified virtual address. */ static __inline vm_page_t pmap_remove_pt_page(pmap_t pmap, vm_offset_t va) { PMAP_LOCK_ASSERT(pmap, MA_OWNED); return (vm_radix_remove(&pmap->pm_root, va >> PDRSHIFT)); } /* * Decrements a page table page's reference count, which is used to record the * number of valid page table entries within the page. If the reference count * drops to zero, then the page table page is unmapped. Returns TRUE if the * page table page was unmapped and FALSE otherwise. */ static inline boolean_t pmap_unwire_ptp(pmap_t pmap, vm_page_t m, struct spglist *free) { --m->ref_count; if (m->ref_count == 0) { _pmap_unwire_ptp(pmap, m, free); return (TRUE); } else return (FALSE); } static void _pmap_unwire_ptp(pmap_t pmap, vm_page_t m, struct spglist *free) { /* * unmap the page table page */ pmap->pm_pdir[m->pindex] = 0; --pmap->pm_stats.resident_count; /* * There is not need to invalidate the recursive mapping since * we never instantiate such mapping for the usermode pmaps, * and never remove page table pages from the kernel pmap. * Put page on a list so that it is released since all TLB * shootdown is done. */ MPASS(pmap != kernel_pmap); pmap_add_delayed_free_list(m, free, TRUE); } /* * After removing a page table entry, this routine is used to * conditionally free the page, and manage the reference count. */ static int pmap_unuse_pt(pmap_t pmap, vm_offset_t va, struct spglist *free) { pd_entry_t ptepde; vm_page_t mpte; if (pmap == kernel_pmap) return (0); ptepde = *pmap_pde(pmap, va); mpte = PHYS_TO_VM_PAGE(ptepde & PG_FRAME); return (pmap_unwire_ptp(pmap, mpte, free)); } /* * Release a page table page reference after a failed attempt to create a * mapping. */ static void pmap_abort_ptp(pmap_t pmap, vm_offset_t va, vm_page_t mpte) { struct spglist free; SLIST_INIT(&free); if (pmap_unwire_ptp(pmap, mpte, &free)) { /* * Although "va" was never mapped, paging-structure caches * could nonetheless have entries that refer to the freed * page table pages. Invalidate those entries. */ pmap_invalidate_page_int(pmap, va); vm_page_free_pages_toq(&free, true); } } /* * Initialize the pmap for the swapper process. */ static void __CONCAT(PMTYPE, pinit0)(pmap_t pmap) { PMAP_LOCK_INIT(pmap); pmap->pm_pdir = IdlePTD; #ifdef PMAP_PAE_COMP pmap->pm_pdpt = IdlePDPT; #endif vm_radix_init(&pmap->pm_root); CPU_ZERO(&pmap->pm_active); TAILQ_INIT(&pmap->pm_pvchunk); bzero(&pmap->pm_stats, sizeof pmap->pm_stats); pmap_activate_boot(pmap); } /* * Initialize a preallocated and zeroed pmap structure, * such as one in a vmspace structure. */ static int __CONCAT(PMTYPE, pinit)(pmap_t pmap) { int i; /* * No need to allocate page table space yet but we do need a valid * page directory table. */ if (pmap->pm_pdir == NULL) { pmap->pm_pdir = (pd_entry_t *)kva_alloc(NBPTD); if (pmap->pm_pdir == NULL) return (0); #ifdef PMAP_PAE_COMP pmap->pm_pdpt = uma_zalloc(pdptzone, M_WAITOK | M_ZERO); KASSERT(((vm_offset_t)pmap->pm_pdpt & ((NPGPTD * sizeof(pdpt_entry_t)) - 1)) == 0, ("pmap_pinit: pdpt misaligned")); KASSERT(pmap_kextract((vm_offset_t)pmap->pm_pdpt) < (4ULL<<30), ("pmap_pinit: pdpt above 4g")); #endif vm_radix_init(&pmap->pm_root); } KASSERT(vm_radix_is_empty(&pmap->pm_root), ("pmap_pinit: pmap has reserved page table page(s)")); /* * allocate the page directory page(s) */ for (i = 0; i < NPGPTD; i++) { pmap->pm_ptdpg[i] = vm_page_alloc_noobj(VM_ALLOC_WIRED | VM_ALLOC_ZERO | VM_ALLOC_WAITOK); #ifdef PMAP_PAE_COMP pmap->pm_pdpt[i] = VM_PAGE_TO_PHYS(pmap->pm_ptdpg[i]) | PG_V; #endif } pmap_qenter((vm_offset_t)pmap->pm_pdir, pmap->pm_ptdpg, NPGPTD); #ifdef PMAP_PAE_COMP if ((cpu_feature & CPUID_PAT) == 0) { pmap_invalidate_cache_range( trunc_page((vm_offset_t)pmap->pm_pdpt), round_page((vm_offset_t)pmap->pm_pdpt + NPGPTD * sizeof(pdpt_entry_t))); } #endif /* Install the trampoline mapping. */ pmap->pm_pdir[TRPTDI] = PTD[TRPTDI]; CPU_ZERO(&pmap->pm_active); TAILQ_INIT(&pmap->pm_pvchunk); bzero(&pmap->pm_stats, sizeof pmap->pm_stats); return (1); } /* * this routine is called if the page table page is not * mapped correctly. */ static vm_page_t _pmap_allocpte(pmap_t pmap, u_int ptepindex, u_int flags) { vm_paddr_t ptepa; vm_page_t m; /* * Allocate a page table page. */ if ((m = vm_page_alloc_noobj(VM_ALLOC_WIRED | VM_ALLOC_ZERO)) == NULL) { if ((flags & PMAP_ENTER_NOSLEEP) == 0) { PMAP_UNLOCK(pmap); rw_wunlock(&pvh_global_lock); vm_wait(NULL); rw_wlock(&pvh_global_lock); PMAP_LOCK(pmap); } /* * Indicate the need to retry. While waiting, the page table * page may have been allocated. */ return (NULL); } m->pindex = ptepindex; /* * Map the pagetable page into the process address space, if * it isn't already there. */ pmap->pm_stats.resident_count++; ptepa = VM_PAGE_TO_PHYS(m); KASSERT((pmap->pm_pdir[ptepindex] & PG_V) == 0, ("%s: page directory entry %#jx is valid", __func__, (uintmax_t)pmap->pm_pdir[ptepindex])); pmap->pm_pdir[ptepindex] = (pd_entry_t)(ptepa | PG_U | PG_RW | PG_V | PG_A | PG_M); return (m); } static vm_page_t pmap_allocpte(pmap_t pmap, vm_offset_t va, u_int flags) { u_int ptepindex; pd_entry_t ptepa; vm_page_t m; /* * Calculate pagetable page index */ ptepindex = va >> PDRSHIFT; retry: /* * Get the page directory entry */ ptepa = pmap->pm_pdir[ptepindex]; /* * This supports switching from a 4MB page to a * normal 4K page. */ if (ptepa & PG_PS) { (void)pmap_demote_pde(pmap, &pmap->pm_pdir[ptepindex], va); ptepa = pmap->pm_pdir[ptepindex]; } /* * If the page table page is mapped, we just increment the * hold count, and activate it. */ if (ptepa) { m = PHYS_TO_VM_PAGE(ptepa & PG_FRAME); m->ref_count++; } else { /* * Here if the pte page isn't mapped, or if it has * been deallocated. */ m = _pmap_allocpte(pmap, ptepindex, flags); if (m == NULL && (flags & PMAP_ENTER_NOSLEEP) == 0) goto retry; } return (m); } /*************************************************** * Pmap allocation/deallocation routines. ***************************************************/ /* * Release any resources held by the given physical map. * Called when a pmap initialized by pmap_pinit is being released. * Should only be called if the map contains no valid mappings. */ static void __CONCAT(PMTYPE, release)(pmap_t pmap) { vm_page_t m; int i; KASSERT(pmap->pm_stats.resident_count == 0, ("pmap_release: pmap resident count %ld != 0", pmap->pm_stats.resident_count)); KASSERT(vm_radix_is_empty(&pmap->pm_root), ("pmap_release: pmap has reserved page table page(s)")); KASSERT(CPU_EMPTY(&pmap->pm_active), ("releasing active pmap %p", pmap)); pmap_qremove((vm_offset_t)pmap->pm_pdir, NPGPTD); for (i = 0; i < NPGPTD; i++) { m = pmap->pm_ptdpg[i]; #ifdef PMAP_PAE_COMP KASSERT(VM_PAGE_TO_PHYS(m) == (pmap->pm_pdpt[i] & PG_FRAME), ("pmap_release: got wrong ptd page")); #endif vm_page_unwire_noq(m); vm_page_free(m); } } /* * grow the number of kernel page table entries, if needed */ static void __CONCAT(PMTYPE, growkernel)(vm_offset_t addr) { vm_paddr_t ptppaddr; vm_page_t nkpg; pd_entry_t newpdir; mtx_assert(&kernel_map->system_mtx, MA_OWNED); addr = roundup2(addr, NBPDR); if (addr - 1 >= vm_map_max(kernel_map)) addr = vm_map_max(kernel_map); while (kernel_vm_end < addr) { if (pdir_pde(PTD, kernel_vm_end)) { kernel_vm_end = (kernel_vm_end + NBPDR) & ~PDRMASK; if (kernel_vm_end - 1 >= vm_map_max(kernel_map)) { kernel_vm_end = vm_map_max(kernel_map); break; } continue; } nkpg = vm_page_alloc_noobj(VM_ALLOC_INTERRUPT | VM_ALLOC_WIRED | VM_ALLOC_ZERO); if (nkpg == NULL) panic("pmap_growkernel: no memory to grow kernel"); nkpg->pindex = kernel_vm_end >> PDRSHIFT; nkpt++; ptppaddr = VM_PAGE_TO_PHYS(nkpg); newpdir = (pd_entry_t) (ptppaddr | PG_V | PG_RW | PG_A | PG_M); pdir_pde(KPTD, kernel_vm_end) = newpdir; pmap_kenter_pde(kernel_vm_end, newpdir); kernel_vm_end = (kernel_vm_end + NBPDR) & ~PDRMASK; if (kernel_vm_end - 1 >= vm_map_max(kernel_map)) { kernel_vm_end = vm_map_max(kernel_map); break; } } } /*************************************************** * page management routines. ***************************************************/ CTASSERT(sizeof(struct pv_chunk) == PAGE_SIZE); CTASSERT(_NPCM == 11); CTASSERT(_NPCPV == 336); static __inline struct pv_chunk * pv_to_chunk(pv_entry_t pv) { return ((struct pv_chunk *)((uintptr_t)pv & ~(uintptr_t)PAGE_MASK)); } #define PV_PMAP(pv) (pv_to_chunk(pv)->pc_pmap) #define PC_FREE0_9 0xfffffffful /* Free values for index 0 through 9 */ #define PC_FREE10 0x0000fffful /* Free values for index 10 */ static const uint32_t pc_freemask[_NPCM] = { PC_FREE0_9, PC_FREE0_9, PC_FREE0_9, PC_FREE0_9, PC_FREE0_9, PC_FREE0_9, PC_FREE0_9, PC_FREE0_9, PC_FREE0_9, PC_FREE0_9, PC_FREE10 }; #ifdef PV_STATS extern int pc_chunk_count, pc_chunk_allocs, pc_chunk_frees, pc_chunk_tryfail; extern long pv_entry_frees, pv_entry_allocs; extern int pv_entry_spare; #endif /* * We are in a serious low memory condition. Resort to * drastic measures to free some pages so we can allocate * another pv entry chunk. */ static vm_page_t pmap_pv_reclaim(pmap_t locked_pmap) { struct pch newtail; struct pv_chunk *pc; struct md_page *pvh; pd_entry_t *pde; pmap_t pmap; pt_entry_t *pte, tpte; pv_entry_t pv; vm_offset_t va; vm_page_t m, m_pc; struct spglist free; uint32_t inuse; int bit, field, freed; PMAP_LOCK_ASSERT(locked_pmap, MA_OWNED); pmap = NULL; m_pc = NULL; SLIST_INIT(&free); TAILQ_INIT(&newtail); while ((pc = TAILQ_FIRST(&pv_chunks)) != NULL && (pv_vafree == 0 || SLIST_EMPTY(&free))) { TAILQ_REMOVE(&pv_chunks, pc, pc_lru); if (pmap != pc->pc_pmap) { if (pmap != NULL) { pmap_invalidate_all_int(pmap); if (pmap != locked_pmap) PMAP_UNLOCK(pmap); } pmap = pc->pc_pmap; /* Avoid deadlock and lock recursion. */ if (pmap > locked_pmap) PMAP_LOCK(pmap); else if (pmap != locked_pmap && !PMAP_TRYLOCK(pmap)) { pmap = NULL; TAILQ_INSERT_TAIL(&newtail, pc, pc_lru); continue; } } /* * Destroy every non-wired, 4 KB page mapping in the chunk. */ freed = 0; for (field = 0; field < _NPCM; field++) { for (inuse = ~pc->pc_map[field] & pc_freemask[field]; inuse != 0; inuse &= ~(1UL << bit)) { bit = bsfl(inuse); pv = &pc->pc_pventry[field * 32 + bit]; va = pv->pv_va; pde = pmap_pde(pmap, va); if ((*pde & PG_PS) != 0) continue; pte = __CONCAT(PMTYPE, pte)(pmap, va); tpte = *pte; if ((tpte & PG_W) == 0) tpte = pte_load_clear(pte); pmap_pte_release(pte); if ((tpte & PG_W) != 0) continue; KASSERT(tpte != 0, ("pmap_pv_reclaim: pmap %p va %x zero pte", pmap, va)); if ((tpte & PG_G) != 0) pmap_invalidate_page_int(pmap, va); m = PHYS_TO_VM_PAGE(tpte & PG_FRAME); if ((tpte & (PG_M | PG_RW)) == (PG_M | PG_RW)) vm_page_dirty(m); if ((tpte & PG_A) != 0) vm_page_aflag_set(m, PGA_REFERENCED); TAILQ_REMOVE(&m->md.pv_list, pv, pv_next); if (TAILQ_EMPTY(&m->md.pv_list) && (m->flags & PG_FICTITIOUS) == 0) { pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); if (TAILQ_EMPTY(&pvh->pv_list)) { vm_page_aflag_clear(m, PGA_WRITEABLE); } } pc->pc_map[field] |= 1UL << bit; pmap_unuse_pt(pmap, va, &free); freed++; } } if (freed == 0) { TAILQ_INSERT_TAIL(&newtail, pc, pc_lru); continue; } /* Every freed mapping is for a 4 KB page. */ pmap->pm_stats.resident_count -= freed; PV_STAT(pv_entry_frees += freed); PV_STAT(pv_entry_spare += freed); pv_entry_count -= freed; TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); for (field = 0; field < _NPCM; field++) if (pc->pc_map[field] != pc_freemask[field]) { TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list); TAILQ_INSERT_TAIL(&newtail, pc, pc_lru); /* * One freed pv entry in locked_pmap is * sufficient. */ if (pmap == locked_pmap) goto out; break; } if (field == _NPCM) { PV_STAT(pv_entry_spare -= _NPCPV); PV_STAT(pc_chunk_count--); PV_STAT(pc_chunk_frees++); /* Entire chunk is free; return it. */ m_pc = PHYS_TO_VM_PAGE(pmap_kextract((vm_offset_t)pc)); pmap_qremove((vm_offset_t)pc, 1); pmap_ptelist_free(&pv_vafree, (vm_offset_t)pc); break; } } out: TAILQ_CONCAT(&pv_chunks, &newtail, pc_lru); if (pmap != NULL) { pmap_invalidate_all_int(pmap); if (pmap != locked_pmap) PMAP_UNLOCK(pmap); } if (m_pc == NULL && pv_vafree != 0 && SLIST_EMPTY(&free)) { m_pc = SLIST_FIRST(&free); SLIST_REMOVE_HEAD(&free, plinks.s.ss); /* Recycle a freed page table page. */ m_pc->ref_count = 1; } vm_page_free_pages_toq(&free, true); return (m_pc); } /* * free the pv_entry back to the free list */ static void free_pv_entry(pmap_t pmap, pv_entry_t pv) { struct pv_chunk *pc; int idx, field, bit; rw_assert(&pvh_global_lock, RA_WLOCKED); PMAP_LOCK_ASSERT(pmap, MA_OWNED); PV_STAT(pv_entry_frees++); PV_STAT(pv_entry_spare++); pv_entry_count--; pc = pv_to_chunk(pv); idx = pv - &pc->pc_pventry[0]; field = idx / 32; bit = idx % 32; pc->pc_map[field] |= 1ul << bit; for (idx = 0; idx < _NPCM; idx++) if (pc->pc_map[idx] != pc_freemask[idx]) { /* * 98% of the time, pc is already at the head of the * list. If it isn't already, move it to the head. */ if (__predict_false(TAILQ_FIRST(&pmap->pm_pvchunk) != pc)) { TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list); } return; } TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); free_pv_chunk(pc); } static void free_pv_chunk(struct pv_chunk *pc) { vm_page_t m; TAILQ_REMOVE(&pv_chunks, pc, pc_lru); PV_STAT(pv_entry_spare -= _NPCPV); PV_STAT(pc_chunk_count--); PV_STAT(pc_chunk_frees++); /* entire chunk is free, return it */ m = PHYS_TO_VM_PAGE(pmap_kextract((vm_offset_t)pc)); pmap_qremove((vm_offset_t)pc, 1); vm_page_unwire_noq(m); vm_page_free(m); pmap_ptelist_free(&pv_vafree, (vm_offset_t)pc); } /* * get a new pv_entry, allocating a block from the system * when needed. */ static pv_entry_t get_pv_entry(pmap_t pmap, boolean_t try) { static const struct timeval printinterval = { 60, 0 }; static struct timeval lastprint; int bit, field; pv_entry_t pv; struct pv_chunk *pc; vm_page_t m; rw_assert(&pvh_global_lock, RA_WLOCKED); PMAP_LOCK_ASSERT(pmap, MA_OWNED); PV_STAT(pv_entry_allocs++); pv_entry_count++; if (pv_entry_count > pv_entry_high_water) if (ratecheck(&lastprint, &printinterval)) printf("Approaching the limit on PV entries, consider " "increasing either the vm.pmap.shpgperproc or the " "vm.pmap.pv_entries tunable.\n"); retry: pc = TAILQ_FIRST(&pmap->pm_pvchunk); if (pc != NULL) { for (field = 0; field < _NPCM; field++) { if (pc->pc_map[field]) { bit = bsfl(pc->pc_map[field]); break; } } if (field < _NPCM) { pv = &pc->pc_pventry[field * 32 + bit]; pc->pc_map[field] &= ~(1ul << bit); /* If this was the last item, move it to tail */ for (field = 0; field < _NPCM; field++) if (pc->pc_map[field] != 0) { PV_STAT(pv_entry_spare--); return (pv); /* not full, return */ } TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); TAILQ_INSERT_TAIL(&pmap->pm_pvchunk, pc, pc_list); PV_STAT(pv_entry_spare--); return (pv); } } /* * Access to the ptelist "pv_vafree" is synchronized by the pvh * global lock. If "pv_vafree" is currently non-empty, it will * remain non-empty until pmap_ptelist_alloc() completes. */ if (pv_vafree == 0 || (m = vm_page_alloc_noobj(VM_ALLOC_WIRED)) == NULL) { if (try) { pv_entry_count--; PV_STAT(pc_chunk_tryfail++); return (NULL); } m = pmap_pv_reclaim(pmap); if (m == NULL) goto retry; } PV_STAT(pc_chunk_count++); PV_STAT(pc_chunk_allocs++); pc = (struct pv_chunk *)pmap_ptelist_alloc(&pv_vafree); pmap_qenter((vm_offset_t)pc, &m, 1); pc->pc_pmap = pmap; pc->pc_map[0] = pc_freemask[0] & ~1ul; /* preallocated bit 0 */ for (field = 1; field < _NPCM; field++) pc->pc_map[field] = pc_freemask[field]; TAILQ_INSERT_TAIL(&pv_chunks, pc, pc_lru); pv = &pc->pc_pventry[0]; TAILQ_INSERT_HEAD(&pmap->pm_pvchunk, pc, pc_list); PV_STAT(pv_entry_spare += _NPCPV - 1); return (pv); } static __inline pv_entry_t pmap_pvh_remove(struct md_page *pvh, pmap_t pmap, vm_offset_t va) { pv_entry_t pv; rw_assert(&pvh_global_lock, RA_WLOCKED); TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) { if (pmap == PV_PMAP(pv) && va == pv->pv_va) { TAILQ_REMOVE(&pvh->pv_list, pv, pv_next); break; } } return (pv); } static void pmap_pv_demote_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa) { struct md_page *pvh; pv_entry_t pv; vm_offset_t va_last; vm_page_t m; rw_assert(&pvh_global_lock, RA_WLOCKED); KASSERT((pa & PDRMASK) == 0, ("pmap_pv_demote_pde: pa is not 4mpage aligned")); /* * Transfer the 4mpage's pv entry for this mapping to the first * page's pv list. */ pvh = pa_to_pvh(pa); va = trunc_4mpage(va); pv = pmap_pvh_remove(pvh, pmap, va); KASSERT(pv != NULL, ("pmap_pv_demote_pde: pv not found")); m = PHYS_TO_VM_PAGE(pa); TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next); /* Instantiate the remaining NPTEPG - 1 pv entries. */ va_last = va + NBPDR - PAGE_SIZE; do { m++; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_pv_demote_pde: page %p is not managed", m)); va += PAGE_SIZE; pmap_insert_entry(pmap, va, m); } while (va < va_last); } #if VM_NRESERVLEVEL > 0 static void pmap_pv_promote_pde(pmap_t pmap, vm_offset_t va, vm_paddr_t pa) { struct md_page *pvh; pv_entry_t pv; vm_offset_t va_last; vm_page_t m; rw_assert(&pvh_global_lock, RA_WLOCKED); KASSERT((pa & PDRMASK) == 0, ("pmap_pv_promote_pde: pa is not 4mpage aligned")); /* * Transfer the first page's pv entry for this mapping to the * 4mpage's pv list. Aside from avoiding the cost of a call * to get_pv_entry(), a transfer avoids the possibility that * get_pv_entry() calls pmap_collect() and that pmap_collect() * removes one of the mappings that is being promoted. */ m = PHYS_TO_VM_PAGE(pa); va = trunc_4mpage(va); pv = pmap_pvh_remove(&m->md, pmap, va); KASSERT(pv != NULL, ("pmap_pv_promote_pde: pv not found")); pvh = pa_to_pvh(pa); TAILQ_INSERT_TAIL(&pvh->pv_list, pv, pv_next); /* Free the remaining NPTEPG - 1 pv entries. */ va_last = va + NBPDR - PAGE_SIZE; do { m++; va += PAGE_SIZE; pmap_pvh_free(&m->md, pmap, va); } while (va < va_last); } #endif /* VM_NRESERVLEVEL > 0 */ static void pmap_pvh_free(struct md_page *pvh, pmap_t pmap, vm_offset_t va) { pv_entry_t pv; pv = pmap_pvh_remove(pvh, pmap, va); KASSERT(pv != NULL, ("pmap_pvh_free: pv not found")); free_pv_entry(pmap, pv); } static void pmap_remove_entry(pmap_t pmap, vm_page_t m, vm_offset_t va) { struct md_page *pvh; rw_assert(&pvh_global_lock, RA_WLOCKED); pmap_pvh_free(&m->md, pmap, va); if (TAILQ_EMPTY(&m->md.pv_list) && (m->flags & PG_FICTITIOUS) == 0) { pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); if (TAILQ_EMPTY(&pvh->pv_list)) vm_page_aflag_clear(m, PGA_WRITEABLE); } } /* * Create a pv entry for page at pa for * (pmap, va). */ static void pmap_insert_entry(pmap_t pmap, vm_offset_t va, vm_page_t m) { pv_entry_t pv; rw_assert(&pvh_global_lock, RA_WLOCKED); PMAP_LOCK_ASSERT(pmap, MA_OWNED); pv = get_pv_entry(pmap, FALSE); pv->pv_va = va; TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next); } /* * Conditionally create a pv entry. */ static boolean_t pmap_try_insert_pv_entry(pmap_t pmap, vm_offset_t va, vm_page_t m) { pv_entry_t pv; rw_assert(&pvh_global_lock, RA_WLOCKED); PMAP_LOCK_ASSERT(pmap, MA_OWNED); if (pv_entry_count < pv_entry_high_water && (pv = get_pv_entry(pmap, TRUE)) != NULL) { pv->pv_va = va; TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next); return (TRUE); } else return (FALSE); } /* * Create the pv entries for each of the pages within a superpage. */ static bool pmap_pv_insert_pde(pmap_t pmap, vm_offset_t va, pd_entry_t pde, u_int flags) { struct md_page *pvh; pv_entry_t pv; bool noreclaim; rw_assert(&pvh_global_lock, RA_WLOCKED); noreclaim = (flags & PMAP_ENTER_NORECLAIM) != 0; if ((noreclaim && pv_entry_count >= pv_entry_high_water) || (pv = get_pv_entry(pmap, noreclaim)) == NULL) return (false); pv->pv_va = va; pvh = pa_to_pvh(pde & PG_PS_FRAME); TAILQ_INSERT_TAIL(&pvh->pv_list, pv, pv_next); return (true); } /* * Fills a page table page with mappings to consecutive physical pages. */ static void pmap_fill_ptp(pt_entry_t *firstpte, pt_entry_t newpte) { pt_entry_t *pte; for (pte = firstpte; pte < firstpte + NPTEPG; pte++) { *pte = newpte; newpte += PAGE_SIZE; } } /* * Tries to demote a 2- or 4MB page mapping. If demotion fails, the * 2- or 4MB page mapping is invalidated. */ static boolean_t pmap_demote_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va) { pd_entry_t newpde, oldpde; pt_entry_t *firstpte, newpte; vm_paddr_t mptepa; vm_page_t mpte; struct spglist free; vm_offset_t sva; PMAP_LOCK_ASSERT(pmap, MA_OWNED); oldpde = *pde; KASSERT((oldpde & (PG_PS | PG_V)) == (PG_PS | PG_V), ("pmap_demote_pde: oldpde is missing PG_PS and/or PG_V")); if ((oldpde & PG_A) == 0 || (mpte = pmap_remove_pt_page(pmap, va)) == NULL) { KASSERT((oldpde & PG_W) == 0, ("pmap_demote_pde: page table page for a wired mapping" " is missing")); /* * Invalidate the 2- or 4MB page mapping and return * "failure" if the mapping was never accessed or the * allocation of the new page table page fails. */ if ((oldpde & PG_A) == 0 || (mpte = vm_page_alloc_noobj(VM_ALLOC_WIRED)) == NULL) { SLIST_INIT(&free); sva = trunc_4mpage(va); pmap_remove_pde(pmap, pde, sva, &free); if ((oldpde & PG_G) == 0) pmap_invalidate_pde_page(pmap, sva, oldpde); vm_page_free_pages_toq(&free, true); CTR2(KTR_PMAP, "pmap_demote_pde: failure for va %#x" " in pmap %p", va, pmap); return (FALSE); } mpte->pindex = va >> PDRSHIFT; if (pmap != kernel_pmap) { mpte->ref_count = NPTEPG; pmap->pm_stats.resident_count++; } } mptepa = VM_PAGE_TO_PHYS(mpte); /* * If the page mapping is in the kernel's address space, then the * KPTmap can provide access to the page table page. Otherwise, * temporarily map the page table page (mpte) into the kernel's * address space at either PADDR1 or PADDR2. */ if (pmap == kernel_pmap) firstpte = &KPTmap[i386_btop(trunc_4mpage(va))]; else if (curthread->td_pinned > 0 && rw_wowned(&pvh_global_lock)) { if ((*PMAP1 & PG_FRAME) != mptepa) { *PMAP1 = mptepa | PG_RW | PG_V | PG_A | PG_M; #ifdef SMP PMAP1cpu = PCPU_GET(cpuid); #endif invlcaddr(PADDR1); PMAP1changed++; } else #ifdef SMP if (PMAP1cpu != PCPU_GET(cpuid)) { PMAP1cpu = PCPU_GET(cpuid); invlcaddr(PADDR1); PMAP1changedcpu++; } else #endif PMAP1unchanged++; firstpte = PADDR1; } else { mtx_lock(&PMAP2mutex); if ((*PMAP2 & PG_FRAME) != mptepa) { *PMAP2 = mptepa | PG_RW | PG_V | PG_A | PG_M; pmap_invalidate_page_int(kernel_pmap, (vm_offset_t)PADDR2); } firstpte = PADDR2; } newpde = mptepa | PG_M | PG_A | (oldpde & PG_U) | PG_RW | PG_V; KASSERT((oldpde & PG_A) != 0, ("pmap_demote_pde: oldpde is missing PG_A")); KASSERT((oldpde & (PG_M | PG_RW)) != PG_RW, ("pmap_demote_pde: oldpde is missing PG_M")); newpte = oldpde & ~PG_PS; if ((newpte & PG_PDE_PAT) != 0) newpte ^= PG_PDE_PAT | PG_PTE_PAT; /* * If the page table page is not leftover from an earlier promotion, * initialize it. */ if (mpte->valid == 0) pmap_fill_ptp(firstpte, newpte); KASSERT((*firstpte & PG_FRAME) == (newpte & PG_FRAME), ("pmap_demote_pde: firstpte and newpte map different physical" " addresses")); /* * If the mapping has changed attributes, update the page table * entries. */ if ((*firstpte & PG_PTE_PROMOTE) != (newpte & PG_PTE_PROMOTE)) pmap_fill_ptp(firstpte, newpte); /* * Demote the mapping. This pmap is locked. The old PDE has * PG_A set. If the old PDE has PG_RW set, it also has PG_M * set. Thus, there is no danger of a race with another * processor changing the setting of PG_A and/or PG_M between * the read above and the store below. */ if (workaround_erratum383) pmap_update_pde(pmap, va, pde, newpde); else if (pmap == kernel_pmap) pmap_kenter_pde(va, newpde); else pde_store(pde, newpde); if (firstpte == PADDR2) mtx_unlock(&PMAP2mutex); /* * Invalidate the recursive mapping of the page table page. */ pmap_invalidate_page_int(pmap, (vm_offset_t)vtopte(va)); /* * Demote the pv entry. This depends on the earlier demotion * of the mapping. Specifically, the (re)creation of a per- * page pv entry might trigger the execution of pmap_collect(), * which might reclaim a newly (re)created per-page pv entry * and destroy the associated mapping. In order to destroy * the mapping, the PDE must have already changed from mapping * the 2mpage to referencing the page table page. */ if ((oldpde & PG_MANAGED) != 0) pmap_pv_demote_pde(pmap, va, oldpde & PG_PS_FRAME); pmap_pde_demotions++; CTR2(KTR_PMAP, "pmap_demote_pde: success for va %#x" " in pmap %p", va, pmap); return (TRUE); } /* * Removes a 2- or 4MB page mapping from the kernel pmap. */ static void pmap_remove_kernel_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va) { pd_entry_t newpde; vm_paddr_t mptepa; vm_page_t mpte; PMAP_LOCK_ASSERT(pmap, MA_OWNED); mpte = pmap_remove_pt_page(pmap, va); if (mpte == NULL) panic("pmap_remove_kernel_pde: Missing pt page."); mptepa = VM_PAGE_TO_PHYS(mpte); newpde = mptepa | PG_M | PG_A | PG_RW | PG_V; /* * If this page table page was unmapped by a promotion, then it * contains valid mappings. Zero it to invalidate those mappings. */ if (mpte->valid != 0) pagezero((void *)&KPTmap[i386_btop(trunc_4mpage(va))]); /* * Remove the mapping. */ if (workaround_erratum383) pmap_update_pde(pmap, va, pde, newpde); else pmap_kenter_pde(va, newpde); /* * Invalidate the recursive mapping of the page table page. */ pmap_invalidate_page_int(pmap, (vm_offset_t)vtopte(va)); } /* * pmap_remove_pde: do the things to unmap a superpage in a process */ static void pmap_remove_pde(pmap_t pmap, pd_entry_t *pdq, vm_offset_t sva, struct spglist *free) { struct md_page *pvh; pd_entry_t oldpde; vm_offset_t eva, va; vm_page_t m, mpte; PMAP_LOCK_ASSERT(pmap, MA_OWNED); KASSERT((sva & PDRMASK) == 0, ("pmap_remove_pde: sva is not 4mpage aligned")); oldpde = pte_load_clear(pdq); if (oldpde & PG_W) pmap->pm_stats.wired_count -= NBPDR / PAGE_SIZE; /* * Machines that don't support invlpg, also don't support * PG_G. */ if ((oldpde & PG_G) != 0) pmap_invalidate_pde_page(kernel_pmap, sva, oldpde); pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; if (oldpde & PG_MANAGED) { pvh = pa_to_pvh(oldpde & PG_PS_FRAME); pmap_pvh_free(pvh, pmap, sva); eva = sva + NBPDR; for (va = sva, m = PHYS_TO_VM_PAGE(oldpde & PG_PS_FRAME); va < eva; va += PAGE_SIZE, m++) { if ((oldpde & (PG_M | PG_RW)) == (PG_M | PG_RW)) vm_page_dirty(m); if (oldpde & PG_A) vm_page_aflag_set(m, PGA_REFERENCED); if (TAILQ_EMPTY(&m->md.pv_list) && TAILQ_EMPTY(&pvh->pv_list)) vm_page_aflag_clear(m, PGA_WRITEABLE); } } if (pmap == kernel_pmap) { pmap_remove_kernel_pde(pmap, pdq, sva); } else { mpte = pmap_remove_pt_page(pmap, sva); if (mpte != NULL) { KASSERT(mpte->valid == VM_PAGE_BITS_ALL, ("pmap_remove_pde: pte page not promoted")); pmap->pm_stats.resident_count--; KASSERT(mpte->ref_count == NPTEPG, ("pmap_remove_pde: pte page ref count error")); mpte->ref_count = 0; pmap_add_delayed_free_list(mpte, free, FALSE); } } } /* * pmap_remove_pte: do the things to unmap a page in a process */ static int pmap_remove_pte(pmap_t pmap, pt_entry_t *ptq, vm_offset_t va, struct spglist *free) { pt_entry_t oldpte; vm_page_t m; rw_assert(&pvh_global_lock, RA_WLOCKED); PMAP_LOCK_ASSERT(pmap, MA_OWNED); oldpte = pte_load_clear(ptq); KASSERT(oldpte != 0, ("pmap_remove_pte: pmap %p va %x zero pte", pmap, va)); if (oldpte & PG_W) pmap->pm_stats.wired_count -= 1; /* * Machines that don't support invlpg, also don't support * PG_G. */ if (oldpte & PG_G) pmap_invalidate_page_int(kernel_pmap, va); pmap->pm_stats.resident_count -= 1; if (oldpte & PG_MANAGED) { m = PHYS_TO_VM_PAGE(oldpte & PG_FRAME); if ((oldpte & (PG_M | PG_RW)) == (PG_M | PG_RW)) vm_page_dirty(m); if (oldpte & PG_A) vm_page_aflag_set(m, PGA_REFERENCED); pmap_remove_entry(pmap, m, va); } return (pmap_unuse_pt(pmap, va, free)); } /* * Remove a single page from a process address space */ static void pmap_remove_page(pmap_t pmap, vm_offset_t va, struct spglist *free) { pt_entry_t *pte; rw_assert(&pvh_global_lock, RA_WLOCKED); KASSERT(curthread->td_pinned > 0, ("curthread not pinned")); PMAP_LOCK_ASSERT(pmap, MA_OWNED); if ((pte = pmap_pte_quick(pmap, va)) == NULL || *pte == 0) return; pmap_remove_pte(pmap, pte, va, free); pmap_invalidate_page_int(pmap, va); } /* * Removes the specified range of addresses from the page table page. */ static bool pmap_remove_ptes(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, struct spglist *free) { pt_entry_t *pte; bool anyvalid; rw_assert(&pvh_global_lock, RA_WLOCKED); KASSERT(curthread->td_pinned > 0, ("curthread not pinned")); PMAP_LOCK_ASSERT(pmap, MA_OWNED); anyvalid = false; for (pte = pmap_pte_quick(pmap, sva); sva != eva; pte++, sva += PAGE_SIZE) { if (*pte == 0) continue; /* * The TLB entry for a PG_G mapping is invalidated by * pmap_remove_pte(). */ if ((*pte & PG_G) == 0) anyvalid = true; if (pmap_remove_pte(pmap, pte, sva, free)) break; } return (anyvalid); } /* * Remove the given range of addresses from the specified map. * * It is assumed that the start and end are properly * rounded to the page size. */ static void __CONCAT(PMTYPE, remove)(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { vm_offset_t pdnxt; pd_entry_t ptpaddr; struct spglist free; int anyvalid; /* * Perform an unsynchronized read. This is, however, safe. */ if (pmap->pm_stats.resident_count == 0) return; anyvalid = 0; SLIST_INIT(&free); rw_wlock(&pvh_global_lock); sched_pin(); PMAP_LOCK(pmap); /* * special handling of removing one page. a very * common operation and easy to short circuit some * code. */ if ((sva + PAGE_SIZE == eva) && ((pmap->pm_pdir[(sva >> PDRSHIFT)] & PG_PS) == 0)) { pmap_remove_page(pmap, sva, &free); goto out; } for (; sva < eva; sva = pdnxt) { u_int pdirindex; /* * Calculate index for next page table. */ pdnxt = (sva + NBPDR) & ~PDRMASK; if (pdnxt < sva) pdnxt = eva; if (pmap->pm_stats.resident_count == 0) break; pdirindex = sva >> PDRSHIFT; ptpaddr = pmap->pm_pdir[pdirindex]; /* * Weed out invalid mappings. Note: we assume that the page * directory table is always allocated, and in kernel virtual. */ if (ptpaddr == 0) continue; /* * Check for large page. */ if ((ptpaddr & PG_PS) != 0) { /* * Are we removing the entire large page? If not, * demote the mapping and fall through. */ if (sva + NBPDR == pdnxt && eva >= pdnxt) { /* * The TLB entry for a PG_G mapping is * invalidated by pmap_remove_pde(). */ if ((ptpaddr & PG_G) == 0) anyvalid = 1; pmap_remove_pde(pmap, &pmap->pm_pdir[pdirindex], sva, &free); continue; } else if (!pmap_demote_pde(pmap, &pmap->pm_pdir[pdirindex], sva)) { /* The large page mapping was destroyed. */ continue; } } /* * Limit our scan to either the end of the va represented * by the current page table page, or to the end of the * range being removed. */ if (pdnxt > eva) pdnxt = eva; if (pmap_remove_ptes(pmap, sva, pdnxt, &free)) anyvalid = 1; } out: sched_unpin(); if (anyvalid) pmap_invalidate_all_int(pmap); rw_wunlock(&pvh_global_lock); PMAP_UNLOCK(pmap); vm_page_free_pages_toq(&free, true); } /* * Routine: pmap_remove_all * Function: * Removes this physical page from * all physical maps in which it resides. * Reflects back modify bits to the pager. * * Notes: * Original versions of this routine were very * inefficient because they iteratively called * pmap_remove (slow...) */ static void __CONCAT(PMTYPE, remove_all)(vm_page_t m) { struct md_page *pvh; pv_entry_t pv; pmap_t pmap; pt_entry_t *pte, tpte; pd_entry_t *pde; vm_offset_t va; struct spglist free; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_remove_all: page %p is not managed", m)); SLIST_INIT(&free); rw_wlock(&pvh_global_lock); sched_pin(); if ((m->flags & PG_FICTITIOUS) != 0) goto small_mappings; pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); while ((pv = TAILQ_FIRST(&pvh->pv_list)) != NULL) { va = pv->pv_va; pmap = PV_PMAP(pv); PMAP_LOCK(pmap); pde = pmap_pde(pmap, va); (void)pmap_demote_pde(pmap, pde, va); PMAP_UNLOCK(pmap); } small_mappings: while ((pv = TAILQ_FIRST(&m->md.pv_list)) != NULL) { pmap = PV_PMAP(pv); PMAP_LOCK(pmap); pmap->pm_stats.resident_count--; pde = pmap_pde(pmap, pv->pv_va); KASSERT((*pde & PG_PS) == 0, ("pmap_remove_all: found" " a 4mpage in page %p's pv list", m)); pte = pmap_pte_quick(pmap, pv->pv_va); tpte = pte_load_clear(pte); KASSERT(tpte != 0, ("pmap_remove_all: pmap %p va %x zero pte", pmap, pv->pv_va)); if (tpte & PG_W) pmap->pm_stats.wired_count--; if (tpte & PG_A) vm_page_aflag_set(m, PGA_REFERENCED); /* * Update the vm_page_t clean and reference bits. */ if ((tpte & (PG_M | PG_RW)) == (PG_M | PG_RW)) vm_page_dirty(m); pmap_unuse_pt(pmap, pv->pv_va, &free); pmap_invalidate_page_int(pmap, pv->pv_va); TAILQ_REMOVE(&m->md.pv_list, pv, pv_next); free_pv_entry(pmap, pv); PMAP_UNLOCK(pmap); } vm_page_aflag_clear(m, PGA_WRITEABLE); sched_unpin(); rw_wunlock(&pvh_global_lock); vm_page_free_pages_toq(&free, true); } /* * pmap_protect_pde: do the things to protect a 4mpage in a process */ static boolean_t pmap_protect_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t sva, vm_prot_t prot) { pd_entry_t newpde, oldpde; vm_page_t m, mt; boolean_t anychanged; PMAP_LOCK_ASSERT(pmap, MA_OWNED); KASSERT((sva & PDRMASK) == 0, ("pmap_protect_pde: sva is not 4mpage aligned")); anychanged = FALSE; retry: oldpde = newpde = *pde; if ((prot & VM_PROT_WRITE) == 0) { if ((oldpde & (PG_MANAGED | PG_M | PG_RW)) == (PG_MANAGED | PG_M | PG_RW)) { m = PHYS_TO_VM_PAGE(oldpde & PG_PS_FRAME); for (mt = m; mt < &m[NBPDR / PAGE_SIZE]; mt++) vm_page_dirty(mt); } newpde &= ~(PG_RW | PG_M); } #ifdef PMAP_PAE_COMP if ((prot & VM_PROT_EXECUTE) == 0 && !i386_read_exec) newpde |= pg_nx; #endif if (newpde != oldpde) { /* * As an optimization to future operations on this PDE, clear * PG_PROMOTED. The impending invalidation will remove any * lingering 4KB page mappings from the TLB. */ if (!pde_cmpset(pde, oldpde, newpde & ~PG_PROMOTED)) goto retry; if ((oldpde & PG_G) != 0) pmap_invalidate_pde_page(kernel_pmap, sva, oldpde); else anychanged = TRUE; } return (anychanged); } /* * Set the physical protection on the * specified range of this map as requested. */ static void __CONCAT(PMTYPE, protect)(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, vm_prot_t prot) { vm_offset_t pdnxt; pd_entry_t ptpaddr; pt_entry_t *pte; boolean_t anychanged, pv_lists_locked; KASSERT((prot & ~VM_PROT_ALL) == 0, ("invalid prot %x", prot)); if (prot == VM_PROT_NONE) { pmap_remove(pmap, sva, eva); return; } #ifdef PMAP_PAE_COMP if ((prot & (VM_PROT_WRITE | VM_PROT_EXECUTE)) == (VM_PROT_WRITE | VM_PROT_EXECUTE)) return; #else if (prot & VM_PROT_WRITE) return; #endif if (pmap_is_current(pmap)) pv_lists_locked = FALSE; else { pv_lists_locked = TRUE; resume: rw_wlock(&pvh_global_lock); sched_pin(); } anychanged = FALSE; PMAP_LOCK(pmap); for (; sva < eva; sva = pdnxt) { pt_entry_t obits, pbits; u_int pdirindex; pdnxt = (sva + NBPDR) & ~PDRMASK; if (pdnxt < sva) pdnxt = eva; pdirindex = sva >> PDRSHIFT; ptpaddr = pmap->pm_pdir[pdirindex]; /* * Weed out invalid mappings. Note: we assume that the page * directory table is always allocated, and in kernel virtual. */ if (ptpaddr == 0) continue; /* * Check for large page. */ if ((ptpaddr & PG_PS) != 0) { /* * Are we protecting the entire large page? If not, * demote the mapping and fall through. */ if (sva + NBPDR == pdnxt && eva >= pdnxt) { /* * The TLB entry for a PG_G mapping is * invalidated by pmap_protect_pde(). */ if (pmap_protect_pde(pmap, &pmap->pm_pdir[pdirindex], sva, prot)) anychanged = TRUE; continue; } else { if (!pv_lists_locked) { pv_lists_locked = TRUE; if (!rw_try_wlock(&pvh_global_lock)) { if (anychanged) pmap_invalidate_all_int( pmap); PMAP_UNLOCK(pmap); goto resume; } sched_pin(); } if (!pmap_demote_pde(pmap, &pmap->pm_pdir[pdirindex], sva)) { /* * The large page mapping was * destroyed. */ continue; } } } if (pdnxt > eva) pdnxt = eva; for (pte = pmap_pte_quick(pmap, sva); sva != pdnxt; pte++, sva += PAGE_SIZE) { vm_page_t m; retry: /* * Regardless of whether a pte is 32 or 64 bits in * size, PG_RW, PG_A, and PG_M are among the least * significant 32 bits. */ obits = pbits = *pte; if ((pbits & PG_V) == 0) continue; if ((prot & VM_PROT_WRITE) == 0) { if ((pbits & (PG_MANAGED | PG_M | PG_RW)) == (PG_MANAGED | PG_M | PG_RW)) { m = PHYS_TO_VM_PAGE(pbits & PG_FRAME); vm_page_dirty(m); } pbits &= ~(PG_RW | PG_M); } #ifdef PMAP_PAE_COMP if ((prot & VM_PROT_EXECUTE) == 0 && !i386_read_exec) pbits |= pg_nx; #endif if (pbits != obits) { #ifdef PMAP_PAE_COMP if (!atomic_cmpset_64(pte, obits, pbits)) goto retry; #else if (!atomic_cmpset_int((u_int *)pte, obits, pbits)) goto retry; #endif if (obits & PG_G) pmap_invalidate_page_int(pmap, sva); else anychanged = TRUE; } } } if (anychanged) pmap_invalidate_all_int(pmap); if (pv_lists_locked) { sched_unpin(); rw_wunlock(&pvh_global_lock); } PMAP_UNLOCK(pmap); } #if VM_NRESERVLEVEL > 0 /* * Tries to promote the 512 or 1024, contiguous 4KB page mappings that are * within a single page table page (PTP) to a single 2- or 4MB page mapping. * For promotion to occur, two conditions must be met: (1) the 4KB page * mappings must map aligned, contiguous physical memory and (2) the 4KB page * mappings must have identical characteristics. * * Managed (PG_MANAGED) mappings within the kernel address space are not * promoted. The reason is that kernel PDEs are replicated in each pmap but * pmap_clear_ptes() and pmap_ts_referenced() only read the PDE from the kernel * pmap. */ static void pmap_promote_pde(pmap_t pmap, pd_entry_t *pde, vm_offset_t va) { pd_entry_t newpde; pt_entry_t *firstpte, oldpte, pa, *pte; vm_offset_t oldpteva __diagused; vm_page_t mpte; PMAP_LOCK_ASSERT(pmap, MA_OWNED); /* * Examine the first PTE in the specified PTP. Abort if this PTE is * either invalid, unused, or does not map the first 4KB physical page * within a 2- or 4MB page. */ firstpte = pmap_pte_quick(pmap, trunc_4mpage(va)); setpde: newpde = *firstpte; if ((newpde & ((PG_FRAME & PDRMASK) | PG_A | PG_V)) != (PG_A | PG_V)) { pmap_pde_p_failures++; CTR2(KTR_PMAP, "pmap_promote_pde: failure for va %#x" " in pmap %p", va, pmap); return; } if ((*firstpte & PG_MANAGED) != 0 && pmap == kernel_pmap) { pmap_pde_p_failures++; CTR2(KTR_PMAP, "pmap_promote_pde: failure for va %#x" " in pmap %p", va, pmap); return; } if ((newpde & (PG_M | PG_RW)) == PG_RW) { /* * When PG_M is already clear, PG_RW can be cleared without * a TLB invalidation. */ if (!atomic_cmpset_int((u_int *)firstpte, newpde, newpde & ~PG_RW)) goto setpde; newpde &= ~PG_RW; } /* * Examine each of the other PTEs in the specified PTP. Abort if this * PTE maps an unexpected 4KB physical page or does not have identical * characteristics to the first PTE. */ pa = (newpde & (PG_PS_FRAME | PG_A | PG_V)) + NBPDR - PAGE_SIZE; for (pte = firstpte + NPTEPG - 1; pte > firstpte; pte--) { setpte: oldpte = *pte; if ((oldpte & (PG_FRAME | PG_A | PG_V)) != pa) { pmap_pde_p_failures++; CTR2(KTR_PMAP, "pmap_promote_pde: failure for va %#x" " in pmap %p", va, pmap); return; } if ((oldpte & (PG_M | PG_RW)) == PG_RW) { /* * When PG_M is already clear, PG_RW can be cleared * without a TLB invalidation. */ if (!atomic_cmpset_int((u_int *)pte, oldpte, oldpte & ~PG_RW)) goto setpte; oldpte &= ~PG_RW; oldpteva = (oldpte & PG_FRAME & PDRMASK) | (va & ~PDRMASK); CTR2(KTR_PMAP, "pmap_promote_pde: protect for va %#x" " in pmap %p", oldpteva, pmap); } if ((oldpte & PG_PTE_PROMOTE) != (newpde & PG_PTE_PROMOTE)) { pmap_pde_p_failures++; CTR2(KTR_PMAP, "pmap_promote_pde: failure for va %#x" " in pmap %p", va, pmap); return; } pa -= PAGE_SIZE; } /* * Save the page table page in its current state until the PDE * mapping the superpage is demoted by pmap_demote_pde() or * destroyed by pmap_remove_pde(). */ mpte = PHYS_TO_VM_PAGE(*pde & PG_FRAME); KASSERT(mpte >= vm_page_array && mpte < &vm_page_array[vm_page_array_size], ("pmap_promote_pde: page table page is out of range")); KASSERT(mpte->pindex == va >> PDRSHIFT, ("pmap_promote_pde: page table page's pindex is wrong")); if (pmap_insert_pt_page(pmap, mpte, true)) { pmap_pde_p_failures++; CTR2(KTR_PMAP, "pmap_promote_pde: failure for va %#x in pmap %p", va, pmap); return; } /* * Promote the pv entries. */ if ((newpde & PG_MANAGED) != 0) pmap_pv_promote_pde(pmap, va, newpde & PG_PS_FRAME); /* * Propagate the PAT index to its proper position. */ if ((newpde & PG_PTE_PAT) != 0) newpde ^= PG_PDE_PAT | PG_PTE_PAT; /* * Map the superpage. */ if (workaround_erratum383) pmap_update_pde(pmap, va, pde, PG_PS | newpde); else if (pmap == kernel_pmap) pmap_kenter_pde(va, PG_PROMOTED | PG_PS | newpde); else pde_store(pde, PG_PROMOTED | PG_PS | newpde); pmap_pde_promotions++; CTR2(KTR_PMAP, "pmap_promote_pde: success for va %#x" " in pmap %p", va, pmap); } #endif /* VM_NRESERVLEVEL > 0 */ /* * Insert the given physical page (p) at * the specified virtual address (v) in the * target physical map with the protection requested. * * If specified, the page will be wired down, meaning * that the related pte can not be reclaimed. * * NB: This is the only routine which MAY NOT lazy-evaluate * or lose information. That is, this routine must actually * insert this page into the given map NOW. */ static int __CONCAT(PMTYPE, enter)(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, u_int flags, int8_t psind) { pd_entry_t *pde; pt_entry_t *pte; pt_entry_t newpte, origpte; pv_entry_t pv; vm_paddr_t opa, pa; vm_page_t mpte, om; int rv; va = trunc_page(va); KASSERT((pmap == kernel_pmap && va < VM_MAX_KERNEL_ADDRESS) || (pmap != kernel_pmap && va < VM_MAXUSER_ADDRESS), ("pmap_enter: toobig k%d %#x", pmap == kernel_pmap, va)); KASSERT(va < PMAP_TRM_MIN_ADDRESS, ("pmap_enter: invalid to pmap_enter into trampoline (va: 0x%x)", va)); KASSERT(pmap != kernel_pmap || (m->oflags & VPO_UNMANAGED) != 0 || va < kmi.clean_sva || va >= kmi.clean_eva, ("pmap_enter: managed mapping within the clean submap")); if ((m->oflags & VPO_UNMANAGED) == 0) VM_PAGE_OBJECT_BUSY_ASSERT(m); KASSERT((flags & PMAP_ENTER_RESERVED) == 0, ("pmap_enter: flags %u has reserved bits set", flags)); pa = VM_PAGE_TO_PHYS(m); newpte = (pt_entry_t)(pa | PG_A | PG_V); if ((flags & VM_PROT_WRITE) != 0) newpte |= PG_M; if ((prot & VM_PROT_WRITE) != 0) newpte |= PG_RW; KASSERT((newpte & (PG_M | PG_RW)) != PG_M, ("pmap_enter: flags includes VM_PROT_WRITE but prot doesn't")); #ifdef PMAP_PAE_COMP if ((prot & VM_PROT_EXECUTE) == 0 && !i386_read_exec) newpte |= pg_nx; #endif if ((flags & PMAP_ENTER_WIRED) != 0) newpte |= PG_W; if (pmap != kernel_pmap) newpte |= PG_U; newpte |= pmap_cache_bits(pmap, m->md.pat_mode, psind > 0); if ((m->oflags & VPO_UNMANAGED) == 0) newpte |= PG_MANAGED; rw_wlock(&pvh_global_lock); PMAP_LOCK(pmap); sched_pin(); if (psind == 1) { /* Assert the required virtual and physical alignment. */ KASSERT((va & PDRMASK) == 0, ("pmap_enter: va unaligned")); KASSERT(m->psind > 0, ("pmap_enter: m->psind < psind")); rv = pmap_enter_pde(pmap, va, newpte | PG_PS, flags, m); goto out; } pde = pmap_pde(pmap, va); if (pmap != kernel_pmap) { /* * va is for UVA. * In the case that a page table page is not resident, * we are creating it here. pmap_allocpte() handles * demotion. */ mpte = pmap_allocpte(pmap, va, flags); if (mpte == NULL) { KASSERT((flags & PMAP_ENTER_NOSLEEP) != 0, ("pmap_allocpte failed with sleep allowed")); rv = KERN_RESOURCE_SHORTAGE; goto out; } } else { /* * va is for KVA, so pmap_demote_pde() will never fail * to install a page table page. PG_V is also * asserted by pmap_demote_pde(). */ mpte = NULL; KASSERT(pde != NULL && (*pde & PG_V) != 0, ("KVA %#x invalid pde pdir %#jx", va, (uintmax_t)pmap->pm_pdir[PTDPTDI])); if ((*pde & PG_PS) != 0) pmap_demote_pde(pmap, pde, va); } pte = pmap_pte_quick(pmap, va); /* * Page Directory table entry is not valid, which should not * happen. We should have either allocated the page table * page or demoted the existing mapping above. */ if (pte == NULL) { panic("pmap_enter: invalid page directory pdir=%#jx, va=%#x", (uintmax_t)pmap->pm_pdir[PTDPTDI], va); } origpte = *pte; pv = NULL; /* * Is the specified virtual address already mapped? */ if ((origpte & PG_V) != 0) { /* * Wiring change, just update stats. We don't worry about * wiring PT pages as they remain resident as long as there * are valid mappings in them. Hence, if a user page is wired, * the PT page will be also. */ if ((newpte & PG_W) != 0 && (origpte & PG_W) == 0) pmap->pm_stats.wired_count++; else if ((newpte & PG_W) == 0 && (origpte & PG_W) != 0) pmap->pm_stats.wired_count--; /* * Remove the extra PT page reference. */ if (mpte != NULL) { mpte->ref_count--; KASSERT(mpte->ref_count > 0, ("pmap_enter: missing reference to page table page," " va: 0x%x", va)); } /* * Has the physical page changed? */ opa = origpte & PG_FRAME; if (opa == pa) { /* * No, might be a protection or wiring change. */ if ((origpte & PG_MANAGED) != 0 && (newpte & PG_RW) != 0) vm_page_aflag_set(m, PGA_WRITEABLE); if (((origpte ^ newpte) & ~(PG_M | PG_A)) == 0) goto unchanged; goto validate; } /* * The physical page has changed. Temporarily invalidate * the mapping. This ensures that all threads sharing the * pmap keep a consistent view of the mapping, which is * necessary for the correct handling of COW faults. It * also permits reuse of the old mapping's PV entry, * avoiding an allocation. * * For consistency, handle unmanaged mappings the same way. */ origpte = pte_load_clear(pte); KASSERT((origpte & PG_FRAME) == opa, ("pmap_enter: unexpected pa update for %#x", va)); if ((origpte & PG_MANAGED) != 0) { om = PHYS_TO_VM_PAGE(opa); /* * The pmap lock is sufficient to synchronize with * concurrent calls to pmap_page_test_mappings() and * pmap_ts_referenced(). */ if ((origpte & (PG_M | PG_RW)) == (PG_M | PG_RW)) vm_page_dirty(om); if ((origpte & PG_A) != 0) { pmap_invalidate_page_int(pmap, va); vm_page_aflag_set(om, PGA_REFERENCED); } pv = pmap_pvh_remove(&om->md, pmap, va); KASSERT(pv != NULL, ("pmap_enter: no PV entry for %#x", va)); if ((newpte & PG_MANAGED) == 0) free_pv_entry(pmap, pv); if ((om->a.flags & PGA_WRITEABLE) != 0 && TAILQ_EMPTY(&om->md.pv_list) && ((om->flags & PG_FICTITIOUS) != 0 || TAILQ_EMPTY(&pa_to_pvh(opa)->pv_list))) vm_page_aflag_clear(om, PGA_WRITEABLE); } else { /* * Since this mapping is unmanaged, assume that PG_A * is set. */ pmap_invalidate_page_int(pmap, va); } origpte = 0; } else { /* * Increment the counters. */ if ((newpte & PG_W) != 0) pmap->pm_stats.wired_count++; pmap->pm_stats.resident_count++; } /* * Enter on the PV list if part of our managed memory. */ if ((newpte & PG_MANAGED) != 0) { if (pv == NULL) { pv = get_pv_entry(pmap, FALSE); pv->pv_va = va; } TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next); if ((newpte & PG_RW) != 0) vm_page_aflag_set(m, PGA_WRITEABLE); } /* * Update the PTE. */ if ((origpte & PG_V) != 0) { validate: origpte = pte_load_store(pte, newpte); KASSERT((origpte & PG_FRAME) == pa, ("pmap_enter: unexpected pa update for %#x", va)); if ((newpte & PG_M) == 0 && (origpte & (PG_M | PG_RW)) == (PG_M | PG_RW)) { if ((origpte & PG_MANAGED) != 0) vm_page_dirty(m); /* * Although the PTE may still have PG_RW set, TLB * invalidation may nonetheless be required because * the PTE no longer has PG_M set. */ } #ifdef PMAP_PAE_COMP else if ((origpte & PG_NX) != 0 || (newpte & PG_NX) == 0) { /* * This PTE change does not require TLB invalidation. */ goto unchanged; } #endif if ((origpte & PG_A) != 0) pmap_invalidate_page_int(pmap, va); } else pte_store_zero(pte, newpte); unchanged: #if VM_NRESERVLEVEL > 0 /* * If both the page table page and the reservation are fully * populated, then attempt promotion. */ if ((mpte == NULL || mpte->ref_count == NPTEPG) && pg_ps_enabled && (m->flags & PG_FICTITIOUS) == 0 && vm_reserv_level_iffullpop(m) == 0) pmap_promote_pde(pmap, pde, va); #endif rv = KERN_SUCCESS; out: sched_unpin(); rw_wunlock(&pvh_global_lock); PMAP_UNLOCK(pmap); return (rv); } /* * Tries to create a read- and/or execute-only 2 or 4 MB page mapping. Returns * true if successful. Returns false if (1) a mapping already exists at the * specified virtual address or (2) a PV entry cannot be allocated without * reclaiming another PV entry. */ static bool pmap_enter_4mpage(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot) { pd_entry_t newpde; PMAP_LOCK_ASSERT(pmap, MA_OWNED); newpde = VM_PAGE_TO_PHYS(m) | pmap_cache_bits(pmap, m->md.pat_mode, 1) | PG_PS | PG_V; if ((m->oflags & VPO_UNMANAGED) == 0) newpde |= PG_MANAGED; #ifdef PMAP_PAE_COMP if ((prot & VM_PROT_EXECUTE) == 0 && !i386_read_exec) newpde |= pg_nx; #endif if (pmap != kernel_pmap) newpde |= PG_U; return (pmap_enter_pde(pmap, va, newpde, PMAP_ENTER_NOSLEEP | PMAP_ENTER_NOREPLACE | PMAP_ENTER_NORECLAIM, NULL) == KERN_SUCCESS); } /* * Returns true if every page table entry in the page table page that maps * the specified kernel virtual address is zero. */ static bool pmap_every_pte_zero(vm_offset_t va) { pt_entry_t *pt_end, *pte; KASSERT((va & PDRMASK) == 0, ("va is misaligned")); pte = vtopte(va); for (pt_end = pte + NPTEPG; pte < pt_end; pte++) { if (*pte != 0) return (false); } return (true); } /* * Tries to create the specified 2 or 4 MB page mapping. Returns KERN_SUCCESS * if the mapping was created, and either KERN_FAILURE or * KERN_RESOURCE_SHORTAGE otherwise. Returns KERN_FAILURE if * PMAP_ENTER_NOREPLACE was specified and a mapping already exists at the * specified virtual address. Returns KERN_RESOURCE_SHORTAGE if * PMAP_ENTER_NORECLAIM was specified and a PV entry allocation failed. * * The parameter "m" is only used when creating a managed, writeable mapping. */ static int pmap_enter_pde(pmap_t pmap, vm_offset_t va, pd_entry_t newpde, u_int flags, vm_page_t m) { struct spglist free; pd_entry_t oldpde, *pde; vm_page_t mt; rw_assert(&pvh_global_lock, RA_WLOCKED); KASSERT((newpde & (PG_M | PG_RW)) != PG_RW, ("pmap_enter_pde: newpde is missing PG_M")); KASSERT(pmap == kernel_pmap || (newpde & PG_W) == 0, ("pmap_enter_pde: cannot create wired user mapping")); PMAP_LOCK_ASSERT(pmap, MA_OWNED); pde = pmap_pde(pmap, va); oldpde = *pde; if ((oldpde & PG_V) != 0) { if ((flags & PMAP_ENTER_NOREPLACE) != 0 && (pmap != kernel_pmap || (oldpde & PG_PS) != 0 || !pmap_every_pte_zero(va))) { CTR2(KTR_PMAP, "pmap_enter_pde: failure for va %#lx" " in pmap %p", va, pmap); return (KERN_FAILURE); } /* Break the existing mapping(s). */ SLIST_INIT(&free); if ((oldpde & PG_PS) != 0) { /* * If the PDE resulted from a promotion, then a * reserved PT page could be freed. */ (void)pmap_remove_pde(pmap, pde, va, &free); if ((oldpde & PG_G) == 0) pmap_invalidate_pde_page(pmap, va, oldpde); } else { if (pmap_remove_ptes(pmap, va, va + NBPDR, &free)) pmap_invalidate_all_int(pmap); } if (pmap != kernel_pmap) { vm_page_free_pages_toq(&free, true); KASSERT(*pde == 0, ("pmap_enter_pde: non-zero pde %p", pde)); } else { KASSERT(SLIST_EMPTY(&free), ("pmap_enter_pde: freed kernel page table page")); /* * Both pmap_remove_pde() and pmap_remove_ptes() will * leave the kernel page table page zero filled. */ mt = PHYS_TO_VM_PAGE(*pde & PG_FRAME); if (pmap_insert_pt_page(pmap, mt, false)) panic("pmap_enter_pde: trie insert failed"); } } if ((newpde & PG_MANAGED) != 0) { /* * Abort this mapping if its PV entry could not be created. */ if (!pmap_pv_insert_pde(pmap, va, newpde, flags)) { CTR2(KTR_PMAP, "pmap_enter_pde: failure for va %#lx" " in pmap %p", va, pmap); return (KERN_RESOURCE_SHORTAGE); } if ((newpde & PG_RW) != 0) { for (mt = m; mt < &m[NBPDR / PAGE_SIZE]; mt++) vm_page_aflag_set(mt, PGA_WRITEABLE); } } /* * Increment counters. */ if ((newpde & PG_W) != 0) pmap->pm_stats.wired_count += NBPDR / PAGE_SIZE; pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE; /* * Map the superpage. (This is not a promoted mapping; there will not * be any lingering 4KB page mappings in the TLB.) */ pde_store(pde, newpde); pmap_pde_mappings++; CTR2(KTR_PMAP, "pmap_enter_pde: success for va %#lx in pmap %p", va, pmap); return (KERN_SUCCESS); } /* * Maps a sequence of resident pages belonging to the same object. * The sequence begins with the given page m_start. This page is * mapped at the given virtual address start. Each subsequent page is * mapped at a virtual address that is offset from start by the same * amount as the page is offset from m_start within the object. The * last page in the sequence is the page with the largest offset from * m_start that can be mapped at a virtual address less than the given * virtual address end. Not every virtual page between start and end * is mapped; only those for which a resident page exists with the * corresponding offset from m_start are mapped. */ static void __CONCAT(PMTYPE, enter_object)(pmap_t pmap, vm_offset_t start, vm_offset_t end, vm_page_t m_start, vm_prot_t prot) { vm_offset_t va; vm_page_t m, mpte; vm_pindex_t diff, psize; VM_OBJECT_ASSERT_LOCKED(m_start->object); psize = atop(end - start); mpte = NULL; m = m_start; rw_wlock(&pvh_global_lock); PMAP_LOCK(pmap); while (m != NULL && (diff = m->pindex - m_start->pindex) < psize) { va = start + ptoa(diff); if ((va & PDRMASK) == 0 && va + NBPDR <= end && m->psind == 1 && pg_ps_enabled && pmap_enter_4mpage(pmap, va, m, prot)) m = &m[NBPDR / PAGE_SIZE - 1]; else mpte = pmap_enter_quick_locked(pmap, va, m, prot, mpte); m = TAILQ_NEXT(m, listq); } rw_wunlock(&pvh_global_lock); PMAP_UNLOCK(pmap); } /* * this code makes some *MAJOR* assumptions: * 1. Current pmap & pmap exists. * 2. Not wired. * 3. Read access. * 4. No page table pages. * but is *MUCH* faster than pmap_enter... */ static void __CONCAT(PMTYPE, enter_quick)(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot) { rw_wlock(&pvh_global_lock); PMAP_LOCK(pmap); (void)pmap_enter_quick_locked(pmap, va, m, prot, NULL); rw_wunlock(&pvh_global_lock); PMAP_UNLOCK(pmap); } static vm_page_t pmap_enter_quick_locked(pmap_t pmap, vm_offset_t va, vm_page_t m, vm_prot_t prot, vm_page_t mpte) { pt_entry_t newpte, *pte; KASSERT(pmap != kernel_pmap || va < kmi.clean_sva || va >= kmi.clean_eva || (m->oflags & VPO_UNMANAGED) != 0, ("pmap_enter_quick_locked: managed mapping within the clean submap")); rw_assert(&pvh_global_lock, RA_WLOCKED); PMAP_LOCK_ASSERT(pmap, MA_OWNED); /* * In the case that a page table page is not * resident, we are creating it here. */ if (pmap != kernel_pmap) { u_int ptepindex; pd_entry_t ptepa; /* * Calculate pagetable page index */ ptepindex = va >> PDRSHIFT; if (mpte && (mpte->pindex == ptepindex)) { mpte->ref_count++; } else { /* * Get the page directory entry */ ptepa = pmap->pm_pdir[ptepindex]; /* * If the page table page is mapped, we just increment * the hold count, and activate it. */ if (ptepa) { if (ptepa & PG_PS) return (NULL); mpte = PHYS_TO_VM_PAGE(ptepa & PG_FRAME); mpte->ref_count++; } else { mpte = _pmap_allocpte(pmap, ptepindex, PMAP_ENTER_NOSLEEP); if (mpte == NULL) return (mpte); } } } else { mpte = NULL; } sched_pin(); pte = pmap_pte_quick(pmap, va); if (*pte) { if (mpte != NULL) mpte->ref_count--; sched_unpin(); return (NULL); } /* * Enter on the PV list if part of our managed memory. */ if ((m->oflags & VPO_UNMANAGED) == 0 && !pmap_try_insert_pv_entry(pmap, va, m)) { if (mpte != NULL) pmap_abort_ptp(pmap, va, mpte); sched_unpin(); return (NULL); } /* * Increment counters */ pmap->pm_stats.resident_count++; newpte = VM_PAGE_TO_PHYS(m) | PG_V | pmap_cache_bits(pmap, m->md.pat_mode, 0); if ((m->oflags & VPO_UNMANAGED) == 0) newpte |= PG_MANAGED; #ifdef PMAP_PAE_COMP if ((prot & VM_PROT_EXECUTE) == 0 && !i386_read_exec) newpte |= pg_nx; #endif if (pmap != kernel_pmap) newpte |= PG_U; pte_store_zero(pte, newpte); sched_unpin(); return (mpte); } /* * Make a temporary mapping for a physical address. This is only intended * to be used for panic dumps. */ static void * __CONCAT(PMTYPE, kenter_temporary)(vm_paddr_t pa, int i) { vm_offset_t va; va = (vm_offset_t)crashdumpmap + (i * PAGE_SIZE); pmap_kenter(va, pa); invlpg(va); return ((void *)crashdumpmap); } /* * This code maps large physical mmap regions into the * processor address space. Note that some shortcuts * are taken, but the code works. */ static void __CONCAT(PMTYPE, object_init_pt)(pmap_t pmap, vm_offset_t addr, vm_object_t object, vm_pindex_t pindex, vm_size_t size) { pd_entry_t *pde; vm_paddr_t pa, ptepa; vm_page_t p; int pat_mode; VM_OBJECT_ASSERT_WLOCKED(object); KASSERT(object->type == OBJT_DEVICE || object->type == OBJT_SG, ("pmap_object_init_pt: non-device object")); if (pg_ps_enabled && (addr & (NBPDR - 1)) == 0 && (size & (NBPDR - 1)) == 0) { if (!vm_object_populate(object, pindex, pindex + atop(size))) return; p = vm_page_lookup(object, pindex); KASSERT(p->valid == VM_PAGE_BITS_ALL, ("pmap_object_init_pt: invalid page %p", p)); pat_mode = p->md.pat_mode; /* * Abort the mapping if the first page is not physically * aligned to a 2/4MB page boundary. */ ptepa = VM_PAGE_TO_PHYS(p); if (ptepa & (NBPDR - 1)) return; /* * Skip the first page. Abort the mapping if the rest of * the pages are not physically contiguous or have differing * memory attributes. */ p = TAILQ_NEXT(p, listq); for (pa = ptepa + PAGE_SIZE; pa < ptepa + size; pa += PAGE_SIZE) { KASSERT(p->valid == VM_PAGE_BITS_ALL, ("pmap_object_init_pt: invalid page %p", p)); if (pa != VM_PAGE_TO_PHYS(p) || pat_mode != p->md.pat_mode) return; p = TAILQ_NEXT(p, listq); } /* * Map using 2/4MB pages. Since "ptepa" is 2/4M aligned and * "size" is a multiple of 2/4M, adding the PAT setting to * "pa" will not affect the termination of this loop. */ PMAP_LOCK(pmap); for (pa = ptepa | pmap_cache_bits(pmap, pat_mode, 1); pa < ptepa + size; pa += NBPDR) { pde = pmap_pde(pmap, addr); if (*pde == 0) { pde_store(pde, pa | PG_PS | PG_M | PG_A | PG_U | PG_RW | PG_V); pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE; pmap_pde_mappings++; } /* Else continue on if the PDE is already valid. */ addr += NBPDR; } PMAP_UNLOCK(pmap); } } /* * Clear the wired attribute from the mappings for the specified range of * addresses in the given pmap. Every valid mapping within that range * must have the wired attribute set. In contrast, invalid mappings * cannot have the wired attribute set, so they are ignored. * * The wired attribute of the page table entry is not a hardware feature, * so there is no need to invalidate any TLB entries. */ static void __CONCAT(PMTYPE, unwire)(pmap_t pmap, vm_offset_t sva, vm_offset_t eva) { vm_offset_t pdnxt; pd_entry_t *pde; pt_entry_t *pte; boolean_t pv_lists_locked; if (pmap_is_current(pmap)) pv_lists_locked = FALSE; else { pv_lists_locked = TRUE; resume: rw_wlock(&pvh_global_lock); sched_pin(); } PMAP_LOCK(pmap); for (; sva < eva; sva = pdnxt) { pdnxt = (sva + NBPDR) & ~PDRMASK; if (pdnxt < sva) pdnxt = eva; pde = pmap_pde(pmap, sva); if ((*pde & PG_V) == 0) continue; if ((*pde & PG_PS) != 0) { if ((*pde & PG_W) == 0) panic("pmap_unwire: pde %#jx is missing PG_W", (uintmax_t)*pde); /* * Are we unwiring the entire large page? If not, * demote the mapping and fall through. */ if (sva + NBPDR == pdnxt && eva >= pdnxt) { /* * Regardless of whether a pde (or pte) is 32 * or 64 bits in size, PG_W is among the least * significant 32 bits. */ atomic_clear_int((u_int *)pde, PG_W); pmap->pm_stats.wired_count -= NBPDR / PAGE_SIZE; continue; } else { if (!pv_lists_locked) { pv_lists_locked = TRUE; if (!rw_try_wlock(&pvh_global_lock)) { PMAP_UNLOCK(pmap); /* Repeat sva. */ goto resume; } sched_pin(); } if (!pmap_demote_pde(pmap, pde, sva)) panic("pmap_unwire: demotion failed"); } } if (pdnxt > eva) pdnxt = eva; for (pte = pmap_pte_quick(pmap, sva); sva != pdnxt; pte++, sva += PAGE_SIZE) { if ((*pte & PG_V) == 0) continue; if ((*pte & PG_W) == 0) panic("pmap_unwire: pte %#jx is missing PG_W", (uintmax_t)*pte); /* * PG_W must be cleared atomically. Although the pmap * lock synchronizes access to PG_W, another processor * could be setting PG_M and/or PG_A concurrently. * * PG_W is among the least significant 32 bits. */ atomic_clear_int((u_int *)pte, PG_W); pmap->pm_stats.wired_count--; } } if (pv_lists_locked) { sched_unpin(); rw_wunlock(&pvh_global_lock); } PMAP_UNLOCK(pmap); } /* * Copy the range specified by src_addr/len * from the source map to the range dst_addr/len * in the destination map. * * This routine is only advisory and need not do anything. Since * current pmap is always the kernel pmap when executing in * kernel, and we do not copy from the kernel pmap to a user * pmap, this optimization is not usable in 4/4G full split i386 * world. */ static void __CONCAT(PMTYPE, copy)(pmap_t dst_pmap, pmap_t src_pmap, vm_offset_t dst_addr, vm_size_t len, vm_offset_t src_addr) { pt_entry_t *src_pte, *dst_pte, ptetemp; pd_entry_t srcptepaddr; vm_page_t dstmpte, srcmpte; vm_offset_t addr, end_addr, pdnxt; u_int ptepindex; if (dst_addr != src_addr) return; end_addr = src_addr + len; rw_wlock(&pvh_global_lock); if (dst_pmap < src_pmap) { PMAP_LOCK(dst_pmap); PMAP_LOCK(src_pmap); } else { PMAP_LOCK(src_pmap); PMAP_LOCK(dst_pmap); } sched_pin(); for (addr = src_addr; addr < end_addr; addr = pdnxt) { KASSERT(addr < PMAP_TRM_MIN_ADDRESS, ("pmap_copy: invalid to pmap_copy the trampoline")); pdnxt = (addr + NBPDR) & ~PDRMASK; if (pdnxt < addr) pdnxt = end_addr; ptepindex = addr >> PDRSHIFT; srcptepaddr = src_pmap->pm_pdir[ptepindex]; if (srcptepaddr == 0) continue; if (srcptepaddr & PG_PS) { if ((addr & PDRMASK) != 0 || addr + NBPDR > end_addr) continue; if (dst_pmap->pm_pdir[ptepindex] == 0 && ((srcptepaddr & PG_MANAGED) == 0 || pmap_pv_insert_pde(dst_pmap, addr, srcptepaddr, PMAP_ENTER_NORECLAIM))) { dst_pmap->pm_pdir[ptepindex] = srcptepaddr & ~PG_W; dst_pmap->pm_stats.resident_count += NBPDR / PAGE_SIZE; pmap_pde_mappings++; } continue; } srcmpte = PHYS_TO_VM_PAGE(srcptepaddr & PG_FRAME); KASSERT(srcmpte->ref_count > 0, ("pmap_copy: source page table page is unused")); if (pdnxt > end_addr) pdnxt = end_addr; src_pte = pmap_pte_quick3(src_pmap, addr); while (addr < pdnxt) { ptetemp = *src_pte; /* * we only virtual copy managed pages */ if ((ptetemp & PG_MANAGED) != 0) { dstmpte = pmap_allocpte(dst_pmap, addr, PMAP_ENTER_NOSLEEP); if (dstmpte == NULL) goto out; dst_pte = pmap_pte_quick(dst_pmap, addr); if (*dst_pte == 0 && pmap_try_insert_pv_entry(dst_pmap, addr, PHYS_TO_VM_PAGE(ptetemp & PG_FRAME))) { /* * Clear the wired, modified, and * accessed (referenced) bits * during the copy. */ *dst_pte = ptetemp & ~(PG_W | PG_M | PG_A); dst_pmap->pm_stats.resident_count++; } else { pmap_abort_ptp(dst_pmap, addr, dstmpte); goto out; } if (dstmpte->ref_count >= srcmpte->ref_count) break; } addr += PAGE_SIZE; src_pte++; } } out: sched_unpin(); rw_wunlock(&pvh_global_lock); PMAP_UNLOCK(src_pmap); PMAP_UNLOCK(dst_pmap); } /* * Zero 1 page of virtual memory mapped from a hardware page by the caller. */ static __inline void pagezero(void *page) { #if defined(I686_CPU) if (cpu_class == CPUCLASS_686) { if (cpu_feature & CPUID_SSE2) sse2_pagezero(page); else i686_pagezero(page); } else #endif bzero(page, PAGE_SIZE); } /* * Zero the specified hardware page. */ static void __CONCAT(PMTYPE, zero_page)(vm_page_t m) { pt_entry_t *cmap_pte2; struct pcpu *pc; sched_pin(); pc = get_pcpu(); cmap_pte2 = pc->pc_cmap_pte2; mtx_lock(&pc->pc_cmap_lock); if (*cmap_pte2) panic("pmap_zero_page: CMAP2 busy"); *cmap_pte2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M | pmap_cache_bits(kernel_pmap, m->md.pat_mode, 0); invlcaddr(pc->pc_cmap_addr2); pagezero(pc->pc_cmap_addr2); *cmap_pte2 = 0; /* * Unpin the thread before releasing the lock. Otherwise the thread * could be rescheduled while still bound to the current CPU, only * to unpin itself immediately upon resuming execution. */ sched_unpin(); mtx_unlock(&pc->pc_cmap_lock); } /* * Zero an an area within a single hardware page. off and size must not * cover an area beyond a single hardware page. */ static void __CONCAT(PMTYPE, zero_page_area)(vm_page_t m, int off, int size) { pt_entry_t *cmap_pte2; struct pcpu *pc; sched_pin(); pc = get_pcpu(); cmap_pte2 = pc->pc_cmap_pte2; mtx_lock(&pc->pc_cmap_lock); if (*cmap_pte2) panic("pmap_zero_page_area: CMAP2 busy"); *cmap_pte2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M | pmap_cache_bits(kernel_pmap, m->md.pat_mode, 0); invlcaddr(pc->pc_cmap_addr2); if (off == 0 && size == PAGE_SIZE) pagezero(pc->pc_cmap_addr2); else bzero(pc->pc_cmap_addr2 + off, size); *cmap_pte2 = 0; sched_unpin(); mtx_unlock(&pc->pc_cmap_lock); } /* * Copy 1 specified hardware page to another. */ static void __CONCAT(PMTYPE, copy_page)(vm_page_t src, vm_page_t dst) { pt_entry_t *cmap_pte1, *cmap_pte2; struct pcpu *pc; sched_pin(); pc = get_pcpu(); cmap_pte1 = pc->pc_cmap_pte1; cmap_pte2 = pc->pc_cmap_pte2; mtx_lock(&pc->pc_cmap_lock); if (*cmap_pte1) panic("pmap_copy_page: CMAP1 busy"); if (*cmap_pte2) panic("pmap_copy_page: CMAP2 busy"); *cmap_pte1 = PG_V | VM_PAGE_TO_PHYS(src) | PG_A | pmap_cache_bits(kernel_pmap, src->md.pat_mode, 0); invlcaddr(pc->pc_cmap_addr1); *cmap_pte2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(dst) | PG_A | PG_M | pmap_cache_bits(kernel_pmap, dst->md.pat_mode, 0); invlcaddr(pc->pc_cmap_addr2); bcopy(pc->pc_cmap_addr1, pc->pc_cmap_addr2, PAGE_SIZE); *cmap_pte1 = 0; *cmap_pte2 = 0; sched_unpin(); mtx_unlock(&pc->pc_cmap_lock); } static void __CONCAT(PMTYPE, copy_pages)(vm_page_t ma[], vm_offset_t a_offset, vm_page_t mb[], vm_offset_t b_offset, int xfersize) { vm_page_t a_pg, b_pg; char *a_cp, *b_cp; vm_offset_t a_pg_offset, b_pg_offset; pt_entry_t *cmap_pte1, *cmap_pte2; struct pcpu *pc; int cnt; sched_pin(); pc = get_pcpu(); cmap_pte1 = pc->pc_cmap_pte1; cmap_pte2 = pc->pc_cmap_pte2; mtx_lock(&pc->pc_cmap_lock); if (*cmap_pte1 != 0) panic("pmap_copy_pages: CMAP1 busy"); if (*cmap_pte2 != 0) panic("pmap_copy_pages: CMAP2 busy"); while (xfersize > 0) { a_pg = ma[a_offset >> PAGE_SHIFT]; a_pg_offset = a_offset & PAGE_MASK; cnt = min(xfersize, PAGE_SIZE - a_pg_offset); b_pg = mb[b_offset >> PAGE_SHIFT]; b_pg_offset = b_offset & PAGE_MASK; cnt = min(cnt, PAGE_SIZE - b_pg_offset); *cmap_pte1 = PG_V | VM_PAGE_TO_PHYS(a_pg) | PG_A | pmap_cache_bits(kernel_pmap, a_pg->md.pat_mode, 0); invlcaddr(pc->pc_cmap_addr1); *cmap_pte2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(b_pg) | PG_A | PG_M | pmap_cache_bits(kernel_pmap, b_pg->md.pat_mode, 0); invlcaddr(pc->pc_cmap_addr2); a_cp = pc->pc_cmap_addr1 + a_pg_offset; b_cp = pc->pc_cmap_addr2 + b_pg_offset; bcopy(a_cp, b_cp, cnt); a_offset += cnt; b_offset += cnt; xfersize -= cnt; } *cmap_pte1 = 0; *cmap_pte2 = 0; sched_unpin(); mtx_unlock(&pc->pc_cmap_lock); } /* * Returns true if the pmap's pv is one of the first * 16 pvs linked to from this page. This count may * be changed upwards or downwards in the future; it * is only necessary that true be returned for a small * subset of pmaps for proper page aging. */ static boolean_t __CONCAT(PMTYPE, page_exists_quick)(pmap_t pmap, vm_page_t m) { struct md_page *pvh; pv_entry_t pv; int loops = 0; boolean_t rv; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_page_exists_quick: page %p is not managed", m)); rv = FALSE; rw_wlock(&pvh_global_lock); TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) { if (PV_PMAP(pv) == pmap) { rv = TRUE; break; } loops++; if (loops >= 16) break; } if (!rv && loops < 16 && (m->flags & PG_FICTITIOUS) == 0) { pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) { if (PV_PMAP(pv) == pmap) { rv = TRUE; break; } loops++; if (loops >= 16) break; } } rw_wunlock(&pvh_global_lock); return (rv); } /* * pmap_page_wired_mappings: * * Return the number of managed mappings to the given physical page * that are wired. */ static int __CONCAT(PMTYPE, page_wired_mappings)(vm_page_t m) { int count; count = 0; if ((m->oflags & VPO_UNMANAGED) != 0) return (count); rw_wlock(&pvh_global_lock); count = pmap_pvh_wired_mappings(&m->md, count); if ((m->flags & PG_FICTITIOUS) == 0) { count = pmap_pvh_wired_mappings(pa_to_pvh(VM_PAGE_TO_PHYS(m)), count); } rw_wunlock(&pvh_global_lock); return (count); } /* * pmap_pvh_wired_mappings: * * Return the updated number "count" of managed mappings that are wired. */ static int pmap_pvh_wired_mappings(struct md_page *pvh, int count) { pmap_t pmap; pt_entry_t *pte; pv_entry_t pv; rw_assert(&pvh_global_lock, RA_WLOCKED); sched_pin(); TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) { pmap = PV_PMAP(pv); PMAP_LOCK(pmap); pte = pmap_pte_quick(pmap, pv->pv_va); if ((*pte & PG_W) != 0) count++; PMAP_UNLOCK(pmap); } sched_unpin(); return (count); } /* * Returns TRUE if the given page is mapped individually or as part of * a 4mpage. Otherwise, returns FALSE. */ static boolean_t __CONCAT(PMTYPE, page_is_mapped)(vm_page_t m) { boolean_t rv; if ((m->oflags & VPO_UNMANAGED) != 0) return (FALSE); rw_wlock(&pvh_global_lock); rv = !TAILQ_EMPTY(&m->md.pv_list) || ((m->flags & PG_FICTITIOUS) == 0 && !TAILQ_EMPTY(&pa_to_pvh(VM_PAGE_TO_PHYS(m))->pv_list)); rw_wunlock(&pvh_global_lock); return (rv); } /* * Remove all pages from specified address space * this aids process exit speeds. Also, this code * is special cased for current process only, but * can have the more generic (and slightly slower) * mode enabled. This is much faster than pmap_remove * in the case of running down an entire address space. */ static void __CONCAT(PMTYPE, remove_pages)(pmap_t pmap) { pt_entry_t *pte, tpte; vm_page_t m, mpte, mt; pv_entry_t pv; struct md_page *pvh; struct pv_chunk *pc, *npc; struct spglist free; int field, idx; int32_t bit; uint32_t inuse, bitmask; int allfree; if (pmap != PCPU_GET(curpmap)) { printf("warning: pmap_remove_pages called with non-current pmap\n"); return; } SLIST_INIT(&free); rw_wlock(&pvh_global_lock); PMAP_LOCK(pmap); sched_pin(); TAILQ_FOREACH_SAFE(pc, &pmap->pm_pvchunk, pc_list, npc) { KASSERT(pc->pc_pmap == pmap, ("Wrong pmap %p %p", pmap, pc->pc_pmap)); allfree = 1; for (field = 0; field < _NPCM; field++) { inuse = ~pc->pc_map[field] & pc_freemask[field]; while (inuse != 0) { bit = bsfl(inuse); bitmask = 1UL << bit; idx = field * 32 + bit; pv = &pc->pc_pventry[idx]; inuse &= ~bitmask; pte = pmap_pde(pmap, pv->pv_va); tpte = *pte; if ((tpte & PG_PS) == 0) { pte = pmap_pte_quick(pmap, pv->pv_va); tpte = *pte & ~PG_PTE_PAT; } if (tpte == 0) { printf( "TPTE at %p IS ZERO @ VA %08x\n", pte, pv->pv_va); panic("bad pte"); } /* * We cannot remove wired pages from a process' mapping at this time */ if (tpte & PG_W) { allfree = 0; continue; } m = PHYS_TO_VM_PAGE(tpte & PG_FRAME); KASSERT(m->phys_addr == (tpte & PG_FRAME), ("vm_page_t %p phys_addr mismatch %016jx %016jx", m, (uintmax_t)m->phys_addr, (uintmax_t)tpte)); KASSERT((m->flags & PG_FICTITIOUS) != 0 || m < &vm_page_array[vm_page_array_size], ("pmap_remove_pages: bad tpte %#jx", (uintmax_t)tpte)); pte_clear(pte); /* * Update the vm_page_t clean/reference bits. */ if ((tpte & (PG_M | PG_RW)) == (PG_M | PG_RW)) { if ((tpte & PG_PS) != 0) { for (mt = m; mt < &m[NBPDR / PAGE_SIZE]; mt++) vm_page_dirty(mt); } else vm_page_dirty(m); } /* Mark free */ PV_STAT(pv_entry_frees++); PV_STAT(pv_entry_spare++); pv_entry_count--; pc->pc_map[field] |= bitmask; if ((tpte & PG_PS) != 0) { pmap->pm_stats.resident_count -= NBPDR / PAGE_SIZE; pvh = pa_to_pvh(tpte & PG_PS_FRAME); TAILQ_REMOVE(&pvh->pv_list, pv, pv_next); if (TAILQ_EMPTY(&pvh->pv_list)) { for (mt = m; mt < &m[NBPDR / PAGE_SIZE]; mt++) if (TAILQ_EMPTY(&mt->md.pv_list)) vm_page_aflag_clear(mt, PGA_WRITEABLE); } mpte = pmap_remove_pt_page(pmap, pv->pv_va); if (mpte != NULL) { KASSERT(mpte->valid == VM_PAGE_BITS_ALL, ("pmap_remove_pages: pte page not promoted")); pmap->pm_stats.resident_count--; KASSERT(mpte->ref_count == NPTEPG, ("pmap_remove_pages: pte page ref count error")); mpte->ref_count = 0; pmap_add_delayed_free_list(mpte, &free, FALSE); } } else { pmap->pm_stats.resident_count--; TAILQ_REMOVE(&m->md.pv_list, pv, pv_next); if (TAILQ_EMPTY(&m->md.pv_list) && (m->flags & PG_FICTITIOUS) == 0) { pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); if (TAILQ_EMPTY(&pvh->pv_list)) vm_page_aflag_clear(m, PGA_WRITEABLE); } pmap_unuse_pt(pmap, pv->pv_va, &free); } } } if (allfree) { TAILQ_REMOVE(&pmap->pm_pvchunk, pc, pc_list); free_pv_chunk(pc); } } sched_unpin(); pmap_invalidate_all_int(pmap); rw_wunlock(&pvh_global_lock); PMAP_UNLOCK(pmap); vm_page_free_pages_toq(&free, true); } /* * pmap_is_modified: * * Return whether or not the specified physical page was modified * in any physical maps. */ static boolean_t __CONCAT(PMTYPE, is_modified)(vm_page_t m) { boolean_t rv; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_is_modified: page %p is not managed", m)); /* * If the page is not busied then this check is racy. */ if (!pmap_page_is_write_mapped(m)) return (FALSE); rw_wlock(&pvh_global_lock); rv = pmap_is_modified_pvh(&m->md) || ((m->flags & PG_FICTITIOUS) == 0 && pmap_is_modified_pvh(pa_to_pvh(VM_PAGE_TO_PHYS(m)))); rw_wunlock(&pvh_global_lock); return (rv); } /* * Returns TRUE if any of the given mappings were used to modify * physical memory. Otherwise, returns FALSE. Both page and 2mpage * mappings are supported. */ static boolean_t pmap_is_modified_pvh(struct md_page *pvh) { pv_entry_t pv; pt_entry_t *pte; pmap_t pmap; boolean_t rv; rw_assert(&pvh_global_lock, RA_WLOCKED); rv = FALSE; sched_pin(); TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) { pmap = PV_PMAP(pv); PMAP_LOCK(pmap); pte = pmap_pte_quick(pmap, pv->pv_va); rv = (*pte & (PG_M | PG_RW)) == (PG_M | PG_RW); PMAP_UNLOCK(pmap); if (rv) break; } sched_unpin(); return (rv); } /* * pmap_is_prefaultable: * * Return whether or not the specified virtual address is elgible * for prefault. */ static boolean_t __CONCAT(PMTYPE, is_prefaultable)(pmap_t pmap, vm_offset_t addr) { pd_entry_t pde; boolean_t rv; rv = FALSE; PMAP_LOCK(pmap); pde = *pmap_pde(pmap, addr); if (pde != 0 && (pde & PG_PS) == 0) rv = pmap_pte_ufast(pmap, addr, pde) == 0; PMAP_UNLOCK(pmap); return (rv); } /* * pmap_is_referenced: * * Return whether or not the specified physical page was referenced * in any physical maps. */ static boolean_t __CONCAT(PMTYPE, is_referenced)(vm_page_t m) { boolean_t rv; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_is_referenced: page %p is not managed", m)); rw_wlock(&pvh_global_lock); rv = pmap_is_referenced_pvh(&m->md) || ((m->flags & PG_FICTITIOUS) == 0 && pmap_is_referenced_pvh(pa_to_pvh(VM_PAGE_TO_PHYS(m)))); rw_wunlock(&pvh_global_lock); return (rv); } /* * Returns TRUE if any of the given mappings were referenced and FALSE * otherwise. Both page and 4mpage mappings are supported. */ static boolean_t pmap_is_referenced_pvh(struct md_page *pvh) { pv_entry_t pv; pt_entry_t *pte; pmap_t pmap; boolean_t rv; rw_assert(&pvh_global_lock, RA_WLOCKED); rv = FALSE; sched_pin(); TAILQ_FOREACH(pv, &pvh->pv_list, pv_next) { pmap = PV_PMAP(pv); PMAP_LOCK(pmap); pte = pmap_pte_quick(pmap, pv->pv_va); rv = (*pte & (PG_A | PG_V)) == (PG_A | PG_V); PMAP_UNLOCK(pmap); if (rv) break; } sched_unpin(); return (rv); } /* * Clear the write and modified bits in each of the given page's mappings. */ static void __CONCAT(PMTYPE, remove_write)(vm_page_t m) { struct md_page *pvh; pv_entry_t next_pv, pv; pmap_t pmap; pd_entry_t *pde; pt_entry_t oldpte, *pte; vm_offset_t va; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_remove_write: page %p is not managed", m)); vm_page_assert_busied(m); if (!pmap_page_is_write_mapped(m)) return; rw_wlock(&pvh_global_lock); sched_pin(); if ((m->flags & PG_FICTITIOUS) != 0) goto small_mappings; pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); TAILQ_FOREACH_SAFE(pv, &pvh->pv_list, pv_next, next_pv) { va = pv->pv_va; pmap = PV_PMAP(pv); PMAP_LOCK(pmap); pde = pmap_pde(pmap, va); if ((*pde & PG_RW) != 0) (void)pmap_demote_pde(pmap, pde, va); PMAP_UNLOCK(pmap); } small_mappings: TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) { pmap = PV_PMAP(pv); PMAP_LOCK(pmap); pde = pmap_pde(pmap, pv->pv_va); KASSERT((*pde & PG_PS) == 0, ("pmap_clear_write: found" " a 4mpage in page %p's pv list", m)); pte = pmap_pte_quick(pmap, pv->pv_va); retry: oldpte = *pte; if ((oldpte & PG_RW) != 0) { /* * Regardless of whether a pte is 32 or 64 bits * in size, PG_RW and PG_M are among the least * significant 32 bits. */ if (!atomic_cmpset_int((u_int *)pte, oldpte, oldpte & ~(PG_RW | PG_M))) goto retry; if ((oldpte & PG_M) != 0) vm_page_dirty(m); pmap_invalidate_page_int(pmap, pv->pv_va); } PMAP_UNLOCK(pmap); } vm_page_aflag_clear(m, PGA_WRITEABLE); sched_unpin(); rw_wunlock(&pvh_global_lock); } /* * pmap_ts_referenced: * * Return a count of reference bits for a page, clearing those bits. * It is not necessary for every reference bit to be cleared, but it * is necessary that 0 only be returned when there are truly no * reference bits set. * * As an optimization, update the page's dirty field if a modified bit is * found while counting reference bits. This opportunistic update can be * performed at low cost and can eliminate the need for some future calls * to pmap_is_modified(). However, since this function stops after * finding PMAP_TS_REFERENCED_MAX reference bits, it may not detect some * dirty pages. Those dirty pages will only be detected by a future call * to pmap_is_modified(). */ static int __CONCAT(PMTYPE, ts_referenced)(vm_page_t m) { struct md_page *pvh; pv_entry_t pv, pvf; pmap_t pmap; pd_entry_t *pde; pt_entry_t *pte; vm_paddr_t pa; int rtval = 0; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_ts_referenced: page %p is not managed", m)); pa = VM_PAGE_TO_PHYS(m); pvh = pa_to_pvh(pa); rw_wlock(&pvh_global_lock); sched_pin(); if ((m->flags & PG_FICTITIOUS) != 0 || (pvf = TAILQ_FIRST(&pvh->pv_list)) == NULL) goto small_mappings; pv = pvf; do { pmap = PV_PMAP(pv); PMAP_LOCK(pmap); pde = pmap_pde(pmap, pv->pv_va); if ((*pde & (PG_M | PG_RW)) == (PG_M | PG_RW)) { /* * Although "*pde" is mapping a 2/4MB page, because * this function is called at a 4KB page granularity, * we only update the 4KB page under test. */ vm_page_dirty(m); } if ((*pde & PG_A) != 0) { /* * Since this reference bit is shared by either 1024 * or 512 4KB pages, it should not be cleared every * time it is tested. Apply a simple "hash" function * on the physical page number, the virtual superpage * number, and the pmap address to select one 4KB page * out of the 1024 or 512 on which testing the * reference bit will result in clearing that bit. * This function is designed to avoid the selection of * the same 4KB page for every 2- or 4MB page mapping. * * On demotion, a mapping that hasn't been referenced * is simply destroyed. To avoid the possibility of a * subsequent page fault on a demoted wired mapping, * always leave its reference bit set. Moreover, * since the superpage is wired, the current state of * its reference bit won't affect page replacement. */ if ((((pa >> PAGE_SHIFT) ^ (pv->pv_va >> PDRSHIFT) ^ (uintptr_t)pmap) & (NPTEPG - 1)) == 0 && (*pde & PG_W) == 0) { atomic_clear_int((u_int *)pde, PG_A); pmap_invalidate_page_int(pmap, pv->pv_va); } rtval++; } PMAP_UNLOCK(pmap); /* Rotate the PV list if it has more than one entry. */ if (TAILQ_NEXT(pv, pv_next) != NULL) { TAILQ_REMOVE(&pvh->pv_list, pv, pv_next); TAILQ_INSERT_TAIL(&pvh->pv_list, pv, pv_next); } if (rtval >= PMAP_TS_REFERENCED_MAX) goto out; } while ((pv = TAILQ_FIRST(&pvh->pv_list)) != pvf); small_mappings: if ((pvf = TAILQ_FIRST(&m->md.pv_list)) == NULL) goto out; pv = pvf; do { pmap = PV_PMAP(pv); PMAP_LOCK(pmap); pde = pmap_pde(pmap, pv->pv_va); KASSERT((*pde & PG_PS) == 0, ("pmap_ts_referenced: found a 4mpage in page %p's pv list", m)); pte = pmap_pte_quick(pmap, pv->pv_va); if ((*pte & (PG_M | PG_RW)) == (PG_M | PG_RW)) vm_page_dirty(m); if ((*pte & PG_A) != 0) { atomic_clear_int((u_int *)pte, PG_A); pmap_invalidate_page_int(pmap, pv->pv_va); rtval++; } PMAP_UNLOCK(pmap); /* Rotate the PV list if it has more than one entry. */ if (TAILQ_NEXT(pv, pv_next) != NULL) { TAILQ_REMOVE(&m->md.pv_list, pv, pv_next); TAILQ_INSERT_TAIL(&m->md.pv_list, pv, pv_next); } } while ((pv = TAILQ_FIRST(&m->md.pv_list)) != pvf && rtval < PMAP_TS_REFERENCED_MAX); out: sched_unpin(); rw_wunlock(&pvh_global_lock); return (rtval); } /* * Apply the given advice to the specified range of addresses within the * given pmap. Depending on the advice, clear the referenced and/or * modified flags in each mapping and set the mapped page's dirty field. */ static void __CONCAT(PMTYPE, advise)(pmap_t pmap, vm_offset_t sva, vm_offset_t eva, int advice) { pd_entry_t oldpde, *pde; pt_entry_t *pte; vm_offset_t va, pdnxt; vm_page_t m; bool anychanged, pv_lists_locked; if (advice != MADV_DONTNEED && advice != MADV_FREE) return; if (pmap_is_current(pmap)) pv_lists_locked = false; else { pv_lists_locked = true; resume: rw_wlock(&pvh_global_lock); sched_pin(); } anychanged = false; PMAP_LOCK(pmap); for (; sva < eva; sva = pdnxt) { pdnxt = (sva + NBPDR) & ~PDRMASK; if (pdnxt < sva) pdnxt = eva; pde = pmap_pde(pmap, sva); oldpde = *pde; if ((oldpde & PG_V) == 0) continue; else if ((oldpde & PG_PS) != 0) { if ((oldpde & PG_MANAGED) == 0) continue; if (!pv_lists_locked) { pv_lists_locked = true; if (!rw_try_wlock(&pvh_global_lock)) { if (anychanged) pmap_invalidate_all_int(pmap); PMAP_UNLOCK(pmap); goto resume; } sched_pin(); } if (!pmap_demote_pde(pmap, pde, sva)) { /* * The large page mapping was destroyed. */ continue; } /* * Unless the page mappings are wired, remove the * mapping to a single page so that a subsequent * access may repromote. Choosing the last page * within the address range [sva, min(pdnxt, eva)) * generally results in more repromotions. Since the * underlying page table page is fully populated, this * removal never frees a page table page. */ if ((oldpde & PG_W) == 0) { va = eva; if (va > pdnxt) va = pdnxt; va -= PAGE_SIZE; KASSERT(va >= sva, ("pmap_advise: no address gap")); pte = pmap_pte_quick(pmap, va); KASSERT((*pte & PG_V) != 0, ("pmap_advise: invalid PTE")); pmap_remove_pte(pmap, pte, va, NULL); anychanged = true; } } if (pdnxt > eva) pdnxt = eva; va = pdnxt; for (pte = pmap_pte_quick(pmap, sva); sva != pdnxt; pte++, sva += PAGE_SIZE) { if ((*pte & (PG_MANAGED | PG_V)) != (PG_MANAGED | PG_V)) goto maybe_invlrng; else if ((*pte & (PG_M | PG_RW)) == (PG_M | PG_RW)) { if (advice == MADV_DONTNEED) { /* * Future calls to pmap_is_modified() * can be avoided by making the page * dirty now. */ m = PHYS_TO_VM_PAGE(*pte & PG_FRAME); vm_page_dirty(m); } atomic_clear_int((u_int *)pte, PG_M | PG_A); } else if ((*pte & PG_A) != 0) atomic_clear_int((u_int *)pte, PG_A); else goto maybe_invlrng; if ((*pte & PG_G) != 0) { if (va == pdnxt) va = sva; } else anychanged = true; continue; maybe_invlrng: if (va != pdnxt) { pmap_invalidate_range_int(pmap, va, sva); va = pdnxt; } } if (va != pdnxt) pmap_invalidate_range_int(pmap, va, sva); } if (anychanged) pmap_invalidate_all_int(pmap); if (pv_lists_locked) { sched_unpin(); rw_wunlock(&pvh_global_lock); } PMAP_UNLOCK(pmap); } /* * Clear the modify bits on the specified physical page. */ static void __CONCAT(PMTYPE, clear_modify)(vm_page_t m) { struct md_page *pvh; pv_entry_t next_pv, pv; pmap_t pmap; pd_entry_t oldpde, *pde; pt_entry_t *pte; vm_offset_t va; KASSERT((m->oflags & VPO_UNMANAGED) == 0, ("pmap_clear_modify: page %p is not managed", m)); vm_page_assert_busied(m); if (!pmap_page_is_write_mapped(m)) return; rw_wlock(&pvh_global_lock); sched_pin(); if ((m->flags & PG_FICTITIOUS) != 0) goto small_mappings; pvh = pa_to_pvh(VM_PAGE_TO_PHYS(m)); TAILQ_FOREACH_SAFE(pv, &pvh->pv_list, pv_next, next_pv) { va = pv->pv_va; pmap = PV_PMAP(pv); PMAP_LOCK(pmap); pde = pmap_pde(pmap, va); oldpde = *pde; /* If oldpde has PG_RW set, then it also has PG_M set. */ if ((oldpde & PG_RW) != 0 && pmap_demote_pde(pmap, pde, va) && (oldpde & PG_W) == 0) { /* * Write protect the mapping to a single page so that * a subsequent write access may repromote. */ va += VM_PAGE_TO_PHYS(m) - (oldpde & PG_PS_FRAME); pte = pmap_pte_quick(pmap, va); /* * Regardless of whether a pte is 32 or 64 bits * in size, PG_RW and PG_M are among the least * significant 32 bits. */ atomic_clear_int((u_int *)pte, PG_M | PG_RW); vm_page_dirty(m); pmap_invalidate_page_int(pmap, va); } PMAP_UNLOCK(pmap); } small_mappings: TAILQ_FOREACH(pv, &m->md.pv_list, pv_next) { pmap = PV_PMAP(pv); PMAP_LOCK(pmap); pde = pmap_pde(pmap, pv->pv_va); KASSERT((*pde & PG_PS) == 0, ("pmap_clear_modify: found" " a 4mpage in page %p's pv list", m)); pte = pmap_pte_quick(pmap, pv->pv_va); if ((*pte & (PG_M | PG_RW)) == (PG_M | PG_RW)) { /* * Regardless of whether a pte is 32 or 64 bits * in size, PG_M is among the least significant * 32 bits. */ atomic_clear_int((u_int *)pte, PG_M); pmap_invalidate_page_int(pmap, pv->pv_va); } PMAP_UNLOCK(pmap); } sched_unpin(); rw_wunlock(&pvh_global_lock); } /* * Miscellaneous support routines follow */ /* Adjust the cache mode for a 4KB page mapped via a PTE. */ static __inline void pmap_pte_attr(pt_entry_t *pte, int cache_bits) { u_int opte, npte; /* * The cache mode bits are all in the low 32-bits of the * PTE, so we can just spin on updating the low 32-bits. */ do { opte = *(u_int *)pte; npte = opte & ~PG_PTE_CACHE; npte |= cache_bits; } while (npte != opte && !atomic_cmpset_int((u_int *)pte, opte, npte)); } /* Adjust the cache mode for a 2/4MB page mapped via a PDE. */ static __inline void pmap_pde_attr(pd_entry_t *pde, int cache_bits) { u_int opde, npde; /* * The cache mode bits are all in the low 32-bits of the * PDE, so we can just spin on updating the low 32-bits. */ do { opde = *(u_int *)pde; npde = opde & ~PG_PDE_CACHE; npde |= cache_bits; } while (npde != opde && !atomic_cmpset_int((u_int *)pde, opde, npde)); } /* * Map a set of physical memory pages into the kernel virtual * address space. Return a pointer to where it is mapped. This * routine is intended to be used for mapping device memory, * NOT real memory. */ static void * __CONCAT(PMTYPE, mapdev_attr)(vm_paddr_t pa, vm_size_t size, int mode, int flags) { struct pmap_preinit_mapping *ppim; vm_offset_t va, offset; vm_page_t m; vm_size_t tmpsize; int i; offset = pa & PAGE_MASK; size = round_page(offset + size); pa = pa & PG_FRAME; if (pa < PMAP_MAP_LOW && pa + size <= PMAP_MAP_LOW) { va = pa + PMAP_MAP_LOW; if ((flags & MAPDEV_SETATTR) == 0) return ((void *)(va + offset)); } else if (!pmap_initialized) { va = 0; for (i = 0; i < PMAP_PREINIT_MAPPING_COUNT; i++) { ppim = pmap_preinit_mapping + i; if (ppim->va == 0) { ppim->pa = pa; ppim->sz = size; ppim->mode = mode; ppim->va = virtual_avail; virtual_avail += size; va = ppim->va; break; } } if (va == 0) panic("%s: too many preinit mappings", __func__); } else { /* * If we have a preinit mapping, re-use it. */ for (i = 0; i < PMAP_PREINIT_MAPPING_COUNT; i++) { ppim = pmap_preinit_mapping + i; if (ppim->pa == pa && ppim->sz == size && (ppim->mode == mode || (flags & MAPDEV_SETATTR) == 0)) return ((void *)(ppim->va + offset)); } va = kva_alloc(size); if (va == 0) panic("%s: Couldn't allocate KVA", __func__); } for (tmpsize = 0; tmpsize < size; tmpsize += PAGE_SIZE) { if ((flags & MAPDEV_SETATTR) == 0 && pmap_initialized) { m = PHYS_TO_VM_PAGE(pa); if (m != NULL && VM_PAGE_TO_PHYS(m) == pa) { pmap_kenter_attr(va + tmpsize, pa + tmpsize, m->md.pat_mode); continue; } } pmap_kenter_attr(va + tmpsize, pa + tmpsize, mode); } pmap_invalidate_range_int(kernel_pmap, va, va + tmpsize); pmap_invalidate_cache_range(va, va + size); return ((void *)(va + offset)); } static void __CONCAT(PMTYPE, unmapdev)(vm_offset_t va, vm_size_t size) { struct pmap_preinit_mapping *ppim; vm_offset_t offset; int i; if (va >= PMAP_MAP_LOW && va <= KERNBASE && va + size <= KERNBASE) return; offset = va & PAGE_MASK; size = round_page(offset + size); va = trunc_page(va); for (i = 0; i < PMAP_PREINIT_MAPPING_COUNT; i++) { ppim = pmap_preinit_mapping + i; if (ppim->va == va && ppim->sz == size) { if (pmap_initialized) return; ppim->pa = 0; ppim->va = 0; ppim->sz = 0; ppim->mode = 0; if (va + size == virtual_avail) virtual_avail = va; return; } } if (pmap_initialized) { pmap_qremove(va, atop(size)); kva_free(va, size); } } /* * Sets the memory attribute for the specified page. */ static void __CONCAT(PMTYPE, page_set_memattr)(vm_page_t m, vm_memattr_t ma) { m->md.pat_mode = ma; if ((m->flags & PG_FICTITIOUS) != 0) return; /* * If "m" is a normal page, flush it from the cache. * See pmap_invalidate_cache_range(). * * First, try to find an existing mapping of the page by sf * buffer. sf_buf_invalidate_cache() modifies mapping and * flushes the cache. */ if (sf_buf_invalidate_cache(m)) return; /* * If page is not mapped by sf buffer, but CPU does not * support self snoop, map the page transient and do * invalidation. In the worst case, whole cache is flushed by * pmap_invalidate_cache_range(). */ if ((cpu_feature & CPUID_SS) == 0) pmap_flush_page(m); } static void __CONCAT(PMTYPE, flush_page)(vm_page_t m) { pt_entry_t *cmap_pte2; struct pcpu *pc; vm_offset_t sva, eva; bool useclflushopt; useclflushopt = (cpu_stdext_feature & CPUID_STDEXT_CLFLUSHOPT) != 0; if (useclflushopt || (cpu_feature & CPUID_CLFSH) != 0) { sched_pin(); pc = get_pcpu(); cmap_pte2 = pc->pc_cmap_pte2; mtx_lock(&pc->pc_cmap_lock); if (*cmap_pte2) panic("pmap_flush_page: CMAP2 busy"); *cmap_pte2 = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M | pmap_cache_bits(kernel_pmap, m->md.pat_mode, 0); invlcaddr(pc->pc_cmap_addr2); sva = (vm_offset_t)pc->pc_cmap_addr2; eva = sva + PAGE_SIZE; /* * Use mfence or sfence despite the ordering implied by * mtx_{un,}lock() because clflush on non-Intel CPUs * and clflushopt are not guaranteed to be ordered by * any other instruction. */ if (useclflushopt) sfence(); else if (cpu_vendor_id != CPU_VENDOR_INTEL) mfence(); for (; sva < eva; sva += cpu_clflush_line_size) { if (useclflushopt) clflushopt(sva); else clflush(sva); } if (useclflushopt) sfence(); else if (cpu_vendor_id != CPU_VENDOR_INTEL) mfence(); *cmap_pte2 = 0; sched_unpin(); mtx_unlock(&pc->pc_cmap_lock); } else pmap_invalidate_cache(); } /* * Changes the specified virtual address range's memory type to that given by * the parameter "mode". The specified virtual address range must be * completely contained within either the kernel map. * * Returns zero if the change completed successfully, and either EINVAL or * ENOMEM if the change failed. Specifically, EINVAL is returned if some part * of the virtual address range was not mapped, and ENOMEM is returned if * there was insufficient memory available to complete the change. */ static int __CONCAT(PMTYPE, change_attr)(vm_offset_t va, vm_size_t size, int mode) { vm_offset_t base, offset, tmpva; pd_entry_t *pde; pt_entry_t *pte; int cache_bits_pte, cache_bits_pde; boolean_t changed; base = trunc_page(va); offset = va & PAGE_MASK; size = round_page(offset + size); /* * Only supported on kernel virtual addresses above the recursive map. */ if (base < VM_MIN_KERNEL_ADDRESS) return (EINVAL); cache_bits_pde = pmap_cache_bits(kernel_pmap, mode, 1); cache_bits_pte = pmap_cache_bits(kernel_pmap, mode, 0); changed = FALSE; /* * Pages that aren't mapped aren't supported. Also break down * 2/4MB pages into 4KB pages if required. */ PMAP_LOCK(kernel_pmap); for (tmpva = base; tmpva < base + size; ) { pde = pmap_pde(kernel_pmap, tmpva); if (*pde == 0) { PMAP_UNLOCK(kernel_pmap); return (EINVAL); } if (*pde & PG_PS) { /* * If the current 2/4MB page already has * the required memory type, then we need not * demote this page. Just increment tmpva to * the next 2/4MB page frame. */ if ((*pde & PG_PDE_CACHE) == cache_bits_pde) { tmpva = trunc_4mpage(tmpva) + NBPDR; continue; } /* * If the current offset aligns with a 2/4MB * page frame and there is at least 2/4MB left * within the range, then we need not break * down this page into 4KB pages. */ if ((tmpva & PDRMASK) == 0 && tmpva + PDRMASK < base + size) { tmpva += NBPDR; continue; } if (!pmap_demote_pde(kernel_pmap, pde, tmpva)) { PMAP_UNLOCK(kernel_pmap); return (ENOMEM); } } pte = vtopte(tmpva); if (*pte == 0) { PMAP_UNLOCK(kernel_pmap); return (EINVAL); } tmpva += PAGE_SIZE; } PMAP_UNLOCK(kernel_pmap); /* * Ok, all the pages exist, so run through them updating their * cache mode if required. */ for (tmpva = base; tmpva < base + size; ) { pde = pmap_pde(kernel_pmap, tmpva); if (*pde & PG_PS) { if ((*pde & PG_PDE_CACHE) != cache_bits_pde) { pmap_pde_attr(pde, cache_bits_pde); changed = TRUE; } tmpva = trunc_4mpage(tmpva) + NBPDR; } else { pte = vtopte(tmpva); if ((*pte & PG_PTE_CACHE) != cache_bits_pte) { pmap_pte_attr(pte, cache_bits_pte); changed = TRUE; } tmpva += PAGE_SIZE; } } /* * Flush CPU caches to make sure any data isn't cached that * shouldn't be, etc. */ if (changed) { pmap_invalidate_range_int(kernel_pmap, base, tmpva); pmap_invalidate_cache_range(base, tmpva); } return (0); } /* * Perform the pmap work for mincore(2). If the page is not both referenced and * modified by this pmap, returns its physical address so that the caller can * find other mappings. */ static int __CONCAT(PMTYPE, mincore)(pmap_t pmap, vm_offset_t addr, vm_paddr_t *pap) { pd_entry_t pde; pt_entry_t pte; vm_paddr_t pa; int val; PMAP_LOCK(pmap); pde = *pmap_pde(pmap, addr); if (pde != 0) { if ((pde & PG_PS) != 0) { pte = pde; /* Compute the physical address of the 4KB page. */ pa = ((pde & PG_PS_FRAME) | (addr & PDRMASK)) & PG_FRAME; val = MINCORE_PSIND(1); } else { pte = pmap_pte_ufast(pmap, addr, pde); pa = pte & PG_FRAME; val = 0; } } else { pte = 0; pa = 0; val = 0; } if ((pte & PG_V) != 0) { val |= MINCORE_INCORE; if ((pte & (PG_M | PG_RW)) == (PG_M | PG_RW)) val |= MINCORE_MODIFIED | MINCORE_MODIFIED_OTHER; if ((pte & PG_A) != 0) val |= MINCORE_REFERENCED | MINCORE_REFERENCED_OTHER; } if ((val & (MINCORE_MODIFIED_OTHER | MINCORE_REFERENCED_OTHER)) != (MINCORE_MODIFIED_OTHER | MINCORE_REFERENCED_OTHER) && (pte & (PG_MANAGED | PG_V)) == (PG_MANAGED | PG_V)) { *pap = pa; } PMAP_UNLOCK(pmap); return (val); } static void __CONCAT(PMTYPE, activate)(struct thread *td) { pmap_t pmap, oldpmap; u_int cpuid; u_int32_t cr3; critical_enter(); pmap = vmspace_pmap(td->td_proc->p_vmspace); oldpmap = PCPU_GET(curpmap); cpuid = PCPU_GET(cpuid); #if defined(SMP) CPU_CLR_ATOMIC(cpuid, &oldpmap->pm_active); CPU_SET_ATOMIC(cpuid, &pmap->pm_active); #else CPU_CLR(cpuid, &oldpmap->pm_active); CPU_SET(cpuid, &pmap->pm_active); #endif #ifdef PMAP_PAE_COMP cr3 = vtophys(pmap->pm_pdpt); #else cr3 = vtophys(pmap->pm_pdir); #endif /* * pmap_activate is for the current thread on the current cpu */ td->td_pcb->pcb_cr3 = cr3; PCPU_SET(curpmap, pmap); critical_exit(); } static void __CONCAT(PMTYPE, activate_boot)(pmap_t pmap) { u_int cpuid; cpuid = PCPU_GET(cpuid); #if defined(SMP) CPU_SET_ATOMIC(cpuid, &pmap->pm_active); #else CPU_SET(cpuid, &pmap->pm_active); #endif PCPU_SET(curpmap, pmap); } /* * Increase the starting virtual address of the given mapping if a * different alignment might result in more superpage mappings. */ static void __CONCAT(PMTYPE, align_superpage)(vm_object_t object, vm_ooffset_t offset, vm_offset_t *addr, vm_size_t size) { vm_offset_t superpage_offset; if (size < NBPDR) return; if (object != NULL && (object->flags & OBJ_COLORED) != 0) offset += ptoa(object->pg_color); superpage_offset = offset & PDRMASK; if (size - ((NBPDR - superpage_offset) & PDRMASK) < NBPDR || (*addr & PDRMASK) == superpage_offset) return; if ((*addr & PDRMASK) < superpage_offset) *addr = (*addr & ~PDRMASK) + superpage_offset; else *addr = ((*addr + PDRMASK) & ~PDRMASK) + superpage_offset; } static vm_offset_t __CONCAT(PMTYPE, quick_enter_page)(vm_page_t m) { vm_offset_t qaddr; pt_entry_t *pte; critical_enter(); qaddr = PCPU_GET(qmap_addr); pte = vtopte(qaddr); KASSERT(*pte == 0, ("pmap_quick_enter_page: PTE busy %#jx", (uintmax_t)*pte)); *pte = PG_V | PG_RW | VM_PAGE_TO_PHYS(m) | PG_A | PG_M | pmap_cache_bits(kernel_pmap, pmap_page_get_memattr(m), 0); invlpg(qaddr); return (qaddr); } static void __CONCAT(PMTYPE, quick_remove_page)(vm_offset_t addr) { vm_offset_t qaddr; pt_entry_t *pte; qaddr = PCPU_GET(qmap_addr); pte = vtopte(qaddr); KASSERT(*pte != 0, ("pmap_quick_remove_page: PTE not in use")); KASSERT(addr == qaddr, ("pmap_quick_remove_page: invalid address")); *pte = 0; critical_exit(); } static vmem_t *pmap_trm_arena; static vmem_addr_t pmap_trm_arena_last = PMAP_TRM_MIN_ADDRESS; static int trm_guard = PAGE_SIZE; static int pmap_trm_import(void *unused __unused, vmem_size_t size, int flags, vmem_addr_t *addrp) { vm_page_t m; vmem_addr_t af, addr, prev_addr; pt_entry_t *trm_pte; prev_addr = atomic_load_long(&pmap_trm_arena_last); size = round_page(size) + trm_guard; for (;;) { if (prev_addr + size < prev_addr || prev_addr + size < size || prev_addr + size > PMAP_TRM_MAX_ADDRESS) return (ENOMEM); addr = prev_addr + size; if (atomic_fcmpset_int(&pmap_trm_arena_last, &prev_addr, addr)) break; } prev_addr += trm_guard; trm_pte = PTmap + atop(prev_addr); for (af = prev_addr; af < addr; af += PAGE_SIZE) { m = vm_page_alloc_noobj(VM_ALLOC_WIRED | VM_ALLOC_WAITOK); pte_store(&trm_pte[atop(af - prev_addr)], VM_PAGE_TO_PHYS(m) | PG_M | PG_A | PG_RW | PG_V | pgeflag | pmap_cache_bits(kernel_pmap, VM_MEMATTR_DEFAULT, FALSE)); } *addrp = prev_addr; return (0); } void pmap_init_trm(void) { vm_page_t pd_m; TUNABLE_INT_FETCH("machdep.trm_guard", &trm_guard); if ((trm_guard & PAGE_MASK) != 0) trm_guard = 0; pmap_trm_arena = vmem_create("i386trampoline", 0, 0, 1, 0, M_WAITOK); vmem_set_import(pmap_trm_arena, pmap_trm_import, NULL, NULL, PAGE_SIZE); pd_m = vm_page_alloc_noobj(VM_ALLOC_WIRED | VM_ALLOC_WAITOK | VM_ALLOC_ZERO); PTD[TRPTDI] = VM_PAGE_TO_PHYS(pd_m) | PG_M | PG_A | PG_RW | PG_V | pmap_cache_bits(kernel_pmap, VM_MEMATTR_DEFAULT, TRUE); } static void * __CONCAT(PMTYPE, trm_alloc)(size_t size, int flags) { vmem_addr_t res; int error; MPASS((flags & ~(M_WAITOK | M_NOWAIT | M_ZERO)) == 0); error = vmem_xalloc(pmap_trm_arena, roundup2(size, 4), sizeof(int), 0, 0, VMEM_ADDR_MIN, VMEM_ADDR_MAX, flags | M_FIRSTFIT, &res); if (error != 0) return (NULL); if ((flags & M_ZERO) != 0) bzero((void *)res, size); return ((void *)res); } static void __CONCAT(PMTYPE, trm_free)(void *addr, size_t size) { vmem_free(pmap_trm_arena, (uintptr_t)addr, roundup2(size, 4)); } static void __CONCAT(PMTYPE, ksetrw)(vm_offset_t va) { *vtopte(va) |= PG_RW; } static void __CONCAT(PMTYPE, remap_lowptdi)(bool enable) { PTD[KPTDI] = enable ? PTD[LOWPTDI] : 0; invltlb_glob(); } static vm_offset_t __CONCAT(PMTYPE, get_map_low)(void) { return (PMAP_MAP_LOW); } static vm_offset_t __CONCAT(PMTYPE, get_vm_maxuser_address)(void) { return (VM_MAXUSER_ADDRESS); } static vm_paddr_t __CONCAT(PMTYPE, pg_frame)(vm_paddr_t pa) { return (pa & PG_FRAME); } static void __CONCAT(PMTYPE, sf_buf_map)(struct sf_buf *sf) { pt_entry_t opte, *ptep; /* * Update the sf_buf's virtual-to-physical mapping, flushing the * virtual address from the TLB. Since the reference count for * the sf_buf's old mapping was zero, that mapping is not * currently in use. Consequently, there is no need to exchange * the old and new PTEs atomically, even under PAE. */ ptep = vtopte(sf->kva); opte = *ptep; *ptep = VM_PAGE_TO_PHYS(sf->m) | PG_RW | PG_V | pmap_cache_bits(kernel_pmap, sf->m->md.pat_mode, 0); /* * Avoid unnecessary TLB invalidations: If the sf_buf's old * virtual-to-physical mapping was not used, then any processor * that has invalidated the sf_buf's virtual address from its TLB * since the last used mapping need not invalidate again. */ #ifdef SMP if ((opte & (PG_V | PG_A)) == (PG_V | PG_A)) CPU_ZERO(&sf->cpumask); #else if ((opte & (PG_V | PG_A)) == (PG_V | PG_A)) pmap_invalidate_page_int(kernel_pmap, sf->kva); #endif } static void __CONCAT(PMTYPE, cp_slow0_map)(vm_offset_t kaddr, int plen, vm_page_t *ma) { pt_entry_t *pte; int i; for (i = 0, pte = vtopte(kaddr); i < plen; i++, pte++) { *pte = PG_V | PG_RW | PG_A | PG_M | VM_PAGE_TO_PHYS(ma[i]) | pmap_cache_bits(kernel_pmap, pmap_page_get_memattr(ma[i]), FALSE); invlpg(kaddr + ptoa(i)); } } static u_int __CONCAT(PMTYPE, get_kcr3)(void) { #ifdef PMAP_PAE_COMP return ((u_int)IdlePDPT); #else return ((u_int)IdlePTD); #endif } static u_int __CONCAT(PMTYPE, get_cr3)(pmap_t pmap) { #ifdef PMAP_PAE_COMP return ((u_int)vtophys(pmap->pm_pdpt)); #else return ((u_int)vtophys(pmap->pm_pdir)); #endif } static caddr_t __CONCAT(PMTYPE, cmap3)(vm_paddr_t pa, u_int pte_bits) { pt_entry_t *pte; pte = CMAP3; *pte = pa | pte_bits; invltlb(); return (CADDR3); } static void __CONCAT(PMTYPE, basemem_setup)(u_int basemem) { pt_entry_t *pte; int i; /* * Map pages between basemem and ISA_HOLE_START, if any, r/w into * the vm86 page table so that vm86 can scribble on them using * the vm86 map too. XXX: why 2 ways for this and only 1 way for * page 0, at least as initialized here? */ pte = (pt_entry_t *)vm86paddr; for (i = basemem / 4; i < 160; i++) pte[i] = (i << PAGE_SHIFT) | PG_V | PG_RW | PG_U; } struct bios16_pmap_handle { pt_entry_t *pte; pd_entry_t *ptd; pt_entry_t orig_ptd; }; static void * __CONCAT(PMTYPE, bios16_enter)(void) { struct bios16_pmap_handle *h; /* * no page table, so create one and install it. */ h = malloc(sizeof(struct bios16_pmap_handle), M_TEMP, M_WAITOK); h->pte = (pt_entry_t *)malloc(PAGE_SIZE, M_TEMP, M_WAITOK); h->ptd = IdlePTD; *h->pte = vm86phystk | PG_RW | PG_V; h->orig_ptd = *h->ptd; *h->ptd = vtophys(h->pte) | PG_RW | PG_V; pmap_invalidate_all_int(kernel_pmap); /* XXX insurance for now */ return (h); } static void __CONCAT(PMTYPE, bios16_leave)(void *arg) { struct bios16_pmap_handle *h; h = arg; *h->ptd = h->orig_ptd; /* remove page table */ /* * XXX only needs to be invlpg(0) but that doesn't work on the 386 */ pmap_invalidate_all_int(kernel_pmap); free(h->pte, M_TEMP); /* ... and free it */ } struct pmap_kernel_map_range { vm_offset_t sva; pt_entry_t attrs; int ptes; int pdes; int pdpes; }; static void sysctl_kmaps_dump(struct sbuf *sb, struct pmap_kernel_map_range *range, vm_offset_t eva) { const char *mode; int i, pat_idx; if (eva <= range->sva) return; pat_idx = pmap_pat_index(kernel_pmap, range->attrs, true); for (i = 0; i < PAT_INDEX_SIZE; i++) if (pat_index[i] == pat_idx) break; switch (i) { case PAT_WRITE_BACK: mode = "WB"; break; case PAT_WRITE_THROUGH: mode = "WT"; break; case PAT_UNCACHEABLE: mode = "UC"; break; case PAT_UNCACHED: mode = "U-"; break; case PAT_WRITE_PROTECTED: mode = "WP"; break; case PAT_WRITE_COMBINING: mode = "WC"; break; default: printf("%s: unknown PAT mode %#x for range 0x%08x-0x%08x\n", __func__, pat_idx, range->sva, eva); mode = "??"; break; } sbuf_printf(sb, "0x%08x-0x%08x r%c%c%c%c %s %d %d %d\n", range->sva, eva, (range->attrs & PG_RW) != 0 ? 'w' : '-', (range->attrs & pg_nx) != 0 ? '-' : 'x', (range->attrs & PG_U) != 0 ? 'u' : 's', (range->attrs & PG_G) != 0 ? 'g' : '-', mode, range->pdpes, range->pdes, range->ptes); /* Reset to sentinel value. */ range->sva = 0xffffffff; } /* * Determine whether the attributes specified by a page table entry match those * being tracked by the current range. This is not quite as simple as a direct * flag comparison since some PAT modes have multiple representations. */ static bool sysctl_kmaps_match(struct pmap_kernel_map_range *range, pt_entry_t attrs) { pt_entry_t diff, mask; mask = pg_nx | PG_G | PG_RW | PG_U | PG_PDE_CACHE; diff = (range->attrs ^ attrs) & mask; if (diff == 0) return (true); if ((diff & ~PG_PDE_PAT) == 0 && pmap_pat_index(kernel_pmap, range->attrs, true) == pmap_pat_index(kernel_pmap, attrs, true)) return (true); return (false); } static void sysctl_kmaps_reinit(struct pmap_kernel_map_range *range, vm_offset_t va, pt_entry_t attrs) { memset(range, 0, sizeof(*range)); range->sva = va; range->attrs = attrs; } /* * Given a leaf PTE, derive the mapping's attributes. If they do not match * those of the current run, dump the address range and its attributes, and * begin a new run. */ static void sysctl_kmaps_check(struct sbuf *sb, struct pmap_kernel_map_range *range, vm_offset_t va, pd_entry_t pde, pt_entry_t pte) { pt_entry_t attrs; attrs = pde & (PG_RW | PG_U | pg_nx); if ((pde & PG_PS) != 0) { attrs |= pde & (PG_G | PG_PDE_CACHE); } else if (pte != 0) { attrs |= pte & pg_nx; attrs &= pg_nx | (pte & (PG_RW | PG_U)); attrs |= pte & (PG_G | PG_PTE_CACHE); /* Canonicalize by always using the PDE PAT bit. */ if ((attrs & PG_PTE_PAT) != 0) attrs ^= PG_PDE_PAT | PG_PTE_PAT; } if (range->sva > va || !sysctl_kmaps_match(range, attrs)) { sysctl_kmaps_dump(sb, range, va); sysctl_kmaps_reinit(range, va, attrs); } } static int __CONCAT(PMTYPE, sysctl_kmaps)(SYSCTL_HANDLER_ARGS) { struct pmap_kernel_map_range range; struct sbuf sbuf, *sb; pd_entry_t pde; pt_entry_t *pt, pte; vm_offset_t sva; int error; u_int i, k; error = sysctl_wire_old_buffer(req, 0); if (error != 0) return (error); sb = &sbuf; sbuf_new_for_sysctl(sb, NULL, PAGE_SIZE, req); /* Sentinel value. */ range.sva = 0xffffffff; /* * Iterate over the kernel page tables without holding the * kernel pmap lock. Kernel page table pages are never freed, * so at worst we will observe inconsistencies in the output. */ for (sva = 0, i = 0; i < NPTEPG * NPGPTD * NPDEPG ;) { if (i == 0) sbuf_printf(sb, "\nLow PDE:\n"); else if (i == LOWPTDI * NPTEPG) sbuf_printf(sb, "Low PDE dup:\n"); else if (i == PTDPTDI * NPTEPG) sbuf_printf(sb, "Recursive map:\n"); else if (i == KERNPTDI * NPTEPG) sbuf_printf(sb, "Kernel base:\n"); else if (i == TRPTDI * NPTEPG) sbuf_printf(sb, "Trampoline:\n"); pde = IdlePTD[sva >> PDRSHIFT]; if ((pde & PG_V) == 0) { sva = rounddown2(sva, NBPDR); sysctl_kmaps_dump(sb, &range, sva); sva += NBPDR; i += NPTEPG; continue; } if ((pde & PG_PS) != 0) { sysctl_kmaps_check(sb, &range, sva, pde, 0); range.pdes++; sva += NBPDR; i += NPTEPG; continue; } for (pt = vtopte(sva), k = 0; k < NPTEPG; i++, k++, pt++, sva += PAGE_SIZE) { pte = *pt; if ((pte & PG_V) == 0) { sysctl_kmaps_dump(sb, &range, sva); continue; } sysctl_kmaps_check(sb, &range, sva, pde, pte); range.ptes++; } } error = sbuf_finish(sb); sbuf_delete(sb); return (error); } #define PMM(a) \ .pm_##a = __CONCAT(PMTYPE, a), struct pmap_methods __CONCAT(PMTYPE, methods) = { PMM(ksetrw) PMM(remap_lower) PMM(remap_lowptdi) PMM(align_superpage) PMM(quick_enter_page) PMM(quick_remove_page) PMM(trm_alloc) PMM(trm_free) PMM(get_map_low) PMM(get_vm_maxuser_address) PMM(kextract) PMM(pg_frame) PMM(sf_buf_map) PMM(cp_slow0_map) PMM(get_kcr3) PMM(get_cr3) PMM(cmap3) PMM(basemem_setup) PMM(set_nx) PMM(bios16_enter) PMM(bios16_leave) PMM(bootstrap) PMM(is_valid_memattr) PMM(cache_bits) PMM(ps_enabled) PMM(pinit0) PMM(pinit) PMM(activate) PMM(activate_boot) PMM(advise) PMM(clear_modify) PMM(change_attr) PMM(mincore) PMM(copy) PMM(copy_page) PMM(copy_pages) PMM(zero_page) PMM(zero_page_area) PMM(enter) PMM(enter_object) PMM(enter_quick) PMM(kenter_temporary) PMM(object_init_pt) PMM(unwire) PMM(page_exists_quick) PMM(page_wired_mappings) PMM(page_is_mapped) PMM(remove_pages) PMM(is_modified) PMM(is_prefaultable) PMM(is_referenced) PMM(remove_write) PMM(ts_referenced) PMM(mapdev_attr) PMM(unmapdev) PMM(page_set_memattr) PMM(extract) PMM(extract_and_hold) PMM(map) PMM(qenter) PMM(qremove) PMM(release) PMM(remove) PMM(protect) PMM(remove_all) PMM(init) PMM(init_pat) PMM(growkernel) PMM(invalidate_page) PMM(invalidate_range) PMM(invalidate_all) PMM(invalidate_cache) PMM(flush_page) PMM(kenter) PMM(kremove) PMM(sysctl_kmaps) }; diff --git a/sys/i386/i386/vm_machdep.c b/sys/i386/i386/vm_machdep.c index 57377575b9bf..0ca6a63eba76 100644 --- a/sys/i386/i386/vm_machdep.c +++ b/sys/i386/i386/vm_machdep.c @@ -1,678 +1,678 @@ /*- * SPDX-License-Identifier: BSD-4-Clause * * Copyright (c) 1982, 1986 The Regents of the University of California. * Copyright (c) 1989, 1990 William Jolitz * Copyright (c) 1994 John Dyson * All rights reserved. * * This code is derived from software contributed to Berkeley by * the Systems Programming Group of the University of Utah Computer * Science Department, and William Jolitz. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. All advertising materials mentioning features or use of this software * must display the following acknowledgement: * This product includes software developed by the University of * California, Berkeley and its contributors. * 4. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)vm_machdep.c 7.3 (Berkeley) 5/13/91 * Utah $Hdr: vm_machdep.c 1.16.1.1 89/06/23$ */ #include __FBSDID("$FreeBSD$"); #include "opt_isa.h" #include "opt_npx.h" #include "opt_reset.h" #include "opt_cpu.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include _Static_assert(__OFFSETOF_MONITORBUF == offsetof(struct pcpu, pc_monitorbuf), "__OFFSETOF_MONITORBUF does not correspond with offset of pc_monitorbuf."); union savefpu * get_pcb_user_save_td(struct thread *td) { vm_offset_t p; p = td->td_kstack + td->td_kstack_pages * PAGE_SIZE - roundup2(cpu_max_ext_state_size, XSAVE_AREA_ALIGN); KASSERT((p % XSAVE_AREA_ALIGN) == 0, ("Unaligned pcb_user_save area")); return ((union savefpu *)p); } union savefpu * get_pcb_user_save_pcb(struct pcb *pcb) { vm_offset_t p; p = (vm_offset_t)(pcb + 1); return ((union savefpu *)p); } struct pcb * get_pcb_td(struct thread *td) { vm_offset_t p; p = td->td_kstack + td->td_kstack_pages * PAGE_SIZE - roundup2(cpu_max_ext_state_size, XSAVE_AREA_ALIGN) - sizeof(struct pcb); return ((struct pcb *)p); } void * alloc_fpusave(int flags) { void *res; struct savefpu_ymm *sf; res = malloc(cpu_max_ext_state_size, M_DEVBUF, flags); if (use_xsave) { sf = (struct savefpu_ymm *)res; bzero(&sf->sv_xstate.sx_hd, sizeof(sf->sv_xstate.sx_hd)); sf->sv_xstate.sx_hd.xstate_bv = xsave_mask; } return (res); } /* * Common code shared between cpu_fork() and cpu_copy_thread() for * initializing a thread. */ static void copy_thread(struct thread *td1, struct thread *td2) { struct pcb *pcb2; pcb2 = td2->td_pcb; /* Ensure that td1's pcb is up to date for user threads. */ if ((td2->td_pflags & TDP_KTHREAD) == 0) { MPASS(td1 == curthread); td1->td_pcb->pcb_gs = rgs(); critical_enter(); if (PCPU_GET(fpcurthread) == td1) npxsave(td1->td_pcb->pcb_save); critical_exit(); } /* Copy td1's pcb */ bcopy(td1->td_pcb, pcb2, sizeof(*pcb2)); /* Properly initialize pcb_save */ pcb2->pcb_save = get_pcb_user_save_pcb(pcb2); /* Kernel threads start with clean NPX and segment bases. */ if ((td2->td_pflags & TDP_KTHREAD) != 0) { pcb2->pcb_gs = _udatasel; set_fsbase(td2, 0); set_gsbase(td2, 0); pcb2->pcb_flags &= ~(PCB_NPXINITDONE | PCB_NPXUSERINITDONE | PCB_KERNNPX | PCB_KERNNPX_THR); } else { MPASS((pcb2->pcb_flags & (PCB_KERNNPX | PCB_KERNNPX_THR)) == 0); bcopy(get_pcb_user_save_td(td1), get_pcb_user_save_pcb(pcb2), cpu_max_ext_state_size); } /* * Set registers for trampoline to user mode. Leave space for the * return address on stack. These are the kernel mode register values. */ pcb2->pcb_edi = 0; pcb2->pcb_esi = (int)fork_return; /* trampoline arg */ pcb2->pcb_ebp = 0; pcb2->pcb_esp = (int)td2->td_frame - sizeof(void *); /* trampoline arg */ pcb2->pcb_ebx = (int)td2; /* trampoline arg */ pcb2->pcb_eip = (int)fork_trampoline + setidt_disp; /* * If we didn't copy the pcb, we'd need to do the following registers: * pcb2->pcb_cr3: cloned above. * pcb2->pcb_dr*: cloned above. * pcb2->pcb_savefpu: cloned above. * pcb2->pcb_flags: cloned above. * pcb2->pcb_onfault: cloned above (always NULL here?). * pcb2->pcb_gs: cloned above. * pcb2->pcb_ext: cleared below. */ pcb2->pcb_ext = NULL; /* Setup to release spin count in fork_exit(). */ td2->td_md.md_spinlock_count = 1; td2->td_md.md_saved_flags = PSL_KERNEL | PSL_I; } /* * Finish a fork operation, with process p2 nearly set up. * Copy and update the pcb, set up the stack so that the child * ready to run and return to user mode. */ void cpu_fork(struct thread *td1, struct proc *p2, struct thread *td2, int flags) { struct proc *p1; struct pcb *pcb2; struct mdproc *mdp2; p1 = td1->td_proc; if ((flags & RFPROC) == 0) { if ((flags & RFMEM) == 0) { /* unshare user LDT */ struct mdproc *mdp1 = &p1->p_md; struct proc_ldt *pldt, *pldt1; mtx_lock_spin(&dt_lock); if ((pldt1 = mdp1->md_ldt) != NULL && pldt1->ldt_refcnt > 1) { pldt = user_ldt_alloc(mdp1, pldt1->ldt_len); if (pldt == NULL) panic("could not copy LDT"); mdp1->md_ldt = pldt; set_user_ldt(mdp1); user_ldt_deref(pldt1); } else mtx_unlock_spin(&dt_lock); } return; } /* Point the pcb to the top of the stack */ pcb2 = get_pcb_td(td2); td2->td_pcb = pcb2; copy_thread(td1, td2); /* Reset debug registers in the new process */ x86_clear_dbregs(pcb2); /* Point mdproc and then copy over td1's contents */ mdp2 = &p2->p_md; bcopy(&p1->p_md, mdp2, sizeof(*mdp2)); /* * Copy the trap frame for the return to user mode as if from a * syscall. This copies most of the user mode register values. * The -VM86_STACK_SPACE (-16) is so we can expand the trapframe * if we go to vm86. */ td2->td_frame = (struct trapframe *)((caddr_t)td2->td_pcb - VM86_STACK_SPACE) - 1; bcopy(td1->td_frame, td2->td_frame, sizeof(struct trapframe)); td2->td_frame->tf_eax = 0; /* Child returns zero */ td2->td_frame->tf_eflags &= ~PSL_C; /* success */ td2->td_frame->tf_edx = 1; /* * If the parent process has the trap bit set (i.e. a debugger * had single stepped the process to the system call), we need * to clear the trap flag from the new frame. */ td2->td_frame->tf_eflags &= ~PSL_T; /* Set cr3 for the new process. */ pcb2->pcb_cr3 = pmap_get_cr3(vmspace_pmap(p2->p_vmspace)); /* * XXX don't copy the i/o pages. this should probably be fixed. */ pcb2->pcb_ext = NULL; /* Copy the LDT, if necessary. */ mtx_lock_spin(&dt_lock); if (mdp2->md_ldt != NULL) { if (flags & RFMEM) { mdp2->md_ldt->ldt_refcnt++; } else { mdp2->md_ldt = user_ldt_alloc(mdp2, mdp2->md_ldt->ldt_len); if (mdp2->md_ldt == NULL) panic("could not copy LDT"); } } mtx_unlock_spin(&dt_lock); /* * Now, cpu_switch() can schedule the new process. * pcb_esp is loaded pointing to the cpu_switch() stack frame * containing the return address when exiting cpu_switch. * This will normally be to fork_trampoline(), which will have * %ebx loaded with the new proc's pointer. fork_trampoline() * will set up a stack to call fork_return(p, frame); to complete * the return to user-mode. */ } /* * Intercept the return address from a freshly forked process that has NOT * been scheduled yet. * * This is needed to make kernel threads stay in kernel mode. */ void cpu_fork_kthread_handler(struct thread *td, void (*func)(void *), void *arg) { /* * Note that the trap frame follows the args, so the function * is really called like this: func(arg, frame); */ td->td_pcb->pcb_esi = (int) func; /* function */ td->td_pcb->pcb_ebx = (int) arg; /* first arg */ } void cpu_exit(struct thread *td) { /* * If this process has a custom LDT, release it. Reset pc->pcb_gs * and %gs before we free it in case they refer to an LDT entry. */ mtx_lock_spin(&dt_lock); if (td->td_proc->p_md.md_ldt) { td->td_pcb->pcb_gs = _udatasel; load_gs(_udatasel); user_ldt_free(td); } else mtx_unlock_spin(&dt_lock); } void cpu_thread_exit(struct thread *td) { critical_enter(); if (td == PCPU_GET(fpcurthread)) npxdrop(); critical_exit(); /* Disable any hardware breakpoints. */ if (td->td_pcb->pcb_flags & PCB_DBREGS) { reset_dbregs(); td->td_pcb->pcb_flags &= ~PCB_DBREGS; } } void cpu_thread_clean(struct thread *td) { struct pcb *pcb; pcb = td->td_pcb; if (pcb->pcb_ext != NULL) { /* if (pcb->pcb_ext->ext_refcount-- == 1) ?? */ /* * XXX do we need to move the TSS off the allocated pages * before freeing them? (not done here) */ pmap_trm_free(pcb->pcb_ext, ctob(IOPAGES + 1)); pcb->pcb_ext = NULL; } } void cpu_thread_swapin(struct thread *td) { } void cpu_thread_swapout(struct thread *td) { } void cpu_thread_alloc(struct thread *td) { struct pcb *pcb; struct xstate_hdr *xhdr; td->td_pcb = pcb = get_pcb_td(td); td->td_frame = (struct trapframe *)((caddr_t)pcb - VM86_STACK_SPACE) - 1; pcb->pcb_ext = NULL; pcb->pcb_save = get_pcb_user_save_pcb(pcb); if (use_xsave) { xhdr = (struct xstate_hdr *)(pcb->pcb_save + 1); bzero(xhdr, sizeof(*xhdr)); xhdr->xstate_bv = xsave_mask; } } void cpu_thread_free(struct thread *td) { cpu_thread_clean(td); } bool cpu_exec_vmspace_reuse(struct proc *p __unused, vm_map_t map __unused) { return (true); } int cpu_procctl(struct thread *td __unused, int idtype __unused, id_t id __unused, int com __unused, void *data __unused) { return (EINVAL); } void cpu_set_syscall_retval(struct thread *td, int error) { switch (error) { case 0: td->td_frame->tf_eax = td->td_retval[0]; td->td_frame->tf_edx = td->td_retval[1]; td->td_frame->tf_eflags &= ~PSL_C; break; case ERESTART: /* * Reconstruct pc, assuming lcall $X,y is 7 bytes, int * 0x80 is 2 bytes. We saved this in tf_err. */ td->td_frame->tf_eip -= td->td_frame->tf_err; break; case EJUSTRETURN: break; default: td->td_frame->tf_eax = error; td->td_frame->tf_eflags |= PSL_C; break; } } /* * Initialize machine state, mostly pcb and trap frame for a new * thread, about to return to userspace. Put enough state in the new * thread's PCB to get it to go back to the fork_return(), which * finalizes the thread state and handles peculiarities of the first * return to userspace for the new thread. */ void cpu_copy_thread(struct thread *td, struct thread *td0) { copy_thread(td0, td); /* * Copy user general-purpose registers. * * Some of these registers are rewritten by cpu_set_upcall() * and linux_set_upcall(). */ bcopy(td0->td_frame, td->td_frame, sizeof(struct trapframe)); /* If the current thread has the trap bit set (i.e. a debugger had * single stepped the process to the system call), we need to clear * the trap flag from the new frame. Otherwise, the new thread will * receive a (likely unexpected) SIGTRAP when it executes the first * instruction after returning to userland. */ td->td_frame->tf_eflags &= ~PSL_T; } /* * Set that machine state for performing an upcall that starts * the entry function with the given argument. */ void cpu_set_upcall(struct thread *td, void (*entry)(void *), void *arg, stack_t *stack) { /* * Do any extra cleaning that needs to be done. * The thread may have optional components * that are not present in a fresh thread. * This may be a recycled thread so make it look * as though it's newly allocated. */ cpu_thread_clean(td); /* * Set the trap frame to point at the beginning of the entry * function. */ td->td_frame->tf_ebp = 0; td->td_frame->tf_esp = (((int)stack->ss_sp + stack->ss_size - 4) & ~0x0f) - 4; td->td_frame->tf_eip = (int)entry; /* Return address sentinel value to stop stack unwinding. */ suword((void *)td->td_frame->tf_esp, 0); /* Pass the argument to the entry point. */ suword((void *)(td->td_frame->tf_esp + sizeof(void *)), (int)arg); } int cpu_set_user_tls(struct thread *td, void *tls_base) { struct segment_descriptor sd; uint32_t base; /* * Construct a descriptor and store it in the pcb for * the next context switch. Also store it in the gdt * so that the load of tf_fs into %fs will activate it * at return to userland. */ base = (uint32_t)tls_base; sd.sd_lobase = base & 0xffffff; sd.sd_hibase = (base >> 24) & 0xff; sd.sd_lolimit = 0xffff; /* 4GB limit, wraps around */ sd.sd_hilimit = 0xf; sd.sd_type = SDT_MEMRWA; sd.sd_dpl = SEL_UPL; sd.sd_p = 1; sd.sd_xx = 0; sd.sd_def32 = 1; sd.sd_gran = 1; critical_enter(); /* set %gs */ td->td_pcb->pcb_gsd = sd; if (td == curthread) { PCPU_GET(fsgs_gdt)[1] = sd; load_gs(GSEL(GUGS_SEL, SEL_UPL)); } critical_exit(); return (0); } /* * Convert kernel VA to physical address */ vm_paddr_t kvtop(void *addr) { vm_paddr_t pa; pa = pmap_kextract((vm_offset_t)addr); if (pa == 0) panic("kvtop: zero page frame"); return (pa); } /* * Get an sf_buf from the freelist. May block if none are available. */ void sf_buf_map(struct sf_buf *sf, int flags) { pmap_sf_buf_map(sf); #ifdef SMP sf_buf_shootdown(sf, flags); #endif } #ifdef SMP static void sf_buf_shootdown_curcpu_cb(pmap_t pmap __unused, vm_offset_t addr1 __unused, vm_offset_t addr2 __unused) { } void sf_buf_shootdown(struct sf_buf *sf, int flags) { cpuset_t other_cpus; u_int cpuid; sched_pin(); cpuid = PCPU_GET(cpuid); if (!CPU_ISSET(cpuid, &sf->cpumask)) { CPU_SET(cpuid, &sf->cpumask); invlpg(sf->kva); } if ((flags & SFB_CPUPRIVATE) == 0) { other_cpus = all_cpus; CPU_CLR(cpuid, &other_cpus); - CPU_ANDNOT(&other_cpus, &sf->cpumask); + CPU_ANDNOT(&other_cpus, &other_cpus, &sf->cpumask); if (!CPU_EMPTY(&other_cpus)) { - CPU_OR(&sf->cpumask, &other_cpus); + CPU_OR(&sf->cpumask, &sf->cpumask, &other_cpus); smp_masked_invlpg(other_cpus, sf->kva, kernel_pmap, sf_buf_shootdown_curcpu_cb); } } sched_unpin(); } #endif /* * MD part of sf_buf_free(). */ int sf_buf_unmap(struct sf_buf *sf) { return (0); } static void sf_buf_invalidate(struct sf_buf *sf) { vm_page_t m = sf->m; /* * Use pmap_qenter to update the pte for * existing mapping, in particular, the PAT * settings are recalculated. */ pmap_qenter(sf->kva, &m, 1); pmap_invalidate_cache_range(sf->kva, sf->kva + PAGE_SIZE); } /* * Invalidate the cache lines that may belong to the page, if * (possibly old) mapping of the page by sf buffer exists. Returns * TRUE when mapping was found and cache invalidated. */ boolean_t sf_buf_invalidate_cache(vm_page_t m) { return (sf_buf_process_page(m, sf_buf_invalidate)); } /* * Software interrupt handler for queued VM system processing. */ void swi_vm(void *dummy) { if (busdma_swi_pending != 0) busdma_swi(); } /* * Tell whether this address is in some physical memory region. * Currently used by the kernel coredump code in order to avoid * dumping the ``ISA memory hole'' which could cause indefinite hangs, * or other unpredictable behaviour. */ int is_physical_memory(vm_paddr_t addr) { #ifdef DEV_ISA /* The ISA ``memory hole''. */ if (addr >= 0xa0000 && addr < 0x100000) return 0; #endif /* * stuff other tests for known memory-mapped devices (PCI?) * here */ return 1; } diff --git a/sys/kern/kern_cpuset.c b/sys/kern/kern_cpuset.c index 119df57c7b02..962e7abb44b7 100644 --- a/sys/kern/kern_cpuset.c +++ b/sys/kern/kern_cpuset.c @@ -1,2459 +1,2456 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2008, Jeffrey Roberson * All rights reserved. * * Copyright (c) 2008 Nokia Corporation * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice unmodified, this list of conditions, and the following * disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. * */ #include __FBSDID("$FreeBSD$"); #include "opt_ddb.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef DDB #include #endif /* DDB */ /* * cpusets provide a mechanism for creating and manipulating sets of * processors for the purpose of constraining the scheduling of threads to * specific processors. * * Each process belongs to an identified set, by default this is set 1. Each * thread may further restrict the cpus it may run on to a subset of this * named set. This creates an anonymous set which other threads and processes * may not join by number. * * The named set is referred to herein as the 'base' set to avoid ambiguity. * This set is usually a child of a 'root' set while the anonymous set may * simply be referred to as a mask. In the syscall api these are referred to * as the ROOT, CPUSET, and MASK levels where CPUSET is called 'base' here. * * Threads inherit their set from their creator whether it be anonymous or * not. This means that anonymous sets are immutable because they may be * shared. To modify an anonymous set a new set is created with the desired * mask and the same parent as the existing anonymous set. This gives the * illusion of each thread having a private mask. * * Via the syscall apis a user may ask to retrieve or modify the root, base, * or mask that is discovered via a pid, tid, or setid. Modifying a set * modifies all numbered and anonymous child sets to comply with the new mask. * Modifying a pid or tid's mask applies only to that tid but must still * exist within the assigned parent set. * * A thread may not be assigned to a group separate from other threads in * the process. This is to remove ambiguity when the setid is queried with * a pid argument. There is no other technical limitation. * * This somewhat complex arrangement is intended to make it easy for * applications to query available processors and bind their threads to * specific processors while also allowing administrators to dynamically * reprovision by changing sets which apply to groups of processes. * * A simple application should not concern itself with sets at all and * rather apply masks to its own threads via CPU_WHICH_TID and a -1 id * meaning 'curthread'. It may query available cpus for that tid with a * getaffinity call using (CPU_LEVEL_CPUSET, CPU_WHICH_PID, -1, ...). */ LIST_HEAD(domainlist, domainset); struct domainset __read_mostly domainset_firsttouch; struct domainset __read_mostly domainset_fixed[MAXMEMDOM]; struct domainset __read_mostly domainset_interleave; struct domainset __read_mostly domainset_prefer[MAXMEMDOM]; struct domainset __read_mostly domainset_roundrobin; static uma_zone_t cpuset_zone; static uma_zone_t domainset_zone; static struct mtx cpuset_lock; static struct setlist cpuset_ids; static struct domainlist cpuset_domains; static struct unrhdr *cpuset_unr; static struct cpuset *cpuset_zero, *cpuset_default, *cpuset_kernel; static struct domainset *domainset0, *domainset2; /* Return the size of cpuset_t at the kernel level */ SYSCTL_INT(_kern_sched, OID_AUTO, cpusetsize, CTLFLAG_RD | CTLFLAG_CAPRD, SYSCTL_NULL_INT_PTR, sizeof(cpuset_t), "sizeof(cpuset_t)"); cpuset_t *cpuset_root; cpuset_t cpuset_domain[MAXMEMDOM]; static int domainset_valid(const struct domainset *, const struct domainset *); /* * Find the first non-anonymous set starting from 'set'. */ static struct cpuset * cpuset_getbase(struct cpuset *set) { if (set->cs_id == CPUSET_INVALID) set = set->cs_parent; return (set); } /* * Walks up the tree from 'set' to find the root. */ static struct cpuset * cpuset_getroot(struct cpuset *set) { while ((set->cs_flags & CPU_SET_ROOT) == 0 && set->cs_parent != NULL) set = set->cs_parent; return (set); } /* * Acquire a reference to a cpuset, all pointers must be tracked with refs. */ struct cpuset * cpuset_ref(struct cpuset *set) { refcount_acquire(&set->cs_ref); return (set); } /* * Walks up the tree from 'set' to find the root. Returns the root * referenced. */ static struct cpuset * cpuset_refroot(struct cpuset *set) { return (cpuset_ref(cpuset_getroot(set))); } /* * Find the first non-anonymous set starting from 'set'. Returns this set * referenced. May return the passed in set with an extra ref if it is * not anonymous. */ static struct cpuset * cpuset_refbase(struct cpuset *set) { return (cpuset_ref(cpuset_getbase(set))); } /* * Release a reference in a context where it is safe to allocate. */ void cpuset_rel(struct cpuset *set) { cpusetid_t id; if (refcount_release_if_not_last(&set->cs_ref)) return; mtx_lock_spin(&cpuset_lock); if (!refcount_release(&set->cs_ref)) { mtx_unlock_spin(&cpuset_lock); return; } LIST_REMOVE(set, cs_siblings); id = set->cs_id; if (id != CPUSET_INVALID) LIST_REMOVE(set, cs_link); mtx_unlock_spin(&cpuset_lock); cpuset_rel(set->cs_parent); uma_zfree(cpuset_zone, set); if (id != CPUSET_INVALID) free_unr(cpuset_unr, id); } /* * Deferred release must be used when in a context that is not safe to * allocate/free. This places any unreferenced sets on the list 'head'. */ static void cpuset_rel_defer(struct setlist *head, struct cpuset *set) { if (refcount_release_if_not_last(&set->cs_ref)) return; mtx_lock_spin(&cpuset_lock); if (!refcount_release(&set->cs_ref)) { mtx_unlock_spin(&cpuset_lock); return; } LIST_REMOVE(set, cs_siblings); if (set->cs_id != CPUSET_INVALID) LIST_REMOVE(set, cs_link); LIST_INSERT_HEAD(head, set, cs_link); mtx_unlock_spin(&cpuset_lock); } /* * Complete a deferred release. Removes the set from the list provided to * cpuset_rel_defer. */ static void cpuset_rel_complete(struct cpuset *set) { cpusetid_t id; id = set->cs_id; LIST_REMOVE(set, cs_link); cpuset_rel(set->cs_parent); uma_zfree(cpuset_zone, set); if (id != CPUSET_INVALID) free_unr(cpuset_unr, id); } /* * Find a set based on an id. Returns it with a ref. */ static struct cpuset * cpuset_lookup(cpusetid_t setid, struct thread *td) { struct cpuset *set; if (setid == CPUSET_INVALID) return (NULL); mtx_lock_spin(&cpuset_lock); LIST_FOREACH(set, &cpuset_ids, cs_link) if (set->cs_id == setid) break; if (set) cpuset_ref(set); mtx_unlock_spin(&cpuset_lock); KASSERT(td != NULL, ("[%s:%d] td is NULL", __func__, __LINE__)); if (set != NULL && jailed(td->td_ucred)) { struct cpuset *jset, *tset; jset = td->td_ucred->cr_prison->pr_cpuset; for (tset = set; tset != NULL; tset = tset->cs_parent) if (tset == jset) break; if (tset == NULL) { cpuset_rel(set); set = NULL; } } return (set); } /* * Initialize a set in the space provided in 'set' with the provided parameters. * The set is returned with a single ref. May return EDEADLK if the set * will have no valid cpu based on restrictions from the parent. */ static int cpuset_init(struct cpuset *set, struct cpuset *parent, const cpuset_t *mask, struct domainset *domain, cpusetid_t id) { if (domain == NULL) domain = parent->cs_domain; if (mask == NULL) mask = &parent->cs_mask; if (!CPU_OVERLAP(&parent->cs_mask, mask)) return (EDEADLK); /* The domain must be prepared ahead of time. */ if (!domainset_valid(parent->cs_domain, domain)) return (EDEADLK); CPU_COPY(mask, &set->cs_mask); LIST_INIT(&set->cs_children); refcount_init(&set->cs_ref, 1); set->cs_flags = 0; mtx_lock_spin(&cpuset_lock); set->cs_domain = domain; - CPU_AND(&set->cs_mask, &parent->cs_mask); + CPU_AND(&set->cs_mask, &set->cs_mask, &parent->cs_mask); set->cs_id = id; set->cs_parent = cpuset_ref(parent); LIST_INSERT_HEAD(&parent->cs_children, set, cs_siblings); if (set->cs_id != CPUSET_INVALID) LIST_INSERT_HEAD(&cpuset_ids, set, cs_link); mtx_unlock_spin(&cpuset_lock); return (0); } /* * Create a new non-anonymous set with the requested parent and mask. May * return failures if the mask is invalid or a new number can not be * allocated. * * If *setp is not NULL, then it will be used as-is. The caller must take * into account that *setp will be inserted at the head of cpuset_ids and * plan any potentially conflicting cs_link usage accordingly. */ static int cpuset_create(struct cpuset **setp, struct cpuset *parent, const cpuset_t *mask) { struct cpuset *set; cpusetid_t id; int error; bool dofree; id = alloc_unr(cpuset_unr); if (id == -1) return (ENFILE); dofree = (*setp == NULL); if (*setp != NULL) set = *setp; else *setp = set = uma_zalloc(cpuset_zone, M_WAITOK | M_ZERO); error = cpuset_init(set, parent, mask, NULL, id); if (error == 0) return (0); free_unr(cpuset_unr, id); if (dofree) uma_zfree(cpuset_zone, set); return (error); } static void cpuset_freelist_add(struct setlist *list, int count) { struct cpuset *set; int i; for (i = 0; i < count; i++) { set = uma_zalloc(cpuset_zone, M_ZERO | M_WAITOK); LIST_INSERT_HEAD(list, set, cs_link); } } static void cpuset_freelist_init(struct setlist *list, int count) { LIST_INIT(list); cpuset_freelist_add(list, count); } static void cpuset_freelist_free(struct setlist *list) { struct cpuset *set; while ((set = LIST_FIRST(list)) != NULL) { LIST_REMOVE(set, cs_link); uma_zfree(cpuset_zone, set); } } static void domainset_freelist_add(struct domainlist *list, int count) { struct domainset *set; int i; for (i = 0; i < count; i++) { set = uma_zalloc(domainset_zone, M_ZERO | M_WAITOK); LIST_INSERT_HEAD(list, set, ds_link); } } static void domainset_freelist_init(struct domainlist *list, int count) { LIST_INIT(list); domainset_freelist_add(list, count); } static void domainset_freelist_free(struct domainlist *list) { struct domainset *set; while ((set = LIST_FIRST(list)) != NULL) { LIST_REMOVE(set, ds_link); uma_zfree(domainset_zone, set); } } /* Copy a domainset preserving mask and policy. */ static void domainset_copy(const struct domainset *from, struct domainset *to) { DOMAINSET_COPY(&from->ds_mask, &to->ds_mask); to->ds_policy = from->ds_policy; to->ds_prefer = from->ds_prefer; } /* Return 1 if mask and policy are equal, otherwise 0. */ static int domainset_equal(const struct domainset *one, const struct domainset *two) { return (DOMAINSET_CMP(&one->ds_mask, &two->ds_mask) == 0 && one->ds_policy == two->ds_policy && one->ds_prefer == two->ds_prefer); } /* Return 1 if child is a valid subset of parent. */ static int domainset_valid(const struct domainset *parent, const struct domainset *child) { if (child->ds_policy != DOMAINSET_POLICY_PREFER) return (DOMAINSET_SUBSET(&parent->ds_mask, &child->ds_mask)); return (DOMAINSET_ISSET(child->ds_prefer, &parent->ds_mask)); } static int domainset_restrict(const struct domainset *parent, const struct domainset *child) { if (child->ds_policy != DOMAINSET_POLICY_PREFER) return (DOMAINSET_OVERLAP(&parent->ds_mask, &child->ds_mask)); return (DOMAINSET_ISSET(child->ds_prefer, &parent->ds_mask)); } /* * Lookup or create a domainset. The key is provided in ds_mask and * ds_policy. If the domainset does not yet exist the storage in * 'domain' is used to insert. Otherwise this storage is freed to the * domainset_zone and the existing domainset is returned. */ static struct domainset * _domainset_create(struct domainset *domain, struct domainlist *freelist) { struct domainset *ndomain; int i, j; KASSERT(domain->ds_cnt <= vm_ndomains, ("invalid domain count in domainset %p", domain)); KASSERT(domain->ds_policy != DOMAINSET_POLICY_PREFER || domain->ds_prefer < vm_ndomains, ("invalid preferred domain in domains %p", domain)); mtx_lock_spin(&cpuset_lock); LIST_FOREACH(ndomain, &cpuset_domains, ds_link) if (domainset_equal(ndomain, domain)) break; /* * If the domain does not yet exist we insert it and initialize * various iteration helpers which are not part of the key. */ if (ndomain == NULL) { LIST_INSERT_HEAD(&cpuset_domains, domain, ds_link); domain->ds_cnt = DOMAINSET_COUNT(&domain->ds_mask); for (i = 0, j = 0; i < DOMAINSET_FLS(&domain->ds_mask); i++) if (DOMAINSET_ISSET(i, &domain->ds_mask)) domain->ds_order[j++] = i; } mtx_unlock_spin(&cpuset_lock); if (ndomain == NULL) return (domain); if (freelist != NULL) LIST_INSERT_HEAD(freelist, domain, ds_link); else uma_zfree(domainset_zone, domain); return (ndomain); } /* * Are any of the domains in the mask empty? If so, silently * remove them and update the domainset accordingly. If only empty * domains are present, we must return failure. */ static bool domainset_empty_vm(struct domainset *domain) { domainset_t empty; int i, j; DOMAINSET_ZERO(&empty); for (i = 0; i < vm_ndomains; i++) if (VM_DOMAIN_EMPTY(i)) DOMAINSET_SET(i, &empty); if (DOMAINSET_SUBSET(&empty, &domain->ds_mask)) return (true); /* Remove empty domains from the set and recompute. */ DOMAINSET_ANDNOT(&domain->ds_mask, &empty); domain->ds_cnt = DOMAINSET_COUNT(&domain->ds_mask); for (i = j = 0; i < DOMAINSET_FLS(&domain->ds_mask); i++) if (DOMAINSET_ISSET(i, &domain->ds_mask)) domain->ds_order[j++] = i; /* Convert a PREFER policy referencing an empty domain to RR. */ if (domain->ds_policy == DOMAINSET_POLICY_PREFER && DOMAINSET_ISSET(domain->ds_prefer, &empty)) { domain->ds_policy = DOMAINSET_POLICY_ROUNDROBIN; domain->ds_prefer = -1; } return (false); } /* * Create or lookup a domainset based on the key held in 'domain'. */ struct domainset * domainset_create(const struct domainset *domain) { struct domainset *ndomain; /* * Validate the policy. It must specify a useable policy number with * only valid domains. Preferred must include the preferred domain * in the mask. */ if (domain->ds_policy <= DOMAINSET_POLICY_INVALID || domain->ds_policy > DOMAINSET_POLICY_MAX) return (NULL); if (domain->ds_policy == DOMAINSET_POLICY_PREFER && !DOMAINSET_ISSET(domain->ds_prefer, &domain->ds_mask)) return (NULL); if (!DOMAINSET_SUBSET(&domainset0->ds_mask, &domain->ds_mask)) return (NULL); ndomain = uma_zalloc(domainset_zone, M_WAITOK | M_ZERO); domainset_copy(domain, ndomain); return _domainset_create(ndomain, NULL); } /* * Update thread domainset pointers. */ static void domainset_notify(void) { struct thread *td; struct proc *p; sx_slock(&allproc_lock); FOREACH_PROC_IN_SYSTEM(p) { PROC_LOCK(p); if (p->p_state == PRS_NEW) { PROC_UNLOCK(p); continue; } FOREACH_THREAD_IN_PROC(p, td) { thread_lock(td); td->td_domain.dr_policy = td->td_cpuset->cs_domain; thread_unlock(td); } PROC_UNLOCK(p); } sx_sunlock(&allproc_lock); kernel_object->domain.dr_policy = cpuset_kernel->cs_domain; } /* * Create a new set that is a subset of a parent. */ static struct domainset * domainset_shadow(const struct domainset *pdomain, const struct domainset *domain, struct domainlist *freelist) { struct domainset *ndomain; ndomain = LIST_FIRST(freelist); LIST_REMOVE(ndomain, ds_link); /* * Initialize the key from the request. */ domainset_copy(domain, ndomain); /* * Restrict the key by the parent. */ DOMAINSET_AND(&ndomain->ds_mask, &pdomain->ds_mask); return _domainset_create(ndomain, freelist); } /* * Recursively check for errors that would occur from applying mask to * the tree of sets starting at 'set'. Checks for sets that would become * empty as well as RDONLY flags. */ static int cpuset_testupdate(struct cpuset *set, cpuset_t *mask, int augment_mask) { struct cpuset *nset; cpuset_t newmask; int error; mtx_assert(&cpuset_lock, MA_OWNED); if (set->cs_flags & CPU_SET_RDONLY) return (EPERM); if (augment_mask) { - CPU_COPY(&set->cs_mask, &newmask); - CPU_AND(&newmask, mask); + CPU_AND(&newmask, &set->cs_mask, mask); } else CPU_COPY(mask, &newmask); if (CPU_EMPTY(&newmask)) return (EDEADLK); error = 0; LIST_FOREACH(nset, &set->cs_children, cs_siblings) if ((error = cpuset_testupdate(nset, &newmask, 1)) != 0) break; return (error); } /* * Applies the mask 'mask' without checking for empty sets or permissions. */ static void cpuset_update(struct cpuset *set, cpuset_t *mask) { struct cpuset *nset; mtx_assert(&cpuset_lock, MA_OWNED); - CPU_AND(&set->cs_mask, mask); + CPU_AND(&set->cs_mask, &set->cs_mask, mask); LIST_FOREACH(nset, &set->cs_children, cs_siblings) cpuset_update(nset, &set->cs_mask); return; } /* * Modify the set 'set' to use a copy of the mask provided. Apply this new * mask to restrict all children in the tree. Checks for validity before * applying the changes. */ static int cpuset_modify(struct cpuset *set, cpuset_t *mask) { struct cpuset *root; int error; error = priv_check(curthread, PRIV_SCHED_CPUSET); if (error) return (error); /* * In case we are called from within the jail, * we do not allow modifying the dedicated root * cpuset of the jail but may still allow to * change child sets, including subordinate jails' * roots. */ if ((set->cs_flags & CPU_SET_ROOT) != 0 && jailed(curthread->td_ucred) && set == curthread->td_ucred->cr_prison->pr_cpuset) return (EPERM); /* * Verify that we have access to this set of * cpus. */ if ((set->cs_flags & (CPU_SET_ROOT | CPU_SET_RDONLY)) == CPU_SET_ROOT) { KASSERT(set->cs_parent != NULL, ("jail.cpuset=%d is not a proper child of parent jail's root.", set->cs_id)); /* * cpuset_getroot() cannot work here due to how top-level jail * roots are constructed. Top-level jails are parented to * thread0's cpuset (i.e. cpuset 1) rather than the system root. */ root = set->cs_parent; } else { root = cpuset_getroot(set); } mtx_lock_spin(&cpuset_lock); if (root && !CPU_SUBSET(&root->cs_mask, mask)) { error = EINVAL; goto out; } error = cpuset_testupdate(set, mask, 0); if (error) goto out; CPU_COPY(mask, &set->cs_mask); cpuset_update(set, mask); out: mtx_unlock_spin(&cpuset_lock); return (error); } /* * Recursively check for errors that would occur from applying mask to * the tree of sets starting at 'set'. Checks for sets that would become * empty as well as RDONLY flags. */ static int cpuset_testupdate_domain(struct cpuset *set, struct domainset *dset, struct domainset *orig, int *count, int augment_mask __unused) { struct cpuset *nset; struct domainset *domain; struct domainset newset; int error; mtx_assert(&cpuset_lock, MA_OWNED); if (set->cs_flags & CPU_SET_RDONLY) return (EPERM); domain = set->cs_domain; domainset_copy(domain, &newset); if (!domainset_equal(domain, orig)) { if (!domainset_restrict(domain, dset)) return (EDEADLK); DOMAINSET_AND(&newset.ds_mask, &dset->ds_mask); /* Count the number of domains that are changing. */ (*count)++; } error = 0; LIST_FOREACH(nset, &set->cs_children, cs_siblings) if ((error = cpuset_testupdate_domain(nset, &newset, domain, count, 1)) != 0) break; return (error); } /* * Applies the mask 'mask' without checking for empty sets or permissions. */ static void cpuset_update_domain(struct cpuset *set, struct domainset *domain, struct domainset *orig, struct domainlist *domains) { struct cpuset *nset; mtx_assert(&cpuset_lock, MA_OWNED); /* * If this domainset has changed from the parent we must calculate * a new set. Otherwise it simply inherits from the parent. When * we inherit from the parent we get a new mask and policy. If the * set is modified from the parent we keep the policy and only * update the mask. */ if (set->cs_domain != orig) { orig = set->cs_domain; set->cs_domain = domainset_shadow(domain, orig, domains); } else set->cs_domain = domain; LIST_FOREACH(nset, &set->cs_children, cs_siblings) cpuset_update_domain(nset, set->cs_domain, orig, domains); return; } /* * Modify the set 'set' to use a copy the domainset provided. Apply this new * mask to restrict all children in the tree. Checks for validity before * applying the changes. */ static int cpuset_modify_domain(struct cpuset *set, struct domainset *domain) { struct domainlist domains; struct domainset temp; struct domainset *dset; struct cpuset *root; int ndomains, needed; int error; error = priv_check(curthread, PRIV_SCHED_CPUSET); if (error) return (error); /* * In case we are called from within the jail * we do not allow modifying the dedicated root * cpuset of the jail but may still allow to * change child sets. */ if (jailed(curthread->td_ucred) && set->cs_flags & CPU_SET_ROOT) return (EPERM); domainset_freelist_init(&domains, 0); domain = domainset_create(domain); ndomains = 0; mtx_lock_spin(&cpuset_lock); for (;;) { root = cpuset_getroot(set); dset = root->cs_domain; /* * Verify that we have access to this set of domains. */ if (!domainset_valid(dset, domain)) { error = EINVAL; goto out; } /* * If applying prefer we keep the current set as the fallback. */ if (domain->ds_policy == DOMAINSET_POLICY_PREFER) DOMAINSET_COPY(&set->cs_domain->ds_mask, &domain->ds_mask); /* * Determine whether we can apply this set of domains and * how many new domain structures it will require. */ domainset_copy(domain, &temp); needed = 0; error = cpuset_testupdate_domain(set, &temp, set->cs_domain, &needed, 0); if (error) goto out; if (ndomains >= needed) break; /* Dropping the lock; we'll need to re-evaluate again. */ mtx_unlock_spin(&cpuset_lock); domainset_freelist_add(&domains, needed - ndomains); ndomains = needed; mtx_lock_spin(&cpuset_lock); } dset = set->cs_domain; cpuset_update_domain(set, domain, dset, &domains); out: mtx_unlock_spin(&cpuset_lock); domainset_freelist_free(&domains); if (error == 0) domainset_notify(); return (error); } /* * Resolve the 'which' parameter of several cpuset apis. * * For WHICH_PID and WHICH_TID return a locked proc and valid proc/tid. Also * checks for permission via p_cansched(). * * For WHICH_SET returns a valid set with a new reference. * * -1 may be supplied for any argument to mean the current proc/thread or * the base set of the current thread. May fail with ESRCH/EPERM. */ int cpuset_which(cpuwhich_t which, id_t id, struct proc **pp, struct thread **tdp, struct cpuset **setp) { struct cpuset *set; struct thread *td; struct proc *p; int error; *pp = p = NULL; *tdp = td = NULL; *setp = set = NULL; switch (which) { case CPU_WHICH_PID: if (id == -1) { PROC_LOCK(curproc); p = curproc; break; } if ((p = pfind(id)) == NULL) return (ESRCH); break; case CPU_WHICH_TID: if (id == -1) { PROC_LOCK(curproc); p = curproc; td = curthread; break; } td = tdfind(id, -1); if (td == NULL) return (ESRCH); p = td->td_proc; break; case CPU_WHICH_CPUSET: if (id == -1) { thread_lock(curthread); set = cpuset_refbase(curthread->td_cpuset); thread_unlock(curthread); } else set = cpuset_lookup(id, curthread); if (set) { *setp = set; return (0); } return (ESRCH); case CPU_WHICH_JAIL: { /* Find `set' for prison with given id. */ struct prison *pr; sx_slock(&allprison_lock); pr = prison_find_child(curthread->td_ucred->cr_prison, id); sx_sunlock(&allprison_lock); if (pr == NULL) return (ESRCH); cpuset_ref(pr->pr_cpuset); *setp = pr->pr_cpuset; mtx_unlock(&pr->pr_mtx); return (0); } case CPU_WHICH_IRQ: case CPU_WHICH_DOMAIN: return (0); default: return (EINVAL); } error = p_cansched(curthread, p); if (error) { PROC_UNLOCK(p); return (error); } if (td == NULL) td = FIRST_THREAD_IN_PROC(p); *pp = p; *tdp = td; return (0); } static int cpuset_testshadow(struct cpuset *set, const cpuset_t *mask, const struct domainset *domain) { struct cpuset *parent; struct domainset *dset; parent = cpuset_getbase(set); /* * If we are restricting a cpu mask it must be a subset of the * parent or invalid CPUs have been specified. */ if (mask != NULL && !CPU_SUBSET(&parent->cs_mask, mask)) return (EINVAL); /* * If we are restricting a domain mask it must be a subset of the * parent or invalid domains have been specified. */ dset = parent->cs_domain; if (domain != NULL && !domainset_valid(dset, domain)) return (EINVAL); return (0); } /* * Create an anonymous set with the provided mask in the space provided by * 'nset'. If the passed in set is anonymous we use its parent otherwise * the new set is a child of 'set'. */ static int cpuset_shadow(struct cpuset *set, struct cpuset **nsetp, const cpuset_t *mask, const struct domainset *domain, struct setlist *cpusets, struct domainlist *domains) { struct cpuset *parent; struct cpuset *nset; struct domainset *dset; struct domainset *d; int error; error = cpuset_testshadow(set, mask, domain); if (error) return (error); parent = cpuset_getbase(set); dset = parent->cs_domain; if (mask == NULL) mask = &set->cs_mask; if (domain != NULL) d = domainset_shadow(dset, domain, domains); else d = set->cs_domain; nset = LIST_FIRST(cpusets); error = cpuset_init(nset, parent, mask, d, CPUSET_INVALID); if (error == 0) { LIST_REMOVE(nset, cs_link); *nsetp = nset; } return (error); } static struct cpuset * cpuset_update_thread(struct thread *td, struct cpuset *nset) { struct cpuset *tdset; tdset = td->td_cpuset; td->td_cpuset = nset; td->td_domain.dr_policy = nset->cs_domain; sched_affinity(td); return (tdset); } static int cpuset_setproc_test_maskthread(struct cpuset *tdset, cpuset_t *mask, struct domainset *domain) { struct cpuset *parent; parent = cpuset_getbase(tdset); if (mask == NULL) mask = &tdset->cs_mask; if (domain == NULL) domain = tdset->cs_domain; return cpuset_testshadow(parent, mask, domain); } static int cpuset_setproc_maskthread(struct cpuset *tdset, cpuset_t *mask, struct domainset *domain, struct cpuset **nsetp, struct setlist *freelist, struct domainlist *domainlist) { struct cpuset *parent; parent = cpuset_getbase(tdset); if (mask == NULL) mask = &tdset->cs_mask; if (domain == NULL) domain = tdset->cs_domain; return cpuset_shadow(parent, nsetp, mask, domain, freelist, domainlist); } static int cpuset_setproc_setthread_mask(struct cpuset *tdset, struct cpuset *set, cpuset_t *mask, struct domainset *domain) { struct cpuset *parent; parent = cpuset_getbase(tdset); /* * If the thread restricted its mask then apply that same * restriction to the new set, otherwise take it wholesale. */ if (CPU_CMP(&tdset->cs_mask, &parent->cs_mask) != 0) { - CPU_COPY(&tdset->cs_mask, mask); - CPU_AND(mask, &set->cs_mask); + CPU_AND(mask, &tdset->cs_mask, &set->cs_mask); } else CPU_COPY(&set->cs_mask, mask); /* * If the thread restricted the domain then we apply the * restriction to the new set but retain the policy. */ if (tdset->cs_domain != parent->cs_domain) { domainset_copy(tdset->cs_domain, domain); DOMAINSET_AND(&domain->ds_mask, &set->cs_domain->ds_mask); } else domainset_copy(set->cs_domain, domain); if (CPU_EMPTY(mask) || DOMAINSET_EMPTY(&domain->ds_mask)) return (EDEADLK); return (0); } static int cpuset_setproc_test_setthread(struct cpuset *tdset, struct cpuset *set) { struct domainset domain; cpuset_t mask; if (tdset->cs_id != CPUSET_INVALID) return (0); return cpuset_setproc_setthread_mask(tdset, set, &mask, &domain); } static int cpuset_setproc_setthread(struct cpuset *tdset, struct cpuset *set, struct cpuset **nsetp, struct setlist *freelist, struct domainlist *domainlist) { struct domainset domain; cpuset_t mask; int error; /* * If we're replacing on a thread that has not constrained the * original set we can simply accept the new set. */ if (tdset->cs_id != CPUSET_INVALID) { *nsetp = cpuset_ref(set); return (0); } error = cpuset_setproc_setthread_mask(tdset, set, &mask, &domain); if (error) return (error); return cpuset_shadow(set, nsetp, &mask, &domain, freelist, domainlist); } static int cpuset_setproc_newbase(struct thread *td, struct cpuset *set, struct cpuset *nroot, struct cpuset **nsetp, struct setlist *cpusets, struct domainlist *domainlist) { struct domainset ndomain; cpuset_t nmask; struct cpuset *pbase; int error; pbase = cpuset_getbase(td->td_cpuset); /* Copy process mask, then further apply the new root mask. */ - CPU_COPY(&pbase->cs_mask, &nmask); - CPU_AND(&nmask, &nroot->cs_mask); + CPU_AND(&nmask, &pbase->cs_mask, &nroot->cs_mask); domainset_copy(pbase->cs_domain, &ndomain); DOMAINSET_AND(&ndomain.ds_mask, &set->cs_domain->ds_mask); /* Policy is too restrictive, will not work. */ if (CPU_EMPTY(&nmask) || DOMAINSET_EMPTY(&ndomain.ds_mask)) return (EDEADLK); /* * Remove pbase from the freelist in advance, it'll be pushed to * cpuset_ids on success. We assume here that cpuset_create() will not * touch pbase on failure, and we just enqueue it back to the freelist * to remain in a consistent state. */ pbase = LIST_FIRST(cpusets); LIST_REMOVE(pbase, cs_link); error = cpuset_create(&pbase, set, &nmask); if (error != 0) { LIST_INSERT_HEAD(cpusets, pbase, cs_link); return (error); } /* Duplicates some work from above... oh well. */ pbase->cs_domain = domainset_shadow(set->cs_domain, &ndomain, domainlist); *nsetp = pbase; return (0); } /* * Handle four cases for updating an entire process. * * 1) Set is non-null and the process is not rebasing onto a new root. This * reparents all anonymous sets to the provided set and replaces all * non-anonymous td_cpusets with the provided set. * 2) Set is non-null and the process is rebasing onto a new root. This * creates a new base set if the process previously had its own base set, * then reparents all anonymous sets either to that set or the provided set * if one was not created. Non-anonymous sets are similarly replaced. * 3) Mask is non-null. This replaces or creates anonymous sets for every * thread with the existing base as a parent. * 4) domain is non-null. This creates anonymous sets for every thread * and replaces the domain set. * * This is overly complicated because we can't allocate while holding a * spinlock and spinlocks must be held while changing and examining thread * state. */ static int cpuset_setproc(pid_t pid, struct cpuset *set, cpuset_t *mask, struct domainset *domain, bool rebase) { struct setlist freelist; struct setlist droplist; struct domainlist domainlist; struct cpuset *base, *nset, *nroot, *tdroot; struct thread *td; struct proc *p; int needed; int nfree; int error; /* * The algorithm requires two passes due to locking considerations. * * 1) Lookup the process and acquire the locks in the required order. * 2) If enough cpusets have not been allocated release the locks and * allocate them. Loop. */ cpuset_freelist_init(&freelist, 1); domainset_freelist_init(&domainlist, 1); nfree = 1; LIST_INIT(&droplist); nfree = 0; base = set; nroot = NULL; if (set != NULL) nroot = cpuset_getroot(set); for (;;) { error = cpuset_which(CPU_WHICH_PID, pid, &p, &td, &nset); if (error) goto out; tdroot = cpuset_getroot(td->td_cpuset); needed = p->p_numthreads; if (set != NULL && rebase && tdroot != nroot) needed++; if (nfree >= needed) break; PROC_UNLOCK(p); if (nfree < needed) { cpuset_freelist_add(&freelist, needed - nfree); domainset_freelist_add(&domainlist, needed - nfree); nfree = needed; } } PROC_LOCK_ASSERT(p, MA_OWNED); /* * If we're changing roots and the root set is what has been specified * as the parent, then we'll check if the process was previously using * the root set and, if it wasn't, create a new base with the process's * mask applied to it. * * If the new root is incompatible with the existing mask, then we allow * the process to take on the new root if and only if they have * privilege to widen their mask anyways. Unprivileged processes get * rejected with EDEADLK. */ if (set != NULL && rebase && nroot != tdroot) { cpusetid_t base_id, root_id; root_id = td->td_ucred->cr_prison->pr_cpuset->cs_id; base_id = cpuset_getbase(td->td_cpuset)->cs_id; if (base_id != root_id) { error = cpuset_setproc_newbase(td, set, nroot, &base, &freelist, &domainlist); if (error == EDEADLK && priv_check(td, PRIV_SCHED_CPUSET) == 0) error = 0; if (error != 0) goto unlock_out; } } /* * Now that the appropriate locks are held and we have enough cpusets, * make sure the operation will succeed before applying changes. The * proc lock prevents td_cpuset from changing between calls. */ error = 0; FOREACH_THREAD_IN_PROC(p, td) { thread_lock(td); if (set != NULL) error = cpuset_setproc_test_setthread(td->td_cpuset, base); else error = cpuset_setproc_test_maskthread(td->td_cpuset, mask, domain); thread_unlock(td); if (error) goto unlock_out; } /* * Replace each thread's cpuset while using deferred release. We * must do this because the thread lock must be held while operating * on the thread and this limits the type of operations allowed. */ FOREACH_THREAD_IN_PROC(p, td) { thread_lock(td); if (set != NULL) error = cpuset_setproc_setthread(td->td_cpuset, base, &nset, &freelist, &domainlist); else error = cpuset_setproc_maskthread(td->td_cpuset, mask, domain, &nset, &freelist, &domainlist); if (error) { thread_unlock(td); break; } cpuset_rel_defer(&droplist, cpuset_update_thread(td, nset)); thread_unlock(td); } unlock_out: PROC_UNLOCK(p); out: if (base != NULL && base != set) cpuset_rel(base); while ((nset = LIST_FIRST(&droplist)) != NULL) cpuset_rel_complete(nset); cpuset_freelist_free(&freelist); domainset_freelist_free(&domainlist); return (error); } static int bitset_strprint(char *buf, size_t bufsiz, const struct bitset *set, int setlen) { size_t bytes; int i, once; char *p; once = 0; p = buf; for (i = 0; i < __bitset_words(setlen); i++) { if (once != 0) { if (bufsiz < 1) return (0); *p = ','; p++; bufsiz--; } else once = 1; if (bufsiz < sizeof(__STRING(ULONG_MAX))) return (0); bytes = snprintf(p, bufsiz, "%lx", set->__bits[i]); p += bytes; bufsiz -= bytes; } return (p - buf); } static int bitset_strscan(struct bitset *set, int setlen, const char *buf) { int i, ret; const char *p; BIT_ZERO(setlen, set); p = buf; for (i = 0; i < __bitset_words(setlen); i++) { if (*p == ',') { p++; continue; } ret = sscanf(p, "%lx", &set->__bits[i]); if (ret == 0 || ret == -1) break; while (isxdigit(*p)) p++; } return (p - buf); } /* * Return a string representing a valid layout for a cpuset_t object. * It expects an incoming buffer at least sized as CPUSETBUFSIZ. */ char * cpusetobj_strprint(char *buf, const cpuset_t *set) { bitset_strprint(buf, CPUSETBUFSIZ, (const struct bitset *)set, CPU_SETSIZE); return (buf); } /* * Build a valid cpuset_t object from a string representation. * It expects an incoming buffer at least sized as CPUSETBUFSIZ. */ int cpusetobj_strscan(cpuset_t *set, const char *buf) { char p; if (strlen(buf) > CPUSETBUFSIZ - 1) return (-1); p = buf[bitset_strscan((struct bitset *)set, CPU_SETSIZE, buf)]; if (p != '\0') return (-1); return (0); } /* * Handle a domainset specifier in the sysctl tree. A poiner to a pointer to * a domainset is in arg1. If the user specifies a valid domainset the * pointer is updated. * * Format is: * hex mask word 0,hex mask word 1,...:decimal policy:decimal preferred */ int sysctl_handle_domainset(SYSCTL_HANDLER_ARGS) { char buf[DOMAINSETBUFSIZ]; struct domainset *dset; struct domainset key; int policy, prefer, error; char *p; dset = *(struct domainset **)arg1; error = 0; if (dset != NULL) { p = buf + bitset_strprint(buf, DOMAINSETBUFSIZ, (const struct bitset *)&dset->ds_mask, DOMAINSET_SETSIZE); sprintf(p, ":%d:%d", dset->ds_policy, dset->ds_prefer); } else sprintf(buf, ""); error = sysctl_handle_string(oidp, buf, sizeof(buf), req); if (error != 0 || req->newptr == NULL) return (error); /* * Read in and validate the string. */ memset(&key, 0, sizeof(key)); p = &buf[bitset_strscan((struct bitset *)&key.ds_mask, DOMAINSET_SETSIZE, buf)]; if (p == buf) return (EINVAL); if (sscanf(p, ":%d:%d", &policy, &prefer) != 2) return (EINVAL); key.ds_policy = policy; key.ds_prefer = prefer; /* Domainset_create() validates the policy.*/ dset = domainset_create(&key); if (dset == NULL) return (EINVAL); *(struct domainset **)arg1 = dset; return (error); } /* * Apply an anonymous mask or a domain to a single thread. */ static int _cpuset_setthread(lwpid_t id, cpuset_t *mask, struct domainset *domain) { struct setlist cpusets; struct domainlist domainlist; struct cpuset *nset; struct cpuset *set; struct thread *td; struct proc *p; int error; cpuset_freelist_init(&cpusets, 1); domainset_freelist_init(&domainlist, domain != NULL); error = cpuset_which(CPU_WHICH_TID, id, &p, &td, &set); if (error) goto out; set = NULL; thread_lock(td); error = cpuset_shadow(td->td_cpuset, &nset, mask, domain, &cpusets, &domainlist); if (error == 0) set = cpuset_update_thread(td, nset); thread_unlock(td); PROC_UNLOCK(p); if (set) cpuset_rel(set); out: cpuset_freelist_free(&cpusets); domainset_freelist_free(&domainlist); return (error); } /* * Apply an anonymous mask to a single thread. */ int cpuset_setthread(lwpid_t id, cpuset_t *mask) { return _cpuset_setthread(id, mask, NULL); } /* * Apply new cpumask to the ithread. */ int cpuset_setithread(lwpid_t id, int cpu) { cpuset_t mask; CPU_ZERO(&mask); if (cpu == NOCPU) CPU_COPY(cpuset_root, &mask); else CPU_SET(cpu, &mask); return _cpuset_setthread(id, &mask, NULL); } /* * Initialize static domainsets after NUMA information is available. This is * called before memory allocators are initialized. */ void domainset_init(void) { struct domainset *dset; int i; dset = &domainset_firsttouch; DOMAINSET_COPY(&all_domains, &dset->ds_mask); dset->ds_policy = DOMAINSET_POLICY_FIRSTTOUCH; dset->ds_prefer = -1; _domainset_create(dset, NULL); dset = &domainset_interleave; DOMAINSET_COPY(&all_domains, &dset->ds_mask); dset->ds_policy = DOMAINSET_POLICY_INTERLEAVE; dset->ds_prefer = -1; _domainset_create(dset, NULL); dset = &domainset_roundrobin; DOMAINSET_COPY(&all_domains, &dset->ds_mask); dset->ds_policy = DOMAINSET_POLICY_ROUNDROBIN; dset->ds_prefer = -1; _domainset_create(dset, NULL); for (i = 0; i < vm_ndomains; i++) { dset = &domainset_fixed[i]; DOMAINSET_ZERO(&dset->ds_mask); DOMAINSET_SET(i, &dset->ds_mask); dset->ds_policy = DOMAINSET_POLICY_ROUNDROBIN; _domainset_create(dset, NULL); dset = &domainset_prefer[i]; DOMAINSET_COPY(&all_domains, &dset->ds_mask); dset->ds_policy = DOMAINSET_POLICY_PREFER; dset->ds_prefer = i; _domainset_create(dset, NULL); } } /* * Define the domainsets for cpuset 0, 1 and cpuset 2. */ void domainset_zero(void) { struct domainset *dset, *tmp; mtx_init(&cpuset_lock, "cpuset", NULL, MTX_SPIN | MTX_RECURSE); domainset0 = &domainset_firsttouch; curthread->td_domain.dr_policy = domainset0; domainset2 = &domainset_interleave; kernel_object->domain.dr_policy = domainset2; /* Remove empty domains from the global policies. */ LIST_FOREACH_SAFE(dset, &cpuset_domains, ds_link, tmp) if (domainset_empty_vm(dset)) LIST_REMOVE(dset, ds_link); } /* * Creates system-wide cpusets and the cpuset for thread0 including three * sets: * * 0 - The root set which should represent all valid processors in the * system. This set is immutable. * 1 - The default set which all processes are a member of until changed. * This allows an administrator to move all threads off of given cpus to * dedicate them to high priority tasks or save power etc. * 2 - The kernel set which allows restriction and policy to be applied only * to kernel threads and the kernel_object. */ struct cpuset * cpuset_thread0(void) { struct cpuset *set; int i; int error __unused; cpuset_zone = uma_zcreate("cpuset", sizeof(struct cpuset), NULL, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0); domainset_zone = uma_zcreate("domainset", sizeof(struct domainset), NULL, NULL, NULL, NULL, UMA_ALIGN_CACHE, 0); /* * Create the root system set (0) for the whole machine. Doesn't use * cpuset_create() due to NULL parent. */ set = uma_zalloc(cpuset_zone, M_WAITOK | M_ZERO); CPU_COPY(&all_cpus, &set->cs_mask); LIST_INIT(&set->cs_children); LIST_INSERT_HEAD(&cpuset_ids, set, cs_link); refcount_init(&set->cs_ref, 1); set->cs_flags = CPU_SET_ROOT | CPU_SET_RDONLY; set->cs_domain = domainset0; cpuset_zero = set; cpuset_root = &set->cs_mask; /* * Now derive a default (1), modifiable set from that to give out. */ set = uma_zalloc(cpuset_zone, M_WAITOK | M_ZERO); error = cpuset_init(set, cpuset_zero, NULL, NULL, 1); KASSERT(error == 0, ("Error creating default set: %d\n", error)); cpuset_default = set; /* * Create the kernel set (2). */ set = uma_zalloc(cpuset_zone, M_WAITOK | M_ZERO); error = cpuset_init(set, cpuset_zero, NULL, NULL, 2); KASSERT(error == 0, ("Error creating kernel set: %d\n", error)); set->cs_domain = domainset2; cpuset_kernel = set; /* * Initialize the unit allocator. 0 and 1 are allocated above. */ cpuset_unr = new_unrhdr(3, INT_MAX, NULL); /* * If MD code has not initialized per-domain cpusets, place all * CPUs in domain 0. */ for (i = 0; i < MAXMEMDOM; i++) if (!CPU_EMPTY(&cpuset_domain[i])) goto domains_set; CPU_COPY(&all_cpus, &cpuset_domain[0]); domains_set: return (cpuset_default); } void cpuset_kernthread(struct thread *td) { struct cpuset *set; thread_lock(td); set = td->td_cpuset; td->td_cpuset = cpuset_ref(cpuset_kernel); thread_unlock(td); cpuset_rel(set); } /* * Create a cpuset, which would be cpuset_create() but * mark the new 'set' as root. * * We are not going to reparent the td to it. Use cpuset_setproc_update_set() * for that. * * In case of no error, returns the set in *setp locked with a reference. */ int cpuset_create_root(struct prison *pr, struct cpuset **setp) { struct cpuset *set; int error; KASSERT(pr != NULL, ("[%s:%d] invalid pr", __func__, __LINE__)); KASSERT(setp != NULL, ("[%s:%d] invalid setp", __func__, __LINE__)); set = NULL; error = cpuset_create(&set, pr->pr_cpuset, &pr->pr_cpuset->cs_mask); if (error) return (error); KASSERT(set != NULL, ("[%s:%d] cpuset_create returned invalid data", __func__, __LINE__)); /* Mark the set as root. */ set->cs_flags |= CPU_SET_ROOT; *setp = set; return (0); } int cpuset_setproc_update_set(struct proc *p, struct cpuset *set) { int error; KASSERT(p != NULL, ("[%s:%d] invalid proc", __func__, __LINE__)); KASSERT(set != NULL, ("[%s:%d] invalid set", __func__, __LINE__)); cpuset_ref(set); error = cpuset_setproc(p->p_pid, set, NULL, NULL, true); if (error) return (error); cpuset_rel(set); return (0); } /* * In Capability mode, the only accesses that are permitted are to the current * thread and process' CPU and domain sets. */ static int cpuset_check_capabilities(struct thread *td, cpulevel_t level, cpuwhich_t which, id_t id) { if (IN_CAPABILITY_MODE(td)) { if (level != CPU_LEVEL_WHICH) return (ECAPMODE); if (which != CPU_WHICH_TID && which != CPU_WHICH_PID) return (ECAPMODE); if (id != -1 && !(which == CPU_WHICH_TID && id == td->td_tid) && !(which == CPU_WHICH_PID && id == td->td_proc->p_pid)) return (ECAPMODE); } return (0); } #ifndef _SYS_SYSPROTO_H_ struct cpuset_args { cpusetid_t *setid; }; #endif int sys_cpuset(struct thread *td, struct cpuset_args *uap) { struct cpuset *root; struct cpuset *set; int error; thread_lock(td); root = cpuset_refroot(td->td_cpuset); thread_unlock(td); set = NULL; error = cpuset_create(&set, root, &root->cs_mask); cpuset_rel(root); if (error) return (error); error = copyout(&set->cs_id, uap->setid, sizeof(set->cs_id)); if (error == 0) error = cpuset_setproc(-1, set, NULL, NULL, false); cpuset_rel(set); return (error); } #ifndef _SYS_SYSPROTO_H_ struct cpuset_setid_args { cpuwhich_t which; id_t id; cpusetid_t setid; }; #endif int sys_cpuset_setid(struct thread *td, struct cpuset_setid_args *uap) { return (kern_cpuset_setid(td, uap->which, uap->id, uap->setid)); } int kern_cpuset_setid(struct thread *td, cpuwhich_t which, id_t id, cpusetid_t setid) { struct cpuset *set; int error; /* * Presently we only support per-process sets. */ if (which != CPU_WHICH_PID) return (EINVAL); set = cpuset_lookup(setid, td); if (set == NULL) return (ESRCH); error = cpuset_setproc(id, set, NULL, NULL, false); cpuset_rel(set); return (error); } #ifndef _SYS_SYSPROTO_H_ struct cpuset_getid_args { cpulevel_t level; cpuwhich_t which; id_t id; cpusetid_t *setid; }; #endif int sys_cpuset_getid(struct thread *td, struct cpuset_getid_args *uap) { return (kern_cpuset_getid(td, uap->level, uap->which, uap->id, uap->setid)); } int kern_cpuset_getid(struct thread *td, cpulevel_t level, cpuwhich_t which, id_t id, cpusetid_t *setid) { struct cpuset *nset; struct cpuset *set; struct thread *ttd; struct proc *p; cpusetid_t tmpid; int error; if (level == CPU_LEVEL_WHICH && which != CPU_WHICH_CPUSET) return (EINVAL); error = cpuset_which(which, id, &p, &ttd, &set); if (error) return (error); switch (which) { case CPU_WHICH_TID: case CPU_WHICH_PID: thread_lock(ttd); set = cpuset_refbase(ttd->td_cpuset); thread_unlock(ttd); PROC_UNLOCK(p); break; case CPU_WHICH_CPUSET: case CPU_WHICH_JAIL: break; case CPU_WHICH_IRQ: case CPU_WHICH_DOMAIN: return (EINVAL); } switch (level) { case CPU_LEVEL_ROOT: nset = cpuset_refroot(set); cpuset_rel(set); set = nset; break; case CPU_LEVEL_CPUSET: break; case CPU_LEVEL_WHICH: break; } tmpid = set->cs_id; cpuset_rel(set); if (error == 0) error = copyout(&tmpid, setid, sizeof(tmpid)); return (error); } #ifndef _SYS_SYSPROTO_H_ struct cpuset_getaffinity_args { cpulevel_t level; cpuwhich_t which; id_t id; size_t cpusetsize; cpuset_t *mask; }; #endif int sys_cpuset_getaffinity(struct thread *td, struct cpuset_getaffinity_args *uap) { return (kern_cpuset_getaffinity(td, uap->level, uap->which, uap->id, uap->cpusetsize, uap->mask)); } int kern_cpuset_getaffinity(struct thread *td, cpulevel_t level, cpuwhich_t which, id_t id, size_t cpusetsize, cpuset_t *maskp) { struct thread *ttd; struct cpuset *nset; struct cpuset *set; struct proc *p; cpuset_t *mask; int error; size_t size; if (cpusetsize < sizeof(cpuset_t) || cpusetsize > CPU_MAXSIZE / NBBY) return (ERANGE); error = cpuset_check_capabilities(td, level, which, id); if (error != 0) return (error); size = cpusetsize; mask = malloc(size, M_TEMP, M_WAITOK | M_ZERO); error = cpuset_which(which, id, &p, &ttd, &set); if (error) goto out; switch (level) { case CPU_LEVEL_ROOT: case CPU_LEVEL_CPUSET: switch (which) { case CPU_WHICH_TID: case CPU_WHICH_PID: thread_lock(ttd); set = cpuset_ref(ttd->td_cpuset); thread_unlock(ttd); break; case CPU_WHICH_CPUSET: case CPU_WHICH_JAIL: break; case CPU_WHICH_IRQ: case CPU_WHICH_INTRHANDLER: case CPU_WHICH_ITHREAD: case CPU_WHICH_DOMAIN: error = EINVAL; goto out; } if (level == CPU_LEVEL_ROOT) nset = cpuset_refroot(set); else nset = cpuset_refbase(set); CPU_COPY(&nset->cs_mask, mask); cpuset_rel(nset); break; case CPU_LEVEL_WHICH: switch (which) { case CPU_WHICH_TID: thread_lock(ttd); CPU_COPY(&ttd->td_cpuset->cs_mask, mask); thread_unlock(ttd); break; case CPU_WHICH_PID: FOREACH_THREAD_IN_PROC(p, ttd) { thread_lock(ttd); - CPU_OR(mask, &ttd->td_cpuset->cs_mask); + CPU_OR(mask, mask, &ttd->td_cpuset->cs_mask); thread_unlock(ttd); } break; case CPU_WHICH_CPUSET: case CPU_WHICH_JAIL: CPU_COPY(&set->cs_mask, mask); break; case CPU_WHICH_IRQ: case CPU_WHICH_INTRHANDLER: case CPU_WHICH_ITHREAD: error = intr_getaffinity(id, which, mask); break; case CPU_WHICH_DOMAIN: if (id < 0 || id >= MAXMEMDOM) error = ESRCH; else CPU_COPY(&cpuset_domain[id], mask); break; } break; default: error = EINVAL; break; } if (set) cpuset_rel(set); if (p) PROC_UNLOCK(p); if (error == 0) error = copyout(mask, maskp, size); out: free(mask, M_TEMP); return (error); } #ifndef _SYS_SYSPROTO_H_ struct cpuset_setaffinity_args { cpulevel_t level; cpuwhich_t which; id_t id; size_t cpusetsize; const cpuset_t *mask; }; #endif int sys_cpuset_setaffinity(struct thread *td, struct cpuset_setaffinity_args *uap) { return (kern_cpuset_setaffinity(td, uap->level, uap->which, uap->id, uap->cpusetsize, uap->mask)); } int kern_cpuset_setaffinity(struct thread *td, cpulevel_t level, cpuwhich_t which, id_t id, size_t cpusetsize, const cpuset_t *maskp) { struct cpuset *nset; struct cpuset *set; struct thread *ttd; struct proc *p; cpuset_t *mask; int error; if (cpusetsize < sizeof(cpuset_t) || cpusetsize > CPU_MAXSIZE / NBBY) return (ERANGE); error = cpuset_check_capabilities(td, level, which, id); if (error != 0) return (error); mask = malloc(cpusetsize, M_TEMP, M_WAITOK | M_ZERO); error = copyin(maskp, mask, cpusetsize); if (error) goto out; /* * Verify that no high bits are set. */ if (cpusetsize > sizeof(cpuset_t)) { char *end; char *cp; end = cp = (char *)&mask->__bits; end += cpusetsize; cp += sizeof(cpuset_t); while (cp != end) if (*cp++ != 0) { error = EINVAL; goto out; } } if (CPU_EMPTY(mask)) { error = EDEADLK; goto out; } switch (level) { case CPU_LEVEL_ROOT: case CPU_LEVEL_CPUSET: error = cpuset_which(which, id, &p, &ttd, &set); if (error) break; switch (which) { case CPU_WHICH_TID: case CPU_WHICH_PID: thread_lock(ttd); set = cpuset_ref(ttd->td_cpuset); thread_unlock(ttd); PROC_UNLOCK(p); break; case CPU_WHICH_CPUSET: case CPU_WHICH_JAIL: break; case CPU_WHICH_IRQ: case CPU_WHICH_INTRHANDLER: case CPU_WHICH_ITHREAD: case CPU_WHICH_DOMAIN: error = EINVAL; goto out; } if (level == CPU_LEVEL_ROOT) nset = cpuset_refroot(set); else nset = cpuset_refbase(set); error = cpuset_modify(nset, mask); cpuset_rel(nset); cpuset_rel(set); break; case CPU_LEVEL_WHICH: switch (which) { case CPU_WHICH_TID: error = cpuset_setthread(id, mask); break; case CPU_WHICH_PID: error = cpuset_setproc(id, NULL, mask, NULL, false); break; case CPU_WHICH_CPUSET: case CPU_WHICH_JAIL: error = cpuset_which(which, id, &p, &ttd, &set); if (error == 0) { error = cpuset_modify(set, mask); cpuset_rel(set); } break; case CPU_WHICH_IRQ: case CPU_WHICH_INTRHANDLER: case CPU_WHICH_ITHREAD: error = intr_setaffinity(id, which, mask); break; default: error = EINVAL; break; } break; default: error = EINVAL; break; } out: free(mask, M_TEMP); return (error); } #ifndef _SYS_SYSPROTO_H_ struct cpuset_getdomain_args { cpulevel_t level; cpuwhich_t which; id_t id; size_t domainsetsize; domainset_t *mask; int *policy; }; #endif int sys_cpuset_getdomain(struct thread *td, struct cpuset_getdomain_args *uap) { return (kern_cpuset_getdomain(td, uap->level, uap->which, uap->id, uap->domainsetsize, uap->mask, uap->policy)); } int kern_cpuset_getdomain(struct thread *td, cpulevel_t level, cpuwhich_t which, id_t id, size_t domainsetsize, domainset_t *maskp, int *policyp) { struct domainset outset; struct thread *ttd; struct cpuset *nset; struct cpuset *set; struct domainset *dset; struct proc *p; domainset_t *mask; int error; if (domainsetsize < sizeof(domainset_t) || domainsetsize > DOMAINSET_MAXSIZE / NBBY) return (ERANGE); error = cpuset_check_capabilities(td, level, which, id); if (error != 0) return (error); mask = malloc(domainsetsize, M_TEMP, M_WAITOK | M_ZERO); bzero(&outset, sizeof(outset)); error = cpuset_which(which, id, &p, &ttd, &set); if (error) goto out; switch (level) { case CPU_LEVEL_ROOT: case CPU_LEVEL_CPUSET: switch (which) { case CPU_WHICH_TID: case CPU_WHICH_PID: thread_lock(ttd); set = cpuset_ref(ttd->td_cpuset); thread_unlock(ttd); break; case CPU_WHICH_CPUSET: case CPU_WHICH_JAIL: break; case CPU_WHICH_IRQ: case CPU_WHICH_INTRHANDLER: case CPU_WHICH_ITHREAD: case CPU_WHICH_DOMAIN: error = EINVAL; goto out; } if (level == CPU_LEVEL_ROOT) nset = cpuset_refroot(set); else nset = cpuset_refbase(set); domainset_copy(nset->cs_domain, &outset); cpuset_rel(nset); break; case CPU_LEVEL_WHICH: switch (which) { case CPU_WHICH_TID: thread_lock(ttd); domainset_copy(ttd->td_cpuset->cs_domain, &outset); thread_unlock(ttd); break; case CPU_WHICH_PID: FOREACH_THREAD_IN_PROC(p, ttd) { thread_lock(ttd); dset = ttd->td_cpuset->cs_domain; /* Show all domains in the proc. */ DOMAINSET_OR(&outset.ds_mask, &dset->ds_mask); /* Last policy wins. */ outset.ds_policy = dset->ds_policy; outset.ds_prefer = dset->ds_prefer; thread_unlock(ttd); } break; case CPU_WHICH_CPUSET: case CPU_WHICH_JAIL: domainset_copy(set->cs_domain, &outset); break; case CPU_WHICH_IRQ: case CPU_WHICH_INTRHANDLER: case CPU_WHICH_ITHREAD: case CPU_WHICH_DOMAIN: error = EINVAL; break; } break; default: error = EINVAL; break; } if (set) cpuset_rel(set); if (p) PROC_UNLOCK(p); /* * Translate prefer into a set containing only the preferred domain, * not the entire fallback set. */ if (outset.ds_policy == DOMAINSET_POLICY_PREFER) { DOMAINSET_ZERO(&outset.ds_mask); DOMAINSET_SET(outset.ds_prefer, &outset.ds_mask); } DOMAINSET_COPY(&outset.ds_mask, mask); if (error == 0) error = copyout(mask, maskp, domainsetsize); if (error == 0) if (suword32(policyp, outset.ds_policy) != 0) error = EFAULT; out: free(mask, M_TEMP); return (error); } #ifndef _SYS_SYSPROTO_H_ struct cpuset_setdomain_args { cpulevel_t level; cpuwhich_t which; id_t id; size_t domainsetsize; domainset_t *mask; int policy; }; #endif int sys_cpuset_setdomain(struct thread *td, struct cpuset_setdomain_args *uap) { return (kern_cpuset_setdomain(td, uap->level, uap->which, uap->id, uap->domainsetsize, uap->mask, uap->policy)); } int kern_cpuset_setdomain(struct thread *td, cpulevel_t level, cpuwhich_t which, id_t id, size_t domainsetsize, const domainset_t *maskp, int policy) { struct cpuset *nset; struct cpuset *set; struct thread *ttd; struct proc *p; struct domainset domain; domainset_t *mask; int error; if (domainsetsize < sizeof(domainset_t) || domainsetsize > DOMAINSET_MAXSIZE / NBBY) return (ERANGE); if (policy <= DOMAINSET_POLICY_INVALID || policy > DOMAINSET_POLICY_MAX) return (EINVAL); error = cpuset_check_capabilities(td, level, which, id); if (error != 0) return (error); memset(&domain, 0, sizeof(domain)); mask = malloc(domainsetsize, M_TEMP, M_WAITOK | M_ZERO); error = copyin(maskp, mask, domainsetsize); if (error) goto out; /* * Verify that no high bits are set. */ if (domainsetsize > sizeof(domainset_t)) { char *end; char *cp; end = cp = (char *)&mask->__bits; end += domainsetsize; cp += sizeof(domainset_t); while (cp != end) if (*cp++ != 0) { error = EINVAL; goto out; } } if (DOMAINSET_EMPTY(mask)) { error = EDEADLK; goto out; } DOMAINSET_COPY(mask, &domain.ds_mask); domain.ds_policy = policy; /* * Sanitize the provided mask. */ if (!DOMAINSET_SUBSET(&all_domains, &domain.ds_mask)) { error = EINVAL; goto out; } /* Translate preferred policy into a mask and fallback. */ if (policy == DOMAINSET_POLICY_PREFER) { /* Only support a single preferred domain. */ if (DOMAINSET_COUNT(&domain.ds_mask) != 1) { error = EINVAL; goto out; } domain.ds_prefer = DOMAINSET_FFS(&domain.ds_mask) - 1; /* This will be constrained by domainset_shadow(). */ DOMAINSET_COPY(&all_domains, &domain.ds_mask); } /* * When given an impossible policy, fall back to interleaving * across all domains. */ if (domainset_empty_vm(&domain)) domainset_copy(domainset2, &domain); switch (level) { case CPU_LEVEL_ROOT: case CPU_LEVEL_CPUSET: error = cpuset_which(which, id, &p, &ttd, &set); if (error) break; switch (which) { case CPU_WHICH_TID: case CPU_WHICH_PID: thread_lock(ttd); set = cpuset_ref(ttd->td_cpuset); thread_unlock(ttd); PROC_UNLOCK(p); break; case CPU_WHICH_CPUSET: case CPU_WHICH_JAIL: break; case CPU_WHICH_IRQ: case CPU_WHICH_INTRHANDLER: case CPU_WHICH_ITHREAD: case CPU_WHICH_DOMAIN: error = EINVAL; goto out; } if (level == CPU_LEVEL_ROOT) nset = cpuset_refroot(set); else nset = cpuset_refbase(set); error = cpuset_modify_domain(nset, &domain); cpuset_rel(nset); cpuset_rel(set); break; case CPU_LEVEL_WHICH: switch (which) { case CPU_WHICH_TID: error = _cpuset_setthread(id, NULL, &domain); break; case CPU_WHICH_PID: error = cpuset_setproc(id, NULL, NULL, &domain, false); break; case CPU_WHICH_CPUSET: case CPU_WHICH_JAIL: error = cpuset_which(which, id, &p, &ttd, &set); if (error == 0) { error = cpuset_modify_domain(set, &domain); cpuset_rel(set); } break; case CPU_WHICH_IRQ: case CPU_WHICH_INTRHANDLER: case CPU_WHICH_ITHREAD: default: error = EINVAL; break; } break; default: error = EINVAL; break; } out: free(mask, M_TEMP); return (error); } #ifdef DDB static void ddb_display_bitset(const struct bitset *set, int size) { int bit, once; for (once = 0, bit = 0; bit < size; bit++) { if (CPU_ISSET(bit, set)) { if (once == 0) { db_printf("%d", bit); once = 1; } else db_printf(",%d", bit); } } if (once == 0) db_printf(""); } void ddb_display_cpuset(const cpuset_t *set) { ddb_display_bitset((const struct bitset *)set, CPU_SETSIZE); } static void ddb_display_domainset(const domainset_t *set) { ddb_display_bitset((const struct bitset *)set, DOMAINSET_SETSIZE); } DB_SHOW_COMMAND(cpusets, db_show_cpusets) { struct cpuset *set; LIST_FOREACH(set, &cpuset_ids, cs_link) { db_printf("set=%p id=%-6u ref=%-6d flags=0x%04x parent id=%d\n", set, set->cs_id, refcount_load(&set->cs_ref), set->cs_flags, (set->cs_parent != NULL) ? set->cs_parent->cs_id : 0); db_printf(" cpu mask="); ddb_display_cpuset(&set->cs_mask); db_printf("\n"); db_printf(" domain policy %d prefer %d mask=", set->cs_domain->ds_policy, set->cs_domain->ds_prefer); ddb_display_domainset(&set->cs_domain->ds_mask); db_printf("\n"); if (db_pager_quit) break; } } DB_SHOW_COMMAND(domainsets, db_show_domainsets) { struct domainset *set; LIST_FOREACH(set, &cpuset_domains, ds_link) { db_printf("set=%p policy %d prefer %d cnt %d\n", set, set->ds_policy, set->ds_prefer, set->ds_cnt); db_printf(" mask ="); ddb_display_domainset(&set->ds_mask); db_printf("\n"); } } #endif /* DDB */ diff --git a/sys/kern/kern_rmlock.c b/sys/kern/kern_rmlock.c index f9b5559a648c..5e64f79e3836 100644 --- a/sys/kern/kern_rmlock.c +++ b/sys/kern/kern_rmlock.c @@ -1,1259 +1,1258 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 2007 Stephan Uphoff * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the author nor the names of any co-contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ /* * Machine independent bits of reader/writer lock implementation. */ #include __FBSDID("$FreeBSD$"); #include "opt_ddb.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef DDB #include #endif /* * A cookie to mark destroyed rmlocks. This is stored in the head of * rm_activeReaders. */ #define RM_DESTROYED ((void *)0xdead) #define rm_destroyed(rm) \ (LIST_FIRST(&(rm)->rm_activeReaders) == RM_DESTROYED) #define RMPF_ONQUEUE 1 #define RMPF_SIGNAL 2 #ifndef INVARIANTS #define _rm_assert(c, what, file, line) #endif static void assert_rm(const struct lock_object *lock, int what); #ifdef DDB static void db_show_rm(const struct lock_object *lock); #endif static void lock_rm(struct lock_object *lock, uintptr_t how); #ifdef KDTRACE_HOOKS static int owner_rm(const struct lock_object *lock, struct thread **owner); #endif static uintptr_t unlock_rm(struct lock_object *lock); struct lock_class lock_class_rm = { .lc_name = "rm", .lc_flags = LC_SLEEPLOCK | LC_RECURSABLE, .lc_assert = assert_rm, #ifdef DDB .lc_ddb_show = db_show_rm, #endif .lc_lock = lock_rm, .lc_unlock = unlock_rm, #ifdef KDTRACE_HOOKS .lc_owner = owner_rm, #endif }; struct lock_class lock_class_rm_sleepable = { .lc_name = "sleepable rm", .lc_flags = LC_SLEEPLOCK | LC_SLEEPABLE | LC_RECURSABLE, .lc_assert = assert_rm, #ifdef DDB .lc_ddb_show = db_show_rm, #endif .lc_lock = lock_rm, .lc_unlock = unlock_rm, #ifdef KDTRACE_HOOKS .lc_owner = owner_rm, #endif }; static void assert_rm(const struct lock_object *lock, int what) { rm_assert((const struct rmlock *)lock, what); } static void lock_rm(struct lock_object *lock, uintptr_t how) { struct rmlock *rm; struct rm_priotracker *tracker; rm = (struct rmlock *)lock; if (how == 0) rm_wlock(rm); else { tracker = (struct rm_priotracker *)how; rm_rlock(rm, tracker); } } static uintptr_t unlock_rm(struct lock_object *lock) { struct thread *td; struct pcpu *pc; struct rmlock *rm; struct rm_queue *queue; struct rm_priotracker *tracker; uintptr_t how; rm = (struct rmlock *)lock; tracker = NULL; how = 0; rm_assert(rm, RA_LOCKED | RA_NOTRECURSED); if (rm_wowned(rm)) rm_wunlock(rm); else { /* * Find the right rm_priotracker structure for curthread. * The guarantee about its uniqueness is given by the fact * we already asserted the lock wasn't recursively acquired. */ critical_enter(); td = curthread; pc = get_pcpu(); for (queue = pc->pc_rm_queue.rmq_next; queue != &pc->pc_rm_queue; queue = queue->rmq_next) { tracker = (struct rm_priotracker *)queue; if ((tracker->rmp_rmlock == rm) && (tracker->rmp_thread == td)) { how = (uintptr_t)tracker; break; } } KASSERT(tracker != NULL, ("rm_priotracker is non-NULL when lock held in read mode")); critical_exit(); rm_runlock(rm, tracker); } return (how); } #ifdef KDTRACE_HOOKS static int owner_rm(const struct lock_object *lock, struct thread **owner) { const struct rmlock *rm; struct lock_class *lc; rm = (const struct rmlock *)lock; lc = LOCK_CLASS(&rm->rm_wlock_object); return (lc->lc_owner(&rm->rm_wlock_object, owner)); } #endif static struct mtx rm_spinlock; MTX_SYSINIT(rm_spinlock, &rm_spinlock, "rm_spinlock", MTX_SPIN); /* * Add or remove tracker from per-cpu list. * * The per-cpu list can be traversed at any time in forward direction from an * interrupt on the *local* cpu. */ static void inline rm_tracker_add(struct pcpu *pc, struct rm_priotracker *tracker) { struct rm_queue *next; /* Initialize all tracker pointers */ tracker->rmp_cpuQueue.rmq_prev = &pc->pc_rm_queue; next = pc->pc_rm_queue.rmq_next; tracker->rmp_cpuQueue.rmq_next = next; /* rmq_prev is not used during froward traversal. */ next->rmq_prev = &tracker->rmp_cpuQueue; /* Update pointer to first element. */ pc->pc_rm_queue.rmq_next = &tracker->rmp_cpuQueue; } /* * Return a count of the number of trackers the thread 'td' already * has on this CPU for the lock 'rm'. */ static int rm_trackers_present(const struct pcpu *pc, const struct rmlock *rm, const struct thread *td) { struct rm_queue *queue; struct rm_priotracker *tracker; int count; count = 0; for (queue = pc->pc_rm_queue.rmq_next; queue != &pc->pc_rm_queue; queue = queue->rmq_next) { tracker = (struct rm_priotracker *)queue; if ((tracker->rmp_rmlock == rm) && (tracker->rmp_thread == td)) count++; } return (count); } static void inline rm_tracker_remove(struct pcpu *pc, struct rm_priotracker *tracker) { struct rm_queue *next, *prev; next = tracker->rmp_cpuQueue.rmq_next; prev = tracker->rmp_cpuQueue.rmq_prev; /* Not used during forward traversal. */ next->rmq_prev = prev; /* Remove from list. */ prev->rmq_next = next; } static void rm_cleanIPI(void *arg) { struct pcpu *pc; struct rmlock *rm = arg; struct rm_priotracker *tracker; struct rm_queue *queue; pc = get_pcpu(); for (queue = pc->pc_rm_queue.rmq_next; queue != &pc->pc_rm_queue; queue = queue->rmq_next) { tracker = (struct rm_priotracker *)queue; if (tracker->rmp_rmlock == rm && tracker->rmp_flags == 0) { tracker->rmp_flags = RMPF_ONQUEUE; mtx_lock_spin(&rm_spinlock); LIST_INSERT_HEAD(&rm->rm_activeReaders, tracker, rmp_qentry); mtx_unlock_spin(&rm_spinlock); } } } void rm_init_flags(struct rmlock *rm, const char *name, int opts) { struct lock_class *lc; int liflags, xflags; liflags = 0; if (!(opts & RM_NOWITNESS)) liflags |= LO_WITNESS; if (opts & RM_RECURSE) liflags |= LO_RECURSABLE; if (opts & RM_NEW) liflags |= LO_NEW; if (opts & RM_DUPOK) liflags |= LO_DUPOK; rm->rm_writecpus = all_cpus; LIST_INIT(&rm->rm_activeReaders); if (opts & RM_SLEEPABLE) { liflags |= LO_SLEEPABLE; lc = &lock_class_rm_sleepable; xflags = (opts & RM_NEW ? SX_NEW : 0); sx_init_flags(&rm->rm_lock_sx, "rmlock_sx", xflags | SX_NOWITNESS); } else { lc = &lock_class_rm; xflags = (opts & RM_NEW ? MTX_NEW : 0); mtx_init(&rm->rm_lock_mtx, name, "rmlock_mtx", xflags | MTX_NOWITNESS); } lock_init(&rm->lock_object, lc, name, NULL, liflags); } void rm_init(struct rmlock *rm, const char *name) { rm_init_flags(rm, name, 0); } void rm_destroy(struct rmlock *rm) { rm_assert(rm, RA_UNLOCKED); LIST_FIRST(&rm->rm_activeReaders) = RM_DESTROYED; if (rm->lock_object.lo_flags & LO_SLEEPABLE) sx_destroy(&rm->rm_lock_sx); else mtx_destroy(&rm->rm_lock_mtx); lock_destroy(&rm->lock_object); } int rm_wowned(const struct rmlock *rm) { if (rm->lock_object.lo_flags & LO_SLEEPABLE) return (sx_xlocked(&rm->rm_lock_sx)); else return (mtx_owned(&rm->rm_lock_mtx)); } void rm_sysinit(void *arg) { struct rm_args *args; args = arg; rm_init_flags(args->ra_rm, args->ra_desc, args->ra_flags); } static __noinline int _rm_rlock_hard(struct rmlock *rm, struct rm_priotracker *tracker, int trylock) { struct pcpu *pc; critical_enter(); pc = get_pcpu(); /* Check if we just need to do a proper critical_exit. */ if (!CPU_ISSET(pc->pc_cpuid, &rm->rm_writecpus)) { critical_exit(); return (1); } /* Remove our tracker from the per-cpu list. */ rm_tracker_remove(pc, tracker); /* * Check to see if the IPI granted us the lock after all. The load of * rmp_flags must happen after the tracker is removed from the list. */ atomic_interrupt_fence(); if (tracker->rmp_flags) { /* Just add back tracker - we hold the lock. */ rm_tracker_add(pc, tracker); critical_exit(); return (1); } /* * We allow readers to acquire a lock even if a writer is blocked if * the lock is recursive and the reader already holds the lock. */ if ((rm->lock_object.lo_flags & LO_RECURSABLE) != 0) { /* * Just grant the lock if this thread already has a tracker * for this lock on the per-cpu queue. */ if (rm_trackers_present(pc, rm, curthread) != 0) { mtx_lock_spin(&rm_spinlock); LIST_INSERT_HEAD(&rm->rm_activeReaders, tracker, rmp_qentry); tracker->rmp_flags = RMPF_ONQUEUE; mtx_unlock_spin(&rm_spinlock); rm_tracker_add(pc, tracker); critical_exit(); return (1); } } sched_unpin(); critical_exit(); if (trylock) { if (rm->lock_object.lo_flags & LO_SLEEPABLE) { if (!sx_try_xlock(&rm->rm_lock_sx)) return (0); } else { if (!mtx_trylock(&rm->rm_lock_mtx)) return (0); } } else { if (rm->lock_object.lo_flags & LO_SLEEPABLE) { THREAD_SLEEPING_OK(); sx_xlock(&rm->rm_lock_sx); THREAD_NO_SLEEPING(); } else mtx_lock(&rm->rm_lock_mtx); } critical_enter(); pc = get_pcpu(); CPU_CLR(pc->pc_cpuid, &rm->rm_writecpus); rm_tracker_add(pc, tracker); sched_pin(); critical_exit(); if (rm->lock_object.lo_flags & LO_SLEEPABLE) sx_xunlock(&rm->rm_lock_sx); else mtx_unlock(&rm->rm_lock_mtx); return (1); } int _rm_rlock(struct rmlock *rm, struct rm_priotracker *tracker, int trylock) { struct thread *td = curthread; struct pcpu *pc; if (SCHEDULER_STOPPED()) return (1); tracker->rmp_flags = 0; tracker->rmp_thread = td; tracker->rmp_rmlock = rm; if (rm->lock_object.lo_flags & LO_SLEEPABLE) THREAD_NO_SLEEPING(); td->td_critnest++; /* critical_enter(); */ atomic_interrupt_fence(); pc = cpuid_to_pcpu[td->td_oncpu]; /* pcpu_find(td->td_oncpu); */ rm_tracker_add(pc, tracker); sched_pin(); atomic_interrupt_fence(); td->td_critnest--; /* * Fast path to combine two common conditions into a single * conditional jump. */ if (__predict_true(0 == (td->td_owepreempt | CPU_ISSET(pc->pc_cpuid, &rm->rm_writecpus)))) return (1); /* We do not have a read token and need to acquire one. */ return _rm_rlock_hard(rm, tracker, trylock); } static __noinline void _rm_unlock_hard(struct thread *td,struct rm_priotracker *tracker) { if (td->td_owepreempt) { td->td_critnest++; critical_exit(); } if (!tracker->rmp_flags) return; mtx_lock_spin(&rm_spinlock); LIST_REMOVE(tracker, rmp_qentry); if (tracker->rmp_flags & RMPF_SIGNAL) { struct rmlock *rm; struct turnstile *ts; rm = tracker->rmp_rmlock; turnstile_chain_lock(&rm->lock_object); mtx_unlock_spin(&rm_spinlock); ts = turnstile_lookup(&rm->lock_object); turnstile_signal(ts, TS_EXCLUSIVE_QUEUE); turnstile_unpend(ts); turnstile_chain_unlock(&rm->lock_object); } else mtx_unlock_spin(&rm_spinlock); } void _rm_runlock(struct rmlock *rm, struct rm_priotracker *tracker) { struct pcpu *pc; struct thread *td = tracker->rmp_thread; if (SCHEDULER_STOPPED()) return; td->td_critnest++; /* critical_enter(); */ pc = cpuid_to_pcpu[td->td_oncpu]; /* pcpu_find(td->td_oncpu); */ rm_tracker_remove(pc, tracker); td->td_critnest--; sched_unpin(); if (rm->lock_object.lo_flags & LO_SLEEPABLE) THREAD_SLEEPING_OK(); if (__predict_true(0 == (td->td_owepreempt | tracker->rmp_flags))) return; _rm_unlock_hard(td, tracker); } void _rm_wlock(struct rmlock *rm) { struct rm_priotracker *prio; struct turnstile *ts; cpuset_t readcpus; if (SCHEDULER_STOPPED()) return; if (rm->lock_object.lo_flags & LO_SLEEPABLE) sx_xlock(&rm->rm_lock_sx); else mtx_lock(&rm->rm_lock_mtx); if (CPU_CMP(&rm->rm_writecpus, &all_cpus)) { /* Get all read tokens back */ - readcpus = all_cpus; - CPU_ANDNOT(&readcpus, &rm->rm_writecpus); + CPU_ANDNOT(&readcpus, &all_cpus, &rm->rm_writecpus); rm->rm_writecpus = all_cpus; /* * Assumes rm->rm_writecpus update is visible on other CPUs * before rm_cleanIPI is called. */ #ifdef SMP smp_rendezvous_cpus(readcpus, smp_no_rendezvous_barrier, rm_cleanIPI, smp_no_rendezvous_barrier, rm); #else rm_cleanIPI(rm); #endif mtx_lock_spin(&rm_spinlock); while ((prio = LIST_FIRST(&rm->rm_activeReaders)) != NULL) { ts = turnstile_trywait(&rm->lock_object); prio->rmp_flags = RMPF_ONQUEUE | RMPF_SIGNAL; mtx_unlock_spin(&rm_spinlock); turnstile_wait(ts, prio->rmp_thread, TS_EXCLUSIVE_QUEUE); mtx_lock_spin(&rm_spinlock); } mtx_unlock_spin(&rm_spinlock); } } void _rm_wunlock(struct rmlock *rm) { if (rm->lock_object.lo_flags & LO_SLEEPABLE) sx_xunlock(&rm->rm_lock_sx); else mtx_unlock(&rm->rm_lock_mtx); } #if LOCK_DEBUG > 0 void _rm_wlock_debug(struct rmlock *rm, const char *file, int line) { if (SCHEDULER_STOPPED()) return; KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread), ("rm_wlock() by idle thread %p on rmlock %s @ %s:%d", curthread, rm->lock_object.lo_name, file, line)); KASSERT(!rm_destroyed(rm), ("rm_wlock() of destroyed rmlock @ %s:%d", file, line)); _rm_assert(rm, RA_UNLOCKED, file, line); WITNESS_CHECKORDER(&rm->lock_object, LOP_NEWORDER | LOP_EXCLUSIVE, file, line, NULL); _rm_wlock(rm); LOCK_LOG_LOCK("RMWLOCK", &rm->lock_object, 0, 0, file, line); WITNESS_LOCK(&rm->lock_object, LOP_EXCLUSIVE, file, line); TD_LOCKS_INC(curthread); } void _rm_wunlock_debug(struct rmlock *rm, const char *file, int line) { if (SCHEDULER_STOPPED()) return; KASSERT(!rm_destroyed(rm), ("rm_wunlock() of destroyed rmlock @ %s:%d", file, line)); _rm_assert(rm, RA_WLOCKED, file, line); WITNESS_UNLOCK(&rm->lock_object, LOP_EXCLUSIVE, file, line); LOCK_LOG_LOCK("RMWUNLOCK", &rm->lock_object, 0, 0, file, line); _rm_wunlock(rm); TD_LOCKS_DEC(curthread); } int _rm_rlock_debug(struct rmlock *rm, struct rm_priotracker *tracker, int trylock, const char *file, int line) { if (SCHEDULER_STOPPED()) return (1); #ifdef INVARIANTS if (!(rm->lock_object.lo_flags & LO_RECURSABLE) && !trylock) { critical_enter(); KASSERT(rm_trackers_present(get_pcpu(), rm, curthread) == 0, ("rm_rlock: recursed on non-recursive rmlock %s @ %s:%d\n", rm->lock_object.lo_name, file, line)); critical_exit(); } #endif KASSERT(kdb_active != 0 || !TD_IS_IDLETHREAD(curthread), ("rm_rlock() by idle thread %p on rmlock %s @ %s:%d", curthread, rm->lock_object.lo_name, file, line)); KASSERT(!rm_destroyed(rm), ("rm_rlock() of destroyed rmlock @ %s:%d", file, line)); if (!trylock) { KASSERT(!rm_wowned(rm), ("rm_rlock: wlock already held for %s @ %s:%d", rm->lock_object.lo_name, file, line)); WITNESS_CHECKORDER(&rm->lock_object, LOP_NEWORDER | LOP_NOSLEEP, file, line, NULL); } if (_rm_rlock(rm, tracker, trylock)) { if (trylock) LOCK_LOG_TRY("RMRLOCK", &rm->lock_object, 0, 1, file, line); else LOCK_LOG_LOCK("RMRLOCK", &rm->lock_object, 0, 0, file, line); WITNESS_LOCK(&rm->lock_object, LOP_NOSLEEP, file, line); TD_LOCKS_INC(curthread); return (1); } else if (trylock) LOCK_LOG_TRY("RMRLOCK", &rm->lock_object, 0, 0, file, line); return (0); } void _rm_runlock_debug(struct rmlock *rm, struct rm_priotracker *tracker, const char *file, int line) { if (SCHEDULER_STOPPED()) return; KASSERT(!rm_destroyed(rm), ("rm_runlock() of destroyed rmlock @ %s:%d", file, line)); _rm_assert(rm, RA_RLOCKED, file, line); WITNESS_UNLOCK(&rm->lock_object, 0, file, line); LOCK_LOG_LOCK("RMRUNLOCK", &rm->lock_object, 0, 0, file, line); _rm_runlock(rm, tracker); TD_LOCKS_DEC(curthread); } #else /* * Just strip out file and line arguments if no lock debugging is enabled in * the kernel - we are called from a kernel module. */ void _rm_wlock_debug(struct rmlock *rm, const char *file, int line) { _rm_wlock(rm); } void _rm_wunlock_debug(struct rmlock *rm, const char *file, int line) { _rm_wunlock(rm); } int _rm_rlock_debug(struct rmlock *rm, struct rm_priotracker *tracker, int trylock, const char *file, int line) { return _rm_rlock(rm, tracker, trylock); } void _rm_runlock_debug(struct rmlock *rm, struct rm_priotracker *tracker, const char *file, int line) { _rm_runlock(rm, tracker); } #endif #ifdef INVARIANT_SUPPORT #ifndef INVARIANTS #undef _rm_assert #endif /* * Note that this does not need to use witness_assert() for read lock * assertions since an exact count of read locks held by this thread * is computable. */ void _rm_assert(const struct rmlock *rm, int what, const char *file, int line) { int count; if (SCHEDULER_STOPPED()) return; switch (what) { case RA_LOCKED: case RA_LOCKED | RA_RECURSED: case RA_LOCKED | RA_NOTRECURSED: case RA_RLOCKED: case RA_RLOCKED | RA_RECURSED: case RA_RLOCKED | RA_NOTRECURSED: /* * Handle the write-locked case. Unlike other * primitives, writers can never recurse. */ if (rm_wowned(rm)) { if (what & RA_RLOCKED) panic("Lock %s exclusively locked @ %s:%d\n", rm->lock_object.lo_name, file, line); if (what & RA_RECURSED) panic("Lock %s not recursed @ %s:%d\n", rm->lock_object.lo_name, file, line); break; } critical_enter(); count = rm_trackers_present(get_pcpu(), rm, curthread); critical_exit(); if (count == 0) panic("Lock %s not %slocked @ %s:%d\n", rm->lock_object.lo_name, (what & RA_RLOCKED) ? "read " : "", file, line); if (count > 1) { if (what & RA_NOTRECURSED) panic("Lock %s recursed @ %s:%d\n", rm->lock_object.lo_name, file, line); } else if (what & RA_RECURSED) panic("Lock %s not recursed @ %s:%d\n", rm->lock_object.lo_name, file, line); break; case RA_WLOCKED: if (!rm_wowned(rm)) panic("Lock %s not exclusively locked @ %s:%d\n", rm->lock_object.lo_name, file, line); break; case RA_UNLOCKED: if (rm_wowned(rm)) panic("Lock %s exclusively locked @ %s:%d\n", rm->lock_object.lo_name, file, line); critical_enter(); count = rm_trackers_present(get_pcpu(), rm, curthread); critical_exit(); if (count != 0) panic("Lock %s read locked @ %s:%d\n", rm->lock_object.lo_name, file, line); break; default: panic("Unknown rm lock assertion: %d @ %s:%d", what, file, line); } } #endif /* INVARIANT_SUPPORT */ #ifdef DDB static void print_tracker(struct rm_priotracker *tr) { struct thread *td; td = tr->rmp_thread; db_printf(" thread %p (tid %d, pid %d, \"%s\") {", td, td->td_tid, td->td_proc->p_pid, td->td_name); if (tr->rmp_flags & RMPF_ONQUEUE) { db_printf("ONQUEUE"); if (tr->rmp_flags & RMPF_SIGNAL) db_printf(",SIGNAL"); } else db_printf("0"); db_printf("}\n"); } static void db_show_rm(const struct lock_object *lock) { struct rm_priotracker *tr; struct rm_queue *queue; const struct rmlock *rm; struct lock_class *lc; struct pcpu *pc; rm = (const struct rmlock *)lock; db_printf(" writecpus: "); ddb_display_cpuset(__DEQUALIFY(const cpuset_t *, &rm->rm_writecpus)); db_printf("\n"); db_printf(" per-CPU readers:\n"); STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) for (queue = pc->pc_rm_queue.rmq_next; queue != &pc->pc_rm_queue; queue = queue->rmq_next) { tr = (struct rm_priotracker *)queue; if (tr->rmp_rmlock == rm) print_tracker(tr); } db_printf(" active readers:\n"); LIST_FOREACH(tr, &rm->rm_activeReaders, rmp_qentry) print_tracker(tr); lc = LOCK_CLASS(&rm->rm_wlock_object); db_printf("Backing write-lock (%s):\n", lc->lc_name); lc->lc_ddb_show(&rm->rm_wlock_object); } #endif /* * Read-mostly sleepable locks. * * These primitives allow both readers and writers to sleep. However, neither * readers nor writers are tracked and subsequently there is no priority * propagation. * * They are intended to be only used when write-locking is almost never needed * (e.g., they can guard against unloading a kernel module) while read-locking * happens all the time. * * Concurrent writers take turns taking the lock while going off cpu. If this is * of concern for your usecase, this is not the right primitive. * * Neither rms_rlock nor rms_runlock use thread fences. Instead interrupt * fences are inserted to ensure ordering with the code executed in the IPI * handler. * * No attempt is made to track which CPUs read locked at least once, * consequently write locking sends IPIs to all of them. This will become a * problem at some point. The easiest way to lessen it is to provide a bitmap. */ #define RMS_NOOWNER ((void *)0x1) #define RMS_TRANSIENT ((void *)0x2) #define RMS_FLAGMASK 0xf struct rmslock_pcpu { int influx; int readers; }; _Static_assert(sizeof(struct rmslock_pcpu) == 8, "bad size"); /* * Internal routines */ static struct rmslock_pcpu * rms_int_pcpu(struct rmslock *rms) { CRITICAL_ASSERT(curthread); return (zpcpu_get(rms->pcpu)); } static struct rmslock_pcpu * rms_int_remote_pcpu(struct rmslock *rms, int cpu) { return (zpcpu_get_cpu(rms->pcpu, cpu)); } static void rms_int_influx_enter(struct rmslock *rms, struct rmslock_pcpu *pcpu) { CRITICAL_ASSERT(curthread); MPASS(pcpu->influx == 0); pcpu->influx = 1; } static void rms_int_influx_exit(struct rmslock *rms, struct rmslock_pcpu *pcpu) { CRITICAL_ASSERT(curthread); MPASS(pcpu->influx == 1); pcpu->influx = 0; } #ifdef INVARIANTS static void rms_int_debug_readers_inc(struct rmslock *rms) { int old; old = atomic_fetchadd_int(&rms->debug_readers, 1); KASSERT(old >= 0, ("%s: bad readers count %d\n", __func__, old)); } static void rms_int_debug_readers_dec(struct rmslock *rms) { int old; old = atomic_fetchadd_int(&rms->debug_readers, -1); KASSERT(old > 0, ("%s: bad readers count %d\n", __func__, old)); } #else static void rms_int_debug_readers_inc(struct rmslock *rms) { } static void rms_int_debug_readers_dec(struct rmslock *rms) { } #endif static void rms_int_readers_inc(struct rmslock *rms, struct rmslock_pcpu *pcpu) { CRITICAL_ASSERT(curthread); rms_int_debug_readers_inc(rms); pcpu->readers++; } static void rms_int_readers_dec(struct rmslock *rms, struct rmslock_pcpu *pcpu) { CRITICAL_ASSERT(curthread); rms_int_debug_readers_dec(rms); pcpu->readers--; } /* * Public API */ void rms_init(struct rmslock *rms, const char *name) { rms->owner = RMS_NOOWNER; rms->writers = 0; rms->readers = 0; rms->debug_readers = 0; mtx_init(&rms->mtx, name, NULL, MTX_DEF | MTX_NEW); rms->pcpu = uma_zalloc_pcpu(pcpu_zone_8, M_WAITOK | M_ZERO); } void rms_destroy(struct rmslock *rms) { MPASS(rms->writers == 0); MPASS(rms->readers == 0); mtx_destroy(&rms->mtx); uma_zfree_pcpu(pcpu_zone_8, rms->pcpu); } static void __noinline rms_rlock_fallback(struct rmslock *rms) { rms_int_influx_exit(rms, rms_int_pcpu(rms)); critical_exit(); mtx_lock(&rms->mtx); while (rms->writers > 0) msleep(&rms->readers, &rms->mtx, PUSER - 1, mtx_name(&rms->mtx), 0); critical_enter(); rms_int_readers_inc(rms, rms_int_pcpu(rms)); mtx_unlock(&rms->mtx); critical_exit(); TD_LOCKS_INC(curthread); } void rms_rlock(struct rmslock *rms) { struct rmslock_pcpu *pcpu; WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, __func__); MPASS(atomic_load_ptr(&rms->owner) != curthread); critical_enter(); pcpu = rms_int_pcpu(rms); rms_int_influx_enter(rms, pcpu); atomic_interrupt_fence(); if (__predict_false(rms->writers > 0)) { rms_rlock_fallback(rms); return; } atomic_interrupt_fence(); rms_int_readers_inc(rms, pcpu); atomic_interrupt_fence(); rms_int_influx_exit(rms, pcpu); critical_exit(); TD_LOCKS_INC(curthread); } int rms_try_rlock(struct rmslock *rms) { struct rmslock_pcpu *pcpu; MPASS(atomic_load_ptr(&rms->owner) != curthread); critical_enter(); pcpu = rms_int_pcpu(rms); rms_int_influx_enter(rms, pcpu); atomic_interrupt_fence(); if (__predict_false(rms->writers > 0)) { rms_int_influx_exit(rms, pcpu); critical_exit(); return (0); } atomic_interrupt_fence(); rms_int_readers_inc(rms, pcpu); atomic_interrupt_fence(); rms_int_influx_exit(rms, pcpu); critical_exit(); TD_LOCKS_INC(curthread); return (1); } static void __noinline rms_runlock_fallback(struct rmslock *rms) { rms_int_influx_exit(rms, rms_int_pcpu(rms)); critical_exit(); mtx_lock(&rms->mtx); MPASS(rms->writers > 0); MPASS(rms->readers > 0); MPASS(rms->debug_readers == rms->readers); rms_int_debug_readers_dec(rms); rms->readers--; if (rms->readers == 0) wakeup_one(&rms->writers); mtx_unlock(&rms->mtx); TD_LOCKS_DEC(curthread); } void rms_runlock(struct rmslock *rms) { struct rmslock_pcpu *pcpu; critical_enter(); pcpu = rms_int_pcpu(rms); rms_int_influx_enter(rms, pcpu); atomic_interrupt_fence(); if (__predict_false(rms->writers > 0)) { rms_runlock_fallback(rms); return; } atomic_interrupt_fence(); rms_int_readers_dec(rms, pcpu); atomic_interrupt_fence(); rms_int_influx_exit(rms, pcpu); critical_exit(); TD_LOCKS_DEC(curthread); } struct rmslock_ipi { struct rmslock *rms; struct smp_rendezvous_cpus_retry_arg srcra; }; static void rms_action_func(void *arg) { struct rmslock_ipi *rmsipi; struct rmslock_pcpu *pcpu; struct rmslock *rms; rmsipi = __containerof(arg, struct rmslock_ipi, srcra); rms = rmsipi->rms; pcpu = rms_int_pcpu(rms); if (pcpu->influx) return; if (pcpu->readers != 0) { atomic_add_int(&rms->readers, pcpu->readers); pcpu->readers = 0; } smp_rendezvous_cpus_done(arg); } static void rms_wait_func(void *arg, int cpu) { struct rmslock_ipi *rmsipi; struct rmslock_pcpu *pcpu; struct rmslock *rms; rmsipi = __containerof(arg, struct rmslock_ipi, srcra); rms = rmsipi->rms; pcpu = rms_int_remote_pcpu(rms, cpu); while (atomic_load_int(&pcpu->influx)) cpu_spinwait(); } #ifdef INVARIANTS static void rms_assert_no_pcpu_readers(struct rmslock *rms) { struct rmslock_pcpu *pcpu; int cpu; CPU_FOREACH(cpu) { pcpu = rms_int_remote_pcpu(rms, cpu); if (pcpu->readers != 0) { panic("%s: got %d readers on cpu %d\n", __func__, pcpu->readers, cpu); } } } #else static void rms_assert_no_pcpu_readers(struct rmslock *rms) { } #endif static void rms_wlock_switch(struct rmslock *rms) { struct rmslock_ipi rmsipi; MPASS(rms->readers == 0); MPASS(rms->writers == 1); rmsipi.rms = rms; smp_rendezvous_cpus_retry(all_cpus, smp_no_rendezvous_barrier, rms_action_func, smp_no_rendezvous_barrier, rms_wait_func, &rmsipi.srcra); } void rms_wlock(struct rmslock *rms) { WITNESS_WARN(WARN_GIANTOK | WARN_SLEEPOK, NULL, __func__); MPASS(atomic_load_ptr(&rms->owner) != curthread); mtx_lock(&rms->mtx); rms->writers++; if (rms->writers > 1) { msleep(&rms->owner, &rms->mtx, (PUSER - 1), mtx_name(&rms->mtx), 0); MPASS(rms->readers == 0); KASSERT(rms->owner == RMS_TRANSIENT, ("%s: unexpected owner value %p\n", __func__, rms->owner)); goto out_grab; } KASSERT(rms->owner == RMS_NOOWNER, ("%s: unexpected owner value %p\n", __func__, rms->owner)); rms_wlock_switch(rms); rms_assert_no_pcpu_readers(rms); if (rms->readers > 0) { msleep(&rms->writers, &rms->mtx, (PUSER - 1), mtx_name(&rms->mtx), 0); } out_grab: rms->owner = curthread; rms_assert_no_pcpu_readers(rms); mtx_unlock(&rms->mtx); MPASS(rms->readers == 0); TD_LOCKS_INC(curthread); } void rms_wunlock(struct rmslock *rms) { mtx_lock(&rms->mtx); KASSERT(rms->owner == curthread, ("%s: unexpected owner value %p\n", __func__, rms->owner)); MPASS(rms->writers >= 1); MPASS(rms->readers == 0); rms->writers--; if (rms->writers > 0) { wakeup_one(&rms->owner); rms->owner = RMS_TRANSIENT; } else { wakeup(&rms->readers); rms->owner = RMS_NOOWNER; } mtx_unlock(&rms->mtx); TD_LOCKS_DEC(curthread); } void rms_unlock(struct rmslock *rms) { if (rms_wowned(rms)) rms_wunlock(rms); else rms_runlock(rms); } diff --git a/sys/kern/sched_4bsd.c b/sys/kern/sched_4bsd.c index a3290de9f460..7e6fe03c80e4 100644 --- a/sys/kern/sched_4bsd.c +++ b/sys/kern/sched_4bsd.c @@ -1,1797 +1,1795 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1982, 1986, 1990, 1991, 1993 * The Regents of the University of California. All rights reserved. * (c) UNIX System Laboratories, Inc. * All or some portions of this file are derived from material licensed * to the University of California by American Telephone and Telegraph * Co. or Unix System Laboratories, Inc. and are reproduced herein with * the permission of UNIX System Laboratories, Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include "opt_hwpmc_hooks.h" #include "opt_sched.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef HWPMC_HOOKS #include #endif #ifdef KDTRACE_HOOKS #include int __read_mostly dtrace_vtime_active; dtrace_vtime_switch_func_t dtrace_vtime_switch_func; #endif /* * INVERSE_ESTCPU_WEIGHT is only suitable for statclock() frequencies in * the range 100-256 Hz (approximately). */ #define ESTCPULIM(e) \ min((e), INVERSE_ESTCPU_WEIGHT * (NICE_WEIGHT * (PRIO_MAX - PRIO_MIN) - \ RQ_PPQ) + INVERSE_ESTCPU_WEIGHT - 1) #ifdef SMP #define INVERSE_ESTCPU_WEIGHT (8 * smp_cpus) #else #define INVERSE_ESTCPU_WEIGHT 8 /* 1 / (priorities per estcpu level). */ #endif #define NICE_WEIGHT 1 /* Priorities per nice level. */ #define TS_NAME_LEN (MAXCOMLEN + sizeof(" td ") + sizeof(__XSTRING(UINT_MAX))) /* * The schedulable entity that runs a context. * This is an extension to the thread structure and is tailored to * the requirements of this scheduler. * All fields are protected by the scheduler lock. */ struct td_sched { fixpt_t ts_pctcpu; /* %cpu during p_swtime. */ u_int ts_estcpu; /* Estimated cpu utilization. */ int ts_cpticks; /* Ticks of cpu time. */ int ts_slptime; /* Seconds !RUNNING. */ int ts_slice; /* Remaining part of time slice. */ int ts_flags; struct runq *ts_runq; /* runq the thread is currently on */ #ifdef KTR char ts_name[TS_NAME_LEN]; #endif }; /* flags kept in td_flags */ #define TDF_DIDRUN TDF_SCHED0 /* thread actually ran. */ #define TDF_BOUND TDF_SCHED1 /* Bound to one CPU. */ #define TDF_SLICEEND TDF_SCHED2 /* Thread time slice is over. */ /* flags kept in ts_flags */ #define TSF_AFFINITY 0x0001 /* Has a non-"full" CPU set. */ #define SKE_RUNQ_PCPU(ts) \ ((ts)->ts_runq != 0 && (ts)->ts_runq != &runq) #define THREAD_CAN_SCHED(td, cpu) \ CPU_ISSET((cpu), &(td)->td_cpuset->cs_mask) _Static_assert(sizeof(struct thread) + sizeof(struct td_sched) <= sizeof(struct thread0_storage), "increase struct thread0_storage.t0st_sched size"); static struct mtx sched_lock; static int realstathz = 127; /* stathz is sometimes 0 and run off of hz. */ static int sched_tdcnt; /* Total runnable threads in the system. */ static int sched_slice = 12; /* Thread run time before rescheduling. */ static void setup_runqs(void); static void schedcpu(void); static void schedcpu_thread(void); static void sched_priority(struct thread *td, u_char prio); static void sched_setup(void *dummy); static void maybe_resched(struct thread *td); static void updatepri(struct thread *td); static void resetpriority(struct thread *td); static void resetpriority_thread(struct thread *td); #ifdef SMP static int sched_pickcpu(struct thread *td); static int forward_wakeup(int cpunum); static void kick_other_cpu(int pri, int cpuid); #endif static struct kproc_desc sched_kp = { "schedcpu", schedcpu_thread, NULL }; SYSINIT(schedcpu, SI_SUB_LAST, SI_ORDER_FIRST, kproc_start, &sched_kp); SYSINIT(sched_setup, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, sched_setup, NULL); static void sched_initticks(void *dummy); SYSINIT(sched_initticks, SI_SUB_CLOCKS, SI_ORDER_THIRD, sched_initticks, NULL); /* * Global run queue. */ static struct runq runq; #ifdef SMP /* * Per-CPU run queues */ static struct runq runq_pcpu[MAXCPU]; long runq_length[MAXCPU]; static cpuset_t idle_cpus_mask; #endif struct pcpuidlestat { u_int idlecalls; u_int oldidlecalls; }; DPCPU_DEFINE_STATIC(struct pcpuidlestat, idlestat); static void setup_runqs(void) { #ifdef SMP int i; for (i = 0; i < MAXCPU; ++i) runq_init(&runq_pcpu[i]); #endif runq_init(&runq); } static int sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) { int error, new_val, period; period = 1000000 / realstathz; new_val = period * sched_slice; error = sysctl_handle_int(oidp, &new_val, 0, req); if (error != 0 || req->newptr == NULL) return (error); if (new_val <= 0) return (EINVAL); sched_slice = imax(1, (new_val + period / 2) / period); hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) / realstathz); return (0); } SYSCTL_NODE(_kern, OID_AUTO, sched, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "Scheduler"); SYSCTL_STRING(_kern_sched, OID_AUTO, name, CTLFLAG_RD, "4BSD", 0, "Scheduler name"); SYSCTL_PROC(_kern_sched, OID_AUTO, quantum, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 0, sysctl_kern_quantum, "I", "Quantum for timeshare threads in microseconds"); SYSCTL_INT(_kern_sched, OID_AUTO, slice, CTLFLAG_RW, &sched_slice, 0, "Quantum for timeshare threads in stathz ticks"); #ifdef SMP /* Enable forwarding of wakeups to all other cpus */ static SYSCTL_NODE(_kern_sched, OID_AUTO, ipiwakeup, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Kernel SMP"); static int runq_fuzz = 1; SYSCTL_INT(_kern_sched, OID_AUTO, runq_fuzz, CTLFLAG_RW, &runq_fuzz, 0, ""); static int forward_wakeup_enabled = 1; SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, enabled, CTLFLAG_RW, &forward_wakeup_enabled, 0, "Forwarding of wakeup to idle CPUs"); static int forward_wakeups_requested = 0; SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, requested, CTLFLAG_RD, &forward_wakeups_requested, 0, "Requests for Forwarding of wakeup to idle CPUs"); static int forward_wakeups_delivered = 0; SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, delivered, CTLFLAG_RD, &forward_wakeups_delivered, 0, "Completed Forwarding of wakeup to idle CPUs"); static int forward_wakeup_use_mask = 1; SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, usemask, CTLFLAG_RW, &forward_wakeup_use_mask, 0, "Use the mask of idle cpus"); static int forward_wakeup_use_loop = 0; SYSCTL_INT(_kern_sched_ipiwakeup, OID_AUTO, useloop, CTLFLAG_RW, &forward_wakeup_use_loop, 0, "Use a loop to find idle cpus"); #endif #if 0 static int sched_followon = 0; SYSCTL_INT(_kern_sched, OID_AUTO, followon, CTLFLAG_RW, &sched_followon, 0, "allow threads to share a quantum"); #endif SDT_PROVIDER_DEFINE(sched); SDT_PROBE_DEFINE3(sched, , , change__pri, "struct thread *", "struct proc *", "uint8_t"); SDT_PROBE_DEFINE3(sched, , , dequeue, "struct thread *", "struct proc *", "void *"); SDT_PROBE_DEFINE4(sched, , , enqueue, "struct thread *", "struct proc *", "void *", "int"); SDT_PROBE_DEFINE4(sched, , , lend__pri, "struct thread *", "struct proc *", "uint8_t", "struct thread *"); SDT_PROBE_DEFINE2(sched, , , load__change, "int", "int"); SDT_PROBE_DEFINE2(sched, , , off__cpu, "struct thread *", "struct proc *"); SDT_PROBE_DEFINE(sched, , , on__cpu); SDT_PROBE_DEFINE(sched, , , remain__cpu); SDT_PROBE_DEFINE2(sched, , , surrender, "struct thread *", "struct proc *"); static __inline void sched_load_add(void) { sched_tdcnt++; KTR_COUNTER0(KTR_SCHED, "load", "global load", sched_tdcnt); SDT_PROBE2(sched, , , load__change, NOCPU, sched_tdcnt); } static __inline void sched_load_rem(void) { sched_tdcnt--; KTR_COUNTER0(KTR_SCHED, "load", "global load", sched_tdcnt); SDT_PROBE2(sched, , , load__change, NOCPU, sched_tdcnt); } /* * Arrange to reschedule if necessary, taking the priorities and * schedulers into account. */ static void maybe_resched(struct thread *td) { THREAD_LOCK_ASSERT(td, MA_OWNED); if (td->td_priority < curthread->td_priority) curthread->td_flags |= TDF_NEEDRESCHED; } /* * This function is called when a thread is about to be put on run queue * because it has been made runnable or its priority has been adjusted. It * determines if the new thread should preempt the current thread. If so, * it sets td_owepreempt to request a preemption. */ int maybe_preempt(struct thread *td) { #ifdef PREEMPTION struct thread *ctd; int cpri, pri; /* * The new thread should not preempt the current thread if any of the * following conditions are true: * * - The kernel is in the throes of crashing (panicstr). * - The current thread has a higher (numerically lower) or * equivalent priority. Note that this prevents curthread from * trying to preempt to itself. * - The current thread has an inhibitor set or is in the process of * exiting. In this case, the current thread is about to switch * out anyways, so there's no point in preempting. If we did, * the current thread would not be properly resumed as well, so * just avoid that whole landmine. * - If the new thread's priority is not a realtime priority and * the current thread's priority is not an idle priority and * FULL_PREEMPTION is disabled. * * If all of these conditions are false, but the current thread is in * a nested critical section, then we have to defer the preemption * until we exit the critical section. Otherwise, switch immediately * to the new thread. */ ctd = curthread; THREAD_LOCK_ASSERT(td, MA_OWNED); KASSERT((td->td_inhibitors == 0), ("maybe_preempt: trying to run inhibited thread")); pri = td->td_priority; cpri = ctd->td_priority; if (KERNEL_PANICKED() || pri >= cpri /* || dumping */ || TD_IS_INHIBITED(ctd)) return (0); #ifndef FULL_PREEMPTION if (pri > PRI_MAX_ITHD && cpri < PRI_MIN_IDLE) return (0); #endif CTR0(KTR_PROC, "maybe_preempt: scheduling preemption"); ctd->td_owepreempt = 1; return (1); #else return (0); #endif } /* * Constants for digital decay and forget: * 90% of (ts_estcpu) usage in 5 * loadav time * 95% of (ts_pctcpu) usage in 60 seconds (load insensitive) * Note that, as ps(1) mentions, this can let percentages * total over 100% (I've seen 137.9% for 3 processes). * * Note that schedclock() updates ts_estcpu and p_cpticks asynchronously. * * We wish to decay away 90% of ts_estcpu in (5 * loadavg) seconds. * That is, the system wants to compute a value of decay such * that the following for loop: * for (i = 0; i < (5 * loadavg); i++) * ts_estcpu *= decay; * will compute * ts_estcpu *= 0.1; * for all values of loadavg: * * Mathematically this loop can be expressed by saying: * decay ** (5 * loadavg) ~= .1 * * The system computes decay as: * decay = (2 * loadavg) / (2 * loadavg + 1) * * We wish to prove that the system's computation of decay * will always fulfill the equation: * decay ** (5 * loadavg) ~= .1 * * If we compute b as: * b = 2 * loadavg * then * decay = b / (b + 1) * * We now need to prove two things: * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1) * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg) * * Facts: * For x close to zero, exp(x) =~ 1 + x, since * exp(x) = 0! + x**1/1! + x**2/2! + ... . * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b. * For x close to zero, ln(1+x) =~ x, since * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1). * ln(.1) =~ -2.30 * * Proof of (1): * Solve (factor)**(power) =~ .1 given power (5*loadav): * solving for factor, * ln(factor) =~ (-2.30/5*loadav), or * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) = * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED * * Proof of (2): * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)): * solving for power, * power*ln(b/(b+1)) =~ -2.30, or * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED * * Actual power values for the implemented algorithm are as follows: * loadav: 1 2 3 4 * power: 5.68 10.32 14.94 19.55 */ /* calculations for digital decay to forget 90% of usage in 5*loadav sec */ #define loadfactor(loadav) (2 * (loadav)) #define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE)) /* decay 95% of `ts_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */ static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ SYSCTL_UINT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, "Decay factor used for updating %CPU"); /* * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). * * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). * * If you don't want to bother with the faster/more-accurate formula, you * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate * (more general) method of calculating the %age of CPU used by a process. */ #define CCPU_SHIFT 11 /* * Recompute process priorities, every hz ticks. * MP-safe, called without the Giant mutex. */ /* ARGSUSED */ static void schedcpu(void) { fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); struct thread *td; struct proc *p; struct td_sched *ts; int awake; sx_slock(&allproc_lock); FOREACH_PROC_IN_SYSTEM(p) { PROC_LOCK(p); if (p->p_state == PRS_NEW) { PROC_UNLOCK(p); continue; } FOREACH_THREAD_IN_PROC(p, td) { awake = 0; ts = td_get_sched(td); thread_lock(td); /* * Increment sleep time (if sleeping). We * ignore overflow, as above. */ /* * The td_sched slptimes are not touched in wakeup * because the thread may not HAVE everything in * memory? XXX I think this is out of date. */ if (TD_ON_RUNQ(td)) { awake = 1; td->td_flags &= ~TDF_DIDRUN; } else if (TD_IS_RUNNING(td)) { awake = 1; /* Do not clear TDF_DIDRUN */ } else if (td->td_flags & TDF_DIDRUN) { awake = 1; td->td_flags &= ~TDF_DIDRUN; } /* * ts_pctcpu is only for ps and ttyinfo(). */ ts->ts_pctcpu = (ts->ts_pctcpu * ccpu) >> FSHIFT; /* * If the td_sched has been idle the entire second, * stop recalculating its priority until * it wakes up. */ if (ts->ts_cpticks != 0) { #if (FSHIFT >= CCPU_SHIFT) ts->ts_pctcpu += (realstathz == 100) ? ((fixpt_t) ts->ts_cpticks) << (FSHIFT - CCPU_SHIFT) : 100 * (((fixpt_t) ts->ts_cpticks) << (FSHIFT - CCPU_SHIFT)) / realstathz; #else ts->ts_pctcpu += ((FSCALE - ccpu) * (ts->ts_cpticks * FSCALE / realstathz)) >> FSHIFT; #endif ts->ts_cpticks = 0; } /* * If there are ANY running threads in this process, * then don't count it as sleeping. * XXX: this is broken. */ if (awake) { if (ts->ts_slptime > 1) { /* * In an ideal world, this should not * happen, because whoever woke us * up from the long sleep should have * unwound the slptime and reset our * priority before we run at the stale * priority. Should KASSERT at some * point when all the cases are fixed. */ updatepri(td); } ts->ts_slptime = 0; } else ts->ts_slptime++; if (ts->ts_slptime > 1) { thread_unlock(td); continue; } ts->ts_estcpu = decay_cpu(loadfac, ts->ts_estcpu); resetpriority(td); resetpriority_thread(td); thread_unlock(td); } PROC_UNLOCK(p); } sx_sunlock(&allproc_lock); } /* * Main loop for a kthread that executes schedcpu once a second. */ static void schedcpu_thread(void) { for (;;) { schedcpu(); pause("-", hz); } } /* * Recalculate the priority of a process after it has slept for a while. * For all load averages >= 1 and max ts_estcpu of 255, sleeping for at * least six times the loadfactor will decay ts_estcpu to zero. */ static void updatepri(struct thread *td) { struct td_sched *ts; fixpt_t loadfac; unsigned int newcpu; ts = td_get_sched(td); loadfac = loadfactor(averunnable.ldavg[0]); if (ts->ts_slptime > 5 * loadfac) ts->ts_estcpu = 0; else { newcpu = ts->ts_estcpu; ts->ts_slptime--; /* was incremented in schedcpu() */ while (newcpu && --ts->ts_slptime) newcpu = decay_cpu(loadfac, newcpu); ts->ts_estcpu = newcpu; } } /* * Compute the priority of a process when running in user mode. * Arrange to reschedule if the resulting priority is better * than that of the current process. */ static void resetpriority(struct thread *td) { u_int newpriority; if (td->td_pri_class != PRI_TIMESHARE) return; newpriority = PUSER + td_get_sched(td)->ts_estcpu / INVERSE_ESTCPU_WEIGHT + NICE_WEIGHT * (td->td_proc->p_nice - PRIO_MIN); newpriority = min(max(newpriority, PRI_MIN_TIMESHARE), PRI_MAX_TIMESHARE); sched_user_prio(td, newpriority); } /* * Update the thread's priority when the associated process's user * priority changes. */ static void resetpriority_thread(struct thread *td) { /* Only change threads with a time sharing user priority. */ if (td->td_priority < PRI_MIN_TIMESHARE || td->td_priority > PRI_MAX_TIMESHARE) return; /* XXX the whole needresched thing is broken, but not silly. */ maybe_resched(td); sched_prio(td, td->td_user_pri); } /* ARGSUSED */ static void sched_setup(void *dummy) { setup_runqs(); /* Account for thread0. */ sched_load_add(); } /* * This routine determines time constants after stathz and hz are setup. */ static void sched_initticks(void *dummy) { realstathz = stathz ? stathz : hz; sched_slice = realstathz / 10; /* ~100ms */ hogticks = imax(1, (2 * hz * sched_slice + realstathz / 2) / realstathz); } /* External interfaces start here */ /* * Very early in the boot some setup of scheduler-specific * parts of proc0 and of some scheduler resources needs to be done. * Called from: * proc0_init() */ void schedinit(void) { /* * Set up the scheduler specific parts of thread0. */ thread0.td_lock = &sched_lock; td_get_sched(&thread0)->ts_slice = sched_slice; mtx_init(&sched_lock, "sched lock", NULL, MTX_SPIN); } int sched_runnable(void) { #ifdef SMP return runq_check(&runq) + runq_check(&runq_pcpu[PCPU_GET(cpuid)]); #else return runq_check(&runq); #endif } int sched_rr_interval(void) { /* Convert sched_slice from stathz to hz. */ return (imax(1, (sched_slice * hz + realstathz / 2) / realstathz)); } /* * We adjust the priority of the current process. The priority of a * process gets worse as it accumulates CPU time. The cpu usage * estimator (ts_estcpu) is increased here. resetpriority() will * compute a different priority each time ts_estcpu increases by * INVERSE_ESTCPU_WEIGHT (until PRI_MAX_TIMESHARE is reached). The * cpu usage estimator ramps up quite quickly when the process is * running (linearly), and decays away exponentially, at a rate which * is proportionally slower when the system is busy. The basic * principle is that the system will 90% forget that the process used * a lot of CPU time in 5 * loadav seconds. This causes the system to * favor processes which haven't run much recently, and to round-robin * among other processes. */ static void sched_clock_tick(struct thread *td) { struct pcpuidlestat *stat; struct td_sched *ts; THREAD_LOCK_ASSERT(td, MA_OWNED); ts = td_get_sched(td); ts->ts_cpticks++; ts->ts_estcpu = ESTCPULIM(ts->ts_estcpu + 1); if ((ts->ts_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { resetpriority(td); resetpriority_thread(td); } /* * Force a context switch if the current thread has used up a full * time slice (default is 100ms). */ if (!TD_IS_IDLETHREAD(td) && --ts->ts_slice <= 0) { ts->ts_slice = sched_slice; td->td_flags |= TDF_NEEDRESCHED | TDF_SLICEEND; } stat = DPCPU_PTR(idlestat); stat->oldidlecalls = stat->idlecalls; stat->idlecalls = 0; } void sched_clock(struct thread *td, int cnt) { for ( ; cnt > 0; cnt--) sched_clock_tick(td); } /* * Charge child's scheduling CPU usage to parent. */ void sched_exit(struct proc *p, struct thread *td) { KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "proc exit", "prio:%d", td->td_priority); PROC_LOCK_ASSERT(p, MA_OWNED); sched_exit_thread(FIRST_THREAD_IN_PROC(p), td); } void sched_exit_thread(struct thread *td, struct thread *child) { KTR_STATE1(KTR_SCHED, "thread", sched_tdname(child), "exit", "prio:%d", child->td_priority); thread_lock(td); td_get_sched(td)->ts_estcpu = ESTCPULIM(td_get_sched(td)->ts_estcpu + td_get_sched(child)->ts_estcpu); thread_unlock(td); thread_lock(child); if ((child->td_flags & TDF_NOLOAD) == 0) sched_load_rem(); thread_unlock(child); } void sched_fork(struct thread *td, struct thread *childtd) { sched_fork_thread(td, childtd); } void sched_fork_thread(struct thread *td, struct thread *childtd) { struct td_sched *ts, *tsc; childtd->td_oncpu = NOCPU; childtd->td_lastcpu = NOCPU; childtd->td_lock = &sched_lock; childtd->td_cpuset = cpuset_ref(td->td_cpuset); childtd->td_domain.dr_policy = td->td_cpuset->cs_domain; childtd->td_priority = childtd->td_base_pri; ts = td_get_sched(childtd); bzero(ts, sizeof(*ts)); tsc = td_get_sched(td); ts->ts_estcpu = tsc->ts_estcpu; ts->ts_flags |= (tsc->ts_flags & TSF_AFFINITY); ts->ts_slice = 1; } void sched_nice(struct proc *p, int nice) { struct thread *td; PROC_LOCK_ASSERT(p, MA_OWNED); p->p_nice = nice; FOREACH_THREAD_IN_PROC(p, td) { thread_lock(td); resetpriority(td); resetpriority_thread(td); thread_unlock(td); } } void sched_class(struct thread *td, int class) { THREAD_LOCK_ASSERT(td, MA_OWNED); td->td_pri_class = class; } /* * Adjust the priority of a thread. */ static void sched_priority(struct thread *td, u_char prio) { KTR_POINT3(KTR_SCHED, "thread", sched_tdname(td), "priority change", "prio:%d", td->td_priority, "new prio:%d", prio, KTR_ATTR_LINKED, sched_tdname(curthread)); SDT_PROBE3(sched, , , change__pri, td, td->td_proc, prio); if (td != curthread && prio > td->td_priority) { KTR_POINT3(KTR_SCHED, "thread", sched_tdname(curthread), "lend prio", "prio:%d", td->td_priority, "new prio:%d", prio, KTR_ATTR_LINKED, sched_tdname(td)); SDT_PROBE4(sched, , , lend__pri, td, td->td_proc, prio, curthread); } THREAD_LOCK_ASSERT(td, MA_OWNED); if (td->td_priority == prio) return; td->td_priority = prio; if (TD_ON_RUNQ(td) && td->td_rqindex != (prio / RQ_PPQ)) { sched_rem(td); sched_add(td, SRQ_BORING | SRQ_HOLDTD); } } /* * Update a thread's priority when it is lent another thread's * priority. */ void sched_lend_prio(struct thread *td, u_char prio) { td->td_flags |= TDF_BORROWING; sched_priority(td, prio); } /* * Restore a thread's priority when priority propagation is * over. The prio argument is the minimum priority the thread * needs to have to satisfy other possible priority lending * requests. If the thread's regulary priority is less * important than prio the thread will keep a priority boost * of prio. */ void sched_unlend_prio(struct thread *td, u_char prio) { u_char base_pri; if (td->td_base_pri >= PRI_MIN_TIMESHARE && td->td_base_pri <= PRI_MAX_TIMESHARE) base_pri = td->td_user_pri; else base_pri = td->td_base_pri; if (prio >= base_pri) { td->td_flags &= ~TDF_BORROWING; sched_prio(td, base_pri); } else sched_lend_prio(td, prio); } void sched_prio(struct thread *td, u_char prio) { u_char oldprio; /* First, update the base priority. */ td->td_base_pri = prio; /* * If the thread is borrowing another thread's priority, don't ever * lower the priority. */ if (td->td_flags & TDF_BORROWING && td->td_priority < prio) return; /* Change the real priority. */ oldprio = td->td_priority; sched_priority(td, prio); /* * If the thread is on a turnstile, then let the turnstile update * its state. */ if (TD_ON_LOCK(td) && oldprio != prio) turnstile_adjust(td, oldprio); } void sched_user_prio(struct thread *td, u_char prio) { THREAD_LOCK_ASSERT(td, MA_OWNED); td->td_base_user_pri = prio; if (td->td_lend_user_pri <= prio) return; td->td_user_pri = prio; } void sched_lend_user_prio(struct thread *td, u_char prio) { THREAD_LOCK_ASSERT(td, MA_OWNED); td->td_lend_user_pri = prio; td->td_user_pri = min(prio, td->td_base_user_pri); if (td->td_priority > td->td_user_pri) sched_prio(td, td->td_user_pri); else if (td->td_priority != td->td_user_pri) td->td_flags |= TDF_NEEDRESCHED; } /* * Like the above but first check if there is anything to do. */ void sched_lend_user_prio_cond(struct thread *td, u_char prio) { if (td->td_lend_user_pri != prio) goto lend; if (td->td_user_pri != min(prio, td->td_base_user_pri)) goto lend; if (td->td_priority != td->td_user_pri) goto lend; return; lend: thread_lock(td); sched_lend_user_prio(td, prio); thread_unlock(td); } void sched_sleep(struct thread *td, int pri) { THREAD_LOCK_ASSERT(td, MA_OWNED); td->td_slptick = ticks; td_get_sched(td)->ts_slptime = 0; if (pri != 0 && PRI_BASE(td->td_pri_class) == PRI_TIMESHARE) sched_prio(td, pri); if (TD_IS_SUSPENDED(td) || pri >= PSOCK) td->td_flags |= TDF_CANSWAP; } void sched_switch(struct thread *td, int flags) { struct thread *newtd; struct mtx *tmtx; struct td_sched *ts; struct proc *p; int preempted; tmtx = &sched_lock; ts = td_get_sched(td); p = td->td_proc; THREAD_LOCK_ASSERT(td, MA_OWNED); td->td_lastcpu = td->td_oncpu; preempted = (td->td_flags & TDF_SLICEEND) == 0 && (flags & SW_PREEMPT) != 0; td->td_flags &= ~(TDF_NEEDRESCHED | TDF_SLICEEND); td->td_owepreempt = 0; td->td_oncpu = NOCPU; /* * At the last moment, if this thread is still marked RUNNING, * then put it back on the run queue as it has not been suspended * or stopped or any thing else similar. We never put the idle * threads on the run queue, however. */ if (td->td_flags & TDF_IDLETD) { TD_SET_CAN_RUN(td); #ifdef SMP CPU_CLR(PCPU_GET(cpuid), &idle_cpus_mask); #endif } else { if (TD_IS_RUNNING(td)) { /* Put us back on the run queue. */ sched_add(td, preempted ? SRQ_HOLDTD|SRQ_OURSELF|SRQ_YIELDING|SRQ_PREEMPTED : SRQ_HOLDTD|SRQ_OURSELF|SRQ_YIELDING); } } /* * Switch to the sched lock to fix things up and pick * a new thread. Block the td_lock in order to avoid * breaking the critical path. */ if (td->td_lock != &sched_lock) { mtx_lock_spin(&sched_lock); tmtx = thread_lock_block(td); mtx_unlock_spin(tmtx); } if ((td->td_flags & TDF_NOLOAD) == 0) sched_load_rem(); newtd = choosethread(); MPASS(newtd->td_lock == &sched_lock); #if (KTR_COMPILE & KTR_SCHED) != 0 if (TD_IS_IDLETHREAD(td)) KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "idle", "prio:%d", td->td_priority); else KTR_STATE3(KTR_SCHED, "thread", sched_tdname(td), KTDSTATE(td), "prio:%d", td->td_priority, "wmesg:\"%s\"", td->td_wmesg, "lockname:\"%s\"", td->td_lockname); #endif if (td != newtd) { #ifdef HWPMC_HOOKS if (PMC_PROC_IS_USING_PMCS(td->td_proc)) PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_OUT); #endif SDT_PROBE2(sched, , , off__cpu, newtd, newtd->td_proc); /* I feel sleepy */ lock_profile_release_lock(&sched_lock.lock_object, true); #ifdef KDTRACE_HOOKS /* * If DTrace has set the active vtime enum to anything * other than INACTIVE (0), then it should have set the * function to call. */ if (dtrace_vtime_active) (*dtrace_vtime_switch_func)(newtd); #endif cpu_switch(td, newtd, tmtx); lock_profile_obtain_lock_success(&sched_lock.lock_object, true, 0, 0, __FILE__, __LINE__); /* * Where am I? What year is it? * We are in the same thread that went to sleep above, * but any amount of time may have passed. All our context * will still be available as will local variables. * PCPU values however may have changed as we may have * changed CPU so don't trust cached values of them. * New threads will go to fork_exit() instead of here * so if you change things here you may need to change * things there too. * * If the thread above was exiting it will never wake * up again here, so either it has saved everything it * needed to, or the thread_wait() or wait() will * need to reap it. */ SDT_PROBE0(sched, , , on__cpu); #ifdef HWPMC_HOOKS if (PMC_PROC_IS_USING_PMCS(td->td_proc)) PMC_SWITCH_CONTEXT(td, PMC_FN_CSW_IN); #endif } else { td->td_lock = &sched_lock; SDT_PROBE0(sched, , , remain__cpu); } KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running", "prio:%d", td->td_priority); #ifdef SMP if (td->td_flags & TDF_IDLETD) CPU_SET(PCPU_GET(cpuid), &idle_cpus_mask); #endif sched_lock.mtx_lock = (uintptr_t)td; td->td_oncpu = PCPU_GET(cpuid); spinlock_enter(); mtx_unlock_spin(&sched_lock); } void sched_wakeup(struct thread *td, int srqflags) { struct td_sched *ts; THREAD_LOCK_ASSERT(td, MA_OWNED); ts = td_get_sched(td); td->td_flags &= ~TDF_CANSWAP; if (ts->ts_slptime > 1) { updatepri(td); resetpriority(td); } td->td_slptick = 0; ts->ts_slptime = 0; ts->ts_slice = sched_slice; sched_add(td, srqflags); } #ifdef SMP static int forward_wakeup(int cpunum) { struct pcpu *pc; cpuset_t dontuse, map, map2; u_int id, me; int iscpuset; mtx_assert(&sched_lock, MA_OWNED); CTR0(KTR_RUNQ, "forward_wakeup()"); if ((!forward_wakeup_enabled) || (forward_wakeup_use_mask == 0 && forward_wakeup_use_loop == 0)) return (0); if (!smp_started || KERNEL_PANICKED()) return (0); forward_wakeups_requested++; /* * Check the idle mask we received against what we calculated * before in the old version. */ me = PCPU_GET(cpuid); /* Don't bother if we should be doing it ourself. */ if (CPU_ISSET(me, &idle_cpus_mask) && (cpunum == NOCPU || me == cpunum)) return (0); CPU_SETOF(me, &dontuse); - CPU_OR(&dontuse, &stopped_cpus); - CPU_OR(&dontuse, &hlt_cpus_mask); + CPU_OR(&dontuse, &dontuse, &stopped_cpus); + CPU_OR(&dontuse, &dontuse, &hlt_cpus_mask); CPU_ZERO(&map2); if (forward_wakeup_use_loop) { STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) { id = pc->pc_cpuid; if (!CPU_ISSET(id, &dontuse) && pc->pc_curthread == pc->pc_idlethread) { CPU_SET(id, &map2); } } } if (forward_wakeup_use_mask) { - map = idle_cpus_mask; - CPU_ANDNOT(&map, &dontuse); + CPU_ANDNOT(&map, &idle_cpus_mask, &dontuse); /* If they are both on, compare and use loop if different. */ if (forward_wakeup_use_loop) { if (CPU_CMP(&map, &map2)) { printf("map != map2, loop method preferred\n"); map = map2; } } } else { map = map2; } /* If we only allow a specific CPU, then mask off all the others. */ if (cpunum != NOCPU) { KASSERT((cpunum <= mp_maxcpus),("forward_wakeup: bad cpunum.")); iscpuset = CPU_ISSET(cpunum, &map); if (iscpuset == 0) CPU_ZERO(&map); else CPU_SETOF(cpunum, &map); } if (!CPU_EMPTY(&map)) { forward_wakeups_delivered++; STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) { id = pc->pc_cpuid; if (!CPU_ISSET(id, &map)) continue; if (cpu_idle_wakeup(pc->pc_cpuid)) CPU_CLR(id, &map); } if (!CPU_EMPTY(&map)) ipi_selected(map, IPI_AST); return (1); } if (cpunum == NOCPU) printf("forward_wakeup: Idle processor not found\n"); return (0); } static void kick_other_cpu(int pri, int cpuid) { struct pcpu *pcpu; int cpri; pcpu = pcpu_find(cpuid); if (CPU_ISSET(cpuid, &idle_cpus_mask)) { forward_wakeups_delivered++; if (!cpu_idle_wakeup(cpuid)) ipi_cpu(cpuid, IPI_AST); return; } cpri = pcpu->pc_curthread->td_priority; if (pri >= cpri) return; #if defined(IPI_PREEMPTION) && defined(PREEMPTION) #if !defined(FULL_PREEMPTION) if (pri <= PRI_MAX_ITHD) #endif /* ! FULL_PREEMPTION */ { ipi_cpu(cpuid, IPI_PREEMPT); return; } #endif /* defined(IPI_PREEMPTION) && defined(PREEMPTION) */ pcpu->pc_curthread->td_flags |= TDF_NEEDRESCHED; ipi_cpu(cpuid, IPI_AST); return; } #endif /* SMP */ #ifdef SMP static int sched_pickcpu(struct thread *td) { int best, cpu; mtx_assert(&sched_lock, MA_OWNED); if (td->td_lastcpu != NOCPU && THREAD_CAN_SCHED(td, td->td_lastcpu)) best = td->td_lastcpu; else best = NOCPU; CPU_FOREACH(cpu) { if (!THREAD_CAN_SCHED(td, cpu)) continue; if (best == NOCPU) best = cpu; else if (runq_length[cpu] < runq_length[best]) best = cpu; } KASSERT(best != NOCPU, ("no valid CPUs")); return (best); } #endif void sched_add(struct thread *td, int flags) #ifdef SMP { cpuset_t tidlemsk; struct td_sched *ts; u_int cpu, cpuid; int forwarded = 0; int single_cpu = 0; ts = td_get_sched(td); THREAD_LOCK_ASSERT(td, MA_OWNED); KASSERT((td->td_inhibitors == 0), ("sched_add: trying to run inhibited thread")); KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), ("sched_add: bad thread state")); KASSERT(td->td_flags & TDF_INMEM, ("sched_add: thread swapped out")); KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add", "prio:%d", td->td_priority, KTR_ATTR_LINKED, sched_tdname(curthread)); KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup", KTR_ATTR_LINKED, sched_tdname(td)); SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL, flags & SRQ_PREEMPTED); /* * Now that the thread is moving to the run-queue, set the lock * to the scheduler's lock. */ if (td->td_lock != &sched_lock) { mtx_lock_spin(&sched_lock); if ((flags & SRQ_HOLD) != 0) td->td_lock = &sched_lock; else thread_lock_set(td, &sched_lock); } TD_SET_RUNQ(td); /* * If SMP is started and the thread is pinned or otherwise limited to * a specific set of CPUs, queue the thread to a per-CPU run queue. * Otherwise, queue the thread to the global run queue. * * If SMP has not yet been started we must use the global run queue * as per-CPU state may not be initialized yet and we may crash if we * try to access the per-CPU run queues. */ if (smp_started && (td->td_pinned != 0 || td->td_flags & TDF_BOUND || ts->ts_flags & TSF_AFFINITY)) { if (td->td_pinned != 0) cpu = td->td_lastcpu; else if (td->td_flags & TDF_BOUND) { /* Find CPU from bound runq. */ KASSERT(SKE_RUNQ_PCPU(ts), ("sched_add: bound td_sched not on cpu runq")); cpu = ts->ts_runq - &runq_pcpu[0]; } else /* Find a valid CPU for our cpuset */ cpu = sched_pickcpu(td); ts->ts_runq = &runq_pcpu[cpu]; single_cpu = 1; CTR3(KTR_RUNQ, "sched_add: Put td_sched:%p(td:%p) on cpu%d runq", ts, td, cpu); } else { CTR2(KTR_RUNQ, "sched_add: adding td_sched:%p (td:%p) to gbl runq", ts, td); cpu = NOCPU; ts->ts_runq = &runq; } if ((td->td_flags & TDF_NOLOAD) == 0) sched_load_add(); runq_add(ts->ts_runq, td, flags); if (cpu != NOCPU) runq_length[cpu]++; cpuid = PCPU_GET(cpuid); if (single_cpu && cpu != cpuid) { kick_other_cpu(td->td_priority, cpu); } else { if (!single_cpu) { - tidlemsk = idle_cpus_mask; - CPU_ANDNOT(&tidlemsk, &hlt_cpus_mask); + CPU_ANDNOT(&tidlemsk, &idle_cpus_mask, &hlt_cpus_mask); CPU_CLR(cpuid, &tidlemsk); if (!CPU_ISSET(cpuid, &idle_cpus_mask) && ((flags & SRQ_INTR) == 0) && !CPU_EMPTY(&tidlemsk)) forwarded = forward_wakeup(cpu); } if (!forwarded) { if (!maybe_preempt(td)) maybe_resched(td); } } if ((flags & SRQ_HOLDTD) == 0) thread_unlock(td); } #else /* SMP */ { struct td_sched *ts; ts = td_get_sched(td); THREAD_LOCK_ASSERT(td, MA_OWNED); KASSERT((td->td_inhibitors == 0), ("sched_add: trying to run inhibited thread")); KASSERT((TD_CAN_RUN(td) || TD_IS_RUNNING(td)), ("sched_add: bad thread state")); KASSERT(td->td_flags & TDF_INMEM, ("sched_add: thread swapped out")); KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq add", "prio:%d", td->td_priority, KTR_ATTR_LINKED, sched_tdname(curthread)); KTR_POINT1(KTR_SCHED, "thread", sched_tdname(curthread), "wokeup", KTR_ATTR_LINKED, sched_tdname(td)); SDT_PROBE4(sched, , , enqueue, td, td->td_proc, NULL, flags & SRQ_PREEMPTED); /* * Now that the thread is moving to the run-queue, set the lock * to the scheduler's lock. */ if (td->td_lock != &sched_lock) { mtx_lock_spin(&sched_lock); if ((flags & SRQ_HOLD) != 0) td->td_lock = &sched_lock; else thread_lock_set(td, &sched_lock); } TD_SET_RUNQ(td); CTR2(KTR_RUNQ, "sched_add: adding td_sched:%p (td:%p) to runq", ts, td); ts->ts_runq = &runq; if ((td->td_flags & TDF_NOLOAD) == 0) sched_load_add(); runq_add(ts->ts_runq, td, flags); if (!maybe_preempt(td)) maybe_resched(td); if ((flags & SRQ_HOLDTD) == 0) thread_unlock(td); } #endif /* SMP */ void sched_rem(struct thread *td) { struct td_sched *ts; ts = td_get_sched(td); KASSERT(td->td_flags & TDF_INMEM, ("sched_rem: thread swapped out")); KASSERT(TD_ON_RUNQ(td), ("sched_rem: thread not on run queue")); mtx_assert(&sched_lock, MA_OWNED); KTR_STATE2(KTR_SCHED, "thread", sched_tdname(td), "runq rem", "prio:%d", td->td_priority, KTR_ATTR_LINKED, sched_tdname(curthread)); SDT_PROBE3(sched, , , dequeue, td, td->td_proc, NULL); if ((td->td_flags & TDF_NOLOAD) == 0) sched_load_rem(); #ifdef SMP if (ts->ts_runq != &runq) runq_length[ts->ts_runq - runq_pcpu]--; #endif runq_remove(ts->ts_runq, td); TD_SET_CAN_RUN(td); } /* * Select threads to run. Note that running threads still consume a * slot. */ struct thread * sched_choose(void) { struct thread *td; struct runq *rq; mtx_assert(&sched_lock, MA_OWNED); #ifdef SMP struct thread *tdcpu; rq = &runq; td = runq_choose_fuzz(&runq, runq_fuzz); tdcpu = runq_choose(&runq_pcpu[PCPU_GET(cpuid)]); if (td == NULL || (tdcpu != NULL && tdcpu->td_priority < td->td_priority)) { CTR2(KTR_RUNQ, "choosing td %p from pcpu runq %d", tdcpu, PCPU_GET(cpuid)); td = tdcpu; rq = &runq_pcpu[PCPU_GET(cpuid)]; } else { CTR1(KTR_RUNQ, "choosing td_sched %p from main runq", td); } #else rq = &runq; td = runq_choose(&runq); #endif if (td) { #ifdef SMP if (td == tdcpu) runq_length[PCPU_GET(cpuid)]--; #endif runq_remove(rq, td); td->td_flags |= TDF_DIDRUN; KASSERT(td->td_flags & TDF_INMEM, ("sched_choose: thread swapped out")); return (td); } return (PCPU_GET(idlethread)); } void sched_preempt(struct thread *td) { SDT_PROBE2(sched, , , surrender, td, td->td_proc); if (td->td_critnest > 1) { td->td_owepreempt = 1; } else { thread_lock(td); mi_switch(SW_INVOL | SW_PREEMPT | SWT_PREEMPT); } } void sched_userret_slowpath(struct thread *td) { thread_lock(td); td->td_priority = td->td_user_pri; td->td_base_pri = td->td_user_pri; thread_unlock(td); } void sched_bind(struct thread *td, int cpu) { struct td_sched *ts; THREAD_LOCK_ASSERT(td, MA_OWNED|MA_NOTRECURSED); KASSERT(td == curthread, ("sched_bind: can only bind curthread")); ts = td_get_sched(td); td->td_flags |= TDF_BOUND; #ifdef SMP ts->ts_runq = &runq_pcpu[cpu]; if (PCPU_GET(cpuid) == cpu) return; mi_switch(SW_VOL); thread_lock(td); #endif } void sched_unbind(struct thread* td) { THREAD_LOCK_ASSERT(td, MA_OWNED); KASSERT(td == curthread, ("sched_unbind: can only bind curthread")); td->td_flags &= ~TDF_BOUND; } int sched_is_bound(struct thread *td) { THREAD_LOCK_ASSERT(td, MA_OWNED); return (td->td_flags & TDF_BOUND); } void sched_relinquish(struct thread *td) { thread_lock(td); mi_switch(SW_VOL | SWT_RELINQUISH); } int sched_load(void) { return (sched_tdcnt); } int sched_sizeof_proc(void) { return (sizeof(struct proc)); } int sched_sizeof_thread(void) { return (sizeof(struct thread) + sizeof(struct td_sched)); } fixpt_t sched_pctcpu(struct thread *td) { struct td_sched *ts; THREAD_LOCK_ASSERT(td, MA_OWNED); ts = td_get_sched(td); return (ts->ts_pctcpu); } #ifdef RACCT /* * Calculates the contribution to the thread cpu usage for the latest * (unfinished) second. */ fixpt_t sched_pctcpu_delta(struct thread *td) { struct td_sched *ts; fixpt_t delta; int realstathz; THREAD_LOCK_ASSERT(td, MA_OWNED); ts = td_get_sched(td); delta = 0; realstathz = stathz ? stathz : hz; if (ts->ts_cpticks != 0) { #if (FSHIFT >= CCPU_SHIFT) delta = (realstathz == 100) ? ((fixpt_t) ts->ts_cpticks) << (FSHIFT - CCPU_SHIFT) : 100 * (((fixpt_t) ts->ts_cpticks) << (FSHIFT - CCPU_SHIFT)) / realstathz; #else delta = ((FSCALE - ccpu) * (ts->ts_cpticks * FSCALE / realstathz)) >> FSHIFT; #endif } return (delta); } #endif u_int sched_estcpu(struct thread *td) { return (td_get_sched(td)->ts_estcpu); } /* * The actual idle process. */ void sched_idletd(void *dummy) { struct pcpuidlestat *stat; THREAD_NO_SLEEPING(); stat = DPCPU_PTR(idlestat); for (;;) { mtx_assert(&Giant, MA_NOTOWNED); while (sched_runnable() == 0) { cpu_idle(stat->idlecalls + stat->oldidlecalls > 64); stat->idlecalls++; } mtx_lock_spin(&sched_lock); mi_switch(SW_VOL | SWT_IDLE); } } /* * A CPU is entering for the first time or a thread is exiting. */ void sched_throw(struct thread *td) { /* * Correct spinlock nesting. The idle thread context that we are * borrowing was created so that it would start out with a single * spin lock (sched_lock) held in fork_trampoline(). Since we've * explicitly acquired locks in this function, the nesting count * is now 2 rather than 1. Since we are nested, calling * spinlock_exit() will simply adjust the counts without allowing * spin lock using code to interrupt us. */ if (td == NULL) { mtx_lock_spin(&sched_lock); spinlock_exit(); PCPU_SET(switchtime, cpu_ticks()); PCPU_SET(switchticks, ticks); } else { lock_profile_release_lock(&sched_lock.lock_object, true); MPASS(td->td_lock == &sched_lock); td->td_lastcpu = td->td_oncpu; td->td_oncpu = NOCPU; } mtx_assert(&sched_lock, MA_OWNED); KASSERT(curthread->td_md.md_spinlock_count == 1, ("invalid count")); cpu_throw(td, choosethread()); /* doesn't return */ } void sched_fork_exit(struct thread *td) { /* * Finish setting up thread glue so that it begins execution in a * non-nested critical section with sched_lock held but not recursed. */ td->td_oncpu = PCPU_GET(cpuid); sched_lock.mtx_lock = (uintptr_t)td; lock_profile_obtain_lock_success(&sched_lock.lock_object, true, 0, 0, __FILE__, __LINE__); THREAD_LOCK_ASSERT(td, MA_OWNED | MA_NOTRECURSED); KTR_STATE1(KTR_SCHED, "thread", sched_tdname(td), "running", "prio:%d", td->td_priority); SDT_PROBE0(sched, , , on__cpu); } char * sched_tdname(struct thread *td) { #ifdef KTR struct td_sched *ts; ts = td_get_sched(td); if (ts->ts_name[0] == '\0') snprintf(ts->ts_name, sizeof(ts->ts_name), "%s tid %d", td->td_name, td->td_tid); return (ts->ts_name); #else return (td->td_name); #endif } #ifdef KTR void sched_clear_tdname(struct thread *td) { struct td_sched *ts; ts = td_get_sched(td); ts->ts_name[0] = '\0'; } #endif void sched_affinity(struct thread *td) { #ifdef SMP struct td_sched *ts; int cpu; THREAD_LOCK_ASSERT(td, MA_OWNED); /* * Set the TSF_AFFINITY flag if there is at least one CPU this * thread can't run on. */ ts = td_get_sched(td); ts->ts_flags &= ~TSF_AFFINITY; CPU_FOREACH(cpu) { if (!THREAD_CAN_SCHED(td, cpu)) { ts->ts_flags |= TSF_AFFINITY; break; } } /* * If this thread can run on all CPUs, nothing else to do. */ if (!(ts->ts_flags & TSF_AFFINITY)) return; /* Pinned threads and bound threads should be left alone. */ if (td->td_pinned != 0 || td->td_flags & TDF_BOUND) return; switch (td->td_state) { case TDS_RUNQ: /* * If we are on a per-CPU runqueue that is in the set, * then nothing needs to be done. */ if (ts->ts_runq != &runq && THREAD_CAN_SCHED(td, ts->ts_runq - runq_pcpu)) return; /* Put this thread on a valid per-CPU runqueue. */ sched_rem(td); sched_add(td, SRQ_HOLDTD | SRQ_BORING); break; case TDS_RUNNING: /* * See if our current CPU is in the set. If not, force a * context switch. */ if (THREAD_CAN_SCHED(td, td->td_oncpu)) return; td->td_flags |= TDF_NEEDRESCHED; if (td != curthread) ipi_cpu(cpu, IPI_AST); break; default: break; } #endif } diff --git a/sys/kern/subr_kdb.c b/sys/kern/subr_kdb.c index 4f439ff103d7..2218c20586f2 100644 --- a/sys/kern/subr_kdb.c +++ b/sys/kern/subr_kdb.c @@ -1,757 +1,756 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2004 The FreeBSD Project * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHORS ``AS IS'' AND ANY EXPRESS OR * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. * IN NO EVENT SHALL THE AUTHORS BE LIABLE FOR ANY DIRECT, INDIRECT, * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include "opt_kdb.h" #include "opt_stack.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef SMP #include #endif u_char __read_frequently kdb_active = 0; static void *kdb_jmpbufp = NULL; struct kdb_dbbe *kdb_dbbe = NULL; static struct pcb kdb_pcb; struct pcb *kdb_thrctx = NULL; struct thread *kdb_thread = NULL; struct trapframe *kdb_frame = NULL; #ifdef BREAK_TO_DEBUGGER #define KDB_BREAK_TO_DEBUGGER 1 #else #define KDB_BREAK_TO_DEBUGGER 0 #endif #ifdef ALT_BREAK_TO_DEBUGGER #define KDB_ALT_BREAK_TO_DEBUGGER 1 #else #define KDB_ALT_BREAK_TO_DEBUGGER 0 #endif static int kdb_break_to_debugger = KDB_BREAK_TO_DEBUGGER; static int kdb_alt_break_to_debugger = KDB_ALT_BREAK_TO_DEBUGGER; KDB_BACKEND(null, NULL, NULL, NULL, NULL); static int kdb_sysctl_available(SYSCTL_HANDLER_ARGS); static int kdb_sysctl_current(SYSCTL_HANDLER_ARGS); static int kdb_sysctl_enter(SYSCTL_HANDLER_ARGS); static int kdb_sysctl_panic(SYSCTL_HANDLER_ARGS); static int kdb_sysctl_panic_str(SYSCTL_HANDLER_ARGS); static int kdb_sysctl_trap(SYSCTL_HANDLER_ARGS); static int kdb_sysctl_trap_code(SYSCTL_HANDLER_ARGS); static int kdb_sysctl_stack_overflow(SYSCTL_HANDLER_ARGS); static SYSCTL_NODE(_debug, OID_AUTO, kdb, CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, "KDB nodes"); SYSCTL_PROC(_debug_kdb, OID_AUTO, available, CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0, kdb_sysctl_available, "A", "list of available KDB backends"); SYSCTL_PROC(_debug_kdb, OID_AUTO, current, CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE, NULL, 0, kdb_sysctl_current, "A", "currently selected KDB backend"); SYSCTL_PROC(_debug_kdb, OID_AUTO, enter, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_SECURE | CTLFLAG_MPSAFE, NULL, 0, kdb_sysctl_enter, "I", "set to enter the debugger"); SYSCTL_PROC(_debug_kdb, OID_AUTO, panic, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_SECURE | CTLFLAG_MPSAFE, NULL, 0, kdb_sysctl_panic, "I", "set to panic the kernel"); SYSCTL_PROC(_debug_kdb, OID_AUTO, panic_str, CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_SECURE | CTLFLAG_MPSAFE, NULL, 0, kdb_sysctl_panic_str, "A", "trigger a kernel panic, using the provided string as the panic message"); SYSCTL_PROC(_debug_kdb, OID_AUTO, trap, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_SECURE | CTLFLAG_MPSAFE, NULL, 0, kdb_sysctl_trap, "I", "set to cause a page fault via data access"); SYSCTL_PROC(_debug_kdb, OID_AUTO, trap_code, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_SECURE | CTLFLAG_MPSAFE, NULL, 0, kdb_sysctl_trap_code, "I", "set to cause a page fault via code access"); SYSCTL_PROC(_debug_kdb, OID_AUTO, stack_overflow, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_SECURE | CTLFLAG_MPSAFE, NULL, 0, kdb_sysctl_stack_overflow, "I", "set to cause a stack overflow"); SYSCTL_INT(_debug_kdb, OID_AUTO, break_to_debugger, CTLFLAG_RWTUN | CTLFLAG_SECURE, &kdb_break_to_debugger, 0, "Enable break to debugger"); SYSCTL_INT(_debug_kdb, OID_AUTO, alt_break_to_debugger, CTLFLAG_RWTUN | CTLFLAG_SECURE, &kdb_alt_break_to_debugger, 0, "Enable alternative break to debugger"); /* * Flag to indicate to debuggers why the debugger was entered. */ const char * volatile kdb_why = KDB_WHY_UNSET; static int kdb_sysctl_available(SYSCTL_HANDLER_ARGS) { struct kdb_dbbe **iter; struct sbuf sbuf; int error; sbuf_new_for_sysctl(&sbuf, NULL, 64, req); SET_FOREACH(iter, kdb_dbbe_set) { if ((*iter)->dbbe_active == 0) sbuf_printf(&sbuf, "%s ", (*iter)->dbbe_name); } error = sbuf_finish(&sbuf); sbuf_delete(&sbuf); return (error); } static int kdb_sysctl_current(SYSCTL_HANDLER_ARGS) { char buf[16]; int error; if (kdb_dbbe != NULL) strlcpy(buf, kdb_dbbe->dbbe_name, sizeof(buf)); else *buf = '\0'; error = sysctl_handle_string(oidp, buf, sizeof(buf), req); if (error != 0 || req->newptr == NULL) return (error); if (kdb_active) return (EBUSY); return (kdb_dbbe_select(buf)); } static int kdb_sysctl_enter(SYSCTL_HANDLER_ARGS) { int error, i; error = sysctl_wire_old_buffer(req, sizeof(int)); if (error == 0) { i = 0; error = sysctl_handle_int(oidp, &i, 0, req); } if (error != 0 || req->newptr == NULL) return (error); if (kdb_active) return (EBUSY); kdb_enter(KDB_WHY_SYSCTL, "sysctl debug.kdb.enter"); return (0); } static int kdb_sysctl_panic(SYSCTL_HANDLER_ARGS) { int error, i; error = sysctl_wire_old_buffer(req, sizeof(int)); if (error == 0) { i = 0; error = sysctl_handle_int(oidp, &i, 0, req); } if (error != 0 || req->newptr == NULL) return (error); panic("kdb_sysctl_panic"); return (0); } static int kdb_sysctl_panic_str(SYSCTL_HANDLER_ARGS) { int error; static char buf[256]; /* static buffer to limit mallocs when panicing */ *buf = '\0'; error = sysctl_handle_string(oidp, buf, sizeof(buf), req); if (error != 0 || req->newptr == NULL) return (error); panic("kdb_sysctl_panic: %s", buf); return (0); } static int kdb_sysctl_trap(SYSCTL_HANDLER_ARGS) { int error, i; int *addr = (int *)0x10; error = sysctl_wire_old_buffer(req, sizeof(int)); if (error == 0) { i = 0; error = sysctl_handle_int(oidp, &i, 0, req); } if (error != 0 || req->newptr == NULL) return (error); return (*addr); } static int kdb_sysctl_trap_code(SYSCTL_HANDLER_ARGS) { int error, i; void (*fp)(u_int, u_int, u_int) = (void *)0xdeadc0de; error = sysctl_wire_old_buffer(req, sizeof(int)); if (error == 0) { i = 0; error = sysctl_handle_int(oidp, &i, 0, req); } if (error != 0 || req->newptr == NULL) return (error); (*fp)(0x11111111, 0x22222222, 0x33333333); return (0); } static void kdb_stack_overflow(volatile int *x) __noinline; static void kdb_stack_overflow(volatile int *x) { if (*x > 10000000) return; kdb_stack_overflow(x); *x += PCPU_GET(cpuid) / 1000000; } static int kdb_sysctl_stack_overflow(SYSCTL_HANDLER_ARGS) { int error, i; volatile int x; error = sysctl_wire_old_buffer(req, sizeof(int)); if (error == 0) { i = 0; error = sysctl_handle_int(oidp, &i, 0, req); } if (error != 0 || req->newptr == NULL) return (error); x = 0; kdb_stack_overflow(&x); return (0); } void kdb_panic(const char *msg) { printf("KDB: panic\n"); panic("%s", msg); } void kdb_reboot(void) { printf("KDB: reboot requested\n"); shutdown_nice(0); } /* * Solaris implements a new BREAK which is initiated by a character sequence * CR ~ ^b which is similar to a familiar pattern used on Sun servers by the * Remote Console. * * Note that this function may be called from almost anywhere, with interrupts * disabled and with unknown locks held, so it must not access data other than * its arguments. Its up to the caller to ensure that the state variable is * consistent. */ #define KEY_CR 13 /* CR '\r' */ #define KEY_TILDE 126 /* ~ */ #define KEY_CRTLB 2 /* ^B */ #define KEY_CRTLP 16 /* ^P */ #define KEY_CRTLR 18 /* ^R */ /* States of th KDB "alternate break sequence" detecting state machine. */ enum { KDB_ALT_BREAK_SEEN_NONE, KDB_ALT_BREAK_SEEN_CR, KDB_ALT_BREAK_SEEN_CR_TILDE, }; int kdb_break(void) { if (!kdb_break_to_debugger) return (0); kdb_enter(KDB_WHY_BREAK, "Break to debugger"); return (KDB_REQ_DEBUGGER); } static int kdb_alt_break_state(int key, int *state) { int brk; /* All states transition to KDB_ALT_BREAK_SEEN_CR on a CR. */ if (key == KEY_CR) { *state = KDB_ALT_BREAK_SEEN_CR; return (0); } brk = 0; switch (*state) { case KDB_ALT_BREAK_SEEN_CR: *state = KDB_ALT_BREAK_SEEN_NONE; if (key == KEY_TILDE) *state = KDB_ALT_BREAK_SEEN_CR_TILDE; break; case KDB_ALT_BREAK_SEEN_CR_TILDE: *state = KDB_ALT_BREAK_SEEN_NONE; if (key == KEY_CRTLB) brk = KDB_REQ_DEBUGGER; else if (key == KEY_CRTLP) brk = KDB_REQ_PANIC; else if (key == KEY_CRTLR) brk = KDB_REQ_REBOOT; break; case KDB_ALT_BREAK_SEEN_NONE: default: *state = KDB_ALT_BREAK_SEEN_NONE; break; } return (brk); } static int kdb_alt_break_internal(int key, int *state, int force_gdb) { int brk; if (!kdb_alt_break_to_debugger) return (0); brk = kdb_alt_break_state(key, state); switch (brk) { case KDB_REQ_DEBUGGER: if (force_gdb) kdb_dbbe_select("gdb"); kdb_enter(KDB_WHY_BREAK, "Break to debugger"); break; case KDB_REQ_PANIC: if (force_gdb) kdb_dbbe_select("gdb"); kdb_panic("Panic sequence on console"); break; case KDB_REQ_REBOOT: kdb_reboot(); break; } return (0); } int kdb_alt_break(int key, int *state) { return (kdb_alt_break_internal(key, state, 0)); } /* * This variation on kdb_alt_break() is used only by dcons, which has its own * configuration flag to force GDB use regardless of the global KDB * configuration. */ int kdb_alt_break_gdb(int key, int *state) { return (kdb_alt_break_internal(key, state, 1)); } /* * Print a backtrace of the calling thread. The backtrace is generated by * the selected debugger, provided it supports backtraces. If no debugger * is selected or the current debugger does not support backtraces, this * function silently returns. */ void kdb_backtrace(void) { if (kdb_dbbe != NULL && kdb_dbbe->dbbe_trace != NULL) { printf("KDB: stack backtrace:\n"); kdb_dbbe->dbbe_trace(); } #ifdef STACK else { struct stack st; printf("KDB: stack backtrace:\n"); stack_zero(&st); stack_save(&st); stack_print_ddb(&st); } #endif } /* * Similar to kdb_backtrace() except that it prints a backtrace of an * arbitrary thread rather than the calling thread. */ void kdb_backtrace_thread(struct thread *td) { if (kdb_dbbe != NULL && kdb_dbbe->dbbe_trace_thread != NULL) { printf("KDB: stack backtrace of thread %d:\n", td->td_tid); kdb_dbbe->dbbe_trace_thread(td); } #ifdef STACK else { struct stack st; printf("KDB: stack backtrace of thread %d:\n", td->td_tid); if (stack_save_td(&st, td) == 0) stack_print_ddb(&st); } #endif } /* * Set/change the current backend. */ int kdb_dbbe_select(const char *name) { struct kdb_dbbe *be, **iter; SET_FOREACH(iter, kdb_dbbe_set) { be = *iter; if (be->dbbe_active == 0 && strcmp(be->dbbe_name, name) == 0) { kdb_dbbe = be; return (0); } } return (EINVAL); } /* * Enter the currently selected debugger. If a message has been provided, * it is printed first. If the debugger does not support the enter method, * it is entered by using breakpoint(), which enters the debugger through * kdb_trap(). The 'why' argument will contain a more mechanically usable * string than 'msg', and is relied upon by DDB scripting to identify the * reason for entering the debugger so that the right script can be run. */ void kdb_enter(const char *why, const char *msg) { if (kdb_dbbe != NULL && kdb_active == 0) { if (msg != NULL) printf("KDB: enter: %s\n", msg); kdb_why = why; breakpoint(); kdb_why = KDB_WHY_UNSET; } } /* * Initialize the kernel debugger interface. */ void kdb_init(void) { struct kdb_dbbe *be, **iter; int cur_pri, pri; kdb_active = 0; kdb_dbbe = NULL; cur_pri = -1; SET_FOREACH(iter, kdb_dbbe_set) { be = *iter; pri = (be->dbbe_init != NULL) ? be->dbbe_init() : -1; be->dbbe_active = (pri >= 0) ? 0 : -1; if (pri > cur_pri) { cur_pri = pri; kdb_dbbe = be; } } if (kdb_dbbe != NULL) { printf("KDB: debugger backends:"); SET_FOREACH(iter, kdb_dbbe_set) { be = *iter; if (be->dbbe_active == 0) printf(" %s", be->dbbe_name); } printf("\n"); printf("KDB: current backend: %s\n", kdb_dbbe->dbbe_name); } } /* * Handle contexts. */ void * kdb_jmpbuf(jmp_buf new) { void *old; old = kdb_jmpbufp; kdb_jmpbufp = new; return (old); } void kdb_reenter(void) { if (!kdb_active || kdb_jmpbufp == NULL) return; printf("KDB: reentering\n"); kdb_backtrace(); longjmp(kdb_jmpbufp, 1); /* NOTREACHED */ } void kdb_reenter_silent(void) { if (!kdb_active || kdb_jmpbufp == NULL) return; longjmp(kdb_jmpbufp, 1); /* NOTREACHED */ } /* * Thread-related support functions. */ struct pcb * kdb_thr_ctx(struct thread *thr) { #if defined(SMP) && defined(KDB_STOPPEDPCB) struct pcpu *pc; #endif if (thr == curthread) return (&kdb_pcb); #if defined(SMP) && defined(KDB_STOPPEDPCB) STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) { if (pc->pc_curthread == thr && CPU_ISSET(pc->pc_cpuid, &stopped_cpus)) return (KDB_STOPPEDPCB(pc)); } #endif return (thr->td_pcb); } struct thread * kdb_thr_first(void) { struct proc *p; struct thread *thr; u_int i; /* This function may be called early. */ if (pidhashtbl == NULL) return (&thread0); for (i = 0; i <= pidhash; i++) { LIST_FOREACH(p, &pidhashtbl[i], p_hash) { thr = FIRST_THREAD_IN_PROC(p); if (thr != NULL) return (thr); } } return (NULL); } struct thread * kdb_thr_from_pid(pid_t pid) { struct proc *p; LIST_FOREACH(p, PIDHASH(pid), p_hash) { if (p->p_pid == pid) return (FIRST_THREAD_IN_PROC(p)); } return (NULL); } struct thread * kdb_thr_lookup(lwpid_t tid) { struct thread *thr; thr = kdb_thr_first(); while (thr != NULL && thr->td_tid != tid) thr = kdb_thr_next(thr); return (thr); } struct thread * kdb_thr_next(struct thread *thr) { struct proc *p; u_int hash; p = thr->td_proc; thr = TAILQ_NEXT(thr, td_plist); if (thr != NULL) return (thr); if (pidhashtbl == NULL) return (NULL); hash = p->p_pid & pidhash; for (;;) { p = LIST_NEXT(p, p_hash); while (p == NULL) { if (++hash > pidhash) return (NULL); p = LIST_FIRST(&pidhashtbl[hash]); } thr = FIRST_THREAD_IN_PROC(p); if (thr != NULL) return (thr); } } int kdb_thr_select(struct thread *thr) { if (thr == NULL) return (EINVAL); kdb_thread = thr; kdb_thrctx = kdb_thr_ctx(thr); return (0); } /* * Enter the debugger due to a trap. */ int kdb_trap(int type, int code, struct trapframe *tf) { #ifdef SMP cpuset_t other_cpus; #endif struct kdb_dbbe *be; register_t intr; int handled; int did_stop_cpus; be = kdb_dbbe; if (be == NULL || be->dbbe_trap == NULL) return (0); /* We reenter the debugger through kdb_reenter(). */ if (kdb_active) return (0); intr = intr_disable(); if (!SCHEDULER_STOPPED()) { #ifdef SMP - other_cpus = all_cpus; - CPU_ANDNOT(&other_cpus, &stopped_cpus); + CPU_ANDNOT(&other_cpus, &all_cpus, &stopped_cpus); CPU_CLR(PCPU_GET(cpuid), &other_cpus); stop_cpus_hard(other_cpus); #endif curthread->td_stopsched = 1; did_stop_cpus = 1; } else did_stop_cpus = 0; kdb_active++; kdb_frame = tf; /* Let MD code do its thing first... */ kdb_cpu_trap(type, code); makectx(tf, &kdb_pcb); kdb_thr_select(curthread); cngrab(); for (;;) { handled = be->dbbe_trap(type, code); if (be == kdb_dbbe) break; be = kdb_dbbe; if (be == NULL || be->dbbe_trap == NULL) break; printf("Switching to %s back-end\n", be->dbbe_name); } cnungrab(); kdb_active--; if (did_stop_cpus) { curthread->td_stopsched = 0; #ifdef SMP - CPU_AND(&other_cpus, &stopped_cpus); + CPU_AND(&other_cpus, &other_cpus, &stopped_cpus); restart_cpus(other_cpus); #endif } intr_restore(intr); return (handled); } diff --git a/sys/kern/subr_smp.c b/sys/kern/subr_smp.c index bfe890d773f9..a677075937e6 100644 --- a/sys/kern/subr_smp.c +++ b/sys/kern/subr_smp.c @@ -1,1342 +1,1342 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2001, John Baldwin . * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. */ /* * This module holds the global variables and machine independent functions * used for the kernel SMP support. */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include "opt_sched.h" #ifdef SMP MALLOC_DEFINE(M_TOPO, "toponodes", "SMP topology data"); volatile cpuset_t stopped_cpus; volatile cpuset_t started_cpus; volatile cpuset_t suspended_cpus; cpuset_t hlt_cpus_mask; cpuset_t logical_cpus_mask; void (*cpustop_restartfunc)(void); #endif static int sysctl_kern_smp_active(SYSCTL_HANDLER_ARGS); /* This is used in modules that need to work in both SMP and UP. */ cpuset_t all_cpus; int mp_ncpus; /* export this for libkvm consumers. */ int mp_maxcpus = MAXCPU; volatile int smp_started; u_int mp_maxid; static SYSCTL_NODE(_kern, OID_AUTO, smp, CTLFLAG_RD | CTLFLAG_CAPRD | CTLFLAG_MPSAFE, NULL, "Kernel SMP"); SYSCTL_INT(_kern_smp, OID_AUTO, maxid, CTLFLAG_RD|CTLFLAG_CAPRD, &mp_maxid, 0, "Max CPU ID."); SYSCTL_INT(_kern_smp, OID_AUTO, maxcpus, CTLFLAG_RD|CTLFLAG_CAPRD, &mp_maxcpus, 0, "Max number of CPUs that the system was compiled for."); SYSCTL_PROC(_kern_smp, OID_AUTO, active, CTLFLAG_RD|CTLTYPE_INT|CTLFLAG_MPSAFE, NULL, 0, sysctl_kern_smp_active, "I", "Indicates system is running in SMP mode"); int smp_disabled = 0; /* has smp been disabled? */ SYSCTL_INT(_kern_smp, OID_AUTO, disabled, CTLFLAG_RDTUN|CTLFLAG_CAPRD, &smp_disabled, 0, "SMP has been disabled from the loader"); int smp_cpus = 1; /* how many cpu's running */ SYSCTL_INT(_kern_smp, OID_AUTO, cpus, CTLFLAG_RD|CTLFLAG_CAPRD, &smp_cpus, 0, "Number of CPUs online"); int smp_threads_per_core = 1; /* how many SMT threads are running per core */ SYSCTL_INT(_kern_smp, OID_AUTO, threads_per_core, CTLFLAG_RD|CTLFLAG_CAPRD, &smp_threads_per_core, 0, "Number of SMT threads online per core"); int mp_ncores = -1; /* how many physical cores running */ SYSCTL_INT(_kern_smp, OID_AUTO, cores, CTLFLAG_RD|CTLFLAG_CAPRD, &mp_ncores, 0, "Number of physical cores online"); int smp_topology = 0; /* Which topology we're using. */ SYSCTL_INT(_kern_smp, OID_AUTO, topology, CTLFLAG_RDTUN, &smp_topology, 0, "Topology override setting; 0 is default provided by hardware."); #ifdef SMP /* Enable forwarding of a signal to a process running on a different CPU */ static int forward_signal_enabled = 1; SYSCTL_INT(_kern_smp, OID_AUTO, forward_signal_enabled, CTLFLAG_RW, &forward_signal_enabled, 0, "Forwarding of a signal to a process on a different CPU"); /* Variables needed for SMP rendezvous. */ static volatile int smp_rv_ncpus; static void (*volatile smp_rv_setup_func)(void *arg); static void (*volatile smp_rv_action_func)(void *arg); static void (*volatile smp_rv_teardown_func)(void *arg); static void *volatile smp_rv_func_arg; static volatile int smp_rv_waiters[4]; /* * Shared mutex to restrict busywaits between smp_rendezvous() and * smp(_targeted)_tlb_shootdown(). A deadlock occurs if both of these * functions trigger at once and cause multiple CPUs to busywait with * interrupts disabled. */ struct mtx smp_ipi_mtx; /* * Let the MD SMP code initialize mp_maxid very early if it can. */ static void mp_setmaxid(void *dummy) { cpu_mp_setmaxid(); KASSERT(mp_ncpus >= 1, ("%s: CPU count < 1", __func__)); KASSERT(mp_ncpus > 1 || mp_maxid == 0, ("%s: one CPU but mp_maxid is not zero", __func__)); KASSERT(mp_maxid >= mp_ncpus - 1, ("%s: counters out of sync: max %d, count %d", __func__, mp_maxid, mp_ncpus)); } SYSINIT(cpu_mp_setmaxid, SI_SUB_TUNABLES, SI_ORDER_FIRST, mp_setmaxid, NULL); /* * Call the MD SMP initialization code. */ static void mp_start(void *dummy) { mtx_init(&smp_ipi_mtx, "smp rendezvous", NULL, MTX_SPIN); /* Probe for MP hardware. */ if (smp_disabled != 0 || cpu_mp_probe() == 0) { mp_ncores = 1; mp_ncpus = 1; CPU_SETOF(PCPU_GET(cpuid), &all_cpus); return; } cpu_mp_start(); printf("FreeBSD/SMP: Multiprocessor System Detected: %d CPUs\n", mp_ncpus); /* Provide a default for most architectures that don't have SMT/HTT. */ if (mp_ncores < 0) mp_ncores = mp_ncpus; cpu_mp_announce(); } SYSINIT(cpu_mp, SI_SUB_CPU, SI_ORDER_THIRD, mp_start, NULL); void forward_signal(struct thread *td) { int id; /* * signotify() has already set TDF_ASTPENDING and TDF_NEEDSIGCHECK on * this thread, so all we need to do is poke it if it is currently * executing so that it executes ast(). */ THREAD_LOCK_ASSERT(td, MA_OWNED); KASSERT(TD_IS_RUNNING(td), ("forward_signal: thread is not TDS_RUNNING")); CTR1(KTR_SMP, "forward_signal(%p)", td->td_proc); if (!smp_started || cold || KERNEL_PANICKED()) return; if (!forward_signal_enabled) return; /* No need to IPI ourself. */ if (td == curthread) return; id = td->td_oncpu; if (id == NOCPU) return; ipi_cpu(id, IPI_AST); } /* * When called the executing CPU will send an IPI to all other CPUs * requesting that they halt execution. * * Usually (but not necessarily) called with 'other_cpus' as its arg. * * - Signals all CPUs in map to stop. * - Waits for each to stop. * * Returns: * -1: error * 0: NA * 1: ok * */ #if defined(__amd64__) || defined(__i386__) #define X86 1 #else #define X86 0 #endif static int generic_stop_cpus(cpuset_t map, u_int type) { #ifdef KTR char cpusetbuf[CPUSETBUFSIZ]; #endif static volatile u_int stopping_cpu = NOCPU; int i; volatile cpuset_t *cpus; KASSERT( type == IPI_STOP || type == IPI_STOP_HARD #if X86 || type == IPI_SUSPEND #endif , ("%s: invalid stop type", __func__)); if (!smp_started) return (0); CTR2(KTR_SMP, "stop_cpus(%s) with %u type", cpusetobj_strprint(cpusetbuf, &map), type); #if X86 /* * When suspending, ensure there are are no IPIs in progress. * IPIs that have been issued, but not yet delivered (e.g. * not pending on a vCPU when running under virtualization) * will be lost, violating FreeBSD's assumption of reliable * IPI delivery. */ if (type == IPI_SUSPEND) mtx_lock_spin(&smp_ipi_mtx); #endif #if X86 if (!nmi_is_broadcast || nmi_kdb_lock == 0) { #endif if (stopping_cpu != PCPU_GET(cpuid)) while (atomic_cmpset_int(&stopping_cpu, NOCPU, PCPU_GET(cpuid)) == 0) while (stopping_cpu != NOCPU) cpu_spinwait(); /* spin */ /* send the stop IPI to all CPUs in map */ ipi_selected(map, type); #if X86 } #endif #if X86 if (type == IPI_SUSPEND) cpus = &suspended_cpus; else #endif cpus = &stopped_cpus; i = 0; while (!CPU_SUBSET(cpus, &map)) { /* spin */ cpu_spinwait(); i++; if (i == 100000000) { printf("timeout stopping cpus\n"); break; } } #if X86 if (type == IPI_SUSPEND) mtx_unlock_spin(&smp_ipi_mtx); #endif stopping_cpu = NOCPU; return (1); } int stop_cpus(cpuset_t map) { return (generic_stop_cpus(map, IPI_STOP)); } int stop_cpus_hard(cpuset_t map) { return (generic_stop_cpus(map, IPI_STOP_HARD)); } #if X86 int suspend_cpus(cpuset_t map) { return (generic_stop_cpus(map, IPI_SUSPEND)); } #endif /* * Called by a CPU to restart stopped CPUs. * * Usually (but not necessarily) called with 'stopped_cpus' as its arg. * * - Signals all CPUs in map to restart. * - Waits for each to restart. * * Returns: * -1: error * 0: NA * 1: ok */ static int generic_restart_cpus(cpuset_t map, u_int type) { #ifdef KTR char cpusetbuf[CPUSETBUFSIZ]; #endif volatile cpuset_t *cpus; #if X86 KASSERT(type == IPI_STOP || type == IPI_STOP_HARD || type == IPI_SUSPEND, ("%s: invalid stop type", __func__)); if (!smp_started) return (0); CTR1(KTR_SMP, "restart_cpus(%s)", cpusetobj_strprint(cpusetbuf, &map)); if (type == IPI_SUSPEND) cpus = &resuming_cpus; else cpus = &stopped_cpus; /* signal other cpus to restart */ if (type == IPI_SUSPEND) CPU_COPY_STORE_REL(&map, &toresume_cpus); else CPU_COPY_STORE_REL(&map, &started_cpus); /* * Wake up any CPUs stopped with MWAIT. From MI code we can't tell if * MONITOR/MWAIT is enabled, but the potentially redundant writes are * relatively inexpensive. */ if (type == IPI_STOP) { struct monitorbuf *mb; u_int id; CPU_FOREACH(id) { if (!CPU_ISSET(id, &map)) continue; mb = &pcpu_find(id)->pc_monitorbuf; atomic_store_int(&mb->stop_state, MONITOR_STOPSTATE_RUNNING); } } if (!nmi_is_broadcast || nmi_kdb_lock == 0) { /* wait for each to clear its bit */ while (CPU_OVERLAP(cpus, &map)) cpu_spinwait(); } #else /* !X86 */ KASSERT(type == IPI_STOP || type == IPI_STOP_HARD, ("%s: invalid stop type", __func__)); if (!smp_started) return (0); CTR1(KTR_SMP, "restart_cpus(%s)", cpusetobj_strprint(cpusetbuf, &map)); cpus = &stopped_cpus; /* signal other cpus to restart */ CPU_COPY_STORE_REL(&map, &started_cpus); /* wait for each to clear its bit */ while (CPU_OVERLAP(cpus, &map)) cpu_spinwait(); #endif return (1); } int restart_cpus(cpuset_t map) { return (generic_restart_cpus(map, IPI_STOP)); } #if X86 int resume_cpus(cpuset_t map) { return (generic_restart_cpus(map, IPI_SUSPEND)); } #endif #undef X86 /* * All-CPU rendezvous. CPUs are signalled, all execute the setup function * (if specified), rendezvous, execute the action function (if specified), * rendezvous again, execute the teardown function (if specified), and then * resume. * * Note that the supplied external functions _must_ be reentrant and aware * that they are running in parallel and in an unknown lock context. */ void smp_rendezvous_action(void) { struct thread *td; void *local_func_arg; void (*local_setup_func)(void*); void (*local_action_func)(void*); void (*local_teardown_func)(void*); #ifdef INVARIANTS int owepreempt; #endif /* Ensure we have up-to-date values. */ atomic_add_acq_int(&smp_rv_waiters[0], 1); while (smp_rv_waiters[0] < smp_rv_ncpus) cpu_spinwait(); /* Fetch rendezvous parameters after acquire barrier. */ local_func_arg = smp_rv_func_arg; local_setup_func = smp_rv_setup_func; local_action_func = smp_rv_action_func; local_teardown_func = smp_rv_teardown_func; /* * Use a nested critical section to prevent any preemptions * from occurring during a rendezvous action routine. * Specifically, if a rendezvous handler is invoked via an IPI * and the interrupted thread was in the critical_exit() * function after setting td_critnest to 0 but before * performing a deferred preemption, this routine can be * invoked with td_critnest set to 0 and td_owepreempt true. * In that case, a critical_exit() during the rendezvous * action would trigger a preemption which is not permitted in * a rendezvous action. To fix this, wrap all of the * rendezvous action handlers in a critical section. We * cannot use a regular critical section however as having * critical_exit() preempt from this routine would also be * problematic (the preemption must not occur before the IPI * has been acknowledged via an EOI). Instead, we * intentionally ignore td_owepreempt when leaving the * critical section. This should be harmless because we do * not permit rendezvous action routines to schedule threads, * and thus td_owepreempt should never transition from 0 to 1 * during this routine. */ td = curthread; td->td_critnest++; #ifdef INVARIANTS owepreempt = td->td_owepreempt; #endif /* * If requested, run a setup function before the main action * function. Ensure all CPUs have completed the setup * function before moving on to the action function. */ if (local_setup_func != smp_no_rendezvous_barrier) { if (smp_rv_setup_func != NULL) smp_rv_setup_func(smp_rv_func_arg); atomic_add_int(&smp_rv_waiters[1], 1); while (smp_rv_waiters[1] < smp_rv_ncpus) cpu_spinwait(); } if (local_action_func != NULL) local_action_func(local_func_arg); if (local_teardown_func != smp_no_rendezvous_barrier) { /* * Signal that the main action has been completed. If a * full exit rendezvous is requested, then all CPUs will * wait here until all CPUs have finished the main action. */ atomic_add_int(&smp_rv_waiters[2], 1); while (smp_rv_waiters[2] < smp_rv_ncpus) cpu_spinwait(); if (local_teardown_func != NULL) local_teardown_func(local_func_arg); } /* * Signal that the rendezvous is fully completed by this CPU. * This means that no member of smp_rv_* pseudo-structure will be * accessed by this target CPU after this point; in particular, * memory pointed by smp_rv_func_arg. * * The release semantic ensures that all accesses performed by * the current CPU are visible when smp_rendezvous_cpus() * returns, by synchronizing with the * atomic_load_acq_int(&smp_rv_waiters[3]). */ atomic_add_rel_int(&smp_rv_waiters[3], 1); td->td_critnest--; KASSERT(owepreempt == td->td_owepreempt, ("rendezvous action changed td_owepreempt")); } void smp_rendezvous_cpus(cpuset_t map, void (* setup_func)(void *), void (* action_func)(void *), void (* teardown_func)(void *), void *arg) { int curcpumap, i, ncpus = 0; /* See comments in the !SMP case. */ if (!smp_started) { spinlock_enter(); if (setup_func != NULL) setup_func(arg); if (action_func != NULL) action_func(arg); if (teardown_func != NULL) teardown_func(arg); spinlock_exit(); return; } /* * Make sure we come here with interrupts enabled. Otherwise we * livelock if smp_ipi_mtx is owned by a thread which sent us an IPI. */ MPASS(curthread->td_md.md_spinlock_count == 0); CPU_FOREACH(i) { if (CPU_ISSET(i, &map)) ncpus++; } if (ncpus == 0) panic("ncpus is 0 with non-zero map"); mtx_lock_spin(&smp_ipi_mtx); /* Pass rendezvous parameters via global variables. */ smp_rv_ncpus = ncpus; smp_rv_setup_func = setup_func; smp_rv_action_func = action_func; smp_rv_teardown_func = teardown_func; smp_rv_func_arg = arg; smp_rv_waiters[1] = 0; smp_rv_waiters[2] = 0; smp_rv_waiters[3] = 0; atomic_store_rel_int(&smp_rv_waiters[0], 0); /* * Signal other processors, which will enter the IPI with * interrupts off. */ curcpumap = CPU_ISSET(curcpu, &map); CPU_CLR(curcpu, &map); ipi_selected(map, IPI_RENDEZVOUS); /* Check if the current CPU is in the map */ if (curcpumap != 0) smp_rendezvous_action(); /* * Ensure that the master CPU waits for all the other * CPUs to finish the rendezvous, so that smp_rv_* * pseudo-structure and the arg are guaranteed to not * be in use. * * Load acquire synchronizes with the release add in * smp_rendezvous_action(), which ensures that our caller sees * all memory actions done by the called functions on other * CPUs. */ while (atomic_load_acq_int(&smp_rv_waiters[3]) < ncpus) cpu_spinwait(); mtx_unlock_spin(&smp_ipi_mtx); } void smp_rendezvous(void (* setup_func)(void *), void (* action_func)(void *), void (* teardown_func)(void *), void *arg) { smp_rendezvous_cpus(all_cpus, setup_func, action_func, teardown_func, arg); } static struct cpu_group group[MAXCPU * MAX_CACHE_LEVELS + 1]; static void smp_topo_fill(struct cpu_group *cg) { int c; for (c = 0; c < cg->cg_children; c++) smp_topo_fill(&cg->cg_child[c]); cg->cg_first = CPU_FFS(&cg->cg_mask) - 1; cg->cg_last = CPU_FLS(&cg->cg_mask) - 1; } struct cpu_group * smp_topo(void) { char cpusetbuf[CPUSETBUFSIZ], cpusetbuf2[CPUSETBUFSIZ]; struct cpu_group *top; /* * Check for a fake topology request for debugging purposes. */ switch (smp_topology) { case 1: /* Dual core with no sharing. */ top = smp_topo_1level(CG_SHARE_NONE, 2, 0); break; case 2: /* No topology, all cpus are equal. */ top = smp_topo_none(); break; case 3: /* Dual core with shared L2. */ top = smp_topo_1level(CG_SHARE_L2, 2, 0); break; case 4: /* quad core, shared l3 among each package, private l2. */ top = smp_topo_1level(CG_SHARE_L3, 4, 0); break; case 5: /* quad core, 2 dualcore parts on each package share l2. */ top = smp_topo_2level(CG_SHARE_NONE, 2, CG_SHARE_L2, 2, 0); break; case 6: /* Single-core 2xHTT */ top = smp_topo_1level(CG_SHARE_L1, 2, CG_FLAG_HTT); break; case 7: /* quad core with a shared l3, 8 threads sharing L2. */ top = smp_topo_2level(CG_SHARE_L3, 4, CG_SHARE_L2, 8, CG_FLAG_SMT); break; default: /* Default, ask the system what it wants. */ top = cpu_topo(); break; } /* * Verify the returned topology. */ if (top->cg_count != mp_ncpus) panic("Built bad topology at %p. CPU count %d != %d", top, top->cg_count, mp_ncpus); if (CPU_CMP(&top->cg_mask, &all_cpus)) panic("Built bad topology at %p. CPU mask (%s) != (%s)", top, cpusetobj_strprint(cpusetbuf, &top->cg_mask), cpusetobj_strprint(cpusetbuf2, &all_cpus)); /* * Collapse nonsense levels that may be created out of convenience by * the MD layers. They cause extra work in the search functions. */ while (top->cg_children == 1) { top = &top->cg_child[0]; top->cg_parent = NULL; } smp_topo_fill(top); return (top); } struct cpu_group * smp_topo_alloc(u_int count) { static u_int index; u_int curr; curr = index; index += count; return (&group[curr]); } struct cpu_group * smp_topo_none(void) { struct cpu_group *top; top = &group[0]; top->cg_parent = NULL; top->cg_child = NULL; top->cg_mask = all_cpus; top->cg_count = mp_ncpus; top->cg_children = 0; top->cg_level = CG_SHARE_NONE; top->cg_flags = 0; return (top); } static int smp_topo_addleaf(struct cpu_group *parent, struct cpu_group *child, int share, int count, int flags, int start) { char cpusetbuf[CPUSETBUFSIZ], cpusetbuf2[CPUSETBUFSIZ]; cpuset_t mask; int i; CPU_ZERO(&mask); for (i = 0; i < count; i++, start++) CPU_SET(start, &mask); child->cg_parent = parent; child->cg_child = NULL; child->cg_children = 0; child->cg_level = share; child->cg_count = count; child->cg_flags = flags; child->cg_mask = mask; parent->cg_children++; for (; parent != NULL; parent = parent->cg_parent) { if (CPU_OVERLAP(&parent->cg_mask, &child->cg_mask)) panic("Duplicate children in %p. mask (%s) child (%s)", parent, cpusetobj_strprint(cpusetbuf, &parent->cg_mask), cpusetobj_strprint(cpusetbuf2, &child->cg_mask)); - CPU_OR(&parent->cg_mask, &child->cg_mask); + CPU_OR(&parent->cg_mask, &parent->cg_mask, &child->cg_mask); parent->cg_count += child->cg_count; } return (start); } struct cpu_group * smp_topo_1level(int share, int count, int flags) { struct cpu_group *child; struct cpu_group *top; int packages; int cpu; int i; cpu = 0; top = &group[0]; packages = mp_ncpus / count; top->cg_child = child = &group[1]; top->cg_level = CG_SHARE_NONE; for (i = 0; i < packages; i++, child++) cpu = smp_topo_addleaf(top, child, share, count, flags, cpu); return (top); } struct cpu_group * smp_topo_2level(int l2share, int l2count, int l1share, int l1count, int l1flags) { struct cpu_group *top; struct cpu_group *l1g; struct cpu_group *l2g; int cpu; int i; int j; cpu = 0; top = &group[0]; l2g = &group[1]; top->cg_child = l2g; top->cg_level = CG_SHARE_NONE; top->cg_children = mp_ncpus / (l2count * l1count); l1g = l2g + top->cg_children; for (i = 0; i < top->cg_children; i++, l2g++) { l2g->cg_parent = top; l2g->cg_child = l1g; l2g->cg_level = l2share; for (j = 0; j < l2count; j++, l1g++) cpu = smp_topo_addleaf(l2g, l1g, l1share, l1count, l1flags, cpu); } return (top); } struct cpu_group * smp_topo_find(struct cpu_group *top, int cpu) { struct cpu_group *cg; cpuset_t mask; int children; int i; CPU_SETOF(cpu, &mask); cg = top; for (;;) { if (!CPU_OVERLAP(&cg->cg_mask, &mask)) return (NULL); if (cg->cg_children == 0) return (cg); children = cg->cg_children; for (i = 0, cg = cg->cg_child; i < children; cg++, i++) if (CPU_OVERLAP(&cg->cg_mask, &mask)) break; } return (NULL); } #else /* !SMP */ void smp_rendezvous_cpus(cpuset_t map, void (*setup_func)(void *), void (*action_func)(void *), void (*teardown_func)(void *), void *arg) { /* * In the !SMP case we just need to ensure the same initial conditions * as the SMP case. */ spinlock_enter(); if (setup_func != NULL) setup_func(arg); if (action_func != NULL) action_func(arg); if (teardown_func != NULL) teardown_func(arg); spinlock_exit(); } void smp_rendezvous(void (*setup_func)(void *), void (*action_func)(void *), void (*teardown_func)(void *), void *arg) { smp_rendezvous_cpus(all_cpus, setup_func, action_func, teardown_func, arg); } /* * Provide dummy SMP support for UP kernels. Modules that need to use SMP * APIs will still work using this dummy support. */ static void mp_setvariables_for_up(void *dummy) { mp_ncpus = 1; mp_ncores = 1; mp_maxid = PCPU_GET(cpuid); CPU_SETOF(mp_maxid, &all_cpus); KASSERT(PCPU_GET(cpuid) == 0, ("UP must have a CPU ID of zero")); } SYSINIT(cpu_mp_setvariables, SI_SUB_TUNABLES, SI_ORDER_FIRST, mp_setvariables_for_up, NULL); #endif /* SMP */ void smp_no_rendezvous_barrier(void *dummy) { #ifdef SMP KASSERT((!smp_started),("smp_no_rendezvous called and smp is started")); #endif } void smp_rendezvous_cpus_retry(cpuset_t map, void (* setup_func)(void *), void (* action_func)(void *), void (* teardown_func)(void *), void (* wait_func)(void *, int), struct smp_rendezvous_cpus_retry_arg *arg) { int cpu; CPU_COPY(&map, &arg->cpus); /* * Only one CPU to execute on. */ if (!smp_started) { spinlock_enter(); if (setup_func != NULL) setup_func(arg); if (action_func != NULL) action_func(arg); if (teardown_func != NULL) teardown_func(arg); spinlock_exit(); return; } /* * Execute an action on all specified CPUs while retrying until they * all acknowledge completion. */ for (;;) { smp_rendezvous_cpus( arg->cpus, setup_func, action_func, teardown_func, arg); if (CPU_EMPTY(&arg->cpus)) break; CPU_FOREACH(cpu) { if (!CPU_ISSET(cpu, &arg->cpus)) continue; wait_func(arg, cpu); } } } void smp_rendezvous_cpus_done(struct smp_rendezvous_cpus_retry_arg *arg) { CPU_CLR_ATOMIC(curcpu, &arg->cpus); } /* * If (prio & PDROP) == 0: * Wait for specified idle threads to switch once. This ensures that even * preempted threads have cycled through the switch function once, * exiting their codepaths. This allows us to change global pointers * with no other synchronization. * If (prio & PDROP) != 0: * Force the specified CPUs to switch context at least once. */ int quiesce_cpus(cpuset_t map, const char *wmesg, int prio) { struct pcpu *pcpu; u_int *gen; int error; int cpu; error = 0; if ((prio & PDROP) == 0) { gen = malloc(sizeof(u_int) * MAXCPU, M_TEMP, M_WAITOK); for (cpu = 0; cpu <= mp_maxid; cpu++) { if (!CPU_ISSET(cpu, &map) || CPU_ABSENT(cpu)) continue; pcpu = pcpu_find(cpu); gen[cpu] = pcpu->pc_idlethread->td_generation; } } for (cpu = 0; cpu <= mp_maxid; cpu++) { if (!CPU_ISSET(cpu, &map) || CPU_ABSENT(cpu)) continue; pcpu = pcpu_find(cpu); thread_lock(curthread); sched_bind(curthread, cpu); thread_unlock(curthread); if ((prio & PDROP) != 0) continue; while (gen[cpu] == pcpu->pc_idlethread->td_generation) { error = tsleep(quiesce_cpus, prio & ~PDROP, wmesg, 1); if (error != EWOULDBLOCK) goto out; error = 0; } } out: thread_lock(curthread); sched_unbind(curthread); thread_unlock(curthread); if ((prio & PDROP) == 0) free(gen, M_TEMP); return (error); } int quiesce_all_cpus(const char *wmesg, int prio) { return quiesce_cpus(all_cpus, wmesg, prio); } /* * Observe all CPUs not executing in critical section. * We are not in one so the check for us is safe. If the found * thread changes to something else we know the section was * exited as well. */ void quiesce_all_critical(void) { struct thread *td, *newtd; struct pcpu *pcpu; int cpu; MPASS(curthread->td_critnest == 0); CPU_FOREACH(cpu) { pcpu = cpuid_to_pcpu[cpu]; td = pcpu->pc_curthread; for (;;) { if (td->td_critnest == 0) break; cpu_spinwait(); newtd = (struct thread *) atomic_load_acq_ptr((void *)pcpu->pc_curthread); if (td != newtd) break; } } } static void cpus_fence_seq_cst_issue(void *arg __unused) { atomic_thread_fence_seq_cst(); } /* * Send an IPI forcing a sequentially consistent fence. * * Allows replacement of an explicitly fence with a compiler barrier. * Trades speed up during normal execution for a significant slowdown when * the barrier is needed. */ void cpus_fence_seq_cst(void) { #ifdef SMP smp_rendezvous( smp_no_rendezvous_barrier, cpus_fence_seq_cst_issue, smp_no_rendezvous_barrier, NULL ); #else cpus_fence_seq_cst_issue(NULL); #endif } /* Extra care is taken with this sysctl because the data type is volatile */ static int sysctl_kern_smp_active(SYSCTL_HANDLER_ARGS) { int error, active; active = smp_started; error = SYSCTL_OUT(req, &active, sizeof(active)); return (error); } #ifdef SMP void topo_init_node(struct topo_node *node) { bzero(node, sizeof(*node)); TAILQ_INIT(&node->children); } void topo_init_root(struct topo_node *root) { topo_init_node(root); root->type = TOPO_TYPE_SYSTEM; } /* * Add a child node with the given ID under the given parent. * Do nothing if there is already a child with that ID. */ struct topo_node * topo_add_node_by_hwid(struct topo_node *parent, int hwid, topo_node_type type, uintptr_t subtype) { struct topo_node *node; TAILQ_FOREACH_REVERSE(node, &parent->children, topo_children, siblings) { if (node->hwid == hwid && node->type == type && node->subtype == subtype) { return (node); } } node = malloc(sizeof(*node), M_TOPO, M_WAITOK); topo_init_node(node); node->parent = parent; node->hwid = hwid; node->type = type; node->subtype = subtype; TAILQ_INSERT_TAIL(&parent->children, node, siblings); parent->nchildren++; return (node); } /* * Find a child node with the given ID under the given parent. */ struct topo_node * topo_find_node_by_hwid(struct topo_node *parent, int hwid, topo_node_type type, uintptr_t subtype) { struct topo_node *node; TAILQ_FOREACH(node, &parent->children, siblings) { if (node->hwid == hwid && node->type == type && node->subtype == subtype) { return (node); } } return (NULL); } /* * Given a node change the order of its parent's child nodes such * that the node becomes the firt child while preserving the cyclic * order of the children. In other words, the given node is promoted * by rotation. */ void topo_promote_child(struct topo_node *child) { struct topo_node *next; struct topo_node *node; struct topo_node *parent; parent = child->parent; next = TAILQ_NEXT(child, siblings); TAILQ_REMOVE(&parent->children, child, siblings); TAILQ_INSERT_HEAD(&parent->children, child, siblings); while (next != NULL) { node = next; next = TAILQ_NEXT(node, siblings); TAILQ_REMOVE(&parent->children, node, siblings); TAILQ_INSERT_AFTER(&parent->children, child, node, siblings); child = node; } } /* * Iterate to the next node in the depth-first search (traversal) of * the topology tree. */ struct topo_node * topo_next_node(struct topo_node *top, struct topo_node *node) { struct topo_node *next; if ((next = TAILQ_FIRST(&node->children)) != NULL) return (next); if ((next = TAILQ_NEXT(node, siblings)) != NULL) return (next); while (node != top && (node = node->parent) != top) if ((next = TAILQ_NEXT(node, siblings)) != NULL) return (next); return (NULL); } /* * Iterate to the next node in the depth-first search of the topology tree, * but without descending below the current node. */ struct topo_node * topo_next_nonchild_node(struct topo_node *top, struct topo_node *node) { struct topo_node *next; if ((next = TAILQ_NEXT(node, siblings)) != NULL) return (next); while (node != top && (node = node->parent) != top) if ((next = TAILQ_NEXT(node, siblings)) != NULL) return (next); return (NULL); } /* * Assign the given ID to the given topology node that represents a logical * processor. */ void topo_set_pu_id(struct topo_node *node, cpuid_t id) { KASSERT(node->type == TOPO_TYPE_PU, ("topo_set_pu_id: wrong node type: %u", node->type)); KASSERT(CPU_EMPTY(&node->cpuset) && node->cpu_count == 0, ("topo_set_pu_id: cpuset already not empty")); node->id = id; CPU_SET(id, &node->cpuset); node->cpu_count = 1; node->subtype = 1; while ((node = node->parent) != NULL) { KASSERT(!CPU_ISSET(id, &node->cpuset), ("logical ID %u is already set in node %p", id, node)); CPU_SET(id, &node->cpuset); node->cpu_count++; } } static struct topology_spec { topo_node_type type; bool match_subtype; uintptr_t subtype; } topology_level_table[TOPO_LEVEL_COUNT] = { [TOPO_LEVEL_PKG] = { .type = TOPO_TYPE_PKG, }, [TOPO_LEVEL_GROUP] = { .type = TOPO_TYPE_GROUP, }, [TOPO_LEVEL_CACHEGROUP] = { .type = TOPO_TYPE_CACHE, .match_subtype = true, .subtype = CG_SHARE_L3, }, [TOPO_LEVEL_CORE] = { .type = TOPO_TYPE_CORE, }, [TOPO_LEVEL_THREAD] = { .type = TOPO_TYPE_PU, }, }; static bool topo_analyze_table(struct topo_node *root, int all, enum topo_level level, struct topo_analysis *results) { struct topology_spec *spec; struct topo_node *node; int count; if (level >= TOPO_LEVEL_COUNT) return (true); spec = &topology_level_table[level]; count = 0; node = topo_next_node(root, root); while (node != NULL) { if (node->type != spec->type || (spec->match_subtype && node->subtype != spec->subtype)) { node = topo_next_node(root, node); continue; } if (!all && CPU_EMPTY(&node->cpuset)) { node = topo_next_nonchild_node(root, node); continue; } count++; if (!topo_analyze_table(node, all, level + 1, results)) return (false); node = topo_next_nonchild_node(root, node); } /* No explicit subgroups is essentially one subgroup. */ if (count == 0) { count = 1; if (!topo_analyze_table(root, all, level + 1, results)) return (false); } if (results->entities[level] == -1) results->entities[level] = count; else if (results->entities[level] != count) return (false); return (true); } /* * Check if the topology is uniform, that is, each package has the same number * of cores in it and each core has the same number of threads (logical * processors) in it. If so, calculate the number of packages, the number of * groups per package, the number of cachegroups per group, and the number of * logical processors per cachegroup. 'all' parameter tells whether to include * administratively disabled logical processors into the analysis. */ int topo_analyze(struct topo_node *topo_root, int all, struct topo_analysis *results) { results->entities[TOPO_LEVEL_PKG] = -1; results->entities[TOPO_LEVEL_CORE] = -1; results->entities[TOPO_LEVEL_THREAD] = -1; results->entities[TOPO_LEVEL_GROUP] = -1; results->entities[TOPO_LEVEL_CACHEGROUP] = -1; if (!topo_analyze_table(topo_root, all, TOPO_LEVEL_PKG, results)) return (0); KASSERT(results->entities[TOPO_LEVEL_PKG] > 0, ("bug in topology or analysis")); return (1); } #endif /* SMP */ diff --git a/sys/net/iflib.c b/sys/net/iflib.c index c852e31c0aac..6dcf7de8256f 100644 --- a/sys/net/iflib.c +++ b/sys/net/iflib.c @@ -1,7162 +1,7162 @@ /*- * Copyright (c) 2014-2018, Matthew Macy * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions are met: * * 1. Redistributions of source code must retain the above copyright notice, * this list of conditions and the following disclaimer. * * 2. Neither the name of Matthew Macy nor the names of its * contributors may be used to endorse or promote products derived from * this software without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" * AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE * LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE * POSSIBILITY OF SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include "opt_inet.h" #include "opt_inet6.h" #include "opt_acpi.h" #include "opt_sched.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include "ifdi_if.h" #ifdef PCI_IOV #include #endif #include /* * enable accounting of every mbuf as it comes in to and goes out of * iflib's software descriptor references */ #define MEMORY_LOGGING 0 /* * Enable mbuf vectors for compressing long mbuf chains */ /* * NB: * - Prefetching in tx cleaning should perhaps be a tunable. The distance ahead * we prefetch needs to be determined by the time spent in m_free vis a vis * the cost of a prefetch. This will of course vary based on the workload: * - NFLX's m_free path is dominated by vm-based M_EXT manipulation which * is quite expensive, thus suggesting very little prefetch. * - small packet forwarding which is just returning a single mbuf to * UMA will typically be very fast vis a vis the cost of a memory * access. */ /* * File organization: * - private structures * - iflib private utility functions * - ifnet functions * - vlan registry and other exported functions * - iflib public core functions * * */ MALLOC_DEFINE(M_IFLIB, "iflib", "ifnet library"); #define IFLIB_RXEOF_MORE (1U << 0) #define IFLIB_RXEOF_EMPTY (2U << 0) struct iflib_txq; typedef struct iflib_txq *iflib_txq_t; struct iflib_rxq; typedef struct iflib_rxq *iflib_rxq_t; struct iflib_fl; typedef struct iflib_fl *iflib_fl_t; struct iflib_ctx; static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid); static void iflib_timer(void *arg); static void iflib_tqg_detach(if_ctx_t ctx); typedef struct iflib_filter_info { driver_filter_t *ifi_filter; void *ifi_filter_arg; struct grouptask *ifi_task; void *ifi_ctx; } *iflib_filter_info_t; struct iflib_ctx { KOBJ_FIELDS; /* * Pointer to hardware driver's softc */ void *ifc_softc; device_t ifc_dev; if_t ifc_ifp; cpuset_t ifc_cpus; if_shared_ctx_t ifc_sctx; struct if_softc_ctx ifc_softc_ctx; struct sx ifc_ctx_sx; struct mtx ifc_state_mtx; iflib_txq_t ifc_txqs; iflib_rxq_t ifc_rxqs; uint32_t ifc_if_flags; uint32_t ifc_flags; uint32_t ifc_max_fl_buf_size; uint32_t ifc_rx_mbuf_sz; int ifc_link_state; int ifc_watchdog_events; struct cdev *ifc_led_dev; struct resource *ifc_msix_mem; struct if_irq ifc_legacy_irq; struct grouptask ifc_admin_task; struct grouptask ifc_vflr_task; struct iflib_filter_info ifc_filter_info; struct ifmedia ifc_media; struct ifmedia *ifc_mediap; struct sysctl_oid *ifc_sysctl_node; uint16_t ifc_sysctl_ntxqs; uint16_t ifc_sysctl_nrxqs; uint16_t ifc_sysctl_qs_eq_override; uint16_t ifc_sysctl_rx_budget; uint16_t ifc_sysctl_tx_abdicate; uint16_t ifc_sysctl_core_offset; #define CORE_OFFSET_UNSPECIFIED 0xffff uint8_t ifc_sysctl_separate_txrx; uint8_t ifc_sysctl_use_logical_cores; bool ifc_cpus_are_physical_cores; qidx_t ifc_sysctl_ntxds[8]; qidx_t ifc_sysctl_nrxds[8]; struct if_txrx ifc_txrx; #define isc_txd_encap ifc_txrx.ift_txd_encap #define isc_txd_flush ifc_txrx.ift_txd_flush #define isc_txd_credits_update ifc_txrx.ift_txd_credits_update #define isc_rxd_available ifc_txrx.ift_rxd_available #define isc_rxd_pkt_get ifc_txrx.ift_rxd_pkt_get #define isc_rxd_refill ifc_txrx.ift_rxd_refill #define isc_rxd_flush ifc_txrx.ift_rxd_flush #define isc_legacy_intr ifc_txrx.ift_legacy_intr eventhandler_tag ifc_vlan_attach_event; eventhandler_tag ifc_vlan_detach_event; struct ether_addr ifc_mac; }; void * iflib_get_softc(if_ctx_t ctx) { return (ctx->ifc_softc); } device_t iflib_get_dev(if_ctx_t ctx) { return (ctx->ifc_dev); } if_t iflib_get_ifp(if_ctx_t ctx) { return (ctx->ifc_ifp); } struct ifmedia * iflib_get_media(if_ctx_t ctx) { return (ctx->ifc_mediap); } uint32_t iflib_get_flags(if_ctx_t ctx) { return (ctx->ifc_flags); } void iflib_set_mac(if_ctx_t ctx, uint8_t mac[ETHER_ADDR_LEN]) { bcopy(mac, ctx->ifc_mac.octet, ETHER_ADDR_LEN); } if_softc_ctx_t iflib_get_softc_ctx(if_ctx_t ctx) { return (&ctx->ifc_softc_ctx); } if_shared_ctx_t iflib_get_sctx(if_ctx_t ctx) { return (ctx->ifc_sctx); } #define IP_ALIGNED(m) ((((uintptr_t)(m)->m_data) & 0x3) == 0x2) #define CACHE_PTR_INCREMENT (CACHE_LINE_SIZE/sizeof(void*)) #define CACHE_PTR_NEXT(ptr) ((void *)(((uintptr_t)(ptr)+CACHE_LINE_SIZE-1) & (CACHE_LINE_SIZE-1))) #define LINK_ACTIVE(ctx) ((ctx)->ifc_link_state == LINK_STATE_UP) #define CTX_IS_VF(ctx) ((ctx)->ifc_sctx->isc_flags & IFLIB_IS_VF) typedef struct iflib_sw_rx_desc_array { bus_dmamap_t *ifsd_map; /* bus_dma maps for packet */ struct mbuf **ifsd_m; /* pkthdr mbufs */ caddr_t *ifsd_cl; /* direct cluster pointer for rx */ bus_addr_t *ifsd_ba; /* bus addr of cluster for rx */ } iflib_rxsd_array_t; typedef struct iflib_sw_tx_desc_array { bus_dmamap_t *ifsd_map; /* bus_dma maps for packet */ bus_dmamap_t *ifsd_tso_map; /* bus_dma maps for TSO packet */ struct mbuf **ifsd_m; /* pkthdr mbufs */ } if_txsd_vec_t; /* magic number that should be high enough for any hardware */ #define IFLIB_MAX_TX_SEGS 128 #define IFLIB_RX_COPY_THRESH 128 #define IFLIB_MAX_RX_REFRESH 32 /* The minimum descriptors per second before we start coalescing */ #define IFLIB_MIN_DESC_SEC 16384 #define IFLIB_DEFAULT_TX_UPDATE_FREQ 16 #define IFLIB_QUEUE_IDLE 0 #define IFLIB_QUEUE_HUNG 1 #define IFLIB_QUEUE_WORKING 2 /* maximum number of txqs that can share an rx interrupt */ #define IFLIB_MAX_TX_SHARED_INTR 4 /* this should really scale with ring size - this is a fairly arbitrary value */ #define TX_BATCH_SIZE 32 #define IFLIB_RESTART_BUDGET 8 #define CSUM_OFFLOAD (CSUM_IP_TSO|CSUM_IP6_TSO|CSUM_IP| \ CSUM_IP_UDP|CSUM_IP_TCP|CSUM_IP_SCTP| \ CSUM_IP6_UDP|CSUM_IP6_TCP|CSUM_IP6_SCTP) struct iflib_txq { qidx_t ift_in_use; qidx_t ift_cidx; qidx_t ift_cidx_processed; qidx_t ift_pidx; uint8_t ift_gen; uint8_t ift_br_offset; uint16_t ift_npending; uint16_t ift_db_pending; uint16_t ift_rs_pending; /* implicit pad */ uint8_t ift_txd_size[8]; uint64_t ift_processed; uint64_t ift_cleaned; uint64_t ift_cleaned_prev; #if MEMORY_LOGGING uint64_t ift_enqueued; uint64_t ift_dequeued; #endif uint64_t ift_no_tx_dma_setup; uint64_t ift_no_desc_avail; uint64_t ift_mbuf_defrag_failed; uint64_t ift_mbuf_defrag; uint64_t ift_map_failed; uint64_t ift_txd_encap_efbig; uint64_t ift_pullups; uint64_t ift_last_timer_tick; struct mtx ift_mtx; struct mtx ift_db_mtx; /* constant values */ if_ctx_t ift_ctx; struct ifmp_ring *ift_br; struct grouptask ift_task; qidx_t ift_size; uint16_t ift_id; struct callout ift_timer; #ifdef DEV_NETMAP struct callout ift_netmap_timer; #endif /* DEV_NETMAP */ if_txsd_vec_t ift_sds; uint8_t ift_qstatus; uint8_t ift_closed; uint8_t ift_update_freq; struct iflib_filter_info ift_filter_info; bus_dma_tag_t ift_buf_tag; bus_dma_tag_t ift_tso_buf_tag; iflib_dma_info_t ift_ifdi; #define MTX_NAME_LEN 32 char ift_mtx_name[MTX_NAME_LEN]; bus_dma_segment_t ift_segs[IFLIB_MAX_TX_SEGS] __aligned(CACHE_LINE_SIZE); #ifdef IFLIB_DIAGNOSTICS uint64_t ift_cpu_exec_count[256]; #endif } __aligned(CACHE_LINE_SIZE); struct iflib_fl { qidx_t ifl_cidx; qidx_t ifl_pidx; qidx_t ifl_credits; uint8_t ifl_gen; uint8_t ifl_rxd_size; #if MEMORY_LOGGING uint64_t ifl_m_enqueued; uint64_t ifl_m_dequeued; uint64_t ifl_cl_enqueued; uint64_t ifl_cl_dequeued; #endif /* implicit pad */ bitstr_t *ifl_rx_bitmap; qidx_t ifl_fragidx; /* constant */ qidx_t ifl_size; uint16_t ifl_buf_size; uint16_t ifl_cltype; uma_zone_t ifl_zone; iflib_rxsd_array_t ifl_sds; iflib_rxq_t ifl_rxq; uint8_t ifl_id; bus_dma_tag_t ifl_buf_tag; iflib_dma_info_t ifl_ifdi; uint64_t ifl_bus_addrs[IFLIB_MAX_RX_REFRESH] __aligned(CACHE_LINE_SIZE); qidx_t ifl_rxd_idxs[IFLIB_MAX_RX_REFRESH]; } __aligned(CACHE_LINE_SIZE); static inline qidx_t get_inuse(int size, qidx_t cidx, qidx_t pidx, uint8_t gen) { qidx_t used; if (pidx > cidx) used = pidx - cidx; else if (pidx < cidx) used = size - cidx + pidx; else if (gen == 0 && pidx == cidx) used = 0; else if (gen == 1 && pidx == cidx) used = size; else panic("bad state"); return (used); } #define TXQ_AVAIL(txq) (txq->ift_size - get_inuse(txq->ift_size, txq->ift_cidx, txq->ift_pidx, txq->ift_gen)) #define IDXDIFF(head, tail, wrap) \ ((head) >= (tail) ? (head) - (tail) : (wrap) - (tail) + (head)) struct iflib_rxq { if_ctx_t ifr_ctx; iflib_fl_t ifr_fl; uint64_t ifr_rx_irq; struct pfil_head *pfil; /* * If there is a separate completion queue (IFLIB_HAS_RXCQ), this is * the completion queue consumer index. Otherwise it's unused. */ qidx_t ifr_cq_cidx; uint16_t ifr_id; uint8_t ifr_nfl; uint8_t ifr_ntxqirq; uint8_t ifr_txqid[IFLIB_MAX_TX_SHARED_INTR]; uint8_t ifr_fl_offset; struct lro_ctrl ifr_lc; struct grouptask ifr_task; struct callout ifr_watchdog; struct iflib_filter_info ifr_filter_info; iflib_dma_info_t ifr_ifdi; /* dynamically allocate if any drivers need a value substantially larger than this */ struct if_rxd_frag ifr_frags[IFLIB_MAX_RX_SEGS] __aligned(CACHE_LINE_SIZE); #ifdef IFLIB_DIAGNOSTICS uint64_t ifr_cpu_exec_count[256]; #endif } __aligned(CACHE_LINE_SIZE); typedef struct if_rxsd { caddr_t *ifsd_cl; iflib_fl_t ifsd_fl; } *if_rxsd_t; /* multiple of word size */ #ifdef __LP64__ #define PKT_INFO_SIZE 6 #define RXD_INFO_SIZE 5 #define PKT_TYPE uint64_t #else #define PKT_INFO_SIZE 11 #define RXD_INFO_SIZE 8 #define PKT_TYPE uint32_t #endif #define PKT_LOOP_BOUND ((PKT_INFO_SIZE/3)*3) #define RXD_LOOP_BOUND ((RXD_INFO_SIZE/4)*4) typedef struct if_pkt_info_pad { PKT_TYPE pkt_val[PKT_INFO_SIZE]; } *if_pkt_info_pad_t; typedef struct if_rxd_info_pad { PKT_TYPE rxd_val[RXD_INFO_SIZE]; } *if_rxd_info_pad_t; CTASSERT(sizeof(struct if_pkt_info_pad) == sizeof(struct if_pkt_info)); CTASSERT(sizeof(struct if_rxd_info_pad) == sizeof(struct if_rxd_info)); static inline void pkt_info_zero(if_pkt_info_t pi) { if_pkt_info_pad_t pi_pad; pi_pad = (if_pkt_info_pad_t)pi; pi_pad->pkt_val[0] = 0; pi_pad->pkt_val[1] = 0; pi_pad->pkt_val[2] = 0; pi_pad->pkt_val[3] = 0; pi_pad->pkt_val[4] = 0; pi_pad->pkt_val[5] = 0; #ifndef __LP64__ pi_pad->pkt_val[6] = 0; pi_pad->pkt_val[7] = 0; pi_pad->pkt_val[8] = 0; pi_pad->pkt_val[9] = 0; pi_pad->pkt_val[10] = 0; #endif } static device_method_t iflib_pseudo_methods[] = { DEVMETHOD(device_attach, noop_attach), DEVMETHOD(device_detach, iflib_pseudo_detach), DEVMETHOD_END }; driver_t iflib_pseudodriver = { "iflib_pseudo", iflib_pseudo_methods, sizeof(struct iflib_ctx), }; static inline void rxd_info_zero(if_rxd_info_t ri) { if_rxd_info_pad_t ri_pad; int i; ri_pad = (if_rxd_info_pad_t)ri; for (i = 0; i < RXD_LOOP_BOUND; i += 4) { ri_pad->rxd_val[i] = 0; ri_pad->rxd_val[i+1] = 0; ri_pad->rxd_val[i+2] = 0; ri_pad->rxd_val[i+3] = 0; } #ifdef __LP64__ ri_pad->rxd_val[RXD_INFO_SIZE-1] = 0; #endif } /* * Only allow a single packet to take up most 1/nth of the tx ring */ #define MAX_SINGLE_PACKET_FRACTION 12 #define IF_BAD_DMA (bus_addr_t)-1 #define CTX_ACTIVE(ctx) ((if_getdrvflags((ctx)->ifc_ifp) & IFF_DRV_RUNNING)) #define CTX_LOCK_INIT(_sc) sx_init(&(_sc)->ifc_ctx_sx, "iflib ctx lock") #define CTX_LOCK(ctx) sx_xlock(&(ctx)->ifc_ctx_sx) #define CTX_UNLOCK(ctx) sx_xunlock(&(ctx)->ifc_ctx_sx) #define CTX_LOCK_DESTROY(ctx) sx_destroy(&(ctx)->ifc_ctx_sx) #define STATE_LOCK_INIT(_sc, _name) mtx_init(&(_sc)->ifc_state_mtx, _name, "iflib state lock", MTX_DEF) #define STATE_LOCK(ctx) mtx_lock(&(ctx)->ifc_state_mtx) #define STATE_UNLOCK(ctx) mtx_unlock(&(ctx)->ifc_state_mtx) #define STATE_LOCK_DESTROY(ctx) mtx_destroy(&(ctx)->ifc_state_mtx) #define CALLOUT_LOCK(txq) mtx_lock(&txq->ift_mtx) #define CALLOUT_UNLOCK(txq) mtx_unlock(&txq->ift_mtx) void iflib_set_detach(if_ctx_t ctx) { STATE_LOCK(ctx); ctx->ifc_flags |= IFC_IN_DETACH; STATE_UNLOCK(ctx); } /* Our boot-time initialization hook */ static int iflib_module_event_handler(module_t, int, void *); static moduledata_t iflib_moduledata = { "iflib", iflib_module_event_handler, NULL }; DECLARE_MODULE(iflib, iflib_moduledata, SI_SUB_INIT_IF, SI_ORDER_ANY); MODULE_VERSION(iflib, 1); MODULE_DEPEND(iflib, pci, 1, 1, 1); MODULE_DEPEND(iflib, ether, 1, 1, 1); TASKQGROUP_DEFINE(if_io_tqg, mp_ncpus, 1); TASKQGROUP_DEFINE(if_config_tqg, 1, 1); #ifndef IFLIB_DEBUG_COUNTERS #ifdef INVARIANTS #define IFLIB_DEBUG_COUNTERS 1 #else #define IFLIB_DEBUG_COUNTERS 0 #endif /* !INVARIANTS */ #endif static SYSCTL_NODE(_net, OID_AUTO, iflib, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "iflib driver parameters"); /* * XXX need to ensure that this can't accidentally cause the head to be moved backwards */ static int iflib_min_tx_latency = 0; SYSCTL_INT(_net_iflib, OID_AUTO, min_tx_latency, CTLFLAG_RW, &iflib_min_tx_latency, 0, "minimize transmit latency at the possible expense of throughput"); static int iflib_no_tx_batch = 0; SYSCTL_INT(_net_iflib, OID_AUTO, no_tx_batch, CTLFLAG_RW, &iflib_no_tx_batch, 0, "minimize transmit latency at the possible expense of throughput"); static int iflib_timer_default = 1000; SYSCTL_INT(_net_iflib, OID_AUTO, timer_default, CTLFLAG_RW, &iflib_timer_default, 0, "number of ticks between iflib_timer calls"); #if IFLIB_DEBUG_COUNTERS static int iflib_tx_seen; static int iflib_tx_sent; static int iflib_tx_encap; static int iflib_rx_allocs; static int iflib_fl_refills; static int iflib_fl_refills_large; static int iflib_tx_frees; SYSCTL_INT(_net_iflib, OID_AUTO, tx_seen, CTLFLAG_RD, &iflib_tx_seen, 0, "# TX mbufs seen"); SYSCTL_INT(_net_iflib, OID_AUTO, tx_sent, CTLFLAG_RD, &iflib_tx_sent, 0, "# TX mbufs sent"); SYSCTL_INT(_net_iflib, OID_AUTO, tx_encap, CTLFLAG_RD, &iflib_tx_encap, 0, "# TX mbufs encapped"); SYSCTL_INT(_net_iflib, OID_AUTO, tx_frees, CTLFLAG_RD, &iflib_tx_frees, 0, "# TX frees"); SYSCTL_INT(_net_iflib, OID_AUTO, rx_allocs, CTLFLAG_RD, &iflib_rx_allocs, 0, "# RX allocations"); SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills, CTLFLAG_RD, &iflib_fl_refills, 0, "# refills"); SYSCTL_INT(_net_iflib, OID_AUTO, fl_refills_large, CTLFLAG_RD, &iflib_fl_refills_large, 0, "# large refills"); static int iflib_txq_drain_flushing; static int iflib_txq_drain_oactive; static int iflib_txq_drain_notready; SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_flushing, CTLFLAG_RD, &iflib_txq_drain_flushing, 0, "# drain flushes"); SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_oactive, CTLFLAG_RD, &iflib_txq_drain_oactive, 0, "# drain oactives"); SYSCTL_INT(_net_iflib, OID_AUTO, txq_drain_notready, CTLFLAG_RD, &iflib_txq_drain_notready, 0, "# drain notready"); static int iflib_encap_load_mbuf_fail; static int iflib_encap_pad_mbuf_fail; static int iflib_encap_txq_avail_fail; static int iflib_encap_txd_encap_fail; SYSCTL_INT(_net_iflib, OID_AUTO, encap_load_mbuf_fail, CTLFLAG_RD, &iflib_encap_load_mbuf_fail, 0, "# busdma load failures"); SYSCTL_INT(_net_iflib, OID_AUTO, encap_pad_mbuf_fail, CTLFLAG_RD, &iflib_encap_pad_mbuf_fail, 0, "# runt frame pad failures"); SYSCTL_INT(_net_iflib, OID_AUTO, encap_txq_avail_fail, CTLFLAG_RD, &iflib_encap_txq_avail_fail, 0, "# txq avail failures"); SYSCTL_INT(_net_iflib, OID_AUTO, encap_txd_encap_fail, CTLFLAG_RD, &iflib_encap_txd_encap_fail, 0, "# driver encap failures"); static int iflib_task_fn_rxs; static int iflib_rx_intr_enables; static int iflib_fast_intrs; static int iflib_rx_unavail; static int iflib_rx_ctx_inactive; static int iflib_rx_if_input; static int iflib_rxd_flush; static int iflib_verbose_debug; SYSCTL_INT(_net_iflib, OID_AUTO, task_fn_rx, CTLFLAG_RD, &iflib_task_fn_rxs, 0, "# task_fn_rx calls"); SYSCTL_INT(_net_iflib, OID_AUTO, rx_intr_enables, CTLFLAG_RD, &iflib_rx_intr_enables, 0, "# RX intr enables"); SYSCTL_INT(_net_iflib, OID_AUTO, fast_intrs, CTLFLAG_RD, &iflib_fast_intrs, 0, "# fast_intr calls"); SYSCTL_INT(_net_iflib, OID_AUTO, rx_unavail, CTLFLAG_RD, &iflib_rx_unavail, 0, "# times rxeof called with no available data"); SYSCTL_INT(_net_iflib, OID_AUTO, rx_ctx_inactive, CTLFLAG_RD, &iflib_rx_ctx_inactive, 0, "# times rxeof called with inactive context"); SYSCTL_INT(_net_iflib, OID_AUTO, rx_if_input, CTLFLAG_RD, &iflib_rx_if_input, 0, "# times rxeof called if_input"); SYSCTL_INT(_net_iflib, OID_AUTO, rxd_flush, CTLFLAG_RD, &iflib_rxd_flush, 0, "# times rxd_flush called"); SYSCTL_INT(_net_iflib, OID_AUTO, verbose_debug, CTLFLAG_RW, &iflib_verbose_debug, 0, "enable verbose debugging"); #define DBG_COUNTER_INC(name) atomic_add_int(&(iflib_ ## name), 1) static void iflib_debug_reset(void) { iflib_tx_seen = iflib_tx_sent = iflib_tx_encap = iflib_rx_allocs = iflib_fl_refills = iflib_fl_refills_large = iflib_tx_frees = iflib_txq_drain_flushing = iflib_txq_drain_oactive = iflib_txq_drain_notready = iflib_encap_load_mbuf_fail = iflib_encap_pad_mbuf_fail = iflib_encap_txq_avail_fail = iflib_encap_txd_encap_fail = iflib_task_fn_rxs = iflib_rx_intr_enables = iflib_fast_intrs = iflib_rx_unavail = iflib_rx_ctx_inactive = iflib_rx_if_input = iflib_rxd_flush = 0; } #else #define DBG_COUNTER_INC(name) static void iflib_debug_reset(void) {} #endif #define IFLIB_DEBUG 0 static void iflib_tx_structures_free(if_ctx_t ctx); static void iflib_rx_structures_free(if_ctx_t ctx); static int iflib_queues_alloc(if_ctx_t ctx); static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq); static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget); static int iflib_qset_structures_setup(if_ctx_t ctx); static int iflib_msix_init(if_ctx_t ctx); static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filterarg, int *rid, const char *str); static void iflib_txq_check_drain(iflib_txq_t txq, int budget); static uint32_t iflib_txq_can_drain(struct ifmp_ring *); #ifdef ALTQ static void iflib_altq_if_start(if_t ifp); static int iflib_altq_if_transmit(if_t ifp, struct mbuf *m); #endif static int iflib_register(if_ctx_t); static void iflib_deregister(if_ctx_t); static void iflib_unregister_vlan_handlers(if_ctx_t ctx); static uint16_t iflib_get_mbuf_size_for(unsigned int size); static void iflib_init_locked(if_ctx_t ctx); static void iflib_add_device_sysctl_pre(if_ctx_t ctx); static void iflib_add_device_sysctl_post(if_ctx_t ctx); static void iflib_ifmp_purge(iflib_txq_t txq); static void _iflib_pre_assert(if_softc_ctx_t scctx); static void iflib_if_init_locked(if_ctx_t ctx); static void iflib_free_intr_mem(if_ctx_t ctx); #ifndef __NO_STRICT_ALIGNMENT static struct mbuf * iflib_fixup_rx(struct mbuf *m); #endif static SLIST_HEAD(cpu_offset_list, cpu_offset) cpu_offsets = SLIST_HEAD_INITIALIZER(cpu_offsets); struct cpu_offset { SLIST_ENTRY(cpu_offset) entries; cpuset_t set; unsigned int refcount; uint16_t next_cpuid; }; static struct mtx cpu_offset_mtx; MTX_SYSINIT(iflib_cpu_offset, &cpu_offset_mtx, "iflib_cpu_offset lock", MTX_DEF); DEBUGNET_DEFINE(iflib); static int iflib_num_rx_descs(if_ctx_t ctx) { if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; if_shared_ctx_t sctx = ctx->ifc_sctx; uint16_t first_rxq = (sctx->isc_flags & IFLIB_HAS_RXCQ) ? 1 : 0; return scctx->isc_nrxd[first_rxq]; } static int iflib_num_tx_descs(if_ctx_t ctx) { if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; if_shared_ctx_t sctx = ctx->ifc_sctx; uint16_t first_txq = (sctx->isc_flags & IFLIB_HAS_TXCQ) ? 1 : 0; return scctx->isc_ntxd[first_txq]; } #ifdef DEV_NETMAP #include #include #include MODULE_DEPEND(iflib, netmap, 1, 1, 1); static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init); static void iflib_netmap_timer(void *arg); /* * device-specific sysctl variables: * * iflib_crcstrip: 0: keep CRC in rx frames (default), 1: strip it. * During regular operations the CRC is stripped, but on some * hardware reception of frames not multiple of 64 is slower, * so using crcstrip=0 helps in benchmarks. * * iflib_rx_miss, iflib_rx_miss_bufs: * count packets that might be missed due to lost interrupts. */ SYSCTL_DECL(_dev_netmap); /* * The xl driver by default strips CRCs and we do not override it. */ int iflib_crcstrip = 1; SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_crcstrip, CTLFLAG_RW, &iflib_crcstrip, 1, "strip CRC on RX frames"); int iflib_rx_miss, iflib_rx_miss_bufs; SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss, CTLFLAG_RW, &iflib_rx_miss, 0, "potentially missed RX intr"); SYSCTL_INT(_dev_netmap, OID_AUTO, iflib_rx_miss_bufs, CTLFLAG_RW, &iflib_rx_miss_bufs, 0, "potentially missed RX intr bufs"); /* * Register/unregister. We are already under netmap lock. * Only called on the first register or the last unregister. */ static int iflib_netmap_register(struct netmap_adapter *na, int onoff) { if_t ifp = na->ifp; if_ctx_t ctx = ifp->if_softc; int status; CTX_LOCK(ctx); if (!CTX_IS_VF(ctx)) IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); iflib_stop(ctx); /* * Enable (or disable) netmap flags, and intercept (or restore) * ifp->if_transmit. This is done once the device has been stopped * to prevent race conditions. Also, this must be done after * calling netmap_disable_all_rings() and before calling * netmap_enable_all_rings(), so that these two functions see the * updated state of the NAF_NETMAP_ON bit. */ if (onoff) { nm_set_native_flags(na); } else { nm_clear_native_flags(na); } iflib_init_locked(ctx); IFDI_CRCSTRIP_SET(ctx, onoff, iflib_crcstrip); // XXX why twice ? status = ifp->if_drv_flags & IFF_DRV_RUNNING ? 0 : 1; if (status) nm_clear_native_flags(na); CTX_UNLOCK(ctx); return (status); } static int iflib_netmap_config(struct netmap_adapter *na, struct nm_config_info *info) { if_t ifp = na->ifp; if_ctx_t ctx = ifp->if_softc; iflib_rxq_t rxq = &ctx->ifc_rxqs[0]; iflib_fl_t fl = &rxq->ifr_fl[0]; info->num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets; info->num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets; info->num_tx_descs = iflib_num_tx_descs(ctx); info->num_rx_descs = iflib_num_rx_descs(ctx); info->rx_buf_maxsize = fl->ifl_buf_size; nm_prinf("txr %u rxr %u txd %u rxd %u rbufsz %u", info->num_tx_rings, info->num_rx_rings, info->num_tx_descs, info->num_rx_descs, info->rx_buf_maxsize); return 0; } static int netmap_fl_refill(iflib_rxq_t rxq, struct netmap_kring *kring, bool init) { struct netmap_adapter *na = kring->na; u_int const lim = kring->nkr_num_slots - 1; struct netmap_ring *ring = kring->ring; bus_dmamap_t *map; struct if_rxd_update iru; if_ctx_t ctx = rxq->ifr_ctx; iflib_fl_t fl = &rxq->ifr_fl[0]; u_int nic_i_first, nic_i; u_int nm_i; int i, n; #if IFLIB_DEBUG_COUNTERS int rf_count = 0; #endif /* * This function is used both at initialization and in rxsync. * At initialization we need to prepare (with isc_rxd_refill()) * all the netmap buffers currently owned by the kernel, in * such a way to keep fl->ifl_pidx and kring->nr_hwcur in sync * (except for kring->nkr_hwofs). These may be less than * kring->nkr_num_slots if netmap_reset() was called while * an application using the kring that still owned some * buffers. * At rxsync time, both indexes point to the next buffer to be * refilled. * In any case we publish (with isc_rxd_flush()) up to * (fl->ifl_pidx - 1) % N (included), to avoid the NIC tail/prod * pointer to overrun the head/cons pointer, although this is * not necessary for some NICs (e.g. vmx). */ if (__predict_false(init)) { n = kring->nkr_num_slots - nm_kr_rxspace(kring); } else { n = kring->rhead - kring->nr_hwcur; if (n == 0) return (0); /* Nothing to do. */ if (n < 0) n += kring->nkr_num_slots; } iru_init(&iru, rxq, 0 /* flid */); map = fl->ifl_sds.ifsd_map; nic_i = fl->ifl_pidx; nm_i = netmap_idx_n2k(kring, nic_i); if (__predict_false(init)) { /* * On init/reset, nic_i must be 0, and we must * start to refill from hwtail (see netmap_reset()). */ MPASS(nic_i == 0); MPASS(nm_i == kring->nr_hwtail); } else MPASS(nm_i == kring->nr_hwcur); DBG_COUNTER_INC(fl_refills); while (n > 0) { #if IFLIB_DEBUG_COUNTERS if (++rf_count == 9) DBG_COUNTER_INC(fl_refills_large); #endif nic_i_first = nic_i; for (i = 0; n > 0 && i < IFLIB_MAX_RX_REFRESH; n--, i++) { struct netmap_slot *slot = &ring->slot[nm_i]; void *addr = PNMB(na, slot, &fl->ifl_bus_addrs[i]); MPASS(i < IFLIB_MAX_RX_REFRESH); if (addr == NETMAP_BUF_BASE(na)) /* bad buf */ return netmap_ring_reinit(kring); fl->ifl_rxd_idxs[i] = nic_i; if (__predict_false(init)) { netmap_load_map(na, fl->ifl_buf_tag, map[nic_i], addr); } else if (slot->flags & NS_BUF_CHANGED) { /* buffer has changed, reload map */ netmap_reload_map(na, fl->ifl_buf_tag, map[nic_i], addr); } bus_dmamap_sync(fl->ifl_buf_tag, map[nic_i], BUS_DMASYNC_PREREAD); slot->flags &= ~NS_BUF_CHANGED; nm_i = nm_next(nm_i, lim); nic_i = nm_next(nic_i, lim); } iru.iru_pidx = nic_i_first; iru.iru_count = i; ctx->isc_rxd_refill(ctx->ifc_softc, &iru); } fl->ifl_pidx = nic_i; /* * At the end of the loop we must have refilled everything * we could possibly refill. */ MPASS(nm_i == kring->rhead); kring->nr_hwcur = nm_i; bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map, BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE); ctx->isc_rxd_flush(ctx->ifc_softc, rxq->ifr_id, fl->ifl_id, nm_prev(nic_i, lim)); DBG_COUNTER_INC(rxd_flush); return (0); } #define NETMAP_TX_TIMER_US 90 /* * Reconcile kernel and user view of the transmit ring. * * All information is in the kring. * Userspace wants to send packets up to the one before kring->rhead, * kernel knows kring->nr_hwcur is the first unsent packet. * * Here we push packets out (as many as possible), and possibly * reclaim buffers from previously completed transmission. * * The caller (netmap) guarantees that there is only one instance * running at any time. Any interference with other driver * methods should be handled by the individual drivers. */ static int iflib_netmap_txsync(struct netmap_kring *kring, int flags) { struct netmap_adapter *na = kring->na; if_t ifp = na->ifp; struct netmap_ring *ring = kring->ring; u_int nm_i; /* index into the netmap kring */ u_int nic_i; /* index into the NIC ring */ u_int n; u_int const lim = kring->nkr_num_slots - 1; u_int const head = kring->rhead; struct if_pkt_info pi; int tx_pkts = 0, tx_bytes = 0; /* * interrupts on every tx packet are expensive so request * them every half ring, or where NS_REPORT is set */ u_int report_frequency = kring->nkr_num_slots >> 1; /* device-specific */ if_ctx_t ctx = ifp->if_softc; iflib_txq_t txq = &ctx->ifc_txqs[kring->ring_id]; bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map, BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE); /* * First part: process new packets to send. * nm_i is the current index in the netmap kring, * nic_i is the corresponding index in the NIC ring. * * If we have packets to send (nm_i != head) * iterate over the netmap ring, fetch length and update * the corresponding slot in the NIC ring. Some drivers also * need to update the buffer's physical address in the NIC slot * even NS_BUF_CHANGED is not set (PNMB computes the addresses). * * The netmap_reload_map() calls is especially expensive, * even when (as in this case) the tag is 0, so do only * when the buffer has actually changed. * * If possible do not set the report/intr bit on all slots, * but only a few times per ring or when NS_REPORT is set. * * Finally, on 10G and faster drivers, it might be useful * to prefetch the next slot and txr entry. */ nm_i = kring->nr_hwcur; if (nm_i != head) { /* we have new packets to send */ uint32_t pkt_len = 0, seg_idx = 0; int nic_i_start = -1, flags = 0; pkt_info_zero(&pi); pi.ipi_segs = txq->ift_segs; pi.ipi_qsidx = kring->ring_id; nic_i = netmap_idx_k2n(kring, nm_i); __builtin_prefetch(&ring->slot[nm_i]); __builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i]); __builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i]); for (n = 0; nm_i != head; n++) { struct netmap_slot *slot = &ring->slot[nm_i]; u_int len = slot->len; uint64_t paddr; void *addr = PNMB(na, slot, &paddr); flags |= (slot->flags & NS_REPORT || nic_i == 0 || nic_i == report_frequency) ? IPI_TX_INTR : 0; /* * If this is the first packet fragment, save the * index of the first NIC slot for later. */ if (nic_i_start < 0) nic_i_start = nic_i; pi.ipi_segs[seg_idx].ds_addr = paddr; pi.ipi_segs[seg_idx].ds_len = len; if (len) { pkt_len += len; seg_idx++; } if (!(slot->flags & NS_MOREFRAG)) { pi.ipi_len = pkt_len; pi.ipi_nsegs = seg_idx; pi.ipi_pidx = nic_i_start; pi.ipi_ndescs = 0; pi.ipi_flags = flags; /* Prepare the NIC TX ring. */ ctx->isc_txd_encap(ctx->ifc_softc, &pi); DBG_COUNTER_INC(tx_encap); /* Update transmit counters */ tx_bytes += pi.ipi_len; tx_pkts++; /* Reinit per-packet info for the next one. */ flags = seg_idx = pkt_len = 0; nic_i_start = -1; } /* prefetch for next round */ __builtin_prefetch(&ring->slot[nm_i + 1]); __builtin_prefetch(&txq->ift_sds.ifsd_m[nic_i + 1]); __builtin_prefetch(&txq->ift_sds.ifsd_map[nic_i + 1]); NM_CHECK_ADDR_LEN(na, addr, len); if (slot->flags & NS_BUF_CHANGED) { /* buffer has changed, reload map */ netmap_reload_map(na, txq->ift_buf_tag, txq->ift_sds.ifsd_map[nic_i], addr); } /* make sure changes to the buffer are synced */ bus_dmamap_sync(txq->ift_buf_tag, txq->ift_sds.ifsd_map[nic_i], BUS_DMASYNC_PREWRITE); slot->flags &= ~(NS_REPORT | NS_BUF_CHANGED | NS_MOREFRAG); nm_i = nm_next(nm_i, lim); nic_i = nm_next(nic_i, lim); } kring->nr_hwcur = nm_i; /* synchronize the NIC ring */ bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map, BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE); /* (re)start the tx unit up to slot nic_i (excluded) */ ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, nic_i); } /* * Second part: reclaim buffers for completed transmissions. * * If there are unclaimed buffers, attempt to reclaim them. * If we don't manage to reclaim them all, and TX IRQs are not in use, * trigger a per-tx-queue timer to try again later. */ if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) { if (iflib_tx_credits_update(ctx, txq)) { /* some tx completed, increment avail */ nic_i = txq->ift_cidx_processed; kring->nr_hwtail = nm_prev(netmap_idx_n2k(kring, nic_i), lim); } } if (!(ctx->ifc_flags & IFC_NETMAP_TX_IRQ)) if (kring->nr_hwtail != nm_prev(kring->nr_hwcur, lim)) { callout_reset_sbt_on(&txq->ift_netmap_timer, NETMAP_TX_TIMER_US * SBT_1US, SBT_1US, iflib_netmap_timer, txq, txq->ift_netmap_timer.c_cpu, 0); } if_inc_counter(ifp, IFCOUNTER_OBYTES, tx_bytes); if_inc_counter(ifp, IFCOUNTER_OPACKETS, tx_pkts); return (0); } /* * Reconcile kernel and user view of the receive ring. * Same as for the txsync, this routine must be efficient. * The caller guarantees a single invocations, but races against * the rest of the driver should be handled here. * * On call, kring->rhead is the first packet that userspace wants * to keep, and kring->rcur is the wakeup point. * The kernel has previously reported packets up to kring->rtail. * * If (flags & NAF_FORCE_READ) also check for incoming packets irrespective * of whether or not we received an interrupt. */ static int iflib_netmap_rxsync(struct netmap_kring *kring, int flags) { struct netmap_adapter *na = kring->na; struct netmap_ring *ring = kring->ring; if_t ifp = na->ifp; uint32_t nm_i; /* index into the netmap ring */ uint32_t nic_i; /* index into the NIC ring */ u_int n; u_int const lim = kring->nkr_num_slots - 1; int force_update = (flags & NAF_FORCE_READ) || kring->nr_kflags & NKR_PENDINTR; int i = 0, rx_bytes = 0, rx_pkts = 0; if_ctx_t ctx = ifp->if_softc; if_shared_ctx_t sctx = ctx->ifc_sctx; if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; iflib_rxq_t rxq = &ctx->ifc_rxqs[kring->ring_id]; iflib_fl_t fl = &rxq->ifr_fl[0]; struct if_rxd_info ri; qidx_t *cidxp; /* * netmap only uses free list 0, to avoid out of order consumption * of receive buffers */ bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map, BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE); /* * First part: import newly received packets. * * nm_i is the index of the next free slot in the netmap ring, * nic_i is the index of the next received packet in the NIC ring * (or in the free list 0 if IFLIB_HAS_RXCQ is set), and they may * differ in case if_init() has been called while * in netmap mode. For the receive ring we have * * nic_i = fl->ifl_cidx; * nm_i = kring->nr_hwtail (previous) * and * nm_i == (nic_i + kring->nkr_hwofs) % ring_size * * fl->ifl_cidx is set to 0 on a ring reinit */ if (netmap_no_pendintr || force_update) { uint32_t hwtail_lim = nm_prev(kring->nr_hwcur, lim); bool have_rxcq = sctx->isc_flags & IFLIB_HAS_RXCQ; int crclen = iflib_crcstrip ? 0 : 4; int error, avail; /* * For the free list consumer index, we use the same * logic as in iflib_rxeof(). */ if (have_rxcq) cidxp = &rxq->ifr_cq_cidx; else cidxp = &fl->ifl_cidx; avail = ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, *cidxp, USHRT_MAX); nic_i = fl->ifl_cidx; nm_i = netmap_idx_n2k(kring, nic_i); MPASS(nm_i == kring->nr_hwtail); for (n = 0; avail > 0 && nm_i != hwtail_lim; n++, avail--) { rxd_info_zero(&ri); ri.iri_frags = rxq->ifr_frags; ri.iri_qsidx = kring->ring_id; ri.iri_ifp = ctx->ifc_ifp; ri.iri_cidx = *cidxp; error = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri); for (i = 0; i < ri.iri_nfrags; i++) { if (error) { ring->slot[nm_i].len = 0; ring->slot[nm_i].flags = 0; } else { ring->slot[nm_i].len = ri.iri_frags[i].irf_len; if (i == (ri.iri_nfrags - 1)) { ring->slot[nm_i].len -= crclen; ring->slot[nm_i].flags = 0; /* Update receive counters */ rx_bytes += ri.iri_len; rx_pkts++; } else ring->slot[nm_i].flags = NS_MOREFRAG; } bus_dmamap_sync(fl->ifl_buf_tag, fl->ifl_sds.ifsd_map[nic_i], BUS_DMASYNC_POSTREAD); nm_i = nm_next(nm_i, lim); fl->ifl_cidx = nic_i = nm_next(nic_i, lim); } if (have_rxcq) { *cidxp = ri.iri_cidx; while (*cidxp >= scctx->isc_nrxd[0]) *cidxp -= scctx->isc_nrxd[0]; } } if (n) { /* update the state variables */ if (netmap_no_pendintr && !force_update) { /* diagnostics */ iflib_rx_miss ++; iflib_rx_miss_bufs += n; } kring->nr_hwtail = nm_i; } kring->nr_kflags &= ~NKR_PENDINTR; } /* * Second part: skip past packets that userspace has released. * (kring->nr_hwcur to head excluded), * and make the buffers available for reception. * As usual nm_i is the index in the netmap ring, * nic_i is the index in the NIC ring, and * nm_i == (nic_i + kring->nkr_hwofs) % ring_size */ netmap_fl_refill(rxq, kring, false); if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes); if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts); return (0); } static void iflib_netmap_intr(struct netmap_adapter *na, int onoff) { if_ctx_t ctx = na->ifp->if_softc; CTX_LOCK(ctx); if (onoff) { IFDI_INTR_ENABLE(ctx); } else { IFDI_INTR_DISABLE(ctx); } CTX_UNLOCK(ctx); } static int iflib_netmap_attach(if_ctx_t ctx) { struct netmap_adapter na; bzero(&na, sizeof(na)); na.ifp = ctx->ifc_ifp; na.na_flags = NAF_BDG_MAYSLEEP | NAF_MOREFRAG; MPASS(ctx->ifc_softc_ctx.isc_ntxqsets); MPASS(ctx->ifc_softc_ctx.isc_nrxqsets); na.num_tx_desc = iflib_num_tx_descs(ctx); na.num_rx_desc = iflib_num_rx_descs(ctx); na.nm_txsync = iflib_netmap_txsync; na.nm_rxsync = iflib_netmap_rxsync; na.nm_register = iflib_netmap_register; na.nm_intr = iflib_netmap_intr; na.nm_config = iflib_netmap_config; na.num_tx_rings = ctx->ifc_softc_ctx.isc_ntxqsets; na.num_rx_rings = ctx->ifc_softc_ctx.isc_nrxqsets; return (netmap_attach(&na)); } static int iflib_netmap_txq_init(if_ctx_t ctx, iflib_txq_t txq) { struct netmap_adapter *na = NA(ctx->ifc_ifp); struct netmap_slot *slot; slot = netmap_reset(na, NR_TX, txq->ift_id, 0); if (slot == NULL) return (0); for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxd[0]; i++) { /* * In netmap mode, set the map for the packet buffer. * NOTE: Some drivers (not this one) also need to set * the physical buffer address in the NIC ring. * netmap_idx_n2k() maps a nic index, i, into the corresponding * netmap slot index, si */ int si = netmap_idx_n2k(na->tx_rings[txq->ift_id], i); netmap_load_map(na, txq->ift_buf_tag, txq->ift_sds.ifsd_map[i], NMB(na, slot + si)); } return (1); } static int iflib_netmap_rxq_init(if_ctx_t ctx, iflib_rxq_t rxq) { struct netmap_adapter *na = NA(ctx->ifc_ifp); struct netmap_kring *kring; struct netmap_slot *slot; slot = netmap_reset(na, NR_RX, rxq->ifr_id, 0); if (slot == NULL) return (0); kring = na->rx_rings[rxq->ifr_id]; netmap_fl_refill(rxq, kring, true); return (1); } static void iflib_netmap_timer(void *arg) { iflib_txq_t txq = arg; if_ctx_t ctx = txq->ift_ctx; /* * Wake up the netmap application, to give it a chance to * call txsync and reclaim more completed TX buffers. */ netmap_tx_irq(ctx->ifc_ifp, txq->ift_id); } #define iflib_netmap_detach(ifp) netmap_detach(ifp) #else #define iflib_netmap_txq_init(ctx, txq) (0) #define iflib_netmap_rxq_init(ctx, rxq) (0) #define iflib_netmap_detach(ifp) #define netmap_enable_all_rings(ifp) #define netmap_disable_all_rings(ifp) #define iflib_netmap_attach(ctx) (0) #define netmap_rx_irq(ifp, qid, budget) (0) #endif #if defined(__i386__) || defined(__amd64__) static __inline void prefetch(void *x) { __asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x)); } static __inline void prefetch2cachelines(void *x) { __asm volatile("prefetcht0 %0" :: "m" (*(unsigned long *)x)); #if (CACHE_LINE_SIZE < 128) __asm volatile("prefetcht0 %0" :: "m" (*(((unsigned long *)x)+CACHE_LINE_SIZE/(sizeof(unsigned long))))); #endif } #else #define prefetch(x) #define prefetch2cachelines(x) #endif static void iru_init(if_rxd_update_t iru, iflib_rxq_t rxq, uint8_t flid) { iflib_fl_t fl; fl = &rxq->ifr_fl[flid]; iru->iru_paddrs = fl->ifl_bus_addrs; iru->iru_idxs = fl->ifl_rxd_idxs; iru->iru_qsidx = rxq->ifr_id; iru->iru_buf_size = fl->ifl_buf_size; iru->iru_flidx = fl->ifl_id; } static void _iflib_dmamap_cb(void *arg, bus_dma_segment_t *segs, int nseg, int err) { if (err) return; *(bus_addr_t *) arg = segs[0].ds_addr; } int iflib_dma_alloc_align(if_ctx_t ctx, int size, int align, iflib_dma_info_t dma, int mapflags) { int err; device_t dev = ctx->ifc_dev; err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */ align, 0, /* alignment, bounds */ BUS_SPACE_MAXADDR, /* lowaddr */ BUS_SPACE_MAXADDR, /* highaddr */ NULL, NULL, /* filter, filterarg */ size, /* maxsize */ 1, /* nsegments */ size, /* maxsegsize */ BUS_DMA_ALLOCNOW, /* flags */ NULL, /* lockfunc */ NULL, /* lockarg */ &dma->idi_tag); if (err) { device_printf(dev, "%s: bus_dma_tag_create failed: %d\n", __func__, err); goto fail_0; } err = bus_dmamem_alloc(dma->idi_tag, (void**) &dma->idi_vaddr, BUS_DMA_NOWAIT | BUS_DMA_COHERENT | BUS_DMA_ZERO, &dma->idi_map); if (err) { device_printf(dev, "%s: bus_dmamem_alloc(%ju) failed: %d\n", __func__, (uintmax_t)size, err); goto fail_1; } dma->idi_paddr = IF_BAD_DMA; err = bus_dmamap_load(dma->idi_tag, dma->idi_map, dma->idi_vaddr, size, _iflib_dmamap_cb, &dma->idi_paddr, mapflags | BUS_DMA_NOWAIT); if (err || dma->idi_paddr == IF_BAD_DMA) { device_printf(dev, "%s: bus_dmamap_load failed: %d\n", __func__, err); goto fail_2; } dma->idi_size = size; return (0); fail_2: bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map); fail_1: bus_dma_tag_destroy(dma->idi_tag); fail_0: dma->idi_tag = NULL; return (err); } int iflib_dma_alloc(if_ctx_t ctx, int size, iflib_dma_info_t dma, int mapflags) { if_shared_ctx_t sctx = ctx->ifc_sctx; KASSERT(sctx->isc_q_align != 0, ("alignment value not initialized")); return (iflib_dma_alloc_align(ctx, size, sctx->isc_q_align, dma, mapflags)); } int iflib_dma_alloc_multi(if_ctx_t ctx, int *sizes, iflib_dma_info_t *dmalist, int mapflags, int count) { int i, err; iflib_dma_info_t *dmaiter; dmaiter = dmalist; for (i = 0; i < count; i++, dmaiter++) { if ((err = iflib_dma_alloc(ctx, sizes[i], *dmaiter, mapflags)) != 0) break; } if (err) iflib_dma_free_multi(dmalist, i); return (err); } void iflib_dma_free(iflib_dma_info_t dma) { if (dma->idi_tag == NULL) return; if (dma->idi_paddr != IF_BAD_DMA) { bus_dmamap_sync(dma->idi_tag, dma->idi_map, BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE); bus_dmamap_unload(dma->idi_tag, dma->idi_map); dma->idi_paddr = IF_BAD_DMA; } if (dma->idi_vaddr != NULL) { bus_dmamem_free(dma->idi_tag, dma->idi_vaddr, dma->idi_map); dma->idi_vaddr = NULL; } bus_dma_tag_destroy(dma->idi_tag); dma->idi_tag = NULL; } void iflib_dma_free_multi(iflib_dma_info_t *dmalist, int count) { int i; iflib_dma_info_t *dmaiter = dmalist; for (i = 0; i < count; i++, dmaiter++) iflib_dma_free(*dmaiter); } static int iflib_fast_intr(void *arg) { iflib_filter_info_t info = arg; struct grouptask *gtask = info->ifi_task; int result; DBG_COUNTER_INC(fast_intrs); if (info->ifi_filter != NULL) { result = info->ifi_filter(info->ifi_filter_arg); if ((result & FILTER_SCHEDULE_THREAD) == 0) return (result); } GROUPTASK_ENQUEUE(gtask); return (FILTER_HANDLED); } static int iflib_fast_intr_rxtx(void *arg) { iflib_filter_info_t info = arg; struct grouptask *gtask = info->ifi_task; if_ctx_t ctx; iflib_rxq_t rxq = (iflib_rxq_t)info->ifi_ctx; iflib_txq_t txq; void *sc; int i, cidx, result; qidx_t txqid; bool intr_enable, intr_legacy; DBG_COUNTER_INC(fast_intrs); if (info->ifi_filter != NULL) { result = info->ifi_filter(info->ifi_filter_arg); if ((result & FILTER_SCHEDULE_THREAD) == 0) return (result); } ctx = rxq->ifr_ctx; sc = ctx->ifc_softc; intr_enable = false; intr_legacy = !!(ctx->ifc_flags & IFC_LEGACY); MPASS(rxq->ifr_ntxqirq); for (i = 0; i < rxq->ifr_ntxqirq; i++) { txqid = rxq->ifr_txqid[i]; txq = &ctx->ifc_txqs[txqid]; bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map, BUS_DMASYNC_POSTREAD); if (!ctx->isc_txd_credits_update(sc, txqid, false)) { if (intr_legacy) intr_enable = true; else IFDI_TX_QUEUE_INTR_ENABLE(ctx, txqid); continue; } GROUPTASK_ENQUEUE(&txq->ift_task); } if (ctx->ifc_sctx->isc_flags & IFLIB_HAS_RXCQ) cidx = rxq->ifr_cq_cidx; else cidx = rxq->ifr_fl[0].ifl_cidx; if (iflib_rxd_avail(ctx, rxq, cidx, 1)) GROUPTASK_ENQUEUE(gtask); else { if (intr_legacy) intr_enable = true; else IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id); DBG_COUNTER_INC(rx_intr_enables); } if (intr_enable) IFDI_INTR_ENABLE(ctx); return (FILTER_HANDLED); } static int iflib_fast_intr_ctx(void *arg) { iflib_filter_info_t info = arg; struct grouptask *gtask = info->ifi_task; int result; DBG_COUNTER_INC(fast_intrs); if (info->ifi_filter != NULL) { result = info->ifi_filter(info->ifi_filter_arg); if ((result & FILTER_SCHEDULE_THREAD) == 0) return (result); } GROUPTASK_ENQUEUE(gtask); return (FILTER_HANDLED); } static int _iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid, driver_filter_t filter, driver_intr_t handler, void *arg, const char *name) { struct resource *res; void *tag = NULL; device_t dev = ctx->ifc_dev; int flags, i, rc; flags = RF_ACTIVE; if (ctx->ifc_flags & IFC_LEGACY) flags |= RF_SHAREABLE; MPASS(rid < 512); i = rid; res = bus_alloc_resource_any(dev, SYS_RES_IRQ, &i, flags); if (res == NULL) { device_printf(dev, "failed to allocate IRQ for rid %d, name %s.\n", rid, name); return (ENOMEM); } irq->ii_res = res; KASSERT(filter == NULL || handler == NULL, ("filter and handler can't both be non-NULL")); rc = bus_setup_intr(dev, res, INTR_MPSAFE | INTR_TYPE_NET, filter, handler, arg, &tag); if (rc != 0) { device_printf(dev, "failed to setup interrupt for rid %d, name %s: %d\n", rid, name ? name : "unknown", rc); return (rc); } else if (name) bus_describe_intr(dev, res, tag, "%s", name); irq->ii_tag = tag; return (0); } /********************************************************************* * * Allocate DMA resources for TX buffers as well as memory for the TX * mbuf map. TX DMA maps (non-TSO/TSO) and TX mbuf map are kept in a * iflib_sw_tx_desc_array structure, storing all the information that * is needed to transmit a packet on the wire. This is called only * once at attach, setup is done every reset. * **********************************************************************/ static int iflib_txsd_alloc(iflib_txq_t txq) { if_ctx_t ctx = txq->ift_ctx; if_shared_ctx_t sctx = ctx->ifc_sctx; if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; device_t dev = ctx->ifc_dev; bus_size_t tsomaxsize; int err, nsegments, ntsosegments; bool tso; nsegments = scctx->isc_tx_nsegments; ntsosegments = scctx->isc_tx_tso_segments_max; tsomaxsize = scctx->isc_tx_tso_size_max; if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_VLAN_MTU) tsomaxsize += sizeof(struct ether_vlan_header); MPASS(scctx->isc_ntxd[0] > 0); MPASS(scctx->isc_ntxd[txq->ift_br_offset] > 0); MPASS(nsegments > 0); if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) { MPASS(ntsosegments > 0); MPASS(sctx->isc_tso_maxsize >= tsomaxsize); } /* * Set up DMA tags for TX buffers. */ if ((err = bus_dma_tag_create(bus_get_dma_tag(dev), 1, 0, /* alignment, bounds */ BUS_SPACE_MAXADDR, /* lowaddr */ BUS_SPACE_MAXADDR, /* highaddr */ NULL, NULL, /* filter, filterarg */ sctx->isc_tx_maxsize, /* maxsize */ nsegments, /* nsegments */ sctx->isc_tx_maxsegsize, /* maxsegsize */ 0, /* flags */ NULL, /* lockfunc */ NULL, /* lockfuncarg */ &txq->ift_buf_tag))) { device_printf(dev,"Unable to allocate TX DMA tag: %d\n", err); device_printf(dev,"maxsize: %ju nsegments: %d maxsegsize: %ju\n", (uintmax_t)sctx->isc_tx_maxsize, nsegments, (uintmax_t)sctx->isc_tx_maxsegsize); goto fail; } tso = (if_getcapabilities(ctx->ifc_ifp) & IFCAP_TSO) != 0; if (tso && (err = bus_dma_tag_create(bus_get_dma_tag(dev), 1, 0, /* alignment, bounds */ BUS_SPACE_MAXADDR, /* lowaddr */ BUS_SPACE_MAXADDR, /* highaddr */ NULL, NULL, /* filter, filterarg */ tsomaxsize, /* maxsize */ ntsosegments, /* nsegments */ sctx->isc_tso_maxsegsize,/* maxsegsize */ 0, /* flags */ NULL, /* lockfunc */ NULL, /* lockfuncarg */ &txq->ift_tso_buf_tag))) { device_printf(dev, "Unable to allocate TSO TX DMA tag: %d\n", err); goto fail; } /* Allocate memory for the TX mbuf map. */ if (!(txq->ift_sds.ifsd_m = (struct mbuf **) malloc(sizeof(struct mbuf *) * scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO))) { device_printf(dev, "Unable to allocate TX mbuf map memory\n"); err = ENOMEM; goto fail; } /* * Create the DMA maps for TX buffers. */ if ((txq->ift_sds.ifsd_map = (bus_dmamap_t *)malloc( sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) { device_printf(dev, "Unable to allocate TX buffer DMA map memory\n"); err = ENOMEM; goto fail; } if (tso && (txq->ift_sds.ifsd_tso_map = (bus_dmamap_t *)malloc( sizeof(bus_dmamap_t) * scctx->isc_ntxd[txq->ift_br_offset], M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) { device_printf(dev, "Unable to allocate TSO TX buffer map memory\n"); err = ENOMEM; goto fail; } for (int i = 0; i < scctx->isc_ntxd[txq->ift_br_offset]; i++) { err = bus_dmamap_create(txq->ift_buf_tag, 0, &txq->ift_sds.ifsd_map[i]); if (err != 0) { device_printf(dev, "Unable to create TX DMA map\n"); goto fail; } if (!tso) continue; err = bus_dmamap_create(txq->ift_tso_buf_tag, 0, &txq->ift_sds.ifsd_tso_map[i]); if (err != 0) { device_printf(dev, "Unable to create TSO TX DMA map\n"); goto fail; } } return (0); fail: /* We free all, it handles case where we are in the middle */ iflib_tx_structures_free(ctx); return (err); } static void iflib_txsd_destroy(if_ctx_t ctx, iflib_txq_t txq, int i) { bus_dmamap_t map; if (txq->ift_sds.ifsd_map != NULL) { map = txq->ift_sds.ifsd_map[i]; bus_dmamap_sync(txq->ift_buf_tag, map, BUS_DMASYNC_POSTWRITE); bus_dmamap_unload(txq->ift_buf_tag, map); bus_dmamap_destroy(txq->ift_buf_tag, map); txq->ift_sds.ifsd_map[i] = NULL; } if (txq->ift_sds.ifsd_tso_map != NULL) { map = txq->ift_sds.ifsd_tso_map[i]; bus_dmamap_sync(txq->ift_tso_buf_tag, map, BUS_DMASYNC_POSTWRITE); bus_dmamap_unload(txq->ift_tso_buf_tag, map); bus_dmamap_destroy(txq->ift_tso_buf_tag, map); txq->ift_sds.ifsd_tso_map[i] = NULL; } } static void iflib_txq_destroy(iflib_txq_t txq) { if_ctx_t ctx = txq->ift_ctx; for (int i = 0; i < txq->ift_size; i++) iflib_txsd_destroy(ctx, txq, i); if (txq->ift_br != NULL) { ifmp_ring_free(txq->ift_br); txq->ift_br = NULL; } mtx_destroy(&txq->ift_mtx); if (txq->ift_sds.ifsd_map != NULL) { free(txq->ift_sds.ifsd_map, M_IFLIB); txq->ift_sds.ifsd_map = NULL; } if (txq->ift_sds.ifsd_tso_map != NULL) { free(txq->ift_sds.ifsd_tso_map, M_IFLIB); txq->ift_sds.ifsd_tso_map = NULL; } if (txq->ift_sds.ifsd_m != NULL) { free(txq->ift_sds.ifsd_m, M_IFLIB); txq->ift_sds.ifsd_m = NULL; } if (txq->ift_buf_tag != NULL) { bus_dma_tag_destroy(txq->ift_buf_tag); txq->ift_buf_tag = NULL; } if (txq->ift_tso_buf_tag != NULL) { bus_dma_tag_destroy(txq->ift_tso_buf_tag); txq->ift_tso_buf_tag = NULL; } if (txq->ift_ifdi != NULL) { free(txq->ift_ifdi, M_IFLIB); } } static void iflib_txsd_free(if_ctx_t ctx, iflib_txq_t txq, int i) { struct mbuf **mp; mp = &txq->ift_sds.ifsd_m[i]; if (*mp == NULL) return; if (txq->ift_sds.ifsd_map != NULL) { bus_dmamap_sync(txq->ift_buf_tag, txq->ift_sds.ifsd_map[i], BUS_DMASYNC_POSTWRITE); bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[i]); } if (txq->ift_sds.ifsd_tso_map != NULL) { bus_dmamap_sync(txq->ift_tso_buf_tag, txq->ift_sds.ifsd_tso_map[i], BUS_DMASYNC_POSTWRITE); bus_dmamap_unload(txq->ift_tso_buf_tag, txq->ift_sds.ifsd_tso_map[i]); } m_freem(*mp); DBG_COUNTER_INC(tx_frees); *mp = NULL; } static int iflib_txq_setup(iflib_txq_t txq) { if_ctx_t ctx = txq->ift_ctx; if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; if_shared_ctx_t sctx = ctx->ifc_sctx; iflib_dma_info_t di; int i; /* Set number of descriptors available */ txq->ift_qstatus = IFLIB_QUEUE_IDLE; /* XXX make configurable */ txq->ift_update_freq = IFLIB_DEFAULT_TX_UPDATE_FREQ; /* Reset indices */ txq->ift_cidx_processed = 0; txq->ift_pidx = txq->ift_cidx = txq->ift_npending = 0; txq->ift_size = scctx->isc_ntxd[txq->ift_br_offset]; for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++) bzero((void *)di->idi_vaddr, di->idi_size); IFDI_TXQ_SETUP(ctx, txq->ift_id); for (i = 0, di = txq->ift_ifdi; i < sctx->isc_ntxqs; i++, di++) bus_dmamap_sync(di->idi_tag, di->idi_map, BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE); return (0); } /********************************************************************* * * Allocate DMA resources for RX buffers as well as memory for the RX * mbuf map, direct RX cluster pointer map and RX cluster bus address * map. RX DMA map, RX mbuf map, direct RX cluster pointer map and * RX cluster map are kept in a iflib_sw_rx_desc_array structure. * Since we use use one entry in iflib_sw_rx_desc_array per received * packet, the maximum number of entries we'll need is equal to the * number of hardware receive descriptors that we've allocated. * **********************************************************************/ static int iflib_rxsd_alloc(iflib_rxq_t rxq) { if_ctx_t ctx = rxq->ifr_ctx; if_shared_ctx_t sctx = ctx->ifc_sctx; if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; device_t dev = ctx->ifc_dev; iflib_fl_t fl; int err; MPASS(scctx->isc_nrxd[0] > 0); MPASS(scctx->isc_nrxd[rxq->ifr_fl_offset] > 0); fl = rxq->ifr_fl; for (int i = 0; i < rxq->ifr_nfl; i++, fl++) { fl->ifl_size = scctx->isc_nrxd[rxq->ifr_fl_offset]; /* this isn't necessarily the same */ /* Set up DMA tag for RX buffers. */ err = bus_dma_tag_create(bus_get_dma_tag(dev), /* parent */ 1, 0, /* alignment, bounds */ BUS_SPACE_MAXADDR, /* lowaddr */ BUS_SPACE_MAXADDR, /* highaddr */ NULL, NULL, /* filter, filterarg */ sctx->isc_rx_maxsize, /* maxsize */ sctx->isc_rx_nsegments, /* nsegments */ sctx->isc_rx_maxsegsize, /* maxsegsize */ 0, /* flags */ NULL, /* lockfunc */ NULL, /* lockarg */ &fl->ifl_buf_tag); if (err) { device_printf(dev, "Unable to allocate RX DMA tag: %d\n", err); goto fail; } /* Allocate memory for the RX mbuf map. */ if (!(fl->ifl_sds.ifsd_m = (struct mbuf **) malloc(sizeof(struct mbuf *) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) { device_printf(dev, "Unable to allocate RX mbuf map memory\n"); err = ENOMEM; goto fail; } /* Allocate memory for the direct RX cluster pointer map. */ if (!(fl->ifl_sds.ifsd_cl = (caddr_t *) malloc(sizeof(caddr_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) { device_printf(dev, "Unable to allocate RX cluster map memory\n"); err = ENOMEM; goto fail; } /* Allocate memory for the RX cluster bus address map. */ if (!(fl->ifl_sds.ifsd_ba = (bus_addr_t *) malloc(sizeof(bus_addr_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) { device_printf(dev, "Unable to allocate RX bus address map memory\n"); err = ENOMEM; goto fail; } /* * Create the DMA maps for RX buffers. */ if (!(fl->ifl_sds.ifsd_map = (bus_dmamap_t *) malloc(sizeof(bus_dmamap_t) * scctx->isc_nrxd[rxq->ifr_fl_offset], M_IFLIB, M_NOWAIT | M_ZERO))) { device_printf(dev, "Unable to allocate RX buffer DMA map memory\n"); err = ENOMEM; goto fail; } for (int i = 0; i < scctx->isc_nrxd[rxq->ifr_fl_offset]; i++) { err = bus_dmamap_create(fl->ifl_buf_tag, 0, &fl->ifl_sds.ifsd_map[i]); if (err != 0) { device_printf(dev, "Unable to create RX buffer DMA map\n"); goto fail; } } } return (0); fail: iflib_rx_structures_free(ctx); return (err); } /* * Internal service routines */ struct rxq_refill_cb_arg { int error; bus_dma_segment_t seg; int nseg; }; static void _rxq_refill_cb(void *arg, bus_dma_segment_t *segs, int nseg, int error) { struct rxq_refill_cb_arg *cb_arg = arg; cb_arg->error = error; cb_arg->seg = segs[0]; cb_arg->nseg = nseg; } /** * iflib_fl_refill - refill an rxq free-buffer list * @ctx: the iflib context * @fl: the free list to refill * @count: the number of new buffers to allocate * * (Re)populate an rxq free-buffer list with up to @count new packet buffers. * The caller must assure that @count does not exceed the queue's capacity * minus one (since we always leave a descriptor unavailable). */ static uint8_t iflib_fl_refill(if_ctx_t ctx, iflib_fl_t fl, int count) { struct if_rxd_update iru; struct rxq_refill_cb_arg cb_arg; struct mbuf *m; caddr_t cl, *sd_cl; struct mbuf **sd_m; bus_dmamap_t *sd_map; bus_addr_t bus_addr, *sd_ba; int err, frag_idx, i, idx, n, pidx; qidx_t credits; MPASS(count <= fl->ifl_size - fl->ifl_credits - 1); sd_m = fl->ifl_sds.ifsd_m; sd_map = fl->ifl_sds.ifsd_map; sd_cl = fl->ifl_sds.ifsd_cl; sd_ba = fl->ifl_sds.ifsd_ba; pidx = fl->ifl_pidx; idx = pidx; frag_idx = fl->ifl_fragidx; credits = fl->ifl_credits; i = 0; n = count; MPASS(n > 0); MPASS(credits + n <= fl->ifl_size); if (pidx < fl->ifl_cidx) MPASS(pidx + n <= fl->ifl_cidx); if (pidx == fl->ifl_cidx && (credits < fl->ifl_size)) MPASS(fl->ifl_gen == 0); if (pidx > fl->ifl_cidx) MPASS(n <= fl->ifl_size - pidx + fl->ifl_cidx); DBG_COUNTER_INC(fl_refills); if (n > 8) DBG_COUNTER_INC(fl_refills_large); iru_init(&iru, fl->ifl_rxq, fl->ifl_id); while (n-- > 0) { /* * We allocate an uninitialized mbuf + cluster, mbuf is * initialized after rx. * * If the cluster is still set then we know a minimum sized * packet was received */ bit_ffc_at(fl->ifl_rx_bitmap, frag_idx, fl->ifl_size, &frag_idx); if (frag_idx < 0) bit_ffc(fl->ifl_rx_bitmap, fl->ifl_size, &frag_idx); MPASS(frag_idx >= 0); if ((cl = sd_cl[frag_idx]) == NULL) { cl = uma_zalloc(fl->ifl_zone, M_NOWAIT); if (__predict_false(cl == NULL)) break; cb_arg.error = 0; MPASS(sd_map != NULL); err = bus_dmamap_load(fl->ifl_buf_tag, sd_map[frag_idx], cl, fl->ifl_buf_size, _rxq_refill_cb, &cb_arg, BUS_DMA_NOWAIT); if (__predict_false(err != 0 || cb_arg.error)) { uma_zfree(fl->ifl_zone, cl); break; } sd_ba[frag_idx] = bus_addr = cb_arg.seg.ds_addr; sd_cl[frag_idx] = cl; #if MEMORY_LOGGING fl->ifl_cl_enqueued++; #endif } else { bus_addr = sd_ba[frag_idx]; } bus_dmamap_sync(fl->ifl_buf_tag, sd_map[frag_idx], BUS_DMASYNC_PREREAD); if (sd_m[frag_idx] == NULL) { m = m_gethdr(M_NOWAIT, MT_NOINIT); if (__predict_false(m == NULL)) break; sd_m[frag_idx] = m; } bit_set(fl->ifl_rx_bitmap, frag_idx); #if MEMORY_LOGGING fl->ifl_m_enqueued++; #endif DBG_COUNTER_INC(rx_allocs); fl->ifl_rxd_idxs[i] = frag_idx; fl->ifl_bus_addrs[i] = bus_addr; credits++; i++; MPASS(credits <= fl->ifl_size); if (++idx == fl->ifl_size) { #ifdef INVARIANTS fl->ifl_gen = 1; #endif idx = 0; } if (n == 0 || i == IFLIB_MAX_RX_REFRESH) { iru.iru_pidx = pidx; iru.iru_count = i; ctx->isc_rxd_refill(ctx->ifc_softc, &iru); fl->ifl_pidx = idx; fl->ifl_credits = credits; pidx = idx; i = 0; } } if (n < count - 1) { if (i != 0) { iru.iru_pidx = pidx; iru.iru_count = i; ctx->isc_rxd_refill(ctx->ifc_softc, &iru); fl->ifl_pidx = idx; fl->ifl_credits = credits; } DBG_COUNTER_INC(rxd_flush); bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map, BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE); ctx->isc_rxd_flush(ctx->ifc_softc, fl->ifl_rxq->ifr_id, fl->ifl_id, fl->ifl_pidx); if (__predict_true(bit_test(fl->ifl_rx_bitmap, frag_idx))) { fl->ifl_fragidx = frag_idx + 1; if (fl->ifl_fragidx == fl->ifl_size) fl->ifl_fragidx = 0; } else { fl->ifl_fragidx = frag_idx; } } return (n == -1 ? 0 : IFLIB_RXEOF_EMPTY); } static inline uint8_t iflib_fl_refill_all(if_ctx_t ctx, iflib_fl_t fl) { /* * We leave an unused descriptor to avoid pidx to catch up with cidx. * This is important as it confuses most NICs. For instance, * Intel NICs have (per receive ring) RDH and RDT registers, where * RDH points to the next receive descriptor to be used by the NIC, * and RDT for the next receive descriptor to be published by the * driver to the NIC (RDT - 1 is thus the last valid one). * The condition RDH == RDT means no descriptors are available to * the NIC, and thus it would be ambiguous if it also meant that * all the descriptors are available to the NIC. */ int32_t reclaimable = fl->ifl_size - fl->ifl_credits - 1; #ifdef INVARIANTS int32_t delta = fl->ifl_size - get_inuse(fl->ifl_size, fl->ifl_cidx, fl->ifl_pidx, fl->ifl_gen) - 1; #endif MPASS(fl->ifl_credits <= fl->ifl_size); MPASS(reclaimable == delta); if (reclaimable > 0) return (iflib_fl_refill(ctx, fl, reclaimable)); return (0); } uint8_t iflib_in_detach(if_ctx_t ctx) { bool in_detach; STATE_LOCK(ctx); in_detach = !!(ctx->ifc_flags & IFC_IN_DETACH); STATE_UNLOCK(ctx); return (in_detach); } static void iflib_fl_bufs_free(iflib_fl_t fl) { iflib_dma_info_t idi = fl->ifl_ifdi; bus_dmamap_t sd_map; uint32_t i; for (i = 0; i < fl->ifl_size; i++) { struct mbuf **sd_m = &fl->ifl_sds.ifsd_m[i]; caddr_t *sd_cl = &fl->ifl_sds.ifsd_cl[i]; if (*sd_cl != NULL) { sd_map = fl->ifl_sds.ifsd_map[i]; bus_dmamap_sync(fl->ifl_buf_tag, sd_map, BUS_DMASYNC_POSTREAD); bus_dmamap_unload(fl->ifl_buf_tag, sd_map); uma_zfree(fl->ifl_zone, *sd_cl); *sd_cl = NULL; if (*sd_m != NULL) { m_init(*sd_m, M_NOWAIT, MT_DATA, 0); m_free_raw(*sd_m); *sd_m = NULL; } } else { MPASS(*sd_m == NULL); } #if MEMORY_LOGGING fl->ifl_m_dequeued++; fl->ifl_cl_dequeued++; #endif } #ifdef INVARIANTS for (i = 0; i < fl->ifl_size; i++) { MPASS(fl->ifl_sds.ifsd_cl[i] == NULL); MPASS(fl->ifl_sds.ifsd_m[i] == NULL); } #endif /* * Reset free list values */ fl->ifl_credits = fl->ifl_cidx = fl->ifl_pidx = fl->ifl_gen = fl->ifl_fragidx = 0; bzero(idi->idi_vaddr, idi->idi_size); } /********************************************************************* * * Initialize a free list and its buffers. * **********************************************************************/ static int iflib_fl_setup(iflib_fl_t fl) { iflib_rxq_t rxq = fl->ifl_rxq; if_ctx_t ctx = rxq->ifr_ctx; if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; int qidx; bit_nclear(fl->ifl_rx_bitmap, 0, fl->ifl_size - 1); /* ** Free current RX buffer structs and their mbufs */ iflib_fl_bufs_free(fl); /* Now replenish the mbufs */ MPASS(fl->ifl_credits == 0); qidx = rxq->ifr_fl_offset + fl->ifl_id; if (scctx->isc_rxd_buf_size[qidx] != 0) fl->ifl_buf_size = scctx->isc_rxd_buf_size[qidx]; else fl->ifl_buf_size = ctx->ifc_rx_mbuf_sz; /* * ifl_buf_size may be a driver-supplied value, so pull it up * to the selected mbuf size. */ fl->ifl_buf_size = iflib_get_mbuf_size_for(fl->ifl_buf_size); if (fl->ifl_buf_size > ctx->ifc_max_fl_buf_size) ctx->ifc_max_fl_buf_size = fl->ifl_buf_size; fl->ifl_cltype = m_gettype(fl->ifl_buf_size); fl->ifl_zone = m_getzone(fl->ifl_buf_size); /* * Avoid pre-allocating zillions of clusters to an idle card * potentially speeding up attach. In any case make sure * to leave a descriptor unavailable. See the comment in * iflib_fl_refill_all(). */ MPASS(fl->ifl_size > 0); (void)iflib_fl_refill(ctx, fl, min(128, fl->ifl_size - 1)); if (min(128, fl->ifl_size - 1) != fl->ifl_credits) return (ENOBUFS); /* * handle failure */ MPASS(rxq != NULL); MPASS(fl->ifl_ifdi != NULL); bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map, BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE); return (0); } /********************************************************************* * * Free receive ring data structures * **********************************************************************/ static void iflib_rx_sds_free(iflib_rxq_t rxq) { iflib_fl_t fl; int i, j; if (rxq->ifr_fl != NULL) { for (i = 0; i < rxq->ifr_nfl; i++) { fl = &rxq->ifr_fl[i]; if (fl->ifl_buf_tag != NULL) { if (fl->ifl_sds.ifsd_map != NULL) { for (j = 0; j < fl->ifl_size; j++) { bus_dmamap_sync( fl->ifl_buf_tag, fl->ifl_sds.ifsd_map[j], BUS_DMASYNC_POSTREAD); bus_dmamap_unload( fl->ifl_buf_tag, fl->ifl_sds.ifsd_map[j]); bus_dmamap_destroy( fl->ifl_buf_tag, fl->ifl_sds.ifsd_map[j]); } } bus_dma_tag_destroy(fl->ifl_buf_tag); fl->ifl_buf_tag = NULL; } free(fl->ifl_sds.ifsd_m, M_IFLIB); free(fl->ifl_sds.ifsd_cl, M_IFLIB); free(fl->ifl_sds.ifsd_ba, M_IFLIB); free(fl->ifl_sds.ifsd_map, M_IFLIB); free(fl->ifl_rx_bitmap, M_IFLIB); fl->ifl_sds.ifsd_m = NULL; fl->ifl_sds.ifsd_cl = NULL; fl->ifl_sds.ifsd_ba = NULL; fl->ifl_sds.ifsd_map = NULL; fl->ifl_rx_bitmap = NULL; } free(rxq->ifr_fl, M_IFLIB); rxq->ifr_fl = NULL; free(rxq->ifr_ifdi, M_IFLIB); rxq->ifr_ifdi = NULL; rxq->ifr_cq_cidx = 0; } } /* * Timer routine */ static void iflib_timer(void *arg) { iflib_txq_t txq = arg; if_ctx_t ctx = txq->ift_ctx; if_softc_ctx_t sctx = &ctx->ifc_softc_ctx; uint64_t this_tick = ticks; if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING)) return; /* ** Check on the state of the TX queue(s), this ** can be done without the lock because its RO ** and the HUNG state will be static if set. */ if (this_tick - txq->ift_last_timer_tick >= iflib_timer_default) { txq->ift_last_timer_tick = this_tick; IFDI_TIMER(ctx, txq->ift_id); if ((txq->ift_qstatus == IFLIB_QUEUE_HUNG) && ((txq->ift_cleaned_prev == txq->ift_cleaned) || (sctx->isc_pause_frames == 0))) goto hung; if (txq->ift_qstatus != IFLIB_QUEUE_IDLE && ifmp_ring_is_stalled(txq->ift_br)) { KASSERT(ctx->ifc_link_state == LINK_STATE_UP, ("queue can't be marked as hung if interface is down")); txq->ift_qstatus = IFLIB_QUEUE_HUNG; } txq->ift_cleaned_prev = txq->ift_cleaned; } /* handle any laggards */ if (txq->ift_db_pending) GROUPTASK_ENQUEUE(&txq->ift_task); sctx->isc_pause_frames = 0; if (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer, txq, txq->ift_timer.c_cpu); return; hung: device_printf(ctx->ifc_dev, "Watchdog timeout (TX: %d desc avail: %d pidx: %d) -- resetting\n", txq->ift_id, TXQ_AVAIL(txq), txq->ift_pidx); STATE_LOCK(ctx); if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING); ctx->ifc_flags |= (IFC_DO_WATCHDOG|IFC_DO_RESET); iflib_admin_intr_deferred(ctx); STATE_UNLOCK(ctx); } static uint16_t iflib_get_mbuf_size_for(unsigned int size) { if (size <= MCLBYTES) return (MCLBYTES); else return (MJUMPAGESIZE); } static void iflib_calc_rx_mbuf_sz(if_ctx_t ctx) { if_softc_ctx_t sctx = &ctx->ifc_softc_ctx; /* * XXX don't set the max_frame_size to larger * than the hardware can handle */ ctx->ifc_rx_mbuf_sz = iflib_get_mbuf_size_for(sctx->isc_max_frame_size); } uint32_t iflib_get_rx_mbuf_sz(if_ctx_t ctx) { return (ctx->ifc_rx_mbuf_sz); } static void iflib_init_locked(if_ctx_t ctx) { if_softc_ctx_t sctx = &ctx->ifc_softc_ctx; if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; if_t ifp = ctx->ifc_ifp; iflib_fl_t fl; iflib_txq_t txq; iflib_rxq_t rxq; int i, j, tx_ip_csum_flags, tx_ip6_csum_flags; if_setdrvflagbits(ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING); IFDI_INTR_DISABLE(ctx); /* * See iflib_stop(). Useful in case iflib_init_locked() is * called without first calling iflib_stop(). */ netmap_disable_all_rings(ifp); tx_ip_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP | CSUM_TCP | CSUM_UDP | CSUM_SCTP); tx_ip6_csum_flags = scctx->isc_tx_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_UDP | CSUM_IP6_SCTP); /* Set hardware offload abilities */ if_clearhwassist(ifp); if (if_getcapenable(ifp) & IFCAP_TXCSUM) if_sethwassistbits(ifp, tx_ip_csum_flags, 0); if (if_getcapenable(ifp) & IFCAP_TXCSUM_IPV6) if_sethwassistbits(ifp, tx_ip6_csum_flags, 0); if (if_getcapenable(ifp) & IFCAP_TSO4) if_sethwassistbits(ifp, CSUM_IP_TSO, 0); if (if_getcapenable(ifp) & IFCAP_TSO6) if_sethwassistbits(ifp, CSUM_IP6_TSO, 0); for (i = 0, txq = ctx->ifc_txqs; i < sctx->isc_ntxqsets; i++, txq++) { CALLOUT_LOCK(txq); callout_stop(&txq->ift_timer); #ifdef DEV_NETMAP callout_stop(&txq->ift_netmap_timer); #endif /* DEV_NETMAP */ CALLOUT_UNLOCK(txq); iflib_netmap_txq_init(ctx, txq); } /* * Calculate a suitable Rx mbuf size prior to calling IFDI_INIT, so * that drivers can use the value when setting up the hardware receive * buffers. */ iflib_calc_rx_mbuf_sz(ctx); #ifdef INVARIANTS i = if_getdrvflags(ifp); #endif IFDI_INIT(ctx); MPASS(if_getdrvflags(ifp) == i); for (i = 0, rxq = ctx->ifc_rxqs; i < sctx->isc_nrxqsets; i++, rxq++) { if (iflib_netmap_rxq_init(ctx, rxq) > 0) { /* This rxq is in netmap mode. Skip normal init. */ continue; } for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) { if (iflib_fl_setup(fl)) { device_printf(ctx->ifc_dev, "setting up free list %d failed - " "check cluster settings\n", j); goto done; } } } done: if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_RUNNING, IFF_DRV_OACTIVE); IFDI_INTR_ENABLE(ctx); txq = ctx->ifc_txqs; for (i = 0; i < sctx->isc_ntxqsets; i++, txq++) callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer, txq, txq->ift_timer.c_cpu); /* Re-enable txsync/rxsync. */ netmap_enable_all_rings(ifp); } static int iflib_media_change(if_t ifp) { if_ctx_t ctx = if_getsoftc(ifp); int err; CTX_LOCK(ctx); if ((err = IFDI_MEDIA_CHANGE(ctx)) == 0) iflib_if_init_locked(ctx); CTX_UNLOCK(ctx); return (err); } static void iflib_media_status(if_t ifp, struct ifmediareq *ifmr) { if_ctx_t ctx = if_getsoftc(ifp); CTX_LOCK(ctx); IFDI_UPDATE_ADMIN_STATUS(ctx); IFDI_MEDIA_STATUS(ctx, ifmr); CTX_UNLOCK(ctx); } void iflib_stop(if_ctx_t ctx) { iflib_txq_t txq = ctx->ifc_txqs; iflib_rxq_t rxq = ctx->ifc_rxqs; if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; if_shared_ctx_t sctx = ctx->ifc_sctx; iflib_dma_info_t di; iflib_fl_t fl; int i, j; /* Tell the stack that the interface is no longer active */ if_setdrvflagbits(ctx->ifc_ifp, IFF_DRV_OACTIVE, IFF_DRV_RUNNING); IFDI_INTR_DISABLE(ctx); DELAY(1000); IFDI_STOP(ctx); DELAY(1000); /* * Stop any pending txsync/rxsync and prevent new ones * form starting. Processes blocked in poll() will get * POLLERR. */ netmap_disable_all_rings(ctx->ifc_ifp); iflib_debug_reset(); /* Wait for current tx queue users to exit to disarm watchdog timer. */ for (i = 0; i < scctx->isc_ntxqsets; i++, txq++) { /* make sure all transmitters have completed before proceeding XXX */ CALLOUT_LOCK(txq); callout_stop(&txq->ift_timer); #ifdef DEV_NETMAP callout_stop(&txq->ift_netmap_timer); #endif /* DEV_NETMAP */ CALLOUT_UNLOCK(txq); /* clean any enqueued buffers */ iflib_ifmp_purge(txq); /* Free any existing tx buffers. */ for (j = 0; j < txq->ift_size; j++) { iflib_txsd_free(ctx, txq, j); } txq->ift_processed = txq->ift_cleaned = txq->ift_cidx_processed = 0; txq->ift_in_use = txq->ift_gen = txq->ift_cidx = txq->ift_pidx = txq->ift_no_desc_avail = 0; txq->ift_closed = txq->ift_mbuf_defrag = txq->ift_mbuf_defrag_failed = 0; txq->ift_no_tx_dma_setup = txq->ift_txd_encap_efbig = txq->ift_map_failed = 0; txq->ift_pullups = 0; ifmp_ring_reset_stats(txq->ift_br); for (j = 0, di = txq->ift_ifdi; j < sctx->isc_ntxqs; j++, di++) bzero((void *)di->idi_vaddr, di->idi_size); } for (i = 0; i < scctx->isc_nrxqsets; i++, rxq++) { gtaskqueue_drain(rxq->ifr_task.gt_taskqueue, &rxq->ifr_task.gt_task); rxq->ifr_cq_cidx = 0; for (j = 0, di = rxq->ifr_ifdi; j < sctx->isc_nrxqs; j++, di++) bzero((void *)di->idi_vaddr, di->idi_size); /* also resets the free lists pidx/cidx */ for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) iflib_fl_bufs_free(fl); } } static inline caddr_t calc_next_rxd(iflib_fl_t fl, int cidx) { qidx_t size; int nrxd; caddr_t start, end, cur, next; nrxd = fl->ifl_size; size = fl->ifl_rxd_size; start = fl->ifl_ifdi->idi_vaddr; if (__predict_false(size == 0)) return (start); cur = start + size*cidx; end = start + size*nrxd; next = CACHE_PTR_NEXT(cur); return (next < end ? next : start); } static inline void prefetch_pkts(iflib_fl_t fl, int cidx) { int nextptr; int nrxd = fl->ifl_size; caddr_t next_rxd; nextptr = (cidx + CACHE_PTR_INCREMENT) & (nrxd-1); prefetch(&fl->ifl_sds.ifsd_m[nextptr]); prefetch(&fl->ifl_sds.ifsd_cl[nextptr]); next_rxd = calc_next_rxd(fl, cidx); prefetch(next_rxd); prefetch(fl->ifl_sds.ifsd_m[(cidx + 1) & (nrxd-1)]); prefetch(fl->ifl_sds.ifsd_m[(cidx + 2) & (nrxd-1)]); prefetch(fl->ifl_sds.ifsd_m[(cidx + 3) & (nrxd-1)]); prefetch(fl->ifl_sds.ifsd_m[(cidx + 4) & (nrxd-1)]); prefetch(fl->ifl_sds.ifsd_cl[(cidx + 1) & (nrxd-1)]); prefetch(fl->ifl_sds.ifsd_cl[(cidx + 2) & (nrxd-1)]); prefetch(fl->ifl_sds.ifsd_cl[(cidx + 3) & (nrxd-1)]); prefetch(fl->ifl_sds.ifsd_cl[(cidx + 4) & (nrxd-1)]); } static struct mbuf * rxd_frag_to_sd(iflib_rxq_t rxq, if_rxd_frag_t irf, bool unload, if_rxsd_t sd, int *pf_rv, if_rxd_info_t ri) { bus_dmamap_t map; iflib_fl_t fl; caddr_t payload; struct mbuf *m; int flid, cidx, len, next; map = NULL; flid = irf->irf_flid; cidx = irf->irf_idx; fl = &rxq->ifr_fl[flid]; sd->ifsd_fl = fl; m = fl->ifl_sds.ifsd_m[cidx]; sd->ifsd_cl = &fl->ifl_sds.ifsd_cl[cidx]; fl->ifl_credits--; #if MEMORY_LOGGING fl->ifl_m_dequeued++; #endif if (rxq->ifr_ctx->ifc_flags & IFC_PREFETCH) prefetch_pkts(fl, cidx); next = (cidx + CACHE_PTR_INCREMENT) & (fl->ifl_size-1); prefetch(&fl->ifl_sds.ifsd_map[next]); map = fl->ifl_sds.ifsd_map[cidx]; bus_dmamap_sync(fl->ifl_buf_tag, map, BUS_DMASYNC_POSTREAD); if (rxq->pfil != NULL && PFIL_HOOKED_IN(rxq->pfil) && pf_rv != NULL && irf->irf_len != 0) { payload = *sd->ifsd_cl; payload += ri->iri_pad; len = ri->iri_len - ri->iri_pad; *pf_rv = pfil_run_hooks(rxq->pfil, payload, ri->iri_ifp, len | PFIL_MEMPTR | PFIL_IN, NULL); switch (*pf_rv) { case PFIL_DROPPED: case PFIL_CONSUMED: /* * The filter ate it. Everything is recycled. */ m = NULL; unload = 0; break; case PFIL_REALLOCED: /* * The filter copied it. Everything is recycled. */ m = pfil_mem2mbuf(payload); unload = 0; break; case PFIL_PASS: /* * Filter said it was OK, so receive like * normal */ fl->ifl_sds.ifsd_m[cidx] = NULL; break; default: MPASS(0); } } else { fl->ifl_sds.ifsd_m[cidx] = NULL; if (pf_rv != NULL) *pf_rv = PFIL_PASS; } if (unload && irf->irf_len != 0) bus_dmamap_unload(fl->ifl_buf_tag, map); fl->ifl_cidx = (fl->ifl_cidx + 1) & (fl->ifl_size-1); if (__predict_false(fl->ifl_cidx == 0)) fl->ifl_gen = 0; bit_clear(fl->ifl_rx_bitmap, cidx); return (m); } static struct mbuf * assemble_segments(iflib_rxq_t rxq, if_rxd_info_t ri, if_rxsd_t sd, int *pf_rv) { struct mbuf *m, *mh, *mt; caddr_t cl; int *pf_rv_ptr, flags, i, padlen; bool consumed; i = 0; mh = NULL; consumed = false; *pf_rv = PFIL_PASS; pf_rv_ptr = pf_rv; do { m = rxd_frag_to_sd(rxq, &ri->iri_frags[i], !consumed, sd, pf_rv_ptr, ri); MPASS(*sd->ifsd_cl != NULL); /* * Exclude zero-length frags & frags from * packets the filter has consumed or dropped */ if (ri->iri_frags[i].irf_len == 0 || consumed || *pf_rv == PFIL_CONSUMED || *pf_rv == PFIL_DROPPED) { if (mh == NULL) { /* everything saved here */ consumed = true; pf_rv_ptr = NULL; continue; } /* XXX we can save the cluster here, but not the mbuf */ m_init(m, M_NOWAIT, MT_DATA, 0); m_free(m); continue; } if (mh == NULL) { flags = M_PKTHDR|M_EXT; mh = mt = m; padlen = ri->iri_pad; } else { flags = M_EXT; mt->m_next = m; mt = m; /* assuming padding is only on the first fragment */ padlen = 0; } cl = *sd->ifsd_cl; *sd->ifsd_cl = NULL; /* Can these two be made one ? */ m_init(m, M_NOWAIT, MT_DATA, flags); m_cljset(m, cl, sd->ifsd_fl->ifl_cltype); /* * These must follow m_init and m_cljset */ m->m_data += padlen; ri->iri_len -= padlen; m->m_len = ri->iri_frags[i].irf_len; } while (++i < ri->iri_nfrags); return (mh); } /* * Process one software descriptor */ static struct mbuf * iflib_rxd_pkt_get(iflib_rxq_t rxq, if_rxd_info_t ri) { struct if_rxsd sd; struct mbuf *m; int pf_rv; /* should I merge this back in now that the two paths are basically duplicated? */ if (ri->iri_nfrags == 1 && ri->iri_frags[0].irf_len != 0 && ri->iri_frags[0].irf_len <= MIN(IFLIB_RX_COPY_THRESH, MHLEN)) { m = rxd_frag_to_sd(rxq, &ri->iri_frags[0], false, &sd, &pf_rv, ri); if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED) return (m); if (pf_rv == PFIL_PASS) { m_init(m, M_NOWAIT, MT_DATA, M_PKTHDR); #ifndef __NO_STRICT_ALIGNMENT if (!IP_ALIGNED(m)) m->m_data += 2; #endif memcpy(m->m_data, *sd.ifsd_cl, ri->iri_len); m->m_len = ri->iri_frags[0].irf_len; } } else { m = assemble_segments(rxq, ri, &sd, &pf_rv); if (m == NULL) return (NULL); if (pf_rv != PFIL_PASS && pf_rv != PFIL_REALLOCED) return (m); } m->m_pkthdr.len = ri->iri_len; m->m_pkthdr.rcvif = ri->iri_ifp; m->m_flags |= ri->iri_flags; m->m_pkthdr.ether_vtag = ri->iri_vtag; m->m_pkthdr.flowid = ri->iri_flowid; M_HASHTYPE_SET(m, ri->iri_rsstype); m->m_pkthdr.csum_flags = ri->iri_csum_flags; m->m_pkthdr.csum_data = ri->iri_csum_data; return (m); } #if defined(INET6) || defined(INET) static void iflib_get_ip_forwarding(struct lro_ctrl *lc, bool *v4, bool *v6) { CURVNET_SET(lc->ifp->if_vnet); #if defined(INET6) *v6 = V_ip6_forwarding; #endif #if defined(INET) *v4 = V_ipforwarding; #endif CURVNET_RESTORE(); } /* * Returns true if it's possible this packet could be LROed. * if it returns false, it is guaranteed that tcp_lro_rx() * would not return zero. */ static bool iflib_check_lro_possible(struct mbuf *m, bool v4_forwarding, bool v6_forwarding) { struct ether_header *eh; eh = mtod(m, struct ether_header *); switch (eh->ether_type) { #if defined(INET6) case htons(ETHERTYPE_IPV6): return (!v6_forwarding); #endif #if defined (INET) case htons(ETHERTYPE_IP): return (!v4_forwarding); #endif } return false; } #else static void iflib_get_ip_forwarding(struct lro_ctrl *lc __unused, bool *v4 __unused, bool *v6 __unused) { } #endif static void _task_fn_rx_watchdog(void *context) { iflib_rxq_t rxq = context; GROUPTASK_ENQUEUE(&rxq->ifr_task); } static uint8_t iflib_rxeof(iflib_rxq_t rxq, qidx_t budget) { if_t ifp; if_ctx_t ctx = rxq->ifr_ctx; if_shared_ctx_t sctx = ctx->ifc_sctx; if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; int avail, i; qidx_t *cidxp; struct if_rxd_info ri; int err, budget_left, rx_bytes, rx_pkts; iflib_fl_t fl; int lro_enabled; bool v4_forwarding, v6_forwarding, lro_possible; uint8_t retval = 0; /* * XXX early demux data packets so that if_input processing only handles * acks in interrupt context */ struct mbuf *m, *mh, *mt, *mf; NET_EPOCH_ASSERT(); lro_possible = v4_forwarding = v6_forwarding = false; ifp = ctx->ifc_ifp; mh = mt = NULL; MPASS(budget > 0); rx_pkts = rx_bytes = 0; if (sctx->isc_flags & IFLIB_HAS_RXCQ) cidxp = &rxq->ifr_cq_cidx; else cidxp = &rxq->ifr_fl[0].ifl_cidx; if ((avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget)) == 0) { for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++) retval |= iflib_fl_refill_all(ctx, fl); DBG_COUNTER_INC(rx_unavail); return (retval); } /* pfil needs the vnet to be set */ CURVNET_SET_QUIET(ifp->if_vnet); for (budget_left = budget; budget_left > 0 && avail > 0;) { if (__predict_false(!CTX_ACTIVE(ctx))) { DBG_COUNTER_INC(rx_ctx_inactive); break; } /* * Reset client set fields to their default values */ rxd_info_zero(&ri); ri.iri_qsidx = rxq->ifr_id; ri.iri_cidx = *cidxp; ri.iri_ifp = ifp; ri.iri_frags = rxq->ifr_frags; err = ctx->isc_rxd_pkt_get(ctx->ifc_softc, &ri); if (err) goto err; rx_pkts += 1; rx_bytes += ri.iri_len; if (sctx->isc_flags & IFLIB_HAS_RXCQ) { *cidxp = ri.iri_cidx; /* Update our consumer index */ /* XXX NB: shurd - check if this is still safe */ while (rxq->ifr_cq_cidx >= scctx->isc_nrxd[0]) rxq->ifr_cq_cidx -= scctx->isc_nrxd[0]; /* was this only a completion queue message? */ if (__predict_false(ri.iri_nfrags == 0)) continue; } MPASS(ri.iri_nfrags != 0); MPASS(ri.iri_len != 0); /* will advance the cidx on the corresponding free lists */ m = iflib_rxd_pkt_get(rxq, &ri); avail--; budget_left--; if (avail == 0 && budget_left) avail = iflib_rxd_avail(ctx, rxq, *cidxp, budget_left); if (__predict_false(m == NULL)) continue; /* imm_pkt: -- cxgb */ if (mh == NULL) mh = mt = m; else { mt->m_nextpkt = m; mt = m; } } CURVNET_RESTORE(); /* make sure that we can refill faster than drain */ for (i = 0, fl = &rxq->ifr_fl[0]; i < sctx->isc_nfl; i++, fl++) retval |= iflib_fl_refill_all(ctx, fl); lro_enabled = (if_getcapenable(ifp) & IFCAP_LRO); if (lro_enabled) iflib_get_ip_forwarding(&rxq->ifr_lc, &v4_forwarding, &v6_forwarding); mt = mf = NULL; while (mh != NULL) { m = mh; mh = mh->m_nextpkt; m->m_nextpkt = NULL; #ifndef __NO_STRICT_ALIGNMENT if (!IP_ALIGNED(m) && (m = iflib_fixup_rx(m)) == NULL) continue; #endif #if defined(INET6) || defined(INET) if (lro_enabled) { if (!lro_possible) { lro_possible = iflib_check_lro_possible(m, v4_forwarding, v6_forwarding); if (lro_possible && mf != NULL) { ifp->if_input(ifp, mf); DBG_COUNTER_INC(rx_if_input); mt = mf = NULL; } } if ((m->m_pkthdr.csum_flags & (CSUM_L4_CALC|CSUM_L4_VALID)) == (CSUM_L4_CALC|CSUM_L4_VALID)) { if (lro_possible && tcp_lro_rx(&rxq->ifr_lc, m, 0) == 0) continue; } } #endif if (lro_possible) { ifp->if_input(ifp, m); DBG_COUNTER_INC(rx_if_input); continue; } if (mf == NULL) mf = m; if (mt != NULL) mt->m_nextpkt = m; mt = m; } if (mf != NULL) { ifp->if_input(ifp, mf); DBG_COUNTER_INC(rx_if_input); } if_inc_counter(ifp, IFCOUNTER_IBYTES, rx_bytes); if_inc_counter(ifp, IFCOUNTER_IPACKETS, rx_pkts); /* * Flush any outstanding LRO work */ #if defined(INET6) || defined(INET) tcp_lro_flush_all(&rxq->ifr_lc); #endif if (avail != 0 || iflib_rxd_avail(ctx, rxq, *cidxp, 1) != 0) retval |= IFLIB_RXEOF_MORE; return (retval); err: STATE_LOCK(ctx); ctx->ifc_flags |= IFC_DO_RESET; iflib_admin_intr_deferred(ctx); STATE_UNLOCK(ctx); return (0); } #define TXD_NOTIFY_COUNT(txq) (((txq)->ift_size / (txq)->ift_update_freq)-1) static inline qidx_t txq_max_db_deferred(iflib_txq_t txq, qidx_t in_use) { qidx_t notify_count = TXD_NOTIFY_COUNT(txq); qidx_t minthresh = txq->ift_size / 8; if (in_use > 4*minthresh) return (notify_count); if (in_use > 2*minthresh) return (notify_count >> 1); if (in_use > minthresh) return (notify_count >> 3); return (0); } static inline qidx_t txq_max_rs_deferred(iflib_txq_t txq) { qidx_t notify_count = TXD_NOTIFY_COUNT(txq); qidx_t minthresh = txq->ift_size / 8; if (txq->ift_in_use > 4*minthresh) return (notify_count); if (txq->ift_in_use > 2*minthresh) return (notify_count >> 1); if (txq->ift_in_use > minthresh) return (notify_count >> 2); return (2); } #define M_CSUM_FLAGS(m) ((m)->m_pkthdr.csum_flags) #define M_HAS_VLANTAG(m) (m->m_flags & M_VLANTAG) #define TXQ_MAX_DB_DEFERRED(txq, in_use) txq_max_db_deferred((txq), (in_use)) #define TXQ_MAX_RS_DEFERRED(txq) txq_max_rs_deferred(txq) #define TXQ_MAX_DB_CONSUMED(size) (size >> 4) /* forward compatibility for cxgb */ #define FIRST_QSET(ctx) 0 #define NTXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_ntxqsets) #define NRXQSETS(ctx) ((ctx)->ifc_softc_ctx.isc_nrxqsets) #define QIDX(ctx, m) ((((m)->m_pkthdr.flowid & ctx->ifc_softc_ctx.isc_rss_table_mask) % NTXQSETS(ctx)) + FIRST_QSET(ctx)) #define DESC_RECLAIMABLE(q) ((int)((q)->ift_processed - (q)->ift_cleaned - (q)->ift_ctx->ifc_softc_ctx.isc_tx_nsegments)) /* XXX we should be setting this to something other than zero */ #define RECLAIM_THRESH(ctx) ((ctx)->ifc_sctx->isc_tx_reclaim_thresh) #define MAX_TX_DESC(ctx) MAX((ctx)->ifc_softc_ctx.isc_tx_tso_segments_max, \ (ctx)->ifc_softc_ctx.isc_tx_nsegments) static inline bool iflib_txd_db_check(iflib_txq_t txq, int ring) { if_ctx_t ctx = txq->ift_ctx; qidx_t dbval, max; max = TXQ_MAX_DB_DEFERRED(txq, txq->ift_in_use); /* force || threshold exceeded || at the edge of the ring */ if (ring || (txq->ift_db_pending >= max) || (TXQ_AVAIL(txq) <= MAX_TX_DESC(ctx) + 2)) { /* * 'npending' is used if the card's doorbell is in terms of the number of descriptors * pending flush (BRCM). 'pidx' is used in cases where the card's doorbeel uses the * producer index explicitly (INTC). */ dbval = txq->ift_npending ? txq->ift_npending : txq->ift_pidx; bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map, BUS_DMASYNC_PREREAD | BUS_DMASYNC_PREWRITE); ctx->isc_txd_flush(ctx->ifc_softc, txq->ift_id, dbval); /* * Absent bugs there are zero packets pending so reset pending counts to zero. */ txq->ift_db_pending = txq->ift_npending = 0; return (true); } return (false); } #ifdef PKT_DEBUG static void print_pkt(if_pkt_info_t pi) { printf("pi len: %d qsidx: %d nsegs: %d ndescs: %d flags: %x pidx: %d\n", pi->ipi_len, pi->ipi_qsidx, pi->ipi_nsegs, pi->ipi_ndescs, pi->ipi_flags, pi->ipi_pidx); printf("pi new_pidx: %d csum_flags: %lx tso_segsz: %d mflags: %x vtag: %d\n", pi->ipi_new_pidx, pi->ipi_csum_flags, pi->ipi_tso_segsz, pi->ipi_mflags, pi->ipi_vtag); printf("pi etype: %d ehdrlen: %d ip_hlen: %d ipproto: %d\n", pi->ipi_etype, pi->ipi_ehdrlen, pi->ipi_ip_hlen, pi->ipi_ipproto); } #endif #define IS_TSO4(pi) ((pi)->ipi_csum_flags & CSUM_IP_TSO) #define IS_TX_OFFLOAD4(pi) ((pi)->ipi_csum_flags & (CSUM_IP_TCP | CSUM_IP_TSO)) #define IS_TSO6(pi) ((pi)->ipi_csum_flags & CSUM_IP6_TSO) #define IS_TX_OFFLOAD6(pi) ((pi)->ipi_csum_flags & (CSUM_IP6_TCP | CSUM_IP6_TSO)) static int iflib_parse_header(iflib_txq_t txq, if_pkt_info_t pi, struct mbuf **mp) { if_shared_ctx_t sctx = txq->ift_ctx->ifc_sctx; struct ether_vlan_header *eh; struct mbuf *m; m = *mp; if ((sctx->isc_flags & IFLIB_NEED_SCRATCH) && M_WRITABLE(m) == 0) { if ((m = m_dup(m, M_NOWAIT)) == NULL) { return (ENOMEM); } else { m_freem(*mp); DBG_COUNTER_INC(tx_frees); *mp = m; } } /* * Determine where frame payload starts. * Jump over vlan headers if already present, * helpful for QinQ too. */ if (__predict_false(m->m_len < sizeof(*eh))) { txq->ift_pullups++; if (__predict_false((m = m_pullup(m, sizeof(*eh))) == NULL)) return (ENOMEM); } eh = mtod(m, struct ether_vlan_header *); if (eh->evl_encap_proto == htons(ETHERTYPE_VLAN)) { pi->ipi_etype = ntohs(eh->evl_proto); pi->ipi_ehdrlen = ETHER_HDR_LEN + ETHER_VLAN_ENCAP_LEN; } else { pi->ipi_etype = ntohs(eh->evl_encap_proto); pi->ipi_ehdrlen = ETHER_HDR_LEN; } switch (pi->ipi_etype) { #ifdef INET case ETHERTYPE_IP: { struct mbuf *n; struct ip *ip = NULL; struct tcphdr *th = NULL; int minthlen; minthlen = min(m->m_pkthdr.len, pi->ipi_ehdrlen + sizeof(*ip) + sizeof(*th)); if (__predict_false(m->m_len < minthlen)) { /* * if this code bloat is causing too much of a hit * move it to a separate function and mark it noinline */ if (m->m_len == pi->ipi_ehdrlen) { n = m->m_next; MPASS(n); if (n->m_len >= sizeof(*ip)) { ip = (struct ip *)n->m_data; if (n->m_len >= (ip->ip_hl << 2) + sizeof(*th)) th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2)); } else { txq->ift_pullups++; if (__predict_false((m = m_pullup(m, minthlen)) == NULL)) return (ENOMEM); ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen); } } else { txq->ift_pullups++; if (__predict_false((m = m_pullup(m, minthlen)) == NULL)) return (ENOMEM); ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen); if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th)) th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2)); } } else { ip = (struct ip *)(m->m_data + pi->ipi_ehdrlen); if (m->m_len >= (ip->ip_hl << 2) + sizeof(*th)) th = (struct tcphdr *)((caddr_t)ip + (ip->ip_hl << 2)); } pi->ipi_ip_hlen = ip->ip_hl << 2; pi->ipi_ipproto = ip->ip_p; pi->ipi_flags |= IPI_TX_IPV4; /* TCP checksum offload may require TCP header length */ if (IS_TX_OFFLOAD4(pi)) { if (__predict_true(pi->ipi_ipproto == IPPROTO_TCP)) { if (__predict_false(th == NULL)) { txq->ift_pullups++; if (__predict_false((m = m_pullup(m, (ip->ip_hl << 2) + sizeof(*th))) == NULL)) return (ENOMEM); th = (struct tcphdr *)((caddr_t)ip + pi->ipi_ip_hlen); } pi->ipi_tcp_hflags = th->th_flags; pi->ipi_tcp_hlen = th->th_off << 2; pi->ipi_tcp_seq = th->th_seq; } if (IS_TSO4(pi)) { if (__predict_false(ip->ip_p != IPPROTO_TCP)) return (ENXIO); /* * TSO always requires hardware checksum offload. */ pi->ipi_csum_flags |= (CSUM_IP_TCP | CSUM_IP); th->th_sum = in_pseudo(ip->ip_src.s_addr, ip->ip_dst.s_addr, htons(IPPROTO_TCP)); pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz; if (sctx->isc_flags & IFLIB_TSO_INIT_IP) { ip->ip_sum = 0; ip->ip_len = htons(pi->ipi_ip_hlen + pi->ipi_tcp_hlen + pi->ipi_tso_segsz); } } } if ((sctx->isc_flags & IFLIB_NEED_ZERO_CSUM) && (pi->ipi_csum_flags & CSUM_IP)) ip->ip_sum = 0; break; } #endif #ifdef INET6 case ETHERTYPE_IPV6: { struct ip6_hdr *ip6 = (struct ip6_hdr *)(m->m_data + pi->ipi_ehdrlen); struct tcphdr *th; pi->ipi_ip_hlen = sizeof(struct ip6_hdr); if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) { txq->ift_pullups++; if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr))) == NULL)) return (ENOMEM); } th = (struct tcphdr *)((caddr_t)ip6 + pi->ipi_ip_hlen); /* XXX-BZ this will go badly in case of ext hdrs. */ pi->ipi_ipproto = ip6->ip6_nxt; pi->ipi_flags |= IPI_TX_IPV6; /* TCP checksum offload may require TCP header length */ if (IS_TX_OFFLOAD6(pi)) { if (pi->ipi_ipproto == IPPROTO_TCP) { if (__predict_false(m->m_len < pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) { txq->ift_pullups++; if (__predict_false((m = m_pullup(m, pi->ipi_ehdrlen + sizeof(struct ip6_hdr) + sizeof(struct tcphdr))) == NULL)) return (ENOMEM); } pi->ipi_tcp_hflags = th->th_flags; pi->ipi_tcp_hlen = th->th_off << 2; pi->ipi_tcp_seq = th->th_seq; } if (IS_TSO6(pi)) { if (__predict_false(ip6->ip6_nxt != IPPROTO_TCP)) return (ENXIO); /* * TSO always requires hardware checksum offload. */ pi->ipi_csum_flags |= CSUM_IP6_TCP; th->th_sum = in6_cksum_pseudo(ip6, 0, IPPROTO_TCP, 0); pi->ipi_tso_segsz = m->m_pkthdr.tso_segsz; } } break; } #endif default: pi->ipi_csum_flags &= ~CSUM_OFFLOAD; pi->ipi_ip_hlen = 0; break; } *mp = m; return (0); } /* * If dodgy hardware rejects the scatter gather chain we've handed it * we'll need to remove the mbuf chain from ifsg_m[] before we can add the * m_defrag'd mbufs */ static __noinline struct mbuf * iflib_remove_mbuf(iflib_txq_t txq) { int ntxd, pidx; struct mbuf *m, **ifsd_m; ifsd_m = txq->ift_sds.ifsd_m; ntxd = txq->ift_size; pidx = txq->ift_pidx & (ntxd - 1); ifsd_m = txq->ift_sds.ifsd_m; m = ifsd_m[pidx]; ifsd_m[pidx] = NULL; bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[pidx]); if (txq->ift_sds.ifsd_tso_map != NULL) bus_dmamap_unload(txq->ift_tso_buf_tag, txq->ift_sds.ifsd_tso_map[pidx]); #if MEMORY_LOGGING txq->ift_dequeued++; #endif return (m); } static inline caddr_t calc_next_txd(iflib_txq_t txq, int cidx, uint8_t qid) { qidx_t size; int ntxd; caddr_t start, end, cur, next; ntxd = txq->ift_size; size = txq->ift_txd_size[qid]; start = txq->ift_ifdi[qid].idi_vaddr; if (__predict_false(size == 0)) return (start); cur = start + size*cidx; end = start + size*ntxd; next = CACHE_PTR_NEXT(cur); return (next < end ? next : start); } /* * Pad an mbuf to ensure a minimum ethernet frame size. * min_frame_size is the frame size (less CRC) to pad the mbuf to */ static __noinline int iflib_ether_pad(device_t dev, struct mbuf **m_head, uint16_t min_frame_size) { /* * 18 is enough bytes to pad an ARP packet to 46 bytes, and * and ARP message is the smallest common payload I can think of */ static char pad[18]; /* just zeros */ int n; struct mbuf *new_head; if (!M_WRITABLE(*m_head)) { new_head = m_dup(*m_head, M_NOWAIT); if (new_head == NULL) { m_freem(*m_head); device_printf(dev, "cannot pad short frame, m_dup() failed"); DBG_COUNTER_INC(encap_pad_mbuf_fail); DBG_COUNTER_INC(tx_frees); return ENOMEM; } m_freem(*m_head); *m_head = new_head; } for (n = min_frame_size - (*m_head)->m_pkthdr.len; n > 0; n -= sizeof(pad)) if (!m_append(*m_head, min(n, sizeof(pad)), pad)) break; if (n > 0) { m_freem(*m_head); device_printf(dev, "cannot pad short frame\n"); DBG_COUNTER_INC(encap_pad_mbuf_fail); DBG_COUNTER_INC(tx_frees); return (ENOBUFS); } return 0; } static int iflib_encap(iflib_txq_t txq, struct mbuf **m_headp) { if_ctx_t ctx; if_shared_ctx_t sctx; if_softc_ctx_t scctx; bus_dma_tag_t buf_tag; bus_dma_segment_t *segs; struct mbuf *m_head, **ifsd_m; void *next_txd; bus_dmamap_t map; struct if_pkt_info pi; int remap = 0; int err, nsegs, ndesc, max_segs, pidx, cidx, next, ntxd; ctx = txq->ift_ctx; sctx = ctx->ifc_sctx; scctx = &ctx->ifc_softc_ctx; segs = txq->ift_segs; ntxd = txq->ift_size; m_head = *m_headp; map = NULL; /* * If we're doing TSO the next descriptor to clean may be quite far ahead */ cidx = txq->ift_cidx; pidx = txq->ift_pidx; if (ctx->ifc_flags & IFC_PREFETCH) { next = (cidx + CACHE_PTR_INCREMENT) & (ntxd-1); if (!(ctx->ifc_flags & IFLIB_HAS_TXCQ)) { next_txd = calc_next_txd(txq, cidx, 0); prefetch(next_txd); } /* prefetch the next cache line of mbuf pointers and flags */ prefetch(&txq->ift_sds.ifsd_m[next]); prefetch(&txq->ift_sds.ifsd_map[next]); next = (cidx + CACHE_LINE_SIZE) & (ntxd-1); } map = txq->ift_sds.ifsd_map[pidx]; ifsd_m = txq->ift_sds.ifsd_m; if (m_head->m_pkthdr.csum_flags & CSUM_TSO) { buf_tag = txq->ift_tso_buf_tag; max_segs = scctx->isc_tx_tso_segments_max; map = txq->ift_sds.ifsd_tso_map[pidx]; MPASS(buf_tag != NULL); MPASS(max_segs > 0); } else { buf_tag = txq->ift_buf_tag; max_segs = scctx->isc_tx_nsegments; map = txq->ift_sds.ifsd_map[pidx]; } if ((sctx->isc_flags & IFLIB_NEED_ETHER_PAD) && __predict_false(m_head->m_pkthdr.len < scctx->isc_min_frame_size)) { err = iflib_ether_pad(ctx->ifc_dev, m_headp, scctx->isc_min_frame_size); if (err) { DBG_COUNTER_INC(encap_txd_encap_fail); return err; } } m_head = *m_headp; pkt_info_zero(&pi); pi.ipi_mflags = (m_head->m_flags & (M_VLANTAG|M_BCAST|M_MCAST)); pi.ipi_pidx = pidx; pi.ipi_qsidx = txq->ift_id; pi.ipi_len = m_head->m_pkthdr.len; pi.ipi_csum_flags = m_head->m_pkthdr.csum_flags; pi.ipi_vtag = M_HAS_VLANTAG(m_head) ? m_head->m_pkthdr.ether_vtag : 0; /* deliberate bitwise OR to make one condition */ if (__predict_true((pi.ipi_csum_flags | pi.ipi_vtag))) { if (__predict_false((err = iflib_parse_header(txq, &pi, m_headp)) != 0)) { DBG_COUNTER_INC(encap_txd_encap_fail); return (err); } m_head = *m_headp; } retry: err = bus_dmamap_load_mbuf_sg(buf_tag, map, m_head, segs, &nsegs, BUS_DMA_NOWAIT); defrag: if (__predict_false(err)) { switch (err) { case EFBIG: /* try collapse once and defrag once */ if (remap == 0) { m_head = m_collapse(*m_headp, M_NOWAIT, max_segs); /* try defrag if collapsing fails */ if (m_head == NULL) remap++; } if (remap == 1) { txq->ift_mbuf_defrag++; m_head = m_defrag(*m_headp, M_NOWAIT); } /* * remap should never be >1 unless bus_dmamap_load_mbuf_sg * failed to map an mbuf that was run through m_defrag */ MPASS(remap <= 1); if (__predict_false(m_head == NULL || remap > 1)) goto defrag_failed; remap++; *m_headp = m_head; goto retry; break; case ENOMEM: txq->ift_no_tx_dma_setup++; break; default: txq->ift_no_tx_dma_setup++; m_freem(*m_headp); DBG_COUNTER_INC(tx_frees); *m_headp = NULL; break; } txq->ift_map_failed++; DBG_COUNTER_INC(encap_load_mbuf_fail); DBG_COUNTER_INC(encap_txd_encap_fail); return (err); } ifsd_m[pidx] = m_head; /* * XXX assumes a 1 to 1 relationship between segments and * descriptors - this does not hold true on all drivers, e.g. * cxgb */ if (__predict_false(nsegs + 2 > TXQ_AVAIL(txq))) { txq->ift_no_desc_avail++; bus_dmamap_unload(buf_tag, map); DBG_COUNTER_INC(encap_txq_avail_fail); DBG_COUNTER_INC(encap_txd_encap_fail); if ((txq->ift_task.gt_task.ta_flags & TASK_ENQUEUED) == 0) GROUPTASK_ENQUEUE(&txq->ift_task); return (ENOBUFS); } /* * On Intel cards we can greatly reduce the number of TX interrupts * we see by only setting report status on every Nth descriptor. * However, this also means that the driver will need to keep track * of the descriptors that RS was set on to check them for the DD bit. */ txq->ift_rs_pending += nsegs + 1; if (txq->ift_rs_pending > TXQ_MAX_RS_DEFERRED(txq) || iflib_no_tx_batch || (TXQ_AVAIL(txq) - nsegs) <= MAX_TX_DESC(ctx) + 2) { pi.ipi_flags |= IPI_TX_INTR; txq->ift_rs_pending = 0; } pi.ipi_segs = segs; pi.ipi_nsegs = nsegs; MPASS(pidx >= 0 && pidx < txq->ift_size); #ifdef PKT_DEBUG print_pkt(&pi); #endif if ((err = ctx->isc_txd_encap(ctx->ifc_softc, &pi)) == 0) { bus_dmamap_sync(buf_tag, map, BUS_DMASYNC_PREWRITE); DBG_COUNTER_INC(tx_encap); MPASS(pi.ipi_new_pidx < txq->ift_size); ndesc = pi.ipi_new_pidx - pi.ipi_pidx; if (pi.ipi_new_pidx < pi.ipi_pidx) { ndesc += txq->ift_size; txq->ift_gen = 1; } /* * drivers can need as many as * two sentinels */ MPASS(ndesc <= pi.ipi_nsegs + 2); MPASS(pi.ipi_new_pidx != pidx); MPASS(ndesc > 0); txq->ift_in_use += ndesc; txq->ift_db_pending += ndesc; /* * We update the last software descriptor again here because there may * be a sentinel and/or there may be more mbufs than segments */ txq->ift_pidx = pi.ipi_new_pidx; txq->ift_npending += pi.ipi_ndescs; } else { *m_headp = m_head = iflib_remove_mbuf(txq); if (err == EFBIG) { txq->ift_txd_encap_efbig++; if (remap < 2) { remap = 1; goto defrag; } } goto defrag_failed; } /* * err can't possibly be non-zero here, so we don't neet to test it * to see if we need to DBG_COUNTER_INC(encap_txd_encap_fail). */ return (err); defrag_failed: txq->ift_mbuf_defrag_failed++; txq->ift_map_failed++; m_freem(*m_headp); DBG_COUNTER_INC(tx_frees); *m_headp = NULL; DBG_COUNTER_INC(encap_txd_encap_fail); return (ENOMEM); } static void iflib_tx_desc_free(iflib_txq_t txq, int n) { uint32_t qsize, cidx, mask, gen; struct mbuf *m, **ifsd_m; bool do_prefetch; cidx = txq->ift_cidx; gen = txq->ift_gen; qsize = txq->ift_size; mask = qsize-1; ifsd_m = txq->ift_sds.ifsd_m; do_prefetch = (txq->ift_ctx->ifc_flags & IFC_PREFETCH); while (n-- > 0) { if (do_prefetch) { prefetch(ifsd_m[(cidx + 3) & mask]); prefetch(ifsd_m[(cidx + 4) & mask]); } if ((m = ifsd_m[cidx]) != NULL) { prefetch(&ifsd_m[(cidx + CACHE_PTR_INCREMENT) & mask]); if (m->m_pkthdr.csum_flags & CSUM_TSO) { bus_dmamap_sync(txq->ift_tso_buf_tag, txq->ift_sds.ifsd_tso_map[cidx], BUS_DMASYNC_POSTWRITE); bus_dmamap_unload(txq->ift_tso_buf_tag, txq->ift_sds.ifsd_tso_map[cidx]); } else { bus_dmamap_sync(txq->ift_buf_tag, txq->ift_sds.ifsd_map[cidx], BUS_DMASYNC_POSTWRITE); bus_dmamap_unload(txq->ift_buf_tag, txq->ift_sds.ifsd_map[cidx]); } /* XXX we don't support any drivers that batch packets yet */ MPASS(m->m_nextpkt == NULL); m_freem(m); ifsd_m[cidx] = NULL; #if MEMORY_LOGGING txq->ift_dequeued++; #endif DBG_COUNTER_INC(tx_frees); } if (__predict_false(++cidx == qsize)) { cidx = 0; gen = 0; } } txq->ift_cidx = cidx; txq->ift_gen = gen; } static __inline int iflib_completed_tx_reclaim(iflib_txq_t txq, int thresh) { int reclaim; if_ctx_t ctx = txq->ift_ctx; KASSERT(thresh >= 0, ("invalid threshold to reclaim")); MPASS(thresh /*+ MAX_TX_DESC(txq->ift_ctx) */ < txq->ift_size); /* * Need a rate-limiting check so that this isn't called every time */ iflib_tx_credits_update(ctx, txq); reclaim = DESC_RECLAIMABLE(txq); if (reclaim <= thresh /* + MAX_TX_DESC(txq->ift_ctx) */) { #ifdef INVARIANTS if (iflib_verbose_debug) { printf("%s processed=%ju cleaned=%ju tx_nsegments=%d reclaim=%d thresh=%d\n", __FUNCTION__, txq->ift_processed, txq->ift_cleaned, txq->ift_ctx->ifc_softc_ctx.isc_tx_nsegments, reclaim, thresh); } #endif return (0); } iflib_tx_desc_free(txq, reclaim); txq->ift_cleaned += reclaim; txq->ift_in_use -= reclaim; return (reclaim); } static struct mbuf ** _ring_peek_one(struct ifmp_ring *r, int cidx, int offset, int remaining) { int next, size; struct mbuf **items; size = r->size; next = (cidx + CACHE_PTR_INCREMENT) & (size-1); items = __DEVOLATILE(struct mbuf **, &r->items[0]); prefetch(items[(cidx + offset) & (size-1)]); if (remaining > 1) { prefetch2cachelines(&items[next]); prefetch2cachelines(items[(cidx + offset + 1) & (size-1)]); prefetch2cachelines(items[(cidx + offset + 2) & (size-1)]); prefetch2cachelines(items[(cidx + offset + 3) & (size-1)]); } return (__DEVOLATILE(struct mbuf **, &r->items[(cidx + offset) & (size-1)])); } static void iflib_txq_check_drain(iflib_txq_t txq, int budget) { ifmp_ring_check_drainage(txq->ift_br, budget); } static uint32_t iflib_txq_can_drain(struct ifmp_ring *r) { iflib_txq_t txq = r->cookie; if_ctx_t ctx = txq->ift_ctx; if (TXQ_AVAIL(txq) > MAX_TX_DESC(ctx) + 2) return (1); bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map, BUS_DMASYNC_POSTREAD); return (ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, false)); } static uint32_t iflib_txq_drain(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx) { iflib_txq_t txq = r->cookie; if_ctx_t ctx = txq->ift_ctx; if_t ifp = ctx->ifc_ifp; struct mbuf *m, **mp; int avail, bytes_sent, skipped, count, err, i; int mcast_sent, pkt_sent, reclaimed; bool do_prefetch, rang, ring; if (__predict_false(!(if_getdrvflags(ifp) & IFF_DRV_RUNNING) || !LINK_ACTIVE(ctx))) { DBG_COUNTER_INC(txq_drain_notready); return (0); } reclaimed = iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx)); rang = iflib_txd_db_check(txq, reclaimed && txq->ift_db_pending); avail = IDXDIFF(pidx, cidx, r->size); if (__predict_false(ctx->ifc_flags & IFC_QFLUSH)) { /* * The driver is unloading so we need to free all pending packets. */ DBG_COUNTER_INC(txq_drain_flushing); for (i = 0; i < avail; i++) { if (__predict_true(r->items[(cidx + i) & (r->size-1)] != (void *)txq)) m_freem(r->items[(cidx + i) & (r->size-1)]); r->items[(cidx + i) & (r->size-1)] = NULL; } return (avail); } if (__predict_false(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE)) { txq->ift_qstatus = IFLIB_QUEUE_IDLE; CALLOUT_LOCK(txq); callout_stop(&txq->ift_timer); CALLOUT_UNLOCK(txq); DBG_COUNTER_INC(txq_drain_oactive); return (0); } /* * If we've reclaimed any packets this queue cannot be hung. */ if (reclaimed) txq->ift_qstatus = IFLIB_QUEUE_IDLE; skipped = mcast_sent = bytes_sent = pkt_sent = 0; count = MIN(avail, TX_BATCH_SIZE); #ifdef INVARIANTS if (iflib_verbose_debug) printf("%s avail=%d ifc_flags=%x txq_avail=%d ", __FUNCTION__, avail, ctx->ifc_flags, TXQ_AVAIL(txq)); #endif do_prefetch = (ctx->ifc_flags & IFC_PREFETCH); err = 0; for (i = 0; i < count && TXQ_AVAIL(txq) >= MAX_TX_DESC(ctx) + 2; i++) { int rem = do_prefetch ? count - i : 0; mp = _ring_peek_one(r, cidx, i, rem); MPASS(mp != NULL && *mp != NULL); /* * Completion interrupts will use the address of the txq * as a sentinel to enqueue _something_ in order to acquire * the lock on the mp_ring (there's no direct lock call). * We obviously whave to check for these sentinel cases * and skip them. */ if (__predict_false(*mp == (struct mbuf *)txq)) { skipped++; continue; } err = iflib_encap(txq, mp); if (__predict_false(err)) { /* no room - bail out */ if (err == ENOBUFS) break; skipped++; /* we can't send this packet - skip it */ continue; } pkt_sent++; m = *mp; DBG_COUNTER_INC(tx_sent); bytes_sent += m->m_pkthdr.len; mcast_sent += !!(m->m_flags & M_MCAST); if (__predict_false(!(ifp->if_drv_flags & IFF_DRV_RUNNING))) break; ETHER_BPF_MTAP(ifp, m); rang = iflib_txd_db_check(txq, false); } /* deliberate use of bitwise or to avoid gratuitous short-circuit */ ring = rang ? false : (iflib_min_tx_latency | err); iflib_txd_db_check(txq, ring); if_inc_counter(ifp, IFCOUNTER_OBYTES, bytes_sent); if_inc_counter(ifp, IFCOUNTER_OPACKETS, pkt_sent); if (mcast_sent) if_inc_counter(ifp, IFCOUNTER_OMCASTS, mcast_sent); #ifdef INVARIANTS if (iflib_verbose_debug) printf("consumed=%d\n", skipped + pkt_sent); #endif return (skipped + pkt_sent); } static uint32_t iflib_txq_drain_always(struct ifmp_ring *r) { return (1); } static uint32_t iflib_txq_drain_free(struct ifmp_ring *r, uint32_t cidx, uint32_t pidx) { int i, avail; struct mbuf **mp; iflib_txq_t txq; txq = r->cookie; txq->ift_qstatus = IFLIB_QUEUE_IDLE; CALLOUT_LOCK(txq); callout_stop(&txq->ift_timer); CALLOUT_UNLOCK(txq); avail = IDXDIFF(pidx, cidx, r->size); for (i = 0; i < avail; i++) { mp = _ring_peek_one(r, cidx, i, avail - i); if (__predict_false(*mp == (struct mbuf *)txq)) continue; m_freem(*mp); DBG_COUNTER_INC(tx_frees); } MPASS(ifmp_ring_is_stalled(r) == 0); return (avail); } static void iflib_ifmp_purge(iflib_txq_t txq) { struct ifmp_ring *r; r = txq->ift_br; r->drain = iflib_txq_drain_free; r->can_drain = iflib_txq_drain_always; ifmp_ring_check_drainage(r, r->size); r->drain = iflib_txq_drain; r->can_drain = iflib_txq_can_drain; } static void _task_fn_tx(void *context) { iflib_txq_t txq = context; if_ctx_t ctx = txq->ift_ctx; if_t ifp = ctx->ifc_ifp; int abdicate = ctx->ifc_sysctl_tx_abdicate; #ifdef IFLIB_DIAGNOSTICS txq->ift_cpu_exec_count[curcpu]++; #endif if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING)) return; #ifdef DEV_NETMAP if ((if_getcapenable(ifp) & IFCAP_NETMAP) && netmap_tx_irq(ifp, txq->ift_id)) goto skip_ifmp; #endif #ifdef ALTQ if (ALTQ_IS_ENABLED(&ifp->if_snd)) iflib_altq_if_start(ifp); #endif if (txq->ift_db_pending) ifmp_ring_enqueue(txq->ift_br, (void **)&txq, 1, TX_BATCH_SIZE, abdicate); else if (!abdicate) ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE); /* * When abdicating, we always need to check drainage, not just when we don't enqueue */ if (abdicate) ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE); #ifdef DEV_NETMAP skip_ifmp: #endif if (ctx->ifc_flags & IFC_LEGACY) IFDI_INTR_ENABLE(ctx); else IFDI_TX_QUEUE_INTR_ENABLE(ctx, txq->ift_id); } static void _task_fn_rx(void *context) { iflib_rxq_t rxq = context; if_ctx_t ctx = rxq->ifr_ctx; uint8_t more; uint16_t budget; #ifdef DEV_NETMAP u_int work = 0; int nmirq; #endif #ifdef IFLIB_DIAGNOSTICS rxq->ifr_cpu_exec_count[curcpu]++; #endif DBG_COUNTER_INC(task_fn_rxs); if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))) return; #ifdef DEV_NETMAP nmirq = netmap_rx_irq(ctx->ifc_ifp, rxq->ifr_id, &work); if (nmirq != NM_IRQ_PASS) { more = (nmirq == NM_IRQ_RESCHED) ? IFLIB_RXEOF_MORE : 0; goto skip_rxeof; } #endif budget = ctx->ifc_sysctl_rx_budget; if (budget == 0) budget = 16; /* XXX */ more = iflib_rxeof(rxq, budget); #ifdef DEV_NETMAP skip_rxeof: #endif if ((more & IFLIB_RXEOF_MORE) == 0) { if (ctx->ifc_flags & IFC_LEGACY) IFDI_INTR_ENABLE(ctx); else IFDI_RX_QUEUE_INTR_ENABLE(ctx, rxq->ifr_id); DBG_COUNTER_INC(rx_intr_enables); } if (__predict_false(!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING))) return; if (more & IFLIB_RXEOF_MORE) GROUPTASK_ENQUEUE(&rxq->ifr_task); else if (more & IFLIB_RXEOF_EMPTY) callout_reset_curcpu(&rxq->ifr_watchdog, 1, &_task_fn_rx_watchdog, rxq); } static void _task_fn_admin(void *context) { if_ctx_t ctx = context; if_softc_ctx_t sctx = &ctx->ifc_softc_ctx; iflib_txq_t txq; int i; bool oactive, running, do_reset, do_watchdog, in_detach; STATE_LOCK(ctx); running = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING); oactive = (if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_OACTIVE); do_reset = (ctx->ifc_flags & IFC_DO_RESET); do_watchdog = (ctx->ifc_flags & IFC_DO_WATCHDOG); in_detach = (ctx->ifc_flags & IFC_IN_DETACH); ctx->ifc_flags &= ~(IFC_DO_RESET|IFC_DO_WATCHDOG); STATE_UNLOCK(ctx); if ((!running && !oactive) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN)) return; if (in_detach) return; CTX_LOCK(ctx); for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) { CALLOUT_LOCK(txq); callout_stop(&txq->ift_timer); CALLOUT_UNLOCK(txq); } if (do_watchdog) { ctx->ifc_watchdog_events++; IFDI_WATCHDOG_RESET(ctx); } IFDI_UPDATE_ADMIN_STATUS(ctx); for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) { callout_reset_on(&txq->ift_timer, iflib_timer_default, iflib_timer, txq, txq->ift_timer.c_cpu); } IFDI_LINK_INTR_ENABLE(ctx); if (do_reset) iflib_if_init_locked(ctx); CTX_UNLOCK(ctx); if (LINK_ACTIVE(ctx) == 0) return; for (txq = ctx->ifc_txqs, i = 0; i < sctx->isc_ntxqsets; i++, txq++) iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET); } static void _task_fn_iov(void *context) { if_ctx_t ctx = context; if (!(if_getdrvflags(ctx->ifc_ifp) & IFF_DRV_RUNNING) && !(ctx->ifc_sctx->isc_flags & IFLIB_ADMIN_ALWAYS_RUN)) return; CTX_LOCK(ctx); IFDI_VFLR_HANDLE(ctx); CTX_UNLOCK(ctx); } static int iflib_sysctl_int_delay(SYSCTL_HANDLER_ARGS) { int err; if_int_delay_info_t info; if_ctx_t ctx; info = (if_int_delay_info_t)arg1; ctx = info->iidi_ctx; info->iidi_req = req; info->iidi_oidp = oidp; CTX_LOCK(ctx); err = IFDI_SYSCTL_INT_DELAY(ctx, info); CTX_UNLOCK(ctx); return (err); } /********************************************************************* * * IFNET FUNCTIONS * **********************************************************************/ static void iflib_if_init_locked(if_ctx_t ctx) { iflib_stop(ctx); iflib_init_locked(ctx); } static void iflib_if_init(void *arg) { if_ctx_t ctx = arg; CTX_LOCK(ctx); iflib_if_init_locked(ctx); CTX_UNLOCK(ctx); } static int iflib_if_transmit(if_t ifp, struct mbuf *m) { if_ctx_t ctx = if_getsoftc(ifp); iflib_txq_t txq; int err, qidx; int abdicate = ctx->ifc_sysctl_tx_abdicate; if (__predict_false((ifp->if_drv_flags & IFF_DRV_RUNNING) == 0 || !LINK_ACTIVE(ctx))) { DBG_COUNTER_INC(tx_frees); m_freem(m); return (ENETDOWN); } MPASS(m->m_nextpkt == NULL); /* ALTQ-enabled interfaces always use queue 0. */ qidx = 0; if ((NTXQSETS(ctx) > 1) && M_HASHTYPE_GET(m) && !ALTQ_IS_ENABLED(&ifp->if_snd)) qidx = QIDX(ctx, m); /* * XXX calculate buf_ring based on flowid (divvy up bits?) */ txq = &ctx->ifc_txqs[qidx]; #ifdef DRIVER_BACKPRESSURE if (txq->ift_closed) { while (m != NULL) { next = m->m_nextpkt; m->m_nextpkt = NULL; m_freem(m); DBG_COUNTER_INC(tx_frees); m = next; } return (ENOBUFS); } #endif #ifdef notyet qidx = count = 0; mp = marr; next = m; do { count++; next = next->m_nextpkt; } while (next != NULL); if (count > nitems(marr)) if ((mp = malloc(count*sizeof(struct mbuf *), M_IFLIB, M_NOWAIT)) == NULL) { /* XXX check nextpkt */ m_freem(m); /* XXX simplify for now */ DBG_COUNTER_INC(tx_frees); return (ENOBUFS); } for (next = m, i = 0; next != NULL; i++) { mp[i] = next; next = next->m_nextpkt; mp[i]->m_nextpkt = NULL; } #endif DBG_COUNTER_INC(tx_seen); err = ifmp_ring_enqueue(txq->ift_br, (void **)&m, 1, TX_BATCH_SIZE, abdicate); if (abdicate) GROUPTASK_ENQUEUE(&txq->ift_task); if (err) { if (!abdicate) GROUPTASK_ENQUEUE(&txq->ift_task); /* support forthcoming later */ #ifdef DRIVER_BACKPRESSURE txq->ift_closed = TRUE; #endif ifmp_ring_check_drainage(txq->ift_br, TX_BATCH_SIZE); m_freem(m); DBG_COUNTER_INC(tx_frees); } return (err); } #ifdef ALTQ /* * The overall approach to integrating iflib with ALTQ is to continue to use * the iflib mp_ring machinery between the ALTQ queue(s) and the hardware * ring. Technically, when using ALTQ, queueing to an intermediate mp_ring * is redundant/unnecessary, but doing so minimizes the amount of * ALTQ-specific code required in iflib. It is assumed that the overhead of * redundantly queueing to an intermediate mp_ring is swamped by the * performance limitations inherent in using ALTQ. * * When ALTQ support is compiled in, all iflib drivers will use a transmit * routine, iflib_altq_if_transmit(), that checks if ALTQ is enabled for the * given interface. If ALTQ is enabled for an interface, then all * transmitted packets for that interface will be submitted to the ALTQ * subsystem via IFQ_ENQUEUE(). We don't use the legacy if_transmit() * implementation because it uses IFQ_HANDOFF(), which will duplicatively * update stats that the iflib machinery handles, and which is sensitve to * the disused IFF_DRV_OACTIVE flag. Additionally, iflib_altq_if_start() * will be installed as the start routine for use by ALTQ facilities that * need to trigger queue drains on a scheduled basis. * */ static void iflib_altq_if_start(if_t ifp) { struct ifaltq *ifq = &ifp->if_snd; struct mbuf *m; IFQ_LOCK(ifq); IFQ_DEQUEUE_NOLOCK(ifq, m); while (m != NULL) { iflib_if_transmit(ifp, m); IFQ_DEQUEUE_NOLOCK(ifq, m); } IFQ_UNLOCK(ifq); } static int iflib_altq_if_transmit(if_t ifp, struct mbuf *m) { int err; if (ALTQ_IS_ENABLED(&ifp->if_snd)) { IFQ_ENQUEUE(&ifp->if_snd, m, err); if (err == 0) iflib_altq_if_start(ifp); } else err = iflib_if_transmit(ifp, m); return (err); } #endif /* ALTQ */ static void iflib_if_qflush(if_t ifp) { if_ctx_t ctx = if_getsoftc(ifp); iflib_txq_t txq = ctx->ifc_txqs; int i; STATE_LOCK(ctx); ctx->ifc_flags |= IFC_QFLUSH; STATE_UNLOCK(ctx); for (i = 0; i < NTXQSETS(ctx); i++, txq++) while (!(ifmp_ring_is_idle(txq->ift_br) || ifmp_ring_is_stalled(txq->ift_br))) iflib_txq_check_drain(txq, 0); STATE_LOCK(ctx); ctx->ifc_flags &= ~IFC_QFLUSH; STATE_UNLOCK(ctx); /* * When ALTQ is enabled, this will also take care of purging the * ALTQ queue(s). */ if_qflush(ifp); } #define IFCAP_FLAGS (IFCAP_HWCSUM_IPV6 | IFCAP_HWCSUM | IFCAP_LRO | \ IFCAP_TSO | IFCAP_VLAN_HWTAGGING | IFCAP_HWSTATS | \ IFCAP_VLAN_MTU | IFCAP_VLAN_HWFILTER | \ IFCAP_VLAN_HWTSO | IFCAP_VLAN_HWCSUM | IFCAP_MEXTPG) static int iflib_if_ioctl(if_t ifp, u_long command, caddr_t data) { if_ctx_t ctx = if_getsoftc(ifp); struct ifreq *ifr = (struct ifreq *)data; #if defined(INET) || defined(INET6) struct ifaddr *ifa = (struct ifaddr *)data; #endif bool avoid_reset = false; int err = 0, reinit = 0, bits; switch (command) { case SIOCSIFADDR: #ifdef INET if (ifa->ifa_addr->sa_family == AF_INET) avoid_reset = true; #endif #ifdef INET6 if (ifa->ifa_addr->sa_family == AF_INET6) avoid_reset = true; #endif /* ** Calling init results in link renegotiation, ** so we avoid doing it when possible. */ if (avoid_reset) { if_setflagbits(ifp, IFF_UP,0); if (!(if_getdrvflags(ifp) & IFF_DRV_RUNNING)) reinit = 1; #ifdef INET if (!(if_getflags(ifp) & IFF_NOARP)) arp_ifinit(ifp, ifa); #endif } else err = ether_ioctl(ifp, command, data); break; case SIOCSIFMTU: CTX_LOCK(ctx); if (ifr->ifr_mtu == if_getmtu(ifp)) { CTX_UNLOCK(ctx); break; } bits = if_getdrvflags(ifp); /* stop the driver and free any clusters before proceeding */ iflib_stop(ctx); if ((err = IFDI_MTU_SET(ctx, ifr->ifr_mtu)) == 0) { STATE_LOCK(ctx); if (ifr->ifr_mtu > ctx->ifc_max_fl_buf_size) ctx->ifc_flags |= IFC_MULTISEG; else ctx->ifc_flags &= ~IFC_MULTISEG; STATE_UNLOCK(ctx); err = if_setmtu(ifp, ifr->ifr_mtu); } iflib_init_locked(ctx); STATE_LOCK(ctx); if_setdrvflags(ifp, bits); STATE_UNLOCK(ctx); CTX_UNLOCK(ctx); break; case SIOCSIFFLAGS: CTX_LOCK(ctx); if (if_getflags(ifp) & IFF_UP) { if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) { if ((if_getflags(ifp) ^ ctx->ifc_if_flags) & (IFF_PROMISC | IFF_ALLMULTI)) { CTX_UNLOCK(ctx); err = IFDI_PROMISC_SET(ctx, if_getflags(ifp)); CTX_LOCK(ctx); } } else reinit = 1; } else if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) { iflib_stop(ctx); } ctx->ifc_if_flags = if_getflags(ifp); CTX_UNLOCK(ctx); break; case SIOCADDMULTI: case SIOCDELMULTI: if (if_getdrvflags(ifp) & IFF_DRV_RUNNING) { CTX_LOCK(ctx); IFDI_INTR_DISABLE(ctx); IFDI_MULTI_SET(ctx); IFDI_INTR_ENABLE(ctx); CTX_UNLOCK(ctx); } break; case SIOCSIFMEDIA: CTX_LOCK(ctx); IFDI_MEDIA_SET(ctx); CTX_UNLOCK(ctx); /* FALLTHROUGH */ case SIOCGIFMEDIA: case SIOCGIFXMEDIA: err = ifmedia_ioctl(ifp, ifr, ctx->ifc_mediap, command); break; case SIOCGI2C: { struct ifi2creq i2c; err = copyin(ifr_data_get_ptr(ifr), &i2c, sizeof(i2c)); if (err != 0) break; if (i2c.dev_addr != 0xA0 && i2c.dev_addr != 0xA2) { err = EINVAL; break; } if (i2c.len > sizeof(i2c.data)) { err = EINVAL; break; } if ((err = IFDI_I2C_REQ(ctx, &i2c)) == 0) err = copyout(&i2c, ifr_data_get_ptr(ifr), sizeof(i2c)); break; } case SIOCSIFCAP: { int mask, setmask, oldmask; oldmask = if_getcapenable(ifp); mask = ifr->ifr_reqcap ^ oldmask; mask &= ctx->ifc_softc_ctx.isc_capabilities | IFCAP_MEXTPG; setmask = 0; #ifdef TCP_OFFLOAD setmask |= mask & (IFCAP_TOE4|IFCAP_TOE6); #endif setmask |= (mask & IFCAP_FLAGS); setmask |= (mask & IFCAP_WOL); /* * If any RX csum has changed, change all the ones that * are supported by the driver. */ if (setmask & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6)) { setmask |= ctx->ifc_softc_ctx.isc_capabilities & (IFCAP_RXCSUM | IFCAP_RXCSUM_IPV6); } /* * want to ensure that traffic has stopped before we change any of the flags */ if (setmask) { CTX_LOCK(ctx); bits = if_getdrvflags(ifp); if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL) iflib_stop(ctx); STATE_LOCK(ctx); if_togglecapenable(ifp, setmask); ctx->ifc_softc_ctx.isc_capenable ^= setmask; STATE_UNLOCK(ctx); if (bits & IFF_DRV_RUNNING && setmask & ~IFCAP_WOL) iflib_init_locked(ctx); STATE_LOCK(ctx); if_setdrvflags(ifp, bits); STATE_UNLOCK(ctx); CTX_UNLOCK(ctx); } if_vlancap(ifp); break; } case SIOCGPRIVATE_0: case SIOCSDRVSPEC: case SIOCGDRVSPEC: CTX_LOCK(ctx); err = IFDI_PRIV_IOCTL(ctx, command, data); CTX_UNLOCK(ctx); break; default: err = ether_ioctl(ifp, command, data); break; } if (reinit) iflib_if_init(ctx); return (err); } static uint64_t iflib_if_get_counter(if_t ifp, ift_counter cnt) { if_ctx_t ctx = if_getsoftc(ifp); return (IFDI_GET_COUNTER(ctx, cnt)); } /********************************************************************* * * OTHER FUNCTIONS EXPORTED TO THE STACK * **********************************************************************/ static void iflib_vlan_register(void *arg, if_t ifp, uint16_t vtag) { if_ctx_t ctx = if_getsoftc(ifp); if ((void *)ctx != arg) return; if ((vtag == 0) || (vtag > 4095)) return; if (iflib_in_detach(ctx)) return; CTX_LOCK(ctx); /* Driver may need all untagged packets to be flushed */ if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG)) iflib_stop(ctx); IFDI_VLAN_REGISTER(ctx, vtag); /* Re-init to load the changes, if required */ if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG)) iflib_init_locked(ctx); CTX_UNLOCK(ctx); } static void iflib_vlan_unregister(void *arg, if_t ifp, uint16_t vtag) { if_ctx_t ctx = if_getsoftc(ifp); if ((void *)ctx != arg) return; if ((vtag == 0) || (vtag > 4095)) return; CTX_LOCK(ctx); /* Driver may need all tagged packets to be flushed */ if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG)) iflib_stop(ctx); IFDI_VLAN_UNREGISTER(ctx, vtag); /* Re-init to load the changes, if required */ if (IFDI_NEEDS_RESTART(ctx, IFLIB_RESTART_VLAN_CONFIG)) iflib_init_locked(ctx); CTX_UNLOCK(ctx); } static void iflib_led_func(void *arg, int onoff) { if_ctx_t ctx = arg; CTX_LOCK(ctx); IFDI_LED_FUNC(ctx, onoff); CTX_UNLOCK(ctx); } /********************************************************************* * * BUS FUNCTION DEFINITIONS * **********************************************************************/ int iflib_device_probe(device_t dev) { const pci_vendor_info_t *ent; if_shared_ctx_t sctx; uint16_t pci_device_id, pci_rev_id, pci_subdevice_id, pci_subvendor_id; uint16_t pci_vendor_id; if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC) return (ENOTSUP); pci_vendor_id = pci_get_vendor(dev); pci_device_id = pci_get_device(dev); pci_subvendor_id = pci_get_subvendor(dev); pci_subdevice_id = pci_get_subdevice(dev); pci_rev_id = pci_get_revid(dev); if (sctx->isc_parse_devinfo != NULL) sctx->isc_parse_devinfo(&pci_device_id, &pci_subvendor_id, &pci_subdevice_id, &pci_rev_id); ent = sctx->isc_vendor_info; while (ent->pvi_vendor_id != 0) { if (pci_vendor_id != ent->pvi_vendor_id) { ent++; continue; } if ((pci_device_id == ent->pvi_device_id) && ((pci_subvendor_id == ent->pvi_subvendor_id) || (ent->pvi_subvendor_id == 0)) && ((pci_subdevice_id == ent->pvi_subdevice_id) || (ent->pvi_subdevice_id == 0)) && ((pci_rev_id == ent->pvi_rev_id) || (ent->pvi_rev_id == 0))) { device_set_desc_copy(dev, ent->pvi_name); /* this needs to be changed to zero if the bus probing code * ever stops re-probing on best match because the sctx * may have its values over written by register calls * in subsequent probes */ return (BUS_PROBE_DEFAULT); } ent++; } return (ENXIO); } int iflib_device_probe_vendor(device_t dev) { int probe; probe = iflib_device_probe(dev); if (probe == BUS_PROBE_DEFAULT) return (BUS_PROBE_VENDOR); else return (probe); } static void iflib_reset_qvalues(if_ctx_t ctx) { if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; if_shared_ctx_t sctx = ctx->ifc_sctx; device_t dev = ctx->ifc_dev; int i; if (ctx->ifc_sysctl_ntxqs != 0) scctx->isc_ntxqsets = ctx->ifc_sysctl_ntxqs; if (ctx->ifc_sysctl_nrxqs != 0) scctx->isc_nrxqsets = ctx->ifc_sysctl_nrxqs; for (i = 0; i < sctx->isc_ntxqs; i++) { if (ctx->ifc_sysctl_ntxds[i] != 0) scctx->isc_ntxd[i] = ctx->ifc_sysctl_ntxds[i]; else scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i]; } for (i = 0; i < sctx->isc_nrxqs; i++) { if (ctx->ifc_sysctl_nrxds[i] != 0) scctx->isc_nrxd[i] = ctx->ifc_sysctl_nrxds[i]; else scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i]; } for (i = 0; i < sctx->isc_nrxqs; i++) { if (scctx->isc_nrxd[i] < sctx->isc_nrxd_min[i]) { device_printf(dev, "nrxd%d: %d less than nrxd_min %d - resetting to min\n", i, scctx->isc_nrxd[i], sctx->isc_nrxd_min[i]); scctx->isc_nrxd[i] = sctx->isc_nrxd_min[i]; } if (scctx->isc_nrxd[i] > sctx->isc_nrxd_max[i]) { device_printf(dev, "nrxd%d: %d greater than nrxd_max %d - resetting to max\n", i, scctx->isc_nrxd[i], sctx->isc_nrxd_max[i]); scctx->isc_nrxd[i] = sctx->isc_nrxd_max[i]; } if (!powerof2(scctx->isc_nrxd[i])) { device_printf(dev, "nrxd%d: %d is not a power of 2 - using default value of %d\n", i, scctx->isc_nrxd[i], sctx->isc_nrxd_default[i]); scctx->isc_nrxd[i] = sctx->isc_nrxd_default[i]; } } for (i = 0; i < sctx->isc_ntxqs; i++) { if (scctx->isc_ntxd[i] < sctx->isc_ntxd_min[i]) { device_printf(dev, "ntxd%d: %d less than ntxd_min %d - resetting to min\n", i, scctx->isc_ntxd[i], sctx->isc_ntxd_min[i]); scctx->isc_ntxd[i] = sctx->isc_ntxd_min[i]; } if (scctx->isc_ntxd[i] > sctx->isc_ntxd_max[i]) { device_printf(dev, "ntxd%d: %d greater than ntxd_max %d - resetting to max\n", i, scctx->isc_ntxd[i], sctx->isc_ntxd_max[i]); scctx->isc_ntxd[i] = sctx->isc_ntxd_max[i]; } if (!powerof2(scctx->isc_ntxd[i])) { device_printf(dev, "ntxd%d: %d is not a power of 2 - using default value of %d\n", i, scctx->isc_ntxd[i], sctx->isc_ntxd_default[i]); scctx->isc_ntxd[i] = sctx->isc_ntxd_default[i]; } } } static void iflib_add_pfil(if_ctx_t ctx) { struct pfil_head *pfil; struct pfil_head_args pa; iflib_rxq_t rxq; int i; pa.pa_version = PFIL_VERSION; pa.pa_flags = PFIL_IN; pa.pa_type = PFIL_TYPE_ETHERNET; pa.pa_headname = ctx->ifc_ifp->if_xname; pfil = pfil_head_register(&pa); for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) { rxq->pfil = pfil; } } static void iflib_rem_pfil(if_ctx_t ctx) { struct pfil_head *pfil; iflib_rxq_t rxq; int i; rxq = ctx->ifc_rxqs; pfil = rxq->pfil; for (i = 0; i < NRXQSETS(ctx); i++, rxq++) { rxq->pfil = NULL; } pfil_head_unregister(pfil); } /* * Advance forward by n members of the cpuset ctx->ifc_cpus starting from * cpuid and wrapping as necessary. */ static unsigned int cpuid_advance(if_ctx_t ctx, unsigned int cpuid, unsigned int n) { unsigned int first_valid; unsigned int last_valid; /* cpuid should always be in the valid set */ MPASS(CPU_ISSET(cpuid, &ctx->ifc_cpus)); /* valid set should never be empty */ MPASS(!CPU_EMPTY(&ctx->ifc_cpus)); first_valid = CPU_FFS(&ctx->ifc_cpus) - 1; last_valid = CPU_FLS(&ctx->ifc_cpus) - 1; n = n % CPU_COUNT(&ctx->ifc_cpus); while (n > 0) { do { cpuid++; if (cpuid > last_valid) cpuid = first_valid; } while (!CPU_ISSET(cpuid, &ctx->ifc_cpus)); n--; } return (cpuid); } #if defined(SMP) && defined(SCHED_ULE) extern struct cpu_group *cpu_top; /* CPU topology */ static int find_child_with_core(int cpu, struct cpu_group *grp) { int i; if (grp->cg_children == 0) return -1; MPASS(grp->cg_child); for (i = 0; i < grp->cg_children; i++) { if (CPU_ISSET(cpu, &grp->cg_child[i].cg_mask)) return i; } return -1; } /* * Find an L2 neighbor of the given CPU or return -1 if none found. This * does not distinguish among multiple L2 neighbors if the given CPU has * more than one (it will always return the same result in that case). */ static int find_l2_neighbor(int cpu) { struct cpu_group *grp; int i; grp = cpu_top; if (grp == NULL) return -1; /* * Find the smallest CPU group that contains the given core. */ i = 0; while ((i = find_child_with_core(cpu, grp)) != -1) { /* * If the smallest group containing the given CPU has less * than two members, we conclude the given CPU has no * L2 neighbor. */ if (grp->cg_child[i].cg_count <= 1) return (-1); grp = &grp->cg_child[i]; } /* Must share L2. */ if (grp->cg_level > CG_SHARE_L2 || grp->cg_level == CG_SHARE_NONE) return -1; /* * Select the first member of the set that isn't the reference * CPU, which at this point is guaranteed to exist. */ for (i = 0; i < CPU_SETSIZE; i++) { if (CPU_ISSET(i, &grp->cg_mask) && i != cpu) return (i); } /* Should never be reached */ return (-1); } #else static int find_l2_neighbor(int cpu) { return (-1); } #endif /* * CPU mapping behaviors * --------------------- * 'separate txrx' refers to the separate_txrx sysctl * 'use logical' refers to the use_logical_cores sysctl * 'INTR CPUS' indicates whether bus_get_cpus(INTR_CPUS) succeeded * * separate use INTR * txrx logical CPUS result * ---------- --------- ------ ------------------------------------------------ * - - X RX and TX queues mapped to consecutive physical * cores with RX/TX pairs on same core and excess * of either following * - X X RX and TX queues mapped to consecutive cores * of any type with RX/TX pairs on same core and * excess of either following * X - X RX and TX queues mapped to consecutive physical * cores; all RX then all TX * X X X RX queues mapped to consecutive physical cores * first, then TX queues mapped to L2 neighbor of * the corresponding RX queue if one exists, * otherwise to consecutive physical cores * - n/a - RX and TX queues mapped to consecutive cores of * any type with RX/TX pairs on same core and excess * of either following * X n/a - RX and TX queues mapped to consecutive cores of * any type; all RX then all TX */ static unsigned int get_cpuid_for_queue(if_ctx_t ctx, unsigned int base_cpuid, unsigned int qid, bool is_tx) { if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; unsigned int core_index; if (ctx->ifc_sysctl_separate_txrx) { /* * When using separate CPUs for TX and RX, the assignment * will always be of a consecutive CPU out of the set of * context CPUs, except for the specific case where the * context CPUs are phsyical cores, the use of logical cores * has been enabled, the assignment is for TX, the TX qid * corresponds to an RX qid, and the CPU assigned to the * corresponding RX queue has an L2 neighbor. */ if (ctx->ifc_sysctl_use_logical_cores && ctx->ifc_cpus_are_physical_cores && is_tx && qid < scctx->isc_nrxqsets) { int l2_neighbor; unsigned int rx_cpuid; rx_cpuid = cpuid_advance(ctx, base_cpuid, qid); l2_neighbor = find_l2_neighbor(rx_cpuid); if (l2_neighbor != -1) { return (l2_neighbor); } /* * ... else fall through to the normal * consecutive-after-RX assignment scheme. * * Note that we are assuming that all RX queue CPUs * have an L2 neighbor, or all do not. If a mixed * scenario is possible, we will have to keep track * separately of how many queues prior to this one * were not able to be assigned to an L2 neighbor. */ } if (is_tx) core_index = scctx->isc_nrxqsets + qid; else core_index = qid; } else { core_index = qid; } return (cpuid_advance(ctx, base_cpuid, core_index)); } static uint16_t get_ctx_core_offset(if_ctx_t ctx) { if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; struct cpu_offset *op; cpuset_t assigned_cpus; unsigned int cores_consumed; unsigned int base_cpuid = ctx->ifc_sysctl_core_offset; unsigned int first_valid; unsigned int last_valid; unsigned int i; first_valid = CPU_FFS(&ctx->ifc_cpus) - 1; last_valid = CPU_FLS(&ctx->ifc_cpus) - 1; if (base_cpuid != CORE_OFFSET_UNSPECIFIED) { /* * Align the user-chosen base CPU ID to the next valid CPU * for this device. If the chosen base CPU ID is smaller * than the first valid CPU or larger than the last valid * CPU, we assume the user does not know what the valid * range is for this device and is thinking in terms of a * zero-based reference frame, and so we shift the given * value into the valid range (and wrap accordingly) so the * intent is translated to the proper frame of reference. * If the base CPU ID is within the valid first/last, but * does not correspond to a valid CPU, it is advanced to the * next valid CPU (wrapping if necessary). */ if (base_cpuid < first_valid || base_cpuid > last_valid) { /* shift from zero-based to first_valid-based */ base_cpuid += first_valid; /* wrap to range [first_valid, last_valid] */ base_cpuid = (base_cpuid - first_valid) % (last_valid - first_valid + 1); } if (!CPU_ISSET(base_cpuid, &ctx->ifc_cpus)) { /* * base_cpuid is in [first_valid, last_valid], but * not a member of the valid set. In this case, * there will always be a member of the valid set * with a CPU ID that is greater than base_cpuid, * and we simply advance to it. */ while (!CPU_ISSET(base_cpuid, &ctx->ifc_cpus)) base_cpuid++; } return (base_cpuid); } /* * Determine how many cores will be consumed by performing the CPU * assignments and counting how many of the assigned CPUs correspond * to CPUs in the set of context CPUs. This is done using the CPU * ID first_valid as the base CPU ID, as the base CPU must be within * the set of context CPUs. * * Note not all assigned CPUs will be in the set of context CPUs * when separate CPUs are being allocated to TX and RX queues, * assignment to logical cores has been enabled, the set of context * CPUs contains only physical CPUs, and TX queues are mapped to L2 * neighbors of CPUs that RX queues have been mapped to - in this * case we do only want to count how many CPUs in the set of context * CPUs have been consumed, as that determines the next CPU in that * set to start allocating at for the next device for which * core_offset is not set. */ CPU_ZERO(&assigned_cpus); for (i = 0; i < scctx->isc_ntxqsets; i++) CPU_SET(get_cpuid_for_queue(ctx, first_valid, i, true), &assigned_cpus); for (i = 0; i < scctx->isc_nrxqsets; i++) CPU_SET(get_cpuid_for_queue(ctx, first_valid, i, false), &assigned_cpus); - CPU_AND(&assigned_cpus, &ctx->ifc_cpus); + CPU_AND(&assigned_cpus, &assigned_cpus, &ctx->ifc_cpus); cores_consumed = CPU_COUNT(&assigned_cpus); mtx_lock(&cpu_offset_mtx); SLIST_FOREACH(op, &cpu_offsets, entries) { if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) { base_cpuid = op->next_cpuid; op->next_cpuid = cpuid_advance(ctx, op->next_cpuid, cores_consumed); MPASS(op->refcount < UINT_MAX); op->refcount++; break; } } if (base_cpuid == CORE_OFFSET_UNSPECIFIED) { base_cpuid = first_valid; op = malloc(sizeof(struct cpu_offset), M_IFLIB, M_NOWAIT | M_ZERO); if (op == NULL) { device_printf(ctx->ifc_dev, "allocation for cpu offset failed.\n"); } else { op->next_cpuid = cpuid_advance(ctx, base_cpuid, cores_consumed); op->refcount = 1; CPU_COPY(&ctx->ifc_cpus, &op->set); SLIST_INSERT_HEAD(&cpu_offsets, op, entries); } } mtx_unlock(&cpu_offset_mtx); return (base_cpuid); } static void unref_ctx_core_offset(if_ctx_t ctx) { struct cpu_offset *op, *top; mtx_lock(&cpu_offset_mtx); SLIST_FOREACH_SAFE(op, &cpu_offsets, entries, top) { if (CPU_CMP(&ctx->ifc_cpus, &op->set) == 0) { MPASS(op->refcount > 0); op->refcount--; if (op->refcount == 0) { SLIST_REMOVE(&cpu_offsets, op, cpu_offset, entries); free(op, M_IFLIB); } break; } } mtx_unlock(&cpu_offset_mtx); } int iflib_device_register(device_t dev, void *sc, if_shared_ctx_t sctx, if_ctx_t *ctxp) { if_ctx_t ctx; if_t ifp; if_softc_ctx_t scctx; kobjop_desc_t kobj_desc; kobj_method_t *kobj_method; int err, msix, rid; int num_txd, num_rxd; ctx = malloc(sizeof(* ctx), M_IFLIB, M_WAITOK|M_ZERO); if (sc == NULL) { sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO); device_set_softc(dev, ctx); ctx->ifc_flags |= IFC_SC_ALLOCATED; } ctx->ifc_sctx = sctx; ctx->ifc_dev = dev; ctx->ifc_softc = sc; if ((err = iflib_register(ctx)) != 0) { device_printf(dev, "iflib_register failed %d\n", err); goto fail_ctx_free; } iflib_add_device_sysctl_pre(ctx); scctx = &ctx->ifc_softc_ctx; ifp = ctx->ifc_ifp; iflib_reset_qvalues(ctx); CTX_LOCK(ctx); if ((err = IFDI_ATTACH_PRE(ctx)) != 0) { device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err); goto fail_unlock; } _iflib_pre_assert(scctx); ctx->ifc_txrx = *scctx->isc_txrx; if (sctx->isc_flags & IFLIB_DRIVER_MEDIA) ctx->ifc_mediap = scctx->isc_media; #ifdef INVARIANTS if (scctx->isc_capabilities & IFCAP_TXCSUM) MPASS(scctx->isc_tx_csum_flags); #endif if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_MEXTPG); if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_MEXTPG); if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets)) scctx->isc_ntxqsets = scctx->isc_ntxqsets_max; if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets)) scctx->isc_nrxqsets = scctx->isc_nrxqsets_max; num_txd = iflib_num_tx_descs(ctx); num_rxd = iflib_num_rx_descs(ctx); /* XXX change for per-queue sizes */ device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n", num_txd, num_rxd); if (scctx->isc_tx_nsegments > num_txd / MAX_SINGLE_PACKET_FRACTION) scctx->isc_tx_nsegments = max(1, num_txd / MAX_SINGLE_PACKET_FRACTION); if (scctx->isc_tx_tso_segments_max > num_txd / MAX_SINGLE_PACKET_FRACTION) scctx->isc_tx_tso_segments_max = max(1, num_txd / MAX_SINGLE_PACKET_FRACTION); /* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */ if (if_getcapabilities(ifp) & IFCAP_TSO) { /* * The stack can't handle a TSO size larger than IP_MAXPACKET, * but some MACs do. */ if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max, IP_MAXPACKET)); /* * Take maximum number of m_pullup(9)'s in iflib_parse_header() * into account. In the worst case, each of these calls will * add another mbuf and, thus, the requirement for another DMA * segment. So for best performance, it doesn't make sense to * advertize a maximum of TSO segments that typically will * require defragmentation in iflib_encap(). */ if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3); if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max); } if (scctx->isc_rss_table_size == 0) scctx->isc_rss_table_size = 64; scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1; GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx); /* XXX format name */ taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, NULL, NULL, "admin"); /* Set up cpu set. If it fails, use the set of all CPUs. */ if (bus_get_cpus(dev, INTR_CPUS, sizeof(ctx->ifc_cpus), &ctx->ifc_cpus) != 0) { device_printf(dev, "Unable to fetch CPU list\n"); CPU_COPY(&all_cpus, &ctx->ifc_cpus); ctx->ifc_cpus_are_physical_cores = false; } else ctx->ifc_cpus_are_physical_cores = true; MPASS(CPU_COUNT(&ctx->ifc_cpus) > 0); /* ** Now set up MSI or MSI-X, should return us the number of supported ** vectors (will be 1 for a legacy interrupt and MSI). */ if (sctx->isc_flags & IFLIB_SKIP_MSIX) { msix = scctx->isc_vectors; } else if (scctx->isc_msix_bar != 0) /* * The simple fact that isc_msix_bar is not 0 does not mean we * we have a good value there that is known to work. */ msix = iflib_msix_init(ctx); else { scctx->isc_vectors = 1; scctx->isc_ntxqsets = 1; scctx->isc_nrxqsets = 1; scctx->isc_intr = IFLIB_INTR_LEGACY; msix = 0; } /* Get memory for the station queues */ if ((err = iflib_queues_alloc(ctx))) { device_printf(dev, "Unable to allocate queue memory\n"); goto fail_intr_free; } if ((err = iflib_qset_structures_setup(ctx))) goto fail_queues; /* * Now that we know how many queues there are, get the core offset. */ ctx->ifc_sysctl_core_offset = get_ctx_core_offset(ctx); if (msix > 1) { /* * When using MSI-X, ensure that ifdi_{r,t}x_queue_intr_enable * aren't the default NULL implementation. */ kobj_desc = &ifdi_rx_queue_intr_enable_desc; kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL, kobj_desc); if (kobj_method == &kobj_desc->deflt) { device_printf(dev, "MSI-X requires ifdi_rx_queue_intr_enable method"); err = EOPNOTSUPP; goto fail_queues; } kobj_desc = &ifdi_tx_queue_intr_enable_desc; kobj_method = kobj_lookup_method(((kobj_t)ctx)->ops->cls, NULL, kobj_desc); if (kobj_method == &kobj_desc->deflt) { device_printf(dev, "MSI-X requires ifdi_tx_queue_intr_enable method"); err = EOPNOTSUPP; goto fail_queues; } /* * Assign the MSI-X vectors. * Note that the default NULL ifdi_msix_intr_assign method will * fail here, too. */ err = IFDI_MSIX_INTR_ASSIGN(ctx, msix); if (err != 0) { device_printf(dev, "IFDI_MSIX_INTR_ASSIGN failed %d\n", err); goto fail_queues; } } else if (scctx->isc_intr != IFLIB_INTR_MSIX) { rid = 0; if (scctx->isc_intr == IFLIB_INTR_MSI) { MPASS(msix == 1); rid = 1; } if ((err = iflib_legacy_setup(ctx, ctx->isc_legacy_intr, ctx->ifc_softc, &rid, "irq0")) != 0) { device_printf(dev, "iflib_legacy_setup failed %d\n", err); goto fail_queues; } } else { device_printf(dev, "Cannot use iflib with only 1 MSI-X interrupt!\n"); err = ENODEV; goto fail_queues; } ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet); if ((err = IFDI_ATTACH_POST(ctx)) != 0) { device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err); goto fail_detach; } /* * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported. * This must appear after the call to ether_ifattach() because * ether_ifattach() sets if_hdrlen to the default value. */ if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU) if_setifheaderlen(ifp, sizeof(struct ether_vlan_header)); if ((err = iflib_netmap_attach(ctx))) { device_printf(ctx->ifc_dev, "netmap attach failed: %d\n", err); goto fail_detach; } *ctxp = ctx; DEBUGNET_SET(ctx->ifc_ifp, iflib); if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter); iflib_add_device_sysctl_post(ctx); iflib_add_pfil(ctx); ctx->ifc_flags |= IFC_INIT_DONE; CTX_UNLOCK(ctx); return (0); fail_detach: ether_ifdetach(ctx->ifc_ifp); fail_queues: iflib_tqg_detach(ctx); iflib_tx_structures_free(ctx); iflib_rx_structures_free(ctx); IFDI_DETACH(ctx); IFDI_QUEUES_FREE(ctx); fail_intr_free: iflib_free_intr_mem(ctx); fail_unlock: CTX_UNLOCK(ctx); iflib_deregister(ctx); fail_ctx_free: device_set_softc(ctx->ifc_dev, NULL); if (ctx->ifc_flags & IFC_SC_ALLOCATED) free(ctx->ifc_softc, M_IFLIB); free(ctx, M_IFLIB); return (err); } int iflib_pseudo_register(device_t dev, if_shared_ctx_t sctx, if_ctx_t *ctxp, struct iflib_cloneattach_ctx *clctx) { int num_txd, num_rxd; int err; if_ctx_t ctx; if_t ifp; if_softc_ctx_t scctx; int i; void *sc; ctx = malloc(sizeof(*ctx), M_IFLIB, M_WAITOK|M_ZERO); sc = malloc(sctx->isc_driver->size, M_IFLIB, M_WAITOK|M_ZERO); ctx->ifc_flags |= IFC_SC_ALLOCATED; if (sctx->isc_flags & (IFLIB_PSEUDO|IFLIB_VIRTUAL)) ctx->ifc_flags |= IFC_PSEUDO; ctx->ifc_sctx = sctx; ctx->ifc_softc = sc; ctx->ifc_dev = dev; if ((err = iflib_register(ctx)) != 0) { device_printf(dev, "%s: iflib_register failed %d\n", __func__, err); goto fail_ctx_free; } iflib_add_device_sysctl_pre(ctx); scctx = &ctx->ifc_softc_ctx; ifp = ctx->ifc_ifp; iflib_reset_qvalues(ctx); CTX_LOCK(ctx); if ((err = IFDI_ATTACH_PRE(ctx)) != 0) { device_printf(dev, "IFDI_ATTACH_PRE failed %d\n", err); goto fail_unlock; } if (sctx->isc_flags & IFLIB_GEN_MAC) ether_gen_addr(ifp, &ctx->ifc_mac); if ((err = IFDI_CLONEATTACH(ctx, clctx->cc_ifc, clctx->cc_name, clctx->cc_params)) != 0) { device_printf(dev, "IFDI_CLONEATTACH failed %d\n", err); goto fail_unlock; } #ifdef INVARIANTS if (scctx->isc_capabilities & IFCAP_TXCSUM) MPASS(scctx->isc_tx_csum_flags); #endif if_setcapabilities(ifp, scctx->isc_capabilities | IFCAP_HWSTATS | IFCAP_LINKSTATE); if_setcapenable(ifp, scctx->isc_capenable | IFCAP_HWSTATS | IFCAP_LINKSTATE); ifp->if_flags |= IFF_NOGROUP; if (sctx->isc_flags & IFLIB_PSEUDO) { ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO, 0, NULL); ifmedia_set(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO); if (sctx->isc_flags & IFLIB_PSEUDO_ETHER) { ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet); } else { if_attach(ctx->ifc_ifp); bpfattach(ctx->ifc_ifp, DLT_NULL, sizeof(u_int32_t)); } if ((err = IFDI_ATTACH_POST(ctx)) != 0) { device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err); goto fail_detach; } *ctxp = ctx; /* * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported. * This must appear after the call to ether_ifattach() because * ether_ifattach() sets if_hdrlen to the default value. */ if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU) if_setifheaderlen(ifp, sizeof(struct ether_vlan_header)); if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter); iflib_add_device_sysctl_post(ctx); ctx->ifc_flags |= IFC_INIT_DONE; CTX_UNLOCK(ctx); return (0); } ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_1000_T | IFM_FDX, 0, NULL); ifmedia_add(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO, 0, NULL); ifmedia_set(ctx->ifc_mediap, IFM_ETHER | IFM_AUTO); _iflib_pre_assert(scctx); ctx->ifc_txrx = *scctx->isc_txrx; if (scctx->isc_ntxqsets == 0 || (scctx->isc_ntxqsets_max && scctx->isc_ntxqsets_max < scctx->isc_ntxqsets)) scctx->isc_ntxqsets = scctx->isc_ntxqsets_max; if (scctx->isc_nrxqsets == 0 || (scctx->isc_nrxqsets_max && scctx->isc_nrxqsets_max < scctx->isc_nrxqsets)) scctx->isc_nrxqsets = scctx->isc_nrxqsets_max; num_txd = iflib_num_tx_descs(ctx); num_rxd = iflib_num_rx_descs(ctx); /* XXX change for per-queue sizes */ device_printf(dev, "Using %d TX descriptors and %d RX descriptors\n", num_txd, num_rxd); if (scctx->isc_tx_nsegments > num_txd / MAX_SINGLE_PACKET_FRACTION) scctx->isc_tx_nsegments = max(1, num_txd / MAX_SINGLE_PACKET_FRACTION); if (scctx->isc_tx_tso_segments_max > num_txd / MAX_SINGLE_PACKET_FRACTION) scctx->isc_tx_tso_segments_max = max(1, num_txd / MAX_SINGLE_PACKET_FRACTION); /* TSO parameters - dig these out of the data sheet - simply correspond to tag setup */ if (if_getcapabilities(ifp) & IFCAP_TSO) { /* * The stack can't handle a TSO size larger than IP_MAXPACKET, * but some MACs do. */ if_sethwtsomax(ifp, min(scctx->isc_tx_tso_size_max, IP_MAXPACKET)); /* * Take maximum number of m_pullup(9)'s in iflib_parse_header() * into account. In the worst case, each of these calls will * add another mbuf and, thus, the requirement for another DMA * segment. So for best performance, it doesn't make sense to * advertize a maximum of TSO segments that typically will * require defragmentation in iflib_encap(). */ if_sethwtsomaxsegcount(ifp, scctx->isc_tx_tso_segments_max - 3); if_sethwtsomaxsegsize(ifp, scctx->isc_tx_tso_segsize_max); } if (scctx->isc_rss_table_size == 0) scctx->isc_rss_table_size = 64; scctx->isc_rss_table_mask = scctx->isc_rss_table_size-1; GROUPTASK_INIT(&ctx->ifc_admin_task, 0, _task_fn_admin, ctx); /* XXX format name */ taskqgroup_attach(qgroup_if_config_tqg, &ctx->ifc_admin_task, ctx, NULL, NULL, "admin"); /* XXX --- can support > 1 -- but keep it simple for now */ scctx->isc_intr = IFLIB_INTR_LEGACY; /* Get memory for the station queues */ if ((err = iflib_queues_alloc(ctx))) { device_printf(dev, "Unable to allocate queue memory\n"); goto fail_iflib_detach; } if ((err = iflib_qset_structures_setup(ctx))) { device_printf(dev, "qset structure setup failed %d\n", err); goto fail_queues; } /* * XXX What if anything do we want to do about interrupts? */ ether_ifattach(ctx->ifc_ifp, ctx->ifc_mac.octet); if ((err = IFDI_ATTACH_POST(ctx)) != 0) { device_printf(dev, "IFDI_ATTACH_POST failed %d\n", err); goto fail_detach; } /* * Tell the upper layer(s) if IFCAP_VLAN_MTU is supported. * This must appear after the call to ether_ifattach() because * ether_ifattach() sets if_hdrlen to the default value. */ if (if_getcapabilities(ifp) & IFCAP_VLAN_MTU) if_setifheaderlen(ifp, sizeof(struct ether_vlan_header)); /* XXX handle more than one queue */ for (i = 0; i < scctx->isc_nrxqsets; i++) IFDI_RX_CLSET(ctx, 0, i, ctx->ifc_rxqs[i].ifr_fl[0].ifl_sds.ifsd_cl); *ctxp = ctx; if_setgetcounterfn(ctx->ifc_ifp, iflib_if_get_counter); iflib_add_device_sysctl_post(ctx); ctx->ifc_flags |= IFC_INIT_DONE; CTX_UNLOCK(ctx); return (0); fail_detach: ether_ifdetach(ctx->ifc_ifp); fail_queues: iflib_tqg_detach(ctx); iflib_tx_structures_free(ctx); iflib_rx_structures_free(ctx); fail_iflib_detach: IFDI_DETACH(ctx); IFDI_QUEUES_FREE(ctx); fail_unlock: CTX_UNLOCK(ctx); iflib_deregister(ctx); fail_ctx_free: free(ctx->ifc_softc, M_IFLIB); free(ctx, M_IFLIB); return (err); } int iflib_pseudo_deregister(if_ctx_t ctx) { if_t ifp = ctx->ifc_ifp; if_shared_ctx_t sctx = ctx->ifc_sctx; /* Unregister VLAN event handlers early */ iflib_unregister_vlan_handlers(ctx); if ((sctx->isc_flags & IFLIB_PSEUDO) && (sctx->isc_flags & IFLIB_PSEUDO_ETHER) == 0) { bpfdetach(ifp); if_detach(ifp); } else { ether_ifdetach(ifp); } iflib_tqg_detach(ctx); iflib_tx_structures_free(ctx); iflib_rx_structures_free(ctx); IFDI_DETACH(ctx); IFDI_QUEUES_FREE(ctx); iflib_deregister(ctx); if (ctx->ifc_flags & IFC_SC_ALLOCATED) free(ctx->ifc_softc, M_IFLIB); free(ctx, M_IFLIB); return (0); } int iflib_device_attach(device_t dev) { if_ctx_t ctx; if_shared_ctx_t sctx; if ((sctx = DEVICE_REGISTER(dev)) == NULL || sctx->isc_magic != IFLIB_MAGIC) return (ENOTSUP); pci_enable_busmaster(dev); return (iflib_device_register(dev, NULL, sctx, &ctx)); } int iflib_device_deregister(if_ctx_t ctx) { if_t ifp = ctx->ifc_ifp; device_t dev = ctx->ifc_dev; /* Make sure VLANS are not using driver */ if (if_vlantrunkinuse(ifp)) { device_printf(dev, "Vlan in use, detach first\n"); return (EBUSY); } #ifdef PCI_IOV if (!CTX_IS_VF(ctx) && pci_iov_detach(dev) != 0) { device_printf(dev, "SR-IOV in use; detach first.\n"); return (EBUSY); } #endif STATE_LOCK(ctx); ctx->ifc_flags |= IFC_IN_DETACH; STATE_UNLOCK(ctx); /* Unregister VLAN handlers before calling iflib_stop() */ iflib_unregister_vlan_handlers(ctx); iflib_netmap_detach(ifp); ether_ifdetach(ifp); CTX_LOCK(ctx); iflib_stop(ctx); CTX_UNLOCK(ctx); iflib_rem_pfil(ctx); if (ctx->ifc_led_dev != NULL) led_destroy(ctx->ifc_led_dev); iflib_tqg_detach(ctx); iflib_tx_structures_free(ctx); iflib_rx_structures_free(ctx); CTX_LOCK(ctx); IFDI_DETACH(ctx); IFDI_QUEUES_FREE(ctx); CTX_UNLOCK(ctx); /* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/ iflib_free_intr_mem(ctx); bus_generic_detach(dev); iflib_deregister(ctx); device_set_softc(ctx->ifc_dev, NULL); if (ctx->ifc_flags & IFC_SC_ALLOCATED) free(ctx->ifc_softc, M_IFLIB); unref_ctx_core_offset(ctx); free(ctx, M_IFLIB); return (0); } static void iflib_tqg_detach(if_ctx_t ctx) { iflib_txq_t txq; iflib_rxq_t rxq; int i; struct taskqgroup *tqg; /* XXX drain any dependent tasks */ tqg = qgroup_if_io_tqg; for (txq = ctx->ifc_txqs, i = 0; i < NTXQSETS(ctx); i++, txq++) { callout_drain(&txq->ift_timer); #ifdef DEV_NETMAP callout_drain(&txq->ift_netmap_timer); #endif /* DEV_NETMAP */ if (txq->ift_task.gt_uniq != NULL) taskqgroup_detach(tqg, &txq->ift_task); } for (i = 0, rxq = ctx->ifc_rxqs; i < NRXQSETS(ctx); i++, rxq++) { if (rxq->ifr_task.gt_uniq != NULL) taskqgroup_detach(tqg, &rxq->ifr_task); } tqg = qgroup_if_config_tqg; if (ctx->ifc_admin_task.gt_uniq != NULL) taskqgroup_detach(tqg, &ctx->ifc_admin_task); if (ctx->ifc_vflr_task.gt_uniq != NULL) taskqgroup_detach(tqg, &ctx->ifc_vflr_task); } static void iflib_free_intr_mem(if_ctx_t ctx) { if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_MSIX) { iflib_irq_free(ctx, &ctx->ifc_legacy_irq); } if (ctx->ifc_softc_ctx.isc_intr != IFLIB_INTR_LEGACY) { pci_release_msi(ctx->ifc_dev); } if (ctx->ifc_msix_mem != NULL) { bus_release_resource(ctx->ifc_dev, SYS_RES_MEMORY, rman_get_rid(ctx->ifc_msix_mem), ctx->ifc_msix_mem); ctx->ifc_msix_mem = NULL; } } int iflib_device_detach(device_t dev) { if_ctx_t ctx = device_get_softc(dev); return (iflib_device_deregister(ctx)); } int iflib_device_suspend(device_t dev) { if_ctx_t ctx = device_get_softc(dev); CTX_LOCK(ctx); IFDI_SUSPEND(ctx); CTX_UNLOCK(ctx); return bus_generic_suspend(dev); } int iflib_device_shutdown(device_t dev) { if_ctx_t ctx = device_get_softc(dev); CTX_LOCK(ctx); IFDI_SHUTDOWN(ctx); CTX_UNLOCK(ctx); return bus_generic_suspend(dev); } int iflib_device_resume(device_t dev) { if_ctx_t ctx = device_get_softc(dev); iflib_txq_t txq = ctx->ifc_txqs; CTX_LOCK(ctx); IFDI_RESUME(ctx); iflib_if_init_locked(ctx); CTX_UNLOCK(ctx); for (int i = 0; i < NTXQSETS(ctx); i++, txq++) iflib_txq_check_drain(txq, IFLIB_RESTART_BUDGET); return (bus_generic_resume(dev)); } int iflib_device_iov_init(device_t dev, uint16_t num_vfs, const nvlist_t *params) { int error; if_ctx_t ctx = device_get_softc(dev); CTX_LOCK(ctx); error = IFDI_IOV_INIT(ctx, num_vfs, params); CTX_UNLOCK(ctx); return (error); } void iflib_device_iov_uninit(device_t dev) { if_ctx_t ctx = device_get_softc(dev); CTX_LOCK(ctx); IFDI_IOV_UNINIT(ctx); CTX_UNLOCK(ctx); } int iflib_device_iov_add_vf(device_t dev, uint16_t vfnum, const nvlist_t *params) { int error; if_ctx_t ctx = device_get_softc(dev); CTX_LOCK(ctx); error = IFDI_IOV_VF_ADD(ctx, vfnum, params); CTX_UNLOCK(ctx); return (error); } /********************************************************************* * * MODULE FUNCTION DEFINITIONS * **********************************************************************/ /* * - Start a fast taskqueue thread for each core * - Start a taskqueue for control operations */ static int iflib_module_init(void) { iflib_timer_default = hz / 2; return (0); } static int iflib_module_event_handler(module_t mod, int what, void *arg) { int err; switch (what) { case MOD_LOAD: if ((err = iflib_module_init()) != 0) return (err); break; case MOD_UNLOAD: return (EBUSY); default: return (EOPNOTSUPP); } return (0); } /********************************************************************* * * PUBLIC FUNCTION DEFINITIONS * ordered as in iflib.h * **********************************************************************/ static void _iflib_assert(if_shared_ctx_t sctx) { int i; MPASS(sctx->isc_tx_maxsize); MPASS(sctx->isc_tx_maxsegsize); MPASS(sctx->isc_rx_maxsize); MPASS(sctx->isc_rx_nsegments); MPASS(sctx->isc_rx_maxsegsize); MPASS(sctx->isc_nrxqs >= 1 && sctx->isc_nrxqs <= 8); for (i = 0; i < sctx->isc_nrxqs; i++) { MPASS(sctx->isc_nrxd_min[i]); MPASS(powerof2(sctx->isc_nrxd_min[i])); MPASS(sctx->isc_nrxd_max[i]); MPASS(powerof2(sctx->isc_nrxd_max[i])); MPASS(sctx->isc_nrxd_default[i]); MPASS(powerof2(sctx->isc_nrxd_default[i])); } MPASS(sctx->isc_ntxqs >= 1 && sctx->isc_ntxqs <= 8); for (i = 0; i < sctx->isc_ntxqs; i++) { MPASS(sctx->isc_ntxd_min[i]); MPASS(powerof2(sctx->isc_ntxd_min[i])); MPASS(sctx->isc_ntxd_max[i]); MPASS(powerof2(sctx->isc_ntxd_max[i])); MPASS(sctx->isc_ntxd_default[i]); MPASS(powerof2(sctx->isc_ntxd_default[i])); } } static void _iflib_pre_assert(if_softc_ctx_t scctx) { MPASS(scctx->isc_txrx->ift_txd_encap); MPASS(scctx->isc_txrx->ift_txd_flush); MPASS(scctx->isc_txrx->ift_txd_credits_update); MPASS(scctx->isc_txrx->ift_rxd_available); MPASS(scctx->isc_txrx->ift_rxd_pkt_get); MPASS(scctx->isc_txrx->ift_rxd_refill); MPASS(scctx->isc_txrx->ift_rxd_flush); } static int iflib_register(if_ctx_t ctx) { if_shared_ctx_t sctx = ctx->ifc_sctx; driver_t *driver = sctx->isc_driver; device_t dev = ctx->ifc_dev; if_t ifp; u_char type; int iflags; if ((sctx->isc_flags & IFLIB_PSEUDO) == 0) _iflib_assert(sctx); CTX_LOCK_INIT(ctx); STATE_LOCK_INIT(ctx, device_get_nameunit(ctx->ifc_dev)); if (sctx->isc_flags & IFLIB_PSEUDO) { if (sctx->isc_flags & IFLIB_PSEUDO_ETHER) type = IFT_ETHER; else type = IFT_PPP; } else type = IFT_ETHER; ifp = ctx->ifc_ifp = if_alloc(type); if (ifp == NULL) { device_printf(dev, "can not allocate ifnet structure\n"); return (ENOMEM); } /* * Initialize our context's device specific methods */ kobj_init((kobj_t) ctx, (kobj_class_t) driver); kobj_class_compile((kobj_class_t) driver); if_initname(ifp, device_get_name(dev), device_get_unit(dev)); if_setsoftc(ifp, ctx); if_setdev(ifp, dev); if_setinitfn(ifp, iflib_if_init); if_setioctlfn(ifp, iflib_if_ioctl); #ifdef ALTQ if_setstartfn(ifp, iflib_altq_if_start); if_settransmitfn(ifp, iflib_altq_if_transmit); if_setsendqready(ifp); #else if_settransmitfn(ifp, iflib_if_transmit); #endif if_setqflushfn(ifp, iflib_if_qflush); iflags = IFF_MULTICAST | IFF_KNOWSEPOCH; if ((sctx->isc_flags & IFLIB_PSEUDO) && (sctx->isc_flags & IFLIB_PSEUDO_ETHER) == 0) iflags |= IFF_POINTOPOINT; else iflags |= IFF_BROADCAST | IFF_SIMPLEX; if_setflags(ifp, iflags); ctx->ifc_vlan_attach_event = EVENTHANDLER_REGISTER(vlan_config, iflib_vlan_register, ctx, EVENTHANDLER_PRI_FIRST); ctx->ifc_vlan_detach_event = EVENTHANDLER_REGISTER(vlan_unconfig, iflib_vlan_unregister, ctx, EVENTHANDLER_PRI_FIRST); if ((sctx->isc_flags & IFLIB_DRIVER_MEDIA) == 0) { ctx->ifc_mediap = &ctx->ifc_media; ifmedia_init(ctx->ifc_mediap, IFM_IMASK, iflib_media_change, iflib_media_status); } return (0); } static void iflib_unregister_vlan_handlers(if_ctx_t ctx) { /* Unregister VLAN events */ if (ctx->ifc_vlan_attach_event != NULL) { EVENTHANDLER_DEREGISTER(vlan_config, ctx->ifc_vlan_attach_event); ctx->ifc_vlan_attach_event = NULL; } if (ctx->ifc_vlan_detach_event != NULL) { EVENTHANDLER_DEREGISTER(vlan_unconfig, ctx->ifc_vlan_detach_event); ctx->ifc_vlan_detach_event = NULL; } } static void iflib_deregister(if_ctx_t ctx) { if_t ifp = ctx->ifc_ifp; /* Remove all media */ ifmedia_removeall(&ctx->ifc_media); /* Ensure that VLAN event handlers are unregistered */ iflib_unregister_vlan_handlers(ctx); /* Release kobject reference */ kobj_delete((kobj_t) ctx, NULL); /* Free the ifnet structure */ if_free(ifp); STATE_LOCK_DESTROY(ctx); /* ether_ifdetach calls if_qflush - lock must be destroy afterwards*/ CTX_LOCK_DESTROY(ctx); } static int iflib_queues_alloc(if_ctx_t ctx) { if_shared_ctx_t sctx = ctx->ifc_sctx; if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; device_t dev = ctx->ifc_dev; int nrxqsets = scctx->isc_nrxqsets; int ntxqsets = scctx->isc_ntxqsets; iflib_txq_t txq; iflib_rxq_t rxq; iflib_fl_t fl = NULL; int i, j, cpu, err, txconf, rxconf; iflib_dma_info_t ifdip; uint32_t *rxqsizes = scctx->isc_rxqsizes; uint32_t *txqsizes = scctx->isc_txqsizes; uint8_t nrxqs = sctx->isc_nrxqs; uint8_t ntxqs = sctx->isc_ntxqs; int nfree_lists = sctx->isc_nfl ? sctx->isc_nfl : 1; int fl_offset = (sctx->isc_flags & IFLIB_HAS_RXCQ ? 1 : 0); caddr_t *vaddrs; uint64_t *paddrs; KASSERT(ntxqs > 0, ("number of queues per qset must be at least 1")); KASSERT(nrxqs > 0, ("number of queues per qset must be at least 1")); KASSERT(nrxqs >= fl_offset + nfree_lists, ("there must be at least a rxq for each free list")); /* Allocate the TX ring struct memory */ if (!(ctx->ifc_txqs = (iflib_txq_t) malloc(sizeof(struct iflib_txq) * ntxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) { device_printf(dev, "Unable to allocate TX ring memory\n"); err = ENOMEM; goto fail; } /* Now allocate the RX */ if (!(ctx->ifc_rxqs = (iflib_rxq_t) malloc(sizeof(struct iflib_rxq) * nrxqsets, M_IFLIB, M_NOWAIT | M_ZERO))) { device_printf(dev, "Unable to allocate RX ring memory\n"); err = ENOMEM; goto rx_fail; } txq = ctx->ifc_txqs; rxq = ctx->ifc_rxqs; /* * XXX handle allocation failure */ for (txconf = i = 0, cpu = CPU_FIRST(); i < ntxqsets; i++, txconf++, txq++, cpu = CPU_NEXT(cpu)) { /* Set up some basics */ if ((ifdip = malloc(sizeof(struct iflib_dma_info) * ntxqs, M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) { device_printf(dev, "Unable to allocate TX DMA info memory\n"); err = ENOMEM; goto err_tx_desc; } txq->ift_ifdi = ifdip; for (j = 0; j < ntxqs; j++, ifdip++) { if (iflib_dma_alloc(ctx, txqsizes[j], ifdip, 0)) { device_printf(dev, "Unable to allocate TX descriptors\n"); err = ENOMEM; goto err_tx_desc; } txq->ift_txd_size[j] = scctx->isc_txd_size[j]; bzero((void *)ifdip->idi_vaddr, txqsizes[j]); } txq->ift_ctx = ctx; txq->ift_id = i; if (sctx->isc_flags & IFLIB_HAS_TXCQ) { txq->ift_br_offset = 1; } else { txq->ift_br_offset = 0; } if (iflib_txsd_alloc(txq)) { device_printf(dev, "Critical Failure setting up TX buffers\n"); err = ENOMEM; goto err_tx_desc; } /* Initialize the TX lock */ snprintf(txq->ift_mtx_name, MTX_NAME_LEN, "%s:TX(%d):callout", device_get_nameunit(dev), txq->ift_id); mtx_init(&txq->ift_mtx, txq->ift_mtx_name, NULL, MTX_DEF); callout_init_mtx(&txq->ift_timer, &txq->ift_mtx, 0); txq->ift_timer.c_cpu = cpu; #ifdef DEV_NETMAP callout_init_mtx(&txq->ift_netmap_timer, &txq->ift_mtx, 0); txq->ift_netmap_timer.c_cpu = cpu; #endif /* DEV_NETMAP */ err = ifmp_ring_alloc(&txq->ift_br, 2048, txq, iflib_txq_drain, iflib_txq_can_drain, M_IFLIB, M_WAITOK); if (err) { /* XXX free any allocated rings */ device_printf(dev, "Unable to allocate buf_ring\n"); goto err_tx_desc; } } for (rxconf = i = 0; i < nrxqsets; i++, rxconf++, rxq++) { /* Set up some basics */ callout_init(&rxq->ifr_watchdog, 1); if ((ifdip = malloc(sizeof(struct iflib_dma_info) * nrxqs, M_IFLIB, M_NOWAIT | M_ZERO)) == NULL) { device_printf(dev, "Unable to allocate RX DMA info memory\n"); err = ENOMEM; goto err_tx_desc; } rxq->ifr_ifdi = ifdip; /* XXX this needs to be changed if #rx queues != #tx queues */ rxq->ifr_ntxqirq = 1; rxq->ifr_txqid[0] = i; for (j = 0; j < nrxqs; j++, ifdip++) { if (iflib_dma_alloc(ctx, rxqsizes[j], ifdip, 0)) { device_printf(dev, "Unable to allocate RX descriptors\n"); err = ENOMEM; goto err_tx_desc; } bzero((void *)ifdip->idi_vaddr, rxqsizes[j]); } rxq->ifr_ctx = ctx; rxq->ifr_id = i; rxq->ifr_fl_offset = fl_offset; rxq->ifr_nfl = nfree_lists; if (!(fl = (iflib_fl_t) malloc(sizeof(struct iflib_fl) * nfree_lists, M_IFLIB, M_NOWAIT | M_ZERO))) { device_printf(dev, "Unable to allocate free list memory\n"); err = ENOMEM; goto err_tx_desc; } rxq->ifr_fl = fl; for (j = 0; j < nfree_lists; j++) { fl[j].ifl_rxq = rxq; fl[j].ifl_id = j; fl[j].ifl_ifdi = &rxq->ifr_ifdi[j + rxq->ifr_fl_offset]; fl[j].ifl_rxd_size = scctx->isc_rxd_size[j]; } /* Allocate receive buffers for the ring */ if (iflib_rxsd_alloc(rxq)) { device_printf(dev, "Critical Failure setting up receive buffers\n"); err = ENOMEM; goto err_rx_desc; } for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) fl->ifl_rx_bitmap = bit_alloc(fl->ifl_size, M_IFLIB, M_WAITOK); } /* TXQs */ vaddrs = malloc(sizeof(caddr_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK); paddrs = malloc(sizeof(uint64_t)*ntxqsets*ntxqs, M_IFLIB, M_WAITOK); for (i = 0; i < ntxqsets; i++) { iflib_dma_info_t di = ctx->ifc_txqs[i].ift_ifdi; for (j = 0; j < ntxqs; j++, di++) { vaddrs[i*ntxqs + j] = di->idi_vaddr; paddrs[i*ntxqs + j] = di->idi_paddr; } } if ((err = IFDI_TX_QUEUES_ALLOC(ctx, vaddrs, paddrs, ntxqs, ntxqsets)) != 0) { device_printf(ctx->ifc_dev, "Unable to allocate device TX queue\n"); iflib_tx_structures_free(ctx); free(vaddrs, M_IFLIB); free(paddrs, M_IFLIB); goto err_rx_desc; } free(vaddrs, M_IFLIB); free(paddrs, M_IFLIB); /* RXQs */ vaddrs = malloc(sizeof(caddr_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK); paddrs = malloc(sizeof(uint64_t)*nrxqsets*nrxqs, M_IFLIB, M_WAITOK); for (i = 0; i < nrxqsets; i++) { iflib_dma_info_t di = ctx->ifc_rxqs[i].ifr_ifdi; for (j = 0; j < nrxqs; j++, di++) { vaddrs[i*nrxqs + j] = di->idi_vaddr; paddrs[i*nrxqs + j] = di->idi_paddr; } } if ((err = IFDI_RX_QUEUES_ALLOC(ctx, vaddrs, paddrs, nrxqs, nrxqsets)) != 0) { device_printf(ctx->ifc_dev, "Unable to allocate device RX queue\n"); iflib_tx_structures_free(ctx); free(vaddrs, M_IFLIB); free(paddrs, M_IFLIB); goto err_rx_desc; } free(vaddrs, M_IFLIB); free(paddrs, M_IFLIB); return (0); /* XXX handle allocation failure changes */ err_rx_desc: err_tx_desc: rx_fail: if (ctx->ifc_rxqs != NULL) free(ctx->ifc_rxqs, M_IFLIB); ctx->ifc_rxqs = NULL; if (ctx->ifc_txqs != NULL) free(ctx->ifc_txqs, M_IFLIB); ctx->ifc_txqs = NULL; fail: return (err); } static int iflib_tx_structures_setup(if_ctx_t ctx) { iflib_txq_t txq = ctx->ifc_txqs; int i; for (i = 0; i < NTXQSETS(ctx); i++, txq++) iflib_txq_setup(txq); return (0); } static void iflib_tx_structures_free(if_ctx_t ctx) { iflib_txq_t txq = ctx->ifc_txqs; if_shared_ctx_t sctx = ctx->ifc_sctx; int i, j; for (i = 0; i < NTXQSETS(ctx); i++, txq++) { for (j = 0; j < sctx->isc_ntxqs; j++) iflib_dma_free(&txq->ift_ifdi[j]); iflib_txq_destroy(txq); } free(ctx->ifc_txqs, M_IFLIB); ctx->ifc_txqs = NULL; } /********************************************************************* * * Initialize all receive rings. * **********************************************************************/ static int iflib_rx_structures_setup(if_ctx_t ctx) { iflib_rxq_t rxq = ctx->ifc_rxqs; int q; #if defined(INET6) || defined(INET) int err, i; #endif for (q = 0; q < ctx->ifc_softc_ctx.isc_nrxqsets; q++, rxq++) { #if defined(INET6) || defined(INET) if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO) { err = tcp_lro_init_args(&rxq->ifr_lc, ctx->ifc_ifp, TCP_LRO_ENTRIES, min(1024, ctx->ifc_softc_ctx.isc_nrxd[rxq->ifr_fl_offset])); if (err != 0) { device_printf(ctx->ifc_dev, "LRO Initialization failed!\n"); goto fail; } } #endif IFDI_RXQ_SETUP(ctx, rxq->ifr_id); } return (0); #if defined(INET6) || defined(INET) fail: /* * Free LRO resources allocated so far, we will only handle * the rings that completed, the failing case will have * cleaned up for itself. 'q' failed, so its the terminus. */ rxq = ctx->ifc_rxqs; for (i = 0; i < q; ++i, rxq++) { if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO) tcp_lro_free(&rxq->ifr_lc); } return (err); #endif } /********************************************************************* * * Free all receive rings. * **********************************************************************/ static void iflib_rx_structures_free(if_ctx_t ctx) { iflib_rxq_t rxq = ctx->ifc_rxqs; if_shared_ctx_t sctx = ctx->ifc_sctx; int i, j; for (i = 0; i < ctx->ifc_softc_ctx.isc_nrxqsets; i++, rxq++) { for (j = 0; j < sctx->isc_nrxqs; j++) iflib_dma_free(&rxq->ifr_ifdi[j]); iflib_rx_sds_free(rxq); #if defined(INET6) || defined(INET) if (if_getcapabilities(ctx->ifc_ifp) & IFCAP_LRO) tcp_lro_free(&rxq->ifr_lc); #endif } free(ctx->ifc_rxqs, M_IFLIB); ctx->ifc_rxqs = NULL; } static int iflib_qset_structures_setup(if_ctx_t ctx) { int err; /* * It is expected that the caller takes care of freeing queues if this * fails. */ if ((err = iflib_tx_structures_setup(ctx)) != 0) { device_printf(ctx->ifc_dev, "iflib_tx_structures_setup failed: %d\n", err); return (err); } if ((err = iflib_rx_structures_setup(ctx)) != 0) device_printf(ctx->ifc_dev, "iflib_rx_structures_setup failed: %d\n", err); return (err); } int iflib_irq_alloc(if_ctx_t ctx, if_irq_t irq, int rid, driver_filter_t filter, void *filter_arg, driver_intr_t handler, void *arg, const char *name) { return (_iflib_irq_alloc(ctx, irq, rid, filter, handler, arg, name)); } /* Just to avoid copy/paste */ static inline int iflib_irq_set_affinity(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type, int qid, struct grouptask *gtask, struct taskqgroup *tqg, void *uniq, const char *name) { device_t dev; unsigned int base_cpuid, cpuid; int err; dev = ctx->ifc_dev; base_cpuid = ctx->ifc_sysctl_core_offset; cpuid = get_cpuid_for_queue(ctx, base_cpuid, qid, type == IFLIB_INTR_TX); err = taskqgroup_attach_cpu(tqg, gtask, uniq, cpuid, dev, irq ? irq->ii_res : NULL, name); if (err) { device_printf(dev, "taskqgroup_attach_cpu failed %d\n", err); return (err); } #ifdef notyet if (cpuid > ctx->ifc_cpuid_highest) ctx->ifc_cpuid_highest = cpuid; #endif return (0); } int iflib_irq_alloc_generic(if_ctx_t ctx, if_irq_t irq, int rid, iflib_intr_type_t type, driver_filter_t *filter, void *filter_arg, int qid, const char *name) { device_t dev; struct grouptask *gtask; struct taskqgroup *tqg; iflib_filter_info_t info; gtask_fn_t *fn; int tqrid, err; driver_filter_t *intr_fast; void *q; info = &ctx->ifc_filter_info; tqrid = rid; switch (type) { /* XXX merge tx/rx for netmap? */ case IFLIB_INTR_TX: q = &ctx->ifc_txqs[qid]; info = &ctx->ifc_txqs[qid].ift_filter_info; gtask = &ctx->ifc_txqs[qid].ift_task; tqg = qgroup_if_io_tqg; fn = _task_fn_tx; intr_fast = iflib_fast_intr; GROUPTASK_INIT(gtask, 0, fn, q); ctx->ifc_flags |= IFC_NETMAP_TX_IRQ; break; case IFLIB_INTR_RX: q = &ctx->ifc_rxqs[qid]; info = &ctx->ifc_rxqs[qid].ifr_filter_info; gtask = &ctx->ifc_rxqs[qid].ifr_task; tqg = qgroup_if_io_tqg; fn = _task_fn_rx; intr_fast = iflib_fast_intr; NET_GROUPTASK_INIT(gtask, 0, fn, q); break; case IFLIB_INTR_RXTX: q = &ctx->ifc_rxqs[qid]; info = &ctx->ifc_rxqs[qid].ifr_filter_info; gtask = &ctx->ifc_rxqs[qid].ifr_task; tqg = qgroup_if_io_tqg; fn = _task_fn_rx; intr_fast = iflib_fast_intr_rxtx; NET_GROUPTASK_INIT(gtask, 0, fn, q); break; case IFLIB_INTR_ADMIN: q = ctx; tqrid = -1; info = &ctx->ifc_filter_info; gtask = &ctx->ifc_admin_task; tqg = qgroup_if_config_tqg; fn = _task_fn_admin; intr_fast = iflib_fast_intr_ctx; break; default: device_printf(ctx->ifc_dev, "%s: unknown net intr type\n", __func__); return (EINVAL); } info->ifi_filter = filter; info->ifi_filter_arg = filter_arg; info->ifi_task = gtask; info->ifi_ctx = q; dev = ctx->ifc_dev; err = _iflib_irq_alloc(ctx, irq, rid, intr_fast, NULL, info, name); if (err != 0) { device_printf(dev, "_iflib_irq_alloc failed %d\n", err); return (err); } if (type == IFLIB_INTR_ADMIN) return (0); if (tqrid != -1) { err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg, q, name); if (err) return (err); } else { taskqgroup_attach(tqg, gtask, q, dev, irq->ii_res, name); } return (0); } void iflib_softirq_alloc_generic(if_ctx_t ctx, if_irq_t irq, iflib_intr_type_t type, void *arg, int qid, const char *name) { device_t dev; struct grouptask *gtask; struct taskqgroup *tqg; gtask_fn_t *fn; void *q; int err; switch (type) { case IFLIB_INTR_TX: q = &ctx->ifc_txqs[qid]; gtask = &ctx->ifc_txqs[qid].ift_task; tqg = qgroup_if_io_tqg; fn = _task_fn_tx; GROUPTASK_INIT(gtask, 0, fn, q); break; case IFLIB_INTR_RX: q = &ctx->ifc_rxqs[qid]; gtask = &ctx->ifc_rxqs[qid].ifr_task; tqg = qgroup_if_io_tqg; fn = _task_fn_rx; NET_GROUPTASK_INIT(gtask, 0, fn, q); break; case IFLIB_INTR_IOV: q = ctx; gtask = &ctx->ifc_vflr_task; tqg = qgroup_if_config_tqg; fn = _task_fn_iov; GROUPTASK_INIT(gtask, 0, fn, q); break; default: panic("unknown net intr type"); } err = iflib_irq_set_affinity(ctx, irq, type, qid, gtask, tqg, q, name); if (err) { dev = ctx->ifc_dev; taskqgroup_attach(tqg, gtask, q, dev, irq ? irq->ii_res : NULL, name); } } void iflib_irq_free(if_ctx_t ctx, if_irq_t irq) { if (irq->ii_tag) bus_teardown_intr(ctx->ifc_dev, irq->ii_res, irq->ii_tag); if (irq->ii_res) bus_release_resource(ctx->ifc_dev, SYS_RES_IRQ, rman_get_rid(irq->ii_res), irq->ii_res); } static int iflib_legacy_setup(if_ctx_t ctx, driver_filter_t filter, void *filter_arg, int *rid, const char *name) { iflib_txq_t txq = ctx->ifc_txqs; iflib_rxq_t rxq = ctx->ifc_rxqs; if_irq_t irq = &ctx->ifc_legacy_irq; iflib_filter_info_t info; device_t dev; struct grouptask *gtask; struct resource *res; struct taskqgroup *tqg; void *q; int err, tqrid; bool rx_only; q = &ctx->ifc_rxqs[0]; info = &rxq[0].ifr_filter_info; gtask = &rxq[0].ifr_task; tqg = qgroup_if_io_tqg; tqrid = *rid; rx_only = (ctx->ifc_sctx->isc_flags & IFLIB_SINGLE_IRQ_RX_ONLY) != 0; ctx->ifc_flags |= IFC_LEGACY; info->ifi_filter = filter; info->ifi_filter_arg = filter_arg; info->ifi_task = gtask; info->ifi_ctx = rx_only ? ctx : q; dev = ctx->ifc_dev; /* We allocate a single interrupt resource */ err = _iflib_irq_alloc(ctx, irq, tqrid, rx_only ? iflib_fast_intr_ctx : iflib_fast_intr_rxtx, NULL, info, name); if (err != 0) return (err); NET_GROUPTASK_INIT(gtask, 0, _task_fn_rx, q); res = irq->ii_res; taskqgroup_attach(tqg, gtask, q, dev, res, name); GROUPTASK_INIT(&txq->ift_task, 0, _task_fn_tx, txq); taskqgroup_attach(qgroup_if_io_tqg, &txq->ift_task, txq, dev, res, "tx"); return (0); } void iflib_led_create(if_ctx_t ctx) { ctx->ifc_led_dev = led_create(iflib_led_func, ctx, device_get_nameunit(ctx->ifc_dev)); } void iflib_tx_intr_deferred(if_ctx_t ctx, int txqid) { GROUPTASK_ENQUEUE(&ctx->ifc_txqs[txqid].ift_task); } void iflib_rx_intr_deferred(if_ctx_t ctx, int rxqid) { GROUPTASK_ENQUEUE(&ctx->ifc_rxqs[rxqid].ifr_task); } void iflib_admin_intr_deferred(if_ctx_t ctx) { MPASS(ctx->ifc_admin_task.gt_taskqueue != NULL); GROUPTASK_ENQUEUE(&ctx->ifc_admin_task); } void iflib_iov_intr_deferred(if_ctx_t ctx) { GROUPTASK_ENQUEUE(&ctx->ifc_vflr_task); } void iflib_io_tqg_attach(struct grouptask *gt, void *uniq, int cpu, const char *name) { taskqgroup_attach_cpu(qgroup_if_io_tqg, gt, uniq, cpu, NULL, NULL, name); } void iflib_config_gtask_init(void *ctx, struct grouptask *gtask, gtask_fn_t *fn, const char *name) { GROUPTASK_INIT(gtask, 0, fn, ctx); taskqgroup_attach(qgroup_if_config_tqg, gtask, gtask, NULL, NULL, name); } void iflib_config_gtask_deinit(struct grouptask *gtask) { taskqgroup_detach(qgroup_if_config_tqg, gtask); } void iflib_link_state_change(if_ctx_t ctx, int link_state, uint64_t baudrate) { if_t ifp = ctx->ifc_ifp; iflib_txq_t txq = ctx->ifc_txqs; if_setbaudrate(ifp, baudrate); if (baudrate >= IF_Gbps(10)) { STATE_LOCK(ctx); ctx->ifc_flags |= IFC_PREFETCH; STATE_UNLOCK(ctx); } /* If link down, disable watchdog */ if ((ctx->ifc_link_state == LINK_STATE_UP) && (link_state == LINK_STATE_DOWN)) { for (int i = 0; i < ctx->ifc_softc_ctx.isc_ntxqsets; i++, txq++) txq->ift_qstatus = IFLIB_QUEUE_IDLE; } ctx->ifc_link_state = link_state; if_link_state_change(ifp, link_state); } static int iflib_tx_credits_update(if_ctx_t ctx, iflib_txq_t txq) { int credits; #ifdef INVARIANTS int credits_pre = txq->ift_cidx_processed; #endif bus_dmamap_sync(txq->ift_ifdi->idi_tag, txq->ift_ifdi->idi_map, BUS_DMASYNC_POSTREAD); if ((credits = ctx->isc_txd_credits_update(ctx->ifc_softc, txq->ift_id, true)) == 0) return (0); txq->ift_processed += credits; txq->ift_cidx_processed += credits; MPASS(credits_pre + credits == txq->ift_cidx_processed); if (txq->ift_cidx_processed >= txq->ift_size) txq->ift_cidx_processed -= txq->ift_size; return (credits); } static int iflib_rxd_avail(if_ctx_t ctx, iflib_rxq_t rxq, qidx_t cidx, qidx_t budget) { iflib_fl_t fl; u_int i; for (i = 0, fl = &rxq->ifr_fl[0]; i < rxq->ifr_nfl; i++, fl++) bus_dmamap_sync(fl->ifl_ifdi->idi_tag, fl->ifl_ifdi->idi_map, BUS_DMASYNC_POSTREAD | BUS_DMASYNC_POSTWRITE); return (ctx->isc_rxd_available(ctx->ifc_softc, rxq->ifr_id, cidx, budget)); } void iflib_add_int_delay_sysctl(if_ctx_t ctx, const char *name, const char *description, if_int_delay_info_t info, int offset, int value) { info->iidi_ctx = ctx; info->iidi_offset = offset; info->iidi_value = value; SYSCTL_ADD_PROC(device_get_sysctl_ctx(ctx->ifc_dev), SYSCTL_CHILDREN(device_get_sysctl_tree(ctx->ifc_dev)), OID_AUTO, name, CTLTYPE_INT | CTLFLAG_RW | CTLFLAG_MPSAFE, info, 0, iflib_sysctl_int_delay, "I", description); } struct sx * iflib_ctx_lock_get(if_ctx_t ctx) { return (&ctx->ifc_ctx_sx); } static int iflib_msix_init(if_ctx_t ctx) { device_t dev = ctx->ifc_dev; if_shared_ctx_t sctx = ctx->ifc_sctx; if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; int admincnt, bar, err, iflib_num_rx_queues, iflib_num_tx_queues; int msgs, queuemsgs, queues, rx_queues, tx_queues, vectors; iflib_num_tx_queues = ctx->ifc_sysctl_ntxqs; iflib_num_rx_queues = ctx->ifc_sysctl_nrxqs; if (bootverbose) device_printf(dev, "msix_init qsets capped at %d\n", imax(scctx->isc_ntxqsets, scctx->isc_nrxqsets)); /* Override by tuneable */ if (scctx->isc_disable_msix) goto msi; /* First try MSI-X */ if ((msgs = pci_msix_count(dev)) == 0) { if (bootverbose) device_printf(dev, "MSI-X not supported or disabled\n"); goto msi; } bar = ctx->ifc_softc_ctx.isc_msix_bar; /* * bar == -1 => "trust me I know what I'm doing" * Some drivers are for hardware that is so shoddily * documented that no one knows which bars are which * so the developer has to map all bars. This hack * allows shoddy garbage to use MSI-X in this framework. */ if (bar != -1) { ctx->ifc_msix_mem = bus_alloc_resource_any(dev, SYS_RES_MEMORY, &bar, RF_ACTIVE); if (ctx->ifc_msix_mem == NULL) { device_printf(dev, "Unable to map MSI-X table\n"); goto msi; } } admincnt = sctx->isc_admin_intrcnt; #if IFLIB_DEBUG /* use only 1 qset in debug mode */ queuemsgs = min(msgs - admincnt, 1); #else queuemsgs = msgs - admincnt; #endif #ifdef RSS queues = imin(queuemsgs, rss_getnumbuckets()); #else queues = queuemsgs; #endif queues = imin(CPU_COUNT(&ctx->ifc_cpus), queues); if (bootverbose) device_printf(dev, "intr CPUs: %d queue msgs: %d admincnt: %d\n", CPU_COUNT(&ctx->ifc_cpus), queuemsgs, admincnt); #ifdef RSS /* If we're doing RSS, clamp at the number of RSS buckets */ if (queues > rss_getnumbuckets()) queues = rss_getnumbuckets(); #endif if (iflib_num_rx_queues > 0 && iflib_num_rx_queues < queuemsgs - admincnt) rx_queues = iflib_num_rx_queues; else rx_queues = queues; if (rx_queues > scctx->isc_nrxqsets) rx_queues = scctx->isc_nrxqsets; /* * We want this to be all logical CPUs by default */ if (iflib_num_tx_queues > 0 && iflib_num_tx_queues < queues) tx_queues = iflib_num_tx_queues; else tx_queues = mp_ncpus; if (tx_queues > scctx->isc_ntxqsets) tx_queues = scctx->isc_ntxqsets; if (ctx->ifc_sysctl_qs_eq_override == 0) { #ifdef INVARIANTS if (tx_queues != rx_queues) device_printf(dev, "queue equality override not set, capping rx_queues at %d and tx_queues at %d\n", min(rx_queues, tx_queues), min(rx_queues, tx_queues)); #endif tx_queues = min(rx_queues, tx_queues); rx_queues = min(rx_queues, tx_queues); } vectors = rx_queues + admincnt; if (msgs < vectors) { device_printf(dev, "insufficient number of MSI-X vectors " "(supported %d, need %d)\n", msgs, vectors); goto msi; } device_printf(dev, "Using %d RX queues %d TX queues\n", rx_queues, tx_queues); msgs = vectors; if ((err = pci_alloc_msix(dev, &vectors)) == 0) { if (vectors != msgs) { device_printf(dev, "Unable to allocate sufficient MSI-X vectors " "(got %d, need %d)\n", vectors, msgs); pci_release_msi(dev); if (bar != -1) { bus_release_resource(dev, SYS_RES_MEMORY, bar, ctx->ifc_msix_mem); ctx->ifc_msix_mem = NULL; } goto msi; } device_printf(dev, "Using MSI-X interrupts with %d vectors\n", vectors); scctx->isc_vectors = vectors; scctx->isc_nrxqsets = rx_queues; scctx->isc_ntxqsets = tx_queues; scctx->isc_intr = IFLIB_INTR_MSIX; return (vectors); } else { device_printf(dev, "failed to allocate %d MSI-X vectors, err: %d\n", vectors, err); if (bar != -1) { bus_release_resource(dev, SYS_RES_MEMORY, bar, ctx->ifc_msix_mem); ctx->ifc_msix_mem = NULL; } } msi: vectors = pci_msi_count(dev); scctx->isc_nrxqsets = 1; scctx->isc_ntxqsets = 1; scctx->isc_vectors = vectors; if (vectors == 1 && pci_alloc_msi(dev, &vectors) == 0) { device_printf(dev,"Using an MSI interrupt\n"); scctx->isc_intr = IFLIB_INTR_MSI; } else { scctx->isc_vectors = 1; device_printf(dev,"Using a Legacy interrupt\n"); scctx->isc_intr = IFLIB_INTR_LEGACY; } return (vectors); } static const char *ring_states[] = { "IDLE", "BUSY", "STALLED", "ABDICATED" }; static int mp_ring_state_handler(SYSCTL_HANDLER_ARGS) { int rc; uint16_t *state = ((uint16_t *)oidp->oid_arg1); struct sbuf *sb; const char *ring_state = "UNKNOWN"; /* XXX needed ? */ rc = sysctl_wire_old_buffer(req, 0); MPASS(rc == 0); if (rc != 0) return (rc); sb = sbuf_new_for_sysctl(NULL, NULL, 80, req); MPASS(sb != NULL); if (sb == NULL) return (ENOMEM); if (state[3] <= 3) ring_state = ring_states[state[3]]; sbuf_printf(sb, "pidx_head: %04hd pidx_tail: %04hd cidx: %04hd state: %s", state[0], state[1], state[2], ring_state); rc = sbuf_finish(sb); sbuf_delete(sb); return(rc); } enum iflib_ndesc_handler { IFLIB_NTXD_HANDLER, IFLIB_NRXD_HANDLER, }; static int mp_ndesc_handler(SYSCTL_HANDLER_ARGS) { if_ctx_t ctx = (void *)arg1; enum iflib_ndesc_handler type = arg2; char buf[256] = {0}; qidx_t *ndesc; char *p, *next; int nqs, rc, i; nqs = 8; switch(type) { case IFLIB_NTXD_HANDLER: ndesc = ctx->ifc_sysctl_ntxds; if (ctx->ifc_sctx) nqs = ctx->ifc_sctx->isc_ntxqs; break; case IFLIB_NRXD_HANDLER: ndesc = ctx->ifc_sysctl_nrxds; if (ctx->ifc_sctx) nqs = ctx->ifc_sctx->isc_nrxqs; break; default: printf("%s: unhandled type\n", __func__); return (EINVAL); } if (nqs == 0) nqs = 8; for (i=0; i<8; i++) { if (i >= nqs) break; if (i) strcat(buf, ","); sprintf(strchr(buf, 0), "%d", ndesc[i]); } rc = sysctl_handle_string(oidp, buf, sizeof(buf), req); if (rc || req->newptr == NULL) return rc; for (i = 0, next = buf, p = strsep(&next, " ,"); i < 8 && p; i++, p = strsep(&next, " ,")) { ndesc[i] = strtoul(p, NULL, 10); } return(rc); } #define NAME_BUFLEN 32 static void iflib_add_device_sysctl_pre(if_ctx_t ctx) { device_t dev = iflib_get_dev(ctx); struct sysctl_oid_list *child, *oid_list; struct sysctl_ctx_list *ctx_list; struct sysctl_oid *node; ctx_list = device_get_sysctl_ctx(dev); child = SYSCTL_CHILDREN(device_get_sysctl_tree(dev)); ctx->ifc_sysctl_node = node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, "iflib", CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "IFLIB fields"); oid_list = SYSCTL_CHILDREN(node); SYSCTL_ADD_CONST_STRING(ctx_list, oid_list, OID_AUTO, "driver_version", CTLFLAG_RD, ctx->ifc_sctx->isc_driver_version, "driver version"); SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_ntxqs", CTLFLAG_RWTUN, &ctx->ifc_sysctl_ntxqs, 0, "# of txqs to use, 0 => use default #"); SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_nrxqs", CTLFLAG_RWTUN, &ctx->ifc_sysctl_nrxqs, 0, "# of rxqs to use, 0 => use default #"); SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "override_qs_enable", CTLFLAG_RWTUN, &ctx->ifc_sysctl_qs_eq_override, 0, "permit #txq != #rxq"); SYSCTL_ADD_INT(ctx_list, oid_list, OID_AUTO, "disable_msix", CTLFLAG_RWTUN, &ctx->ifc_softc_ctx.isc_disable_msix, 0, "disable MSI-X (default 0)"); SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "rx_budget", CTLFLAG_RWTUN, &ctx->ifc_sysctl_rx_budget, 0, "set the RX budget"); SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "tx_abdicate", CTLFLAG_RWTUN, &ctx->ifc_sysctl_tx_abdicate, 0, "cause TX to abdicate instead of running to completion"); ctx->ifc_sysctl_core_offset = CORE_OFFSET_UNSPECIFIED; SYSCTL_ADD_U16(ctx_list, oid_list, OID_AUTO, "core_offset", CTLFLAG_RDTUN, &ctx->ifc_sysctl_core_offset, 0, "offset to start using cores at"); SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "separate_txrx", CTLFLAG_RDTUN, &ctx->ifc_sysctl_separate_txrx, 0, "use separate cores for TX and RX"); SYSCTL_ADD_U8(ctx_list, oid_list, OID_AUTO, "use_logical_cores", CTLFLAG_RDTUN, &ctx->ifc_sysctl_use_logical_cores, 0, "try to make use of logical cores for TX and RX"); /* XXX change for per-queue sizes */ SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_ntxds", CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx, IFLIB_NTXD_HANDLER, mp_ndesc_handler, "A", "list of # of TX descriptors to use, 0 = use default #"); SYSCTL_ADD_PROC(ctx_list, oid_list, OID_AUTO, "override_nrxds", CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NEEDGIANT, ctx, IFLIB_NRXD_HANDLER, mp_ndesc_handler, "A", "list of # of RX descriptors to use, 0 = use default #"); } static void iflib_add_device_sysctl_post(if_ctx_t ctx) { if_shared_ctx_t sctx = ctx->ifc_sctx; if_softc_ctx_t scctx = &ctx->ifc_softc_ctx; device_t dev = iflib_get_dev(ctx); struct sysctl_oid_list *child; struct sysctl_ctx_list *ctx_list; iflib_fl_t fl; iflib_txq_t txq; iflib_rxq_t rxq; int i, j; char namebuf[NAME_BUFLEN]; char *qfmt; struct sysctl_oid *queue_node, *fl_node, *node; struct sysctl_oid_list *queue_list, *fl_list; ctx_list = device_get_sysctl_ctx(dev); node = ctx->ifc_sysctl_node; child = SYSCTL_CHILDREN(node); if (scctx->isc_ntxqsets > 100) qfmt = "txq%03d"; else if (scctx->isc_ntxqsets > 10) qfmt = "txq%02d"; else qfmt = "txq%d"; for (i = 0, txq = ctx->ifc_txqs; i < scctx->isc_ntxqsets; i++, txq++) { snprintf(namebuf, NAME_BUFLEN, qfmt, i); queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name"); queue_list = SYSCTL_CHILDREN(queue_node); SYSCTL_ADD_INT(ctx_list, queue_list, OID_AUTO, "cpu", CTLFLAG_RD, &txq->ift_task.gt_cpu, 0, "cpu this queue is bound to"); #if MEMORY_LOGGING SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_dequeued", CTLFLAG_RD, &txq->ift_dequeued, "total mbufs freed"); SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_enqueued", CTLFLAG_RD, &txq->ift_enqueued, "total mbufs enqueued"); #endif SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag", CTLFLAG_RD, &txq->ift_mbuf_defrag, "# of times m_defrag was called"); SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "m_pullups", CTLFLAG_RD, &txq->ift_pullups, "# of times m_pullup was called"); SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "mbuf_defrag_failed", CTLFLAG_RD, &txq->ift_mbuf_defrag_failed, "# of times m_defrag failed"); SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_desc_avail", CTLFLAG_RD, &txq->ift_no_desc_avail, "# of times no descriptors were available"); SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "tx_map_failed", CTLFLAG_RD, &txq->ift_map_failed, "# of times DMA map failed"); SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txd_encap_efbig", CTLFLAG_RD, &txq->ift_txd_encap_efbig, "# of times txd_encap returned EFBIG"); SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "no_tx_dma_setup", CTLFLAG_RD, &txq->ift_no_tx_dma_setup, "# of times map failed for other than EFBIG"); SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_pidx", CTLFLAG_RD, &txq->ift_pidx, 1, "Producer Index"); SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx", CTLFLAG_RD, &txq->ift_cidx, 1, "Consumer Index"); SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_cidx_processed", CTLFLAG_RD, &txq->ift_cidx_processed, 1, "Consumer Index seen by credit update"); SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "txq_in_use", CTLFLAG_RD, &txq->ift_in_use, 1, "descriptors in use"); SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_processed", CTLFLAG_RD, &txq->ift_processed, "descriptors procesed for clean"); SYSCTL_ADD_QUAD(ctx_list, queue_list, OID_AUTO, "txq_cleaned", CTLFLAG_RD, &txq->ift_cleaned, "total cleaned"); SYSCTL_ADD_PROC(ctx_list, queue_list, OID_AUTO, "ring_state", CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT, __DEVOLATILE(uint64_t *, &txq->ift_br->state), 0, mp_ring_state_handler, "A", "soft ring state"); SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_enqueues", CTLFLAG_RD, &txq->ift_br->enqueues, "# of enqueues to the mp_ring for this queue"); SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_drops", CTLFLAG_RD, &txq->ift_br->drops, "# of drops in the mp_ring for this queue"); SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_starts", CTLFLAG_RD, &txq->ift_br->starts, "# of normal consumer starts in the mp_ring for this queue"); SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_stalls", CTLFLAG_RD, &txq->ift_br->stalls, "# of consumer stalls in the mp_ring for this queue"); SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_restarts", CTLFLAG_RD, &txq->ift_br->restarts, "# of consumer restarts in the mp_ring for this queue"); SYSCTL_ADD_COUNTER_U64(ctx_list, queue_list, OID_AUTO, "r_abdications", CTLFLAG_RD, &txq->ift_br->abdications, "# of consumer abdications in the mp_ring for this queue"); } if (scctx->isc_nrxqsets > 100) qfmt = "rxq%03d"; else if (scctx->isc_nrxqsets > 10) qfmt = "rxq%02d"; else qfmt = "rxq%d"; for (i = 0, rxq = ctx->ifc_rxqs; i < scctx->isc_nrxqsets; i++, rxq++) { snprintf(namebuf, NAME_BUFLEN, qfmt, i); queue_node = SYSCTL_ADD_NODE(ctx_list, child, OID_AUTO, namebuf, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "Queue Name"); queue_list = SYSCTL_CHILDREN(queue_node); SYSCTL_ADD_INT(ctx_list, queue_list, OID_AUTO, "cpu", CTLFLAG_RD, &rxq->ifr_task.gt_cpu, 0, "cpu this queue is bound to"); if (sctx->isc_flags & IFLIB_HAS_RXCQ) { SYSCTL_ADD_U16(ctx_list, queue_list, OID_AUTO, "rxq_cq_cidx", CTLFLAG_RD, &rxq->ifr_cq_cidx, 1, "Consumer Index"); } for (j = 0, fl = rxq->ifr_fl; j < rxq->ifr_nfl; j++, fl++) { snprintf(namebuf, NAME_BUFLEN, "rxq_fl%d", j); fl_node = SYSCTL_ADD_NODE(ctx_list, queue_list, OID_AUTO, namebuf, CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, "freelist Name"); fl_list = SYSCTL_CHILDREN(fl_node); SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "pidx", CTLFLAG_RD, &fl->ifl_pidx, 1, "Producer Index"); SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "cidx", CTLFLAG_RD, &fl->ifl_cidx, 1, "Consumer Index"); SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "credits", CTLFLAG_RD, &fl->ifl_credits, 1, "credits available"); SYSCTL_ADD_U16(ctx_list, fl_list, OID_AUTO, "buf_size", CTLFLAG_RD, &fl->ifl_buf_size, 1, "buffer size"); #if MEMORY_LOGGING SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_enqueued", CTLFLAG_RD, &fl->ifl_m_enqueued, "mbufs allocated"); SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_m_dequeued", CTLFLAG_RD, &fl->ifl_m_dequeued, "mbufs freed"); SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_enqueued", CTLFLAG_RD, &fl->ifl_cl_enqueued, "clusters allocated"); SYSCTL_ADD_QUAD(ctx_list, fl_list, OID_AUTO, "fl_cl_dequeued", CTLFLAG_RD, &fl->ifl_cl_dequeued, "clusters freed"); #endif } } } void iflib_request_reset(if_ctx_t ctx) { STATE_LOCK(ctx); ctx->ifc_flags |= IFC_DO_RESET; STATE_UNLOCK(ctx); } #ifndef __NO_STRICT_ALIGNMENT static struct mbuf * iflib_fixup_rx(struct mbuf *m) { struct mbuf *n; if (m->m_len <= (MCLBYTES - ETHER_HDR_LEN)) { bcopy(m->m_data, m->m_data + ETHER_HDR_LEN, m->m_len); m->m_data += ETHER_HDR_LEN; n = m; } else { MGETHDR(n, M_NOWAIT, MT_DATA); if (n == NULL) { m_freem(m); return (NULL); } bcopy(m->m_data, n->m_data, ETHER_HDR_LEN); m->m_data += ETHER_HDR_LEN; m->m_len -= ETHER_HDR_LEN; n->m_len = ETHER_HDR_LEN; M_MOVE_PKTHDR(n, m); n->m_next = m; } return (n); } #endif #ifdef DEBUGNET static void iflib_debugnet_init(if_t ifp, int *nrxr, int *ncl, int *clsize) { if_ctx_t ctx; ctx = if_getsoftc(ifp); CTX_LOCK(ctx); *nrxr = NRXQSETS(ctx); *ncl = ctx->ifc_rxqs[0].ifr_fl->ifl_size; *clsize = ctx->ifc_rxqs[0].ifr_fl->ifl_buf_size; CTX_UNLOCK(ctx); } static void iflib_debugnet_event(if_t ifp, enum debugnet_ev event) { if_ctx_t ctx; if_softc_ctx_t scctx; iflib_fl_t fl; iflib_rxq_t rxq; int i, j; ctx = if_getsoftc(ifp); scctx = &ctx->ifc_softc_ctx; switch (event) { case DEBUGNET_START: for (i = 0; i < scctx->isc_nrxqsets; i++) { rxq = &ctx->ifc_rxqs[i]; for (j = 0; j < rxq->ifr_nfl; j++) { fl = rxq->ifr_fl; fl->ifl_zone = m_getzone(fl->ifl_buf_size); } } iflib_no_tx_batch = 1; break; default: break; } } static int iflib_debugnet_transmit(if_t ifp, struct mbuf *m) { if_ctx_t ctx; iflib_txq_t txq; int error; ctx = if_getsoftc(ifp); if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) != IFF_DRV_RUNNING) return (EBUSY); txq = &ctx->ifc_txqs[0]; error = iflib_encap(txq, &m); if (error == 0) (void)iflib_txd_db_check(txq, true); return (error); } static int iflib_debugnet_poll(if_t ifp, int count) { struct epoch_tracker et; if_ctx_t ctx; if_softc_ctx_t scctx; iflib_txq_t txq; int i; ctx = if_getsoftc(ifp); scctx = &ctx->ifc_softc_ctx; if ((if_getdrvflags(ifp) & (IFF_DRV_RUNNING | IFF_DRV_OACTIVE)) != IFF_DRV_RUNNING) return (EBUSY); txq = &ctx->ifc_txqs[0]; (void)iflib_completed_tx_reclaim(txq, RECLAIM_THRESH(ctx)); NET_EPOCH_ENTER(et); for (i = 0; i < scctx->isc_nrxqsets; i++) (void)iflib_rxeof(&ctx->ifc_rxqs[i], 16 /* XXX */); NET_EPOCH_EXIT(et); return (0); } #endif /* DEBUGNET */ diff --git a/sys/powerpc/ofw/ofw_pcibus.c b/sys/powerpc/ofw/ofw_pcibus.c index 47cb4665b832..0b771a509c9f 100644 --- a/sys/powerpc/ofw/ofw_pcibus.c +++ b/sys/powerpc/ofw/ofw_pcibus.c @@ -1,430 +1,430 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 1997, Stefan Esser * Copyright (c) 2000, Michael Smith * Copyright (c) 2000, BSDi * Copyright (c) 2003, Thomas Moestl * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice unmodified, this list of conditions, and the following * disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. */ #include __FBSDID("$FreeBSD$"); #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include "ofw_pcibus.h" #include "pcib_if.h" #include "pci_if.h" typedef uint32_t ofw_pci_intr_t; /* Methods */ static device_probe_t ofw_pcibus_probe; static device_attach_t ofw_pcibus_attach; static pci_alloc_devinfo_t ofw_pcibus_alloc_devinfo; static pci_assign_interrupt_t ofw_pcibus_assign_interrupt; static ofw_bus_get_devinfo_t ofw_pcibus_get_devinfo; static bus_child_deleted_t ofw_pcibus_child_deleted; static int ofw_pcibus_child_pnpinfo_str_method(device_t cbdev, device_t child, char *buf, size_t buflen); static void ofw_pcibus_enum_devtree(device_t dev, u_int domain, u_int busno); static void ofw_pcibus_enum_bus(device_t dev, u_int domain, u_int busno); static device_method_t ofw_pcibus_methods[] = { /* Device interface */ DEVMETHOD(device_probe, ofw_pcibus_probe), DEVMETHOD(device_attach, ofw_pcibus_attach), /* Bus interface */ DEVMETHOD(bus_child_deleted, ofw_pcibus_child_deleted), DEVMETHOD(bus_child_pnpinfo_str, ofw_pcibus_child_pnpinfo_str_method), DEVMETHOD(bus_rescan, bus_null_rescan), DEVMETHOD(bus_get_cpus, ofw_pcibus_get_cpus), DEVMETHOD(bus_get_domain, ofw_pcibus_get_domain), /* PCI interface */ DEVMETHOD(pci_alloc_devinfo, ofw_pcibus_alloc_devinfo), DEVMETHOD(pci_assign_interrupt, ofw_pcibus_assign_interrupt), /* ofw_bus interface */ DEVMETHOD(ofw_bus_get_devinfo, ofw_pcibus_get_devinfo), DEVMETHOD(ofw_bus_get_compat, ofw_bus_gen_get_compat), DEVMETHOD(ofw_bus_get_model, ofw_bus_gen_get_model), DEVMETHOD(ofw_bus_get_name, ofw_bus_gen_get_name), DEVMETHOD(ofw_bus_get_node, ofw_bus_gen_get_node), DEVMETHOD(ofw_bus_get_type, ofw_bus_gen_get_type), DEVMETHOD_END }; static devclass_t pci_devclass; DEFINE_CLASS_1(pci, ofw_pcibus_driver, ofw_pcibus_methods, sizeof(struct pci_softc), pci_driver); EARLY_DRIVER_MODULE(ofw_pcibus, pcib, ofw_pcibus_driver, pci_devclass, 0, 0, BUS_PASS_BUS); MODULE_VERSION(ofw_pcibus, 1); MODULE_DEPEND(ofw_pcibus, pci, 1, 1, 1); static int ofw_devices_only = 0; TUNABLE_INT("hw.pci.ofw_devices_only", &ofw_devices_only); static int ofw_pcibus_probe(device_t dev) { if (ofw_bus_get_node(dev) == -1) return (ENXIO); device_set_desc(dev, "OFW PCI bus"); return (BUS_PROBE_DEFAULT); } static int ofw_pcibus_attach(device_t dev) { u_int busno, domain; int error; error = pci_attach_common(dev); if (error) return (error); domain = pcib_get_domain(dev); busno = pcib_get_bus(dev); /* * Attach those children represented in the device tree. */ ofw_pcibus_enum_devtree(dev, domain, busno); /* * We now attach any laggard devices. FDT, for instance, allows * the device tree to enumerate only some PCI devices. Apple's * OF device tree on some Grackle-based hardware can also miss * functions on multi-function cards. */ if (!ofw_devices_only) ofw_pcibus_enum_bus(dev, domain, busno); return (bus_generic_attach(dev)); } struct pci_devinfo * ofw_pcibus_alloc_devinfo(device_t dev) { struct ofw_pcibus_devinfo *dinfo; dinfo = malloc(sizeof(*dinfo), M_DEVBUF, M_WAITOK | M_ZERO); return (&dinfo->opd_dinfo); } static void ofw_pcibus_enum_devtree(device_t dev, u_int domain, u_int busno) { device_t pcib; struct ofw_pci_register pcir; struct ofw_pcibus_devinfo *dinfo; phandle_t node, child; u_int func, slot; int intline; pcib = device_get_parent(dev); node = ofw_bus_get_node(dev); for (child = OF_child(node); child != 0; child = OF_peer(child)) { if (OF_getencprop(child, "reg", (pcell_t *)&pcir, sizeof(pcir)) == -1) continue; slot = OFW_PCI_PHYS_HI_DEVICE(pcir.phys_hi); func = OFW_PCI_PHYS_HI_FUNCTION(pcir.phys_hi); /* Some OFW device trees contain dupes. */ if (pci_find_dbsf(domain, busno, slot, func) != NULL) continue; /* * The preset in the intline register is usually bogus. Reset * it such that the PCI code will reroute the interrupt if * needed. */ intline = PCI_INVALID_IRQ; if (OF_getproplen(child, "interrupts") > 0) intline = 0; PCIB_WRITE_CONFIG(pcib, busno, slot, func, PCIR_INTLINE, intline, 1); /* * Now set up the PCI and OFW bus layer devinfo and add it * to the PCI bus. */ dinfo = (struct ofw_pcibus_devinfo *)pci_read_device(pcib, dev, domain, busno, slot, func); if (dinfo == NULL) continue; if (ofw_bus_gen_setup_devinfo(&dinfo->opd_obdinfo, child) != 0) { pci_freecfg((struct pci_devinfo *)dinfo); continue; } dinfo->opd_dma_tag = NULL; pci_add_child(dev, (struct pci_devinfo *)dinfo); /* * Some devices don't have an intpin set, but do have * interrupts. These are fully specified, and set in the * interrupts property, so add that value to the device's * resource list. */ if (dinfo->opd_dinfo.cfg.intpin == 0) ofw_bus_intr_to_rl(dev, child, &dinfo->opd_dinfo.resources, NULL); } } /* * The following is an almost exact clone of pci_add_children(), with the * addition that it (a) will not add children that have already been added, * and (b) will set up the OFW devinfo to point to invalid values. This is * to handle non-enumerated PCI children as exist in FDT and on the second * function of the Rage 128 in my Blue & White G3. */ static void ofw_pcibus_enum_bus(device_t dev, u_int domain, u_int busno) { device_t pcib; struct ofw_pcibus_devinfo *dinfo; int maxslots; int s, f, pcifunchigh; uint8_t hdrtype; pcib = device_get_parent(dev); maxslots = PCIB_MAXSLOTS(pcib); for (s = 0; s <= maxslots; s++) { pcifunchigh = 0; f = 0; DELAY(1); hdrtype = PCIB_READ_CONFIG(pcib, busno, s, f, PCIR_HDRTYPE, 1); if ((hdrtype & PCIM_HDRTYPE) > PCI_MAXHDRTYPE) continue; if (hdrtype & PCIM_MFDEV) pcifunchigh = PCI_FUNCMAX; for (f = 0; f <= pcifunchigh; f++) { /* Filter devices we have already added */ if (pci_find_dbsf(domain, busno, s, f) != NULL) continue; dinfo = (struct ofw_pcibus_devinfo *)pci_read_device( pcib, dev, domain, busno, s, f); if (dinfo == NULL) continue; dinfo->opd_dma_tag = NULL; dinfo->opd_obdinfo.obd_node = -1; dinfo->opd_obdinfo.obd_name = NULL; dinfo->opd_obdinfo.obd_compat = NULL; dinfo->opd_obdinfo.obd_type = NULL; dinfo->opd_obdinfo.obd_model = NULL; /* * For non OFW-devices, don't believe 0 * for an interrupt. */ if (dinfo->opd_dinfo.cfg.intline == 0) { dinfo->opd_dinfo.cfg.intline = PCI_INVALID_IRQ; PCIB_WRITE_CONFIG(pcib, busno, s, f, PCIR_INTLINE, PCI_INVALID_IRQ, 1); } pci_add_child(dev, (struct pci_devinfo *)dinfo); } } } static void ofw_pcibus_child_deleted(device_t dev, device_t child) { struct ofw_pcibus_devinfo *dinfo; dinfo = device_get_ivars(child); ofw_bus_gen_destroy_devinfo(&dinfo->opd_obdinfo); pci_child_deleted(dev, child); } static int ofw_pcibus_child_pnpinfo_str_method(device_t cbdev, device_t child, char *buf, size_t buflen) { pci_child_pnpinfo_str_method(cbdev, child, buf, buflen); if (ofw_bus_get_node(child) != -1) { strlcat(buf, " ", buflen); /* Separate info */ ofw_bus_gen_child_pnpinfo_str(cbdev, child, buf, buflen); } return (0); } static int ofw_pcibus_assign_interrupt(device_t dev, device_t child) { ofw_pci_intr_t intr[2]; phandle_t node, iparent; int isz, icells; node = ofw_bus_get_node(child); if (node == -1) { /* Non-firmware enumerated child, use standard routing */ intr[0] = pci_get_intpin(child); return (PCIB_ROUTE_INTERRUPT(device_get_parent(dev), child, intr[0])); } /* * Try to determine the node's interrupt parent so we know which * PIC to use. */ iparent = -1; if (OF_getencprop(node, "interrupt-parent", &iparent, sizeof(iparent)) < 0) iparent = -1; icells = 1; if (iparent != -1) OF_getencprop(OF_node_from_xref(iparent), "#interrupt-cells", &icells, sizeof(icells)); /* * Any AAPL,interrupts property gets priority and is * fully specified (i.e. does not need routing) */ isz = OF_getencprop(node, "AAPL,interrupts", intr, sizeof(intr)); if (isz == sizeof(intr[0])*icells) return ((iparent == -1) ? intr[0] : ofw_bus_map_intr(dev, iparent, icells, intr)); isz = OF_getencprop(node, "interrupts", intr, sizeof(intr)); if (isz == sizeof(intr[0])*icells) { if (iparent != -1) intr[0] = ofw_bus_map_intr(dev, iparent, icells, intr); } else { /* No property: our best guess is the intpin. */ intr[0] = pci_get_intpin(child); } /* * If we got intr from a property, it may or may not be an intpin. * For on-board devices, it frequently is not, and is completely out * of the valid intpin range. For PCI slots, it hopefully is, * otherwise we will have trouble interfacing with non-OFW buses * such as cardbus. * Since we cannot tell which it is without violating layering, we * will always use the route_interrupt method, and treat exceptions * on the level they become apparent. */ return (PCIB_ROUTE_INTERRUPT(device_get_parent(dev), child, intr[0])); } static const struct ofw_bus_devinfo * ofw_pcibus_get_devinfo(device_t bus, device_t dev) { struct ofw_pcibus_devinfo *dinfo; dinfo = device_get_ivars(dev); return (&dinfo->opd_obdinfo); } int ofw_pcibus_get_cpus(device_t dev, device_t child, enum cpu_sets op, size_t setsize, cpuset_t *cpuset) { int d, error; d = platform_node_numa_domain(ofw_bus_get_node(dev)); switch (op) { case LOCAL_CPUS: if (setsize != sizeof(cpuset_t)) return (EINVAL); *cpuset = cpuset_domain[d]; return (0); case INTR_CPUS: error = bus_generic_get_cpus(dev, child, op, setsize, cpuset); if (error != 0) return (error); if (setsize != sizeof(cpuset_t)) return (EINVAL); - CPU_AND(cpuset, &cpuset_domain[d]); + CPU_AND(cpuset, cpuset, &cpuset_domain[d]); return (0); default: return (bus_generic_get_cpus(dev, child, op, setsize, cpuset)); } return (0); } /* * Fetch the NUMA domain for the given device 'dev'. * * If a device has a _PXM method, map that to a NUMA domain. * Otherwise, pass the request up to the parent. * If there's no matching domain or the domain cannot be * determined, return ENOENT. */ int ofw_pcibus_get_domain(device_t dev, device_t child, int *domain) { *domain = platform_node_numa_domain(ofw_bus_get_node(child)); return (0); } diff --git a/sys/sys/_cpuset.h b/sys/sys/_cpuset.h index 9033acb9ba14..fb39420238c8 100644 --- a/sys/sys/_cpuset.h +++ b/sys/sys/_cpuset.h @@ -1,52 +1,59 @@ -/*- +#/*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2008, Jeffrey Roberson * All rights reserved. * * Copyright (c) 2008 Nokia Corporation * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice unmodified, this list of conditions, and the following * disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. * * $FreeBSD$ */ #ifndef _SYS__CPUSET_H_ #define _SYS__CPUSET_H_ #include #ifdef _KERNEL #define CPU_SETSIZE MAXCPU #endif #define CPU_MAXSIZE 256 #ifndef CPU_SETSIZE #define CPU_SETSIZE CPU_MAXSIZE #endif __BITSET_DEFINE(_cpuset, CPU_SETSIZE); typedef struct _cpuset cpuset_t; +#ifndef _KERNEL +__BEGIN_DECLS +cpuset_t *__cpuset_alloc(size_t set_size); +void __cpuset_free(cpuset_t *ptr); +__END_DECLS +#endif + #endif /* !_SYS__CPUSET_H_ */ diff --git a/sys/sys/bitset.h b/sys/sys/bitset.h index 47c4106d4127..4952d48b1b3c 100644 --- a/sys/sys/bitset.h +++ b/sys/sys/bitset.h @@ -1,360 +1,362 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2008, Jeffrey Roberson * All rights reserved. * * Copyright (c) 2008 Nokia Corporation * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice unmodified, this list of conditions, and the following * disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. * * $FreeBSD$ */ #ifndef _SYS_BITSET_H_ #define _SYS_BITSET_H_ /* * Whether expr is both constant and true. Result is itself constant. * Used to enable optimizations for sets with a known small size. */ #define __constexpr_cond(expr) (__builtin_constant_p((expr)) && (expr)) #define __bitset_mask(_s, n) \ (1UL << (__constexpr_cond(__bitset_words((_s)) == 1) ? \ (__size_t)(n) : ((n) % _BITSET_BITS))) #define __bitset_word(_s, n) \ (__constexpr_cond(__bitset_words((_s)) == 1) ? \ 0 : ((n) / _BITSET_BITS)) #define __BIT_CLR(_s, n, p) \ ((p)->__bits[__bitset_word(_s, n)] &= ~__bitset_mask((_s), (n))) #define __BIT_COPY(_s, f, t) (void)(*(t) = *(f)) #define __BIT_ISSET(_s, n, p) \ ((((p)->__bits[__bitset_word(_s, n)] & __bitset_mask((_s), (n))) != 0)) #define __BIT_SET(_s, n, p) \ ((p)->__bits[__bitset_word(_s, n)] |= __bitset_mask((_s), (n))) #define __BIT_ZERO(_s, p) do { \ __size_t __i; \ for (__i = 0; __i < __bitset_words((_s)); __i++) \ (p)->__bits[__i] = 0L; \ } while (0) #define __BIT_FILL(_s, p) do { \ __size_t __i; \ for (__i = 0; __i < __bitset_words((_s)); __i++) \ (p)->__bits[__i] = -1L; \ } while (0) #define __BIT_SETOF(_s, n, p) do { \ __BIT_ZERO(_s, p); \ (p)->__bits[__bitset_word(_s, n)] = __bitset_mask((_s), (n)); \ } while (0) /* Is p empty. */ #define __BIT_EMPTY(_s, p) __extension__ ({ \ __size_t __i; \ for (__i = 0; __i < __bitset_words((_s)); __i++) \ if ((p)->__bits[__i]) \ break; \ __i == __bitset_words((_s)); \ }) /* Is p full set. */ #define __BIT_ISFULLSET(_s, p) __extension__ ({ \ __size_t __i; \ for (__i = 0; __i < __bitset_words((_s)); __i++) \ if ((p)->__bits[__i] != (long)-1) \ break; \ __i == __bitset_words((_s)); \ }) /* Is c a subset of p. */ #define __BIT_SUBSET(_s, p, c) __extension__ ({ \ __size_t __i; \ for (__i = 0; __i < __bitset_words((_s)); __i++) \ if (((c)->__bits[__i] & \ (p)->__bits[__i]) != \ (c)->__bits[__i]) \ break; \ __i == __bitset_words((_s)); \ }) /* Are there any common bits between b & c? */ #define __BIT_OVERLAP(_s, p, c) __extension__ ({ \ __size_t __i; \ for (__i = 0; __i < __bitset_words((_s)); __i++) \ if (((c)->__bits[__i] & \ (p)->__bits[__i]) != 0) \ break; \ __i != __bitset_words((_s)); \ }) /* Compare two sets, returns 0 if equal 1 otherwise. */ #define __BIT_CMP(_s, p, c) __extension__ ({ \ __size_t __i; \ for (__i = 0; __i < __bitset_words((_s)); __i++) \ if (((c)->__bits[__i] != \ (p)->__bits[__i])) \ break; \ __i != __bitset_words((_s)); \ }) #define __BIT_OR(_s, d, s) do { \ __size_t __i; \ for (__i = 0; __i < __bitset_words((_s)); __i++) \ (d)->__bits[__i] |= (s)->__bits[__i]; \ } while (0) #define __BIT_OR2(_s, d, s1, s2) do { \ __size_t __i; \ for (__i = 0; __i < __bitset_words((_s)); __i++) \ (d)->__bits[__i] = (s1)->__bits[__i] | (s2)->__bits[__i];\ } while (0) #define __BIT_AND(_s, d, s) do { \ __size_t __i; \ for (__i = 0; __i < __bitset_words((_s)); __i++) \ (d)->__bits[__i] &= (s)->__bits[__i]; \ } while (0) #define __BIT_AND2(_s, d, s1, s2) do { \ __size_t __i; \ for (__i = 0; __i < __bitset_words((_s)); __i++) \ (d)->__bits[__i] = (s1)->__bits[__i] & (s2)->__bits[__i];\ } while (0) #define __BIT_ANDNOT(_s, d, s) do { \ __size_t __i; \ for (__i = 0; __i < __bitset_words((_s)); __i++) \ (d)->__bits[__i] &= ~(s)->__bits[__i]; \ } while (0) #define __BIT_ANDNOT2(_s, d, s1, s2) do { \ __size_t __i; \ for (__i = 0; __i < __bitset_words((_s)); __i++) \ (d)->__bits[__i] = (s1)->__bits[__i] & ~(s2)->__bits[__i];\ } while (0) #define __BIT_XOR(_s, d, s) do { \ __size_t __i; \ for (__i = 0; __i < __bitset_words((_s)); __i++) \ (d)->__bits[__i] ^= (s)->__bits[__i]; \ } while (0) #define __BIT_XOR2(_s, d, s1, s2) do { \ __size_t __i; \ for (__i = 0; __i < __bitset_words((_s)); __i++) \ (d)->__bits[__i] = (s1)->__bits[__i] ^ (s2)->__bits[__i];\ } while (0) /* * Note, the atomic(9) API is not consistent between clear/set and * testandclear/testandset in whether the value argument is a mask * or a bit index. */ #define __BIT_CLR_ATOMIC(_s, n, p) \ atomic_clear_long(&(p)->__bits[__bitset_word(_s, n)], \ __bitset_mask((_s), n)) #define __BIT_SET_ATOMIC(_s, n, p) \ atomic_set_long(&(p)->__bits[__bitset_word(_s, n)], \ __bitset_mask((_s), n)) #define __BIT_SET_ATOMIC_ACQ(_s, n, p) \ atomic_set_acq_long(&(p)->__bits[__bitset_word(_s, n)], \ __bitset_mask((_s), n)) #define __BIT_TEST_CLR_ATOMIC(_s, n, p) \ (atomic_testandclear_long( \ &(p)->__bits[__bitset_word((_s), (n))], (n)) != 0) #define __BIT_TEST_SET_ATOMIC(_s, n, p) \ (atomic_testandset_long( \ &(p)->__bits[__bitset_word((_s), (n))], (n)) != 0) /* Convenience functions catering special cases. */ #define __BIT_AND_ATOMIC(_s, d, s) do { \ __size_t __i; \ for (__i = 0; __i < __bitset_words((_s)); __i++) \ atomic_clear_long(&(d)->__bits[__i], \ ~(s)->__bits[__i]); \ } while (0) #define __BIT_OR_ATOMIC(_s, d, s) do { \ __size_t __i; \ for (__i = 0; __i < __bitset_words((_s)); __i++) \ atomic_set_long(&(d)->__bits[__i], \ (s)->__bits[__i]); \ } while (0) #define __BIT_COPY_STORE_REL(_s, f, t) do { \ __size_t __i; \ for (__i = 0; __i < __bitset_words((_s)); __i++) \ atomic_store_rel_long(&(t)->__bits[__i], \ (f)->__bits[__i]); \ } while (0) /* * Note that `start` and the returned value from __BIT_FFS_AT are * 1-based bit indices. */ #define __BIT_FFS_AT(_s, p, start) __extension__ ({ \ __size_t __i; \ long __bit, __mask; \ \ __mask = ~0UL << ((start) % _BITSET_BITS); \ __bit = 0; \ for (__i = __bitset_word((_s), (start)); \ __i < __bitset_words((_s)); \ __i++) { \ if (((p)->__bits[__i] & __mask) != 0) { \ __bit = ffsl((p)->__bits[__i] & __mask); \ __bit += __i * _BITSET_BITS; \ break; \ } \ __mask = ~0UL; \ } \ __bit; \ }) #define __BIT_FFS(_s, p) __BIT_FFS_AT((_s), (p), 0) #define __BIT_FLS(_s, p) __extension__ ({ \ __size_t __i; \ long __bit; \ \ __bit = 0; \ for (__i = __bitset_words((_s)); __i > 0; __i--) { \ if ((p)->__bits[__i - 1] != 0) { \ __bit = flsl((p)->__bits[__i - 1]); \ __bit += (__i - 1) * _BITSET_BITS; \ break; \ } \ } \ __bit; \ }) #define __BIT_COUNT(_s, p) __extension__ ({ \ __size_t __i; \ long __count; \ \ __count = 0; \ for (__i = 0; __i < __bitset_words((_s)); __i++) \ __count += __bitcountl((p)->__bits[__i]); \ __count; \ }) #define __BIT_FOREACH_ADVANCE(_s, i, p, op) __extension__ ({ \ int __found; \ for (;;) { \ if (__bits != 0) { \ int __bit = ffsl(__bits) - 1; \ __bits &= ~(1ul << __bit); \ (i) = __i * _BITSET_BITS + __bit; \ __found = 1; \ break; \ } \ if (++__i == __bitset_words(_s)) { \ __found = 0; \ break; \ } \ __bits = op((p)->__bits[__i]); \ } \ __found != 0; \ }) /* * Non-destructively loop over all set or clear bits in the set. */ #define __BIT_FOREACH(_s, i, p, op) \ for (long __i = -1, __bits = 0; \ __BIT_FOREACH_ADVANCE(_s, i, p, op); ) #define __BIT_FOREACH_ISSET(_s, i, p) __BIT_FOREACH(_s, i, p, ) #define __BIT_FOREACH_ISCLR(_s, i, p) __BIT_FOREACH(_s, i, p, ~) #define __BITSET_T_INITIALIZER(x) \ { .__bits = { x } } #define __BITSET_FSET(n) \ [ 0 ... ((n) - 1) ] = (-1L) #define __BITSET_SIZE(_s) (__bitset_words((_s)) * sizeof(long)) #if defined(_KERNEL) || defined(_WANT_FREEBSD_BITSET) /* * Dynamically allocate a bitset. */ -#define __BITSET_ALLOC(_s, mt, mf) malloc(__BITSET_SIZE((_s)), mt, (mf)) - #define BIT_AND(_s, d, s) __BIT_AND(_s, d, s) #define BIT_AND2(_s, d, s1, s2) __BIT_AND2(_s, d, s1, s2) #define BIT_ANDNOT(_s, d, s) __BIT_ANDNOT(_s, d, s) #define BIT_ANDNOT2(_s, d, s1, s2) __BIT_ANDNOT2(_s, d, s1, s2) #define BIT_AND_ATOMIC(_s, d, s) __BIT_AND_ATOMIC(_s, d, s) #define BIT_CLR(_s, n, p) __BIT_CLR(_s, n, p) #define BIT_CLR_ATOMIC(_s, n, p) __BIT_CLR_ATOMIC(_s, n, p) #define BIT_CMP(_s, p, c) __BIT_CMP(_s, p, c) #define BIT_COPY(_s, f, t) __BIT_COPY(_s, f, t) #define BIT_COPY_STORE_REL(_s, f, t) __BIT_COPY_STORE_REL(_s, f, t) #define BIT_COUNT(_s, p) __BIT_COUNT(_s, p) #define BIT_EMPTY(_s, p) __BIT_EMPTY(_s, p) #define BIT_FFS(_s, p) __BIT_FFS(_s, p) #define BIT_FFS_AT(_s, p, start) __BIT_FFS_AT(_s, p, start) #define BIT_FILL(_s, p) __BIT_FILL(_s, p) #define BIT_FLS(_s, p) __BIT_FLS(_s, p) #define BIT_FOREACH(_s, i, p, op) __BIT_FOREACH(_s, i, p, op) #define BIT_FOREACH_ADVANCE(_s, i, p, op) __BIT_FOREACH_ADVANCE(_s, i, p, op) #define BIT_FOREACH_ISCLR(_s, i, p) __BIT_FOREACH_ISCLR(_s, i, p) #define BIT_FOREACH_ISSET(_s, i, p) __BIT_FOREACH_ISSET(_s, i, p) #define BIT_ISFULLSET(_s, p) __BIT_ISFULLSET(_s, p) #define BIT_ISSET(_s, n, p) __BIT_ISSET(_s, n, p) #define BIT_OR(_s, d, s) __BIT_OR(_s, d, s) #define BIT_OR2(_s, d, s1, s2) __BIT_OR2(_s, d, s1, s2) #define BIT_OR_ATOMIC(_s, d, s) __BIT_OR_ATOMIC(_s, d, s) #define BIT_OVERLAP(_s, p, c) __BIT_OVERLAP(_s, p, c) #define BIT_SET(_s, n, p) __BIT_SET(_s, n, p) #define BIT_SETOF(_s, n, p) __BIT_SETOF(_s, n, p) #define BIT_SET_ATOMIC(_s, n, p) __BIT_SET_ATOMIC(_s, n, p) #define BIT_SET_ATOMIC_ACQ(_s, n, p) __BIT_SET_ATOMIC_ACQ(_s, n, p) #define BIT_SUBSET(_s, p, c) __BIT_SUBSET(_s, p, c) #define BIT_TEST_CLR_ATOMIC(_s, n, p) __BIT_TEST_CLR_ATOMIC(_s, n, p) #define BIT_TEST_SET_ATOMIC(_s, n, p) __BIT_TEST_SET_ATOMIC(_s, n, p) #define BIT_XOR(_s, d, s) __BIT_XOR(_s, d, s) #define BIT_XOR2(_s, d, s1, s2) __BIT_XOR2(_s, d, s1, s2) #define BIT_ZERO(_s, p) __BIT_ZERO(_s, p) -#define BITSET_ALLOC(_s, mt, mf) __BITSET_ALLOC(_s, mt, mf) +#if defined(_KERNEL) +#define BITSET_ALLOC(_s, mt, mf) malloc(__BITSET_SIZE((_s)), mt, (mf)) +#define BITSET_FREE(p, mt) free(p, mt) +#endif /* _KERNEL */ + #define BITSET_FSET(n) __BITSET_FSET(n) #define BITSET_SIZE(_s) __BITSET_SIZE(_s) #define BITSET_T_INITIALIZER(x) __BITSET_T_INITIALIZER(x) #endif /* defined(_KERNEL) || defined(_WANT_FREEBSD_BITSET) */ #endif /* !_SYS_BITSET_H_ */ diff --git a/sys/sys/cpuset.h b/sys/sys/cpuset.h index 71317670c593..9f777e3af523 100644 --- a/sys/sys/cpuset.h +++ b/sys/sys/cpuset.h @@ -1,163 +1,179 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2008, Jeffrey Roberson * All rights reserved. * * Copyright (c) 2008 Nokia Corporation * All rights reserved. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice unmodified, this list of conditions, and the following * disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. * * $FreeBSD$ */ #ifndef _SYS_CPUSET_H_ #define _SYS_CPUSET_H_ #include +#include #include #define _NCPUBITS _BITSET_BITS #define _NCPUWORDS __bitset_words(CPU_SETSIZE) #define CPUSETBUFSIZ ((2 + sizeof(long) * 2) * _NCPUWORDS) #define CPU_CLR(n, p) __BIT_CLR(CPU_SETSIZE, n, p) #define CPU_COPY(f, t) __BIT_COPY(CPU_SETSIZE, f, t) #define CPU_ISSET(n, p) __BIT_ISSET(CPU_SETSIZE, n, p) #define CPU_SET(n, p) __BIT_SET(CPU_SETSIZE, n, p) #define CPU_ZERO(p) __BIT_ZERO(CPU_SETSIZE, p) #define CPU_FILL(p) __BIT_FILL(CPU_SETSIZE, p) #define CPU_SETOF(n, p) __BIT_SETOF(CPU_SETSIZE, n, p) #define CPU_EQUAL(p, c) (__BIT_CMP(CPU_SETSIZE, p, c) == 0) #define CPU_EMPTY(p) __BIT_EMPTY(CPU_SETSIZE, p) #define CPU_ISFULLSET(p) __BIT_ISFULLSET(CPU_SETSIZE, p) #define CPU_SUBSET(p, c) __BIT_SUBSET(CPU_SETSIZE, p, c) #define CPU_OVERLAP(p, c) __BIT_OVERLAP(CPU_SETSIZE, p, c) #define CPU_CMP(p, c) __BIT_CMP(CPU_SETSIZE, p, c) -#define CPU_OR(d, s) __BIT_OR(CPU_SETSIZE, d, s) -#define CPU_AND(d, s) __BIT_AND(CPU_SETSIZE, d, s) -#define CPU_ANDNOT(d, s) __BIT_ANDNOT(CPU_SETSIZE, d, s) +#define CPU_OR(d, s1, s2) __BIT_OR2(CPU_SETSIZE, d, s1, s2) +#define CPU_AND(d, s1, s2) __BIT_AND2(CPU_SETSIZE, d, s1, s2) +#define CPU_ANDNOT(d, s1, s2) __BIT_ANDNOT2(CPU_SETSIZE, d, s1, s2) +#define CPU_XOR(d, s1, s2) __BIT_XOR2(CPU_SETSIZE, d, s1, s2) #define CPU_CLR_ATOMIC(n, p) __BIT_CLR_ATOMIC(CPU_SETSIZE, n, p) #define CPU_SET_ATOMIC(n, p) __BIT_SET_ATOMIC(CPU_SETSIZE, n, p) #define CPU_SET_ATOMIC_ACQ(n, p) __BIT_SET_ATOMIC_ACQ(CPU_SETSIZE, n, p) #define CPU_AND_ATOMIC(n, p) __BIT_AND_ATOMIC(CPU_SETSIZE, n, p) #define CPU_OR_ATOMIC(d, s) __BIT_OR_ATOMIC(CPU_SETSIZE, d, s) #define CPU_COPY_STORE_REL(f, t) __BIT_COPY_STORE_REL(CPU_SETSIZE, f, t) #define CPU_FFS(p) __BIT_FFS(CPU_SETSIZE, p) #define CPU_FLS(p) __BIT_FLS(CPU_SETSIZE, p) #define CPU_FOREACH_ISSET(i, p) __BIT_FOREACH_ISSET(CPU_SETSIZE, i, p) #define CPU_FOREACH_ISCLR(i, p) __BIT_FOREACH_ISCLR(CPU_SETSIZE, i, p) #define CPU_COUNT(p) ((int)__BIT_COUNT(CPU_SETSIZE, p)) #define CPUSET_FSET __BITSET_FSET(_NCPUWORDS) #define CPUSET_T_INITIALIZER(x) __BITSET_T_INITIALIZER(x) +#if !defined(_KERNEL) +#define CPU_ALLOC_SIZE(_s) __BITSET_SIZE(_s) +#define CPU_ALLOC(_s) __cpuset_alloc(_s) +#define CPU_FREE(p) __cpuset_free(p) + +#define CPU_ISSET_S(n, _s, p) __BIT_ISSET(_s, n, p) +#define CPU_SET_S(n, _s, p) __BIT_SET(_s, n, p) +#define CPU_ZERO_S(_s, p) __BIT_ZERO(_s, p) + +#define CPU_OR_S(_s, d, s1, s2) __BIT_OR2(_s, d, s1, s2) +#define CPU_AND_S(_s, d, s1, s2) __BIT_AND2(_s, d, s1, s2) +#define CPU_XOR_S(_s, d, s1, s2) __BIT_XOR2(_s, d, s1, s2) +#endif + /* * Valid cpulevel_t values. */ #define CPU_LEVEL_ROOT 1 /* All system cpus. */ #define CPU_LEVEL_CPUSET 2 /* Available cpus for which. */ #define CPU_LEVEL_WHICH 3 /* Actual mask/id for which. */ /* * Valid cpuwhich_t values. */ #define CPU_WHICH_TID 1 /* Specifies a thread id. */ #define CPU_WHICH_PID 2 /* Specifies a process id. */ #define CPU_WHICH_CPUSET 3 /* Specifies a set id. */ #define CPU_WHICH_IRQ 4 /* Specifies an irq #. */ #define CPU_WHICH_JAIL 5 /* Specifies a jail id. */ #define CPU_WHICH_DOMAIN 6 /* Specifies a NUMA domain id. */ #define CPU_WHICH_INTRHANDLER 7 /* Specifies an irq # (not ithread). */ #define CPU_WHICH_ITHREAD 8 /* Specifies an irq's ithread. */ /* * Reserved cpuset identifiers. */ #define CPUSET_INVALID -1 #define CPUSET_DEFAULT 0 #ifdef _KERNEL #include LIST_HEAD(setlist, cpuset); /* * cpusets encapsulate cpu binding information for one or more threads. * * a - Accessed with atomics. * s - Set at creation, never modified. Only a ref required to read. * c - Locked internally by a cpuset lock. * * The bitmask is only modified while holding the cpuset lock. It may be * read while only a reference is held but the consumer must be prepared * to deal with inconsistent results. */ struct cpuset { volatile u_int cs_ref; /* (a) Reference count. */ int cs_flags; /* (s) Flags from below. */ LIST_ENTRY(cpuset) cs_link; /* (c) All identified sets. */ LIST_ENTRY(cpuset) cs_siblings; /* (c) Sibling set link. */ struct setlist cs_children; /* (c) List of children. */ struct domainset *cs_domain; /* (c) NUMA policy. */ cpusetid_t cs_id; /* (s) Id or INVALID. */ struct cpuset *cs_parent; /* (s) Pointer to our parent. */ cpuset_t cs_mask; /* bitmask of valid cpus. */ }; #define CPU_SET_ROOT 0x0001 /* Set is a root set. */ #define CPU_SET_RDONLY 0x0002 /* No modification allowed. */ extern cpuset_t *cpuset_root; struct prison; struct proc; struct thread; struct cpuset *cpuset_thread0(void); struct cpuset *cpuset_ref(struct cpuset *); void cpuset_rel(struct cpuset *); int cpuset_setthread(lwpid_t id, cpuset_t *); int cpuset_setithread(lwpid_t id, int cpu); int cpuset_create_root(struct prison *, struct cpuset **); int cpuset_setproc_update_set(struct proc *, struct cpuset *); int cpuset_which(cpuwhich_t, id_t, struct proc **, struct thread **, struct cpuset **); void cpuset_kernthread(struct thread *); char *cpusetobj_strprint(char *, const cpuset_t *); int cpusetobj_strscan(cpuset_t *, const char *); #ifdef DDB void ddb_display_cpuset(const cpuset_t *); #endif #else __BEGIN_DECLS int cpuset(cpusetid_t *); int cpuset_setid(cpuwhich_t, id_t, cpusetid_t); int cpuset_getid(cpulevel_t, cpuwhich_t, id_t, cpusetid_t *); int cpuset_getaffinity(cpulevel_t, cpuwhich_t, id_t, size_t, cpuset_t *); int cpuset_setaffinity(cpulevel_t, cpuwhich_t, id_t, size_t, const cpuset_t *); __END_DECLS #endif #endif /* !_SYS_CPUSET_H_ */ diff --git a/sys/sys/param.h b/sys/sys/param.h index 679f127775b5..f40041e33963 100644 --- a/sys/sys/param.h +++ b/sys/sys/param.h @@ -1,374 +1,374 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1982, 1986, 1989, 1993 * The Regents of the University of California. All rights reserved. * (c) UNIX System Laboratories, Inc. * All or some portions of this file are derived from material licensed * to the University of California by American Telephone and Telegraph * Co. or Unix System Laboratories, Inc. and are reproduced herein with * the permission of UNIX System Laboratories, Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)param.h 8.3 (Berkeley) 4/4/95 * $FreeBSD$ */ #ifndef _SYS_PARAM_H_ #define _SYS_PARAM_H_ #include #define BSD 199506 /* System version (year & month). */ #define BSD4_3 1 #define BSD4_4 1 /* * __FreeBSD_version numbers are documented in the Porter's Handbook. * If you bump the version for any reason, you should update the documentation * there. * Currently this lives here in the doc/ repository: * * documentation/content/en/books/porters-handbook/versions/_index.adoc * * scheme is: Rxx * 'R' is in the range 0 to 4 if this is a release branch or * X.0-CURRENT before releng/X.0 is created, otherwise 'R' is * in the range 5 to 9. */ #undef __FreeBSD_version -#define __FreeBSD_version 1300523 /* Master, propagated to newvers */ +#define __FreeBSD_version 1300524 /* Master, propagated to newvers */ /* * __FreeBSD_kernel__ indicates that this system uses the kernel of FreeBSD, * which by definition is always true on FreeBSD. This macro is also defined * on other systems that use the kernel of FreeBSD, such as GNU/kFreeBSD. * * It is tempting to use this macro in userland code when we want to enable * kernel-specific routines, and in fact it's fine to do this in code that * is part of FreeBSD itself. However, be aware that as presence of this * macro is still not widespread (e.g. older FreeBSD versions, 3rd party * compilers, etc), it is STRONGLY DISCOURAGED to check for this macro in * external applications without also checking for __FreeBSD__ as an * alternative. */ #undef __FreeBSD_kernel__ #define __FreeBSD_kernel__ #if defined(_KERNEL) || defined(IN_RTLD) #define P_OSREL_SIGWAIT 700000 #define P_OSREL_SIGSEGV 700004 #define P_OSREL_MAP_ANON 800104 #define P_OSREL_MAP_FSTRICT 1100036 #define P_OSREL_SHUTDOWN_ENOTCONN 1100077 #define P_OSREL_MAP_GUARD 1200035 #define P_OSREL_WRFSBASE 1200041 #define P_OSREL_CK_CYLGRP 1200046 #define P_OSREL_VMTOTAL64 1200054 #define P_OSREL_CK_SUPERBLOCK 1300000 #define P_OSREL_CK_INODE 1300005 #define P_OSREL_POWERPC_NEW_AUX_ARGS 1300070 #define P_OSREL_MAJOR(x) ((x) / 100000) #endif #ifndef LOCORE #include #endif /* * Machine-independent constants (some used in following include files). * Redefined constants are from POSIX 1003.1 limits file. * * MAXCOMLEN should be >= sizeof(ac_comm) (see ) */ #include #define MAXCOMLEN 19 /* max command name remembered */ #define MAXINTERP PATH_MAX /* max interpreter file name length */ #define MAXLOGNAME 33 /* max login name length (incl. NUL) */ #define MAXUPRC CHILD_MAX /* max simultaneous processes */ #define NCARGS ARG_MAX /* max bytes for an exec function */ #define NGROUPS (NGROUPS_MAX+1) /* max number groups */ #define NOFILE OPEN_MAX /* max open files per process */ #define NOGROUP 65535 /* marker for empty group set member */ #define MAXHOSTNAMELEN 256 /* max hostname size */ #define SPECNAMELEN 255 /* max length of devicename */ /* More types and definitions used throughout the kernel. */ #ifdef _KERNEL #include #include #ifndef LOCORE #include #include #endif #ifndef FALSE #define FALSE 0 #endif #ifndef TRUE #define TRUE 1 #endif #endif #ifndef _KERNEL #ifndef LOCORE /* Signals. */ #include #endif #endif /* Machine type dependent parameters. */ #include #ifndef _KERNEL #include #endif #ifndef DEV_BSHIFT #define DEV_BSHIFT 9 /* log2(DEV_BSIZE) */ #endif #define DEV_BSIZE (1<>PAGE_SHIFT) #endif /* * btodb() is messy and perhaps slow because `bytes' may be an off_t. We * want to shift an unsigned type to avoid sign extension and we don't * want to widen `bytes' unnecessarily. Assume that the result fits in * a daddr_t. */ #ifndef btodb #define btodb(bytes) /* calculates (bytes / DEV_BSIZE) */ \ (sizeof (bytes) > sizeof(long) \ ? (daddr_t)((unsigned long long)(bytes) >> DEV_BSHIFT) \ : (daddr_t)((unsigned long)(bytes) >> DEV_BSHIFT)) #endif #ifndef dbtob #define dbtob(db) /* calculates (db * DEV_BSIZE) */ \ ((off_t)(db) << DEV_BSHIFT) #endif #define PRIMASK 0x0ff #define PCATCH 0x100 /* OR'd with pri for tsleep to check signals */ #define PDROP 0x200 /* OR'd with pri to stop re-entry of interlock mutex */ #define NZERO 0 /* default "nice" */ #define NBBY 8 /* number of bits in a byte */ #define NBPW sizeof(int) /* number of bytes per word (integer) */ #define CMASK 022 /* default file mask: S_IWGRP|S_IWOTH */ #define NODEV (dev_t)(-1) /* non-existent device */ /* * File system parameters and macros. * * MAXBSIZE - Filesystems are made out of blocks of at most MAXBSIZE bytes * per block. MAXBSIZE may be made larger without effecting * any existing filesystems as long as it does not exceed MAXPHYS, * and may be made smaller at the risk of not being able to use * filesystems which require a block size exceeding MAXBSIZE. * * MAXBCACHEBUF - Maximum size of a buffer in the buffer cache. This must * be >= MAXBSIZE and can be set differently for different * architectures by defining it in . * Making this larger allows NFS to do larger reads/writes. * * BKVASIZE - Nominal buffer space per buffer, in bytes. BKVASIZE is the * minimum KVM memory reservation the kernel is willing to make. * Filesystems can of course request smaller chunks. Actual * backing memory uses a chunk size of a page (PAGE_SIZE). * The default value here can be overridden on a per-architecture * basis by defining it in . * * If you make BKVASIZE too small you risk seriously fragmenting * the buffer KVM map which may slow things down a bit. If you * make it too big the kernel will not be able to optimally use * the KVM memory reserved for the buffer cache and will wind * up with too-few buffers. * * The default is 16384, roughly 2x the block size used by a * normal UFS filesystem. */ #define MAXBSIZE 65536 /* must be power of 2 */ #ifndef MAXBCACHEBUF #define MAXBCACHEBUF MAXBSIZE /* must be a power of 2 >= MAXBSIZE */ #endif #ifndef BKVASIZE #define BKVASIZE 16384 /* must be power of 2 */ #endif #define BKVAMASK (BKVASIZE-1) /* * MAXPATHLEN defines the longest permissible path length after expanding * symbolic links. It is used to allocate a temporary buffer from the buffer * pool in which to do the name expansion, hence should be a power of two, * and must be less than or equal to MAXBSIZE. MAXSYMLINKS defines the * maximum number of symbolic links that may be expanded in a path name. * It should be set high enough to allow all legitimate uses, but halt * infinite loops reasonably quickly. */ #define MAXPATHLEN PATH_MAX #define MAXSYMLINKS 32 /* Bit map related macros. */ #define setbit(a,i) (((unsigned char *)(a))[(i)/NBBY] |= 1<<((i)%NBBY)) #define clrbit(a,i) (((unsigned char *)(a))[(i)/NBBY] &= ~(1<<((i)%NBBY))) #define isset(a,i) \ (((const unsigned char *)(a))[(i)/NBBY] & (1<<((i)%NBBY))) #define isclr(a,i) \ ((((const unsigned char *)(a))[(i)/NBBY] & (1<<((i)%NBBY))) == 0) /* Macros for counting and rounding. */ #ifndef howmany #define howmany(x, y) (((x)+((y)-1))/(y)) #endif #define nitems(x) (sizeof((x)) / sizeof((x)[0])) #define rounddown(x, y) (((x)/(y))*(y)) #define rounddown2(x, y) __align_down(x, y) /* if y is power of two */ #define roundup(x, y) ((((x)+((y)-1))/(y))*(y)) /* to any y */ #define roundup2(x, y) __align_up(x, y) /* if y is powers of two */ #define powerof2(x) ((((x)-1)&(x))==0) /* Macros for min/max. */ #define MIN(a,b) (((a)<(b))?(a):(b)) #define MAX(a,b) (((a)>(b))?(a):(b)) #ifdef _KERNEL /* * Basic byte order function prototypes for non-inline functions. */ #ifndef LOCORE #ifndef _BYTEORDER_PROTOTYPED #define _BYTEORDER_PROTOTYPED __BEGIN_DECLS __uint32_t htonl(__uint32_t); __uint16_t htons(__uint16_t); __uint32_t ntohl(__uint32_t); __uint16_t ntohs(__uint16_t); __END_DECLS #endif #endif #ifndef _BYTEORDER_FUNC_DEFINED #define _BYTEORDER_FUNC_DEFINED #define htonl(x) __htonl(x) #define htons(x) __htons(x) #define ntohl(x) __ntohl(x) #define ntohs(x) __ntohs(x) #endif /* !_BYTEORDER_FUNC_DEFINED */ #endif /* _KERNEL */ /* * Scale factor for scaled integers used to count %cpu time and load avgs. * * The number of CPU `tick's that map to a unique `%age' can be expressed * by the formula (1 / (2 ^ (FSHIFT - 11))). The maximum load average that * can be calculated (assuming 32 bits) can be closely approximated using * the formula (2 ^ (2 * (16 - FSHIFT))) for (FSHIFT < 15). * * For the scheduler to maintain a 1:1 mapping of CPU `tick' to `%age', * FSHIFT must be at least 11; this gives us a maximum load avg of ~1024. */ #define FSHIFT 11 /* bits to right of fixed binary point */ #define FSCALE (1<> (PAGE_SHIFT - DEV_BSHIFT)) #define ctodb(db) /* calculates pages to devblks */ \ ((db) << (PAGE_SHIFT - DEV_BSHIFT)) /* * Old spelling of __containerof(). */ #define member2struct(s, m, x) \ ((struct s *)(void *)((char *)(x) - offsetof(struct s, m))) /* * Access a variable length array that has been declared as a fixed * length array. */ #define __PAST_END(array, offset) (((__typeof__(*(array)) *)(array))[offset]) #endif /* _SYS_PARAM_H_ */ diff --git a/sys/x86/x86/cpu_machdep.c b/sys/x86/x86/cpu_machdep.c index c0011a3bc724..53b32672132a 100644 --- a/sys/x86/x86/cpu_machdep.c +++ b/sys/x86/x86/cpu_machdep.c @@ -1,1522 +1,1522 @@ /*- * Copyright (c) 2003 Peter Wemm. * Copyright (c) 1992 Terrence R. Lambert. * Copyright (c) 1982, 1987, 1990 The Regents of the University of California. * All rights reserved. * * This code is derived from software contributed to Berkeley by * William Jolitz. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. All advertising materials mentioning features or use of this software * must display the following acknowledgement: * This product includes software developed by the University of * California, Berkeley and its contributors. * 4. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * from: @(#)machdep.c 7.4 (Berkeley) 6/3/91 */ #include __FBSDID("$FreeBSD$"); #include "opt_acpi.h" #include "opt_atpic.h" #include "opt_cpu.h" #include "opt_ddb.h" #include "opt_inet.h" #include "opt_isa.h" #include "opt_kdb.h" #include "opt_kstack_pages.h" #include "opt_maxmem.h" #include "opt_mp_watchdog.h" #include "opt_platform.h" #ifdef __i386__ #include "opt_apic.h" #endif #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef SMP #include #endif #ifdef CPU_ELAN #include #endif #include #include #include #include #include #include #include #include #include #include #include #include #define STATE_RUNNING 0x0 #define STATE_MWAIT 0x1 #define STATE_SLEEPING 0x2 #ifdef SMP static u_int cpu_reset_proxyid; static volatile u_int cpu_reset_proxy_active; #endif struct msr_op_arg { u_int msr; int op; uint64_t arg1; uint64_t *res; }; static void x86_msr_op_one(void *argp) { struct msr_op_arg *a; uint64_t v; a = argp; switch (a->op) { case MSR_OP_ANDNOT: v = rdmsr(a->msr); v &= ~a->arg1; wrmsr(a->msr, v); break; case MSR_OP_OR: v = rdmsr(a->msr); v |= a->arg1; wrmsr(a->msr, v); break; case MSR_OP_WRITE: wrmsr(a->msr, a->arg1); break; case MSR_OP_READ: v = rdmsr(a->msr); *a->res = v; break; } } #define MSR_OP_EXMODE_MASK 0xf0000000 #define MSR_OP_OP_MASK 0x000000ff #define MSR_OP_GET_CPUID(x) (((x) & ~MSR_OP_EXMODE_MASK) >> 8) void x86_msr_op(u_int msr, u_int op, uint64_t arg1, uint64_t *res) { struct thread *td; struct msr_op_arg a; cpuset_t set; u_int exmode; int bound_cpu, cpu, i, is_bound; a.op = op & MSR_OP_OP_MASK; MPASS(a.op == MSR_OP_ANDNOT || a.op == MSR_OP_OR || a.op == MSR_OP_WRITE || a.op == MSR_OP_READ); exmode = op & MSR_OP_EXMODE_MASK; MPASS(exmode == MSR_OP_LOCAL || exmode == MSR_OP_SCHED_ALL || exmode == MSR_OP_SCHED_ONE || exmode == MSR_OP_RENDEZVOUS_ALL || exmode == MSR_OP_RENDEZVOUS_ONE); a.msr = msr; a.arg1 = arg1; a.res = res; switch (exmode) { case MSR_OP_LOCAL: x86_msr_op_one(&a); break; case MSR_OP_SCHED_ALL: td = curthread; thread_lock(td); is_bound = sched_is_bound(td); bound_cpu = td->td_oncpu; CPU_FOREACH(i) { sched_bind(td, i); x86_msr_op_one(&a); } if (is_bound) sched_bind(td, bound_cpu); else sched_unbind(td); thread_unlock(td); break; case MSR_OP_SCHED_ONE: td = curthread; cpu = MSR_OP_GET_CPUID(op); thread_lock(td); is_bound = sched_is_bound(td); bound_cpu = td->td_oncpu; if (!is_bound || bound_cpu != cpu) sched_bind(td, cpu); x86_msr_op_one(&a); if (is_bound) { if (bound_cpu != cpu) sched_bind(td, bound_cpu); } else { sched_unbind(td); } thread_unlock(td); break; case MSR_OP_RENDEZVOUS_ALL: smp_rendezvous(smp_no_rendezvous_barrier, x86_msr_op_one, smp_no_rendezvous_barrier, &a); break; case MSR_OP_RENDEZVOUS_ONE: cpu = MSR_OP_GET_CPUID(op); CPU_SETOF(cpu, &set); smp_rendezvous_cpus(set, smp_no_rendezvous_barrier, x86_msr_op_one, smp_no_rendezvous_barrier, &a); break; } } /* * Automatically initialized per CPU errata in cpu_idle_tun below. */ bool mwait_cpustop_broken = false; SYSCTL_BOOL(_machdep, OID_AUTO, mwait_cpustop_broken, CTLFLAG_RDTUN, &mwait_cpustop_broken, 0, "Can not reliably wake MONITOR/MWAIT cpus without interrupts"); /* * Flush the D-cache for non-DMA I/O so that the I-cache can * be made coherent later. */ void cpu_flush_dcache(void *ptr, size_t len) { /* Not applicable */ } void acpi_cpu_c1(void) { __asm __volatile("sti; hlt"); } /* * Use mwait to pause execution while waiting for an interrupt or * another thread to signal that there is more work. * * NOTE: Interrupts will cause a wakeup; however, this function does * not enable interrupt handling. The caller is responsible to enable * interrupts. */ void acpi_cpu_idle_mwait(uint32_t mwait_hint) { int *state; uint64_t v; /* * A comment in Linux patch claims that 'CPUs run faster with * speculation protection disabled. All CPU threads in a core * must disable speculation protection for it to be * disabled. Disable it while we are idle so the other * hyperthread can run fast.' * * XXXKIB. Software coordination mode should be supported, * but all Intel CPUs provide hardware coordination. */ state = &PCPU_PTR(monitorbuf)->idle_state; KASSERT(atomic_load_int(state) == STATE_SLEEPING, ("cpu_mwait_cx: wrong monitorbuf state")); atomic_store_int(state, STATE_MWAIT); if (PCPU_GET(ibpb_set) || hw_ssb_active) { v = rdmsr(MSR_IA32_SPEC_CTRL); wrmsr(MSR_IA32_SPEC_CTRL, v & ~(IA32_SPEC_CTRL_IBRS | IA32_SPEC_CTRL_STIBP | IA32_SPEC_CTRL_SSBD)); } else { v = 0; } cpu_monitor(state, 0, 0); if (atomic_load_int(state) == STATE_MWAIT) cpu_mwait(MWAIT_INTRBREAK, mwait_hint); /* * SSB cannot be disabled while we sleep, or rather, if it was * disabled, the sysctl thread will bind to our cpu to tweak * MSR. */ if (v != 0) wrmsr(MSR_IA32_SPEC_CTRL, v); /* * We should exit on any event that interrupts mwait, because * that event might be a wanted interrupt. */ atomic_store_int(state, STATE_RUNNING); } /* Get current clock frequency for the given cpu id. */ int cpu_est_clockrate(int cpu_id, uint64_t *rate) { uint64_t tsc1, tsc2; uint64_t acnt, mcnt, perf; register_t reg; if (pcpu_find(cpu_id) == NULL || rate == NULL) return (EINVAL); #ifdef __i386__ if ((cpu_feature & CPUID_TSC) == 0) return (EOPNOTSUPP); #endif /* * If TSC is P-state invariant and APERF/MPERF MSRs do not exist, * DELAY(9) based logic fails. */ if (tsc_is_invariant && !tsc_perf_stat) return (EOPNOTSUPP); #ifdef SMP if (smp_cpus > 1) { /* Schedule ourselves on the indicated cpu. */ thread_lock(curthread); sched_bind(curthread, cpu_id); thread_unlock(curthread); } #endif /* Calibrate by measuring a short delay. */ reg = intr_disable(); if (tsc_is_invariant) { wrmsr(MSR_MPERF, 0); wrmsr(MSR_APERF, 0); tsc1 = rdtsc(); DELAY(1000); mcnt = rdmsr(MSR_MPERF); acnt = rdmsr(MSR_APERF); tsc2 = rdtsc(); intr_restore(reg); perf = 1000 * acnt / mcnt; *rate = (tsc2 - tsc1) * perf; } else { tsc1 = rdtsc(); DELAY(1000); tsc2 = rdtsc(); intr_restore(reg); *rate = (tsc2 - tsc1) * 1000; } #ifdef SMP if (smp_cpus > 1) { thread_lock(curthread); sched_unbind(curthread); thread_unlock(curthread); } #endif return (0); } /* * Shutdown the CPU as much as possible */ void cpu_halt(void) { for (;;) halt(); } static void cpu_reset_real(void) { struct region_descriptor null_idt; int b; disable_intr(); #ifdef CPU_ELAN if (elan_mmcr != NULL) elan_mmcr->RESCFG = 1; #endif #ifdef __i386__ if (cpu == CPU_GEODE1100) { /* Attempt Geode's own reset */ outl(0xcf8, 0x80009044ul); outl(0xcfc, 0xf); } #endif #if !defined(BROKEN_KEYBOARD_RESET) /* * Attempt to do a CPU reset via the keyboard controller, * do not turn off GateA20, as any machine that fails * to do the reset here would then end up in no man's land. */ outb(IO_KBD + 4, 0xFE); DELAY(500000); /* wait 0.5 sec to see if that did it */ #endif /* * Attempt to force a reset via the Reset Control register at * I/O port 0xcf9. Bit 2 forces a system reset when it * transitions from 0 to 1. Bit 1 selects the type of reset * to attempt: 0 selects a "soft" reset, and 1 selects a * "hard" reset. We try a "hard" reset. The first write sets * bit 1 to select a "hard" reset and clears bit 2. The * second write forces a 0 -> 1 transition in bit 2 to trigger * a reset. */ outb(0xcf9, 0x2); outb(0xcf9, 0x6); DELAY(500000); /* wait 0.5 sec to see if that did it */ /* * Attempt to force a reset via the Fast A20 and Init register * at I/O port 0x92. Bit 1 serves as an alternate A20 gate. * Bit 0 asserts INIT# when set to 1. We are careful to only * preserve bit 1 while setting bit 0. We also must clear bit * 0 before setting it if it isn't already clear. */ b = inb(0x92); if (b != 0xff) { if ((b & 0x1) != 0) outb(0x92, b & 0xfe); outb(0x92, b | 0x1); DELAY(500000); /* wait 0.5 sec to see if that did it */ } printf("No known reset method worked, attempting CPU shutdown\n"); DELAY(1000000); /* wait 1 sec for printf to complete */ /* Wipe the IDT. */ null_idt.rd_limit = 0; null_idt.rd_base = 0; lidt(&null_idt); /* "good night, sweet prince .... " */ breakpoint(); /* NOTREACHED */ while(1); } #ifdef SMP static void cpu_reset_proxy(void) { cpu_reset_proxy_active = 1; while (cpu_reset_proxy_active == 1) ia32_pause(); /* Wait for other cpu to see that we've started */ printf("cpu_reset_proxy: Stopped CPU %d\n", cpu_reset_proxyid); DELAY(1000000); cpu_reset_real(); } #endif void cpu_reset(void) { #ifdef SMP struct monitorbuf *mb; cpuset_t map; u_int cnt; if (smp_started) { map = all_cpus; CPU_CLR(PCPU_GET(cpuid), &map); - CPU_ANDNOT(&map, &stopped_cpus); + CPU_ANDNOT(&map, &map, &stopped_cpus); if (!CPU_EMPTY(&map)) { printf("cpu_reset: Stopping other CPUs\n"); stop_cpus(map); } if (PCPU_GET(cpuid) != 0) { cpu_reset_proxyid = PCPU_GET(cpuid); cpustop_restartfunc = cpu_reset_proxy; cpu_reset_proxy_active = 0; printf("cpu_reset: Restarting BSP\n"); /* Restart CPU #0. */ CPU_SETOF(0, &started_cpus); mb = &pcpu_find(0)->pc_monitorbuf; atomic_store_int(&mb->stop_state, MONITOR_STOPSTATE_RUNNING); cnt = 0; while (cpu_reset_proxy_active == 0 && cnt < 10000000) { ia32_pause(); cnt++; /* Wait for BSP to announce restart */ } if (cpu_reset_proxy_active == 0) { printf("cpu_reset: Failed to restart BSP\n"); } else { cpu_reset_proxy_active = 2; while (1) ia32_pause(); /* NOTREACHED */ } } DELAY(1000000); } #endif cpu_reset_real(); /* NOTREACHED */ } bool cpu_mwait_usable(void) { return ((cpu_feature2 & CPUID2_MON) != 0 && ((cpu_mon_mwait_flags & (CPUID5_MON_MWAIT_EXT | CPUID5_MWAIT_INTRBREAK)) == (CPUID5_MON_MWAIT_EXT | CPUID5_MWAIT_INTRBREAK))); } void (*cpu_idle_hook)(sbintime_t) = NULL; /* ACPI idle hook. */ int cpu_amdc1e_bug = 0; /* AMD C1E APIC workaround required. */ static int idle_mwait = 1; /* Use MONITOR/MWAIT for short idle. */ SYSCTL_INT(_machdep, OID_AUTO, idle_mwait, CTLFLAG_RWTUN, &idle_mwait, 0, "Use MONITOR/MWAIT for short idle"); static void cpu_idle_acpi(sbintime_t sbt) { int *state; state = &PCPU_PTR(monitorbuf)->idle_state; atomic_store_int(state, STATE_SLEEPING); /* See comments in cpu_idle_hlt(). */ disable_intr(); if (sched_runnable()) enable_intr(); else if (cpu_idle_hook) cpu_idle_hook(sbt); else acpi_cpu_c1(); atomic_store_int(state, STATE_RUNNING); } static void cpu_idle_hlt(sbintime_t sbt) { int *state; state = &PCPU_PTR(monitorbuf)->idle_state; atomic_store_int(state, STATE_SLEEPING); /* * Since we may be in a critical section from cpu_idle(), if * an interrupt fires during that critical section we may have * a pending preemption. If the CPU halts, then that thread * may not execute until a later interrupt awakens the CPU. * To handle this race, check for a runnable thread after * disabling interrupts and immediately return if one is * found. Also, we must absolutely guarentee that hlt is * the next instruction after sti. This ensures that any * interrupt that fires after the call to disable_intr() will * immediately awaken the CPU from hlt. Finally, please note * that on x86 this works fine because of interrupts enabled only * after the instruction following sti takes place, while IF is set * to 1 immediately, allowing hlt instruction to acknowledge the * interrupt. */ disable_intr(); if (sched_runnable()) enable_intr(); else acpi_cpu_c1(); atomic_store_int(state, STATE_RUNNING); } static void cpu_idle_mwait(sbintime_t sbt) { int *state; state = &PCPU_PTR(monitorbuf)->idle_state; atomic_store_int(state, STATE_MWAIT); /* See comments in cpu_idle_hlt(). */ disable_intr(); if (sched_runnable()) { atomic_store_int(state, STATE_RUNNING); enable_intr(); return; } cpu_monitor(state, 0, 0); if (atomic_load_int(state) == STATE_MWAIT) __asm __volatile("sti; mwait" : : "a" (MWAIT_C1), "c" (0)); else enable_intr(); atomic_store_int(state, STATE_RUNNING); } static void cpu_idle_spin(sbintime_t sbt) { int *state; int i; state = &PCPU_PTR(monitorbuf)->idle_state; atomic_store_int(state, STATE_RUNNING); /* * The sched_runnable() call is racy but as long as there is * a loop missing it one time will have just a little impact if any * (and it is much better than missing the check at all). */ for (i = 0; i < 1000; i++) { if (sched_runnable()) return; cpu_spinwait(); } } void (*cpu_idle_fn)(sbintime_t) = cpu_idle_acpi; void cpu_idle(int busy) { uint64_t msr; sbintime_t sbt = -1; CTR2(KTR_SPARE2, "cpu_idle(%d) at %d", busy, curcpu); #ifdef MP_WATCHDOG ap_watchdog(PCPU_GET(cpuid)); #endif /* If we are busy - try to use fast methods. */ if (busy) { if ((cpu_feature2 & CPUID2_MON) && idle_mwait) { cpu_idle_mwait(busy); goto out; } } /* If we have time - switch timers into idle mode. */ if (!busy) { critical_enter(); sbt = cpu_idleclock(); } /* Apply AMD APIC timer C1E workaround. */ if (cpu_amdc1e_bug && cpu_disable_c3_sleep) { msr = rdmsr(MSR_AMDK8_IPM); if ((msr & (AMDK8_SMIONCMPHALT | AMDK8_C1EONCMPHALT)) != 0) wrmsr(MSR_AMDK8_IPM, msr & ~(AMDK8_SMIONCMPHALT | AMDK8_C1EONCMPHALT)); } /* Call main idle method. */ cpu_idle_fn(sbt); /* Switch timers back into active mode. */ if (!busy) { cpu_activeclock(); critical_exit(); } out: CTR2(KTR_SPARE2, "cpu_idle(%d) at %d done", busy, curcpu); } static int cpu_idle_apl31_workaround; SYSCTL_INT(_machdep, OID_AUTO, idle_apl31, CTLFLAG_RW, &cpu_idle_apl31_workaround, 0, "Apollo Lake APL31 MWAIT bug workaround"); int cpu_idle_wakeup(int cpu) { struct monitorbuf *mb; int *state; mb = &pcpu_find(cpu)->pc_monitorbuf; state = &mb->idle_state; switch (atomic_load_int(state)) { case STATE_SLEEPING: return (0); case STATE_MWAIT: atomic_store_int(state, STATE_RUNNING); return (cpu_idle_apl31_workaround ? 0 : 1); case STATE_RUNNING: return (1); default: panic("bad monitor state"); return (1); } } /* * Ordered by speed/power consumption. */ static struct { void *id_fn; char *id_name; int id_cpuid2_flag; } idle_tbl[] = { { .id_fn = cpu_idle_spin, .id_name = "spin" }, { .id_fn = cpu_idle_mwait, .id_name = "mwait", .id_cpuid2_flag = CPUID2_MON }, { .id_fn = cpu_idle_hlt, .id_name = "hlt" }, { .id_fn = cpu_idle_acpi, .id_name = "acpi" }, }; static int idle_sysctl_available(SYSCTL_HANDLER_ARGS) { char *avail, *p; int error; int i; avail = malloc(256, M_TEMP, M_WAITOK); p = avail; for (i = 0; i < nitems(idle_tbl); i++) { if (idle_tbl[i].id_cpuid2_flag != 0 && (cpu_feature2 & idle_tbl[i].id_cpuid2_flag) == 0) continue; if (strcmp(idle_tbl[i].id_name, "acpi") == 0 && cpu_idle_hook == NULL) continue; p += sprintf(p, "%s%s", p != avail ? ", " : "", idle_tbl[i].id_name); } error = sysctl_handle_string(oidp, avail, 0, req); free(avail, M_TEMP); return (error); } SYSCTL_PROC(_machdep, OID_AUTO, idle_available, CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 0, idle_sysctl_available, "A", "list of available idle functions"); static bool cpu_idle_selector(const char *new_idle_name) { int i; for (i = 0; i < nitems(idle_tbl); i++) { if (idle_tbl[i].id_cpuid2_flag != 0 && (cpu_feature2 & idle_tbl[i].id_cpuid2_flag) == 0) continue; if (strcmp(idle_tbl[i].id_name, "acpi") == 0 && cpu_idle_hook == NULL) continue; if (strcmp(idle_tbl[i].id_name, new_idle_name)) continue; cpu_idle_fn = idle_tbl[i].id_fn; if (bootverbose) printf("CPU idle set to %s\n", idle_tbl[i].id_name); return (true); } return (false); } static int cpu_idle_sysctl(SYSCTL_HANDLER_ARGS) { char buf[16], *p; int error, i; p = "unknown"; for (i = 0; i < nitems(idle_tbl); i++) { if (idle_tbl[i].id_fn == cpu_idle_fn) { p = idle_tbl[i].id_name; break; } } strncpy(buf, p, sizeof(buf)); error = sysctl_handle_string(oidp, buf, sizeof(buf), req); if (error != 0 || req->newptr == NULL) return (error); return (cpu_idle_selector(buf) ? 0 : EINVAL); } SYSCTL_PROC(_machdep, OID_AUTO, idle, CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 0, cpu_idle_sysctl, "A", "currently selected idle function"); static void cpu_idle_tun(void *unused __unused) { char tunvar[16]; if (TUNABLE_STR_FETCH("machdep.idle", tunvar, sizeof(tunvar))) cpu_idle_selector(tunvar); else if (cpu_vendor_id == CPU_VENDOR_AMD && CPUID_TO_FAMILY(cpu_id) == 0x17 && CPUID_TO_MODEL(cpu_id) == 0x1) { /* Ryzen erratas 1057, 1109. */ cpu_idle_selector("hlt"); idle_mwait = 0; mwait_cpustop_broken = true; } if (cpu_vendor_id == CPU_VENDOR_INTEL && cpu_id == 0x506c9) { /* * Apollo Lake errata APL31 (public errata APL30). * Stores to the armed address range may not trigger * MWAIT to resume execution. OS needs to use * interrupts to wake processors from MWAIT-induced * sleep states. */ cpu_idle_apl31_workaround = 1; mwait_cpustop_broken = true; } TUNABLE_INT_FETCH("machdep.idle_apl31", &cpu_idle_apl31_workaround); } SYSINIT(cpu_idle_tun, SI_SUB_CPU, SI_ORDER_MIDDLE, cpu_idle_tun, NULL); static int panic_on_nmi = 0xff; SYSCTL_INT(_machdep, OID_AUTO, panic_on_nmi, CTLFLAG_RWTUN, &panic_on_nmi, 0, "Panic on NMI: 1 = H/W failure; 2 = unknown; 0xff = all"); int nmi_is_broadcast = 1; SYSCTL_INT(_machdep, OID_AUTO, nmi_is_broadcast, CTLFLAG_RWTUN, &nmi_is_broadcast, 0, "Chipset NMI is broadcast"); int (*apei_nmi)(void); void nmi_call_kdb(u_int cpu, u_int type, struct trapframe *frame) { bool claimed = false; #ifdef DEV_ISA /* machine/parity/power fail/"kitchen sink" faults */ if (isa_nmi(frame->tf_err)) { claimed = true; if ((panic_on_nmi & 1) != 0) panic("NMI indicates hardware failure"); } #endif /* DEV_ISA */ /* ACPI Platform Error Interfaces callback. */ if (apei_nmi != NULL && (*apei_nmi)()) claimed = true; /* * NMIs can be useful for debugging. They can be hooked up to a * pushbutton, usually on an ISA, PCI, or PCIe card. They can also be * generated by an IPMI BMC, either manually or in response to a * watchdog timeout. For example, see the "power diag" command in * ports/sysutils/ipmitool. They can also be generated by a * hypervisor; see "bhyvectl --inject-nmi". */ #ifdef KDB if (!claimed && (panic_on_nmi & 2) != 0) { if (debugger_on_panic) { printf("NMI/cpu%d ... going to debugger\n", cpu); claimed = kdb_trap(type, 0, frame); } } #endif /* KDB */ if (!claimed && panic_on_nmi != 0) panic("NMI"); } void nmi_handle_intr(u_int type, struct trapframe *frame) { #ifdef SMP if (nmi_is_broadcast) { nmi_call_kdb_smp(type, frame); return; } #endif nmi_call_kdb(PCPU_GET(cpuid), type, frame); } static int hw_ibrs_active; int hw_ibrs_ibpb_active; int hw_ibrs_disable = 1; SYSCTL_INT(_hw, OID_AUTO, ibrs_active, CTLFLAG_RD, &hw_ibrs_active, 0, "Indirect Branch Restricted Speculation active"); SYSCTL_NODE(_machdep_mitigations, OID_AUTO, ibrs, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Indirect Branch Restricted Speculation active"); SYSCTL_INT(_machdep_mitigations_ibrs, OID_AUTO, active, CTLFLAG_RD, &hw_ibrs_active, 0, "Indirect Branch Restricted Speculation active"); void hw_ibrs_recalculate(bool for_all_cpus) { if ((cpu_ia32_arch_caps & IA32_ARCH_CAP_IBRS_ALL) != 0) { x86_msr_op(MSR_IA32_SPEC_CTRL, (for_all_cpus ? MSR_OP_RENDEZVOUS_ALL : MSR_OP_LOCAL) | (hw_ibrs_disable != 0 ? MSR_OP_ANDNOT : MSR_OP_OR), IA32_SPEC_CTRL_IBRS, NULL); hw_ibrs_active = hw_ibrs_disable == 0; hw_ibrs_ibpb_active = 0; } else { hw_ibrs_active = hw_ibrs_ibpb_active = (cpu_stdext_feature3 & CPUID_STDEXT3_IBPB) != 0 && !hw_ibrs_disable; } } static int hw_ibrs_disable_handler(SYSCTL_HANDLER_ARGS) { int error, val; val = hw_ibrs_disable; error = sysctl_handle_int(oidp, &val, 0, req); if (error != 0 || req->newptr == NULL) return (error); hw_ibrs_disable = val != 0; hw_ibrs_recalculate(true); return (0); } SYSCTL_PROC(_hw, OID_AUTO, ibrs_disable, CTLTYPE_INT | CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0, hw_ibrs_disable_handler, "I", "Disable Indirect Branch Restricted Speculation"); SYSCTL_PROC(_machdep_mitigations_ibrs, OID_AUTO, disable, CTLTYPE_INT | CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0, hw_ibrs_disable_handler, "I", "Disable Indirect Branch Restricted Speculation"); int hw_ssb_active; int hw_ssb_disable; SYSCTL_INT(_hw, OID_AUTO, spec_store_bypass_disable_active, CTLFLAG_RD, &hw_ssb_active, 0, "Speculative Store Bypass Disable active"); SYSCTL_NODE(_machdep_mitigations, OID_AUTO, ssb, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Speculative Store Bypass Disable active"); SYSCTL_INT(_machdep_mitigations_ssb, OID_AUTO, active, CTLFLAG_RD, &hw_ssb_active, 0, "Speculative Store Bypass Disable active"); static void hw_ssb_set(bool enable, bool for_all_cpus) { if ((cpu_stdext_feature3 & CPUID_STDEXT3_SSBD) == 0) { hw_ssb_active = 0; return; } hw_ssb_active = enable; x86_msr_op(MSR_IA32_SPEC_CTRL, (enable ? MSR_OP_OR : MSR_OP_ANDNOT) | (for_all_cpus ? MSR_OP_SCHED_ALL : MSR_OP_LOCAL), IA32_SPEC_CTRL_SSBD, NULL); } void hw_ssb_recalculate(bool all_cpus) { switch (hw_ssb_disable) { default: hw_ssb_disable = 0; /* FALLTHROUGH */ case 0: /* off */ hw_ssb_set(false, all_cpus); break; case 1: /* on */ hw_ssb_set(true, all_cpus); break; case 2: /* auto */ hw_ssb_set((cpu_ia32_arch_caps & IA32_ARCH_CAP_SSB_NO) != 0 ? false : true, all_cpus); break; } } static int hw_ssb_disable_handler(SYSCTL_HANDLER_ARGS) { int error, val; val = hw_ssb_disable; error = sysctl_handle_int(oidp, &val, 0, req); if (error != 0 || req->newptr == NULL) return (error); hw_ssb_disable = val; hw_ssb_recalculate(true); return (0); } SYSCTL_PROC(_hw, OID_AUTO, spec_store_bypass_disable, CTLTYPE_INT | CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0, hw_ssb_disable_handler, "I", "Speculative Store Bypass Disable (0 - off, 1 - on, 2 - auto)"); SYSCTL_PROC(_machdep_mitigations_ssb, OID_AUTO, disable, CTLTYPE_INT | CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0, hw_ssb_disable_handler, "I", "Speculative Store Bypass Disable (0 - off, 1 - on, 2 - auto)"); int hw_mds_disable; /* * Handler for Microarchitectural Data Sampling issues. Really not a * pointer to C function: on amd64 the code must not change any CPU * architectural state except possibly %rflags. Also, it is always * called with interrupts disabled. */ void mds_handler_void(void); void mds_handler_verw(void); void mds_handler_ivb(void); void mds_handler_bdw(void); void mds_handler_skl_sse(void); void mds_handler_skl_avx(void); void mds_handler_skl_avx512(void); void mds_handler_silvermont(void); void (*mds_handler)(void) = mds_handler_void; static int sysctl_hw_mds_disable_state_handler(SYSCTL_HANDLER_ARGS) { const char *state; if (mds_handler == mds_handler_void) state = "inactive"; else if (mds_handler == mds_handler_verw) state = "VERW"; else if (mds_handler == mds_handler_ivb) state = "software IvyBridge"; else if (mds_handler == mds_handler_bdw) state = "software Broadwell"; else if (mds_handler == mds_handler_skl_sse) state = "software Skylake SSE"; else if (mds_handler == mds_handler_skl_avx) state = "software Skylake AVX"; else if (mds_handler == mds_handler_skl_avx512) state = "software Skylake AVX512"; else if (mds_handler == mds_handler_silvermont) state = "software Silvermont"; else state = "unknown"; return (SYSCTL_OUT(req, state, strlen(state))); } SYSCTL_PROC(_hw, OID_AUTO, mds_disable_state, CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0, sysctl_hw_mds_disable_state_handler, "A", "Microarchitectural Data Sampling Mitigation state"); SYSCTL_NODE(_machdep_mitigations, OID_AUTO, mds, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "Microarchitectural Data Sampling Mitigation state"); SYSCTL_PROC(_machdep_mitigations_mds, OID_AUTO, state, CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0, sysctl_hw_mds_disable_state_handler, "A", "Microarchitectural Data Sampling Mitigation state"); _Static_assert(__offsetof(struct pcpu, pc_mds_tmp) % 64 == 0, "MDS AVX512"); void hw_mds_recalculate(void) { struct pcpu *pc; vm_offset_t b64; u_long xcr0; int i; /* * Allow user to force VERW variant even if MD_CLEAR is not * reported. For instance, hypervisor might unknowingly * filter the cap out. * For the similar reasons, and for testing, allow to enable * mitigation even when MDS_NO cap is set. */ if (cpu_vendor_id != CPU_VENDOR_INTEL || hw_mds_disable == 0 || ((cpu_ia32_arch_caps & IA32_ARCH_CAP_MDS_NO) != 0 && hw_mds_disable == 3)) { mds_handler = mds_handler_void; } else if (((cpu_stdext_feature3 & CPUID_STDEXT3_MD_CLEAR) != 0 && hw_mds_disable == 3) || hw_mds_disable == 1) { mds_handler = mds_handler_verw; } else if (CPUID_TO_FAMILY(cpu_id) == 0x6 && (CPUID_TO_MODEL(cpu_id) == 0x2e || CPUID_TO_MODEL(cpu_id) == 0x1e || CPUID_TO_MODEL(cpu_id) == 0x1f || CPUID_TO_MODEL(cpu_id) == 0x1a || CPUID_TO_MODEL(cpu_id) == 0x2f || CPUID_TO_MODEL(cpu_id) == 0x25 || CPUID_TO_MODEL(cpu_id) == 0x2c || CPUID_TO_MODEL(cpu_id) == 0x2d || CPUID_TO_MODEL(cpu_id) == 0x2a || CPUID_TO_MODEL(cpu_id) == 0x3e || CPUID_TO_MODEL(cpu_id) == 0x3a) && (hw_mds_disable == 2 || hw_mds_disable == 3)) { /* * Nehalem, SandyBridge, IvyBridge */ CPU_FOREACH(i) { pc = pcpu_find(i); if (pc->pc_mds_buf == NULL) { pc->pc_mds_buf = malloc_domainset(672, M_TEMP, DOMAINSET_PREF(pc->pc_domain), M_WAITOK); bzero(pc->pc_mds_buf, 16); } } mds_handler = mds_handler_ivb; } else if (CPUID_TO_FAMILY(cpu_id) == 0x6 && (CPUID_TO_MODEL(cpu_id) == 0x3f || CPUID_TO_MODEL(cpu_id) == 0x3c || CPUID_TO_MODEL(cpu_id) == 0x45 || CPUID_TO_MODEL(cpu_id) == 0x46 || CPUID_TO_MODEL(cpu_id) == 0x56 || CPUID_TO_MODEL(cpu_id) == 0x4f || CPUID_TO_MODEL(cpu_id) == 0x47 || CPUID_TO_MODEL(cpu_id) == 0x3d) && (hw_mds_disable == 2 || hw_mds_disable == 3)) { /* * Haswell, Broadwell */ CPU_FOREACH(i) { pc = pcpu_find(i); if (pc->pc_mds_buf == NULL) { pc->pc_mds_buf = malloc_domainset(1536, M_TEMP, DOMAINSET_PREF(pc->pc_domain), M_WAITOK); bzero(pc->pc_mds_buf, 16); } } mds_handler = mds_handler_bdw; } else if (CPUID_TO_FAMILY(cpu_id) == 0x6 && ((CPUID_TO_MODEL(cpu_id) == 0x55 && (cpu_id & CPUID_STEPPING) <= 5) || CPUID_TO_MODEL(cpu_id) == 0x4e || CPUID_TO_MODEL(cpu_id) == 0x5e || (CPUID_TO_MODEL(cpu_id) == 0x8e && (cpu_id & CPUID_STEPPING) <= 0xb) || (CPUID_TO_MODEL(cpu_id) == 0x9e && (cpu_id & CPUID_STEPPING) <= 0xc)) && (hw_mds_disable == 2 || hw_mds_disable == 3)) { /* * Skylake, KabyLake, CoffeeLake, WhiskeyLake, * CascadeLake */ CPU_FOREACH(i) { pc = pcpu_find(i); if (pc->pc_mds_buf == NULL) { pc->pc_mds_buf = malloc_domainset(6 * 1024, M_TEMP, DOMAINSET_PREF(pc->pc_domain), M_WAITOK); b64 = (vm_offset_t)malloc_domainset(64 + 63, M_TEMP, DOMAINSET_PREF(pc->pc_domain), M_WAITOK); pc->pc_mds_buf64 = (void *)roundup2(b64, 64); bzero(pc->pc_mds_buf64, 64); } } xcr0 = rxcr(0); if ((xcr0 & XFEATURE_ENABLED_ZMM_HI256) != 0 && (cpu_stdext_feature & CPUID_STDEXT_AVX512DQ) != 0) mds_handler = mds_handler_skl_avx512; else if ((xcr0 & XFEATURE_ENABLED_AVX) != 0 && (cpu_feature2 & CPUID2_AVX) != 0) mds_handler = mds_handler_skl_avx; else mds_handler = mds_handler_skl_sse; } else if (CPUID_TO_FAMILY(cpu_id) == 0x6 && ((CPUID_TO_MODEL(cpu_id) == 0x37 || CPUID_TO_MODEL(cpu_id) == 0x4a || CPUID_TO_MODEL(cpu_id) == 0x4c || CPUID_TO_MODEL(cpu_id) == 0x4d || CPUID_TO_MODEL(cpu_id) == 0x5a || CPUID_TO_MODEL(cpu_id) == 0x5d || CPUID_TO_MODEL(cpu_id) == 0x6e || CPUID_TO_MODEL(cpu_id) == 0x65 || CPUID_TO_MODEL(cpu_id) == 0x75 || CPUID_TO_MODEL(cpu_id) == 0x1c || CPUID_TO_MODEL(cpu_id) == 0x26 || CPUID_TO_MODEL(cpu_id) == 0x27 || CPUID_TO_MODEL(cpu_id) == 0x35 || CPUID_TO_MODEL(cpu_id) == 0x36 || CPUID_TO_MODEL(cpu_id) == 0x7a))) { /* Silvermont, Airmont */ CPU_FOREACH(i) { pc = pcpu_find(i); if (pc->pc_mds_buf == NULL) pc->pc_mds_buf = malloc(256, M_TEMP, M_WAITOK); } mds_handler = mds_handler_silvermont; } else { hw_mds_disable = 0; mds_handler = mds_handler_void; } } static void hw_mds_recalculate_boot(void *arg __unused) { hw_mds_recalculate(); } SYSINIT(mds_recalc, SI_SUB_SMP, SI_ORDER_ANY, hw_mds_recalculate_boot, NULL); static int sysctl_mds_disable_handler(SYSCTL_HANDLER_ARGS) { int error, val; val = hw_mds_disable; error = sysctl_handle_int(oidp, &val, 0, req); if (error != 0 || req->newptr == NULL) return (error); if (val < 0 || val > 3) return (EINVAL); hw_mds_disable = val; hw_mds_recalculate(); return (0); } SYSCTL_PROC(_hw, OID_AUTO, mds_disable, CTLTYPE_INT | CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0, sysctl_mds_disable_handler, "I", "Microarchitectural Data Sampling Mitigation " "(0 - off, 1 - on VERW, 2 - on SW, 3 - on AUTO)"); SYSCTL_PROC(_machdep_mitigations_mds, OID_AUTO, disable, CTLTYPE_INT | CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0, sysctl_mds_disable_handler, "I", "Microarchitectural Data Sampling Mitigation " "(0 - off, 1 - on VERW, 2 - on SW, 3 - on AUTO)"); /* * Intel Transactional Memory Asynchronous Abort Mitigation * CVE-2019-11135 */ int x86_taa_enable; int x86_taa_state; enum { TAA_NONE = 0, /* No mitigation enabled */ TAA_TSX_DISABLE = 1, /* Disable TSX via MSR */ TAA_VERW = 2, /* Use VERW mitigation */ TAA_AUTO = 3, /* Automatically select the mitigation */ /* The states below are not selectable by the operator */ TAA_TAA_UC = 4, /* Mitigation present in microcode */ TAA_NOT_PRESENT = 5 /* TSX is not present */ }; static void taa_set(bool enable, bool all) { x86_msr_op(MSR_IA32_TSX_CTRL, (enable ? MSR_OP_OR : MSR_OP_ANDNOT) | (all ? MSR_OP_RENDEZVOUS_ALL : MSR_OP_LOCAL), IA32_TSX_CTRL_RTM_DISABLE | IA32_TSX_CTRL_TSX_CPUID_CLEAR, NULL); } void x86_taa_recalculate(void) { static int taa_saved_mds_disable = 0; int taa_need = 0, taa_state = 0; int mds_disable = 0, need_mds_recalc = 0; /* Check CPUID.07h.EBX.HLE and RTM for the presence of TSX */ if ((cpu_stdext_feature & CPUID_STDEXT_HLE) == 0 || (cpu_stdext_feature & CPUID_STDEXT_RTM) == 0) { /* TSX is not present */ x86_taa_state = TAA_NOT_PRESENT; return; } /* Check to see what mitigation options the CPU gives us */ if (cpu_ia32_arch_caps & IA32_ARCH_CAP_TAA_NO) { /* CPU is not suseptible to TAA */ taa_need = TAA_TAA_UC; } else if (cpu_ia32_arch_caps & IA32_ARCH_CAP_TSX_CTRL) { /* * CPU can turn off TSX. This is the next best option * if TAA_NO hardware mitigation isn't present */ taa_need = TAA_TSX_DISABLE; } else { /* No TSX/TAA specific remedies are available. */ if (x86_taa_enable == TAA_TSX_DISABLE) { if (bootverbose) printf("TSX control not available\n"); return; } else taa_need = TAA_VERW; } /* Can we automatically take action, or are we being forced? */ if (x86_taa_enable == TAA_AUTO) taa_state = taa_need; else taa_state = x86_taa_enable; /* No state change, nothing to do */ if (taa_state == x86_taa_state) { if (bootverbose) printf("No TSX change made\n"); return; } /* Does the MSR need to be turned on or off? */ if (taa_state == TAA_TSX_DISABLE) taa_set(true, true); else if (x86_taa_state == TAA_TSX_DISABLE) taa_set(false, true); /* Does MDS need to be set to turn on VERW? */ if (taa_state == TAA_VERW) { taa_saved_mds_disable = hw_mds_disable; mds_disable = hw_mds_disable = 1; need_mds_recalc = 1; } else if (x86_taa_state == TAA_VERW) { mds_disable = hw_mds_disable = taa_saved_mds_disable; need_mds_recalc = 1; } if (need_mds_recalc) { hw_mds_recalculate(); if (mds_disable != hw_mds_disable) { if (bootverbose) printf("Cannot change MDS state for TAA\n"); /* Don't update our state */ return; } } x86_taa_state = taa_state; return; } static void taa_recalculate_boot(void * arg __unused) { x86_taa_recalculate(); } SYSINIT(taa_recalc, SI_SUB_SMP, SI_ORDER_ANY, taa_recalculate_boot, NULL); SYSCTL_NODE(_machdep_mitigations, OID_AUTO, taa, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "TSX Asynchronous Abort Mitigation"); static int sysctl_taa_handler(SYSCTL_HANDLER_ARGS) { int error, val; val = x86_taa_enable; error = sysctl_handle_int(oidp, &val, 0, req); if (error != 0 || req->newptr == NULL) return (error); if (val < TAA_NONE || val > TAA_AUTO) return (EINVAL); x86_taa_enable = val; x86_taa_recalculate(); return (0); } SYSCTL_PROC(_machdep_mitigations_taa, OID_AUTO, enable, CTLTYPE_INT | CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0, sysctl_taa_handler, "I", "TAA Mitigation enablement control " "(0 - off, 1 - disable TSX, 2 - VERW, 3 - on AUTO)"); static int sysctl_taa_state_handler(SYSCTL_HANDLER_ARGS) { const char *state; switch (x86_taa_state) { case TAA_NONE: state = "inactive"; break; case TAA_TSX_DISABLE: state = "TSX disabled"; break; case TAA_VERW: state = "VERW"; break; case TAA_TAA_UC: state = "Mitigated in microcode"; break; case TAA_NOT_PRESENT: state = "TSX not present"; break; default: state = "unknown"; } return (SYSCTL_OUT(req, state, strlen(state))); } SYSCTL_PROC(_machdep_mitigations_taa, OID_AUTO, state, CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0, sysctl_taa_state_handler, "A", "TAA Mitigation state"); int __read_frequently cpu_flush_rsb_ctxsw; SYSCTL_INT(_machdep_mitigations, OID_AUTO, flush_rsb_ctxsw, CTLFLAG_RW | CTLFLAG_NOFETCH, &cpu_flush_rsb_ctxsw, 0, "Flush Return Stack Buffer on context switch"); SYSCTL_NODE(_machdep_mitigations, OID_AUTO, rngds, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, "MCU Optimization, disable RDSEED mitigation"); int x86_rngds_mitg_enable = 1; void x86_rngds_mitg_recalculate(bool all_cpus) { if ((cpu_stdext_feature3 & CPUID_STDEXT3_MCUOPT) == 0) return; x86_msr_op(MSR_IA32_MCU_OPT_CTRL, (x86_rngds_mitg_enable ? MSR_OP_OR : MSR_OP_ANDNOT) | (all_cpus ? MSR_OP_RENDEZVOUS_ALL : MSR_OP_LOCAL), IA32_RNGDS_MITG_DIS, NULL); } static int sysctl_rngds_mitg_enable_handler(SYSCTL_HANDLER_ARGS) { int error, val; val = x86_rngds_mitg_enable; error = sysctl_handle_int(oidp, &val, 0, req); if (error != 0 || req->newptr == NULL) return (error); x86_rngds_mitg_enable = val; x86_rngds_mitg_recalculate(true); return (0); } SYSCTL_PROC(_machdep_mitigations_rngds, OID_AUTO, enable, CTLTYPE_INT | CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0, sysctl_rngds_mitg_enable_handler, "I", "MCU Optimization, disabling RDSEED mitigation control " "(0 - mitigation disabled (RDSEED optimized), 1 - mitigation enabled)"); static int sysctl_rngds_state_handler(SYSCTL_HANDLER_ARGS) { const char *state; if ((cpu_stdext_feature3 & CPUID_STDEXT3_MCUOPT) == 0) { state = "Not applicable"; } else if (x86_rngds_mitg_enable == 0) { state = "RDSEED not serialized"; } else { state = "Mitigated"; } return (SYSCTL_OUT(req, state, strlen(state))); } SYSCTL_PROC(_machdep_mitigations_rngds, OID_AUTO, state, CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0, sysctl_rngds_state_handler, "A", "MCU Optimization state"); /* * Enable and restore kernel text write permissions. * Callers must ensure that disable_wp()/restore_wp() are executed * without rescheduling on the same core. */ bool disable_wp(void) { u_int cr0; cr0 = rcr0(); if ((cr0 & CR0_WP) == 0) return (false); load_cr0(cr0 & ~CR0_WP); return (true); } void restore_wp(bool old_wp) { if (old_wp) load_cr0(rcr0() | CR0_WP); } bool acpi_get_fadt_bootflags(uint16_t *flagsp) { #ifdef DEV_ACPI ACPI_TABLE_FADT *fadt; vm_paddr_t physaddr; physaddr = acpi_find_table(ACPI_SIG_FADT); if (physaddr == 0) return (false); fadt = acpi_map_table(physaddr, ACPI_SIG_FADT); if (fadt == NULL) return (false); *flagsp = fadt->BootFlags; acpi_unmap_table(fadt); return (true); #else return (false); #endif } DEFINE_IFUNC(, uint64_t, rdtsc_ordered, (void)) { bool cpu_is_amd = cpu_vendor_id == CPU_VENDOR_AMD || cpu_vendor_id == CPU_VENDOR_HYGON; if ((amd_feature & AMDID_RDTSCP) != 0) return (rdtscp); else if ((cpu_feature & CPUID_SSE2) != 0) return (cpu_is_amd ? rdtsc_ordered_mfence : rdtsc_ordered_lfence); else return (rdtsc); }