Index: stable/12/share/man/man4/smp.4 =================================================================== --- stable/12/share/man/man4/smp.4 (revision 354072) +++ stable/12/share/man/man4/smp.4 (revision 354073) @@ -1,168 +1,190 @@ .\" Copyright (c) 1997 .\" Steve Passe . 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. The name of the developer may NOT 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. .\" .\" $FreeBSD$ .\" -.Dd January 6, 2018 +.Dd January 4, 2019 .Dt SMP 4 .Os .Sh NAME .Nm SMP .Nd description of the FreeBSD Symmetric Multi-Processor kernel .Sh SYNOPSIS .Cd options SMP .Sh DESCRIPTION The .Nm kernel implements symmetric multi-processor support. -.Sh COMPATIBILITY -Support for multi-processor systems is present for all Tier-1 -architectures on -.Fx . -Currently, this includes amd64, i386 and sparc64. -Support is enabled using -.Cd options SMP . -It is permissible to use the SMP kernel configuration on non-SMP equipped -motherboards. -.Sh I386 NOTES -For i386 systems, the -.Nm -kernel supports motherboards that follow the Intel MP specification, -version 1.4. -In addition to -.Cd options SMP , -i386 also requires -.Cd device apic . -The -.Xr mptable 1 -command may be used to view the status of multi-processor support. .Pp .Nm support can be disabled by setting the loader tunable .Va kern.smp.disabled to 1. .Pp The number of CPUs detected by the system is available in the read-only sysctl variable .Va hw.ncpu . .Pp +The number of online threads per CPU core is available in the read-only sysctl +variable +.Va kern.smp.threads_per_core . +The number of physical CPU cores detected by the system is available in the +read-only sysctl variable +.Va kern.smp.cores . +.Pp .Fx allows specific CPUs on a multi-processor system to be disabled. This can be done using the .Va hint.lapic.X.disabled tunable, where X is the APIC ID of a CPU. Setting this tunable to 1 will result in the corresponding CPU being disabled. .Pp +.Fx +supports simultaneous multithreading on x86 and powerpc platforms. +On x86, the logical CPUs can be disabled by setting the +.Va machdep.hyperthreading_allowed +tunable to zero. +.Pp The .Xr sched_ule 4 scheduler implements CPU topology detection and adjusts the scheduling algorithms to make better use of modern multi-core CPUs. The sysctl variable .Va kern.sched.topology_spec reflects the detected CPU hardware in a parsable XML format. The top level XML tag is , which encloses one or more tags containing data about individual CPU groups. A CPU group contains CPUs that are detected to be "close" together, usually by being cores in a single multi-core processor. Attributes available in a tag are "level", corresponding to the nesting level of the CPU group and "cache-level", corresponding to the level of CPU caches shared by the CPUs in the group. The tag contains the and tags. The tag describes CPUs in the group. Its attributes are "count", corresponding to the number of CPUs in the group and "mask", corresponding to the integer binary mask in which each bit position set to 1 signifies a CPU belonging to the group. The contents (CDATA) of the tag is the comma-delimited list of CPU indexes (derived from the "mask" attribute). The tag contains special tags (if any) describing the relation of the CPUs in the group. The possible flags are currently "HTT" and "SMT", corresponding to the various implementations of hardware multithreading. An example topology_spec output for a system consisting of two quad-core processors is: .Bd -literal 0, 1, 2, 3, 4, 5, 6, 7 0, 1, 2, 3 4, 5, 6, 7 .Ed .Pp This information is used internally by the kernel to schedule related tasks on CPUs that are closely grouped together. -.Pp -.Fx -supports hyperthreading on Intel CPU's on the i386 and AMD64 platforms. -Because using logical CPUs can cause performance penalties under certain loads, -the logical CPUs can be disabled by setting the -.Va machdep.hyperthreading_allowed -tunable to zero. +.Sh COMPATIBILITY +Support for multi-processor systems is present for all Tier-1 and Tier-2 +architectures on +.Fx . +Currently, this includes x86, powerpc, arm, and sparc64. +Support is enabled using +.Cd options SMP . +It is permissible to use the SMP kernel configuration on non-SMP hardware. +.Sh I386 NOTES +For i386 systems, the +.Nm +kernel supports motherboards that follow the Intel MP specification, +version 1.4. +In addition to +.Cd options SMP , +i386 also requires +.Cd device apic . +The +.Xr mptable 1 +command may be used to view the status of multi-processor support. .Sh SEE ALSO .Xr cpuset 1 , .Xr mptable 1 , .Xr sched_4bsd 4 , .Xr sched_ule 4 , .Xr loader 8 , .Xr sysctl 8 , .Xr condvar 9 , .Xr msleep 9 , .Xr mtx_pool 9 , .Xr mutex 9 , .Xr rwlock 9 , .Xr sema 9 , .Xr sx 9 .Sh HISTORY The .Nm kernel's early history is not (properly) recorded. It was developed in a separate CVS branch until April 26, 1997, at which point it was merged into 3.0-current. By this date 3.0-current had already been merged with Lite2 kernel code. .Pp .Fx 5.0 introduced support for a host of new synchronization primitives, and a move towards fine-grained kernel locking rather than reliance on a Giant kernel lock. The SMPng Project relied heavily on the support of BSDi, who provided reference source code from the fine-grained SMP implementation found in .Bsx . .Pp .Fx 5.0 also introduced support for SMP on the sparc64 architecture. .Sh AUTHORS .An Steve Passe Aq Mt fsmp@FreeBSD.org +.Sh CAVEATS +The +.Va kern.smp.threads_per_core +and +.Va kern.smp.cores +sysctl variables are provided as a best-effort guess. +If an architecture or platform adds SMT and +.Fx +has not yet implemented detection, the reported values may be inaccurate. +In this case, +.Va kern.smp.threads_per_core +will report +.Dv 1 +and +.Va kern.smp.cores +will report the same value as +.Va hw.ncpu . Index: stable/12/sys/kern/subr_smp.c =================================================================== --- stable/12/sys/kern/subr_smp.c (revision 354072) +++ stable/12/sys/kern/subr_smp.c (revision 354073) @@ -1,1159 +1,1174 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2001, John Baldwin . * 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. */ /* * 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, 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 CPUs 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 || panicstr) 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; KASSERT(type == IPI_STOP || type == IPI_STOP_HARD #if X86 || type == IPI_SUSPEND #endif , ("%s: invalid stop type", __func__)); if (!smp_started) return (0); CTR1(KTR_SMP, "restart_cpus(%s)", cpusetobj_strprint(cpusetbuf, &map)); #if X86 if (type == IPI_SUSPEND) cpus = &resuming_cpus; else #endif cpus = &stopped_cpus; /* signal other cpus to restart */ #if X86 if (type == IPI_SUSPEND) CPU_COPY_STORE_REL(&map, &toresume_cpus); else #endif CPU_COPY_STORE_REL(&map, &started_cpus); #if X86 if (!nmi_is_broadcast || nmi_kdb_lock == 0) { #endif /* wait for each to clear its bit */ while (CPU_OVERLAP(cpus, &map)) cpu_spinwait(); #if X86 } #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; /* Look 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; } 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]; 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; } 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); 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 } /* * 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. */ int quiesce_cpus(cpuset_t map, const char *wmesg, int prio) { struct pcpu *pcpu; u_int gen[MAXCPU]; int error; int cpu; error = 0; 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); while (gen[cpu] == pcpu->pc_idlethread->td_generation) { error = tsleep(quiesce_cpus, prio, wmesg, 1); if (error != EWOULDBLOCK) goto out; error = 0; } } out: thread_lock(curthread); sched_unbind(curthread); thread_unlock(curthread); return (error); } int quiesce_all_cpus(const char *wmesg, int prio) { return quiesce_cpus(all_cpus, wmesg, prio); } /* 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 */ Index: stable/12/sys/powerpc/powernv/platform_powernv.c =================================================================== --- stable/12/sys/powerpc/powernv/platform_powernv.c (revision 354072) +++ stable/12/sys/powerpc/powernv/platform_powernv.c (revision 354073) @@ -1,471 +1,475 @@ /*- * Copyright (c) 2015 Nathan Whitehorn * Copyright (c) 2017-2018 Semihalf * 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 ``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 #include #include #include #include "platform_if.h" #include "opal.h" #ifdef SMP extern void *ap_pcpu; #endif extern void xicp_smp_cpu_startup(void); static int powernv_probe(platform_t); static int powernv_attach(platform_t); void powernv_mem_regions(platform_t, struct mem_region *phys, int *physsz, struct mem_region *avail, int *availsz); static u_long powernv_timebase_freq(platform_t, struct cpuref *cpuref); static int powernv_smp_first_cpu(platform_t, struct cpuref *cpuref); static int powernv_smp_next_cpu(platform_t, struct cpuref *cpuref); static int powernv_smp_get_bsp(platform_t, struct cpuref *cpuref); static void powernv_smp_ap_init(platform_t); #ifdef SMP static int powernv_smp_start_cpu(platform_t, struct pcpu *cpu); static struct cpu_group *powernv_smp_topo(platform_t plat); #endif static void powernv_reset(platform_t); static void powernv_cpu_idle(sbintime_t sbt); static int powernv_cpuref_init(void); static platform_method_t powernv_methods[] = { PLATFORMMETHOD(platform_probe, powernv_probe), PLATFORMMETHOD(platform_attach, powernv_attach), PLATFORMMETHOD(platform_mem_regions, powernv_mem_regions), PLATFORMMETHOD(platform_timebase_freq, powernv_timebase_freq), PLATFORMMETHOD(platform_smp_ap_init, powernv_smp_ap_init), PLATFORMMETHOD(platform_smp_first_cpu, powernv_smp_first_cpu), PLATFORMMETHOD(platform_smp_next_cpu, powernv_smp_next_cpu), PLATFORMMETHOD(platform_smp_get_bsp, powernv_smp_get_bsp), #ifdef SMP PLATFORMMETHOD(platform_smp_start_cpu, powernv_smp_start_cpu), PLATFORMMETHOD(platform_smp_topo, powernv_smp_topo), #endif PLATFORMMETHOD(platform_reset, powernv_reset), { 0, 0 } }; static platform_def_t powernv_platform = { "powernv", powernv_methods, 0 }; static struct cpuref platform_cpuref[MAXCPU]; static int platform_cpuref_cnt; static int platform_cpuref_valid; PLATFORM_DEF(powernv_platform); static uint64_t powernv_boot_pir; static int powernv_probe(platform_t plat) { if (opal_check() == 0) return (BUS_PROBE_SPECIFIC); return (ENXIO); } static int powernv_attach(platform_t plat) { uint32_t nptlp, shift = 0, slb_encoding = 0; int32_t lp_size, lp_encoding; char buf[255]; pcell_t prop; phandle_t cpu; int res, len, idx; register_t msr; /* Ping OPAL again just to make sure */ opal_check(); #if BYTE_ORDER == LITTLE_ENDIAN opal_call(OPAL_REINIT_CPUS, 2 /* Little endian */); #else opal_call(OPAL_REINIT_CPUS, 1 /* Big endian */); #endif if (cpu_idle_hook == NULL) cpu_idle_hook = powernv_cpu_idle; powernv_boot_pir = mfspr(SPR_PIR); /* LPID must not be altered when PSL_DR or PSL_IR is set */ msr = mfmsr(); mtmsr(msr & ~(PSL_DR | PSL_IR)); /* Direct interrupts to SRR instead of HSRR and reset LPCR otherwise */ mtspr(SPR_LPID, 0); isync(); if (cpu_features2 & PPC_FEATURE2_ARCH_3_00) lpcr |= LPCR_HVICE; mtspr(SPR_LPCR, lpcr); isync(); mtmsr(msr); powernv_cpuref_init(); /* Set SLB count from device tree */ cpu = OF_peer(0); cpu = OF_child(cpu); while (cpu != 0) { res = OF_getprop(cpu, "name", buf, sizeof(buf)); if (res > 0 && strcmp(buf, "cpus") == 0) break; cpu = OF_peer(cpu); } if (cpu == 0) goto out; cpu = OF_child(cpu); while (cpu != 0) { res = OF_getprop(cpu, "device_type", buf, sizeof(buf)); if (res > 0 && strcmp(buf, "cpu") == 0) break; cpu = OF_peer(cpu); } if (cpu == 0) goto out; res = OF_getencprop(cpu, "ibm,slb-size", &prop, sizeof(prop)); if (res > 0) n_slbs = prop; /* * Scan the large page size property for PAPR compatible machines. * See PAPR D.5 Changes to Section 5.1.4, 'CPU Node Properties' * for the encoding of the property. */ len = OF_getproplen(cpu, "ibm,segment-page-sizes"); if (len > 0) { /* * We have to use a variable length array on the stack * since we have very limited stack space. */ pcell_t arr[len/sizeof(cell_t)]; res = OF_getencprop(cpu, "ibm,segment-page-sizes", arr, sizeof(arr)); len /= 4; idx = 0; while (len > 0) { shift = arr[idx]; slb_encoding = arr[idx + 1]; nptlp = arr[idx + 2]; idx += 3; len -= 3; while (len > 0 && nptlp) { lp_size = arr[idx]; lp_encoding = arr[idx+1]; if (slb_encoding == SLBV_L && lp_encoding == 0) break; idx += 2; len -= 2; nptlp--; } if (nptlp && slb_encoding == SLBV_L && lp_encoding == 0) break; } if (len == 0) panic("Standard large pages (SLB[L] = 1, PTE[LP] = 0) " "not supported by this system."); moea64_large_page_shift = shift; moea64_large_page_size = 1ULL << lp_size; } out: return (0); } void powernv_mem_regions(platform_t plat, struct mem_region *phys, int *physsz, struct mem_region *avail, int *availsz) { ofw_mem_regions(phys, physsz, avail, availsz); } static u_long powernv_timebase_freq(platform_t plat, struct cpuref *cpuref) { char buf[8]; phandle_t cpu, dev, root; int res; int32_t ticks = -1; root = OF_peer(0); dev = OF_child(root); while (dev != 0) { res = OF_getprop(dev, "name", buf, sizeof(buf)); if (res > 0 && strcmp(buf, "cpus") == 0) break; dev = OF_peer(dev); } for (cpu = OF_child(dev); cpu != 0; cpu = OF_peer(cpu)) { res = OF_getprop(cpu, "device_type", buf, sizeof(buf)); if (res > 0 && strcmp(buf, "cpu") == 0) break; } if (cpu == 0) return (512000000); OF_getencprop(cpu, "timebase-frequency", &ticks, sizeof(ticks)); if (ticks <= 0) panic("Unable to determine timebase frequency!"); return (ticks); } static int powernv_cpuref_init(void) { phandle_t cpu, dev; char buf[32]; int a, res, tmp_cpuref_cnt; static struct cpuref tmp_cpuref[MAXCPU]; cell_t interrupt_servers[32]; uint64_t bsp; if (platform_cpuref_valid) return (0); dev = OF_peer(0); dev = OF_child(dev); while (dev != 0) { res = OF_getprop(dev, "name", buf, sizeof(buf)); if (res > 0 && strcmp(buf, "cpus") == 0) break; dev = OF_peer(dev); } bsp = 0; tmp_cpuref_cnt = 0; for (cpu = OF_child(dev); cpu != 0; cpu = OF_peer(cpu)) { res = OF_getprop(cpu, "device_type", buf, sizeof(buf)); if (res > 0 && strcmp(buf, "cpu") == 0) { res = OF_getproplen(cpu, "ibm,ppc-interrupt-server#s"); if (res > 0) { OF_getencprop(cpu, "ibm,ppc-interrupt-server#s", interrupt_servers, res); for (a = 0; a < res/sizeof(cell_t); a++) { tmp_cpuref[tmp_cpuref_cnt].cr_hwref = interrupt_servers[a]; tmp_cpuref[tmp_cpuref_cnt].cr_cpuid = tmp_cpuref_cnt; if (interrupt_servers[a] == (uint32_t)powernv_boot_pir) bsp = tmp_cpuref_cnt; tmp_cpuref_cnt++; } } } } /* Map IDs, so BSP has CPUID 0 regardless of hwref */ for (a = bsp; a < tmp_cpuref_cnt; a++) { platform_cpuref[platform_cpuref_cnt].cr_hwref = tmp_cpuref[a].cr_hwref; platform_cpuref[platform_cpuref_cnt].cr_cpuid = platform_cpuref_cnt; platform_cpuref_cnt++; } for (a = 0; a < bsp; a++) { platform_cpuref[platform_cpuref_cnt].cr_hwref = tmp_cpuref[a].cr_hwref; platform_cpuref[platform_cpuref_cnt].cr_cpuid = platform_cpuref_cnt; platform_cpuref_cnt++; } platform_cpuref_valid = 1; return (0); } static int powernv_smp_first_cpu(platform_t plat, struct cpuref *cpuref) { if (platform_cpuref_valid == 0) return (EINVAL); cpuref->cr_cpuid = 0; cpuref->cr_hwref = platform_cpuref[0].cr_hwref; return (0); } static int powernv_smp_next_cpu(platform_t plat, struct cpuref *cpuref) { int id; if (platform_cpuref_valid == 0) return (EINVAL); id = cpuref->cr_cpuid + 1; if (id >= platform_cpuref_cnt) return (ENOENT); cpuref->cr_cpuid = platform_cpuref[id].cr_cpuid; cpuref->cr_hwref = platform_cpuref[id].cr_hwref; return (0); } static int powernv_smp_get_bsp(platform_t plat, struct cpuref *cpuref) { cpuref->cr_cpuid = platform_cpuref[0].cr_cpuid; cpuref->cr_hwref = platform_cpuref[0].cr_hwref; return (0); } #ifdef SMP static int powernv_smp_start_cpu(platform_t plat, struct pcpu *pc) { int result; ap_pcpu = pc; powerpc_sync(); result = opal_call(OPAL_START_CPU, pc->pc_hwref, EXC_RST); if (result != OPAL_SUCCESS) { printf("OPAL error (%d): unable to start AP %d\n", result, (int)pc->pc_hwref); return (ENXIO); } return (0); } static struct cpu_group * powernv_smp_topo(platform_t plat) { char buf[8]; phandle_t cpu, dev, root; int res, nthreads; root = OF_peer(0); dev = OF_child(root); while (dev != 0) { res = OF_getprop(dev, "name", buf, sizeof(buf)); if (res > 0 && strcmp(buf, "cpus") == 0) break; dev = OF_peer(dev); } nthreads = 1; for (cpu = OF_child(dev); cpu != 0; cpu = OF_peer(cpu)) { res = OF_getprop(cpu, "device_type", buf, sizeof(buf)); if (res <= 0 || strcmp(buf, "cpu") != 0) continue; res = OF_getproplen(cpu, "ibm,ppc-interrupt-server#s"); if (res >= 0) nthreads = res / sizeof(cell_t); else nthreads = 1; break; } + smp_threads_per_core = nthreads; + if (mp_ncpus % nthreads != 0) { printf("WARNING: Irregular SMP topology. Performance may be " "suboptimal (%d threads, %d on first core)\n", mp_ncpus, nthreads); return (smp_topo_none()); } + + mp_ncores = mp_ncpus / nthreads; /* Don't do anything fancier for non-threaded SMP */ if (nthreads == 1) return (smp_topo_none()); return (smp_topo_1level(CG_SHARE_L1, nthreads, CG_FLAG_SMT)); } #endif static void powernv_reset(platform_t platform) { opal_call(OPAL_CEC_REBOOT); } static void powernv_smp_ap_init(platform_t platform) { xicp_smp_cpu_startup(); } static void powernv_cpu_idle(sbintime_t sbt) { } Index: stable/12/sys/powerpc/powerpc/mp_machdep.c =================================================================== --- stable/12/sys/powerpc/powerpc/mp_machdep.c (revision 354072) +++ stable/12/sys/powerpc/powerpc/mp_machdep.c (revision 354073) @@ -1,378 +1,383 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2008 Marcel Moolenaar * 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 ``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 #include #include #include #include #include #include #include #include #include #include "pic_if.h" extern struct pcpu __pcpu[MAXCPU]; volatile static int ap_awake; volatile static u_int ap_letgo; volatile static u_quad_t ap_timebase; static u_int ipi_msg_cnt[32]; static struct mtx ap_boot_mtx; struct pcb stoppcbs[MAXCPU]; void machdep_ap_bootstrap(void) { PCPU_SET(awake, 1); __asm __volatile("msync; isync"); while (ap_letgo == 0) __asm __volatile("or 27,27,27"); __asm __volatile("or 6,6,6"); /* * Set timebase as soon as possible to meet an implicit rendezvous * from cpu_mp_unleash(), which sets ap_letgo and then immediately * sets timebase. * * Note that this is instrinsically racy and is only relevant on * platforms that do not support better mechanisms. */ platform_smp_timebase_sync(ap_timebase, 1); /* Give platform code a chance to do anything else necessary */ platform_smp_ap_init(); /* Initialize decrementer */ decr_ap_init(); /* Serialize console output and AP count increment */ mtx_lock_spin(&ap_boot_mtx); ap_awake++; printf("SMP: AP CPU #%d launched\n", PCPU_GET(cpuid)); mtx_unlock_spin(&ap_boot_mtx); while(smp_started == 0) ; /* Start per-CPU event timers. */ cpu_initclocks_ap(); /* Announce ourselves awake, and enter the scheduler */ sched_throw(NULL); } void cpu_mp_setmaxid(void) { struct cpuref cpuref; int error; mp_ncpus = 0; mp_maxid = 0; error = platform_smp_first_cpu(&cpuref); while (!error) { mp_ncpus++; mp_maxid = max(cpuref.cr_cpuid, mp_maxid); error = platform_smp_next_cpu(&cpuref); } /* Sanity. */ if (mp_ncpus == 0) mp_ncpus = 1; } int cpu_mp_probe(void) { /* * We're not going to enable SMP if there's only 1 processor. */ return (mp_ncpus > 1); } void cpu_mp_start(void) { struct cpuref bsp, cpu; struct pcpu *pc; int error; error = platform_smp_get_bsp(&bsp); KASSERT(error == 0, ("Don't know BSP")); error = platform_smp_first_cpu(&cpu); while (!error) { if (cpu.cr_cpuid >= MAXCPU) { printf("SMP: cpu%d: skipped -- ID out of range\n", cpu.cr_cpuid); goto next; } if (CPU_ISSET(cpu.cr_cpuid, &all_cpus)) { printf("SMP: cpu%d: skipped - duplicate ID\n", cpu.cr_cpuid); goto next; } if (cpu.cr_cpuid != bsp.cr_cpuid) { void *dpcpu; pc = &__pcpu[cpu.cr_cpuid]; dpcpu = (void *)kmem_malloc(DPCPU_SIZE, M_WAITOK | M_ZERO); pcpu_init(pc, cpu.cr_cpuid, sizeof(*pc)); dpcpu_init(dpcpu, cpu.cr_cpuid); } else { pc = pcpup; pc->pc_cpuid = bsp.cr_cpuid; pc->pc_bsp = 1; } pc->pc_hwref = cpu.cr_hwref; CPU_SET(pc->pc_cpuid, &all_cpus); next: error = platform_smp_next_cpu(&cpu); } + +#ifdef SMP + /* Probe mp_ncores and smp_threads_per_core as a side effect. */ + (void)cpu_topo(); +#endif } void cpu_mp_announce(void) { struct pcpu *pc; int i; if (!bootverbose) return; CPU_FOREACH(i) { pc = pcpu_find(i); if (pc == NULL) continue; printf("cpu%d: dev=%x", i, (int)pc->pc_hwref); if (pc->pc_bsp) printf(" (BSP)"); printf("\n"); } } static void cpu_mp_unleash(void *dummy) { struct pcpu *pc; int cpus, timeout; int ret; if (mp_ncpus <= 1) return; mtx_init(&ap_boot_mtx, "ap boot", NULL, MTX_SPIN); cpus = 0; smp_cpus = 0; #ifdef BOOKE tlb1_ap_prep(); #endif STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) { cpus++; if (!pc->pc_bsp) { if (bootverbose) printf("Waking up CPU %d (dev=%x)\n", pc->pc_cpuid, (int)pc->pc_hwref); ret = platform_smp_start_cpu(pc); if (ret == 0) { timeout = 2000; /* wait 2sec for the AP */ while (!pc->pc_awake && --timeout > 0) DELAY(1000); } } else { pc->pc_awake = 1; } if (pc->pc_awake) { if (bootverbose) printf("Adding CPU %d, hwref=%jx, awake=%x\n", pc->pc_cpuid, (uintmax_t)pc->pc_hwref, pc->pc_awake); smp_cpus++; } else CPU_SET(pc->pc_cpuid, &stopped_cpus); } ap_awake = 1; /* Provide our current DEC and TB values for APs */ ap_timebase = mftb() + 10; __asm __volatile("msync; isync"); /* Let APs continue */ atomic_store_rel_int(&ap_letgo, 1); platform_smp_timebase_sync(ap_timebase, 0); while (ap_awake < smp_cpus) ; if (smp_cpus != cpus || cpus != mp_ncpus) { printf("SMP: %d CPUs found; %d CPUs usable; %d CPUs woken\n", mp_ncpus, cpus, smp_cpus); } if (smp_cpus > 1) atomic_store_rel_int(&smp_started, 1); /* Let the APs get into the scheduler */ DELAY(10000); } SYSINIT(start_aps, SI_SUB_SMP, SI_ORDER_FIRST, cpu_mp_unleash, NULL); int powerpc_ipi_handler(void *arg) { u_int cpuid; uint32_t ipimask; int msg; CTR2(KTR_SMP, "%s: MSR 0x%08x", __func__, mfmsr()); ipimask = atomic_readandclear_32(&(pcpup->pc_ipimask)); if (ipimask == 0) return (FILTER_STRAY); while ((msg = ffs(ipimask) - 1) != -1) { ipimask &= ~(1u << msg); ipi_msg_cnt[msg]++; switch (msg) { case IPI_AST: CTR1(KTR_SMP, "%s: IPI_AST", __func__); break; case IPI_PREEMPT: CTR1(KTR_SMP, "%s: IPI_PREEMPT", __func__); sched_preempt(curthread); break; case IPI_RENDEZVOUS: CTR1(KTR_SMP, "%s: IPI_RENDEZVOUS", __func__); smp_rendezvous_action(); break; case IPI_STOP: /* * IPI_STOP_HARD is mapped to IPI_STOP so it is not * necessary to add such case in the switch. */ CTR1(KTR_SMP, "%s: IPI_STOP or IPI_STOP_HARD (stop)", __func__); cpuid = PCPU_GET(cpuid); savectx(&stoppcbs[cpuid]); savectx(PCPU_GET(curpcb)); CPU_SET_ATOMIC(cpuid, &stopped_cpus); while (!CPU_ISSET(cpuid, &started_cpus)) cpu_spinwait(); CPU_CLR_ATOMIC(cpuid, &stopped_cpus); CPU_CLR_ATOMIC(cpuid, &started_cpus); CTR1(KTR_SMP, "%s: IPI_STOP (restart)", __func__); break; case IPI_HARDCLOCK: CTR1(KTR_SMP, "%s: IPI_HARDCLOCK", __func__); hardclockintr(); break; } } return (FILTER_HANDLED); } static void ipi_send(struct pcpu *pc, int ipi) { CTR4(KTR_SMP, "%s: pc=%p, targetcpu=%d, IPI=%d", __func__, pc, pc->pc_cpuid, ipi); atomic_set_32(&pc->pc_ipimask, (1 << ipi)); powerpc_sync(); PIC_IPI(root_pic, pc->pc_cpuid); CTR1(KTR_SMP, "%s: sent", __func__); } /* Send an IPI to a set of cpus. */ void ipi_selected(cpuset_t cpus, int ipi) { struct pcpu *pc; STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) { if (CPU_ISSET(pc->pc_cpuid, &cpus)) ipi_send(pc, ipi); } } /* Send an IPI to a specific CPU. */ void ipi_cpu(int cpu, u_int ipi) { ipi_send(cpuid_to_pcpu[cpu], ipi); } /* Send an IPI to all CPUs EXCEPT myself. */ void ipi_all_but_self(int ipi) { struct pcpu *pc; STAILQ_FOREACH(pc, &cpuhead, pc_allcpu) { if (pc != pcpup) ipi_send(pc, ipi); } } Index: stable/12/sys/powerpc/ps3/platform_ps3.c =================================================================== --- stable/12/sys/powerpc/ps3/platform_ps3.c (revision 354072) +++ stable/12/sys/powerpc/ps3/platform_ps3.c (revision 354073) @@ -1,285 +1,287 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2010 Nathan Whitehorn * 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 ``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 #include #include "platform_if.h" #include "ps3-hvcall.h" #ifdef SMP extern void *ap_pcpu; #endif static int ps3_probe(platform_t); static int ps3_attach(platform_t); static void ps3_mem_regions(platform_t, struct mem_region *phys, int *physsz, struct mem_region *avail, int *availsz); static vm_offset_t ps3_real_maxaddr(platform_t); static u_long ps3_timebase_freq(platform_t, struct cpuref *cpuref); #ifdef SMP static int ps3_smp_first_cpu(platform_t, struct cpuref *cpuref); static int ps3_smp_next_cpu(platform_t, struct cpuref *cpuref); static int ps3_smp_get_bsp(platform_t, struct cpuref *cpuref); static int ps3_smp_start_cpu(platform_t, struct pcpu *cpu); static struct cpu_group *ps3_smp_topo(platform_t); #endif static void ps3_reset(platform_t); static void ps3_cpu_idle(sbintime_t); static platform_method_t ps3_methods[] = { PLATFORMMETHOD(platform_probe, ps3_probe), PLATFORMMETHOD(platform_attach, ps3_attach), PLATFORMMETHOD(platform_mem_regions, ps3_mem_regions), PLATFORMMETHOD(platform_real_maxaddr, ps3_real_maxaddr), PLATFORMMETHOD(platform_timebase_freq, ps3_timebase_freq), #ifdef SMP PLATFORMMETHOD(platform_smp_first_cpu, ps3_smp_first_cpu), PLATFORMMETHOD(platform_smp_next_cpu, ps3_smp_next_cpu), PLATFORMMETHOD(platform_smp_get_bsp, ps3_smp_get_bsp), PLATFORMMETHOD(platform_smp_start_cpu, ps3_smp_start_cpu), PLATFORMMETHOD(platform_smp_topo, ps3_smp_topo), #endif PLATFORMMETHOD(platform_reset, ps3_reset), PLATFORMMETHOD_END }; static platform_def_t ps3_platform = { "ps3", ps3_methods, 0 }; PLATFORM_DEF(ps3_platform); static int ps3_boot_pir = 0; static int ps3_probe(platform_t plat) { phandle_t root; char compatible[64]; root = OF_finddevice("/"); if (OF_getprop(root, "compatible", compatible, sizeof(compatible)) <= 0) return (BUS_PROBE_NOWILDCARD); if (strncmp(compatible, "sony,ps3", sizeof(compatible)) != 0) return (BUS_PROBE_NOWILDCARD); return (BUS_PROBE_SPECIFIC); } static int ps3_attach(platform_t plat) { pmap_mmu_install("mmu_ps3", BUS_PROBE_SPECIFIC); cpu_idle_hook = ps3_cpu_idle; /* Record our PIR at boot for later */ ps3_boot_pir = mfspr(SPR_PIR); return (0); } void ps3_mem_regions(platform_t plat, struct mem_region *phys, int *physsz, struct mem_region *avail_regions, int *availsz) { uint64_t lpar_id, junk; int i; /* Prefer device tree information if available */ if (OF_finddevice("/") != -1) { ofw_mem_regions(phys, physsz, avail_regions, availsz); } else { /* Real mode memory region is first segment */ phys[0].mr_start = 0; phys[0].mr_size = ps3_real_maxaddr(plat); *physsz = *availsz = 1; avail_regions[0] = phys[0]; } /* Now get extended memory region */ lv1_get_logical_partition_id(&lpar_id); lv1_get_repository_node_value(lpar_id, lv1_repository_string("bi") >> 32, lv1_repository_string("rgntotal"), 0, 0, &phys[*physsz].mr_size, &junk); for (i = 0; i < *physsz; i++) phys[*physsz].mr_size -= phys[i].mr_size; /* Convert to maximum amount we can allocate in 16 MB pages */ phys[*physsz].mr_size -= phys[*physsz].mr_size % (16*1024*1024); /* Allocate extended memory region */ lv1_allocate_memory(phys[*physsz].mr_size, 24 /* 16 MB pages */, 0, 0x04 /* any address */, &phys[*physsz].mr_start, &junk); avail_regions[*availsz] = phys[*physsz]; (*physsz)++; (*availsz)++; } static u_long ps3_timebase_freq(platform_t plat, struct cpuref *cpuref) { uint64_t ticks, node_id, junk; lv1_get_repository_node_value(PS3_LPAR_ID_PME, lv1_repository_string("be") >> 32, 0, 0, 0, &node_id, &junk); lv1_get_repository_node_value(PS3_LPAR_ID_PME, lv1_repository_string("be") >> 32, node_id, lv1_repository_string("clock"), 0, &ticks, &junk); return (ticks); } #ifdef SMP static int ps3_smp_first_cpu(platform_t plat, struct cpuref *cpuref) { cpuref->cr_cpuid = 0; cpuref->cr_hwref = ps3_boot_pir; return (0); } static int ps3_smp_next_cpu(platform_t plat, struct cpuref *cpuref) { if (cpuref->cr_cpuid >= 1) return (ENOENT); cpuref->cr_cpuid++; cpuref->cr_hwref = !ps3_boot_pir; return (0); } static int ps3_smp_get_bsp(platform_t plat, struct cpuref *cpuref) { cpuref->cr_cpuid = 0; cpuref->cr_hwref = ps3_boot_pir; return (0); } static int ps3_smp_start_cpu(platform_t plat, struct pcpu *pc) { /* kernel is spinning on 0x40 == -1 right now */ volatile uint32_t *secondary_spin_sem = (uint32_t *)PHYS_TO_DMAP((uintptr_t)0x40); int remote_pir = pc->pc_hwref; int timeout; ap_pcpu = pc; /* Try both PIR values, looping a few times: the HV likes moving us */ timeout = 10000; while (!pc->pc_awake && timeout--) { *secondary_spin_sem = remote_pir; powerpc_sync(); DELAY(100); remote_pir = !remote_pir; } return ((pc->pc_awake) ? 0 : EBUSY); } static struct cpu_group * ps3_smp_topo(platform_t plat) { + mp_ncores = 1; + smp_threads_per_core = 2; return (smp_topo_1level(CG_SHARE_L1, 2, CG_FLAG_SMT)); } #endif static void ps3_reset(platform_t plat) { lv1_panic(1); } static vm_offset_t ps3_real_maxaddr(platform_t plat) { uint64_t lpar_id, junk, ppe_id; static uint64_t rm_maxaddr = 0; if (rm_maxaddr == 0) { /* Get real mode memory region */ lv1_get_logical_partition_id(&lpar_id); lv1_get_logical_ppe_id(&ppe_id); lv1_get_repository_node_value(lpar_id, lv1_repository_string("bi") >> 32, lv1_repository_string("pu"), ppe_id, lv1_repository_string("rm_size"), &rm_maxaddr, &junk); } return (rm_maxaddr); } static void ps3_cpu_idle(sbintime_t sbt) { lv1_pause(0); } Index: stable/12/sys/powerpc/pseries/platform_chrp.c =================================================================== --- stable/12/sys/powerpc/pseries/platform_chrp.c (revision 354072) +++ stable/12/sys/powerpc/pseries/platform_chrp.c (revision 354073) @@ -1,576 +1,579 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2008 Marcel Moolenaar * Copyright (c) 2009 Nathan Whitehorn * 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 ``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 #include #include #include "platform_if.h" #ifdef SMP extern void *ap_pcpu; #endif #ifdef __powerpc64__ static uint8_t splpar_vpa[MAXCPU][640] __aligned(128); /* XXX: dpcpu */ #endif static vm_offset_t realmaxaddr = VM_MAX_ADDRESS; static int chrp_probe(platform_t); static int chrp_attach(platform_t); void chrp_mem_regions(platform_t, struct mem_region *phys, int *physsz, struct mem_region *avail, int *availsz); static vm_offset_t chrp_real_maxaddr(platform_t); static u_long chrp_timebase_freq(platform_t, struct cpuref *cpuref); static int chrp_smp_first_cpu(platform_t, struct cpuref *cpuref); static int chrp_smp_next_cpu(platform_t, struct cpuref *cpuref); static int chrp_smp_get_bsp(platform_t, struct cpuref *cpuref); static void chrp_smp_ap_init(platform_t); static int chrp_cpuref_init(void); #ifdef SMP static int chrp_smp_start_cpu(platform_t, struct pcpu *cpu); static struct cpu_group *chrp_smp_topo(platform_t plat); #endif static void chrp_reset(platform_t); #ifdef __powerpc64__ #include "phyp-hvcall.h" static void phyp_cpu_idle(sbintime_t sbt); #endif static struct cpuref platform_cpuref[MAXCPU]; static int platform_cpuref_cnt; static int platform_cpuref_valid; static platform_method_t chrp_methods[] = { PLATFORMMETHOD(platform_probe, chrp_probe), PLATFORMMETHOD(platform_attach, chrp_attach), PLATFORMMETHOD(platform_mem_regions, chrp_mem_regions), PLATFORMMETHOD(platform_real_maxaddr, chrp_real_maxaddr), PLATFORMMETHOD(platform_timebase_freq, chrp_timebase_freq), PLATFORMMETHOD(platform_smp_ap_init, chrp_smp_ap_init), PLATFORMMETHOD(platform_smp_first_cpu, chrp_smp_first_cpu), PLATFORMMETHOD(platform_smp_next_cpu, chrp_smp_next_cpu), PLATFORMMETHOD(platform_smp_get_bsp, chrp_smp_get_bsp), #ifdef SMP PLATFORMMETHOD(platform_smp_start_cpu, chrp_smp_start_cpu), PLATFORMMETHOD(platform_smp_topo, chrp_smp_topo), #endif PLATFORMMETHOD(platform_reset, chrp_reset), { 0, 0 } }; static platform_def_t chrp_platform = { "chrp", chrp_methods, 0 }; PLATFORM_DEF(chrp_platform); static int chrp_probe(platform_t plat) { if (OF_finddevice("/memory") != -1 || OF_finddevice("/memory@0") != -1) return (BUS_PROBE_GENERIC); return (ENXIO); } static int chrp_attach(platform_t plat) { #ifdef __powerpc64__ int i; /* XXX: check for /rtas/ibm,hypertas-functions? */ if (!(mfmsr() & PSL_HV)) { struct mem_region *phys, *avail; int nphys, navail; mem_regions(&phys, &nphys, &avail, &navail); realmaxaddr = phys[0].mr_size; pmap_mmu_install("mmu_phyp", BUS_PROBE_SPECIFIC); cpu_idle_hook = phyp_cpu_idle; /* Set up important VPA fields */ for (i = 0; i < MAXCPU; i++) { /* First two: VPA size */ splpar_vpa[i][4] = (uint8_t)((sizeof(splpar_vpa[i]) >> 8) & 0xff); splpar_vpa[i][5] = (uint8_t)(sizeof(splpar_vpa[i]) & 0xff); splpar_vpa[i][0xba] = 1; /* Maintain FPRs */ splpar_vpa[i][0xbb] = 1; /* Maintain PMCs */ splpar_vpa[i][0xfc] = 0xff; /* Maintain full SLB */ splpar_vpa[i][0xfd] = 0xff; splpar_vpa[i][0xff] = 1; /* Maintain Altivec */ } mb(); /* Set up hypervisor CPU stuff */ chrp_smp_ap_init(plat); } #endif chrp_cpuref_init(); /* Some systems (e.g. QEMU) need Open Firmware to stand down */ ofw_quiesce(); return (0); } static int parse_drconf_memory(struct mem_region *ofmem, int *msz, struct mem_region *ofavail, int *asz) { phandle_t phandle; vm_offset_t base; int i, idx, len, lasz, lmsz, res; uint32_t flags, lmb_size[2]; uint32_t *dmem; lmsz = *msz; lasz = *asz; phandle = OF_finddevice("/ibm,dynamic-reconfiguration-memory"); if (phandle == -1) /* No drconf node, return. */ return (0); res = OF_getencprop(phandle, "ibm,lmb-size", lmb_size, sizeof(lmb_size)); if (res == -1) return (0); printf("Logical Memory Block size: %d MB\n", lmb_size[1] >> 20); /* Parse the /ibm,dynamic-memory. The first position gives the # of entries. The next two words reflect the address of the memory block. The next four words are the DRC index, reserved, list index and flags. (see PAPR C.6.6.2 ibm,dynamic-reconfiguration-memory) #el Addr DRC-idx res list-idx flags ------------------------------------------------- | 4 | 8 | 4 | 4 | 4 | 4 |.... ------------------------------------------------- */ len = OF_getproplen(phandle, "ibm,dynamic-memory"); if (len > 0) { /* We have to use a variable length array on the stack since we have very limited stack space. */ cell_t arr[len/sizeof(cell_t)]; res = OF_getencprop(phandle, "ibm,dynamic-memory", arr, sizeof(arr)); if (res == -1) return (0); /* Number of elements */ idx = arr[0]; /* First address, in arr[1], arr[2]*/ dmem = &arr[1]; for (i = 0; i < idx; i++) { base = ((uint64_t)dmem[0] << 32) + dmem[1]; dmem += 4; flags = dmem[1]; /* Use region only if available and not reserved. */ if ((flags & 0x8) && !(flags & 0x80)) { ofmem[lmsz].mr_start = base; ofmem[lmsz].mr_size = (vm_size_t)lmb_size[1]; ofavail[lasz].mr_start = base; ofavail[lasz].mr_size = (vm_size_t)lmb_size[1]; lmsz++; lasz++; } dmem += 2; } } *msz = lmsz; *asz = lasz; return (1); } void chrp_mem_regions(platform_t plat, struct mem_region *phys, int *physsz, struct mem_region *avail, int *availsz) { vm_offset_t maxphysaddr; int i; ofw_mem_regions(phys, physsz, avail, availsz); parse_drconf_memory(phys, physsz, avail, availsz); /* * On some firmwares (SLOF), some memory may be marked available that * doesn't actually exist. This manifests as an extension of the last * available segment past the end of physical memory, so truncate that * one. */ maxphysaddr = 0; for (i = 0; i < *physsz; i++) if (phys[i].mr_start + phys[i].mr_size > maxphysaddr) maxphysaddr = phys[i].mr_start + phys[i].mr_size; for (i = 0; i < *availsz; i++) if (avail[i].mr_start + avail[i].mr_size > maxphysaddr) avail[i].mr_size = maxphysaddr - avail[i].mr_start; } static vm_offset_t chrp_real_maxaddr(platform_t plat) { return (realmaxaddr); } static u_long chrp_timebase_freq(platform_t plat, struct cpuref *cpuref) { phandle_t cpus, cpunode; int32_t ticks = -1; int res; char buf[8]; cpus = OF_finddevice("/cpus"); if (cpus == -1) panic("CPU tree not found on Open Firmware\n"); for (cpunode = OF_child(cpus); cpunode != 0; cpunode = OF_peer(cpunode)) { res = OF_getprop(cpunode, "device_type", buf, sizeof(buf)); if (res > 0 && strcmp(buf, "cpu") == 0) break; } if (cpunode <= 0) panic("CPU node not found on Open Firmware\n"); OF_getencprop(cpunode, "timebase-frequency", &ticks, sizeof(ticks)); if (ticks <= 0) panic("Unable to determine timebase frequency!"); return (ticks); } static int chrp_smp_first_cpu(platform_t plat, struct cpuref *cpuref) { if (platform_cpuref_valid == 0) return (EINVAL); cpuref->cr_cpuid = 0; cpuref->cr_hwref = platform_cpuref[0].cr_hwref; return (0); } static int chrp_smp_next_cpu(platform_t plat, struct cpuref *cpuref) { int id; if (platform_cpuref_valid == 0) return (EINVAL); id = cpuref->cr_cpuid + 1; if (id >= platform_cpuref_cnt) return (ENOENT); cpuref->cr_cpuid = platform_cpuref[id].cr_cpuid; cpuref->cr_hwref = platform_cpuref[id].cr_hwref; return (0); } static int chrp_smp_get_bsp(platform_t plat, struct cpuref *cpuref) { cpuref->cr_cpuid = platform_cpuref[0].cr_cpuid; cpuref->cr_hwref = platform_cpuref[0].cr_hwref; return (0); } static void get_cpu_reg(phandle_t cpu, cell_t *reg) { int res; res = OF_getproplen(cpu, "reg"); if (res != sizeof(cell_t)) panic("Unexpected length for CPU property reg on Open Firmware\n"); OF_getencprop(cpu, "reg", reg, res); } static int chrp_cpuref_init(void) { phandle_t cpu, dev, chosen, pbsp; ihandle_t ibsp; char buf[32]; int a, bsp, res, res2, tmp_cpuref_cnt; static struct cpuref tmp_cpuref[MAXCPU]; cell_t interrupt_servers[32], addr_cells, size_cells, reg, bsp_reg; if (platform_cpuref_valid) return (0); dev = OF_peer(0); dev = OF_child(dev); while (dev != 0) { res = OF_getprop(dev, "name", buf, sizeof(buf)); if (res > 0 && strcmp(buf, "cpus") == 0) break; dev = OF_peer(dev); } /* Make sure that cpus reg property have 1 address cell and 0 size cells */ res = OF_getproplen(dev, "#address-cells"); res2 = OF_getproplen(dev, "#size-cells"); if (res != res2 || res != sizeof(cell_t)) panic("CPU properties #address-cells and #size-cells not found on Open Firmware\n"); OF_getencprop(dev, "#address-cells", &addr_cells, sizeof(addr_cells)); OF_getencprop(dev, "#size-cells", &size_cells, sizeof(size_cells)); if (addr_cells != 1 || size_cells != 0) panic("Unexpected values for CPU properties #address-cells and #size-cells on Open Firmware\n"); /* Look for boot CPU in /chosen/cpu and /chosen/fdtbootcpu */ chosen = OF_finddevice("/chosen"); if (chosen == -1) panic("Device /chosen not found on Open Firmware\n"); bsp_reg = -1; /* /chosen/cpu */ if (OF_getproplen(chosen, "cpu") == sizeof(ihandle_t)) { OF_getprop(chosen, "cpu", &ibsp, sizeof(ibsp)); pbsp = OF_instance_to_package(ibsp); if (pbsp != -1) get_cpu_reg(pbsp, &bsp_reg); } /* /chosen/fdtbootcpu */ if (bsp_reg == -1) { if (OF_getproplen(chosen, "fdtbootcpu") == sizeof(cell_t)) OF_getprop(chosen, "fdtbootcpu", &bsp_reg, sizeof(bsp_reg)); } if (bsp_reg == -1) panic("Boot CPU not found on Open Firmware\n"); bsp = -1; tmp_cpuref_cnt = 0; for (cpu = OF_child(dev); cpu != 0; cpu = OF_peer(cpu)) { res = OF_getprop(cpu, "device_type", buf, sizeof(buf)); if (res > 0 && strcmp(buf, "cpu") == 0) { res = OF_getproplen(cpu, "ibm,ppc-interrupt-server#s"); if (res > 0) { OF_getencprop(cpu, "ibm,ppc-interrupt-server#s", interrupt_servers, res); get_cpu_reg(cpu, ®); if (reg == bsp_reg) bsp = tmp_cpuref_cnt; for (a = 0; a < res/sizeof(cell_t); a++) { tmp_cpuref[tmp_cpuref_cnt].cr_hwref = interrupt_servers[a]; tmp_cpuref[tmp_cpuref_cnt].cr_cpuid = tmp_cpuref_cnt; tmp_cpuref_cnt++; } } } } if (bsp == -1) panic("Boot CPU not found\n"); /* Map IDs, so BSP has CPUID 0 regardless of hwref */ for (a = bsp; a < tmp_cpuref_cnt; a++) { platform_cpuref[platform_cpuref_cnt].cr_hwref = tmp_cpuref[a].cr_hwref; platform_cpuref[platform_cpuref_cnt].cr_cpuid = platform_cpuref_cnt; platform_cpuref_cnt++; } for (a = 0; a < bsp; a++) { platform_cpuref[platform_cpuref_cnt].cr_hwref = tmp_cpuref[a].cr_hwref; platform_cpuref[platform_cpuref_cnt].cr_cpuid = platform_cpuref_cnt; platform_cpuref_cnt++; } platform_cpuref_valid = 1; return (0); } #ifdef SMP static int chrp_smp_start_cpu(platform_t plat, struct pcpu *pc) { cell_t start_cpu; int result, err, timeout; if (!rtas_exists()) { printf("RTAS uninitialized: unable to start AP %d\n", pc->pc_cpuid); return (ENXIO); } start_cpu = rtas_token_lookup("start-cpu"); if (start_cpu == -1) { printf("RTAS unknown method: unable to start AP %d\n", pc->pc_cpuid); return (ENXIO); } ap_pcpu = pc; powerpc_sync(); result = rtas_call_method(start_cpu, 3, 1, pc->pc_hwref, EXC_RST, pc, &err); if (result < 0 || err != 0) { printf("RTAS error (%d/%d): unable to start AP %d\n", result, err, pc->pc_cpuid); return (ENXIO); } timeout = 10000; while (!pc->pc_awake && timeout--) DELAY(100); return ((pc->pc_awake) ? 0 : EBUSY); } static struct cpu_group * chrp_smp_topo(platform_t plat) { struct pcpu *pc, *last_pc; int i, ncores, ncpus; ncores = ncpus = 0; last_pc = NULL; for (i = 0; i <= mp_maxid; i++) { pc = pcpu_find(i); if (pc == NULL) continue; if (last_pc == NULL || pc->pc_hwref != last_pc->pc_hwref) ncores++; last_pc = pc; ncpus++; } + mp_ncores = ncores; + if (ncpus % ncores != 0) { printf("WARNING: Irregular SMP topology. Performance may be " "suboptimal (%d CPUS, %d cores)\n", ncpus, ncores); return (smp_topo_none()); } /* Don't do anything fancier for non-threaded SMP */ if (ncpus == ncores) return (smp_topo_none()); + smp_threads_per_core = ncpus / ncores; return (smp_topo_1level(CG_SHARE_L1, ncpus / ncores, CG_FLAG_SMT)); } #endif static void chrp_reset(platform_t platform) { OF_reboot(); } #ifdef __powerpc64__ static void phyp_cpu_idle(sbintime_t sbt) { register_t msr; msr = mfmsr(); mtmsr(msr & ~PSL_EE); if (sched_runnable()) { mtmsr(msr); return; } phyp_hcall(H_CEDE); /* Re-enables interrupts internally */ mtmsr(msr); } static void chrp_smp_ap_init(platform_t platform) { if (!(mfmsr() & PSL_HV)) { /* Register VPA */ phyp_hcall(H_REGISTER_VPA, 1UL, PCPU_GET(hwref), splpar_vpa[PCPU_GET(hwref)]); /* Set interrupt priority */ phyp_hcall(H_CPPR, 0xff); } } #else static void chrp_smp_ap_init(platform_t platform) { } #endif Index: stable/12/sys/sys/smp.h =================================================================== --- stable/12/sys/sys/smp.h (revision 354072) +++ stable/12/sys/sys/smp.h (revision 354073) @@ -1,277 +1,279 @@ /*- * SPDX-License-Identifier: Beerware * * ---------------------------------------------------------------------------- * "THE BEER-WARE LICENSE" (Revision 42): * wrote this file. As long as you retain this notice you * can do whatever you want with this stuff. If we meet some day, and you think * this stuff is worth it, you can buy me a beer in return. Poul-Henning Kamp * ---------------------------------------------------------------------------- * * $FreeBSD$ */ #ifndef _SYS_SMP_H_ #define _SYS_SMP_H_ #ifdef _KERNEL #ifndef LOCORE #include #include /* * Types of nodes in the topological tree. */ typedef enum { /* No node has this type; can be used in topo API calls. */ TOPO_TYPE_DUMMY, /* Processing unit aka computing unit aka logical CPU. */ TOPO_TYPE_PU, /* Physical subdivision of a package. */ TOPO_TYPE_CORE, /* CPU L1/L2/L3 cache. */ TOPO_TYPE_CACHE, /* Package aka chip, equivalent to socket. */ TOPO_TYPE_PKG, /* NUMA node. */ TOPO_TYPE_NODE, /* Other logical or physical grouping of PUs. */ /* E.g. PUs on the same dye, or PUs sharing an FPU. */ TOPO_TYPE_GROUP, /* The whole system. */ TOPO_TYPE_SYSTEM } topo_node_type; /* Hardware indenitifier of a topology component. */ typedef unsigned int hwid_t; /* Logical CPU idenitifier. */ typedef int cpuid_t; /* A node in the topology. */ struct topo_node { struct topo_node *parent; TAILQ_HEAD(topo_children, topo_node) children; TAILQ_ENTRY(topo_node) siblings; cpuset_t cpuset; topo_node_type type; uintptr_t subtype; hwid_t hwid; cpuid_t id; int nchildren; int cpu_count; }; /* * Scheduling topology of a NUMA or SMP system. * * The top level topology is an array of pointers to groups. Each group * contains a bitmask of cpus in its group or subgroups. It may also * contain a pointer to an array of child groups. * * The bitmasks at non leaf groups may be used by consumers who support * a smaller depth than the hardware provides. * * The topology may be omitted by systems where all CPUs are equal. */ struct cpu_group { struct cpu_group *cg_parent; /* Our parent group. */ struct cpu_group *cg_child; /* Optional children groups. */ cpuset_t cg_mask; /* Mask of cpus in this group. */ int32_t cg_count; /* Count of cpus in this group. */ int16_t cg_children; /* Number of children groups. */ int8_t cg_level; /* Shared cache level. */ int8_t cg_flags; /* Traversal modifiers. */ }; typedef struct cpu_group *cpu_group_t; /* * Defines common resources for CPUs in the group. The highest level * resource should be used when multiple are shared. */ #define CG_SHARE_NONE 0 #define CG_SHARE_L1 1 #define CG_SHARE_L2 2 #define CG_SHARE_L3 3 #define MAX_CACHE_LEVELS CG_SHARE_L3 /* * Behavior modifiers for load balancing and affinity. */ #define CG_FLAG_HTT 0x01 /* Schedule the alternate core last. */ #define CG_FLAG_SMT 0x02 /* New age htt, less crippled. */ #define CG_FLAG_THREAD (CG_FLAG_HTT | CG_FLAG_SMT) /* Any threading. */ /* * Convenience routines for building and traversing topologies. */ #ifdef SMP void topo_init_node(struct topo_node *node); void topo_init_root(struct topo_node *root); struct topo_node * topo_add_node_by_hwid(struct topo_node *parent, int hwid, topo_node_type type, uintptr_t subtype); struct topo_node * topo_find_node_by_hwid(struct topo_node *parent, int hwid, topo_node_type type, uintptr_t subtype); void topo_promote_child(struct topo_node *child); struct topo_node * topo_next_node(struct topo_node *top, struct topo_node *node); struct topo_node * topo_next_nonchild_node(struct topo_node *top, struct topo_node *node); void topo_set_pu_id(struct topo_node *node, cpuid_t id); enum topo_level { TOPO_LEVEL_PKG = 0, /* * Some systems have useful sub-package core organizations. On these, * a package has one or more subgroups. Each subgroup contains one or * more cache groups (cores that share a last level cache). */ TOPO_LEVEL_GROUP, TOPO_LEVEL_CACHEGROUP, TOPO_LEVEL_CORE, TOPO_LEVEL_THREAD, TOPO_LEVEL_COUNT /* Must be last */ }; struct topo_analysis { int entities[TOPO_LEVEL_COUNT]; }; int topo_analyze(struct topo_node *topo_root, int all, struct topo_analysis *results); #define TOPO_FOREACH(i, root) \ for (i = root; i != NULL; i = topo_next_node(root, i)) struct cpu_group *smp_topo(void); struct cpu_group *smp_topo_alloc(u_int count); struct cpu_group *smp_topo_none(void); struct cpu_group *smp_topo_1level(int l1share, int l1count, int l1flags); struct cpu_group *smp_topo_2level(int l2share, int l2count, int l1share, int l1count, int l1flags); struct cpu_group *smp_topo_find(struct cpu_group *top, int cpu); extern void (*cpustop_restartfunc)(void); extern int smp_cpus; /* The suspend/resume cpusets are x86 only, but minimize ifdefs. */ extern volatile cpuset_t resuming_cpus; /* woken up cpus in suspend pen */ extern volatile cpuset_t started_cpus; /* cpus to let out of stop pen */ extern volatile cpuset_t stopped_cpus; /* cpus in stop pen */ extern volatile cpuset_t suspended_cpus; /* cpus [near] sleeping in susp pen */ extern volatile cpuset_t toresume_cpus; /* cpus to let out of suspend pen */ extern cpuset_t hlt_cpus_mask; /* XXX 'mask' is detail in old impl */ extern cpuset_t logical_cpus_mask; #endif /* SMP */ extern u_int mp_maxid; extern int mp_maxcpus; +extern int mp_ncores; extern int mp_ncpus; extern volatile int smp_started; +extern int smp_threads_per_core; extern cpuset_t all_cpus; extern cpuset_t cpuset_domain[MAXMEMDOM]; /* CPUs in each NUMA domain. */ /* * Macro allowing us to determine whether a CPU is absent at any given * time, thus permitting us to configure sparse maps of cpuid-dependent * (per-CPU) structures. */ #define CPU_ABSENT(x_cpu) (!CPU_ISSET(x_cpu, &all_cpus)) /* * Macros to iterate over non-absent CPUs. CPU_FOREACH() takes an * integer iterator and iterates over the available set of CPUs. * CPU_FIRST() returns the id of the first non-absent CPU. CPU_NEXT() * returns the id of the next non-absent CPU. It will wrap back to * CPU_FIRST() once the end of the list is reached. The iterators are * currently implemented via inline functions. */ #define CPU_FOREACH(i) \ for ((i) = 0; (i) <= mp_maxid; (i)++) \ if (!CPU_ABSENT((i))) static __inline int cpu_first(void) { int i; for (i = 0;; i++) if (!CPU_ABSENT(i)) return (i); } static __inline int cpu_next(int i) { for (;;) { i++; if (i > mp_maxid) i = 0; if (!CPU_ABSENT(i)) return (i); } } #define CPU_FIRST() cpu_first() #define CPU_NEXT(i) cpu_next((i)) #ifdef SMP /* * Machine dependent functions used to initialize MP support. * * The cpu_mp_probe() should check to see if MP support is present and return * zero if it is not or non-zero if it is. If MP support is present, then * cpu_mp_start() will be called so that MP can be enabled. This function * should do things such as startup secondary processors. It should also * setup mp_ncpus, all_cpus, and smp_cpus. It should also ensure that * smp_started is initialized at the appropriate time. * Once cpu_mp_start() returns, machine independent MP startup code will be * executed and a simple message will be output to the console. Finally, * cpu_mp_announce() will be called so that machine dependent messages about * the MP support may be output to the console if desired. * * The cpu_setmaxid() function is called very early during the boot process * so that the MD code may set mp_maxid to provide an upper bound on CPU IDs * that other subsystems may use. If a platform is not able to determine * the exact maximum ID that early, then it may set mp_maxid to MAXCPU - 1. */ struct thread; struct cpu_group *cpu_topo(void); void cpu_mp_announce(void); int cpu_mp_probe(void); void cpu_mp_setmaxid(void); void cpu_mp_start(void); void forward_signal(struct thread *); int restart_cpus(cpuset_t); int stop_cpus(cpuset_t); int stop_cpus_hard(cpuset_t); #if defined(__amd64__) || defined(__i386__) int suspend_cpus(cpuset_t); int resume_cpus(cpuset_t); #endif void smp_rendezvous_action(void); extern struct mtx smp_ipi_mtx; #endif /* SMP */ int quiesce_all_cpus(const char *, int); int quiesce_cpus(cpuset_t, const char *, int); void smp_no_rendezvous_barrier(void *); void smp_rendezvous(void (*)(void *), void (*)(void *), void (*)(void *), void *arg); void smp_rendezvous_cpus(cpuset_t, void (*)(void *), void (*)(void *), void (*)(void *), void *arg); #endif /* !LOCORE */ #endif /* _KERNEL */ #endif /* _SYS_SMP_H_ */ Index: stable/12/sys/x86/x86/mp_x86.c =================================================================== --- stable/12/sys/x86/x86/mp_x86.c (revision 354072) +++ stable/12/sys/x86/x86/mp_x86.c (revision 354073) @@ -1,1782 +1,1790 @@ /*- * Copyright (c) 1996, by Steve Passe * Copyright (c) 2003, by Peter Wemm * 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. The name of the developer may NOT 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. */ #include __FBSDID("$FreeBSD$"); #ifdef __i386__ #include "opt_apic.h" #endif #include "opt_cpu.h" #include "opt_kstack_pages.h" #include "opt_pmap.h" #include "opt_sched.h" #include "opt_smp.h" #include #include #include #include /* cngetc() */ #include #ifdef GPROF #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 #include #include #include #include #include static MALLOC_DEFINE(M_CPUS, "cpus", "CPU items"); /* lock region used by kernel profiling */ int mcount_lock; int mp_naps; /* # of Applications processors */ int boot_cpu_id = -1; /* designated BSP */ /* AP uses this during bootstrap. Do not staticize. */ char *bootSTK; int bootAP; /* Free these after use */ void *bootstacks[MAXCPU]; void *dpcpu; struct pcb stoppcbs[MAXCPU]; struct susppcb **susppcbs; #ifdef COUNT_IPIS /* Interrupt counts. */ static u_long *ipi_preempt_counts[MAXCPU]; static u_long *ipi_ast_counts[MAXCPU]; u_long *ipi_invltlb_counts[MAXCPU]; u_long *ipi_invlrng_counts[MAXCPU]; u_long *ipi_invlpg_counts[MAXCPU]; u_long *ipi_invlcache_counts[MAXCPU]; u_long *ipi_rendezvous_counts[MAXCPU]; static u_long *ipi_hardclock_counts[MAXCPU]; #endif /* Default cpu_ops implementation. */ struct cpu_ops cpu_ops; /* * Local data and functions. */ static volatile cpuset_t ipi_stop_nmi_pending; volatile cpuset_t resuming_cpus; volatile cpuset_t toresume_cpus; /* used to hold the AP's until we are ready to release them */ struct mtx ap_boot_mtx; /* Set to 1 once we're ready to let the APs out of the pen. */ volatile int aps_ready = 0; /* * Store data from cpu_add() until later in the boot when we actually setup * the APs. */ struct cpu_info *cpu_info; int *apic_cpuids; int cpu_apic_ids[MAXCPU]; _Static_assert(MAXCPU <= MAX_APIC_ID, "MAXCPU cannot be larger that MAX_APIC_ID"); _Static_assert(xAPIC_MAX_APIC_ID <= MAX_APIC_ID, "xAPIC_MAX_APIC_ID cannot be larger that MAX_APIC_ID"); /* Holds pending bitmap based IPIs per CPU */ volatile u_int cpu_ipi_pending[MAXCPU]; static void release_aps(void *dummy); static void cpustop_handler_post(u_int cpu); static int hyperthreading_allowed = 1; SYSCTL_INT(_machdep, OID_AUTO, hyperthreading_allowed, CTLFLAG_RDTUN, &hyperthreading_allowed, 0, "Use Intel HTT logical CPUs"); static struct topo_node topo_root; static int pkg_id_shift; static int node_id_shift; static int core_id_shift; static int disabled_cpus; struct cache_info { int id_shift; int present; } static caches[MAX_CACHE_LEVELS]; unsigned int boot_address; #define MiB(v) (v ## ULL << 20) void mem_range_AP_init(void) { if (mem_range_softc.mr_op && mem_range_softc.mr_op->initAP) mem_range_softc.mr_op->initAP(&mem_range_softc); } /* * Round up to the next power of two, if necessary, and then * take log2. * Returns -1 if argument is zero. */ static __inline int mask_width(u_int x) { return (fls(x << (1 - powerof2(x))) - 1); } /* * Add a cache level to the cache topology description. */ static int add_deterministic_cache(int type, int level, int share_count) { if (type == 0) return (0); if (type > 3) { printf("unexpected cache type %d\n", type); return (1); } if (type == 2) /* ignore instruction cache */ return (1); if (level == 0 || level > MAX_CACHE_LEVELS) { printf("unexpected cache level %d\n", type); return (1); } if (caches[level - 1].present) { printf("WARNING: multiple entries for L%u data cache\n", level); printf("%u => %u\n", caches[level - 1].id_shift, mask_width(share_count)); } caches[level - 1].id_shift = mask_width(share_count); caches[level - 1].present = 1; if (caches[level - 1].id_shift > pkg_id_shift) { printf("WARNING: L%u data cache covers more " "APIC IDs than a package (%u > %u)\n", level, caches[level - 1].id_shift, pkg_id_shift); caches[level - 1].id_shift = pkg_id_shift; } if (caches[level - 1].id_shift < core_id_shift) { printf("WARNING: L%u data cache covers fewer " "APIC IDs than a core (%u < %u)\n", level, caches[level - 1].id_shift, core_id_shift); caches[level - 1].id_shift = core_id_shift; } return (1); } /* * Determine topology of processing units and caches for AMD CPUs. * See: * - AMD CPUID Specification (Publication # 25481) * - BKDG for AMD NPT Family 0Fh Processors (Publication # 32559) * - BKDG For AMD Family 10h Processors (Publication # 31116) * - BKDG For AMD Family 15h Models 00h-0Fh Processors (Publication # 42301) * - BKDG For AMD Family 16h Models 00h-0Fh Processors (Publication # 48751) * - PPR For AMD Family 17h Models 00h-0Fh Processors (Publication # 54945) */ static void topo_probe_amd(void) { u_int p[4]; uint64_t v; int level; int nodes_per_socket; int share_count; int type; int i; /* No multi-core capability. */ if ((amd_feature2 & AMDID2_CMP) == 0) return; /* For families 10h and newer. */ pkg_id_shift = (cpu_procinfo2 & AMDID_COREID_SIZE) >> AMDID_COREID_SIZE_SHIFT; /* For 0Fh family. */ if (pkg_id_shift == 0) pkg_id_shift = mask_width((cpu_procinfo2 & AMDID_CMP_CORES) + 1); /* * Families prior to 16h define the following value as * cores per compute unit and we don't really care about the AMD * compute units at the moment. Perhaps we should treat them as * cores and cores within the compute units as hardware threads, * but that's up for debate. * Later families define the value as threads per compute unit, * so we are following AMD's nomenclature here. */ if ((amd_feature2 & AMDID2_TOPOLOGY) != 0 && CPUID_TO_FAMILY(cpu_id) >= 0x16) { cpuid_count(0x8000001e, 0, p); share_count = ((p[1] >> 8) & 0xff) + 1; core_id_shift = mask_width(share_count); /* * For Zen (17h), gather Nodes per Processor. Each node is a * Zeppelin die; TR and EPYC CPUs will have multiple dies per * package. Communication latency between dies is higher than * within them. */ nodes_per_socket = ((p[2] >> 8) & 0x7) + 1; node_id_shift = pkg_id_shift - mask_width(nodes_per_socket); } if ((amd_feature2 & AMDID2_TOPOLOGY) != 0) { for (i = 0; ; i++) { cpuid_count(0x8000001d, i, p); type = p[0] & 0x1f; level = (p[0] >> 5) & 0x7; share_count = 1 + ((p[0] >> 14) & 0xfff); if (!add_deterministic_cache(type, level, share_count)) break; } } else { if (cpu_exthigh >= 0x80000005) { cpuid_count(0x80000005, 0, p); if (((p[2] >> 24) & 0xff) != 0) { caches[0].id_shift = 0; caches[0].present = 1; } } if (cpu_exthigh >= 0x80000006) { cpuid_count(0x80000006, 0, p); if (((p[2] >> 16) & 0xffff) != 0) { caches[1].id_shift = 0; caches[1].present = 1; } if (((p[3] >> 18) & 0x3fff) != 0) { nodes_per_socket = 1; if ((amd_feature2 & AMDID2_NODE_ID) != 0) { /* * Handle multi-node processors that * have multiple chips, each with its * own L3 cache, on the same die. */ v = rdmsr(0xc001100c); nodes_per_socket = 1 + ((v >> 3) & 0x7); } caches[2].id_shift = pkg_id_shift - mask_width(nodes_per_socket); caches[2].present = 1; } } } } /* * Determine topology of processing units for Intel CPUs * using CPUID Leaf 1 and Leaf 4, if supported. * See: * - Intel 64 Architecture Processor Topology Enumeration * - Intel 64 and IA-32 ArchitecturesSoftware Developer’s Manual, * Volume 3A: System Programming Guide, PROGRAMMING CONSIDERATIONS * FOR HARDWARE MULTI-THREADING CAPABLE PROCESSORS */ static void topo_probe_intel_0x4(void) { u_int p[4]; int max_cores; int max_logical; /* Both zero and one here mean one logical processor per package. */ max_logical = (cpu_feature & CPUID_HTT) != 0 ? (cpu_procinfo & CPUID_HTT_CORES) >> 16 : 1; if (max_logical <= 1) return; if (cpu_high >= 0x4) { cpuid_count(0x04, 0, p); max_cores = ((p[0] >> 26) & 0x3f) + 1; } else max_cores = 1; core_id_shift = mask_width(max_logical/max_cores); KASSERT(core_id_shift >= 0, ("intel topo: max_cores > max_logical\n")); pkg_id_shift = core_id_shift + mask_width(max_cores); } /* * Determine topology of processing units for Intel CPUs * using CPUID Leaf 11, if supported. * See: * - Intel 64 Architecture Processor Topology Enumeration * - Intel 64 and IA-32 ArchitecturesSoftware Developer’s Manual, * Volume 3A: System Programming Guide, PROGRAMMING CONSIDERATIONS * FOR HARDWARE MULTI-THREADING CAPABLE PROCESSORS */ static void topo_probe_intel_0xb(void) { u_int p[4]; int bits; int type; int i; /* Fall back if CPU leaf 11 doesn't really exist. */ cpuid_count(0x0b, 0, p); if (p[1] == 0) { topo_probe_intel_0x4(); return; } /* We only support three levels for now. */ for (i = 0; ; i++) { cpuid_count(0x0b, i, p); bits = p[0] & 0x1f; type = (p[2] >> 8) & 0xff; if (type == 0) break; /* TODO: check for duplicate (re-)assignment */ if (type == CPUID_TYPE_SMT) core_id_shift = bits; else if (type == CPUID_TYPE_CORE) pkg_id_shift = bits; else printf("unknown CPU level type %d\n", type); } if (pkg_id_shift < core_id_shift) { printf("WARNING: core covers more APIC IDs than a package\n"); core_id_shift = pkg_id_shift; } } /* * Determine topology of caches for Intel CPUs. * See: * - Intel 64 Architecture Processor Topology Enumeration * - Intel 64 and IA-32 Architectures Software Developer’s Manual * Volume 2A: Instruction Set Reference, A-M, * CPUID instruction */ static void topo_probe_intel_caches(void) { u_int p[4]; int level; int share_count; int type; int i; if (cpu_high < 0x4) { /* * Available cache level and sizes can be determined * via CPUID leaf 2, but that requires a huge table of hardcoded * values, so for now just assume L1 and L2 caches potentially * shared only by HTT processing units, if HTT is present. */ caches[0].id_shift = pkg_id_shift; caches[0].present = 1; caches[1].id_shift = pkg_id_shift; caches[1].present = 1; return; } for (i = 0; ; i++) { cpuid_count(0x4, i, p); type = p[0] & 0x1f; level = (p[0] >> 5) & 0x7; share_count = 1 + ((p[0] >> 14) & 0xfff); if (!add_deterministic_cache(type, level, share_count)) break; } } /* * Determine topology of processing units and caches for Intel CPUs. * See: * - Intel 64 Architecture Processor Topology Enumeration */ static void topo_probe_intel(void) { /* * Note that 0x1 <= cpu_high < 4 case should be * compatible with topo_probe_intel_0x4() logic when * CPUID.1:EBX[23:16] > 0 (cpu_cores will be 1) * or it should trigger the fallback otherwise. */ if (cpu_high >= 0xb) topo_probe_intel_0xb(); else if (cpu_high >= 0x1) topo_probe_intel_0x4(); topo_probe_intel_caches(); } /* * Topology information is queried only on BSP, on which this * code runs and for which it can query CPUID information. * Then topology is extrapolated on all packages using an * assumption that APIC ID to hardware component ID mapping is * homogenious. * That doesn't necesserily imply that the topology is uniform. */ void topo_probe(void) { static int cpu_topo_probed = 0; struct x86_topo_layer { int type; int subtype; int id_shift; } topo_layers[MAX_CACHE_LEVELS + 4]; struct topo_node *parent; struct topo_node *node; int layer; int nlayers; int node_id; int i; if (cpu_topo_probed) return; CPU_ZERO(&logical_cpus_mask); if (mp_ncpus <= 1) ; /* nothing */ else if (cpu_vendor_id == CPU_VENDOR_AMD) topo_probe_amd(); else if (cpu_vendor_id == CPU_VENDOR_INTEL) topo_probe_intel(); KASSERT(pkg_id_shift >= core_id_shift, ("bug in APIC topology discovery")); nlayers = 0; bzero(topo_layers, sizeof(topo_layers)); topo_layers[nlayers].type = TOPO_TYPE_PKG; topo_layers[nlayers].id_shift = pkg_id_shift; if (bootverbose) printf("Package ID shift: %u\n", topo_layers[nlayers].id_shift); nlayers++; if (pkg_id_shift > node_id_shift && node_id_shift != 0) { topo_layers[nlayers].type = TOPO_TYPE_GROUP; topo_layers[nlayers].id_shift = node_id_shift; if (bootverbose) printf("Node ID shift: %u\n", topo_layers[nlayers].id_shift); nlayers++; } /* * Consider all caches to be within a package/chip * and "in front" of all sub-components like * cores and hardware threads. */ for (i = MAX_CACHE_LEVELS - 1; i >= 0; --i) { if (caches[i].present) { if (node_id_shift != 0) KASSERT(caches[i].id_shift <= node_id_shift, ("bug in APIC topology discovery")); KASSERT(caches[i].id_shift <= pkg_id_shift, ("bug in APIC topology discovery")); KASSERT(caches[i].id_shift >= core_id_shift, ("bug in APIC topology discovery")); topo_layers[nlayers].type = TOPO_TYPE_CACHE; topo_layers[nlayers].subtype = i + 1; topo_layers[nlayers].id_shift = caches[i].id_shift; if (bootverbose) printf("L%u cache ID shift: %u\n", topo_layers[nlayers].subtype, topo_layers[nlayers].id_shift); nlayers++; } } if (pkg_id_shift > core_id_shift) { topo_layers[nlayers].type = TOPO_TYPE_CORE; topo_layers[nlayers].id_shift = core_id_shift; if (bootverbose) printf("Core ID shift: %u\n", topo_layers[nlayers].id_shift); nlayers++; } topo_layers[nlayers].type = TOPO_TYPE_PU; topo_layers[nlayers].id_shift = 0; nlayers++; topo_init_root(&topo_root); for (i = 0; i <= max_apic_id; ++i) { if (!cpu_info[i].cpu_present) continue; parent = &topo_root; for (layer = 0; layer < nlayers; ++layer) { node_id = i >> topo_layers[layer].id_shift; parent = topo_add_node_by_hwid(parent, node_id, topo_layers[layer].type, topo_layers[layer].subtype); } } parent = &topo_root; for (layer = 0; layer < nlayers; ++layer) { node_id = boot_cpu_id >> topo_layers[layer].id_shift; node = topo_find_node_by_hwid(parent, node_id, topo_layers[layer].type, topo_layers[layer].subtype); topo_promote_child(node); parent = node; } cpu_topo_probed = 1; } /* * Assign logical CPU IDs to local APICs. */ void assign_cpu_ids(void) { struct topo_node *node; u_int smt_mask; + int nhyper; smt_mask = (1u << core_id_shift) - 1; /* * Assign CPU IDs to local APIC IDs and disable any CPUs * beyond MAXCPU. CPU 0 is always assigned to the BSP. */ mp_ncpus = 0; + nhyper = 0; TOPO_FOREACH(node, &topo_root) { if (node->type != TOPO_TYPE_PU) continue; if ((node->hwid & smt_mask) != (boot_cpu_id & smt_mask)) cpu_info[node->hwid].cpu_hyperthread = 1; if (resource_disabled("lapic", node->hwid)) { if (node->hwid != boot_cpu_id) cpu_info[node->hwid].cpu_disabled = 1; else printf("Cannot disable BSP, APIC ID = %d\n", node->hwid); } if (!hyperthreading_allowed && cpu_info[node->hwid].cpu_hyperthread) cpu_info[node->hwid].cpu_disabled = 1; if (mp_ncpus >= MAXCPU) cpu_info[node->hwid].cpu_disabled = 1; if (cpu_info[node->hwid].cpu_disabled) { disabled_cpus++; continue; } + if (cpu_info[node->hwid].cpu_hyperthread) + nhyper++; + cpu_apic_ids[mp_ncpus] = node->hwid; apic_cpuids[node->hwid] = mp_ncpus; topo_set_pu_id(node, mp_ncpus); mp_ncpus++; } KASSERT(mp_maxid >= mp_ncpus - 1, ("%s: counters out of sync: max %d, count %d", __func__, mp_maxid, mp_ncpus)); + + mp_ncores = mp_ncpus - nhyper; + smp_threads_per_core = mp_ncpus / mp_ncores; } /* * Print various information about the SMP system hardware and setup. */ void cpu_mp_announce(void) { struct topo_node *node; const char *hyperthread; struct topo_analysis topology; printf("FreeBSD/SMP: "); if (topo_analyze(&topo_root, 1, &topology)) { printf("%d package(s)", topology.entities[TOPO_LEVEL_PKG]); if (topology.entities[TOPO_LEVEL_GROUP] > 1) printf(" x %d groups", topology.entities[TOPO_LEVEL_GROUP]); if (topology.entities[TOPO_LEVEL_CACHEGROUP] > 1) printf(" x %d cache groups", topology.entities[TOPO_LEVEL_CACHEGROUP]); if (topology.entities[TOPO_LEVEL_CORE] > 0) printf(" x %d core(s)", topology.entities[TOPO_LEVEL_CORE]); if (topology.entities[TOPO_LEVEL_THREAD] > 1) printf(" x %d hardware threads", topology.entities[TOPO_LEVEL_THREAD]); } else { printf("Non-uniform topology"); } printf("\n"); if (disabled_cpus) { printf("FreeBSD/SMP Online: "); if (topo_analyze(&topo_root, 0, &topology)) { printf("%d package(s)", topology.entities[TOPO_LEVEL_PKG]); if (topology.entities[TOPO_LEVEL_GROUP] > 1) printf(" x %d groups", topology.entities[TOPO_LEVEL_GROUP]); if (topology.entities[TOPO_LEVEL_CACHEGROUP] > 1) printf(" x %d cache groups", topology.entities[TOPO_LEVEL_CACHEGROUP]); if (topology.entities[TOPO_LEVEL_CORE] > 0) printf(" x %d core(s)", topology.entities[TOPO_LEVEL_CORE]); if (topology.entities[TOPO_LEVEL_THREAD] > 1) printf(" x %d hardware threads", topology.entities[TOPO_LEVEL_THREAD]); } else { printf("Non-uniform topology"); } printf("\n"); } if (!bootverbose) return; TOPO_FOREACH(node, &topo_root) { switch (node->type) { case TOPO_TYPE_PKG: printf("Package HW ID = %u\n", node->hwid); break; case TOPO_TYPE_CORE: printf("\tCore HW ID = %u\n", node->hwid); break; case TOPO_TYPE_PU: if (cpu_info[node->hwid].cpu_hyperthread) hyperthread = "/HT"; else hyperthread = ""; if (node->subtype == 0) printf("\t\tCPU (AP%s): APIC ID: %u" "(disabled)\n", hyperthread, node->hwid); else if (node->id == 0) printf("\t\tCPU0 (BSP): APIC ID: %u\n", node->hwid); else printf("\t\tCPU%u (AP%s): APIC ID: %u\n", node->id, hyperthread, node->hwid); break; default: /* ignored */ break; } } } /* * Add a scheduling group, a group of logical processors sharing * a particular cache (and, thus having an affinity), to the scheduling * topology. * This function recursively works on lower level caches. */ static void x86topo_add_sched_group(struct topo_node *root, struct cpu_group *cg_root) { struct topo_node *node; int nchildren; int ncores; int i; KASSERT(root->type == TOPO_TYPE_SYSTEM || root->type == TOPO_TYPE_CACHE || root->type == TOPO_TYPE_GROUP, ("x86topo_add_sched_group: bad type: %u", root->type)); CPU_COPY(&root->cpuset, &cg_root->cg_mask); cg_root->cg_count = root->cpu_count; if (root->type == TOPO_TYPE_SYSTEM) cg_root->cg_level = CG_SHARE_NONE; else cg_root->cg_level = root->subtype; /* * Check how many core nodes we have under the given root node. * If we have multiple logical processors, but not multiple * cores, then those processors must be hardware threads. */ ncores = 0; node = root; while (node != NULL) { if (node->type != TOPO_TYPE_CORE) { node = topo_next_node(root, node); continue; } ncores++; node = topo_next_nonchild_node(root, node); } if (cg_root->cg_level != CG_SHARE_NONE && root->cpu_count > 1 && ncores < 2) cg_root->cg_flags = CG_FLAG_SMT; /* * Find out how many cache nodes we have under the given root node. * We ignore cache nodes that cover all the same processors as the * root node. Also, we do not descend below found cache nodes. * That is, we count top-level "non-redundant" caches under the root * node. */ nchildren = 0; node = root; while (node != NULL) { if ((node->type != TOPO_TYPE_GROUP && node->type != TOPO_TYPE_CACHE) || (root->type != TOPO_TYPE_SYSTEM && CPU_CMP(&node->cpuset, &root->cpuset) == 0)) { node = topo_next_node(root, node); continue; } nchildren++; node = topo_next_nonchild_node(root, node); } cg_root->cg_child = smp_topo_alloc(nchildren); cg_root->cg_children = nchildren; /* * Now find again the same cache nodes as above and recursively * build scheduling topologies for them. */ node = root; i = 0; while (node != NULL) { if ((node->type != TOPO_TYPE_GROUP && node->type != TOPO_TYPE_CACHE) || (root->type != TOPO_TYPE_SYSTEM && CPU_CMP(&node->cpuset, &root->cpuset) == 0)) { node = topo_next_node(root, node); continue; } cg_root->cg_child[i].cg_parent = cg_root; x86topo_add_sched_group(node, &cg_root->cg_child[i]); i++; node = topo_next_nonchild_node(root, node); } } /* * Build the MI scheduling topology from the discovered hardware topology. */ struct cpu_group * cpu_topo(void) { struct cpu_group *cg_root; if (mp_ncpus <= 1) return (smp_topo_none()); cg_root = smp_topo_alloc(1); x86topo_add_sched_group(&topo_root, cg_root); return (cg_root); } static void cpu_alloc(void *dummy __unused) { /* * Dynamically allocate the arrays that depend on the * maximum APIC ID. */ cpu_info = malloc(sizeof(*cpu_info) * (max_apic_id + 1), M_CPUS, M_WAITOK | M_ZERO); apic_cpuids = malloc(sizeof(*apic_cpuids) * (max_apic_id + 1), M_CPUS, M_WAITOK | M_ZERO); } SYSINIT(cpu_alloc, SI_SUB_CPU, SI_ORDER_FIRST, cpu_alloc, NULL); /* * Add a logical CPU to the topology. */ void cpu_add(u_int apic_id, char boot_cpu) { if (apic_id > max_apic_id) { panic("SMP: APIC ID %d too high", apic_id); return; } KASSERT(cpu_info[apic_id].cpu_present == 0, ("CPU %u added twice", apic_id)); cpu_info[apic_id].cpu_present = 1; if (boot_cpu) { KASSERT(boot_cpu_id == -1, ("CPU %u claims to be BSP, but CPU %u already is", apic_id, boot_cpu_id)); boot_cpu_id = apic_id; cpu_info[apic_id].cpu_bsp = 1; } if (bootverbose) printf("SMP: Added CPU %u (%s)\n", apic_id, boot_cpu ? "BSP" : "AP"); } void cpu_mp_setmaxid(void) { /* * mp_ncpus and mp_maxid should be already set by calls to cpu_add(). * If there were no calls to cpu_add() assume this is a UP system. */ if (mp_ncpus == 0) mp_ncpus = 1; } int cpu_mp_probe(void) { /* * Always record BSP in CPU map so that the mbuf init code works * correctly. */ CPU_SETOF(0, &all_cpus); return (mp_ncpus > 1); } /* Allocate memory for the AP trampoline. */ void alloc_ap_trampoline(vm_paddr_t *physmap, unsigned int *physmap_idx) { unsigned int i; bool allocated; allocated = false; for (i = *physmap_idx; i <= *physmap_idx; i -= 2) { /* * Find a memory region big enough and below the 1MB boundary * for the trampoline code. * NB: needs to be page aligned. */ if (physmap[i] >= MiB(1) || (trunc_page(physmap[i + 1]) - round_page(physmap[i])) < round_page(bootMP_size)) continue; allocated = true; /* * Try to steal from the end of the region to mimic previous * behaviour, else fallback to steal from the start. */ if (physmap[i + 1] < MiB(1)) { boot_address = trunc_page(physmap[i + 1]); if ((physmap[i + 1] - boot_address) < bootMP_size) boot_address -= round_page(bootMP_size); physmap[i + 1] = boot_address; } else { boot_address = round_page(physmap[i]); physmap[i] = boot_address + round_page(bootMP_size); } if (physmap[i] == physmap[i + 1] && *physmap_idx != 0) { memmove(&physmap[i], &physmap[i + 2], sizeof(*physmap) * (*physmap_idx - i + 2)); *physmap_idx -= 2; } break; } if (!allocated) { boot_address = basemem * 1024 - bootMP_size; if (bootverbose) printf( "Cannot find enough space for the boot trampoline, placing it at %#x", boot_address); } } /* * AP CPU's call this to initialize themselves. */ void init_secondary_tail(void) { u_int cpuid; pmap_activate_boot(vmspace_pmap(proc0.p_vmspace)); /* * On real hardware, switch to x2apic mode if possible. Do it * after aps_ready was signalled, to avoid manipulating the * mode while BSP might still want to send some IPI to us * (second startup IPI is ignored on modern hardware etc). */ lapic_xapic_mode(); /* Initialize the PAT MSR. */ pmap_init_pat(); /* set up CPU registers and state */ cpu_setregs(); /* set up SSE/NX */ initializecpu(); /* set up FPU state on the AP */ #ifdef __amd64__ fpuinit(); #else npxinit(false); #endif if (cpu_ops.cpu_init) cpu_ops.cpu_init(); /* A quick check from sanity claus */ cpuid = PCPU_GET(cpuid); if (PCPU_GET(apic_id) != lapic_id()) { printf("SMP: cpuid = %d\n", cpuid); printf("SMP: actual apic_id = %d\n", lapic_id()); printf("SMP: correct apic_id = %d\n", PCPU_GET(apic_id)); panic("cpuid mismatch! boom!!"); } /* Initialize curthread. */ KASSERT(PCPU_GET(idlethread) != NULL, ("no idle thread")); PCPU_SET(curthread, PCPU_GET(idlethread)); mtx_lock_spin(&ap_boot_mtx); mca_init(); /* Init local apic for irq's */ lapic_setup(1); /* Set memory range attributes for this CPU to match the BSP */ mem_range_AP_init(); smp_cpus++; CTR1(KTR_SMP, "SMP: AP CPU #%d Launched", cpuid); if (bootverbose) printf("SMP: AP CPU #%d Launched!\n", cpuid); else printf("%s%d%s", smp_cpus == 2 ? "Launching APs: " : "", cpuid, smp_cpus == mp_ncpus ? "\n" : " "); /* Determine if we are a logical CPU. */ if (cpu_info[PCPU_GET(apic_id)].cpu_hyperthread) CPU_SET(cpuid, &logical_cpus_mask); if (bootverbose) lapic_dump("AP"); if (smp_cpus == mp_ncpus) { /* enable IPI's, tlb shootdown, freezes etc */ atomic_store_rel_int(&smp_started, 1); } #ifdef __amd64__ /* * Enable global pages TLB extension * This also implicitly flushes the TLB */ load_cr4(rcr4() | CR4_PGE); if (pmap_pcid_enabled) load_cr4(rcr4() | CR4_PCIDE); load_ds(_udatasel); load_es(_udatasel); load_fs(_ufssel); #endif mtx_unlock_spin(&ap_boot_mtx); /* Wait until all the AP's are up. */ while (atomic_load_acq_int(&smp_started) == 0) ia32_pause(); #ifndef EARLY_AP_STARTUP /* Start per-CPU event timers. */ cpu_initclocks_ap(); #endif sched_throw(NULL); panic("scheduler returned us to %s", __func__); /* NOTREACHED */ } static void smp_after_idle_runnable(void *arg __unused) { struct thread *idle_td; int cpu; for (cpu = 1; cpu < mp_ncpus; cpu++) { idle_td = pcpu_find(cpu)->pc_idlethread; while (idle_td->td_lastcpu == NOCPU && idle_td->td_oncpu == NOCPU) cpu_spinwait(); kmem_free((vm_offset_t)bootstacks[cpu], kstack_pages * PAGE_SIZE); } } SYSINIT(smp_after_idle_runnable, SI_SUB_SMP, SI_ORDER_ANY, smp_after_idle_runnable, NULL); /* * We tell the I/O APIC code about all the CPUs we want to receive * interrupts. If we don't want certain CPUs to receive IRQs we * can simply not tell the I/O APIC code about them in this function. * We also do not tell it about the BSP since it tells itself about * the BSP internally to work with UP kernels and on UP machines. */ void set_interrupt_apic_ids(void) { u_int i, apic_id; for (i = 0; i < MAXCPU; i++) { apic_id = cpu_apic_ids[i]; if (apic_id == -1) continue; if (cpu_info[apic_id].cpu_bsp) continue; if (cpu_info[apic_id].cpu_disabled) continue; /* Don't let hyperthreads service interrupts. */ if (cpu_info[apic_id].cpu_hyperthread) continue; intr_add_cpu(i); } } #ifdef COUNT_XINVLTLB_HITS u_int xhits_gbl[MAXCPU]; u_int xhits_pg[MAXCPU]; u_int xhits_rng[MAXCPU]; static SYSCTL_NODE(_debug, OID_AUTO, xhits, CTLFLAG_RW, 0, ""); SYSCTL_OPAQUE(_debug_xhits, OID_AUTO, global, CTLFLAG_RW, &xhits_gbl, sizeof(xhits_gbl), "IU", ""); SYSCTL_OPAQUE(_debug_xhits, OID_AUTO, page, CTLFLAG_RW, &xhits_pg, sizeof(xhits_pg), "IU", ""); SYSCTL_OPAQUE(_debug_xhits, OID_AUTO, range, CTLFLAG_RW, &xhits_rng, sizeof(xhits_rng), "IU", ""); u_int ipi_global; u_int ipi_page; u_int ipi_range; u_int ipi_range_size; SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_global, CTLFLAG_RW, &ipi_global, 0, ""); SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_page, CTLFLAG_RW, &ipi_page, 0, ""); SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_range, CTLFLAG_RW, &ipi_range, 0, ""); SYSCTL_INT(_debug_xhits, OID_AUTO, ipi_range_size, CTLFLAG_RW, &ipi_range_size, 0, ""); #endif /* COUNT_XINVLTLB_HITS */ /* * Init and startup IPI. */ void ipi_startup(int apic_id, int vector) { /* * This attempts to follow the algorithm described in the * Intel Multiprocessor Specification v1.4 in section B.4. * For each IPI, we allow the local APIC ~20us to deliver the * IPI. If that times out, we panic. */ /* * first we do an INIT IPI: this INIT IPI might be run, resetting * and running the target CPU. OR this INIT IPI might be latched (P5 * bug), CPU waiting for STARTUP IPI. OR this INIT IPI might be * ignored. */ lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_LEVEL | APIC_LEVEL_ASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_INIT, apic_id); lapic_ipi_wait(100); /* Explicitly deassert the INIT IPI. */ lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_LEVEL | APIC_LEVEL_DEASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_INIT, apic_id); DELAY(10000); /* wait ~10mS */ /* * next we do a STARTUP IPI: the previous INIT IPI might still be * latched, (P5 bug) this 1st STARTUP would then terminate * immediately, and the previously started INIT IPI would continue. OR * the previous INIT IPI has already run. and this STARTUP IPI will * run. OR the previous INIT IPI was ignored. and this STARTUP IPI * will run. */ lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_EDGE | APIC_LEVEL_ASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_STARTUP | vector, apic_id); if (!lapic_ipi_wait(100)) panic("Failed to deliver first STARTUP IPI to APIC %d", apic_id); DELAY(200); /* wait ~200uS */ /* * finally we do a 2nd STARTUP IPI: this 2nd STARTUP IPI should run IF * the previous STARTUP IPI was cancelled by a latched INIT IPI. OR * this STARTUP IPI will be ignored, as only ONE STARTUP IPI is * recognized after hardware RESET or INIT IPI. */ lapic_ipi_raw(APIC_DEST_DESTFLD | APIC_TRIGMOD_EDGE | APIC_LEVEL_ASSERT | APIC_DESTMODE_PHY | APIC_DELMODE_STARTUP | vector, apic_id); if (!lapic_ipi_wait(100)) panic("Failed to deliver second STARTUP IPI to APIC %d", apic_id); DELAY(200); /* wait ~200uS */ } /* * Send an IPI to specified CPU handling the bitmap logic. */ void ipi_send_cpu(int cpu, u_int ipi) { u_int bitmap, old_pending, new_pending; KASSERT(cpu_apic_ids[cpu] != -1, ("IPI to non-existent CPU %d", cpu)); if (IPI_IS_BITMAPED(ipi)) { bitmap = 1 << ipi; ipi = IPI_BITMAP_VECTOR; do { old_pending = cpu_ipi_pending[cpu]; new_pending = old_pending | bitmap; } while (!atomic_cmpset_int(&cpu_ipi_pending[cpu], old_pending, new_pending)); if (old_pending) return; } lapic_ipi_vectored(ipi, cpu_apic_ids[cpu]); } void ipi_bitmap_handler(struct trapframe frame) { struct trapframe *oldframe; struct thread *td; int cpu = PCPU_GET(cpuid); u_int ipi_bitmap; critical_enter(); td = curthread; td->td_intr_nesting_level++; oldframe = td->td_intr_frame; td->td_intr_frame = &frame; ipi_bitmap = atomic_readandclear_int(&cpu_ipi_pending[cpu]); if (ipi_bitmap & (1 << IPI_PREEMPT)) { #ifdef COUNT_IPIS (*ipi_preempt_counts[cpu])++; #endif sched_preempt(td); } if (ipi_bitmap & (1 << IPI_AST)) { #ifdef COUNT_IPIS (*ipi_ast_counts[cpu])++; #endif /* Nothing to do for AST */ } if (ipi_bitmap & (1 << IPI_HARDCLOCK)) { #ifdef COUNT_IPIS (*ipi_hardclock_counts[cpu])++; #endif hardclockintr(); } td->td_intr_frame = oldframe; td->td_intr_nesting_level--; critical_exit(); } /* * send an IPI to a set of cpus. */ void ipi_selected(cpuset_t cpus, u_int ipi) { int cpu; /* * IPI_STOP_HARD maps to a NMI and the trap handler needs a bit * of help in order to understand what is the source. * Set the mask of receiving CPUs for this purpose. */ if (ipi == IPI_STOP_HARD) CPU_OR_ATOMIC(&ipi_stop_nmi_pending, &cpus); while ((cpu = CPU_FFS(&cpus)) != 0) { cpu--; CPU_CLR(cpu, &cpus); CTR3(KTR_SMP, "%s: cpu: %d ipi: %x", __func__, cpu, ipi); ipi_send_cpu(cpu, ipi); } } /* * send an IPI to a specific CPU. */ void ipi_cpu(int cpu, u_int ipi) { /* * IPI_STOP_HARD maps to a NMI and the trap handler needs a bit * of help in order to understand what is the source. * Set the mask of receiving CPUs for this purpose. */ if (ipi == IPI_STOP_HARD) CPU_SET_ATOMIC(cpu, &ipi_stop_nmi_pending); CTR3(KTR_SMP, "%s: cpu: %d ipi: %x", __func__, cpu, ipi); ipi_send_cpu(cpu, ipi); } /* * send an IPI to all CPUs EXCEPT myself */ void ipi_all_but_self(u_int ipi) { cpuset_t other_cpus; other_cpus = all_cpus; CPU_CLR(PCPU_GET(cpuid), &other_cpus); if (IPI_IS_BITMAPED(ipi)) { ipi_selected(other_cpus, ipi); return; } /* * IPI_STOP_HARD maps to a NMI and the trap handler needs a bit * of help in order to understand what is the source. * Set the mask of receiving CPUs for this purpose. */ if (ipi == IPI_STOP_HARD) CPU_OR_ATOMIC(&ipi_stop_nmi_pending, &other_cpus); CTR2(KTR_SMP, "%s: ipi: %x", __func__, ipi); lapic_ipi_vectored(ipi, APIC_IPI_DEST_OTHERS); } int ipi_nmi_handler(void) { u_int cpuid; /* * As long as there is not a simple way to know about a NMI's * source, if the bitmask for the current CPU is present in * the global pending bitword an IPI_STOP_HARD has been issued * and should be handled. */ cpuid = PCPU_GET(cpuid); if (!CPU_ISSET(cpuid, &ipi_stop_nmi_pending)) return (1); CPU_CLR_ATOMIC(cpuid, &ipi_stop_nmi_pending); cpustop_handler(); return (0); } int nmi_kdb_lock; void nmi_call_kdb_smp(u_int type, struct trapframe *frame) { int cpu; bool call_post; cpu = PCPU_GET(cpuid); if (atomic_cmpset_acq_int(&nmi_kdb_lock, 0, 1)) { nmi_call_kdb(cpu, type, frame); call_post = false; } else { savectx(&stoppcbs[cpu]); CPU_SET_ATOMIC(cpu, &stopped_cpus); while (!atomic_cmpset_acq_int(&nmi_kdb_lock, 0, 1)) ia32_pause(); call_post = true; } atomic_store_rel_int(&nmi_kdb_lock, 0); if (call_post) cpustop_handler_post(cpu); } /* * Handle an IPI_STOP by saving our current context and spinning until we * are resumed. */ void cpustop_handler(void) { u_int cpu; cpu = PCPU_GET(cpuid); savectx(&stoppcbs[cpu]); /* Indicate that we are stopped */ CPU_SET_ATOMIC(cpu, &stopped_cpus); /* Wait for restart */ while (!CPU_ISSET(cpu, &started_cpus)) ia32_pause(); cpustop_handler_post(cpu); } static void cpustop_handler_post(u_int cpu) { CPU_CLR_ATOMIC(cpu, &started_cpus); CPU_CLR_ATOMIC(cpu, &stopped_cpus); /* * We don't broadcast TLB invalidations to other CPUs when they are * stopped. Hence, we clear the TLB before resuming. */ invltlb_glob(); #if defined(__amd64__) && defined(DDB) amd64_db_resume_dbreg(); #endif if (cpu == 0 && cpustop_restartfunc != NULL) { cpustop_restartfunc(); cpustop_restartfunc = NULL; } } /* * Handle an IPI_SUSPEND by saving our current context and spinning until we * are resumed. */ void cpususpend_handler(void) { u_int cpu; mtx_assert(&smp_ipi_mtx, MA_NOTOWNED); cpu = PCPU_GET(cpuid); if (savectx(&susppcbs[cpu]->sp_pcb)) { #ifdef __amd64__ fpususpend(susppcbs[cpu]->sp_fpususpend); #else npxsuspend(susppcbs[cpu]->sp_fpususpend); #endif /* * suspended_cpus is cleared shortly after each AP is restarted * by a Startup IPI, so that the BSP can proceed to restarting * the next AP. * * resuming_cpus gets cleared when the AP completes * initialization after having been released by the BSP. * resuming_cpus is probably not the best name for the * variable, because it is actually a set of processors that * haven't resumed yet and haven't necessarily started resuming. * * Note that suspended_cpus is meaningful only for ACPI suspend * as it's not really used for Xen suspend since the APs are * automatically restored to the running state and the correct * context. For the same reason resumectx is never called in * that case. */ CPU_SET_ATOMIC(cpu, &suspended_cpus); CPU_SET_ATOMIC(cpu, &resuming_cpus); /* * Invalidate the cache after setting the global status bits. * The last AP to set its bit may end up being an Owner of the * corresponding cache line in MOESI protocol. The AP may be * stopped before the cache line is written to the main memory. */ wbinvd(); } else { #ifdef __amd64__ fpuresume(susppcbs[cpu]->sp_fpususpend); #else npxresume(susppcbs[cpu]->sp_fpususpend); #endif pmap_init_pat(); initializecpu(); PCPU_SET(switchtime, 0); PCPU_SET(switchticks, ticks); /* Indicate that we have restarted and restored the context. */ CPU_CLR_ATOMIC(cpu, &suspended_cpus); } /* Wait for resume directive */ while (!CPU_ISSET(cpu, &toresume_cpus)) ia32_pause(); /* Re-apply microcode updates. */ ucode_reload(); #ifdef __i386__ /* Finish removing the identity mapping of low memory for this AP. */ invltlb_glob(); #endif if (cpu_ops.cpu_resume) cpu_ops.cpu_resume(); #ifdef __amd64__ if (vmm_resume_p) vmm_resume_p(); #endif /* Resume MCA and local APIC */ lapic_xapic_mode(); mca_resume(); lapic_setup(0); /* Indicate that we are resumed */ CPU_CLR_ATOMIC(cpu, &resuming_cpus); CPU_CLR_ATOMIC(cpu, &suspended_cpus); CPU_CLR_ATOMIC(cpu, &toresume_cpus); } void invlcache_handler(void) { uint32_t generation; #ifdef COUNT_IPIS (*ipi_invlcache_counts[PCPU_GET(cpuid)])++; #endif /* COUNT_IPIS */ /* * Reading the generation here allows greater parallelism * since wbinvd is a serializing instruction. Without the * temporary, we'd wait for wbinvd to complete, then the read * would execute, then the dependent write, which must then * complete before return from interrupt. */ generation = smp_tlb_generation; wbinvd(); PCPU_SET(smp_tlb_done, generation); } /* * This is called once the rest of the system is up and running and we're * ready to let the AP's out of the pen. */ static void release_aps(void *dummy __unused) { if (mp_ncpus == 1) return; atomic_store_rel_int(&aps_ready, 1); while (smp_started == 0) ia32_pause(); } SYSINIT(start_aps, SI_SUB_SMP, SI_ORDER_FIRST, release_aps, NULL); #ifdef COUNT_IPIS /* * Setup interrupt counters for IPI handlers. */ static void mp_ipi_intrcnt(void *dummy) { char buf[64]; int i; CPU_FOREACH(i) { snprintf(buf, sizeof(buf), "cpu%d:invltlb", i); intrcnt_add(buf, &ipi_invltlb_counts[i]); snprintf(buf, sizeof(buf), "cpu%d:invlrng", i); intrcnt_add(buf, &ipi_invlrng_counts[i]); snprintf(buf, sizeof(buf), "cpu%d:invlpg", i); intrcnt_add(buf, &ipi_invlpg_counts[i]); snprintf(buf, sizeof(buf), "cpu%d:invlcache", i); intrcnt_add(buf, &ipi_invlcache_counts[i]); snprintf(buf, sizeof(buf), "cpu%d:preempt", i); intrcnt_add(buf, &ipi_preempt_counts[i]); snprintf(buf, sizeof(buf), "cpu%d:ast", i); intrcnt_add(buf, &ipi_ast_counts[i]); snprintf(buf, sizeof(buf), "cpu%d:rendezvous", i); intrcnt_add(buf, &ipi_rendezvous_counts[i]); snprintf(buf, sizeof(buf), "cpu%d:hardclock", i); intrcnt_add(buf, &ipi_hardclock_counts[i]); } } SYSINIT(mp_ipi_intrcnt, SI_SUB_INTR, SI_ORDER_MIDDLE, mp_ipi_intrcnt, NULL); #endif /* * Flush the TLB on other CPU's */ /* Variables needed for SMP tlb shootdown. */ vm_offset_t smp_tlb_addr1, smp_tlb_addr2; pmap_t smp_tlb_pmap; volatile uint32_t smp_tlb_generation; #ifdef __amd64__ #define read_eflags() read_rflags() #endif static void smp_targeted_tlb_shootdown(cpuset_t mask, u_int vector, pmap_t pmap, vm_offset_t addr1, vm_offset_t addr2) { cpuset_t other_cpus; volatile uint32_t *p_cpudone; uint32_t generation; int cpu; /* It is not necessary to signal other CPUs while in the debugger. */ if (kdb_active || panicstr != NULL) return; /* * Check for other cpus. Return if none. */ if (CPU_ISFULLSET(&mask)) { if (mp_ncpus <= 1) return; } else { CPU_CLR(PCPU_GET(cpuid), &mask); if (CPU_EMPTY(&mask)) return; } if (!(read_eflags() & PSL_I)) panic("%s: interrupts disabled", __func__); mtx_lock_spin(&smp_ipi_mtx); smp_tlb_addr1 = addr1; smp_tlb_addr2 = addr2; smp_tlb_pmap = pmap; generation = ++smp_tlb_generation; if (CPU_ISFULLSET(&mask)) { ipi_all_but_self(vector); other_cpus = all_cpus; CPU_CLR(PCPU_GET(cpuid), &other_cpus); } else { other_cpus = mask; while ((cpu = CPU_FFS(&mask)) != 0) { cpu--; CPU_CLR(cpu, &mask); CTR3(KTR_SMP, "%s: cpu: %d ipi: %x", __func__, cpu, vector); ipi_send_cpu(cpu, vector); } } while ((cpu = CPU_FFS(&other_cpus)) != 0) { cpu--; CPU_CLR(cpu, &other_cpus); p_cpudone = &cpuid_to_pcpu[cpu]->pc_smp_tlb_done; while (*p_cpudone != generation) ia32_pause(); } mtx_unlock_spin(&smp_ipi_mtx); } void smp_masked_invltlb(cpuset_t mask, pmap_t pmap) { if (smp_started) { smp_targeted_tlb_shootdown(mask, IPI_INVLTLB, pmap, 0, 0); #ifdef COUNT_XINVLTLB_HITS ipi_global++; #endif } } void smp_masked_invlpg(cpuset_t mask, vm_offset_t addr, pmap_t pmap) { if (smp_started) { smp_targeted_tlb_shootdown(mask, IPI_INVLPG, pmap, addr, 0); #ifdef COUNT_XINVLTLB_HITS ipi_page++; #endif } } void smp_masked_invlpg_range(cpuset_t mask, vm_offset_t addr1, vm_offset_t addr2, pmap_t pmap) { if (smp_started) { smp_targeted_tlb_shootdown(mask, IPI_INVLRNG, pmap, addr1, addr2); #ifdef COUNT_XINVLTLB_HITS ipi_range++; ipi_range_size += (addr2 - addr1) / PAGE_SIZE; #endif } } void smp_cache_flush(void) { if (smp_started) { smp_targeted_tlb_shootdown(all_cpus, IPI_INVLCACHE, NULL, 0, 0); } } /* * Handlers for TLB related IPIs */ void invltlb_handler(void) { uint32_t generation; #ifdef COUNT_XINVLTLB_HITS xhits_gbl[PCPU_GET(cpuid)]++; #endif /* COUNT_XINVLTLB_HITS */ #ifdef COUNT_IPIS (*ipi_invltlb_counts[PCPU_GET(cpuid)])++; #endif /* COUNT_IPIS */ /* * Reading the generation here allows greater parallelism * since invalidating the TLB is a serializing operation. */ generation = smp_tlb_generation; if (smp_tlb_pmap == kernel_pmap) invltlb_glob(); #ifdef __amd64__ else invltlb(); #endif PCPU_SET(smp_tlb_done, generation); } void invlpg_handler(void) { uint32_t generation; #ifdef COUNT_XINVLTLB_HITS xhits_pg[PCPU_GET(cpuid)]++; #endif /* COUNT_XINVLTLB_HITS */ #ifdef COUNT_IPIS (*ipi_invlpg_counts[PCPU_GET(cpuid)])++; #endif /* COUNT_IPIS */ generation = smp_tlb_generation; /* Overlap with serialization */ #ifdef __i386__ if (smp_tlb_pmap == kernel_pmap) #endif invlpg(smp_tlb_addr1); PCPU_SET(smp_tlb_done, generation); } void invlrng_handler(void) { vm_offset_t addr, addr2; uint32_t generation; #ifdef COUNT_XINVLTLB_HITS xhits_rng[PCPU_GET(cpuid)]++; #endif /* COUNT_XINVLTLB_HITS */ #ifdef COUNT_IPIS (*ipi_invlrng_counts[PCPU_GET(cpuid)])++; #endif /* COUNT_IPIS */ addr = smp_tlb_addr1; addr2 = smp_tlb_addr2; generation = smp_tlb_generation; /* Overlap with serialization */ #ifdef __i386__ if (smp_tlb_pmap == kernel_pmap) #endif do { invlpg(addr); addr += PAGE_SIZE; } while (addr < addr2); PCPU_SET(smp_tlb_done, generation); } Index: stable/12 =================================================================== --- stable/12 (revision 354072) +++ stable/12 (revision 354073) Property changes on: stable/12 ___________________________________________________________________ Modified: svn:mergeinfo ## -0,0 +0,1 ## Merged /head:r342771