Index: head/sys/dev/hwpmc/hwpmc_mod.c =================================================================== --- head/sys/dev/hwpmc/hwpmc_mod.c (revision 334343) +++ head/sys/dev/hwpmc/hwpmc_mod.c (revision 334344) @@ -1,5709 +1,5710 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2003-2008 Joseph Koshy * Copyright (c) 2007 The FreeBSD Foundation * Copyright (c) 2018 Matthew Macy * All rights reserved. * * Portions of this software were developed by A. Joseph Koshy under * sponsorship from the FreeBSD Foundation and Google, Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * */ #include __FBSDID("$FreeBSD$"); #include #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 /* needs to be after */ #include #include #include #include #include #include #include #include "hwpmc_soft.h" #ifdef NUMA #define NDOMAINS vm_ndomains #else #define NDOMAINS 1 #define malloc_domain(size, type, domain, flags) malloc((size), (type), (flags)) #define free_domain(addr, type) free(addr, type) #endif /* * Types */ enum pmc_flags { PMC_FLAG_NONE = 0x00, /* do nothing */ PMC_FLAG_REMOVE = 0x01, /* atomically remove entry from hash */ PMC_FLAG_ALLOCATE = 0x02, /* add entry to hash if not found */ PMC_FLAG_NOWAIT = 0x04, /* do not wait for mallocs */ }; /* * The offset in sysent where the syscall is allocated. */ static int pmc_syscall_num = NO_SYSCALL; struct pmc_cpu **pmc_pcpu; /* per-cpu state */ pmc_value_t *pmc_pcpu_saved; /* saved PMC values: CSW handling */ #define PMC_PCPU_SAVED(C,R) pmc_pcpu_saved[(R) + md->pmd_npmc*(C)] struct mtx_pool *pmc_mtxpool; static int *pmc_pmcdisp; /* PMC row dispositions */ #define PMC_ROW_DISP_IS_FREE(R) (pmc_pmcdisp[(R)] == 0) #define PMC_ROW_DISP_IS_THREAD(R) (pmc_pmcdisp[(R)] > 0) #define PMC_ROW_DISP_IS_STANDALONE(R) (pmc_pmcdisp[(R)] < 0) #define PMC_MARK_ROW_FREE(R) do { \ pmc_pmcdisp[(R)] = 0; \ } while (0) #define PMC_MARK_ROW_STANDALONE(R) do { \ KASSERT(pmc_pmcdisp[(R)] <= 0, ("[pmc,%d] row disposition error", \ __LINE__)); \ atomic_add_int(&pmc_pmcdisp[(R)], -1); \ KASSERT(pmc_pmcdisp[(R)] >= (-pmc_cpu_max_active()), \ ("[pmc,%d] row disposition error", __LINE__)); \ } while (0) #define PMC_UNMARK_ROW_STANDALONE(R) do { \ atomic_add_int(&pmc_pmcdisp[(R)], 1); \ KASSERT(pmc_pmcdisp[(R)] <= 0, ("[pmc,%d] row disposition error", \ __LINE__)); \ } while (0) #define PMC_MARK_ROW_THREAD(R) do { \ KASSERT(pmc_pmcdisp[(R)] >= 0, ("[pmc,%d] row disposition error", \ __LINE__)); \ atomic_add_int(&pmc_pmcdisp[(R)], 1); \ } while (0) #define PMC_UNMARK_ROW_THREAD(R) do { \ atomic_add_int(&pmc_pmcdisp[(R)], -1); \ KASSERT(pmc_pmcdisp[(R)] >= 0, ("[pmc,%d] row disposition error", \ __LINE__)); \ } while (0) /* various event handlers */ static eventhandler_tag pmc_exit_tag, pmc_fork_tag, pmc_kld_load_tag, pmc_kld_unload_tag; /* Module statistics */ struct pmc_driverstats pmc_stats; /* Machine/processor dependent operations */ static struct pmc_mdep *md; /* * Hash tables mapping owner processes and target threads to PMCs. */ struct mtx pmc_processhash_mtx; /* spin mutex */ static u_long pmc_processhashmask; static LIST_HEAD(pmc_processhash, pmc_process) *pmc_processhash; /* * Hash table of PMC owner descriptors. This table is protected by * the shared PMC "sx" lock. */ static u_long pmc_ownerhashmask; static LIST_HEAD(pmc_ownerhash, pmc_owner) *pmc_ownerhash; /* * List of PMC owners with system-wide sampling PMCs. */ static CK_LIST_HEAD(, pmc_owner) pmc_ss_owners; /* * List of free thread entries. This is protected by the spin * mutex. */ static struct mtx pmc_threadfreelist_mtx; /* spin mutex */ static LIST_HEAD(, pmc_thread) pmc_threadfreelist; static int pmc_threadfreelist_entries=0; #define THREADENTRY_SIZE \ (sizeof(struct pmc_thread) + (md->pmd_npmc * sizeof(struct pmc_threadpmcstate))) /* * Task to free thread descriptors */ static struct grouptask free_gtask; /* * A map of row indices to classdep structures. */ static struct pmc_classdep **pmc_rowindex_to_classdep; /* * Prototypes */ #ifdef HWPMC_DEBUG static int pmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS); static int pmc_debugflags_parse(char *newstr, char *fence); #endif static int load(struct module *module, int cmd, void *arg); static void pmc_add_thread_descriptors_from_proc(struct proc *p, struct pmc_process *pp); static int pmc_attach_process(struct proc *p, struct pmc *pm); static struct pmc *pmc_allocate_pmc_descriptor(void); static struct pmc_owner *pmc_allocate_owner_descriptor(struct proc *p); static int pmc_attach_one_process(struct proc *p, struct pmc *pm); static int pmc_can_allocate_rowindex(struct proc *p, unsigned int ri, int cpu); static int pmc_can_attach(struct pmc *pm, struct proc *p); static void pmc_capture_user_callchain(int cpu, int soft, struct trapframe *tf); static void pmc_cleanup(void); static int pmc_detach_process(struct proc *p, struct pmc *pm); static int pmc_detach_one_process(struct proc *p, struct pmc *pm, int flags); static void pmc_destroy_owner_descriptor(struct pmc_owner *po); static void pmc_destroy_pmc_descriptor(struct pmc *pm); static void pmc_destroy_process_descriptor(struct pmc_process *pp); static struct pmc_owner *pmc_find_owner_descriptor(struct proc *p); static int pmc_find_pmc(pmc_id_t pmcid, struct pmc **pm); static struct pmc *pmc_find_pmc_descriptor_in_process(struct pmc_owner *po, pmc_id_t pmc); static struct pmc_process *pmc_find_process_descriptor(struct proc *p, uint32_t mode); static struct pmc_thread *pmc_find_thread_descriptor(struct pmc_process *pp, struct thread *td, uint32_t mode); static void pmc_force_context_switch(void); static void pmc_link_target_process(struct pmc *pm, struct pmc_process *pp); static void pmc_log_all_process_mappings(struct pmc_owner *po); static void pmc_log_kernel_mappings(struct pmc *pm); static void pmc_log_process_mappings(struct pmc_owner *po, struct proc *p); static void pmc_maybe_remove_owner(struct pmc_owner *po); static void pmc_process_csw_in(struct thread *td); static void pmc_process_csw_out(struct thread *td); static void pmc_process_exit(void *arg, struct proc *p); static void pmc_process_fork(void *arg, struct proc *p1, struct proc *p2, int n); static void pmc_process_samples(int cpu, int soft); static void pmc_release_pmc_descriptor(struct pmc *pmc); static void pmc_process_thread_add(struct thread *td); static void pmc_process_thread_delete(struct thread *td); static void pmc_remove_owner(struct pmc_owner *po); static void pmc_remove_process_descriptor(struct pmc_process *pp); static void pmc_restore_cpu_binding(struct pmc_binding *pb); static void pmc_save_cpu_binding(struct pmc_binding *pb); static void pmc_select_cpu(int cpu); static int pmc_start(struct pmc *pm); static int pmc_stop(struct pmc *pm); static int pmc_syscall_handler(struct thread *td, void *syscall_args); static struct pmc_thread *pmc_thread_descriptor_pool_alloc(void); static void pmc_thread_descriptor_pool_drain(void); static void pmc_thread_descriptor_pool_free(struct pmc_thread *pt); static void pmc_unlink_target_process(struct pmc *pmc, struct pmc_process *pp); static int generic_switch_in(struct pmc_cpu *pc, struct pmc_process *pp); static int generic_switch_out(struct pmc_cpu *pc, struct pmc_process *pp); static struct pmc_mdep *pmc_generic_cpu_initialize(void); static void pmc_generic_cpu_finalize(struct pmc_mdep *md); /* * Kernel tunables and sysctl(8) interface. */ SYSCTL_DECL(_kern_hwpmc); SYSCTL_NODE(_kern_hwpmc, OID_AUTO, stats, CTLFLAG_RW, 0, "HWPMC stats"); /* Stats. */ SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, intr_ignored, CTLFLAG_RW, &pmc_stats.pm_intr_ignored, "# of interrupts ignored"); SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, intr_processed, CTLFLAG_RW, &pmc_stats.pm_intr_processed, "# of interrupts processed"); SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, intr_bufferfull, CTLFLAG_RW, &pmc_stats.pm_intr_bufferfull, "# of interrupts where buffer was full"); SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, syscalls, CTLFLAG_RW, &pmc_stats.pm_syscalls, "# of syscalls"); SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, syscall_errors, CTLFLAG_RW, &pmc_stats.pm_syscall_errors, "# of syscall_errors"); SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, buffer_requests, CTLFLAG_RW, &pmc_stats.pm_buffer_requests, "# of buffer requests"); SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, buffer_requests_failed, CTLFLAG_RW, &pmc_stats.pm_buffer_requests_failed, "# of buffer requests which failed"); SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, log_sweeps, CTLFLAG_RW, &pmc_stats.pm_log_sweeps, "# of ?"); static int pmc_callchaindepth = PMC_CALLCHAIN_DEPTH; SYSCTL_INT(_kern_hwpmc, OID_AUTO, callchaindepth, CTLFLAG_RDTUN, &pmc_callchaindepth, 0, "depth of call chain records"); char pmc_cpuid[64]; SYSCTL_STRING(_kern_hwpmc, OID_AUTO, cpuid, CTLFLAG_RD, pmc_cpuid, 0, "cpu version string"); #ifdef HWPMC_DEBUG struct pmc_debugflags pmc_debugflags = PMC_DEBUG_DEFAULT_FLAGS; char pmc_debugstr[PMC_DEBUG_STRSIZE]; TUNABLE_STR(PMC_SYSCTL_NAME_PREFIX "debugflags", pmc_debugstr, sizeof(pmc_debugstr)); SYSCTL_PROC(_kern_hwpmc, OID_AUTO, debugflags, CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NOFETCH, 0, 0, pmc_debugflags_sysctl_handler, "A", "debug flags"); #endif /* * kern.hwpmc.hashrows -- determines the number of rows in the * of the hash table used to look up threads */ static int pmc_hashsize = PMC_HASH_SIZE; SYSCTL_INT(_kern_hwpmc, OID_AUTO, hashsize, CTLFLAG_RDTUN, &pmc_hashsize, 0, "rows in hash tables"); /* * kern.hwpmc.nsamples --- number of PC samples/callchain stacks per CPU */ static int pmc_nsamples = PMC_NSAMPLES; SYSCTL_INT(_kern_hwpmc, OID_AUTO, nsamples, CTLFLAG_RDTUN, &pmc_nsamples, 0, "number of PC samples per CPU"); /* * kern.hwpmc.mtxpoolsize -- number of mutexes in the mutex pool. */ static int pmc_mtxpool_size = PMC_MTXPOOL_SIZE; SYSCTL_INT(_kern_hwpmc, OID_AUTO, mtxpoolsize, CTLFLAG_RDTUN, &pmc_mtxpool_size, 0, "size of spin mutex pool"); /* * kern.hwpmc.threadfreelist_entries -- number of free entries */ SYSCTL_INT(_kern_hwpmc, OID_AUTO, threadfreelist_entries, CTLFLAG_RD, &pmc_threadfreelist_entries, 0, "number of avalable thread entries"); /* * kern.hwpmc.threadfreelist_max -- maximum number of free entries */ static int pmc_threadfreelist_max = PMC_THREADLIST_MAX; SYSCTL_INT(_kern_hwpmc, OID_AUTO, threadfreelist_max, CTLFLAG_RW, &pmc_threadfreelist_max, 0, "maximum number of available thread entries before freeing some"); /* * security.bsd.unprivileged_syspmcs -- allow non-root processes to * allocate system-wide PMCs. * * Allowing unprivileged processes to allocate system PMCs is convenient * if system-wide measurements need to be taken concurrently with other * per-process measurements. This feature is turned off by default. */ static int pmc_unprivileged_syspmcs = 0; SYSCTL_INT(_security_bsd, OID_AUTO, unprivileged_syspmcs, CTLFLAG_RWTUN, &pmc_unprivileged_syspmcs, 0, "allow unprivileged process to allocate system PMCs"); /* * Hash function. Discard the lower 2 bits of the pointer since * these are always zero for our uses. The hash multiplier is * round((2^LONG_BIT) * ((sqrt(5)-1)/2)). */ #if LONG_BIT == 64 #define _PMC_HM 11400714819323198486u #elif LONG_BIT == 32 #define _PMC_HM 2654435769u #else #error Must know the size of 'long' to compile #endif #define PMC_HASH_PTR(P,M) ((((unsigned long) (P) >> 2) * _PMC_HM) & (M)) /* * Syscall structures */ /* The `sysent' for the new syscall */ static struct sysent pmc_sysent = { .sy_narg = 2, .sy_call = pmc_syscall_handler, }; static struct syscall_module_data pmc_syscall_mod = { .chainevh = load, .chainarg = NULL, .offset = &pmc_syscall_num, .new_sysent = &pmc_sysent, .old_sysent = { .sy_narg = 0, .sy_call = NULL }, .flags = SY_THR_STATIC_KLD, }; static moduledata_t pmc_mod = { .name = PMC_MODULE_NAME, .evhand = syscall_module_handler, .priv = &pmc_syscall_mod, }; #ifdef EARLY_AP_STARTUP DECLARE_MODULE(pmc, pmc_mod, SI_SUB_SYSCALLS, SI_ORDER_ANY); #else DECLARE_MODULE(pmc, pmc_mod, SI_SUB_SMP, SI_ORDER_ANY); #endif MODULE_VERSION(pmc, PMC_VERSION); #ifdef HWPMC_DEBUG enum pmc_dbgparse_state { PMCDS_WS, /* in whitespace */ PMCDS_MAJOR, /* seen a major keyword */ PMCDS_MINOR }; static int pmc_debugflags_parse(char *newstr, char *fence) { char c, *p, *q; struct pmc_debugflags *tmpflags; int error, found, *newbits, tmp; size_t kwlen; tmpflags = malloc(sizeof(*tmpflags), M_PMC, M_WAITOK|M_ZERO); p = newstr; error = 0; for (; p < fence && (c = *p); p++) { /* skip white space */ if (c == ' ' || c == '\t') continue; /* look for a keyword followed by "=" */ for (q = p; p < fence && (c = *p) && c != '='; p++) ; if (c != '=') { error = EINVAL; goto done; } kwlen = p - q; newbits = NULL; /* lookup flag group name */ #define DBG_SET_FLAG_MAJ(S,F) \ if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0) \ newbits = &tmpflags->pdb_ ## F; DBG_SET_FLAG_MAJ("cpu", CPU); DBG_SET_FLAG_MAJ("csw", CSW); DBG_SET_FLAG_MAJ("logging", LOG); DBG_SET_FLAG_MAJ("module", MOD); DBG_SET_FLAG_MAJ("md", MDP); DBG_SET_FLAG_MAJ("owner", OWN); DBG_SET_FLAG_MAJ("pmc", PMC); DBG_SET_FLAG_MAJ("process", PRC); DBG_SET_FLAG_MAJ("sampling", SAM); if (newbits == NULL) { error = EINVAL; goto done; } p++; /* skip the '=' */ /* Now parse the individual flags */ tmp = 0; newflag: for (q = p; p < fence && (c = *p); p++) if (c == ' ' || c == '\t' || c == ',') break; /* p == fence or c == ws or c == "," or c == 0 */ if ((kwlen = p - q) == 0) { *newbits = tmp; continue; } found = 0; #define DBG_SET_FLAG_MIN(S,F) \ if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0) \ tmp |= found = (1 << PMC_DEBUG_MIN_ ## F) /* a '*' denotes all possible flags in the group */ if (kwlen == 1 && *q == '*') tmp = found = ~0; /* look for individual flag names */ DBG_SET_FLAG_MIN("allocaterow", ALR); DBG_SET_FLAG_MIN("allocate", ALL); DBG_SET_FLAG_MIN("attach", ATT); DBG_SET_FLAG_MIN("bind", BND); DBG_SET_FLAG_MIN("config", CFG); DBG_SET_FLAG_MIN("exec", EXC); DBG_SET_FLAG_MIN("exit", EXT); DBG_SET_FLAG_MIN("find", FND); DBG_SET_FLAG_MIN("flush", FLS); DBG_SET_FLAG_MIN("fork", FRK); DBG_SET_FLAG_MIN("getbuf", GTB); DBG_SET_FLAG_MIN("hook", PMH); DBG_SET_FLAG_MIN("init", INI); DBG_SET_FLAG_MIN("intr", INT); DBG_SET_FLAG_MIN("linktarget", TLK); DBG_SET_FLAG_MIN("mayberemove", OMR); DBG_SET_FLAG_MIN("ops", OPS); DBG_SET_FLAG_MIN("read", REA); DBG_SET_FLAG_MIN("register", REG); DBG_SET_FLAG_MIN("release", REL); DBG_SET_FLAG_MIN("remove", ORM); DBG_SET_FLAG_MIN("sample", SAM); DBG_SET_FLAG_MIN("scheduleio", SIO); DBG_SET_FLAG_MIN("select", SEL); DBG_SET_FLAG_MIN("signal", SIG); DBG_SET_FLAG_MIN("swi", SWI); DBG_SET_FLAG_MIN("swo", SWO); DBG_SET_FLAG_MIN("start", STA); DBG_SET_FLAG_MIN("stop", STO); DBG_SET_FLAG_MIN("syscall", PMS); DBG_SET_FLAG_MIN("unlinktarget", TUL); DBG_SET_FLAG_MIN("write", WRI); if (found == 0) { /* unrecognized flag name */ error = EINVAL; goto done; } if (c == 0 || c == ' ' || c == '\t') { /* end of flag group */ *newbits = tmp; continue; } p++; goto newflag; } /* save the new flag set */ bcopy(tmpflags, &pmc_debugflags, sizeof(pmc_debugflags)); done: free(tmpflags, M_PMC); return error; } static int pmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS) { char *fence, *newstr; int error; unsigned int n; (void) arg1; (void) arg2; /* unused parameters */ n = sizeof(pmc_debugstr); newstr = malloc(n, M_PMC, M_WAITOK|M_ZERO); (void) strlcpy(newstr, pmc_debugstr, n); error = sysctl_handle_string(oidp, newstr, n, req); /* if there is a new string, parse and copy it */ if (error == 0 && req->newptr != NULL) { fence = newstr + (n < req->newlen ? n : req->newlen + 1); if ((error = pmc_debugflags_parse(newstr, fence)) == 0) (void) strlcpy(pmc_debugstr, newstr, sizeof(pmc_debugstr)); } free(newstr, M_PMC); return error; } #endif /* * Map a row index to a classdep structure and return the adjusted row * index for the PMC class index. */ static struct pmc_classdep * pmc_ri_to_classdep(struct pmc_mdep *md, int ri, int *adjri) { struct pmc_classdep *pcd; (void) md; KASSERT(ri >= 0 && ri < md->pmd_npmc, ("[pmc,%d] illegal row-index %d", __LINE__, ri)); pcd = pmc_rowindex_to_classdep[ri]; KASSERT(pcd != NULL, ("[pmc,%d] ri %d null pcd", __LINE__, ri)); *adjri = ri - pcd->pcd_ri; KASSERT(*adjri >= 0 && *adjri < pcd->pcd_num, ("[pmc,%d] adjusted row-index %d", __LINE__, *adjri)); return (pcd); } /* * Concurrency Control * * The driver manages the following data structures: * * - target process descriptors, one per target process * - owner process descriptors (and attached lists), one per owner process * - lookup hash tables for owner and target processes * - PMC descriptors (and attached lists) * - per-cpu hardware state * - the 'hook' variable through which the kernel calls into * this module * - the machine hardware state (managed by the MD layer) * * These data structures are accessed from: * * - thread context-switch code * - interrupt handlers (possibly on multiple cpus) * - kernel threads on multiple cpus running on behalf of user * processes doing system calls * - this driver's private kernel threads * * = Locks and Locking strategy = * * The driver uses four locking strategies for its operation: * * - The global SX lock "pmc_sx" is used to protect internal * data structures. * * Calls into the module by syscall() start with this lock being * held in exclusive mode. Depending on the requested operation, * the lock may be downgraded to 'shared' mode to allow more * concurrent readers into the module. Calls into the module from * other parts of the kernel acquire the lock in shared mode. * * This SX lock is held in exclusive mode for any operations that * modify the linkages between the driver's internal data structures. * * The 'pmc_hook' function pointer is also protected by this lock. * It is only examined with the sx lock held in exclusive mode. The * kernel module is allowed to be unloaded only with the sx lock held * in exclusive mode. In normal syscall handling, after acquiring the * pmc_sx lock we first check that 'pmc_hook' is non-null before * proceeding. This prevents races between the thread unloading the module * and other threads seeking to use the module. * * - Lookups of target process structures and owner process structures * cannot use the global "pmc_sx" SX lock because these lookups need * to happen during context switches and in other critical sections * where sleeping is not allowed. We protect these lookup tables * with their own private spin-mutexes, "pmc_processhash_mtx" and * "pmc_ownerhash_mtx". * * - Interrupt handlers work in a lock free manner. At interrupt * time, handlers look at the PMC pointer (phw->phw_pmc) configured * when the PMC was started. If this pointer is NULL, the interrupt * is ignored after updating driver statistics. We ensure that this * pointer is set (using an atomic operation if necessary) before the * PMC hardware is started. Conversely, this pointer is unset atomically * only after the PMC hardware is stopped. * * We ensure that everything needed for the operation of an * interrupt handler is available without it needing to acquire any * locks. We also ensure that a PMC's software state is destroyed only * after the PMC is taken off hardware (on all CPUs). * * - Context-switch handling with process-private PMCs needs more * care. * * A given process may be the target of multiple PMCs. For example, * PMCATTACH and PMCDETACH may be requested by a process on one CPU * while the target process is running on another. A PMC could also * be getting released because its owner is exiting. We tackle * these situations in the following manner: * * - each target process structure 'pmc_process' has an array * of 'struct pmc *' pointers, one for each hardware PMC. * * - At context switch IN time, each "target" PMC in RUNNING state * gets started on hardware and a pointer to each PMC is copied into * the per-cpu phw array. The 'runcount' for the PMC is * incremented. * * - At context switch OUT time, all process-virtual PMCs are stopped * on hardware. The saved value is added to the PMCs value field * only if the PMC is in a non-deleted state (the PMCs state could * have changed during the current time slice). * * Note that since in-between a switch IN on a processor and a switch * OUT, the PMC could have been released on another CPU. Therefore * context switch OUT always looks at the hardware state to turn * OFF PMCs and will update a PMC's saved value only if reachable * from the target process record. * * - OP PMCRELEASE could be called on a PMC at any time (the PMC could * be attached to many processes at the time of the call and could * be active on multiple CPUs). * * We prevent further scheduling of the PMC by marking it as in * state 'DELETED'. If the runcount of the PMC is non-zero then * this PMC is currently running on a CPU somewhere. The thread * doing the PMCRELEASE operation waits by repeatedly doing a * pause() till the runcount comes to zero. * * The contents of a PMC descriptor (struct pmc) are protected using * a spin-mutex. In order to save space, we use a mutex pool. * * In terms of lock types used by witness(4), we use: * - Type "pmc-sx", used by the global SX lock. * - Type "pmc-sleep", for sleep mutexes used by logger threads. * - Type "pmc-per-proc", for protecting PMC owner descriptors. * - Type "pmc-leaf", used for all other spin mutexes. */ /* * save the cpu binding of the current kthread */ static void pmc_save_cpu_binding(struct pmc_binding *pb) { PMCDBG0(CPU,BND,2, "save-cpu"); thread_lock(curthread); pb->pb_bound = sched_is_bound(curthread); pb->pb_cpu = curthread->td_oncpu; thread_unlock(curthread); PMCDBG1(CPU,BND,2, "save-cpu cpu=%d", pb->pb_cpu); } /* * restore the cpu binding of the current thread */ static void pmc_restore_cpu_binding(struct pmc_binding *pb) { PMCDBG2(CPU,BND,2, "restore-cpu curcpu=%d restore=%d", curthread->td_oncpu, pb->pb_cpu); thread_lock(curthread); if (pb->pb_bound) sched_bind(curthread, pb->pb_cpu); else sched_unbind(curthread); thread_unlock(curthread); PMCDBG0(CPU,BND,2, "restore-cpu done"); } /* * move execution over the specified cpu and bind it there. */ static void pmc_select_cpu(int cpu) { KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), ("[pmc,%d] bad cpu number %d", __LINE__, cpu)); /* Never move to an inactive CPU. */ KASSERT(pmc_cpu_is_active(cpu), ("[pmc,%d] selecting inactive " "CPU %d", __LINE__, cpu)); PMCDBG1(CPU,SEL,2, "select-cpu cpu=%d", cpu); thread_lock(curthread); sched_bind(curthread, cpu); thread_unlock(curthread); KASSERT(curthread->td_oncpu == cpu, ("[pmc,%d] CPU not bound [cpu=%d, curr=%d]", __LINE__, cpu, curthread->td_oncpu)); PMCDBG1(CPU,SEL,2, "select-cpu cpu=%d ok", cpu); } /* * Force a context switch. * * We do this by pause'ing for 1 tick -- invoking mi_switch() is not * guaranteed to force a context switch. */ static void pmc_force_context_switch(void) { pause("pmcctx", 1); } /* * Get the file name for an executable. This is a simple wrapper * around vn_fullpath(9). */ static void pmc_getfilename(struct vnode *v, char **fullpath, char **freepath) { *fullpath = "unknown"; *freepath = NULL; vn_fullpath(curthread, v, fullpath, freepath); } /* * remove an process owning PMCs */ void pmc_remove_owner(struct pmc_owner *po) { struct pmc *pm, *tmp; sx_assert(&pmc_sx, SX_XLOCKED); PMCDBG1(OWN,ORM,1, "remove-owner po=%p", po); /* Remove descriptor from the owner hash table */ LIST_REMOVE(po, po_next); /* release all owned PMC descriptors */ LIST_FOREACH_SAFE(pm, &po->po_pmcs, pm_next, tmp) { PMCDBG1(OWN,ORM,2, "pmc=%p", pm); KASSERT(pm->pm_owner == po, ("[pmc,%d] owner %p != po %p", __LINE__, pm->pm_owner, po)); pmc_release_pmc_descriptor(pm); /* will unlink from the list */ pmc_destroy_pmc_descriptor(pm); } KASSERT(po->po_sscount == 0, ("[pmc,%d] SS count not zero", __LINE__)); KASSERT(LIST_EMPTY(&po->po_pmcs), ("[pmc,%d] PMC list not empty", __LINE__)); /* de-configure the log file if present */ if (po->po_flags & PMC_PO_OWNS_LOGFILE) pmclog_deconfigure_log(po); } /* * remove an owner process record if all conditions are met. */ static void pmc_maybe_remove_owner(struct pmc_owner *po) { PMCDBG1(OWN,OMR,1, "maybe-remove-owner po=%p", po); /* * Remove owner record if * - this process does not own any PMCs * - this process has not allocated a system-wide sampling buffer */ if (LIST_EMPTY(&po->po_pmcs) && ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)) { pmc_remove_owner(po); pmc_destroy_owner_descriptor(po); } } /* * Add an association between a target process and a PMC. */ static void pmc_link_target_process(struct pmc *pm, struct pmc_process *pp) { int ri; struct pmc_target *pt; #ifdef INVARIANTS struct pmc_thread *pt_td; #endif sx_assert(&pmc_sx, SX_XLOCKED); KASSERT(pm != NULL && pp != NULL, ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp)); KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)), ("[pmc,%d] Attaching a non-process-virtual pmc=%p to pid=%d", __LINE__, pm, pp->pp_proc->p_pid)); KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= ((int) md->pmd_npmc - 1), ("[pmc,%d] Illegal reference count %d for process record %p", __LINE__, pp->pp_refcnt, (void *) pp)); ri = PMC_TO_ROWINDEX(pm); PMCDBG3(PRC,TLK,1, "link-target pmc=%p ri=%d pmc-process=%p", pm, ri, pp); #ifdef HWPMC_DEBUG LIST_FOREACH(pt, &pm->pm_targets, pt_next) if (pt->pt_process == pp) KASSERT(0, ("[pmc,%d] pp %p already in pmc %p targets", __LINE__, pp, pm)); #endif pt = malloc(sizeof(struct pmc_target), M_PMC, M_WAITOK|M_ZERO); pt->pt_process = pp; LIST_INSERT_HEAD(&pm->pm_targets, pt, pt_next); atomic_store_rel_ptr((uintptr_t *)&pp->pp_pmcs[ri].pp_pmc, (uintptr_t)pm); if (pm->pm_owner->po_owner == pp->pp_proc) pm->pm_flags |= PMC_F_ATTACHED_TO_OWNER; /* * Initialize the per-process values at this row index. */ pp->pp_pmcs[ri].pp_pmcval = PMC_TO_MODE(pm) == PMC_MODE_TS ? pm->pm_sc.pm_reloadcount : 0; pp->pp_refcnt++; #ifdef INVARIANTS /* Confirm that the per-thread values at this row index are cleared. */ if (PMC_TO_MODE(pm) == PMC_MODE_TS) { mtx_lock_spin(pp->pp_tdslock); LIST_FOREACH(pt_td, &pp->pp_tds, pt_next) { KASSERT(pt_td->pt_pmcs[ri].pt_pmcval == (pmc_value_t) 0, ("[pmc,%d] pt_pmcval not cleared for pid=%d at " "ri=%d", __LINE__, pp->pp_proc->p_pid, ri)); } mtx_unlock_spin(pp->pp_tdslock); } #endif } /* * Removes the association between a target process and a PMC. */ static void pmc_unlink_target_process(struct pmc *pm, struct pmc_process *pp) { int ri; struct proc *p; struct pmc_target *ptgt; struct pmc_thread *pt; sx_assert(&pmc_sx, SX_XLOCKED); KASSERT(pm != NULL && pp != NULL, ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp)); KASSERT(pp->pp_refcnt >= 1 && pp->pp_refcnt <= (int) md->pmd_npmc, ("[pmc,%d] Illegal ref count %d on process record %p", __LINE__, pp->pp_refcnt, (void *) pp)); ri = PMC_TO_ROWINDEX(pm); PMCDBG3(PRC,TUL,1, "unlink-target pmc=%p ri=%d pmc-process=%p", pm, ri, pp); KASSERT(pp->pp_pmcs[ri].pp_pmc == pm, ("[pmc,%d] PMC ri %d mismatch pmc %p pp->[ri] %p", __LINE__, ri, pm, pp->pp_pmcs[ri].pp_pmc)); pp->pp_pmcs[ri].pp_pmc = NULL; pp->pp_pmcs[ri].pp_pmcval = (pmc_value_t) 0; /* Clear the per-thread values at this row index. */ if (PMC_TO_MODE(pm) == PMC_MODE_TS) { mtx_lock_spin(pp->pp_tdslock); LIST_FOREACH(pt, &pp->pp_tds, pt_next) pt->pt_pmcs[ri].pt_pmcval = (pmc_value_t) 0; mtx_unlock_spin(pp->pp_tdslock); } /* Remove owner-specific flags */ if (pm->pm_owner->po_owner == pp->pp_proc) { pp->pp_flags &= ~PMC_PP_ENABLE_MSR_ACCESS; pm->pm_flags &= ~PMC_F_ATTACHED_TO_OWNER; } pp->pp_refcnt--; /* Remove the target process from the PMC structure */ LIST_FOREACH(ptgt, &pm->pm_targets, pt_next) if (ptgt->pt_process == pp) break; KASSERT(ptgt != NULL, ("[pmc,%d] process %p (pp: %p) not found " "in pmc %p", __LINE__, pp->pp_proc, pp, pm)); LIST_REMOVE(ptgt, pt_next); free(ptgt, M_PMC); /* if the PMC now lacks targets, send the owner a SIGIO */ if (LIST_EMPTY(&pm->pm_targets)) { p = pm->pm_owner->po_owner; PROC_LOCK(p); kern_psignal(p, SIGIO); PROC_UNLOCK(p); PMCDBG2(PRC,SIG,2, "signalling proc=%p signal=%d", p, SIGIO); } } /* * Check if PMC 'pm' may be attached to target process 't'. */ static int pmc_can_attach(struct pmc *pm, struct proc *t) { struct proc *o; /* pmc owner */ struct ucred *oc, *tc; /* owner, target credentials */ int decline_attach, i; /* * A PMC's owner can always attach that PMC to itself. */ if ((o = pm->pm_owner->po_owner) == t) return 0; PROC_LOCK(o); oc = o->p_ucred; crhold(oc); PROC_UNLOCK(o); PROC_LOCK(t); tc = t->p_ucred; crhold(tc); PROC_UNLOCK(t); /* * The effective uid of the PMC owner should match at least one * of the {effective,real,saved} uids of the target process. */ decline_attach = oc->cr_uid != tc->cr_uid && oc->cr_uid != tc->cr_svuid && oc->cr_uid != tc->cr_ruid; /* * Every one of the target's group ids, must be in the owner's * group list. */ for (i = 0; !decline_attach && i < tc->cr_ngroups; i++) decline_attach = !groupmember(tc->cr_groups[i], oc); /* check the read and saved gids too */ if (decline_attach == 0) decline_attach = !groupmember(tc->cr_rgid, oc) || !groupmember(tc->cr_svgid, oc); crfree(tc); crfree(oc); return !decline_attach; } /* * Attach a process to a PMC. */ static int pmc_attach_one_process(struct proc *p, struct pmc *pm) { int ri, error; char *fullpath, *freepath; struct pmc_process *pp; sx_assert(&pmc_sx, SX_XLOCKED); PMCDBG5(PRC,ATT,2, "attach-one pm=%p ri=%d proc=%p (%d, %s)", pm, PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm); /* * Locate the process descriptor corresponding to process 'p', * allocating space as needed. * * Verify that rowindex 'pm_rowindex' is free in the process * descriptor. * * If not, allocate space for a descriptor and link the * process descriptor and PMC. */ ri = PMC_TO_ROWINDEX(pm); /* mark process as using HWPMCs */ PROC_LOCK(p); p->p_flag |= P_HWPMC; PROC_UNLOCK(p); if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_ALLOCATE)) == NULL) { error = ENOMEM; goto fail; } if (pp->pp_pmcs[ri].pp_pmc == pm) {/* already present at slot [ri] */ error = EEXIST; goto fail; } if (pp->pp_pmcs[ri].pp_pmc != NULL) { error = EBUSY; goto fail; } pmc_link_target_process(pm, pp); if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) && (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) == 0) pm->pm_flags |= PMC_F_NEEDS_LOGFILE; pm->pm_flags |= PMC_F_ATTACH_DONE; /* mark as attached */ /* issue an attach event to a configured log file */ if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE) { if (p->p_flag & P_KPROC) { fullpath = kernelname; freepath = NULL; } else { pmc_getfilename(p->p_textvp, &fullpath, &freepath); pmclog_process_pmcattach(pm, p->p_pid, fullpath); } free(freepath, M_TEMP); if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) pmc_log_process_mappings(pm->pm_owner, p); } return (0); fail: PROC_LOCK(p); p->p_flag &= ~P_HWPMC; PROC_UNLOCK(p); return (error); } /* * Attach a process and optionally its children */ static int pmc_attach_process(struct proc *p, struct pmc *pm) { int error; struct proc *top; sx_assert(&pmc_sx, SX_XLOCKED); PMCDBG5(PRC,ATT,1, "attach pm=%p ri=%d proc=%p (%d, %s)", pm, PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm); /* * If this PMC successfully allowed a GETMSR operation * in the past, disallow further ATTACHes. */ if ((pm->pm_flags & PMC_PP_ENABLE_MSR_ACCESS) != 0) return EPERM; if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0) return pmc_attach_one_process(p, pm); /* * Traverse all child processes, attaching them to * this PMC. */ sx_slock(&proctree_lock); top = p; for (;;) { if ((error = pmc_attach_one_process(p, pm)) != 0) break; if (!LIST_EMPTY(&p->p_children)) p = LIST_FIRST(&p->p_children); else for (;;) { if (p == top) goto done; if (LIST_NEXT(p, p_sibling)) { p = LIST_NEXT(p, p_sibling); break; } p = p->p_pptr; } } if (error) (void) pmc_detach_process(top, pm); done: sx_sunlock(&proctree_lock); return error; } /* * Detach a process from a PMC. If there are no other PMCs tracking * this process, remove the process structure from its hash table. If * 'flags' contains PMC_FLAG_REMOVE, then free the process structure. */ static int pmc_detach_one_process(struct proc *p, struct pmc *pm, int flags) { int ri; struct pmc_process *pp; sx_assert(&pmc_sx, SX_XLOCKED); KASSERT(pm != NULL, ("[pmc,%d] null pm pointer", __LINE__)); ri = PMC_TO_ROWINDEX(pm); PMCDBG6(PRC,ATT,2, "detach-one pm=%p ri=%d proc=%p (%d, %s) flags=0x%x", pm, ri, p, p->p_pid, p->p_comm, flags); if ((pp = pmc_find_process_descriptor(p, 0)) == NULL) return ESRCH; if (pp->pp_pmcs[ri].pp_pmc != pm) return EINVAL; pmc_unlink_target_process(pm, pp); /* Issue a detach entry if a log file is configured */ if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE) pmclog_process_pmcdetach(pm, p->p_pid); /* * If there are no PMCs targeting this process, we remove its * descriptor from the target hash table and unset the P_HWPMC * flag in the struct proc. */ KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= (int) md->pmd_npmc, ("[pmc,%d] Illegal refcnt %d for process struct %p", __LINE__, pp->pp_refcnt, pp)); if (pp->pp_refcnt != 0) /* still a target of some PMC */ return 0; pmc_remove_process_descriptor(pp); if (flags & PMC_FLAG_REMOVE) pmc_destroy_process_descriptor(pp); PROC_LOCK(p); p->p_flag &= ~P_HWPMC; PROC_UNLOCK(p); return 0; } /* * Detach a process and optionally its descendants from a PMC. */ static int pmc_detach_process(struct proc *p, struct pmc *pm) { struct proc *top; sx_assert(&pmc_sx, SX_XLOCKED); PMCDBG5(PRC,ATT,1, "detach pm=%p ri=%d proc=%p (%d, %s)", pm, PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm); if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0) return pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE); /* * Traverse all children, detaching them from this PMC. We * ignore errors since we could be detaching a PMC from a * partially attached proc tree. */ sx_slock(&proctree_lock); top = p; for (;;) { (void) pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE); if (!LIST_EMPTY(&p->p_children)) p = LIST_FIRST(&p->p_children); else for (;;) { if (p == top) goto done; if (LIST_NEXT(p, p_sibling)) { p = LIST_NEXT(p, p_sibling); break; } p = p->p_pptr; } } done: sx_sunlock(&proctree_lock); if (LIST_EMPTY(&pm->pm_targets)) pm->pm_flags &= ~PMC_F_ATTACH_DONE; return 0; } /* * Thread context switch IN */ static void pmc_process_csw_in(struct thread *td) { int cpu; unsigned int adjri, ri; struct pmc *pm; struct proc *p; struct pmc_cpu *pc; struct pmc_hw *phw; pmc_value_t newvalue; struct pmc_process *pp; struct pmc_thread *pt; struct pmc_classdep *pcd; p = td->td_proc; pt = NULL; if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE)) == NULL) return; KASSERT(pp->pp_proc == td->td_proc, ("[pmc,%d] not my thread state", __LINE__)); critical_enter(); /* no preemption from this point */ cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */ PMCDBG5(CSW,SWI,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p, p->p_pid, p->p_comm, pp); KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), ("[pmc,%d] weird CPU id %d", __LINE__, cpu)); pc = pmc_pcpu[cpu]; for (ri = 0; ri < md->pmd_npmc; ri++) { if ((pm = pp->pp_pmcs[ri].pp_pmc) == NULL) continue; KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)), ("[pmc,%d] Target PMC in non-virtual mode (%d)", __LINE__, PMC_TO_MODE(pm))); KASSERT(PMC_TO_ROWINDEX(pm) == ri, ("[pmc,%d] Row index mismatch pmc %d != ri %d", __LINE__, PMC_TO_ROWINDEX(pm), ri)); /* * Only PMCs that are marked as 'RUNNING' need * be placed on hardware. */ if (pm->pm_state != PMC_STATE_RUNNING) continue; /* increment PMC runcount */ counter_u64_add(pm->pm_runcount, 1); /* configure the HWPMC we are going to use. */ pcd = pmc_ri_to_classdep(md, ri, &adjri); pcd->pcd_config_pmc(cpu, adjri, pm); phw = pc->pc_hwpmcs[ri]; KASSERT(phw != NULL, ("[pmc,%d] null hw pointer", __LINE__)); KASSERT(phw->phw_pmc == pm, ("[pmc,%d] hw->pmc %p != pmc %p", __LINE__, phw->phw_pmc, pm)); /* * Write out saved value and start the PMC. * * Sampling PMCs use a per-thread value, while * counting mode PMCs use a per-pmc value that is * inherited across descendants. */ if (PMC_TO_MODE(pm) == PMC_MODE_TS) { if (pt == NULL) pt = pmc_find_thread_descriptor(pp, td, PMC_FLAG_NONE); KASSERT(pt != NULL, ("[pmc,%d] No thread found for td=%p", __LINE__, td)); mtx_pool_lock_spin(pmc_mtxpool, pm); /* * If we have a thread descriptor, use the per-thread * counter in the descriptor. If not, we will use * a per-process counter. * * TODO: Remove the per-process "safety net" once * we have thoroughly tested that we don't hit the * above assert. */ if (pt != NULL) { if (pt->pt_pmcs[ri].pt_pmcval > 0) newvalue = pt->pt_pmcs[ri].pt_pmcval; else newvalue = pm->pm_sc.pm_reloadcount; } else { /* * Use the saved value calculated after the most * recent time a thread using the shared counter * switched out. Reset the saved count in case * another thread from this process switches in * before any threads switch out. */ newvalue = pp->pp_pmcs[ri].pp_pmcval; pp->pp_pmcs[ri].pp_pmcval = pm->pm_sc.pm_reloadcount; } mtx_pool_unlock_spin(pmc_mtxpool, pm); KASSERT(newvalue > 0 && newvalue <= pm->pm_sc.pm_reloadcount, ("[pmc,%d] pmcval outside of expected range cpu=%d " "ri=%d pmcval=%jx pm_reloadcount=%jx", __LINE__, cpu, ri, newvalue, pm->pm_sc.pm_reloadcount)); } else { KASSERT(PMC_TO_MODE(pm) == PMC_MODE_TC, ("[pmc,%d] illegal mode=%d", __LINE__, PMC_TO_MODE(pm))); mtx_pool_lock_spin(pmc_mtxpool, pm); newvalue = PMC_PCPU_SAVED(cpu, ri) = pm->pm_gv.pm_savedvalue; mtx_pool_unlock_spin(pmc_mtxpool, pm); } PMCDBG3(CSW,SWI,1,"cpu=%d ri=%d new=%jd", cpu, ri, newvalue); pcd->pcd_write_pmc(cpu, adjri, newvalue); /* If a sampling mode PMC, reset stalled state. */ if (PMC_TO_MODE(pm) == PMC_MODE_TS) pm->pm_pcpu_state[cpu].pps_stalled = 0; /* Indicate that we desire this to run. */ pm->pm_pcpu_state[cpu].pps_cpustate = 1; /* Start the PMC. */ pcd->pcd_start_pmc(cpu, adjri); } /* * perform any other architecture/cpu dependent thread * switch-in actions. */ (void) (*md->pmd_switch_in)(pc, pp); critical_exit(); } /* * Thread context switch OUT. */ static void pmc_process_csw_out(struct thread *td) { int cpu; int64_t tmp; struct pmc *pm; struct proc *p; enum pmc_mode mode; struct pmc_cpu *pc; pmc_value_t newvalue; unsigned int adjri, ri; struct pmc_process *pp; struct pmc_thread *pt = NULL; struct pmc_classdep *pcd; /* * Locate our process descriptor; this may be NULL if * this process is exiting and we have already removed * the process from the target process table. * * Note that due to kernel preemption, multiple * context switches may happen while the process is * exiting. * * Note also that if the target process cannot be * found we still need to deconfigure any PMCs that * are currently running on hardware. */ p = td->td_proc; pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE); /* * save PMCs */ critical_enter(); cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */ PMCDBG5(CSW,SWO,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p, p->p_pid, p->p_comm, pp); KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), ("[pmc,%d weird CPU id %d", __LINE__, cpu)); pc = pmc_pcpu[cpu]; /* * When a PMC gets unlinked from a target PMC, it will * be removed from the target's pp_pmc[] array. * * However, on a MP system, the target could have been * executing on another CPU at the time of the unlink. * So, at context switch OUT time, we need to look at * the hardware to determine if a PMC is scheduled on * it. */ for (ri = 0; ri < md->pmd_npmc; ri++) { pcd = pmc_ri_to_classdep(md, ri, &adjri); pm = NULL; (void) (*pcd->pcd_get_config)(cpu, adjri, &pm); if (pm == NULL) /* nothing at this row index */ continue; mode = PMC_TO_MODE(pm); if (!PMC_IS_VIRTUAL_MODE(mode)) continue; /* not a process virtual PMC */ KASSERT(PMC_TO_ROWINDEX(pm) == ri, ("[pmc,%d] ri mismatch pmc(%d) ri(%d)", __LINE__, PMC_TO_ROWINDEX(pm), ri)); /* * Change desired state, and then stop if not stalled. * This two-step dance should avoid race conditions where * an interrupt re-enables the PMC after this code has * already checked the pm_stalled flag. */ pm->pm_pcpu_state[cpu].pps_cpustate = 0; if (pm->pm_pcpu_state[cpu].pps_stalled == 0) pcd->pcd_stop_pmc(cpu, adjri); /* reduce this PMC's runcount */ counter_u64_add(pm->pm_runcount, -1); /* * If this PMC is associated with this process, * save the reading. */ if (pm->pm_state != PMC_STATE_DELETED && pp != NULL && pp->pp_pmcs[ri].pp_pmc != NULL) { KASSERT(pm == pp->pp_pmcs[ri].pp_pmc, ("[pmc,%d] pm %p != pp_pmcs[%d] %p", __LINE__, pm, ri, pp->pp_pmcs[ri].pp_pmc)); KASSERT(pp->pp_refcnt > 0, ("[pmc,%d] pp refcnt = %d", __LINE__, pp->pp_refcnt)); pcd->pcd_read_pmc(cpu, adjri, &newvalue); if (mode == PMC_MODE_TS) { PMCDBG3(CSW,SWO,1,"cpu=%d ri=%d val=%jd (samp)", cpu, ri, newvalue); if (pt == NULL) pt = pmc_find_thread_descriptor(pp, td, PMC_FLAG_NONE); KASSERT(pt != NULL, ("[pmc,%d] No thread found for td=%p", __LINE__, td)); mtx_pool_lock_spin(pmc_mtxpool, pm); /* * If we have a thread descriptor, save the * per-thread counter in the descriptor. If not, * we will update the per-process counter. * * TODO: Remove the per-process "safety net" * once we have thoroughly tested that we * don't hit the above assert. */ if (pt != NULL) pt->pt_pmcs[ri].pt_pmcval = newvalue; else { /* * For sampling process-virtual PMCs, * newvalue is the number of events to * be seen until the next sampling * interrupt. We can just add the events * left from this invocation to the * counter, then adjust in case we * overflow our range. * * (Recall that we reload the counter * every time we use it.) */ pp->pp_pmcs[ri].pp_pmcval += newvalue; if (pp->pp_pmcs[ri].pp_pmcval > pm->pm_sc.pm_reloadcount) pp->pp_pmcs[ri].pp_pmcval -= pm->pm_sc.pm_reloadcount; } mtx_pool_unlock_spin(pmc_mtxpool, pm); } else { tmp = newvalue - PMC_PCPU_SAVED(cpu,ri); PMCDBG3(CSW,SWO,1,"cpu=%d ri=%d tmp=%jd (count)", cpu, ri, tmp); /* * For counting process-virtual PMCs, * we expect the count to be * increasing monotonically, modulo a 64 * bit wraparound. */ KASSERT(tmp >= 0, ("[pmc,%d] negative increment cpu=%d " "ri=%d newvalue=%jx saved=%jx " "incr=%jx", __LINE__, cpu, ri, newvalue, PMC_PCPU_SAVED(cpu,ri), tmp)); mtx_pool_lock_spin(pmc_mtxpool, pm); pm->pm_gv.pm_savedvalue += tmp; pp->pp_pmcs[ri].pp_pmcval += tmp; mtx_pool_unlock_spin(pmc_mtxpool, pm); if (pm->pm_flags & PMC_F_LOG_PROCCSW) pmclog_process_proccsw(pm, pp, tmp, td); } } /* mark hardware as free */ pcd->pcd_config_pmc(cpu, adjri, NULL); } /* * perform any other architecture/cpu dependent thread * switch out functions. */ (void) (*md->pmd_switch_out)(pc, pp); critical_exit(); } /* * A new thread for a process. */ static void pmc_process_thread_add(struct thread *td) { struct pmc_process *pmc; pmc = pmc_find_process_descriptor(td->td_proc, PMC_FLAG_NONE); if (pmc != NULL) pmc_find_thread_descriptor(pmc, td, PMC_FLAG_ALLOCATE); } /* * A thread delete for a process. */ static void pmc_process_thread_delete(struct thread *td) { struct pmc_process *pmc; pmc = pmc_find_process_descriptor(td->td_proc, PMC_FLAG_NONE); if (pmc != NULL) pmc_thread_descriptor_pool_free(pmc_find_thread_descriptor(pmc, td, PMC_FLAG_REMOVE)); } /* * A mapping change for a process. */ static void pmc_process_mmap(struct thread *td, struct pmckern_map_in *pkm) { int ri; pid_t pid; char *fullpath, *freepath; const struct pmc *pm; struct pmc_owner *po; const struct pmc_process *pp; freepath = fullpath = NULL; - epoch_exit_preempt(global_epoch_preempt); + MPASS(!in_epoch()); pmc_getfilename((struct vnode *) pkm->pm_file, &fullpath, &freepath); pid = td->td_proc->p_pid; epoch_enter_preempt(global_epoch_preempt); /* Inform owners of all system-wide sampling PMCs. */ CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) if (po->po_flags & PMC_PO_OWNS_LOGFILE) pmclog_process_map_in(po, pid, pkm->pm_address, fullpath); if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL) goto done; /* * Inform sampling PMC owners tracking this process. */ for (ri = 0; ri < md->pmd_npmc; ri++) if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL && PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) pmclog_process_map_in(pm->pm_owner, pid, pkm->pm_address, fullpath); done: if (freepath) free(freepath, M_TEMP); + epoch_exit_preempt(global_epoch_preempt); } /* * Log an munmap request. */ static void pmc_process_munmap(struct thread *td, struct pmckern_map_out *pkm) { int ri; pid_t pid; struct pmc_owner *po; const struct pmc *pm; const struct pmc_process *pp; pid = td->td_proc->p_pid; epoch_enter_preempt(global_epoch_preempt); CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) if (po->po_flags & PMC_PO_OWNS_LOGFILE) pmclog_process_map_out(po, pid, pkm->pm_address, pkm->pm_address + pkm->pm_size); epoch_exit_preempt(global_epoch_preempt); if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL) return; for (ri = 0; ri < md->pmd_npmc; ri++) if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL && PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) pmclog_process_map_out(pm->pm_owner, pid, pkm->pm_address, pkm->pm_address + pkm->pm_size); } /* * Log mapping information about the kernel. */ static void pmc_log_kernel_mappings(struct pmc *pm) { struct pmc_owner *po; struct pmckern_map_in *km, *kmbase; MPASS(in_epoch() || sx_xlocked(&pmc_sx)); KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)), ("[pmc,%d] non-sampling PMC (%p) desires mapping information", __LINE__, (void *) pm)); po = pm->pm_owner; if (po->po_flags & PMC_PO_INITIAL_MAPPINGS_DONE) return; /* * Log the current set of kernel modules. */ kmbase = linker_hwpmc_list_objects(); for (km = kmbase; km->pm_file != NULL; km++) { PMCDBG2(LOG,REG,1,"%s %p", (char *) km->pm_file, (void *) km->pm_address); pmclog_process_map_in(po, (pid_t) -1, km->pm_address, km->pm_file); } free(kmbase, M_LINKER); po->po_flags |= PMC_PO_INITIAL_MAPPINGS_DONE; } /* * Log the mappings for a single process. */ static void pmc_log_process_mappings(struct pmc_owner *po, struct proc *p) { vm_map_t map; struct vnode *vp; struct vmspace *vm; vm_map_entry_t entry; vm_offset_t last_end; u_int last_timestamp; struct vnode *last_vp; vm_offset_t start_addr; vm_object_t obj, lobj, tobj; char *fullpath, *freepath; last_vp = NULL; last_end = (vm_offset_t) 0; fullpath = freepath = NULL; if ((vm = vmspace_acquire_ref(p)) == NULL) return; map = &vm->vm_map; vm_map_lock_read(map); for (entry = map->header.next; entry != &map->header; entry = entry->next) { if (entry == NULL) { PMCDBG2(LOG,OPS,2, "hwpmc: vm_map entry unexpectedly " "NULL! pid=%d vm_map=%p\n", p->p_pid, map); break; } /* * We only care about executable map entries. */ if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) || !(entry->protection & VM_PROT_EXECUTE) || (entry->object.vm_object == NULL)) { continue; } obj = entry->object.vm_object; VM_OBJECT_RLOCK(obj); /* * Walk the backing_object list to find the base * (non-shadowed) vm_object. */ for (lobj = tobj = obj; tobj != NULL; tobj = tobj->backing_object) { if (tobj != obj) VM_OBJECT_RLOCK(tobj); if (lobj != obj) VM_OBJECT_RUNLOCK(lobj); lobj = tobj; } /* * At this point lobj is the base vm_object and it is locked. */ if (lobj == NULL) { PMCDBG3(LOG,OPS,2, "hwpmc: lobj unexpectedly NULL! pid=%d " "vm_map=%p vm_obj=%p\n", p->p_pid, map, obj); VM_OBJECT_RUNLOCK(obj); continue; } vp = vm_object_vnode(lobj); if (vp == NULL) { if (lobj != obj) VM_OBJECT_RUNLOCK(lobj); VM_OBJECT_RUNLOCK(obj); continue; } /* * Skip contiguous regions that point to the same * vnode, so we don't emit redundant MAP-IN * directives. */ if (entry->start == last_end && vp == last_vp) { last_end = entry->end; if (lobj != obj) VM_OBJECT_RUNLOCK(lobj); VM_OBJECT_RUNLOCK(obj); continue; } /* * We don't want to keep the proc's vm_map or this * vm_object locked while we walk the pathname, since * vn_fullpath() can sleep. However, if we drop the * lock, it's possible for concurrent activity to * modify the vm_map list. To protect against this, * we save the vm_map timestamp before we release the * lock, and check it after we reacquire the lock * below. */ start_addr = entry->start; last_end = entry->end; last_timestamp = map->timestamp; vm_map_unlock_read(map); vref(vp); if (lobj != obj) VM_OBJECT_RUNLOCK(lobj); VM_OBJECT_RUNLOCK(obj); freepath = NULL; pmc_getfilename(vp, &fullpath, &freepath); last_vp = vp; vrele(vp); vp = NULL; pmclog_process_map_in(po, p->p_pid, start_addr, fullpath); if (freepath) free(freepath, M_TEMP); vm_map_lock_read(map); /* * If our saved timestamp doesn't match, this means * that the vm_map was modified out from under us and * we can't trust our current "entry" pointer. Do a * new lookup for this entry. If there is no entry * for this address range, vm_map_lookup_entry() will * return the previous one, so we always want to go to * entry->next on the next loop iteration. * * There is an edge condition here that can occur if * there is no entry at or before this address. In * this situation, vm_map_lookup_entry returns * &map->header, which would cause our loop to abort * without processing the rest of the map. However, * in practice this will never happen for process * vm_map. This is because the executable's text * segment is the first mapping in the proc's address * space, and this mapping is never removed until the * process exits, so there will always be a non-header * entry at or before the requested address for * vm_map_lookup_entry to return. */ if (map->timestamp != last_timestamp) vm_map_lookup_entry(map, last_end - 1, &entry); } vm_map_unlock_read(map); vmspace_free(vm); return; } /* * Log mappings for all processes in the system. */ static void pmc_log_all_process_mappings(struct pmc_owner *po) { struct proc *p, *top; sx_assert(&pmc_sx, SX_XLOCKED); if ((p = pfind(1)) == NULL) panic("[pmc,%d] Cannot find init", __LINE__); PROC_UNLOCK(p); sx_slock(&proctree_lock); top = p; for (;;) { pmc_log_process_mappings(po, p); if (!LIST_EMPTY(&p->p_children)) p = LIST_FIRST(&p->p_children); else for (;;) { if (p == top) goto done; if (LIST_NEXT(p, p_sibling)) { p = LIST_NEXT(p, p_sibling); break; } p = p->p_pptr; } } done: sx_sunlock(&proctree_lock); } /* * The 'hook' invoked from the kernel proper */ #ifdef HWPMC_DEBUG const char *pmc_hooknames[] = { /* these strings correspond to PMC_FN_* in */ "", "EXEC", "CSW-IN", "CSW-OUT", "SAMPLE", "UNUSED1", "UNUSED2", "MMAP", "MUNMAP", "CALLCHAIN-NMI", "CALLCHAIN-SOFT", "SOFTSAMPLING", "THR-CREATE", "THR-EXIT", }; #endif static int pmc_hook_handler(struct thread *td, int function, void *arg) { int cpu; PMCDBG4(MOD,PMH,1, "hook td=%p func=%d \"%s\" arg=%p", td, function, pmc_hooknames[function], arg); switch (function) { /* * Process exec() */ case PMC_FN_PROCESS_EXEC: { char *fullpath, *freepath; unsigned int ri; int is_using_hwpmcs; struct pmc *pm; struct proc *p; struct pmc_owner *po; struct pmc_process *pp; struct pmckern_procexec *pk; sx_assert(&pmc_sx, SX_XLOCKED); p = td->td_proc; pmc_getfilename(p->p_textvp, &fullpath, &freepath); pk = (struct pmckern_procexec *) arg; epoch_enter_preempt(global_epoch_preempt); /* Inform owners of SS mode PMCs of the exec event. */ CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) if (po->po_flags & PMC_PO_OWNS_LOGFILE) pmclog_process_procexec(po, PMC_ID_INVALID, p->p_pid, pk->pm_entryaddr, fullpath); epoch_exit_preempt(global_epoch_preempt); PROC_LOCK(p); is_using_hwpmcs = p->p_flag & P_HWPMC; PROC_UNLOCK(p); if (!is_using_hwpmcs) { if (freepath) free(freepath, M_TEMP); break; } /* * PMCs are not inherited across an exec(): remove any * PMCs that this process is the owner of. */ if ((po = pmc_find_owner_descriptor(p)) != NULL) { pmc_remove_owner(po); pmc_destroy_owner_descriptor(po); } /* * If the process being exec'ed is not the target of any * PMC, we are done. */ if ((pp = pmc_find_process_descriptor(p, 0)) == NULL) { if (freepath) free(freepath, M_TEMP); break; } /* * Log the exec event to all monitoring owners. Skip * owners who have already received the event because * they had system sampling PMCs active. */ for (ri = 0; ri < md->pmd_npmc; ri++) if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) { po = pm->pm_owner; if (po->po_sscount == 0 && po->po_flags & PMC_PO_OWNS_LOGFILE) pmclog_process_procexec(po, pm->pm_id, p->p_pid, pk->pm_entryaddr, fullpath); } if (freepath) free(freepath, M_TEMP); PMCDBG4(PRC,EXC,1, "exec proc=%p (%d, %s) cred-changed=%d", p, p->p_pid, p->p_comm, pk->pm_credentialschanged); if (pk->pm_credentialschanged == 0) /* no change */ break; /* * If the newly exec()'ed process has a different credential * than before, allow it to be the target of a PMC only if * the PMC's owner has sufficient privilege. */ for (ri = 0; ri < md->pmd_npmc; ri++) if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) if (pmc_can_attach(pm, td->td_proc) != 0) pmc_detach_one_process(td->td_proc, pm, PMC_FLAG_NONE); KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= (int) md->pmd_npmc, ("[pmc,%d] Illegal ref count %d on pp %p", __LINE__, pp->pp_refcnt, pp)); /* * If this process is no longer the target of any * PMCs, we can remove the process entry and free * up space. */ if (pp->pp_refcnt == 0) { pmc_remove_process_descriptor(pp); pmc_destroy_process_descriptor(pp); break; } } break; case PMC_FN_CSW_IN: pmc_process_csw_in(td); break; case PMC_FN_CSW_OUT: pmc_process_csw_out(td); break; /* * Process accumulated PC samples. * * This function is expected to be called by hardclock() for * each CPU that has accumulated PC samples. * * This function is to be executed on the CPU whose samples * are being processed. */ case PMC_FN_DO_SAMPLES: /* * Clear the cpu specific bit in the CPU mask before * do the rest of the processing. If the NMI handler * gets invoked after the "atomic_clear_int()" call * below but before "pmc_process_samples()" gets * around to processing the interrupt, then we will * come back here at the next hardclock() tick (and * may find nothing to do if "pmc_process_samples()" * had already processed the interrupt). We don't * lose the interrupt sample. */ DPCPU_SET(pmc_sampled, 0); cpu = PCPU_GET(cpuid); pmc_process_samples(cpu, PMC_HR); pmc_process_samples(cpu, PMC_SR); break; case PMC_FN_MMAP: MPASS(in_epoch() || sx_xlocked(&pmc_sx)); pmc_process_mmap(td, (struct pmckern_map_in *) arg); break; case PMC_FN_MUNMAP: MPASS(in_epoch() || sx_xlocked(&pmc_sx)); pmc_process_munmap(td, (struct pmckern_map_out *) arg); break; case PMC_FN_USER_CALLCHAIN: /* * Record a call chain. */ KASSERT(td == curthread, ("[pmc,%d] td != curthread", __LINE__)); pmc_capture_user_callchain(PCPU_GET(cpuid), PMC_HR, (struct trapframe *) arg); td->td_pflags &= ~TDP_CALLCHAIN; break; case PMC_FN_USER_CALLCHAIN_SOFT: /* * Record a call chain. */ KASSERT(td == curthread, ("[pmc,%d] td != curthread", __LINE__)); pmc_capture_user_callchain(PCPU_GET(cpuid), PMC_SR, (struct trapframe *) arg); td->td_pflags &= ~TDP_CALLCHAIN; break; case PMC_FN_SOFT_SAMPLING: /* * Call soft PMC sampling intr. */ pmc_soft_intr((struct pmckern_soft *) arg); break; case PMC_FN_THR_CREATE: pmc_process_thread_add(td); break; case PMC_FN_THR_EXIT: KASSERT(td == curthread, ("[pmc,%d] td != curthread", __LINE__)); pmc_process_thread_delete(td); break; default: #ifdef HWPMC_DEBUG KASSERT(0, ("[pmc,%d] unknown hook %d\n", __LINE__, function)); #endif break; } return 0; } /* * allocate a 'struct pmc_owner' descriptor in the owner hash table. */ static struct pmc_owner * pmc_allocate_owner_descriptor(struct proc *p) { uint32_t hindex; struct pmc_owner *po; struct pmc_ownerhash *poh; hindex = PMC_HASH_PTR(p, pmc_ownerhashmask); poh = &pmc_ownerhash[hindex]; /* allocate space for N pointers and one descriptor struct */ po = malloc(sizeof(struct pmc_owner), M_PMC, M_WAITOK|M_ZERO); po->po_owner = p; LIST_INSERT_HEAD(poh, po, po_next); /* insert into hash table */ TAILQ_INIT(&po->po_logbuffers); mtx_init(&po->po_mtx, "pmc-owner-mtx", "pmc-per-proc", MTX_SPIN); PMCDBG4(OWN,ALL,1, "allocate-owner proc=%p (%d, %s) pmc-owner=%p", p, p->p_pid, p->p_comm, po); return po; } static void pmc_destroy_owner_descriptor(struct pmc_owner *po) { PMCDBG4(OWN,REL,1, "destroy-owner po=%p proc=%p (%d, %s)", po, po->po_owner, po->po_owner->p_pid, po->po_owner->p_comm); mtx_destroy(&po->po_mtx); free(po, M_PMC); } /* * Allocate a thread descriptor from the free pool. * * NOTE: This *can* return NULL. */ static struct pmc_thread * pmc_thread_descriptor_pool_alloc(void) { struct pmc_thread *pt; mtx_lock_spin(&pmc_threadfreelist_mtx); if ((pt = LIST_FIRST(&pmc_threadfreelist)) != NULL) { LIST_REMOVE(pt, pt_next); pmc_threadfreelist_entries--; } mtx_unlock_spin(&pmc_threadfreelist_mtx); return (pt); } /* * Add a thread descriptor to the free pool. We use this instead of free() * to maintain a cache of free entries. Additionally, we can safely call * this function when we cannot call free(), such as in a critical section. * */ static void pmc_thread_descriptor_pool_free(struct pmc_thread *pt) { if (pt == NULL) return; memset(pt, 0, THREADENTRY_SIZE); mtx_lock_spin(&pmc_threadfreelist_mtx); LIST_INSERT_HEAD(&pmc_threadfreelist, pt, pt_next); pmc_threadfreelist_entries++; if (pmc_threadfreelist_entries > pmc_threadfreelist_max) GROUPTASK_ENQUEUE(&free_gtask); mtx_unlock_spin(&pmc_threadfreelist_mtx); } /* * A callout to manage the free list. */ static void pmc_thread_descriptor_pool_free_task(void *arg __unused) { struct pmc_thread *pt; LIST_HEAD(, pmc_thread) tmplist; int delta; LIST_INIT(&tmplist); /* Determine what changes, if any, we need to make. */ mtx_lock_spin(&pmc_threadfreelist_mtx); delta = pmc_threadfreelist_entries - pmc_threadfreelist_max; while (delta > 0 && (pt = LIST_FIRST(&pmc_threadfreelist)) != NULL) { delta--; LIST_REMOVE(pt, pt_next); LIST_INSERT_HEAD(&tmplist, pt, pt_next); } mtx_unlock_spin(&pmc_threadfreelist_mtx); /* If there are entries to free, free them. */ while (!LIST_EMPTY(&tmplist)) { pt = LIST_FIRST(&tmplist); LIST_REMOVE(pt, pt_next); free(pt, M_PMC); } } /* * Drain the thread free pool, freeing all allocations. */ static void pmc_thread_descriptor_pool_drain() { struct pmc_thread *pt, *next; LIST_FOREACH_SAFE(pt, &pmc_threadfreelist, pt_next, next) { LIST_REMOVE(pt, pt_next); free(pt, M_PMC); } } /* * find the descriptor corresponding to thread 'td', adding or removing it * as specified by 'mode'. * * Note that this supports additional mode flags in addition to those * supported by pmc_find_process_descriptor(): * PMC_FLAG_NOWAIT: Causes the function to not wait for mallocs. * This makes it safe to call while holding certain other locks. */ static struct pmc_thread * pmc_find_thread_descriptor(struct pmc_process *pp, struct thread *td, uint32_t mode) { struct pmc_thread *pt = NULL, *ptnew = NULL; int wait_flag; KASSERT(td != NULL, ("[pmc,%d] called to add NULL td", __LINE__)); /* * Pre-allocate memory in the PMC_FLAG_ALLOCATE case prior to * acquiring the lock. */ if (mode & PMC_FLAG_ALLOCATE) { if ((ptnew = pmc_thread_descriptor_pool_alloc()) == NULL) { wait_flag = M_WAITOK; if ((mode & PMC_FLAG_NOWAIT) || in_epoch()) wait_flag = M_NOWAIT; ptnew = malloc(THREADENTRY_SIZE, M_PMC, wait_flag|M_ZERO); } } mtx_lock_spin(pp->pp_tdslock); LIST_FOREACH(pt, &pp->pp_tds, pt_next) if (pt->pt_td == td) break; if ((mode & PMC_FLAG_REMOVE) && pt != NULL) LIST_REMOVE(pt, pt_next); if ((mode & PMC_FLAG_ALLOCATE) && pt == NULL && ptnew != NULL) { pt = ptnew; ptnew = NULL; pt->pt_td = td; LIST_INSERT_HEAD(&pp->pp_tds, pt, pt_next); } mtx_unlock_spin(pp->pp_tdslock); if (ptnew != NULL) { free(ptnew, M_PMC); } return pt; } /* * Try to add thread descriptors for each thread in a process. */ static void pmc_add_thread_descriptors_from_proc(struct proc *p, struct pmc_process *pp) { struct thread *curtd; struct pmc_thread **tdlist; int i, tdcnt, tdlistsz; KASSERT(!PROC_LOCKED(p), ("[pmc,%d] proc unexpectedly locked", __LINE__)); tdcnt = 32; restart: tdlistsz = roundup2(tdcnt, 32); tdcnt = 0; tdlist = malloc(sizeof(struct pmc_thread*) * tdlistsz, M_TEMP, M_WAITOK); PROC_LOCK(p); FOREACH_THREAD_IN_PROC(p, curtd) tdcnt++; if (tdcnt >= tdlistsz) { PROC_UNLOCK(p); free(tdlist, M_TEMP); goto restart; } /* * Try to add each thread to the list without sleeping. If unable, * add to a queue to retry after dropping the process lock. */ tdcnt = 0; FOREACH_THREAD_IN_PROC(p, curtd) { tdlist[tdcnt] = pmc_find_thread_descriptor(pp, curtd, PMC_FLAG_ALLOCATE|PMC_FLAG_NOWAIT); if (tdlist[tdcnt] == NULL) { PROC_UNLOCK(p); for (i = 0; i <= tdcnt; i++) pmc_thread_descriptor_pool_free(tdlist[i]); free(tdlist, M_TEMP); goto restart; } tdcnt++; } PROC_UNLOCK(p); free(tdlist, M_TEMP); } /* * find the descriptor corresponding to process 'p', adding or removing it * as specified by 'mode'. */ static struct pmc_process * pmc_find_process_descriptor(struct proc *p, uint32_t mode) { uint32_t hindex; struct pmc_process *pp, *ppnew; struct pmc_processhash *pph; hindex = PMC_HASH_PTR(p, pmc_processhashmask); pph = &pmc_processhash[hindex]; ppnew = NULL; /* * Pre-allocate memory in the PMC_FLAG_ALLOCATE case since we * cannot call malloc(9) once we hold a spin lock. */ if (mode & PMC_FLAG_ALLOCATE) ppnew = malloc(sizeof(struct pmc_process) + md->pmd_npmc * sizeof(struct pmc_targetstate), M_PMC, M_WAITOK|M_ZERO); mtx_lock_spin(&pmc_processhash_mtx); LIST_FOREACH(pp, pph, pp_next) if (pp->pp_proc == p) break; if ((mode & PMC_FLAG_REMOVE) && pp != NULL) LIST_REMOVE(pp, pp_next); if ((mode & PMC_FLAG_ALLOCATE) && pp == NULL && ppnew != NULL) { ppnew->pp_proc = p; LIST_INIT(&ppnew->pp_tds); ppnew->pp_tdslock = mtx_pool_find(pmc_mtxpool, ppnew); LIST_INSERT_HEAD(pph, ppnew, pp_next); mtx_unlock_spin(&pmc_processhash_mtx); pp = ppnew; ppnew = NULL; /* Add thread descriptors for this process' current threads. */ pmc_add_thread_descriptors_from_proc(p, pp); } else mtx_unlock_spin(&pmc_processhash_mtx); if (ppnew != NULL) free(ppnew, M_PMC); return pp; } /* * remove a process descriptor from the process hash table. */ static void pmc_remove_process_descriptor(struct pmc_process *pp) { KASSERT(pp->pp_refcnt == 0, ("[pmc,%d] Removing process descriptor %p with count %d", __LINE__, pp, pp->pp_refcnt)); mtx_lock_spin(&pmc_processhash_mtx); LIST_REMOVE(pp, pp_next); mtx_unlock_spin(&pmc_processhash_mtx); } /* * destroy a process descriptor. */ static void pmc_destroy_process_descriptor(struct pmc_process *pp) { struct pmc_thread *pmc_td; while ((pmc_td = LIST_FIRST(&pp->pp_tds)) != NULL) { LIST_REMOVE(pmc_td, pt_next); pmc_thread_descriptor_pool_free(pmc_td); } free(pp, M_PMC); } /* * find an owner descriptor corresponding to proc 'p' */ static struct pmc_owner * pmc_find_owner_descriptor(struct proc *p) { uint32_t hindex; struct pmc_owner *po; struct pmc_ownerhash *poh; hindex = PMC_HASH_PTR(p, pmc_ownerhashmask); poh = &pmc_ownerhash[hindex]; po = NULL; LIST_FOREACH(po, poh, po_next) if (po->po_owner == p) break; PMCDBG5(OWN,FND,1, "find-owner proc=%p (%d, %s) hindex=0x%x -> " "pmc-owner=%p", p, p->p_pid, p->p_comm, hindex, po); return po; } /* * pmc_allocate_pmc_descriptor * * Allocate a pmc descriptor and initialize its * fields. */ static struct pmc * pmc_allocate_pmc_descriptor(void) { struct pmc *pmc; pmc = malloc(sizeof(struct pmc), M_PMC, M_WAITOK|M_ZERO); pmc->pm_runcount = counter_u64_alloc(M_WAITOK); pmc->pm_pcpu_state = malloc(sizeof(struct pmc_pcpu_state)*mp_ncpus, M_PMC, M_WAITOK|M_ZERO); PMCDBG1(PMC,ALL,1, "allocate-pmc -> pmc=%p", pmc); return pmc; } /* * Destroy a pmc descriptor. */ static void pmc_destroy_pmc_descriptor(struct pmc *pm) { KASSERT(pm->pm_state == PMC_STATE_DELETED || pm->pm_state == PMC_STATE_FREE, ("[pmc,%d] destroying non-deleted PMC", __LINE__)); KASSERT(LIST_EMPTY(&pm->pm_targets), ("[pmc,%d] destroying pmc with targets", __LINE__)); KASSERT(pm->pm_owner == NULL, ("[pmc,%d] destroying pmc attached to an owner", __LINE__)); KASSERT(counter_u64_fetch(pm->pm_runcount) == 0, ("[pmc,%d] pmc has non-zero run count %ld", __LINE__, (unsigned long)counter_u64_fetch(pm->pm_runcount))); counter_u64_free(pm->pm_runcount); free(pm->pm_pcpu_state, M_PMC); free(pm, M_PMC); } static void pmc_wait_for_pmc_idle(struct pmc *pm) { #ifdef HWPMC_DEBUG volatile int maxloop; maxloop = 100 * pmc_cpu_max(); #endif /* * Loop (with a forced context switch) till the PMC's runcount * comes down to zero. */ while (counter_u64_fetch(pm->pm_runcount) > 0) { #ifdef HWPMC_DEBUG maxloop--; KASSERT(maxloop > 0, ("[pmc,%d] (ri%d, rc%ld) waiting too long for " "pmc to be free", __LINE__, PMC_TO_ROWINDEX(pm), (unsigned long)counter_u64_fetch(pm->pm_runcount))); #endif pmc_force_context_switch(); } } /* * This function does the following things: * * - detaches the PMC from hardware * - unlinks all target threads that were attached to it * - removes the PMC from its owner's list * - destroys the PMC private mutex * * Once this function completes, the given pmc pointer can be freed by * calling pmc_destroy_pmc_descriptor(). */ static void pmc_release_pmc_descriptor(struct pmc *pm) { enum pmc_mode mode; struct pmc_hw *phw; u_int adjri, ri, cpu; struct pmc_owner *po; struct pmc_binding pb; struct pmc_process *pp; struct pmc_classdep *pcd; struct pmc_target *ptgt, *tmp; sx_assert(&pmc_sx, SX_XLOCKED); KASSERT(pm, ("[pmc,%d] null pmc", __LINE__)); ri = PMC_TO_ROWINDEX(pm); pcd = pmc_ri_to_classdep(md, ri, &adjri); mode = PMC_TO_MODE(pm); PMCDBG3(PMC,REL,1, "release-pmc pmc=%p ri=%d mode=%d", pm, ri, mode); /* * First, we take the PMC off hardware. */ cpu = 0; if (PMC_IS_SYSTEM_MODE(mode)) { /* * A system mode PMC runs on a specific CPU. Switch * to this CPU and turn hardware off. */ pmc_save_cpu_binding(&pb); cpu = PMC_TO_CPU(pm); pmc_select_cpu(cpu); /* switch off non-stalled CPUs */ pm->pm_pcpu_state[cpu].pps_cpustate = 0; if (pm->pm_state == PMC_STATE_RUNNING && pm->pm_pcpu_state[cpu].pps_stalled == 0) { phw = pmc_pcpu[cpu]->pc_hwpmcs[ri]; KASSERT(phw->phw_pmc == pm, ("[pmc, %d] pmc ptr ri(%d) hw(%p) pm(%p)", __LINE__, ri, phw->phw_pmc, pm)); PMCDBG2(PMC,REL,2, "stopping cpu=%d ri=%d", cpu, ri); critical_enter(); pcd->pcd_stop_pmc(cpu, adjri); critical_exit(); } PMCDBG2(PMC,REL,2, "decfg cpu=%d ri=%d", cpu, ri); critical_enter(); pcd->pcd_config_pmc(cpu, adjri, NULL); critical_exit(); /* adjust the global and process count of SS mode PMCs */ if (mode == PMC_MODE_SS && pm->pm_state == PMC_STATE_RUNNING) { po = pm->pm_owner; po->po_sscount--; if (po->po_sscount == 0) { atomic_subtract_rel_int(&pmc_ss_count, 1); CK_LIST_REMOVE(po, po_ssnext); epoch_wait_preempt(global_epoch_preempt); } } pm->pm_state = PMC_STATE_DELETED; pmc_restore_cpu_binding(&pb); /* * We could have references to this PMC structure in * the per-cpu sample queues. Wait for the queue to * drain. */ pmc_wait_for_pmc_idle(pm); } else if (PMC_IS_VIRTUAL_MODE(mode)) { /* * A virtual PMC could be running on multiple CPUs at * a given instant. * * By marking its state as DELETED, we ensure that * this PMC is never further scheduled on hardware. * * Then we wait till all CPUs are done with this PMC. */ pm->pm_state = PMC_STATE_DELETED; /* Wait for the PMCs runcount to come to zero. */ pmc_wait_for_pmc_idle(pm); /* * At this point the PMC is off all CPUs and cannot be * freshly scheduled onto a CPU. It is now safe to * unlink all targets from this PMC. If a * process-record's refcount falls to zero, we remove * it from the hash table. The module-wide SX lock * protects us from races. */ LIST_FOREACH_SAFE(ptgt, &pm->pm_targets, pt_next, tmp) { pp = ptgt->pt_process; pmc_unlink_target_process(pm, pp); /* frees 'ptgt' */ PMCDBG1(PMC,REL,3, "pp->refcnt=%d", pp->pp_refcnt); /* * If the target process record shows that no * PMCs are attached to it, reclaim its space. */ if (pp->pp_refcnt == 0) { pmc_remove_process_descriptor(pp); pmc_destroy_process_descriptor(pp); } } cpu = curthread->td_oncpu; /* setup cpu for pmd_release() */ } /* * Release any MD resources */ (void) pcd->pcd_release_pmc(cpu, adjri, pm); /* * Update row disposition */ if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm))) PMC_UNMARK_ROW_STANDALONE(ri); else PMC_UNMARK_ROW_THREAD(ri); /* unlink from the owner's list */ if (pm->pm_owner) { LIST_REMOVE(pm, pm_next); pm->pm_owner = NULL; } } /* * Register an owner and a pmc. */ static int pmc_register_owner(struct proc *p, struct pmc *pmc) { struct pmc_owner *po; sx_assert(&pmc_sx, SX_XLOCKED); if ((po = pmc_find_owner_descriptor(p)) == NULL) if ((po = pmc_allocate_owner_descriptor(p)) == NULL) return ENOMEM; KASSERT(pmc->pm_owner == NULL, ("[pmc,%d] attempting to own an initialized PMC", __LINE__)); pmc->pm_owner = po; LIST_INSERT_HEAD(&po->po_pmcs, pmc, pm_next); PROC_LOCK(p); p->p_flag |= P_HWPMC; PROC_UNLOCK(p); if (po->po_flags & PMC_PO_OWNS_LOGFILE) pmclog_process_pmcallocate(pmc); PMCDBG2(PMC,REG,1, "register-owner pmc-owner=%p pmc=%p", po, pmc); return 0; } /* * Return the current row disposition: * == 0 => FREE * > 0 => PROCESS MODE * < 0 => SYSTEM MODE */ int pmc_getrowdisp(int ri) { return pmc_pmcdisp[ri]; } /* * Check if a PMC at row index 'ri' can be allocated to the current * process. * * Allocation can fail if: * - the current process is already being profiled by a PMC at index 'ri', * attached to it via OP_PMCATTACH. * - the current process has already allocated a PMC at index 'ri' * via OP_ALLOCATE. */ static int pmc_can_allocate_rowindex(struct proc *p, unsigned int ri, int cpu) { enum pmc_mode mode; struct pmc *pm; struct pmc_owner *po; struct pmc_process *pp; PMCDBG5(PMC,ALR,1, "can-allocate-rowindex proc=%p (%d, %s) ri=%d " "cpu=%d", p, p->p_pid, p->p_comm, ri, cpu); /* * We shouldn't have already allocated a process-mode PMC at * row index 'ri'. * * We shouldn't have allocated a system-wide PMC on the same * CPU and same RI. */ if ((po = pmc_find_owner_descriptor(p)) != NULL) LIST_FOREACH(pm, &po->po_pmcs, pm_next) { if (PMC_TO_ROWINDEX(pm) == ri) { mode = PMC_TO_MODE(pm); if (PMC_IS_VIRTUAL_MODE(mode)) return EEXIST; if (PMC_IS_SYSTEM_MODE(mode) && (int) PMC_TO_CPU(pm) == cpu) return EEXIST; } } /* * We also shouldn't be the target of any PMC at this index * since otherwise a PMC_ATTACH to ourselves will fail. */ if ((pp = pmc_find_process_descriptor(p, 0)) != NULL) if (pp->pp_pmcs[ri].pp_pmc) return EEXIST; PMCDBG4(PMC,ALR,2, "can-allocate-rowindex proc=%p (%d, %s) ri=%d ok", p, p->p_pid, p->p_comm, ri); return 0; } /* * Check if a given PMC at row index 'ri' can be currently used in * mode 'mode'. */ static int pmc_can_allocate_row(int ri, enum pmc_mode mode) { enum pmc_disp disp; sx_assert(&pmc_sx, SX_XLOCKED); PMCDBG2(PMC,ALR,1, "can-allocate-row ri=%d mode=%d", ri, mode); if (PMC_IS_SYSTEM_MODE(mode)) disp = PMC_DISP_STANDALONE; else disp = PMC_DISP_THREAD; /* * check disposition for PMC row 'ri': * * Expected disposition Row-disposition Result * * STANDALONE STANDALONE or FREE proceed * STANDALONE THREAD fail * THREAD THREAD or FREE proceed * THREAD STANDALONE fail */ if (!PMC_ROW_DISP_IS_FREE(ri) && !(disp == PMC_DISP_THREAD && PMC_ROW_DISP_IS_THREAD(ri)) && !(disp == PMC_DISP_STANDALONE && PMC_ROW_DISP_IS_STANDALONE(ri))) return EBUSY; /* * All OK */ PMCDBG2(PMC,ALR,2, "can-allocate-row ri=%d mode=%d ok", ri, mode); return 0; } /* * Find a PMC descriptor with user handle 'pmcid' for thread 'td'. */ static struct pmc * pmc_find_pmc_descriptor_in_process(struct pmc_owner *po, pmc_id_t pmcid) { struct pmc *pm; KASSERT(PMC_ID_TO_ROWINDEX(pmcid) < md->pmd_npmc, ("[pmc,%d] Illegal pmc index %d (max %d)", __LINE__, PMC_ID_TO_ROWINDEX(pmcid), md->pmd_npmc)); LIST_FOREACH(pm, &po->po_pmcs, pm_next) if (pm->pm_id == pmcid) return pm; return NULL; } static int pmc_find_pmc(pmc_id_t pmcid, struct pmc **pmc) { struct pmc *pm, *opm; struct pmc_owner *po; struct pmc_process *pp; PMCDBG1(PMC,FND,1, "find-pmc id=%d", pmcid); if (PMC_ID_TO_ROWINDEX(pmcid) >= md->pmd_npmc) return (EINVAL); if ((po = pmc_find_owner_descriptor(curthread->td_proc)) == NULL) { /* * In case of PMC_F_DESCENDANTS child processes we will not find * the current process in the owners hash list. Find the owner * process first and from there lookup the po. */ if ((pp = pmc_find_process_descriptor(curthread->td_proc, PMC_FLAG_NONE)) == NULL) { return ESRCH; } else { opm = pp->pp_pmcs[PMC_ID_TO_ROWINDEX(pmcid)].pp_pmc; if (opm == NULL) return ESRCH; if ((opm->pm_flags & (PMC_F_ATTACHED_TO_OWNER| PMC_F_DESCENDANTS)) != (PMC_F_ATTACHED_TO_OWNER| PMC_F_DESCENDANTS)) return ESRCH; po = opm->pm_owner; } } if ((pm = pmc_find_pmc_descriptor_in_process(po, pmcid)) == NULL) return EINVAL; PMCDBG2(PMC,FND,2, "find-pmc id=%d -> pmc=%p", pmcid, pm); *pmc = pm; return 0; } /* * Start a PMC. */ static int pmc_start(struct pmc *pm) { enum pmc_mode mode; struct pmc_owner *po; struct pmc_binding pb; struct pmc_classdep *pcd; int adjri, error, cpu, ri; KASSERT(pm != NULL, ("[pmc,%d] null pm", __LINE__)); mode = PMC_TO_MODE(pm); ri = PMC_TO_ROWINDEX(pm); pcd = pmc_ri_to_classdep(md, ri, &adjri); error = 0; PMCDBG3(PMC,OPS,1, "start pmc=%p mode=%d ri=%d", pm, mode, ri); po = pm->pm_owner; /* * Disallow PMCSTART if a logfile is required but has not been * configured yet. */ if ((pm->pm_flags & PMC_F_NEEDS_LOGFILE) && (po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) return (EDOOFUS); /* programming error */ /* * If this is a sampling mode PMC, log mapping information for * the kernel modules that are currently loaded. */ if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) pmc_log_kernel_mappings(pm); if (PMC_IS_VIRTUAL_MODE(mode)) { /* * If a PMCATTACH has never been done on this PMC, * attach it to its owner process. */ if (LIST_EMPTY(&pm->pm_targets)) error = (pm->pm_flags & PMC_F_ATTACH_DONE) ? ESRCH : pmc_attach_process(po->po_owner, pm); /* * If the PMC is attached to its owner, then force a context * switch to ensure that the MD state gets set correctly. */ if (error == 0) { pm->pm_state = PMC_STATE_RUNNING; if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) pmc_force_context_switch(); } return (error); } /* * A system-wide PMC. * * Add the owner to the global list if this is a system-wide * sampling PMC. */ if (mode == PMC_MODE_SS) { /* * Log mapping information for all existing processes in the * system. Subsequent mappings are logged as they happen; * see pmc_process_mmap(). */ if (po->po_logprocmaps == 0) { pmc_log_all_process_mappings(po); po->po_logprocmaps = 1; } po->po_sscount++; if (po->po_sscount == 1) { atomic_add_rel_int(&pmc_ss_count, 1); CK_LIST_INSERT_HEAD(&pmc_ss_owners, po, po_ssnext); PMCDBG1(PMC,OPS,1, "po=%p in global list", po); } } /* * Move to the CPU associated with this * PMC, and start the hardware. */ pmc_save_cpu_binding(&pb); cpu = PMC_TO_CPU(pm); if (!pmc_cpu_is_active(cpu)) return (ENXIO); pmc_select_cpu(cpu); /* * global PMCs are configured at allocation time * so write out the initial value and start the PMC. */ pm->pm_state = PMC_STATE_RUNNING; critical_enter(); if ((error = pcd->pcd_write_pmc(cpu, adjri, PMC_IS_SAMPLING_MODE(mode) ? pm->pm_sc.pm_reloadcount : pm->pm_sc.pm_initial)) == 0) { /* If a sampling mode PMC, reset stalled state. */ if (PMC_IS_SAMPLING_MODE(mode)) pm->pm_pcpu_state[cpu].pps_stalled = 0; /* Indicate that we desire this to run. Start it. */ pm->pm_pcpu_state[cpu].pps_cpustate = 1; error = pcd->pcd_start_pmc(cpu, adjri); } critical_exit(); pmc_restore_cpu_binding(&pb); return (error); } /* * Stop a PMC. */ static int pmc_stop(struct pmc *pm) { struct pmc_owner *po; struct pmc_binding pb; struct pmc_classdep *pcd; int adjri, cpu, error, ri; KASSERT(pm != NULL, ("[pmc,%d] null pmc", __LINE__)); PMCDBG3(PMC,OPS,1, "stop pmc=%p mode=%d ri=%d", pm, PMC_TO_MODE(pm), PMC_TO_ROWINDEX(pm)); pm->pm_state = PMC_STATE_STOPPED; /* * If the PMC is a virtual mode one, changing the state to * non-RUNNING is enough to ensure that the PMC never gets * scheduled. * * If this PMC is current running on a CPU, then it will * handled correctly at the time its target process is context * switched out. */ if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) return 0; /* * A system-mode PMC. Move to the CPU associated with * this PMC, and stop the hardware. We update the * 'initial count' so that a subsequent PMCSTART will * resume counting from the current hardware count. */ pmc_save_cpu_binding(&pb); cpu = PMC_TO_CPU(pm); KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), ("[pmc,%d] illegal cpu=%d", __LINE__, cpu)); if (!pmc_cpu_is_active(cpu)) return ENXIO; pmc_select_cpu(cpu); ri = PMC_TO_ROWINDEX(pm); pcd = pmc_ri_to_classdep(md, ri, &adjri); pm->pm_pcpu_state[cpu].pps_cpustate = 0; critical_enter(); if ((error = pcd->pcd_stop_pmc(cpu, adjri)) == 0) error = pcd->pcd_read_pmc(cpu, adjri, &pm->pm_sc.pm_initial); critical_exit(); pmc_restore_cpu_binding(&pb); po = pm->pm_owner; /* remove this owner from the global list of SS PMC owners */ if (PMC_TO_MODE(pm) == PMC_MODE_SS) { po->po_sscount--; if (po->po_sscount == 0) { atomic_subtract_rel_int(&pmc_ss_count, 1); CK_LIST_REMOVE(po, po_ssnext); epoch_wait_preempt(global_epoch_preempt); PMCDBG1(PMC,OPS,2,"po=%p removed from global list", po); } } return (error); } static struct pmc_classdep * pmc_class_to_classdep(enum pmc_class class) { int n; for (n = 0; n < md->pmd_nclass; n++) if (md->pmd_classdep[n].pcd_class == class) return (&md->pmd_classdep[n]); return (NULL); } #ifdef HWPMC_DEBUG static const char *pmc_op_to_name[] = { #undef __PMC_OP #define __PMC_OP(N, D) #N , __PMC_OPS() NULL }; #endif /* * The syscall interface */ #define PMC_GET_SX_XLOCK(...) do { \ sx_xlock(&pmc_sx); \ if (pmc_hook == NULL) { \ sx_xunlock(&pmc_sx); \ return __VA_ARGS__; \ } \ } while (0) #define PMC_DOWNGRADE_SX() do { \ sx_downgrade(&pmc_sx); \ is_sx_downgraded = 1; \ } while (0) static int pmc_syscall_handler(struct thread *td, void *syscall_args) { int error, is_sx_downgraded, op; struct pmc_syscall_args *c; void *pmclog_proc_handle; void *arg; c = (struct pmc_syscall_args *)syscall_args; op = c->pmop_code; arg = c->pmop_data; /* PMC isn't set up yet */ if (pmc_hook == NULL) return (EINVAL); if (op == PMC_OP_CONFIGURELOG) { /* * We cannot create the logging process inside * pmclog_configure_log() because there is a LOR * between pmc_sx and process structure locks. * Instead, pre-create the process and ignite the loop * if everything is fine, otherwise direct the process * to exit. */ error = pmclog_proc_create(td, &pmclog_proc_handle); if (error != 0) goto done_syscall; } PMC_GET_SX_XLOCK(ENOSYS); is_sx_downgraded = 0; PMCDBG3(MOD,PMS,1, "syscall op=%d \"%s\" arg=%p", op, pmc_op_to_name[op], arg); error = 0; counter_u64_add(pmc_stats.pm_syscalls, 1); switch (op) { /* * Configure a log file. * * XXX This OP will be reworked. */ case PMC_OP_CONFIGURELOG: { struct proc *p; struct pmc *pm; struct pmc_owner *po; struct pmc_op_configurelog cl; if ((error = copyin(arg, &cl, sizeof(cl))) != 0) { pmclog_proc_ignite(pmclog_proc_handle, NULL); break; } /* mark this process as owning a log file */ p = td->td_proc; if ((po = pmc_find_owner_descriptor(p)) == NULL) if ((po = pmc_allocate_owner_descriptor(p)) == NULL) { pmclog_proc_ignite(pmclog_proc_handle, NULL); error = ENOMEM; break; } /* * If a valid fd was passed in, try to configure that, * otherwise if 'fd' was less than zero and there was * a log file configured, flush its buffers and * de-configure it. */ if (cl.pm_logfd >= 0) { error = pmclog_configure_log(md, po, cl.pm_logfd); pmclog_proc_ignite(pmclog_proc_handle, error == 0 ? po : NULL); } else if (po->po_flags & PMC_PO_OWNS_LOGFILE) { pmclog_proc_ignite(pmclog_proc_handle, NULL); error = pmclog_close(po); if (error == 0) { LIST_FOREACH(pm, &po->po_pmcs, pm_next) if (pm->pm_flags & PMC_F_NEEDS_LOGFILE && pm->pm_state == PMC_STATE_RUNNING) pmc_stop(pm); error = pmclog_deconfigure_log(po); } } else { pmclog_proc_ignite(pmclog_proc_handle, NULL); error = EINVAL; } } break; /* * Flush a log file. */ case PMC_OP_FLUSHLOG: { struct pmc_owner *po; sx_assert(&pmc_sx, SX_XLOCKED); if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) { error = EINVAL; break; } error = pmclog_flush(po); } break; /* * Close a log file. */ case PMC_OP_CLOSELOG: { struct pmc_owner *po; sx_assert(&pmc_sx, SX_XLOCKED); if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) { error = EINVAL; break; } error = pmclog_close(po); } break; /* * Retrieve hardware configuration. */ case PMC_OP_GETCPUINFO: /* CPU information */ { struct pmc_op_getcpuinfo gci; struct pmc_classinfo *pci; struct pmc_classdep *pcd; int cl; gci.pm_cputype = md->pmd_cputype; gci.pm_ncpu = pmc_cpu_max(); gci.pm_npmc = md->pmd_npmc; gci.pm_nclass = md->pmd_nclass; pci = gci.pm_classes; pcd = md->pmd_classdep; for (cl = 0; cl < md->pmd_nclass; cl++, pci++, pcd++) { pci->pm_caps = pcd->pcd_caps; pci->pm_class = pcd->pcd_class; pci->pm_width = pcd->pcd_width; pci->pm_num = pcd->pcd_num; } error = copyout(&gci, arg, sizeof(gci)); } break; /* * Retrieve soft events list. */ case PMC_OP_GETDYNEVENTINFO: { enum pmc_class cl; enum pmc_event ev; struct pmc_op_getdyneventinfo *gei; struct pmc_dyn_event_descr dev; struct pmc_soft *ps; uint32_t nevent; sx_assert(&pmc_sx, SX_LOCKED); gei = (struct pmc_op_getdyneventinfo *) arg; if ((error = copyin(&gei->pm_class, &cl, sizeof(cl))) != 0) break; /* Only SOFT class is dynamic. */ if (cl != PMC_CLASS_SOFT) { error = EINVAL; break; } nevent = 0; for (ev = PMC_EV_SOFT_FIRST; (int)ev <= PMC_EV_SOFT_LAST; ev++) { ps = pmc_soft_ev_acquire(ev); if (ps == NULL) continue; bcopy(&ps->ps_ev, &dev, sizeof(dev)); pmc_soft_ev_release(ps); error = copyout(&dev, &gei->pm_events[nevent], sizeof(struct pmc_dyn_event_descr)); if (error != 0) break; nevent++; } if (error != 0) break; error = copyout(&nevent, &gei->pm_nevent, sizeof(nevent)); } break; /* * Get module statistics */ case PMC_OP_GETDRIVERSTATS: { struct pmc_op_getdriverstats gms; #define CFETCH(a, b, field) a.field = counter_u64_fetch(b.field) CFETCH(gms, pmc_stats, pm_intr_ignored); CFETCH(gms, pmc_stats, pm_intr_processed); CFETCH(gms, pmc_stats, pm_intr_bufferfull); CFETCH(gms, pmc_stats, pm_syscalls); CFETCH(gms, pmc_stats, pm_syscall_errors); CFETCH(gms, pmc_stats, pm_buffer_requests); CFETCH(gms, pmc_stats, pm_buffer_requests_failed); CFETCH(gms, pmc_stats, pm_log_sweeps); #undef CFETCH error = copyout(&gms, arg, sizeof(gms)); } break; /* * Retrieve module version number */ case PMC_OP_GETMODULEVERSION: { uint32_t cv, modv; /* retrieve the client's idea of the ABI version */ if ((error = copyin(arg, &cv, sizeof(uint32_t))) != 0) break; /* don't service clients newer than our driver */ modv = PMC_VERSION; if ((cv & 0xFFFF0000) > (modv & 0xFFFF0000)) { error = EPROGMISMATCH; break; } error = copyout(&modv, arg, sizeof(int)); } break; /* * Retrieve the state of all the PMCs on a given * CPU. */ case PMC_OP_GETPMCINFO: { int ari; struct pmc *pm; size_t pmcinfo_size; uint32_t cpu, n, npmc; struct pmc_owner *po; struct pmc_binding pb; struct pmc_classdep *pcd; struct pmc_info *p, *pmcinfo; struct pmc_op_getpmcinfo *gpi; PMC_DOWNGRADE_SX(); gpi = (struct pmc_op_getpmcinfo *) arg; if ((error = copyin(&gpi->pm_cpu, &cpu, sizeof(cpu))) != 0) break; if (cpu >= pmc_cpu_max()) { error = EINVAL; break; } if (!pmc_cpu_is_active(cpu)) { error = ENXIO; break; } /* switch to CPU 'cpu' */ pmc_save_cpu_binding(&pb); pmc_select_cpu(cpu); npmc = md->pmd_npmc; pmcinfo_size = npmc * sizeof(struct pmc_info); pmcinfo = malloc(pmcinfo_size, M_PMC, M_WAITOK); p = pmcinfo; for (n = 0; n < md->pmd_npmc; n++, p++) { pcd = pmc_ri_to_classdep(md, n, &ari); KASSERT(pcd != NULL, ("[pmc,%d] null pcd ri=%d", __LINE__, n)); if ((error = pcd->pcd_describe(cpu, ari, p, &pm)) != 0) break; if (PMC_ROW_DISP_IS_STANDALONE(n)) p->pm_rowdisp = PMC_DISP_STANDALONE; else if (PMC_ROW_DISP_IS_THREAD(n)) p->pm_rowdisp = PMC_DISP_THREAD; else p->pm_rowdisp = PMC_DISP_FREE; p->pm_ownerpid = -1; if (pm == NULL) /* no PMC associated */ continue; po = pm->pm_owner; KASSERT(po->po_owner != NULL, ("[pmc,%d] pmc_owner had a null proc pointer", __LINE__)); p->pm_ownerpid = po->po_owner->p_pid; p->pm_mode = PMC_TO_MODE(pm); p->pm_event = pm->pm_event; p->pm_flags = pm->pm_flags; if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) p->pm_reloadcount = pm->pm_sc.pm_reloadcount; } pmc_restore_cpu_binding(&pb); /* now copy out the PMC info collected */ if (error == 0) error = copyout(pmcinfo, &gpi->pm_pmcs, pmcinfo_size); free(pmcinfo, M_PMC); } break; /* * Set the administrative state of a PMC. I.e. whether * the PMC is to be used or not. */ case PMC_OP_PMCADMIN: { int cpu, ri; enum pmc_state request; struct pmc_cpu *pc; struct pmc_hw *phw; struct pmc_op_pmcadmin pma; struct pmc_binding pb; sx_assert(&pmc_sx, SX_XLOCKED); KASSERT(td == curthread, ("[pmc,%d] td != curthread", __LINE__)); error = priv_check(td, PRIV_PMC_MANAGE); if (error) break; if ((error = copyin(arg, &pma, sizeof(pma))) != 0) break; cpu = pma.pm_cpu; if (cpu < 0 || cpu >= (int) pmc_cpu_max()) { error = EINVAL; break; } if (!pmc_cpu_is_active(cpu)) { error = ENXIO; break; } request = pma.pm_state; if (request != PMC_STATE_DISABLED && request != PMC_STATE_FREE) { error = EINVAL; break; } ri = pma.pm_pmc; /* pmc id == row index */ if (ri < 0 || ri >= (int) md->pmd_npmc) { error = EINVAL; break; } /* * We can't disable a PMC with a row-index allocated * for process virtual PMCs. */ if (PMC_ROW_DISP_IS_THREAD(ri) && request == PMC_STATE_DISABLED) { error = EBUSY; break; } /* * otherwise, this PMC on this CPU is either free or * in system-wide mode. */ pmc_save_cpu_binding(&pb); pmc_select_cpu(cpu); pc = pmc_pcpu[cpu]; phw = pc->pc_hwpmcs[ri]; /* * XXX do we need some kind of 'forced' disable? */ if (phw->phw_pmc == NULL) { if (request == PMC_STATE_DISABLED && (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED)) { phw->phw_state &= ~PMC_PHW_FLAG_IS_ENABLED; PMC_MARK_ROW_STANDALONE(ri); } else if (request == PMC_STATE_FREE && (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0) { phw->phw_state |= PMC_PHW_FLAG_IS_ENABLED; PMC_UNMARK_ROW_STANDALONE(ri); } /* other cases are a no-op */ } else error = EBUSY; pmc_restore_cpu_binding(&pb); } break; /* * Allocate a PMC. */ case PMC_OP_PMCALLOCATE: { int adjri, n; u_int cpu; uint32_t caps; struct pmc *pmc; enum pmc_mode mode; struct pmc_hw *phw; struct pmc_binding pb; struct pmc_classdep *pcd; struct pmc_op_pmcallocate pa; if ((error = copyin(arg, &pa, sizeof(pa))) != 0) break; caps = pa.pm_caps; mode = pa.pm_mode; cpu = pa.pm_cpu; if ((mode != PMC_MODE_SS && mode != PMC_MODE_SC && mode != PMC_MODE_TS && mode != PMC_MODE_TC) || (cpu != (u_int) PMC_CPU_ANY && cpu >= pmc_cpu_max())) { error = EINVAL; break; } /* * Virtual PMCs should only ask for a default CPU. * System mode PMCs need to specify a non-default CPU. */ if ((PMC_IS_VIRTUAL_MODE(mode) && cpu != (u_int) PMC_CPU_ANY) || (PMC_IS_SYSTEM_MODE(mode) && cpu == (u_int) PMC_CPU_ANY)) { error = EINVAL; break; } /* * Check that an inactive CPU is not being asked for. */ if (PMC_IS_SYSTEM_MODE(mode) && !pmc_cpu_is_active(cpu)) { error = ENXIO; break; } /* * Refuse an allocation for a system-wide PMC if this * process has been jailed, or if this process lacks * super-user credentials and the sysctl tunable * 'security.bsd.unprivileged_syspmcs' is zero. */ if (PMC_IS_SYSTEM_MODE(mode)) { if (jailed(curthread->td_ucred)) { error = EPERM; break; } if (!pmc_unprivileged_syspmcs) { error = priv_check(curthread, PRIV_PMC_SYSTEM); if (error) break; } } /* * Look for valid values for 'pm_flags' */ if ((pa.pm_flags & ~(PMC_F_DESCENDANTS | PMC_F_LOG_PROCCSW | PMC_F_LOG_PROCEXIT | PMC_F_CALLCHAIN)) != 0) { error = EINVAL; break; } /* process logging options are not allowed for system PMCs */ if (PMC_IS_SYSTEM_MODE(mode) && (pa.pm_flags & (PMC_F_LOG_PROCCSW | PMC_F_LOG_PROCEXIT))) { error = EINVAL; break; } /* * All sampling mode PMCs need to be able to interrupt the * CPU. */ if (PMC_IS_SAMPLING_MODE(mode)) caps |= PMC_CAP_INTERRUPT; /* A valid class specifier should have been passed in. */ pcd = pmc_class_to_classdep(pa.pm_class); if (pcd == NULL) { error = EINVAL; break; } /* The requested PMC capabilities should be feasible. */ if ((pcd->pcd_caps & caps) != caps) { error = EOPNOTSUPP; break; } PMCDBG4(PMC,ALL,2, "event=%d caps=0x%x mode=%d cpu=%d", pa.pm_ev, caps, mode, cpu); pmc = pmc_allocate_pmc_descriptor(); pmc->pm_id = PMC_ID_MAKE_ID(cpu,pa.pm_mode,pa.pm_class, PMC_ID_INVALID); pmc->pm_event = pa.pm_ev; pmc->pm_state = PMC_STATE_FREE; pmc->pm_caps = caps; pmc->pm_flags = pa.pm_flags; /* switch thread to CPU 'cpu' */ pmc_save_cpu_binding(&pb); #define PMC_IS_SHAREABLE_PMC(cpu, n) \ (pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_state & \ PMC_PHW_FLAG_IS_SHAREABLE) #define PMC_IS_UNALLOCATED(cpu, n) \ (pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_pmc == NULL) if (PMC_IS_SYSTEM_MODE(mode)) { pmc_select_cpu(cpu); for (n = pcd->pcd_ri; n < (int) md->pmd_npmc; n++) { pcd = pmc_ri_to_classdep(md, n, &adjri); if (pmc_can_allocate_row(n, mode) == 0 && pmc_can_allocate_rowindex( curthread->td_proc, n, cpu) == 0 && (PMC_IS_UNALLOCATED(cpu, n) || PMC_IS_SHAREABLE_PMC(cpu, n)) && pcd->pcd_allocate_pmc(cpu, adjri, pmc, &pa) == 0) break; } } else { /* Process virtual mode */ for (n = pcd->pcd_ri; n < (int) md->pmd_npmc; n++) { pcd = pmc_ri_to_classdep(md, n, &adjri); if (pmc_can_allocate_row(n, mode) == 0 && pmc_can_allocate_rowindex( curthread->td_proc, n, PMC_CPU_ANY) == 0 && pcd->pcd_allocate_pmc(curthread->td_oncpu, adjri, pmc, &pa) == 0) break; } } #undef PMC_IS_UNALLOCATED #undef PMC_IS_SHAREABLE_PMC pmc_restore_cpu_binding(&pb); if (n == (int) md->pmd_npmc) { pmc_destroy_pmc_descriptor(pmc); pmc = NULL; error = EINVAL; break; } /* Fill in the correct value in the ID field */ pmc->pm_id = PMC_ID_MAKE_ID(cpu,mode,pa.pm_class,n); PMCDBG5(PMC,ALL,2, "ev=%d class=%d mode=%d n=%d -> pmcid=%x", pmc->pm_event, pa.pm_class, mode, n, pmc->pm_id); /* Process mode PMCs with logging enabled need log files */ if (pmc->pm_flags & (PMC_F_LOG_PROCEXIT | PMC_F_LOG_PROCCSW)) pmc->pm_flags |= PMC_F_NEEDS_LOGFILE; /* All system mode sampling PMCs require a log file */ if (PMC_IS_SAMPLING_MODE(mode) && PMC_IS_SYSTEM_MODE(mode)) pmc->pm_flags |= PMC_F_NEEDS_LOGFILE; /* * Configure global pmc's immediately */ if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pmc))) { pmc_save_cpu_binding(&pb); pmc_select_cpu(cpu); phw = pmc_pcpu[cpu]->pc_hwpmcs[n]; pcd = pmc_ri_to_classdep(md, n, &adjri); if ((phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0 || (error = pcd->pcd_config_pmc(cpu, adjri, pmc)) != 0) { (void) pcd->pcd_release_pmc(cpu, adjri, pmc); pmc_destroy_pmc_descriptor(pmc); pmc = NULL; pmc_restore_cpu_binding(&pb); error = EPERM; break; } pmc_restore_cpu_binding(&pb); } pmc->pm_state = PMC_STATE_ALLOCATED; pmc->pm_class = pa.pm_class; /* * mark row disposition */ if (PMC_IS_SYSTEM_MODE(mode)) PMC_MARK_ROW_STANDALONE(n); else PMC_MARK_ROW_THREAD(n); /* * Register this PMC with the current thread as its owner. */ if ((error = pmc_register_owner(curthread->td_proc, pmc)) != 0) { pmc_release_pmc_descriptor(pmc); pmc_destroy_pmc_descriptor(pmc); pmc = NULL; break; } /* * Return the allocated index. */ pa.pm_pmcid = pmc->pm_id; error = copyout(&pa, arg, sizeof(pa)); } break; /* * Attach a PMC to a process. */ case PMC_OP_PMCATTACH: { struct pmc *pm; struct proc *p; struct pmc_op_pmcattach a; sx_assert(&pmc_sx, SX_XLOCKED); if ((error = copyin(arg, &a, sizeof(a))) != 0) break; if (a.pm_pid < 0) { error = EINVAL; break; } else if (a.pm_pid == 0) a.pm_pid = td->td_proc->p_pid; if ((error = pmc_find_pmc(a.pm_pmc, &pm)) != 0) break; if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm))) { error = EINVAL; break; } /* PMCs may be (re)attached only when allocated or stopped */ if (pm->pm_state == PMC_STATE_RUNNING) { error = EBUSY; break; } else if (pm->pm_state != PMC_STATE_ALLOCATED && pm->pm_state != PMC_STATE_STOPPED) { error = EINVAL; break; } /* lookup pid */ if ((p = pfind(a.pm_pid)) == NULL) { error = ESRCH; break; } /* * Ignore processes that are working on exiting. */ if (p->p_flag & P_WEXIT) { error = ESRCH; PROC_UNLOCK(p); /* pfind() returns a locked process */ break; } /* * we are allowed to attach a PMC to a process if * we can debug it. */ error = p_candebug(curthread, p); PROC_UNLOCK(p); if (error == 0) error = pmc_attach_process(p, pm); } break; /* * Detach an attached PMC from a process. */ case PMC_OP_PMCDETACH: { struct pmc *pm; struct proc *p; struct pmc_op_pmcattach a; if ((error = copyin(arg, &a, sizeof(a))) != 0) break; if (a.pm_pid < 0) { error = EINVAL; break; } else if (a.pm_pid == 0) a.pm_pid = td->td_proc->p_pid; if ((error = pmc_find_pmc(a.pm_pmc, &pm)) != 0) break; if ((p = pfind(a.pm_pid)) == NULL) { error = ESRCH; break; } /* * Treat processes that are in the process of exiting * as if they were not present. */ if (p->p_flag & P_WEXIT) error = ESRCH; PROC_UNLOCK(p); /* pfind() returns a locked process */ if (error == 0) error = pmc_detach_process(p, pm); } break; /* * Retrieve the MSR number associated with the counter * 'pmc_id'. This allows processes to directly use RDPMC * instructions to read their PMCs, without the overhead of a * system call. */ case PMC_OP_PMCGETMSR: { int adjri, ri; struct pmc *pm; struct pmc_target *pt; struct pmc_op_getmsr gm; struct pmc_classdep *pcd; PMC_DOWNGRADE_SX(); if ((error = copyin(arg, &gm, sizeof(gm))) != 0) break; if ((error = pmc_find_pmc(gm.pm_pmcid, &pm)) != 0) break; /* * The allocated PMC has to be a process virtual PMC, * i.e., of type MODE_T[CS]. Global PMCs can only be * read using the PMCREAD operation since they may be * allocated on a different CPU than the one we could * be running on at the time of the RDPMC instruction. * * The GETMSR operation is not allowed for PMCs that * are inherited across processes. */ if (!PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)) || (pm->pm_flags & PMC_F_DESCENDANTS)) { error = EINVAL; break; } /* * It only makes sense to use a RDPMC (or its * equivalent instruction on non-x86 architectures) on * a process that has allocated and attached a PMC to * itself. Conversely the PMC is only allowed to have * one process attached to it -- its owner. */ if ((pt = LIST_FIRST(&pm->pm_targets)) == NULL || LIST_NEXT(pt, pt_next) != NULL || pt->pt_process->pp_proc != pm->pm_owner->po_owner) { error = EINVAL; break; } ri = PMC_TO_ROWINDEX(pm); pcd = pmc_ri_to_classdep(md, ri, &adjri); /* PMC class has no 'GETMSR' support */ if (pcd->pcd_get_msr == NULL) { error = ENOSYS; break; } if ((error = (*pcd->pcd_get_msr)(adjri, &gm.pm_msr)) < 0) break; if ((error = copyout(&gm, arg, sizeof(gm))) < 0) break; /* * Mark our process as using MSRs. Update machine * state using a forced context switch. */ pt->pt_process->pp_flags |= PMC_PP_ENABLE_MSR_ACCESS; pmc_force_context_switch(); } break; /* * Release an allocated PMC */ case PMC_OP_PMCRELEASE: { pmc_id_t pmcid; struct pmc *pm; struct pmc_owner *po; struct pmc_op_simple sp; /* * Find PMC pointer for the named PMC. * * Use pmc_release_pmc_descriptor() to switch off the * PMC, remove all its target threads, and remove the * PMC from its owner's list. * * Remove the owner record if this is the last PMC * owned. * * Free up space. */ if ((error = copyin(arg, &sp, sizeof(sp))) != 0) break; pmcid = sp.pm_pmcid; if ((error = pmc_find_pmc(pmcid, &pm)) != 0) break; po = pm->pm_owner; pmc_release_pmc_descriptor(pm); pmc_maybe_remove_owner(po); pmc_destroy_pmc_descriptor(pm); } break; /* * Read and/or write a PMC. */ case PMC_OP_PMCRW: { int adjri; struct pmc *pm; uint32_t cpu, ri; pmc_value_t oldvalue; struct pmc_binding pb; struct pmc_op_pmcrw prw; struct pmc_classdep *pcd; struct pmc_op_pmcrw *pprw; PMC_DOWNGRADE_SX(); if ((error = copyin(arg, &prw, sizeof(prw))) != 0) break; ri = 0; PMCDBG2(PMC,OPS,1, "rw id=%d flags=0x%x", prw.pm_pmcid, prw.pm_flags); /* must have at least one flag set */ if ((prw.pm_flags & (PMC_F_OLDVALUE|PMC_F_NEWVALUE)) == 0) { error = EINVAL; break; } /* locate pmc descriptor */ if ((error = pmc_find_pmc(prw.pm_pmcid, &pm)) != 0) break; /* Can't read a PMC that hasn't been started. */ if (pm->pm_state != PMC_STATE_ALLOCATED && pm->pm_state != PMC_STATE_STOPPED && pm->pm_state != PMC_STATE_RUNNING) { error = EINVAL; break; } /* writing a new value is allowed only for 'STOPPED' pmcs */ if (pm->pm_state == PMC_STATE_RUNNING && (prw.pm_flags & PMC_F_NEWVALUE)) { error = EBUSY; break; } if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) { /* * If this PMC is attached to its owner (i.e., * the process requesting this operation) and * is running, then attempt to get an * upto-date reading from hardware for a READ. * Writes are only allowed when the PMC is * stopped, so only update the saved value * field. * * If the PMC is not running, or is not * attached to its owner, read/write to the * savedvalue field. */ ri = PMC_TO_ROWINDEX(pm); pcd = pmc_ri_to_classdep(md, ri, &adjri); mtx_pool_lock_spin(pmc_mtxpool, pm); cpu = curthread->td_oncpu; if (prw.pm_flags & PMC_F_OLDVALUE) { if ((pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) && (pm->pm_state == PMC_STATE_RUNNING)) error = (*pcd->pcd_read_pmc)(cpu, adjri, &oldvalue); else oldvalue = pm->pm_gv.pm_savedvalue; } if (prw.pm_flags & PMC_F_NEWVALUE) pm->pm_gv.pm_savedvalue = prw.pm_value; mtx_pool_unlock_spin(pmc_mtxpool, pm); } else { /* System mode PMCs */ cpu = PMC_TO_CPU(pm); ri = PMC_TO_ROWINDEX(pm); pcd = pmc_ri_to_classdep(md, ri, &adjri); if (!pmc_cpu_is_active(cpu)) { error = ENXIO; break; } /* move this thread to CPU 'cpu' */ pmc_save_cpu_binding(&pb); pmc_select_cpu(cpu); critical_enter(); /* save old value */ if (prw.pm_flags & PMC_F_OLDVALUE) if ((error = (*pcd->pcd_read_pmc)(cpu, adjri, &oldvalue))) goto error; /* write out new value */ if (prw.pm_flags & PMC_F_NEWVALUE) error = (*pcd->pcd_write_pmc)(cpu, adjri, prw.pm_value); error: critical_exit(); pmc_restore_cpu_binding(&pb); if (error) break; } pprw = (struct pmc_op_pmcrw *) arg; #ifdef HWPMC_DEBUG if (prw.pm_flags & PMC_F_NEWVALUE) PMCDBG3(PMC,OPS,2, "rw id=%d new %jx -> old %jx", ri, prw.pm_value, oldvalue); else if (prw.pm_flags & PMC_F_OLDVALUE) PMCDBG2(PMC,OPS,2, "rw id=%d -> old %jx", ri, oldvalue); #endif /* return old value if requested */ if (prw.pm_flags & PMC_F_OLDVALUE) if ((error = copyout(&oldvalue, &pprw->pm_value, sizeof(prw.pm_value)))) break; } break; /* * Set the sampling rate for a sampling mode PMC and the * initial count for a counting mode PMC. */ case PMC_OP_PMCSETCOUNT: { struct pmc *pm; struct pmc_op_pmcsetcount sc; PMC_DOWNGRADE_SX(); if ((error = copyin(arg, &sc, sizeof(sc))) != 0) break; if ((error = pmc_find_pmc(sc.pm_pmcid, &pm)) != 0) break; if (pm->pm_state == PMC_STATE_RUNNING) { error = EBUSY; break; } if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) pm->pm_sc.pm_reloadcount = sc.pm_count; else pm->pm_sc.pm_initial = sc.pm_count; } break; /* * Start a PMC. */ case PMC_OP_PMCSTART: { pmc_id_t pmcid; struct pmc *pm; struct pmc_op_simple sp; sx_assert(&pmc_sx, SX_XLOCKED); if ((error = copyin(arg, &sp, sizeof(sp))) != 0) break; pmcid = sp.pm_pmcid; if ((error = pmc_find_pmc(pmcid, &pm)) != 0) break; KASSERT(pmcid == pm->pm_id, ("[pmc,%d] pmcid %x != id %x", __LINE__, pm->pm_id, pmcid)); if (pm->pm_state == PMC_STATE_RUNNING) /* already running */ break; else if (pm->pm_state != PMC_STATE_STOPPED && pm->pm_state != PMC_STATE_ALLOCATED) { error = EINVAL; break; } error = pmc_start(pm); } break; /* * Stop a PMC. */ case PMC_OP_PMCSTOP: { pmc_id_t pmcid; struct pmc *pm; struct pmc_op_simple sp; PMC_DOWNGRADE_SX(); if ((error = copyin(arg, &sp, sizeof(sp))) != 0) break; pmcid = sp.pm_pmcid; /* * Mark the PMC as inactive and invoke the MD stop * routines if needed. */ if ((error = pmc_find_pmc(pmcid, &pm)) != 0) break; KASSERT(pmcid == pm->pm_id, ("[pmc,%d] pmc id %x != pmcid %x", __LINE__, pm->pm_id, pmcid)); if (pm->pm_state == PMC_STATE_STOPPED) /* already stopped */ break; else if (pm->pm_state != PMC_STATE_RUNNING) { error = EINVAL; break; } error = pmc_stop(pm); } break; /* * Write a user supplied value to the log file. */ case PMC_OP_WRITELOG: { struct pmc_op_writelog wl; struct pmc_owner *po; PMC_DOWNGRADE_SX(); if ((error = copyin(arg, &wl, sizeof(wl))) != 0) break; if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) { error = EINVAL; break; } if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) { error = EINVAL; break; } error = pmclog_process_userlog(po, &wl); } break; default: error = EINVAL; break; } if (is_sx_downgraded) sx_sunlock(&pmc_sx); else sx_xunlock(&pmc_sx); done_syscall: if (error) counter_u64_add(pmc_stats.pm_syscall_errors, 1); return (error); } /* * Helper functions */ /* * Mark the thread as needing callchain capture and post an AST. The * actual callchain capture will be done in a context where it is safe * to take page faults. */ static void pmc_post_callchain_callback(void) { struct thread *td; td = curthread; /* * If there is multiple PMCs for the same interrupt ignore new post */ if (td->td_pflags & TDP_CALLCHAIN) return; /* * Mark this thread as needing callchain capture. * `td->td_pflags' will be safe to touch because this thread * was in user space when it was interrupted. */ td->td_pflags |= TDP_CALLCHAIN; /* * Don't let this thread migrate between CPUs until callchain * capture completes. */ sched_pin(); return; } /* * Interrupt processing. * * Find a free slot in the per-cpu array of samples and capture the * current callchain there. If a sample was successfully added, a bit * is set in mask 'pmc_cpumask' denoting that the DO_SAMPLES hook * needs to be invoked from the clock handler. * * This function is meant to be called from an NMI handler. It cannot * use any of the locking primitives supplied by the OS. */ int pmc_process_interrupt(int cpu, int ring, struct pmc *pm, struct trapframe *tf, int inuserspace) { int error, callchaindepth; struct thread *td; struct pmc_sample *ps; struct pmc_samplebuffer *psb; error = 0; /* * Allocate space for a sample buffer. */ psb = pmc_pcpu[cpu]->pc_sb[ring]; ps = psb->ps_write; if (ps->ps_nsamples) { /* in use, reader hasn't caught up */ pm->pm_pcpu_state[cpu].pps_stalled = 1; counter_u64_add(pmc_stats.pm_intr_bufferfull, 1); PMCDBG6(SAM,INT,1,"(spc) cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d", cpu, pm, (void *) tf, inuserspace, (int) (psb->ps_write - psb->ps_samples), (int) (psb->ps_read - psb->ps_samples)); callchaindepth = 1; error = ENOMEM; goto done; } /* Fill in entry. */ PMCDBG6(SAM,INT,1,"cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d", cpu, pm, (void *) tf, inuserspace, (int) (psb->ps_write - psb->ps_samples), (int) (psb->ps_read - psb->ps_samples)); KASSERT(counter_u64_fetch(pm->pm_runcount) >= 0, ("[pmc,%d] pm=%p runcount %ld", __LINE__, (void *) pm, (unsigned long)counter_u64_fetch(pm->pm_runcount))); counter_u64_add(pm->pm_runcount, 1); /* hold onto PMC */ ps->ps_pmc = pm; ps->ps_pid = -1; ps->ps_tid = -1; if ((td = curthread) != NULL) { ps->ps_tid = td->td_tid; if (td->td_proc) ps->ps_pid = td->td_proc->p_pid; } ps->ps_cpu = cpu; ps->ps_td = td; ps->ps_flags = inuserspace ? PMC_CC_F_USERSPACE : 0; callchaindepth = (pm->pm_flags & PMC_F_CALLCHAIN) ? pmc_callchaindepth : 1; if (callchaindepth == 1) ps->ps_pc[0] = PMC_TRAPFRAME_TO_PC(tf); else { /* * Kernel stack traversals can be done immediately, * while we defer to an AST for user space traversals. */ if (!inuserspace) { callchaindepth = pmc_save_kernel_callchain(ps->ps_pc, callchaindepth, tf); } else { pmc_post_callchain_callback(); callchaindepth = PMC_SAMPLE_INUSE; } } ps->ps_nsamples = callchaindepth; /* mark entry as in use */ /* increment write pointer, modulo ring buffer size */ ps++; if (ps == psb->ps_fence) psb->ps_write = psb->ps_samples; else psb->ps_write = ps; done: /* mark CPU as needing processing */ if (callchaindepth != PMC_SAMPLE_INUSE) DPCPU_SET(pmc_sampled, 1); return (error); } /* * Capture a user call chain. This function will be called from ast() * before control returns to userland and before the process gets * rescheduled. */ static void pmc_capture_user_callchain(int cpu, int ring, struct trapframe *tf) { struct pmc *pm; struct thread *td; struct pmc_sample *ps, *ps_end; struct pmc_samplebuffer *psb; #ifdef INVARIANTS int ncallchains; int nfree; #endif psb = pmc_pcpu[cpu]->pc_sb[ring]; td = curthread; KASSERT(td->td_pflags & TDP_CALLCHAIN, ("[pmc,%d] Retrieving callchain for thread that doesn't want it", __LINE__)); #ifdef INVARIANTS ncallchains = 0; nfree = 0; #endif /* * Iterate through all deferred callchain requests. * Walk from the current read pointer to the current * write pointer. */ ps = psb->ps_read; ps_end = psb->ps_write; do { #ifdef INVARIANTS if ((ps->ps_pmc == NULL) || (ps->ps_pmc->pm_state != PMC_STATE_RUNNING)) nfree++; #endif if (ps->ps_nsamples != PMC_SAMPLE_INUSE) goto next; if (ps->ps_td != td) goto next; KASSERT(ps->ps_cpu == cpu, ("[pmc,%d] cpu mismatch ps_cpu=%d pcpu=%d", __LINE__, ps->ps_cpu, PCPU_GET(cpuid))); pm = ps->ps_pmc; KASSERT(pm->pm_flags & PMC_F_CALLCHAIN, ("[pmc,%d] Retrieving callchain for PMC that doesn't " "want it", __LINE__)); KASSERT(counter_u64_fetch(pm->pm_runcount) > 0, ("[pmc,%d] runcount %ld", __LINE__, (unsigned long)counter_u64_fetch(pm->pm_runcount))); /* * Retrieve the callchain and mark the sample buffer * as 'processable' by the timer tick sweep code. */ ps->ps_nsamples = pmc_save_user_callchain(ps->ps_pc, pmc_callchaindepth, tf); #ifdef INVARIANTS ncallchains++; #endif next: /* increment the pointer, modulo sample ring size */ if (++ps == psb->ps_fence) ps = psb->ps_samples; } while (ps != ps_end); #ifdef INVARIANTS KASSERT(ncallchains > 0 || nfree > 0, ("[pmc,%d] cpu %d didn't find a sample to collect", __LINE__, cpu)); #endif KASSERT(td->td_pinned == 1, ("[pmc,%d] invalid td_pinned value", __LINE__)); sched_unpin(); /* Can migrate safely now. */ /* mark CPU as needing processing */ DPCPU_SET(pmc_sampled, 1); } /* * Process saved PC samples. */ static void pmc_process_samples(int cpu, int ring) { struct pmc *pm; int adjri, n; struct thread *td; struct pmc_owner *po; struct pmc_sample *ps; struct pmc_classdep *pcd; struct pmc_samplebuffer *psb; KASSERT(PCPU_GET(cpuid) == cpu, ("[pmc,%d] not on the correct CPU pcpu=%d cpu=%d", __LINE__, PCPU_GET(cpuid), cpu)); psb = pmc_pcpu[cpu]->pc_sb[ring]; for (n = 0; n < pmc_nsamples; n++) { /* bound on #iterations */ ps = psb->ps_read; if (ps->ps_nsamples == PMC_SAMPLE_FREE) break; pm = ps->ps_pmc; KASSERT(counter_u64_fetch(pm->pm_runcount) > 0, ("[pmc,%d] pm=%p runcount %ld", __LINE__, (void *) pm, (unsigned long)counter_u64_fetch(pm->pm_runcount))); po = pm->pm_owner; KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)), ("[pmc,%d] pmc=%p non-sampling mode=%d", __LINE__, pm, PMC_TO_MODE(pm))); /* Ignore PMCs that have been switched off */ if (pm->pm_state != PMC_STATE_RUNNING) goto entrydone; /* If there is a pending AST wait for completion */ if (ps->ps_nsamples == PMC_SAMPLE_INUSE) { /* Need a rescan at a later time. */ DPCPU_SET(pmc_sampled, 1); break; } PMCDBG6(SAM,OPS,1,"cpu=%d pm=%p n=%d fl=%x wr=%d rd=%d", cpu, pm, ps->ps_nsamples, ps->ps_flags, (int) (psb->ps_write - psb->ps_samples), (int) (psb->ps_read - psb->ps_samples)); /* * If this is a process-mode PMC that is attached to * its owner, and if the PC is in user mode, update * profiling statistics like timer-based profiling * would have done. */ if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) { if (ps->ps_flags & PMC_CC_F_USERSPACE) { td = FIRST_THREAD_IN_PROC(po->po_owner); addupc_intr(td, ps->ps_pc[0], 1); } goto entrydone; } /* * Otherwise, this is either a sampling mode PMC that * is attached to a different process than its owner, * or a system-wide sampling PMC. Dispatch a log * entry to the PMC's owner process. */ pmclog_process_callchain(pm, ps); entrydone: ps->ps_nsamples = 0; /* mark entry as free */ counter_u64_add(pm->pm_runcount, -1); /* increment read pointer, modulo sample size */ if (++ps == psb->ps_fence) psb->ps_read = psb->ps_samples; else psb->ps_read = ps; } counter_u64_add(pmc_stats.pm_log_sweeps, 1); /* Do not re-enable stalled PMCs if we failed to process any samples */ if (n == 0) return; /* * Restart any stalled sampling PMCs on this CPU. * * If the NMI handler sets the pm_stalled field of a PMC after * the check below, we'll end up processing the stalled PMC at * the next hardclock tick. */ for (n = 0; n < md->pmd_npmc; n++) { pcd = pmc_ri_to_classdep(md, n, &adjri); KASSERT(pcd != NULL, ("[pmc,%d] null pcd ri=%d", __LINE__, n)); (void) (*pcd->pcd_get_config)(cpu,adjri,&pm); if (pm == NULL || /* !cfg'ed */ pm->pm_state != PMC_STATE_RUNNING || /* !active */ !PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) || /* !sampling */ !pm->pm_pcpu_state[cpu].pps_cpustate || /* !desired */ !pm->pm_pcpu_state[cpu].pps_stalled) /* !stalled */ continue; pm->pm_pcpu_state[cpu].pps_stalled = 0; (*pcd->pcd_start_pmc)(cpu, adjri); } } /* * Event handlers. */ /* * Handle a process exit. * * Remove this process from all hash tables. If this process * owned any PMCs, turn off those PMCs and deallocate them, * removing any associations with target processes. * * This function will be called by the last 'thread' of a * process. * * XXX This eventhandler gets called early in the exit process. * Consider using a 'hook' invocation from thread_exit() or equivalent * spot. Another negative is that kse_exit doesn't seem to call * exit1() [??]. * */ static void pmc_process_exit(void *arg __unused, struct proc *p) { struct pmc *pm; int adjri, cpu; unsigned int ri; int is_using_hwpmcs; struct pmc_owner *po; struct pmc_process *pp; struct pmc_classdep *pcd; pmc_value_t newvalue, tmp; PROC_LOCK(p); is_using_hwpmcs = p->p_flag & P_HWPMC; PROC_UNLOCK(p); /* * Log a sysexit event to all SS PMC owners. */ epoch_enter_preempt(global_epoch_preempt); CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) if (po->po_flags & PMC_PO_OWNS_LOGFILE) pmclog_process_sysexit(po, p->p_pid); epoch_exit_preempt(global_epoch_preempt); if (!is_using_hwpmcs) return; PMC_GET_SX_XLOCK(); PMCDBG3(PRC,EXT,1,"process-exit proc=%p (%d, %s)", p, p->p_pid, p->p_comm); /* * Since this code is invoked by the last thread in an exiting * process, we would have context switched IN at some prior * point. However, with PREEMPTION, kernel mode context * switches may happen any time, so we want to disable a * context switch OUT till we get any PMCs targeting this * process off the hardware. * * We also need to atomically remove this process' * entry from our target process hash table, using * PMC_FLAG_REMOVE. */ PMCDBG3(PRC,EXT,1, "process-exit proc=%p (%d, %s)", p, p->p_pid, p->p_comm); critical_enter(); /* no preemption */ cpu = curthread->td_oncpu; if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_REMOVE)) != NULL) { PMCDBG2(PRC,EXT,2, "process-exit proc=%p pmc-process=%p", p, pp); /* * The exiting process could the target of * some PMCs which will be running on * currently executing CPU. * * We need to turn these PMCs off like we * would do at context switch OUT time. */ for (ri = 0; ri < md->pmd_npmc; ri++) { /* * Pick up the pmc pointer from hardware * state similar to the CSW_OUT code. */ pm = NULL; pcd = pmc_ri_to_classdep(md, ri, &adjri); (void) (*pcd->pcd_get_config)(cpu, adjri, &pm); PMCDBG2(PRC,EXT,2, "ri=%d pm=%p", ri, pm); if (pm == NULL || !PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) continue; PMCDBG4(PRC,EXT,2, "ppmcs[%d]=%p pm=%p " "state=%d", ri, pp->pp_pmcs[ri].pp_pmc, pm, pm->pm_state); KASSERT(PMC_TO_ROWINDEX(pm) == ri, ("[pmc,%d] ri mismatch pmc(%d) ri(%d)", __LINE__, PMC_TO_ROWINDEX(pm), ri)); KASSERT(pm == pp->pp_pmcs[ri].pp_pmc, ("[pmc,%d] pm %p != pp_pmcs[%d] %p", __LINE__, pm, ri, pp->pp_pmcs[ri].pp_pmc)); KASSERT(counter_u64_fetch(pm->pm_runcount) > 0, ("[pmc,%d] bad runcount ri %d rc %ld", __LINE__, ri, (unsigned long)counter_u64_fetch(pm->pm_runcount))); /* * Change desired state, and then stop if not * stalled. This two-step dance should avoid * race conditions where an interrupt re-enables * the PMC after this code has already checked * the pm_stalled flag. */ if (pm->pm_pcpu_state[cpu].pps_cpustate) { pm->pm_pcpu_state[cpu].pps_cpustate = 0; if (!pm->pm_pcpu_state[cpu].pps_stalled) { (void) pcd->pcd_stop_pmc(cpu, adjri); if (PMC_TO_MODE(pm) == PMC_MODE_TC) { pcd->pcd_read_pmc(cpu, adjri, &newvalue); tmp = newvalue - PMC_PCPU_SAVED(cpu,ri); mtx_pool_lock_spin(pmc_mtxpool, pm); pm->pm_gv.pm_savedvalue += tmp; pp->pp_pmcs[ri].pp_pmcval += tmp; mtx_pool_unlock_spin( pmc_mtxpool, pm); } } } counter_u64_add(pm->pm_runcount, -1); KASSERT((int) counter_u64_fetch(pm->pm_runcount) >= 0, ("[pmc,%d] runcount is %d", __LINE__, ri)); (void) pcd->pcd_config_pmc(cpu, adjri, NULL); } /* * Inform the MD layer of this pseudo "context switch * out" */ (void) md->pmd_switch_out(pmc_pcpu[cpu], pp); critical_exit(); /* ok to be pre-empted now */ /* * Unlink this process from the PMCs that are * targeting it. This will send a signal to * all PMC owner's whose PMCs are orphaned. * * Log PMC value at exit time if requested. */ for (ri = 0; ri < md->pmd_npmc; ri++) if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) { if (pm->pm_flags & PMC_F_NEEDS_LOGFILE && PMC_IS_COUNTING_MODE(PMC_TO_MODE(pm))) pmclog_process_procexit(pm, pp); pmc_unlink_target_process(pm, pp); } free(pp, M_PMC); } else critical_exit(); /* pp == NULL */ /* * If the process owned PMCs, free them up and free up * memory. */ if ((po = pmc_find_owner_descriptor(p)) != NULL) { pmc_remove_owner(po); pmc_destroy_owner_descriptor(po); } sx_xunlock(&pmc_sx); } /* * Handle a process fork. * * If the parent process 'p1' is under HWPMC monitoring, then copy * over any attached PMCs that have 'do_descendants' semantics. */ static void pmc_process_fork(void *arg __unused, struct proc *p1, struct proc *newproc, int flags) { int is_using_hwpmcs; unsigned int ri; uint32_t do_descendants; struct pmc *pm; struct pmc_owner *po; struct pmc_process *ppnew, *ppold; (void) flags; /* unused parameter */ PROC_LOCK(p1); is_using_hwpmcs = p1->p_flag & P_HWPMC; PROC_UNLOCK(p1); /* * If there are system-wide sampling PMCs active, we need to * log all fork events to their owner's logs. */ epoch_enter_preempt(global_epoch_preempt); CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) if (po->po_flags & PMC_PO_OWNS_LOGFILE) pmclog_process_procfork(po, p1->p_pid, newproc->p_pid); epoch_exit_preempt(global_epoch_preempt); if (!is_using_hwpmcs) return; PMC_GET_SX_XLOCK(); PMCDBG4(PMC,FRK,1, "process-fork proc=%p (%d, %s) -> %p", p1, p1->p_pid, p1->p_comm, newproc); /* * If the parent process (curthread->td_proc) is a * target of any PMCs, look for PMCs that are to be * inherited, and link these into the new process * descriptor. */ if ((ppold = pmc_find_process_descriptor(curthread->td_proc, PMC_FLAG_NONE)) == NULL) goto done; /* nothing to do */ do_descendants = 0; for (ri = 0; ri < md->pmd_npmc; ri++) if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL) do_descendants |= pm->pm_flags & PMC_F_DESCENDANTS; if (do_descendants == 0) /* nothing to do */ goto done; /* * Now mark the new process as being tracked by this driver. */ PROC_LOCK(newproc); newproc->p_flag |= P_HWPMC; PROC_UNLOCK(newproc); /* allocate a descriptor for the new process */ if ((ppnew = pmc_find_process_descriptor(newproc, PMC_FLAG_ALLOCATE)) == NULL) goto done; /* * Run through all PMCs that were targeting the old process * and which specified F_DESCENDANTS and attach them to the * new process. * * Log the fork event to all owners of PMCs attached to this * process, if not already logged. */ for (ri = 0; ri < md->pmd_npmc; ri++) if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL && (pm->pm_flags & PMC_F_DESCENDANTS)) { pmc_link_target_process(pm, ppnew); po = pm->pm_owner; if (po->po_sscount == 0 && po->po_flags & PMC_PO_OWNS_LOGFILE) pmclog_process_procfork(po, p1->p_pid, newproc->p_pid); } done: sx_xunlock(&pmc_sx); } static void pmc_kld_load(void *arg __unused, linker_file_t lf) { struct pmc_owner *po; /* * Notify owners of system sampling PMCs about KLD operations. */ epoch_enter_preempt(global_epoch_preempt); CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) if (po->po_flags & PMC_PO_OWNS_LOGFILE) pmclog_process_map_in(po, (pid_t) -1, (uintfptr_t) lf->address, lf->filename); epoch_exit_preempt(global_epoch_preempt); /* * TODO: Notify owners of (all) process-sampling PMCs too. */ } static void pmc_kld_unload(void *arg __unused, const char *filename __unused, caddr_t address, size_t size) { struct pmc_owner *po; epoch_enter_preempt(global_epoch_preempt); CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) if (po->po_flags & PMC_PO_OWNS_LOGFILE) pmclog_process_map_out(po, (pid_t) -1, (uintfptr_t) address, (uintfptr_t) address + size); epoch_exit_preempt(global_epoch_preempt); /* * TODO: Notify owners of process-sampling PMCs. */ } /* * initialization */ static const char * pmc_name_of_pmcclass(enum pmc_class class) { switch (class) { #undef __PMC_CLASS #define __PMC_CLASS(S,V,D) \ case PMC_CLASS_##S: \ return #S; __PMC_CLASSES(); default: return (""); } } /* * Base class initializer: allocate structure and set default classes. */ struct pmc_mdep * pmc_mdep_alloc(int nclasses) { struct pmc_mdep *md; int n; /* SOFT + md classes */ n = 1 + nclasses; md = malloc(sizeof(struct pmc_mdep) + n * sizeof(struct pmc_classdep), M_PMC, M_WAITOK|M_ZERO); md->pmd_nclass = n; /* Add base class. */ pmc_soft_initialize(md); return md; } void pmc_mdep_free(struct pmc_mdep *md) { pmc_soft_finalize(md); free(md, M_PMC); } static int generic_switch_in(struct pmc_cpu *pc, struct pmc_process *pp) { (void) pc; (void) pp; return (0); } static int generic_switch_out(struct pmc_cpu *pc, struct pmc_process *pp) { (void) pc; (void) pp; return (0); } static struct pmc_mdep * pmc_generic_cpu_initialize(void) { struct pmc_mdep *md; md = pmc_mdep_alloc(0); md->pmd_cputype = PMC_CPU_GENERIC; md->pmd_pcpu_init = NULL; md->pmd_pcpu_fini = NULL; md->pmd_switch_in = generic_switch_in; md->pmd_switch_out = generic_switch_out; return (md); } static void pmc_generic_cpu_finalize(struct pmc_mdep *md) { (void) md; } static int pmc_initialize(void) { int c, cpu, error, n, ri; unsigned int maxcpu, domain; struct pcpu *pc; struct pmc_binding pb; struct pmc_sample *ps; struct pmc_classdep *pcd; struct pmc_samplebuffer *sb; md = NULL; error = 0; pmc_stats.pm_intr_ignored = counter_u64_alloc(M_WAITOK); pmc_stats.pm_intr_processed = counter_u64_alloc(M_WAITOK); pmc_stats.pm_intr_bufferfull = counter_u64_alloc(M_WAITOK); pmc_stats.pm_syscalls = counter_u64_alloc(M_WAITOK); pmc_stats.pm_syscall_errors = counter_u64_alloc(M_WAITOK); pmc_stats.pm_buffer_requests = counter_u64_alloc(M_WAITOK); pmc_stats.pm_buffer_requests_failed = counter_u64_alloc(M_WAITOK); pmc_stats.pm_log_sweeps = counter_u64_alloc(M_WAITOK); #ifdef HWPMC_DEBUG /* parse debug flags first */ if (TUNABLE_STR_FETCH(PMC_SYSCTL_NAME_PREFIX "debugflags", pmc_debugstr, sizeof(pmc_debugstr))) pmc_debugflags_parse(pmc_debugstr, pmc_debugstr+strlen(pmc_debugstr)); #endif PMCDBG1(MOD,INI,0, "PMC Initialize (version %x)", PMC_VERSION); /* check kernel version */ if (pmc_kernel_version != PMC_VERSION) { if (pmc_kernel_version == 0) printf("hwpmc: this kernel has not been compiled with " "'options HWPMC_HOOKS'.\n"); else printf("hwpmc: kernel version (0x%x) does not match " "module version (0x%x).\n", pmc_kernel_version, PMC_VERSION); return EPROGMISMATCH; } /* * check sysctl parameters */ if (pmc_hashsize <= 0) { (void) printf("hwpmc: tunable \"hashsize\"=%d must be " "greater than zero.\n", pmc_hashsize); pmc_hashsize = PMC_HASH_SIZE; } if (pmc_nsamples <= 0 || pmc_nsamples > 65535) { (void) printf("hwpmc: tunable \"nsamples\"=%d out of " "range.\n", pmc_nsamples); pmc_nsamples = PMC_NSAMPLES; } if (pmc_callchaindepth <= 0 || pmc_callchaindepth > PMC_CALLCHAIN_DEPTH_MAX) { (void) printf("hwpmc: tunable \"callchaindepth\"=%d out of " "range - using %d.\n", pmc_callchaindepth, PMC_CALLCHAIN_DEPTH_MAX); pmc_callchaindepth = PMC_CALLCHAIN_DEPTH_MAX; } md = pmc_md_initialize(); if (md == NULL) { /* Default to generic CPU. */ md = pmc_generic_cpu_initialize(); if (md == NULL) return (ENOSYS); } KASSERT(md->pmd_nclass >= 1 && md->pmd_npmc >= 1, ("[pmc,%d] no classes or pmcs", __LINE__)); /* Compute the map from row-indices to classdep pointers. */ pmc_rowindex_to_classdep = malloc(sizeof(struct pmc_classdep *) * md->pmd_npmc, M_PMC, M_WAITOK|M_ZERO); for (n = 0; n < md->pmd_npmc; n++) pmc_rowindex_to_classdep[n] = NULL; for (ri = c = 0; c < md->pmd_nclass; c++) { pcd = &md->pmd_classdep[c]; for (n = 0; n < pcd->pcd_num; n++, ri++) pmc_rowindex_to_classdep[ri] = pcd; } KASSERT(ri == md->pmd_npmc, ("[pmc,%d] npmc miscomputed: ri=%d, md->npmc=%d", __LINE__, ri, md->pmd_npmc)); maxcpu = pmc_cpu_max(); /* allocate space for the per-cpu array */ pmc_pcpu = malloc(maxcpu * sizeof(struct pmc_cpu *), M_PMC, M_WAITOK|M_ZERO); /* per-cpu 'saved values' for managing process-mode PMCs */ pmc_pcpu_saved = malloc(sizeof(pmc_value_t) * maxcpu * md->pmd_npmc, M_PMC, M_WAITOK); /* Perform CPU-dependent initialization. */ pmc_save_cpu_binding(&pb); error = 0; for (cpu = 0; error == 0 && cpu < maxcpu; cpu++) { if (!pmc_cpu_is_active(cpu)) continue; pmc_select_cpu(cpu); pmc_pcpu[cpu] = malloc(sizeof(struct pmc_cpu) + md->pmd_npmc * sizeof(struct pmc_hw *), M_PMC, M_WAITOK|M_ZERO); if (md->pmd_pcpu_init) error = md->pmd_pcpu_init(md, cpu); for (n = 0; error == 0 && n < md->pmd_nclass; n++) error = md->pmd_classdep[n].pcd_pcpu_init(md, cpu); } pmc_restore_cpu_binding(&pb); if (error) return (error); /* allocate space for the sample array */ for (cpu = 0; cpu < maxcpu; cpu++) { if (!pmc_cpu_is_active(cpu)) continue; pc = pcpu_find(cpu); domain = pc->pc_domain; sb = malloc_domain(sizeof(struct pmc_samplebuffer) + pmc_nsamples * sizeof(struct pmc_sample), M_PMC, domain, M_WAITOK|M_ZERO); sb->ps_read = sb->ps_write = sb->ps_samples; sb->ps_fence = sb->ps_samples + pmc_nsamples; KASSERT(pmc_pcpu[cpu] != NULL, ("[pmc,%d] cpu=%d Null per-cpu data", __LINE__, cpu)); sb->ps_callchains = malloc_domain(pmc_callchaindepth * pmc_nsamples * sizeof(uintptr_t), M_PMC, domain, M_WAITOK|M_ZERO); for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++) ps->ps_pc = sb->ps_callchains + (n * pmc_callchaindepth); pmc_pcpu[cpu]->pc_sb[PMC_HR] = sb; sb = malloc_domain(sizeof(struct pmc_samplebuffer) + pmc_nsamples * sizeof(struct pmc_sample), M_PMC, domain, M_WAITOK|M_ZERO); sb->ps_read = sb->ps_write = sb->ps_samples; sb->ps_fence = sb->ps_samples + pmc_nsamples; KASSERT(pmc_pcpu[cpu] != NULL, ("[pmc,%d] cpu=%d Null per-cpu data", __LINE__, cpu)); sb->ps_callchains = malloc_domain(pmc_callchaindepth * pmc_nsamples * sizeof(uintptr_t), M_PMC, domain, M_WAITOK|M_ZERO); for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++) ps->ps_pc = sb->ps_callchains + (n * pmc_callchaindepth); pmc_pcpu[cpu]->pc_sb[PMC_SR] = sb; } /* allocate space for the row disposition array */ pmc_pmcdisp = malloc(sizeof(enum pmc_mode) * md->pmd_npmc, M_PMC, M_WAITOK|M_ZERO); /* mark all PMCs as available */ for (n = 0; n < (int) md->pmd_npmc; n++) PMC_MARK_ROW_FREE(n); /* allocate thread hash tables */ pmc_ownerhash = hashinit(pmc_hashsize, M_PMC, &pmc_ownerhashmask); pmc_processhash = hashinit(pmc_hashsize, M_PMC, &pmc_processhashmask); mtx_init(&pmc_processhash_mtx, "pmc-process-hash", "pmc-leaf", MTX_SPIN); CK_LIST_INIT(&pmc_ss_owners); pmc_ss_count = 0; /* allocate a pool of spin mutexes */ pmc_mtxpool = mtx_pool_create("pmc-leaf", pmc_mtxpool_size, MTX_SPIN); PMCDBG4(MOD,INI,1, "pmc_ownerhash=%p, mask=0x%lx " "targethash=%p mask=0x%lx", pmc_ownerhash, pmc_ownerhashmask, pmc_processhash, pmc_processhashmask); /* Initialize a spin mutex for the thread free list. */ mtx_init(&pmc_threadfreelist_mtx, "pmc-threadfreelist", "pmc-leaf", MTX_SPIN); /* * Initialize the callout to monitor the thread free list. * This callout will also handle the initial population of the list. */ taskqgroup_config_gtask_init(NULL, &free_gtask, pmc_thread_descriptor_pool_free_task, "thread descriptor pool free task"); /* register process {exit,fork,exec} handlers */ pmc_exit_tag = EVENTHANDLER_REGISTER(process_exit, pmc_process_exit, NULL, EVENTHANDLER_PRI_ANY); pmc_fork_tag = EVENTHANDLER_REGISTER(process_fork, pmc_process_fork, NULL, EVENTHANDLER_PRI_ANY); /* register kld event handlers */ pmc_kld_load_tag = EVENTHANDLER_REGISTER(kld_load, pmc_kld_load, NULL, EVENTHANDLER_PRI_ANY); pmc_kld_unload_tag = EVENTHANDLER_REGISTER(kld_unload, pmc_kld_unload, NULL, EVENTHANDLER_PRI_ANY); /* initialize logging */ pmclog_initialize(); /* set hook functions */ pmc_intr = md->pmd_intr; wmb(); pmc_hook = pmc_hook_handler; if (error == 0) { printf(PMC_MODULE_NAME ":"); for (n = 0; n < (int) md->pmd_nclass; n++) { pcd = &md->pmd_classdep[n]; printf(" %s/%d/%d/0x%b", pmc_name_of_pmcclass(pcd->pcd_class), pcd->pcd_num, pcd->pcd_width, pcd->pcd_caps, "\20" "\1INT\2USR\3SYS\4EDG\5THR" "\6REA\7WRI\10INV\11QUA\12PRC" "\13TAG\14CSC"); } printf("\n"); } return (error); } /* prepare to be unloaded */ static void pmc_cleanup(void) { int c, cpu; unsigned int maxcpu; struct pmc_ownerhash *ph; struct pmc_owner *po, *tmp; struct pmc_binding pb; #ifdef HWPMC_DEBUG struct pmc_processhash *prh; #endif PMCDBG0(MOD,INI,0, "cleanup"); /* switch off sampling */ CPU_FOREACH(cpu) DPCPU_ID_SET(cpu, pmc_sampled, 0); pmc_intr = NULL; sx_xlock(&pmc_sx); if (pmc_hook == NULL) { /* being unloaded already */ sx_xunlock(&pmc_sx); return; } pmc_hook = NULL; /* prevent new threads from entering module */ /* deregister event handlers */ EVENTHANDLER_DEREGISTER(process_fork, pmc_fork_tag); EVENTHANDLER_DEREGISTER(process_exit, pmc_exit_tag); EVENTHANDLER_DEREGISTER(kld_load, pmc_kld_load_tag); EVENTHANDLER_DEREGISTER(kld_unload, pmc_kld_unload_tag); /* send SIGBUS to all owner threads, free up allocations */ if (pmc_ownerhash) for (ph = pmc_ownerhash; ph <= &pmc_ownerhash[pmc_ownerhashmask]; ph++) { LIST_FOREACH_SAFE(po, ph, po_next, tmp) { pmc_remove_owner(po); /* send SIGBUS to owner processes */ PMCDBG3(MOD,INI,2, "cleanup signal proc=%p " "(%d, %s)", po->po_owner, po->po_owner->p_pid, po->po_owner->p_comm); PROC_LOCK(po->po_owner); kern_psignal(po->po_owner, SIGBUS); PROC_UNLOCK(po->po_owner); pmc_destroy_owner_descriptor(po); } } /* reclaim allocated data structures */ mtx_destroy(&pmc_threadfreelist_mtx); pmc_thread_descriptor_pool_drain(); if (pmc_mtxpool) mtx_pool_destroy(&pmc_mtxpool); mtx_destroy(&pmc_processhash_mtx); taskqgroup_config_gtask_deinit(&free_gtask); if (pmc_processhash) { #ifdef HWPMC_DEBUG struct pmc_process *pp; PMCDBG0(MOD,INI,3, "destroy process hash"); for (prh = pmc_processhash; prh <= &pmc_processhash[pmc_processhashmask]; prh++) LIST_FOREACH(pp, prh, pp_next) PMCDBG1(MOD,INI,3, "pid=%d", pp->pp_proc->p_pid); #endif hashdestroy(pmc_processhash, M_PMC, pmc_processhashmask); pmc_processhash = NULL; } if (pmc_ownerhash) { PMCDBG0(MOD,INI,3, "destroy owner hash"); hashdestroy(pmc_ownerhash, M_PMC, pmc_ownerhashmask); pmc_ownerhash = NULL; } KASSERT(CK_LIST_EMPTY(&pmc_ss_owners), ("[pmc,%d] Global SS owner list not empty", __LINE__)); KASSERT(pmc_ss_count == 0, ("[pmc,%d] Global SS count not empty", __LINE__)); /* do processor and pmc-class dependent cleanup */ maxcpu = pmc_cpu_max(); PMCDBG0(MOD,INI,3, "md cleanup"); if (md) { pmc_save_cpu_binding(&pb); for (cpu = 0; cpu < maxcpu; cpu++) { PMCDBG2(MOD,INI,1,"pmc-cleanup cpu=%d pcs=%p", cpu, pmc_pcpu[cpu]); if (!pmc_cpu_is_active(cpu) || pmc_pcpu[cpu] == NULL) continue; pmc_select_cpu(cpu); for (c = 0; c < md->pmd_nclass; c++) md->pmd_classdep[c].pcd_pcpu_fini(md, cpu); if (md->pmd_pcpu_fini) md->pmd_pcpu_fini(md, cpu); } if (md->pmd_cputype == PMC_CPU_GENERIC) pmc_generic_cpu_finalize(md); else pmc_md_finalize(md); pmc_mdep_free(md); md = NULL; pmc_restore_cpu_binding(&pb); } /* Free per-cpu descriptors. */ for (cpu = 0; cpu < maxcpu; cpu++) { if (!pmc_cpu_is_active(cpu)) continue; KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_HR] != NULL, ("[pmc,%d] Null hw cpu sample buffer cpu=%d", __LINE__, cpu)); KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_SR] != NULL, ("[pmc,%d] Null sw cpu sample buffer cpu=%d", __LINE__, cpu)); free_domain(pmc_pcpu[cpu]->pc_sb[PMC_HR]->ps_callchains, M_PMC); free_domain(pmc_pcpu[cpu]->pc_sb[PMC_HR], M_PMC); free_domain(pmc_pcpu[cpu]->pc_sb[PMC_SR]->ps_callchains, M_PMC); free_domain(pmc_pcpu[cpu]->pc_sb[PMC_SR], M_PMC); free_domain(pmc_pcpu[cpu], M_PMC); } free(pmc_pcpu, M_PMC); pmc_pcpu = NULL; free(pmc_pcpu_saved, M_PMC); pmc_pcpu_saved = NULL; if (pmc_pmcdisp) { free(pmc_pmcdisp, M_PMC); pmc_pmcdisp = NULL; } if (pmc_rowindex_to_classdep) { free(pmc_rowindex_to_classdep, M_PMC); pmc_rowindex_to_classdep = NULL; } pmclog_shutdown(); counter_u64_free(pmc_stats.pm_intr_ignored); counter_u64_free(pmc_stats.pm_intr_processed); counter_u64_free(pmc_stats.pm_intr_bufferfull); counter_u64_free(pmc_stats.pm_syscalls); counter_u64_free(pmc_stats.pm_syscall_errors); counter_u64_free(pmc_stats.pm_buffer_requests); counter_u64_free(pmc_stats.pm_buffer_requests_failed); counter_u64_free(pmc_stats.pm_log_sweeps); sx_xunlock(&pmc_sx); /* we are done */ } /* * The function called at load/unload. */ static int load (struct module *module __unused, int cmd, void *arg __unused) { int error; error = 0; switch (cmd) { case MOD_LOAD : /* initialize the subsystem */ error = pmc_initialize(); if (error != 0) break; PMCDBG2(MOD,INI,1, "syscall=%d maxcpu=%d", pmc_syscall_num, pmc_cpu_max()); break; case MOD_UNLOAD : case MOD_SHUTDOWN: pmc_cleanup(); PMCDBG0(MOD,INI,1, "unloaded"); break; default : error = EINVAL; /* XXX should panic(9) */ break; } return error; } Index: head/sys/kern/vfs_vnops.c =================================================================== --- head/sys/kern/vfs_vnops.c (revision 334343) +++ head/sys/kern/vfs_vnops.c (revision 334344) @@ -1,2513 +1,2513 @@ /*- * SPDX-License-Identifier: BSD-3-Clause * * Copyright (c) 1982, 1986, 1989, 1993 * The Regents of the University of California. All rights reserved. * (c) UNIX System Laboratories, Inc. * All or some portions of this file are derived from material licensed * to the University of California by American Telephone and Telegraph * Co. or Unix System Laboratories, Inc. and are reproduced herein with * the permission of UNIX System Laboratories, Inc. * * Copyright (c) 2012 Konstantin Belousov * Copyright (c) 2013, 2014 The FreeBSD Foundation * * Portions of this software were developed by Konstantin Belousov * under sponsorship from the FreeBSD Foundation. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * 3. Neither the name of the University nor the names of its contributors * may be used to endorse or promote products derived from this software * without specific prior written permission. * * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF * SUCH DAMAGE. * * @(#)vfs_vnops.c 8.2 (Berkeley) 1/21/94 */ #include __FBSDID("$FreeBSD$"); #include "opt_hwpmc_hooks.h" #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #include #ifdef HWPMC_HOOKS #include #endif static fo_rdwr_t vn_read; static fo_rdwr_t vn_write; static fo_rdwr_t vn_io_fault; static fo_truncate_t vn_truncate; static fo_ioctl_t vn_ioctl; static fo_poll_t vn_poll; static fo_kqfilter_t vn_kqfilter; static fo_stat_t vn_statfile; static fo_close_t vn_closefile; static fo_mmap_t vn_mmap; struct fileops vnops = { .fo_read = vn_io_fault, .fo_write = vn_io_fault, .fo_truncate = vn_truncate, .fo_ioctl = vn_ioctl, .fo_poll = vn_poll, .fo_kqfilter = vn_kqfilter, .fo_stat = vn_statfile, .fo_close = vn_closefile, .fo_chmod = vn_chmod, .fo_chown = vn_chown, .fo_sendfile = vn_sendfile, .fo_seek = vn_seek, .fo_fill_kinfo = vn_fill_kinfo, .fo_mmap = vn_mmap, .fo_flags = DFLAG_PASSABLE | DFLAG_SEEKABLE }; static const int io_hold_cnt = 16; static int vn_io_fault_enable = 1; SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_enable, CTLFLAG_RW, &vn_io_fault_enable, 0, "Enable vn_io_fault lock avoidance"); static int vn_io_fault_prefault = 0; SYSCTL_INT(_debug, OID_AUTO, vn_io_fault_prefault, CTLFLAG_RW, &vn_io_fault_prefault, 0, "Enable vn_io_fault prefaulting"); static u_long vn_io_faults_cnt; SYSCTL_ULONG(_debug, OID_AUTO, vn_io_faults, CTLFLAG_RD, &vn_io_faults_cnt, 0, "Count of vn_io_fault lock avoidance triggers"); /* * Returns true if vn_io_fault mode of handling the i/o request should * be used. */ static bool do_vn_io_fault(struct vnode *vp, struct uio *uio) { struct mount *mp; return (uio->uio_segflg == UIO_USERSPACE && vp->v_type == VREG && (mp = vp->v_mount) != NULL && (mp->mnt_kern_flag & MNTK_NO_IOPF) != 0 && vn_io_fault_enable); } /* * Structure used to pass arguments to vn_io_fault1(), to do either * file- or vnode-based I/O calls. */ struct vn_io_fault_args { enum { VN_IO_FAULT_FOP, VN_IO_FAULT_VOP } kind; struct ucred *cred; int flags; union { struct fop_args_tag { struct file *fp; fo_rdwr_t *doio; } fop_args; struct vop_args_tag { struct vnode *vp; } vop_args; } args; }; static int vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args, struct thread *td); int vn_open(ndp, flagp, cmode, fp) struct nameidata *ndp; int *flagp, cmode; struct file *fp; { struct thread *td = ndp->ni_cnd.cn_thread; return (vn_open_cred(ndp, flagp, cmode, 0, td->td_ucred, fp)); } /* * Common code for vnode open operations via a name lookup. * Lookup the vnode and invoke VOP_CREATE if needed. * Check permissions, and call the VOP_OPEN or VOP_CREATE routine. * * Note that this does NOT free nameidata for the successful case, * due to the NDINIT being done elsewhere. */ int vn_open_cred(struct nameidata *ndp, int *flagp, int cmode, u_int vn_open_flags, struct ucred *cred, struct file *fp) { struct vnode *vp; struct mount *mp; struct thread *td = ndp->ni_cnd.cn_thread; struct vattr vat; struct vattr *vap = &vat; int fmode, error; restart: fmode = *flagp; if ((fmode & (O_CREAT | O_EXCL | O_DIRECTORY)) == (O_CREAT | O_EXCL | O_DIRECTORY)) return (EINVAL); else if ((fmode & (O_CREAT | O_DIRECTORY)) == O_CREAT) { ndp->ni_cnd.cn_nameiop = CREATE; /* * Set NOCACHE to avoid flushing the cache when * rolling in many files at once. */ ndp->ni_cnd.cn_flags = ISOPEN | LOCKPARENT | LOCKLEAF | NOCACHE; if ((fmode & O_EXCL) == 0 && (fmode & O_NOFOLLOW) == 0) ndp->ni_cnd.cn_flags |= FOLLOW; if (!(vn_open_flags & VN_OPEN_NOAUDIT)) ndp->ni_cnd.cn_flags |= AUDITVNODE1; if (vn_open_flags & VN_OPEN_NOCAPCHECK) ndp->ni_cnd.cn_flags |= NOCAPCHECK; bwillwrite(); if ((error = namei(ndp)) != 0) return (error); if (ndp->ni_vp == NULL) { VATTR_NULL(vap); vap->va_type = VREG; vap->va_mode = cmode; if (fmode & O_EXCL) vap->va_vaflags |= VA_EXCLUSIVE; if (vn_start_write(ndp->ni_dvp, &mp, V_NOWAIT) != 0) { NDFREE(ndp, NDF_ONLY_PNBUF); vput(ndp->ni_dvp); if ((error = vn_start_write(NULL, &mp, V_XSLEEP | PCATCH)) != 0) return (error); goto restart; } if ((vn_open_flags & VN_OPEN_NAMECACHE) != 0) ndp->ni_cnd.cn_flags |= MAKEENTRY; #ifdef MAC error = mac_vnode_check_create(cred, ndp->ni_dvp, &ndp->ni_cnd, vap); if (error == 0) #endif error = VOP_CREATE(ndp->ni_dvp, &ndp->ni_vp, &ndp->ni_cnd, vap); vput(ndp->ni_dvp); vn_finished_write(mp); if (error) { NDFREE(ndp, NDF_ONLY_PNBUF); return (error); } fmode &= ~O_TRUNC; vp = ndp->ni_vp; } else { if (ndp->ni_dvp == ndp->ni_vp) vrele(ndp->ni_dvp); else vput(ndp->ni_dvp); ndp->ni_dvp = NULL; vp = ndp->ni_vp; if (fmode & O_EXCL) { error = EEXIST; goto bad; } fmode &= ~O_CREAT; } } else { ndp->ni_cnd.cn_nameiop = LOOKUP; ndp->ni_cnd.cn_flags = ISOPEN | ((fmode & O_NOFOLLOW) ? NOFOLLOW : FOLLOW) | LOCKLEAF; if (!(fmode & FWRITE)) ndp->ni_cnd.cn_flags |= LOCKSHARED; if (!(vn_open_flags & VN_OPEN_NOAUDIT)) ndp->ni_cnd.cn_flags |= AUDITVNODE1; if (vn_open_flags & VN_OPEN_NOCAPCHECK) ndp->ni_cnd.cn_flags |= NOCAPCHECK; if ((error = namei(ndp)) != 0) return (error); vp = ndp->ni_vp; } error = vn_open_vnode(vp, fmode, cred, td, fp); if (error) goto bad; *flagp = fmode; return (0); bad: NDFREE(ndp, NDF_ONLY_PNBUF); vput(vp); *flagp = fmode; ndp->ni_vp = NULL; return (error); } /* * Common code for vnode open operations once a vnode is located. * Check permissions, and call the VOP_OPEN routine. */ int vn_open_vnode(struct vnode *vp, int fmode, struct ucred *cred, struct thread *td, struct file *fp) { accmode_t accmode; struct flock lf; int error, lock_flags, type; if (vp->v_type == VLNK) return (EMLINK); if (vp->v_type == VSOCK) return (EOPNOTSUPP); if (vp->v_type != VDIR && fmode & O_DIRECTORY) return (ENOTDIR); accmode = 0; if (fmode & (FWRITE | O_TRUNC)) { if (vp->v_type == VDIR) return (EISDIR); accmode |= VWRITE; } if (fmode & FREAD) accmode |= VREAD; if (fmode & FEXEC) accmode |= VEXEC; if ((fmode & O_APPEND) && (fmode & FWRITE)) accmode |= VAPPEND; #ifdef MAC if (fmode & O_CREAT) accmode |= VCREAT; if (fmode & O_VERIFY) accmode |= VVERIFY; error = mac_vnode_check_open(cred, vp, accmode); if (error) return (error); accmode &= ~(VCREAT | VVERIFY); #endif if ((fmode & O_CREAT) == 0) { if (accmode & VWRITE) { error = vn_writechk(vp); if (error) return (error); } if (accmode) { error = VOP_ACCESS(vp, accmode, cred, td); if (error) return (error); } } if (vp->v_type == VFIFO && VOP_ISLOCKED(vp) != LK_EXCLUSIVE) vn_lock(vp, LK_UPGRADE | LK_RETRY); if ((error = VOP_OPEN(vp, fmode, cred, td, fp)) != 0) return (error); while ((fmode & (O_EXLOCK | O_SHLOCK)) != 0) { KASSERT(fp != NULL, ("open with flock requires fp")); if (fp->f_type != DTYPE_NONE && fp->f_type != DTYPE_VNODE) { error = EOPNOTSUPP; break; } lock_flags = VOP_ISLOCKED(vp); VOP_UNLOCK(vp, 0); lf.l_whence = SEEK_SET; lf.l_start = 0; lf.l_len = 0; if (fmode & O_EXLOCK) lf.l_type = F_WRLCK; else lf.l_type = F_RDLCK; type = F_FLOCK; if ((fmode & FNONBLOCK) == 0) type |= F_WAIT; error = VOP_ADVLOCK(vp, (caddr_t)fp, F_SETLK, &lf, type); if (error == 0) fp->f_flag |= FHASLOCK; vn_lock(vp, lock_flags | LK_RETRY); if (error != 0) break; if ((vp->v_iflag & VI_DOOMED) != 0) { error = ENOENT; break; } /* * Another thread might have used this vnode as an * executable while the vnode lock was dropped. * Ensure the vnode is still able to be opened for * writing after the lock has been obtained. */ if ((accmode & VWRITE) != 0) error = vn_writechk(vp); break; } if (error != 0) { fp->f_flag |= FOPENFAILED; fp->f_vnode = vp; if (fp->f_ops == &badfileops) { fp->f_type = DTYPE_VNODE; fp->f_ops = &vnops; } vref(vp); } else if ((fmode & FWRITE) != 0) { VOP_ADD_WRITECOUNT(vp, 1); CTR3(KTR_VFS, "%s: vp %p v_writecount increased to %d", __func__, vp, vp->v_writecount); } ASSERT_VOP_LOCKED(vp, "vn_open_vnode"); return (error); } /* * Check for write permissions on the specified vnode. * Prototype text segments cannot be written. */ int vn_writechk(struct vnode *vp) { ASSERT_VOP_LOCKED(vp, "vn_writechk"); /* * If there's shared text associated with * the vnode, try to free it up once. If * we fail, we can't allow writing. */ if (VOP_IS_TEXT(vp)) return (ETXTBSY); return (0); } /* * Vnode close call */ static int vn_close1(struct vnode *vp, int flags, struct ucred *file_cred, struct thread *td, bool keep_ref) { struct mount *mp; int error, lock_flags; if (vp->v_type != VFIFO && (flags & FWRITE) == 0 && MNT_EXTENDED_SHARED(vp->v_mount)) lock_flags = LK_SHARED; else lock_flags = LK_EXCLUSIVE; vn_start_write(vp, &mp, V_WAIT); vn_lock(vp, lock_flags | LK_RETRY); AUDIT_ARG_VNODE1(vp); if ((flags & (FWRITE | FOPENFAILED)) == FWRITE) { VNASSERT(vp->v_writecount > 0, vp, ("vn_close: negative writecount")); VOP_ADD_WRITECOUNT(vp, -1); CTR3(KTR_VFS, "%s: vp %p v_writecount decreased to %d", __func__, vp, vp->v_writecount); } error = VOP_CLOSE(vp, flags, file_cred, td); if (keep_ref) VOP_UNLOCK(vp, 0); else vput(vp); vn_finished_write(mp); return (error); } int vn_close(struct vnode *vp, int flags, struct ucred *file_cred, struct thread *td) { return (vn_close1(vp, flags, file_cred, td, false)); } /* * Heuristic to detect sequential operation. */ static int sequential_heuristic(struct uio *uio, struct file *fp) { ASSERT_VOP_LOCKED(fp->f_vnode, __func__); if (fp->f_flag & FRDAHEAD) return (fp->f_seqcount << IO_SEQSHIFT); /* * Offset 0 is handled specially. open() sets f_seqcount to 1 so * that the first I/O is normally considered to be slightly * sequential. Seeking to offset 0 doesn't change sequentiality * unless previous seeks have reduced f_seqcount to 0, in which * case offset 0 is not special. */ if ((uio->uio_offset == 0 && fp->f_seqcount > 0) || uio->uio_offset == fp->f_nextoff) { /* * f_seqcount is in units of fixed-size blocks so that it * depends mainly on the amount of sequential I/O and not * much on the number of sequential I/O's. The fixed size * of 16384 is hard-coded here since it is (not quite) just * a magic size that works well here. This size is more * closely related to the best I/O size for real disks than * to any block size used by software. */ fp->f_seqcount += howmany(uio->uio_resid, 16384); if (fp->f_seqcount > IO_SEQMAX) fp->f_seqcount = IO_SEQMAX; return (fp->f_seqcount << IO_SEQSHIFT); } /* Not sequential. Quickly draw-down sequentiality. */ if (fp->f_seqcount > 1) fp->f_seqcount = 1; else fp->f_seqcount = 0; return (0); } /* * Package up an I/O request on a vnode into a uio and do it. */ int vn_rdwr(enum uio_rw rw, struct vnode *vp, void *base, int len, off_t offset, enum uio_seg segflg, int ioflg, struct ucred *active_cred, struct ucred *file_cred, ssize_t *aresid, struct thread *td) { struct uio auio; struct iovec aiov; struct mount *mp; struct ucred *cred; void *rl_cookie; struct vn_io_fault_args args; int error, lock_flags; auio.uio_iov = &aiov; auio.uio_iovcnt = 1; aiov.iov_base = base; aiov.iov_len = len; auio.uio_resid = len; auio.uio_offset = offset; auio.uio_segflg = segflg; auio.uio_rw = rw; auio.uio_td = td; error = 0; if ((ioflg & IO_NODELOCKED) == 0) { if ((ioflg & IO_RANGELOCKED) == 0) { if (rw == UIO_READ) { rl_cookie = vn_rangelock_rlock(vp, offset, offset + len); } else { rl_cookie = vn_rangelock_wlock(vp, offset, offset + len); } } else rl_cookie = NULL; mp = NULL; if (rw == UIO_WRITE) { if (vp->v_type != VCHR && (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0) goto out; if (MNT_SHARED_WRITES(mp) || ((mp == NULL) && MNT_SHARED_WRITES(vp->v_mount))) lock_flags = LK_SHARED; else lock_flags = LK_EXCLUSIVE; } else lock_flags = LK_SHARED; vn_lock(vp, lock_flags | LK_RETRY); } else rl_cookie = NULL; ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); #ifdef MAC if ((ioflg & IO_NOMACCHECK) == 0) { if (rw == UIO_READ) error = mac_vnode_check_read(active_cred, file_cred, vp); else error = mac_vnode_check_write(active_cred, file_cred, vp); } #endif if (error == 0) { if (file_cred != NULL) cred = file_cred; else cred = active_cred; if (do_vn_io_fault(vp, &auio)) { args.kind = VN_IO_FAULT_VOP; args.cred = cred; args.flags = ioflg; args.args.vop_args.vp = vp; error = vn_io_fault1(vp, &auio, &args, td); } else if (rw == UIO_READ) { error = VOP_READ(vp, &auio, ioflg, cred); } else /* if (rw == UIO_WRITE) */ { error = VOP_WRITE(vp, &auio, ioflg, cred); } } if (aresid) *aresid = auio.uio_resid; else if (auio.uio_resid && error == 0) error = EIO; if ((ioflg & IO_NODELOCKED) == 0) { VOP_UNLOCK(vp, 0); if (mp != NULL) vn_finished_write(mp); } out: if (rl_cookie != NULL) vn_rangelock_unlock(vp, rl_cookie); return (error); } /* * Package up an I/O request on a vnode into a uio and do it. The I/O * request is split up into smaller chunks and we try to avoid saturating * the buffer cache while potentially holding a vnode locked, so we * check bwillwrite() before calling vn_rdwr(). We also call kern_yield() * to give other processes a chance to lock the vnode (either other processes * core'ing the same binary, or unrelated processes scanning the directory). */ int vn_rdwr_inchunks(rw, vp, base, len, offset, segflg, ioflg, active_cred, file_cred, aresid, td) enum uio_rw rw; struct vnode *vp; void *base; size_t len; off_t offset; enum uio_seg segflg; int ioflg; struct ucred *active_cred; struct ucred *file_cred; size_t *aresid; struct thread *td; { int error = 0; ssize_t iaresid; do { int chunk; /* * Force `offset' to a multiple of MAXBSIZE except possibly * for the first chunk, so that filesystems only need to * write full blocks except possibly for the first and last * chunks. */ chunk = MAXBSIZE - (uoff_t)offset % MAXBSIZE; if (chunk > len) chunk = len; if (rw != UIO_READ && vp->v_type == VREG) bwillwrite(); iaresid = 0; error = vn_rdwr(rw, vp, base, chunk, offset, segflg, ioflg, active_cred, file_cred, &iaresid, td); len -= chunk; /* aresid calc already includes length */ if (error) break; offset += chunk; base = (char *)base + chunk; kern_yield(PRI_USER); } while (len); if (aresid) *aresid = len + iaresid; return (error); } off_t foffset_lock(struct file *fp, int flags) { struct mtx *mtxp; off_t res; KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed")); #if OFF_MAX <= LONG_MAX /* * Caller only wants the current f_offset value. Assume that * the long and shorter integer types reads are atomic. */ if ((flags & FOF_NOLOCK) != 0) return (fp->f_offset); #endif /* * According to McKusick the vn lock was protecting f_offset here. * It is now protected by the FOFFSET_LOCKED flag. */ mtxp = mtx_pool_find(mtxpool_sleep, fp); mtx_lock(mtxp); if ((flags & FOF_NOLOCK) == 0) { while (fp->f_vnread_flags & FOFFSET_LOCKED) { fp->f_vnread_flags |= FOFFSET_LOCK_WAITING; msleep(&fp->f_vnread_flags, mtxp, PUSER -1, "vofflock", 0); } fp->f_vnread_flags |= FOFFSET_LOCKED; } res = fp->f_offset; mtx_unlock(mtxp); return (res); } void foffset_unlock(struct file *fp, off_t val, int flags) { struct mtx *mtxp; KASSERT((flags & FOF_OFFSET) == 0, ("FOF_OFFSET passed")); #if OFF_MAX <= LONG_MAX if ((flags & FOF_NOLOCK) != 0) { if ((flags & FOF_NOUPDATE) == 0) fp->f_offset = val; if ((flags & FOF_NEXTOFF) != 0) fp->f_nextoff = val; return; } #endif mtxp = mtx_pool_find(mtxpool_sleep, fp); mtx_lock(mtxp); if ((flags & FOF_NOUPDATE) == 0) fp->f_offset = val; if ((flags & FOF_NEXTOFF) != 0) fp->f_nextoff = val; if ((flags & FOF_NOLOCK) == 0) { KASSERT((fp->f_vnread_flags & FOFFSET_LOCKED) != 0, ("Lost FOFFSET_LOCKED")); if (fp->f_vnread_flags & FOFFSET_LOCK_WAITING) wakeup(&fp->f_vnread_flags); fp->f_vnread_flags = 0; } mtx_unlock(mtxp); } void foffset_lock_uio(struct file *fp, struct uio *uio, int flags) { if ((flags & FOF_OFFSET) == 0) uio->uio_offset = foffset_lock(fp, flags); } void foffset_unlock_uio(struct file *fp, struct uio *uio, int flags) { if ((flags & FOF_OFFSET) == 0) foffset_unlock(fp, uio->uio_offset, flags); } static int get_advice(struct file *fp, struct uio *uio) { struct mtx *mtxp; int ret; ret = POSIX_FADV_NORMAL; if (fp->f_advice == NULL || fp->f_vnode->v_type != VREG) return (ret); mtxp = mtx_pool_find(mtxpool_sleep, fp); mtx_lock(mtxp); if (fp->f_advice != NULL && uio->uio_offset >= fp->f_advice->fa_start && uio->uio_offset + uio->uio_resid <= fp->f_advice->fa_end) ret = fp->f_advice->fa_advice; mtx_unlock(mtxp); return (ret); } /* * File table vnode read routine. */ static int vn_read(fp, uio, active_cred, flags, td) struct file *fp; struct uio *uio; struct ucred *active_cred; int flags; struct thread *td; { struct vnode *vp; off_t orig_offset; int error, ioflag; int advice; KASSERT(uio->uio_td == td, ("uio_td %p is not td %p", uio->uio_td, td)); KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET")); vp = fp->f_vnode; ioflag = 0; if (fp->f_flag & FNONBLOCK) ioflag |= IO_NDELAY; if (fp->f_flag & O_DIRECT) ioflag |= IO_DIRECT; advice = get_advice(fp, uio); vn_lock(vp, LK_SHARED | LK_RETRY); switch (advice) { case POSIX_FADV_NORMAL: case POSIX_FADV_SEQUENTIAL: case POSIX_FADV_NOREUSE: ioflag |= sequential_heuristic(uio, fp); break; case POSIX_FADV_RANDOM: /* Disable read-ahead for random I/O. */ break; } orig_offset = uio->uio_offset; #ifdef MAC error = mac_vnode_check_read(active_cred, fp->f_cred, vp); if (error == 0) #endif error = VOP_READ(vp, uio, ioflag, fp->f_cred); fp->f_nextoff = uio->uio_offset; VOP_UNLOCK(vp, 0); if (error == 0 && advice == POSIX_FADV_NOREUSE && orig_offset != uio->uio_offset) /* * Use POSIX_FADV_DONTNEED to flush pages and buffers * for the backing file after a POSIX_FADV_NOREUSE * read(2). */ error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1, POSIX_FADV_DONTNEED); return (error); } /* * File table vnode write routine. */ static int vn_write(fp, uio, active_cred, flags, td) struct file *fp; struct uio *uio; struct ucred *active_cred; int flags; struct thread *td; { struct vnode *vp; struct mount *mp; off_t orig_offset; int error, ioflag, lock_flags; int advice; KASSERT(uio->uio_td == td, ("uio_td %p is not td %p", uio->uio_td, td)); KASSERT(flags & FOF_OFFSET, ("No FOF_OFFSET")); vp = fp->f_vnode; if (vp->v_type == VREG) bwillwrite(); ioflag = IO_UNIT; if (vp->v_type == VREG && (fp->f_flag & O_APPEND)) ioflag |= IO_APPEND; if (fp->f_flag & FNONBLOCK) ioflag |= IO_NDELAY; if (fp->f_flag & O_DIRECT) ioflag |= IO_DIRECT; if ((fp->f_flag & O_FSYNC) || (vp->v_mount && (vp->v_mount->mnt_flag & MNT_SYNCHRONOUS))) ioflag |= IO_SYNC; mp = NULL; if (vp->v_type != VCHR && (error = vn_start_write(vp, &mp, V_WAIT | PCATCH)) != 0) goto unlock; advice = get_advice(fp, uio); if (MNT_SHARED_WRITES(mp) || (mp == NULL && MNT_SHARED_WRITES(vp->v_mount))) { lock_flags = LK_SHARED; } else { lock_flags = LK_EXCLUSIVE; } vn_lock(vp, lock_flags | LK_RETRY); switch (advice) { case POSIX_FADV_NORMAL: case POSIX_FADV_SEQUENTIAL: case POSIX_FADV_NOREUSE: ioflag |= sequential_heuristic(uio, fp); break; case POSIX_FADV_RANDOM: /* XXX: Is this correct? */ break; } orig_offset = uio->uio_offset; #ifdef MAC error = mac_vnode_check_write(active_cred, fp->f_cred, vp); if (error == 0) #endif error = VOP_WRITE(vp, uio, ioflag, fp->f_cred); fp->f_nextoff = uio->uio_offset; VOP_UNLOCK(vp, 0); if (vp->v_type != VCHR) vn_finished_write(mp); if (error == 0 && advice == POSIX_FADV_NOREUSE && orig_offset != uio->uio_offset) /* * Use POSIX_FADV_DONTNEED to flush pages and buffers * for the backing file after a POSIX_FADV_NOREUSE * write(2). */ error = VOP_ADVISE(vp, orig_offset, uio->uio_offset - 1, POSIX_FADV_DONTNEED); unlock: return (error); } /* * The vn_io_fault() is a wrapper around vn_read() and vn_write() to * prevent the following deadlock: * * Assume that the thread A reads from the vnode vp1 into userspace * buffer buf1 backed by the pages of vnode vp2. If a page in buf1 is * currently not resident, then system ends up with the call chain * vn_read() -> VOP_READ(vp1) -> uiomove() -> [Page Fault] -> * vm_fault(buf1) -> vnode_pager_getpages(vp2) -> VOP_GETPAGES(vp2) * which establishes lock order vp1->vn_lock, then vp2->vn_lock. * If, at the same time, thread B reads from vnode vp2 into buffer buf2 * backed by the pages of vnode vp1, and some page in buf2 is not * resident, we get a reversed order vp2->vn_lock, then vp1->vn_lock. * * To prevent the lock order reversal and deadlock, vn_io_fault() does * not allow page faults to happen during VOP_READ() or VOP_WRITE(). * Instead, it first tries to do the whole range i/o with pagefaults * disabled. If all pages in the i/o buffer are resident and mapped, * VOP will succeed (ignoring the genuine filesystem errors). * Otherwise, we get back EFAULT, and vn_io_fault() falls back to do * i/o in chunks, with all pages in the chunk prefaulted and held * using vm_fault_quick_hold_pages(). * * Filesystems using this deadlock avoidance scheme should use the * array of the held pages from uio, saved in the curthread->td_ma, * instead of doing uiomove(). A helper function * vn_io_fault_uiomove() converts uiomove request into * uiomove_fromphys() over td_ma array. * * Since vnode locks do not cover the whole i/o anymore, rangelocks * make the current i/o request atomic with respect to other i/os and * truncations. */ /* * Decode vn_io_fault_args and perform the corresponding i/o. */ static int vn_io_fault_doio(struct vn_io_fault_args *args, struct uio *uio, struct thread *td) { int error, save; error = 0; save = vm_fault_disable_pagefaults(); switch (args->kind) { case VN_IO_FAULT_FOP: error = (args->args.fop_args.doio)(args->args.fop_args.fp, uio, args->cred, args->flags, td); break; case VN_IO_FAULT_VOP: if (uio->uio_rw == UIO_READ) { error = VOP_READ(args->args.vop_args.vp, uio, args->flags, args->cred); } else if (uio->uio_rw == UIO_WRITE) { error = VOP_WRITE(args->args.vop_args.vp, uio, args->flags, args->cred); } break; default: panic("vn_io_fault_doio: unknown kind of io %d %d", args->kind, uio->uio_rw); } vm_fault_enable_pagefaults(save); return (error); } static int vn_io_fault_touch(char *base, const struct uio *uio) { int r; r = fubyte(base); if (r == -1 || (uio->uio_rw == UIO_READ && subyte(base, r) == -1)) return (EFAULT); return (0); } static int vn_io_fault_prefault_user(const struct uio *uio) { char *base; const struct iovec *iov; size_t len; ssize_t resid; int error, i; KASSERT(uio->uio_segflg == UIO_USERSPACE, ("vn_io_fault_prefault userspace")); error = i = 0; iov = uio->uio_iov; resid = uio->uio_resid; base = iov->iov_base; len = iov->iov_len; while (resid > 0) { error = vn_io_fault_touch(base, uio); if (error != 0) break; if (len < PAGE_SIZE) { if (len != 0) { error = vn_io_fault_touch(base + len - 1, uio); if (error != 0) break; resid -= len; } if (++i >= uio->uio_iovcnt) break; iov = uio->uio_iov + i; base = iov->iov_base; len = iov->iov_len; } else { len -= PAGE_SIZE; base += PAGE_SIZE; resid -= PAGE_SIZE; } } return (error); } /* * Common code for vn_io_fault(), agnostic to the kind of i/o request. * Uses vn_io_fault_doio() to make the call to an actual i/o function. * Used from vn_rdwr() and vn_io_fault(), which encode the i/o request * into args and call vn_io_fault1() to handle faults during the user * mode buffer accesses. */ static int vn_io_fault1(struct vnode *vp, struct uio *uio, struct vn_io_fault_args *args, struct thread *td) { vm_page_t ma[io_hold_cnt + 2]; struct uio *uio_clone, short_uio; struct iovec short_iovec[1]; vm_page_t *prev_td_ma; vm_prot_t prot; vm_offset_t addr, end; size_t len, resid; ssize_t adv; int error, cnt, saveheld, prev_td_ma_cnt; if (vn_io_fault_prefault) { error = vn_io_fault_prefault_user(uio); if (error != 0) return (error); /* Or ignore ? */ } prot = uio->uio_rw == UIO_READ ? VM_PROT_WRITE : VM_PROT_READ; /* * The UFS follows IO_UNIT directive and replays back both * uio_offset and uio_resid if an error is encountered during the * operation. But, since the iovec may be already advanced, * uio is still in an inconsistent state. * * Cache a copy of the original uio, which is advanced to the redo * point using UIO_NOCOPY below. */ uio_clone = cloneuio(uio); resid = uio->uio_resid; short_uio.uio_segflg = UIO_USERSPACE; short_uio.uio_rw = uio->uio_rw; short_uio.uio_td = uio->uio_td; error = vn_io_fault_doio(args, uio, td); if (error != EFAULT) goto out; atomic_add_long(&vn_io_faults_cnt, 1); uio_clone->uio_segflg = UIO_NOCOPY; uiomove(NULL, resid - uio->uio_resid, uio_clone); uio_clone->uio_segflg = uio->uio_segflg; saveheld = curthread_pflags_set(TDP_UIOHELD); prev_td_ma = td->td_ma; prev_td_ma_cnt = td->td_ma_cnt; while (uio_clone->uio_resid != 0) { len = uio_clone->uio_iov->iov_len; if (len == 0) { KASSERT(uio_clone->uio_iovcnt >= 1, ("iovcnt underflow")); uio_clone->uio_iov++; uio_clone->uio_iovcnt--; continue; } if (len > io_hold_cnt * PAGE_SIZE) len = io_hold_cnt * PAGE_SIZE; addr = (uintptr_t)uio_clone->uio_iov->iov_base; end = round_page(addr + len); if (end < addr) { error = EFAULT; break; } cnt = atop(end - trunc_page(addr)); /* * A perfectly misaligned address and length could cause * both the start and the end of the chunk to use partial * page. +2 accounts for such a situation. */ cnt = vm_fault_quick_hold_pages(&td->td_proc->p_vmspace->vm_map, addr, len, prot, ma, io_hold_cnt + 2); if (cnt == -1) { error = EFAULT; break; } short_uio.uio_iov = &short_iovec[0]; short_iovec[0].iov_base = (void *)addr; short_uio.uio_iovcnt = 1; short_uio.uio_resid = short_iovec[0].iov_len = len; short_uio.uio_offset = uio_clone->uio_offset; td->td_ma = ma; td->td_ma_cnt = cnt; error = vn_io_fault_doio(args, &short_uio, td); vm_page_unhold_pages(ma, cnt); adv = len - short_uio.uio_resid; uio_clone->uio_iov->iov_base = (char *)uio_clone->uio_iov->iov_base + adv; uio_clone->uio_iov->iov_len -= adv; uio_clone->uio_resid -= adv; uio_clone->uio_offset += adv; uio->uio_resid -= adv; uio->uio_offset += adv; if (error != 0 || adv == 0) break; } td->td_ma = prev_td_ma; td->td_ma_cnt = prev_td_ma_cnt; curthread_pflags_restore(saveheld); out: free(uio_clone, M_IOV); return (error); } static int vn_io_fault(struct file *fp, struct uio *uio, struct ucred *active_cred, int flags, struct thread *td) { fo_rdwr_t *doio; struct vnode *vp; void *rl_cookie; struct vn_io_fault_args args; int error; doio = uio->uio_rw == UIO_READ ? vn_read : vn_write; vp = fp->f_vnode; foffset_lock_uio(fp, uio, flags); if (do_vn_io_fault(vp, uio)) { args.kind = VN_IO_FAULT_FOP; args.args.fop_args.fp = fp; args.args.fop_args.doio = doio; args.cred = active_cred; args.flags = flags | FOF_OFFSET; if (uio->uio_rw == UIO_READ) { rl_cookie = vn_rangelock_rlock(vp, uio->uio_offset, uio->uio_offset + uio->uio_resid); } else if ((fp->f_flag & O_APPEND) != 0 || (flags & FOF_OFFSET) == 0) { /* For appenders, punt and lock the whole range. */ rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX); } else { rl_cookie = vn_rangelock_wlock(vp, uio->uio_offset, uio->uio_offset + uio->uio_resid); } error = vn_io_fault1(vp, uio, &args, td); vn_rangelock_unlock(vp, rl_cookie); } else { error = doio(fp, uio, active_cred, flags | FOF_OFFSET, td); } foffset_unlock_uio(fp, uio, flags); return (error); } /* * Helper function to perform the requested uiomove operation using * the held pages for io->uio_iov[0].iov_base buffer instead of * copyin/copyout. Access to the pages with uiomove_fromphys() * instead of iov_base prevents page faults that could occur due to * pmap_collect() invalidating the mapping created by * vm_fault_quick_hold_pages(), or pageout daemon, page laundry or * object cleanup revoking the write access from page mappings. * * Filesystems specified MNTK_NO_IOPF shall use vn_io_fault_uiomove() * instead of plain uiomove(). */ int vn_io_fault_uiomove(char *data, int xfersize, struct uio *uio) { struct uio transp_uio; struct iovec transp_iov[1]; struct thread *td; size_t adv; int error, pgadv; td = curthread; if ((td->td_pflags & TDP_UIOHELD) == 0 || uio->uio_segflg != UIO_USERSPACE) return (uiomove(data, xfersize, uio)); KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt)); transp_iov[0].iov_base = data; transp_uio.uio_iov = &transp_iov[0]; transp_uio.uio_iovcnt = 1; if (xfersize > uio->uio_resid) xfersize = uio->uio_resid; transp_uio.uio_resid = transp_iov[0].iov_len = xfersize; transp_uio.uio_offset = 0; transp_uio.uio_segflg = UIO_SYSSPACE; /* * Since transp_iov points to data, and td_ma page array * corresponds to original uio->uio_iov, we need to invert the * direction of the i/o operation as passed to * uiomove_fromphys(). */ switch (uio->uio_rw) { case UIO_WRITE: transp_uio.uio_rw = UIO_READ; break; case UIO_READ: transp_uio.uio_rw = UIO_WRITE; break; } transp_uio.uio_td = uio->uio_td; error = uiomove_fromphys(td->td_ma, ((vm_offset_t)uio->uio_iov->iov_base) & PAGE_MASK, xfersize, &transp_uio); adv = xfersize - transp_uio.uio_resid; pgadv = (((vm_offset_t)uio->uio_iov->iov_base + adv) >> PAGE_SHIFT) - (((vm_offset_t)uio->uio_iov->iov_base) >> PAGE_SHIFT); td->td_ma += pgadv; KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt, pgadv)); td->td_ma_cnt -= pgadv; uio->uio_iov->iov_base = (char *)uio->uio_iov->iov_base + adv; uio->uio_iov->iov_len -= adv; uio->uio_resid -= adv; uio->uio_offset += adv; return (error); } int vn_io_fault_pgmove(vm_page_t ma[], vm_offset_t offset, int xfersize, struct uio *uio) { struct thread *td; vm_offset_t iov_base; int cnt, pgadv; td = curthread; if ((td->td_pflags & TDP_UIOHELD) == 0 || uio->uio_segflg != UIO_USERSPACE) return (uiomove_fromphys(ma, offset, xfersize, uio)); KASSERT(uio->uio_iovcnt == 1, ("uio_iovcnt %d", uio->uio_iovcnt)); cnt = xfersize > uio->uio_resid ? uio->uio_resid : xfersize; iov_base = (vm_offset_t)uio->uio_iov->iov_base; switch (uio->uio_rw) { case UIO_WRITE: pmap_copy_pages(td->td_ma, iov_base & PAGE_MASK, ma, offset, cnt); break; case UIO_READ: pmap_copy_pages(ma, offset, td->td_ma, iov_base & PAGE_MASK, cnt); break; } pgadv = ((iov_base + cnt) >> PAGE_SHIFT) - (iov_base >> PAGE_SHIFT); td->td_ma += pgadv; KASSERT(td->td_ma_cnt >= pgadv, ("consumed pages %d %d", td->td_ma_cnt, pgadv)); td->td_ma_cnt -= pgadv; uio->uio_iov->iov_base = (char *)(iov_base + cnt); uio->uio_iov->iov_len -= cnt; uio->uio_resid -= cnt; uio->uio_offset += cnt; return (0); } /* * File table truncate routine. */ static int vn_truncate(struct file *fp, off_t length, struct ucred *active_cred, struct thread *td) { struct vattr vattr; struct mount *mp; struct vnode *vp; void *rl_cookie; int error; vp = fp->f_vnode; /* * Lock the whole range for truncation. Otherwise split i/o * might happen partly before and partly after the truncation. */ rl_cookie = vn_rangelock_wlock(vp, 0, OFF_MAX); error = vn_start_write(vp, &mp, V_WAIT | PCATCH); if (error) goto out1; vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); AUDIT_ARG_VNODE1(vp); if (vp->v_type == VDIR) { error = EISDIR; goto out; } #ifdef MAC error = mac_vnode_check_write(active_cred, fp->f_cred, vp); if (error) goto out; #endif error = vn_writechk(vp); if (error == 0) { VATTR_NULL(&vattr); vattr.va_size = length; if ((fp->f_flag & O_FSYNC) != 0) vattr.va_vaflags |= VA_SYNC; error = VOP_SETATTR(vp, &vattr, fp->f_cred); } out: VOP_UNLOCK(vp, 0); vn_finished_write(mp); out1: vn_rangelock_unlock(vp, rl_cookie); return (error); } /* * File table vnode stat routine. */ static int vn_statfile(fp, sb, active_cred, td) struct file *fp; struct stat *sb; struct ucred *active_cred; struct thread *td; { struct vnode *vp = fp->f_vnode; int error; vn_lock(vp, LK_SHARED | LK_RETRY); error = vn_stat(vp, sb, active_cred, fp->f_cred, td); VOP_UNLOCK(vp, 0); return (error); } /* * Stat a vnode; implementation for the stat syscall */ int vn_stat(struct vnode *vp, struct stat *sb, struct ucred *active_cred, struct ucred *file_cred, struct thread *td) { struct vattr vattr; struct vattr *vap; int error; u_short mode; AUDIT_ARG_VNODE1(vp); #ifdef MAC error = mac_vnode_check_stat(active_cred, file_cred, vp); if (error) return (error); #endif vap = &vattr; /* * Initialize defaults for new and unusual fields, so that file * systems which don't support these fields don't need to know * about them. */ vap->va_birthtime.tv_sec = -1; vap->va_birthtime.tv_nsec = 0; vap->va_fsid = VNOVAL; vap->va_rdev = NODEV; error = VOP_GETATTR(vp, vap, active_cred); if (error) return (error); /* * Zero the spare stat fields */ bzero(sb, sizeof *sb); /* * Copy from vattr table */ if (vap->va_fsid != VNOVAL) sb->st_dev = vap->va_fsid; else sb->st_dev = vp->v_mount->mnt_stat.f_fsid.val[0]; sb->st_ino = vap->va_fileid; mode = vap->va_mode; switch (vap->va_type) { case VREG: mode |= S_IFREG; break; case VDIR: mode |= S_IFDIR; break; case VBLK: mode |= S_IFBLK; break; case VCHR: mode |= S_IFCHR; break; case VLNK: mode |= S_IFLNK; break; case VSOCK: mode |= S_IFSOCK; break; case VFIFO: mode |= S_IFIFO; break; default: return (EBADF); } sb->st_mode = mode; sb->st_nlink = vap->va_nlink; sb->st_uid = vap->va_uid; sb->st_gid = vap->va_gid; sb->st_rdev = vap->va_rdev; if (vap->va_size > OFF_MAX) return (EOVERFLOW); sb->st_size = vap->va_size; sb->st_atim = vap->va_atime; sb->st_mtim = vap->va_mtime; sb->st_ctim = vap->va_ctime; sb->st_birthtim = vap->va_birthtime; /* * According to www.opengroup.org, the meaning of st_blksize is * "a filesystem-specific preferred I/O block size for this * object. In some filesystem types, this may vary from file * to file" * Use miminum/default of PAGE_SIZE (e.g. for VCHR). */ sb->st_blksize = max(PAGE_SIZE, vap->va_blocksize); sb->st_flags = vap->va_flags; if (priv_check(td, PRIV_VFS_GENERATION)) sb->st_gen = 0; else sb->st_gen = vap->va_gen; sb->st_blocks = vap->va_bytes / S_BLKSIZE; return (0); } /* * File table vnode ioctl routine. */ static int vn_ioctl(struct file *fp, u_long com, void *data, struct ucred *active_cred, struct thread *td) { struct vattr vattr; struct vnode *vp; int error; vp = fp->f_vnode; switch (vp->v_type) { case VDIR: case VREG: switch (com) { case FIONREAD: vn_lock(vp, LK_SHARED | LK_RETRY); error = VOP_GETATTR(vp, &vattr, active_cred); VOP_UNLOCK(vp, 0); if (error == 0) *(int *)data = vattr.va_size - fp->f_offset; return (error); case FIONBIO: case FIOASYNC: return (0); default: return (VOP_IOCTL(vp, com, data, fp->f_flag, active_cred, td)); } break; case VCHR: return (VOP_IOCTL(vp, com, data, fp->f_flag, active_cred, td)); default: return (ENOTTY); } } /* * File table vnode poll routine. */ static int vn_poll(struct file *fp, int events, struct ucred *active_cred, struct thread *td) { struct vnode *vp; int error; vp = fp->f_vnode; #ifdef MAC vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); AUDIT_ARG_VNODE1(vp); error = mac_vnode_check_poll(active_cred, fp->f_cred, vp); VOP_UNLOCK(vp, 0); if (!error) #endif error = VOP_POLL(vp, events, fp->f_cred, td); return (error); } /* * Acquire the requested lock and then check for validity. LK_RETRY * permits vn_lock to return doomed vnodes. */ int _vn_lock(struct vnode *vp, int flags, char *file, int line) { int error; VNASSERT((flags & LK_TYPE_MASK) != 0, vp, ("vn_lock: no locktype")); VNASSERT(vp->v_holdcnt != 0, vp, ("vn_lock: zero hold count")); retry: error = VOP_LOCK1(vp, flags, file, line); flags &= ~LK_INTERLOCK; /* Interlock is always dropped. */ KASSERT((flags & LK_RETRY) == 0 || error == 0, ("vn_lock: error %d incompatible with flags %#x", error, flags)); if ((flags & LK_RETRY) == 0) { if (error == 0 && (vp->v_iflag & VI_DOOMED) != 0) { VOP_UNLOCK(vp, 0); error = ENOENT; } } else if (error != 0) goto retry; return (error); } /* * File table vnode close routine. */ static int vn_closefile(struct file *fp, struct thread *td) { struct vnode *vp; struct flock lf; int error; bool ref; vp = fp->f_vnode; fp->f_ops = &badfileops; ref= (fp->f_flag & FHASLOCK) != 0 && fp->f_type == DTYPE_VNODE; error = vn_close1(vp, fp->f_flag, fp->f_cred, td, ref); if (__predict_false(ref)) { lf.l_whence = SEEK_SET; lf.l_start = 0; lf.l_len = 0; lf.l_type = F_UNLCK; (void) VOP_ADVLOCK(vp, fp, F_UNLCK, &lf, F_FLOCK); vrele(vp); } return (error); } static bool vn_suspendable(struct mount *mp) { return (mp->mnt_op->vfs_susp_clean != NULL); } /* * Preparing to start a filesystem write operation. If the operation is * permitted, then we bump the count of operations in progress and * proceed. If a suspend request is in progress, we wait until the * suspension is over, and then proceed. */ static int vn_start_write_locked(struct mount *mp, int flags) { int error, mflags; mtx_assert(MNT_MTX(mp), MA_OWNED); error = 0; /* * Check on status of suspension. */ if ((curthread->td_pflags & TDP_IGNSUSP) == 0 || mp->mnt_susp_owner != curthread) { mflags = ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0) | (PUSER - 1); while ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) { if (flags & V_NOWAIT) { error = EWOULDBLOCK; goto unlock; } error = msleep(&mp->mnt_flag, MNT_MTX(mp), mflags, "suspfs", 0); if (error) goto unlock; } } if (flags & V_XSLEEP) goto unlock; mp->mnt_writeopcount++; unlock: if (error != 0 || (flags & V_XSLEEP) != 0) MNT_REL(mp); MNT_IUNLOCK(mp); return (error); } int vn_start_write(struct vnode *vp, struct mount **mpp, int flags) { struct mount *mp; int error; KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL), ("V_MNTREF requires mp")); error = 0; /* * If a vnode is provided, get and return the mount point that * to which it will write. */ if (vp != NULL) { if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) { *mpp = NULL; if (error != EOPNOTSUPP) return (error); return (0); } } if ((mp = *mpp) == NULL) return (0); if (!vn_suspendable(mp)) { if (vp != NULL || (flags & V_MNTREF) != 0) vfs_rel(mp); return (0); } /* * VOP_GETWRITEMOUNT() returns with the mp refcount held through * a vfs_ref(). * As long as a vnode is not provided we need to acquire a * refcount for the provided mountpoint too, in order to * emulate a vfs_ref(). */ MNT_ILOCK(mp); if (vp == NULL && (flags & V_MNTREF) == 0) MNT_REF(mp); return (vn_start_write_locked(mp, flags)); } /* * Secondary suspension. Used by operations such as vop_inactive * routines that are needed by the higher level functions. These * are allowed to proceed until all the higher level functions have * completed (indicated by mnt_writeopcount dropping to zero). At that * time, these operations are halted until the suspension is over. */ int vn_start_secondary_write(struct vnode *vp, struct mount **mpp, int flags) { struct mount *mp; int error; KASSERT((flags & V_MNTREF) == 0 || (*mpp != NULL && vp == NULL), ("V_MNTREF requires mp")); retry: if (vp != NULL) { if ((error = VOP_GETWRITEMOUNT(vp, mpp)) != 0) { *mpp = NULL; if (error != EOPNOTSUPP) return (error); return (0); } } /* * If we are not suspended or have not yet reached suspended * mode, then let the operation proceed. */ if ((mp = *mpp) == NULL) return (0); if (!vn_suspendable(mp)) { if (vp != NULL || (flags & V_MNTREF) != 0) vfs_rel(mp); return (0); } /* * VOP_GETWRITEMOUNT() returns with the mp refcount held through * a vfs_ref(). * As long as a vnode is not provided we need to acquire a * refcount for the provided mountpoint too, in order to * emulate a vfs_ref(). */ MNT_ILOCK(mp); if (vp == NULL && (flags & V_MNTREF) == 0) MNT_REF(mp); if ((mp->mnt_kern_flag & (MNTK_SUSPENDED | MNTK_SUSPEND2)) == 0) { mp->mnt_secondary_writes++; mp->mnt_secondary_accwrites++; MNT_IUNLOCK(mp); return (0); } if (flags & V_NOWAIT) { MNT_REL(mp); MNT_IUNLOCK(mp); return (EWOULDBLOCK); } /* * Wait for the suspension to finish. */ error = msleep(&mp->mnt_flag, MNT_MTX(mp), (PUSER - 1) | PDROP | ((mp->mnt_vfc->vfc_flags & VFCF_SBDRY) != 0 ? (flags & PCATCH) : 0), "suspfs", 0); vfs_rel(mp); if (error == 0) goto retry; return (error); } /* * Filesystem write operation has completed. If we are suspending and this * operation is the last one, notify the suspender that the suspension is * now in effect. */ void vn_finished_write(struct mount *mp) { if (mp == NULL || !vn_suspendable(mp)) return; MNT_ILOCK(mp); MNT_REL(mp); mp->mnt_writeopcount--; if (mp->mnt_writeopcount < 0) panic("vn_finished_write: neg cnt"); if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && mp->mnt_writeopcount <= 0) wakeup(&mp->mnt_writeopcount); MNT_IUNLOCK(mp); } /* * Filesystem secondary write operation has completed. If we are * suspending and this operation is the last one, notify the suspender * that the suspension is now in effect. */ void vn_finished_secondary_write(struct mount *mp) { if (mp == NULL || !vn_suspendable(mp)) return; MNT_ILOCK(mp); MNT_REL(mp); mp->mnt_secondary_writes--; if (mp->mnt_secondary_writes < 0) panic("vn_finished_secondary_write: neg cnt"); if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0 && mp->mnt_secondary_writes <= 0) wakeup(&mp->mnt_secondary_writes); MNT_IUNLOCK(mp); } /* * Request a filesystem to suspend write operations. */ int vfs_write_suspend(struct mount *mp, int flags) { int error; MPASS(vn_suspendable(mp)); MNT_ILOCK(mp); if (mp->mnt_susp_owner == curthread) { MNT_IUNLOCK(mp); return (EALREADY); } while (mp->mnt_kern_flag & MNTK_SUSPEND) msleep(&mp->mnt_flag, MNT_MTX(mp), PUSER - 1, "wsuspfs", 0); /* * Unmount holds a write reference on the mount point. If we * own busy reference and drain for writers, we deadlock with * the reference draining in the unmount path. Callers of * vfs_write_suspend() must specify VS_SKIP_UNMOUNT if * vfs_busy() reference is owned and caller is not in the * unmount context. */ if ((flags & VS_SKIP_UNMOUNT) != 0 && (mp->mnt_kern_flag & MNTK_UNMOUNT) != 0) { MNT_IUNLOCK(mp); return (EBUSY); } mp->mnt_kern_flag |= MNTK_SUSPEND; mp->mnt_susp_owner = curthread; if (mp->mnt_writeopcount > 0) (void) msleep(&mp->mnt_writeopcount, MNT_MTX(mp), (PUSER - 1)|PDROP, "suspwt", 0); else MNT_IUNLOCK(mp); if ((error = VFS_SYNC(mp, MNT_SUSPEND)) != 0) vfs_write_resume(mp, 0); return (error); } /* * Request a filesystem to resume write operations. */ void vfs_write_resume(struct mount *mp, int flags) { MPASS(vn_suspendable(mp)); MNT_ILOCK(mp); if ((mp->mnt_kern_flag & MNTK_SUSPEND) != 0) { KASSERT(mp->mnt_susp_owner == curthread, ("mnt_susp_owner")); mp->mnt_kern_flag &= ~(MNTK_SUSPEND | MNTK_SUSPEND2 | MNTK_SUSPENDED); mp->mnt_susp_owner = NULL; wakeup(&mp->mnt_writeopcount); wakeup(&mp->mnt_flag); curthread->td_pflags &= ~TDP_IGNSUSP; if ((flags & VR_START_WRITE) != 0) { MNT_REF(mp); mp->mnt_writeopcount++; } MNT_IUNLOCK(mp); if ((flags & VR_NO_SUSPCLR) == 0) VFS_SUSP_CLEAN(mp); } else if ((flags & VR_START_WRITE) != 0) { MNT_REF(mp); vn_start_write_locked(mp, 0); } else { MNT_IUNLOCK(mp); } } /* * Helper loop around vfs_write_suspend() for filesystem unmount VFS * methods. */ int vfs_write_suspend_umnt(struct mount *mp) { int error; MPASS(vn_suspendable(mp)); KASSERT((curthread->td_pflags & TDP_IGNSUSP) == 0, ("vfs_write_suspend_umnt: recursed")); /* dounmount() already called vn_start_write(). */ for (;;) { vn_finished_write(mp); error = vfs_write_suspend(mp, 0); if (error != 0) { vn_start_write(NULL, &mp, V_WAIT); return (error); } MNT_ILOCK(mp); if ((mp->mnt_kern_flag & MNTK_SUSPENDED) != 0) break; MNT_IUNLOCK(mp); vn_start_write(NULL, &mp, V_WAIT); } mp->mnt_kern_flag &= ~(MNTK_SUSPENDED | MNTK_SUSPEND2); wakeup(&mp->mnt_flag); MNT_IUNLOCK(mp); curthread->td_pflags |= TDP_IGNSUSP; return (0); } /* * Implement kqueues for files by translating it to vnode operation. */ static int vn_kqfilter(struct file *fp, struct knote *kn) { return (VOP_KQFILTER(fp->f_vnode, kn)); } /* * Simplified in-kernel wrapper calls for extended attribute access. * Both calls pass in a NULL credential, authorizing as "kernel" access. * Set IO_NODELOCKED in ioflg if the vnode is already locked. */ int vn_extattr_get(struct vnode *vp, int ioflg, int attrnamespace, const char *attrname, int *buflen, char *buf, struct thread *td) { struct uio auio; struct iovec iov; int error; iov.iov_len = *buflen; iov.iov_base = buf; auio.uio_iov = &iov; auio.uio_iovcnt = 1; auio.uio_rw = UIO_READ; auio.uio_segflg = UIO_SYSSPACE; auio.uio_td = td; auio.uio_offset = 0; auio.uio_resid = *buflen; if ((ioflg & IO_NODELOCKED) == 0) vn_lock(vp, LK_SHARED | LK_RETRY); ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); /* authorize attribute retrieval as kernel */ error = VOP_GETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, NULL, td); if ((ioflg & IO_NODELOCKED) == 0) VOP_UNLOCK(vp, 0); if (error == 0) { *buflen = *buflen - auio.uio_resid; } return (error); } /* * XXX failure mode if partially written? */ int vn_extattr_set(struct vnode *vp, int ioflg, int attrnamespace, const char *attrname, int buflen, char *buf, struct thread *td) { struct uio auio; struct iovec iov; struct mount *mp; int error; iov.iov_len = buflen; iov.iov_base = buf; auio.uio_iov = &iov; auio.uio_iovcnt = 1; auio.uio_rw = UIO_WRITE; auio.uio_segflg = UIO_SYSSPACE; auio.uio_td = td; auio.uio_offset = 0; auio.uio_resid = buflen; if ((ioflg & IO_NODELOCKED) == 0) { if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0) return (error); vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); } ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); /* authorize attribute setting as kernel */ error = VOP_SETEXTATTR(vp, attrnamespace, attrname, &auio, NULL, td); if ((ioflg & IO_NODELOCKED) == 0) { vn_finished_write(mp); VOP_UNLOCK(vp, 0); } return (error); } int vn_extattr_rm(struct vnode *vp, int ioflg, int attrnamespace, const char *attrname, struct thread *td) { struct mount *mp; int error; if ((ioflg & IO_NODELOCKED) == 0) { if ((error = vn_start_write(vp, &mp, V_WAIT)) != 0) return (error); vn_lock(vp, LK_EXCLUSIVE | LK_RETRY); } ASSERT_VOP_LOCKED(vp, "IO_NODELOCKED with no vp lock held"); /* authorize attribute removal as kernel */ error = VOP_DELETEEXTATTR(vp, attrnamespace, attrname, NULL, td); if (error == EOPNOTSUPP) error = VOP_SETEXTATTR(vp, attrnamespace, attrname, NULL, NULL, td); if ((ioflg & IO_NODELOCKED) == 0) { vn_finished_write(mp); VOP_UNLOCK(vp, 0); } return (error); } static int vn_get_ino_alloc_vget(struct mount *mp, void *arg, int lkflags, struct vnode **rvp) { return (VFS_VGET(mp, *(ino_t *)arg, lkflags, rvp)); } int vn_vget_ino(struct vnode *vp, ino_t ino, int lkflags, struct vnode **rvp) { return (vn_vget_ino_gen(vp, vn_get_ino_alloc_vget, &ino, lkflags, rvp)); } int vn_vget_ino_gen(struct vnode *vp, vn_get_ino_t alloc, void *alloc_arg, int lkflags, struct vnode **rvp) { struct mount *mp; int ltype, error; ASSERT_VOP_LOCKED(vp, "vn_vget_ino_get"); mp = vp->v_mount; ltype = VOP_ISLOCKED(vp); KASSERT(ltype == LK_EXCLUSIVE || ltype == LK_SHARED, ("vn_vget_ino: vp not locked")); error = vfs_busy(mp, MBF_NOWAIT); if (error != 0) { vfs_ref(mp); VOP_UNLOCK(vp, 0); error = vfs_busy(mp, 0); vn_lock(vp, ltype | LK_RETRY); vfs_rel(mp); if (error != 0) return (ENOENT); if (vp->v_iflag & VI_DOOMED) { vfs_unbusy(mp); return (ENOENT); } } VOP_UNLOCK(vp, 0); error = alloc(mp, alloc_arg, lkflags, rvp); vfs_unbusy(mp); if (*rvp != vp) vn_lock(vp, ltype | LK_RETRY); if (vp->v_iflag & VI_DOOMED) { if (error == 0) { if (*rvp == vp) vunref(vp); else vput(*rvp); } error = ENOENT; } return (error); } int vn_rlimit_fsize(const struct vnode *vp, const struct uio *uio, struct thread *td) { if (vp->v_type != VREG || td == NULL) return (0); if ((uoff_t)uio->uio_offset + uio->uio_resid > lim_cur(td, RLIMIT_FSIZE)) { PROC_LOCK(td->td_proc); kern_psignal(td->td_proc, SIGXFSZ); PROC_UNLOCK(td->td_proc); return (EFBIG); } return (0); } int vn_chmod(struct file *fp, mode_t mode, struct ucred *active_cred, struct thread *td) { struct vnode *vp; vp = fp->f_vnode; #ifdef AUDIT vn_lock(vp, LK_SHARED | LK_RETRY); AUDIT_ARG_VNODE1(vp); VOP_UNLOCK(vp, 0); #endif return (setfmode(td, active_cred, vp, mode)); } int vn_chown(struct file *fp, uid_t uid, gid_t gid, struct ucred *active_cred, struct thread *td) { struct vnode *vp; vp = fp->f_vnode; #ifdef AUDIT vn_lock(vp, LK_SHARED | LK_RETRY); AUDIT_ARG_VNODE1(vp); VOP_UNLOCK(vp, 0); #endif return (setfown(td, active_cred, vp, uid, gid)); } void vn_pages_remove(struct vnode *vp, vm_pindex_t start, vm_pindex_t end) { vm_object_t object; if ((object = vp->v_object) == NULL) return; VM_OBJECT_WLOCK(object); vm_object_page_remove(object, start, end, 0); VM_OBJECT_WUNLOCK(object); } int vn_bmap_seekhole(struct vnode *vp, u_long cmd, off_t *off, struct ucred *cred) { struct vattr va; daddr_t bn, bnp; uint64_t bsize; off_t noff; int error; KASSERT(cmd == FIOSEEKHOLE || cmd == FIOSEEKDATA, ("Wrong command %lu", cmd)); if (vn_lock(vp, LK_SHARED) != 0) return (EBADF); if (vp->v_type != VREG) { error = ENOTTY; goto unlock; } error = VOP_GETATTR(vp, &va, cred); if (error != 0) goto unlock; noff = *off; if (noff >= va.va_size) { error = ENXIO; goto unlock; } bsize = vp->v_mount->mnt_stat.f_iosize; for (bn = noff / bsize; noff < va.va_size; bn++, noff += bsize) { error = VOP_BMAP(vp, bn, NULL, &bnp, NULL, NULL); if (error == EOPNOTSUPP) { error = ENOTTY; goto unlock; } if ((bnp == -1 && cmd == FIOSEEKHOLE) || (bnp != -1 && cmd == FIOSEEKDATA)) { noff = bn * bsize; if (noff < *off) noff = *off; goto unlock; } } if (noff > va.va_size) noff = va.va_size; /* noff == va.va_size. There is an implicit hole at the end of file. */ if (cmd == FIOSEEKDATA) error = ENXIO; unlock: VOP_UNLOCK(vp, 0); if (error == 0) *off = noff; return (error); } int vn_seek(struct file *fp, off_t offset, int whence, struct thread *td) { struct ucred *cred; struct vnode *vp; struct vattr vattr; off_t foffset, size; int error, noneg; cred = td->td_ucred; vp = fp->f_vnode; foffset = foffset_lock(fp, 0); noneg = (vp->v_type != VCHR); error = 0; switch (whence) { case L_INCR: if (noneg && (foffset < 0 || (offset > 0 && foffset > OFF_MAX - offset))) { error = EOVERFLOW; break; } offset += foffset; break; case L_XTND: vn_lock(vp, LK_SHARED | LK_RETRY); error = VOP_GETATTR(vp, &vattr, cred); VOP_UNLOCK(vp, 0); if (error) break; /* * If the file references a disk device, then fetch * the media size and use that to determine the ending * offset. */ if (vattr.va_size == 0 && vp->v_type == VCHR && fo_ioctl(fp, DIOCGMEDIASIZE, &size, cred, td) == 0) vattr.va_size = size; if (noneg && (vattr.va_size > OFF_MAX || (offset > 0 && vattr.va_size > OFF_MAX - offset))) { error = EOVERFLOW; break; } offset += vattr.va_size; break; case L_SET: break; case SEEK_DATA: error = fo_ioctl(fp, FIOSEEKDATA, &offset, cred, td); break; case SEEK_HOLE: error = fo_ioctl(fp, FIOSEEKHOLE, &offset, cred, td); break; default: error = EINVAL; } if (error == 0 && noneg && offset < 0) error = EINVAL; if (error != 0) goto drop; VFS_KNOTE_UNLOCKED(vp, 0); td->td_uretoff.tdu_off = offset; drop: foffset_unlock(fp, offset, error != 0 ? FOF_NOUPDATE : 0); return (error); } int vn_utimes_perm(struct vnode *vp, struct vattr *vap, struct ucred *cred, struct thread *td) { int error; /* * Grant permission if the caller is the owner of the file, or * the super-user, or has ACL_WRITE_ATTRIBUTES permission on * on the file. If the time pointer is null, then write * permission on the file is also sufficient. * * From NFSv4.1, draft 21, 6.2.1.3.1, Discussion of Mask Attributes: * A user having ACL_WRITE_DATA or ACL_WRITE_ATTRIBUTES * will be allowed to set the times [..] to the current * server time. */ error = VOP_ACCESSX(vp, VWRITE_ATTRIBUTES, cred, td); if (error != 0 && (vap->va_vaflags & VA_UTIMES_NULL) != 0) error = VOP_ACCESS(vp, VWRITE, cred, td); return (error); } int vn_fill_kinfo(struct file *fp, struct kinfo_file *kif, struct filedesc *fdp) { struct vnode *vp; int error; if (fp->f_type == DTYPE_FIFO) kif->kf_type = KF_TYPE_FIFO; else kif->kf_type = KF_TYPE_VNODE; vp = fp->f_vnode; vref(vp); FILEDESC_SUNLOCK(fdp); error = vn_fill_kinfo_vnode(vp, kif); vrele(vp); FILEDESC_SLOCK(fdp); return (error); } static inline void vn_fill_junk(struct kinfo_file *kif) { size_t len, olen; /* * Simulate vn_fullpath returning changing values for a given * vp during e.g. coredump. */ len = (arc4random() % (sizeof(kif->kf_path) - 2)) + 1; olen = strlen(kif->kf_path); if (len < olen) strcpy(&kif->kf_path[len - 1], "$"); else for (; olen < len; olen++) strcpy(&kif->kf_path[olen], "A"); } int vn_fill_kinfo_vnode(struct vnode *vp, struct kinfo_file *kif) { struct vattr va; char *fullpath, *freepath; int error; kif->kf_un.kf_file.kf_file_type = vntype_to_kinfo(vp->v_type); freepath = NULL; fullpath = "-"; error = vn_fullpath(curthread, vp, &fullpath, &freepath); if (error == 0) { strlcpy(kif->kf_path, fullpath, sizeof(kif->kf_path)); } if (freepath != NULL) free(freepath, M_TEMP); KFAIL_POINT_CODE(DEBUG_FP, fill_kinfo_vnode__random_path, vn_fill_junk(kif); ); /* * Retrieve vnode attributes. */ va.va_fsid = VNOVAL; va.va_rdev = NODEV; vn_lock(vp, LK_SHARED | LK_RETRY); error = VOP_GETATTR(vp, &va, curthread->td_ucred); VOP_UNLOCK(vp, 0); if (error != 0) return (error); if (va.va_fsid != VNOVAL) kif->kf_un.kf_file.kf_file_fsid = va.va_fsid; else kif->kf_un.kf_file.kf_file_fsid = vp->v_mount->mnt_stat.f_fsid.val[0]; kif->kf_un.kf_file.kf_file_fsid_freebsd11 = kif->kf_un.kf_file.kf_file_fsid; /* truncate */ kif->kf_un.kf_file.kf_file_fileid = va.va_fileid; kif->kf_un.kf_file.kf_file_mode = MAKEIMODE(va.va_type, va.va_mode); kif->kf_un.kf_file.kf_file_size = va.va_size; kif->kf_un.kf_file.kf_file_rdev = va.va_rdev; kif->kf_un.kf_file.kf_file_rdev_freebsd11 = kif->kf_un.kf_file.kf_file_rdev; /* truncate */ return (0); } int vn_mmap(struct file *fp, vm_map_t map, vm_offset_t *addr, vm_size_t size, vm_prot_t prot, vm_prot_t cap_maxprot, int flags, vm_ooffset_t foff, struct thread *td) { #ifdef HWPMC_HOOKS struct pmckern_map_in pkm; #endif struct mount *mp; struct vnode *vp; vm_object_t object; vm_prot_t maxprot; boolean_t writecounted; int error; #if defined(COMPAT_FREEBSD7) || defined(COMPAT_FREEBSD6) || \ defined(COMPAT_FREEBSD5) || defined(COMPAT_FREEBSD4) /* * POSIX shared-memory objects are defined to have * kernel persistence, and are not defined to support * read(2)/write(2) -- or even open(2). Thus, we can * use MAP_ASYNC to trade on-disk coherence for speed. * The shm_open(3) library routine turns on the FPOSIXSHM * flag to request this behavior. */ if ((fp->f_flag & FPOSIXSHM) != 0) flags |= MAP_NOSYNC; #endif vp = fp->f_vnode; /* * Ensure that file and memory protections are * compatible. Note that we only worry about * writability if mapping is shared; in this case, * current and max prot are dictated by the open file. * XXX use the vnode instead? Problem is: what * credentials do we use for determination? What if * proc does a setuid? */ mp = vp->v_mount; if (mp != NULL && (mp->mnt_flag & MNT_NOEXEC) != 0) { maxprot = VM_PROT_NONE; if ((prot & VM_PROT_EXECUTE) != 0) return (EACCES); } else maxprot = VM_PROT_EXECUTE; if ((fp->f_flag & FREAD) != 0) maxprot |= VM_PROT_READ; else if ((prot & VM_PROT_READ) != 0) return (EACCES); /* * If we are sharing potential changes via MAP_SHARED and we * are trying to get write permission although we opened it * without asking for it, bail out. */ if ((flags & MAP_SHARED) != 0) { if ((fp->f_flag & FWRITE) != 0) maxprot |= VM_PROT_WRITE; else if ((prot & VM_PROT_WRITE) != 0) return (EACCES); } else { maxprot |= VM_PROT_WRITE; cap_maxprot |= VM_PROT_WRITE; } maxprot &= cap_maxprot; /* * For regular files and shared memory, POSIX requires that * the value of foff be a legitimate offset within the data * object. In particular, negative offsets are invalid. * Blocking negative offsets and overflows here avoids * possible wraparound or user-level access into reserved * ranges of the data object later. In contrast, POSIX does * not dictate how offsets are used by device drivers, so in * the case of a device mapping a negative offset is passed * on. */ if ( #ifdef _LP64 size > OFF_MAX || #endif foff < 0 || foff > OFF_MAX - size) return (EINVAL); writecounted = FALSE; error = vm_mmap_vnode(td, size, prot, &maxprot, &flags, vp, &foff, &object, &writecounted); if (error != 0) return (error); error = vm_mmap_object(map, addr, size, prot, maxprot, flags, object, foff, writecounted, td); if (error != 0) { /* * If this mapping was accounted for in the vnode's * writecount, then undo that now. */ if (writecounted) vnode_pager_release_writecount(object, 0, size); vm_object_deallocate(object); } #ifdef HWPMC_HOOKS /* Inform hwpmc(4) if an executable is being mapped. */ if (PMC_HOOK_INSTALLED(PMC_FN_MMAP)) { if ((prot & VM_PROT_EXECUTE) != 0 && error == 0) { pkm.pm_file = vp; pkm.pm_address = (uintptr_t) *addr; - PMC_CALL_HOOK(td, PMC_FN_MMAP, (void *) &pkm); + PMC_CALL_HOOK_UNLOCKED(td, PMC_FN_MMAP, (void *) &pkm); } } #endif return (error); } void vn_fsid(struct vnode *vp, struct vattr *va) { fsid_t *f; f = &vp->v_mount->mnt_stat.f_fsid; va->va_fsid = (uint32_t)f->val[1]; va->va_fsid <<= sizeof(f->val[1]) * NBBY; va->va_fsid += (uint32_t)f->val[0]; } Index: head/sys/sys/pmckern.h =================================================================== --- head/sys/sys/pmckern.h (revision 334343) +++ head/sys/sys/pmckern.h (revision 334344) @@ -1,260 +1,261 @@ /*- * SPDX-License-Identifier: BSD-2-Clause-FreeBSD * * Copyright (c) 2003-2007, Joseph Koshy * Copyright (c) 2007 The FreeBSD Foundation * All rights reserved. * * Portions of this software were developed by A. Joseph Koshy under * sponsorship from the FreeBSD Foundation and Google, Inc. * * Redistribution and use in source and binary forms, with or without * modification, are permitted provided that the following conditions * are met: * 1. Redistributions of source code must retain the above copyright * notice, this list of conditions and the following disclaimer. * 2. Redistributions in binary form must reproduce the above copyright * notice, this list of conditions and the following disclaimer in the * documentation and/or other materials provided with the distribution. * * 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. * * $FreeBSD$ */ /* * PMC interface used by the base kernel. */ #ifndef _SYS_PMCKERN_H_ #define _SYS_PMCKERN_H_ #include #include #include #include #include #include #include #include #define PMC_FN_PROCESS_EXEC 1 #define PMC_FN_CSW_IN 2 #define PMC_FN_CSW_OUT 3 #define PMC_FN_DO_SAMPLES 4 #define PMC_FN_UNUSED1 5 #define PMC_FN_UNUSED2 6 #define PMC_FN_MMAP 7 #define PMC_FN_MUNMAP 8 #define PMC_FN_USER_CALLCHAIN 9 #define PMC_FN_USER_CALLCHAIN_SOFT 10 #define PMC_FN_SOFT_SAMPLING 11 #define PMC_FN_THR_CREATE 12 #define PMC_FN_THR_EXIT 13 #define PMC_HR 0 /* Hardware ring buffer */ #define PMC_SR 1 /* Software ring buffer */ struct pmckern_procexec { int pm_credentialschanged; uintfptr_t pm_entryaddr; }; struct pmckern_map_in { void *pm_file; /* filename or vnode pointer */ uintfptr_t pm_address; /* address object is loaded at */ }; struct pmckern_map_out { uintfptr_t pm_address; /* start address of region */ size_t pm_size; /* size of unmapped region */ }; struct pmckern_soft { enum pmc_event pm_ev; int pm_cpu; struct trapframe *pm_tf; }; /* * Soft PMC. */ #define PMC_SOFT_DEFINE_EX(prov, mod, func, name, alloc, release) \ struct pmc_soft pmc_##prov##_##mod##_##func##_##name = \ { 0, alloc, release, { #prov "_" #mod "_" #func "." #name, 0 } }; \ SYSINIT(pmc_##prov##_##mod##_##func##_##name##_init, SI_SUB_KDTRACE, \ SI_ORDER_SECOND + 1, pmc_soft_ev_register, \ &pmc_##prov##_##mod##_##func##_##name ); \ SYSUNINIT(pmc_##prov##_##mod##_##func##_##name##_uninit, \ SI_SUB_KDTRACE, SI_ORDER_SECOND + 1, pmc_soft_ev_deregister, \ &pmc_##prov##_##mod##_##func##_##name ) #define PMC_SOFT_DEFINE(prov, mod, func, name) \ PMC_SOFT_DEFINE_EX(prov, mod, func, name, NULL, NULL) #define PMC_SOFT_DECLARE(prov, mod, func, name) \ extern struct pmc_soft pmc_##prov##_##mod##_##func##_##name /* * PMC_SOFT_CALL can be used anywhere in the kernel. * Require md defined PMC_FAKE_TRAPFRAME. */ #ifdef PMC_FAKE_TRAPFRAME #define PMC_SOFT_CALL(pr, mo, fu, na) \ do { \ if (__predict_false(pmc_##pr##_##mo##_##fu##_##na.ps_running)) { \ struct pmckern_soft ks; \ register_t intr; \ intr = intr_disable(); \ PMC_FAKE_TRAPFRAME(&pmc_tf[curcpu]); \ ks.pm_ev = pmc_##pr##_##mo##_##fu##_##na.ps_ev.pm_ev_code; \ ks.pm_cpu = PCPU_GET(cpuid); \ ks.pm_tf = &pmc_tf[curcpu]; \ PMC_CALL_HOOK_UNLOCKED(curthread, \ PMC_FN_SOFT_SAMPLING, (void *) &ks); \ intr_restore(intr); \ } \ } while (0) #else #define PMC_SOFT_CALL(pr, mo, fu, na) \ do { \ } while (0) #endif /* * PMC_SOFT_CALL_TF need to be used carefully. * Userland capture will be done during AST processing. */ #define PMC_SOFT_CALL_TF(pr, mo, fu, na, tf) \ do { \ if (__predict_false(pmc_##pr##_##mo##_##fu##_##na.ps_running)) { \ struct pmckern_soft ks; \ register_t intr; \ intr = intr_disable(); \ ks.pm_ev = pmc_##pr##_##mo##_##fu##_##na.ps_ev.pm_ev_code; \ ks.pm_cpu = PCPU_GET(cpuid); \ ks.pm_tf = tf; \ PMC_CALL_HOOK_UNLOCKED(curthread, \ PMC_FN_SOFT_SAMPLING, (void *) &ks); \ intr_restore(intr); \ } \ } while (0) struct pmc_soft { int ps_running; void (*ps_alloc)(void); void (*ps_release)(void); struct pmc_dyn_event_descr ps_ev; }; struct pmclog_buffer; struct pmc_domain_buffer_header { struct mtx pdbh_mtx; TAILQ_HEAD(, pmclog_buffer) pdbh_head; struct pmclog_buffer *pdbh_plbs; int pdbh_ncpus; } __aligned(CACHE_LINE_SIZE); /* hook */ extern int (*pmc_hook)(struct thread *_td, int _function, void *_arg); extern int (*pmc_intr)(int _cpu, struct trapframe *_frame); /* SX lock protecting the hook */ extern struct sx pmc_sx; /* Per-cpu flags indicating availability of sampling data */ DPCPU_DECLARE(uint8_t, pmc_sampled); /* Count of system-wide sampling PMCs in existence */ extern volatile int pmc_ss_count; /* kernel version number */ extern const int pmc_kernel_version; /* PMC soft per cpu trapframe */ extern struct trapframe pmc_tf[MAXCPU]; /* per domain buffer header list */ extern struct pmc_domain_buffer_header *pmc_dom_hdrs[MAXMEMDOM]; /* Quick check if preparatory work is necessary */ #define PMC_HOOK_INSTALLED(cmd) __predict_false(pmc_hook != NULL) /* Hook invocation; for use within the kernel */ #define PMC_CALL_HOOK(t, cmd, arg) \ do { \ epoch_enter_preempt(global_epoch_preempt); \ if (pmc_hook != NULL) \ (pmc_hook)((t), (cmd), (arg)); \ epoch_exit_preempt(global_epoch_preempt); \ } while (0) /* Hook invocation that needs an exclusive lock */ #define PMC_CALL_HOOK_X(t, cmd, arg) \ do { \ sx_xlock(&pmc_sx); \ if (pmc_hook != NULL) \ (pmc_hook)((t), (cmd), (arg)); \ sx_xunlock(&pmc_sx); \ } while (0) /* * Some hook invocations (e.g., from context switch and clock handling * code) need to be lock-free. */ #define PMC_CALL_HOOK_UNLOCKED(t, cmd, arg) \ do { \ - if (pmc_hook != NULL) \ + MPASS(!in_epoch()); \ + if (pmc_hook != NULL) \ (pmc_hook)((t), (cmd), (arg)); \ } while (0) #define PMC_SWITCH_CONTEXT(t,cmd) PMC_CALL_HOOK_UNLOCKED(t,cmd,NULL) /* Check if a process is using HWPMCs.*/ #define PMC_PROC_IS_USING_PMCS(p) \ (__predict_false(p->p_flag & P_HWPMC)) /* Check if a thread have pending user capture. */ #define PMC_IS_PENDING_CALLCHAIN(p) \ (__predict_false((p)->td_pflags & TDP_CALLCHAIN)) #define PMC_SYSTEM_SAMPLING_ACTIVE() (pmc_ss_count > 0) /* Check if a CPU has recorded samples. */ #define PMC_CPU_HAS_SAMPLES(C) (__predict_false(DPCPU_ID_GET((C), pmc_sampled))) /* * Helper functions. */ int pmc_cpu_is_disabled(int _cpu); /* deprecated */ int pmc_cpu_is_active(int _cpu); int pmc_cpu_is_present(int _cpu); int pmc_cpu_is_primary(int _cpu); unsigned int pmc_cpu_max(void); #ifdef INVARIANTS int pmc_cpu_max_active(void); #endif /* * Soft events functions. */ void pmc_soft_ev_register(struct pmc_soft *ps); void pmc_soft_ev_deregister(struct pmc_soft *ps); struct pmc_soft *pmc_soft_ev_acquire(enum pmc_event ev); void pmc_soft_ev_release(struct pmc_soft *ps); #endif /* _SYS_PMCKERN_H_ */