2 * Copyright (c) 2003-2008 Joseph Koshy
3 * Copyright (c) 2007 The FreeBSD Foundation
6 * Portions of this software were developed by A. Joseph Koshy under
7 * sponsorship from the FreeBSD Foundation and Google, Inc.
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
12 * 1. Redistributions of source code must retain the above copyright
13 * notice, this list of conditions and the following disclaimer.
14 * 2. Redistributions in binary form must reproduce the above copyright
15 * notice, this list of conditions and the following disclaimer in the
16 * documentation and/or other materials provided with the distribution.
18 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
19 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
22 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
35 #include <sys/param.h>
36 #include <sys/eventhandler.h>
38 #include <sys/kernel.h>
39 #include <sys/kthread.h>
40 #include <sys/limits.h>
42 #include <sys/malloc.h>
43 #include <sys/module.h>
44 #include <sys/mount.h>
45 #include <sys/mutex.h>
47 #include <sys/pmckern.h>
48 #include <sys/pmclog.h>
51 #include <sys/queue.h>
52 #include <sys/resourcevar.h>
53 #include <sys/rwlock.h>
54 #include <sys/sched.h>
55 #include <sys/signalvar.h>
58 #include <sys/sysctl.h>
59 #include <sys/sysent.h>
60 #include <sys/systm.h>
61 #include <sys/vnode.h>
63 #include <sys/linker.h> /* needs to be after <sys/malloc.h> */
65 #include <machine/atomic.h>
66 #include <machine/md_var.h>
69 #include <vm/vm_extern.h>
71 #include <vm/vm_map.h>
72 #include <vm/vm_object.h>
74 #include "hwpmc_soft.h"
81 PMC_FLAG_NONE = 0x00, /* do nothing */
82 PMC_FLAG_REMOVE = 0x01, /* atomically remove entry from hash */
83 PMC_FLAG_ALLOCATE = 0x02, /* add entry to hash if not found */
87 * The offset in sysent where the syscall is allocated.
90 static int pmc_syscall_num = NO_SYSCALL;
91 struct pmc_cpu **pmc_pcpu; /* per-cpu state */
92 pmc_value_t *pmc_pcpu_saved; /* saved PMC values: CSW handling */
94 #define PMC_PCPU_SAVED(C,R) pmc_pcpu_saved[(R) + md->pmd_npmc*(C)]
96 struct mtx_pool *pmc_mtxpool;
97 static int *pmc_pmcdisp; /* PMC row dispositions */
99 #define PMC_ROW_DISP_IS_FREE(R) (pmc_pmcdisp[(R)] == 0)
100 #define PMC_ROW_DISP_IS_THREAD(R) (pmc_pmcdisp[(R)] > 0)
101 #define PMC_ROW_DISP_IS_STANDALONE(R) (pmc_pmcdisp[(R)] < 0)
103 #define PMC_MARK_ROW_FREE(R) do { \
104 pmc_pmcdisp[(R)] = 0; \
107 #define PMC_MARK_ROW_STANDALONE(R) do { \
108 KASSERT(pmc_pmcdisp[(R)] <= 0, ("[pmc,%d] row disposition error", \
110 atomic_add_int(&pmc_pmcdisp[(R)], -1); \
111 KASSERT(pmc_pmcdisp[(R)] >= (-pmc_cpu_max_active()), \
112 ("[pmc,%d] row disposition error", __LINE__)); \
115 #define PMC_UNMARK_ROW_STANDALONE(R) do { \
116 atomic_add_int(&pmc_pmcdisp[(R)], 1); \
117 KASSERT(pmc_pmcdisp[(R)] <= 0, ("[pmc,%d] row disposition error", \
121 #define PMC_MARK_ROW_THREAD(R) do { \
122 KASSERT(pmc_pmcdisp[(R)] >= 0, ("[pmc,%d] row disposition error", \
124 atomic_add_int(&pmc_pmcdisp[(R)], 1); \
127 #define PMC_UNMARK_ROW_THREAD(R) do { \
128 atomic_add_int(&pmc_pmcdisp[(R)], -1); \
129 KASSERT(pmc_pmcdisp[(R)] >= 0, ("[pmc,%d] row disposition error", \
134 /* various event handlers */
135 static eventhandler_tag pmc_exit_tag, pmc_fork_tag, pmc_kld_load_tag,
138 /* Module statistics */
139 struct pmc_op_getdriverstats pmc_stats;
141 /* Machine/processor dependent operations */
142 static struct pmc_mdep *md;
145 * Hash tables mapping owner processes and target threads to PMCs.
148 struct mtx pmc_processhash_mtx; /* spin mutex */
149 static u_long pmc_processhashmask;
150 static LIST_HEAD(pmc_processhash, pmc_process) *pmc_processhash;
153 * Hash table of PMC owner descriptors. This table is protected by
154 * the shared PMC "sx" lock.
157 static u_long pmc_ownerhashmask;
158 static LIST_HEAD(pmc_ownerhash, pmc_owner) *pmc_ownerhash;
161 * List of PMC owners with system-wide sampling PMCs.
164 static LIST_HEAD(, pmc_owner) pmc_ss_owners;
168 * A map of row indices to classdep structures.
170 static struct pmc_classdep **pmc_rowindex_to_classdep;
177 static int pmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS);
178 static int pmc_debugflags_parse(char *newstr, char *fence);
181 static int load(struct module *module, int cmd, void *arg);
182 static int pmc_attach_process(struct proc *p, struct pmc *pm);
183 static struct pmc *pmc_allocate_pmc_descriptor(void);
184 static struct pmc_owner *pmc_allocate_owner_descriptor(struct proc *p);
185 static int pmc_attach_one_process(struct proc *p, struct pmc *pm);
186 static int pmc_can_allocate_rowindex(struct proc *p, unsigned int ri,
188 static int pmc_can_attach(struct pmc *pm, struct proc *p);
189 static void pmc_capture_user_callchain(int cpu, int soft, struct trapframe *tf);
190 static void pmc_cleanup(void);
191 static int pmc_detach_process(struct proc *p, struct pmc *pm);
192 static int pmc_detach_one_process(struct proc *p, struct pmc *pm,
194 static void pmc_destroy_owner_descriptor(struct pmc_owner *po);
195 static void pmc_destroy_pmc_descriptor(struct pmc *pm);
196 static struct pmc_owner *pmc_find_owner_descriptor(struct proc *p);
197 static int pmc_find_pmc(pmc_id_t pmcid, struct pmc **pm);
198 static struct pmc *pmc_find_pmc_descriptor_in_process(struct pmc_owner *po,
200 static struct pmc_process *pmc_find_process_descriptor(struct proc *p,
202 static void pmc_force_context_switch(void);
203 static void pmc_link_target_process(struct pmc *pm,
204 struct pmc_process *pp);
205 static void pmc_log_all_process_mappings(struct pmc_owner *po);
206 static void pmc_log_kernel_mappings(struct pmc *pm);
207 static void pmc_log_process_mappings(struct pmc_owner *po, struct proc *p);
208 static void pmc_maybe_remove_owner(struct pmc_owner *po);
209 static void pmc_process_csw_in(struct thread *td);
210 static void pmc_process_csw_out(struct thread *td);
211 static void pmc_process_exit(void *arg, struct proc *p);
212 static void pmc_process_fork(void *arg, struct proc *p1,
213 struct proc *p2, int n);
214 static void pmc_process_samples(int cpu, int soft);
215 static void pmc_release_pmc_descriptor(struct pmc *pmc);
216 static void pmc_remove_owner(struct pmc_owner *po);
217 static void pmc_remove_process_descriptor(struct pmc_process *pp);
218 static void pmc_restore_cpu_binding(struct pmc_binding *pb);
219 static void pmc_save_cpu_binding(struct pmc_binding *pb);
220 static void pmc_select_cpu(int cpu);
221 static int pmc_start(struct pmc *pm);
222 static int pmc_stop(struct pmc *pm);
223 static int pmc_syscall_handler(struct thread *td, void *syscall_args);
224 static void pmc_unlink_target_process(struct pmc *pmc,
225 struct pmc_process *pp);
226 static int generic_switch_in(struct pmc_cpu *pc, struct pmc_process *pp);
227 static int generic_switch_out(struct pmc_cpu *pc, struct pmc_process *pp);
228 static struct pmc_mdep *pmc_generic_cpu_initialize(void);
229 static void pmc_generic_cpu_finalize(struct pmc_mdep *md);
232 * Kernel tunables and sysctl(8) interface.
235 SYSCTL_DECL(_kern_hwpmc);
237 static int pmc_callchaindepth = PMC_CALLCHAIN_DEPTH;
238 SYSCTL_INT(_kern_hwpmc, OID_AUTO, callchaindepth, CTLFLAG_RDTUN,
239 &pmc_callchaindepth, 0, "depth of call chain records");
242 struct pmc_debugflags pmc_debugflags = PMC_DEBUG_DEFAULT_FLAGS;
243 char pmc_debugstr[PMC_DEBUG_STRSIZE];
244 TUNABLE_STR(PMC_SYSCTL_NAME_PREFIX "debugflags", pmc_debugstr,
245 sizeof(pmc_debugstr));
246 SYSCTL_PROC(_kern_hwpmc, OID_AUTO, debugflags,
247 CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NOFETCH,
248 0, 0, pmc_debugflags_sysctl_handler, "A", "debug flags");
252 * kern.hwpmc.hashrows -- determines the number of rows in the
253 * of the hash table used to look up threads
256 static int pmc_hashsize = PMC_HASH_SIZE;
257 SYSCTL_INT(_kern_hwpmc, OID_AUTO, hashsize, CTLFLAG_RDTUN,
258 &pmc_hashsize, 0, "rows in hash tables");
261 * kern.hwpmc.nsamples --- number of PC samples/callchain stacks per CPU
264 static int pmc_nsamples = PMC_NSAMPLES;
265 SYSCTL_INT(_kern_hwpmc, OID_AUTO, nsamples, CTLFLAG_RDTUN,
266 &pmc_nsamples, 0, "number of PC samples per CPU");
270 * kern.hwpmc.mtxpoolsize -- number of mutexes in the mutex pool.
273 static int pmc_mtxpool_size = PMC_MTXPOOL_SIZE;
274 SYSCTL_INT(_kern_hwpmc, OID_AUTO, mtxpoolsize, CTLFLAG_RDTUN,
275 &pmc_mtxpool_size, 0, "size of spin mutex pool");
279 * security.bsd.unprivileged_syspmcs -- allow non-root processes to
280 * allocate system-wide PMCs.
282 * Allowing unprivileged processes to allocate system PMCs is convenient
283 * if system-wide measurements need to be taken concurrently with other
284 * per-process measurements. This feature is turned off by default.
287 static int pmc_unprivileged_syspmcs = 0;
288 SYSCTL_INT(_security_bsd, OID_AUTO, unprivileged_syspmcs, CTLFLAG_RWTUN,
289 &pmc_unprivileged_syspmcs, 0,
290 "allow unprivileged process to allocate system PMCs");
293 * Hash function. Discard the lower 2 bits of the pointer since
294 * these are always zero for our uses. The hash multiplier is
295 * round((2^LONG_BIT) * ((sqrt(5)-1)/2)).
299 #define _PMC_HM 11400714819323198486u
301 #define _PMC_HM 2654435769u
303 #error Must know the size of 'long' to compile
306 #define PMC_HASH_PTR(P,M) ((((unsigned long) (P) >> 2) * _PMC_HM) & (M))
312 /* The `sysent' for the new syscall */
313 static struct sysent pmc_sysent = {
315 pmc_syscall_handler /* sy_call */
318 static struct syscall_module_data pmc_syscall_mod = {
323 #if (__FreeBSD_version >= 1100000)
331 static moduledata_t pmc_mod = {
333 syscall_module_handler,
337 #ifdef EARLY_AP_STARTUP
338 DECLARE_MODULE(pmc, pmc_mod, SI_SUB_SYSCALLS, SI_ORDER_ANY);
340 DECLARE_MODULE(pmc, pmc_mod, SI_SUB_SMP, SI_ORDER_ANY);
342 MODULE_VERSION(pmc, PMC_VERSION);
345 enum pmc_dbgparse_state {
346 PMCDS_WS, /* in whitespace */
347 PMCDS_MAJOR, /* seen a major keyword */
352 pmc_debugflags_parse(char *newstr, char *fence)
355 struct pmc_debugflags *tmpflags;
356 int error, found, *newbits, tmp;
359 tmpflags = malloc(sizeof(*tmpflags), M_PMC, M_WAITOK|M_ZERO);
364 for (; p < fence && (c = *p); p++) {
366 /* skip white space */
367 if (c == ' ' || c == '\t')
370 /* look for a keyword followed by "=" */
371 for (q = p; p < fence && (c = *p) && c != '='; p++)
381 /* lookup flag group name */
382 #define DBG_SET_FLAG_MAJ(S,F) \
383 if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0) \
384 newbits = &tmpflags->pdb_ ## F;
386 DBG_SET_FLAG_MAJ("cpu", CPU);
387 DBG_SET_FLAG_MAJ("csw", CSW);
388 DBG_SET_FLAG_MAJ("logging", LOG);
389 DBG_SET_FLAG_MAJ("module", MOD);
390 DBG_SET_FLAG_MAJ("md", MDP);
391 DBG_SET_FLAG_MAJ("owner", OWN);
392 DBG_SET_FLAG_MAJ("pmc", PMC);
393 DBG_SET_FLAG_MAJ("process", PRC);
394 DBG_SET_FLAG_MAJ("sampling", SAM);
396 if (newbits == NULL) {
401 p++; /* skip the '=' */
403 /* Now parse the individual flags */
406 for (q = p; p < fence && (c = *p); p++)
407 if (c == ' ' || c == '\t' || c == ',')
410 /* p == fence or c == ws or c == "," or c == 0 */
412 if ((kwlen = p - q) == 0) {
418 #define DBG_SET_FLAG_MIN(S,F) \
419 if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0) \
420 tmp |= found = (1 << PMC_DEBUG_MIN_ ## F)
422 /* a '*' denotes all possible flags in the group */
423 if (kwlen == 1 && *q == '*')
425 /* look for individual flag names */
426 DBG_SET_FLAG_MIN("allocaterow", ALR);
427 DBG_SET_FLAG_MIN("allocate", ALL);
428 DBG_SET_FLAG_MIN("attach", ATT);
429 DBG_SET_FLAG_MIN("bind", BND);
430 DBG_SET_FLAG_MIN("config", CFG);
431 DBG_SET_FLAG_MIN("exec", EXC);
432 DBG_SET_FLAG_MIN("exit", EXT);
433 DBG_SET_FLAG_MIN("find", FND);
434 DBG_SET_FLAG_MIN("flush", FLS);
435 DBG_SET_FLAG_MIN("fork", FRK);
436 DBG_SET_FLAG_MIN("getbuf", GTB);
437 DBG_SET_FLAG_MIN("hook", PMH);
438 DBG_SET_FLAG_MIN("init", INI);
439 DBG_SET_FLAG_MIN("intr", INT);
440 DBG_SET_FLAG_MIN("linktarget", TLK);
441 DBG_SET_FLAG_MIN("mayberemove", OMR);
442 DBG_SET_FLAG_MIN("ops", OPS);
443 DBG_SET_FLAG_MIN("read", REA);
444 DBG_SET_FLAG_MIN("register", REG);
445 DBG_SET_FLAG_MIN("release", REL);
446 DBG_SET_FLAG_MIN("remove", ORM);
447 DBG_SET_FLAG_MIN("sample", SAM);
448 DBG_SET_FLAG_MIN("scheduleio", SIO);
449 DBG_SET_FLAG_MIN("select", SEL);
450 DBG_SET_FLAG_MIN("signal", SIG);
451 DBG_SET_FLAG_MIN("swi", SWI);
452 DBG_SET_FLAG_MIN("swo", SWO);
453 DBG_SET_FLAG_MIN("start", STA);
454 DBG_SET_FLAG_MIN("stop", STO);
455 DBG_SET_FLAG_MIN("syscall", PMS);
456 DBG_SET_FLAG_MIN("unlinktarget", TUL);
457 DBG_SET_FLAG_MIN("write", WRI);
459 /* unrecognized flag name */
464 if (c == 0 || c == ' ' || c == '\t') { /* end of flag group */
473 /* save the new flag set */
474 bcopy(tmpflags, &pmc_debugflags, sizeof(pmc_debugflags));
477 free(tmpflags, M_PMC);
482 pmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS)
484 char *fence, *newstr;
488 (void) arg1; (void) arg2; /* unused parameters */
490 n = sizeof(pmc_debugstr);
491 newstr = malloc(n, M_PMC, M_WAITOK|M_ZERO);
492 (void) strlcpy(newstr, pmc_debugstr, n);
494 error = sysctl_handle_string(oidp, newstr, n, req);
496 /* if there is a new string, parse and copy it */
497 if (error == 0 && req->newptr != NULL) {
498 fence = newstr + (n < req->newlen ? n : req->newlen + 1);
499 if ((error = pmc_debugflags_parse(newstr, fence)) == 0)
500 (void) strlcpy(pmc_debugstr, newstr,
501 sizeof(pmc_debugstr));
511 * Map a row index to a classdep structure and return the adjusted row
512 * index for the PMC class index.
514 static struct pmc_classdep *
515 pmc_ri_to_classdep(struct pmc_mdep *md, int ri, int *adjri)
517 struct pmc_classdep *pcd;
521 KASSERT(ri >= 0 && ri < md->pmd_npmc,
522 ("[pmc,%d] illegal row-index %d", __LINE__, ri));
524 pcd = pmc_rowindex_to_classdep[ri];
527 ("[pmc,%d] ri %d null pcd", __LINE__, ri));
529 *adjri = ri - pcd->pcd_ri;
531 KASSERT(*adjri >= 0 && *adjri < pcd->pcd_num,
532 ("[pmc,%d] adjusted row-index %d", __LINE__, *adjri));
538 * Concurrency Control
540 * The driver manages the following data structures:
542 * - target process descriptors, one per target process
543 * - owner process descriptors (and attached lists), one per owner process
544 * - lookup hash tables for owner and target processes
545 * - PMC descriptors (and attached lists)
546 * - per-cpu hardware state
547 * - the 'hook' variable through which the kernel calls into
549 * - the machine hardware state (managed by the MD layer)
551 * These data structures are accessed from:
553 * - thread context-switch code
554 * - interrupt handlers (possibly on multiple cpus)
555 * - kernel threads on multiple cpus running on behalf of user
556 * processes doing system calls
557 * - this driver's private kernel threads
559 * = Locks and Locking strategy =
561 * The driver uses four locking strategies for its operation:
563 * - The global SX lock "pmc_sx" is used to protect internal
566 * Calls into the module by syscall() start with this lock being
567 * held in exclusive mode. Depending on the requested operation,
568 * the lock may be downgraded to 'shared' mode to allow more
569 * concurrent readers into the module. Calls into the module from
570 * other parts of the kernel acquire the lock in shared mode.
572 * This SX lock is held in exclusive mode for any operations that
573 * modify the linkages between the driver's internal data structures.
575 * The 'pmc_hook' function pointer is also protected by this lock.
576 * It is only examined with the sx lock held in exclusive mode. The
577 * kernel module is allowed to be unloaded only with the sx lock held
578 * in exclusive mode. In normal syscall handling, after acquiring the
579 * pmc_sx lock we first check that 'pmc_hook' is non-null before
580 * proceeding. This prevents races between the thread unloading the module
581 * and other threads seeking to use the module.
583 * - Lookups of target process structures and owner process structures
584 * cannot use the global "pmc_sx" SX lock because these lookups need
585 * to happen during context switches and in other critical sections
586 * where sleeping is not allowed. We protect these lookup tables
587 * with their own private spin-mutexes, "pmc_processhash_mtx" and
588 * "pmc_ownerhash_mtx".
590 * - Interrupt handlers work in a lock free manner. At interrupt
591 * time, handlers look at the PMC pointer (phw->phw_pmc) configured
592 * when the PMC was started. If this pointer is NULL, the interrupt
593 * is ignored after updating driver statistics. We ensure that this
594 * pointer is set (using an atomic operation if necessary) before the
595 * PMC hardware is started. Conversely, this pointer is unset atomically
596 * only after the PMC hardware is stopped.
598 * We ensure that everything needed for the operation of an
599 * interrupt handler is available without it needing to acquire any
600 * locks. We also ensure that a PMC's software state is destroyed only
601 * after the PMC is taken off hardware (on all CPUs).
603 * - Context-switch handling with process-private PMCs needs more
606 * A given process may be the target of multiple PMCs. For example,
607 * PMCATTACH and PMCDETACH may be requested by a process on one CPU
608 * while the target process is running on another. A PMC could also
609 * be getting released because its owner is exiting. We tackle
610 * these situations in the following manner:
612 * - each target process structure 'pmc_process' has an array
613 * of 'struct pmc *' pointers, one for each hardware PMC.
615 * - At context switch IN time, each "target" PMC in RUNNING state
616 * gets started on hardware and a pointer to each PMC is copied into
617 * the per-cpu phw array. The 'runcount' for the PMC is
620 * - At context switch OUT time, all process-virtual PMCs are stopped
621 * on hardware. The saved value is added to the PMCs value field
622 * only if the PMC is in a non-deleted state (the PMCs state could
623 * have changed during the current time slice).
625 * Note that since in-between a switch IN on a processor and a switch
626 * OUT, the PMC could have been released on another CPU. Therefore
627 * context switch OUT always looks at the hardware state to turn
628 * OFF PMCs and will update a PMC's saved value only if reachable
629 * from the target process record.
631 * - OP PMCRELEASE could be called on a PMC at any time (the PMC could
632 * be attached to many processes at the time of the call and could
633 * be active on multiple CPUs).
635 * We prevent further scheduling of the PMC by marking it as in
636 * state 'DELETED'. If the runcount of the PMC is non-zero then
637 * this PMC is currently running on a CPU somewhere. The thread
638 * doing the PMCRELEASE operation waits by repeatedly doing a
639 * pause() till the runcount comes to zero.
641 * The contents of a PMC descriptor (struct pmc) are protected using
642 * a spin-mutex. In order to save space, we use a mutex pool.
644 * In terms of lock types used by witness(4), we use:
645 * - Type "pmc-sx", used by the global SX lock.
646 * - Type "pmc-sleep", for sleep mutexes used by logger threads.
647 * - Type "pmc-per-proc", for protecting PMC owner descriptors.
648 * - Type "pmc-leaf", used for all other spin mutexes.
652 * save the cpu binding of the current kthread
656 pmc_save_cpu_binding(struct pmc_binding *pb)
658 PMCDBG0(CPU,BND,2, "save-cpu");
659 thread_lock(curthread);
660 pb->pb_bound = sched_is_bound(curthread);
661 pb->pb_cpu = curthread->td_oncpu;
662 thread_unlock(curthread);
663 PMCDBG1(CPU,BND,2, "save-cpu cpu=%d", pb->pb_cpu);
667 * restore the cpu binding of the current thread
671 pmc_restore_cpu_binding(struct pmc_binding *pb)
673 PMCDBG2(CPU,BND,2, "restore-cpu curcpu=%d restore=%d",
674 curthread->td_oncpu, pb->pb_cpu);
675 thread_lock(curthread);
677 sched_bind(curthread, pb->pb_cpu);
679 sched_unbind(curthread);
680 thread_unlock(curthread);
681 PMCDBG0(CPU,BND,2, "restore-cpu done");
685 * move execution over the specified cpu and bind it there.
689 pmc_select_cpu(int cpu)
691 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
692 ("[pmc,%d] bad cpu number %d", __LINE__, cpu));
694 /* Never move to an inactive CPU. */
695 KASSERT(pmc_cpu_is_active(cpu), ("[pmc,%d] selecting inactive "
696 "CPU %d", __LINE__, cpu));
698 PMCDBG1(CPU,SEL,2, "select-cpu cpu=%d", cpu);
699 thread_lock(curthread);
700 sched_bind(curthread, cpu);
701 thread_unlock(curthread);
703 KASSERT(curthread->td_oncpu == cpu,
704 ("[pmc,%d] CPU not bound [cpu=%d, curr=%d]", __LINE__,
705 cpu, curthread->td_oncpu));
707 PMCDBG1(CPU,SEL,2, "select-cpu cpu=%d ok", cpu);
711 * Force a context switch.
713 * We do this by pause'ing for 1 tick -- invoking mi_switch() is not
714 * guaranteed to force a context switch.
718 pmc_force_context_switch(void)
725 * Get the file name for an executable. This is a simple wrapper
726 * around vn_fullpath(9).
730 pmc_getfilename(struct vnode *v, char **fullpath, char **freepath)
733 *fullpath = "unknown";
735 vn_fullpath(curthread, v, fullpath, freepath);
739 * remove an process owning PMCs
743 pmc_remove_owner(struct pmc_owner *po)
745 struct pmc *pm, *tmp;
747 sx_assert(&pmc_sx, SX_XLOCKED);
749 PMCDBG1(OWN,ORM,1, "remove-owner po=%p", po);
751 /* Remove descriptor from the owner hash table */
752 LIST_REMOVE(po, po_next);
754 /* release all owned PMC descriptors */
755 LIST_FOREACH_SAFE(pm, &po->po_pmcs, pm_next, tmp) {
756 PMCDBG1(OWN,ORM,2, "pmc=%p", pm);
757 KASSERT(pm->pm_owner == po,
758 ("[pmc,%d] owner %p != po %p", __LINE__, pm->pm_owner, po));
760 pmc_release_pmc_descriptor(pm); /* will unlink from the list */
761 pmc_destroy_pmc_descriptor(pm);
764 KASSERT(po->po_sscount == 0,
765 ("[pmc,%d] SS count not zero", __LINE__));
766 KASSERT(LIST_EMPTY(&po->po_pmcs),
767 ("[pmc,%d] PMC list not empty", __LINE__));
769 /* de-configure the log file if present */
770 if (po->po_flags & PMC_PO_OWNS_LOGFILE)
771 pmclog_deconfigure_log(po);
775 * remove an owner process record if all conditions are met.
779 pmc_maybe_remove_owner(struct pmc_owner *po)
782 PMCDBG1(OWN,OMR,1, "maybe-remove-owner po=%p", po);
785 * Remove owner record if
786 * - this process does not own any PMCs
787 * - this process has not allocated a system-wide sampling buffer
790 if (LIST_EMPTY(&po->po_pmcs) &&
791 ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)) {
792 pmc_remove_owner(po);
793 pmc_destroy_owner_descriptor(po);
798 * Add an association between a target process and a PMC.
802 pmc_link_target_process(struct pmc *pm, struct pmc_process *pp)
805 struct pmc_target *pt;
807 sx_assert(&pmc_sx, SX_XLOCKED);
809 KASSERT(pm != NULL && pp != NULL,
810 ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp));
811 KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)),
812 ("[pmc,%d] Attaching a non-process-virtual pmc=%p to pid=%d",
813 __LINE__, pm, pp->pp_proc->p_pid));
814 KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= ((int) md->pmd_npmc - 1),
815 ("[pmc,%d] Illegal reference count %d for process record %p",
816 __LINE__, pp->pp_refcnt, (void *) pp));
818 ri = PMC_TO_ROWINDEX(pm);
820 PMCDBG3(PRC,TLK,1, "link-target pmc=%p ri=%d pmc-process=%p",
824 LIST_FOREACH(pt, &pm->pm_targets, pt_next)
825 if (pt->pt_process == pp)
826 KASSERT(0, ("[pmc,%d] pp %p already in pmc %p targets",
830 pt = malloc(sizeof(struct pmc_target), M_PMC, M_WAITOK|M_ZERO);
833 LIST_INSERT_HEAD(&pm->pm_targets, pt, pt_next);
835 atomic_store_rel_ptr((uintptr_t *)&pp->pp_pmcs[ri].pp_pmc,
838 if (pm->pm_owner->po_owner == pp->pp_proc)
839 pm->pm_flags |= PMC_F_ATTACHED_TO_OWNER;
842 * Initialize the per-process values at this row index.
844 pp->pp_pmcs[ri].pp_pmcval = PMC_TO_MODE(pm) == PMC_MODE_TS ?
845 pm->pm_sc.pm_reloadcount : 0;
852 * Removes the association between a target process and a PMC.
856 pmc_unlink_target_process(struct pmc *pm, struct pmc_process *pp)
860 struct pmc_target *ptgt;
862 sx_assert(&pmc_sx, SX_XLOCKED);
864 KASSERT(pm != NULL && pp != NULL,
865 ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp));
867 KASSERT(pp->pp_refcnt >= 1 && pp->pp_refcnt <= (int) md->pmd_npmc,
868 ("[pmc,%d] Illegal ref count %d on process record %p",
869 __LINE__, pp->pp_refcnt, (void *) pp));
871 ri = PMC_TO_ROWINDEX(pm);
873 PMCDBG3(PRC,TUL,1, "unlink-target pmc=%p ri=%d pmc-process=%p",
876 KASSERT(pp->pp_pmcs[ri].pp_pmc == pm,
877 ("[pmc,%d] PMC ri %d mismatch pmc %p pp->[ri] %p", __LINE__,
878 ri, pm, pp->pp_pmcs[ri].pp_pmc));
880 pp->pp_pmcs[ri].pp_pmc = NULL;
881 pp->pp_pmcs[ri].pp_pmcval = (pmc_value_t) 0;
883 /* Remove owner-specific flags */
884 if (pm->pm_owner->po_owner == pp->pp_proc) {
885 pp->pp_flags &= ~PMC_PP_ENABLE_MSR_ACCESS;
886 pm->pm_flags &= ~PMC_F_ATTACHED_TO_OWNER;
891 /* Remove the target process from the PMC structure */
892 LIST_FOREACH(ptgt, &pm->pm_targets, pt_next)
893 if (ptgt->pt_process == pp)
896 KASSERT(ptgt != NULL, ("[pmc,%d] process %p (pp: %p) not found "
897 "in pmc %p", __LINE__, pp->pp_proc, pp, pm));
899 LIST_REMOVE(ptgt, pt_next);
902 /* if the PMC now lacks targets, send the owner a SIGIO */
903 if (LIST_EMPTY(&pm->pm_targets)) {
904 p = pm->pm_owner->po_owner;
906 kern_psignal(p, SIGIO);
909 PMCDBG2(PRC,SIG,2, "signalling proc=%p signal=%d", p,
915 * Check if PMC 'pm' may be attached to target process 't'.
919 pmc_can_attach(struct pmc *pm, struct proc *t)
921 struct proc *o; /* pmc owner */
922 struct ucred *oc, *tc; /* owner, target credentials */
923 int decline_attach, i;
926 * A PMC's owner can always attach that PMC to itself.
929 if ((o = pm->pm_owner->po_owner) == t)
943 * The effective uid of the PMC owner should match at least one
944 * of the {effective,real,saved} uids of the target process.
947 decline_attach = oc->cr_uid != tc->cr_uid &&
948 oc->cr_uid != tc->cr_svuid &&
949 oc->cr_uid != tc->cr_ruid;
952 * Every one of the target's group ids, must be in the owner's
955 for (i = 0; !decline_attach && i < tc->cr_ngroups; i++)
956 decline_attach = !groupmember(tc->cr_groups[i], oc);
958 /* check the read and saved gids too */
959 if (decline_attach == 0)
960 decline_attach = !groupmember(tc->cr_rgid, oc) ||
961 !groupmember(tc->cr_svgid, oc);
966 return !decline_attach;
970 * Attach a process to a PMC.
974 pmc_attach_one_process(struct proc *p, struct pmc *pm)
977 char *fullpath, *freepath;
978 struct pmc_process *pp;
980 sx_assert(&pmc_sx, SX_XLOCKED);
982 PMCDBG5(PRC,ATT,2, "attach-one pm=%p ri=%d proc=%p (%d, %s)", pm,
983 PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm);
986 * Locate the process descriptor corresponding to process 'p',
987 * allocating space as needed.
989 * Verify that rowindex 'pm_rowindex' is free in the process
992 * If not, allocate space for a descriptor and link the
993 * process descriptor and PMC.
995 ri = PMC_TO_ROWINDEX(pm);
997 if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_ALLOCATE)) == NULL)
1000 if (pp->pp_pmcs[ri].pp_pmc == pm) /* already present at slot [ri] */
1003 if (pp->pp_pmcs[ri].pp_pmc != NULL)
1006 pmc_link_target_process(pm, pp);
1008 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) &&
1009 (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) == 0)
1010 pm->pm_flags |= PMC_F_NEEDS_LOGFILE;
1012 pm->pm_flags |= PMC_F_ATTACH_DONE; /* mark as attached */
1014 /* issue an attach event to a configured log file */
1015 if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE) {
1016 if (p->p_flag & P_KPROC) {
1017 fullpath = kernelname;
1020 pmc_getfilename(p->p_textvp, &fullpath, &freepath);
1021 pmclog_process_pmcattach(pm, p->p_pid, fullpath);
1023 free(freepath, M_TEMP);
1024 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1025 pmc_log_process_mappings(pm->pm_owner, p);
1027 /* mark process as using HWPMCs */
1029 p->p_flag |= P_HWPMC;
1036 * Attach a process and optionally its children
1040 pmc_attach_process(struct proc *p, struct pmc *pm)
1045 sx_assert(&pmc_sx, SX_XLOCKED);
1047 PMCDBG5(PRC,ATT,1, "attach pm=%p ri=%d proc=%p (%d, %s)", pm,
1048 PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm);
1052 * If this PMC successfully allowed a GETMSR operation
1053 * in the past, disallow further ATTACHes.
1056 if ((pm->pm_flags & PMC_PP_ENABLE_MSR_ACCESS) != 0)
1059 if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0)
1060 return pmc_attach_one_process(p, pm);
1063 * Traverse all child processes, attaching them to
1067 sx_slock(&proctree_lock);
1072 if ((error = pmc_attach_one_process(p, pm)) != 0)
1074 if (!LIST_EMPTY(&p->p_children))
1075 p = LIST_FIRST(&p->p_children);
1079 if (LIST_NEXT(p, p_sibling)) {
1080 p = LIST_NEXT(p, p_sibling);
1088 (void) pmc_detach_process(top, pm);
1091 sx_sunlock(&proctree_lock);
1096 * Detach a process from a PMC. If there are no other PMCs tracking
1097 * this process, remove the process structure from its hash table. If
1098 * 'flags' contains PMC_FLAG_REMOVE, then free the process structure.
1102 pmc_detach_one_process(struct proc *p, struct pmc *pm, int flags)
1105 struct pmc_process *pp;
1107 sx_assert(&pmc_sx, SX_XLOCKED);
1110 ("[pmc,%d] null pm pointer", __LINE__));
1112 ri = PMC_TO_ROWINDEX(pm);
1114 PMCDBG6(PRC,ATT,2, "detach-one pm=%p ri=%d proc=%p (%d, %s) flags=0x%x",
1115 pm, ri, p, p->p_pid, p->p_comm, flags);
1117 if ((pp = pmc_find_process_descriptor(p, 0)) == NULL)
1120 if (pp->pp_pmcs[ri].pp_pmc != pm)
1123 pmc_unlink_target_process(pm, pp);
1125 /* Issue a detach entry if a log file is configured */
1126 if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE)
1127 pmclog_process_pmcdetach(pm, p->p_pid);
1130 * If there are no PMCs targeting this process, we remove its
1131 * descriptor from the target hash table and unset the P_HWPMC
1132 * flag in the struct proc.
1134 KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= (int) md->pmd_npmc,
1135 ("[pmc,%d] Illegal refcnt %d for process struct %p",
1136 __LINE__, pp->pp_refcnt, pp));
1138 if (pp->pp_refcnt != 0) /* still a target of some PMC */
1141 pmc_remove_process_descriptor(pp);
1143 if (flags & PMC_FLAG_REMOVE)
1147 p->p_flag &= ~P_HWPMC;
1154 * Detach a process and optionally its descendants from a PMC.
1158 pmc_detach_process(struct proc *p, struct pmc *pm)
1162 sx_assert(&pmc_sx, SX_XLOCKED);
1164 PMCDBG5(PRC,ATT,1, "detach pm=%p ri=%d proc=%p (%d, %s)", pm,
1165 PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm);
1167 if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0)
1168 return pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE);
1171 * Traverse all children, detaching them from this PMC. We
1172 * ignore errors since we could be detaching a PMC from a
1173 * partially attached proc tree.
1176 sx_slock(&proctree_lock);
1181 (void) pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE);
1183 if (!LIST_EMPTY(&p->p_children))
1184 p = LIST_FIRST(&p->p_children);
1188 if (LIST_NEXT(p, p_sibling)) {
1189 p = LIST_NEXT(p, p_sibling);
1197 sx_sunlock(&proctree_lock);
1199 if (LIST_EMPTY(&pm->pm_targets))
1200 pm->pm_flags &= ~PMC_F_ATTACH_DONE;
1207 * Thread context switch IN
1211 pmc_process_csw_in(struct thread *td)
1214 unsigned int adjri, ri;
1219 pmc_value_t newvalue;
1220 struct pmc_process *pp;
1221 struct pmc_classdep *pcd;
1225 if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE)) == NULL)
1228 KASSERT(pp->pp_proc == td->td_proc,
1229 ("[pmc,%d] not my thread state", __LINE__));
1231 critical_enter(); /* no preemption from this point */
1233 cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */
1235 PMCDBG5(CSW,SWI,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p,
1236 p->p_pid, p->p_comm, pp);
1238 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
1239 ("[pmc,%d] weird CPU id %d", __LINE__, cpu));
1243 for (ri = 0; ri < md->pmd_npmc; ri++) {
1245 if ((pm = pp->pp_pmcs[ri].pp_pmc) == NULL)
1248 KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)),
1249 ("[pmc,%d] Target PMC in non-virtual mode (%d)",
1250 __LINE__, PMC_TO_MODE(pm)));
1252 KASSERT(PMC_TO_ROWINDEX(pm) == ri,
1253 ("[pmc,%d] Row index mismatch pmc %d != ri %d",
1254 __LINE__, PMC_TO_ROWINDEX(pm), ri));
1257 * Only PMCs that are marked as 'RUNNING' need
1258 * be placed on hardware.
1261 if (pm->pm_state != PMC_STATE_RUNNING)
1264 /* increment PMC runcount */
1265 atomic_add_rel_int(&pm->pm_runcount, 1);
1267 /* configure the HWPMC we are going to use. */
1268 pcd = pmc_ri_to_classdep(md, ri, &adjri);
1269 pcd->pcd_config_pmc(cpu, adjri, pm);
1271 phw = pc->pc_hwpmcs[ri];
1273 KASSERT(phw != NULL,
1274 ("[pmc,%d] null hw pointer", __LINE__));
1276 KASSERT(phw->phw_pmc == pm,
1277 ("[pmc,%d] hw->pmc %p != pmc %p", __LINE__,
1281 * Write out saved value and start the PMC.
1283 * Sampling PMCs use a per-process value, while
1284 * counting mode PMCs use a per-pmc value that is
1285 * inherited across descendants.
1287 if (PMC_TO_MODE(pm) == PMC_MODE_TS) {
1288 mtx_pool_lock_spin(pmc_mtxpool, pm);
1291 * Use the saved value calculated after the most recent
1292 * thread switch out to start this counter. Reset
1293 * the saved count in case another thread from this
1294 * process switches in before any threads switch out.
1296 newvalue = PMC_PCPU_SAVED(cpu,ri) =
1297 pp->pp_pmcs[ri].pp_pmcval;
1298 pp->pp_pmcs[ri].pp_pmcval = pm->pm_sc.pm_reloadcount;
1299 mtx_pool_unlock_spin(pmc_mtxpool, pm);
1301 KASSERT(PMC_TO_MODE(pm) == PMC_MODE_TC,
1302 ("[pmc,%d] illegal mode=%d", __LINE__,
1304 mtx_pool_lock_spin(pmc_mtxpool, pm);
1305 newvalue = PMC_PCPU_SAVED(cpu, ri) =
1306 pm->pm_gv.pm_savedvalue;
1307 mtx_pool_unlock_spin(pmc_mtxpool, pm);
1310 PMCDBG3(CSW,SWI,1,"cpu=%d ri=%d new=%jd", cpu, ri, newvalue);
1312 pcd->pcd_write_pmc(cpu, adjri, newvalue);
1314 /* If a sampling mode PMC, reset stalled state. */
1315 if (PMC_TO_MODE(pm) == PMC_MODE_TS)
1316 CPU_CLR_ATOMIC(cpu, &pm->pm_stalled);
1318 /* Indicate that we desire this to run. */
1319 CPU_SET_ATOMIC(cpu, &pm->pm_cpustate);
1321 /* Start the PMC. */
1322 pcd->pcd_start_pmc(cpu, adjri);
1326 * perform any other architecture/cpu dependent thread
1327 * switch-in actions.
1330 (void) (*md->pmd_switch_in)(pc, pp);
1337 * Thread context switch OUT.
1341 pmc_process_csw_out(struct thread *td)
1349 pmc_value_t newvalue;
1350 unsigned int adjri, ri;
1351 struct pmc_process *pp;
1352 struct pmc_classdep *pcd;
1356 * Locate our process descriptor; this may be NULL if
1357 * this process is exiting and we have already removed
1358 * the process from the target process table.
1360 * Note that due to kernel preemption, multiple
1361 * context switches may happen while the process is
1364 * Note also that if the target process cannot be
1365 * found we still need to deconfigure any PMCs that
1366 * are currently running on hardware.
1370 pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE);
1378 cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */
1380 PMCDBG5(CSW,SWO,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p,
1381 p->p_pid, p->p_comm, pp);
1383 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
1384 ("[pmc,%d weird CPU id %d", __LINE__, cpu));
1389 * When a PMC gets unlinked from a target PMC, it will
1390 * be removed from the target's pp_pmc[] array.
1392 * However, on a MP system, the target could have been
1393 * executing on another CPU at the time of the unlink.
1394 * So, at context switch OUT time, we need to look at
1395 * the hardware to determine if a PMC is scheduled on
1399 for (ri = 0; ri < md->pmd_npmc; ri++) {
1401 pcd = pmc_ri_to_classdep(md, ri, &adjri);
1403 (void) (*pcd->pcd_get_config)(cpu, adjri, &pm);
1405 if (pm == NULL) /* nothing at this row index */
1408 mode = PMC_TO_MODE(pm);
1409 if (!PMC_IS_VIRTUAL_MODE(mode))
1410 continue; /* not a process virtual PMC */
1412 KASSERT(PMC_TO_ROWINDEX(pm) == ri,
1413 ("[pmc,%d] ri mismatch pmc(%d) ri(%d)",
1414 __LINE__, PMC_TO_ROWINDEX(pm), ri));
1417 * Change desired state, and then stop if not stalled.
1418 * This two-step dance should avoid race conditions where
1419 * an interrupt re-enables the PMC after this code has
1420 * already checked the pm_stalled flag.
1422 CPU_CLR_ATOMIC(cpu, &pm->pm_cpustate);
1423 if (!CPU_ISSET(cpu, &pm->pm_stalled))
1424 pcd->pcd_stop_pmc(cpu, adjri);
1426 /* reduce this PMC's runcount */
1427 atomic_subtract_rel_int(&pm->pm_runcount, 1);
1430 * If this PMC is associated with this process,
1434 if (pm->pm_state != PMC_STATE_DELETED && pp != NULL &&
1435 pp->pp_pmcs[ri].pp_pmc != NULL) {
1436 KASSERT(pm == pp->pp_pmcs[ri].pp_pmc,
1437 ("[pmc,%d] pm %p != pp_pmcs[%d] %p", __LINE__,
1438 pm, ri, pp->pp_pmcs[ri].pp_pmc));
1440 KASSERT(pp->pp_refcnt > 0,
1441 ("[pmc,%d] pp refcnt = %d", __LINE__,
1444 pcd->pcd_read_pmc(cpu, adjri, &newvalue);
1446 if (mode == PMC_MODE_TS) {
1447 PMCDBG3(CSW,SWO,1,"cpu=%d ri=%d tmp=%jd (samp)",
1448 cpu, ri, PMC_PCPU_SAVED(cpu,ri) - newvalue);
1451 * For sampling process-virtual PMCs,
1452 * newvalue is the number of events to be seen
1453 * until the next sampling interrupt.
1454 * We can just add the events left from this
1455 * invocation to the counter, then adjust
1456 * in case we overflow our range.
1458 * (Recall that we reload the counter every
1461 mtx_pool_lock_spin(pmc_mtxpool, pm);
1463 pp->pp_pmcs[ri].pp_pmcval += newvalue;
1464 if (pp->pp_pmcs[ri].pp_pmcval >
1465 pm->pm_sc.pm_reloadcount)
1466 pp->pp_pmcs[ri].pp_pmcval -=
1467 pm->pm_sc.pm_reloadcount;
1468 KASSERT(pp->pp_pmcs[ri].pp_pmcval > 0 &&
1469 pp->pp_pmcs[ri].pp_pmcval <=
1470 pm->pm_sc.pm_reloadcount,
1471 ("[pmc,%d] pp_pmcval outside of expected "
1472 "range cpu=%d ri=%d pp_pmcval=%jx "
1473 "pm_reloadcount=%jx", __LINE__, cpu, ri,
1474 pp->pp_pmcs[ri].pp_pmcval,
1475 pm->pm_sc.pm_reloadcount));
1476 mtx_pool_unlock_spin(pmc_mtxpool, pm);
1479 tmp = newvalue - PMC_PCPU_SAVED(cpu,ri);
1481 PMCDBG3(CSW,SWO,1,"cpu=%d ri=%d tmp=%jd (count)",
1485 * For counting process-virtual PMCs,
1486 * we expect the count to be
1487 * increasing monotonically, modulo a 64
1491 ("[pmc,%d] negative increment cpu=%d "
1492 "ri=%d newvalue=%jx saved=%jx "
1493 "incr=%jx", __LINE__, cpu, ri,
1494 newvalue, PMC_PCPU_SAVED(cpu,ri), tmp));
1496 mtx_pool_lock_spin(pmc_mtxpool, pm);
1497 pm->pm_gv.pm_savedvalue += tmp;
1498 pp->pp_pmcs[ri].pp_pmcval += tmp;
1499 mtx_pool_unlock_spin(pmc_mtxpool, pm);
1501 if (pm->pm_flags & PMC_F_LOG_PROCCSW)
1502 pmclog_process_proccsw(pm, pp, tmp);
1506 /* mark hardware as free */
1507 pcd->pcd_config_pmc(cpu, adjri, NULL);
1511 * perform any other architecture/cpu dependent thread
1512 * switch out functions.
1515 (void) (*md->pmd_switch_out)(pc, pp);
1521 * A mapping change for a process.
1525 pmc_process_mmap(struct thread *td, struct pmckern_map_in *pkm)
1529 char *fullpath, *freepath;
1530 const struct pmc *pm;
1531 struct pmc_owner *po;
1532 const struct pmc_process *pp;
1534 freepath = fullpath = NULL;
1535 pmc_getfilename((struct vnode *) pkm->pm_file, &fullpath, &freepath);
1537 pid = td->td_proc->p_pid;
1539 /* Inform owners of all system-wide sampling PMCs. */
1540 LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1541 if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1542 pmclog_process_map_in(po, pid, pkm->pm_address, fullpath);
1544 if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL)
1548 * Inform sampling PMC owners tracking this process.
1550 for (ri = 0; ri < md->pmd_npmc; ri++)
1551 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL &&
1552 PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1553 pmclog_process_map_in(pm->pm_owner,
1554 pid, pkm->pm_address, fullpath);
1558 free(freepath, M_TEMP);
1563 * Log an munmap request.
1567 pmc_process_munmap(struct thread *td, struct pmckern_map_out *pkm)
1571 struct pmc_owner *po;
1572 const struct pmc *pm;
1573 const struct pmc_process *pp;
1575 pid = td->td_proc->p_pid;
1577 LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1578 if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1579 pmclog_process_map_out(po, pid, pkm->pm_address,
1580 pkm->pm_address + pkm->pm_size);
1582 if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL)
1585 for (ri = 0; ri < md->pmd_npmc; ri++)
1586 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL &&
1587 PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1588 pmclog_process_map_out(pm->pm_owner, pid,
1589 pkm->pm_address, pkm->pm_address + pkm->pm_size);
1593 * Log mapping information about the kernel.
1597 pmc_log_kernel_mappings(struct pmc *pm)
1599 struct pmc_owner *po;
1600 struct pmckern_map_in *km, *kmbase;
1602 sx_assert(&pmc_sx, SX_LOCKED);
1603 KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)),
1604 ("[pmc,%d] non-sampling PMC (%p) desires mapping information",
1605 __LINE__, (void *) pm));
1609 if (po->po_flags & PMC_PO_INITIAL_MAPPINGS_DONE)
1613 * Log the current set of kernel modules.
1615 kmbase = linker_hwpmc_list_objects();
1616 for (km = kmbase; km->pm_file != NULL; km++) {
1617 PMCDBG2(LOG,REG,1,"%s %p", (char *) km->pm_file,
1618 (void *) km->pm_address);
1619 pmclog_process_map_in(po, (pid_t) -1, km->pm_address,
1622 free(kmbase, M_LINKER);
1624 po->po_flags |= PMC_PO_INITIAL_MAPPINGS_DONE;
1628 * Log the mappings for a single process.
1632 pmc_log_process_mappings(struct pmc_owner *po, struct proc *p)
1637 vm_map_entry_t entry;
1638 vm_offset_t last_end;
1639 u_int last_timestamp;
1640 struct vnode *last_vp;
1641 vm_offset_t start_addr;
1642 vm_object_t obj, lobj, tobj;
1643 char *fullpath, *freepath;
1646 last_end = (vm_offset_t) 0;
1647 fullpath = freepath = NULL;
1649 if ((vm = vmspace_acquire_ref(p)) == NULL)
1653 vm_map_lock_read(map);
1655 for (entry = map->header.next; entry != &map->header; entry = entry->next) {
1657 if (entry == NULL) {
1658 PMCDBG2(LOG,OPS,2, "hwpmc: vm_map entry unexpectedly "
1659 "NULL! pid=%d vm_map=%p\n", p->p_pid, map);
1664 * We only care about executable map entries.
1666 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) ||
1667 !(entry->protection & VM_PROT_EXECUTE) ||
1668 (entry->object.vm_object == NULL)) {
1672 obj = entry->object.vm_object;
1673 VM_OBJECT_RLOCK(obj);
1676 * Walk the backing_object list to find the base
1677 * (non-shadowed) vm_object.
1679 for (lobj = tobj = obj; tobj != NULL; tobj = tobj->backing_object) {
1681 VM_OBJECT_RLOCK(tobj);
1683 VM_OBJECT_RUNLOCK(lobj);
1688 * At this point lobj is the base vm_object and it is locked.
1691 PMCDBG3(LOG,OPS,2, "hwpmc: lobj unexpectedly NULL! pid=%d "
1692 "vm_map=%p vm_obj=%p\n", p->p_pid, map, obj);
1693 VM_OBJECT_RUNLOCK(obj);
1697 vp = vm_object_vnode(lobj);
1700 VM_OBJECT_RUNLOCK(lobj);
1701 VM_OBJECT_RUNLOCK(obj);
1706 * Skip contiguous regions that point to the same
1707 * vnode, so we don't emit redundant MAP-IN
1710 if (entry->start == last_end && vp == last_vp) {
1711 last_end = entry->end;
1713 VM_OBJECT_RUNLOCK(lobj);
1714 VM_OBJECT_RUNLOCK(obj);
1719 * We don't want to keep the proc's vm_map or this
1720 * vm_object locked while we walk the pathname, since
1721 * vn_fullpath() can sleep. However, if we drop the
1722 * lock, it's possible for concurrent activity to
1723 * modify the vm_map list. To protect against this,
1724 * we save the vm_map timestamp before we release the
1725 * lock, and check it after we reacquire the lock
1728 start_addr = entry->start;
1729 last_end = entry->end;
1730 last_timestamp = map->timestamp;
1731 vm_map_unlock_read(map);
1735 VM_OBJECT_RUNLOCK(lobj);
1737 VM_OBJECT_RUNLOCK(obj);
1740 pmc_getfilename(vp, &fullpath, &freepath);
1746 pmclog_process_map_in(po, p->p_pid, start_addr, fullpath);
1748 free(freepath, M_TEMP);
1750 vm_map_lock_read(map);
1753 * If our saved timestamp doesn't match, this means
1754 * that the vm_map was modified out from under us and
1755 * we can't trust our current "entry" pointer. Do a
1756 * new lookup for this entry. If there is no entry
1757 * for this address range, vm_map_lookup_entry() will
1758 * return the previous one, so we always want to go to
1759 * entry->next on the next loop iteration.
1761 * There is an edge condition here that can occur if
1762 * there is no entry at or before this address. In
1763 * this situation, vm_map_lookup_entry returns
1764 * &map->header, which would cause our loop to abort
1765 * without processing the rest of the map. However,
1766 * in practice this will never happen for process
1767 * vm_map. This is because the executable's text
1768 * segment is the first mapping in the proc's address
1769 * space, and this mapping is never removed until the
1770 * process exits, so there will always be a non-header
1771 * entry at or before the requested address for
1772 * vm_map_lookup_entry to return.
1774 if (map->timestamp != last_timestamp)
1775 vm_map_lookup_entry(map, last_end - 1, &entry);
1778 vm_map_unlock_read(map);
1784 * Log mappings for all processes in the system.
1788 pmc_log_all_process_mappings(struct pmc_owner *po)
1790 struct proc *p, *top;
1792 sx_assert(&pmc_sx, SX_XLOCKED);
1794 if ((p = pfind(1)) == NULL)
1795 panic("[pmc,%d] Cannot find init", __LINE__);
1799 sx_slock(&proctree_lock);
1804 pmc_log_process_mappings(po, p);
1805 if (!LIST_EMPTY(&p->p_children))
1806 p = LIST_FIRST(&p->p_children);
1810 if (LIST_NEXT(p, p_sibling)) {
1811 p = LIST_NEXT(p, p_sibling);
1818 sx_sunlock(&proctree_lock);
1822 * The 'hook' invoked from the kernel proper
1827 const char *pmc_hooknames[] = {
1828 /* these strings correspond to PMC_FN_* in <sys/pmckern.h> */
1845 pmc_hook_handler(struct thread *td, int function, void *arg)
1848 PMCDBG4(MOD,PMH,1, "hook td=%p func=%d \"%s\" arg=%p", td, function,
1849 pmc_hooknames[function], arg);
1858 case PMC_FN_PROCESS_EXEC:
1860 char *fullpath, *freepath;
1862 int is_using_hwpmcs;
1865 struct pmc_owner *po;
1866 struct pmc_process *pp;
1867 struct pmckern_procexec *pk;
1869 sx_assert(&pmc_sx, SX_XLOCKED);
1872 pmc_getfilename(p->p_textvp, &fullpath, &freepath);
1874 pk = (struct pmckern_procexec *) arg;
1876 /* Inform owners of SS mode PMCs of the exec event. */
1877 LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1878 if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1879 pmclog_process_procexec(po, PMC_ID_INVALID,
1880 p->p_pid, pk->pm_entryaddr, fullpath);
1883 is_using_hwpmcs = p->p_flag & P_HWPMC;
1886 if (!is_using_hwpmcs) {
1888 free(freepath, M_TEMP);
1893 * PMCs are not inherited across an exec(): remove any
1894 * PMCs that this process is the owner of.
1897 if ((po = pmc_find_owner_descriptor(p)) != NULL) {
1898 pmc_remove_owner(po);
1899 pmc_destroy_owner_descriptor(po);
1903 * If the process being exec'ed is not the target of any
1906 if ((pp = pmc_find_process_descriptor(p, 0)) == NULL) {
1908 free(freepath, M_TEMP);
1913 * Log the exec event to all monitoring owners. Skip
1914 * owners who have already received the event because
1915 * they had system sampling PMCs active.
1917 for (ri = 0; ri < md->pmd_npmc; ri++)
1918 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) {
1920 if (po->po_sscount == 0 &&
1921 po->po_flags & PMC_PO_OWNS_LOGFILE)
1922 pmclog_process_procexec(po, pm->pm_id,
1923 p->p_pid, pk->pm_entryaddr,
1928 free(freepath, M_TEMP);
1931 PMCDBG4(PRC,EXC,1, "exec proc=%p (%d, %s) cred-changed=%d",
1932 p, p->p_pid, p->p_comm, pk->pm_credentialschanged);
1934 if (pk->pm_credentialschanged == 0) /* no change */
1938 * If the newly exec()'ed process has a different credential
1939 * than before, allow it to be the target of a PMC only if
1940 * the PMC's owner has sufficient privilege.
1943 for (ri = 0; ri < md->pmd_npmc; ri++)
1944 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL)
1945 if (pmc_can_attach(pm, td->td_proc) != 0)
1946 pmc_detach_one_process(td->td_proc,
1949 KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= (int) md->pmd_npmc,
1950 ("[pmc,%d] Illegal ref count %d on pp %p", __LINE__,
1951 pp->pp_refcnt, pp));
1954 * If this process is no longer the target of any
1955 * PMCs, we can remove the process entry and free
1959 if (pp->pp_refcnt == 0) {
1960 pmc_remove_process_descriptor(pp);
1969 pmc_process_csw_in(td);
1972 case PMC_FN_CSW_OUT:
1973 pmc_process_csw_out(td);
1977 * Process accumulated PC samples.
1979 * This function is expected to be called by hardclock() for
1980 * each CPU that has accumulated PC samples.
1982 * This function is to be executed on the CPU whose samples
1983 * are being processed.
1985 case PMC_FN_DO_SAMPLES:
1988 * Clear the cpu specific bit in the CPU mask before
1989 * do the rest of the processing. If the NMI handler
1990 * gets invoked after the "atomic_clear_int()" call
1991 * below but before "pmc_process_samples()" gets
1992 * around to processing the interrupt, then we will
1993 * come back here at the next hardclock() tick (and
1994 * may find nothing to do if "pmc_process_samples()"
1995 * had already processed the interrupt). We don't
1996 * lose the interrupt sample.
1998 CPU_CLR_ATOMIC(PCPU_GET(cpuid), &pmc_cpumask);
1999 pmc_process_samples(PCPU_GET(cpuid), PMC_HR);
2000 pmc_process_samples(PCPU_GET(cpuid), PMC_SR);
2004 sx_assert(&pmc_sx, SX_LOCKED);
2005 pmc_process_mmap(td, (struct pmckern_map_in *) arg);
2009 sx_assert(&pmc_sx, SX_LOCKED);
2010 pmc_process_munmap(td, (struct pmckern_map_out *) arg);
2013 case PMC_FN_USER_CALLCHAIN:
2015 * Record a call chain.
2017 KASSERT(td == curthread, ("[pmc,%d] td != curthread",
2020 pmc_capture_user_callchain(PCPU_GET(cpuid), PMC_HR,
2021 (struct trapframe *) arg);
2022 td->td_pflags &= ~TDP_CALLCHAIN;
2025 case PMC_FN_USER_CALLCHAIN_SOFT:
2027 * Record a call chain.
2029 KASSERT(td == curthread, ("[pmc,%d] td != curthread",
2031 pmc_capture_user_callchain(PCPU_GET(cpuid), PMC_SR,
2032 (struct trapframe *) arg);
2033 td->td_pflags &= ~TDP_CALLCHAIN;
2036 case PMC_FN_SOFT_SAMPLING:
2038 * Call soft PMC sampling intr.
2040 pmc_soft_intr((struct pmckern_soft *) arg);
2045 KASSERT(0, ("[pmc,%d] unknown hook %d\n", __LINE__, function));
2055 * allocate a 'struct pmc_owner' descriptor in the owner hash table.
2058 static struct pmc_owner *
2059 pmc_allocate_owner_descriptor(struct proc *p)
2062 struct pmc_owner *po;
2063 struct pmc_ownerhash *poh;
2065 hindex = PMC_HASH_PTR(p, pmc_ownerhashmask);
2066 poh = &pmc_ownerhash[hindex];
2068 /* allocate space for N pointers and one descriptor struct */
2069 po = malloc(sizeof(struct pmc_owner), M_PMC, M_WAITOK|M_ZERO);
2071 LIST_INSERT_HEAD(poh, po, po_next); /* insert into hash table */
2073 TAILQ_INIT(&po->po_logbuffers);
2074 mtx_init(&po->po_mtx, "pmc-owner-mtx", "pmc-per-proc", MTX_SPIN);
2076 PMCDBG4(OWN,ALL,1, "allocate-owner proc=%p (%d, %s) pmc-owner=%p",
2077 p, p->p_pid, p->p_comm, po);
2083 pmc_destroy_owner_descriptor(struct pmc_owner *po)
2086 PMCDBG4(OWN,REL,1, "destroy-owner po=%p proc=%p (%d, %s)",
2087 po, po->po_owner, po->po_owner->p_pid, po->po_owner->p_comm);
2089 mtx_destroy(&po->po_mtx);
2094 * find the descriptor corresponding to process 'p', adding or removing it
2095 * as specified by 'mode'.
2098 static struct pmc_process *
2099 pmc_find_process_descriptor(struct proc *p, uint32_t mode)
2102 struct pmc_process *pp, *ppnew;
2103 struct pmc_processhash *pph;
2105 hindex = PMC_HASH_PTR(p, pmc_processhashmask);
2106 pph = &pmc_processhash[hindex];
2111 * Pre-allocate memory in the FIND_ALLOCATE case since we
2112 * cannot call malloc(9) once we hold a spin lock.
2114 if (mode & PMC_FLAG_ALLOCATE)
2115 ppnew = malloc(sizeof(struct pmc_process) + md->pmd_npmc *
2116 sizeof(struct pmc_targetstate), M_PMC, M_WAITOK|M_ZERO);
2118 mtx_lock_spin(&pmc_processhash_mtx);
2119 LIST_FOREACH(pp, pph, pp_next)
2120 if (pp->pp_proc == p)
2123 if ((mode & PMC_FLAG_REMOVE) && pp != NULL)
2124 LIST_REMOVE(pp, pp_next);
2126 if ((mode & PMC_FLAG_ALLOCATE) && pp == NULL &&
2129 LIST_INSERT_HEAD(pph, ppnew, pp_next);
2133 mtx_unlock_spin(&pmc_processhash_mtx);
2135 if (pp != NULL && ppnew != NULL)
2142 * remove a process descriptor from the process hash table.
2146 pmc_remove_process_descriptor(struct pmc_process *pp)
2148 KASSERT(pp->pp_refcnt == 0,
2149 ("[pmc,%d] Removing process descriptor %p with count %d",
2150 __LINE__, pp, pp->pp_refcnt));
2152 mtx_lock_spin(&pmc_processhash_mtx);
2153 LIST_REMOVE(pp, pp_next);
2154 mtx_unlock_spin(&pmc_processhash_mtx);
2159 * find an owner descriptor corresponding to proc 'p'
2162 static struct pmc_owner *
2163 pmc_find_owner_descriptor(struct proc *p)
2166 struct pmc_owner *po;
2167 struct pmc_ownerhash *poh;
2169 hindex = PMC_HASH_PTR(p, pmc_ownerhashmask);
2170 poh = &pmc_ownerhash[hindex];
2173 LIST_FOREACH(po, poh, po_next)
2174 if (po->po_owner == p)
2177 PMCDBG5(OWN,FND,1, "find-owner proc=%p (%d, %s) hindex=0x%x -> "
2178 "pmc-owner=%p", p, p->p_pid, p->p_comm, hindex, po);
2184 * pmc_allocate_pmc_descriptor
2186 * Allocate a pmc descriptor and initialize its
2191 pmc_allocate_pmc_descriptor(void)
2195 pmc = malloc(sizeof(struct pmc), M_PMC, M_WAITOK|M_ZERO);
2197 PMCDBG1(PMC,ALL,1, "allocate-pmc -> pmc=%p", pmc);
2203 * Destroy a pmc descriptor.
2207 pmc_destroy_pmc_descriptor(struct pmc *pm)
2210 KASSERT(pm->pm_state == PMC_STATE_DELETED ||
2211 pm->pm_state == PMC_STATE_FREE,
2212 ("[pmc,%d] destroying non-deleted PMC", __LINE__));
2213 KASSERT(LIST_EMPTY(&pm->pm_targets),
2214 ("[pmc,%d] destroying pmc with targets", __LINE__));
2215 KASSERT(pm->pm_owner == NULL,
2216 ("[pmc,%d] destroying pmc attached to an owner", __LINE__));
2217 KASSERT(pm->pm_runcount == 0,
2218 ("[pmc,%d] pmc has non-zero run count %d", __LINE__,
2225 pmc_wait_for_pmc_idle(struct pmc *pm)
2228 volatile int maxloop;
2230 maxloop = 100 * pmc_cpu_max();
2233 * Loop (with a forced context switch) till the PMC's runcount
2234 * comes down to zero.
2236 while (atomic_load_acq_32(&pm->pm_runcount) > 0) {
2239 KASSERT(maxloop > 0,
2240 ("[pmc,%d] (ri%d, rc%d) waiting too long for "
2241 "pmc to be free", __LINE__,
2242 PMC_TO_ROWINDEX(pm), pm->pm_runcount));
2244 pmc_force_context_switch();
2249 * This function does the following things:
2251 * - detaches the PMC from hardware
2252 * - unlinks all target threads that were attached to it
2253 * - removes the PMC from its owner's list
2254 * - destroys the PMC private mutex
2256 * Once this function completes, the given pmc pointer can be freed by
2257 * calling pmc_destroy_pmc_descriptor().
2261 pmc_release_pmc_descriptor(struct pmc *pm)
2265 u_int adjri, ri, cpu;
2266 struct pmc_owner *po;
2267 struct pmc_binding pb;
2268 struct pmc_process *pp;
2269 struct pmc_classdep *pcd;
2270 struct pmc_target *ptgt, *tmp;
2272 sx_assert(&pmc_sx, SX_XLOCKED);
2274 KASSERT(pm, ("[pmc,%d] null pmc", __LINE__));
2276 ri = PMC_TO_ROWINDEX(pm);
2277 pcd = pmc_ri_to_classdep(md, ri, &adjri);
2278 mode = PMC_TO_MODE(pm);
2280 PMCDBG3(PMC,REL,1, "release-pmc pmc=%p ri=%d mode=%d", pm, ri,
2284 * First, we take the PMC off hardware.
2287 if (PMC_IS_SYSTEM_MODE(mode)) {
2290 * A system mode PMC runs on a specific CPU. Switch
2291 * to this CPU and turn hardware off.
2293 pmc_save_cpu_binding(&pb);
2295 cpu = PMC_TO_CPU(pm);
2297 pmc_select_cpu(cpu);
2299 /* switch off non-stalled CPUs */
2300 CPU_CLR_ATOMIC(cpu, &pm->pm_cpustate);
2301 if (pm->pm_state == PMC_STATE_RUNNING &&
2302 !CPU_ISSET(cpu, &pm->pm_stalled)) {
2304 phw = pmc_pcpu[cpu]->pc_hwpmcs[ri];
2306 KASSERT(phw->phw_pmc == pm,
2307 ("[pmc, %d] pmc ptr ri(%d) hw(%p) pm(%p)",
2308 __LINE__, ri, phw->phw_pmc, pm));
2309 PMCDBG2(PMC,REL,2, "stopping cpu=%d ri=%d", cpu, ri);
2312 pcd->pcd_stop_pmc(cpu, adjri);
2316 PMCDBG2(PMC,REL,2, "decfg cpu=%d ri=%d", cpu, ri);
2319 pcd->pcd_config_pmc(cpu, adjri, NULL);
2322 /* adjust the global and process count of SS mode PMCs */
2323 if (mode == PMC_MODE_SS && pm->pm_state == PMC_STATE_RUNNING) {
2326 if (po->po_sscount == 0) {
2327 atomic_subtract_rel_int(&pmc_ss_count, 1);
2328 LIST_REMOVE(po, po_ssnext);
2332 pm->pm_state = PMC_STATE_DELETED;
2334 pmc_restore_cpu_binding(&pb);
2337 * We could have references to this PMC structure in
2338 * the per-cpu sample queues. Wait for the queue to
2341 pmc_wait_for_pmc_idle(pm);
2343 } else if (PMC_IS_VIRTUAL_MODE(mode)) {
2346 * A virtual PMC could be running on multiple CPUs at
2349 * By marking its state as DELETED, we ensure that
2350 * this PMC is never further scheduled on hardware.
2352 * Then we wait till all CPUs are done with this PMC.
2354 pm->pm_state = PMC_STATE_DELETED;
2357 /* Wait for the PMCs runcount to come to zero. */
2358 pmc_wait_for_pmc_idle(pm);
2361 * At this point the PMC is off all CPUs and cannot be
2362 * freshly scheduled onto a CPU. It is now safe to
2363 * unlink all targets from this PMC. If a
2364 * process-record's refcount falls to zero, we remove
2365 * it from the hash table. The module-wide SX lock
2366 * protects us from races.
2368 LIST_FOREACH_SAFE(ptgt, &pm->pm_targets, pt_next, tmp) {
2369 pp = ptgt->pt_process;
2370 pmc_unlink_target_process(pm, pp); /* frees 'ptgt' */
2372 PMCDBG1(PMC,REL,3, "pp->refcnt=%d", pp->pp_refcnt);
2375 * If the target process record shows that no
2376 * PMCs are attached to it, reclaim its space.
2379 if (pp->pp_refcnt == 0) {
2380 pmc_remove_process_descriptor(pp);
2385 cpu = curthread->td_oncpu; /* setup cpu for pmd_release() */
2390 * Release any MD resources
2392 (void) pcd->pcd_release_pmc(cpu, adjri, pm);
2395 * Update row disposition
2398 if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm)))
2399 PMC_UNMARK_ROW_STANDALONE(ri);
2401 PMC_UNMARK_ROW_THREAD(ri);
2403 /* unlink from the owner's list */
2405 LIST_REMOVE(pm, pm_next);
2406 pm->pm_owner = NULL;
2411 * Register an owner and a pmc.
2415 pmc_register_owner(struct proc *p, struct pmc *pmc)
2417 struct pmc_owner *po;
2419 sx_assert(&pmc_sx, SX_XLOCKED);
2421 if ((po = pmc_find_owner_descriptor(p)) == NULL)
2422 if ((po = pmc_allocate_owner_descriptor(p)) == NULL)
2425 KASSERT(pmc->pm_owner == NULL,
2426 ("[pmc,%d] attempting to own an initialized PMC", __LINE__));
2429 LIST_INSERT_HEAD(&po->po_pmcs, pmc, pm_next);
2432 p->p_flag |= P_HWPMC;
2435 if (po->po_flags & PMC_PO_OWNS_LOGFILE)
2436 pmclog_process_pmcallocate(pmc);
2438 PMCDBG2(PMC,REG,1, "register-owner pmc-owner=%p pmc=%p",
2445 * Return the current row disposition:
2447 * > 0 => PROCESS MODE
2448 * < 0 => SYSTEM MODE
2452 pmc_getrowdisp(int ri)
2454 return pmc_pmcdisp[ri];
2458 * Check if a PMC at row index 'ri' can be allocated to the current
2461 * Allocation can fail if:
2462 * - the current process is already being profiled by a PMC at index 'ri',
2463 * attached to it via OP_PMCATTACH.
2464 * - the current process has already allocated a PMC at index 'ri'
2469 pmc_can_allocate_rowindex(struct proc *p, unsigned int ri, int cpu)
2473 struct pmc_owner *po;
2474 struct pmc_process *pp;
2476 PMCDBG5(PMC,ALR,1, "can-allocate-rowindex proc=%p (%d, %s) ri=%d "
2477 "cpu=%d", p, p->p_pid, p->p_comm, ri, cpu);
2480 * We shouldn't have already allocated a process-mode PMC at
2483 * We shouldn't have allocated a system-wide PMC on the same
2486 if ((po = pmc_find_owner_descriptor(p)) != NULL)
2487 LIST_FOREACH(pm, &po->po_pmcs, pm_next) {
2488 if (PMC_TO_ROWINDEX(pm) == ri) {
2489 mode = PMC_TO_MODE(pm);
2490 if (PMC_IS_VIRTUAL_MODE(mode))
2492 if (PMC_IS_SYSTEM_MODE(mode) &&
2493 (int) PMC_TO_CPU(pm) == cpu)
2499 * We also shouldn't be the target of any PMC at this index
2500 * since otherwise a PMC_ATTACH to ourselves will fail.
2502 if ((pp = pmc_find_process_descriptor(p, 0)) != NULL)
2503 if (pp->pp_pmcs[ri].pp_pmc)
2506 PMCDBG4(PMC,ALR,2, "can-allocate-rowindex proc=%p (%d, %s) ri=%d ok",
2507 p, p->p_pid, p->p_comm, ri);
2513 * Check if a given PMC at row index 'ri' can be currently used in
2518 pmc_can_allocate_row(int ri, enum pmc_mode mode)
2522 sx_assert(&pmc_sx, SX_XLOCKED);
2524 PMCDBG2(PMC,ALR,1, "can-allocate-row ri=%d mode=%d", ri, mode);
2526 if (PMC_IS_SYSTEM_MODE(mode))
2527 disp = PMC_DISP_STANDALONE;
2529 disp = PMC_DISP_THREAD;
2532 * check disposition for PMC row 'ri':
2534 * Expected disposition Row-disposition Result
2536 * STANDALONE STANDALONE or FREE proceed
2537 * STANDALONE THREAD fail
2538 * THREAD THREAD or FREE proceed
2539 * THREAD STANDALONE fail
2542 if (!PMC_ROW_DISP_IS_FREE(ri) &&
2543 !(disp == PMC_DISP_THREAD && PMC_ROW_DISP_IS_THREAD(ri)) &&
2544 !(disp == PMC_DISP_STANDALONE && PMC_ROW_DISP_IS_STANDALONE(ri)))
2551 PMCDBG2(PMC,ALR,2, "can-allocate-row ri=%d mode=%d ok", ri, mode);
2558 * Find a PMC descriptor with user handle 'pmcid' for thread 'td'.
2562 pmc_find_pmc_descriptor_in_process(struct pmc_owner *po, pmc_id_t pmcid)
2566 KASSERT(PMC_ID_TO_ROWINDEX(pmcid) < md->pmd_npmc,
2567 ("[pmc,%d] Illegal pmc index %d (max %d)", __LINE__,
2568 PMC_ID_TO_ROWINDEX(pmcid), md->pmd_npmc));
2570 LIST_FOREACH(pm, &po->po_pmcs, pm_next)
2571 if (pm->pm_id == pmcid)
2578 pmc_find_pmc(pmc_id_t pmcid, struct pmc **pmc)
2581 struct pmc *pm, *opm;
2582 struct pmc_owner *po;
2583 struct pmc_process *pp;
2585 KASSERT(PMC_ID_TO_ROWINDEX(pmcid) < md->pmd_npmc,
2586 ("[pmc,%d] Illegal pmc index %d (max %d)", __LINE__,
2587 PMC_ID_TO_ROWINDEX(pmcid), md->pmd_npmc));
2588 PMCDBG1(PMC,FND,1, "find-pmc id=%d", pmcid);
2590 if ((po = pmc_find_owner_descriptor(curthread->td_proc)) == NULL) {
2592 * In case of PMC_F_DESCENDANTS child processes we will not find
2593 * the current process in the owners hash list. Find the owner
2594 * process first and from there lookup the po.
2596 if ((pp = pmc_find_process_descriptor(curthread->td_proc,
2597 PMC_FLAG_NONE)) == NULL) {
2600 opm = pp->pp_pmcs[PMC_ID_TO_ROWINDEX(pmcid)].pp_pmc;
2603 if ((opm->pm_flags & (PMC_F_ATTACHED_TO_OWNER|
2604 PMC_F_DESCENDANTS)) != (PMC_F_ATTACHED_TO_OWNER|
2611 if ((pm = pmc_find_pmc_descriptor_in_process(po, pmcid)) == NULL)
2614 PMCDBG2(PMC,FND,2, "find-pmc id=%d -> pmc=%p", pmcid, pm);
2625 pmc_start(struct pmc *pm)
2628 struct pmc_owner *po;
2629 struct pmc_binding pb;
2630 struct pmc_classdep *pcd;
2631 int adjri, error, cpu, ri;
2634 ("[pmc,%d] null pm", __LINE__));
2636 mode = PMC_TO_MODE(pm);
2637 ri = PMC_TO_ROWINDEX(pm);
2638 pcd = pmc_ri_to_classdep(md, ri, &adjri);
2642 PMCDBG3(PMC,OPS,1, "start pmc=%p mode=%d ri=%d", pm, mode, ri);
2647 * Disallow PMCSTART if a logfile is required but has not been
2650 if ((pm->pm_flags & PMC_F_NEEDS_LOGFILE) &&
2651 (po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
2652 return (EDOOFUS); /* programming error */
2655 * If this is a sampling mode PMC, log mapping information for
2656 * the kernel modules that are currently loaded.
2658 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
2659 pmc_log_kernel_mappings(pm);
2661 if (PMC_IS_VIRTUAL_MODE(mode)) {
2664 * If a PMCATTACH has never been done on this PMC,
2665 * attach it to its owner process.
2668 if (LIST_EMPTY(&pm->pm_targets))
2669 error = (pm->pm_flags & PMC_F_ATTACH_DONE) ? ESRCH :
2670 pmc_attach_process(po->po_owner, pm);
2673 * If the PMC is attached to its owner, then force a context
2674 * switch to ensure that the MD state gets set correctly.
2678 pm->pm_state = PMC_STATE_RUNNING;
2679 if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER)
2680 pmc_force_context_switch();
2688 * A system-wide PMC.
2690 * Add the owner to the global list if this is a system-wide
2694 if (mode == PMC_MODE_SS) {
2695 if (po->po_sscount == 0) {
2696 LIST_INSERT_HEAD(&pmc_ss_owners, po, po_ssnext);
2697 atomic_add_rel_int(&pmc_ss_count, 1);
2698 PMCDBG1(PMC,OPS,1, "po=%p in global list", po);
2703 * Log mapping information for all existing processes in the
2704 * system. Subsequent mappings are logged as they happen;
2705 * see pmc_process_mmap().
2707 if (po->po_logprocmaps == 0) {
2708 pmc_log_all_process_mappings(po);
2709 po->po_logprocmaps = 1;
2714 * Move to the CPU associated with this
2715 * PMC, and start the hardware.
2718 pmc_save_cpu_binding(&pb);
2720 cpu = PMC_TO_CPU(pm);
2722 if (!pmc_cpu_is_active(cpu))
2725 pmc_select_cpu(cpu);
2728 * global PMCs are configured at allocation time
2729 * so write out the initial value and start the PMC.
2732 pm->pm_state = PMC_STATE_RUNNING;
2735 if ((error = pcd->pcd_write_pmc(cpu, adjri,
2736 PMC_IS_SAMPLING_MODE(mode) ?
2737 pm->pm_sc.pm_reloadcount :
2738 pm->pm_sc.pm_initial)) == 0) {
2739 /* If a sampling mode PMC, reset stalled state. */
2740 if (PMC_IS_SAMPLING_MODE(mode))
2741 CPU_CLR_ATOMIC(cpu, &pm->pm_stalled);
2743 /* Indicate that we desire this to run. Start it. */
2744 CPU_SET_ATOMIC(cpu, &pm->pm_cpustate);
2745 error = pcd->pcd_start_pmc(cpu, adjri);
2749 pmc_restore_cpu_binding(&pb);
2759 pmc_stop(struct pmc *pm)
2761 struct pmc_owner *po;
2762 struct pmc_binding pb;
2763 struct pmc_classdep *pcd;
2764 int adjri, cpu, error, ri;
2766 KASSERT(pm != NULL, ("[pmc,%d] null pmc", __LINE__));
2768 PMCDBG3(PMC,OPS,1, "stop pmc=%p mode=%d ri=%d", pm,
2769 PMC_TO_MODE(pm), PMC_TO_ROWINDEX(pm));
2771 pm->pm_state = PMC_STATE_STOPPED;
2774 * If the PMC is a virtual mode one, changing the state to
2775 * non-RUNNING is enough to ensure that the PMC never gets
2778 * If this PMC is current running on a CPU, then it will
2779 * handled correctly at the time its target process is context
2783 if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)))
2787 * A system-mode PMC. Move to the CPU associated with
2788 * this PMC, and stop the hardware. We update the
2789 * 'initial count' so that a subsequent PMCSTART will
2790 * resume counting from the current hardware count.
2793 pmc_save_cpu_binding(&pb);
2795 cpu = PMC_TO_CPU(pm);
2797 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
2798 ("[pmc,%d] illegal cpu=%d", __LINE__, cpu));
2800 if (!pmc_cpu_is_active(cpu))
2803 pmc_select_cpu(cpu);
2805 ri = PMC_TO_ROWINDEX(pm);
2806 pcd = pmc_ri_to_classdep(md, ri, &adjri);
2808 CPU_CLR_ATOMIC(cpu, &pm->pm_cpustate);
2810 if ((error = pcd->pcd_stop_pmc(cpu, adjri)) == 0)
2811 error = pcd->pcd_read_pmc(cpu, adjri, &pm->pm_sc.pm_initial);
2814 pmc_restore_cpu_binding(&pb);
2818 /* remove this owner from the global list of SS PMC owners */
2819 if (PMC_TO_MODE(pm) == PMC_MODE_SS) {
2821 if (po->po_sscount == 0) {
2822 atomic_subtract_rel_int(&pmc_ss_count, 1);
2823 LIST_REMOVE(po, po_ssnext);
2824 PMCDBG1(PMC,OPS,2,"po=%p removed from global list", po);
2833 static const char *pmc_op_to_name[] = {
2835 #define __PMC_OP(N, D) #N ,
2842 * The syscall interface
2845 #define PMC_GET_SX_XLOCK(...) do { \
2846 sx_xlock(&pmc_sx); \
2847 if (pmc_hook == NULL) { \
2848 sx_xunlock(&pmc_sx); \
2849 return __VA_ARGS__; \
2853 #define PMC_DOWNGRADE_SX() do { \
2854 sx_downgrade(&pmc_sx); \
2855 is_sx_downgraded = 1; \
2859 pmc_syscall_handler(struct thread *td, void *syscall_args)
2861 int error, is_sx_downgraded, is_sx_locked, op;
2862 struct pmc_syscall_args *c;
2865 PMC_GET_SX_XLOCK(ENOSYS);
2869 is_sx_downgraded = 0;
2872 c = (struct pmc_syscall_args *) syscall_args;
2877 PMCDBG3(MOD,PMS,1, "syscall op=%d \"%s\" arg=%p", op,
2878 pmc_op_to_name[op], arg);
2881 atomic_add_int(&pmc_stats.pm_syscalls, 1);
2888 * Configure a log file.
2890 * XXX This OP will be reworked.
2893 case PMC_OP_CONFIGURELOG:
2897 struct pmc_owner *po;
2898 struct pmc_op_configurelog cl;
2900 sx_assert(&pmc_sx, SX_XLOCKED);
2902 if ((error = copyin(arg, &cl, sizeof(cl))) != 0)
2905 /* mark this process as owning a log file */
2907 if ((po = pmc_find_owner_descriptor(p)) == NULL)
2908 if ((po = pmc_allocate_owner_descriptor(p)) == NULL) {
2914 * If a valid fd was passed in, try to configure that,
2915 * otherwise if 'fd' was less than zero and there was
2916 * a log file configured, flush its buffers and
2919 if (cl.pm_logfd >= 0) {
2920 sx_xunlock(&pmc_sx);
2922 error = pmclog_configure_log(md, po, cl.pm_logfd);
2923 } else if (po->po_flags & PMC_PO_OWNS_LOGFILE) {
2924 pmclog_process_closelog(po);
2925 error = pmclog_close(po);
2927 LIST_FOREACH(pm, &po->po_pmcs, pm_next)
2928 if (pm->pm_flags & PMC_F_NEEDS_LOGFILE &&
2929 pm->pm_state == PMC_STATE_RUNNING)
2931 error = pmclog_deconfigure_log(po);
2945 case PMC_OP_FLUSHLOG:
2947 struct pmc_owner *po;
2949 sx_assert(&pmc_sx, SX_XLOCKED);
2951 if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) {
2956 error = pmclog_flush(po);
2964 case PMC_OP_CLOSELOG:
2966 struct pmc_owner *po;
2968 sx_assert(&pmc_sx, SX_XLOCKED);
2970 if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) {
2975 error = pmclog_close(po);
2980 * Retrieve hardware configuration.
2983 case PMC_OP_GETCPUINFO: /* CPU information */
2985 struct pmc_op_getcpuinfo gci;
2986 struct pmc_classinfo *pci;
2987 struct pmc_classdep *pcd;
2990 gci.pm_cputype = md->pmd_cputype;
2991 gci.pm_ncpu = pmc_cpu_max();
2992 gci.pm_npmc = md->pmd_npmc;
2993 gci.pm_nclass = md->pmd_nclass;
2994 pci = gci.pm_classes;
2995 pcd = md->pmd_classdep;
2996 for (cl = 0; cl < md->pmd_nclass; cl++, pci++, pcd++) {
2997 pci->pm_caps = pcd->pcd_caps;
2998 pci->pm_class = pcd->pcd_class;
2999 pci->pm_width = pcd->pcd_width;
3000 pci->pm_num = pcd->pcd_num;
3002 error = copyout(&gci, arg, sizeof(gci));
3007 * Retrieve soft events list.
3009 case PMC_OP_GETDYNEVENTINFO:
3013 struct pmc_op_getdyneventinfo *gei;
3014 struct pmc_dyn_event_descr dev;
3015 struct pmc_soft *ps;
3018 sx_assert(&pmc_sx, SX_LOCKED);
3020 gei = (struct pmc_op_getdyneventinfo *) arg;
3022 if ((error = copyin(&gei->pm_class, &cl, sizeof(cl))) != 0)
3025 /* Only SOFT class is dynamic. */
3026 if (cl != PMC_CLASS_SOFT) {
3032 for (ev = PMC_EV_SOFT_FIRST; (int)ev <= PMC_EV_SOFT_LAST; ev++) {
3033 ps = pmc_soft_ev_acquire(ev);
3036 bcopy(&ps->ps_ev, &dev, sizeof(dev));
3037 pmc_soft_ev_release(ps);
3039 error = copyout(&dev,
3040 &gei->pm_events[nevent],
3041 sizeof(struct pmc_dyn_event_descr));
3049 error = copyout(&nevent, &gei->pm_nevent,
3055 * Get module statistics
3058 case PMC_OP_GETDRIVERSTATS:
3060 struct pmc_op_getdriverstats gms;
3062 bcopy(&pmc_stats, &gms, sizeof(gms));
3063 error = copyout(&gms, arg, sizeof(gms));
3069 * Retrieve module version number
3072 case PMC_OP_GETMODULEVERSION:
3076 /* retrieve the client's idea of the ABI version */
3077 if ((error = copyin(arg, &cv, sizeof(uint32_t))) != 0)
3079 /* don't service clients newer than our driver */
3081 if ((cv & 0xFFFF0000) > (modv & 0xFFFF0000)) {
3082 error = EPROGMISMATCH;
3085 error = copyout(&modv, arg, sizeof(int));
3091 * Retrieve the state of all the PMCs on a given
3095 case PMC_OP_GETPMCINFO:
3099 size_t pmcinfo_size;
3100 uint32_t cpu, n, npmc;
3101 struct pmc_owner *po;
3102 struct pmc_binding pb;
3103 struct pmc_classdep *pcd;
3104 struct pmc_info *p, *pmcinfo;
3105 struct pmc_op_getpmcinfo *gpi;
3109 gpi = (struct pmc_op_getpmcinfo *) arg;
3111 if ((error = copyin(&gpi->pm_cpu, &cpu, sizeof(cpu))) != 0)
3114 if (cpu >= pmc_cpu_max()) {
3119 if (!pmc_cpu_is_active(cpu)) {
3124 /* switch to CPU 'cpu' */
3125 pmc_save_cpu_binding(&pb);
3126 pmc_select_cpu(cpu);
3128 npmc = md->pmd_npmc;
3130 pmcinfo_size = npmc * sizeof(struct pmc_info);
3131 pmcinfo = malloc(pmcinfo_size, M_PMC, M_WAITOK);
3135 for (n = 0; n < md->pmd_npmc; n++, p++) {
3137 pcd = pmc_ri_to_classdep(md, n, &ari);
3139 KASSERT(pcd != NULL,
3140 ("[pmc,%d] null pcd ri=%d", __LINE__, n));
3142 if ((error = pcd->pcd_describe(cpu, ari, p, &pm)) != 0)
3145 if (PMC_ROW_DISP_IS_STANDALONE(n))
3146 p->pm_rowdisp = PMC_DISP_STANDALONE;
3147 else if (PMC_ROW_DISP_IS_THREAD(n))
3148 p->pm_rowdisp = PMC_DISP_THREAD;
3150 p->pm_rowdisp = PMC_DISP_FREE;
3152 p->pm_ownerpid = -1;
3154 if (pm == NULL) /* no PMC associated */
3159 KASSERT(po->po_owner != NULL,
3160 ("[pmc,%d] pmc_owner had a null proc pointer",
3163 p->pm_ownerpid = po->po_owner->p_pid;
3164 p->pm_mode = PMC_TO_MODE(pm);
3165 p->pm_event = pm->pm_event;
3166 p->pm_flags = pm->pm_flags;
3168 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
3170 pm->pm_sc.pm_reloadcount;
3173 pmc_restore_cpu_binding(&pb);
3175 /* now copy out the PMC info collected */
3177 error = copyout(pmcinfo, &gpi->pm_pmcs, pmcinfo_size);
3179 free(pmcinfo, M_PMC);
3185 * Set the administrative state of a PMC. I.e. whether
3186 * the PMC is to be used or not.
3189 case PMC_OP_PMCADMIN:
3192 enum pmc_state request;
3195 struct pmc_op_pmcadmin pma;
3196 struct pmc_binding pb;
3198 sx_assert(&pmc_sx, SX_XLOCKED);
3200 KASSERT(td == curthread,
3201 ("[pmc,%d] td != curthread", __LINE__));
3203 error = priv_check(td, PRIV_PMC_MANAGE);
3207 if ((error = copyin(arg, &pma, sizeof(pma))) != 0)
3212 if (cpu < 0 || cpu >= (int) pmc_cpu_max()) {
3217 if (!pmc_cpu_is_active(cpu)) {
3222 request = pma.pm_state;
3224 if (request != PMC_STATE_DISABLED &&
3225 request != PMC_STATE_FREE) {
3230 ri = pma.pm_pmc; /* pmc id == row index */
3231 if (ri < 0 || ri >= (int) md->pmd_npmc) {
3237 * We can't disable a PMC with a row-index allocated
3238 * for process virtual PMCs.
3241 if (PMC_ROW_DISP_IS_THREAD(ri) &&
3242 request == PMC_STATE_DISABLED) {
3248 * otherwise, this PMC on this CPU is either free or
3249 * in system-wide mode.
3252 pmc_save_cpu_binding(&pb);
3253 pmc_select_cpu(cpu);
3256 phw = pc->pc_hwpmcs[ri];
3259 * XXX do we need some kind of 'forced' disable?
3262 if (phw->phw_pmc == NULL) {
3263 if (request == PMC_STATE_DISABLED &&
3264 (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED)) {
3265 phw->phw_state &= ~PMC_PHW_FLAG_IS_ENABLED;
3266 PMC_MARK_ROW_STANDALONE(ri);
3267 } else if (request == PMC_STATE_FREE &&
3268 (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0) {
3269 phw->phw_state |= PMC_PHW_FLAG_IS_ENABLED;
3270 PMC_UNMARK_ROW_STANDALONE(ri);
3272 /* other cases are a no-op */
3276 pmc_restore_cpu_binding(&pb);
3285 case PMC_OP_PMCALLOCATE:
3293 struct pmc_binding pb;
3294 struct pmc_classdep *pcd;
3295 struct pmc_op_pmcallocate pa;
3297 if ((error = copyin(arg, &pa, sizeof(pa))) != 0)
3304 if ((mode != PMC_MODE_SS && mode != PMC_MODE_SC &&
3305 mode != PMC_MODE_TS && mode != PMC_MODE_TC) ||
3306 (cpu != (u_int) PMC_CPU_ANY && cpu >= pmc_cpu_max())) {
3312 * Virtual PMCs should only ask for a default CPU.
3313 * System mode PMCs need to specify a non-default CPU.
3316 if ((PMC_IS_VIRTUAL_MODE(mode) && cpu != (u_int) PMC_CPU_ANY) ||
3317 (PMC_IS_SYSTEM_MODE(mode) && cpu == (u_int) PMC_CPU_ANY)) {
3323 * Check that an inactive CPU is not being asked for.
3326 if (PMC_IS_SYSTEM_MODE(mode) && !pmc_cpu_is_active(cpu)) {
3332 * Refuse an allocation for a system-wide PMC if this
3333 * process has been jailed, or if this process lacks
3334 * super-user credentials and the sysctl tunable
3335 * 'security.bsd.unprivileged_syspmcs' is zero.
3338 if (PMC_IS_SYSTEM_MODE(mode)) {
3339 if (jailed(curthread->td_ucred)) {
3343 if (!pmc_unprivileged_syspmcs) {
3344 error = priv_check(curthread,
3352 * Look for valid values for 'pm_flags'
3355 if ((pa.pm_flags & ~(PMC_F_DESCENDANTS | PMC_F_LOG_PROCCSW |
3356 PMC_F_LOG_PROCEXIT | PMC_F_CALLCHAIN)) != 0) {
3361 /* process logging options are not allowed for system PMCs */
3362 if (PMC_IS_SYSTEM_MODE(mode) && (pa.pm_flags &
3363 (PMC_F_LOG_PROCCSW | PMC_F_LOG_PROCEXIT))) {
3369 * All sampling mode PMCs need to be able to interrupt the
3372 if (PMC_IS_SAMPLING_MODE(mode))
3373 caps |= PMC_CAP_INTERRUPT;
3375 /* A valid class specifier should have been passed in. */
3376 for (n = 0; n < md->pmd_nclass; n++)
3377 if (md->pmd_classdep[n].pcd_class == pa.pm_class)
3379 if (n == md->pmd_nclass) {
3384 /* The requested PMC capabilities should be feasible. */
3385 if ((md->pmd_classdep[n].pcd_caps & caps) != caps) {
3390 PMCDBG4(PMC,ALL,2, "event=%d caps=0x%x mode=%d cpu=%d",
3391 pa.pm_ev, caps, mode, cpu);
3393 pmc = pmc_allocate_pmc_descriptor();
3394 pmc->pm_id = PMC_ID_MAKE_ID(cpu,pa.pm_mode,pa.pm_class,
3396 pmc->pm_event = pa.pm_ev;
3397 pmc->pm_state = PMC_STATE_FREE;
3398 pmc->pm_caps = caps;
3399 pmc->pm_flags = pa.pm_flags;
3401 /* switch thread to CPU 'cpu' */
3402 pmc_save_cpu_binding(&pb);
3404 #define PMC_IS_SHAREABLE_PMC(cpu, n) \
3405 (pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_state & \
3406 PMC_PHW_FLAG_IS_SHAREABLE)
3407 #define PMC_IS_UNALLOCATED(cpu, n) \
3408 (pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_pmc == NULL)
3410 if (PMC_IS_SYSTEM_MODE(mode)) {
3411 pmc_select_cpu(cpu);
3412 for (n = 0; n < (int) md->pmd_npmc; n++) {
3413 pcd = pmc_ri_to_classdep(md, n, &adjri);
3414 if (pmc_can_allocate_row(n, mode) == 0 &&
3415 pmc_can_allocate_rowindex(
3416 curthread->td_proc, n, cpu) == 0 &&
3417 (PMC_IS_UNALLOCATED(cpu, n) ||
3418 PMC_IS_SHAREABLE_PMC(cpu, n)) &&
3419 pcd->pcd_allocate_pmc(cpu, adjri, pmc,
3424 /* Process virtual mode */
3425 for (n = 0; n < (int) md->pmd_npmc; n++) {
3426 pcd = pmc_ri_to_classdep(md, n, &adjri);
3427 if (pmc_can_allocate_row(n, mode) == 0 &&
3428 pmc_can_allocate_rowindex(
3429 curthread->td_proc, n,
3430 PMC_CPU_ANY) == 0 &&
3431 pcd->pcd_allocate_pmc(curthread->td_oncpu,
3432 adjri, pmc, &pa) == 0)
3437 #undef PMC_IS_UNALLOCATED
3438 #undef PMC_IS_SHAREABLE_PMC
3440 pmc_restore_cpu_binding(&pb);
3442 if (n == (int) md->pmd_npmc) {
3443 pmc_destroy_pmc_descriptor(pmc);
3449 /* Fill in the correct value in the ID field */
3450 pmc->pm_id = PMC_ID_MAKE_ID(cpu,mode,pa.pm_class,n);
3452 PMCDBG5(PMC,ALL,2, "ev=%d class=%d mode=%d n=%d -> pmcid=%x",
3453 pmc->pm_event, pa.pm_class, mode, n, pmc->pm_id);
3455 /* Process mode PMCs with logging enabled need log files */
3456 if (pmc->pm_flags & (PMC_F_LOG_PROCEXIT | PMC_F_LOG_PROCCSW))
3457 pmc->pm_flags |= PMC_F_NEEDS_LOGFILE;
3459 /* All system mode sampling PMCs require a log file */
3460 if (PMC_IS_SAMPLING_MODE(mode) && PMC_IS_SYSTEM_MODE(mode))
3461 pmc->pm_flags |= PMC_F_NEEDS_LOGFILE;
3464 * Configure global pmc's immediately
3467 if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pmc))) {
3469 pmc_save_cpu_binding(&pb);
3470 pmc_select_cpu(cpu);
3472 phw = pmc_pcpu[cpu]->pc_hwpmcs[n];
3473 pcd = pmc_ri_to_classdep(md, n, &adjri);
3475 if ((phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0 ||
3476 (error = pcd->pcd_config_pmc(cpu, adjri, pmc)) != 0) {
3477 (void) pcd->pcd_release_pmc(cpu, adjri, pmc);
3478 pmc_destroy_pmc_descriptor(pmc);
3480 pmc_restore_cpu_binding(&pb);
3485 pmc_restore_cpu_binding(&pb);
3488 pmc->pm_state = PMC_STATE_ALLOCATED;
3491 * mark row disposition
3494 if (PMC_IS_SYSTEM_MODE(mode))
3495 PMC_MARK_ROW_STANDALONE(n);
3497 PMC_MARK_ROW_THREAD(n);
3500 * Register this PMC with the current thread as its owner.
3504 pmc_register_owner(curthread->td_proc, pmc)) != 0) {
3505 pmc_release_pmc_descriptor(pmc);
3506 pmc_destroy_pmc_descriptor(pmc);
3512 * Return the allocated index.
3515 pa.pm_pmcid = pmc->pm_id;
3517 error = copyout(&pa, arg, sizeof(pa));
3523 * Attach a PMC to a process.
3526 case PMC_OP_PMCATTACH:
3530 struct pmc_op_pmcattach a;
3532 sx_assert(&pmc_sx, SX_XLOCKED);
3534 if ((error = copyin(arg, &a, sizeof(a))) != 0)
3540 } else if (a.pm_pid == 0)
3541 a.pm_pid = td->td_proc->p_pid;
3543 if ((error = pmc_find_pmc(a.pm_pmc, &pm)) != 0)
3546 if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm))) {
3551 /* PMCs may be (re)attached only when allocated or stopped */
3552 if (pm->pm_state == PMC_STATE_RUNNING) {
3555 } else if (pm->pm_state != PMC_STATE_ALLOCATED &&
3556 pm->pm_state != PMC_STATE_STOPPED) {
3562 if ((p = pfind(a.pm_pid)) == NULL) {
3568 * Ignore processes that are working on exiting.
3570 if (p->p_flag & P_WEXIT) {
3572 PROC_UNLOCK(p); /* pfind() returns a locked process */
3577 * we are allowed to attach a PMC to a process if
3580 error = p_candebug(curthread, p);
3585 error = pmc_attach_process(p, pm);
3591 * Detach an attached PMC from a process.
3594 case PMC_OP_PMCDETACH:
3598 struct pmc_op_pmcattach a;
3600 if ((error = copyin(arg, &a, sizeof(a))) != 0)
3606 } else if (a.pm_pid == 0)
3607 a.pm_pid = td->td_proc->p_pid;
3609 if ((error = pmc_find_pmc(a.pm_pmc, &pm)) != 0)
3612 if ((p = pfind(a.pm_pid)) == NULL) {
3618 * Treat processes that are in the process of exiting
3619 * as if they were not present.
3622 if (p->p_flag & P_WEXIT)
3625 PROC_UNLOCK(p); /* pfind() returns a locked process */
3628 error = pmc_detach_process(p, pm);
3634 * Retrieve the MSR number associated with the counter
3635 * 'pmc_id'. This allows processes to directly use RDPMC
3636 * instructions to read their PMCs, without the overhead of a
3640 case PMC_OP_PMCGETMSR:
3644 struct pmc_target *pt;
3645 struct pmc_op_getmsr gm;
3646 struct pmc_classdep *pcd;
3650 if ((error = copyin(arg, &gm, sizeof(gm))) != 0)
3653 if ((error = pmc_find_pmc(gm.pm_pmcid, &pm)) != 0)
3657 * The allocated PMC has to be a process virtual PMC,
3658 * i.e., of type MODE_T[CS]. Global PMCs can only be
3659 * read using the PMCREAD operation since they may be
3660 * allocated on a different CPU than the one we could
3661 * be running on at the time of the RDPMC instruction.
3663 * The GETMSR operation is not allowed for PMCs that
3664 * are inherited across processes.
3667 if (!PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)) ||
3668 (pm->pm_flags & PMC_F_DESCENDANTS)) {
3674 * It only makes sense to use a RDPMC (or its
3675 * equivalent instruction on non-x86 architectures) on
3676 * a process that has allocated and attached a PMC to
3677 * itself. Conversely the PMC is only allowed to have
3678 * one process attached to it -- its owner.
3681 if ((pt = LIST_FIRST(&pm->pm_targets)) == NULL ||
3682 LIST_NEXT(pt, pt_next) != NULL ||
3683 pt->pt_process->pp_proc != pm->pm_owner->po_owner) {
3688 ri = PMC_TO_ROWINDEX(pm);
3689 pcd = pmc_ri_to_classdep(md, ri, &adjri);
3691 /* PMC class has no 'GETMSR' support */
3692 if (pcd->pcd_get_msr == NULL) {
3697 if ((error = (*pcd->pcd_get_msr)(adjri, &gm.pm_msr)) < 0)
3700 if ((error = copyout(&gm, arg, sizeof(gm))) < 0)
3704 * Mark our process as using MSRs. Update machine
3705 * state using a forced context switch.
3708 pt->pt_process->pp_flags |= PMC_PP_ENABLE_MSR_ACCESS;
3709 pmc_force_context_switch();
3715 * Release an allocated PMC
3718 case PMC_OP_PMCRELEASE:
3722 struct pmc_owner *po;
3723 struct pmc_op_simple sp;
3726 * Find PMC pointer for the named PMC.
3728 * Use pmc_release_pmc_descriptor() to switch off the
3729 * PMC, remove all its target threads, and remove the
3730 * PMC from its owner's list.
3732 * Remove the owner record if this is the last PMC
3738 if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
3741 pmcid = sp.pm_pmcid;
3743 if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
3747 pmc_release_pmc_descriptor(pm);
3748 pmc_maybe_remove_owner(po);
3749 pmc_destroy_pmc_descriptor(pm);
3755 * Read and/or write a PMC.
3763 pmc_value_t oldvalue;
3764 struct pmc_binding pb;
3765 struct pmc_op_pmcrw prw;
3766 struct pmc_classdep *pcd;
3767 struct pmc_op_pmcrw *pprw;
3771 if ((error = copyin(arg, &prw, sizeof(prw))) != 0)
3775 PMCDBG2(PMC,OPS,1, "rw id=%d flags=0x%x", prw.pm_pmcid,
3778 /* must have at least one flag set */
3779 if ((prw.pm_flags & (PMC_F_OLDVALUE|PMC_F_NEWVALUE)) == 0) {
3784 /* locate pmc descriptor */
3785 if ((error = pmc_find_pmc(prw.pm_pmcid, &pm)) != 0)
3788 /* Can't read a PMC that hasn't been started. */
3789 if (pm->pm_state != PMC_STATE_ALLOCATED &&
3790 pm->pm_state != PMC_STATE_STOPPED &&
3791 pm->pm_state != PMC_STATE_RUNNING) {
3796 /* writing a new value is allowed only for 'STOPPED' pmcs */
3797 if (pm->pm_state == PMC_STATE_RUNNING &&
3798 (prw.pm_flags & PMC_F_NEWVALUE)) {
3803 if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) {
3806 * If this PMC is attached to its owner (i.e.,
3807 * the process requesting this operation) and
3808 * is running, then attempt to get an
3809 * upto-date reading from hardware for a READ.
3810 * Writes are only allowed when the PMC is
3811 * stopped, so only update the saved value
3814 * If the PMC is not running, or is not
3815 * attached to its owner, read/write to the
3819 ri = PMC_TO_ROWINDEX(pm);
3820 pcd = pmc_ri_to_classdep(md, ri, &adjri);
3822 mtx_pool_lock_spin(pmc_mtxpool, pm);
3823 cpu = curthread->td_oncpu;
3825 if (prw.pm_flags & PMC_F_OLDVALUE) {
3826 if ((pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) &&
3827 (pm->pm_state == PMC_STATE_RUNNING))
3828 error = (*pcd->pcd_read_pmc)(cpu, adjri,
3831 oldvalue = pm->pm_gv.pm_savedvalue;
3833 if (prw.pm_flags & PMC_F_NEWVALUE)
3834 pm->pm_gv.pm_savedvalue = prw.pm_value;
3836 mtx_pool_unlock_spin(pmc_mtxpool, pm);
3838 } else { /* System mode PMCs */
3839 cpu = PMC_TO_CPU(pm);
3840 ri = PMC_TO_ROWINDEX(pm);
3841 pcd = pmc_ri_to_classdep(md, ri, &adjri);
3843 if (!pmc_cpu_is_active(cpu)) {
3848 /* move this thread to CPU 'cpu' */
3849 pmc_save_cpu_binding(&pb);
3850 pmc_select_cpu(cpu);
3853 /* save old value */
3854 if (prw.pm_flags & PMC_F_OLDVALUE)
3855 if ((error = (*pcd->pcd_read_pmc)(cpu, adjri,
3858 /* write out new value */
3859 if (prw.pm_flags & PMC_F_NEWVALUE)
3860 error = (*pcd->pcd_write_pmc)(cpu, adjri,
3864 pmc_restore_cpu_binding(&pb);
3869 pprw = (struct pmc_op_pmcrw *) arg;
3872 if (prw.pm_flags & PMC_F_NEWVALUE)
3873 PMCDBG3(PMC,OPS,2, "rw id=%d new %jx -> old %jx",
3874 ri, prw.pm_value, oldvalue);
3875 else if (prw.pm_flags & PMC_F_OLDVALUE)
3876 PMCDBG2(PMC,OPS,2, "rw id=%d -> old %jx", ri, oldvalue);
3879 /* return old value if requested */
3880 if (prw.pm_flags & PMC_F_OLDVALUE)
3881 if ((error = copyout(&oldvalue, &pprw->pm_value,
3882 sizeof(prw.pm_value))))
3890 * Set the sampling rate for a sampling mode PMC and the
3891 * initial count for a counting mode PMC.
3894 case PMC_OP_PMCSETCOUNT:
3897 struct pmc_op_pmcsetcount sc;
3901 if ((error = copyin(arg, &sc, sizeof(sc))) != 0)
3904 if ((error = pmc_find_pmc(sc.pm_pmcid, &pm)) != 0)
3907 if (pm->pm_state == PMC_STATE_RUNNING) {
3912 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
3913 pm->pm_sc.pm_reloadcount = sc.pm_count;
3915 pm->pm_sc.pm_initial = sc.pm_count;
3924 case PMC_OP_PMCSTART:
3928 struct pmc_op_simple sp;
3930 sx_assert(&pmc_sx, SX_XLOCKED);
3932 if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
3935 pmcid = sp.pm_pmcid;
3937 if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
3940 KASSERT(pmcid == pm->pm_id,
3941 ("[pmc,%d] pmcid %x != id %x", __LINE__,
3944 if (pm->pm_state == PMC_STATE_RUNNING) /* already running */
3946 else if (pm->pm_state != PMC_STATE_STOPPED &&
3947 pm->pm_state != PMC_STATE_ALLOCATED) {
3952 error = pmc_start(pm);
3961 case PMC_OP_PMCSTOP:
3965 struct pmc_op_simple sp;
3969 if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
3972 pmcid = sp.pm_pmcid;
3975 * Mark the PMC as inactive and invoke the MD stop
3976 * routines if needed.
3979 if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
3982 KASSERT(pmcid == pm->pm_id,
3983 ("[pmc,%d] pmc id %x != pmcid %x", __LINE__,
3986 if (pm->pm_state == PMC_STATE_STOPPED) /* already stopped */
3988 else if (pm->pm_state != PMC_STATE_RUNNING) {
3993 error = pmc_stop(pm);
3999 * Write a user supplied value to the log file.
4002 case PMC_OP_WRITELOG:
4004 struct pmc_op_writelog wl;
4005 struct pmc_owner *po;
4009 if ((error = copyin(arg, &wl, sizeof(wl))) != 0)
4012 if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) {
4017 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) {
4022 error = pmclog_process_userlog(po, &wl);
4032 if (is_sx_locked != 0) {
4033 if (is_sx_downgraded)
4034 sx_sunlock(&pmc_sx);
4036 sx_xunlock(&pmc_sx);
4040 atomic_add_int(&pmc_stats.pm_syscall_errors, 1);
4053 * Mark the thread as needing callchain capture and post an AST. The
4054 * actual callchain capture will be done in a context where it is safe
4055 * to take page faults.
4059 pmc_post_callchain_callback(void)
4066 * If there is multiple PMCs for the same interrupt ignore new post
4068 if (td->td_pflags & TDP_CALLCHAIN)
4072 * Mark this thread as needing callchain capture.
4073 * `td->td_pflags' will be safe to touch because this thread
4074 * was in user space when it was interrupted.
4076 td->td_pflags |= TDP_CALLCHAIN;
4079 * Don't let this thread migrate between CPUs until callchain
4080 * capture completes.
4088 * Interrupt processing.
4090 * Find a free slot in the per-cpu array of samples and capture the
4091 * current callchain there. If a sample was successfully added, a bit
4092 * is set in mask 'pmc_cpumask' denoting that the DO_SAMPLES hook
4093 * needs to be invoked from the clock handler.
4095 * This function is meant to be called from an NMI handler. It cannot
4096 * use any of the locking primitives supplied by the OS.
4100 pmc_process_interrupt(int cpu, int ring, struct pmc *pm, struct trapframe *tf,
4103 int error, callchaindepth;
4105 struct pmc_sample *ps;
4106 struct pmc_samplebuffer *psb;
4111 * Allocate space for a sample buffer.
4113 psb = pmc_pcpu[cpu]->pc_sb[ring];
4116 if (ps->ps_nsamples) { /* in use, reader hasn't caught up */
4117 CPU_SET_ATOMIC(cpu, &pm->pm_stalled);
4118 atomic_add_int(&pmc_stats.pm_intr_bufferfull, 1);
4119 PMCDBG6(SAM,INT,1,"(spc) cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d",
4120 cpu, pm, (void *) tf, inuserspace,
4121 (int) (psb->ps_write - psb->ps_samples),
4122 (int) (psb->ps_read - psb->ps_samples));
4129 /* Fill in entry. */
4130 PMCDBG6(SAM,INT,1,"cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d", cpu, pm,
4131 (void *) tf, inuserspace,
4132 (int) (psb->ps_write - psb->ps_samples),
4133 (int) (psb->ps_read - psb->ps_samples));
4135 KASSERT(pm->pm_runcount >= 0,
4136 ("[pmc,%d] pm=%p runcount %d", __LINE__, (void *) pm,
4139 atomic_add_rel_int(&pm->pm_runcount, 1); /* hold onto PMC */
4142 if ((td = curthread) && td->td_proc)
4143 ps->ps_pid = td->td_proc->p_pid;
4148 ps->ps_flags = inuserspace ? PMC_CC_F_USERSPACE : 0;
4150 callchaindepth = (pm->pm_flags & PMC_F_CALLCHAIN) ?
4151 pmc_callchaindepth : 1;
4153 if (callchaindepth == 1)
4154 ps->ps_pc[0] = PMC_TRAPFRAME_TO_PC(tf);
4157 * Kernel stack traversals can be done immediately,
4158 * while we defer to an AST for user space traversals.
4162 pmc_save_kernel_callchain(ps->ps_pc,
4163 callchaindepth, tf);
4165 pmc_post_callchain_callback();
4166 callchaindepth = PMC_SAMPLE_INUSE;
4170 ps->ps_nsamples = callchaindepth; /* mark entry as in use */
4172 /* increment write pointer, modulo ring buffer size */
4174 if (ps == psb->ps_fence)
4175 psb->ps_write = psb->ps_samples;
4180 /* mark CPU as needing processing */
4181 if (callchaindepth != PMC_SAMPLE_INUSE)
4182 CPU_SET_ATOMIC(cpu, &pmc_cpumask);
4188 * Capture a user call chain. This function will be called from ast()
4189 * before control returns to userland and before the process gets
4194 pmc_capture_user_callchain(int cpu, int ring, struct trapframe *tf)
4198 struct pmc_sample *ps, *ps_end;
4199 struct pmc_samplebuffer *psb;
4205 psb = pmc_pcpu[cpu]->pc_sb[ring];
4208 KASSERT(td->td_pflags & TDP_CALLCHAIN,
4209 ("[pmc,%d] Retrieving callchain for thread that doesn't want it",
4218 * Iterate through all deferred callchain requests.
4219 * Walk from the current read pointer to the current
4224 ps_end = psb->ps_write;
4227 if (ps->ps_pmc->pm_state != PMC_STATE_RUNNING)
4230 if (ps->ps_nsamples != PMC_SAMPLE_INUSE)
4232 if (ps->ps_td != td)
4235 KASSERT(ps->ps_cpu == cpu,
4236 ("[pmc,%d] cpu mismatch ps_cpu=%d pcpu=%d", __LINE__,
4237 ps->ps_cpu, PCPU_GET(cpuid)));
4241 KASSERT(pm->pm_flags & PMC_F_CALLCHAIN,
4242 ("[pmc,%d] Retrieving callchain for PMC that doesn't "
4243 "want it", __LINE__));
4245 KASSERT(pm->pm_runcount > 0,
4246 ("[pmc,%d] runcount %d", __LINE__, pm->pm_runcount));
4249 * Retrieve the callchain and mark the sample buffer
4250 * as 'processable' by the timer tick sweep code.
4252 ps->ps_nsamples = pmc_save_user_callchain(ps->ps_pc,
4253 pmc_callchaindepth, tf);
4260 /* increment the pointer, modulo sample ring size */
4261 if (++ps == psb->ps_fence)
4262 ps = psb->ps_samples;
4263 } while (ps != ps_end);
4265 KASSERT(ncallchains > 0 || nfree > 0,
4266 ("[pmc,%d] cpu %d didn't find a sample to collect", __LINE__,
4269 KASSERT(td->td_pinned == 1,
4270 ("[pmc,%d] invalid td_pinned value", __LINE__));
4271 sched_unpin(); /* Can migrate safely now. */
4273 /* mark CPU as needing processing */
4274 CPU_SET_ATOMIC(cpu, &pmc_cpumask);
4280 * Process saved PC samples.
4284 pmc_process_samples(int cpu, int ring)
4289 struct pmc_owner *po;
4290 struct pmc_sample *ps;
4291 struct pmc_classdep *pcd;
4292 struct pmc_samplebuffer *psb;
4294 KASSERT(PCPU_GET(cpuid) == cpu,
4295 ("[pmc,%d] not on the correct CPU pcpu=%d cpu=%d", __LINE__,
4296 PCPU_GET(cpuid), cpu));
4298 psb = pmc_pcpu[cpu]->pc_sb[ring];
4300 for (n = 0; n < pmc_nsamples; n++) { /* bound on #iterations */
4303 if (ps->ps_nsamples == PMC_SAMPLE_FREE)
4308 KASSERT(pm->pm_runcount > 0,
4309 ("[pmc,%d] pm=%p runcount %d", __LINE__, (void *) pm,
4314 KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)),
4315 ("[pmc,%d] pmc=%p non-sampling mode=%d", __LINE__,
4316 pm, PMC_TO_MODE(pm)));
4318 /* Ignore PMCs that have been switched off */
4319 if (pm->pm_state != PMC_STATE_RUNNING)
4322 /* If there is a pending AST wait for completion */
4323 if (ps->ps_nsamples == PMC_SAMPLE_INUSE) {
4324 /* Need a rescan at a later time. */
4325 CPU_SET_ATOMIC(cpu, &pmc_cpumask);
4329 PMCDBG6(SAM,OPS,1,"cpu=%d pm=%p n=%d fl=%x wr=%d rd=%d", cpu,
4330 pm, ps->ps_nsamples, ps->ps_flags,
4331 (int) (psb->ps_write - psb->ps_samples),
4332 (int) (psb->ps_read - psb->ps_samples));
4335 * If this is a process-mode PMC that is attached to
4336 * its owner, and if the PC is in user mode, update
4337 * profiling statistics like timer-based profiling
4340 if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) {
4341 if (ps->ps_flags & PMC_CC_F_USERSPACE) {
4342 td = FIRST_THREAD_IN_PROC(po->po_owner);
4343 addupc_intr(td, ps->ps_pc[0], 1);
4349 * Otherwise, this is either a sampling mode PMC that
4350 * is attached to a different process than its owner,
4351 * or a system-wide sampling PMC. Dispatch a log
4352 * entry to the PMC's owner process.
4354 pmclog_process_callchain(pm, ps);
4357 ps->ps_nsamples = 0; /* mark entry as free */
4358 atomic_subtract_rel_int(&pm->pm_runcount, 1);
4360 /* increment read pointer, modulo sample size */
4361 if (++ps == psb->ps_fence)
4362 psb->ps_read = psb->ps_samples;
4367 atomic_add_int(&pmc_stats.pm_log_sweeps, 1);
4369 /* Do not re-enable stalled PMCs if we failed to process any samples */
4374 * Restart any stalled sampling PMCs on this CPU.
4376 * If the NMI handler sets the pm_stalled field of a PMC after
4377 * the check below, we'll end up processing the stalled PMC at
4378 * the next hardclock tick.
4380 for (n = 0; n < md->pmd_npmc; n++) {
4381 pcd = pmc_ri_to_classdep(md, n, &adjri);
4382 KASSERT(pcd != NULL,
4383 ("[pmc,%d] null pcd ri=%d", __LINE__, n));
4384 (void) (*pcd->pcd_get_config)(cpu,adjri,&pm);
4386 if (pm == NULL || /* !cfg'ed */
4387 pm->pm_state != PMC_STATE_RUNNING || /* !active */
4388 !PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) || /* !sampling */
4389 !CPU_ISSET(cpu, &pm->pm_cpustate) || /* !desired */
4390 !CPU_ISSET(cpu, &pm->pm_stalled)) /* !stalled */
4393 CPU_CLR_ATOMIC(cpu, &pm->pm_stalled);
4394 (*pcd->pcd_start_pmc)(cpu, adjri);
4403 * Handle a process exit.
4405 * Remove this process from all hash tables. If this process
4406 * owned any PMCs, turn off those PMCs and deallocate them,
4407 * removing any associations with target processes.
4409 * This function will be called by the last 'thread' of a
4412 * XXX This eventhandler gets called early in the exit process.
4413 * Consider using a 'hook' invocation from thread_exit() or equivalent
4414 * spot. Another negative is that kse_exit doesn't seem to call
4420 pmc_process_exit(void *arg __unused, struct proc *p)
4425 int is_using_hwpmcs;
4426 struct pmc_owner *po;
4427 struct pmc_process *pp;
4428 struct pmc_classdep *pcd;
4429 pmc_value_t newvalue, tmp;
4432 is_using_hwpmcs = p->p_flag & P_HWPMC;
4436 * Log a sysexit event to all SS PMC owners.
4438 LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
4439 if (po->po_flags & PMC_PO_OWNS_LOGFILE)
4440 pmclog_process_sysexit(po, p->p_pid);
4442 if (!is_using_hwpmcs)
4446 PMCDBG3(PRC,EXT,1,"process-exit proc=%p (%d, %s)", p, p->p_pid,
4450 * Since this code is invoked by the last thread in an exiting
4451 * process, we would have context switched IN at some prior
4452 * point. However, with PREEMPTION, kernel mode context
4453 * switches may happen any time, so we want to disable a
4454 * context switch OUT till we get any PMCs targeting this
4455 * process off the hardware.
4457 * We also need to atomically remove this process'
4458 * entry from our target process hash table, using
4461 PMCDBG3(PRC,EXT,1, "process-exit proc=%p (%d, %s)", p, p->p_pid,
4464 critical_enter(); /* no preemption */
4466 cpu = curthread->td_oncpu;
4468 if ((pp = pmc_find_process_descriptor(p,
4469 PMC_FLAG_REMOVE)) != NULL) {
4472 "process-exit proc=%p pmc-process=%p", p, pp);
4475 * The exiting process could the target of
4476 * some PMCs which will be running on
4477 * currently executing CPU.
4479 * We need to turn these PMCs off like we
4480 * would do at context switch OUT time.
4482 for (ri = 0; ri < md->pmd_npmc; ri++) {
4485 * Pick up the pmc pointer from hardware
4486 * state similar to the CSW_OUT code.
4490 pcd = pmc_ri_to_classdep(md, ri, &adjri);
4492 (void) (*pcd->pcd_get_config)(cpu, adjri, &pm);
4494 PMCDBG2(PRC,EXT,2, "ri=%d pm=%p", ri, pm);
4497 !PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)))
4500 PMCDBG4(PRC,EXT,2, "ppmcs[%d]=%p pm=%p "
4501 "state=%d", ri, pp->pp_pmcs[ri].pp_pmc,
4504 KASSERT(PMC_TO_ROWINDEX(pm) == ri,
4505 ("[pmc,%d] ri mismatch pmc(%d) ri(%d)",
4506 __LINE__, PMC_TO_ROWINDEX(pm), ri));
4508 KASSERT(pm == pp->pp_pmcs[ri].pp_pmc,
4509 ("[pmc,%d] pm %p != pp_pmcs[%d] %p",
4510 __LINE__, pm, ri, pp->pp_pmcs[ri].pp_pmc));
4512 KASSERT(pm->pm_runcount > 0,
4513 ("[pmc,%d] bad runcount ri %d rc %d",
4514 __LINE__, ri, pm->pm_runcount));
4517 * Change desired state, and then stop if not
4518 * stalled. This two-step dance should avoid
4519 * race conditions where an interrupt re-enables
4520 * the PMC after this code has already checked
4521 * the pm_stalled flag.
4523 if (CPU_ISSET(cpu, &pm->pm_cpustate)) {
4524 CPU_CLR_ATOMIC(cpu, &pm->pm_cpustate);
4525 if (!CPU_ISSET(cpu, &pm->pm_stalled)) {
4526 (void) pcd->pcd_stop_pmc(cpu, adjri);
4527 pcd->pcd_read_pmc(cpu, adjri,
4530 PMC_PCPU_SAVED(cpu,ri);
4532 mtx_pool_lock_spin(pmc_mtxpool, pm);
4533 pm->pm_gv.pm_savedvalue += tmp;
4534 pp->pp_pmcs[ri].pp_pmcval += tmp;
4535 mtx_pool_unlock_spin(pmc_mtxpool, pm);
4539 atomic_subtract_rel_int(&pm->pm_runcount,1);
4541 KASSERT((int) pm->pm_runcount >= 0,
4542 ("[pmc,%d] runcount is %d", __LINE__, ri));
4544 (void) pcd->pcd_config_pmc(cpu, adjri, NULL);
4548 * Inform the MD layer of this pseudo "context switch
4551 (void) md->pmd_switch_out(pmc_pcpu[cpu], pp);
4553 critical_exit(); /* ok to be pre-empted now */
4556 * Unlink this process from the PMCs that are
4557 * targeting it. This will send a signal to
4558 * all PMC owner's whose PMCs are orphaned.
4560 * Log PMC value at exit time if requested.
4562 for (ri = 0; ri < md->pmd_npmc; ri++)
4563 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) {
4564 if (pm->pm_flags & PMC_F_NEEDS_LOGFILE &&
4565 PMC_IS_COUNTING_MODE(PMC_TO_MODE(pm)))
4566 pmclog_process_procexit(pm, pp);
4567 pmc_unlink_target_process(pm, pp);
4572 critical_exit(); /* pp == NULL */
4576 * If the process owned PMCs, free them up and free up
4579 if ((po = pmc_find_owner_descriptor(p)) != NULL) {
4580 pmc_remove_owner(po);
4581 pmc_destroy_owner_descriptor(po);
4584 sx_xunlock(&pmc_sx);
4588 * Handle a process fork.
4590 * If the parent process 'p1' is under HWPMC monitoring, then copy
4591 * over any attached PMCs that have 'do_descendants' semantics.
4595 pmc_process_fork(void *arg __unused, struct proc *p1, struct proc *newproc,
4598 int is_using_hwpmcs;
4600 uint32_t do_descendants;
4602 struct pmc_owner *po;
4603 struct pmc_process *ppnew, *ppold;
4605 (void) flags; /* unused parameter */
4608 is_using_hwpmcs = p1->p_flag & P_HWPMC;
4612 * If there are system-wide sampling PMCs active, we need to
4613 * log all fork events to their owner's logs.
4616 LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
4617 if (po->po_flags & PMC_PO_OWNS_LOGFILE)
4618 pmclog_process_procfork(po, p1->p_pid, newproc->p_pid);
4620 if (!is_using_hwpmcs)
4624 PMCDBG4(PMC,FRK,1, "process-fork proc=%p (%d, %s) -> %p", p1,
4625 p1->p_pid, p1->p_comm, newproc);
4628 * If the parent process (curthread->td_proc) is a
4629 * target of any PMCs, look for PMCs that are to be
4630 * inherited, and link these into the new process
4633 if ((ppold = pmc_find_process_descriptor(curthread->td_proc,
4634 PMC_FLAG_NONE)) == NULL)
4635 goto done; /* nothing to do */
4638 for (ri = 0; ri < md->pmd_npmc; ri++)
4639 if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL)
4640 do_descendants |= pm->pm_flags & PMC_F_DESCENDANTS;
4641 if (do_descendants == 0) /* nothing to do */
4644 /* allocate a descriptor for the new process */
4645 if ((ppnew = pmc_find_process_descriptor(newproc,
4646 PMC_FLAG_ALLOCATE)) == NULL)
4650 * Run through all PMCs that were targeting the old process
4651 * and which specified F_DESCENDANTS and attach them to the
4654 * Log the fork event to all owners of PMCs attached to this
4655 * process, if not already logged.
4657 for (ri = 0; ri < md->pmd_npmc; ri++)
4658 if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL &&
4659 (pm->pm_flags & PMC_F_DESCENDANTS)) {
4660 pmc_link_target_process(pm, ppnew);
4662 if (po->po_sscount == 0 &&
4663 po->po_flags & PMC_PO_OWNS_LOGFILE)
4664 pmclog_process_procfork(po, p1->p_pid,
4669 * Now mark the new process as being tracked by this driver.
4672 newproc->p_flag |= P_HWPMC;
4673 PROC_UNLOCK(newproc);
4676 sx_xunlock(&pmc_sx);
4680 pmc_kld_load(void *arg __unused, linker_file_t lf)
4682 struct pmc_owner *po;
4687 * Notify owners of system sampling PMCs about KLD operations.
4689 LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
4690 if (po->po_flags & PMC_PO_OWNS_LOGFILE)
4691 pmclog_process_map_in(po, (pid_t) -1,
4692 (uintfptr_t) lf->address, lf->filename);
4695 * TODO: Notify owners of (all) process-sampling PMCs too.
4698 sx_sunlock(&pmc_sx);
4702 pmc_kld_unload(void *arg __unused, const char *filename __unused,
4703 caddr_t address, size_t size)
4705 struct pmc_owner *po;
4709 LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
4710 if (po->po_flags & PMC_PO_OWNS_LOGFILE)
4711 pmclog_process_map_out(po, (pid_t) -1,
4712 (uintfptr_t) address, (uintfptr_t) address + size);
4715 * TODO: Notify owners of process-sampling PMCs.
4718 sx_sunlock(&pmc_sx);
4725 pmc_name_of_pmcclass(enum pmc_class class)
4730 #define __PMC_CLASS(S,V,D) \
4731 case PMC_CLASS_##S: \
4735 return ("<unknown>");
4740 * Base class initializer: allocate structure and set default classes.
4743 pmc_mdep_alloc(int nclasses)
4745 struct pmc_mdep *md;
4748 /* SOFT + md classes */
4750 md = malloc(sizeof(struct pmc_mdep) + n *
4751 sizeof(struct pmc_classdep), M_PMC, M_WAITOK|M_ZERO);
4754 /* Add base class. */
4755 pmc_soft_initialize(md);
4760 pmc_mdep_free(struct pmc_mdep *md)
4762 pmc_soft_finalize(md);
4767 generic_switch_in(struct pmc_cpu *pc, struct pmc_process *pp)
4769 (void) pc; (void) pp;
4775 generic_switch_out(struct pmc_cpu *pc, struct pmc_process *pp)
4777 (void) pc; (void) pp;
4782 static struct pmc_mdep *
4783 pmc_generic_cpu_initialize(void)
4785 struct pmc_mdep *md;
4787 md = pmc_mdep_alloc(0);
4789 md->pmd_cputype = PMC_CPU_GENERIC;
4791 md->pmd_pcpu_init = NULL;
4792 md->pmd_pcpu_fini = NULL;
4793 md->pmd_switch_in = generic_switch_in;
4794 md->pmd_switch_out = generic_switch_out;
4800 pmc_generic_cpu_finalize(struct pmc_mdep *md)
4807 pmc_initialize(void)
4809 int c, cpu, error, n, ri;
4810 unsigned int maxcpu;
4811 struct pmc_binding pb;
4812 struct pmc_sample *ps;
4813 struct pmc_classdep *pcd;
4814 struct pmc_samplebuffer *sb;
4820 /* parse debug flags first */
4821 if (TUNABLE_STR_FETCH(PMC_SYSCTL_NAME_PREFIX "debugflags",
4822 pmc_debugstr, sizeof(pmc_debugstr)))
4823 pmc_debugflags_parse(pmc_debugstr,
4824 pmc_debugstr+strlen(pmc_debugstr));
4827 PMCDBG1(MOD,INI,0, "PMC Initialize (version %x)", PMC_VERSION);
4829 /* check kernel version */
4830 if (pmc_kernel_version != PMC_VERSION) {
4831 if (pmc_kernel_version == 0)
4832 printf("hwpmc: this kernel has not been compiled with "
4833 "'options HWPMC_HOOKS'.\n");
4835 printf("hwpmc: kernel version (0x%x) does not match "
4836 "module version (0x%x).\n", pmc_kernel_version,
4838 return EPROGMISMATCH;
4842 * check sysctl parameters
4845 if (pmc_hashsize <= 0) {
4846 (void) printf("hwpmc: tunable \"hashsize\"=%d must be "
4847 "greater than zero.\n", pmc_hashsize);
4848 pmc_hashsize = PMC_HASH_SIZE;
4851 if (pmc_nsamples <= 0 || pmc_nsamples > 65535) {
4852 (void) printf("hwpmc: tunable \"nsamples\"=%d out of "
4853 "range.\n", pmc_nsamples);
4854 pmc_nsamples = PMC_NSAMPLES;
4857 if (pmc_callchaindepth <= 0 ||
4858 pmc_callchaindepth > PMC_CALLCHAIN_DEPTH_MAX) {
4859 (void) printf("hwpmc: tunable \"callchaindepth\"=%d out of "
4860 "range - using %d.\n", pmc_callchaindepth,
4861 PMC_CALLCHAIN_DEPTH_MAX);
4862 pmc_callchaindepth = PMC_CALLCHAIN_DEPTH_MAX;
4865 md = pmc_md_initialize();
4867 /* Default to generic CPU. */
4868 md = pmc_generic_cpu_initialize();
4873 KASSERT(md->pmd_nclass >= 1 && md->pmd_npmc >= 1,
4874 ("[pmc,%d] no classes or pmcs", __LINE__));
4876 /* Compute the map from row-indices to classdep pointers. */
4877 pmc_rowindex_to_classdep = malloc(sizeof(struct pmc_classdep *) *
4878 md->pmd_npmc, M_PMC, M_WAITOK|M_ZERO);
4880 for (n = 0; n < md->pmd_npmc; n++)
4881 pmc_rowindex_to_classdep[n] = NULL;
4882 for (ri = c = 0; c < md->pmd_nclass; c++) {
4883 pcd = &md->pmd_classdep[c];
4884 for (n = 0; n < pcd->pcd_num; n++, ri++)
4885 pmc_rowindex_to_classdep[ri] = pcd;
4888 KASSERT(ri == md->pmd_npmc,
4889 ("[pmc,%d] npmc miscomputed: ri=%d, md->npmc=%d", __LINE__,
4892 maxcpu = pmc_cpu_max();
4894 /* allocate space for the per-cpu array */
4895 pmc_pcpu = malloc(maxcpu * sizeof(struct pmc_cpu *), M_PMC,
4898 /* per-cpu 'saved values' for managing process-mode PMCs */
4899 pmc_pcpu_saved = malloc(sizeof(pmc_value_t) * maxcpu * md->pmd_npmc,
4902 /* Perform CPU-dependent initialization. */
4903 pmc_save_cpu_binding(&pb);
4905 for (cpu = 0; error == 0 && cpu < maxcpu; cpu++) {
4906 if (!pmc_cpu_is_active(cpu))
4908 pmc_select_cpu(cpu);
4909 pmc_pcpu[cpu] = malloc(sizeof(struct pmc_cpu) +
4910 md->pmd_npmc * sizeof(struct pmc_hw *), M_PMC,
4912 if (md->pmd_pcpu_init)
4913 error = md->pmd_pcpu_init(md, cpu);
4914 for (n = 0; error == 0 && n < md->pmd_nclass; n++)
4915 error = md->pmd_classdep[n].pcd_pcpu_init(md, cpu);
4917 pmc_restore_cpu_binding(&pb);
4922 /* allocate space for the sample array */
4923 for (cpu = 0; cpu < maxcpu; cpu++) {
4924 if (!pmc_cpu_is_active(cpu))
4927 sb = malloc(sizeof(struct pmc_samplebuffer) +
4928 pmc_nsamples * sizeof(struct pmc_sample), M_PMC,
4930 sb->ps_read = sb->ps_write = sb->ps_samples;
4931 sb->ps_fence = sb->ps_samples + pmc_nsamples;
4933 KASSERT(pmc_pcpu[cpu] != NULL,
4934 ("[pmc,%d] cpu=%d Null per-cpu data", __LINE__, cpu));
4936 sb->ps_callchains = malloc(pmc_callchaindepth * pmc_nsamples *
4937 sizeof(uintptr_t), M_PMC, M_WAITOK|M_ZERO);
4939 for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++)
4940 ps->ps_pc = sb->ps_callchains +
4941 (n * pmc_callchaindepth);
4943 pmc_pcpu[cpu]->pc_sb[PMC_HR] = sb;
4945 sb = malloc(sizeof(struct pmc_samplebuffer) +
4946 pmc_nsamples * sizeof(struct pmc_sample), M_PMC,
4948 sb->ps_read = sb->ps_write = sb->ps_samples;
4949 sb->ps_fence = sb->ps_samples + pmc_nsamples;
4951 KASSERT(pmc_pcpu[cpu] != NULL,
4952 ("[pmc,%d] cpu=%d Null per-cpu data", __LINE__, cpu));
4954 sb->ps_callchains = malloc(pmc_callchaindepth * pmc_nsamples *
4955 sizeof(uintptr_t), M_PMC, M_WAITOK|M_ZERO);
4957 for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++)
4958 ps->ps_pc = sb->ps_callchains +
4959 (n * pmc_callchaindepth);
4961 pmc_pcpu[cpu]->pc_sb[PMC_SR] = sb;
4964 /* allocate space for the row disposition array */
4965 pmc_pmcdisp = malloc(sizeof(enum pmc_mode) * md->pmd_npmc,
4966 M_PMC, M_WAITOK|M_ZERO);
4968 /* mark all PMCs as available */
4969 for (n = 0; n < (int) md->pmd_npmc; n++)
4970 PMC_MARK_ROW_FREE(n);
4972 /* allocate thread hash tables */
4973 pmc_ownerhash = hashinit(pmc_hashsize, M_PMC,
4974 &pmc_ownerhashmask);
4976 pmc_processhash = hashinit(pmc_hashsize, M_PMC,
4977 &pmc_processhashmask);
4978 mtx_init(&pmc_processhash_mtx, "pmc-process-hash", "pmc-leaf",
4981 LIST_INIT(&pmc_ss_owners);
4984 /* allocate a pool of spin mutexes */
4985 pmc_mtxpool = mtx_pool_create("pmc-leaf", pmc_mtxpool_size,
4988 PMCDBG4(MOD,INI,1, "pmc_ownerhash=%p, mask=0x%lx "
4989 "targethash=%p mask=0x%lx", pmc_ownerhash, pmc_ownerhashmask,
4990 pmc_processhash, pmc_processhashmask);
4992 /* register process {exit,fork,exec} handlers */
4993 pmc_exit_tag = EVENTHANDLER_REGISTER(process_exit,
4994 pmc_process_exit, NULL, EVENTHANDLER_PRI_ANY);
4995 pmc_fork_tag = EVENTHANDLER_REGISTER(process_fork,
4996 pmc_process_fork, NULL, EVENTHANDLER_PRI_ANY);
4998 /* register kld event handlers */
4999 pmc_kld_load_tag = EVENTHANDLER_REGISTER(kld_load, pmc_kld_load,
5000 NULL, EVENTHANDLER_PRI_ANY);
5001 pmc_kld_unload_tag = EVENTHANDLER_REGISTER(kld_unload, pmc_kld_unload,
5002 NULL, EVENTHANDLER_PRI_ANY);
5004 /* initialize logging */
5005 pmclog_initialize();
5007 /* set hook functions */
5008 pmc_intr = md->pmd_intr;
5009 pmc_hook = pmc_hook_handler;
5012 printf(PMC_MODULE_NAME ":");
5013 for (n = 0; n < (int) md->pmd_nclass; n++) {
5014 pcd = &md->pmd_classdep[n];
5015 printf(" %s/%d/%d/0x%b",
5016 pmc_name_of_pmcclass(pcd->pcd_class),
5021 "\1INT\2USR\3SYS\4EDG\5THR"
5022 "\6REA\7WRI\10INV\11QUA\12PRC"
5031 /* prepare to be unloaded */
5036 unsigned int maxcpu;
5037 struct pmc_ownerhash *ph;
5038 struct pmc_owner *po, *tmp;
5039 struct pmc_binding pb;
5041 struct pmc_processhash *prh;
5044 PMCDBG0(MOD,INI,0, "cleanup");
5046 /* switch off sampling */
5047 CPU_ZERO(&pmc_cpumask);
5051 if (pmc_hook == NULL) { /* being unloaded already */
5052 sx_xunlock(&pmc_sx);
5056 pmc_hook = NULL; /* prevent new threads from entering module */
5058 /* deregister event handlers */
5059 EVENTHANDLER_DEREGISTER(process_fork, pmc_fork_tag);
5060 EVENTHANDLER_DEREGISTER(process_exit, pmc_exit_tag);
5061 EVENTHANDLER_DEREGISTER(kld_load, pmc_kld_load_tag);
5062 EVENTHANDLER_DEREGISTER(kld_unload, pmc_kld_unload_tag);
5064 /* send SIGBUS to all owner threads, free up allocations */
5066 for (ph = pmc_ownerhash;
5067 ph <= &pmc_ownerhash[pmc_ownerhashmask];
5069 LIST_FOREACH_SAFE(po, ph, po_next, tmp) {
5070 pmc_remove_owner(po);
5072 /* send SIGBUS to owner processes */
5073 PMCDBG3(MOD,INI,2, "cleanup signal proc=%p "
5074 "(%d, %s)", po->po_owner,
5075 po->po_owner->p_pid,
5076 po->po_owner->p_comm);
5078 PROC_LOCK(po->po_owner);
5079 kern_psignal(po->po_owner, SIGBUS);
5080 PROC_UNLOCK(po->po_owner);
5082 pmc_destroy_owner_descriptor(po);
5086 /* reclaim allocated data structures */
5088 mtx_pool_destroy(&pmc_mtxpool);
5090 mtx_destroy(&pmc_processhash_mtx);
5091 if (pmc_processhash) {
5093 struct pmc_process *pp;
5095 PMCDBG0(MOD,INI,3, "destroy process hash");
5096 for (prh = pmc_processhash;
5097 prh <= &pmc_processhash[pmc_processhashmask];
5099 LIST_FOREACH(pp, prh, pp_next)
5100 PMCDBG1(MOD,INI,3, "pid=%d", pp->pp_proc->p_pid);
5103 hashdestroy(pmc_processhash, M_PMC, pmc_processhashmask);
5104 pmc_processhash = NULL;
5107 if (pmc_ownerhash) {
5108 PMCDBG0(MOD,INI,3, "destroy owner hash");
5109 hashdestroy(pmc_ownerhash, M_PMC, pmc_ownerhashmask);
5110 pmc_ownerhash = NULL;
5113 KASSERT(LIST_EMPTY(&pmc_ss_owners),
5114 ("[pmc,%d] Global SS owner list not empty", __LINE__));
5115 KASSERT(pmc_ss_count == 0,
5116 ("[pmc,%d] Global SS count not empty", __LINE__));
5118 /* do processor and pmc-class dependent cleanup */
5119 maxcpu = pmc_cpu_max();
5121 PMCDBG0(MOD,INI,3, "md cleanup");
5123 pmc_save_cpu_binding(&pb);
5124 for (cpu = 0; cpu < maxcpu; cpu++) {
5125 PMCDBG2(MOD,INI,1,"pmc-cleanup cpu=%d pcs=%p",
5126 cpu, pmc_pcpu[cpu]);
5127 if (!pmc_cpu_is_active(cpu) || pmc_pcpu[cpu] == NULL)
5129 pmc_select_cpu(cpu);
5130 for (c = 0; c < md->pmd_nclass; c++)
5131 md->pmd_classdep[c].pcd_pcpu_fini(md, cpu);
5132 if (md->pmd_pcpu_fini)
5133 md->pmd_pcpu_fini(md, cpu);
5136 if (md->pmd_cputype == PMC_CPU_GENERIC)
5137 pmc_generic_cpu_finalize(md);
5139 pmc_md_finalize(md);
5143 pmc_restore_cpu_binding(&pb);
5146 /* Free per-cpu descriptors. */
5147 for (cpu = 0; cpu < maxcpu; cpu++) {
5148 if (!pmc_cpu_is_active(cpu))
5150 KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_HR] != NULL,
5151 ("[pmc,%d] Null hw cpu sample buffer cpu=%d", __LINE__,
5153 KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_SR] != NULL,
5154 ("[pmc,%d] Null sw cpu sample buffer cpu=%d", __LINE__,
5156 free(pmc_pcpu[cpu]->pc_sb[PMC_HR]->ps_callchains, M_PMC);
5157 free(pmc_pcpu[cpu]->pc_sb[PMC_HR], M_PMC);
5158 free(pmc_pcpu[cpu]->pc_sb[PMC_SR]->ps_callchains, M_PMC);
5159 free(pmc_pcpu[cpu]->pc_sb[PMC_SR], M_PMC);
5160 free(pmc_pcpu[cpu], M_PMC);
5163 free(pmc_pcpu, M_PMC);
5166 free(pmc_pcpu_saved, M_PMC);
5167 pmc_pcpu_saved = NULL;
5170 free(pmc_pmcdisp, M_PMC);
5174 if (pmc_rowindex_to_classdep) {
5175 free(pmc_rowindex_to_classdep, M_PMC);
5176 pmc_rowindex_to_classdep = NULL;
5181 sx_xunlock(&pmc_sx); /* we are done */
5185 * The function called at load/unload.
5189 load (struct module *module __unused, int cmd, void *arg __unused)
5197 /* initialize the subsystem */
5198 error = pmc_initialize();
5201 PMCDBG2(MOD,INI,1, "syscall=%d maxcpu=%d",
5202 pmc_syscall_num, pmc_cpu_max());
5209 PMCDBG0(MOD,INI,1, "unloaded");
5213 error = EINVAL; /* XXX should panic(9) */
5221 MALLOC_DEFINE(M_PMC, "pmc", "Memory space for the PMC module");