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 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "callchaindepth", &pmc_callchaindepth);
239 SYSCTL_INT(_kern_hwpmc, OID_AUTO, callchaindepth, CTLFLAG_TUN|CTLFLAG_RD,
240 &pmc_callchaindepth, 0, "depth of call chain records");
243 struct pmc_debugflags pmc_debugflags = PMC_DEBUG_DEFAULT_FLAGS;
244 char pmc_debugstr[PMC_DEBUG_STRSIZE];
245 TUNABLE_STR(PMC_SYSCTL_NAME_PREFIX "debugflags", pmc_debugstr,
246 sizeof(pmc_debugstr));
247 SYSCTL_PROC(_kern_hwpmc, OID_AUTO, debugflags,
248 CTLTYPE_STRING|CTLFLAG_RW|CTLFLAG_TUN,
249 0, 0, pmc_debugflags_sysctl_handler, "A", "debug flags");
253 * kern.hwpmc.hashrows -- determines the number of rows in the
254 * of the hash table used to look up threads
257 static int pmc_hashsize = PMC_HASH_SIZE;
258 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "hashsize", &pmc_hashsize);
259 SYSCTL_INT(_kern_hwpmc, OID_AUTO, hashsize, CTLFLAG_TUN|CTLFLAG_RD,
260 &pmc_hashsize, 0, "rows in hash tables");
263 * kern.hwpmc.nsamples --- number of PC samples/callchain stacks per CPU
266 static int pmc_nsamples = PMC_NSAMPLES;
267 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "nsamples", &pmc_nsamples);
268 SYSCTL_INT(_kern_hwpmc, OID_AUTO, nsamples, CTLFLAG_TUN|CTLFLAG_RD,
269 &pmc_nsamples, 0, "number of PC samples per CPU");
273 * kern.hwpmc.mtxpoolsize -- number of mutexes in the mutex pool.
276 static int pmc_mtxpool_size = PMC_MTXPOOL_SIZE;
277 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "mtxpoolsize", &pmc_mtxpool_size);
278 SYSCTL_INT(_kern_hwpmc, OID_AUTO, mtxpoolsize, CTLFLAG_TUN|CTLFLAG_RD,
279 &pmc_mtxpool_size, 0, "size of spin mutex pool");
283 * security.bsd.unprivileged_syspmcs -- allow non-root processes to
284 * allocate system-wide PMCs.
286 * Allowing unprivileged processes to allocate system PMCs is convenient
287 * if system-wide measurements need to be taken concurrently with other
288 * per-process measurements. This feature is turned off by default.
291 static int pmc_unprivileged_syspmcs = 0;
292 TUNABLE_INT("security.bsd.unprivileged_syspmcs", &pmc_unprivileged_syspmcs);
293 SYSCTL_INT(_security_bsd, OID_AUTO, unprivileged_syspmcs, CTLFLAG_RW,
294 &pmc_unprivileged_syspmcs, 0,
295 "allow unprivileged process to allocate system PMCs");
298 * Hash function. Discard the lower 2 bits of the pointer since
299 * these are always zero for our uses. The hash multiplier is
300 * round((2^LONG_BIT) * ((sqrt(5)-1)/2)).
304 #define _PMC_HM 11400714819323198486u
306 #define _PMC_HM 2654435769u
308 #error Must know the size of 'long' to compile
311 #define PMC_HASH_PTR(P,M) ((((unsigned long) (P) >> 2) * _PMC_HM) & (M))
317 /* The `sysent' for the new syscall */
318 static struct sysent pmc_sysent = {
320 pmc_syscall_handler /* sy_call */
323 static struct syscall_module_data pmc_syscall_mod = {
331 static moduledata_t pmc_mod = {
333 syscall_module_handler,
337 DECLARE_MODULE(pmc, pmc_mod, SI_SUB_SMP, SI_ORDER_ANY);
338 MODULE_VERSION(pmc, PMC_VERSION);
341 enum pmc_dbgparse_state {
342 PMCDS_WS, /* in whitespace */
343 PMCDS_MAJOR, /* seen a major keyword */
348 pmc_debugflags_parse(char *newstr, char *fence)
351 struct pmc_debugflags *tmpflags;
352 int error, found, *newbits, tmp;
355 tmpflags = malloc(sizeof(*tmpflags), M_PMC, M_WAITOK|M_ZERO);
360 for (; p < fence && (c = *p); p++) {
362 /* skip white space */
363 if (c == ' ' || c == '\t')
366 /* look for a keyword followed by "=" */
367 for (q = p; p < fence && (c = *p) && c != '='; p++)
377 /* lookup flag group name */
378 #define DBG_SET_FLAG_MAJ(S,F) \
379 if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0) \
380 newbits = &tmpflags->pdb_ ## F;
382 DBG_SET_FLAG_MAJ("cpu", CPU);
383 DBG_SET_FLAG_MAJ("csw", CSW);
384 DBG_SET_FLAG_MAJ("logging", LOG);
385 DBG_SET_FLAG_MAJ("module", MOD);
386 DBG_SET_FLAG_MAJ("md", MDP);
387 DBG_SET_FLAG_MAJ("owner", OWN);
388 DBG_SET_FLAG_MAJ("pmc", PMC);
389 DBG_SET_FLAG_MAJ("process", PRC);
390 DBG_SET_FLAG_MAJ("sampling", SAM);
392 if (newbits == NULL) {
397 p++; /* skip the '=' */
399 /* Now parse the individual flags */
402 for (q = p; p < fence && (c = *p); p++)
403 if (c == ' ' || c == '\t' || c == ',')
406 /* p == fence or c == ws or c == "," or c == 0 */
408 if ((kwlen = p - q) == 0) {
414 #define DBG_SET_FLAG_MIN(S,F) \
415 if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0) \
416 tmp |= found = (1 << PMC_DEBUG_MIN_ ## F)
418 /* a '*' denotes all possible flags in the group */
419 if (kwlen == 1 && *q == '*')
421 /* look for individual flag names */
422 DBG_SET_FLAG_MIN("allocaterow", ALR);
423 DBG_SET_FLAG_MIN("allocate", ALL);
424 DBG_SET_FLAG_MIN("attach", ATT);
425 DBG_SET_FLAG_MIN("bind", BND);
426 DBG_SET_FLAG_MIN("config", CFG);
427 DBG_SET_FLAG_MIN("exec", EXC);
428 DBG_SET_FLAG_MIN("exit", EXT);
429 DBG_SET_FLAG_MIN("find", FND);
430 DBG_SET_FLAG_MIN("flush", FLS);
431 DBG_SET_FLAG_MIN("fork", FRK);
432 DBG_SET_FLAG_MIN("getbuf", GTB);
433 DBG_SET_FLAG_MIN("hook", PMH);
434 DBG_SET_FLAG_MIN("init", INI);
435 DBG_SET_FLAG_MIN("intr", INT);
436 DBG_SET_FLAG_MIN("linktarget", TLK);
437 DBG_SET_FLAG_MIN("mayberemove", OMR);
438 DBG_SET_FLAG_MIN("ops", OPS);
439 DBG_SET_FLAG_MIN("read", REA);
440 DBG_SET_FLAG_MIN("register", REG);
441 DBG_SET_FLAG_MIN("release", REL);
442 DBG_SET_FLAG_MIN("remove", ORM);
443 DBG_SET_FLAG_MIN("sample", SAM);
444 DBG_SET_FLAG_MIN("scheduleio", SIO);
445 DBG_SET_FLAG_MIN("select", SEL);
446 DBG_SET_FLAG_MIN("signal", SIG);
447 DBG_SET_FLAG_MIN("swi", SWI);
448 DBG_SET_FLAG_MIN("swo", SWO);
449 DBG_SET_FLAG_MIN("start", STA);
450 DBG_SET_FLAG_MIN("stop", STO);
451 DBG_SET_FLAG_MIN("syscall", PMS);
452 DBG_SET_FLAG_MIN("unlinktarget", TUL);
453 DBG_SET_FLAG_MIN("write", WRI);
455 /* unrecognized flag name */
460 if (c == 0 || c == ' ' || c == '\t') { /* end of flag group */
469 /* save the new flag set */
470 bcopy(tmpflags, &pmc_debugflags, sizeof(pmc_debugflags));
473 free(tmpflags, M_PMC);
478 pmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS)
480 char *fence, *newstr;
484 (void) arg1; (void) arg2; /* unused parameters */
486 n = sizeof(pmc_debugstr);
487 newstr = malloc(n, M_PMC, M_WAITOK|M_ZERO);
488 (void) strlcpy(newstr, pmc_debugstr, n);
490 error = sysctl_handle_string(oidp, newstr, n, req);
492 /* if there is a new string, parse and copy it */
493 if (error == 0 && req->newptr != NULL) {
494 fence = newstr + (n < req->newlen ? n : req->newlen + 1);
495 if ((error = pmc_debugflags_parse(newstr, fence)) == 0)
496 (void) strlcpy(pmc_debugstr, newstr,
497 sizeof(pmc_debugstr));
507 * Map a row index to a classdep structure and return the adjusted row
508 * index for the PMC class index.
510 static struct pmc_classdep *
511 pmc_ri_to_classdep(struct pmc_mdep *md, int ri, int *adjri)
513 struct pmc_classdep *pcd;
517 KASSERT(ri >= 0 && ri < md->pmd_npmc,
518 ("[pmc,%d] illegal row-index %d", __LINE__, ri));
520 pcd = pmc_rowindex_to_classdep[ri];
523 ("[pmc,%d] ri %d null pcd", __LINE__, ri));
525 *adjri = ri - pcd->pcd_ri;
527 KASSERT(*adjri >= 0 && *adjri < pcd->pcd_num,
528 ("[pmc,%d] adjusted row-index %d", __LINE__, *adjri));
534 * Concurrency Control
536 * The driver manages the following data structures:
538 * - target process descriptors, one per target process
539 * - owner process descriptors (and attached lists), one per owner process
540 * - lookup hash tables for owner and target processes
541 * - PMC descriptors (and attached lists)
542 * - per-cpu hardware state
543 * - the 'hook' variable through which the kernel calls into
545 * - the machine hardware state (managed by the MD layer)
547 * These data structures are accessed from:
549 * - thread context-switch code
550 * - interrupt handlers (possibly on multiple cpus)
551 * - kernel threads on multiple cpus running on behalf of user
552 * processes doing system calls
553 * - this driver's private kernel threads
555 * = Locks and Locking strategy =
557 * The driver uses four locking strategies for its operation:
559 * - The global SX lock "pmc_sx" is used to protect internal
562 * Calls into the module by syscall() start with this lock being
563 * held in exclusive mode. Depending on the requested operation,
564 * the lock may be downgraded to 'shared' mode to allow more
565 * concurrent readers into the module. Calls into the module from
566 * other parts of the kernel acquire the lock in shared mode.
568 * This SX lock is held in exclusive mode for any operations that
569 * modify the linkages between the driver's internal data structures.
571 * The 'pmc_hook' function pointer is also protected by this lock.
572 * It is only examined with the sx lock held in exclusive mode. The
573 * kernel module is allowed to be unloaded only with the sx lock held
574 * in exclusive mode. In normal syscall handling, after acquiring the
575 * pmc_sx lock we first check that 'pmc_hook' is non-null before
576 * proceeding. This prevents races between the thread unloading the module
577 * and other threads seeking to use the module.
579 * - Lookups of target process structures and owner process structures
580 * cannot use the global "pmc_sx" SX lock because these lookups need
581 * to happen during context switches and in other critical sections
582 * where sleeping is not allowed. We protect these lookup tables
583 * with their own private spin-mutexes, "pmc_processhash_mtx" and
584 * "pmc_ownerhash_mtx".
586 * - Interrupt handlers work in a lock free manner. At interrupt
587 * time, handlers look at the PMC pointer (phw->phw_pmc) configured
588 * when the PMC was started. If this pointer is NULL, the interrupt
589 * is ignored after updating driver statistics. We ensure that this
590 * pointer is set (using an atomic operation if necessary) before the
591 * PMC hardware is started. Conversely, this pointer is unset atomically
592 * only after the PMC hardware is stopped.
594 * We ensure that everything needed for the operation of an
595 * interrupt handler is available without it needing to acquire any
596 * locks. We also ensure that a PMC's software state is destroyed only
597 * after the PMC is taken off hardware (on all CPUs).
599 * - Context-switch handling with process-private PMCs needs more
602 * A given process may be the target of multiple PMCs. For example,
603 * PMCATTACH and PMCDETACH may be requested by a process on one CPU
604 * while the target process is running on another. A PMC could also
605 * be getting released because its owner is exiting. We tackle
606 * these situations in the following manner:
608 * - each target process structure 'pmc_process' has an array
609 * of 'struct pmc *' pointers, one for each hardware PMC.
611 * - At context switch IN time, each "target" PMC in RUNNING state
612 * gets started on hardware and a pointer to each PMC is copied into
613 * the per-cpu phw array. The 'runcount' for the PMC is
616 * - At context switch OUT time, all process-virtual PMCs are stopped
617 * on hardware. The saved value is added to the PMCs value field
618 * only if the PMC is in a non-deleted state (the PMCs state could
619 * have changed during the current time slice).
621 * Note that since in-between a switch IN on a processor and a switch
622 * OUT, the PMC could have been released on another CPU. Therefore
623 * context switch OUT always looks at the hardware state to turn
624 * OFF PMCs and will update a PMC's saved value only if reachable
625 * from the target process record.
627 * - OP PMCRELEASE could be called on a PMC at any time (the PMC could
628 * be attached to many processes at the time of the call and could
629 * be active on multiple CPUs).
631 * We prevent further scheduling of the PMC by marking it as in
632 * state 'DELETED'. If the runcount of the PMC is non-zero then
633 * this PMC is currently running on a CPU somewhere. The thread
634 * doing the PMCRELEASE operation waits by repeatedly doing a
635 * pause() till the runcount comes to zero.
637 * The contents of a PMC descriptor (struct pmc) are protected using
638 * a spin-mutex. In order to save space, we use a mutex pool.
640 * In terms of lock types used by witness(4), we use:
641 * - Type "pmc-sx", used by the global SX lock.
642 * - Type "pmc-sleep", for sleep mutexes used by logger threads.
643 * - Type "pmc-per-proc", for protecting PMC owner descriptors.
644 * - Type "pmc-leaf", used for all other spin mutexes.
648 * save the cpu binding of the current kthread
652 pmc_save_cpu_binding(struct pmc_binding *pb)
654 PMCDBG(CPU,BND,2, "%s", "save-cpu");
655 thread_lock(curthread);
656 pb->pb_bound = sched_is_bound(curthread);
657 pb->pb_cpu = curthread->td_oncpu;
658 thread_unlock(curthread);
659 PMCDBG(CPU,BND,2, "save-cpu cpu=%d", pb->pb_cpu);
663 * restore the cpu binding of the current thread
667 pmc_restore_cpu_binding(struct pmc_binding *pb)
669 PMCDBG(CPU,BND,2, "restore-cpu curcpu=%d restore=%d",
670 curthread->td_oncpu, pb->pb_cpu);
671 thread_lock(curthread);
673 sched_bind(curthread, pb->pb_cpu);
675 sched_unbind(curthread);
676 thread_unlock(curthread);
677 PMCDBG(CPU,BND,2, "%s", "restore-cpu done");
681 * move execution over the specified cpu and bind it there.
685 pmc_select_cpu(int cpu)
687 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
688 ("[pmc,%d] bad cpu number %d", __LINE__, cpu));
690 /* Never move to an inactive CPU. */
691 KASSERT(pmc_cpu_is_active(cpu), ("[pmc,%d] selecting inactive "
692 "CPU %d", __LINE__, cpu));
694 PMCDBG(CPU,SEL,2, "select-cpu cpu=%d", cpu);
695 thread_lock(curthread);
696 sched_bind(curthread, cpu);
697 thread_unlock(curthread);
699 KASSERT(curthread->td_oncpu == cpu,
700 ("[pmc,%d] CPU not bound [cpu=%d, curr=%d]", __LINE__,
701 cpu, curthread->td_oncpu));
703 PMCDBG(CPU,SEL,2, "select-cpu cpu=%d ok", cpu);
707 * Force a context switch.
709 * We do this by pause'ing for 1 tick -- invoking mi_switch() is not
710 * guaranteed to force a context switch.
714 pmc_force_context_switch(void)
721 * Get the file name for an executable. This is a simple wrapper
722 * around vn_fullpath(9).
726 pmc_getfilename(struct vnode *v, char **fullpath, char **freepath)
729 *fullpath = "unknown";
731 vn_fullpath(curthread, v, fullpath, freepath);
735 * remove an process owning PMCs
739 pmc_remove_owner(struct pmc_owner *po)
741 struct pmc *pm, *tmp;
743 sx_assert(&pmc_sx, SX_XLOCKED);
745 PMCDBG(OWN,ORM,1, "remove-owner po=%p", po);
747 /* Remove descriptor from the owner hash table */
748 LIST_REMOVE(po, po_next);
750 /* release all owned PMC descriptors */
751 LIST_FOREACH_SAFE(pm, &po->po_pmcs, pm_next, tmp) {
752 PMCDBG(OWN,ORM,2, "pmc=%p", pm);
753 KASSERT(pm->pm_owner == po,
754 ("[pmc,%d] owner %p != po %p", __LINE__, pm->pm_owner, po));
756 pmc_release_pmc_descriptor(pm); /* will unlink from the list */
757 pmc_destroy_pmc_descriptor(pm);
760 KASSERT(po->po_sscount == 0,
761 ("[pmc,%d] SS count not zero", __LINE__));
762 KASSERT(LIST_EMPTY(&po->po_pmcs),
763 ("[pmc,%d] PMC list not empty", __LINE__));
765 /* de-configure the log file if present */
766 if (po->po_flags & PMC_PO_OWNS_LOGFILE)
767 pmclog_deconfigure_log(po);
771 * remove an owner process record if all conditions are met.
775 pmc_maybe_remove_owner(struct pmc_owner *po)
778 PMCDBG(OWN,OMR,1, "maybe-remove-owner po=%p", po);
781 * Remove owner record if
782 * - this process does not own any PMCs
783 * - this process has not allocated a system-wide sampling buffer
786 if (LIST_EMPTY(&po->po_pmcs) &&
787 ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)) {
788 pmc_remove_owner(po);
789 pmc_destroy_owner_descriptor(po);
794 * Add an association between a target process and a PMC.
798 pmc_link_target_process(struct pmc *pm, struct pmc_process *pp)
801 struct pmc_target *pt;
803 sx_assert(&pmc_sx, SX_XLOCKED);
805 KASSERT(pm != NULL && pp != NULL,
806 ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp));
807 KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)),
808 ("[pmc,%d] Attaching a non-process-virtual pmc=%p to pid=%d",
809 __LINE__, pm, pp->pp_proc->p_pid));
810 KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= ((int) md->pmd_npmc - 1),
811 ("[pmc,%d] Illegal reference count %d for process record %p",
812 __LINE__, pp->pp_refcnt, (void *) pp));
814 ri = PMC_TO_ROWINDEX(pm);
816 PMCDBG(PRC,TLK,1, "link-target pmc=%p ri=%d pmc-process=%p",
820 LIST_FOREACH(pt, &pm->pm_targets, pt_next)
821 if (pt->pt_process == pp)
822 KASSERT(0, ("[pmc,%d] pp %p already in pmc %p targets",
826 pt = malloc(sizeof(struct pmc_target), M_PMC, M_WAITOK|M_ZERO);
829 LIST_INSERT_HEAD(&pm->pm_targets, pt, pt_next);
831 atomic_store_rel_ptr((uintptr_t *)&pp->pp_pmcs[ri].pp_pmc,
834 if (pm->pm_owner->po_owner == pp->pp_proc)
835 pm->pm_flags |= PMC_F_ATTACHED_TO_OWNER;
838 * Initialize the per-process values at this row index.
840 pp->pp_pmcs[ri].pp_pmcval = PMC_TO_MODE(pm) == PMC_MODE_TS ?
841 pm->pm_sc.pm_reloadcount : 0;
848 * Removes the association between a target process and a PMC.
852 pmc_unlink_target_process(struct pmc *pm, struct pmc_process *pp)
856 struct pmc_target *ptgt;
858 sx_assert(&pmc_sx, SX_XLOCKED);
860 KASSERT(pm != NULL && pp != NULL,
861 ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp));
863 KASSERT(pp->pp_refcnt >= 1 && pp->pp_refcnt <= (int) md->pmd_npmc,
864 ("[pmc,%d] Illegal ref count %d on process record %p",
865 __LINE__, pp->pp_refcnt, (void *) pp));
867 ri = PMC_TO_ROWINDEX(pm);
869 PMCDBG(PRC,TUL,1, "unlink-target pmc=%p ri=%d pmc-process=%p",
872 KASSERT(pp->pp_pmcs[ri].pp_pmc == pm,
873 ("[pmc,%d] PMC ri %d mismatch pmc %p pp->[ri] %p", __LINE__,
874 ri, pm, pp->pp_pmcs[ri].pp_pmc));
876 pp->pp_pmcs[ri].pp_pmc = NULL;
877 pp->pp_pmcs[ri].pp_pmcval = (pmc_value_t) 0;
879 /* Remove owner-specific flags */
880 if (pm->pm_owner->po_owner == pp->pp_proc) {
881 pp->pp_flags &= ~PMC_PP_ENABLE_MSR_ACCESS;
882 pm->pm_flags &= ~PMC_F_ATTACHED_TO_OWNER;
887 /* Remove the target process from the PMC structure */
888 LIST_FOREACH(ptgt, &pm->pm_targets, pt_next)
889 if (ptgt->pt_process == pp)
892 KASSERT(ptgt != NULL, ("[pmc,%d] process %p (pp: %p) not found "
893 "in pmc %p", __LINE__, pp->pp_proc, pp, pm));
895 LIST_REMOVE(ptgt, pt_next);
898 /* if the PMC now lacks targets, send the owner a SIGIO */
899 if (LIST_EMPTY(&pm->pm_targets)) {
900 p = pm->pm_owner->po_owner;
902 kern_psignal(p, SIGIO);
905 PMCDBG(PRC,SIG,2, "signalling proc=%p signal=%d", p,
911 * Check if PMC 'pm' may be attached to target process 't'.
915 pmc_can_attach(struct pmc *pm, struct proc *t)
917 struct proc *o; /* pmc owner */
918 struct ucred *oc, *tc; /* owner, target credentials */
919 int decline_attach, i;
922 * A PMC's owner can always attach that PMC to itself.
925 if ((o = pm->pm_owner->po_owner) == t)
939 * The effective uid of the PMC owner should match at least one
940 * of the {effective,real,saved} uids of the target process.
943 decline_attach = oc->cr_uid != tc->cr_uid &&
944 oc->cr_uid != tc->cr_svuid &&
945 oc->cr_uid != tc->cr_ruid;
948 * Every one of the target's group ids, must be in the owner's
951 for (i = 0; !decline_attach && i < tc->cr_ngroups; i++)
952 decline_attach = !groupmember(tc->cr_groups[i], oc);
954 /* check the read and saved gids too */
955 if (decline_attach == 0)
956 decline_attach = !groupmember(tc->cr_rgid, oc) ||
957 !groupmember(tc->cr_svgid, oc);
962 return !decline_attach;
966 * Attach a process to a PMC.
970 pmc_attach_one_process(struct proc *p, struct pmc *pm)
973 char *fullpath, *freepath;
974 struct pmc_process *pp;
976 sx_assert(&pmc_sx, SX_XLOCKED);
978 PMCDBG(PRC,ATT,2, "attach-one pm=%p ri=%d proc=%p (%d, %s)", pm,
979 PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm);
982 * Locate the process descriptor corresponding to process 'p',
983 * allocating space as needed.
985 * Verify that rowindex 'pm_rowindex' is free in the process
988 * If not, allocate space for a descriptor and link the
989 * process descriptor and PMC.
991 ri = PMC_TO_ROWINDEX(pm);
993 if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_ALLOCATE)) == NULL)
996 if (pp->pp_pmcs[ri].pp_pmc == pm) /* already present at slot [ri] */
999 if (pp->pp_pmcs[ri].pp_pmc != NULL)
1002 pmc_link_target_process(pm, pp);
1004 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) &&
1005 (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) == 0)
1006 pm->pm_flags |= PMC_F_NEEDS_LOGFILE;
1008 pm->pm_flags |= PMC_F_ATTACH_DONE; /* mark as attached */
1010 /* issue an attach event to a configured log file */
1011 if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE) {
1012 pmc_getfilename(p->p_textvp, &fullpath, &freepath);
1013 if (p->p_flag & P_KTHREAD) {
1014 fullpath = kernelname;
1017 pmclog_process_pmcattach(pm, p->p_pid, fullpath);
1019 free(freepath, M_TEMP);
1020 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1021 pmc_log_process_mappings(pm->pm_owner, p);
1023 /* mark process as using HWPMCs */
1025 p->p_flag |= P_HWPMC;
1032 * Attach a process and optionally its children
1036 pmc_attach_process(struct proc *p, struct pmc *pm)
1041 sx_assert(&pmc_sx, SX_XLOCKED);
1043 PMCDBG(PRC,ATT,1, "attach pm=%p ri=%d proc=%p (%d, %s)", pm,
1044 PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm);
1048 * If this PMC successfully allowed a GETMSR operation
1049 * in the past, disallow further ATTACHes.
1052 if ((pm->pm_flags & PMC_PP_ENABLE_MSR_ACCESS) != 0)
1055 if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0)
1056 return pmc_attach_one_process(p, pm);
1059 * Traverse all child processes, attaching them to
1063 sx_slock(&proctree_lock);
1068 if ((error = pmc_attach_one_process(p, pm)) != 0)
1070 if (!LIST_EMPTY(&p->p_children))
1071 p = LIST_FIRST(&p->p_children);
1075 if (LIST_NEXT(p, p_sibling)) {
1076 p = LIST_NEXT(p, p_sibling);
1084 (void) pmc_detach_process(top, pm);
1087 sx_sunlock(&proctree_lock);
1092 * Detach a process from a PMC. If there are no other PMCs tracking
1093 * this process, remove the process structure from its hash table. If
1094 * 'flags' contains PMC_FLAG_REMOVE, then free the process structure.
1098 pmc_detach_one_process(struct proc *p, struct pmc *pm, int flags)
1101 struct pmc_process *pp;
1103 sx_assert(&pmc_sx, SX_XLOCKED);
1106 ("[pmc,%d] null pm pointer", __LINE__));
1108 ri = PMC_TO_ROWINDEX(pm);
1110 PMCDBG(PRC,ATT,2, "detach-one pm=%p ri=%d proc=%p (%d, %s) flags=0x%x",
1111 pm, ri, p, p->p_pid, p->p_comm, flags);
1113 if ((pp = pmc_find_process_descriptor(p, 0)) == NULL)
1116 if (pp->pp_pmcs[ri].pp_pmc != pm)
1119 pmc_unlink_target_process(pm, pp);
1121 /* Issue a detach entry if a log file is configured */
1122 if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE)
1123 pmclog_process_pmcdetach(pm, p->p_pid);
1126 * If there are no PMCs targetting this process, we remove its
1127 * descriptor from the target hash table and unset the P_HWPMC
1128 * flag in the struct proc.
1130 KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= (int) md->pmd_npmc,
1131 ("[pmc,%d] Illegal refcnt %d for process struct %p",
1132 __LINE__, pp->pp_refcnt, pp));
1134 if (pp->pp_refcnt != 0) /* still a target of some PMC */
1137 pmc_remove_process_descriptor(pp);
1139 if (flags & PMC_FLAG_REMOVE)
1143 p->p_flag &= ~P_HWPMC;
1150 * Detach a process and optionally its descendants from a PMC.
1154 pmc_detach_process(struct proc *p, struct pmc *pm)
1158 sx_assert(&pmc_sx, SX_XLOCKED);
1160 PMCDBG(PRC,ATT,1, "detach pm=%p ri=%d proc=%p (%d, %s)", pm,
1161 PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm);
1163 if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0)
1164 return pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE);
1167 * Traverse all children, detaching them from this PMC. We
1168 * ignore errors since we could be detaching a PMC from a
1169 * partially attached proc tree.
1172 sx_slock(&proctree_lock);
1177 (void) pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE);
1179 if (!LIST_EMPTY(&p->p_children))
1180 p = LIST_FIRST(&p->p_children);
1184 if (LIST_NEXT(p, p_sibling)) {
1185 p = LIST_NEXT(p, p_sibling);
1193 sx_sunlock(&proctree_lock);
1195 if (LIST_EMPTY(&pm->pm_targets))
1196 pm->pm_flags &= ~PMC_F_ATTACH_DONE;
1203 * Thread context switch IN
1207 pmc_process_csw_in(struct thread *td)
1210 unsigned int adjri, ri;
1215 pmc_value_t newvalue;
1216 struct pmc_process *pp;
1217 struct pmc_classdep *pcd;
1221 if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE)) == NULL)
1224 KASSERT(pp->pp_proc == td->td_proc,
1225 ("[pmc,%d] not my thread state", __LINE__));
1227 critical_enter(); /* no preemption from this point */
1229 cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */
1231 PMCDBG(CSW,SWI,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p,
1232 p->p_pid, p->p_comm, pp);
1234 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
1235 ("[pmc,%d] wierd CPU id %d", __LINE__, cpu));
1239 for (ri = 0; ri < md->pmd_npmc; ri++) {
1241 if ((pm = pp->pp_pmcs[ri].pp_pmc) == NULL)
1244 KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)),
1245 ("[pmc,%d] Target PMC in non-virtual mode (%d)",
1246 __LINE__, PMC_TO_MODE(pm)));
1248 KASSERT(PMC_TO_ROWINDEX(pm) == ri,
1249 ("[pmc,%d] Row index mismatch pmc %d != ri %d",
1250 __LINE__, PMC_TO_ROWINDEX(pm), ri));
1253 * Only PMCs that are marked as 'RUNNING' need
1254 * be placed on hardware.
1257 if (pm->pm_state != PMC_STATE_RUNNING)
1260 /* increment PMC runcount */
1261 atomic_add_rel_int(&pm->pm_runcount, 1);
1263 /* configure the HWPMC we are going to use. */
1264 pcd = pmc_ri_to_classdep(md, ri, &adjri);
1265 pcd->pcd_config_pmc(cpu, adjri, pm);
1267 phw = pc->pc_hwpmcs[ri];
1269 KASSERT(phw != NULL,
1270 ("[pmc,%d] null hw pointer", __LINE__));
1272 KASSERT(phw->phw_pmc == pm,
1273 ("[pmc,%d] hw->pmc %p != pmc %p", __LINE__,
1277 * Write out saved value and start the PMC.
1279 * Sampling PMCs use a per-process value, while
1280 * counting mode PMCs use a per-pmc value that is
1281 * inherited across descendants.
1283 if (PMC_TO_MODE(pm) == PMC_MODE_TS) {
1284 mtx_pool_lock_spin(pmc_mtxpool, pm);
1285 newvalue = PMC_PCPU_SAVED(cpu,ri) =
1286 pp->pp_pmcs[ri].pp_pmcval;
1287 mtx_pool_unlock_spin(pmc_mtxpool, pm);
1289 KASSERT(PMC_TO_MODE(pm) == PMC_MODE_TC,
1290 ("[pmc,%d] illegal mode=%d", __LINE__,
1292 mtx_pool_lock_spin(pmc_mtxpool, pm);
1293 newvalue = PMC_PCPU_SAVED(cpu, ri) =
1294 pm->pm_gv.pm_savedvalue;
1295 mtx_pool_unlock_spin(pmc_mtxpool, pm);
1298 PMCDBG(CSW,SWI,1,"cpu=%d ri=%d new=%jd", cpu, ri, newvalue);
1300 pcd->pcd_write_pmc(cpu, adjri, newvalue);
1301 pcd->pcd_start_pmc(cpu, adjri);
1305 * perform any other architecture/cpu dependent thread
1306 * switch-in actions.
1309 (void) (*md->pmd_switch_in)(pc, pp);
1316 * Thread context switch OUT.
1320 pmc_process_csw_out(struct thread *td)
1328 pmc_value_t newvalue;
1329 unsigned int adjri, ri;
1330 struct pmc_process *pp;
1331 struct pmc_classdep *pcd;
1335 * Locate our process descriptor; this may be NULL if
1336 * this process is exiting and we have already removed
1337 * the process from the target process table.
1339 * Note that due to kernel preemption, multiple
1340 * context switches may happen while the process is
1343 * Note also that if the target process cannot be
1344 * found we still need to deconfigure any PMCs that
1345 * are currently running on hardware.
1349 pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE);
1357 cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */
1359 PMCDBG(CSW,SWO,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p,
1360 p->p_pid, p->p_comm, pp);
1362 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
1363 ("[pmc,%d wierd CPU id %d", __LINE__, cpu));
1368 * When a PMC gets unlinked from a target PMC, it will
1369 * be removed from the target's pp_pmc[] array.
1371 * However, on a MP system, the target could have been
1372 * executing on another CPU at the time of the unlink.
1373 * So, at context switch OUT time, we need to look at
1374 * the hardware to determine if a PMC is scheduled on
1378 for (ri = 0; ri < md->pmd_npmc; ri++) {
1380 pcd = pmc_ri_to_classdep(md, ri, &adjri);
1382 (void) (*pcd->pcd_get_config)(cpu, adjri, &pm);
1384 if (pm == NULL) /* nothing at this row index */
1387 mode = PMC_TO_MODE(pm);
1388 if (!PMC_IS_VIRTUAL_MODE(mode))
1389 continue; /* not a process virtual PMC */
1391 KASSERT(PMC_TO_ROWINDEX(pm) == ri,
1392 ("[pmc,%d] ri mismatch pmc(%d) ri(%d)",
1393 __LINE__, PMC_TO_ROWINDEX(pm), ri));
1395 /* Stop hardware if not already stopped */
1396 if (pm->pm_stalled == 0)
1397 pcd->pcd_stop_pmc(cpu, adjri);
1399 /* reduce this PMC's runcount */
1400 atomic_subtract_rel_int(&pm->pm_runcount, 1);
1403 * If this PMC is associated with this process,
1407 if (pp != NULL && pp->pp_pmcs[ri].pp_pmc != NULL) {
1409 KASSERT(pm == pp->pp_pmcs[ri].pp_pmc,
1410 ("[pmc,%d] pm %p != pp_pmcs[%d] %p", __LINE__,
1411 pm, ri, pp->pp_pmcs[ri].pp_pmc));
1413 KASSERT(pp->pp_refcnt > 0,
1414 ("[pmc,%d] pp refcnt = %d", __LINE__,
1417 pcd->pcd_read_pmc(cpu, adjri, &newvalue);
1419 tmp = newvalue - PMC_PCPU_SAVED(cpu,ri);
1421 PMCDBG(CSW,SWO,1,"cpu=%d ri=%d tmp=%jd", cpu, ri,
1424 if (mode == PMC_MODE_TS) {
1427 * For sampling process-virtual PMCs,
1428 * we expect the count to be
1429 * decreasing as the 'value'
1430 * programmed into the PMC is the
1431 * number of events to be seen till
1432 * the next sampling interrupt.
1435 tmp += pm->pm_sc.pm_reloadcount;
1436 mtx_pool_lock_spin(pmc_mtxpool, pm);
1437 pp->pp_pmcs[ri].pp_pmcval -= tmp;
1438 if ((int64_t) pp->pp_pmcs[ri].pp_pmcval < 0)
1439 pp->pp_pmcs[ri].pp_pmcval +=
1440 pm->pm_sc.pm_reloadcount;
1441 mtx_pool_unlock_spin(pmc_mtxpool, pm);
1446 * For counting process-virtual PMCs,
1447 * we expect the count to be
1448 * increasing monotonically, modulo a 64
1451 KASSERT((int64_t) tmp >= 0,
1452 ("[pmc,%d] negative increment cpu=%d "
1453 "ri=%d newvalue=%jx saved=%jx "
1454 "incr=%jx", __LINE__, cpu, ri,
1455 newvalue, PMC_PCPU_SAVED(cpu,ri), tmp));
1457 mtx_pool_lock_spin(pmc_mtxpool, pm);
1458 pm->pm_gv.pm_savedvalue += tmp;
1459 pp->pp_pmcs[ri].pp_pmcval += tmp;
1460 mtx_pool_unlock_spin(pmc_mtxpool, pm);
1462 if (pm->pm_flags & PMC_F_LOG_PROCCSW)
1463 pmclog_process_proccsw(pm, pp, tmp);
1467 /* mark hardware as free */
1468 pcd->pcd_config_pmc(cpu, adjri, NULL);
1472 * perform any other architecture/cpu dependent thread
1473 * switch out functions.
1476 (void) (*md->pmd_switch_out)(pc, pp);
1482 * A mapping change for a process.
1486 pmc_process_mmap(struct thread *td, struct pmckern_map_in *pkm)
1490 char *fullpath, *freepath;
1491 const struct pmc *pm;
1492 struct pmc_owner *po;
1493 const struct pmc_process *pp;
1495 freepath = fullpath = NULL;
1496 pmc_getfilename((struct vnode *) pkm->pm_file, &fullpath, &freepath);
1498 pid = td->td_proc->p_pid;
1500 /* Inform owners of all system-wide sampling PMCs. */
1501 LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1502 if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1503 pmclog_process_map_in(po, pid, pkm->pm_address, fullpath);
1505 if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL)
1509 * Inform sampling PMC owners tracking this process.
1511 for (ri = 0; ri < md->pmd_npmc; ri++)
1512 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL &&
1513 PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1514 pmclog_process_map_in(pm->pm_owner,
1515 pid, pkm->pm_address, fullpath);
1519 free(freepath, M_TEMP);
1524 * Log an munmap request.
1528 pmc_process_munmap(struct thread *td, struct pmckern_map_out *pkm)
1532 struct pmc_owner *po;
1533 const struct pmc *pm;
1534 const struct pmc_process *pp;
1536 pid = td->td_proc->p_pid;
1538 LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1539 if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1540 pmclog_process_map_out(po, pid, pkm->pm_address,
1541 pkm->pm_address + pkm->pm_size);
1543 if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL)
1546 for (ri = 0; ri < md->pmd_npmc; ri++)
1547 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL &&
1548 PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1549 pmclog_process_map_out(pm->pm_owner, pid,
1550 pkm->pm_address, pkm->pm_address + pkm->pm_size);
1554 * Log mapping information about the kernel.
1558 pmc_log_kernel_mappings(struct pmc *pm)
1560 struct pmc_owner *po;
1561 struct pmckern_map_in *km, *kmbase;
1563 sx_assert(&pmc_sx, SX_LOCKED);
1564 KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)),
1565 ("[pmc,%d] non-sampling PMC (%p) desires mapping information",
1566 __LINE__, (void *) pm));
1570 if (po->po_flags & PMC_PO_INITIAL_MAPPINGS_DONE)
1574 * Log the current set of kernel modules.
1576 kmbase = linker_hwpmc_list_objects();
1577 for (km = kmbase; km->pm_file != NULL; km++) {
1578 PMCDBG(LOG,REG,1,"%s %p", (char *) km->pm_file,
1579 (void *) km->pm_address);
1580 pmclog_process_map_in(po, (pid_t) -1, km->pm_address,
1583 free(kmbase, M_LINKER);
1585 po->po_flags |= PMC_PO_INITIAL_MAPPINGS_DONE;
1589 * Log the mappings for a single process.
1593 pmc_log_process_mappings(struct pmc_owner *po, struct proc *p)
1598 vm_map_entry_t entry;
1599 vm_offset_t last_end;
1600 u_int last_timestamp;
1601 struct vnode *last_vp;
1602 vm_offset_t start_addr;
1603 vm_object_t obj, lobj, tobj;
1604 char *fullpath, *freepath;
1607 last_end = (vm_offset_t) 0;
1608 fullpath = freepath = NULL;
1610 if ((vm = vmspace_acquire_ref(p)) == NULL)
1614 vm_map_lock_read(map);
1616 for (entry = map->header.next; entry != &map->header; entry = entry->next) {
1618 if (entry == NULL) {
1619 PMCDBG(LOG,OPS,2, "hwpmc: vm_map entry unexpectedly "
1620 "NULL! pid=%d vm_map=%p\n", p->p_pid, map);
1625 * We only care about executable map entries.
1627 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) ||
1628 !(entry->protection & VM_PROT_EXECUTE) ||
1629 (entry->object.vm_object == NULL)) {
1633 obj = entry->object.vm_object;
1634 VM_OBJECT_RLOCK(obj);
1637 * Walk the backing_object list to find the base
1638 * (non-shadowed) vm_object.
1640 for (lobj = tobj = obj; tobj != NULL; tobj = tobj->backing_object) {
1642 VM_OBJECT_RLOCK(tobj);
1644 VM_OBJECT_RUNLOCK(lobj);
1649 * At this point lobj is the base vm_object and it is locked.
1652 PMCDBG(LOG,OPS,2, "hwpmc: lobj unexpectedly NULL! pid=%d "
1653 "vm_map=%p vm_obj=%p\n", p->p_pid, map, obj);
1654 VM_OBJECT_RUNLOCK(obj);
1658 if (lobj->type != OBJT_VNODE || lobj->handle == NULL) {
1660 VM_OBJECT_RUNLOCK(lobj);
1661 VM_OBJECT_RUNLOCK(obj);
1666 * Skip contiguous regions that point to the same
1667 * vnode, so we don't emit redundant MAP-IN
1670 if (entry->start == last_end && lobj->handle == last_vp) {
1671 last_end = entry->end;
1673 VM_OBJECT_RUNLOCK(lobj);
1674 VM_OBJECT_RUNLOCK(obj);
1679 * We don't want to keep the proc's vm_map or this
1680 * vm_object locked while we walk the pathname, since
1681 * vn_fullpath() can sleep. However, if we drop the
1682 * lock, it's possible for concurrent activity to
1683 * modify the vm_map list. To protect against this,
1684 * we save the vm_map timestamp before we release the
1685 * lock, and check it after we reacquire the lock
1688 start_addr = entry->start;
1689 last_end = entry->end;
1690 last_timestamp = map->timestamp;
1691 vm_map_unlock_read(map);
1696 VM_OBJECT_RUNLOCK(lobj);
1698 VM_OBJECT_RUNLOCK(obj);
1701 pmc_getfilename(vp, &fullpath, &freepath);
1707 pmclog_process_map_in(po, p->p_pid, start_addr, fullpath);
1709 free(freepath, M_TEMP);
1711 vm_map_lock_read(map);
1714 * If our saved timestamp doesn't match, this means
1715 * that the vm_map was modified out from under us and
1716 * we can't trust our current "entry" pointer. Do a
1717 * new lookup for this entry. If there is no entry
1718 * for this address range, vm_map_lookup_entry() will
1719 * return the previous one, so we always want to go to
1720 * entry->next on the next loop iteration.
1722 * There is an edge condition here that can occur if
1723 * there is no entry at or before this address. In
1724 * this situation, vm_map_lookup_entry returns
1725 * &map->header, which would cause our loop to abort
1726 * without processing the rest of the map. However,
1727 * in practice this will never happen for process
1728 * vm_map. This is because the executable's text
1729 * segment is the first mapping in the proc's address
1730 * space, and this mapping is never removed until the
1731 * process exits, so there will always be a non-header
1732 * entry at or before the requested address for
1733 * vm_map_lookup_entry to return.
1735 if (map->timestamp != last_timestamp)
1736 vm_map_lookup_entry(map, last_end - 1, &entry);
1739 vm_map_unlock_read(map);
1745 * Log mappings for all processes in the system.
1749 pmc_log_all_process_mappings(struct pmc_owner *po)
1751 struct proc *p, *top;
1753 sx_assert(&pmc_sx, SX_XLOCKED);
1755 if ((p = pfind(1)) == NULL)
1756 panic("[pmc,%d] Cannot find init", __LINE__);
1760 sx_slock(&proctree_lock);
1765 pmc_log_process_mappings(po, p);
1766 if (!LIST_EMPTY(&p->p_children))
1767 p = LIST_FIRST(&p->p_children);
1771 if (LIST_NEXT(p, p_sibling)) {
1772 p = LIST_NEXT(p, p_sibling);
1779 sx_sunlock(&proctree_lock);
1783 * The 'hook' invoked from the kernel proper
1788 const char *pmc_hooknames[] = {
1789 /* these strings correspond to PMC_FN_* in <sys/pmckern.h> */
1806 pmc_hook_handler(struct thread *td, int function, void *arg)
1809 PMCDBG(MOD,PMH,1, "hook td=%p func=%d \"%s\" arg=%p", td, function,
1810 pmc_hooknames[function], arg);
1819 case PMC_FN_PROCESS_EXEC:
1821 char *fullpath, *freepath;
1823 int is_using_hwpmcs;
1826 struct pmc_owner *po;
1827 struct pmc_process *pp;
1828 struct pmckern_procexec *pk;
1830 sx_assert(&pmc_sx, SX_XLOCKED);
1833 pmc_getfilename(p->p_textvp, &fullpath, &freepath);
1835 pk = (struct pmckern_procexec *) arg;
1837 /* Inform owners of SS mode PMCs of the exec event. */
1838 LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1839 if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1840 pmclog_process_procexec(po, PMC_ID_INVALID,
1841 p->p_pid, pk->pm_entryaddr, fullpath);
1844 is_using_hwpmcs = p->p_flag & P_HWPMC;
1847 if (!is_using_hwpmcs) {
1849 free(freepath, M_TEMP);
1854 * PMCs are not inherited across an exec(): remove any
1855 * PMCs that this process is the owner of.
1858 if ((po = pmc_find_owner_descriptor(p)) != NULL) {
1859 pmc_remove_owner(po);
1860 pmc_destroy_owner_descriptor(po);
1864 * If the process being exec'ed is not the target of any
1867 if ((pp = pmc_find_process_descriptor(p, 0)) == NULL) {
1869 free(freepath, M_TEMP);
1874 * Log the exec event to all monitoring owners. Skip
1875 * owners who have already recieved the event because
1876 * they had system sampling PMCs active.
1878 for (ri = 0; ri < md->pmd_npmc; ri++)
1879 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) {
1881 if (po->po_sscount == 0 &&
1882 po->po_flags & PMC_PO_OWNS_LOGFILE)
1883 pmclog_process_procexec(po, pm->pm_id,
1884 p->p_pid, pk->pm_entryaddr,
1889 free(freepath, M_TEMP);
1892 PMCDBG(PRC,EXC,1, "exec proc=%p (%d, %s) cred-changed=%d",
1893 p, p->p_pid, p->p_comm, pk->pm_credentialschanged);
1895 if (pk->pm_credentialschanged == 0) /* no change */
1899 * If the newly exec()'ed process has a different credential
1900 * than before, allow it to be the target of a PMC only if
1901 * the PMC's owner has sufficient priviledge.
1904 for (ri = 0; ri < md->pmd_npmc; ri++)
1905 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL)
1906 if (pmc_can_attach(pm, td->td_proc) != 0)
1907 pmc_detach_one_process(td->td_proc,
1910 KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= (int) md->pmd_npmc,
1911 ("[pmc,%d] Illegal ref count %d on pp %p", __LINE__,
1912 pp->pp_refcnt, pp));
1915 * If this process is no longer the target of any
1916 * PMCs, we can remove the process entry and free
1920 if (pp->pp_refcnt == 0) {
1921 pmc_remove_process_descriptor(pp);
1930 pmc_process_csw_in(td);
1933 case PMC_FN_CSW_OUT:
1934 pmc_process_csw_out(td);
1938 * Process accumulated PC samples.
1940 * This function is expected to be called by hardclock() for
1941 * each CPU that has accumulated PC samples.
1943 * This function is to be executed on the CPU whose samples
1944 * are being processed.
1946 case PMC_FN_DO_SAMPLES:
1949 * Clear the cpu specific bit in the CPU mask before
1950 * do the rest of the processing. If the NMI handler
1951 * gets invoked after the "atomic_clear_int()" call
1952 * below but before "pmc_process_samples()" gets
1953 * around to processing the interrupt, then we will
1954 * come back here at the next hardclock() tick (and
1955 * may find nothing to do if "pmc_process_samples()"
1956 * had already processed the interrupt). We don't
1957 * lose the interrupt sample.
1959 CPU_CLR_ATOMIC(PCPU_GET(cpuid), &pmc_cpumask);
1960 pmc_process_samples(PCPU_GET(cpuid), PMC_HR);
1961 pmc_process_samples(PCPU_GET(cpuid), PMC_SR);
1965 sx_assert(&pmc_sx, SX_LOCKED);
1966 pmc_process_mmap(td, (struct pmckern_map_in *) arg);
1970 sx_assert(&pmc_sx, SX_LOCKED);
1971 pmc_process_munmap(td, (struct pmckern_map_out *) arg);
1974 case PMC_FN_USER_CALLCHAIN:
1976 * Record a call chain.
1978 KASSERT(td == curthread, ("[pmc,%d] td != curthread",
1981 pmc_capture_user_callchain(PCPU_GET(cpuid), PMC_HR,
1982 (struct trapframe *) arg);
1983 td->td_pflags &= ~TDP_CALLCHAIN;
1986 case PMC_FN_USER_CALLCHAIN_SOFT:
1988 * Record a call chain.
1990 KASSERT(td == curthread, ("[pmc,%d] td != curthread",
1992 pmc_capture_user_callchain(PCPU_GET(cpuid), PMC_SR,
1993 (struct trapframe *) arg);
1994 td->td_pflags &= ~TDP_CALLCHAIN;
1997 case PMC_FN_SOFT_SAMPLING:
1999 * Call soft PMC sampling intr.
2001 pmc_soft_intr((struct pmckern_soft *) arg);
2006 KASSERT(0, ("[pmc,%d] unknown hook %d\n", __LINE__, function));
2016 * allocate a 'struct pmc_owner' descriptor in the owner hash table.
2019 static struct pmc_owner *
2020 pmc_allocate_owner_descriptor(struct proc *p)
2023 struct pmc_owner *po;
2024 struct pmc_ownerhash *poh;
2026 hindex = PMC_HASH_PTR(p, pmc_ownerhashmask);
2027 poh = &pmc_ownerhash[hindex];
2029 /* allocate space for N pointers and one descriptor struct */
2030 po = malloc(sizeof(struct pmc_owner), M_PMC, M_WAITOK|M_ZERO);
2032 LIST_INSERT_HEAD(poh, po, po_next); /* insert into hash table */
2034 TAILQ_INIT(&po->po_logbuffers);
2035 mtx_init(&po->po_mtx, "pmc-owner-mtx", "pmc-per-proc", MTX_SPIN);
2037 PMCDBG(OWN,ALL,1, "allocate-owner proc=%p (%d, %s) pmc-owner=%p",
2038 p, p->p_pid, p->p_comm, po);
2044 pmc_destroy_owner_descriptor(struct pmc_owner *po)
2047 PMCDBG(OWN,REL,1, "destroy-owner po=%p proc=%p (%d, %s)",
2048 po, po->po_owner, po->po_owner->p_pid, po->po_owner->p_comm);
2050 mtx_destroy(&po->po_mtx);
2055 * find the descriptor corresponding to process 'p', adding or removing it
2056 * as specified by 'mode'.
2059 static struct pmc_process *
2060 pmc_find_process_descriptor(struct proc *p, uint32_t mode)
2063 struct pmc_process *pp, *ppnew;
2064 struct pmc_processhash *pph;
2066 hindex = PMC_HASH_PTR(p, pmc_processhashmask);
2067 pph = &pmc_processhash[hindex];
2072 * Pre-allocate memory in the FIND_ALLOCATE case since we
2073 * cannot call malloc(9) once we hold a spin lock.
2075 if (mode & PMC_FLAG_ALLOCATE)
2076 ppnew = malloc(sizeof(struct pmc_process) + md->pmd_npmc *
2077 sizeof(struct pmc_targetstate), M_PMC, M_WAITOK|M_ZERO);
2079 mtx_lock_spin(&pmc_processhash_mtx);
2080 LIST_FOREACH(pp, pph, pp_next)
2081 if (pp->pp_proc == p)
2084 if ((mode & PMC_FLAG_REMOVE) && pp != NULL)
2085 LIST_REMOVE(pp, pp_next);
2087 if ((mode & PMC_FLAG_ALLOCATE) && pp == NULL &&
2090 LIST_INSERT_HEAD(pph, ppnew, pp_next);
2094 mtx_unlock_spin(&pmc_processhash_mtx);
2096 if (pp != NULL && ppnew != NULL)
2103 * remove a process descriptor from the process hash table.
2107 pmc_remove_process_descriptor(struct pmc_process *pp)
2109 KASSERT(pp->pp_refcnt == 0,
2110 ("[pmc,%d] Removing process descriptor %p with count %d",
2111 __LINE__, pp, pp->pp_refcnt));
2113 mtx_lock_spin(&pmc_processhash_mtx);
2114 LIST_REMOVE(pp, pp_next);
2115 mtx_unlock_spin(&pmc_processhash_mtx);
2120 * find an owner descriptor corresponding to proc 'p'
2123 static struct pmc_owner *
2124 pmc_find_owner_descriptor(struct proc *p)
2127 struct pmc_owner *po;
2128 struct pmc_ownerhash *poh;
2130 hindex = PMC_HASH_PTR(p, pmc_ownerhashmask);
2131 poh = &pmc_ownerhash[hindex];
2134 LIST_FOREACH(po, poh, po_next)
2135 if (po->po_owner == p)
2138 PMCDBG(OWN,FND,1, "find-owner proc=%p (%d, %s) hindex=0x%x -> "
2139 "pmc-owner=%p", p, p->p_pid, p->p_comm, hindex, po);
2145 * pmc_allocate_pmc_descriptor
2147 * Allocate a pmc descriptor and initialize its
2152 pmc_allocate_pmc_descriptor(void)
2156 pmc = malloc(sizeof(struct pmc), M_PMC, M_WAITOK|M_ZERO);
2158 PMCDBG(PMC,ALL,1, "allocate-pmc -> pmc=%p", pmc);
2164 * Destroy a pmc descriptor.
2168 pmc_destroy_pmc_descriptor(struct pmc *pm)
2171 KASSERT(pm->pm_state == PMC_STATE_DELETED ||
2172 pm->pm_state == PMC_STATE_FREE,
2173 ("[pmc,%d] destroying non-deleted PMC", __LINE__));
2174 KASSERT(LIST_EMPTY(&pm->pm_targets),
2175 ("[pmc,%d] destroying pmc with targets", __LINE__));
2176 KASSERT(pm->pm_owner == NULL,
2177 ("[pmc,%d] destroying pmc attached to an owner", __LINE__));
2178 KASSERT(pm->pm_runcount == 0,
2179 ("[pmc,%d] pmc has non-zero run count %d", __LINE__,
2186 pmc_wait_for_pmc_idle(struct pmc *pm)
2189 volatile int maxloop;
2191 maxloop = 100 * pmc_cpu_max();
2194 * Loop (with a forced context switch) till the PMC's runcount
2195 * comes down to zero.
2197 while (atomic_load_acq_32(&pm->pm_runcount) > 0) {
2200 KASSERT(maxloop > 0,
2201 ("[pmc,%d] (ri%d, rc%d) waiting too long for "
2202 "pmc to be free", __LINE__,
2203 PMC_TO_ROWINDEX(pm), pm->pm_runcount));
2205 pmc_force_context_switch();
2210 * This function does the following things:
2212 * - detaches the PMC from hardware
2213 * - unlinks all target threads that were attached to it
2214 * - removes the PMC from its owner's list
2215 * - destroys the PMC private mutex
2217 * Once this function completes, the given pmc pointer can be freed by
2218 * calling pmc_destroy_pmc_descriptor().
2222 pmc_release_pmc_descriptor(struct pmc *pm)
2226 u_int adjri, ri, cpu;
2227 struct pmc_owner *po;
2228 struct pmc_binding pb;
2229 struct pmc_process *pp;
2230 struct pmc_classdep *pcd;
2231 struct pmc_target *ptgt, *tmp;
2233 sx_assert(&pmc_sx, SX_XLOCKED);
2235 KASSERT(pm, ("[pmc,%d] null pmc", __LINE__));
2237 ri = PMC_TO_ROWINDEX(pm);
2238 pcd = pmc_ri_to_classdep(md, ri, &adjri);
2239 mode = PMC_TO_MODE(pm);
2241 PMCDBG(PMC,REL,1, "release-pmc pmc=%p ri=%d mode=%d", pm, ri,
2245 * First, we take the PMC off hardware.
2248 if (PMC_IS_SYSTEM_MODE(mode)) {
2251 * A system mode PMC runs on a specific CPU. Switch
2252 * to this CPU and turn hardware off.
2254 pmc_save_cpu_binding(&pb);
2256 cpu = PMC_TO_CPU(pm);
2258 pmc_select_cpu(cpu);
2260 /* switch off non-stalled CPUs */
2261 if (pm->pm_state == PMC_STATE_RUNNING &&
2262 pm->pm_stalled == 0) {
2264 phw = pmc_pcpu[cpu]->pc_hwpmcs[ri];
2266 KASSERT(phw->phw_pmc == pm,
2267 ("[pmc, %d] pmc ptr ri(%d) hw(%p) pm(%p)",
2268 __LINE__, ri, phw->phw_pmc, pm));
2269 PMCDBG(PMC,REL,2, "stopping cpu=%d ri=%d", cpu, ri);
2272 pcd->pcd_stop_pmc(cpu, adjri);
2276 PMCDBG(PMC,REL,2, "decfg cpu=%d ri=%d", cpu, ri);
2279 pcd->pcd_config_pmc(cpu, adjri, NULL);
2282 /* adjust the global and process count of SS mode PMCs */
2283 if (mode == PMC_MODE_SS && pm->pm_state == PMC_STATE_RUNNING) {
2286 if (po->po_sscount == 0) {
2287 atomic_subtract_rel_int(&pmc_ss_count, 1);
2288 LIST_REMOVE(po, po_ssnext);
2292 pm->pm_state = PMC_STATE_DELETED;
2294 pmc_restore_cpu_binding(&pb);
2297 * We could have references to this PMC structure in
2298 * the per-cpu sample queues. Wait for the queue to
2301 pmc_wait_for_pmc_idle(pm);
2303 } else if (PMC_IS_VIRTUAL_MODE(mode)) {
2306 * A virtual PMC could be running on multiple CPUs at
2309 * By marking its state as DELETED, we ensure that
2310 * this PMC is never further scheduled on hardware.
2312 * Then we wait till all CPUs are done with this PMC.
2314 pm->pm_state = PMC_STATE_DELETED;
2317 /* Wait for the PMCs runcount to come to zero. */
2318 pmc_wait_for_pmc_idle(pm);
2321 * At this point the PMC is off all CPUs and cannot be
2322 * freshly scheduled onto a CPU. It is now safe to
2323 * unlink all targets from this PMC. If a
2324 * process-record's refcount falls to zero, we remove
2325 * it from the hash table. The module-wide SX lock
2326 * protects us from races.
2328 LIST_FOREACH_SAFE(ptgt, &pm->pm_targets, pt_next, tmp) {
2329 pp = ptgt->pt_process;
2330 pmc_unlink_target_process(pm, pp); /* frees 'ptgt' */
2332 PMCDBG(PMC,REL,3, "pp->refcnt=%d", pp->pp_refcnt);
2335 * If the target process record shows that no
2336 * PMCs are attached to it, reclaim its space.
2339 if (pp->pp_refcnt == 0) {
2340 pmc_remove_process_descriptor(pp);
2345 cpu = curthread->td_oncpu; /* setup cpu for pmd_release() */
2350 * Release any MD resources
2352 (void) pcd->pcd_release_pmc(cpu, adjri, pm);
2355 * Update row disposition
2358 if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm)))
2359 PMC_UNMARK_ROW_STANDALONE(ri);
2361 PMC_UNMARK_ROW_THREAD(ri);
2363 /* unlink from the owner's list */
2365 LIST_REMOVE(pm, pm_next);
2366 pm->pm_owner = NULL;
2371 * Register an owner and a pmc.
2375 pmc_register_owner(struct proc *p, struct pmc *pmc)
2377 struct pmc_owner *po;
2379 sx_assert(&pmc_sx, SX_XLOCKED);
2381 if ((po = pmc_find_owner_descriptor(p)) == NULL)
2382 if ((po = pmc_allocate_owner_descriptor(p)) == NULL)
2385 KASSERT(pmc->pm_owner == NULL,
2386 ("[pmc,%d] attempting to own an initialized PMC", __LINE__));
2389 LIST_INSERT_HEAD(&po->po_pmcs, pmc, pm_next);
2392 p->p_flag |= P_HWPMC;
2395 if (po->po_flags & PMC_PO_OWNS_LOGFILE)
2396 pmclog_process_pmcallocate(pmc);
2398 PMCDBG(PMC,REG,1, "register-owner pmc-owner=%p pmc=%p",
2405 * Return the current row disposition:
2407 * > 0 => PROCESS MODE
2408 * < 0 => SYSTEM MODE
2412 pmc_getrowdisp(int ri)
2414 return pmc_pmcdisp[ri];
2418 * Check if a PMC at row index 'ri' can be allocated to the current
2421 * Allocation can fail if:
2422 * - the current process is already being profiled by a PMC at index 'ri',
2423 * attached to it via OP_PMCATTACH.
2424 * - the current process has already allocated a PMC at index 'ri'
2429 pmc_can_allocate_rowindex(struct proc *p, unsigned int ri, int cpu)
2433 struct pmc_owner *po;
2434 struct pmc_process *pp;
2436 PMCDBG(PMC,ALR,1, "can-allocate-rowindex proc=%p (%d, %s) ri=%d "
2437 "cpu=%d", p, p->p_pid, p->p_comm, ri, cpu);
2440 * We shouldn't have already allocated a process-mode PMC at
2443 * We shouldn't have allocated a system-wide PMC on the same
2446 if ((po = pmc_find_owner_descriptor(p)) != NULL)
2447 LIST_FOREACH(pm, &po->po_pmcs, pm_next) {
2448 if (PMC_TO_ROWINDEX(pm) == ri) {
2449 mode = PMC_TO_MODE(pm);
2450 if (PMC_IS_VIRTUAL_MODE(mode))
2452 if (PMC_IS_SYSTEM_MODE(mode) &&
2453 (int) PMC_TO_CPU(pm) == cpu)
2459 * We also shouldn't be the target of any PMC at this index
2460 * since otherwise a PMC_ATTACH to ourselves will fail.
2462 if ((pp = pmc_find_process_descriptor(p, 0)) != NULL)
2463 if (pp->pp_pmcs[ri].pp_pmc)
2466 PMCDBG(PMC,ALR,2, "can-allocate-rowindex proc=%p (%d, %s) ri=%d ok",
2467 p, p->p_pid, p->p_comm, ri);
2473 * Check if a given PMC at row index 'ri' can be currently used in
2478 pmc_can_allocate_row(int ri, enum pmc_mode mode)
2482 sx_assert(&pmc_sx, SX_XLOCKED);
2484 PMCDBG(PMC,ALR,1, "can-allocate-row ri=%d mode=%d", ri, mode);
2486 if (PMC_IS_SYSTEM_MODE(mode))
2487 disp = PMC_DISP_STANDALONE;
2489 disp = PMC_DISP_THREAD;
2492 * check disposition for PMC row 'ri':
2494 * Expected disposition Row-disposition Result
2496 * STANDALONE STANDALONE or FREE proceed
2497 * STANDALONE THREAD fail
2498 * THREAD THREAD or FREE proceed
2499 * THREAD STANDALONE fail
2502 if (!PMC_ROW_DISP_IS_FREE(ri) &&
2503 !(disp == PMC_DISP_THREAD && PMC_ROW_DISP_IS_THREAD(ri)) &&
2504 !(disp == PMC_DISP_STANDALONE && PMC_ROW_DISP_IS_STANDALONE(ri)))
2511 PMCDBG(PMC,ALR,2, "can-allocate-row ri=%d mode=%d ok", ri, mode);
2518 * Find a PMC descriptor with user handle 'pmcid' for thread 'td'.
2522 pmc_find_pmc_descriptor_in_process(struct pmc_owner *po, pmc_id_t pmcid)
2526 KASSERT(PMC_ID_TO_ROWINDEX(pmcid) < md->pmd_npmc,
2527 ("[pmc,%d] Illegal pmc index %d (max %d)", __LINE__,
2528 PMC_ID_TO_ROWINDEX(pmcid), md->pmd_npmc));
2530 LIST_FOREACH(pm, &po->po_pmcs, pm_next)
2531 if (pm->pm_id == pmcid)
2538 pmc_find_pmc(pmc_id_t pmcid, struct pmc **pmc)
2542 struct pmc_owner *po;
2544 PMCDBG(PMC,FND,1, "find-pmc id=%d", pmcid);
2546 if ((po = pmc_find_owner_descriptor(curthread->td_proc)) == NULL)
2549 if ((pm = pmc_find_pmc_descriptor_in_process(po, pmcid)) == NULL)
2552 PMCDBG(PMC,FND,2, "find-pmc id=%d -> pmc=%p", pmcid, pm);
2563 pmc_start(struct pmc *pm)
2566 struct pmc_owner *po;
2567 struct pmc_binding pb;
2568 struct pmc_classdep *pcd;
2569 int adjri, error, cpu, ri;
2572 ("[pmc,%d] null pm", __LINE__));
2574 mode = PMC_TO_MODE(pm);
2575 ri = PMC_TO_ROWINDEX(pm);
2576 pcd = pmc_ri_to_classdep(md, ri, &adjri);
2580 PMCDBG(PMC,OPS,1, "start pmc=%p mode=%d ri=%d", pm, mode, ri);
2585 * Disallow PMCSTART if a logfile is required but has not been
2588 if ((pm->pm_flags & PMC_F_NEEDS_LOGFILE) &&
2589 (po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
2590 return (EDOOFUS); /* programming error */
2593 * If this is a sampling mode PMC, log mapping information for
2594 * the kernel modules that are currently loaded.
2596 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
2597 pmc_log_kernel_mappings(pm);
2599 if (PMC_IS_VIRTUAL_MODE(mode)) {
2602 * If a PMCATTACH has never been done on this PMC,
2603 * attach it to its owner process.
2606 if (LIST_EMPTY(&pm->pm_targets))
2607 error = (pm->pm_flags & PMC_F_ATTACH_DONE) ? ESRCH :
2608 pmc_attach_process(po->po_owner, pm);
2611 * If the PMC is attached to its owner, then force a context
2612 * switch to ensure that the MD state gets set correctly.
2616 pm->pm_state = PMC_STATE_RUNNING;
2617 if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER)
2618 pmc_force_context_switch();
2626 * A system-wide PMC.
2628 * Add the owner to the global list if this is a system-wide
2632 if (mode == PMC_MODE_SS) {
2633 if (po->po_sscount == 0) {
2634 LIST_INSERT_HEAD(&pmc_ss_owners, po, po_ssnext);
2635 atomic_add_rel_int(&pmc_ss_count, 1);
2636 PMCDBG(PMC,OPS,1, "po=%p in global list", po);
2641 * Log mapping information for all existing processes in the
2642 * system. Subsequent mappings are logged as they happen;
2643 * see pmc_process_mmap().
2645 if (po->po_logprocmaps == 0) {
2646 pmc_log_all_process_mappings(po);
2647 po->po_logprocmaps = 1;
2652 * Move to the CPU associated with this
2653 * PMC, and start the hardware.
2656 pmc_save_cpu_binding(&pb);
2658 cpu = PMC_TO_CPU(pm);
2660 if (!pmc_cpu_is_active(cpu))
2663 pmc_select_cpu(cpu);
2666 * global PMCs are configured at allocation time
2667 * so write out the initial value and start the PMC.
2670 pm->pm_state = PMC_STATE_RUNNING;
2673 if ((error = pcd->pcd_write_pmc(cpu, adjri,
2674 PMC_IS_SAMPLING_MODE(mode) ?
2675 pm->pm_sc.pm_reloadcount :
2676 pm->pm_sc.pm_initial)) == 0)
2677 error = pcd->pcd_start_pmc(cpu, adjri);
2680 pmc_restore_cpu_binding(&pb);
2690 pmc_stop(struct pmc *pm)
2692 struct pmc_owner *po;
2693 struct pmc_binding pb;
2694 struct pmc_classdep *pcd;
2695 int adjri, cpu, error, ri;
2697 KASSERT(pm != NULL, ("[pmc,%d] null pmc", __LINE__));
2699 PMCDBG(PMC,OPS,1, "stop pmc=%p mode=%d ri=%d", pm,
2700 PMC_TO_MODE(pm), PMC_TO_ROWINDEX(pm));
2702 pm->pm_state = PMC_STATE_STOPPED;
2705 * If the PMC is a virtual mode one, changing the state to
2706 * non-RUNNING is enough to ensure that the PMC never gets
2709 * If this PMC is current running on a CPU, then it will
2710 * handled correctly at the time its target process is context
2714 if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)))
2718 * A system-mode PMC. Move to the CPU associated with
2719 * this PMC, and stop the hardware. We update the
2720 * 'initial count' so that a subsequent PMCSTART will
2721 * resume counting from the current hardware count.
2724 pmc_save_cpu_binding(&pb);
2726 cpu = PMC_TO_CPU(pm);
2728 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
2729 ("[pmc,%d] illegal cpu=%d", __LINE__, cpu));
2731 if (!pmc_cpu_is_active(cpu))
2734 pmc_select_cpu(cpu);
2736 ri = PMC_TO_ROWINDEX(pm);
2737 pcd = pmc_ri_to_classdep(md, ri, &adjri);
2740 if ((error = pcd->pcd_stop_pmc(cpu, adjri)) == 0)
2741 error = pcd->pcd_read_pmc(cpu, adjri, &pm->pm_sc.pm_initial);
2744 pmc_restore_cpu_binding(&pb);
2748 /* remove this owner from the global list of SS PMC owners */
2749 if (PMC_TO_MODE(pm) == PMC_MODE_SS) {
2751 if (po->po_sscount == 0) {
2752 atomic_subtract_rel_int(&pmc_ss_count, 1);
2753 LIST_REMOVE(po, po_ssnext);
2754 PMCDBG(PMC,OPS,2,"po=%p removed from global list", po);
2763 static const char *pmc_op_to_name[] = {
2765 #define __PMC_OP(N, D) #N ,
2772 * The syscall interface
2775 #define PMC_GET_SX_XLOCK(...) do { \
2776 sx_xlock(&pmc_sx); \
2777 if (pmc_hook == NULL) { \
2778 sx_xunlock(&pmc_sx); \
2779 return __VA_ARGS__; \
2783 #define PMC_DOWNGRADE_SX() do { \
2784 sx_downgrade(&pmc_sx); \
2785 is_sx_downgraded = 1; \
2789 pmc_syscall_handler(struct thread *td, void *syscall_args)
2791 int error, is_sx_downgraded, is_sx_locked, op;
2792 struct pmc_syscall_args *c;
2795 PMC_GET_SX_XLOCK(ENOSYS);
2799 is_sx_downgraded = 0;
2802 c = (struct pmc_syscall_args *) syscall_args;
2807 PMCDBG(MOD,PMS,1, "syscall op=%d \"%s\" arg=%p", op,
2808 pmc_op_to_name[op], arg);
2811 atomic_add_int(&pmc_stats.pm_syscalls, 1);
2818 * Configure a log file.
2820 * XXX This OP will be reworked.
2823 case PMC_OP_CONFIGURELOG:
2827 struct pmc_owner *po;
2828 struct pmc_op_configurelog cl;
2830 sx_assert(&pmc_sx, SX_XLOCKED);
2832 if ((error = copyin(arg, &cl, sizeof(cl))) != 0)
2835 /* mark this process as owning a log file */
2837 if ((po = pmc_find_owner_descriptor(p)) == NULL)
2838 if ((po = pmc_allocate_owner_descriptor(p)) == NULL) {
2844 * If a valid fd was passed in, try to configure that,
2845 * otherwise if 'fd' was less than zero and there was
2846 * a log file configured, flush its buffers and
2849 if (cl.pm_logfd >= 0) {
2850 sx_xunlock(&pmc_sx);
2852 error = pmclog_configure_log(md, po, cl.pm_logfd);
2853 } else if (po->po_flags & PMC_PO_OWNS_LOGFILE) {
2854 pmclog_process_closelog(po);
2855 error = pmclog_close(po);
2857 LIST_FOREACH(pm, &po->po_pmcs, pm_next)
2858 if (pm->pm_flags & PMC_F_NEEDS_LOGFILE &&
2859 pm->pm_state == PMC_STATE_RUNNING)
2861 error = pmclog_deconfigure_log(po);
2875 case PMC_OP_FLUSHLOG:
2877 struct pmc_owner *po;
2879 sx_assert(&pmc_sx, SX_XLOCKED);
2881 if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) {
2886 error = pmclog_flush(po);
2894 case PMC_OP_CLOSELOG:
2896 struct pmc_owner *po;
2898 sx_assert(&pmc_sx, SX_XLOCKED);
2900 if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) {
2905 error = pmclog_close(po);
2910 * Retrieve hardware configuration.
2913 case PMC_OP_GETCPUINFO: /* CPU information */
2915 struct pmc_op_getcpuinfo gci;
2916 struct pmc_classinfo *pci;
2917 struct pmc_classdep *pcd;
2920 gci.pm_cputype = md->pmd_cputype;
2921 gci.pm_ncpu = pmc_cpu_max();
2922 gci.pm_npmc = md->pmd_npmc;
2923 gci.pm_nclass = md->pmd_nclass;
2924 pci = gci.pm_classes;
2925 pcd = md->pmd_classdep;
2926 for (cl = 0; cl < md->pmd_nclass; cl++, pci++, pcd++) {
2927 pci->pm_caps = pcd->pcd_caps;
2928 pci->pm_class = pcd->pcd_class;
2929 pci->pm_width = pcd->pcd_width;
2930 pci->pm_num = pcd->pcd_num;
2932 error = copyout(&gci, arg, sizeof(gci));
2937 * Retrieve soft events list.
2939 case PMC_OP_GETDYNEVENTINFO:
2943 struct pmc_op_getdyneventinfo *gei;
2944 struct pmc_dyn_event_descr dev;
2945 struct pmc_soft *ps;
2948 sx_assert(&pmc_sx, SX_LOCKED);
2950 gei = (struct pmc_op_getdyneventinfo *) arg;
2952 if ((error = copyin(&gei->pm_class, &cl, sizeof(cl))) != 0)
2955 /* Only SOFT class is dynamic. */
2956 if (cl != PMC_CLASS_SOFT) {
2962 for (ev = PMC_EV_SOFT_FIRST; (int)ev <= PMC_EV_SOFT_LAST; ev++) {
2963 ps = pmc_soft_ev_acquire(ev);
2966 bcopy(&ps->ps_ev, &dev, sizeof(dev));
2967 pmc_soft_ev_release(ps);
2969 error = copyout(&dev,
2970 &gei->pm_events[nevent],
2971 sizeof(struct pmc_dyn_event_descr));
2979 error = copyout(&nevent, &gei->pm_nevent,
2985 * Get module statistics
2988 case PMC_OP_GETDRIVERSTATS:
2990 struct pmc_op_getdriverstats gms;
2992 bcopy(&pmc_stats, &gms, sizeof(gms));
2993 error = copyout(&gms, arg, sizeof(gms));
2999 * Retrieve module version number
3002 case PMC_OP_GETMODULEVERSION:
3006 /* retrieve the client's idea of the ABI version */
3007 if ((error = copyin(arg, &cv, sizeof(uint32_t))) != 0)
3009 /* don't service clients newer than our driver */
3011 if ((cv & 0xFFFF0000) > (modv & 0xFFFF0000)) {
3012 error = EPROGMISMATCH;
3015 error = copyout(&modv, arg, sizeof(int));
3021 * Retrieve the state of all the PMCs on a given
3025 case PMC_OP_GETPMCINFO:
3029 size_t pmcinfo_size;
3030 uint32_t cpu, n, npmc;
3031 struct pmc_owner *po;
3032 struct pmc_binding pb;
3033 struct pmc_classdep *pcd;
3034 struct pmc_info *p, *pmcinfo;
3035 struct pmc_op_getpmcinfo *gpi;
3039 gpi = (struct pmc_op_getpmcinfo *) arg;
3041 if ((error = copyin(&gpi->pm_cpu, &cpu, sizeof(cpu))) != 0)
3044 if (cpu >= pmc_cpu_max()) {
3049 if (!pmc_cpu_is_active(cpu)) {
3054 /* switch to CPU 'cpu' */
3055 pmc_save_cpu_binding(&pb);
3056 pmc_select_cpu(cpu);
3058 npmc = md->pmd_npmc;
3060 pmcinfo_size = npmc * sizeof(struct pmc_info);
3061 pmcinfo = malloc(pmcinfo_size, M_PMC, M_WAITOK);
3065 for (n = 0; n < md->pmd_npmc; n++, p++) {
3067 pcd = pmc_ri_to_classdep(md, n, &ari);
3069 KASSERT(pcd != NULL,
3070 ("[pmc,%d] null pcd ri=%d", __LINE__, n));
3072 if ((error = pcd->pcd_describe(cpu, ari, p, &pm)) != 0)
3075 if (PMC_ROW_DISP_IS_STANDALONE(n))
3076 p->pm_rowdisp = PMC_DISP_STANDALONE;
3077 else if (PMC_ROW_DISP_IS_THREAD(n))
3078 p->pm_rowdisp = PMC_DISP_THREAD;
3080 p->pm_rowdisp = PMC_DISP_FREE;
3082 p->pm_ownerpid = -1;
3084 if (pm == NULL) /* no PMC associated */
3089 KASSERT(po->po_owner != NULL,
3090 ("[pmc,%d] pmc_owner had a null proc pointer",
3093 p->pm_ownerpid = po->po_owner->p_pid;
3094 p->pm_mode = PMC_TO_MODE(pm);
3095 p->pm_event = pm->pm_event;
3096 p->pm_flags = pm->pm_flags;
3098 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
3100 pm->pm_sc.pm_reloadcount;
3103 pmc_restore_cpu_binding(&pb);
3105 /* now copy out the PMC info collected */
3107 error = copyout(pmcinfo, &gpi->pm_pmcs, pmcinfo_size);
3109 free(pmcinfo, M_PMC);
3115 * Set the administrative state of a PMC. I.e. whether
3116 * the PMC is to be used or not.
3119 case PMC_OP_PMCADMIN:
3122 enum pmc_state request;
3125 struct pmc_op_pmcadmin pma;
3126 struct pmc_binding pb;
3128 sx_assert(&pmc_sx, SX_XLOCKED);
3130 KASSERT(td == curthread,
3131 ("[pmc,%d] td != curthread", __LINE__));
3133 error = priv_check(td, PRIV_PMC_MANAGE);
3137 if ((error = copyin(arg, &pma, sizeof(pma))) != 0)
3142 if (cpu < 0 || cpu >= (int) pmc_cpu_max()) {
3147 if (!pmc_cpu_is_active(cpu)) {
3152 request = pma.pm_state;
3154 if (request != PMC_STATE_DISABLED &&
3155 request != PMC_STATE_FREE) {
3160 ri = pma.pm_pmc; /* pmc id == row index */
3161 if (ri < 0 || ri >= (int) md->pmd_npmc) {
3167 * We can't disable a PMC with a row-index allocated
3168 * for process virtual PMCs.
3171 if (PMC_ROW_DISP_IS_THREAD(ri) &&
3172 request == PMC_STATE_DISABLED) {
3178 * otherwise, this PMC on this CPU is either free or
3179 * in system-wide mode.
3182 pmc_save_cpu_binding(&pb);
3183 pmc_select_cpu(cpu);
3186 phw = pc->pc_hwpmcs[ri];
3189 * XXX do we need some kind of 'forced' disable?
3192 if (phw->phw_pmc == NULL) {
3193 if (request == PMC_STATE_DISABLED &&
3194 (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED)) {
3195 phw->phw_state &= ~PMC_PHW_FLAG_IS_ENABLED;
3196 PMC_MARK_ROW_STANDALONE(ri);
3197 } else if (request == PMC_STATE_FREE &&
3198 (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0) {
3199 phw->phw_state |= PMC_PHW_FLAG_IS_ENABLED;
3200 PMC_UNMARK_ROW_STANDALONE(ri);
3202 /* other cases are a no-op */
3206 pmc_restore_cpu_binding(&pb);
3215 case PMC_OP_PMCALLOCATE:
3223 struct pmc_binding pb;
3224 struct pmc_classdep *pcd;
3225 struct pmc_op_pmcallocate pa;
3227 if ((error = copyin(arg, &pa, sizeof(pa))) != 0)
3234 if ((mode != PMC_MODE_SS && mode != PMC_MODE_SC &&
3235 mode != PMC_MODE_TS && mode != PMC_MODE_TC) ||
3236 (cpu != (u_int) PMC_CPU_ANY && cpu >= pmc_cpu_max())) {
3242 * Virtual PMCs should only ask for a default CPU.
3243 * System mode PMCs need to specify a non-default CPU.
3246 if ((PMC_IS_VIRTUAL_MODE(mode) && cpu != (u_int) PMC_CPU_ANY) ||
3247 (PMC_IS_SYSTEM_MODE(mode) && cpu == (u_int) PMC_CPU_ANY)) {
3253 * Check that an inactive CPU is not being asked for.
3256 if (PMC_IS_SYSTEM_MODE(mode) && !pmc_cpu_is_active(cpu)) {
3262 * Refuse an allocation for a system-wide PMC if this
3263 * process has been jailed, or if this process lacks
3264 * super-user credentials and the sysctl tunable
3265 * 'security.bsd.unprivileged_syspmcs' is zero.
3268 if (PMC_IS_SYSTEM_MODE(mode)) {
3269 if (jailed(curthread->td_ucred)) {
3273 if (!pmc_unprivileged_syspmcs) {
3274 error = priv_check(curthread,
3282 * Look for valid values for 'pm_flags'
3285 if ((pa.pm_flags & ~(PMC_F_DESCENDANTS | PMC_F_LOG_PROCCSW |
3286 PMC_F_LOG_PROCEXIT | PMC_F_CALLCHAIN)) != 0) {
3291 /* process logging options are not allowed for system PMCs */
3292 if (PMC_IS_SYSTEM_MODE(mode) && (pa.pm_flags &
3293 (PMC_F_LOG_PROCCSW | PMC_F_LOG_PROCEXIT))) {
3299 * All sampling mode PMCs need to be able to interrupt the
3302 if (PMC_IS_SAMPLING_MODE(mode))
3303 caps |= PMC_CAP_INTERRUPT;
3305 /* A valid class specifier should have been passed in. */
3306 for (n = 0; n < md->pmd_nclass; n++)
3307 if (md->pmd_classdep[n].pcd_class == pa.pm_class)
3309 if (n == md->pmd_nclass) {
3314 /* The requested PMC capabilities should be feasible. */
3315 if ((md->pmd_classdep[n].pcd_caps & caps) != caps) {
3320 PMCDBG(PMC,ALL,2, "event=%d caps=0x%x mode=%d cpu=%d",
3321 pa.pm_ev, caps, mode, cpu);
3323 pmc = pmc_allocate_pmc_descriptor();
3324 pmc->pm_id = PMC_ID_MAKE_ID(cpu,pa.pm_mode,pa.pm_class,
3326 pmc->pm_event = pa.pm_ev;
3327 pmc->pm_state = PMC_STATE_FREE;
3328 pmc->pm_caps = caps;
3329 pmc->pm_flags = pa.pm_flags;
3331 /* switch thread to CPU 'cpu' */
3332 pmc_save_cpu_binding(&pb);
3334 #define PMC_IS_SHAREABLE_PMC(cpu, n) \
3335 (pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_state & \
3336 PMC_PHW_FLAG_IS_SHAREABLE)
3337 #define PMC_IS_UNALLOCATED(cpu, n) \
3338 (pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_pmc == NULL)
3340 if (PMC_IS_SYSTEM_MODE(mode)) {
3341 pmc_select_cpu(cpu);
3342 for (n = 0; n < (int) md->pmd_npmc; n++) {
3343 pcd = pmc_ri_to_classdep(md, n, &adjri);
3344 if (pmc_can_allocate_row(n, mode) == 0 &&
3345 pmc_can_allocate_rowindex(
3346 curthread->td_proc, n, cpu) == 0 &&
3347 (PMC_IS_UNALLOCATED(cpu, n) ||
3348 PMC_IS_SHAREABLE_PMC(cpu, n)) &&
3349 pcd->pcd_allocate_pmc(cpu, adjri, pmc,
3354 /* Process virtual mode */
3355 for (n = 0; n < (int) md->pmd_npmc; n++) {
3356 pcd = pmc_ri_to_classdep(md, n, &adjri);
3357 if (pmc_can_allocate_row(n, mode) == 0 &&
3358 pmc_can_allocate_rowindex(
3359 curthread->td_proc, n,
3360 PMC_CPU_ANY) == 0 &&
3361 pcd->pcd_allocate_pmc(curthread->td_oncpu,
3362 adjri, pmc, &pa) == 0)
3367 #undef PMC_IS_UNALLOCATED
3368 #undef PMC_IS_SHAREABLE_PMC
3370 pmc_restore_cpu_binding(&pb);
3372 if (n == (int) md->pmd_npmc) {
3373 pmc_destroy_pmc_descriptor(pmc);
3379 /* Fill in the correct value in the ID field */
3380 pmc->pm_id = PMC_ID_MAKE_ID(cpu,mode,pa.pm_class,n);
3382 PMCDBG(PMC,ALL,2, "ev=%d class=%d mode=%d n=%d -> pmcid=%x",
3383 pmc->pm_event, pa.pm_class, mode, n, pmc->pm_id);
3385 /* Process mode PMCs with logging enabled need log files */
3386 if (pmc->pm_flags & (PMC_F_LOG_PROCEXIT | PMC_F_LOG_PROCCSW))
3387 pmc->pm_flags |= PMC_F_NEEDS_LOGFILE;
3389 /* All system mode sampling PMCs require a log file */
3390 if (PMC_IS_SAMPLING_MODE(mode) && PMC_IS_SYSTEM_MODE(mode))
3391 pmc->pm_flags |= PMC_F_NEEDS_LOGFILE;
3394 * Configure global pmc's immediately
3397 if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pmc))) {
3399 pmc_save_cpu_binding(&pb);
3400 pmc_select_cpu(cpu);
3402 phw = pmc_pcpu[cpu]->pc_hwpmcs[n];
3403 pcd = pmc_ri_to_classdep(md, n, &adjri);
3405 if ((phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0 ||
3406 (error = pcd->pcd_config_pmc(cpu, adjri, pmc)) != 0) {
3407 (void) pcd->pcd_release_pmc(cpu, adjri, pmc);
3408 pmc_destroy_pmc_descriptor(pmc);
3410 pmc_restore_cpu_binding(&pb);
3415 pmc_restore_cpu_binding(&pb);
3418 pmc->pm_state = PMC_STATE_ALLOCATED;
3421 * mark row disposition
3424 if (PMC_IS_SYSTEM_MODE(mode))
3425 PMC_MARK_ROW_STANDALONE(n);
3427 PMC_MARK_ROW_THREAD(n);
3430 * Register this PMC with the current thread as its owner.
3434 pmc_register_owner(curthread->td_proc, pmc)) != 0) {
3435 pmc_release_pmc_descriptor(pmc);
3436 pmc_destroy_pmc_descriptor(pmc);
3442 * Return the allocated index.
3445 pa.pm_pmcid = pmc->pm_id;
3447 error = copyout(&pa, arg, sizeof(pa));
3453 * Attach a PMC to a process.
3456 case PMC_OP_PMCATTACH:
3460 struct pmc_op_pmcattach a;
3462 sx_assert(&pmc_sx, SX_XLOCKED);
3464 if ((error = copyin(arg, &a, sizeof(a))) != 0)
3470 } else if (a.pm_pid == 0)
3471 a.pm_pid = td->td_proc->p_pid;
3473 if ((error = pmc_find_pmc(a.pm_pmc, &pm)) != 0)
3476 if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm))) {
3481 /* PMCs may be (re)attached only when allocated or stopped */
3482 if (pm->pm_state == PMC_STATE_RUNNING) {
3485 } else if (pm->pm_state != PMC_STATE_ALLOCATED &&
3486 pm->pm_state != PMC_STATE_STOPPED) {
3492 if ((p = pfind(a.pm_pid)) == NULL) {
3498 * Ignore processes that are working on exiting.
3500 if (p->p_flag & P_WEXIT) {
3502 PROC_UNLOCK(p); /* pfind() returns a locked process */
3507 * we are allowed to attach a PMC to a process if
3510 error = p_candebug(curthread, p);
3515 error = pmc_attach_process(p, pm);
3521 * Detach an attached PMC from a process.
3524 case PMC_OP_PMCDETACH:
3528 struct pmc_op_pmcattach a;
3530 if ((error = copyin(arg, &a, sizeof(a))) != 0)
3536 } else if (a.pm_pid == 0)
3537 a.pm_pid = td->td_proc->p_pid;
3539 if ((error = pmc_find_pmc(a.pm_pmc, &pm)) != 0)
3542 if ((p = pfind(a.pm_pid)) == NULL) {
3548 * Treat processes that are in the process of exiting
3549 * as if they were not present.
3552 if (p->p_flag & P_WEXIT)
3555 PROC_UNLOCK(p); /* pfind() returns a locked process */
3558 error = pmc_detach_process(p, pm);
3564 * Retrieve the MSR number associated with the counter
3565 * 'pmc_id'. This allows processes to directly use RDPMC
3566 * instructions to read their PMCs, without the overhead of a
3570 case PMC_OP_PMCGETMSR:
3574 struct pmc_target *pt;
3575 struct pmc_op_getmsr gm;
3576 struct pmc_classdep *pcd;
3580 if ((error = copyin(arg, &gm, sizeof(gm))) != 0)
3583 if ((error = pmc_find_pmc(gm.pm_pmcid, &pm)) != 0)
3587 * The allocated PMC has to be a process virtual PMC,
3588 * i.e., of type MODE_T[CS]. Global PMCs can only be
3589 * read using the PMCREAD operation since they may be
3590 * allocated on a different CPU than the one we could
3591 * be running on at the time of the RDPMC instruction.
3593 * The GETMSR operation is not allowed for PMCs that
3594 * are inherited across processes.
3597 if (!PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)) ||
3598 (pm->pm_flags & PMC_F_DESCENDANTS)) {
3604 * It only makes sense to use a RDPMC (or its
3605 * equivalent instruction on non-x86 architectures) on
3606 * a process that has allocated and attached a PMC to
3607 * itself. Conversely the PMC is only allowed to have
3608 * one process attached to it -- its owner.
3611 if ((pt = LIST_FIRST(&pm->pm_targets)) == NULL ||
3612 LIST_NEXT(pt, pt_next) != NULL ||
3613 pt->pt_process->pp_proc != pm->pm_owner->po_owner) {
3618 ri = PMC_TO_ROWINDEX(pm);
3619 pcd = pmc_ri_to_classdep(md, ri, &adjri);
3621 /* PMC class has no 'GETMSR' support */
3622 if (pcd->pcd_get_msr == NULL) {
3627 if ((error = (*pcd->pcd_get_msr)(adjri, &gm.pm_msr)) < 0)
3630 if ((error = copyout(&gm, arg, sizeof(gm))) < 0)
3634 * Mark our process as using MSRs. Update machine
3635 * state using a forced context switch.
3638 pt->pt_process->pp_flags |= PMC_PP_ENABLE_MSR_ACCESS;
3639 pmc_force_context_switch();
3645 * Release an allocated PMC
3648 case PMC_OP_PMCRELEASE:
3652 struct pmc_owner *po;
3653 struct pmc_op_simple sp;
3656 * Find PMC pointer for the named PMC.
3658 * Use pmc_release_pmc_descriptor() to switch off the
3659 * PMC, remove all its target threads, and remove the
3660 * PMC from its owner's list.
3662 * Remove the owner record if this is the last PMC
3668 if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
3671 pmcid = sp.pm_pmcid;
3673 if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
3677 pmc_release_pmc_descriptor(pm);
3678 pmc_maybe_remove_owner(po);
3679 pmc_destroy_pmc_descriptor(pm);
3685 * Read and/or write a PMC.
3693 pmc_value_t oldvalue;
3694 struct pmc_binding pb;
3695 struct pmc_op_pmcrw prw;
3696 struct pmc_classdep *pcd;
3697 struct pmc_op_pmcrw *pprw;
3701 if ((error = copyin(arg, &prw, sizeof(prw))) != 0)
3705 PMCDBG(PMC,OPS,1, "rw id=%d flags=0x%x", prw.pm_pmcid,
3708 /* must have at least one flag set */
3709 if ((prw.pm_flags & (PMC_F_OLDVALUE|PMC_F_NEWVALUE)) == 0) {
3714 /* locate pmc descriptor */
3715 if ((error = pmc_find_pmc(prw.pm_pmcid, &pm)) != 0)
3718 /* Can't read a PMC that hasn't been started. */
3719 if (pm->pm_state != PMC_STATE_ALLOCATED &&
3720 pm->pm_state != PMC_STATE_STOPPED &&
3721 pm->pm_state != PMC_STATE_RUNNING) {
3726 /* writing a new value is allowed only for 'STOPPED' pmcs */
3727 if (pm->pm_state == PMC_STATE_RUNNING &&
3728 (prw.pm_flags & PMC_F_NEWVALUE)) {
3733 if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) {
3736 * If this PMC is attached to its owner (i.e.,
3737 * the process requesting this operation) and
3738 * is running, then attempt to get an
3739 * upto-date reading from hardware for a READ.
3740 * Writes are only allowed when the PMC is
3741 * stopped, so only update the saved value
3744 * If the PMC is not running, or is not
3745 * attached to its owner, read/write to the
3749 ri = PMC_TO_ROWINDEX(pm);
3750 pcd = pmc_ri_to_classdep(md, ri, &adjri);
3752 mtx_pool_lock_spin(pmc_mtxpool, pm);
3753 cpu = curthread->td_oncpu;
3755 if (prw.pm_flags & PMC_F_OLDVALUE) {
3756 if ((pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) &&
3757 (pm->pm_state == PMC_STATE_RUNNING))
3758 error = (*pcd->pcd_read_pmc)(cpu, adjri,
3761 oldvalue = pm->pm_gv.pm_savedvalue;
3763 if (prw.pm_flags & PMC_F_NEWVALUE)
3764 pm->pm_gv.pm_savedvalue = prw.pm_value;
3766 mtx_pool_unlock_spin(pmc_mtxpool, pm);
3768 } else { /* System mode PMCs */
3769 cpu = PMC_TO_CPU(pm);
3770 ri = PMC_TO_ROWINDEX(pm);
3771 pcd = pmc_ri_to_classdep(md, ri, &adjri);
3773 if (!pmc_cpu_is_active(cpu)) {
3778 /* move this thread to CPU 'cpu' */
3779 pmc_save_cpu_binding(&pb);
3780 pmc_select_cpu(cpu);
3783 /* save old value */
3784 if (prw.pm_flags & PMC_F_OLDVALUE)
3785 if ((error = (*pcd->pcd_read_pmc)(cpu, adjri,
3788 /* write out new value */
3789 if (prw.pm_flags & PMC_F_NEWVALUE)
3790 error = (*pcd->pcd_write_pmc)(cpu, adjri,
3794 pmc_restore_cpu_binding(&pb);
3799 pprw = (struct pmc_op_pmcrw *) arg;
3802 if (prw.pm_flags & PMC_F_NEWVALUE)
3803 PMCDBG(PMC,OPS,2, "rw id=%d new %jx -> old %jx",
3804 ri, prw.pm_value, oldvalue);
3805 else if (prw.pm_flags & PMC_F_OLDVALUE)
3806 PMCDBG(PMC,OPS,2, "rw id=%d -> old %jx", ri, oldvalue);
3809 /* return old value if requested */
3810 if (prw.pm_flags & PMC_F_OLDVALUE)
3811 if ((error = copyout(&oldvalue, &pprw->pm_value,
3812 sizeof(prw.pm_value))))
3820 * Set the sampling rate for a sampling mode PMC and the
3821 * initial count for a counting mode PMC.
3824 case PMC_OP_PMCSETCOUNT:
3827 struct pmc_op_pmcsetcount sc;
3831 if ((error = copyin(arg, &sc, sizeof(sc))) != 0)
3834 if ((error = pmc_find_pmc(sc.pm_pmcid, &pm)) != 0)
3837 if (pm->pm_state == PMC_STATE_RUNNING) {
3842 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
3843 pm->pm_sc.pm_reloadcount = sc.pm_count;
3845 pm->pm_sc.pm_initial = sc.pm_count;
3854 case PMC_OP_PMCSTART:
3858 struct pmc_op_simple sp;
3860 sx_assert(&pmc_sx, SX_XLOCKED);
3862 if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
3865 pmcid = sp.pm_pmcid;
3867 if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
3870 KASSERT(pmcid == pm->pm_id,
3871 ("[pmc,%d] pmcid %x != id %x", __LINE__,
3874 if (pm->pm_state == PMC_STATE_RUNNING) /* already running */
3876 else if (pm->pm_state != PMC_STATE_STOPPED &&
3877 pm->pm_state != PMC_STATE_ALLOCATED) {
3882 error = pmc_start(pm);
3891 case PMC_OP_PMCSTOP:
3895 struct pmc_op_simple sp;
3899 if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
3902 pmcid = sp.pm_pmcid;
3905 * Mark the PMC as inactive and invoke the MD stop
3906 * routines if needed.
3909 if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
3912 KASSERT(pmcid == pm->pm_id,
3913 ("[pmc,%d] pmc id %x != pmcid %x", __LINE__,
3916 if (pm->pm_state == PMC_STATE_STOPPED) /* already stopped */
3918 else if (pm->pm_state != PMC_STATE_RUNNING) {
3923 error = pmc_stop(pm);
3929 * Write a user supplied value to the log file.
3932 case PMC_OP_WRITELOG:
3934 struct pmc_op_writelog wl;
3935 struct pmc_owner *po;
3939 if ((error = copyin(arg, &wl, sizeof(wl))) != 0)
3942 if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) {
3947 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) {
3952 error = pmclog_process_userlog(po, &wl);
3962 if (is_sx_locked != 0) {
3963 if (is_sx_downgraded)
3964 sx_sunlock(&pmc_sx);
3966 sx_xunlock(&pmc_sx);
3970 atomic_add_int(&pmc_stats.pm_syscall_errors, 1);
3983 * Mark the thread as needing callchain capture and post an AST. The
3984 * actual callchain capture will be done in a context where it is safe
3985 * to take page faults.
3989 pmc_post_callchain_callback(void)
3996 * If there is multiple PMCs for the same interrupt ignore new post
3998 if (td->td_pflags & TDP_CALLCHAIN)
4002 * Mark this thread as needing callchain capture.
4003 * `td->td_pflags' will be safe to touch because this thread
4004 * was in user space when it was interrupted.
4006 td->td_pflags |= TDP_CALLCHAIN;
4009 * Don't let this thread migrate between CPUs until callchain
4010 * capture completes.
4018 * Interrupt processing.
4020 * Find a free slot in the per-cpu array of samples and capture the
4021 * current callchain there. If a sample was successfully added, a bit
4022 * is set in mask 'pmc_cpumask' denoting that the DO_SAMPLES hook
4023 * needs to be invoked from the clock handler.
4025 * This function is meant to be called from an NMI handler. It cannot
4026 * use any of the locking primitives supplied by the OS.
4030 pmc_process_interrupt(int cpu, int ring, struct pmc *pm, struct trapframe *tf,
4033 int error, callchaindepth;
4035 struct pmc_sample *ps;
4036 struct pmc_samplebuffer *psb;
4041 * Allocate space for a sample buffer.
4043 psb = pmc_pcpu[cpu]->pc_sb[ring];
4046 if (ps->ps_nsamples) { /* in use, reader hasn't caught up */
4048 atomic_add_int(&pmc_stats.pm_intr_bufferfull, 1);
4049 PMCDBG(SAM,INT,1,"(spc) cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d",
4050 cpu, pm, (void *) tf, inuserspace,
4051 (int) (psb->ps_write - psb->ps_samples),
4052 (int) (psb->ps_read - psb->ps_samples));
4058 /* Fill in entry. */
4059 PMCDBG(SAM,INT,1,"cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d", cpu, pm,
4060 (void *) tf, inuserspace,
4061 (int) (psb->ps_write - psb->ps_samples),
4062 (int) (psb->ps_read - psb->ps_samples));
4064 KASSERT(pm->pm_runcount >= 0,
4065 ("[pmc,%d] pm=%p runcount %d", __LINE__, (void *) pm,
4068 atomic_add_rel_int(&pm->pm_runcount, 1); /* hold onto PMC */
4071 if ((td = curthread) && td->td_proc)
4072 ps->ps_pid = td->td_proc->p_pid;
4077 ps->ps_flags = inuserspace ? PMC_CC_F_USERSPACE : 0;
4079 callchaindepth = (pm->pm_flags & PMC_F_CALLCHAIN) ?
4080 pmc_callchaindepth : 1;
4082 if (callchaindepth == 1)
4083 ps->ps_pc[0] = PMC_TRAPFRAME_TO_PC(tf);
4086 * Kernel stack traversals can be done immediately,
4087 * while we defer to an AST for user space traversals.
4091 pmc_save_kernel_callchain(ps->ps_pc,
4092 callchaindepth, tf);
4094 pmc_post_callchain_callback();
4095 callchaindepth = PMC_SAMPLE_INUSE;
4099 ps->ps_nsamples = callchaindepth; /* mark entry as in use */
4101 /* increment write pointer, modulo ring buffer size */
4103 if (ps == psb->ps_fence)
4104 psb->ps_write = psb->ps_samples;
4109 /* mark CPU as needing processing */
4110 CPU_SET_ATOMIC(cpu, &pmc_cpumask);
4116 * Capture a user call chain. This function will be called from ast()
4117 * before control returns to userland and before the process gets
4122 pmc_capture_user_callchain(int cpu, int ring, struct trapframe *tf)
4127 struct pmc_sample *ps;
4128 struct pmc_samplebuffer *psb;
4133 psb = pmc_pcpu[cpu]->pc_sb[ring];
4136 KASSERT(td->td_pflags & TDP_CALLCHAIN,
4137 ("[pmc,%d] Retrieving callchain for thread that doesn't want it",
4145 * Iterate through all deferred callchain requests.
4148 ps = psb->ps_samples;
4149 for (i = 0; i < pmc_nsamples; i++, ps++) {
4151 if (ps->ps_nsamples != PMC_SAMPLE_INUSE)
4153 if (ps->ps_td != td)
4156 KASSERT(ps->ps_cpu == cpu,
4157 ("[pmc,%d] cpu mismatch ps_cpu=%d pcpu=%d", __LINE__,
4158 ps->ps_cpu, PCPU_GET(cpuid)));
4162 KASSERT(pm->pm_flags & PMC_F_CALLCHAIN,
4163 ("[pmc,%d] Retrieving callchain for PMC that doesn't "
4164 "want it", __LINE__));
4166 KASSERT(pm->pm_runcount > 0,
4167 ("[pmc,%d] runcount %d", __LINE__, pm->pm_runcount));
4170 * Retrieve the callchain and mark the sample buffer
4171 * as 'processable' by the timer tick sweep code.
4173 ps->ps_nsamples = pmc_save_user_callchain(ps->ps_pc,
4174 pmc_callchaindepth, tf);
4181 KASSERT(ncallchains > 0,
4182 ("[pmc,%d] cpu %d didn't find a sample to collect", __LINE__,
4185 KASSERT(td->td_pinned == 1,
4186 ("[pmc,%d] invalid td_pinned value", __LINE__));
4187 sched_unpin(); /* Can migrate safely now. */
4193 * Process saved PC samples.
4197 pmc_process_samples(int cpu, int ring)
4202 struct pmc_owner *po;
4203 struct pmc_sample *ps;
4204 struct pmc_classdep *pcd;
4205 struct pmc_samplebuffer *psb;
4207 KASSERT(PCPU_GET(cpuid) == cpu,
4208 ("[pmc,%d] not on the correct CPU pcpu=%d cpu=%d", __LINE__,
4209 PCPU_GET(cpuid), cpu));
4211 psb = pmc_pcpu[cpu]->pc_sb[ring];
4213 for (n = 0; n < pmc_nsamples; n++) { /* bound on #iterations */
4216 if (ps->ps_nsamples == PMC_SAMPLE_FREE)
4221 KASSERT(pm->pm_runcount > 0,
4222 ("[pmc,%d] pm=%p runcount %d", __LINE__, (void *) pm,
4227 KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)),
4228 ("[pmc,%d] pmc=%p non-sampling mode=%d", __LINE__,
4229 pm, PMC_TO_MODE(pm)));
4231 /* Ignore PMCs that have been switched off */
4232 if (pm->pm_state != PMC_STATE_RUNNING)
4235 /* If there is a pending AST wait for completion */
4236 if (ps->ps_nsamples == PMC_SAMPLE_INUSE) {
4237 /* Need a rescan at a later time. */
4238 CPU_SET_ATOMIC(cpu, &pmc_cpumask);
4242 PMCDBG(SAM,OPS,1,"cpu=%d pm=%p n=%d fl=%x wr=%d rd=%d", cpu,
4243 pm, ps->ps_nsamples, ps->ps_flags,
4244 (int) (psb->ps_write - psb->ps_samples),
4245 (int) (psb->ps_read - psb->ps_samples));
4248 * If this is a process-mode PMC that is attached to
4249 * its owner, and if the PC is in user mode, update
4250 * profiling statistics like timer-based profiling
4253 if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) {
4254 if (ps->ps_flags & PMC_CC_F_USERSPACE) {
4255 td = FIRST_THREAD_IN_PROC(po->po_owner);
4256 addupc_intr(td, ps->ps_pc[0], 1);
4262 * Otherwise, this is either a sampling mode PMC that
4263 * is attached to a different process than its owner,
4264 * or a system-wide sampling PMC. Dispatch a log
4265 * entry to the PMC's owner process.
4267 pmclog_process_callchain(pm, ps);
4270 ps->ps_nsamples = 0; /* mark entry as free */
4271 atomic_subtract_rel_int(&pm->pm_runcount, 1);
4273 /* increment read pointer, modulo sample size */
4274 if (++ps == psb->ps_fence)
4275 psb->ps_read = psb->ps_samples;
4280 atomic_add_int(&pmc_stats.pm_log_sweeps, 1);
4282 /* Do not re-enable stalled PMCs if we failed to process any samples */
4287 * Restart any stalled sampling PMCs on this CPU.
4289 * If the NMI handler sets the pm_stalled field of a PMC after
4290 * the check below, we'll end up processing the stalled PMC at
4291 * the next hardclock tick.
4293 for (n = 0; n < md->pmd_npmc; n++) {
4294 pcd = pmc_ri_to_classdep(md, n, &adjri);
4295 KASSERT(pcd != NULL,
4296 ("[pmc,%d] null pcd ri=%d", __LINE__, n));
4297 (void) (*pcd->pcd_get_config)(cpu,adjri,&pm);
4299 if (pm == NULL || /* !cfg'ed */
4300 pm->pm_state != PMC_STATE_RUNNING || /* !active */
4301 !PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) || /* !sampling */
4302 pm->pm_stalled == 0) /* !stalled */
4306 (*pcd->pcd_start_pmc)(cpu, adjri);
4315 * Handle a process exit.
4317 * Remove this process from all hash tables. If this process
4318 * owned any PMCs, turn off those PMCs and deallocate them,
4319 * removing any associations with target processes.
4321 * This function will be called by the last 'thread' of a
4324 * XXX This eventhandler gets called early in the exit process.
4325 * Consider using a 'hook' invocation from thread_exit() or equivalent
4326 * spot. Another negative is that kse_exit doesn't seem to call
4332 pmc_process_exit(void *arg __unused, struct proc *p)
4337 int is_using_hwpmcs;
4338 struct pmc_owner *po;
4339 struct pmc_process *pp;
4340 struct pmc_classdep *pcd;
4341 pmc_value_t newvalue, tmp;
4344 is_using_hwpmcs = p->p_flag & P_HWPMC;
4348 * Log a sysexit event to all SS PMC owners.
4350 LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
4351 if (po->po_flags & PMC_PO_OWNS_LOGFILE)
4352 pmclog_process_sysexit(po, p->p_pid);
4354 if (!is_using_hwpmcs)
4358 PMCDBG(PRC,EXT,1,"process-exit proc=%p (%d, %s)", p, p->p_pid,
4362 * Since this code is invoked by the last thread in an exiting
4363 * process, we would have context switched IN at some prior
4364 * point. However, with PREEMPTION, kernel mode context
4365 * switches may happen any time, so we want to disable a
4366 * context switch OUT till we get any PMCs targetting this
4367 * process off the hardware.
4369 * We also need to atomically remove this process'
4370 * entry from our target process hash table, using
4373 PMCDBG(PRC,EXT,1, "process-exit proc=%p (%d, %s)", p, p->p_pid,
4376 critical_enter(); /* no preemption */
4378 cpu = curthread->td_oncpu;
4380 if ((pp = pmc_find_process_descriptor(p,
4381 PMC_FLAG_REMOVE)) != NULL) {
4384 "process-exit proc=%p pmc-process=%p", p, pp);
4387 * The exiting process could the target of
4388 * some PMCs which will be running on
4389 * currently executing CPU.
4391 * We need to turn these PMCs off like we
4392 * would do at context switch OUT time.
4394 for (ri = 0; ri < md->pmd_npmc; ri++) {
4397 * Pick up the pmc pointer from hardware
4398 * state similar to the CSW_OUT code.
4402 pcd = pmc_ri_to_classdep(md, ri, &adjri);
4404 (void) (*pcd->pcd_get_config)(cpu, adjri, &pm);
4406 PMCDBG(PRC,EXT,2, "ri=%d pm=%p", ri, pm);
4409 !PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)))
4412 PMCDBG(PRC,EXT,2, "ppmcs[%d]=%p pm=%p "
4413 "state=%d", ri, pp->pp_pmcs[ri].pp_pmc,
4416 KASSERT(PMC_TO_ROWINDEX(pm) == ri,
4417 ("[pmc,%d] ri mismatch pmc(%d) ri(%d)",
4418 __LINE__, PMC_TO_ROWINDEX(pm), ri));
4420 KASSERT(pm == pp->pp_pmcs[ri].pp_pmc,
4421 ("[pmc,%d] pm %p != pp_pmcs[%d] %p",
4422 __LINE__, pm, ri, pp->pp_pmcs[ri].pp_pmc));
4424 (void) pcd->pcd_stop_pmc(cpu, adjri);
4426 KASSERT(pm->pm_runcount > 0,
4427 ("[pmc,%d] bad runcount ri %d rc %d",
4428 __LINE__, ri, pm->pm_runcount));
4430 /* Stop hardware only if it is actually running */
4431 if (pm->pm_state == PMC_STATE_RUNNING &&
4432 pm->pm_stalled == 0) {
4433 pcd->pcd_read_pmc(cpu, adjri, &newvalue);
4435 PMC_PCPU_SAVED(cpu,ri);
4437 mtx_pool_lock_spin(pmc_mtxpool, pm);
4438 pm->pm_gv.pm_savedvalue += tmp;
4439 pp->pp_pmcs[ri].pp_pmcval += tmp;
4440 mtx_pool_unlock_spin(pmc_mtxpool, pm);
4443 atomic_subtract_rel_int(&pm->pm_runcount,1);
4445 KASSERT((int) pm->pm_runcount >= 0,
4446 ("[pmc,%d] runcount is %d", __LINE__, ri));
4448 (void) pcd->pcd_config_pmc(cpu, adjri, NULL);
4452 * Inform the MD layer of this pseudo "context switch
4455 (void) md->pmd_switch_out(pmc_pcpu[cpu], pp);
4457 critical_exit(); /* ok to be pre-empted now */
4460 * Unlink this process from the PMCs that are
4461 * targetting it. This will send a signal to
4462 * all PMC owner's whose PMCs are orphaned.
4464 * Log PMC value at exit time if requested.
4466 for (ri = 0; ri < md->pmd_npmc; ri++)
4467 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) {
4468 if (pm->pm_flags & PMC_F_NEEDS_LOGFILE &&
4469 PMC_IS_COUNTING_MODE(PMC_TO_MODE(pm)))
4470 pmclog_process_procexit(pm, pp);
4471 pmc_unlink_target_process(pm, pp);
4476 critical_exit(); /* pp == NULL */
4480 * If the process owned PMCs, free them up and free up
4483 if ((po = pmc_find_owner_descriptor(p)) != NULL) {
4484 pmc_remove_owner(po);
4485 pmc_destroy_owner_descriptor(po);
4488 sx_xunlock(&pmc_sx);
4492 * Handle a process fork.
4494 * If the parent process 'p1' is under HWPMC monitoring, then copy
4495 * over any attached PMCs that have 'do_descendants' semantics.
4499 pmc_process_fork(void *arg __unused, struct proc *p1, struct proc *newproc,
4502 int is_using_hwpmcs;
4504 uint32_t do_descendants;
4506 struct pmc_owner *po;
4507 struct pmc_process *ppnew, *ppold;
4509 (void) flags; /* unused parameter */
4512 is_using_hwpmcs = p1->p_flag & P_HWPMC;
4516 * If there are system-wide sampling PMCs active, we need to
4517 * log all fork events to their owner's logs.
4520 LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
4521 if (po->po_flags & PMC_PO_OWNS_LOGFILE)
4522 pmclog_process_procfork(po, p1->p_pid, newproc->p_pid);
4524 if (!is_using_hwpmcs)
4528 PMCDBG(PMC,FRK,1, "process-fork proc=%p (%d, %s) -> %p", p1,
4529 p1->p_pid, p1->p_comm, newproc);
4532 * If the parent process (curthread->td_proc) is a
4533 * target of any PMCs, look for PMCs that are to be
4534 * inherited, and link these into the new process
4537 if ((ppold = pmc_find_process_descriptor(curthread->td_proc,
4538 PMC_FLAG_NONE)) == NULL)
4539 goto done; /* nothing to do */
4542 for (ri = 0; ri < md->pmd_npmc; ri++)
4543 if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL)
4544 do_descendants |= pm->pm_flags & PMC_F_DESCENDANTS;
4545 if (do_descendants == 0) /* nothing to do */
4548 /* allocate a descriptor for the new process */
4549 if ((ppnew = pmc_find_process_descriptor(newproc,
4550 PMC_FLAG_ALLOCATE)) == NULL)
4554 * Run through all PMCs that were targeting the old process
4555 * and which specified F_DESCENDANTS and attach them to the
4558 * Log the fork event to all owners of PMCs attached to this
4559 * process, if not already logged.
4561 for (ri = 0; ri < md->pmd_npmc; ri++)
4562 if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL &&
4563 (pm->pm_flags & PMC_F_DESCENDANTS)) {
4564 pmc_link_target_process(pm, ppnew);
4566 if (po->po_sscount == 0 &&
4567 po->po_flags & PMC_PO_OWNS_LOGFILE)
4568 pmclog_process_procfork(po, p1->p_pid,
4573 * Now mark the new process as being tracked by this driver.
4576 newproc->p_flag |= P_HWPMC;
4577 PROC_UNLOCK(newproc);
4580 sx_xunlock(&pmc_sx);
4584 pmc_kld_load(void *arg __unused, linker_file_t lf)
4586 struct pmc_owner *po;
4591 * Notify owners of system sampling PMCs about KLD operations.
4593 LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
4594 if (po->po_flags & PMC_PO_OWNS_LOGFILE)
4595 pmclog_process_map_in(po, (pid_t) -1,
4596 (uintfptr_t) lf->address, lf->filename);
4599 * TODO: Notify owners of (all) process-sampling PMCs too.
4602 sx_sunlock(&pmc_sx);
4606 pmc_kld_unload(void *arg __unused, const char *filename __unused,
4607 caddr_t address, size_t size)
4609 struct pmc_owner *po;
4613 LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
4614 if (po->po_flags & PMC_PO_OWNS_LOGFILE)
4615 pmclog_process_map_out(po, (pid_t) -1,
4616 (uintfptr_t) address, (uintfptr_t) address + size);
4619 * TODO: Notify owners of process-sampling PMCs.
4622 sx_sunlock(&pmc_sx);
4629 static const char *pmc_name_of_pmcclass[] = {
4631 #define __PMC_CLASS(N) #N ,
4636 * Base class initializer: allocate structure and set default classes.
4639 pmc_mdep_alloc(int nclasses)
4641 struct pmc_mdep *md;
4644 /* SOFT + md classes */
4646 md = malloc(sizeof(struct pmc_mdep) + n *
4647 sizeof(struct pmc_classdep), M_PMC, M_WAITOK|M_ZERO);
4650 /* Add base class. */
4651 pmc_soft_initialize(md);
4656 pmc_mdep_free(struct pmc_mdep *md)
4658 pmc_soft_finalize(md);
4663 generic_switch_in(struct pmc_cpu *pc, struct pmc_process *pp)
4665 (void) pc; (void) pp;
4671 generic_switch_out(struct pmc_cpu *pc, struct pmc_process *pp)
4673 (void) pc; (void) pp;
4678 static struct pmc_mdep *
4679 pmc_generic_cpu_initialize(void)
4681 struct pmc_mdep *md;
4683 md = pmc_mdep_alloc(0);
4685 md->pmd_cputype = PMC_CPU_GENERIC;
4687 md->pmd_pcpu_init = NULL;
4688 md->pmd_pcpu_fini = NULL;
4689 md->pmd_switch_in = generic_switch_in;
4690 md->pmd_switch_out = generic_switch_out;
4696 pmc_generic_cpu_finalize(struct pmc_mdep *md)
4703 pmc_initialize(void)
4705 int c, cpu, error, n, ri;
4706 unsigned int maxcpu;
4707 struct pmc_binding pb;
4708 struct pmc_sample *ps;
4709 struct pmc_classdep *pcd;
4710 struct pmc_samplebuffer *sb;
4716 /* parse debug flags first */
4717 if (TUNABLE_STR_FETCH(PMC_SYSCTL_NAME_PREFIX "debugflags",
4718 pmc_debugstr, sizeof(pmc_debugstr)))
4719 pmc_debugflags_parse(pmc_debugstr,
4720 pmc_debugstr+strlen(pmc_debugstr));
4723 PMCDBG(MOD,INI,0, "PMC Initialize (version %x)", PMC_VERSION);
4725 /* check kernel version */
4726 if (pmc_kernel_version != PMC_VERSION) {
4727 if (pmc_kernel_version == 0)
4728 printf("hwpmc: this kernel has not been compiled with "
4729 "'options HWPMC_HOOKS'.\n");
4731 printf("hwpmc: kernel version (0x%x) does not match "
4732 "module version (0x%x).\n", pmc_kernel_version,
4734 return EPROGMISMATCH;
4738 * check sysctl parameters
4741 if (pmc_hashsize <= 0) {
4742 (void) printf("hwpmc: tunable \"hashsize\"=%d must be "
4743 "greater than zero.\n", pmc_hashsize);
4744 pmc_hashsize = PMC_HASH_SIZE;
4747 if (pmc_nsamples <= 0 || pmc_nsamples > 65535) {
4748 (void) printf("hwpmc: tunable \"nsamples\"=%d out of "
4749 "range.\n", pmc_nsamples);
4750 pmc_nsamples = PMC_NSAMPLES;
4753 if (pmc_callchaindepth <= 0 ||
4754 pmc_callchaindepth > PMC_CALLCHAIN_DEPTH_MAX) {
4755 (void) printf("hwpmc: tunable \"callchaindepth\"=%d out of "
4756 "range.\n", pmc_callchaindepth);
4757 pmc_callchaindepth = PMC_CALLCHAIN_DEPTH;
4760 md = pmc_md_initialize();
4762 /* Default to generic CPU. */
4763 md = pmc_generic_cpu_initialize();
4768 KASSERT(md->pmd_nclass >= 1 && md->pmd_npmc >= 1,
4769 ("[pmc,%d] no classes or pmcs", __LINE__));
4771 /* Compute the map from row-indices to classdep pointers. */
4772 pmc_rowindex_to_classdep = malloc(sizeof(struct pmc_classdep *) *
4773 md->pmd_npmc, M_PMC, M_WAITOK|M_ZERO);
4775 for (n = 0; n < md->pmd_npmc; n++)
4776 pmc_rowindex_to_classdep[n] = NULL;
4777 for (ri = c = 0; c < md->pmd_nclass; c++) {
4778 pcd = &md->pmd_classdep[c];
4779 for (n = 0; n < pcd->pcd_num; n++, ri++)
4780 pmc_rowindex_to_classdep[ri] = pcd;
4783 KASSERT(ri == md->pmd_npmc,
4784 ("[pmc,%d] npmc miscomputed: ri=%d, md->npmc=%d", __LINE__,
4787 maxcpu = pmc_cpu_max();
4789 /* allocate space for the per-cpu array */
4790 pmc_pcpu = malloc(maxcpu * sizeof(struct pmc_cpu *), M_PMC,
4793 /* per-cpu 'saved values' for managing process-mode PMCs */
4794 pmc_pcpu_saved = malloc(sizeof(pmc_value_t) * maxcpu * md->pmd_npmc,
4797 /* Perform CPU-dependent initialization. */
4798 pmc_save_cpu_binding(&pb);
4800 for (cpu = 0; error == 0 && cpu < maxcpu; cpu++) {
4801 if (!pmc_cpu_is_active(cpu))
4803 pmc_select_cpu(cpu);
4804 pmc_pcpu[cpu] = malloc(sizeof(struct pmc_cpu) +
4805 md->pmd_npmc * sizeof(struct pmc_hw *), M_PMC,
4807 if (md->pmd_pcpu_init)
4808 error = md->pmd_pcpu_init(md, cpu);
4809 for (n = 0; error == 0 && n < md->pmd_nclass; n++)
4810 error = md->pmd_classdep[n].pcd_pcpu_init(md, cpu);
4812 pmc_restore_cpu_binding(&pb);
4817 /* allocate space for the sample array */
4818 for (cpu = 0; cpu < maxcpu; cpu++) {
4819 if (!pmc_cpu_is_active(cpu))
4822 sb = malloc(sizeof(struct pmc_samplebuffer) +
4823 pmc_nsamples * sizeof(struct pmc_sample), M_PMC,
4825 sb->ps_read = sb->ps_write = sb->ps_samples;
4826 sb->ps_fence = sb->ps_samples + pmc_nsamples;
4828 KASSERT(pmc_pcpu[cpu] != NULL,
4829 ("[pmc,%d] cpu=%d Null per-cpu data", __LINE__, cpu));
4831 sb->ps_callchains = malloc(pmc_callchaindepth * pmc_nsamples *
4832 sizeof(uintptr_t), M_PMC, M_WAITOK|M_ZERO);
4834 for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++)
4835 ps->ps_pc = sb->ps_callchains +
4836 (n * pmc_callchaindepth);
4838 pmc_pcpu[cpu]->pc_sb[PMC_HR] = sb;
4840 sb = malloc(sizeof(struct pmc_samplebuffer) +
4841 pmc_nsamples * sizeof(struct pmc_sample), M_PMC,
4843 sb->ps_read = sb->ps_write = sb->ps_samples;
4844 sb->ps_fence = sb->ps_samples + pmc_nsamples;
4846 KASSERT(pmc_pcpu[cpu] != NULL,
4847 ("[pmc,%d] cpu=%d Null per-cpu data", __LINE__, cpu));
4849 sb->ps_callchains = malloc(pmc_callchaindepth * pmc_nsamples *
4850 sizeof(uintptr_t), M_PMC, M_WAITOK|M_ZERO);
4852 for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++)
4853 ps->ps_pc = sb->ps_callchains +
4854 (n * pmc_callchaindepth);
4856 pmc_pcpu[cpu]->pc_sb[PMC_SR] = sb;
4859 /* allocate space for the row disposition array */
4860 pmc_pmcdisp = malloc(sizeof(enum pmc_mode) * md->pmd_npmc,
4861 M_PMC, M_WAITOK|M_ZERO);
4863 /* mark all PMCs as available */
4864 for (n = 0; n < (int) md->pmd_npmc; n++)
4865 PMC_MARK_ROW_FREE(n);
4867 /* allocate thread hash tables */
4868 pmc_ownerhash = hashinit(pmc_hashsize, M_PMC,
4869 &pmc_ownerhashmask);
4871 pmc_processhash = hashinit(pmc_hashsize, M_PMC,
4872 &pmc_processhashmask);
4873 mtx_init(&pmc_processhash_mtx, "pmc-process-hash", "pmc-leaf",
4876 LIST_INIT(&pmc_ss_owners);
4879 /* allocate a pool of spin mutexes */
4880 pmc_mtxpool = mtx_pool_create("pmc-leaf", pmc_mtxpool_size,
4883 PMCDBG(MOD,INI,1, "pmc_ownerhash=%p, mask=0x%lx "
4884 "targethash=%p mask=0x%lx", pmc_ownerhash, pmc_ownerhashmask,
4885 pmc_processhash, pmc_processhashmask);
4887 /* register process {exit,fork,exec} handlers */
4888 pmc_exit_tag = EVENTHANDLER_REGISTER(process_exit,
4889 pmc_process_exit, NULL, EVENTHANDLER_PRI_ANY);
4890 pmc_fork_tag = EVENTHANDLER_REGISTER(process_fork,
4891 pmc_process_fork, NULL, EVENTHANDLER_PRI_ANY);
4893 /* register kld event handlers */
4894 pmc_kld_load_tag = EVENTHANDLER_REGISTER(kld_load, pmc_kld_load,
4895 NULL, EVENTHANDLER_PRI_ANY);
4896 pmc_kld_unload_tag = EVENTHANDLER_REGISTER(kld_unload, pmc_kld_unload,
4897 NULL, EVENTHANDLER_PRI_ANY);
4899 /* initialize logging */
4900 pmclog_initialize();
4902 /* set hook functions */
4903 pmc_intr = md->pmd_intr;
4904 pmc_hook = pmc_hook_handler;
4907 printf(PMC_MODULE_NAME ":");
4908 for (n = 0; n < (int) md->pmd_nclass; n++) {
4909 pcd = &md->pmd_classdep[n];
4910 printf(" %s/%d/%d/0x%b",
4911 pmc_name_of_pmcclass[pcd->pcd_class],
4916 "\1INT\2USR\3SYS\4EDG\5THR"
4917 "\6REA\7WRI\10INV\11QUA\12PRC"
4926 /* prepare to be unloaded */
4931 unsigned int maxcpu;
4932 struct pmc_ownerhash *ph;
4933 struct pmc_owner *po, *tmp;
4934 struct pmc_binding pb;
4936 struct pmc_processhash *prh;
4939 PMCDBG(MOD,INI,0, "%s", "cleanup");
4941 /* switch off sampling */
4942 CPU_ZERO(&pmc_cpumask);
4946 if (pmc_hook == NULL) { /* being unloaded already */
4947 sx_xunlock(&pmc_sx);
4951 pmc_hook = NULL; /* prevent new threads from entering module */
4953 /* deregister event handlers */
4954 EVENTHANDLER_DEREGISTER(process_fork, pmc_fork_tag);
4955 EVENTHANDLER_DEREGISTER(process_exit, pmc_exit_tag);
4956 EVENTHANDLER_DEREGISTER(kld_load, pmc_kld_load_tag);
4957 EVENTHANDLER_DEREGISTER(kld_unload, pmc_kld_unload_tag);
4959 /* send SIGBUS to all owner threads, free up allocations */
4961 for (ph = pmc_ownerhash;
4962 ph <= &pmc_ownerhash[pmc_ownerhashmask];
4964 LIST_FOREACH_SAFE(po, ph, po_next, tmp) {
4965 pmc_remove_owner(po);
4967 /* send SIGBUS to owner processes */
4968 PMCDBG(MOD,INI,2, "cleanup signal proc=%p "
4969 "(%d, %s)", po->po_owner,
4970 po->po_owner->p_pid,
4971 po->po_owner->p_comm);
4973 PROC_LOCK(po->po_owner);
4974 kern_psignal(po->po_owner, SIGBUS);
4975 PROC_UNLOCK(po->po_owner);
4977 pmc_destroy_owner_descriptor(po);
4981 /* reclaim allocated data structures */
4983 mtx_pool_destroy(&pmc_mtxpool);
4985 mtx_destroy(&pmc_processhash_mtx);
4986 if (pmc_processhash) {
4988 struct pmc_process *pp;
4990 PMCDBG(MOD,INI,3, "%s", "destroy process hash");
4991 for (prh = pmc_processhash;
4992 prh <= &pmc_processhash[pmc_processhashmask];
4994 LIST_FOREACH(pp, prh, pp_next)
4995 PMCDBG(MOD,INI,3, "pid=%d", pp->pp_proc->p_pid);
4998 hashdestroy(pmc_processhash, M_PMC, pmc_processhashmask);
4999 pmc_processhash = NULL;
5002 if (pmc_ownerhash) {
5003 PMCDBG(MOD,INI,3, "%s", "destroy owner hash");
5004 hashdestroy(pmc_ownerhash, M_PMC, pmc_ownerhashmask);
5005 pmc_ownerhash = NULL;
5008 KASSERT(LIST_EMPTY(&pmc_ss_owners),
5009 ("[pmc,%d] Global SS owner list not empty", __LINE__));
5010 KASSERT(pmc_ss_count == 0,
5011 ("[pmc,%d] Global SS count not empty", __LINE__));
5013 /* do processor and pmc-class dependent cleanup */
5014 maxcpu = pmc_cpu_max();
5016 PMCDBG(MOD,INI,3, "%s", "md cleanup");
5018 pmc_save_cpu_binding(&pb);
5019 for (cpu = 0; cpu < maxcpu; cpu++) {
5020 PMCDBG(MOD,INI,1,"pmc-cleanup cpu=%d pcs=%p",
5021 cpu, pmc_pcpu[cpu]);
5022 if (!pmc_cpu_is_active(cpu) || pmc_pcpu[cpu] == NULL)
5024 pmc_select_cpu(cpu);
5025 for (c = 0; c < md->pmd_nclass; c++)
5026 md->pmd_classdep[c].pcd_pcpu_fini(md, cpu);
5027 if (md->pmd_pcpu_fini)
5028 md->pmd_pcpu_fini(md, cpu);
5031 if (md->pmd_cputype == PMC_CPU_GENERIC)
5032 pmc_generic_cpu_finalize(md);
5034 pmc_md_finalize(md);
5038 pmc_restore_cpu_binding(&pb);
5041 /* Free per-cpu descriptors. */
5042 for (cpu = 0; cpu < maxcpu; cpu++) {
5043 if (!pmc_cpu_is_active(cpu))
5045 KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_HR] != NULL,
5046 ("[pmc,%d] Null hw cpu sample buffer cpu=%d", __LINE__,
5048 KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_SR] != NULL,
5049 ("[pmc,%d] Null sw cpu sample buffer cpu=%d", __LINE__,
5051 free(pmc_pcpu[cpu]->pc_sb[PMC_HR]->ps_callchains, M_PMC);
5052 free(pmc_pcpu[cpu]->pc_sb[PMC_HR], M_PMC);
5053 free(pmc_pcpu[cpu]->pc_sb[PMC_SR]->ps_callchains, M_PMC);
5054 free(pmc_pcpu[cpu]->pc_sb[PMC_SR], M_PMC);
5055 free(pmc_pcpu[cpu], M_PMC);
5058 free(pmc_pcpu, M_PMC);
5061 free(pmc_pcpu_saved, M_PMC);
5062 pmc_pcpu_saved = NULL;
5065 free(pmc_pmcdisp, M_PMC);
5069 if (pmc_rowindex_to_classdep) {
5070 free(pmc_rowindex_to_classdep, M_PMC);
5071 pmc_rowindex_to_classdep = NULL;
5076 sx_xunlock(&pmc_sx); /* we are done */
5080 * The function called at load/unload.
5084 load (struct module *module __unused, int cmd, void *arg __unused)
5092 /* initialize the subsystem */
5093 error = pmc_initialize();
5096 PMCDBG(MOD,INI,1, "syscall=%d maxcpu=%d",
5097 pmc_syscall_num, pmc_cpu_max());
5104 PMCDBG(MOD,INI,1, "%s", "unloaded");
5108 error = EINVAL; /* XXX should panic(9) */
5116 MALLOC_DEFINE(M_PMC, "pmc", "Memory space for the PMC module");