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/sched.h>
54 #include <sys/signalvar.h>
57 #include <sys/sysctl.h>
58 #include <sys/sysent.h>
59 #include <sys/systm.h>
60 #include <sys/vnode.h>
62 #include <sys/linker.h> /* needs to be after <sys/malloc.h> */
64 #include <machine/atomic.h>
65 #include <machine/md_var.h>
68 #include <vm/vm_extern.h>
70 #include <vm/vm_map.h>
71 #include <vm/vm_object.h>
78 PMC_FLAG_NONE = 0x00, /* do nothing */
79 PMC_FLAG_REMOVE = 0x01, /* atomically remove entry from hash */
80 PMC_FLAG_ALLOCATE = 0x02, /* add entry to hash if not found */
84 * The offset in sysent where the syscall is allocated.
87 static int pmc_syscall_num = NO_SYSCALL;
88 struct pmc_cpu **pmc_pcpu; /* per-cpu state */
89 pmc_value_t *pmc_pcpu_saved; /* saved PMC values: CSW handling */
91 #define PMC_PCPU_SAVED(C,R) pmc_pcpu_saved[(R) + md->pmd_npmc*(C)]
93 struct mtx_pool *pmc_mtxpool;
94 static int *pmc_pmcdisp; /* PMC row dispositions */
96 #define PMC_ROW_DISP_IS_FREE(R) (pmc_pmcdisp[(R)] == 0)
97 #define PMC_ROW_DISP_IS_THREAD(R) (pmc_pmcdisp[(R)] > 0)
98 #define PMC_ROW_DISP_IS_STANDALONE(R) (pmc_pmcdisp[(R)] < 0)
100 #define PMC_MARK_ROW_FREE(R) do { \
101 pmc_pmcdisp[(R)] = 0; \
104 #define PMC_MARK_ROW_STANDALONE(R) do { \
105 KASSERT(pmc_pmcdisp[(R)] <= 0, ("[pmc,%d] row disposition error", \
107 atomic_add_int(&pmc_pmcdisp[(R)], -1); \
108 KASSERT(pmc_pmcdisp[(R)] >= (-pmc_cpu_max_active()), \
109 ("[pmc,%d] row disposition error", __LINE__)); \
112 #define PMC_UNMARK_ROW_STANDALONE(R) do { \
113 atomic_add_int(&pmc_pmcdisp[(R)], 1); \
114 KASSERT(pmc_pmcdisp[(R)] <= 0, ("[pmc,%d] row disposition error", \
118 #define PMC_MARK_ROW_THREAD(R) do { \
119 KASSERT(pmc_pmcdisp[(R)] >= 0, ("[pmc,%d] row disposition error", \
121 atomic_add_int(&pmc_pmcdisp[(R)], 1); \
124 #define PMC_UNMARK_ROW_THREAD(R) do { \
125 atomic_add_int(&pmc_pmcdisp[(R)], -1); \
126 KASSERT(pmc_pmcdisp[(R)] >= 0, ("[pmc,%d] row disposition error", \
131 /* various event handlers */
132 static eventhandler_tag pmc_exit_tag, pmc_fork_tag;
134 /* Module statistics */
135 struct pmc_op_getdriverstats pmc_stats;
137 /* Machine/processor dependent operations */
138 static struct pmc_mdep *md;
141 * Hash tables mapping owner processes and target threads to PMCs.
144 struct mtx pmc_processhash_mtx; /* spin mutex */
145 static u_long pmc_processhashmask;
146 static LIST_HEAD(pmc_processhash, pmc_process) *pmc_processhash;
149 * Hash table of PMC owner descriptors. This table is protected by
150 * the shared PMC "sx" lock.
153 static u_long pmc_ownerhashmask;
154 static LIST_HEAD(pmc_ownerhash, pmc_owner) *pmc_ownerhash;
157 * List of PMC owners with system-wide sampling PMCs.
160 static LIST_HEAD(, pmc_owner) pmc_ss_owners;
164 * A map of row indices to classdep structures.
166 static struct pmc_classdep **pmc_rowindex_to_classdep;
173 static int pmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS);
174 static int pmc_debugflags_parse(char *newstr, char *fence);
177 static int load(struct module *module, int cmd, void *arg);
178 static int pmc_attach_process(struct proc *p, struct pmc *pm);
179 static struct pmc *pmc_allocate_pmc_descriptor(void);
180 static struct pmc_owner *pmc_allocate_owner_descriptor(struct proc *p);
181 static int pmc_attach_one_process(struct proc *p, struct pmc *pm);
182 static int pmc_can_allocate_rowindex(struct proc *p, unsigned int ri,
184 static int pmc_can_attach(struct pmc *pm, struct proc *p);
185 static void pmc_capture_user_callchain(int cpu, struct trapframe *tf);
186 static void pmc_cleanup(void);
187 static int pmc_detach_process(struct proc *p, struct pmc *pm);
188 static int pmc_detach_one_process(struct proc *p, struct pmc *pm,
190 static void pmc_destroy_owner_descriptor(struct pmc_owner *po);
191 static struct pmc_owner *pmc_find_owner_descriptor(struct proc *p);
192 static int pmc_find_pmc(pmc_id_t pmcid, struct pmc **pm);
193 static struct pmc *pmc_find_pmc_descriptor_in_process(struct pmc_owner *po,
195 static struct pmc_process *pmc_find_process_descriptor(struct proc *p,
197 static void pmc_force_context_switch(void);
198 static void pmc_link_target_process(struct pmc *pm,
199 struct pmc_process *pp);
200 static void pmc_log_all_process_mappings(struct pmc_owner *po);
201 static void pmc_log_kernel_mappings(struct pmc *pm);
202 static void pmc_log_process_mappings(struct pmc_owner *po, struct proc *p);
203 static void pmc_maybe_remove_owner(struct pmc_owner *po);
204 static void pmc_process_csw_in(struct thread *td);
205 static void pmc_process_csw_out(struct thread *td);
206 static void pmc_process_exit(void *arg, struct proc *p);
207 static void pmc_process_fork(void *arg, struct proc *p1,
208 struct proc *p2, int n);
209 static void pmc_process_samples(int cpu);
210 static void pmc_release_pmc_descriptor(struct pmc *pmc);
211 static void pmc_remove_owner(struct pmc_owner *po);
212 static void pmc_remove_process_descriptor(struct pmc_process *pp);
213 static void pmc_restore_cpu_binding(struct pmc_binding *pb);
214 static void pmc_save_cpu_binding(struct pmc_binding *pb);
215 static void pmc_select_cpu(int cpu);
216 static int pmc_start(struct pmc *pm);
217 static int pmc_stop(struct pmc *pm);
218 static int pmc_syscall_handler(struct thread *td, void *syscall_args);
219 static void pmc_unlink_target_process(struct pmc *pmc,
220 struct pmc_process *pp);
223 * Kernel tunables and sysctl(8) interface.
226 SYSCTL_NODE(_kern, OID_AUTO, hwpmc, CTLFLAG_RW, 0, "HWPMC parameters");
228 static int pmc_callchaindepth = PMC_CALLCHAIN_DEPTH;
229 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "callchaindepth", &pmc_callchaindepth);
230 SYSCTL_INT(_kern_hwpmc, OID_AUTO, callchaindepth, CTLFLAG_TUN|CTLFLAG_RD,
231 &pmc_callchaindepth, 0, "depth of call chain records");
234 struct pmc_debugflags pmc_debugflags = PMC_DEBUG_DEFAULT_FLAGS;
235 char pmc_debugstr[PMC_DEBUG_STRSIZE];
236 TUNABLE_STR(PMC_SYSCTL_NAME_PREFIX "debugflags", pmc_debugstr,
237 sizeof(pmc_debugstr));
238 SYSCTL_PROC(_kern_hwpmc, OID_AUTO, debugflags,
239 CTLTYPE_STRING|CTLFLAG_RW|CTLFLAG_TUN,
240 0, 0, pmc_debugflags_sysctl_handler, "A", "debug flags");
244 * kern.hwpmc.hashrows -- determines the number of rows in the
245 * of the hash table used to look up threads
248 static int pmc_hashsize = PMC_HASH_SIZE;
249 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "hashsize", &pmc_hashsize);
250 SYSCTL_INT(_kern_hwpmc, OID_AUTO, hashsize, CTLFLAG_TUN|CTLFLAG_RD,
251 &pmc_hashsize, 0, "rows in hash tables");
254 * kern.hwpmc.nsamples --- number of PC samples/callchain stacks per CPU
257 static int pmc_nsamples = PMC_NSAMPLES;
258 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "nsamples", &pmc_nsamples);
259 SYSCTL_INT(_kern_hwpmc, OID_AUTO, nsamples, CTLFLAG_TUN|CTLFLAG_RD,
260 &pmc_nsamples, 0, "number of PC samples per CPU");
264 * kern.hwpmc.mtxpoolsize -- number of mutexes in the mutex pool.
267 static int pmc_mtxpool_size = PMC_MTXPOOL_SIZE;
268 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "mtxpoolsize", &pmc_mtxpool_size);
269 SYSCTL_INT(_kern_hwpmc, OID_AUTO, mtxpoolsize, CTLFLAG_TUN|CTLFLAG_RD,
270 &pmc_mtxpool_size, 0, "size of spin mutex pool");
274 * security.bsd.unprivileged_syspmcs -- allow non-root processes to
275 * allocate system-wide PMCs.
277 * Allowing unprivileged processes to allocate system PMCs is convenient
278 * if system-wide measurements need to be taken concurrently with other
279 * per-process measurements. This feature is turned off by default.
282 static int pmc_unprivileged_syspmcs = 0;
283 TUNABLE_INT("security.bsd.unprivileged_syspmcs", &pmc_unprivileged_syspmcs);
284 SYSCTL_INT(_security_bsd, OID_AUTO, unprivileged_syspmcs, CTLFLAG_RW,
285 &pmc_unprivileged_syspmcs, 0,
286 "allow unprivileged process to allocate system PMCs");
289 * Hash function. Discard the lower 2 bits of the pointer since
290 * these are always zero for our uses. The hash multiplier is
291 * round((2^LONG_BIT) * ((sqrt(5)-1)/2)).
295 #define _PMC_HM 11400714819323198486u
297 #define _PMC_HM 2654435769u
299 #error Must know the size of 'long' to compile
302 #define PMC_HASH_PTR(P,M) ((((unsigned long) (P) >> 2) * _PMC_HM) & (M))
308 /* The `sysent' for the new syscall */
309 static struct sysent pmc_sysent = {
311 pmc_syscall_handler /* sy_call */
314 static struct syscall_module_data pmc_syscall_mod = {
322 static moduledata_t pmc_mod = {
324 syscall_module_handler,
328 DECLARE_MODULE(pmc, pmc_mod, SI_SUB_SMP, SI_ORDER_ANY);
329 MODULE_VERSION(pmc, PMC_VERSION);
332 enum pmc_dbgparse_state {
333 PMCDS_WS, /* in whitespace */
334 PMCDS_MAJOR, /* seen a major keyword */
339 pmc_debugflags_parse(char *newstr, char *fence)
342 struct pmc_debugflags *tmpflags;
343 int error, found, *newbits, tmp;
346 tmpflags = malloc(sizeof(*tmpflags), M_PMC, M_WAITOK|M_ZERO);
351 for (; p < fence && (c = *p); p++) {
353 /* skip white space */
354 if (c == ' ' || c == '\t')
357 /* look for a keyword followed by "=" */
358 for (q = p; p < fence && (c = *p) && c != '='; p++)
368 /* lookup flag group name */
369 #define DBG_SET_FLAG_MAJ(S,F) \
370 if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0) \
371 newbits = &tmpflags->pdb_ ## F;
373 DBG_SET_FLAG_MAJ("cpu", CPU);
374 DBG_SET_FLAG_MAJ("csw", CSW);
375 DBG_SET_FLAG_MAJ("logging", LOG);
376 DBG_SET_FLAG_MAJ("module", MOD);
377 DBG_SET_FLAG_MAJ("md", MDP);
378 DBG_SET_FLAG_MAJ("owner", OWN);
379 DBG_SET_FLAG_MAJ("pmc", PMC);
380 DBG_SET_FLAG_MAJ("process", PRC);
381 DBG_SET_FLAG_MAJ("sampling", SAM);
383 if (newbits == NULL) {
388 p++; /* skip the '=' */
390 /* Now parse the individual flags */
393 for (q = p; p < fence && (c = *p); p++)
394 if (c == ' ' || c == '\t' || c == ',')
397 /* p == fence or c == ws or c == "," or c == 0 */
399 if ((kwlen = p - q) == 0) {
405 #define DBG_SET_FLAG_MIN(S,F) \
406 if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0) \
407 tmp |= found = (1 << PMC_DEBUG_MIN_ ## F)
409 /* a '*' denotes all possible flags in the group */
410 if (kwlen == 1 && *q == '*')
412 /* look for individual flag names */
413 DBG_SET_FLAG_MIN("allocaterow", ALR);
414 DBG_SET_FLAG_MIN("allocate", ALL);
415 DBG_SET_FLAG_MIN("attach", ATT);
416 DBG_SET_FLAG_MIN("bind", BND);
417 DBG_SET_FLAG_MIN("config", CFG);
418 DBG_SET_FLAG_MIN("exec", EXC);
419 DBG_SET_FLAG_MIN("exit", EXT);
420 DBG_SET_FLAG_MIN("find", FND);
421 DBG_SET_FLAG_MIN("flush", FLS);
422 DBG_SET_FLAG_MIN("fork", FRK);
423 DBG_SET_FLAG_MIN("getbuf", GTB);
424 DBG_SET_FLAG_MIN("hook", PMH);
425 DBG_SET_FLAG_MIN("init", INI);
426 DBG_SET_FLAG_MIN("intr", INT);
427 DBG_SET_FLAG_MIN("linktarget", TLK);
428 DBG_SET_FLAG_MIN("mayberemove", OMR);
429 DBG_SET_FLAG_MIN("ops", OPS);
430 DBG_SET_FLAG_MIN("read", REA);
431 DBG_SET_FLAG_MIN("register", REG);
432 DBG_SET_FLAG_MIN("release", REL);
433 DBG_SET_FLAG_MIN("remove", ORM);
434 DBG_SET_FLAG_MIN("sample", SAM);
435 DBG_SET_FLAG_MIN("scheduleio", SIO);
436 DBG_SET_FLAG_MIN("select", SEL);
437 DBG_SET_FLAG_MIN("signal", SIG);
438 DBG_SET_FLAG_MIN("swi", SWI);
439 DBG_SET_FLAG_MIN("swo", SWO);
440 DBG_SET_FLAG_MIN("start", STA);
441 DBG_SET_FLAG_MIN("stop", STO);
442 DBG_SET_FLAG_MIN("syscall", PMS);
443 DBG_SET_FLAG_MIN("unlinktarget", TUL);
444 DBG_SET_FLAG_MIN("write", WRI);
446 /* unrecognized flag name */
451 if (c == 0 || c == ' ' || c == '\t') { /* end of flag group */
460 /* save the new flag set */
461 bcopy(tmpflags, &pmc_debugflags, sizeof(pmc_debugflags));
464 free(tmpflags, M_PMC);
469 pmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS)
471 char *fence, *newstr;
475 (void) arg1; (void) arg2; /* unused parameters */
477 n = sizeof(pmc_debugstr);
478 newstr = malloc(n, M_PMC, M_WAITOK|M_ZERO);
479 (void) strlcpy(newstr, pmc_debugstr, n);
481 error = sysctl_handle_string(oidp, newstr, n, req);
483 /* if there is a new string, parse and copy it */
484 if (error == 0 && req->newptr != NULL) {
485 fence = newstr + (n < req->newlen ? n : req->newlen + 1);
486 if ((error = pmc_debugflags_parse(newstr, fence)) == 0)
487 (void) strlcpy(pmc_debugstr, newstr,
488 sizeof(pmc_debugstr));
498 * Map a row index to a classdep structure and return the adjusted row
499 * index for the PMC class index.
501 static struct pmc_classdep *
502 pmc_ri_to_classdep(struct pmc_mdep *md, int ri, int *adjri)
504 struct pmc_classdep *pcd;
508 KASSERT(ri >= 0 && ri < md->pmd_npmc,
509 ("[pmc,%d] illegal row-index %d", __LINE__, ri));
511 pcd = pmc_rowindex_to_classdep[ri];
514 ("[amd,%d] ri %d null pcd", __LINE__, ri));
516 *adjri = ri - pcd->pcd_ri;
518 KASSERT(*adjri >= 0 && *adjri < pcd->pcd_num,
519 ("[pmc,%d] adjusted row-index %d", __LINE__, *adjri));
525 * Concurrency Control
527 * The driver manages the following data structures:
529 * - target process descriptors, one per target process
530 * - owner process descriptors (and attached lists), one per owner process
531 * - lookup hash tables for owner and target processes
532 * - PMC descriptors (and attached lists)
533 * - per-cpu hardware state
534 * - the 'hook' variable through which the kernel calls into
536 * - the machine hardware state (managed by the MD layer)
538 * These data structures are accessed from:
540 * - thread context-switch code
541 * - interrupt handlers (possibly on multiple cpus)
542 * - kernel threads on multiple cpus running on behalf of user
543 * processes doing system calls
544 * - this driver's private kernel threads
546 * = Locks and Locking strategy =
548 * The driver uses four locking strategies for its operation:
550 * - The global SX lock "pmc_sx" is used to protect internal
553 * Calls into the module by syscall() start with this lock being
554 * held in exclusive mode. Depending on the requested operation,
555 * the lock may be downgraded to 'shared' mode to allow more
556 * concurrent readers into the module. Calls into the module from
557 * other parts of the kernel acquire the lock in shared mode.
559 * This SX lock is held in exclusive mode for any operations that
560 * modify the linkages between the driver's internal data structures.
562 * The 'pmc_hook' function pointer is also protected by this lock.
563 * It is only examined with the sx lock held in exclusive mode. The
564 * kernel module is allowed to be unloaded only with the sx lock held
565 * in exclusive mode. In normal syscall handling, after acquiring the
566 * pmc_sx lock we first check that 'pmc_hook' is non-null before
567 * proceeding. This prevents races between the thread unloading the module
568 * and other threads seeking to use the module.
570 * - Lookups of target process structures and owner process structures
571 * cannot use the global "pmc_sx" SX lock because these lookups need
572 * to happen during context switches and in other critical sections
573 * where sleeping is not allowed. We protect these lookup tables
574 * with their own private spin-mutexes, "pmc_processhash_mtx" and
575 * "pmc_ownerhash_mtx".
577 * - Interrupt handlers work in a lock free manner. At interrupt
578 * time, handlers look at the PMC pointer (phw->phw_pmc) configured
579 * when the PMC was started. If this pointer is NULL, the interrupt
580 * is ignored after updating driver statistics. We ensure that this
581 * pointer is set (using an atomic operation if necessary) before the
582 * PMC hardware is started. Conversely, this pointer is unset atomically
583 * only after the PMC hardware is stopped.
585 * We ensure that everything needed for the operation of an
586 * interrupt handler is available without it needing to acquire any
587 * locks. We also ensure that a PMC's software state is destroyed only
588 * after the PMC is taken off hardware (on all CPUs).
590 * - Context-switch handling with process-private PMCs needs more
593 * A given process may be the target of multiple PMCs. For example,
594 * PMCATTACH and PMCDETACH may be requested by a process on one CPU
595 * while the target process is running on another. A PMC could also
596 * be getting released because its owner is exiting. We tackle
597 * these situations in the following manner:
599 * - each target process structure 'pmc_process' has an array
600 * of 'struct pmc *' pointers, one for each hardware PMC.
602 * - At context switch IN time, each "target" PMC in RUNNING state
603 * gets started on hardware and a pointer to each PMC is copied into
604 * the per-cpu phw array. The 'runcount' for the PMC is
607 * - At context switch OUT time, all process-virtual PMCs are stopped
608 * on hardware. The saved value is added to the PMCs value field
609 * only if the PMC is in a non-deleted state (the PMCs state could
610 * have changed during the current time slice).
612 * Note that since in-between a switch IN on a processor and a switch
613 * OUT, the PMC could have been released on another CPU. Therefore
614 * context switch OUT always looks at the hardware state to turn
615 * OFF PMCs and will update a PMC's saved value only if reachable
616 * from the target process record.
618 * - OP PMCRELEASE could be called on a PMC at any time (the PMC could
619 * be attached to many processes at the time of the call and could
620 * be active on multiple CPUs).
622 * We prevent further scheduling of the PMC by marking it as in
623 * state 'DELETED'. If the runcount of the PMC is non-zero then
624 * this PMC is currently running on a CPU somewhere. The thread
625 * doing the PMCRELEASE operation waits by repeatedly doing a
626 * pause() till the runcount comes to zero.
628 * The contents of a PMC descriptor (struct pmc) are protected using
629 * a spin-mutex. In order to save space, we use a mutex pool.
631 * In terms of lock types used by witness(4), we use:
632 * - Type "pmc-sx", used by the global SX lock.
633 * - Type "pmc-sleep", for sleep mutexes used by logger threads.
634 * - Type "pmc-per-proc", for protecting PMC owner descriptors.
635 * - Type "pmc-leaf", used for all other spin mutexes.
639 * save the cpu binding of the current kthread
643 pmc_save_cpu_binding(struct pmc_binding *pb)
645 PMCDBG(CPU,BND,2, "%s", "save-cpu");
646 thread_lock(curthread);
647 pb->pb_bound = sched_is_bound(curthread);
648 pb->pb_cpu = curthread->td_oncpu;
649 thread_unlock(curthread);
650 PMCDBG(CPU,BND,2, "save-cpu cpu=%d", pb->pb_cpu);
654 * restore the cpu binding of the current thread
658 pmc_restore_cpu_binding(struct pmc_binding *pb)
660 PMCDBG(CPU,BND,2, "restore-cpu curcpu=%d restore=%d",
661 curthread->td_oncpu, pb->pb_cpu);
662 thread_lock(curthread);
664 sched_bind(curthread, pb->pb_cpu);
666 sched_unbind(curthread);
667 thread_unlock(curthread);
668 PMCDBG(CPU,BND,2, "%s", "restore-cpu done");
672 * move execution over the specified cpu and bind it there.
676 pmc_select_cpu(int cpu)
678 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
679 ("[pmc,%d] bad cpu number %d", __LINE__, cpu));
681 /* Never move to an inactive CPU. */
682 KASSERT(pmc_cpu_is_active(cpu), ("[pmc,%d] selecting inactive "
683 "CPU %d", __LINE__, cpu));
685 PMCDBG(CPU,SEL,2, "select-cpu cpu=%d", cpu);
686 thread_lock(curthread);
687 sched_bind(curthread, cpu);
688 thread_unlock(curthread);
690 KASSERT(curthread->td_oncpu == cpu,
691 ("[pmc,%d] CPU not bound [cpu=%d, curr=%d]", __LINE__,
692 cpu, curthread->td_oncpu));
694 PMCDBG(CPU,SEL,2, "select-cpu cpu=%d ok", cpu);
698 * Force a context switch.
700 * We do this by pause'ing for 1 tick -- invoking mi_switch() is not
701 * guaranteed to force a context switch.
705 pmc_force_context_switch(void)
712 * Get the file name for an executable. This is a simple wrapper
713 * around vn_fullpath(9).
717 pmc_getfilename(struct vnode *v, char **fullpath, char **freepath)
720 *fullpath = "unknown";
722 vn_fullpath(curthread, v, fullpath, freepath);
726 * remove an process owning PMCs
730 pmc_remove_owner(struct pmc_owner *po)
732 struct pmc *pm, *tmp;
734 sx_assert(&pmc_sx, SX_XLOCKED);
736 PMCDBG(OWN,ORM,1, "remove-owner po=%p", po);
738 /* Remove descriptor from the owner hash table */
739 LIST_REMOVE(po, po_next);
741 /* release all owned PMC descriptors */
742 LIST_FOREACH_SAFE(pm, &po->po_pmcs, pm_next, tmp) {
743 PMCDBG(OWN,ORM,2, "pmc=%p", pm);
744 KASSERT(pm->pm_owner == po,
745 ("[pmc,%d] owner %p != po %p", __LINE__, pm->pm_owner, po));
747 pmc_release_pmc_descriptor(pm); /* will unlink from the list */
750 KASSERT(po->po_sscount == 0,
751 ("[pmc,%d] SS count not zero", __LINE__));
752 KASSERT(LIST_EMPTY(&po->po_pmcs),
753 ("[pmc,%d] PMC list not empty", __LINE__));
755 /* de-configure the log file if present */
756 if (po->po_flags & PMC_PO_OWNS_LOGFILE)
757 pmclog_deconfigure_log(po);
761 * remove an owner process record if all conditions are met.
765 pmc_maybe_remove_owner(struct pmc_owner *po)
768 PMCDBG(OWN,OMR,1, "maybe-remove-owner po=%p", po);
771 * Remove owner record if
772 * - this process does not own any PMCs
773 * - this process has not allocated a system-wide sampling buffer
776 if (LIST_EMPTY(&po->po_pmcs) &&
777 ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)) {
778 pmc_remove_owner(po);
779 pmc_destroy_owner_descriptor(po);
784 * Add an association between a target process and a PMC.
788 pmc_link_target_process(struct pmc *pm, struct pmc_process *pp)
791 struct pmc_target *pt;
793 sx_assert(&pmc_sx, SX_XLOCKED);
795 KASSERT(pm != NULL && pp != NULL,
796 ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp));
797 KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)),
798 ("[pmc,%d] Attaching a non-process-virtual pmc=%p to pid=%d",
799 __LINE__, pm, pp->pp_proc->p_pid));
800 KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= ((int) md->pmd_npmc - 1),
801 ("[pmc,%d] Illegal reference count %d for process record %p",
802 __LINE__, pp->pp_refcnt, (void *) pp));
804 ri = PMC_TO_ROWINDEX(pm);
806 PMCDBG(PRC,TLK,1, "link-target pmc=%p ri=%d pmc-process=%p",
810 LIST_FOREACH(pt, &pm->pm_targets, pt_next)
811 if (pt->pt_process == pp)
812 KASSERT(0, ("[pmc,%d] pp %p already in pmc %p targets",
816 pt = malloc(sizeof(struct pmc_target), M_PMC, M_WAITOK|M_ZERO);
819 LIST_INSERT_HEAD(&pm->pm_targets, pt, pt_next);
821 atomic_store_rel_ptr((uintptr_t *)&pp->pp_pmcs[ri].pp_pmc,
824 if (pm->pm_owner->po_owner == pp->pp_proc)
825 pm->pm_flags |= PMC_F_ATTACHED_TO_OWNER;
828 * Initialize the per-process values at this row index.
830 pp->pp_pmcs[ri].pp_pmcval = PMC_TO_MODE(pm) == PMC_MODE_TS ?
831 pm->pm_sc.pm_reloadcount : 0;
838 * Removes the association between a target process and a PMC.
842 pmc_unlink_target_process(struct pmc *pm, struct pmc_process *pp)
846 struct pmc_target *ptgt;
848 sx_assert(&pmc_sx, SX_XLOCKED);
850 KASSERT(pm != NULL && pp != NULL,
851 ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp));
853 KASSERT(pp->pp_refcnt >= 1 && pp->pp_refcnt <= (int) md->pmd_npmc,
854 ("[pmc,%d] Illegal ref count %d on process record %p",
855 __LINE__, pp->pp_refcnt, (void *) pp));
857 ri = PMC_TO_ROWINDEX(pm);
859 PMCDBG(PRC,TUL,1, "unlink-target pmc=%p ri=%d pmc-process=%p",
862 KASSERT(pp->pp_pmcs[ri].pp_pmc == pm,
863 ("[pmc,%d] PMC ri %d mismatch pmc %p pp->[ri] %p", __LINE__,
864 ri, pm, pp->pp_pmcs[ri].pp_pmc));
866 pp->pp_pmcs[ri].pp_pmc = NULL;
867 pp->pp_pmcs[ri].pp_pmcval = (pmc_value_t) 0;
869 /* Remove owner-specific flags */
870 if (pm->pm_owner->po_owner == pp->pp_proc) {
871 pp->pp_flags &= ~PMC_PP_ENABLE_MSR_ACCESS;
872 pm->pm_flags &= ~PMC_F_ATTACHED_TO_OWNER;
877 /* Remove the target process from the PMC structure */
878 LIST_FOREACH(ptgt, &pm->pm_targets, pt_next)
879 if (ptgt->pt_process == pp)
882 KASSERT(ptgt != NULL, ("[pmc,%d] process %p (pp: %p) not found "
883 "in pmc %p", __LINE__, pp->pp_proc, pp, pm));
885 LIST_REMOVE(ptgt, pt_next);
888 /* if the PMC now lacks targets, send the owner a SIGIO */
889 if (LIST_EMPTY(&pm->pm_targets)) {
890 p = pm->pm_owner->po_owner;
895 PMCDBG(PRC,SIG,2, "signalling proc=%p signal=%d", p,
901 * Check if PMC 'pm' may be attached to target process 't'.
905 pmc_can_attach(struct pmc *pm, struct proc *t)
907 struct proc *o; /* pmc owner */
908 struct ucred *oc, *tc; /* owner, target credentials */
909 int decline_attach, i;
912 * A PMC's owner can always attach that PMC to itself.
915 if ((o = pm->pm_owner->po_owner) == t)
929 * The effective uid of the PMC owner should match at least one
930 * of the {effective,real,saved} uids of the target process.
933 decline_attach = oc->cr_uid != tc->cr_uid &&
934 oc->cr_uid != tc->cr_svuid &&
935 oc->cr_uid != tc->cr_ruid;
938 * Every one of the target's group ids, must be in the owner's
941 for (i = 0; !decline_attach && i < tc->cr_ngroups; i++)
942 decline_attach = !groupmember(tc->cr_groups[i], oc);
944 /* check the read and saved gids too */
945 if (decline_attach == 0)
946 decline_attach = !groupmember(tc->cr_rgid, oc) ||
947 !groupmember(tc->cr_svgid, oc);
952 return !decline_attach;
956 * Attach a process to a PMC.
960 pmc_attach_one_process(struct proc *p, struct pmc *pm)
963 char *fullpath, *freepath;
964 struct pmc_process *pp;
966 sx_assert(&pmc_sx, SX_XLOCKED);
968 PMCDBG(PRC,ATT,2, "attach-one pm=%p ri=%d proc=%p (%d, %s)", pm,
969 PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm);
972 * Locate the process descriptor corresponding to process 'p',
973 * allocating space as needed.
975 * Verify that rowindex 'pm_rowindex' is free in the process
978 * If not, allocate space for a descriptor and link the
979 * process descriptor and PMC.
981 ri = PMC_TO_ROWINDEX(pm);
983 if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_ALLOCATE)) == NULL)
986 if (pp->pp_pmcs[ri].pp_pmc == pm) /* already present at slot [ri] */
989 if (pp->pp_pmcs[ri].pp_pmc != NULL)
992 pmc_link_target_process(pm, pp);
994 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) &&
995 (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) == 0)
996 pm->pm_flags |= PMC_F_NEEDS_LOGFILE;
998 pm->pm_flags |= PMC_F_ATTACH_DONE; /* mark as attached */
1000 /* issue an attach event to a configured log file */
1001 if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE) {
1002 pmc_getfilename(p->p_textvp, &fullpath, &freepath);
1003 if (p->p_flag & P_KTHREAD) {
1004 fullpath = kernelname;
1007 pmclog_process_pmcattach(pm, p->p_pid, fullpath);
1009 free(freepath, M_TEMP);
1010 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1011 pmc_log_process_mappings(pm->pm_owner, p);
1013 /* mark process as using HWPMCs */
1015 p->p_flag |= P_HWPMC;
1022 * Attach a process and optionally its children
1026 pmc_attach_process(struct proc *p, struct pmc *pm)
1031 sx_assert(&pmc_sx, SX_XLOCKED);
1033 PMCDBG(PRC,ATT,1, "attach pm=%p ri=%d proc=%p (%d, %s)", pm,
1034 PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm);
1038 * If this PMC successfully allowed a GETMSR operation
1039 * in the past, disallow further ATTACHes.
1042 if ((pm->pm_flags & PMC_PP_ENABLE_MSR_ACCESS) != 0)
1045 if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0)
1046 return pmc_attach_one_process(p, pm);
1049 * Traverse all child processes, attaching them to
1053 sx_slock(&proctree_lock);
1058 if ((error = pmc_attach_one_process(p, pm)) != 0)
1060 if (!LIST_EMPTY(&p->p_children))
1061 p = LIST_FIRST(&p->p_children);
1065 if (LIST_NEXT(p, p_sibling)) {
1066 p = LIST_NEXT(p, p_sibling);
1074 (void) pmc_detach_process(top, pm);
1077 sx_sunlock(&proctree_lock);
1082 * Detach a process from a PMC. If there are no other PMCs tracking
1083 * this process, remove the process structure from its hash table. If
1084 * 'flags' contains PMC_FLAG_REMOVE, then free the process structure.
1088 pmc_detach_one_process(struct proc *p, struct pmc *pm, int flags)
1091 struct pmc_process *pp;
1093 sx_assert(&pmc_sx, SX_XLOCKED);
1096 ("[pmc,%d] null pm pointer", __LINE__));
1098 ri = PMC_TO_ROWINDEX(pm);
1100 PMCDBG(PRC,ATT,2, "detach-one pm=%p ri=%d proc=%p (%d, %s) flags=0x%x",
1101 pm, ri, p, p->p_pid, p->p_comm, flags);
1103 if ((pp = pmc_find_process_descriptor(p, 0)) == NULL)
1106 if (pp->pp_pmcs[ri].pp_pmc != pm)
1109 pmc_unlink_target_process(pm, pp);
1111 /* Issue a detach entry if a log file is configured */
1112 if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE)
1113 pmclog_process_pmcdetach(pm, p->p_pid);
1116 * If there are no PMCs targetting this process, we remove its
1117 * descriptor from the target hash table and unset the P_HWPMC
1118 * flag in the struct proc.
1120 KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= (int) md->pmd_npmc,
1121 ("[pmc,%d] Illegal refcnt %d for process struct %p",
1122 __LINE__, pp->pp_refcnt, pp));
1124 if (pp->pp_refcnt != 0) /* still a target of some PMC */
1127 pmc_remove_process_descriptor(pp);
1129 if (flags & PMC_FLAG_REMOVE)
1133 p->p_flag &= ~P_HWPMC;
1140 * Detach a process and optionally its descendants from a PMC.
1144 pmc_detach_process(struct proc *p, struct pmc *pm)
1148 sx_assert(&pmc_sx, SX_XLOCKED);
1150 PMCDBG(PRC,ATT,1, "detach pm=%p ri=%d proc=%p (%d, %s)", pm,
1151 PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm);
1153 if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0)
1154 return pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE);
1157 * Traverse all children, detaching them from this PMC. We
1158 * ignore errors since we could be detaching a PMC from a
1159 * partially attached proc tree.
1162 sx_slock(&proctree_lock);
1167 (void) pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE);
1169 if (!LIST_EMPTY(&p->p_children))
1170 p = LIST_FIRST(&p->p_children);
1174 if (LIST_NEXT(p, p_sibling)) {
1175 p = LIST_NEXT(p, p_sibling);
1183 sx_sunlock(&proctree_lock);
1185 if (LIST_EMPTY(&pm->pm_targets))
1186 pm->pm_flags &= ~PMC_F_ATTACH_DONE;
1193 * Thread context switch IN
1197 pmc_process_csw_in(struct thread *td)
1200 unsigned int adjri, ri;
1205 pmc_value_t newvalue;
1206 struct pmc_process *pp;
1207 struct pmc_classdep *pcd;
1211 if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE)) == NULL)
1214 KASSERT(pp->pp_proc == td->td_proc,
1215 ("[pmc,%d] not my thread state", __LINE__));
1217 critical_enter(); /* no preemption from this point */
1219 cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */
1221 PMCDBG(CSW,SWI,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p,
1222 p->p_pid, p->p_comm, pp);
1224 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
1225 ("[pmc,%d] wierd CPU id %d", __LINE__, cpu));
1229 for (ri = 0; ri < md->pmd_npmc; ri++) {
1231 if ((pm = pp->pp_pmcs[ri].pp_pmc) == NULL)
1234 KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)),
1235 ("[pmc,%d] Target PMC in non-virtual mode (%d)",
1236 __LINE__, PMC_TO_MODE(pm)));
1238 KASSERT(PMC_TO_ROWINDEX(pm) == ri,
1239 ("[pmc,%d] Row index mismatch pmc %d != ri %d",
1240 __LINE__, PMC_TO_ROWINDEX(pm), ri));
1243 * Only PMCs that are marked as 'RUNNING' need
1244 * be placed on hardware.
1247 if (pm->pm_state != PMC_STATE_RUNNING)
1250 /* increment PMC runcount */
1251 atomic_add_rel_32(&pm->pm_runcount, 1);
1253 /* configure the HWPMC we are going to use. */
1254 pcd = pmc_ri_to_classdep(md, ri, &adjri);
1255 pcd->pcd_config_pmc(cpu, adjri, pm);
1257 phw = pc->pc_hwpmcs[ri];
1259 KASSERT(phw != NULL,
1260 ("[pmc,%d] null hw pointer", __LINE__));
1262 KASSERT(phw->phw_pmc == pm,
1263 ("[pmc,%d] hw->pmc %p != pmc %p", __LINE__,
1267 * Write out saved value and start the PMC.
1269 * Sampling PMCs use a per-process value, while
1270 * counting mode PMCs use a per-pmc value that is
1271 * inherited across descendants.
1273 if (PMC_TO_MODE(pm) == PMC_MODE_TS) {
1274 mtx_pool_lock_spin(pmc_mtxpool, pm);
1275 newvalue = PMC_PCPU_SAVED(cpu,ri) =
1276 pp->pp_pmcs[ri].pp_pmcval;
1277 mtx_pool_unlock_spin(pmc_mtxpool, pm);
1279 KASSERT(PMC_TO_MODE(pm) == PMC_MODE_TC,
1280 ("[pmc,%d] illegal mode=%d", __LINE__,
1282 mtx_pool_lock_spin(pmc_mtxpool, pm);
1283 newvalue = PMC_PCPU_SAVED(cpu, ri) =
1284 pm->pm_gv.pm_savedvalue;
1285 mtx_pool_unlock_spin(pmc_mtxpool, pm);
1288 PMCDBG(CSW,SWI,1,"cpu=%d ri=%d new=%jd", cpu, ri, newvalue);
1290 pcd->pcd_write_pmc(cpu, adjri, newvalue);
1291 pcd->pcd_start_pmc(cpu, adjri);
1295 * perform any other architecture/cpu dependent thread
1296 * switch-in actions.
1299 (void) (*md->pmd_switch_in)(pc, pp);
1306 * Thread context switch OUT.
1310 pmc_process_csw_out(struct thread *td)
1318 pmc_value_t newvalue;
1319 unsigned int adjri, ri;
1320 struct pmc_process *pp;
1321 struct pmc_classdep *pcd;
1325 * Locate our process descriptor; this may be NULL if
1326 * this process is exiting and we have already removed
1327 * the process from the target process table.
1329 * Note that due to kernel preemption, multiple
1330 * context switches may happen while the process is
1333 * Note also that if the target process cannot be
1334 * found we still need to deconfigure any PMCs that
1335 * are currently running on hardware.
1339 pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE);
1347 cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */
1349 PMCDBG(CSW,SWO,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p,
1350 p->p_pid, p->p_comm, pp);
1352 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
1353 ("[pmc,%d wierd CPU id %d", __LINE__, cpu));
1358 * When a PMC gets unlinked from a target PMC, it will
1359 * be removed from the target's pp_pmc[] array.
1361 * However, on a MP system, the target could have been
1362 * executing on another CPU at the time of the unlink.
1363 * So, at context switch OUT time, we need to look at
1364 * the hardware to determine if a PMC is scheduled on
1368 for (ri = 0; ri < md->pmd_npmc; ri++) {
1370 pcd = pmc_ri_to_classdep(md, ri, &adjri);
1372 (void) (*pcd->pcd_get_config)(cpu, adjri, &pm);
1374 if (pm == NULL) /* nothing at this row index */
1377 mode = PMC_TO_MODE(pm);
1378 if (!PMC_IS_VIRTUAL_MODE(mode))
1379 continue; /* not a process virtual PMC */
1381 KASSERT(PMC_TO_ROWINDEX(pm) == ri,
1382 ("[pmc,%d] ri mismatch pmc(%d) ri(%d)",
1383 __LINE__, PMC_TO_ROWINDEX(pm), ri));
1385 /* Stop hardware if not already stopped */
1386 if (pm->pm_stalled == 0)
1387 pcd->pcd_stop_pmc(cpu, adjri);
1389 /* reduce this PMC's runcount */
1390 atomic_subtract_rel_32(&pm->pm_runcount, 1);
1393 * If this PMC is associated with this process,
1397 if (pp != NULL && pp->pp_pmcs[ri].pp_pmc != NULL) {
1399 KASSERT(pm == pp->pp_pmcs[ri].pp_pmc,
1400 ("[pmc,%d] pm %p != pp_pmcs[%d] %p", __LINE__,
1401 pm, ri, pp->pp_pmcs[ri].pp_pmc));
1403 KASSERT(pp->pp_refcnt > 0,
1404 ("[pmc,%d] pp refcnt = %d", __LINE__,
1407 pcd->pcd_read_pmc(cpu, adjri, &newvalue);
1409 tmp = newvalue - PMC_PCPU_SAVED(cpu,ri);
1411 PMCDBG(CSW,SWI,1,"cpu=%d ri=%d tmp=%jd", cpu, ri,
1414 if (mode == PMC_MODE_TS) {
1417 * For sampling process-virtual PMCs,
1418 * we expect the count to be
1419 * decreasing as the 'value'
1420 * programmed into the PMC is the
1421 * number of events to be seen till
1422 * the next sampling interrupt.
1425 tmp += pm->pm_sc.pm_reloadcount;
1426 mtx_pool_lock_spin(pmc_mtxpool, pm);
1427 pp->pp_pmcs[ri].pp_pmcval -= tmp;
1428 if ((int64_t) pp->pp_pmcs[ri].pp_pmcval < 0)
1429 pp->pp_pmcs[ri].pp_pmcval +=
1430 pm->pm_sc.pm_reloadcount;
1431 mtx_pool_unlock_spin(pmc_mtxpool, pm);
1436 * For counting process-virtual PMCs,
1437 * we expect the count to be
1438 * increasing monotonically, modulo a 64
1441 KASSERT((int64_t) tmp >= 0,
1442 ("[pmc,%d] negative increment cpu=%d "
1443 "ri=%d newvalue=%jx saved=%jx "
1444 "incr=%jx", __LINE__, cpu, ri,
1445 newvalue, PMC_PCPU_SAVED(cpu,ri), tmp));
1447 mtx_pool_lock_spin(pmc_mtxpool, pm);
1448 pm->pm_gv.pm_savedvalue += tmp;
1449 pp->pp_pmcs[ri].pp_pmcval += tmp;
1450 mtx_pool_unlock_spin(pmc_mtxpool, pm);
1452 if (pm->pm_flags & PMC_F_LOG_PROCCSW)
1453 pmclog_process_proccsw(pm, pp, tmp);
1457 /* mark hardware as free */
1458 pcd->pcd_config_pmc(cpu, adjri, NULL);
1462 * perform any other architecture/cpu dependent thread
1463 * switch out functions.
1466 (void) (*md->pmd_switch_out)(pc, pp);
1472 * Log a KLD operation.
1476 pmc_process_kld_load(struct pmckern_map_in *pkm)
1478 struct pmc_owner *po;
1480 sx_assert(&pmc_sx, SX_LOCKED);
1483 * Notify owners of system sampling PMCs about KLD operations.
1486 LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1487 if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1488 pmclog_process_map_in(po, (pid_t) -1, pkm->pm_address,
1489 (char *) pkm->pm_file);
1492 * TODO: Notify owners of (all) process-sampling PMCs too.
1499 pmc_process_kld_unload(struct pmckern_map_out *pkm)
1501 struct pmc_owner *po;
1503 sx_assert(&pmc_sx, SX_LOCKED);
1505 LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1506 if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1507 pmclog_process_map_out(po, (pid_t) -1,
1508 pkm->pm_address, pkm->pm_address + pkm->pm_size);
1511 * TODO: Notify owners of process-sampling PMCs.
1516 * A mapping change for a process.
1520 pmc_process_mmap(struct thread *td, struct pmckern_map_in *pkm)
1524 char *fullpath, *freepath;
1525 const struct pmc *pm;
1526 struct pmc_owner *po;
1527 const struct pmc_process *pp;
1529 freepath = fullpath = NULL;
1530 pmc_getfilename((struct vnode *) pkm->pm_file, &fullpath, &freepath);
1532 pid = td->td_proc->p_pid;
1534 /* Inform owners of all system-wide sampling PMCs. */
1535 LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1536 if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1537 pmclog_process_map_in(po, pid, pkm->pm_address, fullpath);
1539 if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL)
1543 * Inform sampling PMC owners tracking this process.
1545 for (ri = 0; ri < md->pmd_npmc; ri++)
1546 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL &&
1547 PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1548 pmclog_process_map_in(pm->pm_owner,
1549 pid, pkm->pm_address, fullpath);
1553 free(freepath, M_TEMP);
1558 * Log an munmap request.
1562 pmc_process_munmap(struct thread *td, struct pmckern_map_out *pkm)
1566 struct pmc_owner *po;
1567 const struct pmc *pm;
1568 const struct pmc_process *pp;
1570 pid = td->td_proc->p_pid;
1572 LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1573 if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1574 pmclog_process_map_out(po, pid, pkm->pm_address,
1575 pkm->pm_address + pkm->pm_size);
1577 if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL)
1580 for (ri = 0; ri < md->pmd_npmc; ri++)
1581 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL &&
1582 PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1583 pmclog_process_map_out(pm->pm_owner, pid,
1584 pkm->pm_address, pkm->pm_address + pkm->pm_size);
1588 * Log mapping information about the kernel.
1592 pmc_log_kernel_mappings(struct pmc *pm)
1594 struct pmc_owner *po;
1595 struct pmckern_map_in *km, *kmbase;
1597 sx_assert(&pmc_sx, SX_LOCKED);
1598 KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)),
1599 ("[pmc,%d] non-sampling PMC (%p) desires mapping information",
1600 __LINE__, (void *) pm));
1604 if (po->po_flags & PMC_PO_INITIAL_MAPPINGS_DONE)
1608 * Log the current set of kernel modules.
1610 kmbase = linker_hwpmc_list_objects();
1611 for (km = kmbase; km->pm_file != NULL; km++) {
1612 PMCDBG(LOG,REG,1,"%s %p", (char *) km->pm_file,
1613 (void *) km->pm_address);
1614 pmclog_process_map_in(po, (pid_t) -1, km->pm_address,
1617 free(kmbase, M_LINKER);
1619 po->po_flags |= PMC_PO_INITIAL_MAPPINGS_DONE;
1623 * Log the mappings for a single process.
1627 pmc_log_process_mappings(struct pmc_owner *po, struct proc *p)
1633 vm_map_entry_t entry;
1634 vm_offset_t last_end;
1635 u_int last_timestamp;
1636 struct vnode *last_vp;
1637 vm_offset_t start_addr;
1638 vm_object_t obj, lobj, tobj;
1639 char *fullpath, *freepath;
1642 last_end = (vm_offset_t) 0;
1643 fullpath = freepath = NULL;
1645 if ((vm = vmspace_acquire_ref(p)) == NULL)
1649 vm_map_lock_read(map);
1651 for (entry = map->header.next; entry != &map->header; entry = entry->next) {
1653 if (entry == NULL) {
1654 PMCDBG(LOG,OPS,2, "hwpmc: vm_map entry unexpectedly "
1655 "NULL! pid=%d vm_map=%p\n", p->p_pid, map);
1660 * We only care about executable map entries.
1662 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) ||
1663 !(entry->protection & VM_PROT_EXECUTE) ||
1664 (entry->object.vm_object == NULL)) {
1668 obj = entry->object.vm_object;
1669 VM_OBJECT_LOCK(obj);
1672 * Walk the backing_object list to find the base
1673 * (non-shadowed) vm_object.
1675 for (lobj = tobj = obj; tobj != NULL; tobj = tobj->backing_object) {
1677 VM_OBJECT_LOCK(tobj);
1679 VM_OBJECT_UNLOCK(lobj);
1684 * At this point lobj is the base vm_object and it is locked.
1687 PMCDBG(LOG,OPS,2, "hwpmc: lobj unexpectedly NULL! pid=%d "
1688 "vm_map=%p vm_obj=%p\n", p->p_pid, map, obj);
1689 VM_OBJECT_UNLOCK(obj);
1693 if (lobj->type != OBJT_VNODE || lobj->handle == NULL) {
1695 VM_OBJECT_UNLOCK(lobj);
1696 VM_OBJECT_UNLOCK(obj);
1701 * Skip contiguous regions that point to the same
1702 * vnode, so we don't emit redundant MAP-IN
1705 if (entry->start == last_end && lobj->handle == last_vp) {
1706 last_end = entry->end;
1708 VM_OBJECT_UNLOCK(lobj);
1709 VM_OBJECT_UNLOCK(obj);
1714 * We don't want to keep the proc's vm_map or this
1715 * vm_object locked while we walk the pathname, since
1716 * vn_fullpath() can sleep. However, if we drop the
1717 * lock, it's possible for concurrent activity to
1718 * modify the vm_map list. To protect against this,
1719 * we save the vm_map timestamp before we release the
1720 * lock, and check it after we reacquire the lock
1723 start_addr = entry->start;
1724 last_end = entry->end;
1725 last_timestamp = map->timestamp;
1726 vm_map_unlock_read(map);
1731 VM_OBJECT_UNLOCK(lobj);
1733 VM_OBJECT_UNLOCK(obj);
1736 pmc_getfilename(vp, &fullpath, &freepath);
1739 locked = VFS_LOCK_GIANT(vp->v_mount);
1741 VFS_UNLOCK_GIANT(locked);
1744 pmclog_process_map_in(po, p->p_pid, start_addr, fullpath);
1746 free(freepath, M_TEMP);
1748 vm_map_lock_read(map);
1751 * If our saved timestamp doesn't match, this means
1752 * that the vm_map was modified out from under us and
1753 * we can't trust our current "entry" pointer. Do a
1754 * new lookup for this entry. If there is no entry
1755 * for this address range, vm_map_lookup_entry() will
1756 * return the previous one, so we always want to go to
1757 * entry->next on the next loop iteration.
1759 * There is an edge condition here that can occur if
1760 * there is no entry at or before this address. In
1761 * this situation, vm_map_lookup_entry returns
1762 * &map->header, which would cause our loop to abort
1763 * without processing the rest of the map. However,
1764 * in practice this will never happen for process
1765 * vm_map. This is because the executable's text
1766 * segment is the first mapping in the proc's address
1767 * space, and this mapping is never removed until the
1768 * process exits, so there will always be a non-header
1769 * entry at or before the requested address for
1770 * vm_map_lookup_entry to return.
1772 if (map->timestamp != last_timestamp)
1773 vm_map_lookup_entry(map, last_end - 1, &entry);
1776 vm_map_unlock_read(map);
1782 * Log mappings for all processes in the system.
1786 pmc_log_all_process_mappings(struct pmc_owner *po)
1788 struct proc *p, *top;
1790 sx_assert(&pmc_sx, SX_XLOCKED);
1792 if ((p = pfind(1)) == NULL)
1793 panic("[pmc,%d] Cannot find init", __LINE__);
1797 sx_slock(&proctree_lock);
1802 pmc_log_process_mappings(po, p);
1803 if (!LIST_EMPTY(&p->p_children))
1804 p = LIST_FIRST(&p->p_children);
1808 if (LIST_NEXT(p, p_sibling)) {
1809 p = LIST_NEXT(p, p_sibling);
1816 sx_sunlock(&proctree_lock);
1820 * The 'hook' invoked from the kernel proper
1825 const char *pmc_hooknames[] = {
1826 /* these strings correspond to PMC_FN_* in <sys/pmckern.h> */
1841 pmc_hook_handler(struct thread *td, int function, void *arg)
1844 PMCDBG(MOD,PMH,1, "hook td=%p func=%d \"%s\" arg=%p", td, function,
1845 pmc_hooknames[function], arg);
1854 case PMC_FN_PROCESS_EXEC:
1856 char *fullpath, *freepath;
1858 int is_using_hwpmcs;
1861 struct pmc_owner *po;
1862 struct pmc_process *pp;
1863 struct pmckern_procexec *pk;
1865 sx_assert(&pmc_sx, SX_XLOCKED);
1868 pmc_getfilename(p->p_textvp, &fullpath, &freepath);
1870 pk = (struct pmckern_procexec *) arg;
1872 /* Inform owners of SS mode PMCs of the exec event. */
1873 LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1874 if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1875 pmclog_process_procexec(po, PMC_ID_INVALID,
1876 p->p_pid, pk->pm_entryaddr, fullpath);
1879 is_using_hwpmcs = p->p_flag & P_HWPMC;
1882 if (!is_using_hwpmcs) {
1884 free(freepath, M_TEMP);
1889 * PMCs are not inherited across an exec(): remove any
1890 * PMCs that this process is the owner of.
1893 if ((po = pmc_find_owner_descriptor(p)) != NULL) {
1894 pmc_remove_owner(po);
1895 pmc_destroy_owner_descriptor(po);
1899 * If the process being exec'ed is not the target of any
1902 if ((pp = pmc_find_process_descriptor(p, 0)) == NULL) {
1904 free(freepath, M_TEMP);
1909 * Log the exec event to all monitoring owners. Skip
1910 * owners who have already recieved the event because
1911 * they had system sampling PMCs active.
1913 for (ri = 0; ri < md->pmd_npmc; ri++)
1914 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) {
1916 if (po->po_sscount == 0 &&
1917 po->po_flags & PMC_PO_OWNS_LOGFILE)
1918 pmclog_process_procexec(po, pm->pm_id,
1919 p->p_pid, pk->pm_entryaddr,
1924 free(freepath, M_TEMP);
1927 PMCDBG(PRC,EXC,1, "exec proc=%p (%d, %s) cred-changed=%d",
1928 p, p->p_pid, p->p_comm, pk->pm_credentialschanged);
1930 if (pk->pm_credentialschanged == 0) /* no change */
1934 * If the newly exec()'ed process has a different credential
1935 * than before, allow it to be the target of a PMC only if
1936 * the PMC's owner has sufficient priviledge.
1939 for (ri = 0; ri < md->pmd_npmc; ri++)
1940 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL)
1941 if (pmc_can_attach(pm, td->td_proc) != 0)
1942 pmc_detach_one_process(td->td_proc,
1945 KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= (int) md->pmd_npmc,
1946 ("[pmc,%d] Illegal ref count %d on pp %p", __LINE__,
1947 pp->pp_refcnt, pp));
1950 * If this process is no longer the target of any
1951 * PMCs, we can remove the process entry and free
1955 if (pp->pp_refcnt == 0) {
1956 pmc_remove_process_descriptor(pp);
1965 pmc_process_csw_in(td);
1968 case PMC_FN_CSW_OUT:
1969 pmc_process_csw_out(td);
1973 * Process accumulated PC samples.
1975 * This function is expected to be called by hardclock() for
1976 * each CPU that has accumulated PC samples.
1978 * This function is to be executed on the CPU whose samples
1979 * are being processed.
1981 case PMC_FN_DO_SAMPLES:
1984 * Clear the cpu specific bit in the CPU mask before
1985 * do the rest of the processing. If the NMI handler
1986 * gets invoked after the "atomic_clear_int()" call
1987 * below but before "pmc_process_samples()" gets
1988 * around to processing the interrupt, then we will
1989 * come back here at the next hardclock() tick (and
1990 * may find nothing to do if "pmc_process_samples()"
1991 * had already processed the interrupt). We don't
1992 * lose the interrupt sample.
1994 atomic_clear_int(&pmc_cpumask, (1 << PCPU_GET(cpuid)));
1995 pmc_process_samples(PCPU_GET(cpuid));
1999 case PMC_FN_KLD_LOAD:
2000 sx_assert(&pmc_sx, SX_LOCKED);
2001 pmc_process_kld_load((struct pmckern_map_in *) arg);
2004 case PMC_FN_KLD_UNLOAD:
2005 sx_assert(&pmc_sx, SX_LOCKED);
2006 pmc_process_kld_unload((struct pmckern_map_out *) arg);
2010 sx_assert(&pmc_sx, SX_LOCKED);
2011 pmc_process_mmap(td, (struct pmckern_map_in *) arg);
2015 sx_assert(&pmc_sx, SX_LOCKED);
2016 pmc_process_munmap(td, (struct pmckern_map_out *) arg);
2019 case PMC_FN_USER_CALLCHAIN:
2021 * Record a call chain.
2023 KASSERT(td == curthread, ("[pmc,%d] td != curthread",
2025 pmc_capture_user_callchain(PCPU_GET(cpuid),
2026 (struct trapframe *) arg);
2027 td->td_pflags &= ~TDP_CALLCHAIN;
2032 KASSERT(0, ("[pmc,%d] unknown hook %d\n", __LINE__, function));
2042 * allocate a 'struct pmc_owner' descriptor in the owner hash table.
2045 static struct pmc_owner *
2046 pmc_allocate_owner_descriptor(struct proc *p)
2049 struct pmc_owner *po;
2050 struct pmc_ownerhash *poh;
2052 hindex = PMC_HASH_PTR(p, pmc_ownerhashmask);
2053 poh = &pmc_ownerhash[hindex];
2055 /* allocate space for N pointers and one descriptor struct */
2056 po = malloc(sizeof(struct pmc_owner), M_PMC, M_WAITOK|M_ZERO);
2057 po->po_sscount = po->po_error = po->po_flags = po->po_logprocmaps = 0;
2060 po->po_kthread = NULL;
2061 LIST_INIT(&po->po_pmcs);
2062 LIST_INSERT_HEAD(poh, po, po_next); /* insert into hash table */
2064 TAILQ_INIT(&po->po_logbuffers);
2065 mtx_init(&po->po_mtx, "pmc-owner-mtx", "pmc-per-proc", MTX_SPIN);
2067 PMCDBG(OWN,ALL,1, "allocate-owner proc=%p (%d, %s) pmc-owner=%p",
2068 p, p->p_pid, p->p_comm, po);
2074 pmc_destroy_owner_descriptor(struct pmc_owner *po)
2077 PMCDBG(OWN,REL,1, "destroy-owner po=%p proc=%p (%d, %s)",
2078 po, po->po_owner, po->po_owner->p_pid, po->po_owner->p_comm);
2080 mtx_destroy(&po->po_mtx);
2085 * find the descriptor corresponding to process 'p', adding or removing it
2086 * as specified by 'mode'.
2089 static struct pmc_process *
2090 pmc_find_process_descriptor(struct proc *p, uint32_t mode)
2093 struct pmc_process *pp, *ppnew;
2094 struct pmc_processhash *pph;
2096 hindex = PMC_HASH_PTR(p, pmc_processhashmask);
2097 pph = &pmc_processhash[hindex];
2102 * Pre-allocate memory in the FIND_ALLOCATE case since we
2103 * cannot call malloc(9) once we hold a spin lock.
2105 if (mode & PMC_FLAG_ALLOCATE)
2106 ppnew = malloc(sizeof(struct pmc_process) + md->pmd_npmc *
2107 sizeof(struct pmc_targetstate), M_PMC, M_WAITOK|M_ZERO);
2109 mtx_lock_spin(&pmc_processhash_mtx);
2110 LIST_FOREACH(pp, pph, pp_next)
2111 if (pp->pp_proc == p)
2114 if ((mode & PMC_FLAG_REMOVE) && pp != NULL)
2115 LIST_REMOVE(pp, pp_next);
2117 if ((mode & PMC_FLAG_ALLOCATE) && pp == NULL &&
2120 LIST_INSERT_HEAD(pph, ppnew, pp_next);
2124 mtx_unlock_spin(&pmc_processhash_mtx);
2126 if (pp != NULL && ppnew != NULL)
2133 * remove a process descriptor from the process hash table.
2137 pmc_remove_process_descriptor(struct pmc_process *pp)
2139 KASSERT(pp->pp_refcnt == 0,
2140 ("[pmc,%d] Removing process descriptor %p with count %d",
2141 __LINE__, pp, pp->pp_refcnt));
2143 mtx_lock_spin(&pmc_processhash_mtx);
2144 LIST_REMOVE(pp, pp_next);
2145 mtx_unlock_spin(&pmc_processhash_mtx);
2150 * find an owner descriptor corresponding to proc 'p'
2153 static struct pmc_owner *
2154 pmc_find_owner_descriptor(struct proc *p)
2157 struct pmc_owner *po;
2158 struct pmc_ownerhash *poh;
2160 hindex = PMC_HASH_PTR(p, pmc_ownerhashmask);
2161 poh = &pmc_ownerhash[hindex];
2164 LIST_FOREACH(po, poh, po_next)
2165 if (po->po_owner == p)
2168 PMCDBG(OWN,FND,1, "find-owner proc=%p (%d, %s) hindex=0x%x -> "
2169 "pmc-owner=%p", p, p->p_pid, p->p_comm, hindex, po);
2175 * pmc_allocate_pmc_descriptor
2177 * Allocate a pmc descriptor and initialize its
2182 pmc_allocate_pmc_descriptor(void)
2186 pmc = malloc(sizeof(struct pmc), M_PMC, M_WAITOK|M_ZERO);
2189 pmc->pm_owner = NULL;
2190 LIST_INIT(&pmc->pm_targets);
2193 PMCDBG(PMC,ALL,1, "allocate-pmc -> pmc=%p", pmc);
2199 * Destroy a pmc descriptor.
2203 pmc_destroy_pmc_descriptor(struct pmc *pm)
2208 KASSERT(pm->pm_state == PMC_STATE_DELETED ||
2209 pm->pm_state == PMC_STATE_FREE,
2210 ("[pmc,%d] destroying non-deleted PMC", __LINE__));
2211 KASSERT(LIST_EMPTY(&pm->pm_targets),
2212 ("[pmc,%d] destroying pmc with targets", __LINE__));
2213 KASSERT(pm->pm_owner == NULL,
2214 ("[pmc,%d] destroying pmc attached to an owner", __LINE__));
2215 KASSERT(pm->pm_runcount == 0,
2216 ("[pmc,%d] pmc has non-zero run count %d", __LINE__,
2222 pmc_wait_for_pmc_idle(struct pmc *pm)
2225 volatile int maxloop;
2227 maxloop = 100 * pmc_cpu_max();
2231 * Loop (with a forced context switch) till the PMC's runcount
2232 * comes down to zero.
2234 while (atomic_load_acq_32(&pm->pm_runcount) > 0) {
2237 KASSERT(maxloop > 0,
2238 ("[pmc,%d] (ri%d, rc%d) waiting too long for "
2239 "pmc to be free", __LINE__,
2240 PMC_TO_ROWINDEX(pm), pm->pm_runcount));
2242 pmc_force_context_switch();
2247 * This function does the following things:
2249 * - detaches the PMC from hardware
2250 * - unlinks all target threads that were attached to it
2251 * - removes the PMC from its owner's list
2252 * - destroy's the PMC private mutex
2254 * Once this function completes, the given pmc pointer can be safely
2255 * FREE'd by the caller.
2259 pmc_release_pmc_descriptor(struct pmc *pm)
2263 u_int adjri, ri, cpu;
2264 struct pmc_owner *po;
2265 struct pmc_binding pb;
2266 struct pmc_process *pp;
2267 struct pmc_classdep *pcd;
2268 struct pmc_target *ptgt, *tmp;
2270 sx_assert(&pmc_sx, SX_XLOCKED);
2272 KASSERT(pm, ("[pmc,%d] null pmc", __LINE__));
2274 ri = PMC_TO_ROWINDEX(pm);
2275 pcd = pmc_ri_to_classdep(md, ri, &adjri);
2276 mode = PMC_TO_MODE(pm);
2278 PMCDBG(PMC,REL,1, "release-pmc pmc=%p ri=%d mode=%d", pm, ri,
2282 * First, we take the PMC off hardware.
2285 if (PMC_IS_SYSTEM_MODE(mode)) {
2288 * A system mode PMC runs on a specific CPU. Switch
2289 * to this CPU and turn hardware off.
2291 pmc_save_cpu_binding(&pb);
2293 cpu = PMC_TO_CPU(pm);
2295 pmc_select_cpu(cpu);
2297 /* switch off non-stalled CPUs */
2298 if (pm->pm_state == PMC_STATE_RUNNING &&
2299 pm->pm_stalled == 0) {
2301 phw = pmc_pcpu[cpu]->pc_hwpmcs[ri];
2303 KASSERT(phw->phw_pmc == pm,
2304 ("[pmc, %d] pmc ptr ri(%d) hw(%p) pm(%p)",
2305 __LINE__, ri, phw->phw_pmc, pm));
2306 PMCDBG(PMC,REL,2, "stopping cpu=%d ri=%d", cpu, ri);
2309 pcd->pcd_stop_pmc(cpu, adjri);
2313 PMCDBG(PMC,REL,2, "decfg cpu=%d ri=%d", cpu, ri);
2316 pcd->pcd_config_pmc(cpu, adjri, NULL);
2319 /* adjust the global and process count of SS mode PMCs */
2320 if (mode == PMC_MODE_SS && pm->pm_state == PMC_STATE_RUNNING) {
2323 if (po->po_sscount == 0) {
2324 atomic_subtract_rel_int(&pmc_ss_count, 1);
2325 LIST_REMOVE(po, po_ssnext);
2329 pm->pm_state = PMC_STATE_DELETED;
2331 pmc_restore_cpu_binding(&pb);
2334 * We could have references to this PMC structure in
2335 * the per-cpu sample queues. Wait for the queue to
2338 pmc_wait_for_pmc_idle(pm);
2340 } else if (PMC_IS_VIRTUAL_MODE(mode)) {
2343 * A virtual PMC could be running on multiple CPUs at
2346 * By marking its state as DELETED, we ensure that
2347 * this PMC is never further scheduled on hardware.
2349 * Then we wait till all CPUs are done with this PMC.
2351 pm->pm_state = PMC_STATE_DELETED;
2354 /* Wait for the PMCs runcount to come to zero. */
2355 pmc_wait_for_pmc_idle(pm);
2358 * At this point the PMC is off all CPUs and cannot be
2359 * freshly scheduled onto a CPU. It is now safe to
2360 * unlink all targets from this PMC. If a
2361 * process-record's refcount falls to zero, we remove
2362 * it from the hash table. The module-wide SX lock
2363 * protects us from races.
2365 LIST_FOREACH_SAFE(ptgt, &pm->pm_targets, pt_next, tmp) {
2366 pp = ptgt->pt_process;
2367 pmc_unlink_target_process(pm, pp); /* frees 'ptgt' */
2369 PMCDBG(PMC,REL,3, "pp->refcnt=%d", pp->pp_refcnt);
2372 * If the target process record shows that no
2373 * PMCs are attached to it, reclaim its space.
2376 if (pp->pp_refcnt == 0) {
2377 pmc_remove_process_descriptor(pp);
2382 cpu = curthread->td_oncpu; /* setup cpu for pmd_release() */
2387 * Release any MD resources
2389 (void) pcd->pcd_release_pmc(cpu, adjri, pm);
2392 * Update row disposition
2395 if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm)))
2396 PMC_UNMARK_ROW_STANDALONE(ri);
2398 PMC_UNMARK_ROW_THREAD(ri);
2400 /* unlink from the owner's list */
2402 LIST_REMOVE(pm, pm_next);
2403 pm->pm_owner = NULL;
2406 pmc_destroy_pmc_descriptor(pm);
2410 * Register an owner and a pmc.
2414 pmc_register_owner(struct proc *p, struct pmc *pmc)
2416 struct pmc_owner *po;
2418 sx_assert(&pmc_sx, SX_XLOCKED);
2420 if ((po = pmc_find_owner_descriptor(p)) == NULL)
2421 if ((po = pmc_allocate_owner_descriptor(p)) == NULL)
2424 KASSERT(pmc->pm_owner == NULL,
2425 ("[pmc,%d] attempting to own an initialized PMC", __LINE__));
2428 LIST_INSERT_HEAD(&po->po_pmcs, pmc, pm_next);
2431 p->p_flag |= P_HWPMC;
2434 if (po->po_flags & PMC_PO_OWNS_LOGFILE)
2435 pmclog_process_pmcallocate(pmc);
2437 PMCDBG(PMC,REG,1, "register-owner pmc-owner=%p pmc=%p",
2444 * Return the current row disposition:
2446 * > 0 => PROCESS MODE
2447 * < 0 => SYSTEM MODE
2451 pmc_getrowdisp(int ri)
2453 return pmc_pmcdisp[ri];
2457 * Check if a PMC at row index 'ri' can be allocated to the current
2460 * Allocation can fail if:
2461 * - the current process is already being profiled by a PMC at index 'ri',
2462 * attached to it via OP_PMCATTACH.
2463 * - the current process has already allocated a PMC at index 'ri'
2468 pmc_can_allocate_rowindex(struct proc *p, unsigned int ri, int cpu)
2472 struct pmc_owner *po;
2473 struct pmc_process *pp;
2475 PMCDBG(PMC,ALR,1, "can-allocate-rowindex proc=%p (%d, %s) ri=%d "
2476 "cpu=%d", p, p->p_pid, p->p_comm, ri, cpu);
2479 * We shouldn't have already allocated a process-mode PMC at
2482 * We shouldn't have allocated a system-wide PMC on the same
2485 if ((po = pmc_find_owner_descriptor(p)) != NULL)
2486 LIST_FOREACH(pm, &po->po_pmcs, pm_next) {
2487 if (PMC_TO_ROWINDEX(pm) == ri) {
2488 mode = PMC_TO_MODE(pm);
2489 if (PMC_IS_VIRTUAL_MODE(mode))
2491 if (PMC_IS_SYSTEM_MODE(mode) &&
2492 (int) PMC_TO_CPU(pm) == cpu)
2498 * We also shouldn't be the target of any PMC at this index
2499 * since otherwise a PMC_ATTACH to ourselves will fail.
2501 if ((pp = pmc_find_process_descriptor(p, 0)) != NULL)
2502 if (pp->pp_pmcs[ri].pp_pmc)
2505 PMCDBG(PMC,ALR,2, "can-allocate-rowindex proc=%p (%d, %s) ri=%d ok",
2506 p, p->p_pid, p->p_comm, ri);
2512 * Check if a given PMC at row index 'ri' can be currently used in
2517 pmc_can_allocate_row(int ri, enum pmc_mode mode)
2521 sx_assert(&pmc_sx, SX_XLOCKED);
2523 PMCDBG(PMC,ALR,1, "can-allocate-row ri=%d mode=%d", ri, mode);
2525 if (PMC_IS_SYSTEM_MODE(mode))
2526 disp = PMC_DISP_STANDALONE;
2528 disp = PMC_DISP_THREAD;
2531 * check disposition for PMC row 'ri':
2533 * Expected disposition Row-disposition Result
2535 * STANDALONE STANDALONE or FREE proceed
2536 * STANDALONE THREAD fail
2537 * THREAD THREAD or FREE proceed
2538 * THREAD STANDALONE fail
2541 if (!PMC_ROW_DISP_IS_FREE(ri) &&
2542 !(disp == PMC_DISP_THREAD && PMC_ROW_DISP_IS_THREAD(ri)) &&
2543 !(disp == PMC_DISP_STANDALONE && PMC_ROW_DISP_IS_STANDALONE(ri)))
2550 PMCDBG(PMC,ALR,2, "can-allocate-row ri=%d mode=%d ok", ri, mode);
2557 * Find a PMC descriptor with user handle 'pmcid' for thread 'td'.
2561 pmc_find_pmc_descriptor_in_process(struct pmc_owner *po, pmc_id_t pmcid)
2565 KASSERT(PMC_ID_TO_ROWINDEX(pmcid) < md->pmd_npmc,
2566 ("[pmc,%d] Illegal pmc index %d (max %d)", __LINE__,
2567 PMC_ID_TO_ROWINDEX(pmcid), md->pmd_npmc));
2569 LIST_FOREACH(pm, &po->po_pmcs, pm_next)
2570 if (pm->pm_id == pmcid)
2577 pmc_find_pmc(pmc_id_t pmcid, struct pmc **pmc)
2581 struct pmc_owner *po;
2583 PMCDBG(PMC,FND,1, "find-pmc id=%d", pmcid);
2585 if ((po = pmc_find_owner_descriptor(curthread->td_proc)) == NULL)
2588 if ((pm = pmc_find_pmc_descriptor_in_process(po, pmcid)) == NULL)
2591 PMCDBG(PMC,FND,2, "find-pmc id=%d -> pmc=%p", pmcid, pm);
2602 pmc_start(struct pmc *pm)
2605 struct pmc_owner *po;
2606 struct pmc_binding pb;
2607 struct pmc_classdep *pcd;
2608 int adjri, error, cpu, ri;
2611 ("[pmc,%d] null pm", __LINE__));
2613 mode = PMC_TO_MODE(pm);
2614 ri = PMC_TO_ROWINDEX(pm);
2615 pcd = pmc_ri_to_classdep(md, ri, &adjri);
2619 PMCDBG(PMC,OPS,1, "start pmc=%p mode=%d ri=%d", pm, mode, ri);
2624 * Disallow PMCSTART if a logfile is required but has not been
2627 if ((pm->pm_flags & PMC_F_NEEDS_LOGFILE) &&
2628 (po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
2629 return (EDOOFUS); /* programming error */
2632 * If this is a sampling mode PMC, log mapping information for
2633 * the kernel modules that are currently loaded.
2635 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
2636 pmc_log_kernel_mappings(pm);
2638 if (PMC_IS_VIRTUAL_MODE(mode)) {
2641 * If a PMCATTACH has never been done on this PMC,
2642 * attach it to its owner process.
2645 if (LIST_EMPTY(&pm->pm_targets))
2646 error = (pm->pm_flags & PMC_F_ATTACH_DONE) ? ESRCH :
2647 pmc_attach_process(po->po_owner, pm);
2650 * If the PMC is attached to its owner, then force a context
2651 * switch to ensure that the MD state gets set correctly.
2655 pm->pm_state = PMC_STATE_RUNNING;
2656 if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER)
2657 pmc_force_context_switch();
2665 * A system-wide PMC.
2667 * Add the owner to the global list if this is a system-wide
2671 if (mode == PMC_MODE_SS) {
2672 if (po->po_sscount == 0) {
2673 LIST_INSERT_HEAD(&pmc_ss_owners, po, po_ssnext);
2674 atomic_add_rel_int(&pmc_ss_count, 1);
2675 PMCDBG(PMC,OPS,1, "po=%p in global list", po);
2680 * Log mapping information for all existing processes in the
2681 * system. Subsequent mappings are logged as they happen;
2682 * see pmc_process_mmap().
2684 if (po->po_logprocmaps == 0) {
2685 pmc_log_all_process_mappings(po);
2686 po->po_logprocmaps = 1;
2691 * Move to the CPU associated with this
2692 * PMC, and start the hardware.
2695 pmc_save_cpu_binding(&pb);
2697 cpu = PMC_TO_CPU(pm);
2699 if (!pmc_cpu_is_active(cpu))
2702 pmc_select_cpu(cpu);
2705 * global PMCs are configured at allocation time
2706 * so write out the initial value and start the PMC.
2709 pm->pm_state = PMC_STATE_RUNNING;
2712 if ((error = pcd->pcd_write_pmc(cpu, adjri,
2713 PMC_IS_SAMPLING_MODE(mode) ?
2714 pm->pm_sc.pm_reloadcount :
2715 pm->pm_sc.pm_initial)) == 0)
2716 error = pcd->pcd_start_pmc(cpu, adjri);
2719 pmc_restore_cpu_binding(&pb);
2729 pmc_stop(struct pmc *pm)
2731 struct pmc_owner *po;
2732 struct pmc_binding pb;
2733 struct pmc_classdep *pcd;
2734 int adjri, cpu, error, ri;
2736 KASSERT(pm != NULL, ("[pmc,%d] null pmc", __LINE__));
2738 PMCDBG(PMC,OPS,1, "stop pmc=%p mode=%d ri=%d", pm,
2739 PMC_TO_MODE(pm), PMC_TO_ROWINDEX(pm));
2741 pm->pm_state = PMC_STATE_STOPPED;
2744 * If the PMC is a virtual mode one, changing the state to
2745 * non-RUNNING is enough to ensure that the PMC never gets
2748 * If this PMC is current running on a CPU, then it will
2749 * handled correctly at the time its target process is context
2753 if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)))
2757 * A system-mode PMC. Move to the CPU associated with
2758 * this PMC, and stop the hardware. We update the
2759 * 'initial count' so that a subsequent PMCSTART will
2760 * resume counting from the current hardware count.
2763 pmc_save_cpu_binding(&pb);
2765 cpu = PMC_TO_CPU(pm);
2767 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
2768 ("[pmc,%d] illegal cpu=%d", __LINE__, cpu));
2770 if (!pmc_cpu_is_active(cpu))
2773 pmc_select_cpu(cpu);
2775 ri = PMC_TO_ROWINDEX(pm);
2776 pcd = pmc_ri_to_classdep(md, ri, &adjri);
2779 if ((error = pcd->pcd_stop_pmc(cpu, adjri)) == 0)
2780 error = pcd->pcd_read_pmc(cpu, adjri, &pm->pm_sc.pm_initial);
2783 pmc_restore_cpu_binding(&pb);
2787 /* remove this owner from the global list of SS PMC owners */
2788 if (PMC_TO_MODE(pm) == PMC_MODE_SS) {
2790 if (po->po_sscount == 0) {
2791 atomic_subtract_rel_int(&pmc_ss_count, 1);
2792 LIST_REMOVE(po, po_ssnext);
2793 PMCDBG(PMC,OPS,2,"po=%p removed from global list", po);
2802 static const char *pmc_op_to_name[] = {
2804 #define __PMC_OP(N, D) #N ,
2811 * The syscall interface
2814 #define PMC_GET_SX_XLOCK(...) do { \
2815 sx_xlock(&pmc_sx); \
2816 if (pmc_hook == NULL) { \
2817 sx_xunlock(&pmc_sx); \
2818 return __VA_ARGS__; \
2822 #define PMC_DOWNGRADE_SX() do { \
2823 sx_downgrade(&pmc_sx); \
2824 is_sx_downgraded = 1; \
2828 pmc_syscall_handler(struct thread *td, void *syscall_args)
2830 int error, is_sx_downgraded, is_sx_locked, op;
2831 struct pmc_syscall_args *c;
2834 PMC_GET_SX_XLOCK(ENOSYS);
2838 is_sx_downgraded = 0;
2841 c = (struct pmc_syscall_args *) syscall_args;
2846 PMCDBG(MOD,PMS,1, "syscall op=%d \"%s\" arg=%p", op,
2847 pmc_op_to_name[op], arg);
2850 atomic_add_int(&pmc_stats.pm_syscalls, 1);
2857 * Configure a log file.
2859 * XXX This OP will be reworked.
2862 case PMC_OP_CONFIGURELOG:
2866 struct pmc_owner *po;
2867 struct pmc_op_configurelog cl;
2869 sx_assert(&pmc_sx, SX_XLOCKED);
2871 if ((error = copyin(arg, &cl, sizeof(cl))) != 0)
2874 /* mark this process as owning a log file */
2876 if ((po = pmc_find_owner_descriptor(p)) == NULL)
2877 if ((po = pmc_allocate_owner_descriptor(p)) == NULL) {
2883 * If a valid fd was passed in, try to configure that,
2884 * otherwise if 'fd' was less than zero and there was
2885 * a log file configured, flush its buffers and
2888 if (cl.pm_logfd >= 0) {
2889 sx_xunlock(&pmc_sx);
2891 error = pmclog_configure_log(md, po, cl.pm_logfd);
2892 } else if (po->po_flags & PMC_PO_OWNS_LOGFILE) {
2893 pmclog_process_closelog(po);
2894 error = pmclog_flush(po);
2896 LIST_FOREACH(pm, &po->po_pmcs, pm_next)
2897 if (pm->pm_flags & PMC_F_NEEDS_LOGFILE &&
2898 pm->pm_state == PMC_STATE_RUNNING)
2900 error = pmclog_deconfigure_log(po);
2915 case PMC_OP_FLUSHLOG:
2917 struct pmc_owner *po;
2919 sx_assert(&pmc_sx, SX_XLOCKED);
2921 if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) {
2926 error = pmclog_flush(po);
2931 * Retrieve hardware configuration.
2934 case PMC_OP_GETCPUINFO: /* CPU information */
2936 struct pmc_op_getcpuinfo gci;
2937 struct pmc_classinfo *pci;
2938 struct pmc_classdep *pcd;
2941 gci.pm_cputype = md->pmd_cputype;
2942 gci.pm_ncpu = pmc_cpu_max();
2943 gci.pm_npmc = md->pmd_npmc;
2944 gci.pm_nclass = md->pmd_nclass;
2945 pci = gci.pm_classes;
2946 pcd = md->pmd_classdep;
2947 for (cl = 0; cl < md->pmd_nclass; cl++, pci++, pcd++) {
2948 pci->pm_caps = pcd->pcd_caps;
2949 pci->pm_class = pcd->pcd_class;
2950 pci->pm_width = pcd->pcd_width;
2951 pci->pm_num = pcd->pcd_num;
2953 error = copyout(&gci, arg, sizeof(gci));
2959 * Get module statistics
2962 case PMC_OP_GETDRIVERSTATS:
2964 struct pmc_op_getdriverstats gms;
2966 bcopy(&pmc_stats, &gms, sizeof(gms));
2967 error = copyout(&gms, arg, sizeof(gms));
2973 * Retrieve module version number
2976 case PMC_OP_GETMODULEVERSION:
2980 /* retrieve the client's idea of the ABI version */
2981 if ((error = copyin(arg, &cv, sizeof(uint32_t))) != 0)
2983 /* don't service clients newer than our driver */
2985 if ((cv & 0xFFFF0000) > (modv & 0xFFFF0000)) {
2986 error = EPROGMISMATCH;
2989 error = copyout(&modv, arg, sizeof(int));
2995 * Retrieve the state of all the PMCs on a given
2999 case PMC_OP_GETPMCINFO:
3003 size_t pmcinfo_size;
3004 uint32_t cpu, n, npmc;
3005 struct pmc_owner *po;
3006 struct pmc_binding pb;
3007 struct pmc_classdep *pcd;
3008 struct pmc_info *p, *pmcinfo;
3009 struct pmc_op_getpmcinfo *gpi;
3013 gpi = (struct pmc_op_getpmcinfo *) arg;
3015 if ((error = copyin(&gpi->pm_cpu, &cpu, sizeof(cpu))) != 0)
3018 if (cpu >= pmc_cpu_max()) {
3023 if (!pmc_cpu_is_active(cpu)) {
3028 /* switch to CPU 'cpu' */
3029 pmc_save_cpu_binding(&pb);
3030 pmc_select_cpu(cpu);
3032 npmc = md->pmd_npmc;
3034 pmcinfo_size = npmc * sizeof(struct pmc_info);
3035 pmcinfo = malloc(pmcinfo_size, M_PMC, M_WAITOK);
3039 for (n = 0; n < md->pmd_npmc; n++, p++) {
3041 pcd = pmc_ri_to_classdep(md, n, &ari);
3043 KASSERT(pcd != NULL,
3044 ("[pmc,%d] null pcd ri=%d", __LINE__, n));
3046 if ((error = pcd->pcd_describe(cpu, ari, p, &pm)) != 0)
3049 if (PMC_ROW_DISP_IS_STANDALONE(n))
3050 p->pm_rowdisp = PMC_DISP_STANDALONE;
3051 else if (PMC_ROW_DISP_IS_THREAD(n))
3052 p->pm_rowdisp = PMC_DISP_THREAD;
3054 p->pm_rowdisp = PMC_DISP_FREE;
3056 p->pm_ownerpid = -1;
3058 if (pm == NULL) /* no PMC associated */
3063 KASSERT(po->po_owner != NULL,
3064 ("[pmc,%d] pmc_owner had a null proc pointer",
3067 p->pm_ownerpid = po->po_owner->p_pid;
3068 p->pm_mode = PMC_TO_MODE(pm);
3069 p->pm_event = pm->pm_event;
3070 p->pm_flags = pm->pm_flags;
3072 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
3074 pm->pm_sc.pm_reloadcount;
3077 pmc_restore_cpu_binding(&pb);
3079 /* now copy out the PMC info collected */
3081 error = copyout(pmcinfo, &gpi->pm_pmcs, pmcinfo_size);
3083 free(pmcinfo, M_PMC);
3089 * Set the administrative state of a PMC. I.e. whether
3090 * the PMC is to be used or not.
3093 case PMC_OP_PMCADMIN:
3096 enum pmc_state request;
3099 struct pmc_op_pmcadmin pma;
3100 struct pmc_binding pb;
3102 sx_assert(&pmc_sx, SX_XLOCKED);
3104 KASSERT(td == curthread,
3105 ("[pmc,%d] td != curthread", __LINE__));
3107 error = priv_check(td, PRIV_PMC_MANAGE);
3111 if ((error = copyin(arg, &pma, sizeof(pma))) != 0)
3116 if (cpu < 0 || cpu >= (int) pmc_cpu_max()) {
3121 if (!pmc_cpu_is_active(cpu)) {
3126 request = pma.pm_state;
3128 if (request != PMC_STATE_DISABLED &&
3129 request != PMC_STATE_FREE) {
3134 ri = pma.pm_pmc; /* pmc id == row index */
3135 if (ri < 0 || ri >= (int) md->pmd_npmc) {
3141 * We can't disable a PMC with a row-index allocated
3142 * for process virtual PMCs.
3145 if (PMC_ROW_DISP_IS_THREAD(ri) &&
3146 request == PMC_STATE_DISABLED) {
3152 * otherwise, this PMC on this CPU is either free or
3153 * in system-wide mode.
3156 pmc_save_cpu_binding(&pb);
3157 pmc_select_cpu(cpu);
3160 phw = pc->pc_hwpmcs[ri];
3163 * XXX do we need some kind of 'forced' disable?
3166 if (phw->phw_pmc == NULL) {
3167 if (request == PMC_STATE_DISABLED &&
3168 (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED)) {
3169 phw->phw_state &= ~PMC_PHW_FLAG_IS_ENABLED;
3170 PMC_MARK_ROW_STANDALONE(ri);
3171 } else if (request == PMC_STATE_FREE &&
3172 (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0) {
3173 phw->phw_state |= PMC_PHW_FLAG_IS_ENABLED;
3174 PMC_UNMARK_ROW_STANDALONE(ri);
3176 /* other cases are a no-op */
3180 pmc_restore_cpu_binding(&pb);
3189 case PMC_OP_PMCALLOCATE:
3197 struct pmc_binding pb;
3198 struct pmc_classdep *pcd;
3199 struct pmc_op_pmcallocate pa;
3201 if ((error = copyin(arg, &pa, sizeof(pa))) != 0)
3208 if ((mode != PMC_MODE_SS && mode != PMC_MODE_SC &&
3209 mode != PMC_MODE_TS && mode != PMC_MODE_TC) ||
3210 (cpu != (u_int) PMC_CPU_ANY && cpu >= pmc_cpu_max())) {
3216 * Virtual PMCs should only ask for a default CPU.
3217 * System mode PMCs need to specify a non-default CPU.
3220 if ((PMC_IS_VIRTUAL_MODE(mode) && cpu != (u_int) PMC_CPU_ANY) ||
3221 (PMC_IS_SYSTEM_MODE(mode) && cpu == (u_int) PMC_CPU_ANY)) {
3227 * Check that an inactive CPU is not being asked for.
3230 if (PMC_IS_SYSTEM_MODE(mode) && !pmc_cpu_is_active(cpu)) {
3236 * Refuse an allocation for a system-wide PMC if this
3237 * process has been jailed, or if this process lacks
3238 * super-user credentials and the sysctl tunable
3239 * 'security.bsd.unprivileged_syspmcs' is zero.
3242 if (PMC_IS_SYSTEM_MODE(mode)) {
3243 if (jailed(curthread->td_ucred)) {
3247 if (!pmc_unprivileged_syspmcs) {
3248 error = priv_check(curthread,
3259 * Look for valid values for 'pm_flags'
3262 if ((pa.pm_flags & ~(PMC_F_DESCENDANTS | PMC_F_LOG_PROCCSW |
3263 PMC_F_LOG_PROCEXIT | PMC_F_CALLCHAIN)) != 0) {
3268 /* process logging options are not allowed for system PMCs */
3269 if (PMC_IS_SYSTEM_MODE(mode) && (pa.pm_flags &
3270 (PMC_F_LOG_PROCCSW | PMC_F_LOG_PROCEXIT))) {
3276 * All sampling mode PMCs need to be able to interrupt the
3279 if (PMC_IS_SAMPLING_MODE(mode))
3280 caps |= PMC_CAP_INTERRUPT;
3282 /* A valid class specifier should have been passed in. */
3283 for (n = 0; n < md->pmd_nclass; n++)
3284 if (md->pmd_classdep[n].pcd_class == pa.pm_class)
3286 if (n == md->pmd_nclass) {
3291 /* The requested PMC capabilities should be feasible. */
3292 if ((md->pmd_classdep[n].pcd_caps & caps) != caps) {
3297 PMCDBG(PMC,ALL,2, "event=%d caps=0x%x mode=%d cpu=%d",
3298 pa.pm_ev, caps, mode, cpu);
3300 pmc = pmc_allocate_pmc_descriptor();
3301 pmc->pm_id = PMC_ID_MAKE_ID(cpu,pa.pm_mode,pa.pm_class,
3303 pmc->pm_event = pa.pm_ev;
3304 pmc->pm_state = PMC_STATE_FREE;
3305 pmc->pm_caps = caps;
3306 pmc->pm_flags = pa.pm_flags;
3308 /* switch thread to CPU 'cpu' */
3309 pmc_save_cpu_binding(&pb);
3311 #define PMC_IS_SHAREABLE_PMC(cpu, n) \
3312 (pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_state & \
3313 PMC_PHW_FLAG_IS_SHAREABLE)
3314 #define PMC_IS_UNALLOCATED(cpu, n) \
3315 (pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_pmc == NULL)
3317 if (PMC_IS_SYSTEM_MODE(mode)) {
3318 pmc_select_cpu(cpu);
3319 for (n = 0; n < (int) md->pmd_npmc; n++) {
3320 pcd = pmc_ri_to_classdep(md, n, &adjri);
3321 if (pmc_can_allocate_row(n, mode) == 0 &&
3322 pmc_can_allocate_rowindex(
3323 curthread->td_proc, n, cpu) == 0 &&
3324 (PMC_IS_UNALLOCATED(cpu, n) ||
3325 PMC_IS_SHAREABLE_PMC(cpu, n)) &&
3326 pcd->pcd_allocate_pmc(cpu, adjri, pmc,
3331 /* Process virtual mode */
3332 for (n = 0; n < (int) md->pmd_npmc; n++) {
3333 pcd = pmc_ri_to_classdep(md, n, &adjri);
3334 if (pmc_can_allocate_row(n, mode) == 0 &&
3335 pmc_can_allocate_rowindex(
3336 curthread->td_proc, n,
3337 PMC_CPU_ANY) == 0 &&
3338 pcd->pcd_allocate_pmc(curthread->td_oncpu,
3339 adjri, pmc, &pa) == 0)
3344 #undef PMC_IS_UNALLOCATED
3345 #undef PMC_IS_SHAREABLE_PMC
3347 pmc_restore_cpu_binding(&pb);
3349 if (n == (int) md->pmd_npmc) {
3350 pmc_destroy_pmc_descriptor(pmc);
3357 /* Fill in the correct value in the ID field */
3358 pmc->pm_id = PMC_ID_MAKE_ID(cpu,mode,pa.pm_class,n);
3360 PMCDBG(PMC,ALL,2, "ev=%d class=%d mode=%d n=%d -> pmcid=%x",
3361 pmc->pm_event, pa.pm_class, mode, n, pmc->pm_id);
3363 /* Process mode PMCs with logging enabled need log files */
3364 if (pmc->pm_flags & (PMC_F_LOG_PROCEXIT | PMC_F_LOG_PROCCSW))
3365 pmc->pm_flags |= PMC_F_NEEDS_LOGFILE;
3367 /* All system mode sampling PMCs require a log file */
3368 if (PMC_IS_SAMPLING_MODE(mode) && PMC_IS_SYSTEM_MODE(mode))
3369 pmc->pm_flags |= PMC_F_NEEDS_LOGFILE;
3372 * Configure global pmc's immediately
3375 if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pmc))) {
3377 pmc_save_cpu_binding(&pb);
3378 pmc_select_cpu(cpu);
3380 phw = pmc_pcpu[cpu]->pc_hwpmcs[n];
3381 pcd = pmc_ri_to_classdep(md, n, &adjri);
3383 if ((phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0 ||
3384 (error = pcd->pcd_config_pmc(cpu, adjri, pmc)) != 0) {
3385 (void) pcd->pcd_release_pmc(cpu, adjri, pmc);
3386 pmc_destroy_pmc_descriptor(pmc);
3389 pmc_restore_cpu_binding(&pb);
3394 pmc_restore_cpu_binding(&pb);
3397 pmc->pm_state = PMC_STATE_ALLOCATED;
3400 * mark row disposition
3403 if (PMC_IS_SYSTEM_MODE(mode))
3404 PMC_MARK_ROW_STANDALONE(n);
3406 PMC_MARK_ROW_THREAD(n);
3409 * Register this PMC with the current thread as its owner.
3413 pmc_register_owner(curthread->td_proc, pmc)) != 0) {
3414 pmc_release_pmc_descriptor(pmc);
3421 * Return the allocated index.
3424 pa.pm_pmcid = pmc->pm_id;
3426 error = copyout(&pa, arg, sizeof(pa));
3432 * Attach a PMC to a process.
3435 case PMC_OP_PMCATTACH:
3439 struct pmc_op_pmcattach a;
3441 sx_assert(&pmc_sx, SX_XLOCKED);
3443 if ((error = copyin(arg, &a, sizeof(a))) != 0)
3449 } else if (a.pm_pid == 0)
3450 a.pm_pid = td->td_proc->p_pid;
3452 if ((error = pmc_find_pmc(a.pm_pmc, &pm)) != 0)
3455 if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm))) {
3460 /* PMCs may be (re)attached only when allocated or stopped */
3461 if (pm->pm_state == PMC_STATE_RUNNING) {
3464 } else if (pm->pm_state != PMC_STATE_ALLOCATED &&
3465 pm->pm_state != PMC_STATE_STOPPED) {
3471 if ((p = pfind(a.pm_pid)) == NULL) {
3477 * Ignore processes that are working on exiting.
3479 if (p->p_flag & P_WEXIT) {
3481 PROC_UNLOCK(p); /* pfind() returns a locked process */
3486 * we are allowed to attach a PMC to a process if
3489 error = p_candebug(curthread, p);
3494 error = pmc_attach_process(p, pm);
3500 * Detach an attached PMC from a process.
3503 case PMC_OP_PMCDETACH:
3507 struct pmc_op_pmcattach a;
3509 if ((error = copyin(arg, &a, sizeof(a))) != 0)
3515 } else if (a.pm_pid == 0)
3516 a.pm_pid = td->td_proc->p_pid;
3518 if ((error = pmc_find_pmc(a.pm_pmc, &pm)) != 0)
3521 if ((p = pfind(a.pm_pid)) == NULL) {
3527 * Treat processes that are in the process of exiting
3528 * as if they were not present.
3531 if (p->p_flag & P_WEXIT)
3534 PROC_UNLOCK(p); /* pfind() returns a locked process */
3537 error = pmc_detach_process(p, pm);
3543 * Retrieve the MSR number associated with the counter
3544 * 'pmc_id'. This allows processes to directly use RDPMC
3545 * instructions to read their PMCs, without the overhead of a
3549 case PMC_OP_PMCGETMSR:
3553 struct pmc_target *pt;
3554 struct pmc_op_getmsr gm;
3555 struct pmc_classdep *pcd;
3559 if ((error = copyin(arg, &gm, sizeof(gm))) != 0)
3562 if ((error = pmc_find_pmc(gm.pm_pmcid, &pm)) != 0)
3566 * The allocated PMC has to be a process virtual PMC,
3567 * i.e., of type MODE_T[CS]. Global PMCs can only be
3568 * read using the PMCREAD operation since they may be
3569 * allocated on a different CPU than the one we could
3570 * be running on at the time of the RDPMC instruction.
3572 * The GETMSR operation is not allowed for PMCs that
3573 * are inherited across processes.
3576 if (!PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)) ||
3577 (pm->pm_flags & PMC_F_DESCENDANTS)) {
3583 * It only makes sense to use a RDPMC (or its
3584 * equivalent instruction on non-x86 architectures) on
3585 * a process that has allocated and attached a PMC to
3586 * itself. Conversely the PMC is only allowed to have
3587 * one process attached to it -- its owner.
3590 if ((pt = LIST_FIRST(&pm->pm_targets)) == NULL ||
3591 LIST_NEXT(pt, pt_next) != NULL ||
3592 pt->pt_process->pp_proc != pm->pm_owner->po_owner) {
3597 ri = PMC_TO_ROWINDEX(pm);
3598 pcd = pmc_ri_to_classdep(md, ri, &adjri);
3600 /* PMC class has no 'GETMSR' support */
3601 if (pcd->pcd_get_msr == NULL) {
3606 if ((error = (*pcd->pcd_get_msr)(adjri, &gm.pm_msr)) < 0)
3609 if ((error = copyout(&gm, arg, sizeof(gm))) < 0)
3613 * Mark our process as using MSRs. Update machine
3614 * state using a forced context switch.
3617 pt->pt_process->pp_flags |= PMC_PP_ENABLE_MSR_ACCESS;
3618 pmc_force_context_switch();
3624 * Release an allocated PMC
3627 case PMC_OP_PMCRELEASE:
3631 struct pmc_owner *po;
3632 struct pmc_op_simple sp;
3635 * Find PMC pointer for the named PMC.
3637 * Use pmc_release_pmc_descriptor() to switch off the
3638 * PMC, remove all its target threads, and remove the
3639 * PMC from its owner's list.
3641 * Remove the owner record if this is the last PMC
3647 if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
3650 pmcid = sp.pm_pmcid;
3652 if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
3656 pmc_release_pmc_descriptor(pm);
3657 pmc_maybe_remove_owner(po);
3665 * Read and/or write a PMC.
3673 pmc_value_t oldvalue;
3674 struct pmc_binding pb;
3675 struct pmc_op_pmcrw prw;
3676 struct pmc_classdep *pcd;
3677 struct pmc_op_pmcrw *pprw;
3681 if ((error = copyin(arg, &prw, sizeof(prw))) != 0)
3685 PMCDBG(PMC,OPS,1, "rw id=%d flags=0x%x", prw.pm_pmcid,
3688 /* must have at least one flag set */
3689 if ((prw.pm_flags & (PMC_F_OLDVALUE|PMC_F_NEWVALUE)) == 0) {
3694 /* locate pmc descriptor */
3695 if ((error = pmc_find_pmc(prw.pm_pmcid, &pm)) != 0)
3698 /* Can't read a PMC that hasn't been started. */
3699 if (pm->pm_state != PMC_STATE_ALLOCATED &&
3700 pm->pm_state != PMC_STATE_STOPPED &&
3701 pm->pm_state != PMC_STATE_RUNNING) {
3706 /* writing a new value is allowed only for 'STOPPED' pmcs */
3707 if (pm->pm_state == PMC_STATE_RUNNING &&
3708 (prw.pm_flags & PMC_F_NEWVALUE)) {
3713 if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) {
3716 * If this PMC is attached to its owner (i.e.,
3717 * the process requesting this operation) and
3718 * is running, then attempt to get an
3719 * upto-date reading from hardware for a READ.
3720 * Writes are only allowed when the PMC is
3721 * stopped, so only update the saved value
3724 * If the PMC is not running, or is not
3725 * attached to its owner, read/write to the
3729 ri = PMC_TO_ROWINDEX(pm);
3730 pcd = pmc_ri_to_classdep(md, ri, &adjri);
3732 mtx_pool_lock_spin(pmc_mtxpool, pm);
3733 cpu = curthread->td_oncpu;
3735 if (prw.pm_flags & PMC_F_OLDVALUE) {
3736 if ((pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) &&
3737 (pm->pm_state == PMC_STATE_RUNNING))
3738 error = (*pcd->pcd_read_pmc)(cpu, adjri,
3741 oldvalue = pm->pm_gv.pm_savedvalue;
3743 if (prw.pm_flags & PMC_F_NEWVALUE)
3744 pm->pm_gv.pm_savedvalue = prw.pm_value;
3746 mtx_pool_unlock_spin(pmc_mtxpool, pm);
3748 } else { /* System mode PMCs */
3749 cpu = PMC_TO_CPU(pm);
3750 ri = PMC_TO_ROWINDEX(pm);
3751 pcd = pmc_ri_to_classdep(md, ri, &adjri);
3753 if (!pmc_cpu_is_active(cpu)) {
3758 /* move this thread to CPU 'cpu' */
3759 pmc_save_cpu_binding(&pb);
3760 pmc_select_cpu(cpu);
3763 /* save old value */
3764 if (prw.pm_flags & PMC_F_OLDVALUE)
3765 if ((error = (*pcd->pcd_read_pmc)(cpu, adjri,
3768 /* write out new value */
3769 if (prw.pm_flags & PMC_F_NEWVALUE)
3770 error = (*pcd->pcd_write_pmc)(cpu, adjri,
3774 pmc_restore_cpu_binding(&pb);
3779 pprw = (struct pmc_op_pmcrw *) arg;
3782 if (prw.pm_flags & PMC_F_NEWVALUE)
3783 PMCDBG(PMC,OPS,2, "rw id=%d new %jx -> old %jx",
3784 ri, prw.pm_value, oldvalue);
3785 else if (prw.pm_flags & PMC_F_OLDVALUE)
3786 PMCDBG(PMC,OPS,2, "rw id=%d -> old %jx", ri, oldvalue);
3789 /* return old value if requested */
3790 if (prw.pm_flags & PMC_F_OLDVALUE)
3791 if ((error = copyout(&oldvalue, &pprw->pm_value,
3792 sizeof(prw.pm_value))))
3800 * Set the sampling rate for a sampling mode PMC and the
3801 * initial count for a counting mode PMC.
3804 case PMC_OP_PMCSETCOUNT:
3807 struct pmc_op_pmcsetcount sc;
3811 if ((error = copyin(arg, &sc, sizeof(sc))) != 0)
3814 if ((error = pmc_find_pmc(sc.pm_pmcid, &pm)) != 0)
3817 if (pm->pm_state == PMC_STATE_RUNNING) {
3822 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
3823 pm->pm_sc.pm_reloadcount = sc.pm_count;
3825 pm->pm_sc.pm_initial = sc.pm_count;
3834 case PMC_OP_PMCSTART:
3838 struct pmc_op_simple sp;
3840 sx_assert(&pmc_sx, SX_XLOCKED);
3842 if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
3845 pmcid = sp.pm_pmcid;
3847 if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
3850 KASSERT(pmcid == pm->pm_id,
3851 ("[pmc,%d] pmcid %x != id %x", __LINE__,
3854 if (pm->pm_state == PMC_STATE_RUNNING) /* already running */
3856 else if (pm->pm_state != PMC_STATE_STOPPED &&
3857 pm->pm_state != PMC_STATE_ALLOCATED) {
3862 error = pmc_start(pm);
3871 case PMC_OP_PMCSTOP:
3875 struct pmc_op_simple sp;
3879 if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
3882 pmcid = sp.pm_pmcid;
3885 * Mark the PMC as inactive and invoke the MD stop
3886 * routines if needed.
3889 if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
3892 KASSERT(pmcid == pm->pm_id,
3893 ("[pmc,%d] pmc id %x != pmcid %x", __LINE__,
3896 if (pm->pm_state == PMC_STATE_STOPPED) /* already stopped */
3898 else if (pm->pm_state != PMC_STATE_RUNNING) {
3903 error = pmc_stop(pm);
3909 * Write a user supplied value to the log file.
3912 case PMC_OP_WRITELOG:
3914 struct pmc_op_writelog wl;
3915 struct pmc_owner *po;
3919 if ((error = copyin(arg, &wl, sizeof(wl))) != 0)
3922 if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) {
3927 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) {
3932 error = pmclog_process_userlog(po, &wl);
3942 if (is_sx_locked != 0) {
3943 if (is_sx_downgraded)
3944 sx_sunlock(&pmc_sx);
3946 sx_xunlock(&pmc_sx);
3950 atomic_add_int(&pmc_stats.pm_syscall_errors, 1);
3963 * Mark the thread as needing callchain capture and post an AST. The
3964 * actual callchain capture will be done in a context where it is safe
3965 * to take page faults.
3969 pmc_post_callchain_callback(void)
3976 * If there is multiple PMCs for the same interrupt ignore new post
3978 if (td->td_pflags & TDP_CALLCHAIN)
3982 * Mark this thread as needing callchain capture.
3983 * `td->td_pflags' will be safe to touch because this thread
3984 * was in user space when it was interrupted.
3986 td->td_pflags |= TDP_CALLCHAIN;
3989 * Don't let this thread migrate between CPUs until callchain
3990 * capture completes.
3998 * Interrupt processing.
4000 * Find a free slot in the per-cpu array of samples and capture the
4001 * current callchain there. If a sample was successfully added, a bit
4002 * is set in mask 'pmc_cpumask' denoting that the DO_SAMPLES hook
4003 * needs to be invoked from the clock handler.
4005 * This function is meant to be called from an NMI handler. It cannot
4006 * use any of the locking primitives supplied by the OS.
4010 pmc_process_interrupt(int cpu, struct pmc *pm, struct trapframe *tf,
4013 int error, callchaindepth;
4015 struct pmc_sample *ps;
4016 struct pmc_samplebuffer *psb;
4021 * Allocate space for a sample buffer.
4023 psb = pmc_pcpu[cpu]->pc_sb;
4026 if (ps->ps_nsamples) { /* in use, reader hasn't caught up */
4028 atomic_add_int(&pmc_stats.pm_intr_bufferfull, 1);
4029 PMCDBG(SAM,INT,1,"(spc) cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d",
4030 cpu, pm, (void *) tf, inuserspace,
4031 (int) (psb->ps_write - psb->ps_samples),
4032 (int) (psb->ps_read - psb->ps_samples));
4038 /* Fill in entry. */
4039 PMCDBG(SAM,INT,1,"cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d", cpu, pm,
4040 (void *) tf, inuserspace,
4041 (int) (psb->ps_write - psb->ps_samples),
4042 (int) (psb->ps_read - psb->ps_samples));
4044 KASSERT(pm->pm_runcount >= 0,
4045 ("[pmc,%d] pm=%p runcount %d", __LINE__, (void *) pm,
4048 atomic_add_rel_32(&pm->pm_runcount, 1); /* hold onto PMC */
4050 if ((td = curthread) && td->td_proc)
4051 ps->ps_pid = td->td_proc->p_pid;
4056 ps->ps_flags = inuserspace ? PMC_CC_F_USERSPACE : 0;
4058 callchaindepth = (pm->pm_flags & PMC_F_CALLCHAIN) ?
4059 pmc_callchaindepth : 1;
4061 if (callchaindepth == 1)
4062 ps->ps_pc[0] = PMC_TRAPFRAME_TO_PC(tf);
4065 * Kernel stack traversals can be done immediately,
4066 * while we defer to an AST for user space traversals.
4070 pmc_save_kernel_callchain(ps->ps_pc,
4071 callchaindepth, tf);
4073 pmc_post_callchain_callback();
4074 callchaindepth = PMC_SAMPLE_INUSE;
4078 ps->ps_nsamples = callchaindepth; /* mark entry as in use */
4080 /* increment write pointer, modulo ring buffer size */
4082 if (ps == psb->ps_fence)
4083 psb->ps_write = psb->ps_samples;
4088 /* mark CPU as needing processing */
4089 atomic_set_rel_int(&pmc_cpumask, (1 << cpu));
4095 * Capture a user call chain. This function will be called from ast()
4096 * before control returns to userland and before the process gets
4101 pmc_capture_user_callchain(int cpu, struct trapframe *tf)
4106 struct pmc_sample *ps;
4107 struct pmc_samplebuffer *psb;
4112 sched_unpin(); /* Can migrate safely now. */
4114 psb = pmc_pcpu[cpu]->pc_sb;
4117 KASSERT(td->td_pflags & TDP_CALLCHAIN,
4118 ("[pmc,%d] Retrieving callchain for thread that doesn't want it",
4126 * Iterate through all deferred callchain requests.
4129 ps = psb->ps_samples;
4130 for (i = 0; i < pmc_nsamples; i++, ps++) {
4132 if (ps->ps_nsamples != PMC_SAMPLE_INUSE)
4134 if (ps->ps_td != td)
4137 KASSERT(ps->ps_cpu == cpu,
4138 ("[pmc,%d] cpu mismatch ps_cpu=%d pcpu=%d", __LINE__,
4139 ps->ps_cpu, PCPU_GET(cpuid)));
4143 KASSERT(pm->pm_flags & PMC_F_CALLCHAIN,
4144 ("[pmc,%d] Retrieving callchain for PMC that doesn't "
4145 "want it", __LINE__));
4147 KASSERT(pm->pm_runcount > 0,
4148 ("[pmc,%d] runcount %d", __LINE__, pm->pm_runcount));
4151 * Retrieve the callchain and mark the sample buffer
4152 * as 'processable' by the timer tick sweep code.
4154 ps->ps_nsamples = pmc_save_user_callchain(ps->ps_pc,
4155 pmc_callchaindepth, tf);
4163 KASSERT(ncallchains > 0,
4164 ("[pmc,%d] cpu %d didn't find a sample to collect", __LINE__,
4172 * Process saved PC samples.
4176 pmc_process_samples(int cpu)
4181 struct pmc_owner *po;
4182 struct pmc_sample *ps;
4183 struct pmc_classdep *pcd;
4184 struct pmc_samplebuffer *psb;
4186 KASSERT(PCPU_GET(cpuid) == cpu,
4187 ("[pmc,%d] not on the correct CPU pcpu=%d cpu=%d", __LINE__,
4188 PCPU_GET(cpuid), cpu));
4190 psb = pmc_pcpu[cpu]->pc_sb;
4192 for (n = 0; n < pmc_nsamples; n++) { /* bound on #iterations */
4195 if (ps->ps_nsamples == PMC_SAMPLE_FREE)
4197 if (ps->ps_nsamples == PMC_SAMPLE_INUSE) {
4198 /* Need a rescan at a later time. */
4199 atomic_set_rel_int(&pmc_cpumask, (1 << cpu));
4205 KASSERT(pm->pm_runcount > 0,
4206 ("[pmc,%d] pm=%p runcount %d", __LINE__, (void *) pm,
4211 KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)),
4212 ("[pmc,%d] pmc=%p non-sampling mode=%d", __LINE__,
4213 pm, PMC_TO_MODE(pm)));
4215 /* Ignore PMCs that have been switched off */
4216 if (pm->pm_state != PMC_STATE_RUNNING)
4219 PMCDBG(SAM,OPS,1,"cpu=%d pm=%p n=%d fl=%x wr=%d rd=%d", cpu,
4220 pm, ps->ps_nsamples, ps->ps_flags,
4221 (int) (psb->ps_write - psb->ps_samples),
4222 (int) (psb->ps_read - psb->ps_samples));
4225 * If this is a process-mode PMC that is attached to
4226 * its owner, and if the PC is in user mode, update
4227 * profiling statistics like timer-based profiling
4230 if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) {
4231 if (ps->ps_flags & PMC_CC_F_USERSPACE) {
4232 td = FIRST_THREAD_IN_PROC(po->po_owner);
4233 addupc_intr(td, ps->ps_pc[0], 1);
4239 * Otherwise, this is either a sampling mode PMC that
4240 * is attached to a different process than its owner,
4241 * or a system-wide sampling PMC. Dispatch a log
4242 * entry to the PMC's owner process.
4245 pmclog_process_callchain(pm, ps);
4248 ps->ps_nsamples = 0; /* mark entry as free */
4249 atomic_subtract_rel_32(&pm->pm_runcount, 1);
4251 /* increment read pointer, modulo sample size */
4252 if (++ps == psb->ps_fence)
4253 psb->ps_read = psb->ps_samples;
4258 atomic_add_int(&pmc_stats.pm_log_sweeps, 1);
4260 /* Do not re-enable stalled PMCs if we failed to process any samples */
4265 * Restart any stalled sampling PMCs on this CPU.
4267 * If the NMI handler sets the pm_stalled field of a PMC after
4268 * the check below, we'll end up processing the stalled PMC at
4269 * the next hardclock tick.
4271 for (n = 0; n < md->pmd_npmc; n++) {
4272 pcd = pmc_ri_to_classdep(md, n, &adjri);
4273 KASSERT(pcd != NULL,
4274 ("[pmc,%d] null pcd ri=%d", __LINE__, n));
4275 (void) (*pcd->pcd_get_config)(cpu,adjri,&pm);
4277 if (pm == NULL || /* !cfg'ed */
4278 pm->pm_state != PMC_STATE_RUNNING || /* !active */
4279 !PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) || /* !sampling */
4280 pm->pm_stalled == 0) /* !stalled */
4284 (*pcd->pcd_start_pmc)(cpu, adjri);
4293 * Handle a process exit.
4295 * Remove this process from all hash tables. If this process
4296 * owned any PMCs, turn off those PMCs and deallocate them,
4297 * removing any associations with target processes.
4299 * This function will be called by the last 'thread' of a
4302 * XXX This eventhandler gets called early in the exit process.
4303 * Consider using a 'hook' invocation from thread_exit() or equivalent
4304 * spot. Another negative is that kse_exit doesn't seem to call
4310 pmc_process_exit(void *arg __unused, struct proc *p)
4315 int is_using_hwpmcs;
4316 struct pmc_owner *po;
4317 struct pmc_process *pp;
4318 struct pmc_classdep *pcd;
4319 pmc_value_t newvalue, tmp;
4322 is_using_hwpmcs = p->p_flag & P_HWPMC;
4326 * Log a sysexit event to all SS PMC owners.
4328 LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
4329 if (po->po_flags & PMC_PO_OWNS_LOGFILE)
4330 pmclog_process_sysexit(po, p->p_pid);
4332 if (!is_using_hwpmcs)
4336 PMCDBG(PRC,EXT,1,"process-exit proc=%p (%d, %s)", p, p->p_pid,
4340 * Since this code is invoked by the last thread in an exiting
4341 * process, we would have context switched IN at some prior
4342 * point. However, with PREEMPTION, kernel mode context
4343 * switches may happen any time, so we want to disable a
4344 * context switch OUT till we get any PMCs targetting this
4345 * process off the hardware.
4347 * We also need to atomically remove this process'
4348 * entry from our target process hash table, using
4351 PMCDBG(PRC,EXT,1, "process-exit proc=%p (%d, %s)", p, p->p_pid,
4354 critical_enter(); /* no preemption */
4356 cpu = curthread->td_oncpu;
4358 if ((pp = pmc_find_process_descriptor(p,
4359 PMC_FLAG_REMOVE)) != NULL) {
4362 "process-exit proc=%p pmc-process=%p", p, pp);
4365 * The exiting process could the target of
4366 * some PMCs which will be running on
4367 * currently executing CPU.
4369 * We need to turn these PMCs off like we
4370 * would do at context switch OUT time.
4372 for (ri = 0; ri < md->pmd_npmc; ri++) {
4375 * Pick up the pmc pointer from hardware
4376 * state similar to the CSW_OUT code.
4380 pcd = pmc_ri_to_classdep(md, ri, &adjri);
4382 (void) (*pcd->pcd_get_config)(cpu, adjri, &pm);
4384 PMCDBG(PRC,EXT,2, "ri=%d pm=%p", ri, pm);
4387 !PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)))
4390 PMCDBG(PRC,EXT,2, "ppmcs[%d]=%p pm=%p "
4391 "state=%d", ri, pp->pp_pmcs[ri].pp_pmc,
4394 KASSERT(PMC_TO_ROWINDEX(pm) == ri,
4395 ("[pmc,%d] ri mismatch pmc(%d) ri(%d)",
4396 __LINE__, PMC_TO_ROWINDEX(pm), ri));
4398 KASSERT(pm == pp->pp_pmcs[ri].pp_pmc,
4399 ("[pmc,%d] pm %p != pp_pmcs[%d] %p",
4400 __LINE__, pm, ri, pp->pp_pmcs[ri].pp_pmc));
4402 (void) pcd->pcd_stop_pmc(cpu, adjri);
4404 KASSERT(pm->pm_runcount > 0,
4405 ("[pmc,%d] bad runcount ri %d rc %d",
4406 __LINE__, ri, pm->pm_runcount));
4408 /* Stop hardware only if it is actually running */
4409 if (pm->pm_state == PMC_STATE_RUNNING &&
4410 pm->pm_stalled == 0) {
4411 pcd->pcd_read_pmc(cpu, adjri, &newvalue);
4413 PMC_PCPU_SAVED(cpu,ri);
4415 mtx_pool_lock_spin(pmc_mtxpool, pm);
4416 pm->pm_gv.pm_savedvalue += tmp;
4417 pp->pp_pmcs[ri].pp_pmcval += tmp;
4418 mtx_pool_unlock_spin(pmc_mtxpool, pm);
4421 atomic_subtract_rel_32(&pm->pm_runcount,1);
4423 KASSERT((int) pm->pm_runcount >= 0,
4424 ("[pmc,%d] runcount is %d", __LINE__, ri));
4426 (void) pcd->pcd_config_pmc(cpu, adjri, NULL);
4430 * Inform the MD layer of this pseudo "context switch
4433 (void) md->pmd_switch_out(pmc_pcpu[cpu], pp);
4435 critical_exit(); /* ok to be pre-empted now */
4438 * Unlink this process from the PMCs that are
4439 * targetting it. This will send a signal to
4440 * all PMC owner's whose PMCs are orphaned.
4442 * Log PMC value at exit time if requested.
4444 for (ri = 0; ri < md->pmd_npmc; ri++)
4445 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) {
4446 if (pm->pm_flags & PMC_F_NEEDS_LOGFILE &&
4447 PMC_IS_COUNTING_MODE(PMC_TO_MODE(pm)))
4448 pmclog_process_procexit(pm, pp);
4449 pmc_unlink_target_process(pm, pp);
4454 critical_exit(); /* pp == NULL */
4458 * If the process owned PMCs, free them up and free up
4461 if ((po = pmc_find_owner_descriptor(p)) != NULL) {
4462 pmc_remove_owner(po);
4463 pmc_destroy_owner_descriptor(po);
4466 sx_xunlock(&pmc_sx);
4470 * Handle a process fork.
4472 * If the parent process 'p1' is under HWPMC monitoring, then copy
4473 * over any attached PMCs that have 'do_descendants' semantics.
4477 pmc_process_fork(void *arg __unused, struct proc *p1, struct proc *newproc,
4480 int is_using_hwpmcs;
4482 uint32_t do_descendants;
4484 struct pmc_owner *po;
4485 struct pmc_process *ppnew, *ppold;
4487 (void) flags; /* unused parameter */
4490 is_using_hwpmcs = p1->p_flag & P_HWPMC;
4494 * If there are system-wide sampling PMCs active, we need to
4495 * log all fork events to their owner's logs.
4498 LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
4499 if (po->po_flags & PMC_PO_OWNS_LOGFILE)
4500 pmclog_process_procfork(po, p1->p_pid, newproc->p_pid);
4502 if (!is_using_hwpmcs)
4506 PMCDBG(PMC,FRK,1, "process-fork proc=%p (%d, %s) -> %p", p1,
4507 p1->p_pid, p1->p_comm, newproc);
4510 * If the parent process (curthread->td_proc) is a
4511 * target of any PMCs, look for PMCs that are to be
4512 * inherited, and link these into the new process
4515 if ((ppold = pmc_find_process_descriptor(curthread->td_proc,
4516 PMC_FLAG_NONE)) == NULL)
4517 goto done; /* nothing to do */
4520 for (ri = 0; ri < md->pmd_npmc; ri++)
4521 if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL)
4522 do_descendants |= pm->pm_flags & PMC_F_DESCENDANTS;
4523 if (do_descendants == 0) /* nothing to do */
4526 /* allocate a descriptor for the new process */
4527 if ((ppnew = pmc_find_process_descriptor(newproc,
4528 PMC_FLAG_ALLOCATE)) == NULL)
4532 * Run through all PMCs that were targeting the old process
4533 * and which specified F_DESCENDANTS and attach them to the
4536 * Log the fork event to all owners of PMCs attached to this
4537 * process, if not already logged.
4539 for (ri = 0; ri < md->pmd_npmc; ri++)
4540 if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL &&
4541 (pm->pm_flags & PMC_F_DESCENDANTS)) {
4542 pmc_link_target_process(pm, ppnew);
4544 if (po->po_sscount == 0 &&
4545 po->po_flags & PMC_PO_OWNS_LOGFILE)
4546 pmclog_process_procfork(po, p1->p_pid,
4551 * Now mark the new process as being tracked by this driver.
4554 newproc->p_flag |= P_HWPMC;
4555 PROC_UNLOCK(newproc);
4558 sx_xunlock(&pmc_sx);
4566 static const char *pmc_name_of_pmcclass[] = {
4568 #define __PMC_CLASS(N) #N ,
4573 pmc_initialize(void)
4575 int c, cpu, error, n, ri;
4576 unsigned int maxcpu;
4577 struct pmc_binding pb;
4578 struct pmc_sample *ps;
4579 struct pmc_classdep *pcd;
4580 struct pmc_samplebuffer *sb;
4586 /* parse debug flags first */
4587 if (TUNABLE_STR_FETCH(PMC_SYSCTL_NAME_PREFIX "debugflags",
4588 pmc_debugstr, sizeof(pmc_debugstr)))
4589 pmc_debugflags_parse(pmc_debugstr,
4590 pmc_debugstr+strlen(pmc_debugstr));
4593 PMCDBG(MOD,INI,0, "PMC Initialize (version %x)", PMC_VERSION);
4595 /* check kernel version */
4596 if (pmc_kernel_version != PMC_VERSION) {
4597 if (pmc_kernel_version == 0)
4598 printf("hwpmc: this kernel has not been compiled with "
4599 "'options HWPMC_HOOKS'.\n");
4601 printf("hwpmc: kernel version (0x%x) does not match "
4602 "module version (0x%x).\n", pmc_kernel_version,
4604 return EPROGMISMATCH;
4608 * check sysctl parameters
4611 if (pmc_hashsize <= 0) {
4612 (void) printf("hwpmc: tunable \"hashsize\"=%d must be "
4613 "greater than zero.\n", pmc_hashsize);
4614 pmc_hashsize = PMC_HASH_SIZE;
4617 if (pmc_nsamples <= 0 || pmc_nsamples > 65535) {
4618 (void) printf("hwpmc: tunable \"nsamples\"=%d out of "
4619 "range.\n", pmc_nsamples);
4620 pmc_nsamples = PMC_NSAMPLES;
4623 if (pmc_callchaindepth <= 0 ||
4624 pmc_callchaindepth > PMC_CALLCHAIN_DEPTH_MAX) {
4625 (void) printf("hwpmc: tunable \"callchaindepth\"=%d out of "
4626 "range.\n", pmc_callchaindepth);
4627 pmc_callchaindepth = PMC_CALLCHAIN_DEPTH;
4630 md = pmc_md_initialize();
4635 KASSERT(md->pmd_nclass >= 1 && md->pmd_npmc >= 1,
4636 ("[pmc,%d] no classes or pmcs", __LINE__));
4638 /* Compute the map from row-indices to classdep pointers. */
4639 pmc_rowindex_to_classdep = malloc(sizeof(struct pmc_classdep *) *
4640 md->pmd_npmc, M_PMC, M_WAITOK|M_ZERO);
4642 for (n = 0; n < md->pmd_npmc; n++)
4643 pmc_rowindex_to_classdep[n] = NULL;
4644 for (ri = c = 0; c < md->pmd_nclass; c++) {
4645 pcd = &md->pmd_classdep[c];
4646 for (n = 0; n < pcd->pcd_num; n++, ri++)
4647 pmc_rowindex_to_classdep[ri] = pcd;
4650 KASSERT(ri == md->pmd_npmc,
4651 ("[pmc,%d] npmc miscomputed: ri=%d, md->npmc=%d", __LINE__,
4654 maxcpu = pmc_cpu_max();
4656 /* allocate space for the per-cpu array */
4657 pmc_pcpu = malloc(maxcpu * sizeof(struct pmc_cpu *), M_PMC,
4660 /* per-cpu 'saved values' for managing process-mode PMCs */
4661 pmc_pcpu_saved = malloc(sizeof(pmc_value_t) * maxcpu * md->pmd_npmc,
4664 /* Perform CPU-dependent initialization. */
4665 pmc_save_cpu_binding(&pb);
4667 for (cpu = 0; error == 0 && cpu < maxcpu; cpu++) {
4668 if (!pmc_cpu_is_active(cpu))
4670 pmc_select_cpu(cpu);
4671 pmc_pcpu[cpu] = malloc(sizeof(struct pmc_cpu) +
4672 md->pmd_npmc * sizeof(struct pmc_hw *), M_PMC,
4674 if (md->pmd_pcpu_init)
4675 error = md->pmd_pcpu_init(md, cpu);
4676 for (n = 0; error == 0 && n < md->pmd_nclass; n++)
4677 error = md->pmd_classdep[n].pcd_pcpu_init(md, cpu);
4679 pmc_restore_cpu_binding(&pb);
4684 /* allocate space for the sample array */
4685 for (cpu = 0; cpu < maxcpu; cpu++) {
4686 if (!pmc_cpu_is_active(cpu))
4689 sb = malloc(sizeof(struct pmc_samplebuffer) +
4690 pmc_nsamples * sizeof(struct pmc_sample), M_PMC,
4692 sb->ps_read = sb->ps_write = sb->ps_samples;
4693 sb->ps_fence = sb->ps_samples + pmc_nsamples;
4695 KASSERT(pmc_pcpu[cpu] != NULL,
4696 ("[pmc,%d] cpu=%d Null per-cpu data", __LINE__, cpu));
4698 sb->ps_callchains = malloc(pmc_callchaindepth * pmc_nsamples *
4699 sizeof(uintptr_t), M_PMC, M_WAITOK|M_ZERO);
4701 for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++)
4702 ps->ps_pc = sb->ps_callchains +
4703 (n * pmc_callchaindepth);
4705 pmc_pcpu[cpu]->pc_sb = sb;
4708 /* allocate space for the row disposition array */
4709 pmc_pmcdisp = malloc(sizeof(enum pmc_mode) * md->pmd_npmc,
4710 M_PMC, M_WAITOK|M_ZERO);
4712 KASSERT(pmc_pmcdisp != NULL,
4713 ("[pmc,%d] pmcdisp allocation returned NULL", __LINE__));
4715 /* mark all PMCs as available */
4716 for (n = 0; n < (int) md->pmd_npmc; n++)
4717 PMC_MARK_ROW_FREE(n);
4719 /* allocate thread hash tables */
4720 pmc_ownerhash = hashinit(pmc_hashsize, M_PMC,
4721 &pmc_ownerhashmask);
4723 pmc_processhash = hashinit(pmc_hashsize, M_PMC,
4724 &pmc_processhashmask);
4725 mtx_init(&pmc_processhash_mtx, "pmc-process-hash", "pmc-leaf",
4728 LIST_INIT(&pmc_ss_owners);
4731 /* allocate a pool of spin mutexes */
4732 pmc_mtxpool = mtx_pool_create("pmc-leaf", pmc_mtxpool_size,
4735 PMCDBG(MOD,INI,1, "pmc_ownerhash=%p, mask=0x%lx "
4736 "targethash=%p mask=0x%lx", pmc_ownerhash, pmc_ownerhashmask,
4737 pmc_processhash, pmc_processhashmask);
4739 /* register process {exit,fork,exec} handlers */
4740 pmc_exit_tag = EVENTHANDLER_REGISTER(process_exit,
4741 pmc_process_exit, NULL, EVENTHANDLER_PRI_ANY);
4742 pmc_fork_tag = EVENTHANDLER_REGISTER(process_fork,
4743 pmc_process_fork, NULL, EVENTHANDLER_PRI_ANY);
4745 /* initialize logging */
4746 pmclog_initialize();
4748 /* set hook functions */
4749 pmc_intr = md->pmd_intr;
4750 pmc_hook = pmc_hook_handler;
4753 printf(PMC_MODULE_NAME ":");
4754 for (n = 0; n < (int) md->pmd_nclass; n++) {
4755 pcd = &md->pmd_classdep[n];
4756 printf(" %s/%d/%d/0x%b",
4757 pmc_name_of_pmcclass[pcd->pcd_class],
4762 "\1INT\2USR\3SYS\4EDG\5THR"
4763 "\6REA\7WRI\10INV\11QUA\12PRC"
4772 /* prepare to be unloaded */
4777 unsigned int maxcpu;
4778 struct pmc_ownerhash *ph;
4779 struct pmc_owner *po, *tmp;
4780 struct pmc_binding pb;
4782 struct pmc_processhash *prh;
4785 PMCDBG(MOD,INI,0, "%s", "cleanup");
4787 /* switch off sampling */
4788 atomic_store_rel_int(&pmc_cpumask, 0);
4792 if (pmc_hook == NULL) { /* being unloaded already */
4793 sx_xunlock(&pmc_sx);
4797 pmc_hook = NULL; /* prevent new threads from entering module */
4799 /* deregister event handlers */
4800 EVENTHANDLER_DEREGISTER(process_fork, pmc_fork_tag);
4801 EVENTHANDLER_DEREGISTER(process_exit, pmc_exit_tag);
4803 /* send SIGBUS to all owner threads, free up allocations */
4805 for (ph = pmc_ownerhash;
4806 ph <= &pmc_ownerhash[pmc_ownerhashmask];
4808 LIST_FOREACH_SAFE(po, ph, po_next, tmp) {
4809 pmc_remove_owner(po);
4811 /* send SIGBUS to owner processes */
4812 PMCDBG(MOD,INI,2, "cleanup signal proc=%p "
4813 "(%d, %s)", po->po_owner,
4814 po->po_owner->p_pid,
4815 po->po_owner->p_comm);
4817 PROC_LOCK(po->po_owner);
4818 psignal(po->po_owner, SIGBUS);
4819 PROC_UNLOCK(po->po_owner);
4821 pmc_destroy_owner_descriptor(po);
4825 /* reclaim allocated data structures */
4827 mtx_pool_destroy(&pmc_mtxpool);
4829 mtx_destroy(&pmc_processhash_mtx);
4830 if (pmc_processhash) {
4832 struct pmc_process *pp;
4834 PMCDBG(MOD,INI,3, "%s", "destroy process hash");
4835 for (prh = pmc_processhash;
4836 prh <= &pmc_processhash[pmc_processhashmask];
4838 LIST_FOREACH(pp, prh, pp_next)
4839 PMCDBG(MOD,INI,3, "pid=%d", pp->pp_proc->p_pid);
4842 hashdestroy(pmc_processhash, M_PMC, pmc_processhashmask);
4843 pmc_processhash = NULL;
4846 if (pmc_ownerhash) {
4847 PMCDBG(MOD,INI,3, "%s", "destroy owner hash");
4848 hashdestroy(pmc_ownerhash, M_PMC, pmc_ownerhashmask);
4849 pmc_ownerhash = NULL;
4852 KASSERT(LIST_EMPTY(&pmc_ss_owners),
4853 ("[pmc,%d] Global SS owner list not empty", __LINE__));
4854 KASSERT(pmc_ss_count == 0,
4855 ("[pmc,%d] Global SS count not empty", __LINE__));
4857 /* do processor and pmc-class dependent cleanup */
4858 maxcpu = pmc_cpu_max();
4860 PMCDBG(MOD,INI,3, "%s", "md cleanup");
4862 pmc_save_cpu_binding(&pb);
4863 for (cpu = 0; cpu < maxcpu; cpu++) {
4864 PMCDBG(MOD,INI,1,"pmc-cleanup cpu=%d pcs=%p",
4865 cpu, pmc_pcpu[cpu]);
4866 if (!pmc_cpu_is_active(cpu) || pmc_pcpu[cpu] == NULL)
4868 pmc_select_cpu(cpu);
4869 for (c = 0; c < md->pmd_nclass; c++)
4870 md->pmd_classdep[c].pcd_pcpu_fini(md, cpu);
4871 if (md->pmd_pcpu_fini)
4872 md->pmd_pcpu_fini(md, cpu);
4875 pmc_md_finalize(md);
4879 pmc_restore_cpu_binding(&pb);
4882 /* Free per-cpu descriptors. */
4883 for (cpu = 0; cpu < maxcpu; cpu++) {
4884 if (!pmc_cpu_is_active(cpu))
4886 KASSERT(pmc_pcpu[cpu]->pc_sb != NULL,
4887 ("[pmc,%d] Null cpu sample buffer cpu=%d", __LINE__,
4889 free(pmc_pcpu[cpu]->pc_sb->ps_callchains, M_PMC);
4890 free(pmc_pcpu[cpu]->pc_sb, M_PMC);
4891 free(pmc_pcpu[cpu], M_PMC);
4894 free(pmc_pcpu, M_PMC);
4897 free(pmc_pcpu_saved, M_PMC);
4898 pmc_pcpu_saved = NULL;
4901 free(pmc_pmcdisp, M_PMC);
4905 if (pmc_rowindex_to_classdep) {
4906 free(pmc_rowindex_to_classdep, M_PMC);
4907 pmc_rowindex_to_classdep = NULL;
4912 sx_xunlock(&pmc_sx); /* we are done */
4916 * The function called at load/unload.
4920 load (struct module *module __unused, int cmd, void *arg __unused)
4928 /* initialize the subsystem */
4929 error = pmc_initialize();
4932 PMCDBG(MOD,INI,1, "syscall=%d maxcpu=%d",
4933 pmc_syscall_num, pmc_cpu_max());
4940 PMCDBG(MOD,INI,1, "%s", "unloaded");
4944 error = EINVAL; /* XXX should panic(9) */
4952 MALLOC_DEFINE(M_PMC, "pmc", "Memory space for the PMC module");