xref: /freebsd/sys/dev/hwpmc/hwpmc_mod.c (revision 9336e0699bda8a301cd2bfa37106b6ec5e32012e)
1 /*-
2  * Copyright (c) 2003-2007 Joseph Koshy
3  * Copyright (c) 2007 The FreeBSD Foundation
4  * All rights reserved.
5  *
6  * Portions of this software were developed by A. Joseph Koshy under
7  * sponsorship from the FreeBSD Foundation and Google, Inc.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
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.
17  *
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
28  * SUCH DAMAGE.
29  *
30  */
31 
32 #include <sys/cdefs.h>
33 __FBSDID("$FreeBSD$");
34 
35 #include <sys/param.h>
36 #include <sys/eventhandler.h>
37 #include <sys/jail.h>
38 #include <sys/kernel.h>
39 #include <sys/kthread.h>
40 #include <sys/limits.h>
41 #include <sys/lock.h>
42 #include <sys/malloc.h>
43 #include <sys/module.h>
44 #include <sys/mutex.h>
45 #include <sys/pmc.h>
46 #include <sys/pmckern.h>
47 #include <sys/pmclog.h>
48 #include <sys/priv.h>
49 #include <sys/proc.h>
50 #include <sys/queue.h>
51 #include <sys/resourcevar.h>
52 #include <sys/sched.h>
53 #include <sys/signalvar.h>
54 #include <sys/smp.h>
55 #include <sys/sx.h>
56 #include <sys/sysctl.h>
57 #include <sys/sysent.h>
58 #include <sys/systm.h>
59 #include <sys/vnode.h>
60 
61 #include <sys/linker.h>		/* needs to be after <sys/malloc.h> */
62 
63 #include <machine/atomic.h>
64 #include <machine/md_var.h>
65 
66 /*
67  * Types
68  */
69 
70 enum pmc_flags {
71 	PMC_FLAG_NONE	  = 0x00, /* do nothing */
72 	PMC_FLAG_REMOVE   = 0x01, /* atomically remove entry from hash */
73 	PMC_FLAG_ALLOCATE = 0x02, /* add entry to hash if not found */
74 };
75 
76 /*
77  * The offset in sysent where the syscall is allocated.
78  */
79 
80 static int pmc_syscall_num = NO_SYSCALL;
81 struct pmc_cpu		**pmc_pcpu;	 /* per-cpu state */
82 pmc_value_t		*pmc_pcpu_saved; /* saved PMC values: CSW handling */
83 
84 #define	PMC_PCPU_SAVED(C,R)	pmc_pcpu_saved[(R) + md->pmd_npmc*(C)]
85 
86 struct mtx_pool		*pmc_mtxpool;
87 static int		*pmc_pmcdisp;	 /* PMC row dispositions */
88 
89 #define	PMC_ROW_DISP_IS_FREE(R)		(pmc_pmcdisp[(R)] == 0)
90 #define	PMC_ROW_DISP_IS_THREAD(R)	(pmc_pmcdisp[(R)] > 0)
91 #define	PMC_ROW_DISP_IS_STANDALONE(R)	(pmc_pmcdisp[(R)] < 0)
92 
93 #define	PMC_MARK_ROW_FREE(R) do {					  \
94 	pmc_pmcdisp[(R)] = 0;						  \
95 } while (0)
96 
97 #define	PMC_MARK_ROW_STANDALONE(R) do {					  \
98 	KASSERT(pmc_pmcdisp[(R)] <= 0, ("[pmc,%d] row disposition error", \
99 		    __LINE__));						  \
100 	atomic_add_int(&pmc_pmcdisp[(R)], -1);				  \
101 	KASSERT(pmc_pmcdisp[(R)] >= (-mp_ncpus), ("[pmc,%d] row "	  \
102 		"disposition error", __LINE__));			  \
103 } while (0)
104 
105 #define	PMC_UNMARK_ROW_STANDALONE(R) do { 				  \
106 	atomic_add_int(&pmc_pmcdisp[(R)], 1);				  \
107 	KASSERT(pmc_pmcdisp[(R)] <= 0, ("[pmc,%d] row disposition error", \
108 		    __LINE__));						  \
109 } while (0)
110 
111 #define	PMC_MARK_ROW_THREAD(R) do {					  \
112 	KASSERT(pmc_pmcdisp[(R)] >= 0, ("[pmc,%d] row disposition error", \
113 		    __LINE__));						  \
114 	atomic_add_int(&pmc_pmcdisp[(R)], 1);				  \
115 } while (0)
116 
117 #define	PMC_UNMARK_ROW_THREAD(R) do {					  \
118 	atomic_add_int(&pmc_pmcdisp[(R)], -1);				  \
119 	KASSERT(pmc_pmcdisp[(R)] >= 0, ("[pmc,%d] row disposition error", \
120 		    __LINE__));						  \
121 } while (0)
122 
123 
124 /* various event handlers */
125 static eventhandler_tag	pmc_exit_tag, pmc_fork_tag;
126 
127 /* Module statistics */
128 struct pmc_op_getdriverstats pmc_stats;
129 
130 /* Machine/processor dependent operations */
131 struct pmc_mdep  *md;
132 
133 /*
134  * Hash tables mapping owner processes and target threads to PMCs.
135  */
136 
137 struct mtx pmc_processhash_mtx;		/* spin mutex */
138 static u_long pmc_processhashmask;
139 static LIST_HEAD(pmc_processhash, pmc_process)	*pmc_processhash;
140 
141 /*
142  * Hash table of PMC owner descriptors.  This table is protected by
143  * the shared PMC "sx" lock.
144  */
145 
146 static u_long pmc_ownerhashmask;
147 static LIST_HEAD(pmc_ownerhash, pmc_owner)	*pmc_ownerhash;
148 
149 /*
150  * List of PMC owners with system-wide sampling PMCs.
151  */
152 
153 static LIST_HEAD(, pmc_owner)			pmc_ss_owners;
154 
155 
156 /*
157  * Prototypes
158  */
159 
160 #ifdef	DEBUG
161 static int	pmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS);
162 static int	pmc_debugflags_parse(char *newstr, char *fence);
163 #endif
164 
165 static int	load(struct module *module, int cmd, void *arg);
166 static int	pmc_attach_process(struct proc *p, struct pmc *pm);
167 static struct pmc *pmc_allocate_pmc_descriptor(void);
168 static struct pmc_owner *pmc_allocate_owner_descriptor(struct proc *p);
169 static int	pmc_attach_one_process(struct proc *p, struct pmc *pm);
170 static int	pmc_can_allocate_rowindex(struct proc *p, unsigned int ri,
171     int cpu);
172 static int	pmc_can_attach(struct pmc *pm, struct proc *p);
173 static void	pmc_capture_user_callchain(int cpu, struct trapframe *tf);
174 static void	pmc_cleanup(void);
175 static int	pmc_detach_process(struct proc *p, struct pmc *pm);
176 static int	pmc_detach_one_process(struct proc *p, struct pmc *pm,
177     int flags);
178 static void	pmc_destroy_owner_descriptor(struct pmc_owner *po);
179 static struct pmc_owner *pmc_find_owner_descriptor(struct proc *p);
180 static int	pmc_find_pmc(pmc_id_t pmcid, struct pmc **pm);
181 static struct pmc *pmc_find_pmc_descriptor_in_process(struct pmc_owner *po,
182     pmc_id_t pmc);
183 static struct pmc_process *pmc_find_process_descriptor(struct proc *p,
184     uint32_t mode);
185 static void	pmc_force_context_switch(void);
186 static void	pmc_link_target_process(struct pmc *pm,
187     struct pmc_process *pp);
188 static void	pmc_log_all_process_mappings(struct pmc_owner *po);
189 static void	pmc_log_kernel_mappings(struct pmc *pm);
190 static void	pmc_log_process_mappings(struct pmc_owner *po, struct proc *p);
191 static void	pmc_maybe_remove_owner(struct pmc_owner *po);
192 static void	pmc_process_csw_in(struct thread *td);
193 static void	pmc_process_csw_out(struct thread *td);
194 static void	pmc_process_exit(void *arg, struct proc *p);
195 static void	pmc_process_fork(void *arg, struct proc *p1,
196     struct proc *p2, int n);
197 static void	pmc_process_samples(int cpu);
198 static void	pmc_release_pmc_descriptor(struct pmc *pmc);
199 static void	pmc_remove_owner(struct pmc_owner *po);
200 static void	pmc_remove_process_descriptor(struct pmc_process *pp);
201 static void	pmc_restore_cpu_binding(struct pmc_binding *pb);
202 static void	pmc_save_cpu_binding(struct pmc_binding *pb);
203 static void	pmc_select_cpu(int cpu);
204 static int	pmc_start(struct pmc *pm);
205 static int	pmc_stop(struct pmc *pm);
206 static int	pmc_syscall_handler(struct thread *td, void *syscall_args);
207 static void	pmc_unlink_target_process(struct pmc *pmc,
208     struct pmc_process *pp);
209 
210 /*
211  * Kernel tunables and sysctl(8) interface.
212  */
213 
214 SYSCTL_NODE(_kern, OID_AUTO, hwpmc, CTLFLAG_RW, 0, "HWPMC parameters");
215 
216 static int pmc_callchaindepth = PMC_CALLCHAIN_DEPTH;
217 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "callchaindepth", &pmc_callchaindepth);
218 SYSCTL_INT(_kern_hwpmc, OID_AUTO, callchaindepth, CTLFLAG_TUN|CTLFLAG_RD,
219     &pmc_callchaindepth, 0, "depth of call chain records");
220 
221 #ifdef	DEBUG
222 struct pmc_debugflags pmc_debugflags = PMC_DEBUG_DEFAULT_FLAGS;
223 char	pmc_debugstr[PMC_DEBUG_STRSIZE];
224 TUNABLE_STR(PMC_SYSCTL_NAME_PREFIX "debugflags", pmc_debugstr,
225     sizeof(pmc_debugstr));
226 SYSCTL_PROC(_kern_hwpmc, OID_AUTO, debugflags,
227     CTLTYPE_STRING|CTLFLAG_RW|CTLFLAG_TUN,
228     0, 0, pmc_debugflags_sysctl_handler, "A", "debug flags");
229 #endif
230 
231 /*
232  * kern.hwpmc.hashrows -- determines the number of rows in the
233  * of the hash table used to look up threads
234  */
235 
236 static int pmc_hashsize = PMC_HASH_SIZE;
237 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "hashsize", &pmc_hashsize);
238 SYSCTL_INT(_kern_hwpmc, OID_AUTO, hashsize, CTLFLAG_TUN|CTLFLAG_RD,
239     &pmc_hashsize, 0, "rows in hash tables");
240 
241 /*
242  * kern.hwpmc.nsamples --- number of PC samples/callchain stacks per CPU
243  */
244 
245 static int pmc_nsamples = PMC_NSAMPLES;
246 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "nsamples", &pmc_nsamples);
247 SYSCTL_INT(_kern_hwpmc, OID_AUTO, nsamples, CTLFLAG_TUN|CTLFLAG_RD,
248     &pmc_nsamples, 0, "number of PC samples per CPU");
249 
250 
251 /*
252  * kern.hwpmc.mtxpoolsize -- number of mutexes in the mutex pool.
253  */
254 
255 static int pmc_mtxpool_size = PMC_MTXPOOL_SIZE;
256 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "mtxpoolsize", &pmc_mtxpool_size);
257 SYSCTL_INT(_kern_hwpmc, OID_AUTO, mtxpoolsize, CTLFLAG_TUN|CTLFLAG_RD,
258     &pmc_mtxpool_size, 0, "size of spin mutex pool");
259 
260 
261 /*
262  * security.bsd.unprivileged_syspmcs -- allow non-root processes to
263  * allocate system-wide PMCs.
264  *
265  * Allowing unprivileged processes to allocate system PMCs is convenient
266  * if system-wide measurements need to be taken concurrently with other
267  * per-process measurements.  This feature is turned off by default.
268  */
269 
270 static int pmc_unprivileged_syspmcs = 0;
271 TUNABLE_INT("security.bsd.unprivileged_syspmcs", &pmc_unprivileged_syspmcs);
272 SYSCTL_INT(_security_bsd, OID_AUTO, unprivileged_syspmcs, CTLFLAG_RW,
273     &pmc_unprivileged_syspmcs, 0,
274     "allow unprivileged process to allocate system PMCs");
275 
276 /*
277  * Hash function.  Discard the lower 2 bits of the pointer since
278  * these are always zero for our uses.  The hash multiplier is
279  * round((2^LONG_BIT) * ((sqrt(5)-1)/2)).
280  */
281 
282 #if	LONG_BIT == 64
283 #define	_PMC_HM		11400714819323198486u
284 #elif	LONG_BIT == 32
285 #define	_PMC_HM		2654435769u
286 #else
287 #error 	Must know the size of 'long' to compile
288 #endif
289 
290 #define	PMC_HASH_PTR(P,M)	((((unsigned long) (P) >> 2) * _PMC_HM) & (M))
291 
292 /*
293  * Syscall structures
294  */
295 
296 /* The `sysent' for the new syscall */
297 static struct sysent pmc_sysent = {
298 	2,			/* sy_narg */
299 	pmc_syscall_handler	/* sy_call */
300 };
301 
302 static struct syscall_module_data pmc_syscall_mod = {
303 	load,
304 	NULL,
305 	&pmc_syscall_num,
306 	&pmc_sysent,
307 	{ 0, NULL }
308 };
309 
310 static moduledata_t pmc_mod = {
311 	PMC_MODULE_NAME,
312 	syscall_module_handler,
313 	&pmc_syscall_mod
314 };
315 
316 DECLARE_MODULE(pmc, pmc_mod, SI_SUB_SMP, SI_ORDER_ANY);
317 MODULE_VERSION(pmc, PMC_VERSION);
318 
319 #ifdef	DEBUG
320 enum pmc_dbgparse_state {
321 	PMCDS_WS,		/* in whitespace */
322 	PMCDS_MAJOR,		/* seen a major keyword */
323 	PMCDS_MINOR
324 };
325 
326 static int
327 pmc_debugflags_parse(char *newstr, char *fence)
328 {
329 	char c, *p, *q;
330 	struct pmc_debugflags *tmpflags;
331 	int error, found, *newbits, tmp;
332 	size_t kwlen;
333 
334 	MALLOC(tmpflags, struct pmc_debugflags *, sizeof(*tmpflags),
335 	    M_PMC, M_WAITOK|M_ZERO);
336 
337 	p = newstr;
338 	error = 0;
339 
340 	for (; p < fence && (c = *p); p++) {
341 
342 		/* skip white space */
343 		if (c == ' ' || c == '\t')
344 			continue;
345 
346 		/* look for a keyword followed by "=" */
347 		for (q = p; p < fence && (c = *p) && c != '='; p++)
348 			;
349 		if (c != '=') {
350 			error = EINVAL;
351 			goto done;
352 		}
353 
354 		kwlen = p - q;
355 		newbits = NULL;
356 
357 		/* lookup flag group name */
358 #define	DBG_SET_FLAG_MAJ(S,F)						\
359 		if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0)	\
360 			newbits = &tmpflags->pdb_ ## F;
361 
362 		DBG_SET_FLAG_MAJ("cpu",		CPU);
363 		DBG_SET_FLAG_MAJ("csw",		CSW);
364 		DBG_SET_FLAG_MAJ("logging",	LOG);
365 		DBG_SET_FLAG_MAJ("module",	MOD);
366 		DBG_SET_FLAG_MAJ("md", 		MDP);
367 		DBG_SET_FLAG_MAJ("owner",	OWN);
368 		DBG_SET_FLAG_MAJ("pmc",		PMC);
369 		DBG_SET_FLAG_MAJ("process",	PRC);
370 		DBG_SET_FLAG_MAJ("sampling", 	SAM);
371 
372 		if (newbits == NULL) {
373 			error = EINVAL;
374 			goto done;
375 		}
376 
377 		p++;		/* skip the '=' */
378 
379 		/* Now parse the individual flags */
380 		tmp = 0;
381 	newflag:
382 		for (q = p; p < fence && (c = *p); p++)
383 			if (c == ' ' || c == '\t' || c == ',')
384 				break;
385 
386 		/* p == fence or c == ws or c == "," or c == 0 */
387 
388 		if ((kwlen = p - q) == 0) {
389 			*newbits = tmp;
390 			continue;
391 		}
392 
393 		found = 0;
394 #define	DBG_SET_FLAG_MIN(S,F)						\
395 		if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0)	\
396 			tmp |= found = (1 << PMC_DEBUG_MIN_ ## F)
397 
398 		/* a '*' denotes all possible flags in the group */
399 		if (kwlen == 1 && *q == '*')
400 			tmp = found = ~0;
401 		/* look for individual flag names */
402 		DBG_SET_FLAG_MIN("allocaterow", ALR);
403 		DBG_SET_FLAG_MIN("allocate",	ALL);
404 		DBG_SET_FLAG_MIN("attach",	ATT);
405 		DBG_SET_FLAG_MIN("bind",	BND);
406 		DBG_SET_FLAG_MIN("config",	CFG);
407 		DBG_SET_FLAG_MIN("exec",	EXC);
408 		DBG_SET_FLAG_MIN("exit",	EXT);
409 		DBG_SET_FLAG_MIN("find",	FND);
410 		DBG_SET_FLAG_MIN("flush",	FLS);
411 		DBG_SET_FLAG_MIN("fork",	FRK);
412 		DBG_SET_FLAG_MIN("getbuf",	GTB);
413 		DBG_SET_FLAG_MIN("hook",	PMH);
414 		DBG_SET_FLAG_MIN("init",	INI);
415 		DBG_SET_FLAG_MIN("intr",	INT);
416 		DBG_SET_FLAG_MIN("linktarget",	TLK);
417 		DBG_SET_FLAG_MIN("mayberemove", OMR);
418 		DBG_SET_FLAG_MIN("ops",		OPS);
419 		DBG_SET_FLAG_MIN("read",	REA);
420 		DBG_SET_FLAG_MIN("register",	REG);
421 		DBG_SET_FLAG_MIN("release",	REL);
422 		DBG_SET_FLAG_MIN("remove",	ORM);
423 		DBG_SET_FLAG_MIN("sample",	SAM);
424 		DBG_SET_FLAG_MIN("scheduleio",	SIO);
425 		DBG_SET_FLAG_MIN("select",	SEL);
426 		DBG_SET_FLAG_MIN("signal",	SIG);
427 		DBG_SET_FLAG_MIN("swi",		SWI);
428 		DBG_SET_FLAG_MIN("swo",		SWO);
429 		DBG_SET_FLAG_MIN("start",	STA);
430 		DBG_SET_FLAG_MIN("stop",	STO);
431 		DBG_SET_FLAG_MIN("syscall",	PMS);
432 		DBG_SET_FLAG_MIN("unlinktarget", TUL);
433 		DBG_SET_FLAG_MIN("write",	WRI);
434 		if (found == 0) {
435 			/* unrecognized flag name */
436 			error = EINVAL;
437 			goto done;
438 		}
439 
440 		if (c == 0 || c == ' ' || c == '\t') {	/* end of flag group */
441 			*newbits = tmp;
442 			continue;
443 		}
444 
445 		p++;
446 		goto newflag;
447 	}
448 
449 	/* save the new flag set */
450 	bcopy(tmpflags, &pmc_debugflags, sizeof(pmc_debugflags));
451 
452  done:
453 	FREE(tmpflags, M_PMC);
454 	return error;
455 }
456 
457 static int
458 pmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS)
459 {
460 	char *fence, *newstr;
461 	int error;
462 	unsigned int n;
463 
464 	(void) arg1; (void) arg2; /* unused parameters */
465 
466 	n = sizeof(pmc_debugstr);
467 	MALLOC(newstr, char *, n, M_PMC, M_ZERO|M_WAITOK);
468 	(void) strlcpy(newstr, pmc_debugstr, n);
469 
470 	error = sysctl_handle_string(oidp, newstr, n, req);
471 
472 	/* if there is a new string, parse and copy it */
473 	if (error == 0 && req->newptr != NULL) {
474 		fence = newstr + (n < req->newlen ? n : req->newlen + 1);
475 		if ((error = pmc_debugflags_parse(newstr, fence)) == 0)
476 			(void) strlcpy(pmc_debugstr, newstr,
477 			    sizeof(pmc_debugstr));
478 	}
479 
480 	FREE(newstr, M_PMC);
481 
482 	return error;
483 }
484 #endif
485 
486 /*
487  * Concurrency Control
488  *
489  * The driver manages the following data structures:
490  *
491  *   - target process descriptors, one per target process
492  *   - owner process descriptors (and attached lists), one per owner process
493  *   - lookup hash tables for owner and target processes
494  *   - PMC descriptors (and attached lists)
495  *   - per-cpu hardware state
496  *   - the 'hook' variable through which the kernel calls into
497  *     this module
498  *   - the machine hardware state (managed by the MD layer)
499  *
500  * These data structures are accessed from:
501  *
502  * - thread context-switch code
503  * - interrupt handlers (possibly on multiple cpus)
504  * - kernel threads on multiple cpus running on behalf of user
505  *   processes doing system calls
506  * - this driver's private kernel threads
507  *
508  * = Locks and Locking strategy =
509  *
510  * The driver uses four locking strategies for its operation:
511  *
512  * - The global SX lock "pmc_sx" is used to protect internal
513  *   data structures.
514  *
515  *   Calls into the module by syscall() start with this lock being
516  *   held in exclusive mode.  Depending on the requested operation,
517  *   the lock may be downgraded to 'shared' mode to allow more
518  *   concurrent readers into the module.  Calls into the module from
519  *   other parts of the kernel acquire the lock in shared mode.
520  *
521  *   This SX lock is held in exclusive mode for any operations that
522  *   modify the linkages between the driver's internal data structures.
523  *
524  *   The 'pmc_hook' function pointer is also protected by this lock.
525  *   It is only examined with the sx lock held in exclusive mode.  The
526  *   kernel module is allowed to be unloaded only with the sx lock held
527  *   in exclusive mode.  In normal syscall handling, after acquiring the
528  *   pmc_sx lock we first check that 'pmc_hook' is non-null before
529  *   proceeding.  This prevents races between the thread unloading the module
530  *   and other threads seeking to use the module.
531  *
532  * - Lookups of target process structures and owner process structures
533  *   cannot use the global "pmc_sx" SX lock because these lookups need
534  *   to happen during context switches and in other critical sections
535  *   where sleeping is not allowed.  We protect these lookup tables
536  *   with their own private spin-mutexes, "pmc_processhash_mtx" and
537  *   "pmc_ownerhash_mtx".
538  *
539  * - Interrupt handlers work in a lock free manner.  At interrupt
540  *   time, handlers look at the PMC pointer (phw->phw_pmc) configured
541  *   when the PMC was started.  If this pointer is NULL, the interrupt
542  *   is ignored after updating driver statistics.  We ensure that this
543  *   pointer is set (using an atomic operation if necessary) before the
544  *   PMC hardware is started.  Conversely, this pointer is unset atomically
545  *   only after the PMC hardware is stopped.
546  *
547  *   We ensure that everything needed for the operation of an
548  *   interrupt handler is available without it needing to acquire any
549  *   locks.  We also ensure that a PMC's software state is destroyed only
550  *   after the PMC is taken off hardware (on all CPUs).
551  *
552  * - Context-switch handling with process-private PMCs needs more
553  *   care.
554  *
555  *   A given process may be the target of multiple PMCs.  For example,
556  *   PMCATTACH and PMCDETACH may be requested by a process on one CPU
557  *   while the target process is running on another.  A PMC could also
558  *   be getting released because its owner is exiting.  We tackle
559  *   these situations in the following manner:
560  *
561  *   - each target process structure 'pmc_process' has an array
562  *     of 'struct pmc *' pointers, one for each hardware PMC.
563  *
564  *   - At context switch IN time, each "target" PMC in RUNNING state
565  *     gets started on hardware and a pointer to each PMC is copied into
566  *     the per-cpu phw array.  The 'runcount' for the PMC is
567  *     incremented.
568  *
569  *   - At context switch OUT time, all process-virtual PMCs are stopped
570  *     on hardware.  The saved value is added to the PMCs value field
571  *     only if the PMC is in a non-deleted state (the PMCs state could
572  *     have changed during the current time slice).
573  *
574  *     Note that since in-between a switch IN on a processor and a switch
575  *     OUT, the PMC could have been released on another CPU.  Therefore
576  *     context switch OUT always looks at the hardware state to turn
577  *     OFF PMCs and will update a PMC's saved value only if reachable
578  *     from the target process record.
579  *
580  *   - OP PMCRELEASE could be called on a PMC at any time (the PMC could
581  *     be attached to many processes at the time of the call and could
582  *     be active on multiple CPUs).
583  *
584  *     We prevent further scheduling of the PMC by marking it as in
585  *     state 'DELETED'.  If the runcount of the PMC is non-zero then
586  *     this PMC is currently running on a CPU somewhere.  The thread
587  *     doing the PMCRELEASE operation waits by repeatedly doing a
588  *     pause() till the runcount comes to zero.
589  *
590  * The contents of a PMC descriptor (struct pmc) are protected using
591  * a spin-mutex.  In order to save space, we use a mutex pool.
592  *
593  * In terms of lock types used by witness(4), we use:
594  * - Type "pmc-sx", used by the global SX lock.
595  * - Type "pmc-sleep", for sleep mutexes used by logger threads.
596  * - Type "pmc-per-proc", for protecting PMC owner descriptors.
597  * - Type "pmc-leaf", used for all other spin mutexes.
598  */
599 
600 /*
601  * save the cpu binding of the current kthread
602  */
603 
604 static void
605 pmc_save_cpu_binding(struct pmc_binding *pb)
606 {
607 	PMCDBG(CPU,BND,2, "%s", "save-cpu");
608 	thread_lock(curthread);
609 	pb->pb_bound = sched_is_bound(curthread);
610 	pb->pb_cpu   = curthread->td_oncpu;
611 	thread_unlock(curthread);
612 	PMCDBG(CPU,BND,2, "save-cpu cpu=%d", pb->pb_cpu);
613 }
614 
615 /*
616  * restore the cpu binding of the current thread
617  */
618 
619 static void
620 pmc_restore_cpu_binding(struct pmc_binding *pb)
621 {
622 	PMCDBG(CPU,BND,2, "restore-cpu curcpu=%d restore=%d",
623 	    curthread->td_oncpu, pb->pb_cpu);
624 	thread_lock(curthread);
625 	if (pb->pb_bound)
626 		sched_bind(curthread, pb->pb_cpu);
627 	else
628 		sched_unbind(curthread);
629 	thread_unlock(curthread);
630 	PMCDBG(CPU,BND,2, "%s", "restore-cpu done");
631 }
632 
633 /*
634  * move execution over the specified cpu and bind it there.
635  */
636 
637 static void
638 pmc_select_cpu(int cpu)
639 {
640 	KASSERT(cpu >= 0 && cpu < mp_ncpus,
641 	    ("[pmc,%d] bad cpu number %d", __LINE__, cpu));
642 
643 	/* never move to a disabled CPU */
644 	KASSERT(pmc_cpu_is_disabled(cpu) == 0, ("[pmc,%d] selecting "
645 	    "disabled CPU %d", __LINE__, cpu));
646 
647 	PMCDBG(CPU,SEL,2, "select-cpu cpu=%d", cpu);
648 	thread_lock(curthread);
649 	sched_bind(curthread, cpu);
650 	thread_unlock(curthread);
651 
652 	KASSERT(curthread->td_oncpu == cpu,
653 	    ("[pmc,%d] CPU not bound [cpu=%d, curr=%d]", __LINE__,
654 		cpu, curthread->td_oncpu));
655 
656 	PMCDBG(CPU,SEL,2, "select-cpu cpu=%d ok", cpu);
657 }
658 
659 /*
660  * Force a context switch.
661  *
662  * We do this by pause'ing for 1 tick -- invoking mi_switch() is not
663  * guaranteed to force a context switch.
664  */
665 
666 static void
667 pmc_force_context_switch(void)
668 {
669 
670 	pause("pmcctx", 1);
671 }
672 
673 /*
674  * Get the file name for an executable.  This is a simple wrapper
675  * around vn_fullpath(9).
676  */
677 
678 static void
679 pmc_getfilename(struct vnode *v, char **fullpath, char **freepath)
680 {
681 	struct thread *td;
682 
683 	td = curthread;
684 	*fullpath = "unknown";
685 	*freepath = NULL;
686 	vn_lock(v, LK_CANRECURSE | LK_EXCLUSIVE | LK_RETRY, td);
687 	vn_fullpath(td, v, fullpath, freepath);
688 	VOP_UNLOCK(v, 0, td);
689 }
690 
691 /*
692  * remove an process owning PMCs
693  */
694 
695 void
696 pmc_remove_owner(struct pmc_owner *po)
697 {
698 	struct pmc *pm, *tmp;
699 
700 	sx_assert(&pmc_sx, SX_XLOCKED);
701 
702 	PMCDBG(OWN,ORM,1, "remove-owner po=%p", po);
703 
704 	/* Remove descriptor from the owner hash table */
705 	LIST_REMOVE(po, po_next);
706 
707 	/* release all owned PMC descriptors */
708 	LIST_FOREACH_SAFE(pm, &po->po_pmcs, pm_next, tmp) {
709 		PMCDBG(OWN,ORM,2, "pmc=%p", pm);
710 		KASSERT(pm->pm_owner == po,
711 		    ("[pmc,%d] owner %p != po %p", __LINE__, pm->pm_owner, po));
712 
713 		pmc_release_pmc_descriptor(pm);	/* will unlink from the list */
714 	}
715 
716 	KASSERT(po->po_sscount == 0,
717 	    ("[pmc,%d] SS count not zero", __LINE__));
718 	KASSERT(LIST_EMPTY(&po->po_pmcs),
719 	    ("[pmc,%d] PMC list not empty", __LINE__));
720 
721 	/* de-configure the log file if present */
722 	if (po->po_flags & PMC_PO_OWNS_LOGFILE)
723 		pmclog_deconfigure_log(po);
724 }
725 
726 /*
727  * remove an owner process record if all conditions are met.
728  */
729 
730 static void
731 pmc_maybe_remove_owner(struct pmc_owner *po)
732 {
733 
734 	PMCDBG(OWN,OMR,1, "maybe-remove-owner po=%p", po);
735 
736 	/*
737 	 * Remove owner record if
738 	 * - this process does not own any PMCs
739 	 * - this process has not allocated a system-wide sampling buffer
740 	 */
741 
742 	if (LIST_EMPTY(&po->po_pmcs) &&
743 	    ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)) {
744 		pmc_remove_owner(po);
745 		pmc_destroy_owner_descriptor(po);
746 	}
747 }
748 
749 /*
750  * Add an association between a target process and a PMC.
751  */
752 
753 static void
754 pmc_link_target_process(struct pmc *pm, struct pmc_process *pp)
755 {
756 	int ri;
757 	struct pmc_target *pt;
758 
759 	sx_assert(&pmc_sx, SX_XLOCKED);
760 
761 	KASSERT(pm != NULL && pp != NULL,
762 	    ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp));
763 	KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)),
764 	    ("[pmc,%d] Attaching a non-process-virtual pmc=%p to pid=%d",
765 		__LINE__, pm, pp->pp_proc->p_pid));
766 	KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt < ((int) md->pmd_npmc - 1),
767 	    ("[pmc,%d] Illegal reference count %d for process record %p",
768 		__LINE__, pp->pp_refcnt, (void *) pp));
769 
770 	ri = PMC_TO_ROWINDEX(pm);
771 
772 	PMCDBG(PRC,TLK,1, "link-target pmc=%p ri=%d pmc-process=%p",
773 	    pm, ri, pp);
774 
775 #ifdef	DEBUG
776 	LIST_FOREACH(pt, &pm->pm_targets, pt_next)
777 	    if (pt->pt_process == pp)
778 		    KASSERT(0, ("[pmc,%d] pp %p already in pmc %p targets",
779 				__LINE__, pp, pm));
780 #endif
781 
782 	MALLOC(pt, struct pmc_target *, sizeof(struct pmc_target),
783 	    M_PMC, M_ZERO|M_WAITOK);
784 
785 	pt->pt_process = pp;
786 
787 	LIST_INSERT_HEAD(&pm->pm_targets, pt, pt_next);
788 
789 	atomic_store_rel_ptr((uintptr_t *)&pp->pp_pmcs[ri].pp_pmc,
790 	    (uintptr_t)pm);
791 
792 	if (pm->pm_owner->po_owner == pp->pp_proc)
793 		pm->pm_flags |= PMC_F_ATTACHED_TO_OWNER;
794 
795 	/*
796 	 * Initialize the per-process values at this row index.
797 	 */
798 	pp->pp_pmcs[ri].pp_pmcval = PMC_TO_MODE(pm) == PMC_MODE_TS ?
799 	    pm->pm_sc.pm_reloadcount : 0;
800 
801 	pp->pp_refcnt++;
802 
803 }
804 
805 /*
806  * Removes the association between a target process and a PMC.
807  */
808 
809 static void
810 pmc_unlink_target_process(struct pmc *pm, struct pmc_process *pp)
811 {
812 	int ri;
813 	struct proc *p;
814 	struct pmc_target *ptgt;
815 
816 	sx_assert(&pmc_sx, SX_XLOCKED);
817 
818 	KASSERT(pm != NULL && pp != NULL,
819 	    ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp));
820 
821 	KASSERT(pp->pp_refcnt >= 1 && pp->pp_refcnt < (int) md->pmd_npmc,
822 	    ("[pmc,%d] Illegal ref count %d on process record %p",
823 		__LINE__, pp->pp_refcnt, (void *) pp));
824 
825 	ri = PMC_TO_ROWINDEX(pm);
826 
827 	PMCDBG(PRC,TUL,1, "unlink-target pmc=%p ri=%d pmc-process=%p",
828 	    pm, ri, pp);
829 
830 	KASSERT(pp->pp_pmcs[ri].pp_pmc == pm,
831 	    ("[pmc,%d] PMC ri %d mismatch pmc %p pp->[ri] %p", __LINE__,
832 		ri, pm, pp->pp_pmcs[ri].pp_pmc));
833 
834 	pp->pp_pmcs[ri].pp_pmc = NULL;
835 	pp->pp_pmcs[ri].pp_pmcval = (pmc_value_t) 0;
836 
837 	/* Remove owner-specific flags */
838 	if (pm->pm_owner->po_owner == pp->pp_proc) {
839 		pp->pp_flags &= ~PMC_PP_ENABLE_MSR_ACCESS;
840 		pm->pm_flags &= ~PMC_F_ATTACHED_TO_OWNER;
841 	}
842 
843 	pp->pp_refcnt--;
844 
845 	/* Remove the target process from the PMC structure */
846 	LIST_FOREACH(ptgt, &pm->pm_targets, pt_next)
847 		if (ptgt->pt_process == pp)
848 			break;
849 
850 	KASSERT(ptgt != NULL, ("[pmc,%d] process %p (pp: %p) not found "
851 		    "in pmc %p", __LINE__, pp->pp_proc, pp, pm));
852 
853 	LIST_REMOVE(ptgt, pt_next);
854 	FREE(ptgt, M_PMC);
855 
856 	/* if the PMC now lacks targets, send the owner a SIGIO */
857 	if (LIST_EMPTY(&pm->pm_targets)) {
858 		p = pm->pm_owner->po_owner;
859 		PROC_LOCK(p);
860 		psignal(p, SIGIO);
861 		PROC_UNLOCK(p);
862 
863 		PMCDBG(PRC,SIG,2, "signalling proc=%p signal=%d", p,
864 		    SIGIO);
865 	}
866 }
867 
868 /*
869  * Check if PMC 'pm' may be attached to target process 't'.
870  */
871 
872 static int
873 pmc_can_attach(struct pmc *pm, struct proc *t)
874 {
875 	struct proc *o;		/* pmc owner */
876 	struct ucred *oc, *tc;	/* owner, target credentials */
877 	int decline_attach, i;
878 
879 	/*
880 	 * A PMC's owner can always attach that PMC to itself.
881 	 */
882 
883 	if ((o = pm->pm_owner->po_owner) == t)
884 		return 0;
885 
886 	PROC_LOCK(o);
887 	oc = o->p_ucred;
888 	crhold(oc);
889 	PROC_UNLOCK(o);
890 
891 	PROC_LOCK(t);
892 	tc = t->p_ucred;
893 	crhold(tc);
894 	PROC_UNLOCK(t);
895 
896 	/*
897 	 * The effective uid of the PMC owner should match at least one
898 	 * of the {effective,real,saved} uids of the target process.
899 	 */
900 
901 	decline_attach = oc->cr_uid != tc->cr_uid &&
902 	    oc->cr_uid != tc->cr_svuid &&
903 	    oc->cr_uid != tc->cr_ruid;
904 
905 	/*
906 	 * Every one of the target's group ids, must be in the owner's
907 	 * group list.
908 	 */
909 	for (i = 0; !decline_attach && i < tc->cr_ngroups; i++)
910 		decline_attach = !groupmember(tc->cr_groups[i], oc);
911 
912 	/* check the read and saved gids too */
913 	if (decline_attach == 0)
914 		decline_attach = !groupmember(tc->cr_rgid, oc) ||
915 		    !groupmember(tc->cr_svgid, oc);
916 
917 	crfree(tc);
918 	crfree(oc);
919 
920 	return !decline_attach;
921 }
922 
923 /*
924  * Attach a process to a PMC.
925  */
926 
927 static int
928 pmc_attach_one_process(struct proc *p, struct pmc *pm)
929 {
930 	int ri;
931 	char *fullpath, *freepath;
932 	struct pmc_process	*pp;
933 
934 	sx_assert(&pmc_sx, SX_XLOCKED);
935 
936 	PMCDBG(PRC,ATT,2, "attach-one pm=%p ri=%d proc=%p (%d, %s)", pm,
937 	    PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm);
938 
939 	/*
940 	 * Locate the process descriptor corresponding to process 'p',
941 	 * allocating space as needed.
942 	 *
943 	 * Verify that rowindex 'pm_rowindex' is free in the process
944 	 * descriptor.
945 	 *
946 	 * If not, allocate space for a descriptor and link the
947 	 * process descriptor and PMC.
948 	 */
949 	ri = PMC_TO_ROWINDEX(pm);
950 
951 	if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_ALLOCATE)) == NULL)
952 		return ENOMEM;
953 
954 	if (pp->pp_pmcs[ri].pp_pmc == pm) /* already present at slot [ri] */
955 		return EEXIST;
956 
957 	if (pp->pp_pmcs[ri].pp_pmc != NULL)
958 		return EBUSY;
959 
960 	pmc_link_target_process(pm, pp);
961 
962 	if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) &&
963 	    (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) == 0)
964 		pm->pm_flags |= PMC_F_NEEDS_LOGFILE;
965 
966 	pm->pm_flags |= PMC_F_ATTACH_DONE; /* mark as attached */
967 
968 	/* issue an attach event to a configured log file */
969 	if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE) {
970 		pmc_getfilename(p->p_textvp, &fullpath, &freepath);
971 		pmclog_process_pmcattach(pm, p->p_pid, fullpath);
972 		if (freepath)
973 			FREE(freepath, M_TEMP);
974 		if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
975 			pmc_log_process_mappings(pm->pm_owner, p);
976 	}
977 	/* mark process as using HWPMCs */
978 	PROC_LOCK(p);
979 	p->p_flag |= P_HWPMC;
980 	PROC_UNLOCK(p);
981 
982 	return 0;
983 }
984 
985 /*
986  * Attach a process and optionally its children
987  */
988 
989 static int
990 pmc_attach_process(struct proc *p, struct pmc *pm)
991 {
992 	int error;
993 	struct proc *top;
994 
995 	sx_assert(&pmc_sx, SX_XLOCKED);
996 
997 	PMCDBG(PRC,ATT,1, "attach pm=%p ri=%d proc=%p (%d, %s)", pm,
998 	    PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm);
999 
1000 
1001 	/*
1002 	 * If this PMC successfully allowed a GETMSR operation
1003 	 * in the past, disallow further ATTACHes.
1004 	 */
1005 
1006 	if ((pm->pm_flags & PMC_PP_ENABLE_MSR_ACCESS) != 0)
1007 		return EPERM;
1008 
1009 	if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0)
1010 		return pmc_attach_one_process(p, pm);
1011 
1012 	/*
1013 	 * Traverse all child processes, attaching them to
1014 	 * this PMC.
1015 	 */
1016 
1017 	sx_slock(&proctree_lock);
1018 
1019 	top = p;
1020 
1021 	for (;;) {
1022 		if ((error = pmc_attach_one_process(p, pm)) != 0)
1023 			break;
1024 		if (!LIST_EMPTY(&p->p_children))
1025 			p = LIST_FIRST(&p->p_children);
1026 		else for (;;) {
1027 			if (p == top)
1028 				goto done;
1029 			if (LIST_NEXT(p, p_sibling)) {
1030 				p = LIST_NEXT(p, p_sibling);
1031 				break;
1032 			}
1033 			p = p->p_pptr;
1034 		}
1035 	}
1036 
1037 	if (error)
1038 		(void) pmc_detach_process(top, pm);
1039 
1040  done:
1041 	sx_sunlock(&proctree_lock);
1042 	return error;
1043 }
1044 
1045 /*
1046  * Detach a process from a PMC.  If there are no other PMCs tracking
1047  * this process, remove the process structure from its hash table.  If
1048  * 'flags' contains PMC_FLAG_REMOVE, then free the process structure.
1049  */
1050 
1051 static int
1052 pmc_detach_one_process(struct proc *p, struct pmc *pm, int flags)
1053 {
1054 	int ri;
1055 	struct pmc_process *pp;
1056 
1057 	sx_assert(&pmc_sx, SX_XLOCKED);
1058 
1059 	KASSERT(pm != NULL,
1060 	    ("[pmc,%d] null pm pointer", __LINE__));
1061 
1062 	ri = PMC_TO_ROWINDEX(pm);
1063 
1064 	PMCDBG(PRC,ATT,2, "detach-one pm=%p ri=%d proc=%p (%d, %s) flags=0x%x",
1065 	    pm, ri, p, p->p_pid, p->p_comm, flags);
1066 
1067 	if ((pp = pmc_find_process_descriptor(p, 0)) == NULL)
1068 		return ESRCH;
1069 
1070 	if (pp->pp_pmcs[ri].pp_pmc != pm)
1071 		return EINVAL;
1072 
1073 	pmc_unlink_target_process(pm, pp);
1074 
1075 	/* Issue a detach entry if a log file is configured */
1076 	if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE)
1077 		pmclog_process_pmcdetach(pm, p->p_pid);
1078 
1079 	/*
1080 	 * If there are no PMCs targetting this process, we remove its
1081 	 * descriptor from the target hash table and unset the P_HWPMC
1082 	 * flag in the struct proc.
1083 	 */
1084 	KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt < (int) md->pmd_npmc,
1085 	    ("[pmc,%d] Illegal refcnt %d for process struct %p",
1086 		__LINE__, pp->pp_refcnt, pp));
1087 
1088 	if (pp->pp_refcnt != 0)	/* still a target of some PMC */
1089 		return 0;
1090 
1091 	pmc_remove_process_descriptor(pp);
1092 
1093 	if (flags & PMC_FLAG_REMOVE)
1094 		FREE(pp, M_PMC);
1095 
1096 	PROC_LOCK(p);
1097 	p->p_flag &= ~P_HWPMC;
1098 	PROC_UNLOCK(p);
1099 
1100 	return 0;
1101 }
1102 
1103 /*
1104  * Detach a process and optionally its descendants from a PMC.
1105  */
1106 
1107 static int
1108 pmc_detach_process(struct proc *p, struct pmc *pm)
1109 {
1110 	struct proc *top;
1111 
1112 	sx_assert(&pmc_sx, SX_XLOCKED);
1113 
1114 	PMCDBG(PRC,ATT,1, "detach pm=%p ri=%d proc=%p (%d, %s)", pm,
1115 	    PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm);
1116 
1117 	if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0)
1118 		return pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE);
1119 
1120 	/*
1121 	 * Traverse all children, detaching them from this PMC.  We
1122 	 * ignore errors since we could be detaching a PMC from a
1123 	 * partially attached proc tree.
1124 	 */
1125 
1126 	sx_slock(&proctree_lock);
1127 
1128 	top = p;
1129 
1130 	for (;;) {
1131 		(void) pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE);
1132 
1133 		if (!LIST_EMPTY(&p->p_children))
1134 			p = LIST_FIRST(&p->p_children);
1135 		else for (;;) {
1136 			if (p == top)
1137 				goto done;
1138 			if (LIST_NEXT(p, p_sibling)) {
1139 				p = LIST_NEXT(p, p_sibling);
1140 				break;
1141 			}
1142 			p = p->p_pptr;
1143 		}
1144 	}
1145 
1146  done:
1147 	sx_sunlock(&proctree_lock);
1148 
1149 	if (LIST_EMPTY(&pm->pm_targets))
1150 		pm->pm_flags &= ~PMC_F_ATTACH_DONE;
1151 
1152 	return 0;
1153 }
1154 
1155 
1156 /*
1157  * Thread context switch IN
1158  */
1159 
1160 static void
1161 pmc_process_csw_in(struct thread *td)
1162 {
1163 	int cpu;
1164 	unsigned int ri;
1165 	struct pmc *pm;
1166 	struct proc *p;
1167 	struct pmc_cpu *pc;
1168 	struct pmc_hw *phw;
1169 	struct pmc_process *pp;
1170 	pmc_value_t newvalue;
1171 
1172 	p = td->td_proc;
1173 
1174 	if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE)) == NULL)
1175 		return;
1176 
1177 	KASSERT(pp->pp_proc == td->td_proc,
1178 	    ("[pmc,%d] not my thread state", __LINE__));
1179 
1180 	critical_enter(); /* no preemption from this point */
1181 
1182 	cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */
1183 
1184 	PMCDBG(CSW,SWI,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p,
1185 	    p->p_pid, p->p_comm, pp);
1186 
1187 	KASSERT(cpu >= 0 && cpu < mp_ncpus,
1188 	    ("[pmc,%d] wierd CPU id %d", __LINE__, cpu));
1189 
1190 	pc = pmc_pcpu[cpu];
1191 
1192 	for (ri = 0; ri < md->pmd_npmc; ri++) {
1193 
1194 		if ((pm = pp->pp_pmcs[ri].pp_pmc) == NULL)
1195 			continue;
1196 
1197 		KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)),
1198 		    ("[pmc,%d] Target PMC in non-virtual mode (%d)",
1199 			__LINE__, PMC_TO_MODE(pm)));
1200 
1201 		KASSERT(PMC_TO_ROWINDEX(pm) == ri,
1202 		    ("[pmc,%d] Row index mismatch pmc %d != ri %d",
1203 			__LINE__, PMC_TO_ROWINDEX(pm), ri));
1204 
1205 		/*
1206 		 * Only PMCs that are marked as 'RUNNING' need
1207 		 * be placed on hardware.
1208 		 */
1209 
1210 		if (pm->pm_state != PMC_STATE_RUNNING)
1211 			continue;
1212 
1213 		/* increment PMC runcount */
1214 		atomic_add_rel_32(&pm->pm_runcount, 1);
1215 
1216 		/* configure the HWPMC we are going to use. */
1217 		md->pmd_config_pmc(cpu, ri, pm);
1218 
1219 		phw = pc->pc_hwpmcs[ri];
1220 
1221 		KASSERT(phw != NULL,
1222 		    ("[pmc,%d] null hw pointer", __LINE__));
1223 
1224 		KASSERT(phw->phw_pmc == pm,
1225 		    ("[pmc,%d] hw->pmc %p != pmc %p", __LINE__,
1226 			phw->phw_pmc, pm));
1227 
1228 		/*
1229 		 * Write out saved value and start the PMC.
1230 		 *
1231 		 * Sampling PMCs use a per-process value, while
1232 		 * counting mode PMCs use a per-pmc value that is
1233 		 * inherited across descendants.
1234 		 */
1235 		if (PMC_TO_MODE(pm) == PMC_MODE_TS) {
1236 			mtx_pool_lock_spin(pmc_mtxpool, pm);
1237 			newvalue = PMC_PCPU_SAVED(cpu,ri) =
1238 			    pp->pp_pmcs[ri].pp_pmcval;
1239 			mtx_pool_unlock_spin(pmc_mtxpool, pm);
1240 		} else {
1241 			KASSERT(PMC_TO_MODE(pm) == PMC_MODE_TC,
1242 			    ("[pmc,%d] illegal mode=%d", __LINE__,
1243 			    PMC_TO_MODE(pm)));
1244 			mtx_pool_lock_spin(pmc_mtxpool, pm);
1245 			newvalue = PMC_PCPU_SAVED(cpu, ri) =
1246 			    pm->pm_gv.pm_savedvalue;
1247 			mtx_pool_unlock_spin(pmc_mtxpool, pm);
1248 		}
1249 
1250 		PMCDBG(CSW,SWI,1,"cpu=%d ri=%d new=%jd", cpu, ri, newvalue);
1251 
1252 		md->pmd_write_pmc(cpu, ri, newvalue);
1253 		md->pmd_start_pmc(cpu, ri);
1254 	}
1255 
1256 	/*
1257 	 * perform any other architecture/cpu dependent thread
1258 	 * switch-in actions.
1259 	 */
1260 
1261 	(void) (*md->pmd_switch_in)(pc, pp);
1262 
1263 	critical_exit();
1264 
1265 }
1266 
1267 /*
1268  * Thread context switch OUT.
1269  */
1270 
1271 static void
1272 pmc_process_csw_out(struct thread *td)
1273 {
1274 	int cpu;
1275 	enum pmc_mode mode;
1276 	unsigned int ri;
1277 	struct pmc *pm;
1278 	struct proc *p;
1279 	struct pmc_cpu *pc;
1280 	struct pmc_process *pp;
1281 	int64_t tmp;
1282 	pmc_value_t newvalue;
1283 
1284 	/*
1285 	 * Locate our process descriptor; this may be NULL if
1286 	 * this process is exiting and we have already removed
1287 	 * the process from the target process table.
1288 	 *
1289 	 * Note that due to kernel preemption, multiple
1290 	 * context switches may happen while the process is
1291 	 * exiting.
1292 	 *
1293 	 * Note also that if the target process cannot be
1294 	 * found we still need to deconfigure any PMCs that
1295 	 * are currently running on hardware.
1296 	 */
1297 
1298 	p = td->td_proc;
1299 	pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE);
1300 
1301 	/*
1302 	 * save PMCs
1303 	 */
1304 
1305 	critical_enter();
1306 
1307 	cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */
1308 
1309 	PMCDBG(CSW,SWO,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p,
1310 	    p->p_pid, p->p_comm, pp);
1311 
1312 	KASSERT(cpu >= 0 && cpu < mp_ncpus,
1313 	    ("[pmc,%d wierd CPU id %d", __LINE__, cpu));
1314 
1315 	pc = pmc_pcpu[cpu];
1316 
1317 	/*
1318 	 * When a PMC gets unlinked from a target PMC, it will
1319 	 * be removed from the target's pp_pmc[] array.
1320 	 *
1321 	 * However, on a MP system, the target could have been
1322 	 * executing on another CPU at the time of the unlink.
1323 	 * So, at context switch OUT time, we need to look at
1324 	 * the hardware to determine if a PMC is scheduled on
1325 	 * it.
1326 	 */
1327 
1328 	for (ri = 0; ri < md->pmd_npmc; ri++) {
1329 
1330 		pm = NULL;
1331 		(void) (*md->pmd_get_config)(cpu, ri, &pm);
1332 
1333 		if (pm == NULL)	/* nothing at this row index */
1334 			continue;
1335 
1336 		mode = PMC_TO_MODE(pm);
1337 		if (!PMC_IS_VIRTUAL_MODE(mode))
1338 			continue; /* not a process virtual PMC */
1339 
1340 		KASSERT(PMC_TO_ROWINDEX(pm) == ri,
1341 		    ("[pmc,%d] ri mismatch pmc(%d) ri(%d)",
1342 			__LINE__, PMC_TO_ROWINDEX(pm), ri));
1343 
1344 		/* Stop hardware if not already stopped */
1345 		if (pm->pm_stalled == 0)
1346 			md->pmd_stop_pmc(cpu, ri);
1347 
1348 		/* reduce this PMC's runcount */
1349 		atomic_subtract_rel_32(&pm->pm_runcount, 1);
1350 
1351 		/*
1352 		 * If this PMC is associated with this process,
1353 		 * save the reading.
1354 		 */
1355 
1356 		if (pp != NULL && pp->pp_pmcs[ri].pp_pmc != NULL) {
1357 
1358 			KASSERT(pm == pp->pp_pmcs[ri].pp_pmc,
1359 			    ("[pmc,%d] pm %p != pp_pmcs[%d] %p", __LINE__,
1360 				pm, ri, pp->pp_pmcs[ri].pp_pmc));
1361 
1362 			KASSERT(pp->pp_refcnt > 0,
1363 			    ("[pmc,%d] pp refcnt = %d", __LINE__,
1364 				pp->pp_refcnt));
1365 
1366 			md->pmd_read_pmc(cpu, ri, &newvalue);
1367 
1368 			tmp = newvalue - PMC_PCPU_SAVED(cpu,ri);
1369 
1370 			PMCDBG(CSW,SWI,1,"cpu=%d ri=%d tmp=%jd", cpu, ri,
1371 			    tmp);
1372 
1373 			if (mode == PMC_MODE_TS) {
1374 
1375 				/*
1376 				 * For sampling process-virtual PMCs,
1377 				 * we expect the count to be
1378 				 * decreasing as the 'value'
1379 				 * programmed into the PMC is the
1380 				 * number of events to be seen till
1381 				 * the next sampling interrupt.
1382 				 */
1383 				if (tmp < 0)
1384 					tmp += pm->pm_sc.pm_reloadcount;
1385 				mtx_pool_lock_spin(pmc_mtxpool, pm);
1386 				pp->pp_pmcs[ri].pp_pmcval -= tmp;
1387 				if ((int64_t) pp->pp_pmcs[ri].pp_pmcval < 0)
1388 					pp->pp_pmcs[ri].pp_pmcval +=
1389 					    pm->pm_sc.pm_reloadcount;
1390 				mtx_pool_unlock_spin(pmc_mtxpool, pm);
1391 
1392 			} else {
1393 
1394 				/*
1395 				 * For counting process-virtual PMCs,
1396 				 * we expect the count to be
1397 				 * increasing monotonically, modulo a 64
1398 				 * bit wraparound.
1399 				 */
1400 				KASSERT((int64_t) tmp >= 0,
1401 				    ("[pmc,%d] negative increment cpu=%d "
1402 				     "ri=%d newvalue=%jx saved=%jx "
1403 				     "incr=%jx", __LINE__, cpu, ri,
1404 				     newvalue, PMC_PCPU_SAVED(cpu,ri), tmp));
1405 
1406 				mtx_pool_lock_spin(pmc_mtxpool, pm);
1407 				pm->pm_gv.pm_savedvalue += tmp;
1408 				pp->pp_pmcs[ri].pp_pmcval += tmp;
1409 				mtx_pool_unlock_spin(pmc_mtxpool, pm);
1410 
1411 				if (pm->pm_flags & PMC_F_LOG_PROCCSW)
1412 					pmclog_process_proccsw(pm, pp, tmp);
1413 			}
1414 		}
1415 
1416 		/* mark hardware as free */
1417 		md->pmd_config_pmc(cpu, ri, NULL);
1418 	}
1419 
1420 	/*
1421 	 * perform any other architecture/cpu dependent thread
1422 	 * switch out functions.
1423 	 */
1424 
1425 	(void) (*md->pmd_switch_out)(pc, pp);
1426 
1427 	critical_exit();
1428 }
1429 
1430 /*
1431  * Log a KLD operation.
1432  */
1433 
1434 static void
1435 pmc_process_kld_load(struct pmckern_map_in *pkm)
1436 {
1437 	struct pmc_owner *po;
1438 
1439 	sx_assert(&pmc_sx, SX_LOCKED);
1440 
1441 	/*
1442 	 * Notify owners of system sampling PMCs about KLD operations.
1443 	 */
1444 
1445 	LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1446 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1447 	    	pmclog_process_map_in(po, (pid_t) -1, pkm->pm_address,
1448 		    (char *) pkm->pm_file);
1449 
1450 	/*
1451 	 * TODO: Notify owners of (all) process-sampling PMCs too.
1452 	 */
1453 
1454 	return;
1455 }
1456 
1457 static void
1458 pmc_process_kld_unload(struct pmckern_map_out *pkm)
1459 {
1460 	struct pmc_owner *po;
1461 
1462 	sx_assert(&pmc_sx, SX_LOCKED);
1463 
1464 	LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1465 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1466 		pmclog_process_map_out(po, (pid_t) -1,
1467 		    pkm->pm_address, pkm->pm_address + pkm->pm_size);
1468 
1469 	/*
1470 	 * TODO: Notify owners of process-sampling PMCs.
1471 	 */
1472 }
1473 
1474 /*
1475  * A mapping change for a process.
1476  */
1477 
1478 static void
1479 pmc_process_mmap(struct thread *td, struct pmckern_map_in *pkm)
1480 {
1481 	int ri;
1482 	pid_t pid;
1483 	char *fullpath, *freepath;
1484 	const struct pmc *pm;
1485 	struct pmc_owner *po;
1486 	const struct pmc_process *pp;
1487 
1488 	freepath = fullpath = NULL;
1489 	pmc_getfilename((struct vnode *) pkm->pm_file, &fullpath, &freepath);
1490 
1491 	pid = td->td_proc->p_pid;
1492 
1493 	/* Inform owners of all system-wide sampling PMCs. */
1494 	LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1495 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1496 		pmclog_process_map_in(po, pid, pkm->pm_address, fullpath);
1497 
1498 	if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL)
1499 		goto done;
1500 
1501 	/*
1502 	 * Inform sampling PMC owners tracking this process.
1503 	 */
1504 	for (ri = 0; ri < md->pmd_npmc; ri++)
1505 		if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL &&
1506 		    PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1507 			pmclog_process_map_in(pm->pm_owner,
1508 			    pid, pkm->pm_address, fullpath);
1509 
1510   done:
1511 	if (freepath)
1512 		FREE(freepath, M_TEMP);
1513 }
1514 
1515 
1516 /*
1517  * Log an munmap request.
1518  */
1519 
1520 static void
1521 pmc_process_munmap(struct thread *td, struct pmckern_map_out *pkm)
1522 {
1523 	int ri;
1524 	pid_t pid;
1525 	struct pmc_owner *po;
1526 	const struct pmc *pm;
1527 	const struct pmc_process *pp;
1528 
1529 	pid = td->td_proc->p_pid;
1530 
1531 	LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1532 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1533 		pmclog_process_map_out(po, pid, pkm->pm_address,
1534 		    pkm->pm_address + pkm->pm_size);
1535 
1536 	if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL)
1537 		return;
1538 
1539 	for (ri = 0; ri < md->pmd_npmc; ri++)
1540 		if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL &&
1541 		    PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1542 			pmclog_process_map_out(pm->pm_owner, pid,
1543 			    pkm->pm_address, pkm->pm_address + pkm->pm_size);
1544 }
1545 
1546 /*
1547  * Log mapping information about the kernel.
1548  */
1549 
1550 static void
1551 pmc_log_kernel_mappings(struct pmc *pm)
1552 {
1553 	struct pmc_owner *po;
1554 	struct pmckern_map_in *km, *kmbase;
1555 
1556 	sx_assert(&pmc_sx, SX_LOCKED);
1557 	KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)),
1558 	    ("[pmc,%d] non-sampling PMC (%p) desires mapping information",
1559 		__LINE__, (void *) pm));
1560 
1561 	po = pm->pm_owner;
1562 
1563 	if (po->po_flags & PMC_PO_INITIAL_MAPPINGS_DONE)
1564 		return;
1565 
1566 	/*
1567 	 * Log the current set of kernel modules.
1568 	 */
1569 	kmbase = linker_hwpmc_list_objects();
1570 	for (km = kmbase; km->pm_file != NULL; km++) {
1571 		PMCDBG(LOG,REG,1,"%s %p", (char *) km->pm_file,
1572 		    (void *) km->pm_address);
1573 		pmclog_process_map_in(po, (pid_t) -1, km->pm_address,
1574 		    km->pm_file);
1575 	}
1576 	FREE(kmbase, M_LINKER);
1577 
1578 	po->po_flags |= PMC_PO_INITIAL_MAPPINGS_DONE;
1579 }
1580 
1581 /*
1582  * Log the mappings for a single process.
1583  */
1584 
1585 static void
1586 pmc_log_process_mappings(struct pmc_owner *po, struct proc *p)
1587 {
1588 }
1589 
1590 /*
1591  * Log mappings for all processes in the system.
1592  */
1593 
1594 static void
1595 pmc_log_all_process_mappings(struct pmc_owner *po)
1596 {
1597 	struct proc *p, *top;
1598 
1599 	sx_assert(&pmc_sx, SX_XLOCKED);
1600 
1601 	if ((p = pfind(1)) == NULL)
1602 		panic("[pmc,%d] Cannot find init", __LINE__);
1603 
1604 	PROC_UNLOCK(p);
1605 
1606 	sx_slock(&proctree_lock);
1607 
1608 	top = p;
1609 
1610 	for (;;) {
1611 		pmc_log_process_mappings(po, p);
1612 		if (!LIST_EMPTY(&p->p_children))
1613 			p = LIST_FIRST(&p->p_children);
1614 		else for (;;) {
1615 			if (p == top)
1616 				goto done;
1617 			if (LIST_NEXT(p, p_sibling)) {
1618 				p = LIST_NEXT(p, p_sibling);
1619 				break;
1620 			}
1621 			p = p->p_pptr;
1622 		}
1623 	}
1624  done:
1625 	sx_sunlock(&proctree_lock);
1626 }
1627 
1628 /*
1629  * The 'hook' invoked from the kernel proper
1630  */
1631 
1632 
1633 #ifdef	DEBUG
1634 const char *pmc_hooknames[] = {
1635 	/* these strings correspond to PMC_FN_* in <sys/pmckern.h> */
1636 	"",
1637 	"EXEC",
1638 	"CSW-IN",
1639 	"CSW-OUT",
1640 	"SAMPLE",
1641 	"KLDLOAD",
1642 	"KLDUNLOAD",
1643 	"MMAP",
1644 	"MUNMAP",
1645 	"CALLCHAIN"
1646 };
1647 #endif
1648 
1649 static int
1650 pmc_hook_handler(struct thread *td, int function, void *arg)
1651 {
1652 
1653 	PMCDBG(MOD,PMH,1, "hook td=%p func=%d \"%s\" arg=%p", td, function,
1654 	    pmc_hooknames[function], arg);
1655 
1656 	switch (function)
1657 	{
1658 
1659 	/*
1660 	 * Process exec()
1661 	 */
1662 
1663 	case PMC_FN_PROCESS_EXEC:
1664 	{
1665 		char *fullpath, *freepath;
1666 		unsigned int ri;
1667 		int is_using_hwpmcs;
1668 		struct pmc *pm;
1669 		struct proc *p;
1670 		struct pmc_owner *po;
1671 		struct pmc_process *pp;
1672 		struct pmckern_procexec *pk;
1673 
1674 		sx_assert(&pmc_sx, SX_XLOCKED);
1675 
1676 		p = td->td_proc;
1677 		pmc_getfilename(p->p_textvp, &fullpath, &freepath);
1678 
1679 		pk = (struct pmckern_procexec *) arg;
1680 
1681 		/* Inform owners of SS mode PMCs of the exec event. */
1682 		LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1683 		    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1684 			    pmclog_process_procexec(po, PMC_ID_INVALID,
1685 				p->p_pid, pk->pm_entryaddr, fullpath);
1686 
1687 		PROC_LOCK(p);
1688 		is_using_hwpmcs = p->p_flag & P_HWPMC;
1689 		PROC_UNLOCK(p);
1690 
1691 		if (!is_using_hwpmcs) {
1692 			if (freepath)
1693 				FREE(freepath, M_TEMP);
1694 			break;
1695 		}
1696 
1697 		/*
1698 		 * PMCs are not inherited across an exec():  remove any
1699 		 * PMCs that this process is the owner of.
1700 		 */
1701 
1702 		if ((po = pmc_find_owner_descriptor(p)) != NULL) {
1703 			pmc_remove_owner(po);
1704 			pmc_destroy_owner_descriptor(po);
1705 		}
1706 
1707 		/*
1708 		 * If the process being exec'ed is not the target of any
1709 		 * PMC, we are done.
1710 		 */
1711 		if ((pp = pmc_find_process_descriptor(p, 0)) == NULL) {
1712 			if (freepath)
1713 				FREE(freepath, M_TEMP);
1714 			break;
1715 		}
1716 
1717 		/*
1718 		 * Log the exec event to all monitoring owners.  Skip
1719 		 * owners who have already recieved the event because
1720 		 * they had system sampling PMCs active.
1721 		 */
1722 		for (ri = 0; ri < md->pmd_npmc; ri++)
1723 			if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) {
1724 				po = pm->pm_owner;
1725 				if (po->po_sscount == 0 &&
1726 				    po->po_flags & PMC_PO_OWNS_LOGFILE)
1727 					pmclog_process_procexec(po, pm->pm_id,
1728 					    p->p_pid, pk->pm_entryaddr,
1729 					    fullpath);
1730 			}
1731 
1732 		if (freepath)
1733 			FREE(freepath, M_TEMP);
1734 
1735 
1736 		PMCDBG(PRC,EXC,1, "exec proc=%p (%d, %s) cred-changed=%d",
1737 		    p, p->p_pid, p->p_comm, pk->pm_credentialschanged);
1738 
1739 		if (pk->pm_credentialschanged == 0) /* no change */
1740 			break;
1741 
1742 		/*
1743 		 * If the newly exec()'ed process has a different credential
1744 		 * than before, allow it to be the target of a PMC only if
1745 		 * the PMC's owner has sufficient priviledge.
1746 		 */
1747 
1748 		for (ri = 0; ri < md->pmd_npmc; ri++)
1749 			if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL)
1750 				if (pmc_can_attach(pm, td->td_proc) != 0)
1751 					pmc_detach_one_process(td->td_proc,
1752 					    pm, PMC_FLAG_NONE);
1753 
1754 		KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt < (int) md->pmd_npmc,
1755 		    ("[pmc,%d] Illegal ref count %d on pp %p", __LINE__,
1756 			pp->pp_refcnt, pp));
1757 
1758 		/*
1759 		 * If this process is no longer the target of any
1760 		 * PMCs, we can remove the process entry and free
1761 		 * up space.
1762 		 */
1763 
1764 		if (pp->pp_refcnt == 0) {
1765 			pmc_remove_process_descriptor(pp);
1766 			FREE(pp, M_PMC);
1767 			break;
1768 		}
1769 
1770 	}
1771 	break;
1772 
1773 	case PMC_FN_CSW_IN:
1774 		pmc_process_csw_in(td);
1775 		break;
1776 
1777 	case PMC_FN_CSW_OUT:
1778 		pmc_process_csw_out(td);
1779 		break;
1780 
1781 	/*
1782 	 * Process accumulated PC samples.
1783 	 *
1784 	 * This function is expected to be called by hardclock() for
1785 	 * each CPU that has accumulated PC samples.
1786 	 *
1787 	 * This function is to be executed on the CPU whose samples
1788 	 * are being processed.
1789 	 */
1790 	case PMC_FN_DO_SAMPLES:
1791 
1792 		/*
1793 		 * Clear the cpu specific bit in the CPU mask before
1794 		 * do the rest of the processing.  If the NMI handler
1795 		 * gets invoked after the "atomic_clear_int()" call
1796 		 * below but before "pmc_process_samples()" gets
1797 		 * around to processing the interrupt, then we will
1798 		 * come back here at the next hardclock() tick (and
1799 		 * may find nothing to do if "pmc_process_samples()"
1800 		 * had already processed the interrupt).  We don't
1801 		 * lose the interrupt sample.
1802 		 */
1803 		atomic_clear_int(&pmc_cpumask, (1 << PCPU_GET(cpuid)));
1804 		pmc_process_samples(PCPU_GET(cpuid));
1805 		break;
1806 
1807 
1808 	case PMC_FN_KLD_LOAD:
1809 		sx_assert(&pmc_sx, SX_LOCKED);
1810 		pmc_process_kld_load((struct pmckern_map_in *) arg);
1811 		break;
1812 
1813 	case PMC_FN_KLD_UNLOAD:
1814 		sx_assert(&pmc_sx, SX_LOCKED);
1815 		pmc_process_kld_unload((struct pmckern_map_out *) arg);
1816 		break;
1817 
1818 	case PMC_FN_MMAP:
1819 		sx_assert(&pmc_sx, SX_LOCKED);
1820 		pmc_process_mmap(td, (struct pmckern_map_in *) arg);
1821 		break;
1822 
1823 	case PMC_FN_MUNMAP:
1824 		sx_assert(&pmc_sx, SX_LOCKED);
1825 		pmc_process_munmap(td, (struct pmckern_map_out *) arg);
1826 		break;
1827 
1828 	case PMC_FN_USER_CALLCHAIN:
1829 		/*
1830 		 * Record a call chain.
1831 		 */
1832 		pmc_capture_user_callchain(PCPU_GET(cpuid),
1833 		    (struct trapframe *) arg);
1834 		break;
1835 
1836 	default:
1837 #ifdef	DEBUG
1838 		KASSERT(0, ("[pmc,%d] unknown hook %d\n", __LINE__, function));
1839 #endif
1840 		break;
1841 
1842 	}
1843 
1844 	return 0;
1845 }
1846 
1847 /*
1848  * allocate a 'struct pmc_owner' descriptor in the owner hash table.
1849  */
1850 
1851 static struct pmc_owner *
1852 pmc_allocate_owner_descriptor(struct proc *p)
1853 {
1854 	uint32_t hindex;
1855 	struct pmc_owner *po;
1856 	struct pmc_ownerhash *poh;
1857 
1858 	hindex = PMC_HASH_PTR(p, pmc_ownerhashmask);
1859 	poh = &pmc_ownerhash[hindex];
1860 
1861 	/* allocate space for N pointers and one descriptor struct */
1862 	MALLOC(po, struct pmc_owner *, sizeof(struct pmc_owner),
1863 	    M_PMC, M_ZERO|M_WAITOK);
1864 
1865 	po->po_sscount = po->po_error = po->po_flags = 0;
1866 	po->po_file  = NULL;
1867 	po->po_owner = p;
1868 	po->po_kthread = NULL;
1869 	LIST_INIT(&po->po_pmcs);
1870 	LIST_INSERT_HEAD(poh, po, po_next); /* insert into hash table */
1871 
1872 	TAILQ_INIT(&po->po_logbuffers);
1873 	mtx_init(&po->po_mtx, "pmc-owner-mtx", "pmc-per-proc", MTX_SPIN);
1874 
1875 	PMCDBG(OWN,ALL,1, "allocate-owner proc=%p (%d, %s) pmc-owner=%p",
1876 	    p, p->p_pid, p->p_comm, po);
1877 
1878 	return po;
1879 }
1880 
1881 static void
1882 pmc_destroy_owner_descriptor(struct pmc_owner *po)
1883 {
1884 
1885 	PMCDBG(OWN,REL,1, "destroy-owner po=%p proc=%p (%d, %s)",
1886 	    po, po->po_owner, po->po_owner->p_pid, po->po_owner->p_comm);
1887 
1888 	mtx_destroy(&po->po_mtx);
1889 	FREE(po, M_PMC);
1890 }
1891 
1892 /*
1893  * find the descriptor corresponding to process 'p', adding or removing it
1894  * as specified by 'mode'.
1895  */
1896 
1897 static struct pmc_process *
1898 pmc_find_process_descriptor(struct proc *p, uint32_t mode)
1899 {
1900 	uint32_t hindex;
1901 	struct pmc_process *pp, *ppnew;
1902 	struct pmc_processhash *pph;
1903 
1904 	hindex = PMC_HASH_PTR(p, pmc_processhashmask);
1905 	pph = &pmc_processhash[hindex];
1906 
1907 	ppnew = NULL;
1908 
1909 	/*
1910 	 * Pre-allocate memory in the FIND_ALLOCATE case since we
1911 	 * cannot call malloc(9) once we hold a spin lock.
1912 	 */
1913 
1914 	if (mode & PMC_FLAG_ALLOCATE) {
1915 		/* allocate additional space for 'n' pmc pointers */
1916 		MALLOC(ppnew, struct pmc_process *,
1917 		    sizeof(struct pmc_process) + md->pmd_npmc *
1918 		    sizeof(struct pmc_targetstate), M_PMC, M_ZERO|M_WAITOK);
1919 	}
1920 
1921 	mtx_lock_spin(&pmc_processhash_mtx);
1922 	LIST_FOREACH(pp, pph, pp_next)
1923 	    if (pp->pp_proc == p)
1924 		    break;
1925 
1926 	if ((mode & PMC_FLAG_REMOVE) && pp != NULL)
1927 		LIST_REMOVE(pp, pp_next);
1928 
1929 	if ((mode & PMC_FLAG_ALLOCATE) && pp == NULL &&
1930 	    ppnew != NULL) {
1931 		ppnew->pp_proc = p;
1932 		LIST_INSERT_HEAD(pph, ppnew, pp_next);
1933 		pp = ppnew;
1934 		ppnew = NULL;
1935 	}
1936 	mtx_unlock_spin(&pmc_processhash_mtx);
1937 
1938 	if (pp != NULL && ppnew != NULL)
1939 		FREE(ppnew, M_PMC);
1940 
1941 	return pp;
1942 }
1943 
1944 /*
1945  * remove a process descriptor from the process hash table.
1946  */
1947 
1948 static void
1949 pmc_remove_process_descriptor(struct pmc_process *pp)
1950 {
1951 	KASSERT(pp->pp_refcnt == 0,
1952 	    ("[pmc,%d] Removing process descriptor %p with count %d",
1953 		__LINE__, pp, pp->pp_refcnt));
1954 
1955 	mtx_lock_spin(&pmc_processhash_mtx);
1956 	LIST_REMOVE(pp, pp_next);
1957 	mtx_unlock_spin(&pmc_processhash_mtx);
1958 }
1959 
1960 
1961 /*
1962  * find an owner descriptor corresponding to proc 'p'
1963  */
1964 
1965 static struct pmc_owner *
1966 pmc_find_owner_descriptor(struct proc *p)
1967 {
1968 	uint32_t hindex;
1969 	struct pmc_owner *po;
1970 	struct pmc_ownerhash *poh;
1971 
1972 	hindex = PMC_HASH_PTR(p, pmc_ownerhashmask);
1973 	poh = &pmc_ownerhash[hindex];
1974 
1975 	po = NULL;
1976 	LIST_FOREACH(po, poh, po_next)
1977 	    if (po->po_owner == p)
1978 		    break;
1979 
1980 	PMCDBG(OWN,FND,1, "find-owner proc=%p (%d, %s) hindex=0x%x -> "
1981 	    "pmc-owner=%p", p, p->p_pid, p->p_comm, hindex, po);
1982 
1983 	return po;
1984 }
1985 
1986 /*
1987  * pmc_allocate_pmc_descriptor
1988  *
1989  * Allocate a pmc descriptor and initialize its
1990  * fields.
1991  */
1992 
1993 static struct pmc *
1994 pmc_allocate_pmc_descriptor(void)
1995 {
1996 	struct pmc *pmc;
1997 
1998 	MALLOC(pmc, struct pmc *, sizeof(struct pmc), M_PMC, M_ZERO|M_WAITOK);
1999 
2000 	if (pmc != NULL) {
2001 		pmc->pm_owner = NULL;
2002 		LIST_INIT(&pmc->pm_targets);
2003 	}
2004 
2005 	PMCDBG(PMC,ALL,1, "allocate-pmc -> pmc=%p", pmc);
2006 
2007 	return pmc;
2008 }
2009 
2010 /*
2011  * Destroy a pmc descriptor.
2012  */
2013 
2014 static void
2015 pmc_destroy_pmc_descriptor(struct pmc *pm)
2016 {
2017 	(void) pm;
2018 
2019 #ifdef	DEBUG
2020 	KASSERT(pm->pm_state == PMC_STATE_DELETED ||
2021 	    pm->pm_state == PMC_STATE_FREE,
2022 	    ("[pmc,%d] destroying non-deleted PMC", __LINE__));
2023 	KASSERT(LIST_EMPTY(&pm->pm_targets),
2024 	    ("[pmc,%d] destroying pmc with targets", __LINE__));
2025 	KASSERT(pm->pm_owner == NULL,
2026 	    ("[pmc,%d] destroying pmc attached to an owner", __LINE__));
2027 	KASSERT(pm->pm_runcount == 0,
2028 	    ("[pmc,%d] pmc has non-zero run count %d", __LINE__,
2029 		pm->pm_runcount));
2030 #endif
2031 }
2032 
2033 static void
2034 pmc_wait_for_pmc_idle(struct pmc *pm)
2035 {
2036 #ifdef	DEBUG
2037 	volatile int maxloop;
2038 
2039 	maxloop = 100 * mp_ncpus;
2040 #endif
2041 
2042 	/*
2043 	 * Loop (with a forced context switch) till the PMC's runcount
2044 	 * comes down to zero.
2045 	 */
2046 	while (atomic_load_acq_32(&pm->pm_runcount) > 0) {
2047 #ifdef	DEBUG
2048 		maxloop--;
2049 		KASSERT(maxloop > 0,
2050 		    ("[pmc,%d] (ri%d, rc%d) waiting too long for "
2051 			"pmc to be free", __LINE__,
2052 			PMC_TO_ROWINDEX(pm), pm->pm_runcount));
2053 #endif
2054 		pmc_force_context_switch();
2055 	}
2056 }
2057 
2058 /*
2059  * This function does the following things:
2060  *
2061  *  - detaches the PMC from hardware
2062  *  - unlinks all target threads that were attached to it
2063  *  - removes the PMC from its owner's list
2064  *  - destroy's the PMC private mutex
2065  *
2066  * Once this function completes, the given pmc pointer can be safely
2067  * FREE'd by the caller.
2068  */
2069 
2070 static void
2071 pmc_release_pmc_descriptor(struct pmc *pm)
2072 {
2073 	u_int ri, cpu;
2074 	enum pmc_mode mode;
2075 	struct pmc_hw *phw;
2076 	struct pmc_owner *po;
2077 	struct pmc_process *pp;
2078 	struct pmc_target *ptgt, *tmp;
2079 	struct pmc_binding pb;
2080 
2081 	sx_assert(&pmc_sx, SX_XLOCKED);
2082 
2083 	KASSERT(pm, ("[pmc,%d] null pmc", __LINE__));
2084 
2085 	ri   = PMC_TO_ROWINDEX(pm);
2086 	mode = PMC_TO_MODE(pm);
2087 
2088 	PMCDBG(PMC,REL,1, "release-pmc pmc=%p ri=%d mode=%d", pm, ri,
2089 	    mode);
2090 
2091 	/*
2092 	 * First, we take the PMC off hardware.
2093 	 */
2094 	cpu = 0;
2095 	if (PMC_IS_SYSTEM_MODE(mode)) {
2096 
2097 		/*
2098 		 * A system mode PMC runs on a specific CPU.  Switch
2099 		 * to this CPU and turn hardware off.
2100 		 */
2101 		pmc_save_cpu_binding(&pb);
2102 
2103 		cpu = PMC_TO_CPU(pm);
2104 
2105 		pmc_select_cpu(cpu);
2106 
2107 		/* switch off non-stalled CPUs */
2108 		if (pm->pm_state == PMC_STATE_RUNNING &&
2109 		    pm->pm_stalled == 0) {
2110 
2111 			phw = pmc_pcpu[cpu]->pc_hwpmcs[ri];
2112 
2113 			KASSERT(phw->phw_pmc == pm,
2114 			    ("[pmc, %d] pmc ptr ri(%d) hw(%p) pm(%p)",
2115 				__LINE__, ri, phw->phw_pmc, pm));
2116 			PMCDBG(PMC,REL,2, "stopping cpu=%d ri=%d", cpu, ri);
2117 
2118 			critical_enter();
2119 			md->pmd_stop_pmc(cpu, ri);
2120 			critical_exit();
2121 		}
2122 
2123 		PMCDBG(PMC,REL,2, "decfg cpu=%d ri=%d", cpu, ri);
2124 
2125 		critical_enter();
2126 		md->pmd_config_pmc(cpu, ri, NULL);
2127 		critical_exit();
2128 
2129 		/* adjust the global and process count of SS mode PMCs */
2130 		if (mode == PMC_MODE_SS && pm->pm_state == PMC_STATE_RUNNING) {
2131 			po = pm->pm_owner;
2132 			po->po_sscount--;
2133 			if (po->po_sscount == 0) {
2134 				atomic_subtract_rel_int(&pmc_ss_count, 1);
2135 				LIST_REMOVE(po, po_ssnext);
2136 			}
2137 		}
2138 
2139 		pm->pm_state = PMC_STATE_DELETED;
2140 
2141 		pmc_restore_cpu_binding(&pb);
2142 
2143 		/*
2144 		 * We could have references to this PMC structure in
2145 		 * the per-cpu sample queues.  Wait for the queue to
2146 		 * drain.
2147 		 */
2148 		pmc_wait_for_pmc_idle(pm);
2149 
2150 	} else if (PMC_IS_VIRTUAL_MODE(mode)) {
2151 
2152 		/*
2153 		 * A virtual PMC could be running on multiple CPUs at
2154 		 * a given instant.
2155 		 *
2156 		 * By marking its state as DELETED, we ensure that
2157 		 * this PMC is never further scheduled on hardware.
2158 		 *
2159 		 * Then we wait till all CPUs are done with this PMC.
2160 		 */
2161 		pm->pm_state = PMC_STATE_DELETED;
2162 
2163 
2164 		/* Wait for the PMCs runcount to come to zero. */
2165 		pmc_wait_for_pmc_idle(pm);
2166 
2167 		/*
2168 		 * At this point the PMC is off all CPUs and cannot be
2169 		 * freshly scheduled onto a CPU.  It is now safe to
2170 		 * unlink all targets from this PMC.  If a
2171 		 * process-record's refcount falls to zero, we remove
2172 		 * it from the hash table.  The module-wide SX lock
2173 		 * protects us from races.
2174 		 */
2175 		LIST_FOREACH_SAFE(ptgt, &pm->pm_targets, pt_next, tmp) {
2176 			pp = ptgt->pt_process;
2177 			pmc_unlink_target_process(pm, pp); /* frees 'ptgt' */
2178 
2179 			PMCDBG(PMC,REL,3, "pp->refcnt=%d", pp->pp_refcnt);
2180 
2181 			/*
2182 			 * If the target process record shows that no
2183 			 * PMCs are attached to it, reclaim its space.
2184 			 */
2185 
2186 			if (pp->pp_refcnt == 0) {
2187 				pmc_remove_process_descriptor(pp);
2188 				FREE(pp, M_PMC);
2189 			}
2190 		}
2191 
2192 		cpu = curthread->td_oncpu; /* setup cpu for pmd_release() */
2193 
2194 	}
2195 
2196 	/*
2197 	 * Release any MD resources
2198 	 */
2199 
2200 	(void) md->pmd_release_pmc(cpu, ri, pm);
2201 
2202 	/*
2203 	 * Update row disposition
2204 	 */
2205 
2206 	if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm)))
2207 		PMC_UNMARK_ROW_STANDALONE(ri);
2208 	else
2209 		PMC_UNMARK_ROW_THREAD(ri);
2210 
2211 	/* unlink from the owner's list */
2212 	if (pm->pm_owner) {
2213 		LIST_REMOVE(pm, pm_next);
2214 		pm->pm_owner = NULL;
2215 	}
2216 
2217 	pmc_destroy_pmc_descriptor(pm);
2218 }
2219 
2220 /*
2221  * Register an owner and a pmc.
2222  */
2223 
2224 static int
2225 pmc_register_owner(struct proc *p, struct pmc *pmc)
2226 {
2227 	struct pmc_owner *po;
2228 
2229 	sx_assert(&pmc_sx, SX_XLOCKED);
2230 
2231 	if ((po = pmc_find_owner_descriptor(p)) == NULL)
2232 		if ((po = pmc_allocate_owner_descriptor(p)) == NULL)
2233 			return ENOMEM;
2234 
2235 	KASSERT(pmc->pm_owner == NULL,
2236 	    ("[pmc,%d] attempting to own an initialized PMC", __LINE__));
2237 	pmc->pm_owner  = po;
2238 
2239 	LIST_INSERT_HEAD(&po->po_pmcs, pmc, pm_next);
2240 
2241 	PROC_LOCK(p);
2242 	p->p_flag |= P_HWPMC;
2243 	PROC_UNLOCK(p);
2244 
2245 	if (po->po_flags & PMC_PO_OWNS_LOGFILE)
2246 		pmclog_process_pmcallocate(pmc);
2247 
2248 	PMCDBG(PMC,REG,1, "register-owner pmc-owner=%p pmc=%p",
2249 	    po, pmc);
2250 
2251 	return 0;
2252 }
2253 
2254 /*
2255  * Return the current row disposition:
2256  * == 0 => FREE
2257  *  > 0 => PROCESS MODE
2258  *  < 0 => SYSTEM MODE
2259  */
2260 
2261 int
2262 pmc_getrowdisp(int ri)
2263 {
2264 	return pmc_pmcdisp[ri];
2265 }
2266 
2267 /*
2268  * Check if a PMC at row index 'ri' can be allocated to the current
2269  * process.
2270  *
2271  * Allocation can fail if:
2272  *   - the current process is already being profiled by a PMC at index 'ri',
2273  *     attached to it via OP_PMCATTACH.
2274  *   - the current process has already allocated a PMC at index 'ri'
2275  *     via OP_ALLOCATE.
2276  */
2277 
2278 static int
2279 pmc_can_allocate_rowindex(struct proc *p, unsigned int ri, int cpu)
2280 {
2281 	enum pmc_mode mode;
2282 	struct pmc *pm;
2283 	struct pmc_owner *po;
2284 	struct pmc_process *pp;
2285 
2286 	PMCDBG(PMC,ALR,1, "can-allocate-rowindex proc=%p (%d, %s) ri=%d "
2287 	    "cpu=%d", p, p->p_pid, p->p_comm, ri, cpu);
2288 
2289 	/*
2290 	 * We shouldn't have already allocated a process-mode PMC at
2291 	 * row index 'ri'.
2292 	 *
2293 	 * We shouldn't have allocated a system-wide PMC on the same
2294 	 * CPU and same RI.
2295 	 */
2296 	if ((po = pmc_find_owner_descriptor(p)) != NULL)
2297 		LIST_FOREACH(pm, &po->po_pmcs, pm_next) {
2298 		    if (PMC_TO_ROWINDEX(pm) == ri) {
2299 			    mode = PMC_TO_MODE(pm);
2300 			    if (PMC_IS_VIRTUAL_MODE(mode))
2301 				    return EEXIST;
2302 			    if (PMC_IS_SYSTEM_MODE(mode) &&
2303 				(int) PMC_TO_CPU(pm) == cpu)
2304 				    return EEXIST;
2305 		    }
2306 	        }
2307 
2308 	/*
2309 	 * We also shouldn't be the target of any PMC at this index
2310 	 * since otherwise a PMC_ATTACH to ourselves will fail.
2311 	 */
2312 	if ((pp = pmc_find_process_descriptor(p, 0)) != NULL)
2313 		if (pp->pp_pmcs[ri].pp_pmc)
2314 			return EEXIST;
2315 
2316 	PMCDBG(PMC,ALR,2, "can-allocate-rowindex proc=%p (%d, %s) ri=%d ok",
2317 	    p, p->p_pid, p->p_comm, ri);
2318 
2319 	return 0;
2320 }
2321 
2322 /*
2323  * Check if a given PMC at row index 'ri' can be currently used in
2324  * mode 'mode'.
2325  */
2326 
2327 static int
2328 pmc_can_allocate_row(int ri, enum pmc_mode mode)
2329 {
2330 	enum pmc_disp	disp;
2331 
2332 	sx_assert(&pmc_sx, SX_XLOCKED);
2333 
2334 	PMCDBG(PMC,ALR,1, "can-allocate-row ri=%d mode=%d", ri, mode);
2335 
2336 	if (PMC_IS_SYSTEM_MODE(mode))
2337 		disp = PMC_DISP_STANDALONE;
2338 	else
2339 		disp = PMC_DISP_THREAD;
2340 
2341 	/*
2342 	 * check disposition for PMC row 'ri':
2343 	 *
2344 	 * Expected disposition		Row-disposition		Result
2345 	 *
2346 	 * STANDALONE			STANDALONE or FREE	proceed
2347 	 * STANDALONE			THREAD			fail
2348 	 * THREAD			THREAD or FREE		proceed
2349 	 * THREAD			STANDALONE		fail
2350 	 */
2351 
2352 	if (!PMC_ROW_DISP_IS_FREE(ri) &&
2353 	    !(disp == PMC_DISP_THREAD && PMC_ROW_DISP_IS_THREAD(ri)) &&
2354 	    !(disp == PMC_DISP_STANDALONE && PMC_ROW_DISP_IS_STANDALONE(ri)))
2355 		return EBUSY;
2356 
2357 	/*
2358 	 * All OK
2359 	 */
2360 
2361 	PMCDBG(PMC,ALR,2, "can-allocate-row ri=%d mode=%d ok", ri, mode);
2362 
2363 	return 0;
2364 
2365 }
2366 
2367 /*
2368  * Find a PMC descriptor with user handle 'pmcid' for thread 'td'.
2369  */
2370 
2371 static struct pmc *
2372 pmc_find_pmc_descriptor_in_process(struct pmc_owner *po, pmc_id_t pmcid)
2373 {
2374 	struct pmc *pm;
2375 
2376 	KASSERT(PMC_ID_TO_ROWINDEX(pmcid) < md->pmd_npmc,
2377 	    ("[pmc,%d] Illegal pmc index %d (max %d)", __LINE__,
2378 		PMC_ID_TO_ROWINDEX(pmcid), md->pmd_npmc));
2379 
2380 	LIST_FOREACH(pm, &po->po_pmcs, pm_next)
2381 	    if (pm->pm_id == pmcid)
2382 		    return pm;
2383 
2384 	return NULL;
2385 }
2386 
2387 static int
2388 pmc_find_pmc(pmc_id_t pmcid, struct pmc **pmc)
2389 {
2390 
2391 	struct pmc *pm;
2392 	struct pmc_owner *po;
2393 
2394 	PMCDBG(PMC,FND,1, "find-pmc id=%d", pmcid);
2395 
2396 	if ((po = pmc_find_owner_descriptor(curthread->td_proc)) == NULL)
2397 		return ESRCH;
2398 
2399 	if ((pm = pmc_find_pmc_descriptor_in_process(po, pmcid)) == NULL)
2400 		return EINVAL;
2401 
2402 	PMCDBG(PMC,FND,2, "find-pmc id=%d -> pmc=%p", pmcid, pm);
2403 
2404 	*pmc = pm;
2405 	return 0;
2406 }
2407 
2408 /*
2409  * Start a PMC.
2410  */
2411 
2412 static int
2413 pmc_start(struct pmc *pm)
2414 {
2415 	int error, cpu, ri;
2416 	enum pmc_mode mode;
2417 	struct pmc_owner *po;
2418 	struct pmc_binding pb;
2419 
2420 	KASSERT(pm != NULL,
2421 	    ("[pmc,%d] null pm", __LINE__));
2422 
2423 	mode = PMC_TO_MODE(pm);
2424 	ri   = PMC_TO_ROWINDEX(pm);
2425 	error = 0;
2426 
2427 	PMCDBG(PMC,OPS,1, "start pmc=%p mode=%d ri=%d", pm, mode, ri);
2428 
2429 	po = pm->pm_owner;
2430 
2431 	/*
2432 	 * Disallow PMCSTART if a logfile is required but has not been
2433 	 * configured yet.
2434 	 */
2435 	if ((pm->pm_flags & PMC_F_NEEDS_LOGFILE) &&
2436 	    (po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
2437 		return EDOOFUS;	/* programming error */
2438 
2439 	/*
2440 	 * If this is a sampling mode PMC, log mapping information for
2441 	 * the kernel modules that are currently loaded.
2442 	 */
2443 	if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
2444 	    pmc_log_kernel_mappings(pm);
2445 
2446 	if (PMC_IS_VIRTUAL_MODE(mode)) {
2447 
2448 		/*
2449 		 * If a PMCATTACH has never been done on this PMC,
2450 		 * attach it to its owner process.
2451 		 */
2452 
2453 		if (LIST_EMPTY(&pm->pm_targets))
2454 			error = (pm->pm_flags & PMC_F_ATTACH_DONE) ? ESRCH :
2455 			    pmc_attach_process(po->po_owner, pm);
2456 
2457 		/*
2458 		 * If the PMC is attached to its owner, then force a context
2459 		 * switch to ensure that the MD state gets set correctly.
2460 		 */
2461 
2462 		if (error == 0) {
2463 			pm->pm_state = PMC_STATE_RUNNING;
2464 			if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER)
2465 				pmc_force_context_switch();
2466 		}
2467 
2468 		return error;
2469 	}
2470 
2471 
2472 	/*
2473 	 * A system-wide PMC.
2474 	 *
2475 	 * Add the owner to the global list if this is a system-wide
2476 	 * sampling PMC.
2477 	 */
2478 
2479 	if (mode == PMC_MODE_SS) {
2480 		if (po->po_sscount == 0) {
2481 			LIST_INSERT_HEAD(&pmc_ss_owners, po, po_ssnext);
2482 			atomic_add_rel_int(&pmc_ss_count, 1);
2483 			PMCDBG(PMC,OPS,1, "po=%p in global list", po);
2484 		}
2485 		po->po_sscount++;
2486 	}
2487 
2488 	/* Log mapping information for all processes in the system. */
2489 	pmc_log_all_process_mappings(po);
2490 
2491 	/*
2492 	 * Move to the CPU associated with this
2493 	 * PMC, and start the hardware.
2494 	 */
2495 
2496 	pmc_save_cpu_binding(&pb);
2497 
2498 	cpu = PMC_TO_CPU(pm);
2499 
2500 	if (pmc_cpu_is_disabled(cpu))
2501 		return ENXIO;
2502 
2503 	pmc_select_cpu(cpu);
2504 
2505 	/*
2506 	 * global PMCs are configured at allocation time
2507 	 * so write out the initial value and start the PMC.
2508 	 */
2509 
2510 	pm->pm_state = PMC_STATE_RUNNING;
2511 
2512 	critical_enter();
2513 	if ((error = md->pmd_write_pmc(cpu, ri,
2514 		 PMC_IS_SAMPLING_MODE(mode) ?
2515 		 pm->pm_sc.pm_reloadcount :
2516 		 pm->pm_sc.pm_initial)) == 0)
2517 		error = md->pmd_start_pmc(cpu, ri);
2518 	critical_exit();
2519 
2520 	pmc_restore_cpu_binding(&pb);
2521 
2522 	return error;
2523 }
2524 
2525 /*
2526  * Stop a PMC.
2527  */
2528 
2529 static int
2530 pmc_stop(struct pmc *pm)
2531 {
2532 	int cpu, error, ri;
2533 	struct pmc_owner *po;
2534 	struct pmc_binding pb;
2535 
2536 	KASSERT(pm != NULL, ("[pmc,%d] null pmc", __LINE__));
2537 
2538 	PMCDBG(PMC,OPS,1, "stop pmc=%p mode=%d ri=%d", pm,
2539 	    PMC_TO_MODE(pm), PMC_TO_ROWINDEX(pm));
2540 
2541 	pm->pm_state = PMC_STATE_STOPPED;
2542 
2543 	/*
2544 	 * If the PMC is a virtual mode one, changing the state to
2545 	 * non-RUNNING is enough to ensure that the PMC never gets
2546 	 * scheduled.
2547 	 *
2548 	 * If this PMC is current running on a CPU, then it will
2549 	 * handled correctly at the time its target process is context
2550 	 * switched out.
2551 	 */
2552 
2553 	if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)))
2554 		return 0;
2555 
2556 	/*
2557 	 * A system-mode PMC.  Move to the CPU associated with
2558 	 * this PMC, and stop the hardware.  We update the
2559 	 * 'initial count' so that a subsequent PMCSTART will
2560 	 * resume counting from the current hardware count.
2561 	 */
2562 
2563 	pmc_save_cpu_binding(&pb);
2564 
2565 	cpu = PMC_TO_CPU(pm);
2566 
2567 	KASSERT(cpu >= 0 && cpu < mp_ncpus,
2568 	    ("[pmc,%d] illegal cpu=%d", __LINE__, cpu));
2569 
2570 	if (pmc_cpu_is_disabled(cpu))
2571 		return ENXIO;
2572 
2573 	pmc_select_cpu(cpu);
2574 
2575 	ri = PMC_TO_ROWINDEX(pm);
2576 
2577 	critical_enter();
2578 	if ((error = md->pmd_stop_pmc(cpu, ri)) == 0)
2579 		error = md->pmd_read_pmc(cpu, ri, &pm->pm_sc.pm_initial);
2580 	critical_exit();
2581 
2582 	pmc_restore_cpu_binding(&pb);
2583 
2584 	po = pm->pm_owner;
2585 
2586 	/* remove this owner from the global list of SS PMC owners */
2587 	if (PMC_TO_MODE(pm) == PMC_MODE_SS) {
2588 		po->po_sscount--;
2589 		if (po->po_sscount == 0) {
2590 			atomic_subtract_rel_int(&pmc_ss_count, 1);
2591 			LIST_REMOVE(po, po_ssnext);
2592 			PMCDBG(PMC,OPS,2,"po=%p removed from global list", po);
2593 		}
2594 	}
2595 
2596 	return error;
2597 }
2598 
2599 
2600 #ifdef	DEBUG
2601 static const char *pmc_op_to_name[] = {
2602 #undef	__PMC_OP
2603 #define	__PMC_OP(N, D)	#N ,
2604 	__PMC_OPS()
2605 	NULL
2606 };
2607 #endif
2608 
2609 /*
2610  * The syscall interface
2611  */
2612 
2613 #define	PMC_GET_SX_XLOCK(...) do {		\
2614 	sx_xlock(&pmc_sx);			\
2615 	if (pmc_hook == NULL) {			\
2616 		sx_xunlock(&pmc_sx);		\
2617 		return __VA_ARGS__;		\
2618 	}					\
2619 } while (0)
2620 
2621 #define	PMC_DOWNGRADE_SX() do {			\
2622 	sx_downgrade(&pmc_sx);			\
2623 	is_sx_downgraded = 1;			\
2624 } while (0)
2625 
2626 static int
2627 pmc_syscall_handler(struct thread *td, void *syscall_args)
2628 {
2629 	int error, is_sx_downgraded, op;
2630 	struct pmc_syscall_args *c;
2631 	void *arg;
2632 
2633 	PMC_GET_SX_XLOCK(ENOSYS);
2634 
2635 	DROP_GIANT();
2636 
2637 	is_sx_downgraded = 0;
2638 
2639 	c = (struct pmc_syscall_args *) syscall_args;
2640 
2641 	op = c->pmop_code;
2642 	arg = c->pmop_data;
2643 
2644 	PMCDBG(MOD,PMS,1, "syscall op=%d \"%s\" arg=%p", op,
2645 	    pmc_op_to_name[op], arg);
2646 
2647 	error = 0;
2648 	atomic_add_int(&pmc_stats.pm_syscalls, 1);
2649 
2650 	switch(op)
2651 	{
2652 
2653 
2654 	/*
2655 	 * Configure a log file.
2656 	 *
2657 	 * XXX This OP will be reworked.
2658 	 */
2659 
2660 	case PMC_OP_CONFIGURELOG:
2661 	{
2662 		struct proc *p;
2663 		struct pmc *pm;
2664 		struct pmc_owner *po;
2665 		struct pmc_op_configurelog cl;
2666 
2667 		sx_assert(&pmc_sx, SX_XLOCKED);
2668 
2669 		if ((error = copyin(arg, &cl, sizeof(cl))) != 0)
2670 			break;
2671 
2672 		/* mark this process as owning a log file */
2673 		p = td->td_proc;
2674 		if ((po = pmc_find_owner_descriptor(p)) == NULL)
2675 			if ((po = pmc_allocate_owner_descriptor(p)) == NULL) {
2676 				error = ENOMEM;
2677 				break;
2678 			}
2679 
2680 		/*
2681 		 * If a valid fd was passed in, try to configure that,
2682 		 * otherwise if 'fd' was less than zero and there was
2683 		 * a log file configured, flush its buffers and
2684 		 * de-configure it.
2685 		 */
2686 		if (cl.pm_logfd >= 0)
2687 			error = pmclog_configure_log(po, cl.pm_logfd);
2688 		else if (po->po_flags & PMC_PO_OWNS_LOGFILE) {
2689 			pmclog_process_closelog(po);
2690 			error = pmclog_flush(po);
2691 			if (error == 0) {
2692 				LIST_FOREACH(pm, &po->po_pmcs, pm_next)
2693 				    if (pm->pm_flags & PMC_F_NEEDS_LOGFILE &&
2694 					pm->pm_state == PMC_STATE_RUNNING)
2695 					    pmc_stop(pm);
2696 				error = pmclog_deconfigure_log(po);
2697 			}
2698 		} else
2699 			error = EINVAL;
2700 
2701 		if (error)
2702 			break;
2703 	}
2704 	break;
2705 
2706 
2707 	/*
2708 	 * Flush a log file.
2709 	 */
2710 
2711 	case PMC_OP_FLUSHLOG:
2712 	{
2713 		struct pmc_owner *po;
2714 
2715 		sx_assert(&pmc_sx, SX_XLOCKED);
2716 
2717 		if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) {
2718 			error = EINVAL;
2719 			break;
2720 		}
2721 
2722 		error = pmclog_flush(po);
2723 	}
2724 	break;
2725 
2726 	/*
2727 	 * Retrieve hardware configuration.
2728 	 */
2729 
2730 	case PMC_OP_GETCPUINFO:	/* CPU information */
2731 	{
2732 		struct pmc_op_getcpuinfo gci;
2733 
2734 		gci.pm_cputype = md->pmd_cputype;
2735 		gci.pm_ncpu    = mp_ncpus;
2736 		gci.pm_npmc    = md->pmd_npmc;
2737 		gci.pm_nclass  = md->pmd_nclass;
2738 		bcopy(md->pmd_classes, &gci.pm_classes,
2739 		    sizeof(gci.pm_classes));
2740 		error = copyout(&gci, arg, sizeof(gci));
2741 	}
2742 	break;
2743 
2744 
2745 	/*
2746 	 * Get module statistics
2747 	 */
2748 
2749 	case PMC_OP_GETDRIVERSTATS:
2750 	{
2751 		struct pmc_op_getdriverstats gms;
2752 
2753 		bcopy(&pmc_stats, &gms, sizeof(gms));
2754 		error = copyout(&gms, arg, sizeof(gms));
2755 	}
2756 	break;
2757 
2758 
2759 	/*
2760 	 * Retrieve module version number
2761 	 */
2762 
2763 	case PMC_OP_GETMODULEVERSION:
2764 	{
2765 		uint32_t cv, modv;
2766 
2767 		/* retrieve the client's idea of the ABI version */
2768 		if ((error = copyin(arg, &cv, sizeof(uint32_t))) != 0)
2769 			break;
2770 		/* don't service clients newer than our driver */
2771 		modv = PMC_VERSION;
2772 		if ((cv & 0xFFFF0000) > (modv & 0xFFFF0000)) {
2773 			error = EPROGMISMATCH;
2774 			break;
2775 		}
2776 		error = copyout(&modv, arg, sizeof(int));
2777 	}
2778 	break;
2779 
2780 
2781 	/*
2782 	 * Retrieve the state of all the PMCs on a given
2783 	 * CPU.
2784 	 */
2785 
2786 	case PMC_OP_GETPMCINFO:
2787 	{
2788 		uint32_t cpu, n, npmc;
2789 		size_t pmcinfo_size;
2790 		struct pmc *pm;
2791 		struct pmc_info *p, *pmcinfo;
2792 		struct pmc_op_getpmcinfo *gpi;
2793 		struct pmc_owner *po;
2794 		struct pmc_binding pb;
2795 
2796 		PMC_DOWNGRADE_SX();
2797 
2798 		gpi = (struct pmc_op_getpmcinfo *) arg;
2799 
2800 		if ((error = copyin(&gpi->pm_cpu, &cpu, sizeof(cpu))) != 0)
2801 			break;
2802 
2803 		if (cpu >= (unsigned int) mp_ncpus) {
2804 			error = EINVAL;
2805 			break;
2806 		}
2807 
2808 		if (pmc_cpu_is_disabled(cpu)) {
2809 			error = ENXIO;
2810 			break;
2811 		}
2812 
2813 		/* switch to CPU 'cpu' */
2814 		pmc_save_cpu_binding(&pb);
2815 		pmc_select_cpu(cpu);
2816 
2817 		npmc = md->pmd_npmc;
2818 
2819 		pmcinfo_size = npmc * sizeof(struct pmc_info);
2820 		MALLOC(pmcinfo, struct pmc_info *, pmcinfo_size, M_PMC,
2821 		    M_WAITOK);
2822 
2823 		p = pmcinfo;
2824 
2825 		for (n = 0; n < md->pmd_npmc; n++, p++) {
2826 
2827 			if ((error = md->pmd_describe(cpu, n, p, &pm)) != 0)
2828 				break;
2829 
2830 			if (PMC_ROW_DISP_IS_STANDALONE(n))
2831 				p->pm_rowdisp = PMC_DISP_STANDALONE;
2832 			else if (PMC_ROW_DISP_IS_THREAD(n))
2833 				p->pm_rowdisp = PMC_DISP_THREAD;
2834 			else
2835 				p->pm_rowdisp = PMC_DISP_FREE;
2836 
2837 			p->pm_ownerpid = -1;
2838 
2839 			if (pm == NULL)	/* no PMC associated */
2840 				continue;
2841 
2842 			po = pm->pm_owner;
2843 
2844 			KASSERT(po->po_owner != NULL,
2845 			    ("[pmc,%d] pmc_owner had a null proc pointer",
2846 				__LINE__));
2847 
2848 			p->pm_ownerpid = po->po_owner->p_pid;
2849 			p->pm_mode     = PMC_TO_MODE(pm);
2850 			p->pm_event    = pm->pm_event;
2851 			p->pm_flags    = pm->pm_flags;
2852 
2853 			if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
2854 				p->pm_reloadcount =
2855 				    pm->pm_sc.pm_reloadcount;
2856 		}
2857 
2858 		pmc_restore_cpu_binding(&pb);
2859 
2860 		/* now copy out the PMC info collected */
2861 		if (error == 0)
2862 			error = copyout(pmcinfo, &gpi->pm_pmcs, pmcinfo_size);
2863 
2864 		FREE(pmcinfo, M_PMC);
2865 	}
2866 	break;
2867 
2868 
2869 	/*
2870 	 * Set the administrative state of a PMC.  I.e. whether
2871 	 * the PMC is to be used or not.
2872 	 */
2873 
2874 	case PMC_OP_PMCADMIN:
2875 	{
2876 		int cpu, ri;
2877 		enum pmc_state request;
2878 		struct pmc_cpu *pc;
2879 		struct pmc_hw *phw;
2880 		struct pmc_op_pmcadmin pma;
2881 		struct pmc_binding pb;
2882 
2883 		sx_assert(&pmc_sx, SX_XLOCKED);
2884 
2885 		KASSERT(td == curthread,
2886 		    ("[pmc,%d] td != curthread", __LINE__));
2887 
2888 		error = priv_check(td, PRIV_PMC_MANAGE);
2889 		if (error)
2890 			break;
2891 
2892 		if ((error = copyin(arg, &pma, sizeof(pma))) != 0)
2893 			break;
2894 
2895 		cpu = pma.pm_cpu;
2896 
2897 		if (cpu < 0 || cpu >= mp_ncpus) {
2898 			error = EINVAL;
2899 			break;
2900 		}
2901 
2902 		if (pmc_cpu_is_disabled(cpu)) {
2903 			error = ENXIO;
2904 			break;
2905 		}
2906 
2907 		request = pma.pm_state;
2908 
2909 		if (request != PMC_STATE_DISABLED &&
2910 		    request != PMC_STATE_FREE) {
2911 			error = EINVAL;
2912 			break;
2913 		}
2914 
2915 		ri = pma.pm_pmc; /* pmc id == row index */
2916 		if (ri < 0 || ri >= (int) md->pmd_npmc) {
2917 			error = EINVAL;
2918 			break;
2919 		}
2920 
2921 		/*
2922 		 * We can't disable a PMC with a row-index allocated
2923 		 * for process virtual PMCs.
2924 		 */
2925 
2926 		if (PMC_ROW_DISP_IS_THREAD(ri) &&
2927 		    request == PMC_STATE_DISABLED) {
2928 			error = EBUSY;
2929 			break;
2930 		}
2931 
2932 		/*
2933 		 * otherwise, this PMC on this CPU is either free or
2934 		 * in system-wide mode.
2935 		 */
2936 
2937 		pmc_save_cpu_binding(&pb);
2938 		pmc_select_cpu(cpu);
2939 
2940 		pc  = pmc_pcpu[cpu];
2941 		phw = pc->pc_hwpmcs[ri];
2942 
2943 		/*
2944 		 * XXX do we need some kind of 'forced' disable?
2945 		 */
2946 
2947 		if (phw->phw_pmc == NULL) {
2948 			if (request == PMC_STATE_DISABLED &&
2949 			    (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED)) {
2950 				phw->phw_state &= ~PMC_PHW_FLAG_IS_ENABLED;
2951 				PMC_MARK_ROW_STANDALONE(ri);
2952 			} else if (request == PMC_STATE_FREE &&
2953 			    (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0) {
2954 				phw->phw_state |=  PMC_PHW_FLAG_IS_ENABLED;
2955 				PMC_UNMARK_ROW_STANDALONE(ri);
2956 			}
2957 			/* other cases are a no-op */
2958 		} else
2959 			error = EBUSY;
2960 
2961 		pmc_restore_cpu_binding(&pb);
2962 	}
2963 	break;
2964 
2965 
2966 	/*
2967 	 * Allocate a PMC.
2968 	 */
2969 
2970 	case PMC_OP_PMCALLOCATE:
2971 	{
2972 		uint32_t caps;
2973 		u_int cpu;
2974 		int n;
2975 		enum pmc_mode mode;
2976 		struct pmc *pmc;
2977 		struct pmc_hw *phw;
2978 		struct pmc_op_pmcallocate pa;
2979 		struct pmc_binding pb;
2980 
2981 		if ((error = copyin(arg, &pa, sizeof(pa))) != 0)
2982 			break;
2983 
2984 		caps = pa.pm_caps;
2985 		mode = pa.pm_mode;
2986 		cpu  = pa.pm_cpu;
2987 
2988 		if ((mode != PMC_MODE_SS  &&  mode != PMC_MODE_SC  &&
2989 		     mode != PMC_MODE_TS  &&  mode != PMC_MODE_TC) ||
2990 		    (cpu != (u_int) PMC_CPU_ANY && cpu >= (u_int) mp_ncpus)) {
2991 			error = EINVAL;
2992 			break;
2993 		}
2994 
2995 		/*
2996 		 * Virtual PMCs should only ask for a default CPU.
2997 		 * System mode PMCs need to specify a non-default CPU.
2998 		 */
2999 
3000 		if ((PMC_IS_VIRTUAL_MODE(mode) && cpu != (u_int) PMC_CPU_ANY) ||
3001 		    (PMC_IS_SYSTEM_MODE(mode) && cpu == (u_int) PMC_CPU_ANY)) {
3002 			error = EINVAL;
3003 			break;
3004 		}
3005 
3006 		/*
3007 		 * Check that a disabled CPU is not being asked for.
3008 		 */
3009 
3010 		if (PMC_IS_SYSTEM_MODE(mode) && pmc_cpu_is_disabled(cpu)) {
3011 			error = ENXIO;
3012 			break;
3013 		}
3014 
3015 		/*
3016 		 * Refuse an allocation for a system-wide PMC if this
3017 		 * process has been jailed, or if this process lacks
3018 		 * super-user credentials and the sysctl tunable
3019 		 * 'security.bsd.unprivileged_syspmcs' is zero.
3020 		 */
3021 
3022 		if (PMC_IS_SYSTEM_MODE(mode)) {
3023 			if (jailed(curthread->td_ucred)) {
3024 				error = EPERM;
3025 				break;
3026 			}
3027 			if (!pmc_unprivileged_syspmcs) {
3028 				error = priv_check(curthread,
3029 				    PRIV_PMC_SYSTEM);
3030 				if (error)
3031 					break;
3032 			}
3033 		}
3034 
3035 		if (error)
3036 			break;
3037 
3038 		/*
3039 		 * Look for valid values for 'pm_flags'
3040 		 */
3041 
3042 		if ((pa.pm_flags & ~(PMC_F_DESCENDANTS | PMC_F_LOG_PROCCSW |
3043 		    PMC_F_LOG_PROCEXIT | PMC_F_CALLCHAIN)) != 0) {
3044 			error = EINVAL;
3045 			break;
3046 		}
3047 
3048 		/* process logging options are not allowed for system PMCs */
3049 		if (PMC_IS_SYSTEM_MODE(mode) && (pa.pm_flags &
3050 		    (PMC_F_LOG_PROCCSW | PMC_F_LOG_PROCEXIT))) {
3051 			error = EINVAL;
3052 			break;
3053 		}
3054 
3055 		/*
3056 		 * All sampling mode PMCs need to be able to interrupt the
3057 		 * CPU.
3058 		 */
3059 		if (PMC_IS_SAMPLING_MODE(mode))
3060 			caps |= PMC_CAP_INTERRUPT;
3061 
3062 		/* A valid class specifier should have been passed in. */
3063 		for (n = 0; n < md->pmd_nclass; n++)
3064 			if (md->pmd_classes[n].pm_class == pa.pm_class)
3065 				break;
3066 		if (n == md->pmd_nclass) {
3067 			error = EINVAL;
3068 			break;
3069 		}
3070 
3071 		/* The requested PMC capabilities should be feasible. */
3072 		if ((md->pmd_classes[n].pm_caps & caps) != caps) {
3073 			error = EOPNOTSUPP;
3074 			break;
3075 		}
3076 
3077 		PMCDBG(PMC,ALL,2, "event=%d caps=0x%x mode=%d cpu=%d",
3078 		    pa.pm_ev, caps, mode, cpu);
3079 
3080 		pmc = pmc_allocate_pmc_descriptor();
3081 		pmc->pm_id    = PMC_ID_MAKE_ID(cpu,pa.pm_mode,pa.pm_class,
3082 		    PMC_ID_INVALID);
3083 		pmc->pm_event = pa.pm_ev;
3084 		pmc->pm_state = PMC_STATE_FREE;
3085 		pmc->pm_caps  = caps;
3086 		pmc->pm_flags = pa.pm_flags;
3087 
3088 		/* switch thread to CPU 'cpu' */
3089 		pmc_save_cpu_binding(&pb);
3090 
3091 #define	PMC_IS_SHAREABLE_PMC(cpu, n)				\
3092 	(pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_state &		\
3093 	 PMC_PHW_FLAG_IS_SHAREABLE)
3094 #define	PMC_IS_UNALLOCATED(cpu, n)				\
3095 	(pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_pmc == NULL)
3096 
3097 		if (PMC_IS_SYSTEM_MODE(mode)) {
3098 			pmc_select_cpu(cpu);
3099 			for (n = 0; n < (int) md->pmd_npmc; n++)
3100 				if (pmc_can_allocate_row(n, mode) == 0 &&
3101 				    pmc_can_allocate_rowindex(
3102 					    curthread->td_proc, n, cpu) == 0 &&
3103 				    (PMC_IS_UNALLOCATED(cpu, n) ||
3104 				     PMC_IS_SHAREABLE_PMC(cpu, n)) &&
3105 				    md->pmd_allocate_pmc(cpu, n, pmc,
3106 					&pa) == 0)
3107 					break;
3108 		} else {
3109 			/* Process virtual mode */
3110 			for (n = 0; n < (int) md->pmd_npmc; n++) {
3111 				if (pmc_can_allocate_row(n, mode) == 0 &&
3112 				    pmc_can_allocate_rowindex(
3113 					    curthread->td_proc, n,
3114 					    PMC_CPU_ANY) == 0 &&
3115 				    md->pmd_allocate_pmc(curthread->td_oncpu,
3116 					n, pmc, &pa) == 0)
3117 					break;
3118 			}
3119 		}
3120 
3121 #undef	PMC_IS_UNALLOCATED
3122 #undef	PMC_IS_SHAREABLE_PMC
3123 
3124 		pmc_restore_cpu_binding(&pb);
3125 
3126 		if (n == (int) md->pmd_npmc) {
3127 			pmc_destroy_pmc_descriptor(pmc);
3128 			FREE(pmc, M_PMC);
3129 			pmc = NULL;
3130 			error = EINVAL;
3131 			break;
3132 		}
3133 
3134 		/* Fill in the correct value in the ID field */
3135 		pmc->pm_id = PMC_ID_MAKE_ID(cpu,mode,pa.pm_class,n);
3136 
3137 		PMCDBG(PMC,ALL,2, "ev=%d class=%d mode=%d n=%d -> pmcid=%x",
3138 		    pmc->pm_event, pa.pm_class, mode, n, pmc->pm_id);
3139 
3140 		/* Process mode PMCs with logging enabled need log files */
3141 		if (pmc->pm_flags & (PMC_F_LOG_PROCEXIT | PMC_F_LOG_PROCCSW))
3142 			pmc->pm_flags |= PMC_F_NEEDS_LOGFILE;
3143 
3144 		/* All system mode sampling PMCs require a log file */
3145 		if (PMC_IS_SAMPLING_MODE(mode) && PMC_IS_SYSTEM_MODE(mode))
3146 			pmc->pm_flags |= PMC_F_NEEDS_LOGFILE;
3147 
3148 		/*
3149 		 * Configure global pmc's immediately
3150 		 */
3151 
3152 		if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pmc))) {
3153 
3154 			pmc_save_cpu_binding(&pb);
3155 			pmc_select_cpu(cpu);
3156 
3157 			phw = pmc_pcpu[cpu]->pc_hwpmcs[n];
3158 
3159 			if ((phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0 ||
3160 			    (error = md->pmd_config_pmc(cpu, n, pmc)) != 0) {
3161 				(void) md->pmd_release_pmc(cpu, n, pmc);
3162 				pmc_destroy_pmc_descriptor(pmc);
3163 				FREE(pmc, M_PMC);
3164 				pmc = NULL;
3165 				pmc_restore_cpu_binding(&pb);
3166 				error = EPERM;
3167 				break;
3168 			}
3169 
3170 			pmc_restore_cpu_binding(&pb);
3171 		}
3172 
3173 		pmc->pm_state    = PMC_STATE_ALLOCATED;
3174 
3175 		/*
3176 		 * mark row disposition
3177 		 */
3178 
3179 		if (PMC_IS_SYSTEM_MODE(mode))
3180 			PMC_MARK_ROW_STANDALONE(n);
3181 		else
3182 			PMC_MARK_ROW_THREAD(n);
3183 
3184 		/*
3185 		 * Register this PMC with the current thread as its owner.
3186 		 */
3187 
3188 		if ((error =
3189 		    pmc_register_owner(curthread->td_proc, pmc)) != 0) {
3190 			pmc_release_pmc_descriptor(pmc);
3191 			FREE(pmc, M_PMC);
3192 			pmc = NULL;
3193 			break;
3194 		}
3195 
3196 		/*
3197 		 * Return the allocated index.
3198 		 */
3199 
3200 		pa.pm_pmcid = pmc->pm_id;
3201 
3202 		error = copyout(&pa, arg, sizeof(pa));
3203 	}
3204 	break;
3205 
3206 
3207 	/*
3208 	 * Attach a PMC to a process.
3209 	 */
3210 
3211 	case PMC_OP_PMCATTACH:
3212 	{
3213 		struct pmc *pm;
3214 		struct proc *p;
3215 		struct pmc_op_pmcattach a;
3216 
3217 		sx_assert(&pmc_sx, SX_XLOCKED);
3218 
3219 		if ((error = copyin(arg, &a, sizeof(a))) != 0)
3220 			break;
3221 
3222 		if (a.pm_pid < 0) {
3223 			error = EINVAL;
3224 			break;
3225 		} else if (a.pm_pid == 0)
3226 			a.pm_pid = td->td_proc->p_pid;
3227 
3228 		if ((error = pmc_find_pmc(a.pm_pmc, &pm)) != 0)
3229 			break;
3230 
3231 		if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm))) {
3232 			error = EINVAL;
3233 			break;
3234 		}
3235 
3236 		/* PMCs may be (re)attached only when allocated or stopped */
3237 		if (pm->pm_state == PMC_STATE_RUNNING) {
3238 			error = EBUSY;
3239 			break;
3240 		} else if (pm->pm_state != PMC_STATE_ALLOCATED &&
3241 		    pm->pm_state != PMC_STATE_STOPPED) {
3242 			error = EINVAL;
3243 			break;
3244 		}
3245 
3246 		/* lookup pid */
3247 		if ((p = pfind(a.pm_pid)) == NULL) {
3248 			error = ESRCH;
3249 			break;
3250 		}
3251 
3252 		/*
3253 		 * Ignore processes that are working on exiting.
3254 		 */
3255 		if (p->p_flag & P_WEXIT) {
3256 			error = ESRCH;
3257 			PROC_UNLOCK(p);	/* pfind() returns a locked process */
3258 			break;
3259 		}
3260 
3261 		/*
3262 		 * we are allowed to attach a PMC to a process if
3263 		 * we can debug it.
3264 		 */
3265 		error = p_candebug(curthread, p);
3266 
3267 		PROC_UNLOCK(p);
3268 
3269 		if (error == 0)
3270 			error = pmc_attach_process(p, pm);
3271 	}
3272 	break;
3273 
3274 
3275 	/*
3276 	 * Detach an attached PMC from a process.
3277 	 */
3278 
3279 	case PMC_OP_PMCDETACH:
3280 	{
3281 		struct pmc *pm;
3282 		struct proc *p;
3283 		struct pmc_op_pmcattach a;
3284 
3285 		if ((error = copyin(arg, &a, sizeof(a))) != 0)
3286 			break;
3287 
3288 		if (a.pm_pid < 0) {
3289 			error = EINVAL;
3290 			break;
3291 		} else if (a.pm_pid == 0)
3292 			a.pm_pid = td->td_proc->p_pid;
3293 
3294 		if ((error = pmc_find_pmc(a.pm_pmc, &pm)) != 0)
3295 			break;
3296 
3297 		if ((p = pfind(a.pm_pid)) == NULL) {
3298 			error = ESRCH;
3299 			break;
3300 		}
3301 
3302 		/*
3303 		 * Treat processes that are in the process of exiting
3304 		 * as if they were not present.
3305 		 */
3306 
3307 		if (p->p_flag & P_WEXIT)
3308 			error = ESRCH;
3309 
3310 		PROC_UNLOCK(p);	/* pfind() returns a locked process */
3311 
3312 		if (error == 0)
3313 			error = pmc_detach_process(p, pm);
3314 	}
3315 	break;
3316 
3317 
3318 	/*
3319 	 * Retrieve the MSR number associated with the counter
3320 	 * 'pmc_id'.  This allows processes to directly use RDPMC
3321 	 * instructions to read their PMCs, without the overhead of a
3322 	 * system call.
3323 	 */
3324 
3325 	case PMC_OP_PMCGETMSR:
3326 	{
3327 		int ri;
3328 		struct pmc	*pm;
3329 		struct pmc_target *pt;
3330 		struct pmc_op_getmsr gm;
3331 
3332 		PMC_DOWNGRADE_SX();
3333 
3334 		/* CPU has no 'GETMSR' support */
3335 		if (md->pmd_get_msr == NULL) {
3336 			error = ENOSYS;
3337 			break;
3338 		}
3339 
3340 		if ((error = copyin(arg, &gm, sizeof(gm))) != 0)
3341 			break;
3342 
3343 		if ((error = pmc_find_pmc(gm.pm_pmcid, &pm)) != 0)
3344 			break;
3345 
3346 		/*
3347 		 * The allocated PMC has to be a process virtual PMC,
3348 		 * i.e., of type MODE_T[CS].  Global PMCs can only be
3349 		 * read using the PMCREAD operation since they may be
3350 		 * allocated on a different CPU than the one we could
3351 		 * be running on at the time of the RDPMC instruction.
3352 		 *
3353 		 * The GETMSR operation is not allowed for PMCs that
3354 		 * are inherited across processes.
3355 		 */
3356 
3357 		if (!PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)) ||
3358 		    (pm->pm_flags & PMC_F_DESCENDANTS)) {
3359 			error = EINVAL;
3360 			break;
3361 		}
3362 
3363 		/*
3364 		 * It only makes sense to use a RDPMC (or its
3365 		 * equivalent instruction on non-x86 architectures) on
3366 		 * a process that has allocated and attached a PMC to
3367 		 * itself.  Conversely the PMC is only allowed to have
3368 		 * one process attached to it -- its owner.
3369 		 */
3370 
3371 		if ((pt = LIST_FIRST(&pm->pm_targets)) == NULL ||
3372 		    LIST_NEXT(pt, pt_next) != NULL ||
3373 		    pt->pt_process->pp_proc != pm->pm_owner->po_owner) {
3374 			error = EINVAL;
3375 			break;
3376 		}
3377 
3378 		ri = PMC_TO_ROWINDEX(pm);
3379 
3380 		if ((error = (*md->pmd_get_msr)(ri, &gm.pm_msr)) < 0)
3381 			break;
3382 
3383 		if ((error = copyout(&gm, arg, sizeof(gm))) < 0)
3384 			break;
3385 
3386 		/*
3387 		 * Mark our process as using MSRs.  Update machine
3388 		 * state using a forced context switch.
3389 		 */
3390 
3391 		pt->pt_process->pp_flags |= PMC_PP_ENABLE_MSR_ACCESS;
3392 		pmc_force_context_switch();
3393 
3394 	}
3395 	break;
3396 
3397 	/*
3398 	 * Release an allocated PMC
3399 	 */
3400 
3401 	case PMC_OP_PMCRELEASE:
3402 	{
3403 		pmc_id_t pmcid;
3404 		struct pmc *pm;
3405 		struct pmc_owner *po;
3406 		struct pmc_op_simple sp;
3407 
3408 		/*
3409 		 * Find PMC pointer for the named PMC.
3410 		 *
3411 		 * Use pmc_release_pmc_descriptor() to switch off the
3412 		 * PMC, remove all its target threads, and remove the
3413 		 * PMC from its owner's list.
3414 		 *
3415 		 * Remove the owner record if this is the last PMC
3416 		 * owned.
3417 		 *
3418 		 * Free up space.
3419 		 */
3420 
3421 		if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
3422 			break;
3423 
3424 		pmcid = sp.pm_pmcid;
3425 
3426 		if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
3427 			break;
3428 
3429 		po = pm->pm_owner;
3430 		pmc_release_pmc_descriptor(pm);
3431 		pmc_maybe_remove_owner(po);
3432 
3433 		FREE(pm, M_PMC);
3434 	}
3435 	break;
3436 
3437 
3438 	/*
3439 	 * Read and/or write a PMC.
3440 	 */
3441 
3442 	case PMC_OP_PMCRW:
3443 	{
3444 		uint32_t cpu, ri;
3445 		struct pmc *pm;
3446 		struct pmc_op_pmcrw *pprw;
3447 		struct pmc_op_pmcrw prw;
3448 		struct pmc_binding pb;
3449 		pmc_value_t oldvalue;
3450 
3451 		PMC_DOWNGRADE_SX();
3452 
3453 		if ((error = copyin(arg, &prw, sizeof(prw))) != 0)
3454 			break;
3455 
3456 		ri = 0;
3457 		PMCDBG(PMC,OPS,1, "rw id=%d flags=0x%x", prw.pm_pmcid,
3458 		    prw.pm_flags);
3459 
3460 		/* must have at least one flag set */
3461 		if ((prw.pm_flags & (PMC_F_OLDVALUE|PMC_F_NEWVALUE)) == 0) {
3462 			error = EINVAL;
3463 			break;
3464 		}
3465 
3466 		/* locate pmc descriptor */
3467 		if ((error = pmc_find_pmc(prw.pm_pmcid, &pm)) != 0)
3468 			break;
3469 
3470 		/* Can't read a PMC that hasn't been started. */
3471 		if (pm->pm_state != PMC_STATE_ALLOCATED &&
3472 		    pm->pm_state != PMC_STATE_STOPPED &&
3473 		    pm->pm_state != PMC_STATE_RUNNING) {
3474 			error = EINVAL;
3475 			break;
3476 		}
3477 
3478 		/* writing a new value is allowed only for 'STOPPED' pmcs */
3479 		if (pm->pm_state == PMC_STATE_RUNNING &&
3480 		    (prw.pm_flags & PMC_F_NEWVALUE)) {
3481 			error = EBUSY;
3482 			break;
3483 		}
3484 
3485 		if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) {
3486 
3487 			/*
3488 			 * If this PMC is attached to its owner (i.e.,
3489 			 * the process requesting this operation) and
3490 			 * is running, then attempt to get an
3491 			 * upto-date reading from hardware for a READ.
3492 			 * Writes are only allowed when the PMC is
3493 			 * stopped, so only update the saved value
3494 			 * field.
3495 			 *
3496 			 * If the PMC is not running, or is not
3497 			 * attached to its owner, read/write to the
3498 			 * savedvalue field.
3499 			 */
3500 
3501 			ri = PMC_TO_ROWINDEX(pm);
3502 
3503 			mtx_pool_lock_spin(pmc_mtxpool, pm);
3504 			cpu = curthread->td_oncpu;
3505 
3506 			if (prw.pm_flags & PMC_F_OLDVALUE) {
3507 				if ((pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) &&
3508 				    (pm->pm_state == PMC_STATE_RUNNING))
3509 					error = (*md->pmd_read_pmc)(cpu, ri,
3510 					    &oldvalue);
3511 				else
3512 					oldvalue = pm->pm_gv.pm_savedvalue;
3513 			}
3514 			if (prw.pm_flags & PMC_F_NEWVALUE)
3515 				pm->pm_gv.pm_savedvalue = prw.pm_value;
3516 
3517 			mtx_pool_unlock_spin(pmc_mtxpool, pm);
3518 
3519 		} else { /* System mode PMCs */
3520 			cpu = PMC_TO_CPU(pm);
3521 			ri  = PMC_TO_ROWINDEX(pm);
3522 
3523 			if (pmc_cpu_is_disabled(cpu)) {
3524 				error = ENXIO;
3525 				break;
3526 			}
3527 
3528 			/* move this thread to CPU 'cpu' */
3529 			pmc_save_cpu_binding(&pb);
3530 			pmc_select_cpu(cpu);
3531 
3532 			critical_enter();
3533 			/* save old value */
3534 			if (prw.pm_flags & PMC_F_OLDVALUE)
3535 				if ((error = (*md->pmd_read_pmc)(cpu, ri,
3536 					 &oldvalue)))
3537 					goto error;
3538 			/* write out new value */
3539 			if (prw.pm_flags & PMC_F_NEWVALUE)
3540 				error = (*md->pmd_write_pmc)(cpu, ri,
3541 				    prw.pm_value);
3542 		error:
3543 			critical_exit();
3544 			pmc_restore_cpu_binding(&pb);
3545 			if (error)
3546 				break;
3547 		}
3548 
3549 		pprw = (struct pmc_op_pmcrw *) arg;
3550 
3551 #ifdef	DEBUG
3552 		if (prw.pm_flags & PMC_F_NEWVALUE)
3553 			PMCDBG(PMC,OPS,2, "rw id=%d new %jx -> old %jx",
3554 			    ri, prw.pm_value, oldvalue);
3555 		else if (prw.pm_flags & PMC_F_OLDVALUE)
3556 			PMCDBG(PMC,OPS,2, "rw id=%d -> old %jx", ri, oldvalue);
3557 #endif
3558 
3559 		/* return old value if requested */
3560 		if (prw.pm_flags & PMC_F_OLDVALUE)
3561 			if ((error = copyout(&oldvalue, &pprw->pm_value,
3562 				 sizeof(prw.pm_value))))
3563 				break;
3564 
3565 	}
3566 	break;
3567 
3568 
3569 	/*
3570 	 * Set the sampling rate for a sampling mode PMC and the
3571 	 * initial count for a counting mode PMC.
3572 	 */
3573 
3574 	case PMC_OP_PMCSETCOUNT:
3575 	{
3576 		struct pmc *pm;
3577 		struct pmc_op_pmcsetcount sc;
3578 
3579 		PMC_DOWNGRADE_SX();
3580 
3581 		if ((error = copyin(arg, &sc, sizeof(sc))) != 0)
3582 			break;
3583 
3584 		if ((error = pmc_find_pmc(sc.pm_pmcid, &pm)) != 0)
3585 			break;
3586 
3587 		if (pm->pm_state == PMC_STATE_RUNNING) {
3588 			error = EBUSY;
3589 			break;
3590 		}
3591 
3592 		if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
3593 			pm->pm_sc.pm_reloadcount = sc.pm_count;
3594 		else
3595 			pm->pm_sc.pm_initial = sc.pm_count;
3596 	}
3597 	break;
3598 
3599 
3600 	/*
3601 	 * Start a PMC.
3602 	 */
3603 
3604 	case PMC_OP_PMCSTART:
3605 	{
3606 		pmc_id_t pmcid;
3607 		struct pmc *pm;
3608 		struct pmc_op_simple sp;
3609 
3610 		sx_assert(&pmc_sx, SX_XLOCKED);
3611 
3612 		if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
3613 			break;
3614 
3615 		pmcid = sp.pm_pmcid;
3616 
3617 		if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
3618 			break;
3619 
3620 		KASSERT(pmcid == pm->pm_id,
3621 		    ("[pmc,%d] pmcid %x != id %x", __LINE__,
3622 			pm->pm_id, pmcid));
3623 
3624 		if (pm->pm_state == PMC_STATE_RUNNING) /* already running */
3625 			break;
3626 		else if (pm->pm_state != PMC_STATE_STOPPED &&
3627 		    pm->pm_state != PMC_STATE_ALLOCATED) {
3628 			error = EINVAL;
3629 			break;
3630 		}
3631 
3632 		error = pmc_start(pm);
3633 	}
3634 	break;
3635 
3636 
3637 	/*
3638 	 * Stop a PMC.
3639 	 */
3640 
3641 	case PMC_OP_PMCSTOP:
3642 	{
3643 		pmc_id_t pmcid;
3644 		struct pmc *pm;
3645 		struct pmc_op_simple sp;
3646 
3647 		PMC_DOWNGRADE_SX();
3648 
3649 		if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
3650 			break;
3651 
3652 		pmcid = sp.pm_pmcid;
3653 
3654 		/*
3655 		 * Mark the PMC as inactive and invoke the MD stop
3656 		 * routines if needed.
3657 		 */
3658 
3659 		if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
3660 			break;
3661 
3662 		KASSERT(pmcid == pm->pm_id,
3663 		    ("[pmc,%d] pmc id %x != pmcid %x", __LINE__,
3664 			pm->pm_id, pmcid));
3665 
3666 		if (pm->pm_state == PMC_STATE_STOPPED) /* already stopped */
3667 			break;
3668 		else if (pm->pm_state != PMC_STATE_RUNNING) {
3669 			error = EINVAL;
3670 			break;
3671 		}
3672 
3673 		error = pmc_stop(pm);
3674 	}
3675 	break;
3676 
3677 
3678 	/*
3679 	 * Write a user supplied value to the log file.
3680 	 */
3681 
3682 	case PMC_OP_WRITELOG:
3683 	{
3684 		struct pmc_op_writelog wl;
3685 		struct pmc_owner *po;
3686 
3687 		PMC_DOWNGRADE_SX();
3688 
3689 		if ((error = copyin(arg, &wl, sizeof(wl))) != 0)
3690 			break;
3691 
3692 		if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) {
3693 			error = EINVAL;
3694 			break;
3695 		}
3696 
3697 		if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) {
3698 			error = EINVAL;
3699 			break;
3700 		}
3701 
3702 		error = pmclog_process_userlog(po, &wl);
3703 	}
3704 	break;
3705 
3706 
3707 	default:
3708 		error = EINVAL;
3709 		break;
3710 	}
3711 
3712 	if (is_sx_downgraded)
3713 		sx_sunlock(&pmc_sx);
3714 	else
3715 		sx_xunlock(&pmc_sx);
3716 
3717 	if (error)
3718 		atomic_add_int(&pmc_stats.pm_syscall_errors, 1);
3719 
3720 	PICKUP_GIANT();
3721 
3722 	return error;
3723 }
3724 
3725 /*
3726  * Helper functions
3727  */
3728 
3729 
3730 /*
3731  * Mark the thread as needing callchain capture and post an AST.  The
3732  * actual callchain capture will be done in a context where it is safe
3733  * to take page faults.
3734  */
3735 
3736 static void
3737 pmc_post_callchain_ast(void)
3738 {
3739 	struct thread *td;
3740 
3741 	td = curthread;
3742 
3743 	/*
3744 	 * Mark this thread as needing processing in ast().
3745 	 * td->td_pflags will be safe to touch as the process was in
3746 	 * user space when it was interrupted.
3747 	 */
3748 	td->td_pflags |= TDP_CALLCHAIN;
3749 
3750 	/*
3751 	 * Again, since we've entered this function directly from
3752 	 * userland, `td' is guaranteed to be not locked by this CPU,
3753 	 * so its safe to try acquire the thread lock even though we
3754 	 * are executing in an NMI context.  We need to acquire this
3755 	 * lock before touching `td_flags' because other CPUs may be
3756 	 * in the process of touching this field.
3757 	 */
3758 	thread_lock(td);
3759 	td->td_flags |= TDF_ASTPENDING;
3760 	thread_unlock(td);
3761 
3762 	return;
3763 }
3764 
3765 /*
3766  * Interrupt processing.
3767  *
3768  * Find a free slot in the per-cpu array of samples and capture the
3769  * current callchain there.  If a sample was successfully added, a bit
3770  * is set in mask 'pmc_cpumask' denoting that the DO_SAMPLES hook
3771  * needs to be invoked from the clock handler.
3772  *
3773  * This function is meant to be called from an NMI handler.  It cannot
3774  * use any of the locking primitives supplied by the OS.
3775  */
3776 
3777 int
3778 pmc_process_interrupt(int cpu, struct pmc *pm, struct trapframe *tf,
3779     int inuserspace)
3780 {
3781 	int error, callchaindepth;
3782 	struct thread *td;
3783 	struct pmc_sample *ps;
3784 	struct pmc_samplebuffer *psb;
3785 
3786 	error = 0;
3787 
3788 	/*
3789 	 * Allocate space for a sample buffer.
3790 	 */
3791 	psb = pmc_pcpu[cpu]->pc_sb;
3792 
3793 	ps = psb->ps_write;
3794 	if (ps->ps_nsamples) {	/* in use, reader hasn't caught up */
3795 		pm->pm_stalled = 1;
3796 		atomic_add_int(&pmc_stats.pm_intr_bufferfull, 1);
3797 		PMCDBG(SAM,INT,1,"(spc) cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d",
3798 		    cpu, pm, (void *) tf, inuserspace,
3799 		    (int) (psb->ps_write - psb->ps_samples),
3800 		    (int) (psb->ps_read - psb->ps_samples));
3801 		error = ENOMEM;
3802 		goto done;
3803 	}
3804 
3805 
3806 	/* Fill in entry. */
3807 	PMCDBG(SAM,INT,1,"cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d", cpu, pm,
3808 	    (void *) tf, inuserspace,
3809 	    (int) (psb->ps_write - psb->ps_samples),
3810 	    (int) (psb->ps_read - psb->ps_samples));
3811 
3812 	atomic_add_rel_32(&pm->pm_runcount, 1);	/* hold onto PMC */
3813 	ps->ps_pmc = pm;
3814 	if ((td = curthread) && td->td_proc)
3815 		ps->ps_pid = td->td_proc->p_pid;
3816 	else
3817 		ps->ps_pid = -1;
3818 	ps->ps_cpu = cpu;
3819 	ps->ps_flags = inuserspace ? PMC_CC_F_USERSPACE : 0;
3820 
3821 	callchaindepth = (pm->pm_flags & PMC_F_CALLCHAIN) ?
3822 	    pmc_callchaindepth : 1;
3823 
3824 	if (callchaindepth == 1)
3825 		ps->ps_pc[0] = PMC_TRAPFRAME_TO_PC(tf);
3826 	else {
3827 		/*
3828 		 * Kernel stack traversals can be done immediately,
3829 		 * while we defer to an AST for user space traversals.
3830 		 */
3831 		if (!inuserspace)
3832 			callchaindepth =
3833 			    pmc_save_kernel_callchain(ps->ps_pc,
3834 				callchaindepth, tf);
3835 		else {
3836 			pmc_post_callchain_ast();
3837 			callchaindepth = PMC_SAMPLE_INUSE;
3838 		}
3839 	}
3840 
3841 	ps->ps_nsamples = callchaindepth;	/* mark entry as in use */
3842 
3843 	/* increment write pointer, modulo ring buffer size */
3844 	ps++;
3845 	if (ps == psb->ps_fence)
3846 		psb->ps_write = psb->ps_samples;
3847 	else
3848 		psb->ps_write = ps;
3849 
3850  done:
3851 	/* mark CPU as needing processing */
3852 	atomic_set_rel_int(&pmc_cpumask, (1 << cpu));
3853 
3854 	return (error);
3855 }
3856 
3857 /*
3858  * Capture a user call chain.  This function will be called from ast()
3859  * before control returns to userland and before the process gets
3860  * rescheduled.
3861  */
3862 
3863 static void
3864 pmc_capture_user_callchain(int cpu, struct trapframe *tf)
3865 {
3866 	int i;
3867 	struct pmc *pm;
3868 	struct pmc_sample *ps;
3869 	struct pmc_samplebuffer *psb;
3870 
3871 	psb = pmc_pcpu[cpu]->pc_sb;
3872 
3873 	/*
3874 	 * Iterate through all deferred callchain requests.
3875 	 */
3876 
3877 	for (i = 0; i < pmc_nsamples; i++) {
3878 
3879 		ps = &psb->ps_samples[i];
3880 		if (ps->ps_nsamples != PMC_SAMPLE_INUSE)
3881 			continue;
3882 
3883 		pm = ps->ps_pmc;
3884 
3885 		KASSERT(pm->pm_flags & PMC_F_CALLCHAIN,
3886 		    ("[pmc,%d] Retrieving callchain for PMC that doesn't "
3887 			"want it", __LINE__));
3888 
3889 		/*
3890 		 * Retrieve the callchain and mark the sample buffer
3891 		 * as 'processable' by the timer tick sweep code.
3892 		 */
3893 		ps->ps_nsamples = pmc_save_user_callchain(ps->ps_pc,
3894 		    pmc_callchaindepth, tf);
3895 	}
3896 
3897 	return;
3898 }
3899 
3900 
3901 /*
3902  * Process saved PC samples.
3903  */
3904 
3905 static void
3906 pmc_process_samples(int cpu)
3907 {
3908 	int n, ri;
3909 	struct pmc *pm;
3910 	struct thread *td;
3911 	struct pmc_owner *po;
3912 	struct pmc_sample *ps;
3913 	struct pmc_samplebuffer *psb;
3914 
3915 	KASSERT(PCPU_GET(cpuid) == cpu,
3916 	    ("[pmc,%d] not on the correct CPU pcpu=%d cpu=%d", __LINE__,
3917 		PCPU_GET(cpuid), cpu));
3918 
3919 	psb = pmc_pcpu[cpu]->pc_sb;
3920 
3921 	for (n = 0; n < pmc_nsamples; n++) { /* bound on #iterations */
3922 
3923 		ps = psb->ps_read;
3924 		if (ps->ps_nsamples == PMC_SAMPLE_FREE)
3925 			break;
3926 		if (ps->ps_nsamples == PMC_SAMPLE_INUSE) {
3927 			/* Need a rescan at a later time. */
3928 			atomic_set_rel_int(&pmc_cpumask, (1 << cpu));
3929 			break;
3930 		}
3931 
3932 		pm = ps->ps_pmc;
3933 		po = pm->pm_owner;
3934 
3935 		KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)),
3936 		    ("[pmc,%d] pmc=%p non-sampling mode=%d", __LINE__,
3937 			pm, PMC_TO_MODE(pm)));
3938 
3939 		/* Ignore PMCs that have been switched off */
3940 		if (pm->pm_state != PMC_STATE_RUNNING)
3941 			goto entrydone;
3942 
3943 		PMCDBG(SAM,OPS,1,"cpu=%d pm=%p n=%d fl=%x wr=%d rd=%d", cpu,
3944 		    pm, ps->ps_nsamples, ps->ps_flags,
3945 		    (int) (psb->ps_write - psb->ps_samples),
3946 		    (int) (psb->ps_read - psb->ps_samples));
3947 
3948 		/*
3949 		 * If this is a process-mode PMC that is attached to
3950 		 * its owner, and if the PC is in user mode, update
3951 		 * profiling statistics like timer-based profiling
3952 		 * would have done.
3953 		 */
3954 		if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) {
3955 			if (ps->ps_flags & PMC_CC_F_USERSPACE) {
3956 				td = FIRST_THREAD_IN_PROC(po->po_owner);
3957 				addupc_intr(td, ps->ps_pc[0], 1);
3958 			}
3959 			goto entrydone;
3960 		}
3961 
3962 		/*
3963 		 * Otherwise, this is either a sampling mode PMC that
3964 		 * is attached to a different process than its owner,
3965 		 * or a system-wide sampling PMC.  Dispatch a log
3966 		 * entry to the PMC's owner process.
3967 		 */
3968 
3969 		pmclog_process_callchain(pm, ps);
3970 
3971 	entrydone:
3972 		ps->ps_nsamples = 0;	/* mark entry as free */
3973 		atomic_subtract_rel_32(&pm->pm_runcount, 1);
3974 
3975 		/* increment read pointer, modulo sample size */
3976 		if (++ps == psb->ps_fence)
3977 			psb->ps_read = psb->ps_samples;
3978 		else
3979 			psb->ps_read = ps;
3980 	}
3981 
3982 	atomic_add_int(&pmc_stats.pm_log_sweeps, 1);
3983 
3984 	/* Do not re-enable stalled PMCs if we failed to process any samples */
3985 	if (n == 0)
3986 		return;
3987 
3988 	/*
3989 	 * Restart any stalled sampling PMCs on this CPU.
3990 	 *
3991 	 * If the NMI handler sets the pm_stalled field of a PMC after
3992 	 * the check below, we'll end up processing the stalled PMC at
3993 	 * the next hardclock tick.
3994 	 */
3995 	for (n = 0; n < md->pmd_npmc; n++) {
3996 		(void) (*md->pmd_get_config)(cpu,n,&pm);
3997 		if (pm == NULL ||			 /* !cfg'ed */
3998 		    pm->pm_state != PMC_STATE_RUNNING || /* !active */
3999 		    !PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) || /* !sampling */
4000 		    pm->pm_stalled == 0) /* !stalled */
4001 			continue;
4002 
4003 		pm->pm_stalled = 0;
4004 		ri = PMC_TO_ROWINDEX(pm);
4005 		(*md->pmd_start_pmc)(cpu, ri);
4006 	}
4007 }
4008 
4009 /*
4010  * Event handlers.
4011  */
4012 
4013 /*
4014  * Handle a process exit.
4015  *
4016  * Remove this process from all hash tables.  If this process
4017  * owned any PMCs, turn off those PMCs and deallocate them,
4018  * removing any associations with target processes.
4019  *
4020  * This function will be called by the last 'thread' of a
4021  * process.
4022  *
4023  * XXX This eventhandler gets called early in the exit process.
4024  * Consider using a 'hook' invocation from thread_exit() or equivalent
4025  * spot.  Another negative is that kse_exit doesn't seem to call
4026  * exit1() [??].
4027  *
4028  */
4029 
4030 static void
4031 pmc_process_exit(void *arg __unused, struct proc *p)
4032 {
4033 	int is_using_hwpmcs;
4034 	int cpu;
4035 	unsigned int ri;
4036 	struct pmc *pm;
4037 	struct pmc_process *pp;
4038 	struct pmc_owner *po;
4039 	pmc_value_t newvalue, tmp;
4040 
4041 	PROC_LOCK(p);
4042 	is_using_hwpmcs = p->p_flag & P_HWPMC;
4043 	PROC_UNLOCK(p);
4044 
4045 	/*
4046 	 * Log a sysexit event to all SS PMC owners.
4047 	 */
4048 	LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
4049 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
4050 		    pmclog_process_sysexit(po, p->p_pid);
4051 
4052 	if (!is_using_hwpmcs)
4053 		return;
4054 
4055 	PMC_GET_SX_XLOCK();
4056 	PMCDBG(PRC,EXT,1,"process-exit proc=%p (%d, %s)", p, p->p_pid,
4057 	    p->p_comm);
4058 
4059 	/*
4060 	 * Since this code is invoked by the last thread in an exiting
4061 	 * process, we would have context switched IN at some prior
4062 	 * point.  However, with PREEMPTION, kernel mode context
4063 	 * switches may happen any time, so we want to disable a
4064 	 * context switch OUT till we get any PMCs targetting this
4065 	 * process off the hardware.
4066 	 *
4067 	 * We also need to atomically remove this process'
4068 	 * entry from our target process hash table, using
4069 	 * PMC_FLAG_REMOVE.
4070 	 */
4071 	PMCDBG(PRC,EXT,1, "process-exit proc=%p (%d, %s)", p, p->p_pid,
4072 	    p->p_comm);
4073 
4074 	critical_enter(); /* no preemption */
4075 
4076 	cpu = curthread->td_oncpu;
4077 
4078 	if ((pp = pmc_find_process_descriptor(p,
4079 		 PMC_FLAG_REMOVE)) != NULL) {
4080 
4081 		PMCDBG(PRC,EXT,2,
4082 		    "process-exit proc=%p pmc-process=%p", p, pp);
4083 
4084 		/*
4085 		 * The exiting process could the target of
4086 		 * some PMCs which will be running on
4087 		 * currently executing CPU.
4088 		 *
4089 		 * We need to turn these PMCs off like we
4090 		 * would do at context switch OUT time.
4091 		 */
4092 		for (ri = 0; ri < md->pmd_npmc; ri++) {
4093 
4094 			/*
4095 			 * Pick up the pmc pointer from hardware
4096 			 * state similar to the CSW_OUT code.
4097 			 */
4098 			pm = NULL;
4099 			(void) (*md->pmd_get_config)(cpu, ri, &pm);
4100 
4101 			PMCDBG(PRC,EXT,2, "ri=%d pm=%p", ri, pm);
4102 
4103 			if (pm == NULL ||
4104 			    !PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)))
4105 				continue;
4106 
4107 			PMCDBG(PRC,EXT,2, "ppmcs[%d]=%p pm=%p "
4108 			    "state=%d", ri, pp->pp_pmcs[ri].pp_pmc,
4109 			    pm, pm->pm_state);
4110 
4111 			KASSERT(PMC_TO_ROWINDEX(pm) == ri,
4112 			    ("[pmc,%d] ri mismatch pmc(%d) ri(%d)",
4113 				__LINE__, PMC_TO_ROWINDEX(pm), ri));
4114 
4115 			KASSERT(pm == pp->pp_pmcs[ri].pp_pmc,
4116 			    ("[pmc,%d] pm %p != pp_pmcs[%d] %p",
4117 				__LINE__, pm, ri, pp->pp_pmcs[ri].pp_pmc));
4118 
4119 			(void) md->pmd_stop_pmc(cpu, ri);
4120 
4121 			KASSERT(pm->pm_runcount > 0,
4122 			    ("[pmc,%d] bad runcount ri %d rc %d",
4123 				__LINE__, ri, pm->pm_runcount));
4124 
4125 			/* Stop hardware only if it is actually running */
4126 			if (pm->pm_state == PMC_STATE_RUNNING &&
4127 			    pm->pm_stalled == 0) {
4128 				md->pmd_read_pmc(cpu, ri, &newvalue);
4129 				tmp = newvalue -
4130 				    PMC_PCPU_SAVED(cpu,ri);
4131 
4132 				mtx_pool_lock_spin(pmc_mtxpool, pm);
4133 				pm->pm_gv.pm_savedvalue += tmp;
4134 				pp->pp_pmcs[ri].pp_pmcval += tmp;
4135 				mtx_pool_unlock_spin(pmc_mtxpool, pm);
4136 			}
4137 
4138 			atomic_subtract_rel_32(&pm->pm_runcount,1);
4139 
4140 			KASSERT((int) pm->pm_runcount >= 0,
4141 			    ("[pmc,%d] runcount is %d", __LINE__, ri));
4142 
4143 			(void) md->pmd_config_pmc(cpu, ri, NULL);
4144 		}
4145 
4146 		/*
4147 		 * Inform the MD layer of this pseudo "context switch
4148 		 * out"
4149 		 */
4150 		(void) md->pmd_switch_out(pmc_pcpu[cpu], pp);
4151 
4152 		critical_exit(); /* ok to be pre-empted now */
4153 
4154 		/*
4155 		 * Unlink this process from the PMCs that are
4156 		 * targetting it.  This will send a signal to
4157 		 * all PMC owner's whose PMCs are orphaned.
4158 		 *
4159 		 * Log PMC value at exit time if requested.
4160 		 */
4161 		for (ri = 0; ri < md->pmd_npmc; ri++)
4162 			if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) {
4163 				if (pm->pm_flags & PMC_F_NEEDS_LOGFILE &&
4164 				    PMC_IS_COUNTING_MODE(PMC_TO_MODE(pm)))
4165 					pmclog_process_procexit(pm, pp);
4166 				pmc_unlink_target_process(pm, pp);
4167 			}
4168 		FREE(pp, M_PMC);
4169 
4170 	} else
4171 		critical_exit(); /* pp == NULL */
4172 
4173 
4174 	/*
4175 	 * If the process owned PMCs, free them up and free up
4176 	 * memory.
4177 	 */
4178 	if ((po = pmc_find_owner_descriptor(p)) != NULL) {
4179 		pmc_remove_owner(po);
4180 		pmc_destroy_owner_descriptor(po);
4181 	}
4182 
4183 	sx_xunlock(&pmc_sx);
4184 }
4185 
4186 /*
4187  * Handle a process fork.
4188  *
4189  * If the parent process 'p1' is under HWPMC monitoring, then copy
4190  * over any attached PMCs that have 'do_descendants' semantics.
4191  */
4192 
4193 static void
4194 pmc_process_fork(void *arg __unused, struct proc *p1, struct proc *newproc,
4195     int flags)
4196 {
4197 	int is_using_hwpmcs;
4198 	unsigned int ri;
4199 	uint32_t do_descendants;
4200 	struct pmc *pm;
4201 	struct pmc_owner *po;
4202 	struct pmc_process *ppnew, *ppold;
4203 
4204 	(void) flags;		/* unused parameter */
4205 
4206 	PROC_LOCK(p1);
4207 	is_using_hwpmcs = p1->p_flag & P_HWPMC;
4208 	PROC_UNLOCK(p1);
4209 
4210 	/*
4211 	 * If there are system-wide sampling PMCs active, we need to
4212 	 * log all fork events to their owner's logs.
4213 	 */
4214 
4215 	LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
4216 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
4217 		    pmclog_process_procfork(po, p1->p_pid, newproc->p_pid);
4218 
4219 	if (!is_using_hwpmcs)
4220 		return;
4221 
4222 	PMC_GET_SX_XLOCK();
4223 	PMCDBG(PMC,FRK,1, "process-fork proc=%p (%d, %s) -> %p", p1,
4224 	    p1->p_pid, p1->p_comm, newproc);
4225 
4226 	/*
4227 	 * If the parent process (curthread->td_proc) is a
4228 	 * target of any PMCs, look for PMCs that are to be
4229 	 * inherited, and link these into the new process
4230 	 * descriptor.
4231 	 */
4232 	if ((ppold = pmc_find_process_descriptor(curthread->td_proc,
4233 		 PMC_FLAG_NONE)) == NULL)
4234 		goto done;		/* nothing to do */
4235 
4236 	do_descendants = 0;
4237 	for (ri = 0; ri < md->pmd_npmc; ri++)
4238 		if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL)
4239 			do_descendants |= pm->pm_flags & PMC_F_DESCENDANTS;
4240 	if (do_descendants == 0) /* nothing to do */
4241 		goto done;
4242 
4243 	/* allocate a descriptor for the new process  */
4244 	if ((ppnew = pmc_find_process_descriptor(newproc,
4245 		 PMC_FLAG_ALLOCATE)) == NULL)
4246 		goto done;
4247 
4248 	/*
4249 	 * Run through all PMCs that were targeting the old process
4250 	 * and which specified F_DESCENDANTS and attach them to the
4251 	 * new process.
4252 	 *
4253 	 * Log the fork event to all owners of PMCs attached to this
4254 	 * process, if not already logged.
4255 	 */
4256 	for (ri = 0; ri < md->pmd_npmc; ri++)
4257 		if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL &&
4258 		    (pm->pm_flags & PMC_F_DESCENDANTS)) {
4259 			pmc_link_target_process(pm, ppnew);
4260 			po = pm->pm_owner;
4261 			if (po->po_sscount == 0 &&
4262 			    po->po_flags & PMC_PO_OWNS_LOGFILE)
4263 				pmclog_process_procfork(po, p1->p_pid,
4264 				    newproc->p_pid);
4265 		}
4266 
4267 	/*
4268 	 * Now mark the new process as being tracked by this driver.
4269 	 */
4270 	PROC_LOCK(newproc);
4271 	newproc->p_flag |= P_HWPMC;
4272 	PROC_UNLOCK(newproc);
4273 
4274  done:
4275 	sx_xunlock(&pmc_sx);
4276 }
4277 
4278 
4279 /*
4280  * initialization
4281  */
4282 
4283 static const char *pmc_name_of_pmcclass[] = {
4284 #undef	__PMC_CLASS
4285 #define	__PMC_CLASS(N) #N ,
4286 	__PMC_CLASSES()
4287 };
4288 
4289 static int
4290 pmc_initialize(void)
4291 {
4292 	int cpu, error, n;
4293 	struct pmc_binding pb;
4294 	struct pmc_sample *ps;
4295 	struct pmc_samplebuffer *sb;
4296 
4297 	md = NULL;
4298 	error = 0;
4299 
4300 #ifdef	DEBUG
4301 	/* parse debug flags first */
4302 	if (TUNABLE_STR_FETCH(PMC_SYSCTL_NAME_PREFIX "debugflags",
4303 		pmc_debugstr, sizeof(pmc_debugstr)))
4304 		pmc_debugflags_parse(pmc_debugstr,
4305 		    pmc_debugstr+strlen(pmc_debugstr));
4306 #endif
4307 
4308 	PMCDBG(MOD,INI,0, "PMC Initialize (version %x)", PMC_VERSION);
4309 
4310 	/* check kernel version */
4311 	if (pmc_kernel_version != PMC_VERSION) {
4312 		if (pmc_kernel_version == 0)
4313 			printf("hwpmc: this kernel has not been compiled with "
4314 			    "'options HWPMC_HOOKS'.\n");
4315 		else
4316 			printf("hwpmc: kernel version (0x%x) does not match "
4317 			    "module version (0x%x).\n", pmc_kernel_version,
4318 			    PMC_VERSION);
4319 		return EPROGMISMATCH;
4320 	}
4321 
4322 	/*
4323 	 * check sysctl parameters
4324 	 */
4325 
4326 	if (pmc_hashsize <= 0) {
4327 		(void) printf("hwpmc: tunable \"hashsize\"=%d must be "
4328 		    "greater than zero.\n", pmc_hashsize);
4329 		pmc_hashsize = PMC_HASH_SIZE;
4330 	}
4331 
4332 	if (pmc_nsamples <= 0 || pmc_nsamples > 65535) {
4333 		(void) printf("hwpmc: tunable \"nsamples\"=%d out of "
4334 		    "range.\n", pmc_nsamples);
4335 		pmc_nsamples = PMC_NSAMPLES;
4336 	}
4337 
4338 	if (pmc_callchaindepth <= 0 ||
4339 	    pmc_callchaindepth > PMC_CALLCHAIN_DEPTH_MAX) {
4340 		(void) printf("hwpmc: tunable \"callchaindepth\"=%d out of "
4341 		    "range.\n", pmc_callchaindepth);
4342 		pmc_callchaindepth = PMC_CALLCHAIN_DEPTH;
4343 	}
4344 
4345 	md = pmc_md_initialize();
4346 
4347 	if (md == NULL || md->pmd_init == NULL)
4348 		return ENOSYS;
4349 
4350 	/* allocate space for the per-cpu array */
4351 	MALLOC(pmc_pcpu, struct pmc_cpu **, mp_ncpus * sizeof(struct pmc_cpu *),
4352 	    M_PMC, M_WAITOK|M_ZERO);
4353 
4354 	/* per-cpu 'saved values' for managing process-mode PMCs */
4355 	MALLOC(pmc_pcpu_saved, pmc_value_t *,
4356 	    sizeof(pmc_value_t) * mp_ncpus * md->pmd_npmc, M_PMC, M_WAITOK);
4357 
4358 	/* perform cpu dependent initialization */
4359 	pmc_save_cpu_binding(&pb);
4360 	for (cpu = 0; cpu < mp_ncpus; cpu++) {
4361 		if (pmc_cpu_is_disabled(cpu))
4362 			continue;
4363 		pmc_select_cpu(cpu);
4364 		if ((error = md->pmd_init(cpu)) != 0)
4365 			break;
4366 	}
4367 	pmc_restore_cpu_binding(&pb);
4368 
4369 	if (error != 0)
4370 		return error;
4371 
4372 	/* allocate space for the sample array */
4373 	for (cpu = 0; cpu < mp_ncpus; cpu++) {
4374 		if (pmc_cpu_is_disabled(cpu))
4375 			continue;
4376 		MALLOC(sb, struct pmc_samplebuffer *,
4377 		    sizeof(struct pmc_samplebuffer) +
4378 		    pmc_nsamples * sizeof(struct pmc_sample), M_PMC,
4379 		    M_WAITOK|M_ZERO);
4380 
4381 		sb->ps_read = sb->ps_write = sb->ps_samples;
4382 		sb->ps_fence = sb->ps_samples + pmc_nsamples;
4383 		KASSERT(pmc_pcpu[cpu] != NULL,
4384 		    ("[pmc,%d] cpu=%d Null per-cpu data", __LINE__, cpu));
4385 
4386 		MALLOC(sb->ps_callchains, uintptr_t *,
4387 		    pmc_callchaindepth * pmc_nsamples * sizeof(uintptr_t),
4388 		    M_PMC, M_WAITOK|M_ZERO);
4389 
4390 		for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++)
4391 			ps->ps_pc = sb->ps_callchains +
4392 			    (n * pmc_callchaindepth);
4393 
4394 		pmc_pcpu[cpu]->pc_sb = sb;
4395 	}
4396 
4397 	/* allocate space for the row disposition array */
4398 	pmc_pmcdisp = malloc(sizeof(enum pmc_mode) * md->pmd_npmc,
4399 	    M_PMC, M_WAITOK|M_ZERO);
4400 
4401 	KASSERT(pmc_pmcdisp != NULL,
4402 	    ("[pmc,%d] pmcdisp allocation returned NULL", __LINE__));
4403 
4404 	/* mark all PMCs as available */
4405 	for (n = 0; n < (int) md->pmd_npmc; n++)
4406 		PMC_MARK_ROW_FREE(n);
4407 
4408 	/* allocate thread hash tables */
4409 	pmc_ownerhash = hashinit(pmc_hashsize, M_PMC,
4410 	    &pmc_ownerhashmask);
4411 
4412 	pmc_processhash = hashinit(pmc_hashsize, M_PMC,
4413 	    &pmc_processhashmask);
4414 	mtx_init(&pmc_processhash_mtx, "pmc-process-hash", "pmc-leaf",
4415 	    MTX_SPIN);
4416 
4417 	LIST_INIT(&pmc_ss_owners);
4418 	pmc_ss_count = 0;
4419 
4420 	/* allocate a pool of spin mutexes */
4421 	pmc_mtxpool = mtx_pool_create("pmc-leaf", pmc_mtxpool_size,
4422 	    MTX_SPIN);
4423 
4424 	PMCDBG(MOD,INI,1, "pmc_ownerhash=%p, mask=0x%lx "
4425 	    "targethash=%p mask=0x%lx", pmc_ownerhash, pmc_ownerhashmask,
4426 	    pmc_processhash, pmc_processhashmask);
4427 
4428 	/* register process {exit,fork,exec} handlers */
4429 	pmc_exit_tag = EVENTHANDLER_REGISTER(process_exit,
4430 	    pmc_process_exit, NULL, EVENTHANDLER_PRI_ANY);
4431 	pmc_fork_tag = EVENTHANDLER_REGISTER(process_fork,
4432 	    pmc_process_fork, NULL, EVENTHANDLER_PRI_ANY);
4433 
4434 	/* initialize logging */
4435 	pmclog_initialize();
4436 
4437 	/* set hook functions */
4438 	pmc_intr = md->pmd_intr;
4439 	pmc_hook = pmc_hook_handler;
4440 
4441 	if (error == 0) {
4442 		printf(PMC_MODULE_NAME ":");
4443 		for (n = 0; n < (int) md->pmd_nclass; n++) {
4444 			printf(" %s/%d/0x%b",
4445 			    pmc_name_of_pmcclass[md->pmd_classes[n].pm_class],
4446 			    md->pmd_nclasspmcs[n],
4447 			    md->pmd_classes[n].pm_caps,
4448 			    "\20"
4449 			    "\1INT\2USR\3SYS\4EDG\5THR"
4450 			    "\6REA\7WRI\10INV\11QUA\12PRC"
4451 			    "\13TAG\14CSC");
4452 		}
4453 		printf("\n");
4454 	}
4455 
4456 	return error;
4457 }
4458 
4459 /* prepare to be unloaded */
4460 static void
4461 pmc_cleanup(void)
4462 {
4463 	int cpu;
4464 	struct pmc_ownerhash *ph;
4465 	struct pmc_owner *po, *tmp;
4466 	struct pmc_binding pb;
4467 #ifdef	DEBUG
4468 	struct pmc_processhash *prh;
4469 #endif
4470 
4471 	PMCDBG(MOD,INI,0, "%s", "cleanup");
4472 
4473 	/* switch off sampling */
4474 	atomic_store_rel_int(&pmc_cpumask, 0);
4475 	pmc_intr = NULL;
4476 
4477 	sx_xlock(&pmc_sx);
4478 	if (pmc_hook == NULL) {	/* being unloaded already */
4479 		sx_xunlock(&pmc_sx);
4480 		return;
4481 	}
4482 
4483 	pmc_hook = NULL; /* prevent new threads from entering module */
4484 
4485 	/* deregister event handlers */
4486 	EVENTHANDLER_DEREGISTER(process_fork, pmc_fork_tag);
4487 	EVENTHANDLER_DEREGISTER(process_exit, pmc_exit_tag);
4488 
4489 	/* send SIGBUS to all owner threads, free up allocations */
4490 	if (pmc_ownerhash)
4491 		for (ph = pmc_ownerhash;
4492 		     ph <= &pmc_ownerhash[pmc_ownerhashmask];
4493 		     ph++) {
4494 			LIST_FOREACH_SAFE(po, ph, po_next, tmp) {
4495 				pmc_remove_owner(po);
4496 
4497 				/* send SIGBUS to owner processes */
4498 				PMCDBG(MOD,INI,2, "cleanup signal proc=%p "
4499 				    "(%d, %s)", po->po_owner,
4500 				    po->po_owner->p_pid,
4501 				    po->po_owner->p_comm);
4502 
4503 				PROC_LOCK(po->po_owner);
4504 				psignal(po->po_owner, SIGBUS);
4505 				PROC_UNLOCK(po->po_owner);
4506 
4507 				pmc_destroy_owner_descriptor(po);
4508 			}
4509 		}
4510 
4511 	/* reclaim allocated data structures */
4512 	if (pmc_mtxpool)
4513 		mtx_pool_destroy(&pmc_mtxpool);
4514 
4515 	mtx_destroy(&pmc_processhash_mtx);
4516 	if (pmc_processhash) {
4517 #ifdef	DEBUG
4518 		struct pmc_process *pp;
4519 
4520 		PMCDBG(MOD,INI,3, "%s", "destroy process hash");
4521 		for (prh = pmc_processhash;
4522 		     prh <= &pmc_processhash[pmc_processhashmask];
4523 		     prh++)
4524 			LIST_FOREACH(pp, prh, pp_next)
4525 			    PMCDBG(MOD,INI,3, "pid=%d", pp->pp_proc->p_pid);
4526 #endif
4527 
4528 		hashdestroy(pmc_processhash, M_PMC, pmc_processhashmask);
4529 		pmc_processhash = NULL;
4530 	}
4531 
4532 	if (pmc_ownerhash) {
4533 		PMCDBG(MOD,INI,3, "%s", "destroy owner hash");
4534 		hashdestroy(pmc_ownerhash, M_PMC, pmc_ownerhashmask);
4535 		pmc_ownerhash = NULL;
4536 	}
4537 
4538 	KASSERT(LIST_EMPTY(&pmc_ss_owners),
4539 	    ("[pmc,%d] Global SS owner list not empty", __LINE__));
4540 	KASSERT(pmc_ss_count == 0,
4541 	    ("[pmc,%d] Global SS count not empty", __LINE__));
4542 
4543 	/* free the per-cpu sample buffers */
4544 	for (cpu = 0; cpu < mp_ncpus; cpu++) {
4545 		if (pmc_cpu_is_disabled(cpu))
4546 			continue;
4547 		KASSERT(pmc_pcpu[cpu]->pc_sb != NULL,
4548 		    ("[pmc,%d] Null cpu sample buffer cpu=%d", __LINE__,
4549 			cpu));
4550 		FREE(pmc_pcpu[cpu]->pc_sb->ps_callchains, M_PMC);
4551 		FREE(pmc_pcpu[cpu]->pc_sb, M_PMC);
4552 		pmc_pcpu[cpu]->pc_sb = NULL;
4553 	}
4554 
4555  	/* do processor dependent cleanup */
4556 	PMCDBG(MOD,INI,3, "%s", "md cleanup");
4557 	if (md) {
4558 		pmc_save_cpu_binding(&pb);
4559 		for (cpu = 0; cpu < mp_ncpus; cpu++) {
4560 			PMCDBG(MOD,INI,1,"pmc-cleanup cpu=%d pcs=%p",
4561 			    cpu, pmc_pcpu[cpu]);
4562 			if (pmc_cpu_is_disabled(cpu))
4563 				continue;
4564 			pmc_select_cpu(cpu);
4565 			if (pmc_pcpu[cpu])
4566 				(void) md->pmd_cleanup(cpu);
4567 		}
4568 		FREE(md, M_PMC);
4569 		md = NULL;
4570 		pmc_restore_cpu_binding(&pb);
4571 	}
4572 
4573 	/* deallocate per-cpu structures */
4574 	FREE(pmc_pcpu, M_PMC);
4575 	pmc_pcpu = NULL;
4576 
4577 	FREE(pmc_pcpu_saved, M_PMC);
4578 	pmc_pcpu_saved = NULL;
4579 
4580 	if (pmc_pmcdisp) {
4581 		FREE(pmc_pmcdisp, M_PMC);
4582 		pmc_pmcdisp = NULL;
4583 	}
4584 
4585 	pmclog_shutdown();
4586 
4587 	sx_xunlock(&pmc_sx); 	/* we are done */
4588 }
4589 
4590 /*
4591  * The function called at load/unload.
4592  */
4593 
4594 static int
4595 load (struct module *module __unused, int cmd, void *arg __unused)
4596 {
4597 	int error;
4598 
4599 	error = 0;
4600 
4601 	switch (cmd) {
4602 	case MOD_LOAD :
4603 		/* initialize the subsystem */
4604 		error = pmc_initialize();
4605 		if (error != 0)
4606 			break;
4607 		PMCDBG(MOD,INI,1, "syscall=%d ncpus=%d",
4608 		    pmc_syscall_num, mp_ncpus);
4609 		break;
4610 
4611 
4612 	case MOD_UNLOAD :
4613 	case MOD_SHUTDOWN:
4614 		pmc_cleanup();
4615 		PMCDBG(MOD,INI,1, "%s", "unloaded");
4616 		break;
4617 
4618 	default :
4619 		error = EINVAL;	/* XXX should panic(9) */
4620 		break;
4621 	}
4622 
4623 	return error;
4624 }
4625 
4626 /* memory pool */
4627 MALLOC_DEFINE(M_PMC, "pmc", "Memory space for the PMC module");
4628