xref: /freebsd/sys/dev/hwpmc/hwpmc_mod.c (revision 40f65a4df509c6b29b9e72407c7afefe28838844)
1 /*-
2  * Copyright (c) 2003-2008 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/mount.h>
45 #include <sys/mutex.h>
46 #include <sys/pmc.h>
47 #include <sys/pmckern.h>
48 #include <sys/pmclog.h>
49 #include <sys/priv.h>
50 #include <sys/proc.h>
51 #include <sys/queue.h>
52 #include <sys/resourcevar.h>
53 #include <sys/rwlock.h>
54 #include <sys/sched.h>
55 #include <sys/signalvar.h>
56 #include <sys/smp.h>
57 #include <sys/sx.h>
58 #include <sys/sysctl.h>
59 #include <sys/sysent.h>
60 #include <sys/systm.h>
61 #include <sys/vnode.h>
62 
63 #include <sys/linker.h>		/* needs to be after <sys/malloc.h> */
64 
65 #include <machine/atomic.h>
66 #include <machine/md_var.h>
67 
68 #include <vm/vm.h>
69 #include <vm/vm_extern.h>
70 #include <vm/pmap.h>
71 #include <vm/vm_map.h>
72 #include <vm/vm_object.h>
73 
74 #include "hwpmc_soft.h"
75 
76 /*
77  * Types
78  */
79 
80 enum pmc_flags {
81 	PMC_FLAG_NONE	  = 0x00, /* do nothing */
82 	PMC_FLAG_REMOVE   = 0x01, /* atomically remove entry from hash */
83 	PMC_FLAG_ALLOCATE = 0x02, /* add entry to hash if not found */
84 };
85 
86 /*
87  * The offset in sysent where the syscall is allocated.
88  */
89 
90 static int pmc_syscall_num = NO_SYSCALL;
91 struct pmc_cpu		**pmc_pcpu;	 /* per-cpu state */
92 pmc_value_t		*pmc_pcpu_saved; /* saved PMC values: CSW handling */
93 
94 #define	PMC_PCPU_SAVED(C,R)	pmc_pcpu_saved[(R) + md->pmd_npmc*(C)]
95 
96 struct mtx_pool		*pmc_mtxpool;
97 static int		*pmc_pmcdisp;	 /* PMC row dispositions */
98 
99 #define	PMC_ROW_DISP_IS_FREE(R)		(pmc_pmcdisp[(R)] == 0)
100 #define	PMC_ROW_DISP_IS_THREAD(R)	(pmc_pmcdisp[(R)] > 0)
101 #define	PMC_ROW_DISP_IS_STANDALONE(R)	(pmc_pmcdisp[(R)] < 0)
102 
103 #define	PMC_MARK_ROW_FREE(R) do {					  \
104 	pmc_pmcdisp[(R)] = 0;						  \
105 } while (0)
106 
107 #define	PMC_MARK_ROW_STANDALONE(R) do {					  \
108 	KASSERT(pmc_pmcdisp[(R)] <= 0, ("[pmc,%d] row disposition error", \
109 		    __LINE__));						  \
110 	atomic_add_int(&pmc_pmcdisp[(R)], -1);				  \
111 	KASSERT(pmc_pmcdisp[(R)] >= (-pmc_cpu_max_active()),		  \
112 		("[pmc,%d] row disposition error", __LINE__));		  \
113 } while (0)
114 
115 #define	PMC_UNMARK_ROW_STANDALONE(R) do { 				  \
116 	atomic_add_int(&pmc_pmcdisp[(R)], 1);				  \
117 	KASSERT(pmc_pmcdisp[(R)] <= 0, ("[pmc,%d] row disposition error", \
118 		    __LINE__));						  \
119 } while (0)
120 
121 #define	PMC_MARK_ROW_THREAD(R) do {					  \
122 	KASSERT(pmc_pmcdisp[(R)] >= 0, ("[pmc,%d] row disposition error", \
123 		    __LINE__));						  \
124 	atomic_add_int(&pmc_pmcdisp[(R)], 1);				  \
125 } while (0)
126 
127 #define	PMC_UNMARK_ROW_THREAD(R) do {					  \
128 	atomic_add_int(&pmc_pmcdisp[(R)], -1);				  \
129 	KASSERT(pmc_pmcdisp[(R)] >= 0, ("[pmc,%d] row disposition error", \
130 		    __LINE__));						  \
131 } while (0)
132 
133 
134 /* various event handlers */
135 static eventhandler_tag	pmc_exit_tag, pmc_fork_tag;
136 
137 /* Module statistics */
138 struct pmc_op_getdriverstats pmc_stats;
139 
140 /* Machine/processor dependent operations */
141 static struct pmc_mdep  *md;
142 
143 /*
144  * Hash tables mapping owner processes and target threads to PMCs.
145  */
146 
147 struct mtx pmc_processhash_mtx;		/* spin mutex */
148 static u_long pmc_processhashmask;
149 static LIST_HEAD(pmc_processhash, pmc_process)	*pmc_processhash;
150 
151 /*
152  * Hash table of PMC owner descriptors.  This table is protected by
153  * the shared PMC "sx" lock.
154  */
155 
156 static u_long pmc_ownerhashmask;
157 static LIST_HEAD(pmc_ownerhash, pmc_owner)	*pmc_ownerhash;
158 
159 /*
160  * List of PMC owners with system-wide sampling PMCs.
161  */
162 
163 static LIST_HEAD(, pmc_owner)			pmc_ss_owners;
164 
165 
166 /*
167  * A map of row indices to classdep structures.
168  */
169 static struct pmc_classdep **pmc_rowindex_to_classdep;
170 
171 /*
172  * Prototypes
173  */
174 
175 #ifdef	DEBUG
176 static int	pmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS);
177 static int	pmc_debugflags_parse(char *newstr, char *fence);
178 #endif
179 
180 static int	load(struct module *module, int cmd, void *arg);
181 static int	pmc_attach_process(struct proc *p, struct pmc *pm);
182 static struct pmc *pmc_allocate_pmc_descriptor(void);
183 static struct pmc_owner *pmc_allocate_owner_descriptor(struct proc *p);
184 static int	pmc_attach_one_process(struct proc *p, struct pmc *pm);
185 static int	pmc_can_allocate_rowindex(struct proc *p, unsigned int ri,
186     int cpu);
187 static int	pmc_can_attach(struct pmc *pm, struct proc *p);
188 static void	pmc_capture_user_callchain(int cpu, int soft, struct trapframe *tf);
189 static void	pmc_cleanup(void);
190 static int	pmc_detach_process(struct proc *p, struct pmc *pm);
191 static int	pmc_detach_one_process(struct proc *p, struct pmc *pm,
192     int flags);
193 static void	pmc_destroy_owner_descriptor(struct pmc_owner *po);
194 static struct pmc_owner *pmc_find_owner_descriptor(struct proc *p);
195 static int	pmc_find_pmc(pmc_id_t pmcid, struct pmc **pm);
196 static struct pmc *pmc_find_pmc_descriptor_in_process(struct pmc_owner *po,
197     pmc_id_t pmc);
198 static struct pmc_process *pmc_find_process_descriptor(struct proc *p,
199     uint32_t mode);
200 static void	pmc_force_context_switch(void);
201 static void	pmc_link_target_process(struct pmc *pm,
202     struct pmc_process *pp);
203 static void	pmc_log_all_process_mappings(struct pmc_owner *po);
204 static void	pmc_log_kernel_mappings(struct pmc *pm);
205 static void	pmc_log_process_mappings(struct pmc_owner *po, struct proc *p);
206 static void	pmc_maybe_remove_owner(struct pmc_owner *po);
207 static void	pmc_process_csw_in(struct thread *td);
208 static void	pmc_process_csw_out(struct thread *td);
209 static void	pmc_process_exit(void *arg, struct proc *p);
210 static void	pmc_process_fork(void *arg, struct proc *p1,
211     struct proc *p2, int n);
212 static void	pmc_process_samples(int cpu, int soft);
213 static void	pmc_release_pmc_descriptor(struct pmc *pmc);
214 static void	pmc_remove_owner(struct pmc_owner *po);
215 static void	pmc_remove_process_descriptor(struct pmc_process *pp);
216 static void	pmc_restore_cpu_binding(struct pmc_binding *pb);
217 static void	pmc_save_cpu_binding(struct pmc_binding *pb);
218 static void	pmc_select_cpu(int cpu);
219 static int	pmc_start(struct pmc *pm);
220 static int	pmc_stop(struct pmc *pm);
221 static int	pmc_syscall_handler(struct thread *td, void *syscall_args);
222 static void	pmc_unlink_target_process(struct pmc *pmc,
223     struct pmc_process *pp);
224 static int generic_switch_in(struct pmc_cpu *pc, struct pmc_process *pp);
225 static int generic_switch_out(struct pmc_cpu *pc, struct pmc_process *pp);
226 static struct pmc_mdep *pmc_generic_cpu_initialize(void);
227 static void pmc_generic_cpu_finalize(struct pmc_mdep *md);
228 
229 /*
230  * Kernel tunables and sysctl(8) interface.
231  */
232 
233 SYSCTL_DECL(_kern_hwpmc);
234 
235 static int pmc_callchaindepth = PMC_CALLCHAIN_DEPTH;
236 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "callchaindepth", &pmc_callchaindepth);
237 SYSCTL_INT(_kern_hwpmc, OID_AUTO, callchaindepth, CTLFLAG_TUN|CTLFLAG_RD,
238     &pmc_callchaindepth, 0, "depth of call chain records");
239 
240 #ifdef	DEBUG
241 struct pmc_debugflags pmc_debugflags = PMC_DEBUG_DEFAULT_FLAGS;
242 char	pmc_debugstr[PMC_DEBUG_STRSIZE];
243 TUNABLE_STR(PMC_SYSCTL_NAME_PREFIX "debugflags", pmc_debugstr,
244     sizeof(pmc_debugstr));
245 SYSCTL_PROC(_kern_hwpmc, OID_AUTO, debugflags,
246     CTLTYPE_STRING|CTLFLAG_RW|CTLFLAG_TUN,
247     0, 0, pmc_debugflags_sysctl_handler, "A", "debug flags");
248 #endif
249 
250 /*
251  * kern.hwpmc.hashrows -- determines the number of rows in the
252  * of the hash table used to look up threads
253  */
254 
255 static int pmc_hashsize = PMC_HASH_SIZE;
256 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "hashsize", &pmc_hashsize);
257 SYSCTL_INT(_kern_hwpmc, OID_AUTO, hashsize, CTLFLAG_TUN|CTLFLAG_RD,
258     &pmc_hashsize, 0, "rows in hash tables");
259 
260 /*
261  * kern.hwpmc.nsamples --- number of PC samples/callchain stacks per CPU
262  */
263 
264 static int pmc_nsamples = PMC_NSAMPLES;
265 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "nsamples", &pmc_nsamples);
266 SYSCTL_INT(_kern_hwpmc, OID_AUTO, nsamples, CTLFLAG_TUN|CTLFLAG_RD,
267     &pmc_nsamples, 0, "number of PC samples per CPU");
268 
269 
270 /*
271  * kern.hwpmc.mtxpoolsize -- number of mutexes in the mutex pool.
272  */
273 
274 static int pmc_mtxpool_size = PMC_MTXPOOL_SIZE;
275 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "mtxpoolsize", &pmc_mtxpool_size);
276 SYSCTL_INT(_kern_hwpmc, OID_AUTO, mtxpoolsize, CTLFLAG_TUN|CTLFLAG_RD,
277     &pmc_mtxpool_size, 0, "size of spin mutex pool");
278 
279 
280 /*
281  * security.bsd.unprivileged_syspmcs -- allow non-root processes to
282  * allocate system-wide PMCs.
283  *
284  * Allowing unprivileged processes to allocate system PMCs is convenient
285  * if system-wide measurements need to be taken concurrently with other
286  * per-process measurements.  This feature is turned off by default.
287  */
288 
289 static int pmc_unprivileged_syspmcs = 0;
290 TUNABLE_INT("security.bsd.unprivileged_syspmcs", &pmc_unprivileged_syspmcs);
291 SYSCTL_INT(_security_bsd, OID_AUTO, unprivileged_syspmcs, CTLFLAG_RW,
292     &pmc_unprivileged_syspmcs, 0,
293     "allow unprivileged process to allocate system PMCs");
294 
295 /*
296  * Hash function.  Discard the lower 2 bits of the pointer since
297  * these are always zero for our uses.  The hash multiplier is
298  * round((2^LONG_BIT) * ((sqrt(5)-1)/2)).
299  */
300 
301 #if	LONG_BIT == 64
302 #define	_PMC_HM		11400714819323198486u
303 #elif	LONG_BIT == 32
304 #define	_PMC_HM		2654435769u
305 #else
306 #error 	Must know the size of 'long' to compile
307 #endif
308 
309 #define	PMC_HASH_PTR(P,M)	((((unsigned long) (P) >> 2) * _PMC_HM) & (M))
310 
311 /*
312  * Syscall structures
313  */
314 
315 /* The `sysent' for the new syscall */
316 static struct sysent pmc_sysent = {
317 	2,			/* sy_narg */
318 	pmc_syscall_handler	/* sy_call */
319 };
320 
321 static struct syscall_module_data pmc_syscall_mod = {
322 	load,
323 	NULL,
324 	&pmc_syscall_num,
325 	&pmc_sysent,
326 	{ 0, NULL }
327 };
328 
329 static moduledata_t pmc_mod = {
330 	PMC_MODULE_NAME,
331 	syscall_module_handler,
332 	&pmc_syscall_mod
333 };
334 
335 DECLARE_MODULE(pmc, pmc_mod, SI_SUB_SMP, SI_ORDER_ANY);
336 MODULE_VERSION(pmc, PMC_VERSION);
337 
338 #ifdef	DEBUG
339 enum pmc_dbgparse_state {
340 	PMCDS_WS,		/* in whitespace */
341 	PMCDS_MAJOR,		/* seen a major keyword */
342 	PMCDS_MINOR
343 };
344 
345 static int
346 pmc_debugflags_parse(char *newstr, char *fence)
347 {
348 	char c, *p, *q;
349 	struct pmc_debugflags *tmpflags;
350 	int error, found, *newbits, tmp;
351 	size_t kwlen;
352 
353 	tmpflags = malloc(sizeof(*tmpflags), M_PMC, M_WAITOK|M_ZERO);
354 
355 	p = newstr;
356 	error = 0;
357 
358 	for (; p < fence && (c = *p); p++) {
359 
360 		/* skip white space */
361 		if (c == ' ' || c == '\t')
362 			continue;
363 
364 		/* look for a keyword followed by "=" */
365 		for (q = p; p < fence && (c = *p) && c != '='; p++)
366 			;
367 		if (c != '=') {
368 			error = EINVAL;
369 			goto done;
370 		}
371 
372 		kwlen = p - q;
373 		newbits = NULL;
374 
375 		/* lookup flag group name */
376 #define	DBG_SET_FLAG_MAJ(S,F)						\
377 		if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0)	\
378 			newbits = &tmpflags->pdb_ ## F;
379 
380 		DBG_SET_FLAG_MAJ("cpu",		CPU);
381 		DBG_SET_FLAG_MAJ("csw",		CSW);
382 		DBG_SET_FLAG_MAJ("logging",	LOG);
383 		DBG_SET_FLAG_MAJ("module",	MOD);
384 		DBG_SET_FLAG_MAJ("md", 		MDP);
385 		DBG_SET_FLAG_MAJ("owner",	OWN);
386 		DBG_SET_FLAG_MAJ("pmc",		PMC);
387 		DBG_SET_FLAG_MAJ("process",	PRC);
388 		DBG_SET_FLAG_MAJ("sampling", 	SAM);
389 
390 		if (newbits == NULL) {
391 			error = EINVAL;
392 			goto done;
393 		}
394 
395 		p++;		/* skip the '=' */
396 
397 		/* Now parse the individual flags */
398 		tmp = 0;
399 	newflag:
400 		for (q = p; p < fence && (c = *p); p++)
401 			if (c == ' ' || c == '\t' || c == ',')
402 				break;
403 
404 		/* p == fence or c == ws or c == "," or c == 0 */
405 
406 		if ((kwlen = p - q) == 0) {
407 			*newbits = tmp;
408 			continue;
409 		}
410 
411 		found = 0;
412 #define	DBG_SET_FLAG_MIN(S,F)						\
413 		if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0)	\
414 			tmp |= found = (1 << PMC_DEBUG_MIN_ ## F)
415 
416 		/* a '*' denotes all possible flags in the group */
417 		if (kwlen == 1 && *q == '*')
418 			tmp = found = ~0;
419 		/* look for individual flag names */
420 		DBG_SET_FLAG_MIN("allocaterow", ALR);
421 		DBG_SET_FLAG_MIN("allocate",	ALL);
422 		DBG_SET_FLAG_MIN("attach",	ATT);
423 		DBG_SET_FLAG_MIN("bind",	BND);
424 		DBG_SET_FLAG_MIN("config",	CFG);
425 		DBG_SET_FLAG_MIN("exec",	EXC);
426 		DBG_SET_FLAG_MIN("exit",	EXT);
427 		DBG_SET_FLAG_MIN("find",	FND);
428 		DBG_SET_FLAG_MIN("flush",	FLS);
429 		DBG_SET_FLAG_MIN("fork",	FRK);
430 		DBG_SET_FLAG_MIN("getbuf",	GTB);
431 		DBG_SET_FLAG_MIN("hook",	PMH);
432 		DBG_SET_FLAG_MIN("init",	INI);
433 		DBG_SET_FLAG_MIN("intr",	INT);
434 		DBG_SET_FLAG_MIN("linktarget",	TLK);
435 		DBG_SET_FLAG_MIN("mayberemove", OMR);
436 		DBG_SET_FLAG_MIN("ops",		OPS);
437 		DBG_SET_FLAG_MIN("read",	REA);
438 		DBG_SET_FLAG_MIN("register",	REG);
439 		DBG_SET_FLAG_MIN("release",	REL);
440 		DBG_SET_FLAG_MIN("remove",	ORM);
441 		DBG_SET_FLAG_MIN("sample",	SAM);
442 		DBG_SET_FLAG_MIN("scheduleio",	SIO);
443 		DBG_SET_FLAG_MIN("select",	SEL);
444 		DBG_SET_FLAG_MIN("signal",	SIG);
445 		DBG_SET_FLAG_MIN("swi",		SWI);
446 		DBG_SET_FLAG_MIN("swo",		SWO);
447 		DBG_SET_FLAG_MIN("start",	STA);
448 		DBG_SET_FLAG_MIN("stop",	STO);
449 		DBG_SET_FLAG_MIN("syscall",	PMS);
450 		DBG_SET_FLAG_MIN("unlinktarget", TUL);
451 		DBG_SET_FLAG_MIN("write",	WRI);
452 		if (found == 0) {
453 			/* unrecognized flag name */
454 			error = EINVAL;
455 			goto done;
456 		}
457 
458 		if (c == 0 || c == ' ' || c == '\t') {	/* end of flag group */
459 			*newbits = tmp;
460 			continue;
461 		}
462 
463 		p++;
464 		goto newflag;
465 	}
466 
467 	/* save the new flag set */
468 	bcopy(tmpflags, &pmc_debugflags, sizeof(pmc_debugflags));
469 
470  done:
471 	free(tmpflags, M_PMC);
472 	return error;
473 }
474 
475 static int
476 pmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS)
477 {
478 	char *fence, *newstr;
479 	int error;
480 	unsigned int n;
481 
482 	(void) arg1; (void) arg2; /* unused parameters */
483 
484 	n = sizeof(pmc_debugstr);
485 	newstr = malloc(n, M_PMC, M_WAITOK|M_ZERO);
486 	(void) strlcpy(newstr, pmc_debugstr, n);
487 
488 	error = sysctl_handle_string(oidp, newstr, n, req);
489 
490 	/* if there is a new string, parse and copy it */
491 	if (error == 0 && req->newptr != NULL) {
492 		fence = newstr + (n < req->newlen ? n : req->newlen + 1);
493 		if ((error = pmc_debugflags_parse(newstr, fence)) == 0)
494 			(void) strlcpy(pmc_debugstr, newstr,
495 			    sizeof(pmc_debugstr));
496 	}
497 
498 	free(newstr, M_PMC);
499 
500 	return error;
501 }
502 #endif
503 
504 /*
505  * Map a row index to a classdep structure and return the adjusted row
506  * index for the PMC class index.
507  */
508 static struct pmc_classdep *
509 pmc_ri_to_classdep(struct pmc_mdep *md, int ri, int *adjri)
510 {
511 	struct pmc_classdep *pcd;
512 
513 	(void) md;
514 
515 	KASSERT(ri >= 0 && ri < md->pmd_npmc,
516 	    ("[pmc,%d] illegal row-index %d", __LINE__, ri));
517 
518 	pcd = pmc_rowindex_to_classdep[ri];
519 
520 	KASSERT(pcd != NULL,
521 	    ("[pmc,%d] ri %d null pcd", __LINE__, ri));
522 
523 	*adjri = ri - pcd->pcd_ri;
524 
525 	KASSERT(*adjri >= 0 && *adjri < pcd->pcd_num,
526 	    ("[pmc,%d] adjusted row-index %d", __LINE__, *adjri));
527 
528 	return (pcd);
529 }
530 
531 /*
532  * Concurrency Control
533  *
534  * The driver manages the following data structures:
535  *
536  *   - target process descriptors, one per target process
537  *   - owner process descriptors (and attached lists), one per owner process
538  *   - lookup hash tables for owner and target processes
539  *   - PMC descriptors (and attached lists)
540  *   - per-cpu hardware state
541  *   - the 'hook' variable through which the kernel calls into
542  *     this module
543  *   - the machine hardware state (managed by the MD layer)
544  *
545  * These data structures are accessed from:
546  *
547  * - thread context-switch code
548  * - interrupt handlers (possibly on multiple cpus)
549  * - kernel threads on multiple cpus running on behalf of user
550  *   processes doing system calls
551  * - this driver's private kernel threads
552  *
553  * = Locks and Locking strategy =
554  *
555  * The driver uses four locking strategies for its operation:
556  *
557  * - The global SX lock "pmc_sx" is used to protect internal
558  *   data structures.
559  *
560  *   Calls into the module by syscall() start with this lock being
561  *   held in exclusive mode.  Depending on the requested operation,
562  *   the lock may be downgraded to 'shared' mode to allow more
563  *   concurrent readers into the module.  Calls into the module from
564  *   other parts of the kernel acquire the lock in shared mode.
565  *
566  *   This SX lock is held in exclusive mode for any operations that
567  *   modify the linkages between the driver's internal data structures.
568  *
569  *   The 'pmc_hook' function pointer is also protected by this lock.
570  *   It is only examined with the sx lock held in exclusive mode.  The
571  *   kernel module is allowed to be unloaded only with the sx lock held
572  *   in exclusive mode.  In normal syscall handling, after acquiring the
573  *   pmc_sx lock we first check that 'pmc_hook' is non-null before
574  *   proceeding.  This prevents races between the thread unloading the module
575  *   and other threads seeking to use the module.
576  *
577  * - Lookups of target process structures and owner process structures
578  *   cannot use the global "pmc_sx" SX lock because these lookups need
579  *   to happen during context switches and in other critical sections
580  *   where sleeping is not allowed.  We protect these lookup tables
581  *   with their own private spin-mutexes, "pmc_processhash_mtx" and
582  *   "pmc_ownerhash_mtx".
583  *
584  * - Interrupt handlers work in a lock free manner.  At interrupt
585  *   time, handlers look at the PMC pointer (phw->phw_pmc) configured
586  *   when the PMC was started.  If this pointer is NULL, the interrupt
587  *   is ignored after updating driver statistics.  We ensure that this
588  *   pointer is set (using an atomic operation if necessary) before the
589  *   PMC hardware is started.  Conversely, this pointer is unset atomically
590  *   only after the PMC hardware is stopped.
591  *
592  *   We ensure that everything needed for the operation of an
593  *   interrupt handler is available without it needing to acquire any
594  *   locks.  We also ensure that a PMC's software state is destroyed only
595  *   after the PMC is taken off hardware (on all CPUs).
596  *
597  * - Context-switch handling with process-private PMCs needs more
598  *   care.
599  *
600  *   A given process may be the target of multiple PMCs.  For example,
601  *   PMCATTACH and PMCDETACH may be requested by a process on one CPU
602  *   while the target process is running on another.  A PMC could also
603  *   be getting released because its owner is exiting.  We tackle
604  *   these situations in the following manner:
605  *
606  *   - each target process structure 'pmc_process' has an array
607  *     of 'struct pmc *' pointers, one for each hardware PMC.
608  *
609  *   - At context switch IN time, each "target" PMC in RUNNING state
610  *     gets started on hardware and a pointer to each PMC is copied into
611  *     the per-cpu phw array.  The 'runcount' for the PMC is
612  *     incremented.
613  *
614  *   - At context switch OUT time, all process-virtual PMCs are stopped
615  *     on hardware.  The saved value is added to the PMCs value field
616  *     only if the PMC is in a non-deleted state (the PMCs state could
617  *     have changed during the current time slice).
618  *
619  *     Note that since in-between a switch IN on a processor and a switch
620  *     OUT, the PMC could have been released on another CPU.  Therefore
621  *     context switch OUT always looks at the hardware state to turn
622  *     OFF PMCs and will update a PMC's saved value only if reachable
623  *     from the target process record.
624  *
625  *   - OP PMCRELEASE could be called on a PMC at any time (the PMC could
626  *     be attached to many processes at the time of the call and could
627  *     be active on multiple CPUs).
628  *
629  *     We prevent further scheduling of the PMC by marking it as in
630  *     state 'DELETED'.  If the runcount of the PMC is non-zero then
631  *     this PMC is currently running on a CPU somewhere.  The thread
632  *     doing the PMCRELEASE operation waits by repeatedly doing a
633  *     pause() till the runcount comes to zero.
634  *
635  * The contents of a PMC descriptor (struct pmc) are protected using
636  * a spin-mutex.  In order to save space, we use a mutex pool.
637  *
638  * In terms of lock types used by witness(4), we use:
639  * - Type "pmc-sx", used by the global SX lock.
640  * - Type "pmc-sleep", for sleep mutexes used by logger threads.
641  * - Type "pmc-per-proc", for protecting PMC owner descriptors.
642  * - Type "pmc-leaf", used for all other spin mutexes.
643  */
644 
645 /*
646  * save the cpu binding of the current kthread
647  */
648 
649 static void
650 pmc_save_cpu_binding(struct pmc_binding *pb)
651 {
652 	PMCDBG(CPU,BND,2, "%s", "save-cpu");
653 	thread_lock(curthread);
654 	pb->pb_bound = sched_is_bound(curthread);
655 	pb->pb_cpu   = curthread->td_oncpu;
656 	thread_unlock(curthread);
657 	PMCDBG(CPU,BND,2, "save-cpu cpu=%d", pb->pb_cpu);
658 }
659 
660 /*
661  * restore the cpu binding of the current thread
662  */
663 
664 static void
665 pmc_restore_cpu_binding(struct pmc_binding *pb)
666 {
667 	PMCDBG(CPU,BND,2, "restore-cpu curcpu=%d restore=%d",
668 	    curthread->td_oncpu, pb->pb_cpu);
669 	thread_lock(curthread);
670 	if (pb->pb_bound)
671 		sched_bind(curthread, pb->pb_cpu);
672 	else
673 		sched_unbind(curthread);
674 	thread_unlock(curthread);
675 	PMCDBG(CPU,BND,2, "%s", "restore-cpu done");
676 }
677 
678 /*
679  * move execution over the specified cpu and bind it there.
680  */
681 
682 static void
683 pmc_select_cpu(int cpu)
684 {
685 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
686 	    ("[pmc,%d] bad cpu number %d", __LINE__, cpu));
687 
688 	/* Never move to an inactive CPU. */
689 	KASSERT(pmc_cpu_is_active(cpu), ("[pmc,%d] selecting inactive "
690 	    "CPU %d", __LINE__, cpu));
691 
692 	PMCDBG(CPU,SEL,2, "select-cpu cpu=%d", cpu);
693 	thread_lock(curthread);
694 	sched_bind(curthread, cpu);
695 	thread_unlock(curthread);
696 
697 	KASSERT(curthread->td_oncpu == cpu,
698 	    ("[pmc,%d] CPU not bound [cpu=%d, curr=%d]", __LINE__,
699 		cpu, curthread->td_oncpu));
700 
701 	PMCDBG(CPU,SEL,2, "select-cpu cpu=%d ok", cpu);
702 }
703 
704 /*
705  * Force a context switch.
706  *
707  * We do this by pause'ing for 1 tick -- invoking mi_switch() is not
708  * guaranteed to force a context switch.
709  */
710 
711 static void
712 pmc_force_context_switch(void)
713 {
714 
715 	pause("pmcctx", 1);
716 }
717 
718 /*
719  * Get the file name for an executable.  This is a simple wrapper
720  * around vn_fullpath(9).
721  */
722 
723 static void
724 pmc_getfilename(struct vnode *v, char **fullpath, char **freepath)
725 {
726 
727 	*fullpath = "unknown";
728 	*freepath = NULL;
729 	vn_fullpath(curthread, v, fullpath, freepath);
730 }
731 
732 /*
733  * remove an process owning PMCs
734  */
735 
736 void
737 pmc_remove_owner(struct pmc_owner *po)
738 {
739 	struct pmc *pm, *tmp;
740 
741 	sx_assert(&pmc_sx, SX_XLOCKED);
742 
743 	PMCDBG(OWN,ORM,1, "remove-owner po=%p", po);
744 
745 	/* Remove descriptor from the owner hash table */
746 	LIST_REMOVE(po, po_next);
747 
748 	/* release all owned PMC descriptors */
749 	LIST_FOREACH_SAFE(pm, &po->po_pmcs, pm_next, tmp) {
750 		PMCDBG(OWN,ORM,2, "pmc=%p", pm);
751 		KASSERT(pm->pm_owner == po,
752 		    ("[pmc,%d] owner %p != po %p", __LINE__, pm->pm_owner, po));
753 
754 		pmc_release_pmc_descriptor(pm);	/* will unlink from the list */
755 	}
756 
757 	KASSERT(po->po_sscount == 0,
758 	    ("[pmc,%d] SS count not zero", __LINE__));
759 	KASSERT(LIST_EMPTY(&po->po_pmcs),
760 	    ("[pmc,%d] PMC list not empty", __LINE__));
761 
762 	/* de-configure the log file if present */
763 	if (po->po_flags & PMC_PO_OWNS_LOGFILE)
764 		pmclog_deconfigure_log(po);
765 }
766 
767 /*
768  * remove an owner process record if all conditions are met.
769  */
770 
771 static void
772 pmc_maybe_remove_owner(struct pmc_owner *po)
773 {
774 
775 	PMCDBG(OWN,OMR,1, "maybe-remove-owner po=%p", po);
776 
777 	/*
778 	 * Remove owner record if
779 	 * - this process does not own any PMCs
780 	 * - this process has not allocated a system-wide sampling buffer
781 	 */
782 
783 	if (LIST_EMPTY(&po->po_pmcs) &&
784 	    ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)) {
785 		pmc_remove_owner(po);
786 		pmc_destroy_owner_descriptor(po);
787 	}
788 }
789 
790 /*
791  * Add an association between a target process and a PMC.
792  */
793 
794 static void
795 pmc_link_target_process(struct pmc *pm, struct pmc_process *pp)
796 {
797 	int ri;
798 	struct pmc_target *pt;
799 
800 	sx_assert(&pmc_sx, SX_XLOCKED);
801 
802 	KASSERT(pm != NULL && pp != NULL,
803 	    ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp));
804 	KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)),
805 	    ("[pmc,%d] Attaching a non-process-virtual pmc=%p to pid=%d",
806 		__LINE__, pm, pp->pp_proc->p_pid));
807 	KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= ((int) md->pmd_npmc - 1),
808 	    ("[pmc,%d] Illegal reference count %d for process record %p",
809 		__LINE__, pp->pp_refcnt, (void *) pp));
810 
811 	ri = PMC_TO_ROWINDEX(pm);
812 
813 	PMCDBG(PRC,TLK,1, "link-target pmc=%p ri=%d pmc-process=%p",
814 	    pm, ri, pp);
815 
816 #ifdef	DEBUG
817 	LIST_FOREACH(pt, &pm->pm_targets, pt_next)
818 	    if (pt->pt_process == pp)
819 		    KASSERT(0, ("[pmc,%d] pp %p already in pmc %p targets",
820 				__LINE__, pp, pm));
821 #endif
822 
823 	pt = malloc(sizeof(struct pmc_target), M_PMC, M_WAITOK|M_ZERO);
824 	pt->pt_process = pp;
825 
826 	LIST_INSERT_HEAD(&pm->pm_targets, pt, pt_next);
827 
828 	atomic_store_rel_ptr((uintptr_t *)&pp->pp_pmcs[ri].pp_pmc,
829 	    (uintptr_t)pm);
830 
831 	if (pm->pm_owner->po_owner == pp->pp_proc)
832 		pm->pm_flags |= PMC_F_ATTACHED_TO_OWNER;
833 
834 	/*
835 	 * Initialize the per-process values at this row index.
836 	 */
837 	pp->pp_pmcs[ri].pp_pmcval = PMC_TO_MODE(pm) == PMC_MODE_TS ?
838 	    pm->pm_sc.pm_reloadcount : 0;
839 
840 	pp->pp_refcnt++;
841 
842 }
843 
844 /*
845  * Removes the association between a target process and a PMC.
846  */
847 
848 static void
849 pmc_unlink_target_process(struct pmc *pm, struct pmc_process *pp)
850 {
851 	int ri;
852 	struct proc *p;
853 	struct pmc_target *ptgt;
854 
855 	sx_assert(&pmc_sx, SX_XLOCKED);
856 
857 	KASSERT(pm != NULL && pp != NULL,
858 	    ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp));
859 
860 	KASSERT(pp->pp_refcnt >= 1 && pp->pp_refcnt <= (int) md->pmd_npmc,
861 	    ("[pmc,%d] Illegal ref count %d on process record %p",
862 		__LINE__, pp->pp_refcnt, (void *) pp));
863 
864 	ri = PMC_TO_ROWINDEX(pm);
865 
866 	PMCDBG(PRC,TUL,1, "unlink-target pmc=%p ri=%d pmc-process=%p",
867 	    pm, ri, pp);
868 
869 	KASSERT(pp->pp_pmcs[ri].pp_pmc == pm,
870 	    ("[pmc,%d] PMC ri %d mismatch pmc %p pp->[ri] %p", __LINE__,
871 		ri, pm, pp->pp_pmcs[ri].pp_pmc));
872 
873 	pp->pp_pmcs[ri].pp_pmc = NULL;
874 	pp->pp_pmcs[ri].pp_pmcval = (pmc_value_t) 0;
875 
876 	/* Remove owner-specific flags */
877 	if (pm->pm_owner->po_owner == pp->pp_proc) {
878 		pp->pp_flags &= ~PMC_PP_ENABLE_MSR_ACCESS;
879 		pm->pm_flags &= ~PMC_F_ATTACHED_TO_OWNER;
880 	}
881 
882 	pp->pp_refcnt--;
883 
884 	/* Remove the target process from the PMC structure */
885 	LIST_FOREACH(ptgt, &pm->pm_targets, pt_next)
886 		if (ptgt->pt_process == pp)
887 			break;
888 
889 	KASSERT(ptgt != NULL, ("[pmc,%d] process %p (pp: %p) not found "
890 		    "in pmc %p", __LINE__, pp->pp_proc, pp, pm));
891 
892 	LIST_REMOVE(ptgt, pt_next);
893 	free(ptgt, M_PMC);
894 
895 	/* if the PMC now lacks targets, send the owner a SIGIO */
896 	if (LIST_EMPTY(&pm->pm_targets)) {
897 		p = pm->pm_owner->po_owner;
898 		PROC_LOCK(p);
899 		kern_psignal(p, SIGIO);
900 		PROC_UNLOCK(p);
901 
902 		PMCDBG(PRC,SIG,2, "signalling proc=%p signal=%d", p,
903 		    SIGIO);
904 	}
905 }
906 
907 /*
908  * Check if PMC 'pm' may be attached to target process 't'.
909  */
910 
911 static int
912 pmc_can_attach(struct pmc *pm, struct proc *t)
913 {
914 	struct proc *o;		/* pmc owner */
915 	struct ucred *oc, *tc;	/* owner, target credentials */
916 	int decline_attach, i;
917 
918 	/*
919 	 * A PMC's owner can always attach that PMC to itself.
920 	 */
921 
922 	if ((o = pm->pm_owner->po_owner) == t)
923 		return 0;
924 
925 	PROC_LOCK(o);
926 	oc = o->p_ucred;
927 	crhold(oc);
928 	PROC_UNLOCK(o);
929 
930 	PROC_LOCK(t);
931 	tc = t->p_ucred;
932 	crhold(tc);
933 	PROC_UNLOCK(t);
934 
935 	/*
936 	 * The effective uid of the PMC owner should match at least one
937 	 * of the {effective,real,saved} uids of the target process.
938 	 */
939 
940 	decline_attach = oc->cr_uid != tc->cr_uid &&
941 	    oc->cr_uid != tc->cr_svuid &&
942 	    oc->cr_uid != tc->cr_ruid;
943 
944 	/*
945 	 * Every one of the target's group ids, must be in the owner's
946 	 * group list.
947 	 */
948 	for (i = 0; !decline_attach && i < tc->cr_ngroups; i++)
949 		decline_attach = !groupmember(tc->cr_groups[i], oc);
950 
951 	/* check the read and saved gids too */
952 	if (decline_attach == 0)
953 		decline_attach = !groupmember(tc->cr_rgid, oc) ||
954 		    !groupmember(tc->cr_svgid, oc);
955 
956 	crfree(tc);
957 	crfree(oc);
958 
959 	return !decline_attach;
960 }
961 
962 /*
963  * Attach a process to a PMC.
964  */
965 
966 static int
967 pmc_attach_one_process(struct proc *p, struct pmc *pm)
968 {
969 	int ri;
970 	char *fullpath, *freepath;
971 	struct pmc_process	*pp;
972 
973 	sx_assert(&pmc_sx, SX_XLOCKED);
974 
975 	PMCDBG(PRC,ATT,2, "attach-one pm=%p ri=%d proc=%p (%d, %s)", pm,
976 	    PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm);
977 
978 	/*
979 	 * Locate the process descriptor corresponding to process 'p',
980 	 * allocating space as needed.
981 	 *
982 	 * Verify that rowindex 'pm_rowindex' is free in the process
983 	 * descriptor.
984 	 *
985 	 * If not, allocate space for a descriptor and link the
986 	 * process descriptor and PMC.
987 	 */
988 	ri = PMC_TO_ROWINDEX(pm);
989 
990 	if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_ALLOCATE)) == NULL)
991 		return ENOMEM;
992 
993 	if (pp->pp_pmcs[ri].pp_pmc == pm) /* already present at slot [ri] */
994 		return EEXIST;
995 
996 	if (pp->pp_pmcs[ri].pp_pmc != NULL)
997 		return EBUSY;
998 
999 	pmc_link_target_process(pm, pp);
1000 
1001 	if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) &&
1002 	    (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) == 0)
1003 		pm->pm_flags |= PMC_F_NEEDS_LOGFILE;
1004 
1005 	pm->pm_flags |= PMC_F_ATTACH_DONE; /* mark as attached */
1006 
1007 	/* issue an attach event to a configured log file */
1008 	if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE) {
1009 		pmc_getfilename(p->p_textvp, &fullpath, &freepath);
1010 		if (p->p_flag & P_KTHREAD) {
1011 			fullpath = kernelname;
1012 			freepath = NULL;
1013 		} else
1014 			pmclog_process_pmcattach(pm, p->p_pid, fullpath);
1015 		if (freepath)
1016 			free(freepath, M_TEMP);
1017 		if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1018 			pmc_log_process_mappings(pm->pm_owner, p);
1019 	}
1020 	/* mark process as using HWPMCs */
1021 	PROC_LOCK(p);
1022 	p->p_flag |= P_HWPMC;
1023 	PROC_UNLOCK(p);
1024 
1025 	return 0;
1026 }
1027 
1028 /*
1029  * Attach a process and optionally its children
1030  */
1031 
1032 static int
1033 pmc_attach_process(struct proc *p, struct pmc *pm)
1034 {
1035 	int error;
1036 	struct proc *top;
1037 
1038 	sx_assert(&pmc_sx, SX_XLOCKED);
1039 
1040 	PMCDBG(PRC,ATT,1, "attach pm=%p ri=%d proc=%p (%d, %s)", pm,
1041 	    PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm);
1042 
1043 
1044 	/*
1045 	 * If this PMC successfully allowed a GETMSR operation
1046 	 * in the past, disallow further ATTACHes.
1047 	 */
1048 
1049 	if ((pm->pm_flags & PMC_PP_ENABLE_MSR_ACCESS) != 0)
1050 		return EPERM;
1051 
1052 	if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0)
1053 		return pmc_attach_one_process(p, pm);
1054 
1055 	/*
1056 	 * Traverse all child processes, attaching them to
1057 	 * this PMC.
1058 	 */
1059 
1060 	sx_slock(&proctree_lock);
1061 
1062 	top = p;
1063 
1064 	for (;;) {
1065 		if ((error = pmc_attach_one_process(p, pm)) != 0)
1066 			break;
1067 		if (!LIST_EMPTY(&p->p_children))
1068 			p = LIST_FIRST(&p->p_children);
1069 		else for (;;) {
1070 			if (p == top)
1071 				goto done;
1072 			if (LIST_NEXT(p, p_sibling)) {
1073 				p = LIST_NEXT(p, p_sibling);
1074 				break;
1075 			}
1076 			p = p->p_pptr;
1077 		}
1078 	}
1079 
1080 	if (error)
1081 		(void) pmc_detach_process(top, pm);
1082 
1083  done:
1084 	sx_sunlock(&proctree_lock);
1085 	return error;
1086 }
1087 
1088 /*
1089  * Detach a process from a PMC.  If there are no other PMCs tracking
1090  * this process, remove the process structure from its hash table.  If
1091  * 'flags' contains PMC_FLAG_REMOVE, then free the process structure.
1092  */
1093 
1094 static int
1095 pmc_detach_one_process(struct proc *p, struct pmc *pm, int flags)
1096 {
1097 	int ri;
1098 	struct pmc_process *pp;
1099 
1100 	sx_assert(&pmc_sx, SX_XLOCKED);
1101 
1102 	KASSERT(pm != NULL,
1103 	    ("[pmc,%d] null pm pointer", __LINE__));
1104 
1105 	ri = PMC_TO_ROWINDEX(pm);
1106 
1107 	PMCDBG(PRC,ATT,2, "detach-one pm=%p ri=%d proc=%p (%d, %s) flags=0x%x",
1108 	    pm, ri, p, p->p_pid, p->p_comm, flags);
1109 
1110 	if ((pp = pmc_find_process_descriptor(p, 0)) == NULL)
1111 		return ESRCH;
1112 
1113 	if (pp->pp_pmcs[ri].pp_pmc != pm)
1114 		return EINVAL;
1115 
1116 	pmc_unlink_target_process(pm, pp);
1117 
1118 	/* Issue a detach entry if a log file is configured */
1119 	if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE)
1120 		pmclog_process_pmcdetach(pm, p->p_pid);
1121 
1122 	/*
1123 	 * If there are no PMCs targetting this process, we remove its
1124 	 * descriptor from the target hash table and unset the P_HWPMC
1125 	 * flag in the struct proc.
1126 	 */
1127 	KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= (int) md->pmd_npmc,
1128 	    ("[pmc,%d] Illegal refcnt %d for process struct %p",
1129 		__LINE__, pp->pp_refcnt, pp));
1130 
1131 	if (pp->pp_refcnt != 0)	/* still a target of some PMC */
1132 		return 0;
1133 
1134 	pmc_remove_process_descriptor(pp);
1135 
1136 	if (flags & PMC_FLAG_REMOVE)
1137 		free(pp, M_PMC);
1138 
1139 	PROC_LOCK(p);
1140 	p->p_flag &= ~P_HWPMC;
1141 	PROC_UNLOCK(p);
1142 
1143 	return 0;
1144 }
1145 
1146 /*
1147  * Detach a process and optionally its descendants from a PMC.
1148  */
1149 
1150 static int
1151 pmc_detach_process(struct proc *p, struct pmc *pm)
1152 {
1153 	struct proc *top;
1154 
1155 	sx_assert(&pmc_sx, SX_XLOCKED);
1156 
1157 	PMCDBG(PRC,ATT,1, "detach pm=%p ri=%d proc=%p (%d, %s)", pm,
1158 	    PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm);
1159 
1160 	if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0)
1161 		return pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE);
1162 
1163 	/*
1164 	 * Traverse all children, detaching them from this PMC.  We
1165 	 * ignore errors since we could be detaching a PMC from a
1166 	 * partially attached proc tree.
1167 	 */
1168 
1169 	sx_slock(&proctree_lock);
1170 
1171 	top = p;
1172 
1173 	for (;;) {
1174 		(void) pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE);
1175 
1176 		if (!LIST_EMPTY(&p->p_children))
1177 			p = LIST_FIRST(&p->p_children);
1178 		else for (;;) {
1179 			if (p == top)
1180 				goto done;
1181 			if (LIST_NEXT(p, p_sibling)) {
1182 				p = LIST_NEXT(p, p_sibling);
1183 				break;
1184 			}
1185 			p = p->p_pptr;
1186 		}
1187 	}
1188 
1189  done:
1190 	sx_sunlock(&proctree_lock);
1191 
1192 	if (LIST_EMPTY(&pm->pm_targets))
1193 		pm->pm_flags &= ~PMC_F_ATTACH_DONE;
1194 
1195 	return 0;
1196 }
1197 
1198 
1199 /*
1200  * Thread context switch IN
1201  */
1202 
1203 static void
1204 pmc_process_csw_in(struct thread *td)
1205 {
1206 	int cpu;
1207 	unsigned int adjri, ri;
1208 	struct pmc *pm;
1209 	struct proc *p;
1210 	struct pmc_cpu *pc;
1211 	struct pmc_hw *phw;
1212 	pmc_value_t newvalue;
1213 	struct pmc_process *pp;
1214 	struct pmc_classdep *pcd;
1215 
1216 	p = td->td_proc;
1217 
1218 	if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE)) == NULL)
1219 		return;
1220 
1221 	KASSERT(pp->pp_proc == td->td_proc,
1222 	    ("[pmc,%d] not my thread state", __LINE__));
1223 
1224 	critical_enter(); /* no preemption from this point */
1225 
1226 	cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */
1227 
1228 	PMCDBG(CSW,SWI,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p,
1229 	    p->p_pid, p->p_comm, pp);
1230 
1231 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
1232 	    ("[pmc,%d] wierd CPU id %d", __LINE__, cpu));
1233 
1234 	pc = pmc_pcpu[cpu];
1235 
1236 	for (ri = 0; ri < md->pmd_npmc; ri++) {
1237 
1238 		if ((pm = pp->pp_pmcs[ri].pp_pmc) == NULL)
1239 			continue;
1240 
1241 		KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)),
1242 		    ("[pmc,%d] Target PMC in non-virtual mode (%d)",
1243 			__LINE__, PMC_TO_MODE(pm)));
1244 
1245 		KASSERT(PMC_TO_ROWINDEX(pm) == ri,
1246 		    ("[pmc,%d] Row index mismatch pmc %d != ri %d",
1247 			__LINE__, PMC_TO_ROWINDEX(pm), ri));
1248 
1249 		/*
1250 		 * Only PMCs that are marked as 'RUNNING' need
1251 		 * be placed on hardware.
1252 		 */
1253 
1254 		if (pm->pm_state != PMC_STATE_RUNNING)
1255 			continue;
1256 
1257 		/* increment PMC runcount */
1258 		atomic_add_rel_int(&pm->pm_runcount, 1);
1259 
1260 		/* configure the HWPMC we are going to use. */
1261 		pcd = pmc_ri_to_classdep(md, ri, &adjri);
1262 		pcd->pcd_config_pmc(cpu, adjri, pm);
1263 
1264 		phw = pc->pc_hwpmcs[ri];
1265 
1266 		KASSERT(phw != NULL,
1267 		    ("[pmc,%d] null hw pointer", __LINE__));
1268 
1269 		KASSERT(phw->phw_pmc == pm,
1270 		    ("[pmc,%d] hw->pmc %p != pmc %p", __LINE__,
1271 			phw->phw_pmc, pm));
1272 
1273 		/*
1274 		 * Write out saved value and start the PMC.
1275 		 *
1276 		 * Sampling PMCs use a per-process value, while
1277 		 * counting mode PMCs use a per-pmc value that is
1278 		 * inherited across descendants.
1279 		 */
1280 		if (PMC_TO_MODE(pm) == PMC_MODE_TS) {
1281 			mtx_pool_lock_spin(pmc_mtxpool, pm);
1282 			newvalue = PMC_PCPU_SAVED(cpu,ri) =
1283 			    pp->pp_pmcs[ri].pp_pmcval;
1284 			mtx_pool_unlock_spin(pmc_mtxpool, pm);
1285 		} else {
1286 			KASSERT(PMC_TO_MODE(pm) == PMC_MODE_TC,
1287 			    ("[pmc,%d] illegal mode=%d", __LINE__,
1288 			    PMC_TO_MODE(pm)));
1289 			mtx_pool_lock_spin(pmc_mtxpool, pm);
1290 			newvalue = PMC_PCPU_SAVED(cpu, ri) =
1291 			    pm->pm_gv.pm_savedvalue;
1292 			mtx_pool_unlock_spin(pmc_mtxpool, pm);
1293 		}
1294 
1295 		PMCDBG(CSW,SWI,1,"cpu=%d ri=%d new=%jd", cpu, ri, newvalue);
1296 
1297 		pcd->pcd_write_pmc(cpu, adjri, newvalue);
1298 		pcd->pcd_start_pmc(cpu, adjri);
1299 	}
1300 
1301 	/*
1302 	 * perform any other architecture/cpu dependent thread
1303 	 * switch-in actions.
1304 	 */
1305 
1306 	(void) (*md->pmd_switch_in)(pc, pp);
1307 
1308 	critical_exit();
1309 
1310 }
1311 
1312 /*
1313  * Thread context switch OUT.
1314  */
1315 
1316 static void
1317 pmc_process_csw_out(struct thread *td)
1318 {
1319 	int cpu;
1320 	int64_t tmp;
1321 	struct pmc *pm;
1322 	struct proc *p;
1323 	enum pmc_mode mode;
1324 	struct pmc_cpu *pc;
1325 	pmc_value_t newvalue;
1326 	unsigned int adjri, ri;
1327 	struct pmc_process *pp;
1328 	struct pmc_classdep *pcd;
1329 
1330 
1331 	/*
1332 	 * Locate our process descriptor; this may be NULL if
1333 	 * this process is exiting and we have already removed
1334 	 * the process from the target process table.
1335 	 *
1336 	 * Note that due to kernel preemption, multiple
1337 	 * context switches may happen while the process is
1338 	 * exiting.
1339 	 *
1340 	 * Note also that if the target process cannot be
1341 	 * found we still need to deconfigure any PMCs that
1342 	 * are currently running on hardware.
1343 	 */
1344 
1345 	p = td->td_proc;
1346 	pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE);
1347 
1348 	/*
1349 	 * save PMCs
1350 	 */
1351 
1352 	critical_enter();
1353 
1354 	cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */
1355 
1356 	PMCDBG(CSW,SWO,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p,
1357 	    p->p_pid, p->p_comm, pp);
1358 
1359 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
1360 	    ("[pmc,%d wierd CPU id %d", __LINE__, cpu));
1361 
1362 	pc = pmc_pcpu[cpu];
1363 
1364 	/*
1365 	 * When a PMC gets unlinked from a target PMC, it will
1366 	 * be removed from the target's pp_pmc[] array.
1367 	 *
1368 	 * However, on a MP system, the target could have been
1369 	 * executing on another CPU at the time of the unlink.
1370 	 * So, at context switch OUT time, we need to look at
1371 	 * the hardware to determine if a PMC is scheduled on
1372 	 * it.
1373 	 */
1374 
1375 	for (ri = 0; ri < md->pmd_npmc; ri++) {
1376 
1377 		pcd = pmc_ri_to_classdep(md, ri, &adjri);
1378 		pm  = NULL;
1379 		(void) (*pcd->pcd_get_config)(cpu, adjri, &pm);
1380 
1381 		if (pm == NULL)	/* nothing at this row index */
1382 			continue;
1383 
1384 		mode = PMC_TO_MODE(pm);
1385 		if (!PMC_IS_VIRTUAL_MODE(mode))
1386 			continue; /* not a process virtual PMC */
1387 
1388 		KASSERT(PMC_TO_ROWINDEX(pm) == ri,
1389 		    ("[pmc,%d] ri mismatch pmc(%d) ri(%d)",
1390 			__LINE__, PMC_TO_ROWINDEX(pm), ri));
1391 
1392 		/* Stop hardware if not already stopped */
1393 		if (pm->pm_stalled == 0)
1394 			pcd->pcd_stop_pmc(cpu, adjri);
1395 
1396 		/* reduce this PMC's runcount */
1397 		atomic_subtract_rel_int(&pm->pm_runcount, 1);
1398 
1399 		/*
1400 		 * If this PMC is associated with this process,
1401 		 * save the reading.
1402 		 */
1403 
1404 		if (pp != NULL && pp->pp_pmcs[ri].pp_pmc != NULL) {
1405 
1406 			KASSERT(pm == pp->pp_pmcs[ri].pp_pmc,
1407 			    ("[pmc,%d] pm %p != pp_pmcs[%d] %p", __LINE__,
1408 				pm, ri, pp->pp_pmcs[ri].pp_pmc));
1409 
1410 			KASSERT(pp->pp_refcnt > 0,
1411 			    ("[pmc,%d] pp refcnt = %d", __LINE__,
1412 				pp->pp_refcnt));
1413 
1414 			pcd->pcd_read_pmc(cpu, adjri, &newvalue);
1415 
1416 			tmp = newvalue - PMC_PCPU_SAVED(cpu,ri);
1417 
1418 			PMCDBG(CSW,SWO,1,"cpu=%d ri=%d tmp=%jd", cpu, ri,
1419 			    tmp);
1420 
1421 			if (mode == PMC_MODE_TS) {
1422 
1423 				/*
1424 				 * For sampling process-virtual PMCs,
1425 				 * we expect the count to be
1426 				 * decreasing as the 'value'
1427 				 * programmed into the PMC is the
1428 				 * number of events to be seen till
1429 				 * the next sampling interrupt.
1430 				 */
1431 				if (tmp < 0)
1432 					tmp += pm->pm_sc.pm_reloadcount;
1433 				mtx_pool_lock_spin(pmc_mtxpool, pm);
1434 				pp->pp_pmcs[ri].pp_pmcval -= tmp;
1435 				if ((int64_t) pp->pp_pmcs[ri].pp_pmcval < 0)
1436 					pp->pp_pmcs[ri].pp_pmcval +=
1437 					    pm->pm_sc.pm_reloadcount;
1438 				mtx_pool_unlock_spin(pmc_mtxpool, pm);
1439 
1440 			} else {
1441 
1442 				/*
1443 				 * For counting process-virtual PMCs,
1444 				 * we expect the count to be
1445 				 * increasing monotonically, modulo a 64
1446 				 * bit wraparound.
1447 				 */
1448 				KASSERT((int64_t) tmp >= 0,
1449 				    ("[pmc,%d] negative increment cpu=%d "
1450 				     "ri=%d newvalue=%jx saved=%jx "
1451 				     "incr=%jx", __LINE__, cpu, ri,
1452 				     newvalue, PMC_PCPU_SAVED(cpu,ri), tmp));
1453 
1454 				mtx_pool_lock_spin(pmc_mtxpool, pm);
1455 				pm->pm_gv.pm_savedvalue += tmp;
1456 				pp->pp_pmcs[ri].pp_pmcval += tmp;
1457 				mtx_pool_unlock_spin(pmc_mtxpool, pm);
1458 
1459 				if (pm->pm_flags & PMC_F_LOG_PROCCSW)
1460 					pmclog_process_proccsw(pm, pp, tmp);
1461 			}
1462 		}
1463 
1464 		/* mark hardware as free */
1465 		pcd->pcd_config_pmc(cpu, adjri, NULL);
1466 	}
1467 
1468 	/*
1469 	 * perform any other architecture/cpu dependent thread
1470 	 * switch out functions.
1471 	 */
1472 
1473 	(void) (*md->pmd_switch_out)(pc, pp);
1474 
1475 	critical_exit();
1476 }
1477 
1478 /*
1479  * Log a KLD operation.
1480  */
1481 
1482 static void
1483 pmc_process_kld_load(struct pmckern_map_in *pkm)
1484 {
1485 	struct pmc_owner *po;
1486 
1487 	sx_assert(&pmc_sx, SX_LOCKED);
1488 
1489 	/*
1490 	 * Notify owners of system sampling PMCs about KLD operations.
1491 	 */
1492 
1493 	LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1494 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1495 	    	pmclog_process_map_in(po, (pid_t) -1, pkm->pm_address,
1496 		    (char *) pkm->pm_file);
1497 
1498 	/*
1499 	 * TODO: Notify owners of (all) process-sampling PMCs too.
1500 	 */
1501 
1502 	return;
1503 }
1504 
1505 static void
1506 pmc_process_kld_unload(struct pmckern_map_out *pkm)
1507 {
1508 	struct pmc_owner *po;
1509 
1510 	sx_assert(&pmc_sx, SX_LOCKED);
1511 
1512 	LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1513 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1514 		pmclog_process_map_out(po, (pid_t) -1,
1515 		    pkm->pm_address, pkm->pm_address + pkm->pm_size);
1516 
1517 	/*
1518 	 * TODO: Notify owners of process-sampling PMCs.
1519 	 */
1520 }
1521 
1522 /*
1523  * A mapping change for a process.
1524  */
1525 
1526 static void
1527 pmc_process_mmap(struct thread *td, struct pmckern_map_in *pkm)
1528 {
1529 	int ri;
1530 	pid_t pid;
1531 	char *fullpath, *freepath;
1532 	const struct pmc *pm;
1533 	struct pmc_owner *po;
1534 	const struct pmc_process *pp;
1535 
1536 	freepath = fullpath = NULL;
1537 	pmc_getfilename((struct vnode *) pkm->pm_file, &fullpath, &freepath);
1538 
1539 	pid = td->td_proc->p_pid;
1540 
1541 	/* Inform owners of all system-wide sampling PMCs. */
1542 	LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1543 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1544 		pmclog_process_map_in(po, pid, pkm->pm_address, fullpath);
1545 
1546 	if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL)
1547 		goto done;
1548 
1549 	/*
1550 	 * Inform sampling PMC owners tracking this process.
1551 	 */
1552 	for (ri = 0; ri < md->pmd_npmc; ri++)
1553 		if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL &&
1554 		    PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1555 			pmclog_process_map_in(pm->pm_owner,
1556 			    pid, pkm->pm_address, fullpath);
1557 
1558   done:
1559 	if (freepath)
1560 		free(freepath, M_TEMP);
1561 }
1562 
1563 
1564 /*
1565  * Log an munmap request.
1566  */
1567 
1568 static void
1569 pmc_process_munmap(struct thread *td, struct pmckern_map_out *pkm)
1570 {
1571 	int ri;
1572 	pid_t pid;
1573 	struct pmc_owner *po;
1574 	const struct pmc *pm;
1575 	const struct pmc_process *pp;
1576 
1577 	pid = td->td_proc->p_pid;
1578 
1579 	LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1580 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1581 		pmclog_process_map_out(po, pid, pkm->pm_address,
1582 		    pkm->pm_address + pkm->pm_size);
1583 
1584 	if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL)
1585 		return;
1586 
1587 	for (ri = 0; ri < md->pmd_npmc; ri++)
1588 		if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL &&
1589 		    PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1590 			pmclog_process_map_out(pm->pm_owner, pid,
1591 			    pkm->pm_address, pkm->pm_address + pkm->pm_size);
1592 }
1593 
1594 /*
1595  * Log mapping information about the kernel.
1596  */
1597 
1598 static void
1599 pmc_log_kernel_mappings(struct pmc *pm)
1600 {
1601 	struct pmc_owner *po;
1602 	struct pmckern_map_in *km, *kmbase;
1603 
1604 	sx_assert(&pmc_sx, SX_LOCKED);
1605 	KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)),
1606 	    ("[pmc,%d] non-sampling PMC (%p) desires mapping information",
1607 		__LINE__, (void *) pm));
1608 
1609 	po = pm->pm_owner;
1610 
1611 	if (po->po_flags & PMC_PO_INITIAL_MAPPINGS_DONE)
1612 		return;
1613 
1614 	/*
1615 	 * Log the current set of kernel modules.
1616 	 */
1617 	kmbase = linker_hwpmc_list_objects();
1618 	for (km = kmbase; km->pm_file != NULL; km++) {
1619 		PMCDBG(LOG,REG,1,"%s %p", (char *) km->pm_file,
1620 		    (void *) km->pm_address);
1621 		pmclog_process_map_in(po, (pid_t) -1, km->pm_address,
1622 		    km->pm_file);
1623 	}
1624 	free(kmbase, M_LINKER);
1625 
1626 	po->po_flags |= PMC_PO_INITIAL_MAPPINGS_DONE;
1627 }
1628 
1629 /*
1630  * Log the mappings for a single process.
1631  */
1632 
1633 static void
1634 pmc_log_process_mappings(struct pmc_owner *po, struct proc *p)
1635 {
1636 	vm_map_t map;
1637 	struct vnode *vp;
1638 	struct vmspace *vm;
1639 	vm_map_entry_t entry;
1640 	vm_offset_t last_end;
1641 	u_int last_timestamp;
1642 	struct vnode *last_vp;
1643 	vm_offset_t start_addr;
1644 	vm_object_t obj, lobj, tobj;
1645 	char *fullpath, *freepath;
1646 
1647 	last_vp = NULL;
1648 	last_end = (vm_offset_t) 0;
1649 	fullpath = freepath = NULL;
1650 
1651 	if ((vm = vmspace_acquire_ref(p)) == NULL)
1652 		return;
1653 
1654 	map = &vm->vm_map;
1655 	vm_map_lock_read(map);
1656 
1657 	for (entry = map->header.next; entry != &map->header; entry = entry->next) {
1658 
1659 		if (entry == NULL) {
1660 			PMCDBG(LOG,OPS,2, "hwpmc: vm_map entry unexpectedly "
1661 			    "NULL! pid=%d vm_map=%p\n", p->p_pid, map);
1662 			break;
1663 		}
1664 
1665 		/*
1666 		 * We only care about executable map entries.
1667 		 */
1668 		if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) ||
1669 		    !(entry->protection & VM_PROT_EXECUTE) ||
1670 		    (entry->object.vm_object == NULL)) {
1671 			continue;
1672 		}
1673 
1674 		obj = entry->object.vm_object;
1675 		VM_OBJECT_RLOCK(obj);
1676 
1677 		/*
1678 		 * Walk the backing_object list to find the base
1679 		 * (non-shadowed) vm_object.
1680 		 */
1681 		for (lobj = tobj = obj; tobj != NULL; tobj = tobj->backing_object) {
1682 			if (tobj != obj)
1683 				VM_OBJECT_RLOCK(tobj);
1684 			if (lobj != obj)
1685 				VM_OBJECT_RUNLOCK(lobj);
1686 			lobj = tobj;
1687 		}
1688 
1689 		/*
1690 		 * At this point lobj is the base vm_object and it is locked.
1691 		 */
1692 		if (lobj == NULL) {
1693 			PMCDBG(LOG,OPS,2, "hwpmc: lobj unexpectedly NULL! pid=%d "
1694 			    "vm_map=%p vm_obj=%p\n", p->p_pid, map, obj);
1695 			VM_OBJECT_RUNLOCK(obj);
1696 			continue;
1697 		}
1698 
1699 		if (lobj->type != OBJT_VNODE || lobj->handle == NULL) {
1700 			if (lobj != obj)
1701 				VM_OBJECT_RUNLOCK(lobj);
1702 			VM_OBJECT_RUNLOCK(obj);
1703 			continue;
1704 		}
1705 
1706 		/*
1707 		 * Skip contiguous regions that point to the same
1708 		 * vnode, so we don't emit redundant MAP-IN
1709 		 * directives.
1710 		 */
1711 		if (entry->start == last_end && lobj->handle == last_vp) {
1712 			last_end = entry->end;
1713 			if (lobj != obj)
1714 				VM_OBJECT_RUNLOCK(lobj);
1715 			VM_OBJECT_RUNLOCK(obj);
1716 			continue;
1717 		}
1718 
1719 		/*
1720 		 * We don't want to keep the proc's vm_map or this
1721 		 * vm_object locked while we walk the pathname, since
1722 		 * vn_fullpath() can sleep.  However, if we drop the
1723 		 * lock, it's possible for concurrent activity to
1724 		 * modify the vm_map list.  To protect against this,
1725 		 * we save the vm_map timestamp before we release the
1726 		 * lock, and check it after we reacquire the lock
1727 		 * below.
1728 		 */
1729 		start_addr = entry->start;
1730 		last_end = entry->end;
1731 		last_timestamp = map->timestamp;
1732 		vm_map_unlock_read(map);
1733 
1734 		vp = lobj->handle;
1735 		vref(vp);
1736 		if (lobj != obj)
1737 			VM_OBJECT_RUNLOCK(lobj);
1738 
1739 		VM_OBJECT_RUNLOCK(obj);
1740 
1741 		freepath = NULL;
1742 		pmc_getfilename(vp, &fullpath, &freepath);
1743 		last_vp = vp;
1744 
1745 		vrele(vp);
1746 
1747 		vp = NULL;
1748 		pmclog_process_map_in(po, p->p_pid, start_addr, fullpath);
1749 		if (freepath)
1750 			free(freepath, M_TEMP);
1751 
1752 		vm_map_lock_read(map);
1753 
1754 		/*
1755 		 * If our saved timestamp doesn't match, this means
1756 		 * that the vm_map was modified out from under us and
1757 		 * we can't trust our current "entry" pointer.  Do a
1758 		 * new lookup for this entry.  If there is no entry
1759 		 * for this address range, vm_map_lookup_entry() will
1760 		 * return the previous one, so we always want to go to
1761 		 * entry->next on the next loop iteration.
1762 		 *
1763 		 * There is an edge condition here that can occur if
1764 		 * there is no entry at or before this address.  In
1765 		 * this situation, vm_map_lookup_entry returns
1766 		 * &map->header, which would cause our loop to abort
1767 		 * without processing the rest of the map.  However,
1768 		 * in practice this will never happen for process
1769 		 * vm_map.  This is because the executable's text
1770 		 * segment is the first mapping in the proc's address
1771 		 * space, and this mapping is never removed until the
1772 		 * process exits, so there will always be a non-header
1773 		 * entry at or before the requested address for
1774 		 * vm_map_lookup_entry to return.
1775 		 */
1776 		if (map->timestamp != last_timestamp)
1777 			vm_map_lookup_entry(map, last_end - 1, &entry);
1778 	}
1779 
1780 	vm_map_unlock_read(map);
1781 	vmspace_free(vm);
1782 	return;
1783 }
1784 
1785 /*
1786  * Log mappings for all processes in the system.
1787  */
1788 
1789 static void
1790 pmc_log_all_process_mappings(struct pmc_owner *po)
1791 {
1792 	struct proc *p, *top;
1793 
1794 	sx_assert(&pmc_sx, SX_XLOCKED);
1795 
1796 	if ((p = pfind(1)) == NULL)
1797 		panic("[pmc,%d] Cannot find init", __LINE__);
1798 
1799 	PROC_UNLOCK(p);
1800 
1801 	sx_slock(&proctree_lock);
1802 
1803 	top = p;
1804 
1805 	for (;;) {
1806 		pmc_log_process_mappings(po, p);
1807 		if (!LIST_EMPTY(&p->p_children))
1808 			p = LIST_FIRST(&p->p_children);
1809 		else for (;;) {
1810 			if (p == top)
1811 				goto done;
1812 			if (LIST_NEXT(p, p_sibling)) {
1813 				p = LIST_NEXT(p, p_sibling);
1814 				break;
1815 			}
1816 			p = p->p_pptr;
1817 		}
1818 	}
1819  done:
1820 	sx_sunlock(&proctree_lock);
1821 }
1822 
1823 /*
1824  * The 'hook' invoked from the kernel proper
1825  */
1826 
1827 
1828 #ifdef	DEBUG
1829 const char *pmc_hooknames[] = {
1830 	/* these strings correspond to PMC_FN_* in <sys/pmckern.h> */
1831 	"",
1832 	"EXEC",
1833 	"CSW-IN",
1834 	"CSW-OUT",
1835 	"SAMPLE",
1836 	"KLDLOAD",
1837 	"KLDUNLOAD",
1838 	"MMAP",
1839 	"MUNMAP",
1840 	"CALLCHAIN-NMI",
1841 	"CALLCHAIN-SOFT",
1842 	"SOFTSAMPLING"
1843 };
1844 #endif
1845 
1846 static int
1847 pmc_hook_handler(struct thread *td, int function, void *arg)
1848 {
1849 
1850 	PMCDBG(MOD,PMH,1, "hook td=%p func=%d \"%s\" arg=%p", td, function,
1851 	    pmc_hooknames[function], arg);
1852 
1853 	switch (function)
1854 	{
1855 
1856 	/*
1857 	 * Process exec()
1858 	 */
1859 
1860 	case PMC_FN_PROCESS_EXEC:
1861 	{
1862 		char *fullpath, *freepath;
1863 		unsigned int ri;
1864 		int is_using_hwpmcs;
1865 		struct pmc *pm;
1866 		struct proc *p;
1867 		struct pmc_owner *po;
1868 		struct pmc_process *pp;
1869 		struct pmckern_procexec *pk;
1870 
1871 		sx_assert(&pmc_sx, SX_XLOCKED);
1872 
1873 		p = td->td_proc;
1874 		pmc_getfilename(p->p_textvp, &fullpath, &freepath);
1875 
1876 		pk = (struct pmckern_procexec *) arg;
1877 
1878 		/* Inform owners of SS mode PMCs of the exec event. */
1879 		LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1880 		    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1881 			    pmclog_process_procexec(po, PMC_ID_INVALID,
1882 				p->p_pid, pk->pm_entryaddr, fullpath);
1883 
1884 		PROC_LOCK(p);
1885 		is_using_hwpmcs = p->p_flag & P_HWPMC;
1886 		PROC_UNLOCK(p);
1887 
1888 		if (!is_using_hwpmcs) {
1889 			if (freepath)
1890 				free(freepath, M_TEMP);
1891 			break;
1892 		}
1893 
1894 		/*
1895 		 * PMCs are not inherited across an exec():  remove any
1896 		 * PMCs that this process is the owner of.
1897 		 */
1898 
1899 		if ((po = pmc_find_owner_descriptor(p)) != NULL) {
1900 			pmc_remove_owner(po);
1901 			pmc_destroy_owner_descriptor(po);
1902 		}
1903 
1904 		/*
1905 		 * If the process being exec'ed is not the target of any
1906 		 * PMC, we are done.
1907 		 */
1908 		if ((pp = pmc_find_process_descriptor(p, 0)) == NULL) {
1909 			if (freepath)
1910 				free(freepath, M_TEMP);
1911 			break;
1912 		}
1913 
1914 		/*
1915 		 * Log the exec event to all monitoring owners.  Skip
1916 		 * owners who have already recieved the event because
1917 		 * they had system sampling PMCs active.
1918 		 */
1919 		for (ri = 0; ri < md->pmd_npmc; ri++)
1920 			if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) {
1921 				po = pm->pm_owner;
1922 				if (po->po_sscount == 0 &&
1923 				    po->po_flags & PMC_PO_OWNS_LOGFILE)
1924 					pmclog_process_procexec(po, pm->pm_id,
1925 					    p->p_pid, pk->pm_entryaddr,
1926 					    fullpath);
1927 			}
1928 
1929 		if (freepath)
1930 			free(freepath, M_TEMP);
1931 
1932 
1933 		PMCDBG(PRC,EXC,1, "exec proc=%p (%d, %s) cred-changed=%d",
1934 		    p, p->p_pid, p->p_comm, pk->pm_credentialschanged);
1935 
1936 		if (pk->pm_credentialschanged == 0) /* no change */
1937 			break;
1938 
1939 		/*
1940 		 * If the newly exec()'ed process has a different credential
1941 		 * than before, allow it to be the target of a PMC only if
1942 		 * the PMC's owner has sufficient priviledge.
1943 		 */
1944 
1945 		for (ri = 0; ri < md->pmd_npmc; ri++)
1946 			if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL)
1947 				if (pmc_can_attach(pm, td->td_proc) != 0)
1948 					pmc_detach_one_process(td->td_proc,
1949 					    pm, PMC_FLAG_NONE);
1950 
1951 		KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= (int) md->pmd_npmc,
1952 		    ("[pmc,%d] Illegal ref count %d on pp %p", __LINE__,
1953 			pp->pp_refcnt, pp));
1954 
1955 		/*
1956 		 * If this process is no longer the target of any
1957 		 * PMCs, we can remove the process entry and free
1958 		 * up space.
1959 		 */
1960 
1961 		if (pp->pp_refcnt == 0) {
1962 			pmc_remove_process_descriptor(pp);
1963 			free(pp, M_PMC);
1964 			break;
1965 		}
1966 
1967 	}
1968 	break;
1969 
1970 	case PMC_FN_CSW_IN:
1971 		pmc_process_csw_in(td);
1972 		break;
1973 
1974 	case PMC_FN_CSW_OUT:
1975 		pmc_process_csw_out(td);
1976 		break;
1977 
1978 	/*
1979 	 * Process accumulated PC samples.
1980 	 *
1981 	 * This function is expected to be called by hardclock() for
1982 	 * each CPU that has accumulated PC samples.
1983 	 *
1984 	 * This function is to be executed on the CPU whose samples
1985 	 * are being processed.
1986 	 */
1987 	case PMC_FN_DO_SAMPLES:
1988 
1989 		/*
1990 		 * Clear the cpu specific bit in the CPU mask before
1991 		 * do the rest of the processing.  If the NMI handler
1992 		 * gets invoked after the "atomic_clear_int()" call
1993 		 * below but before "pmc_process_samples()" gets
1994 		 * around to processing the interrupt, then we will
1995 		 * come back here at the next hardclock() tick (and
1996 		 * may find nothing to do if "pmc_process_samples()"
1997 		 * had already processed the interrupt).  We don't
1998 		 * lose the interrupt sample.
1999 		 */
2000 		CPU_CLR_ATOMIC(PCPU_GET(cpuid), &pmc_cpumask);
2001 		pmc_process_samples(PCPU_GET(cpuid), PMC_HR);
2002 		pmc_process_samples(PCPU_GET(cpuid), PMC_SR);
2003 		break;
2004 
2005 
2006 	case PMC_FN_KLD_LOAD:
2007 		sx_assert(&pmc_sx, SX_LOCKED);
2008 		pmc_process_kld_load((struct pmckern_map_in *) arg);
2009 		break;
2010 
2011 	case PMC_FN_KLD_UNLOAD:
2012 		sx_assert(&pmc_sx, SX_LOCKED);
2013 		pmc_process_kld_unload((struct pmckern_map_out *) arg);
2014 		break;
2015 
2016 	case PMC_FN_MMAP:
2017 		sx_assert(&pmc_sx, SX_LOCKED);
2018 		pmc_process_mmap(td, (struct pmckern_map_in *) arg);
2019 		break;
2020 
2021 	case PMC_FN_MUNMAP:
2022 		sx_assert(&pmc_sx, SX_LOCKED);
2023 		pmc_process_munmap(td, (struct pmckern_map_out *) arg);
2024 		break;
2025 
2026 	case PMC_FN_USER_CALLCHAIN:
2027 		/*
2028 		 * Record a call chain.
2029 		 */
2030 		KASSERT(td == curthread, ("[pmc,%d] td != curthread",
2031 		    __LINE__));
2032 
2033 		pmc_capture_user_callchain(PCPU_GET(cpuid), PMC_HR,
2034 		    (struct trapframe *) arg);
2035 		td->td_pflags &= ~TDP_CALLCHAIN;
2036 		break;
2037 
2038 	case PMC_FN_USER_CALLCHAIN_SOFT:
2039 		/*
2040 		 * Record a call chain.
2041 		 */
2042 		KASSERT(td == curthread, ("[pmc,%d] td != curthread",
2043 		    __LINE__));
2044 		pmc_capture_user_callchain(PCPU_GET(cpuid), PMC_SR,
2045 		    (struct trapframe *) arg);
2046 		td->td_pflags &= ~TDP_CALLCHAIN;
2047 		break;
2048 
2049 	case PMC_FN_SOFT_SAMPLING:
2050 		/*
2051 		 * Call soft PMC sampling intr.
2052 		 */
2053 		pmc_soft_intr((struct pmckern_soft *) arg);
2054 		break;
2055 
2056 	default:
2057 #ifdef	DEBUG
2058 		KASSERT(0, ("[pmc,%d] unknown hook %d\n", __LINE__, function));
2059 #endif
2060 		break;
2061 
2062 	}
2063 
2064 	return 0;
2065 }
2066 
2067 /*
2068  * allocate a 'struct pmc_owner' descriptor in the owner hash table.
2069  */
2070 
2071 static struct pmc_owner *
2072 pmc_allocate_owner_descriptor(struct proc *p)
2073 {
2074 	uint32_t hindex;
2075 	struct pmc_owner *po;
2076 	struct pmc_ownerhash *poh;
2077 
2078 	hindex = PMC_HASH_PTR(p, pmc_ownerhashmask);
2079 	poh = &pmc_ownerhash[hindex];
2080 
2081 	/* allocate space for N pointers and one descriptor struct */
2082 	po = malloc(sizeof(struct pmc_owner), M_PMC, M_WAITOK|M_ZERO);
2083 	po->po_sscount = po->po_error = po->po_flags = po->po_logprocmaps = 0;
2084 	po->po_file  = NULL;
2085 	po->po_owner = p;
2086 	po->po_kthread = NULL;
2087 	LIST_INIT(&po->po_pmcs);
2088 	LIST_INSERT_HEAD(poh, po, po_next); /* insert into hash table */
2089 
2090 	TAILQ_INIT(&po->po_logbuffers);
2091 	mtx_init(&po->po_mtx, "pmc-owner-mtx", "pmc-per-proc", MTX_SPIN);
2092 
2093 	PMCDBG(OWN,ALL,1, "allocate-owner proc=%p (%d, %s) pmc-owner=%p",
2094 	    p, p->p_pid, p->p_comm, po);
2095 
2096 	return po;
2097 }
2098 
2099 static void
2100 pmc_destroy_owner_descriptor(struct pmc_owner *po)
2101 {
2102 
2103 	PMCDBG(OWN,REL,1, "destroy-owner po=%p proc=%p (%d, %s)",
2104 	    po, po->po_owner, po->po_owner->p_pid, po->po_owner->p_comm);
2105 
2106 	mtx_destroy(&po->po_mtx);
2107 	free(po, M_PMC);
2108 }
2109 
2110 /*
2111  * find the descriptor corresponding to process 'p', adding or removing it
2112  * as specified by 'mode'.
2113  */
2114 
2115 static struct pmc_process *
2116 pmc_find_process_descriptor(struct proc *p, uint32_t mode)
2117 {
2118 	uint32_t hindex;
2119 	struct pmc_process *pp, *ppnew;
2120 	struct pmc_processhash *pph;
2121 
2122 	hindex = PMC_HASH_PTR(p, pmc_processhashmask);
2123 	pph = &pmc_processhash[hindex];
2124 
2125 	ppnew = NULL;
2126 
2127 	/*
2128 	 * Pre-allocate memory in the FIND_ALLOCATE case since we
2129 	 * cannot call malloc(9) once we hold a spin lock.
2130 	 */
2131 	if (mode & PMC_FLAG_ALLOCATE)
2132 		ppnew = malloc(sizeof(struct pmc_process) + md->pmd_npmc *
2133 		    sizeof(struct pmc_targetstate), M_PMC, M_WAITOK|M_ZERO);
2134 
2135 	mtx_lock_spin(&pmc_processhash_mtx);
2136 	LIST_FOREACH(pp, pph, pp_next)
2137 	    if (pp->pp_proc == p)
2138 		    break;
2139 
2140 	if ((mode & PMC_FLAG_REMOVE) && pp != NULL)
2141 		LIST_REMOVE(pp, pp_next);
2142 
2143 	if ((mode & PMC_FLAG_ALLOCATE) && pp == NULL &&
2144 	    ppnew != NULL) {
2145 		ppnew->pp_proc = p;
2146 		LIST_INSERT_HEAD(pph, ppnew, pp_next);
2147 		pp = ppnew;
2148 		ppnew = NULL;
2149 	}
2150 	mtx_unlock_spin(&pmc_processhash_mtx);
2151 
2152 	if (pp != NULL && ppnew != NULL)
2153 		free(ppnew, M_PMC);
2154 
2155 	return pp;
2156 }
2157 
2158 /*
2159  * remove a process descriptor from the process hash table.
2160  */
2161 
2162 static void
2163 pmc_remove_process_descriptor(struct pmc_process *pp)
2164 {
2165 	KASSERT(pp->pp_refcnt == 0,
2166 	    ("[pmc,%d] Removing process descriptor %p with count %d",
2167 		__LINE__, pp, pp->pp_refcnt));
2168 
2169 	mtx_lock_spin(&pmc_processhash_mtx);
2170 	LIST_REMOVE(pp, pp_next);
2171 	mtx_unlock_spin(&pmc_processhash_mtx);
2172 }
2173 
2174 
2175 /*
2176  * find an owner descriptor corresponding to proc 'p'
2177  */
2178 
2179 static struct pmc_owner *
2180 pmc_find_owner_descriptor(struct proc *p)
2181 {
2182 	uint32_t hindex;
2183 	struct pmc_owner *po;
2184 	struct pmc_ownerhash *poh;
2185 
2186 	hindex = PMC_HASH_PTR(p, pmc_ownerhashmask);
2187 	poh = &pmc_ownerhash[hindex];
2188 
2189 	po = NULL;
2190 	LIST_FOREACH(po, poh, po_next)
2191 	    if (po->po_owner == p)
2192 		    break;
2193 
2194 	PMCDBG(OWN,FND,1, "find-owner proc=%p (%d, %s) hindex=0x%x -> "
2195 	    "pmc-owner=%p", p, p->p_pid, p->p_comm, hindex, po);
2196 
2197 	return po;
2198 }
2199 
2200 /*
2201  * pmc_allocate_pmc_descriptor
2202  *
2203  * Allocate a pmc descriptor and initialize its
2204  * fields.
2205  */
2206 
2207 static struct pmc *
2208 pmc_allocate_pmc_descriptor(void)
2209 {
2210 	struct pmc *pmc;
2211 
2212 	pmc = malloc(sizeof(struct pmc), M_PMC, M_WAITOK|M_ZERO);
2213 	pmc->pm_owner = NULL;
2214 	LIST_INIT(&pmc->pm_targets);
2215 
2216 	PMCDBG(PMC,ALL,1, "allocate-pmc -> pmc=%p", pmc);
2217 
2218 	return pmc;
2219 }
2220 
2221 /*
2222  * Destroy a pmc descriptor.
2223  */
2224 
2225 static void
2226 pmc_destroy_pmc_descriptor(struct pmc *pm)
2227 {
2228 	(void) pm;
2229 
2230 #ifdef	DEBUG
2231 	KASSERT(pm->pm_state == PMC_STATE_DELETED ||
2232 	    pm->pm_state == PMC_STATE_FREE,
2233 	    ("[pmc,%d] destroying non-deleted PMC", __LINE__));
2234 	KASSERT(LIST_EMPTY(&pm->pm_targets),
2235 	    ("[pmc,%d] destroying pmc with targets", __LINE__));
2236 	KASSERT(pm->pm_owner == NULL,
2237 	    ("[pmc,%d] destroying pmc attached to an owner", __LINE__));
2238 	KASSERT(pm->pm_runcount == 0,
2239 	    ("[pmc,%d] pmc has non-zero run count %d", __LINE__,
2240 		pm->pm_runcount));
2241 #endif
2242 }
2243 
2244 static void
2245 pmc_wait_for_pmc_idle(struct pmc *pm)
2246 {
2247 #ifdef DEBUG
2248 	volatile int maxloop;
2249 
2250 	maxloop = 100 * pmc_cpu_max();
2251 #endif
2252 	/*
2253 	 * Loop (with a forced context switch) till the PMC's runcount
2254 	 * comes down to zero.
2255 	 */
2256 	while (atomic_load_acq_32(&pm->pm_runcount) > 0) {
2257 #ifdef DEBUG
2258 		maxloop--;
2259 		KASSERT(maxloop > 0,
2260 		    ("[pmc,%d] (ri%d, rc%d) waiting too long for "
2261 			"pmc to be free", __LINE__,
2262 			PMC_TO_ROWINDEX(pm), pm->pm_runcount));
2263 #endif
2264 		pmc_force_context_switch();
2265 	}
2266 }
2267 
2268 /*
2269  * This function does the following things:
2270  *
2271  *  - detaches the PMC from hardware
2272  *  - unlinks all target threads that were attached to it
2273  *  - removes the PMC from its owner's list
2274  *  - destroy's the PMC private mutex
2275  *
2276  * Once this function completes, the given pmc pointer can be safely
2277  * FREE'd by the caller.
2278  */
2279 
2280 static void
2281 pmc_release_pmc_descriptor(struct pmc *pm)
2282 {
2283 	enum pmc_mode mode;
2284 	struct pmc_hw *phw;
2285 	u_int adjri, ri, cpu;
2286 	struct pmc_owner *po;
2287 	struct pmc_binding pb;
2288 	struct pmc_process *pp;
2289 	struct pmc_classdep *pcd;
2290 	struct pmc_target *ptgt, *tmp;
2291 
2292 	sx_assert(&pmc_sx, SX_XLOCKED);
2293 
2294 	KASSERT(pm, ("[pmc,%d] null pmc", __LINE__));
2295 
2296 	ri   = PMC_TO_ROWINDEX(pm);
2297 	pcd  = pmc_ri_to_classdep(md, ri, &adjri);
2298 	mode = PMC_TO_MODE(pm);
2299 
2300 	PMCDBG(PMC,REL,1, "release-pmc pmc=%p ri=%d mode=%d", pm, ri,
2301 	    mode);
2302 
2303 	/*
2304 	 * First, we take the PMC off hardware.
2305 	 */
2306 	cpu = 0;
2307 	if (PMC_IS_SYSTEM_MODE(mode)) {
2308 
2309 		/*
2310 		 * A system mode PMC runs on a specific CPU.  Switch
2311 		 * to this CPU and turn hardware off.
2312 		 */
2313 		pmc_save_cpu_binding(&pb);
2314 
2315 		cpu = PMC_TO_CPU(pm);
2316 
2317 		pmc_select_cpu(cpu);
2318 
2319 		/* switch off non-stalled CPUs */
2320 		if (pm->pm_state == PMC_STATE_RUNNING &&
2321 		    pm->pm_stalled == 0) {
2322 
2323 			phw = pmc_pcpu[cpu]->pc_hwpmcs[ri];
2324 
2325 			KASSERT(phw->phw_pmc == pm,
2326 			    ("[pmc, %d] pmc ptr ri(%d) hw(%p) pm(%p)",
2327 				__LINE__, ri, phw->phw_pmc, pm));
2328 			PMCDBG(PMC,REL,2, "stopping cpu=%d ri=%d", cpu, ri);
2329 
2330 			critical_enter();
2331 			pcd->pcd_stop_pmc(cpu, adjri);
2332 			critical_exit();
2333 		}
2334 
2335 		PMCDBG(PMC,REL,2, "decfg cpu=%d ri=%d", cpu, ri);
2336 
2337 		critical_enter();
2338 		pcd->pcd_config_pmc(cpu, adjri, NULL);
2339 		critical_exit();
2340 
2341 		/* adjust the global and process count of SS mode PMCs */
2342 		if (mode == PMC_MODE_SS && pm->pm_state == PMC_STATE_RUNNING) {
2343 			po = pm->pm_owner;
2344 			po->po_sscount--;
2345 			if (po->po_sscount == 0) {
2346 				atomic_subtract_rel_int(&pmc_ss_count, 1);
2347 				LIST_REMOVE(po, po_ssnext);
2348 			}
2349 		}
2350 
2351 		pm->pm_state = PMC_STATE_DELETED;
2352 
2353 		pmc_restore_cpu_binding(&pb);
2354 
2355 		/*
2356 		 * We could have references to this PMC structure in
2357 		 * the per-cpu sample queues.  Wait for the queue to
2358 		 * drain.
2359 		 */
2360 		pmc_wait_for_pmc_idle(pm);
2361 
2362 	} else if (PMC_IS_VIRTUAL_MODE(mode)) {
2363 
2364 		/*
2365 		 * A virtual PMC could be running on multiple CPUs at
2366 		 * a given instant.
2367 		 *
2368 		 * By marking its state as DELETED, we ensure that
2369 		 * this PMC is never further scheduled on hardware.
2370 		 *
2371 		 * Then we wait till all CPUs are done with this PMC.
2372 		 */
2373 		pm->pm_state = PMC_STATE_DELETED;
2374 
2375 
2376 		/* Wait for the PMCs runcount to come to zero. */
2377 		pmc_wait_for_pmc_idle(pm);
2378 
2379 		/*
2380 		 * At this point the PMC is off all CPUs and cannot be
2381 		 * freshly scheduled onto a CPU.  It is now safe to
2382 		 * unlink all targets from this PMC.  If a
2383 		 * process-record's refcount falls to zero, we remove
2384 		 * it from the hash table.  The module-wide SX lock
2385 		 * protects us from races.
2386 		 */
2387 		LIST_FOREACH_SAFE(ptgt, &pm->pm_targets, pt_next, tmp) {
2388 			pp = ptgt->pt_process;
2389 			pmc_unlink_target_process(pm, pp); /* frees 'ptgt' */
2390 
2391 			PMCDBG(PMC,REL,3, "pp->refcnt=%d", pp->pp_refcnt);
2392 
2393 			/*
2394 			 * If the target process record shows that no
2395 			 * PMCs are attached to it, reclaim its space.
2396 			 */
2397 
2398 			if (pp->pp_refcnt == 0) {
2399 				pmc_remove_process_descriptor(pp);
2400 				free(pp, M_PMC);
2401 			}
2402 		}
2403 
2404 		cpu = curthread->td_oncpu; /* setup cpu for pmd_release() */
2405 
2406 	}
2407 
2408 	/*
2409 	 * Release any MD resources
2410 	 */
2411 	(void) pcd->pcd_release_pmc(cpu, adjri, pm);
2412 
2413 	/*
2414 	 * Update row disposition
2415 	 */
2416 
2417 	if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm)))
2418 		PMC_UNMARK_ROW_STANDALONE(ri);
2419 	else
2420 		PMC_UNMARK_ROW_THREAD(ri);
2421 
2422 	/* unlink from the owner's list */
2423 	if (pm->pm_owner) {
2424 		LIST_REMOVE(pm, pm_next);
2425 		pm->pm_owner = NULL;
2426 	}
2427 
2428 	pmc_destroy_pmc_descriptor(pm);
2429 }
2430 
2431 /*
2432  * Register an owner and a pmc.
2433  */
2434 
2435 static int
2436 pmc_register_owner(struct proc *p, struct pmc *pmc)
2437 {
2438 	struct pmc_owner *po;
2439 
2440 	sx_assert(&pmc_sx, SX_XLOCKED);
2441 
2442 	if ((po = pmc_find_owner_descriptor(p)) == NULL)
2443 		if ((po = pmc_allocate_owner_descriptor(p)) == NULL)
2444 			return ENOMEM;
2445 
2446 	KASSERT(pmc->pm_owner == NULL,
2447 	    ("[pmc,%d] attempting to own an initialized PMC", __LINE__));
2448 	pmc->pm_owner  = po;
2449 
2450 	LIST_INSERT_HEAD(&po->po_pmcs, pmc, pm_next);
2451 
2452 	PROC_LOCK(p);
2453 	p->p_flag |= P_HWPMC;
2454 	PROC_UNLOCK(p);
2455 
2456 	if (po->po_flags & PMC_PO_OWNS_LOGFILE)
2457 		pmclog_process_pmcallocate(pmc);
2458 
2459 	PMCDBG(PMC,REG,1, "register-owner pmc-owner=%p pmc=%p",
2460 	    po, pmc);
2461 
2462 	return 0;
2463 }
2464 
2465 /*
2466  * Return the current row disposition:
2467  * == 0 => FREE
2468  *  > 0 => PROCESS MODE
2469  *  < 0 => SYSTEM MODE
2470  */
2471 
2472 int
2473 pmc_getrowdisp(int ri)
2474 {
2475 	return pmc_pmcdisp[ri];
2476 }
2477 
2478 /*
2479  * Check if a PMC at row index 'ri' can be allocated to the current
2480  * process.
2481  *
2482  * Allocation can fail if:
2483  *   - the current process is already being profiled by a PMC at index 'ri',
2484  *     attached to it via OP_PMCATTACH.
2485  *   - the current process has already allocated a PMC at index 'ri'
2486  *     via OP_ALLOCATE.
2487  */
2488 
2489 static int
2490 pmc_can_allocate_rowindex(struct proc *p, unsigned int ri, int cpu)
2491 {
2492 	enum pmc_mode mode;
2493 	struct pmc *pm;
2494 	struct pmc_owner *po;
2495 	struct pmc_process *pp;
2496 
2497 	PMCDBG(PMC,ALR,1, "can-allocate-rowindex proc=%p (%d, %s) ri=%d "
2498 	    "cpu=%d", p, p->p_pid, p->p_comm, ri, cpu);
2499 
2500 	/*
2501 	 * We shouldn't have already allocated a process-mode PMC at
2502 	 * row index 'ri'.
2503 	 *
2504 	 * We shouldn't have allocated a system-wide PMC on the same
2505 	 * CPU and same RI.
2506 	 */
2507 	if ((po = pmc_find_owner_descriptor(p)) != NULL)
2508 		LIST_FOREACH(pm, &po->po_pmcs, pm_next) {
2509 		    if (PMC_TO_ROWINDEX(pm) == ri) {
2510 			    mode = PMC_TO_MODE(pm);
2511 			    if (PMC_IS_VIRTUAL_MODE(mode))
2512 				    return EEXIST;
2513 			    if (PMC_IS_SYSTEM_MODE(mode) &&
2514 				(int) PMC_TO_CPU(pm) == cpu)
2515 				    return EEXIST;
2516 		    }
2517 	        }
2518 
2519 	/*
2520 	 * We also shouldn't be the target of any PMC at this index
2521 	 * since otherwise a PMC_ATTACH to ourselves will fail.
2522 	 */
2523 	if ((pp = pmc_find_process_descriptor(p, 0)) != NULL)
2524 		if (pp->pp_pmcs[ri].pp_pmc)
2525 			return EEXIST;
2526 
2527 	PMCDBG(PMC,ALR,2, "can-allocate-rowindex proc=%p (%d, %s) ri=%d ok",
2528 	    p, p->p_pid, p->p_comm, ri);
2529 
2530 	return 0;
2531 }
2532 
2533 /*
2534  * Check if a given PMC at row index 'ri' can be currently used in
2535  * mode 'mode'.
2536  */
2537 
2538 static int
2539 pmc_can_allocate_row(int ri, enum pmc_mode mode)
2540 {
2541 	enum pmc_disp	disp;
2542 
2543 	sx_assert(&pmc_sx, SX_XLOCKED);
2544 
2545 	PMCDBG(PMC,ALR,1, "can-allocate-row ri=%d mode=%d", ri, mode);
2546 
2547 	if (PMC_IS_SYSTEM_MODE(mode))
2548 		disp = PMC_DISP_STANDALONE;
2549 	else
2550 		disp = PMC_DISP_THREAD;
2551 
2552 	/*
2553 	 * check disposition for PMC row 'ri':
2554 	 *
2555 	 * Expected disposition		Row-disposition		Result
2556 	 *
2557 	 * STANDALONE			STANDALONE or FREE	proceed
2558 	 * STANDALONE			THREAD			fail
2559 	 * THREAD			THREAD or FREE		proceed
2560 	 * THREAD			STANDALONE		fail
2561 	 */
2562 
2563 	if (!PMC_ROW_DISP_IS_FREE(ri) &&
2564 	    !(disp == PMC_DISP_THREAD && PMC_ROW_DISP_IS_THREAD(ri)) &&
2565 	    !(disp == PMC_DISP_STANDALONE && PMC_ROW_DISP_IS_STANDALONE(ri)))
2566 		return EBUSY;
2567 
2568 	/*
2569 	 * All OK
2570 	 */
2571 
2572 	PMCDBG(PMC,ALR,2, "can-allocate-row ri=%d mode=%d ok", ri, mode);
2573 
2574 	return 0;
2575 
2576 }
2577 
2578 /*
2579  * Find a PMC descriptor with user handle 'pmcid' for thread 'td'.
2580  */
2581 
2582 static struct pmc *
2583 pmc_find_pmc_descriptor_in_process(struct pmc_owner *po, pmc_id_t pmcid)
2584 {
2585 	struct pmc *pm;
2586 
2587 	KASSERT(PMC_ID_TO_ROWINDEX(pmcid) < md->pmd_npmc,
2588 	    ("[pmc,%d] Illegal pmc index %d (max %d)", __LINE__,
2589 		PMC_ID_TO_ROWINDEX(pmcid), md->pmd_npmc));
2590 
2591 	LIST_FOREACH(pm, &po->po_pmcs, pm_next)
2592 	    if (pm->pm_id == pmcid)
2593 		    return pm;
2594 
2595 	return NULL;
2596 }
2597 
2598 static int
2599 pmc_find_pmc(pmc_id_t pmcid, struct pmc **pmc)
2600 {
2601 
2602 	struct pmc *pm;
2603 	struct pmc_owner *po;
2604 
2605 	PMCDBG(PMC,FND,1, "find-pmc id=%d", pmcid);
2606 
2607 	if ((po = pmc_find_owner_descriptor(curthread->td_proc)) == NULL)
2608 		return ESRCH;
2609 
2610 	if ((pm = pmc_find_pmc_descriptor_in_process(po, pmcid)) == NULL)
2611 		return EINVAL;
2612 
2613 	PMCDBG(PMC,FND,2, "find-pmc id=%d -> pmc=%p", pmcid, pm);
2614 
2615 	*pmc = pm;
2616 	return 0;
2617 }
2618 
2619 /*
2620  * Start a PMC.
2621  */
2622 
2623 static int
2624 pmc_start(struct pmc *pm)
2625 {
2626 	enum pmc_mode mode;
2627 	struct pmc_owner *po;
2628 	struct pmc_binding pb;
2629 	struct pmc_classdep *pcd;
2630 	int adjri, error, cpu, ri;
2631 
2632 	KASSERT(pm != NULL,
2633 	    ("[pmc,%d] null pm", __LINE__));
2634 
2635 	mode = PMC_TO_MODE(pm);
2636 	ri   = PMC_TO_ROWINDEX(pm);
2637 	pcd  = pmc_ri_to_classdep(md, ri, &adjri);
2638 
2639 	error = 0;
2640 
2641 	PMCDBG(PMC,OPS,1, "start pmc=%p mode=%d ri=%d", pm, mode, ri);
2642 
2643 	po = pm->pm_owner;
2644 
2645 	/*
2646 	 * Disallow PMCSTART if a logfile is required but has not been
2647 	 * configured yet.
2648 	 */
2649 	if ((pm->pm_flags & PMC_F_NEEDS_LOGFILE) &&
2650 	    (po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
2651 		return (EDOOFUS);	/* programming error */
2652 
2653 	/*
2654 	 * If this is a sampling mode PMC, log mapping information for
2655 	 * the kernel modules that are currently loaded.
2656 	 */
2657 	if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
2658 	    pmc_log_kernel_mappings(pm);
2659 
2660 	if (PMC_IS_VIRTUAL_MODE(mode)) {
2661 
2662 		/*
2663 		 * If a PMCATTACH has never been done on this PMC,
2664 		 * attach it to its owner process.
2665 		 */
2666 
2667 		if (LIST_EMPTY(&pm->pm_targets))
2668 			error = (pm->pm_flags & PMC_F_ATTACH_DONE) ? ESRCH :
2669 			    pmc_attach_process(po->po_owner, pm);
2670 
2671 		/*
2672 		 * If the PMC is attached to its owner, then force a context
2673 		 * switch to ensure that the MD state gets set correctly.
2674 		 */
2675 
2676 		if (error == 0) {
2677 			pm->pm_state = PMC_STATE_RUNNING;
2678 			if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER)
2679 				pmc_force_context_switch();
2680 		}
2681 
2682 		return (error);
2683 	}
2684 
2685 
2686 	/*
2687 	 * A system-wide PMC.
2688 	 *
2689 	 * Add the owner to the global list if this is a system-wide
2690 	 * sampling PMC.
2691 	 */
2692 
2693 	if (mode == PMC_MODE_SS) {
2694 		if (po->po_sscount == 0) {
2695 			LIST_INSERT_HEAD(&pmc_ss_owners, po, po_ssnext);
2696 			atomic_add_rel_int(&pmc_ss_count, 1);
2697 			PMCDBG(PMC,OPS,1, "po=%p in global list", po);
2698 		}
2699 		po->po_sscount++;
2700 
2701 		/*
2702 		 * Log mapping information for all existing processes in the
2703 		 * system.  Subsequent mappings are logged as they happen;
2704 		 * see pmc_process_mmap().
2705 		 */
2706 		if (po->po_logprocmaps == 0) {
2707 			pmc_log_all_process_mappings(po);
2708 			po->po_logprocmaps = 1;
2709 		}
2710 	}
2711 
2712 	/*
2713 	 * Move to the CPU associated with this
2714 	 * PMC, and start the hardware.
2715 	 */
2716 
2717 	pmc_save_cpu_binding(&pb);
2718 
2719 	cpu = PMC_TO_CPU(pm);
2720 
2721 	if (!pmc_cpu_is_active(cpu))
2722 		return (ENXIO);
2723 
2724 	pmc_select_cpu(cpu);
2725 
2726 	/*
2727 	 * global PMCs are configured at allocation time
2728 	 * so write out the initial value and start the PMC.
2729 	 */
2730 
2731 	pm->pm_state = PMC_STATE_RUNNING;
2732 
2733 	critical_enter();
2734 	if ((error = pcd->pcd_write_pmc(cpu, adjri,
2735 		 PMC_IS_SAMPLING_MODE(mode) ?
2736 		 pm->pm_sc.pm_reloadcount :
2737 		 pm->pm_sc.pm_initial)) == 0)
2738 		error = pcd->pcd_start_pmc(cpu, adjri);
2739 	critical_exit();
2740 
2741 	pmc_restore_cpu_binding(&pb);
2742 
2743 	return (error);
2744 }
2745 
2746 /*
2747  * Stop a PMC.
2748  */
2749 
2750 static int
2751 pmc_stop(struct pmc *pm)
2752 {
2753 	struct pmc_owner *po;
2754 	struct pmc_binding pb;
2755 	struct pmc_classdep *pcd;
2756 	int adjri, cpu, error, ri;
2757 
2758 	KASSERT(pm != NULL, ("[pmc,%d] null pmc", __LINE__));
2759 
2760 	PMCDBG(PMC,OPS,1, "stop pmc=%p mode=%d ri=%d", pm,
2761 	    PMC_TO_MODE(pm), PMC_TO_ROWINDEX(pm));
2762 
2763 	pm->pm_state = PMC_STATE_STOPPED;
2764 
2765 	/*
2766 	 * If the PMC is a virtual mode one, changing the state to
2767 	 * non-RUNNING is enough to ensure that the PMC never gets
2768 	 * scheduled.
2769 	 *
2770 	 * If this PMC is current running on a CPU, then it will
2771 	 * handled correctly at the time its target process is context
2772 	 * switched out.
2773 	 */
2774 
2775 	if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)))
2776 		return 0;
2777 
2778 	/*
2779 	 * A system-mode PMC.  Move to the CPU associated with
2780 	 * this PMC, and stop the hardware.  We update the
2781 	 * 'initial count' so that a subsequent PMCSTART will
2782 	 * resume counting from the current hardware count.
2783 	 */
2784 
2785 	pmc_save_cpu_binding(&pb);
2786 
2787 	cpu = PMC_TO_CPU(pm);
2788 
2789 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
2790 	    ("[pmc,%d] illegal cpu=%d", __LINE__, cpu));
2791 
2792 	if (!pmc_cpu_is_active(cpu))
2793 		return ENXIO;
2794 
2795 	pmc_select_cpu(cpu);
2796 
2797 	ri = PMC_TO_ROWINDEX(pm);
2798 	pcd = pmc_ri_to_classdep(md, ri, &adjri);
2799 
2800 	critical_enter();
2801 	if ((error = pcd->pcd_stop_pmc(cpu, adjri)) == 0)
2802 		error = pcd->pcd_read_pmc(cpu, adjri, &pm->pm_sc.pm_initial);
2803 	critical_exit();
2804 
2805 	pmc_restore_cpu_binding(&pb);
2806 
2807 	po = pm->pm_owner;
2808 
2809 	/* remove this owner from the global list of SS PMC owners */
2810 	if (PMC_TO_MODE(pm) == PMC_MODE_SS) {
2811 		po->po_sscount--;
2812 		if (po->po_sscount == 0) {
2813 			atomic_subtract_rel_int(&pmc_ss_count, 1);
2814 			LIST_REMOVE(po, po_ssnext);
2815 			PMCDBG(PMC,OPS,2,"po=%p removed from global list", po);
2816 		}
2817 	}
2818 
2819 	return (error);
2820 }
2821 
2822 
2823 #ifdef	DEBUG
2824 static const char *pmc_op_to_name[] = {
2825 #undef	__PMC_OP
2826 #define	__PMC_OP(N, D)	#N ,
2827 	__PMC_OPS()
2828 	NULL
2829 };
2830 #endif
2831 
2832 /*
2833  * The syscall interface
2834  */
2835 
2836 #define	PMC_GET_SX_XLOCK(...) do {		\
2837 	sx_xlock(&pmc_sx);			\
2838 	if (pmc_hook == NULL) {			\
2839 		sx_xunlock(&pmc_sx);		\
2840 		return __VA_ARGS__;		\
2841 	}					\
2842 } while (0)
2843 
2844 #define	PMC_DOWNGRADE_SX() do {			\
2845 	sx_downgrade(&pmc_sx);			\
2846 	is_sx_downgraded = 1;			\
2847 } while (0)
2848 
2849 static int
2850 pmc_syscall_handler(struct thread *td, void *syscall_args)
2851 {
2852 	int error, is_sx_downgraded, is_sx_locked, op;
2853 	struct pmc_syscall_args *c;
2854 	void *arg;
2855 
2856 	PMC_GET_SX_XLOCK(ENOSYS);
2857 
2858 	DROP_GIANT();
2859 
2860 	is_sx_downgraded = 0;
2861 	is_sx_locked = 1;
2862 
2863 	c = (struct pmc_syscall_args *) syscall_args;
2864 
2865 	op = c->pmop_code;
2866 	arg = c->pmop_data;
2867 
2868 	PMCDBG(MOD,PMS,1, "syscall op=%d \"%s\" arg=%p", op,
2869 	    pmc_op_to_name[op], arg);
2870 
2871 	error = 0;
2872 	atomic_add_int(&pmc_stats.pm_syscalls, 1);
2873 
2874 	switch(op)
2875 	{
2876 
2877 
2878 	/*
2879 	 * Configure a log file.
2880 	 *
2881 	 * XXX This OP will be reworked.
2882 	 */
2883 
2884 	case PMC_OP_CONFIGURELOG:
2885 	{
2886 		struct proc *p;
2887 		struct pmc *pm;
2888 		struct pmc_owner *po;
2889 		struct pmc_op_configurelog cl;
2890 
2891 		sx_assert(&pmc_sx, SX_XLOCKED);
2892 
2893 		if ((error = copyin(arg, &cl, sizeof(cl))) != 0)
2894 			break;
2895 
2896 		/* mark this process as owning a log file */
2897 		p = td->td_proc;
2898 		if ((po = pmc_find_owner_descriptor(p)) == NULL)
2899 			if ((po = pmc_allocate_owner_descriptor(p)) == NULL) {
2900 				error = ENOMEM;
2901 				break;
2902 			}
2903 
2904 		/*
2905 		 * If a valid fd was passed in, try to configure that,
2906 		 * otherwise if 'fd' was less than zero and there was
2907 		 * a log file configured, flush its buffers and
2908 		 * de-configure it.
2909 		 */
2910 		if (cl.pm_logfd >= 0) {
2911 			sx_xunlock(&pmc_sx);
2912 			is_sx_locked = 0;
2913 			error = pmclog_configure_log(md, po, cl.pm_logfd);
2914 		} else if (po->po_flags & PMC_PO_OWNS_LOGFILE) {
2915 			pmclog_process_closelog(po);
2916 			error = pmclog_close(po);
2917 			if (error == 0) {
2918 				LIST_FOREACH(pm, &po->po_pmcs, pm_next)
2919 				    if (pm->pm_flags & PMC_F_NEEDS_LOGFILE &&
2920 					pm->pm_state == PMC_STATE_RUNNING)
2921 					    pmc_stop(pm);
2922 				error = pmclog_deconfigure_log(po);
2923 			}
2924 		} else
2925 			error = EINVAL;
2926 
2927 		if (error)
2928 			break;
2929 	}
2930 	break;
2931 
2932 	/*
2933 	 * Flush a log file.
2934 	 */
2935 
2936 	case PMC_OP_FLUSHLOG:
2937 	{
2938 		struct pmc_owner *po;
2939 
2940 		sx_assert(&pmc_sx, SX_XLOCKED);
2941 
2942 		if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) {
2943 			error = EINVAL;
2944 			break;
2945 		}
2946 
2947 		error = pmclog_flush(po);
2948 	}
2949 	break;
2950 
2951 	/*
2952 	 * Close a log file.
2953 	 */
2954 
2955 	case PMC_OP_CLOSELOG:
2956 	{
2957 		struct pmc_owner *po;
2958 
2959 		sx_assert(&pmc_sx, SX_XLOCKED);
2960 
2961 		if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) {
2962 			error = EINVAL;
2963 			break;
2964 		}
2965 
2966 		error = pmclog_close(po);
2967 	}
2968 	break;
2969 
2970 	/*
2971 	 * Retrieve hardware configuration.
2972 	 */
2973 
2974 	case PMC_OP_GETCPUINFO:	/* CPU information */
2975 	{
2976 		struct pmc_op_getcpuinfo gci;
2977 		struct pmc_classinfo *pci;
2978 		struct pmc_classdep *pcd;
2979 		int cl;
2980 
2981 		gci.pm_cputype = md->pmd_cputype;
2982 		gci.pm_ncpu    = pmc_cpu_max();
2983 		gci.pm_npmc    = md->pmd_npmc;
2984 		gci.pm_nclass  = md->pmd_nclass;
2985 		pci = gci.pm_classes;
2986 		pcd = md->pmd_classdep;
2987 		for (cl = 0; cl < md->pmd_nclass; cl++, pci++, pcd++) {
2988 			pci->pm_caps  = pcd->pcd_caps;
2989 			pci->pm_class = pcd->pcd_class;
2990 			pci->pm_width = pcd->pcd_width;
2991 			pci->pm_num   = pcd->pcd_num;
2992 		}
2993 		error = copyout(&gci, arg, sizeof(gci));
2994 	}
2995 	break;
2996 
2997 	/*
2998 	 * Retrieve soft events list.
2999 	 */
3000 	case PMC_OP_GETDYNEVENTINFO:
3001 	{
3002 		enum pmc_class			cl;
3003 		enum pmc_event			ev;
3004 		struct pmc_op_getdyneventinfo	*gei;
3005 		struct pmc_dyn_event_descr	dev;
3006 		struct pmc_soft			*ps;
3007 		uint32_t			nevent;
3008 
3009 		sx_assert(&pmc_sx, SX_LOCKED);
3010 
3011 		gei = (struct pmc_op_getdyneventinfo *) arg;
3012 
3013 		if ((error = copyin(&gei->pm_class, &cl, sizeof(cl))) != 0)
3014 			break;
3015 
3016 		/* Only SOFT class is dynamic. */
3017 		if (cl != PMC_CLASS_SOFT) {
3018 			error = EINVAL;
3019 			break;
3020 		}
3021 
3022 		nevent = 0;
3023 		for (ev = PMC_EV_SOFT_FIRST; (int)ev <= PMC_EV_SOFT_LAST; ev++) {
3024 			ps = pmc_soft_ev_acquire(ev);
3025 			if (ps == NULL)
3026 				continue;
3027 			bcopy(&ps->ps_ev, &dev, sizeof(dev));
3028 			pmc_soft_ev_release(ps);
3029 
3030 			error = copyout(&dev,
3031 			    &gei->pm_events[nevent],
3032 			    sizeof(struct pmc_dyn_event_descr));
3033 			if (error != 0)
3034 				break;
3035 			nevent++;
3036 		}
3037 		if (error != 0)
3038 			break;
3039 
3040 		error = copyout(&nevent, &gei->pm_nevent,
3041 		    sizeof(nevent));
3042 	}
3043 	break;
3044 
3045 	/*
3046 	 * Get module statistics
3047 	 */
3048 
3049 	case PMC_OP_GETDRIVERSTATS:
3050 	{
3051 		struct pmc_op_getdriverstats gms;
3052 
3053 		bcopy(&pmc_stats, &gms, sizeof(gms));
3054 		error = copyout(&gms, arg, sizeof(gms));
3055 	}
3056 	break;
3057 
3058 
3059 	/*
3060 	 * Retrieve module version number
3061 	 */
3062 
3063 	case PMC_OP_GETMODULEVERSION:
3064 	{
3065 		uint32_t cv, modv;
3066 
3067 		/* retrieve the client's idea of the ABI version */
3068 		if ((error = copyin(arg, &cv, sizeof(uint32_t))) != 0)
3069 			break;
3070 		/* don't service clients newer than our driver */
3071 		modv = PMC_VERSION;
3072 		if ((cv & 0xFFFF0000) > (modv & 0xFFFF0000)) {
3073 			error = EPROGMISMATCH;
3074 			break;
3075 		}
3076 		error = copyout(&modv, arg, sizeof(int));
3077 	}
3078 	break;
3079 
3080 
3081 	/*
3082 	 * Retrieve the state of all the PMCs on a given
3083 	 * CPU.
3084 	 */
3085 
3086 	case PMC_OP_GETPMCINFO:
3087 	{
3088 		int ari;
3089 		struct pmc *pm;
3090 		size_t pmcinfo_size;
3091 		uint32_t cpu, n, npmc;
3092 		struct pmc_owner *po;
3093 		struct pmc_binding pb;
3094 		struct pmc_classdep *pcd;
3095 		struct pmc_info *p, *pmcinfo;
3096 		struct pmc_op_getpmcinfo *gpi;
3097 
3098 		PMC_DOWNGRADE_SX();
3099 
3100 		gpi = (struct pmc_op_getpmcinfo *) arg;
3101 
3102 		if ((error = copyin(&gpi->pm_cpu, &cpu, sizeof(cpu))) != 0)
3103 			break;
3104 
3105 		if (cpu >= pmc_cpu_max()) {
3106 			error = EINVAL;
3107 			break;
3108 		}
3109 
3110 		if (!pmc_cpu_is_active(cpu)) {
3111 			error = ENXIO;
3112 			break;
3113 		}
3114 
3115 		/* switch to CPU 'cpu' */
3116 		pmc_save_cpu_binding(&pb);
3117 		pmc_select_cpu(cpu);
3118 
3119 		npmc = md->pmd_npmc;
3120 
3121 		pmcinfo_size = npmc * sizeof(struct pmc_info);
3122 		pmcinfo = malloc(pmcinfo_size, M_PMC, M_WAITOK);
3123 
3124 		p = pmcinfo;
3125 
3126 		for (n = 0; n < md->pmd_npmc; n++, p++) {
3127 
3128 			pcd = pmc_ri_to_classdep(md, n, &ari);
3129 
3130 			KASSERT(pcd != NULL,
3131 			    ("[pmc,%d] null pcd ri=%d", __LINE__, n));
3132 
3133 			if ((error = pcd->pcd_describe(cpu, ari, p, &pm)) != 0)
3134 				break;
3135 
3136 			if (PMC_ROW_DISP_IS_STANDALONE(n))
3137 				p->pm_rowdisp = PMC_DISP_STANDALONE;
3138 			else if (PMC_ROW_DISP_IS_THREAD(n))
3139 				p->pm_rowdisp = PMC_DISP_THREAD;
3140 			else
3141 				p->pm_rowdisp = PMC_DISP_FREE;
3142 
3143 			p->pm_ownerpid = -1;
3144 
3145 			if (pm == NULL)	/* no PMC associated */
3146 				continue;
3147 
3148 			po = pm->pm_owner;
3149 
3150 			KASSERT(po->po_owner != NULL,
3151 			    ("[pmc,%d] pmc_owner had a null proc pointer",
3152 				__LINE__));
3153 
3154 			p->pm_ownerpid = po->po_owner->p_pid;
3155 			p->pm_mode     = PMC_TO_MODE(pm);
3156 			p->pm_event    = pm->pm_event;
3157 			p->pm_flags    = pm->pm_flags;
3158 
3159 			if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
3160 				p->pm_reloadcount =
3161 				    pm->pm_sc.pm_reloadcount;
3162 		}
3163 
3164 		pmc_restore_cpu_binding(&pb);
3165 
3166 		/* now copy out the PMC info collected */
3167 		if (error == 0)
3168 			error = copyout(pmcinfo, &gpi->pm_pmcs, pmcinfo_size);
3169 
3170 		free(pmcinfo, M_PMC);
3171 	}
3172 	break;
3173 
3174 
3175 	/*
3176 	 * Set the administrative state of a PMC.  I.e. whether
3177 	 * the PMC is to be used or not.
3178 	 */
3179 
3180 	case PMC_OP_PMCADMIN:
3181 	{
3182 		int cpu, ri;
3183 		enum pmc_state request;
3184 		struct pmc_cpu *pc;
3185 		struct pmc_hw *phw;
3186 		struct pmc_op_pmcadmin pma;
3187 		struct pmc_binding pb;
3188 
3189 		sx_assert(&pmc_sx, SX_XLOCKED);
3190 
3191 		KASSERT(td == curthread,
3192 		    ("[pmc,%d] td != curthread", __LINE__));
3193 
3194 		error = priv_check(td, PRIV_PMC_MANAGE);
3195 		if (error)
3196 			break;
3197 
3198 		if ((error = copyin(arg, &pma, sizeof(pma))) != 0)
3199 			break;
3200 
3201 		cpu = pma.pm_cpu;
3202 
3203 		if (cpu < 0 || cpu >= (int) pmc_cpu_max()) {
3204 			error = EINVAL;
3205 			break;
3206 		}
3207 
3208 		if (!pmc_cpu_is_active(cpu)) {
3209 			error = ENXIO;
3210 			break;
3211 		}
3212 
3213 		request = pma.pm_state;
3214 
3215 		if (request != PMC_STATE_DISABLED &&
3216 		    request != PMC_STATE_FREE) {
3217 			error = EINVAL;
3218 			break;
3219 		}
3220 
3221 		ri = pma.pm_pmc; /* pmc id == row index */
3222 		if (ri < 0 || ri >= (int) md->pmd_npmc) {
3223 			error = EINVAL;
3224 			break;
3225 		}
3226 
3227 		/*
3228 		 * We can't disable a PMC with a row-index allocated
3229 		 * for process virtual PMCs.
3230 		 */
3231 
3232 		if (PMC_ROW_DISP_IS_THREAD(ri) &&
3233 		    request == PMC_STATE_DISABLED) {
3234 			error = EBUSY;
3235 			break;
3236 		}
3237 
3238 		/*
3239 		 * otherwise, this PMC on this CPU is either free or
3240 		 * in system-wide mode.
3241 		 */
3242 
3243 		pmc_save_cpu_binding(&pb);
3244 		pmc_select_cpu(cpu);
3245 
3246 		pc  = pmc_pcpu[cpu];
3247 		phw = pc->pc_hwpmcs[ri];
3248 
3249 		/*
3250 		 * XXX do we need some kind of 'forced' disable?
3251 		 */
3252 
3253 		if (phw->phw_pmc == NULL) {
3254 			if (request == PMC_STATE_DISABLED &&
3255 			    (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED)) {
3256 				phw->phw_state &= ~PMC_PHW_FLAG_IS_ENABLED;
3257 				PMC_MARK_ROW_STANDALONE(ri);
3258 			} else if (request == PMC_STATE_FREE &&
3259 			    (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0) {
3260 				phw->phw_state |=  PMC_PHW_FLAG_IS_ENABLED;
3261 				PMC_UNMARK_ROW_STANDALONE(ri);
3262 			}
3263 			/* other cases are a no-op */
3264 		} else
3265 			error = EBUSY;
3266 
3267 		pmc_restore_cpu_binding(&pb);
3268 	}
3269 	break;
3270 
3271 
3272 	/*
3273 	 * Allocate a PMC.
3274 	 */
3275 
3276 	case PMC_OP_PMCALLOCATE:
3277 	{
3278 		int adjri, n;
3279 		u_int cpu;
3280 		uint32_t caps;
3281 		struct pmc *pmc;
3282 		enum pmc_mode mode;
3283 		struct pmc_hw *phw;
3284 		struct pmc_binding pb;
3285 		struct pmc_classdep *pcd;
3286 		struct pmc_op_pmcallocate pa;
3287 
3288 		if ((error = copyin(arg, &pa, sizeof(pa))) != 0)
3289 			break;
3290 
3291 		caps = pa.pm_caps;
3292 		mode = pa.pm_mode;
3293 		cpu  = pa.pm_cpu;
3294 
3295 		if ((mode != PMC_MODE_SS  &&  mode != PMC_MODE_SC  &&
3296 		     mode != PMC_MODE_TS  &&  mode != PMC_MODE_TC) ||
3297 		    (cpu != (u_int) PMC_CPU_ANY && cpu >= pmc_cpu_max())) {
3298 			error = EINVAL;
3299 			break;
3300 		}
3301 
3302 		/*
3303 		 * Virtual PMCs should only ask for a default CPU.
3304 		 * System mode PMCs need to specify a non-default CPU.
3305 		 */
3306 
3307 		if ((PMC_IS_VIRTUAL_MODE(mode) && cpu != (u_int) PMC_CPU_ANY) ||
3308 		    (PMC_IS_SYSTEM_MODE(mode) && cpu == (u_int) PMC_CPU_ANY)) {
3309 			error = EINVAL;
3310 			break;
3311 		}
3312 
3313 		/*
3314 		 * Check that an inactive CPU is not being asked for.
3315 		 */
3316 
3317 		if (PMC_IS_SYSTEM_MODE(mode) && !pmc_cpu_is_active(cpu)) {
3318 			error = ENXIO;
3319 			break;
3320 		}
3321 
3322 		/*
3323 		 * Refuse an allocation for a system-wide PMC if this
3324 		 * process has been jailed, or if this process lacks
3325 		 * super-user credentials and the sysctl tunable
3326 		 * 'security.bsd.unprivileged_syspmcs' is zero.
3327 		 */
3328 
3329 		if (PMC_IS_SYSTEM_MODE(mode)) {
3330 			if (jailed(curthread->td_ucred)) {
3331 				error = EPERM;
3332 				break;
3333 			}
3334 			if (!pmc_unprivileged_syspmcs) {
3335 				error = priv_check(curthread,
3336 				    PRIV_PMC_SYSTEM);
3337 				if (error)
3338 					break;
3339 			}
3340 		}
3341 
3342 		/*
3343 		 * Look for valid values for 'pm_flags'
3344 		 */
3345 
3346 		if ((pa.pm_flags & ~(PMC_F_DESCENDANTS | PMC_F_LOG_PROCCSW |
3347 		    PMC_F_LOG_PROCEXIT | PMC_F_CALLCHAIN)) != 0) {
3348 			error = EINVAL;
3349 			break;
3350 		}
3351 
3352 		/* process logging options are not allowed for system PMCs */
3353 		if (PMC_IS_SYSTEM_MODE(mode) && (pa.pm_flags &
3354 		    (PMC_F_LOG_PROCCSW | PMC_F_LOG_PROCEXIT))) {
3355 			error = EINVAL;
3356 			break;
3357 		}
3358 
3359 		/*
3360 		 * All sampling mode PMCs need to be able to interrupt the
3361 		 * CPU.
3362 		 */
3363 		if (PMC_IS_SAMPLING_MODE(mode))
3364 			caps |= PMC_CAP_INTERRUPT;
3365 
3366 		/* A valid class specifier should have been passed in. */
3367 		for (n = 0; n < md->pmd_nclass; n++)
3368 			if (md->pmd_classdep[n].pcd_class == pa.pm_class)
3369 				break;
3370 		if (n == md->pmd_nclass) {
3371 			error = EINVAL;
3372 			break;
3373 		}
3374 
3375 		/* The requested PMC capabilities should be feasible. */
3376 		if ((md->pmd_classdep[n].pcd_caps & caps) != caps) {
3377 			error = EOPNOTSUPP;
3378 			break;
3379 		}
3380 
3381 		PMCDBG(PMC,ALL,2, "event=%d caps=0x%x mode=%d cpu=%d",
3382 		    pa.pm_ev, caps, mode, cpu);
3383 
3384 		pmc = pmc_allocate_pmc_descriptor();
3385 		pmc->pm_id    = PMC_ID_MAKE_ID(cpu,pa.pm_mode,pa.pm_class,
3386 		    PMC_ID_INVALID);
3387 		pmc->pm_event = pa.pm_ev;
3388 		pmc->pm_state = PMC_STATE_FREE;
3389 		pmc->pm_caps  = caps;
3390 		pmc->pm_flags = pa.pm_flags;
3391 
3392 		/* switch thread to CPU 'cpu' */
3393 		pmc_save_cpu_binding(&pb);
3394 
3395 #define	PMC_IS_SHAREABLE_PMC(cpu, n)				\
3396 	(pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_state &		\
3397 	 PMC_PHW_FLAG_IS_SHAREABLE)
3398 #define	PMC_IS_UNALLOCATED(cpu, n)				\
3399 	(pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_pmc == NULL)
3400 
3401 		if (PMC_IS_SYSTEM_MODE(mode)) {
3402 			pmc_select_cpu(cpu);
3403 			for (n = 0; n < (int) md->pmd_npmc; n++) {
3404 				pcd = pmc_ri_to_classdep(md, n, &adjri);
3405 				if (pmc_can_allocate_row(n, mode) == 0 &&
3406 				    pmc_can_allocate_rowindex(
3407 					    curthread->td_proc, n, cpu) == 0 &&
3408 				    (PMC_IS_UNALLOCATED(cpu, n) ||
3409 				     PMC_IS_SHAREABLE_PMC(cpu, n)) &&
3410 				    pcd->pcd_allocate_pmc(cpu, adjri, pmc,
3411 					&pa) == 0)
3412 					break;
3413 			}
3414 		} else {
3415 			/* Process virtual mode */
3416 			for (n = 0; n < (int) md->pmd_npmc; n++) {
3417 				pcd = pmc_ri_to_classdep(md, n, &adjri);
3418 				if (pmc_can_allocate_row(n, mode) == 0 &&
3419 				    pmc_can_allocate_rowindex(
3420 					    curthread->td_proc, n,
3421 					    PMC_CPU_ANY) == 0 &&
3422 				    pcd->pcd_allocate_pmc(curthread->td_oncpu,
3423 					adjri, pmc, &pa) == 0)
3424 					break;
3425 			}
3426 		}
3427 
3428 #undef	PMC_IS_UNALLOCATED
3429 #undef	PMC_IS_SHAREABLE_PMC
3430 
3431 		pmc_restore_cpu_binding(&pb);
3432 
3433 		if (n == (int) md->pmd_npmc) {
3434 			pmc_destroy_pmc_descriptor(pmc);
3435 			free(pmc, M_PMC);
3436 			pmc = NULL;
3437 			error = EINVAL;
3438 			break;
3439 		}
3440 
3441 		/* Fill in the correct value in the ID field */
3442 		pmc->pm_id = PMC_ID_MAKE_ID(cpu,mode,pa.pm_class,n);
3443 
3444 		PMCDBG(PMC,ALL,2, "ev=%d class=%d mode=%d n=%d -> pmcid=%x",
3445 		    pmc->pm_event, pa.pm_class, mode, n, pmc->pm_id);
3446 
3447 		/* Process mode PMCs with logging enabled need log files */
3448 		if (pmc->pm_flags & (PMC_F_LOG_PROCEXIT | PMC_F_LOG_PROCCSW))
3449 			pmc->pm_flags |= PMC_F_NEEDS_LOGFILE;
3450 
3451 		/* All system mode sampling PMCs require a log file */
3452 		if (PMC_IS_SAMPLING_MODE(mode) && PMC_IS_SYSTEM_MODE(mode))
3453 			pmc->pm_flags |= PMC_F_NEEDS_LOGFILE;
3454 
3455 		/*
3456 		 * Configure global pmc's immediately
3457 		 */
3458 
3459 		if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pmc))) {
3460 
3461 			pmc_save_cpu_binding(&pb);
3462 			pmc_select_cpu(cpu);
3463 
3464 			phw = pmc_pcpu[cpu]->pc_hwpmcs[n];
3465 			pcd = pmc_ri_to_classdep(md, n, &adjri);
3466 
3467 			if ((phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0 ||
3468 			    (error = pcd->pcd_config_pmc(cpu, adjri, pmc)) != 0) {
3469 				(void) pcd->pcd_release_pmc(cpu, adjri, pmc);
3470 				pmc_destroy_pmc_descriptor(pmc);
3471 				free(pmc, M_PMC);
3472 				pmc = NULL;
3473 				pmc_restore_cpu_binding(&pb);
3474 				error = EPERM;
3475 				break;
3476 			}
3477 
3478 			pmc_restore_cpu_binding(&pb);
3479 		}
3480 
3481 		pmc->pm_state    = PMC_STATE_ALLOCATED;
3482 
3483 		/*
3484 		 * mark row disposition
3485 		 */
3486 
3487 		if (PMC_IS_SYSTEM_MODE(mode))
3488 			PMC_MARK_ROW_STANDALONE(n);
3489 		else
3490 			PMC_MARK_ROW_THREAD(n);
3491 
3492 		/*
3493 		 * Register this PMC with the current thread as its owner.
3494 		 */
3495 
3496 		if ((error =
3497 		    pmc_register_owner(curthread->td_proc, pmc)) != 0) {
3498 			pmc_release_pmc_descriptor(pmc);
3499 			free(pmc, M_PMC);
3500 			pmc = NULL;
3501 			break;
3502 		}
3503 
3504 		/*
3505 		 * Return the allocated index.
3506 		 */
3507 
3508 		pa.pm_pmcid = pmc->pm_id;
3509 
3510 		error = copyout(&pa, arg, sizeof(pa));
3511 	}
3512 	break;
3513 
3514 
3515 	/*
3516 	 * Attach a PMC to a process.
3517 	 */
3518 
3519 	case PMC_OP_PMCATTACH:
3520 	{
3521 		struct pmc *pm;
3522 		struct proc *p;
3523 		struct pmc_op_pmcattach a;
3524 
3525 		sx_assert(&pmc_sx, SX_XLOCKED);
3526 
3527 		if ((error = copyin(arg, &a, sizeof(a))) != 0)
3528 			break;
3529 
3530 		if (a.pm_pid < 0) {
3531 			error = EINVAL;
3532 			break;
3533 		} else if (a.pm_pid == 0)
3534 			a.pm_pid = td->td_proc->p_pid;
3535 
3536 		if ((error = pmc_find_pmc(a.pm_pmc, &pm)) != 0)
3537 			break;
3538 
3539 		if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm))) {
3540 			error = EINVAL;
3541 			break;
3542 		}
3543 
3544 		/* PMCs may be (re)attached only when allocated or stopped */
3545 		if (pm->pm_state == PMC_STATE_RUNNING) {
3546 			error = EBUSY;
3547 			break;
3548 		} else if (pm->pm_state != PMC_STATE_ALLOCATED &&
3549 		    pm->pm_state != PMC_STATE_STOPPED) {
3550 			error = EINVAL;
3551 			break;
3552 		}
3553 
3554 		/* lookup pid */
3555 		if ((p = pfind(a.pm_pid)) == NULL) {
3556 			error = ESRCH;
3557 			break;
3558 		}
3559 
3560 		/*
3561 		 * Ignore processes that are working on exiting.
3562 		 */
3563 		if (p->p_flag & P_WEXIT) {
3564 			error = ESRCH;
3565 			PROC_UNLOCK(p);	/* pfind() returns a locked process */
3566 			break;
3567 		}
3568 
3569 		/*
3570 		 * we are allowed to attach a PMC to a process if
3571 		 * we can debug it.
3572 		 */
3573 		error = p_candebug(curthread, p);
3574 
3575 		PROC_UNLOCK(p);
3576 
3577 		if (error == 0)
3578 			error = pmc_attach_process(p, pm);
3579 	}
3580 	break;
3581 
3582 
3583 	/*
3584 	 * Detach an attached PMC from a process.
3585 	 */
3586 
3587 	case PMC_OP_PMCDETACH:
3588 	{
3589 		struct pmc *pm;
3590 		struct proc *p;
3591 		struct pmc_op_pmcattach a;
3592 
3593 		if ((error = copyin(arg, &a, sizeof(a))) != 0)
3594 			break;
3595 
3596 		if (a.pm_pid < 0) {
3597 			error = EINVAL;
3598 			break;
3599 		} else if (a.pm_pid == 0)
3600 			a.pm_pid = td->td_proc->p_pid;
3601 
3602 		if ((error = pmc_find_pmc(a.pm_pmc, &pm)) != 0)
3603 			break;
3604 
3605 		if ((p = pfind(a.pm_pid)) == NULL) {
3606 			error = ESRCH;
3607 			break;
3608 		}
3609 
3610 		/*
3611 		 * Treat processes that are in the process of exiting
3612 		 * as if they were not present.
3613 		 */
3614 
3615 		if (p->p_flag & P_WEXIT)
3616 			error = ESRCH;
3617 
3618 		PROC_UNLOCK(p);	/* pfind() returns a locked process */
3619 
3620 		if (error == 0)
3621 			error = pmc_detach_process(p, pm);
3622 	}
3623 	break;
3624 
3625 
3626 	/*
3627 	 * Retrieve the MSR number associated with the counter
3628 	 * 'pmc_id'.  This allows processes to directly use RDPMC
3629 	 * instructions to read their PMCs, without the overhead of a
3630 	 * system call.
3631 	 */
3632 
3633 	case PMC_OP_PMCGETMSR:
3634 	{
3635 		int adjri, ri;
3636 		struct pmc *pm;
3637 		struct pmc_target *pt;
3638 		struct pmc_op_getmsr gm;
3639 		struct pmc_classdep *pcd;
3640 
3641 		PMC_DOWNGRADE_SX();
3642 
3643 		if ((error = copyin(arg, &gm, sizeof(gm))) != 0)
3644 			break;
3645 
3646 		if ((error = pmc_find_pmc(gm.pm_pmcid, &pm)) != 0)
3647 			break;
3648 
3649 		/*
3650 		 * The allocated PMC has to be a process virtual PMC,
3651 		 * i.e., of type MODE_T[CS].  Global PMCs can only be
3652 		 * read using the PMCREAD operation since they may be
3653 		 * allocated on a different CPU than the one we could
3654 		 * be running on at the time of the RDPMC instruction.
3655 		 *
3656 		 * The GETMSR operation is not allowed for PMCs that
3657 		 * are inherited across processes.
3658 		 */
3659 
3660 		if (!PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)) ||
3661 		    (pm->pm_flags & PMC_F_DESCENDANTS)) {
3662 			error = EINVAL;
3663 			break;
3664 		}
3665 
3666 		/*
3667 		 * It only makes sense to use a RDPMC (or its
3668 		 * equivalent instruction on non-x86 architectures) on
3669 		 * a process that has allocated and attached a PMC to
3670 		 * itself.  Conversely the PMC is only allowed to have
3671 		 * one process attached to it -- its owner.
3672 		 */
3673 
3674 		if ((pt = LIST_FIRST(&pm->pm_targets)) == NULL ||
3675 		    LIST_NEXT(pt, pt_next) != NULL ||
3676 		    pt->pt_process->pp_proc != pm->pm_owner->po_owner) {
3677 			error = EINVAL;
3678 			break;
3679 		}
3680 
3681 		ri = PMC_TO_ROWINDEX(pm);
3682 		pcd = pmc_ri_to_classdep(md, ri, &adjri);
3683 
3684 		/* PMC class has no 'GETMSR' support */
3685 		if (pcd->pcd_get_msr == NULL) {
3686 			error = ENOSYS;
3687 			break;
3688 		}
3689 
3690 		if ((error = (*pcd->pcd_get_msr)(adjri, &gm.pm_msr)) < 0)
3691 			break;
3692 
3693 		if ((error = copyout(&gm, arg, sizeof(gm))) < 0)
3694 			break;
3695 
3696 		/*
3697 		 * Mark our process as using MSRs.  Update machine
3698 		 * state using a forced context switch.
3699 		 */
3700 
3701 		pt->pt_process->pp_flags |= PMC_PP_ENABLE_MSR_ACCESS;
3702 		pmc_force_context_switch();
3703 
3704 	}
3705 	break;
3706 
3707 	/*
3708 	 * Release an allocated PMC
3709 	 */
3710 
3711 	case PMC_OP_PMCRELEASE:
3712 	{
3713 		pmc_id_t pmcid;
3714 		struct pmc *pm;
3715 		struct pmc_owner *po;
3716 		struct pmc_op_simple sp;
3717 
3718 		/*
3719 		 * Find PMC pointer for the named PMC.
3720 		 *
3721 		 * Use pmc_release_pmc_descriptor() to switch off the
3722 		 * PMC, remove all its target threads, and remove the
3723 		 * PMC from its owner's list.
3724 		 *
3725 		 * Remove the owner record if this is the last PMC
3726 		 * owned.
3727 		 *
3728 		 * Free up space.
3729 		 */
3730 
3731 		if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
3732 			break;
3733 
3734 		pmcid = sp.pm_pmcid;
3735 
3736 		if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
3737 			break;
3738 
3739 		po = pm->pm_owner;
3740 		pmc_release_pmc_descriptor(pm);
3741 		pmc_maybe_remove_owner(po);
3742 
3743 		free(pm, M_PMC);
3744 	}
3745 	break;
3746 
3747 
3748 	/*
3749 	 * Read and/or write a PMC.
3750 	 */
3751 
3752 	case PMC_OP_PMCRW:
3753 	{
3754 		int adjri;
3755 		struct pmc *pm;
3756 		uint32_t cpu, ri;
3757 		pmc_value_t oldvalue;
3758 		struct pmc_binding pb;
3759 		struct pmc_op_pmcrw prw;
3760 		struct pmc_classdep *pcd;
3761 		struct pmc_op_pmcrw *pprw;
3762 
3763 		PMC_DOWNGRADE_SX();
3764 
3765 		if ((error = copyin(arg, &prw, sizeof(prw))) != 0)
3766 			break;
3767 
3768 		ri = 0;
3769 		PMCDBG(PMC,OPS,1, "rw id=%d flags=0x%x", prw.pm_pmcid,
3770 		    prw.pm_flags);
3771 
3772 		/* must have at least one flag set */
3773 		if ((prw.pm_flags & (PMC_F_OLDVALUE|PMC_F_NEWVALUE)) == 0) {
3774 			error = EINVAL;
3775 			break;
3776 		}
3777 
3778 		/* locate pmc descriptor */
3779 		if ((error = pmc_find_pmc(prw.pm_pmcid, &pm)) != 0)
3780 			break;
3781 
3782 		/* Can't read a PMC that hasn't been started. */
3783 		if (pm->pm_state != PMC_STATE_ALLOCATED &&
3784 		    pm->pm_state != PMC_STATE_STOPPED &&
3785 		    pm->pm_state != PMC_STATE_RUNNING) {
3786 			error = EINVAL;
3787 			break;
3788 		}
3789 
3790 		/* writing a new value is allowed only for 'STOPPED' pmcs */
3791 		if (pm->pm_state == PMC_STATE_RUNNING &&
3792 		    (prw.pm_flags & PMC_F_NEWVALUE)) {
3793 			error = EBUSY;
3794 			break;
3795 		}
3796 
3797 		if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) {
3798 
3799 			/*
3800 			 * If this PMC is attached to its owner (i.e.,
3801 			 * the process requesting this operation) and
3802 			 * is running, then attempt to get an
3803 			 * upto-date reading from hardware for a READ.
3804 			 * Writes are only allowed when the PMC is
3805 			 * stopped, so only update the saved value
3806 			 * field.
3807 			 *
3808 			 * If the PMC is not running, or is not
3809 			 * attached to its owner, read/write to the
3810 			 * savedvalue field.
3811 			 */
3812 
3813 			ri = PMC_TO_ROWINDEX(pm);
3814 			pcd = pmc_ri_to_classdep(md, ri, &adjri);
3815 
3816 			mtx_pool_lock_spin(pmc_mtxpool, pm);
3817 			cpu = curthread->td_oncpu;
3818 
3819 			if (prw.pm_flags & PMC_F_OLDVALUE) {
3820 				if ((pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) &&
3821 				    (pm->pm_state == PMC_STATE_RUNNING))
3822 					error = (*pcd->pcd_read_pmc)(cpu, adjri,
3823 					    &oldvalue);
3824 				else
3825 					oldvalue = pm->pm_gv.pm_savedvalue;
3826 			}
3827 			if (prw.pm_flags & PMC_F_NEWVALUE)
3828 				pm->pm_gv.pm_savedvalue = prw.pm_value;
3829 
3830 			mtx_pool_unlock_spin(pmc_mtxpool, pm);
3831 
3832 		} else { /* System mode PMCs */
3833 			cpu = PMC_TO_CPU(pm);
3834 			ri  = PMC_TO_ROWINDEX(pm);
3835 			pcd = pmc_ri_to_classdep(md, ri, &adjri);
3836 
3837 			if (!pmc_cpu_is_active(cpu)) {
3838 				error = ENXIO;
3839 				break;
3840 			}
3841 
3842 			/* move this thread to CPU 'cpu' */
3843 			pmc_save_cpu_binding(&pb);
3844 			pmc_select_cpu(cpu);
3845 
3846 			critical_enter();
3847 			/* save old value */
3848 			if (prw.pm_flags & PMC_F_OLDVALUE)
3849 				if ((error = (*pcd->pcd_read_pmc)(cpu, adjri,
3850 					 &oldvalue)))
3851 					goto error;
3852 			/* write out new value */
3853 			if (prw.pm_flags & PMC_F_NEWVALUE)
3854 				error = (*pcd->pcd_write_pmc)(cpu, adjri,
3855 				    prw.pm_value);
3856 		error:
3857 			critical_exit();
3858 			pmc_restore_cpu_binding(&pb);
3859 			if (error)
3860 				break;
3861 		}
3862 
3863 		pprw = (struct pmc_op_pmcrw *) arg;
3864 
3865 #ifdef	DEBUG
3866 		if (prw.pm_flags & PMC_F_NEWVALUE)
3867 			PMCDBG(PMC,OPS,2, "rw id=%d new %jx -> old %jx",
3868 			    ri, prw.pm_value, oldvalue);
3869 		else if (prw.pm_flags & PMC_F_OLDVALUE)
3870 			PMCDBG(PMC,OPS,2, "rw id=%d -> old %jx", ri, oldvalue);
3871 #endif
3872 
3873 		/* return old value if requested */
3874 		if (prw.pm_flags & PMC_F_OLDVALUE)
3875 			if ((error = copyout(&oldvalue, &pprw->pm_value,
3876 				 sizeof(prw.pm_value))))
3877 				break;
3878 
3879 	}
3880 	break;
3881 
3882 
3883 	/*
3884 	 * Set the sampling rate for a sampling mode PMC and the
3885 	 * initial count for a counting mode PMC.
3886 	 */
3887 
3888 	case PMC_OP_PMCSETCOUNT:
3889 	{
3890 		struct pmc *pm;
3891 		struct pmc_op_pmcsetcount sc;
3892 
3893 		PMC_DOWNGRADE_SX();
3894 
3895 		if ((error = copyin(arg, &sc, sizeof(sc))) != 0)
3896 			break;
3897 
3898 		if ((error = pmc_find_pmc(sc.pm_pmcid, &pm)) != 0)
3899 			break;
3900 
3901 		if (pm->pm_state == PMC_STATE_RUNNING) {
3902 			error = EBUSY;
3903 			break;
3904 		}
3905 
3906 		if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
3907 			pm->pm_sc.pm_reloadcount = sc.pm_count;
3908 		else
3909 			pm->pm_sc.pm_initial = sc.pm_count;
3910 	}
3911 	break;
3912 
3913 
3914 	/*
3915 	 * Start a PMC.
3916 	 */
3917 
3918 	case PMC_OP_PMCSTART:
3919 	{
3920 		pmc_id_t pmcid;
3921 		struct pmc *pm;
3922 		struct pmc_op_simple sp;
3923 
3924 		sx_assert(&pmc_sx, SX_XLOCKED);
3925 
3926 		if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
3927 			break;
3928 
3929 		pmcid = sp.pm_pmcid;
3930 
3931 		if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
3932 			break;
3933 
3934 		KASSERT(pmcid == pm->pm_id,
3935 		    ("[pmc,%d] pmcid %x != id %x", __LINE__,
3936 			pm->pm_id, pmcid));
3937 
3938 		if (pm->pm_state == PMC_STATE_RUNNING) /* already running */
3939 			break;
3940 		else if (pm->pm_state != PMC_STATE_STOPPED &&
3941 		    pm->pm_state != PMC_STATE_ALLOCATED) {
3942 			error = EINVAL;
3943 			break;
3944 		}
3945 
3946 		error = pmc_start(pm);
3947 	}
3948 	break;
3949 
3950 
3951 	/*
3952 	 * Stop a PMC.
3953 	 */
3954 
3955 	case PMC_OP_PMCSTOP:
3956 	{
3957 		pmc_id_t pmcid;
3958 		struct pmc *pm;
3959 		struct pmc_op_simple sp;
3960 
3961 		PMC_DOWNGRADE_SX();
3962 
3963 		if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
3964 			break;
3965 
3966 		pmcid = sp.pm_pmcid;
3967 
3968 		/*
3969 		 * Mark the PMC as inactive and invoke the MD stop
3970 		 * routines if needed.
3971 		 */
3972 
3973 		if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
3974 			break;
3975 
3976 		KASSERT(pmcid == pm->pm_id,
3977 		    ("[pmc,%d] pmc id %x != pmcid %x", __LINE__,
3978 			pm->pm_id, pmcid));
3979 
3980 		if (pm->pm_state == PMC_STATE_STOPPED) /* already stopped */
3981 			break;
3982 		else if (pm->pm_state != PMC_STATE_RUNNING) {
3983 			error = EINVAL;
3984 			break;
3985 		}
3986 
3987 		error = pmc_stop(pm);
3988 	}
3989 	break;
3990 
3991 
3992 	/*
3993 	 * Write a user supplied value to the log file.
3994 	 */
3995 
3996 	case PMC_OP_WRITELOG:
3997 	{
3998 		struct pmc_op_writelog wl;
3999 		struct pmc_owner *po;
4000 
4001 		PMC_DOWNGRADE_SX();
4002 
4003 		if ((error = copyin(arg, &wl, sizeof(wl))) != 0)
4004 			break;
4005 
4006 		if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) {
4007 			error = EINVAL;
4008 			break;
4009 		}
4010 
4011 		if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) {
4012 			error = EINVAL;
4013 			break;
4014 		}
4015 
4016 		error = pmclog_process_userlog(po, &wl);
4017 	}
4018 	break;
4019 
4020 
4021 	default:
4022 		error = EINVAL;
4023 		break;
4024 	}
4025 
4026 	if (is_sx_locked != 0) {
4027 		if (is_sx_downgraded)
4028 			sx_sunlock(&pmc_sx);
4029 		else
4030 			sx_xunlock(&pmc_sx);
4031 	}
4032 
4033 	if (error)
4034 		atomic_add_int(&pmc_stats.pm_syscall_errors, 1);
4035 
4036 	PICKUP_GIANT();
4037 
4038 	return error;
4039 }
4040 
4041 /*
4042  * Helper functions
4043  */
4044 
4045 
4046 /*
4047  * Mark the thread as needing callchain capture and post an AST.  The
4048  * actual callchain capture will be done in a context where it is safe
4049  * to take page faults.
4050  */
4051 
4052 static void
4053 pmc_post_callchain_callback(void)
4054 {
4055 	struct thread *td;
4056 
4057 	td = curthread;
4058 
4059 	/*
4060 	 * If there is multiple PMCs for the same interrupt ignore new post
4061 	 */
4062 	if (td->td_pflags & TDP_CALLCHAIN)
4063 		return;
4064 
4065 	/*
4066 	 * Mark this thread as needing callchain capture.
4067 	 * `td->td_pflags' will be safe to touch because this thread
4068 	 * was in user space when it was interrupted.
4069 	 */
4070 	td->td_pflags |= TDP_CALLCHAIN;
4071 
4072 	/*
4073 	 * Don't let this thread migrate between CPUs until callchain
4074 	 * capture completes.
4075 	 */
4076 	sched_pin();
4077 
4078 	return;
4079 }
4080 
4081 /*
4082  * Interrupt processing.
4083  *
4084  * Find a free slot in the per-cpu array of samples and capture the
4085  * current callchain there.  If a sample was successfully added, a bit
4086  * is set in mask 'pmc_cpumask' denoting that the DO_SAMPLES hook
4087  * needs to be invoked from the clock handler.
4088  *
4089  * This function is meant to be called from an NMI handler.  It cannot
4090  * use any of the locking primitives supplied by the OS.
4091  */
4092 
4093 int
4094 pmc_process_interrupt(int cpu, int ring, struct pmc *pm, struct trapframe *tf,
4095     int inuserspace)
4096 {
4097 	int error, callchaindepth;
4098 	struct thread *td;
4099 	struct pmc_sample *ps;
4100 	struct pmc_samplebuffer *psb;
4101 
4102 	error = 0;
4103 
4104 	/*
4105 	 * Allocate space for a sample buffer.
4106 	 */
4107 	psb = pmc_pcpu[cpu]->pc_sb[ring];
4108 
4109 	ps = psb->ps_write;
4110 	if (ps->ps_nsamples) {	/* in use, reader hasn't caught up */
4111 		pm->pm_stalled = 1;
4112 		atomic_add_int(&pmc_stats.pm_intr_bufferfull, 1);
4113 		PMCDBG(SAM,INT,1,"(spc) cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d",
4114 		    cpu, pm, (void *) tf, inuserspace,
4115 		    (int) (psb->ps_write - psb->ps_samples),
4116 		    (int) (psb->ps_read - psb->ps_samples));
4117 		error = ENOMEM;
4118 		goto done;
4119 	}
4120 
4121 
4122 	/* Fill in entry. */
4123 	PMCDBG(SAM,INT,1,"cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d", cpu, pm,
4124 	    (void *) tf, inuserspace,
4125 	    (int) (psb->ps_write - psb->ps_samples),
4126 	    (int) (psb->ps_read - psb->ps_samples));
4127 
4128 	KASSERT(pm->pm_runcount >= 0,
4129 	    ("[pmc,%d] pm=%p runcount %d", __LINE__, (void *) pm,
4130 		pm->pm_runcount));
4131 
4132 	atomic_add_rel_int(&pm->pm_runcount, 1);	/* hold onto PMC */
4133 
4134 	ps->ps_pmc = pm;
4135 	if ((td = curthread) && td->td_proc)
4136 		ps->ps_pid = td->td_proc->p_pid;
4137 	else
4138 		ps->ps_pid = -1;
4139 	ps->ps_cpu = cpu;
4140 	ps->ps_td = td;
4141 	ps->ps_flags = inuserspace ? PMC_CC_F_USERSPACE : 0;
4142 
4143 	callchaindepth = (pm->pm_flags & PMC_F_CALLCHAIN) ?
4144 	    pmc_callchaindepth : 1;
4145 
4146 	if (callchaindepth == 1)
4147 		ps->ps_pc[0] = PMC_TRAPFRAME_TO_PC(tf);
4148 	else {
4149 		/*
4150 		 * Kernel stack traversals can be done immediately,
4151 		 * while we defer to an AST for user space traversals.
4152 		 */
4153 		if (!inuserspace) {
4154 			callchaindepth =
4155 			    pmc_save_kernel_callchain(ps->ps_pc,
4156 				callchaindepth, tf);
4157 		} else {
4158 			pmc_post_callchain_callback();
4159 			callchaindepth = PMC_SAMPLE_INUSE;
4160 		}
4161 	}
4162 
4163 	ps->ps_nsamples = callchaindepth;	/* mark entry as in use */
4164 
4165 	/* increment write pointer, modulo ring buffer size */
4166 	ps++;
4167 	if (ps == psb->ps_fence)
4168 		psb->ps_write = psb->ps_samples;
4169 	else
4170 		psb->ps_write = ps;
4171 
4172  done:
4173 	/* mark CPU as needing processing */
4174 	CPU_SET_ATOMIC(cpu, &pmc_cpumask);
4175 
4176 	return (error);
4177 }
4178 
4179 /*
4180  * Capture a user call chain.  This function will be called from ast()
4181  * before control returns to userland and before the process gets
4182  * rescheduled.
4183  */
4184 
4185 static void
4186 pmc_capture_user_callchain(int cpu, int ring, struct trapframe *tf)
4187 {
4188 	int i;
4189 	struct pmc *pm;
4190 	struct thread *td;
4191 	struct pmc_sample *ps;
4192 	struct pmc_samplebuffer *psb;
4193 #ifdef	INVARIANTS
4194 	int ncallchains;
4195 #endif
4196 
4197 	psb = pmc_pcpu[cpu]->pc_sb[ring];
4198 	td = curthread;
4199 
4200 	KASSERT(td->td_pflags & TDP_CALLCHAIN,
4201 	    ("[pmc,%d] Retrieving callchain for thread that doesn't want it",
4202 		__LINE__));
4203 
4204 #ifdef	INVARIANTS
4205 	ncallchains = 0;
4206 #endif
4207 
4208 	/*
4209 	 * Iterate through all deferred callchain requests.
4210 	 */
4211 
4212 	ps = psb->ps_samples;
4213 	for (i = 0; i < pmc_nsamples; i++, ps++) {
4214 
4215 		if (ps->ps_nsamples != PMC_SAMPLE_INUSE)
4216 			continue;
4217 		if (ps->ps_td != td)
4218 			continue;
4219 
4220 		KASSERT(ps->ps_cpu == cpu,
4221 		    ("[pmc,%d] cpu mismatch ps_cpu=%d pcpu=%d", __LINE__,
4222 			ps->ps_cpu, PCPU_GET(cpuid)));
4223 
4224 		pm = ps->ps_pmc;
4225 
4226 		KASSERT(pm->pm_flags & PMC_F_CALLCHAIN,
4227 		    ("[pmc,%d] Retrieving callchain for PMC that doesn't "
4228 			"want it", __LINE__));
4229 
4230 		KASSERT(pm->pm_runcount > 0,
4231 		    ("[pmc,%d] runcount %d", __LINE__, pm->pm_runcount));
4232 
4233 		/*
4234 		 * Retrieve the callchain and mark the sample buffer
4235 		 * as 'processable' by the timer tick sweep code.
4236 		 */
4237 		ps->ps_nsamples = pmc_save_user_callchain(ps->ps_pc,
4238 		    pmc_callchaindepth, tf);
4239 
4240 #ifdef	INVARIANTS
4241 		ncallchains++;
4242 #endif
4243 	}
4244 
4245 	KASSERT(ncallchains > 0,
4246 	    ("[pmc,%d] cpu %d didn't find a sample to collect", __LINE__,
4247 		cpu));
4248 
4249 	KASSERT(td->td_pinned == 1,
4250 	    ("[pmc,%d] invalid td_pinned value", __LINE__));
4251 	sched_unpin();	/* Can migrate safely now. */
4252 
4253 	return;
4254 }
4255 
4256 /*
4257  * Process saved PC samples.
4258  */
4259 
4260 static void
4261 pmc_process_samples(int cpu, int ring)
4262 {
4263 	struct pmc *pm;
4264 	int adjri, n;
4265 	struct thread *td;
4266 	struct pmc_owner *po;
4267 	struct pmc_sample *ps;
4268 	struct pmc_classdep *pcd;
4269 	struct pmc_samplebuffer *psb;
4270 
4271 	KASSERT(PCPU_GET(cpuid) == cpu,
4272 	    ("[pmc,%d] not on the correct CPU pcpu=%d cpu=%d", __LINE__,
4273 		PCPU_GET(cpuid), cpu));
4274 
4275 	psb = pmc_pcpu[cpu]->pc_sb[ring];
4276 
4277 	for (n = 0; n < pmc_nsamples; n++) { /* bound on #iterations */
4278 
4279 		ps = psb->ps_read;
4280 		if (ps->ps_nsamples == PMC_SAMPLE_FREE)
4281 			break;
4282 
4283 		pm = ps->ps_pmc;
4284 
4285 		KASSERT(pm->pm_runcount > 0,
4286 		    ("[pmc,%d] pm=%p runcount %d", __LINE__, (void *) pm,
4287 			pm->pm_runcount));
4288 
4289 		po = pm->pm_owner;
4290 
4291 		KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)),
4292 		    ("[pmc,%d] pmc=%p non-sampling mode=%d", __LINE__,
4293 			pm, PMC_TO_MODE(pm)));
4294 
4295 		/* Ignore PMCs that have been switched off */
4296 		if (pm->pm_state != PMC_STATE_RUNNING)
4297 			goto entrydone;
4298 
4299 		/* If there is a pending AST wait for completion */
4300 		if (ps->ps_nsamples == PMC_SAMPLE_INUSE) {
4301 			/* Need a rescan at a later time. */
4302 			CPU_SET_ATOMIC(cpu, &pmc_cpumask);
4303 			break;
4304 		}
4305 
4306 		PMCDBG(SAM,OPS,1,"cpu=%d pm=%p n=%d fl=%x wr=%d rd=%d", cpu,
4307 		    pm, ps->ps_nsamples, ps->ps_flags,
4308 		    (int) (psb->ps_write - psb->ps_samples),
4309 		    (int) (psb->ps_read - psb->ps_samples));
4310 
4311 		/*
4312 		 * If this is a process-mode PMC that is attached to
4313 		 * its owner, and if the PC is in user mode, update
4314 		 * profiling statistics like timer-based profiling
4315 		 * would have done.
4316 		 */
4317 		if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) {
4318 			if (ps->ps_flags & PMC_CC_F_USERSPACE) {
4319 				td = FIRST_THREAD_IN_PROC(po->po_owner);
4320 				addupc_intr(td, ps->ps_pc[0], 1);
4321 			}
4322 			goto entrydone;
4323 		}
4324 
4325 		/*
4326 		 * Otherwise, this is either a sampling mode PMC that
4327 		 * is attached to a different process than its owner,
4328 		 * or a system-wide sampling PMC.  Dispatch a log
4329 		 * entry to the PMC's owner process.
4330 		 */
4331 		pmclog_process_callchain(pm, ps);
4332 
4333 	entrydone:
4334 		ps->ps_nsamples = 0; /* mark entry as free */
4335 		atomic_subtract_rel_int(&pm->pm_runcount, 1);
4336 
4337 		/* increment read pointer, modulo sample size */
4338 		if (++ps == psb->ps_fence)
4339 			psb->ps_read = psb->ps_samples;
4340 		else
4341 			psb->ps_read = ps;
4342 	}
4343 
4344 	atomic_add_int(&pmc_stats.pm_log_sweeps, 1);
4345 
4346 	/* Do not re-enable stalled PMCs if we failed to process any samples */
4347 	if (n == 0)
4348 		return;
4349 
4350 	/*
4351 	 * Restart any stalled sampling PMCs on this CPU.
4352 	 *
4353 	 * If the NMI handler sets the pm_stalled field of a PMC after
4354 	 * the check below, we'll end up processing the stalled PMC at
4355 	 * the next hardclock tick.
4356 	 */
4357 	for (n = 0; n < md->pmd_npmc; n++) {
4358 		pcd = pmc_ri_to_classdep(md, n, &adjri);
4359 		KASSERT(pcd != NULL,
4360 		    ("[pmc,%d] null pcd ri=%d", __LINE__, n));
4361 		(void) (*pcd->pcd_get_config)(cpu,adjri,&pm);
4362 
4363 		if (pm == NULL ||			 /* !cfg'ed */
4364 		    pm->pm_state != PMC_STATE_RUNNING || /* !active */
4365 		    !PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) || /* !sampling */
4366 		    pm->pm_stalled == 0) /* !stalled */
4367 			continue;
4368 
4369 		pm->pm_stalled = 0;
4370 		(*pcd->pcd_start_pmc)(cpu, adjri);
4371 	}
4372 }
4373 
4374 /*
4375  * Event handlers.
4376  */
4377 
4378 /*
4379  * Handle a process exit.
4380  *
4381  * Remove this process from all hash tables.  If this process
4382  * owned any PMCs, turn off those PMCs and deallocate them,
4383  * removing any associations with target processes.
4384  *
4385  * This function will be called by the last 'thread' of a
4386  * process.
4387  *
4388  * XXX This eventhandler gets called early in the exit process.
4389  * Consider using a 'hook' invocation from thread_exit() or equivalent
4390  * spot.  Another negative is that kse_exit doesn't seem to call
4391  * exit1() [??].
4392  *
4393  */
4394 
4395 static void
4396 pmc_process_exit(void *arg __unused, struct proc *p)
4397 {
4398 	struct pmc *pm;
4399 	int adjri, cpu;
4400 	unsigned int ri;
4401 	int is_using_hwpmcs;
4402 	struct pmc_owner *po;
4403 	struct pmc_process *pp;
4404 	struct pmc_classdep *pcd;
4405 	pmc_value_t newvalue, tmp;
4406 
4407 	PROC_LOCK(p);
4408 	is_using_hwpmcs = p->p_flag & P_HWPMC;
4409 	PROC_UNLOCK(p);
4410 
4411 	/*
4412 	 * Log a sysexit event to all SS PMC owners.
4413 	 */
4414 	LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
4415 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
4416 		    pmclog_process_sysexit(po, p->p_pid);
4417 
4418 	if (!is_using_hwpmcs)
4419 		return;
4420 
4421 	PMC_GET_SX_XLOCK();
4422 	PMCDBG(PRC,EXT,1,"process-exit proc=%p (%d, %s)", p, p->p_pid,
4423 	    p->p_comm);
4424 
4425 	/*
4426 	 * Since this code is invoked by the last thread in an exiting
4427 	 * process, we would have context switched IN at some prior
4428 	 * point.  However, with PREEMPTION, kernel mode context
4429 	 * switches may happen any time, so we want to disable a
4430 	 * context switch OUT till we get any PMCs targetting this
4431 	 * process off the hardware.
4432 	 *
4433 	 * We also need to atomically remove this process'
4434 	 * entry from our target process hash table, using
4435 	 * PMC_FLAG_REMOVE.
4436 	 */
4437 	PMCDBG(PRC,EXT,1, "process-exit proc=%p (%d, %s)", p, p->p_pid,
4438 	    p->p_comm);
4439 
4440 	critical_enter(); /* no preemption */
4441 
4442 	cpu = curthread->td_oncpu;
4443 
4444 	if ((pp = pmc_find_process_descriptor(p,
4445 		 PMC_FLAG_REMOVE)) != NULL) {
4446 
4447 		PMCDBG(PRC,EXT,2,
4448 		    "process-exit proc=%p pmc-process=%p", p, pp);
4449 
4450 		/*
4451 		 * The exiting process could the target of
4452 		 * some PMCs which will be running on
4453 		 * currently executing CPU.
4454 		 *
4455 		 * We need to turn these PMCs off like we
4456 		 * would do at context switch OUT time.
4457 		 */
4458 		for (ri = 0; ri < md->pmd_npmc; ri++) {
4459 
4460 			/*
4461 			 * Pick up the pmc pointer from hardware
4462 			 * state similar to the CSW_OUT code.
4463 			 */
4464 			pm = NULL;
4465 
4466 			pcd = pmc_ri_to_classdep(md, ri, &adjri);
4467 
4468 			(void) (*pcd->pcd_get_config)(cpu, adjri, &pm);
4469 
4470 			PMCDBG(PRC,EXT,2, "ri=%d pm=%p", ri, pm);
4471 
4472 			if (pm == NULL ||
4473 			    !PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)))
4474 				continue;
4475 
4476 			PMCDBG(PRC,EXT,2, "ppmcs[%d]=%p pm=%p "
4477 			    "state=%d", ri, pp->pp_pmcs[ri].pp_pmc,
4478 			    pm, pm->pm_state);
4479 
4480 			KASSERT(PMC_TO_ROWINDEX(pm) == ri,
4481 			    ("[pmc,%d] ri mismatch pmc(%d) ri(%d)",
4482 				__LINE__, PMC_TO_ROWINDEX(pm), ri));
4483 
4484 			KASSERT(pm == pp->pp_pmcs[ri].pp_pmc,
4485 			    ("[pmc,%d] pm %p != pp_pmcs[%d] %p",
4486 				__LINE__, pm, ri, pp->pp_pmcs[ri].pp_pmc));
4487 
4488 			(void) pcd->pcd_stop_pmc(cpu, adjri);
4489 
4490 			KASSERT(pm->pm_runcount > 0,
4491 			    ("[pmc,%d] bad runcount ri %d rc %d",
4492 				__LINE__, ri, pm->pm_runcount));
4493 
4494 			/* Stop hardware only if it is actually running */
4495 			if (pm->pm_state == PMC_STATE_RUNNING &&
4496 			    pm->pm_stalled == 0) {
4497 				pcd->pcd_read_pmc(cpu, adjri, &newvalue);
4498 				tmp = newvalue -
4499 				    PMC_PCPU_SAVED(cpu,ri);
4500 
4501 				mtx_pool_lock_spin(pmc_mtxpool, pm);
4502 				pm->pm_gv.pm_savedvalue += tmp;
4503 				pp->pp_pmcs[ri].pp_pmcval += tmp;
4504 				mtx_pool_unlock_spin(pmc_mtxpool, pm);
4505 			}
4506 
4507 			atomic_subtract_rel_int(&pm->pm_runcount,1);
4508 
4509 			KASSERT((int) pm->pm_runcount >= 0,
4510 			    ("[pmc,%d] runcount is %d", __LINE__, ri));
4511 
4512 			(void) pcd->pcd_config_pmc(cpu, adjri, NULL);
4513 		}
4514 
4515 		/*
4516 		 * Inform the MD layer of this pseudo "context switch
4517 		 * out"
4518 		 */
4519 		(void) md->pmd_switch_out(pmc_pcpu[cpu], pp);
4520 
4521 		critical_exit(); /* ok to be pre-empted now */
4522 
4523 		/*
4524 		 * Unlink this process from the PMCs that are
4525 		 * targetting it.  This will send a signal to
4526 		 * all PMC owner's whose PMCs are orphaned.
4527 		 *
4528 		 * Log PMC value at exit time if requested.
4529 		 */
4530 		for (ri = 0; ri < md->pmd_npmc; ri++)
4531 			if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) {
4532 				if (pm->pm_flags & PMC_F_NEEDS_LOGFILE &&
4533 				    PMC_IS_COUNTING_MODE(PMC_TO_MODE(pm)))
4534 					pmclog_process_procexit(pm, pp);
4535 				pmc_unlink_target_process(pm, pp);
4536 			}
4537 		free(pp, M_PMC);
4538 
4539 	} else
4540 		critical_exit(); /* pp == NULL */
4541 
4542 
4543 	/*
4544 	 * If the process owned PMCs, free them up and free up
4545 	 * memory.
4546 	 */
4547 	if ((po = pmc_find_owner_descriptor(p)) != NULL) {
4548 		pmc_remove_owner(po);
4549 		pmc_destroy_owner_descriptor(po);
4550 	}
4551 
4552 	sx_xunlock(&pmc_sx);
4553 }
4554 
4555 /*
4556  * Handle a process fork.
4557  *
4558  * If the parent process 'p1' is under HWPMC monitoring, then copy
4559  * over any attached PMCs that have 'do_descendants' semantics.
4560  */
4561 
4562 static void
4563 pmc_process_fork(void *arg __unused, struct proc *p1, struct proc *newproc,
4564     int flags)
4565 {
4566 	int is_using_hwpmcs;
4567 	unsigned int ri;
4568 	uint32_t do_descendants;
4569 	struct pmc *pm;
4570 	struct pmc_owner *po;
4571 	struct pmc_process *ppnew, *ppold;
4572 
4573 	(void) flags;		/* unused parameter */
4574 
4575 	PROC_LOCK(p1);
4576 	is_using_hwpmcs = p1->p_flag & P_HWPMC;
4577 	PROC_UNLOCK(p1);
4578 
4579 	/*
4580 	 * If there are system-wide sampling PMCs active, we need to
4581 	 * log all fork events to their owner's logs.
4582 	 */
4583 
4584 	LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
4585 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
4586 		    pmclog_process_procfork(po, p1->p_pid, newproc->p_pid);
4587 
4588 	if (!is_using_hwpmcs)
4589 		return;
4590 
4591 	PMC_GET_SX_XLOCK();
4592 	PMCDBG(PMC,FRK,1, "process-fork proc=%p (%d, %s) -> %p", p1,
4593 	    p1->p_pid, p1->p_comm, newproc);
4594 
4595 	/*
4596 	 * If the parent process (curthread->td_proc) is a
4597 	 * target of any PMCs, look for PMCs that are to be
4598 	 * inherited, and link these into the new process
4599 	 * descriptor.
4600 	 */
4601 	if ((ppold = pmc_find_process_descriptor(curthread->td_proc,
4602 		 PMC_FLAG_NONE)) == NULL)
4603 		goto done;		/* nothing to do */
4604 
4605 	do_descendants = 0;
4606 	for (ri = 0; ri < md->pmd_npmc; ri++)
4607 		if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL)
4608 			do_descendants |= pm->pm_flags & PMC_F_DESCENDANTS;
4609 	if (do_descendants == 0) /* nothing to do */
4610 		goto done;
4611 
4612 	/* allocate a descriptor for the new process  */
4613 	if ((ppnew = pmc_find_process_descriptor(newproc,
4614 		 PMC_FLAG_ALLOCATE)) == NULL)
4615 		goto done;
4616 
4617 	/*
4618 	 * Run through all PMCs that were targeting the old process
4619 	 * and which specified F_DESCENDANTS and attach them to the
4620 	 * new process.
4621 	 *
4622 	 * Log the fork event to all owners of PMCs attached to this
4623 	 * process, if not already logged.
4624 	 */
4625 	for (ri = 0; ri < md->pmd_npmc; ri++)
4626 		if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL &&
4627 		    (pm->pm_flags & PMC_F_DESCENDANTS)) {
4628 			pmc_link_target_process(pm, ppnew);
4629 			po = pm->pm_owner;
4630 			if (po->po_sscount == 0 &&
4631 			    po->po_flags & PMC_PO_OWNS_LOGFILE)
4632 				pmclog_process_procfork(po, p1->p_pid,
4633 				    newproc->p_pid);
4634 		}
4635 
4636 	/*
4637 	 * Now mark the new process as being tracked by this driver.
4638 	 */
4639 	PROC_LOCK(newproc);
4640 	newproc->p_flag |= P_HWPMC;
4641 	PROC_UNLOCK(newproc);
4642 
4643  done:
4644 	sx_xunlock(&pmc_sx);
4645 }
4646 
4647 
4648 /*
4649  * initialization
4650  */
4651 
4652 static const char *pmc_name_of_pmcclass[] = {
4653 #undef	__PMC_CLASS
4654 #define	__PMC_CLASS(N) #N ,
4655 	__PMC_CLASSES()
4656 };
4657 
4658 /*
4659  * Base class initializer: allocate structure and set default classes.
4660  */
4661 struct pmc_mdep *
4662 pmc_mdep_alloc(int nclasses)
4663 {
4664 	struct pmc_mdep *md;
4665 	int	n;
4666 
4667 	/* SOFT + md classes */
4668 	n = 1 + nclasses;
4669 	md = malloc(sizeof(struct pmc_mdep) + n *
4670 	    sizeof(struct pmc_classdep), M_PMC, M_WAITOK|M_ZERO);
4671 	md->pmd_nclass = n;
4672 
4673 	/* Add base class. */
4674 	pmc_soft_initialize(md);
4675 	return md;
4676 }
4677 
4678 void
4679 pmc_mdep_free(struct pmc_mdep *md)
4680 {
4681 	pmc_soft_finalize(md);
4682 	free(md, M_PMC);
4683 }
4684 
4685 static int
4686 generic_switch_in(struct pmc_cpu *pc, struct pmc_process *pp)
4687 {
4688 	(void) pc; (void) pp;
4689 
4690 	return (0);
4691 }
4692 
4693 static int
4694 generic_switch_out(struct pmc_cpu *pc, struct pmc_process *pp)
4695 {
4696 	(void) pc; (void) pp;
4697 
4698 	return (0);
4699 }
4700 
4701 static struct pmc_mdep *
4702 pmc_generic_cpu_initialize(void)
4703 {
4704 	struct pmc_mdep *md;
4705 
4706 	md = pmc_mdep_alloc(0);
4707 
4708 	md->pmd_cputype    = PMC_CPU_GENERIC;
4709 
4710 	md->pmd_pcpu_init  = NULL;
4711 	md->pmd_pcpu_fini  = NULL;
4712 	md->pmd_switch_in  = generic_switch_in;
4713 	md->pmd_switch_out = generic_switch_out;
4714 
4715 	return (md);
4716 }
4717 
4718 static void
4719 pmc_generic_cpu_finalize(struct pmc_mdep *md)
4720 {
4721 	(void) md;
4722 }
4723 
4724 
4725 static int
4726 pmc_initialize(void)
4727 {
4728 	int c, cpu, error, n, ri;
4729 	unsigned int maxcpu;
4730 	struct pmc_binding pb;
4731 	struct pmc_sample *ps;
4732 	struct pmc_classdep *pcd;
4733 	struct pmc_samplebuffer *sb;
4734 
4735 	md = NULL;
4736 	error = 0;
4737 
4738 #ifdef	DEBUG
4739 	/* parse debug flags first */
4740 	if (TUNABLE_STR_FETCH(PMC_SYSCTL_NAME_PREFIX "debugflags",
4741 		pmc_debugstr, sizeof(pmc_debugstr)))
4742 		pmc_debugflags_parse(pmc_debugstr,
4743 		    pmc_debugstr+strlen(pmc_debugstr));
4744 #endif
4745 
4746 	PMCDBG(MOD,INI,0, "PMC Initialize (version %x)", PMC_VERSION);
4747 
4748 	/* check kernel version */
4749 	if (pmc_kernel_version != PMC_VERSION) {
4750 		if (pmc_kernel_version == 0)
4751 			printf("hwpmc: this kernel has not been compiled with "
4752 			    "'options HWPMC_HOOKS'.\n");
4753 		else
4754 			printf("hwpmc: kernel version (0x%x) does not match "
4755 			    "module version (0x%x).\n", pmc_kernel_version,
4756 			    PMC_VERSION);
4757 		return EPROGMISMATCH;
4758 	}
4759 
4760 	/*
4761 	 * check sysctl parameters
4762 	 */
4763 
4764 	if (pmc_hashsize <= 0) {
4765 		(void) printf("hwpmc: tunable \"hashsize\"=%d must be "
4766 		    "greater than zero.\n", pmc_hashsize);
4767 		pmc_hashsize = PMC_HASH_SIZE;
4768 	}
4769 
4770 	if (pmc_nsamples <= 0 || pmc_nsamples > 65535) {
4771 		(void) printf("hwpmc: tunable \"nsamples\"=%d out of "
4772 		    "range.\n", pmc_nsamples);
4773 		pmc_nsamples = PMC_NSAMPLES;
4774 	}
4775 
4776 	if (pmc_callchaindepth <= 0 ||
4777 	    pmc_callchaindepth > PMC_CALLCHAIN_DEPTH_MAX) {
4778 		(void) printf("hwpmc: tunable \"callchaindepth\"=%d out of "
4779 		    "range.\n", pmc_callchaindepth);
4780 		pmc_callchaindepth = PMC_CALLCHAIN_DEPTH;
4781 	}
4782 
4783 	md = pmc_md_initialize();
4784 	if (md == NULL) {
4785 		/* Default to generic CPU. */
4786 		md = pmc_generic_cpu_initialize();
4787 		if (md == NULL)
4788 			return (ENOSYS);
4789         }
4790 
4791 	KASSERT(md->pmd_nclass >= 1 && md->pmd_npmc >= 1,
4792 	    ("[pmc,%d] no classes or pmcs", __LINE__));
4793 
4794 	/* Compute the map from row-indices to classdep pointers. */
4795 	pmc_rowindex_to_classdep = malloc(sizeof(struct pmc_classdep *) *
4796 	    md->pmd_npmc, M_PMC, M_WAITOK|M_ZERO);
4797 
4798 	for (n = 0; n < md->pmd_npmc; n++)
4799 		pmc_rowindex_to_classdep[n] = NULL;
4800 	for (ri = c = 0; c < md->pmd_nclass; c++) {
4801 		pcd = &md->pmd_classdep[c];
4802 		for (n = 0; n < pcd->pcd_num; n++, ri++)
4803 			pmc_rowindex_to_classdep[ri] = pcd;
4804 	}
4805 
4806 	KASSERT(ri == md->pmd_npmc,
4807 	    ("[pmc,%d] npmc miscomputed: ri=%d, md->npmc=%d", __LINE__,
4808 	    ri, md->pmd_npmc));
4809 
4810 	maxcpu = pmc_cpu_max();
4811 
4812 	/* allocate space for the per-cpu array */
4813 	pmc_pcpu = malloc(maxcpu * sizeof(struct pmc_cpu *), M_PMC,
4814 	    M_WAITOK|M_ZERO);
4815 
4816 	/* per-cpu 'saved values' for managing process-mode PMCs */
4817 	pmc_pcpu_saved = malloc(sizeof(pmc_value_t) * maxcpu * md->pmd_npmc,
4818 	    M_PMC, M_WAITOK);
4819 
4820 	/* Perform CPU-dependent initialization. */
4821 	pmc_save_cpu_binding(&pb);
4822 	error = 0;
4823 	for (cpu = 0; error == 0 && cpu < maxcpu; cpu++) {
4824 		if (!pmc_cpu_is_active(cpu))
4825 			continue;
4826 		pmc_select_cpu(cpu);
4827 		pmc_pcpu[cpu] = malloc(sizeof(struct pmc_cpu) +
4828 		    md->pmd_npmc * sizeof(struct pmc_hw *), M_PMC,
4829 		    M_WAITOK|M_ZERO);
4830 		if (md->pmd_pcpu_init)
4831 			error = md->pmd_pcpu_init(md, cpu);
4832 		for (n = 0; error == 0 && n < md->pmd_nclass; n++)
4833 			error = md->pmd_classdep[n].pcd_pcpu_init(md, cpu);
4834 	}
4835 	pmc_restore_cpu_binding(&pb);
4836 
4837 	if (error)
4838 		return (error);
4839 
4840 	/* allocate space for the sample array */
4841 	for (cpu = 0; cpu < maxcpu; cpu++) {
4842 		if (!pmc_cpu_is_active(cpu))
4843 			continue;
4844 
4845 		sb = malloc(sizeof(struct pmc_samplebuffer) +
4846 		    pmc_nsamples * sizeof(struct pmc_sample), M_PMC,
4847 		    M_WAITOK|M_ZERO);
4848 		sb->ps_read = sb->ps_write = sb->ps_samples;
4849 		sb->ps_fence = sb->ps_samples + pmc_nsamples;
4850 
4851 		KASSERT(pmc_pcpu[cpu] != NULL,
4852 		    ("[pmc,%d] cpu=%d Null per-cpu data", __LINE__, cpu));
4853 
4854 		sb->ps_callchains = malloc(pmc_callchaindepth * pmc_nsamples *
4855 		    sizeof(uintptr_t), M_PMC, M_WAITOK|M_ZERO);
4856 
4857 		for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++)
4858 			ps->ps_pc = sb->ps_callchains +
4859 			    (n * pmc_callchaindepth);
4860 
4861 		pmc_pcpu[cpu]->pc_sb[PMC_HR] = sb;
4862 
4863 		sb = malloc(sizeof(struct pmc_samplebuffer) +
4864 		    pmc_nsamples * sizeof(struct pmc_sample), M_PMC,
4865 		    M_WAITOK|M_ZERO);
4866 		sb->ps_read = sb->ps_write = sb->ps_samples;
4867 		sb->ps_fence = sb->ps_samples + pmc_nsamples;
4868 
4869 		KASSERT(pmc_pcpu[cpu] != NULL,
4870 		    ("[pmc,%d] cpu=%d Null per-cpu data", __LINE__, cpu));
4871 
4872 		sb->ps_callchains = malloc(pmc_callchaindepth * pmc_nsamples *
4873 		    sizeof(uintptr_t), M_PMC, M_WAITOK|M_ZERO);
4874 
4875 		for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++)
4876 			ps->ps_pc = sb->ps_callchains +
4877 			    (n * pmc_callchaindepth);
4878 
4879 		pmc_pcpu[cpu]->pc_sb[PMC_SR] = sb;
4880 	}
4881 
4882 	/* allocate space for the row disposition array */
4883 	pmc_pmcdisp = malloc(sizeof(enum pmc_mode) * md->pmd_npmc,
4884 	    M_PMC, M_WAITOK|M_ZERO);
4885 
4886 	/* mark all PMCs as available */
4887 	for (n = 0; n < (int) md->pmd_npmc; n++)
4888 		PMC_MARK_ROW_FREE(n);
4889 
4890 	/* allocate thread hash tables */
4891 	pmc_ownerhash = hashinit(pmc_hashsize, M_PMC,
4892 	    &pmc_ownerhashmask);
4893 
4894 	pmc_processhash = hashinit(pmc_hashsize, M_PMC,
4895 	    &pmc_processhashmask);
4896 	mtx_init(&pmc_processhash_mtx, "pmc-process-hash", "pmc-leaf",
4897 	    MTX_SPIN);
4898 
4899 	LIST_INIT(&pmc_ss_owners);
4900 	pmc_ss_count = 0;
4901 
4902 	/* allocate a pool of spin mutexes */
4903 	pmc_mtxpool = mtx_pool_create("pmc-leaf", pmc_mtxpool_size,
4904 	    MTX_SPIN);
4905 
4906 	PMCDBG(MOD,INI,1, "pmc_ownerhash=%p, mask=0x%lx "
4907 	    "targethash=%p mask=0x%lx", pmc_ownerhash, pmc_ownerhashmask,
4908 	    pmc_processhash, pmc_processhashmask);
4909 
4910 	/* register process {exit,fork,exec} handlers */
4911 	pmc_exit_tag = EVENTHANDLER_REGISTER(process_exit,
4912 	    pmc_process_exit, NULL, EVENTHANDLER_PRI_ANY);
4913 	pmc_fork_tag = EVENTHANDLER_REGISTER(process_fork,
4914 	    pmc_process_fork, NULL, EVENTHANDLER_PRI_ANY);
4915 
4916 	/* initialize logging */
4917 	pmclog_initialize();
4918 
4919 	/* set hook functions */
4920 	pmc_intr = md->pmd_intr;
4921 	pmc_hook = pmc_hook_handler;
4922 
4923 	if (error == 0) {
4924 		printf(PMC_MODULE_NAME ":");
4925 		for (n = 0; n < (int) md->pmd_nclass; n++) {
4926 			pcd = &md->pmd_classdep[n];
4927 			printf(" %s/%d/%d/0x%b",
4928 			    pmc_name_of_pmcclass[pcd->pcd_class],
4929 			    pcd->pcd_num,
4930 			    pcd->pcd_width,
4931 			    pcd->pcd_caps,
4932 			    "\20"
4933 			    "\1INT\2USR\3SYS\4EDG\5THR"
4934 			    "\6REA\7WRI\10INV\11QUA\12PRC"
4935 			    "\13TAG\14CSC");
4936 		}
4937 		printf("\n");
4938 	}
4939 
4940 	return (error);
4941 }
4942 
4943 /* prepare to be unloaded */
4944 static void
4945 pmc_cleanup(void)
4946 {
4947 	int c, cpu;
4948 	unsigned int maxcpu;
4949 	struct pmc_ownerhash *ph;
4950 	struct pmc_owner *po, *tmp;
4951 	struct pmc_binding pb;
4952 #ifdef	DEBUG
4953 	struct pmc_processhash *prh;
4954 #endif
4955 
4956 	PMCDBG(MOD,INI,0, "%s", "cleanup");
4957 
4958 	/* switch off sampling */
4959 	CPU_ZERO(&pmc_cpumask);
4960 	pmc_intr = NULL;
4961 
4962 	sx_xlock(&pmc_sx);
4963 	if (pmc_hook == NULL) {	/* being unloaded already */
4964 		sx_xunlock(&pmc_sx);
4965 		return;
4966 	}
4967 
4968 	pmc_hook = NULL; /* prevent new threads from entering module */
4969 
4970 	/* deregister event handlers */
4971 	EVENTHANDLER_DEREGISTER(process_fork, pmc_fork_tag);
4972 	EVENTHANDLER_DEREGISTER(process_exit, pmc_exit_tag);
4973 
4974 	/* send SIGBUS to all owner threads, free up allocations */
4975 	if (pmc_ownerhash)
4976 		for (ph = pmc_ownerhash;
4977 		     ph <= &pmc_ownerhash[pmc_ownerhashmask];
4978 		     ph++) {
4979 			LIST_FOREACH_SAFE(po, ph, po_next, tmp) {
4980 				pmc_remove_owner(po);
4981 
4982 				/* send SIGBUS to owner processes */
4983 				PMCDBG(MOD,INI,2, "cleanup signal proc=%p "
4984 				    "(%d, %s)", po->po_owner,
4985 				    po->po_owner->p_pid,
4986 				    po->po_owner->p_comm);
4987 
4988 				PROC_LOCK(po->po_owner);
4989 				kern_psignal(po->po_owner, SIGBUS);
4990 				PROC_UNLOCK(po->po_owner);
4991 
4992 				pmc_destroy_owner_descriptor(po);
4993 			}
4994 		}
4995 
4996 	/* reclaim allocated data structures */
4997 	if (pmc_mtxpool)
4998 		mtx_pool_destroy(&pmc_mtxpool);
4999 
5000 	mtx_destroy(&pmc_processhash_mtx);
5001 	if (pmc_processhash) {
5002 #ifdef	DEBUG
5003 		struct pmc_process *pp;
5004 
5005 		PMCDBG(MOD,INI,3, "%s", "destroy process hash");
5006 		for (prh = pmc_processhash;
5007 		     prh <= &pmc_processhash[pmc_processhashmask];
5008 		     prh++)
5009 			LIST_FOREACH(pp, prh, pp_next)
5010 			    PMCDBG(MOD,INI,3, "pid=%d", pp->pp_proc->p_pid);
5011 #endif
5012 
5013 		hashdestroy(pmc_processhash, M_PMC, pmc_processhashmask);
5014 		pmc_processhash = NULL;
5015 	}
5016 
5017 	if (pmc_ownerhash) {
5018 		PMCDBG(MOD,INI,3, "%s", "destroy owner hash");
5019 		hashdestroy(pmc_ownerhash, M_PMC, pmc_ownerhashmask);
5020 		pmc_ownerhash = NULL;
5021 	}
5022 
5023 	KASSERT(LIST_EMPTY(&pmc_ss_owners),
5024 	    ("[pmc,%d] Global SS owner list not empty", __LINE__));
5025 	KASSERT(pmc_ss_count == 0,
5026 	    ("[pmc,%d] Global SS count not empty", __LINE__));
5027 
5028  	/* do processor and pmc-class dependent cleanup */
5029 	maxcpu = pmc_cpu_max();
5030 
5031 	PMCDBG(MOD,INI,3, "%s", "md cleanup");
5032 	if (md) {
5033 		pmc_save_cpu_binding(&pb);
5034 		for (cpu = 0; cpu < maxcpu; cpu++) {
5035 			PMCDBG(MOD,INI,1,"pmc-cleanup cpu=%d pcs=%p",
5036 			    cpu, pmc_pcpu[cpu]);
5037 			if (!pmc_cpu_is_active(cpu) || pmc_pcpu[cpu] == NULL)
5038 				continue;
5039 			pmc_select_cpu(cpu);
5040 			for (c = 0; c < md->pmd_nclass; c++)
5041 				md->pmd_classdep[c].pcd_pcpu_fini(md, cpu);
5042 			if (md->pmd_pcpu_fini)
5043 				md->pmd_pcpu_fini(md, cpu);
5044 		}
5045 
5046 		if (md->pmd_cputype == PMC_CPU_GENERIC)
5047 			pmc_generic_cpu_finalize(md);
5048 		else
5049 			pmc_md_finalize(md);
5050 
5051 		pmc_mdep_free(md);
5052 		md = NULL;
5053 		pmc_restore_cpu_binding(&pb);
5054 	}
5055 
5056 	/* Free per-cpu descriptors. */
5057 	for (cpu = 0; cpu < maxcpu; cpu++) {
5058 		if (!pmc_cpu_is_active(cpu))
5059 			continue;
5060 		KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_HR] != NULL,
5061 		    ("[pmc,%d] Null hw cpu sample buffer cpu=%d", __LINE__,
5062 			cpu));
5063 		KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_SR] != NULL,
5064 		    ("[pmc,%d] Null sw cpu sample buffer cpu=%d", __LINE__,
5065 			cpu));
5066 		free(pmc_pcpu[cpu]->pc_sb[PMC_HR]->ps_callchains, M_PMC);
5067 		free(pmc_pcpu[cpu]->pc_sb[PMC_HR], M_PMC);
5068 		free(pmc_pcpu[cpu]->pc_sb[PMC_SR]->ps_callchains, M_PMC);
5069 		free(pmc_pcpu[cpu]->pc_sb[PMC_SR], M_PMC);
5070 		free(pmc_pcpu[cpu], M_PMC);
5071 	}
5072 
5073 	free(pmc_pcpu, M_PMC);
5074 	pmc_pcpu = NULL;
5075 
5076 	free(pmc_pcpu_saved, M_PMC);
5077 	pmc_pcpu_saved = NULL;
5078 
5079 	if (pmc_pmcdisp) {
5080 		free(pmc_pmcdisp, M_PMC);
5081 		pmc_pmcdisp = NULL;
5082 	}
5083 
5084 	if (pmc_rowindex_to_classdep) {
5085 		free(pmc_rowindex_to_classdep, M_PMC);
5086 		pmc_rowindex_to_classdep = NULL;
5087 	}
5088 
5089 	pmclog_shutdown();
5090 
5091 	sx_xunlock(&pmc_sx); 	/* we are done */
5092 }
5093 
5094 /*
5095  * The function called at load/unload.
5096  */
5097 
5098 static int
5099 load (struct module *module __unused, int cmd, void *arg __unused)
5100 {
5101 	int error;
5102 
5103 	error = 0;
5104 
5105 	switch (cmd) {
5106 	case MOD_LOAD :
5107 		/* initialize the subsystem */
5108 		error = pmc_initialize();
5109 		if (error != 0)
5110 			break;
5111 		PMCDBG(MOD,INI,1, "syscall=%d maxcpu=%d",
5112 		    pmc_syscall_num, pmc_cpu_max());
5113 		break;
5114 
5115 
5116 	case MOD_UNLOAD :
5117 	case MOD_SHUTDOWN:
5118 		pmc_cleanup();
5119 		PMCDBG(MOD,INI,1, "%s", "unloaded");
5120 		break;
5121 
5122 	default :
5123 		error = EINVAL;	/* XXX should panic(9) */
5124 		break;
5125 	}
5126 
5127 	return error;
5128 }
5129 
5130 /* memory pool */
5131 MALLOC_DEFINE(M_PMC, "pmc", "Memory space for the PMC module");
5132