xref: /freebsd/sys/dev/hwpmc/hwpmc_mod.c (revision fb47a3769c514735bceb2822a64e5e70c3d2f7a4)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2003-2008 Joseph Koshy
5  * Copyright (c) 2007 The FreeBSD Foundation
6  * Copyright (c) 2018 Matthew Macy
7  * All rights reserved.
8  *
9  * Portions of this software were developed by A. Joseph Koshy under
10  * sponsorship from the FreeBSD Foundation and Google, Inc.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  *
33  */
34 
35 #include <sys/cdefs.h>
36 __FBSDID("$FreeBSD$");
37 
38 #include <sys/param.h>
39 #include <sys/eventhandler.h>
40 #include <sys/gtaskqueue.h>
41 #include <sys/jail.h>
42 #include <sys/kernel.h>
43 #include <sys/kthread.h>
44 #include <sys/limits.h>
45 #include <sys/lock.h>
46 #include <sys/malloc.h>
47 #include <sys/module.h>
48 #include <sys/mount.h>
49 #include <sys/mutex.h>
50 #include <sys/pmc.h>
51 #include <sys/pmckern.h>
52 #include <sys/pmclog.h>
53 #include <sys/priv.h>
54 #include <sys/proc.h>
55 #include <sys/queue.h>
56 #include <sys/resourcevar.h>
57 #include <sys/rwlock.h>
58 #include <sys/sched.h>
59 #include <sys/signalvar.h>
60 #include <sys/smp.h>
61 #include <sys/sx.h>
62 #include <sys/sysctl.h>
63 #include <sys/sysent.h>
64 #include <sys/systm.h>
65 #include <sys/vnode.h>
66 
67 #include <sys/linker.h>		/* needs to be after <sys/malloc.h> */
68 
69 #include <machine/atomic.h>
70 #include <machine/md_var.h>
71 
72 #include <vm/vm.h>
73 #include <vm/vm_extern.h>
74 #include <vm/pmap.h>
75 #include <vm/vm_map.h>
76 #include <vm/vm_object.h>
77 
78 #include "hwpmc_soft.h"
79 
80 #ifdef NUMA
81 #define NDOMAINS vm_ndomains
82 #else
83 #define NDOMAINS 1
84 #define malloc_domain(size, type, domain, flags) malloc((size), (type), (flags))
85 #define free_domain(addr, type) free(addr, type)
86 #endif
87 
88 /*
89  * Types
90  */
91 
92 enum pmc_flags {
93 	PMC_FLAG_NONE	  = 0x00, /* do nothing */
94 	PMC_FLAG_REMOVE   = 0x01, /* atomically remove entry from hash */
95 	PMC_FLAG_ALLOCATE = 0x02, /* add entry to hash if not found */
96 	PMC_FLAG_NOWAIT   = 0x04, /* do not wait for mallocs */
97 };
98 
99 /*
100  * The offset in sysent where the syscall is allocated.
101  */
102 
103 static int pmc_syscall_num = NO_SYSCALL;
104 struct pmc_cpu		**pmc_pcpu;	 /* per-cpu state */
105 pmc_value_t		*pmc_pcpu_saved; /* saved PMC values: CSW handling */
106 
107 #define	PMC_PCPU_SAVED(C,R)	pmc_pcpu_saved[(R) + md->pmd_npmc*(C)]
108 
109 struct mtx_pool		*pmc_mtxpool;
110 static int		*pmc_pmcdisp;	 /* PMC row dispositions */
111 
112 #define	PMC_ROW_DISP_IS_FREE(R)		(pmc_pmcdisp[(R)] == 0)
113 #define	PMC_ROW_DISP_IS_THREAD(R)	(pmc_pmcdisp[(R)] > 0)
114 #define	PMC_ROW_DISP_IS_STANDALONE(R)	(pmc_pmcdisp[(R)] < 0)
115 
116 #define	PMC_MARK_ROW_FREE(R) do {					  \
117 	pmc_pmcdisp[(R)] = 0;						  \
118 } while (0)
119 
120 #define	PMC_MARK_ROW_STANDALONE(R) do {					  \
121 	KASSERT(pmc_pmcdisp[(R)] <= 0, ("[pmc,%d] row disposition error", \
122 		    __LINE__));						  \
123 	atomic_add_int(&pmc_pmcdisp[(R)], -1);				  \
124 	KASSERT(pmc_pmcdisp[(R)] >= (-pmc_cpu_max_active()),		  \
125 		("[pmc,%d] row disposition error", __LINE__));		  \
126 } while (0)
127 
128 #define	PMC_UNMARK_ROW_STANDALONE(R) do { 				  \
129 	atomic_add_int(&pmc_pmcdisp[(R)], 1);				  \
130 	KASSERT(pmc_pmcdisp[(R)] <= 0, ("[pmc,%d] row disposition error", \
131 		    __LINE__));						  \
132 } while (0)
133 
134 #define	PMC_MARK_ROW_THREAD(R) do {					  \
135 	KASSERT(pmc_pmcdisp[(R)] >= 0, ("[pmc,%d] row disposition error", \
136 		    __LINE__));						  \
137 	atomic_add_int(&pmc_pmcdisp[(R)], 1);				  \
138 } while (0)
139 
140 #define	PMC_UNMARK_ROW_THREAD(R) do {					  \
141 	atomic_add_int(&pmc_pmcdisp[(R)], -1);				  \
142 	KASSERT(pmc_pmcdisp[(R)] >= 0, ("[pmc,%d] row disposition error", \
143 		    __LINE__));						  \
144 } while (0)
145 
146 
147 /* various event handlers */
148 static eventhandler_tag	pmc_exit_tag, pmc_fork_tag, pmc_kld_load_tag,
149     pmc_kld_unload_tag;
150 
151 /* Module statistics */
152 struct pmc_driverstats pmc_stats;
153 
154 
155 /* Machine/processor dependent operations */
156 static struct pmc_mdep  *md;
157 
158 /*
159  * Hash tables mapping owner processes and target threads to PMCs.
160  */
161 
162 struct mtx pmc_processhash_mtx;		/* spin mutex */
163 static u_long pmc_processhashmask;
164 static LIST_HEAD(pmc_processhash, pmc_process)	*pmc_processhash;
165 
166 /*
167  * Hash table of PMC owner descriptors.  This table is protected by
168  * the shared PMC "sx" lock.
169  */
170 
171 static u_long pmc_ownerhashmask;
172 static LIST_HEAD(pmc_ownerhash, pmc_owner)	*pmc_ownerhash;
173 
174 /*
175  * List of PMC owners with system-wide sampling PMCs.
176  */
177 
178 static CK_LIST_HEAD(, pmc_owner)			pmc_ss_owners;
179 
180 /*
181  * List of free thread entries. This is protected by the spin
182  * mutex.
183  */
184 static struct mtx pmc_threadfreelist_mtx;	/* spin mutex */
185 static LIST_HEAD(, pmc_thread)			pmc_threadfreelist;
186 static int pmc_threadfreelist_entries=0;
187 #define	THREADENTRY_SIZE						\
188 (sizeof(struct pmc_thread) + (md->pmd_npmc * sizeof(struct pmc_threadpmcstate)))
189 
190 /*
191  * Task to free thread descriptors
192  */
193 static struct grouptask free_gtask;
194 
195 /*
196  * A map of row indices to classdep structures.
197  */
198 static struct pmc_classdep **pmc_rowindex_to_classdep;
199 
200 /*
201  * Prototypes
202  */
203 
204 #ifdef	HWPMC_DEBUG
205 static int	pmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS);
206 static int	pmc_debugflags_parse(char *newstr, char *fence);
207 #endif
208 
209 static int	load(struct module *module, int cmd, void *arg);
210 static int	pmc_add_sample(int ring, struct pmc *pm, struct trapframe *tf);
211 static void	pmc_add_thread_descriptors_from_proc(struct proc *p,
212     struct pmc_process *pp);
213 static int	pmc_attach_process(struct proc *p, struct pmc *pm);
214 static struct pmc *pmc_allocate_pmc_descriptor(void);
215 static struct pmc_owner *pmc_allocate_owner_descriptor(struct proc *p);
216 static int	pmc_attach_one_process(struct proc *p, struct pmc *pm);
217 static int	pmc_can_allocate_rowindex(struct proc *p, unsigned int ri,
218     int cpu);
219 static int	pmc_can_attach(struct pmc *pm, struct proc *p);
220 static void	pmc_capture_user_callchain(int cpu, int soft, struct trapframe *tf);
221 static void	pmc_cleanup(void);
222 static int	pmc_detach_process(struct proc *p, struct pmc *pm);
223 static int	pmc_detach_one_process(struct proc *p, struct pmc *pm,
224     int flags);
225 static void	pmc_destroy_owner_descriptor(struct pmc_owner *po);
226 static void	pmc_destroy_pmc_descriptor(struct pmc *pm);
227 static void	pmc_destroy_process_descriptor(struct pmc_process *pp);
228 static struct pmc_owner *pmc_find_owner_descriptor(struct proc *p);
229 static int	pmc_find_pmc(pmc_id_t pmcid, struct pmc **pm);
230 static struct pmc *pmc_find_pmc_descriptor_in_process(struct pmc_owner *po,
231     pmc_id_t pmc);
232 static struct pmc_process *pmc_find_process_descriptor(struct proc *p,
233     uint32_t mode);
234 static struct pmc_thread *pmc_find_thread_descriptor(struct pmc_process *pp,
235     struct thread *td, uint32_t mode);
236 static void	pmc_force_context_switch(void);
237 static void	pmc_link_target_process(struct pmc *pm,
238     struct pmc_process *pp);
239 static void	pmc_log_all_process_mappings(struct pmc_owner *po);
240 static void	pmc_log_kernel_mappings(struct pmc *pm);
241 static void	pmc_log_process_mappings(struct pmc_owner *po, struct proc *p);
242 static void	pmc_maybe_remove_owner(struct pmc_owner *po);
243 static void	pmc_process_csw_in(struct thread *td);
244 static void	pmc_process_csw_out(struct thread *td);
245 static void	pmc_process_exit(void *arg, struct proc *p);
246 static void	pmc_process_fork(void *arg, struct proc *p1,
247     struct proc *p2, int n);
248 static void	pmc_process_samples(int cpu, int soft);
249 static void	pmc_release_pmc_descriptor(struct pmc *pmc);
250 static void	pmc_process_thread_add(struct thread *td);
251 static void	pmc_process_thread_delete(struct thread *td);
252 static void	pmc_process_thread_userret(struct thread *td);
253 static void	pmc_remove_owner(struct pmc_owner *po);
254 static void	pmc_remove_process_descriptor(struct pmc_process *pp);
255 static void	pmc_restore_cpu_binding(struct pmc_binding *pb);
256 static void	pmc_save_cpu_binding(struct pmc_binding *pb);
257 static void	pmc_select_cpu(int cpu);
258 static int	pmc_start(struct pmc *pm);
259 static int	pmc_stop(struct pmc *pm);
260 static int	pmc_syscall_handler(struct thread *td, void *syscall_args);
261 static struct pmc_thread *pmc_thread_descriptor_pool_alloc(void);
262 static void	pmc_thread_descriptor_pool_drain(void);
263 static void	pmc_thread_descriptor_pool_free(struct pmc_thread *pt);
264 static void	pmc_unlink_target_process(struct pmc *pmc,
265     struct pmc_process *pp);
266 static int generic_switch_in(struct pmc_cpu *pc, struct pmc_process *pp);
267 static int generic_switch_out(struct pmc_cpu *pc, struct pmc_process *pp);
268 static struct pmc_mdep *pmc_generic_cpu_initialize(void);
269 static void pmc_generic_cpu_finalize(struct pmc_mdep *md);
270 static void pmc_post_callchain_callback(void);
271 static void pmc_process_threadcreate(struct thread *td);
272 static void pmc_process_threadexit(struct thread *td);
273 static void pmc_process_proccreate(struct proc *p);
274 static void pmc_process_allproc(struct pmc *pm);
275 
276 /*
277  * Kernel tunables and sysctl(8) interface.
278  */
279 
280 SYSCTL_DECL(_kern_hwpmc);
281 SYSCTL_NODE(_kern_hwpmc, OID_AUTO, stats, CTLFLAG_RW, 0, "HWPMC stats");
282 
283 
284 /* Stats. */
285 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, intr_ignored, CTLFLAG_RW,
286 				   &pmc_stats.pm_intr_ignored, "# of interrupts ignored");
287 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, intr_processed, CTLFLAG_RW,
288 				   &pmc_stats.pm_intr_processed, "# of interrupts processed");
289 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, intr_bufferfull, CTLFLAG_RW,
290 				   &pmc_stats.pm_intr_bufferfull, "# of interrupts where buffer was full");
291 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, syscalls, CTLFLAG_RW,
292 				   &pmc_stats.pm_syscalls, "# of syscalls");
293 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, syscall_errors, CTLFLAG_RW,
294 				   &pmc_stats.pm_syscall_errors, "# of syscall_errors");
295 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, buffer_requests, CTLFLAG_RW,
296 				   &pmc_stats.pm_buffer_requests, "# of buffer requests");
297 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, buffer_requests_failed, CTLFLAG_RW,
298 				   &pmc_stats.pm_buffer_requests_failed, "# of buffer requests which failed");
299 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, log_sweeps, CTLFLAG_RW,
300 				   &pmc_stats.pm_log_sweeps, "# of ?");
301 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, merges, CTLFLAG_RW,
302 				   &pmc_stats.pm_merges, "# of times kernel stack was found for user trace");
303 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, overwrites, CTLFLAG_RW,
304 				   &pmc_stats.pm_overwrites, "# of times a sample was overwritten before being logged");
305 
306 static int pmc_callchaindepth = PMC_CALLCHAIN_DEPTH;
307 SYSCTL_INT(_kern_hwpmc, OID_AUTO, callchaindepth, CTLFLAG_RDTUN,
308     &pmc_callchaindepth, 0, "depth of call chain records");
309 
310 char pmc_cpuid[64];
311 SYSCTL_STRING(_kern_hwpmc, OID_AUTO, cpuid, CTLFLAG_RD,
312 	pmc_cpuid, 0, "cpu version string");
313 #ifdef	HWPMC_DEBUG
314 struct pmc_debugflags pmc_debugflags = PMC_DEBUG_DEFAULT_FLAGS;
315 char	pmc_debugstr[PMC_DEBUG_STRSIZE];
316 TUNABLE_STR(PMC_SYSCTL_NAME_PREFIX "debugflags", pmc_debugstr,
317     sizeof(pmc_debugstr));
318 SYSCTL_PROC(_kern_hwpmc, OID_AUTO, debugflags,
319     CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NOFETCH,
320     0, 0, pmc_debugflags_sysctl_handler, "A", "debug flags");
321 #endif
322 
323 
324 /*
325  * kern.hwpmc.hashrows -- determines the number of rows in the
326  * of the hash table used to look up threads
327  */
328 
329 static int pmc_hashsize = PMC_HASH_SIZE;
330 SYSCTL_INT(_kern_hwpmc, OID_AUTO, hashsize, CTLFLAG_RDTUN,
331     &pmc_hashsize, 0, "rows in hash tables");
332 
333 /*
334  * kern.hwpmc.nsamples --- number of PC samples/callchain stacks per CPU
335  */
336 
337 static int pmc_nsamples = PMC_NSAMPLES;
338 SYSCTL_INT(_kern_hwpmc, OID_AUTO, nsamples, CTLFLAG_RDTUN,
339     &pmc_nsamples, 0, "number of PC samples per CPU");
340 
341 
342 /*
343  * kern.hwpmc.mtxpoolsize -- number of mutexes in the mutex pool.
344  */
345 
346 static int pmc_mtxpool_size = PMC_MTXPOOL_SIZE;
347 SYSCTL_INT(_kern_hwpmc, OID_AUTO, mtxpoolsize, CTLFLAG_RDTUN,
348     &pmc_mtxpool_size, 0, "size of spin mutex pool");
349 
350 
351 /*
352  * kern.hwpmc.threadfreelist_entries -- number of free entries
353  */
354 
355 SYSCTL_INT(_kern_hwpmc, OID_AUTO, threadfreelist_entries, CTLFLAG_RD,
356     &pmc_threadfreelist_entries, 0, "number of avalable thread entries");
357 
358 
359 /*
360  * kern.hwpmc.threadfreelist_max -- maximum number of free entries
361  */
362 
363 static int pmc_threadfreelist_max = PMC_THREADLIST_MAX;
364 SYSCTL_INT(_kern_hwpmc, OID_AUTO, threadfreelist_max, CTLFLAG_RW,
365     &pmc_threadfreelist_max, 0,
366     "maximum number of available thread entries before freeing some");
367 
368 
369 /*
370  * security.bsd.unprivileged_syspmcs -- allow non-root processes to
371  * allocate system-wide PMCs.
372  *
373  * Allowing unprivileged processes to allocate system PMCs is convenient
374  * if system-wide measurements need to be taken concurrently with other
375  * per-process measurements.  This feature is turned off by default.
376  */
377 
378 static int pmc_unprivileged_syspmcs = 0;
379 SYSCTL_INT(_security_bsd, OID_AUTO, unprivileged_syspmcs, CTLFLAG_RWTUN,
380     &pmc_unprivileged_syspmcs, 0,
381     "allow unprivileged process to allocate system PMCs");
382 
383 /*
384  * Hash function.  Discard the lower 2 bits of the pointer since
385  * these are always zero for our uses.  The hash multiplier is
386  * round((2^LONG_BIT) * ((sqrt(5)-1)/2)).
387  */
388 
389 #if	LONG_BIT == 64
390 #define	_PMC_HM		11400714819323198486u
391 #elif	LONG_BIT == 32
392 #define	_PMC_HM		2654435769u
393 #else
394 #error 	Must know the size of 'long' to compile
395 #endif
396 
397 #define	PMC_HASH_PTR(P,M)	((((unsigned long) (P) >> 2) * _PMC_HM) & (M))
398 
399 /*
400  * Syscall structures
401  */
402 
403 /* The `sysent' for the new syscall */
404 static struct sysent pmc_sysent = {
405 	.sy_narg =	2,
406 	.sy_call =	pmc_syscall_handler,
407 };
408 
409 static struct syscall_module_data pmc_syscall_mod = {
410 	.chainevh =	load,
411 	.chainarg =	NULL,
412 	.offset =	&pmc_syscall_num,
413 	.new_sysent =	&pmc_sysent,
414 	.old_sysent =	{ .sy_narg = 0, .sy_call = NULL },
415 	.flags =	SY_THR_STATIC_KLD,
416 };
417 
418 static moduledata_t pmc_mod = {
419 	.name =		PMC_MODULE_NAME,
420 	.evhand =	syscall_module_handler,
421 	.priv =		&pmc_syscall_mod,
422 };
423 
424 #ifdef EARLY_AP_STARTUP
425 DECLARE_MODULE(pmc, pmc_mod, SI_SUB_SYSCALLS, SI_ORDER_ANY);
426 #else
427 DECLARE_MODULE(pmc, pmc_mod, SI_SUB_SMP, SI_ORDER_ANY);
428 #endif
429 MODULE_VERSION(pmc, PMC_VERSION);
430 
431 #ifdef	HWPMC_DEBUG
432 enum pmc_dbgparse_state {
433 	PMCDS_WS,		/* in whitespace */
434 	PMCDS_MAJOR,		/* seen a major keyword */
435 	PMCDS_MINOR
436 };
437 
438 static int
439 pmc_debugflags_parse(char *newstr, char *fence)
440 {
441 	char c, *p, *q;
442 	struct pmc_debugflags *tmpflags;
443 	int error, found, *newbits, tmp;
444 	size_t kwlen;
445 
446 	tmpflags = malloc(sizeof(*tmpflags), M_PMC, M_WAITOK|M_ZERO);
447 
448 	p = newstr;
449 	error = 0;
450 
451 	for (; p < fence && (c = *p); p++) {
452 
453 		/* skip white space */
454 		if (c == ' ' || c == '\t')
455 			continue;
456 
457 		/* look for a keyword followed by "=" */
458 		for (q = p; p < fence && (c = *p) && c != '='; p++)
459 			;
460 		if (c != '=') {
461 			error = EINVAL;
462 			goto done;
463 		}
464 
465 		kwlen = p - q;
466 		newbits = NULL;
467 
468 		/* lookup flag group name */
469 #define	DBG_SET_FLAG_MAJ(S,F)						\
470 		if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0)	\
471 			newbits = &tmpflags->pdb_ ## F;
472 
473 		DBG_SET_FLAG_MAJ("cpu",		CPU);
474 		DBG_SET_FLAG_MAJ("csw",		CSW);
475 		DBG_SET_FLAG_MAJ("logging",	LOG);
476 		DBG_SET_FLAG_MAJ("module",	MOD);
477 		DBG_SET_FLAG_MAJ("md", 		MDP);
478 		DBG_SET_FLAG_MAJ("owner",	OWN);
479 		DBG_SET_FLAG_MAJ("pmc",		PMC);
480 		DBG_SET_FLAG_MAJ("process",	PRC);
481 		DBG_SET_FLAG_MAJ("sampling", 	SAM);
482 
483 		if (newbits == NULL) {
484 			error = EINVAL;
485 			goto done;
486 		}
487 
488 		p++;		/* skip the '=' */
489 
490 		/* Now parse the individual flags */
491 		tmp = 0;
492 	newflag:
493 		for (q = p; p < fence && (c = *p); p++)
494 			if (c == ' ' || c == '\t' || c == ',')
495 				break;
496 
497 		/* p == fence or c == ws or c == "," or c == 0 */
498 
499 		if ((kwlen = p - q) == 0) {
500 			*newbits = tmp;
501 			continue;
502 		}
503 
504 		found = 0;
505 #define	DBG_SET_FLAG_MIN(S,F)						\
506 		if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0)	\
507 			tmp |= found = (1 << PMC_DEBUG_MIN_ ## F)
508 
509 		/* a '*' denotes all possible flags in the group */
510 		if (kwlen == 1 && *q == '*')
511 			tmp = found = ~0;
512 		/* look for individual flag names */
513 		DBG_SET_FLAG_MIN("allocaterow", ALR);
514 		DBG_SET_FLAG_MIN("allocate",	ALL);
515 		DBG_SET_FLAG_MIN("attach",	ATT);
516 		DBG_SET_FLAG_MIN("bind",	BND);
517 		DBG_SET_FLAG_MIN("config",	CFG);
518 		DBG_SET_FLAG_MIN("exec",	EXC);
519 		DBG_SET_FLAG_MIN("exit",	EXT);
520 		DBG_SET_FLAG_MIN("find",	FND);
521 		DBG_SET_FLAG_MIN("flush",	FLS);
522 		DBG_SET_FLAG_MIN("fork",	FRK);
523 		DBG_SET_FLAG_MIN("getbuf",	GTB);
524 		DBG_SET_FLAG_MIN("hook",	PMH);
525 		DBG_SET_FLAG_MIN("init",	INI);
526 		DBG_SET_FLAG_MIN("intr",	INT);
527 		DBG_SET_FLAG_MIN("linktarget",	TLK);
528 		DBG_SET_FLAG_MIN("mayberemove", OMR);
529 		DBG_SET_FLAG_MIN("ops",		OPS);
530 		DBG_SET_FLAG_MIN("read",	REA);
531 		DBG_SET_FLAG_MIN("register",	REG);
532 		DBG_SET_FLAG_MIN("release",	REL);
533 		DBG_SET_FLAG_MIN("remove",	ORM);
534 		DBG_SET_FLAG_MIN("sample",	SAM);
535 		DBG_SET_FLAG_MIN("scheduleio",	SIO);
536 		DBG_SET_FLAG_MIN("select",	SEL);
537 		DBG_SET_FLAG_MIN("signal",	SIG);
538 		DBG_SET_FLAG_MIN("swi",		SWI);
539 		DBG_SET_FLAG_MIN("swo",		SWO);
540 		DBG_SET_FLAG_MIN("start",	STA);
541 		DBG_SET_FLAG_MIN("stop",	STO);
542 		DBG_SET_FLAG_MIN("syscall",	PMS);
543 		DBG_SET_FLAG_MIN("unlinktarget", TUL);
544 		DBG_SET_FLAG_MIN("write",	WRI);
545 		if (found == 0) {
546 			/* unrecognized flag name */
547 			error = EINVAL;
548 			goto done;
549 		}
550 
551 		if (c == 0 || c == ' ' || c == '\t') {	/* end of flag group */
552 			*newbits = tmp;
553 			continue;
554 		}
555 
556 		p++;
557 		goto newflag;
558 	}
559 
560 	/* save the new flag set */
561 	bcopy(tmpflags, &pmc_debugflags, sizeof(pmc_debugflags));
562 
563  done:
564 	free(tmpflags, M_PMC);
565 	return error;
566 }
567 
568 static int
569 pmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS)
570 {
571 	char *fence, *newstr;
572 	int error;
573 	unsigned int n;
574 
575 	(void) arg1; (void) arg2; /* unused parameters */
576 
577 	n = sizeof(pmc_debugstr);
578 	newstr = malloc(n, M_PMC, M_WAITOK|M_ZERO);
579 	(void) strlcpy(newstr, pmc_debugstr, n);
580 
581 	error = sysctl_handle_string(oidp, newstr, n, req);
582 
583 	/* if there is a new string, parse and copy it */
584 	if (error == 0 && req->newptr != NULL) {
585 		fence = newstr + (n < req->newlen ? n : req->newlen + 1);
586 		if ((error = pmc_debugflags_parse(newstr, fence)) == 0)
587 			(void) strlcpy(pmc_debugstr, newstr,
588 			    sizeof(pmc_debugstr));
589 	}
590 
591 	free(newstr, M_PMC);
592 
593 	return error;
594 }
595 #endif
596 
597 /*
598  * Map a row index to a classdep structure and return the adjusted row
599  * index for the PMC class index.
600  */
601 static struct pmc_classdep *
602 pmc_ri_to_classdep(struct pmc_mdep *md, int ri, int *adjri)
603 {
604 	struct pmc_classdep *pcd;
605 
606 	(void) md;
607 
608 	KASSERT(ri >= 0 && ri < md->pmd_npmc,
609 	    ("[pmc,%d] illegal row-index %d", __LINE__, ri));
610 
611 	pcd = pmc_rowindex_to_classdep[ri];
612 
613 	KASSERT(pcd != NULL,
614 	    ("[pmc,%d] ri %d null pcd", __LINE__, ri));
615 
616 	*adjri = ri - pcd->pcd_ri;
617 
618 	KASSERT(*adjri >= 0 && *adjri < pcd->pcd_num,
619 	    ("[pmc,%d] adjusted row-index %d", __LINE__, *adjri));
620 
621 	return (pcd);
622 }
623 
624 /*
625  * Concurrency Control
626  *
627  * The driver manages the following data structures:
628  *
629  *   - target process descriptors, one per target process
630  *   - owner process descriptors (and attached lists), one per owner process
631  *   - lookup hash tables for owner and target processes
632  *   - PMC descriptors (and attached lists)
633  *   - per-cpu hardware state
634  *   - the 'hook' variable through which the kernel calls into
635  *     this module
636  *   - the machine hardware state (managed by the MD layer)
637  *
638  * These data structures are accessed from:
639  *
640  * - thread context-switch code
641  * - interrupt handlers (possibly on multiple cpus)
642  * - kernel threads on multiple cpus running on behalf of user
643  *   processes doing system calls
644  * - this driver's private kernel threads
645  *
646  * = Locks and Locking strategy =
647  *
648  * The driver uses four locking strategies for its operation:
649  *
650  * - The global SX lock "pmc_sx" is used to protect internal
651  *   data structures.
652  *
653  *   Calls into the module by syscall() start with this lock being
654  *   held in exclusive mode.  Depending on the requested operation,
655  *   the lock may be downgraded to 'shared' mode to allow more
656  *   concurrent readers into the module.  Calls into the module from
657  *   other parts of the kernel acquire the lock in shared mode.
658  *
659  *   This SX lock is held in exclusive mode for any operations that
660  *   modify the linkages between the driver's internal data structures.
661  *
662  *   The 'pmc_hook' function pointer is also protected by this lock.
663  *   It is only examined with the sx lock held in exclusive mode.  The
664  *   kernel module is allowed to be unloaded only with the sx lock held
665  *   in exclusive mode.  In normal syscall handling, after acquiring the
666  *   pmc_sx lock we first check that 'pmc_hook' is non-null before
667  *   proceeding.  This prevents races between the thread unloading the module
668  *   and other threads seeking to use the module.
669  *
670  * - Lookups of target process structures and owner process structures
671  *   cannot use the global "pmc_sx" SX lock because these lookups need
672  *   to happen during context switches and in other critical sections
673  *   where sleeping is not allowed.  We protect these lookup tables
674  *   with their own private spin-mutexes, "pmc_processhash_mtx" and
675  *   "pmc_ownerhash_mtx".
676  *
677  * - Interrupt handlers work in a lock free manner.  At interrupt
678  *   time, handlers look at the PMC pointer (phw->phw_pmc) configured
679  *   when the PMC was started.  If this pointer is NULL, the interrupt
680  *   is ignored after updating driver statistics.  We ensure that this
681  *   pointer is set (using an atomic operation if necessary) before the
682  *   PMC hardware is started.  Conversely, this pointer is unset atomically
683  *   only after the PMC hardware is stopped.
684  *
685  *   We ensure that everything needed for the operation of an
686  *   interrupt handler is available without it needing to acquire any
687  *   locks.  We also ensure that a PMC's software state is destroyed only
688  *   after the PMC is taken off hardware (on all CPUs).
689  *
690  * - Context-switch handling with process-private PMCs needs more
691  *   care.
692  *
693  *   A given process may be the target of multiple PMCs.  For example,
694  *   PMCATTACH and PMCDETACH may be requested by a process on one CPU
695  *   while the target process is running on another.  A PMC could also
696  *   be getting released because its owner is exiting.  We tackle
697  *   these situations in the following manner:
698  *
699  *   - each target process structure 'pmc_process' has an array
700  *     of 'struct pmc *' pointers, one for each hardware PMC.
701  *
702  *   - At context switch IN time, each "target" PMC in RUNNING state
703  *     gets started on hardware and a pointer to each PMC is copied into
704  *     the per-cpu phw array.  The 'runcount' for the PMC is
705  *     incremented.
706  *
707  *   - At context switch OUT time, all process-virtual PMCs are stopped
708  *     on hardware.  The saved value is added to the PMCs value field
709  *     only if the PMC is in a non-deleted state (the PMCs state could
710  *     have changed during the current time slice).
711  *
712  *     Note that since in-between a switch IN on a processor and a switch
713  *     OUT, the PMC could have been released on another CPU.  Therefore
714  *     context switch OUT always looks at the hardware state to turn
715  *     OFF PMCs and will update a PMC's saved value only if reachable
716  *     from the target process record.
717  *
718  *   - OP PMCRELEASE could be called on a PMC at any time (the PMC could
719  *     be attached to many processes at the time of the call and could
720  *     be active on multiple CPUs).
721  *
722  *     We prevent further scheduling of the PMC by marking it as in
723  *     state 'DELETED'.  If the runcount of the PMC is non-zero then
724  *     this PMC is currently running on a CPU somewhere.  The thread
725  *     doing the PMCRELEASE operation waits by repeatedly doing a
726  *     pause() till the runcount comes to zero.
727  *
728  * The contents of a PMC descriptor (struct pmc) are protected using
729  * a spin-mutex.  In order to save space, we use a mutex pool.
730  *
731  * In terms of lock types used by witness(4), we use:
732  * - Type "pmc-sx", used by the global SX lock.
733  * - Type "pmc-sleep", for sleep mutexes used by logger threads.
734  * - Type "pmc-per-proc", for protecting PMC owner descriptors.
735  * - Type "pmc-leaf", used for all other spin mutexes.
736  */
737 
738 /*
739  * save the cpu binding of the current kthread
740  */
741 
742 static void
743 pmc_save_cpu_binding(struct pmc_binding *pb)
744 {
745 	PMCDBG0(CPU,BND,2, "save-cpu");
746 	thread_lock(curthread);
747 	pb->pb_bound = sched_is_bound(curthread);
748 	pb->pb_cpu   = curthread->td_oncpu;
749 	thread_unlock(curthread);
750 	PMCDBG1(CPU,BND,2, "save-cpu cpu=%d", pb->pb_cpu);
751 }
752 
753 /*
754  * restore the cpu binding of the current thread
755  */
756 
757 static void
758 pmc_restore_cpu_binding(struct pmc_binding *pb)
759 {
760 	PMCDBG2(CPU,BND,2, "restore-cpu curcpu=%d restore=%d",
761 	    curthread->td_oncpu, pb->pb_cpu);
762 	thread_lock(curthread);
763 	if (pb->pb_bound)
764 		sched_bind(curthread, pb->pb_cpu);
765 	else
766 		sched_unbind(curthread);
767 	thread_unlock(curthread);
768 	PMCDBG0(CPU,BND,2, "restore-cpu done");
769 }
770 
771 /*
772  * move execution over the specified cpu and bind it there.
773  */
774 
775 static void
776 pmc_select_cpu(int cpu)
777 {
778 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
779 	    ("[pmc,%d] bad cpu number %d", __LINE__, cpu));
780 
781 	/* Never move to an inactive CPU. */
782 	KASSERT(pmc_cpu_is_active(cpu), ("[pmc,%d] selecting inactive "
783 	    "CPU %d", __LINE__, cpu));
784 
785 	PMCDBG1(CPU,SEL,2, "select-cpu cpu=%d", cpu);
786 	thread_lock(curthread);
787 	sched_bind(curthread, cpu);
788 	thread_unlock(curthread);
789 
790 	KASSERT(curthread->td_oncpu == cpu,
791 	    ("[pmc,%d] CPU not bound [cpu=%d, curr=%d]", __LINE__,
792 		cpu, curthread->td_oncpu));
793 
794 	PMCDBG1(CPU,SEL,2, "select-cpu cpu=%d ok", cpu);
795 }
796 
797 /*
798  * Force a context switch.
799  *
800  * We do this by pause'ing for 1 tick -- invoking mi_switch() is not
801  * guaranteed to force a context switch.
802  */
803 
804 static void
805 pmc_force_context_switch(void)
806 {
807 
808 	pause("pmcctx", 1);
809 }
810 
811 uint64_t
812 pmc_rdtsc(void)
813 {
814 #if defined(__i386__) || defined(__amd64__)
815 	if (__predict_true(amd_feature & AMDID_RDTSCP))
816 		return rdtscp();
817 	else
818 		return rdtsc();
819 #else
820 	return get_cyclecount();
821 #endif
822 }
823 
824 /*
825  * Get the file name for an executable.  This is a simple wrapper
826  * around vn_fullpath(9).
827  */
828 
829 static void
830 pmc_getfilename(struct vnode *v, char **fullpath, char **freepath)
831 {
832 
833 	*fullpath = "unknown";
834 	*freepath = NULL;
835 	vn_fullpath(curthread, v, fullpath, freepath);
836 }
837 
838 /*
839  * remove an process owning PMCs
840  */
841 
842 void
843 pmc_remove_owner(struct pmc_owner *po)
844 {
845 	struct pmc *pm, *tmp;
846 
847 	sx_assert(&pmc_sx, SX_XLOCKED);
848 
849 	PMCDBG1(OWN,ORM,1, "remove-owner po=%p", po);
850 
851 	/* Remove descriptor from the owner hash table */
852 	LIST_REMOVE(po, po_next);
853 
854 	/* release all owned PMC descriptors */
855 	LIST_FOREACH_SAFE(pm, &po->po_pmcs, pm_next, tmp) {
856 		PMCDBG1(OWN,ORM,2, "pmc=%p", pm);
857 		KASSERT(pm->pm_owner == po,
858 		    ("[pmc,%d] owner %p != po %p", __LINE__, pm->pm_owner, po));
859 
860 		pmc_release_pmc_descriptor(pm);	/* will unlink from the list */
861 		pmc_destroy_pmc_descriptor(pm);
862 	}
863 
864 	KASSERT(po->po_sscount == 0,
865 	    ("[pmc,%d] SS count not zero", __LINE__));
866 	KASSERT(LIST_EMPTY(&po->po_pmcs),
867 	    ("[pmc,%d] PMC list not empty", __LINE__));
868 
869 	/* de-configure the log file if present */
870 	if (po->po_flags & PMC_PO_OWNS_LOGFILE)
871 		pmclog_deconfigure_log(po);
872 }
873 
874 /*
875  * remove an owner process record if all conditions are met.
876  */
877 
878 static void
879 pmc_maybe_remove_owner(struct pmc_owner *po)
880 {
881 
882 	PMCDBG1(OWN,OMR,1, "maybe-remove-owner po=%p", po);
883 
884 	/*
885 	 * Remove owner record if
886 	 * - this process does not own any PMCs
887 	 * - this process has not allocated a system-wide sampling buffer
888 	 */
889 
890 	if (LIST_EMPTY(&po->po_pmcs) &&
891 	    ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)) {
892 		pmc_remove_owner(po);
893 		pmc_destroy_owner_descriptor(po);
894 	}
895 }
896 
897 /*
898  * Add an association between a target process and a PMC.
899  */
900 
901 static void
902 pmc_link_target_process(struct pmc *pm, struct pmc_process *pp)
903 {
904 	int ri;
905 	struct pmc_target *pt;
906 #ifdef INVARIANTS
907 	struct pmc_thread *pt_td;
908 #endif
909 
910 	sx_assert(&pmc_sx, SX_XLOCKED);
911 
912 	KASSERT(pm != NULL && pp != NULL,
913 	    ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp));
914 	KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)),
915 	    ("[pmc,%d] Attaching a non-process-virtual pmc=%p to pid=%d",
916 		__LINE__, pm, pp->pp_proc->p_pid));
917 	KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= ((int) md->pmd_npmc - 1),
918 	    ("[pmc,%d] Illegal reference count %d for process record %p",
919 		__LINE__, pp->pp_refcnt, (void *) pp));
920 
921 	ri = PMC_TO_ROWINDEX(pm);
922 
923 	PMCDBG3(PRC,TLK,1, "link-target pmc=%p ri=%d pmc-process=%p",
924 	    pm, ri, pp);
925 
926 #ifdef	HWPMC_DEBUG
927 	LIST_FOREACH(pt, &pm->pm_targets, pt_next)
928 	    if (pt->pt_process == pp)
929 		    KASSERT(0, ("[pmc,%d] pp %p already in pmc %p targets",
930 				__LINE__, pp, pm));
931 #endif
932 
933 	pt = malloc(sizeof(struct pmc_target), M_PMC, M_WAITOK|M_ZERO);
934 	pt->pt_process = pp;
935 
936 	LIST_INSERT_HEAD(&pm->pm_targets, pt, pt_next);
937 
938 	atomic_store_rel_ptr((uintptr_t *)&pp->pp_pmcs[ri].pp_pmc,
939 	    (uintptr_t)pm);
940 
941 	if (pm->pm_owner->po_owner == pp->pp_proc)
942 		pm->pm_flags |= PMC_F_ATTACHED_TO_OWNER;
943 
944 	/*
945 	 * Initialize the per-process values at this row index.
946 	 */
947 	pp->pp_pmcs[ri].pp_pmcval = PMC_TO_MODE(pm) == PMC_MODE_TS ?
948 	    pm->pm_sc.pm_reloadcount : 0;
949 
950 	pp->pp_refcnt++;
951 
952 #ifdef INVARIANTS
953 	/* Confirm that the per-thread values at this row index are cleared. */
954 	if (PMC_TO_MODE(pm) == PMC_MODE_TS) {
955 		mtx_lock_spin(pp->pp_tdslock);
956 		LIST_FOREACH(pt_td, &pp->pp_tds, pt_next) {
957 			KASSERT(pt_td->pt_pmcs[ri].pt_pmcval == (pmc_value_t) 0,
958 			    ("[pmc,%d] pt_pmcval not cleared for pid=%d at "
959 			    "ri=%d", __LINE__, pp->pp_proc->p_pid, ri));
960 		}
961 		mtx_unlock_spin(pp->pp_tdslock);
962 	}
963 #endif
964 }
965 
966 /*
967  * Removes the association between a target process and a PMC.
968  */
969 
970 static void
971 pmc_unlink_target_process(struct pmc *pm, struct pmc_process *pp)
972 {
973 	int ri;
974 	struct proc *p;
975 	struct pmc_target *ptgt;
976 	struct pmc_thread *pt;
977 
978 	sx_assert(&pmc_sx, SX_XLOCKED);
979 
980 	KASSERT(pm != NULL && pp != NULL,
981 	    ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp));
982 
983 	KASSERT(pp->pp_refcnt >= 1 && pp->pp_refcnt <= (int) md->pmd_npmc,
984 	    ("[pmc,%d] Illegal ref count %d on process record %p",
985 		__LINE__, pp->pp_refcnt, (void *) pp));
986 
987 	ri = PMC_TO_ROWINDEX(pm);
988 
989 	PMCDBG3(PRC,TUL,1, "unlink-target pmc=%p ri=%d pmc-process=%p",
990 	    pm, ri, pp);
991 
992 	KASSERT(pp->pp_pmcs[ri].pp_pmc == pm,
993 	    ("[pmc,%d] PMC ri %d mismatch pmc %p pp->[ri] %p", __LINE__,
994 		ri, pm, pp->pp_pmcs[ri].pp_pmc));
995 
996 	pp->pp_pmcs[ri].pp_pmc = NULL;
997 	pp->pp_pmcs[ri].pp_pmcval = (pmc_value_t) 0;
998 
999 	/* Clear the per-thread values at this row index. */
1000 	if (PMC_TO_MODE(pm) == PMC_MODE_TS) {
1001 		mtx_lock_spin(pp->pp_tdslock);
1002 		LIST_FOREACH(pt, &pp->pp_tds, pt_next)
1003 			pt->pt_pmcs[ri].pt_pmcval = (pmc_value_t) 0;
1004 		mtx_unlock_spin(pp->pp_tdslock);
1005 	}
1006 
1007 	/* Remove owner-specific flags */
1008 	if (pm->pm_owner->po_owner == pp->pp_proc) {
1009 		pp->pp_flags &= ~PMC_PP_ENABLE_MSR_ACCESS;
1010 		pm->pm_flags &= ~PMC_F_ATTACHED_TO_OWNER;
1011 	}
1012 
1013 	pp->pp_refcnt--;
1014 
1015 	/* Remove the target process from the PMC structure */
1016 	LIST_FOREACH(ptgt, &pm->pm_targets, pt_next)
1017 		if (ptgt->pt_process == pp)
1018 			break;
1019 
1020 	KASSERT(ptgt != NULL, ("[pmc,%d] process %p (pp: %p) not found "
1021 		    "in pmc %p", __LINE__, pp->pp_proc, pp, pm));
1022 
1023 	LIST_REMOVE(ptgt, pt_next);
1024 	free(ptgt, M_PMC);
1025 
1026 	/* if the PMC now lacks targets, send the owner a SIGIO */
1027 	if (LIST_EMPTY(&pm->pm_targets)) {
1028 		p = pm->pm_owner->po_owner;
1029 		PROC_LOCK(p);
1030 		kern_psignal(p, SIGIO);
1031 		PROC_UNLOCK(p);
1032 
1033 		PMCDBG2(PRC,SIG,2, "signalling proc=%p signal=%d", p,
1034 		    SIGIO);
1035 	}
1036 }
1037 
1038 /*
1039  * Check if PMC 'pm' may be attached to target process 't'.
1040  */
1041 
1042 static int
1043 pmc_can_attach(struct pmc *pm, struct proc *t)
1044 {
1045 	struct proc *o;		/* pmc owner */
1046 	struct ucred *oc, *tc;	/* owner, target credentials */
1047 	int decline_attach, i;
1048 
1049 	/*
1050 	 * A PMC's owner can always attach that PMC to itself.
1051 	 */
1052 
1053 	if ((o = pm->pm_owner->po_owner) == t)
1054 		return 0;
1055 
1056 	PROC_LOCK(o);
1057 	oc = o->p_ucred;
1058 	crhold(oc);
1059 	PROC_UNLOCK(o);
1060 
1061 	PROC_LOCK(t);
1062 	tc = t->p_ucred;
1063 	crhold(tc);
1064 	PROC_UNLOCK(t);
1065 
1066 	/*
1067 	 * The effective uid of the PMC owner should match at least one
1068 	 * of the {effective,real,saved} uids of the target process.
1069 	 */
1070 
1071 	decline_attach = oc->cr_uid != tc->cr_uid &&
1072 	    oc->cr_uid != tc->cr_svuid &&
1073 	    oc->cr_uid != tc->cr_ruid;
1074 
1075 	/*
1076 	 * Every one of the target's group ids, must be in the owner's
1077 	 * group list.
1078 	 */
1079 	for (i = 0; !decline_attach && i < tc->cr_ngroups; i++)
1080 		decline_attach = !groupmember(tc->cr_groups[i], oc);
1081 
1082 	/* check the read and saved gids too */
1083 	if (decline_attach == 0)
1084 		decline_attach = !groupmember(tc->cr_rgid, oc) ||
1085 		    !groupmember(tc->cr_svgid, oc);
1086 
1087 	crfree(tc);
1088 	crfree(oc);
1089 
1090 	return !decline_attach;
1091 }
1092 
1093 /*
1094  * Attach a process to a PMC.
1095  */
1096 
1097 static int
1098 pmc_attach_one_process(struct proc *p, struct pmc *pm)
1099 {
1100 	int ri, error;
1101 	char *fullpath, *freepath;
1102 	struct pmc_process	*pp;
1103 
1104 	sx_assert(&pmc_sx, SX_XLOCKED);
1105 
1106 	PMCDBG5(PRC,ATT,2, "attach-one pm=%p ri=%d proc=%p (%d, %s)", pm,
1107 	    PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm);
1108 
1109 	/*
1110 	 * Locate the process descriptor corresponding to process 'p',
1111 	 * allocating space as needed.
1112 	 *
1113 	 * Verify that rowindex 'pm_rowindex' is free in the process
1114 	 * descriptor.
1115 	 *
1116 	 * If not, allocate space for a descriptor and link the
1117 	 * process descriptor and PMC.
1118 	 */
1119 	ri = PMC_TO_ROWINDEX(pm);
1120 
1121 	/* mark process as using HWPMCs */
1122 	PROC_LOCK(p);
1123 	p->p_flag |= P_HWPMC;
1124 	PROC_UNLOCK(p);
1125 
1126 	if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_ALLOCATE)) == NULL) {
1127 		error = ENOMEM;
1128 		goto fail;
1129 	}
1130 
1131 	if (pp->pp_pmcs[ri].pp_pmc == pm) {/* already present at slot [ri] */
1132 		error = EEXIST;
1133 		goto fail;
1134 	}
1135 
1136 	if (pp->pp_pmcs[ri].pp_pmc != NULL) {
1137 		error = EBUSY;
1138 		goto fail;
1139 	}
1140 
1141 	pmc_link_target_process(pm, pp);
1142 
1143 	if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) &&
1144 	    (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) == 0)
1145 		pm->pm_flags |= PMC_F_NEEDS_LOGFILE;
1146 
1147 	pm->pm_flags |= PMC_F_ATTACH_DONE; /* mark as attached */
1148 
1149 	/* issue an attach event to a configured log file */
1150 	if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE) {
1151 		if (p->p_flag & P_KPROC) {
1152 			fullpath = kernelname;
1153 			freepath = NULL;
1154 		} else {
1155 			pmc_getfilename(p->p_textvp, &fullpath, &freepath);
1156 			pmclog_process_pmcattach(pm, p->p_pid, fullpath);
1157 		}
1158 		free(freepath, M_TEMP);
1159 		if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1160 			pmc_log_process_mappings(pm->pm_owner, p);
1161 	}
1162 
1163 	return (0);
1164  fail:
1165 	PROC_LOCK(p);
1166 	p->p_flag &= ~P_HWPMC;
1167 	PROC_UNLOCK(p);
1168 	return (error);
1169 }
1170 
1171 /*
1172  * Attach a process and optionally its children
1173  */
1174 
1175 static int
1176 pmc_attach_process(struct proc *p, struct pmc *pm)
1177 {
1178 	int error;
1179 	struct proc *top;
1180 
1181 	sx_assert(&pmc_sx, SX_XLOCKED);
1182 
1183 	PMCDBG5(PRC,ATT,1, "attach pm=%p ri=%d proc=%p (%d, %s)", pm,
1184 	    PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm);
1185 
1186 
1187 	/*
1188 	 * If this PMC successfully allowed a GETMSR operation
1189 	 * in the past, disallow further ATTACHes.
1190 	 */
1191 
1192 	if ((pm->pm_flags & PMC_PP_ENABLE_MSR_ACCESS) != 0)
1193 		return EPERM;
1194 
1195 	if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0)
1196 		return pmc_attach_one_process(p, pm);
1197 
1198 	/*
1199 	 * Traverse all child processes, attaching them to
1200 	 * this PMC.
1201 	 */
1202 
1203 	sx_slock(&proctree_lock);
1204 
1205 	top = p;
1206 
1207 	for (;;) {
1208 		if ((error = pmc_attach_one_process(p, pm)) != 0)
1209 			break;
1210 		if (!LIST_EMPTY(&p->p_children))
1211 			p = LIST_FIRST(&p->p_children);
1212 		else for (;;) {
1213 			if (p == top)
1214 				goto done;
1215 			if (LIST_NEXT(p, p_sibling)) {
1216 				p = LIST_NEXT(p, p_sibling);
1217 				break;
1218 			}
1219 			p = p->p_pptr;
1220 		}
1221 	}
1222 
1223 	if (error)
1224 		(void) pmc_detach_process(top, pm);
1225 
1226  done:
1227 	sx_sunlock(&proctree_lock);
1228 	return error;
1229 }
1230 
1231 /*
1232  * Detach a process from a PMC.  If there are no other PMCs tracking
1233  * this process, remove the process structure from its hash table.  If
1234  * 'flags' contains PMC_FLAG_REMOVE, then free the process structure.
1235  */
1236 
1237 static int
1238 pmc_detach_one_process(struct proc *p, struct pmc *pm, int flags)
1239 {
1240 	int ri;
1241 	struct pmc_process *pp;
1242 
1243 	sx_assert(&pmc_sx, SX_XLOCKED);
1244 
1245 	KASSERT(pm != NULL,
1246 	    ("[pmc,%d] null pm pointer", __LINE__));
1247 
1248 	ri = PMC_TO_ROWINDEX(pm);
1249 
1250 	PMCDBG6(PRC,ATT,2, "detach-one pm=%p ri=%d proc=%p (%d, %s) flags=0x%x",
1251 	    pm, ri, p, p->p_pid, p->p_comm, flags);
1252 
1253 	if ((pp = pmc_find_process_descriptor(p, 0)) == NULL)
1254 		return ESRCH;
1255 
1256 	if (pp->pp_pmcs[ri].pp_pmc != pm)
1257 		return EINVAL;
1258 
1259 	pmc_unlink_target_process(pm, pp);
1260 
1261 	/* Issue a detach entry if a log file is configured */
1262 	if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE)
1263 		pmclog_process_pmcdetach(pm, p->p_pid);
1264 
1265 	/*
1266 	 * If there are no PMCs targeting this process, we remove its
1267 	 * descriptor from the target hash table and unset the P_HWPMC
1268 	 * flag in the struct proc.
1269 	 */
1270 	KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= (int) md->pmd_npmc,
1271 	    ("[pmc,%d] Illegal refcnt %d for process struct %p",
1272 		__LINE__, pp->pp_refcnt, pp));
1273 
1274 	if (pp->pp_refcnt != 0)	/* still a target of some PMC */
1275 		return 0;
1276 
1277 	pmc_remove_process_descriptor(pp);
1278 
1279 	if (flags & PMC_FLAG_REMOVE)
1280 		pmc_destroy_process_descriptor(pp);
1281 
1282 	PROC_LOCK(p);
1283 	p->p_flag &= ~P_HWPMC;
1284 	PROC_UNLOCK(p);
1285 
1286 	return 0;
1287 }
1288 
1289 /*
1290  * Detach a process and optionally its descendants from a PMC.
1291  */
1292 
1293 static int
1294 pmc_detach_process(struct proc *p, struct pmc *pm)
1295 {
1296 	struct proc *top;
1297 
1298 	sx_assert(&pmc_sx, SX_XLOCKED);
1299 
1300 	PMCDBG5(PRC,ATT,1, "detach pm=%p ri=%d proc=%p (%d, %s)", pm,
1301 	    PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm);
1302 
1303 	if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0)
1304 		return pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE);
1305 
1306 	/*
1307 	 * Traverse all children, detaching them from this PMC.  We
1308 	 * ignore errors since we could be detaching a PMC from a
1309 	 * partially attached proc tree.
1310 	 */
1311 
1312 	sx_slock(&proctree_lock);
1313 
1314 	top = p;
1315 
1316 	for (;;) {
1317 		(void) pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE);
1318 
1319 		if (!LIST_EMPTY(&p->p_children))
1320 			p = LIST_FIRST(&p->p_children);
1321 		else for (;;) {
1322 			if (p == top)
1323 				goto done;
1324 			if (LIST_NEXT(p, p_sibling)) {
1325 				p = LIST_NEXT(p, p_sibling);
1326 				break;
1327 			}
1328 			p = p->p_pptr;
1329 		}
1330 	}
1331 
1332  done:
1333 	sx_sunlock(&proctree_lock);
1334 
1335 	if (LIST_EMPTY(&pm->pm_targets))
1336 		pm->pm_flags &= ~PMC_F_ATTACH_DONE;
1337 
1338 	return 0;
1339 }
1340 
1341 
1342 /*
1343  * Thread context switch IN
1344  */
1345 
1346 static void
1347 pmc_process_csw_in(struct thread *td)
1348 {
1349 	int cpu;
1350 	unsigned int adjri, ri;
1351 	struct pmc *pm;
1352 	struct proc *p;
1353 	struct pmc_cpu *pc;
1354 	struct pmc_hw *phw;
1355 	pmc_value_t newvalue;
1356 	struct pmc_process *pp;
1357 	struct pmc_thread *pt;
1358 	struct pmc_classdep *pcd;
1359 
1360 	p = td->td_proc;
1361 	pt = NULL;
1362 	if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE)) == NULL)
1363 		return;
1364 
1365 	KASSERT(pp->pp_proc == td->td_proc,
1366 	    ("[pmc,%d] not my thread state", __LINE__));
1367 
1368 	critical_enter(); /* no preemption from this point */
1369 
1370 	cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */
1371 
1372 	PMCDBG5(CSW,SWI,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p,
1373 	    p->p_pid, p->p_comm, pp);
1374 
1375 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
1376 	    ("[pmc,%d] weird CPU id %d", __LINE__, cpu));
1377 
1378 	pc = pmc_pcpu[cpu];
1379 
1380 	for (ri = 0; ri < md->pmd_npmc; ri++) {
1381 
1382 		if ((pm = pp->pp_pmcs[ri].pp_pmc) == NULL)
1383 			continue;
1384 
1385 		KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)),
1386 		    ("[pmc,%d] Target PMC in non-virtual mode (%d)",
1387 			__LINE__, PMC_TO_MODE(pm)));
1388 
1389 		KASSERT(PMC_TO_ROWINDEX(pm) == ri,
1390 		    ("[pmc,%d] Row index mismatch pmc %d != ri %d",
1391 			__LINE__, PMC_TO_ROWINDEX(pm), ri));
1392 
1393 		/*
1394 		 * Only PMCs that are marked as 'RUNNING' need
1395 		 * be placed on hardware.
1396 		 */
1397 
1398 		if (pm->pm_state != PMC_STATE_RUNNING)
1399 			continue;
1400 
1401 		/* increment PMC runcount */
1402 		counter_u64_add(pm->pm_runcount, 1);
1403 
1404 		/* configure the HWPMC we are going to use. */
1405 		pcd = pmc_ri_to_classdep(md, ri, &adjri);
1406 		pcd->pcd_config_pmc(cpu, adjri, pm);
1407 
1408 		phw = pc->pc_hwpmcs[ri];
1409 
1410 		KASSERT(phw != NULL,
1411 		    ("[pmc,%d] null hw pointer", __LINE__));
1412 
1413 		KASSERT(phw->phw_pmc == pm,
1414 		    ("[pmc,%d] hw->pmc %p != pmc %p", __LINE__,
1415 			phw->phw_pmc, pm));
1416 
1417 		/*
1418 		 * Write out saved value and start the PMC.
1419 		 *
1420 		 * Sampling PMCs use a per-thread value, while
1421 		 * counting mode PMCs use a per-pmc value that is
1422 		 * inherited across descendants.
1423 		 */
1424 		if (PMC_TO_MODE(pm) == PMC_MODE_TS) {
1425 			if (pt == NULL)
1426 				pt = pmc_find_thread_descriptor(pp, td,
1427 				    PMC_FLAG_NONE);
1428 
1429 			KASSERT(pt != NULL,
1430 			    ("[pmc,%d] No thread found for td=%p", __LINE__,
1431 			    td));
1432 
1433 			mtx_pool_lock_spin(pmc_mtxpool, pm);
1434 
1435 			/*
1436 			 * If we have a thread descriptor, use the per-thread
1437 			 * counter in the descriptor. If not, we will use
1438 			 * a per-process counter.
1439 			 *
1440 			 * TODO: Remove the per-process "safety net" once
1441 			 * we have thoroughly tested that we don't hit the
1442 			 * above assert.
1443 			 */
1444 			if (pt != NULL) {
1445 				if (pt->pt_pmcs[ri].pt_pmcval > 0)
1446 					newvalue = pt->pt_pmcs[ri].pt_pmcval;
1447 				else
1448 					newvalue = pm->pm_sc.pm_reloadcount;
1449 			} else {
1450 				/*
1451 				 * Use the saved value calculated after the most
1452 				 * recent time a thread using the shared counter
1453 				 * switched out. Reset the saved count in case
1454 				 * another thread from this process switches in
1455 				 * before any threads switch out.
1456 				 */
1457 
1458 				newvalue = pp->pp_pmcs[ri].pp_pmcval;
1459 				pp->pp_pmcs[ri].pp_pmcval =
1460 				    pm->pm_sc.pm_reloadcount;
1461 			}
1462 			mtx_pool_unlock_spin(pmc_mtxpool, pm);
1463 			KASSERT(newvalue > 0 && newvalue <=
1464 			    pm->pm_sc.pm_reloadcount,
1465 			    ("[pmc,%d] pmcval outside of expected range cpu=%d "
1466 			    "ri=%d pmcval=%jx pm_reloadcount=%jx", __LINE__,
1467 			    cpu, ri, newvalue, pm->pm_sc.pm_reloadcount));
1468 		} else {
1469 			KASSERT(PMC_TO_MODE(pm) == PMC_MODE_TC,
1470 			    ("[pmc,%d] illegal mode=%d", __LINE__,
1471 			    PMC_TO_MODE(pm)));
1472 			mtx_pool_lock_spin(pmc_mtxpool, pm);
1473 			newvalue = PMC_PCPU_SAVED(cpu, ri) =
1474 			    pm->pm_gv.pm_savedvalue;
1475 			mtx_pool_unlock_spin(pmc_mtxpool, pm);
1476 		}
1477 
1478 		PMCDBG3(CSW,SWI,1,"cpu=%d ri=%d new=%jd", cpu, ri, newvalue);
1479 
1480 		pcd->pcd_write_pmc(cpu, adjri, newvalue);
1481 
1482 		/* If a sampling mode PMC, reset stalled state. */
1483 		if (PMC_TO_MODE(pm) == PMC_MODE_TS)
1484 			pm->pm_pcpu_state[cpu].pps_stalled = 0;
1485 
1486 		/* Indicate that we desire this to run. */
1487 		pm->pm_pcpu_state[cpu].pps_cpustate = 1;
1488 
1489 		/* Start the PMC. */
1490 		pcd->pcd_start_pmc(cpu, adjri);
1491 	}
1492 
1493 	/*
1494 	 * perform any other architecture/cpu dependent thread
1495 	 * switch-in actions.
1496 	 */
1497 
1498 	(void) (*md->pmd_switch_in)(pc, pp);
1499 
1500 	critical_exit();
1501 
1502 }
1503 
1504 /*
1505  * Thread context switch OUT.
1506  */
1507 
1508 static void
1509 pmc_process_csw_out(struct thread *td)
1510 {
1511 	int cpu;
1512 	int64_t tmp;
1513 	struct pmc *pm;
1514 	struct proc *p;
1515 	enum pmc_mode mode;
1516 	struct pmc_cpu *pc;
1517 	pmc_value_t newvalue;
1518 	unsigned int adjri, ri;
1519 	struct pmc_process *pp;
1520 	struct pmc_thread *pt = NULL;
1521 	struct pmc_classdep *pcd;
1522 
1523 
1524 	/*
1525 	 * Locate our process descriptor; this may be NULL if
1526 	 * this process is exiting and we have already removed
1527 	 * the process from the target process table.
1528 	 *
1529 	 * Note that due to kernel preemption, multiple
1530 	 * context switches may happen while the process is
1531 	 * exiting.
1532 	 *
1533 	 * Note also that if the target process cannot be
1534 	 * found we still need to deconfigure any PMCs that
1535 	 * are currently running on hardware.
1536 	 */
1537 
1538 	p = td->td_proc;
1539 	pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE);
1540 
1541 	/*
1542 	 * save PMCs
1543 	 */
1544 
1545 	critical_enter();
1546 
1547 	cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */
1548 
1549 	PMCDBG5(CSW,SWO,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p,
1550 	    p->p_pid, p->p_comm, pp);
1551 
1552 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
1553 	    ("[pmc,%d weird CPU id %d", __LINE__, cpu));
1554 
1555 	pc = pmc_pcpu[cpu];
1556 
1557 	/*
1558 	 * When a PMC gets unlinked from a target PMC, it will
1559 	 * be removed from the target's pp_pmc[] array.
1560 	 *
1561 	 * However, on a MP system, the target could have been
1562 	 * executing on another CPU at the time of the unlink.
1563 	 * So, at context switch OUT time, we need to look at
1564 	 * the hardware to determine if a PMC is scheduled on
1565 	 * it.
1566 	 */
1567 
1568 	for (ri = 0; ri < md->pmd_npmc; ri++) {
1569 
1570 		pcd = pmc_ri_to_classdep(md, ri, &adjri);
1571 		pm  = NULL;
1572 		(void) (*pcd->pcd_get_config)(cpu, adjri, &pm);
1573 
1574 		if (pm == NULL)	/* nothing at this row index */
1575 			continue;
1576 
1577 		mode = PMC_TO_MODE(pm);
1578 		if (!PMC_IS_VIRTUAL_MODE(mode))
1579 			continue; /* not a process virtual PMC */
1580 
1581 		KASSERT(PMC_TO_ROWINDEX(pm) == ri,
1582 		    ("[pmc,%d] ri mismatch pmc(%d) ri(%d)",
1583 			__LINE__, PMC_TO_ROWINDEX(pm), ri));
1584 
1585 		/*
1586 		 * Change desired state, and then stop if not stalled.
1587 		 * This two-step dance should avoid race conditions where
1588 		 * an interrupt re-enables the PMC after this code has
1589 		 * already checked the pm_stalled flag.
1590 		 */
1591 		pm->pm_pcpu_state[cpu].pps_cpustate = 0;
1592 		if (pm->pm_pcpu_state[cpu].pps_stalled == 0)
1593 			pcd->pcd_stop_pmc(cpu, adjri);
1594 
1595 		/* reduce this PMC's runcount */
1596 		counter_u64_add(pm->pm_runcount, -1);
1597 
1598 		/*
1599 		 * If this PMC is associated with this process,
1600 		 * save the reading.
1601 		 */
1602 
1603 		if (pm->pm_state != PMC_STATE_DELETED && pp != NULL &&
1604 		    pp->pp_pmcs[ri].pp_pmc != NULL) {
1605 			KASSERT(pm == pp->pp_pmcs[ri].pp_pmc,
1606 			    ("[pmc,%d] pm %p != pp_pmcs[%d] %p", __LINE__,
1607 				pm, ri, pp->pp_pmcs[ri].pp_pmc));
1608 
1609 			KASSERT(pp->pp_refcnt > 0,
1610 			    ("[pmc,%d] pp refcnt = %d", __LINE__,
1611 				pp->pp_refcnt));
1612 
1613 			pcd->pcd_read_pmc(cpu, adjri, &newvalue);
1614 
1615 			if (mode == PMC_MODE_TS) {
1616 				PMCDBG3(CSW,SWO,1,"cpu=%d ri=%d val=%jd (samp)",
1617 				    cpu, ri, newvalue);
1618 
1619 				if (pt == NULL)
1620 					pt = pmc_find_thread_descriptor(pp, td,
1621 					    PMC_FLAG_NONE);
1622 
1623 				KASSERT(pt != NULL,
1624 				    ("[pmc,%d] No thread found for td=%p",
1625 				    __LINE__, td));
1626 
1627 				mtx_pool_lock_spin(pmc_mtxpool, pm);
1628 
1629 				/*
1630 				 * If we have a thread descriptor, save the
1631 				 * per-thread counter in the descriptor. If not,
1632 				 * we will update the per-process counter.
1633 				 *
1634 				 * TODO: Remove the per-process "safety net"
1635 				 * once we have thoroughly tested that we
1636 				 * don't hit the above assert.
1637 				 */
1638 				if (pt != NULL)
1639 					pt->pt_pmcs[ri].pt_pmcval = newvalue;
1640 				else {
1641 					/*
1642 					 * For sampling process-virtual PMCs,
1643 					 * newvalue is the number of events to
1644 					 * be seen until the next sampling
1645 					 * interrupt. We can just add the events
1646 					 * left from this invocation to the
1647 					 * counter, then adjust in case we
1648 					 * overflow our range.
1649 					 *
1650 					 * (Recall that we reload the counter
1651 					 * every time we use it.)
1652 					 */
1653 					pp->pp_pmcs[ri].pp_pmcval += newvalue;
1654 					if (pp->pp_pmcs[ri].pp_pmcval >
1655 					    pm->pm_sc.pm_reloadcount)
1656 						pp->pp_pmcs[ri].pp_pmcval -=
1657 						    pm->pm_sc.pm_reloadcount;
1658 				}
1659 				mtx_pool_unlock_spin(pmc_mtxpool, pm);
1660 			} else {
1661 				tmp = newvalue - PMC_PCPU_SAVED(cpu,ri);
1662 
1663 				PMCDBG3(CSW,SWO,1,"cpu=%d ri=%d tmp=%jd (count)",
1664 				    cpu, ri, tmp);
1665 
1666 				/*
1667 				 * For counting process-virtual PMCs,
1668 				 * we expect the count to be
1669 				 * increasing monotonically, modulo a 64
1670 				 * bit wraparound.
1671 				 */
1672 				KASSERT(tmp >= 0,
1673 				    ("[pmc,%d] negative increment cpu=%d "
1674 				     "ri=%d newvalue=%jx saved=%jx "
1675 				     "incr=%jx", __LINE__, cpu, ri,
1676 				     newvalue, PMC_PCPU_SAVED(cpu,ri), tmp));
1677 
1678 				mtx_pool_lock_spin(pmc_mtxpool, pm);
1679 				pm->pm_gv.pm_savedvalue += tmp;
1680 				pp->pp_pmcs[ri].pp_pmcval += tmp;
1681 				mtx_pool_unlock_spin(pmc_mtxpool, pm);
1682 
1683 				if (pm->pm_flags & PMC_F_LOG_PROCCSW)
1684 					pmclog_process_proccsw(pm, pp, tmp, td);
1685 			}
1686 		}
1687 
1688 		/* mark hardware as free */
1689 		pcd->pcd_config_pmc(cpu, adjri, NULL);
1690 	}
1691 
1692 	/*
1693 	 * perform any other architecture/cpu dependent thread
1694 	 * switch out functions.
1695 	 */
1696 
1697 	(void) (*md->pmd_switch_out)(pc, pp);
1698 
1699 	critical_exit();
1700 }
1701 
1702 /*
1703  * A new thread for a process.
1704  */
1705 static void
1706 pmc_process_thread_add(struct thread *td)
1707 {
1708 	struct pmc_process *pmc;
1709 
1710 	pmc = pmc_find_process_descriptor(td->td_proc, PMC_FLAG_NONE);
1711 	if (pmc != NULL)
1712 		pmc_find_thread_descriptor(pmc, td, PMC_FLAG_ALLOCATE);
1713 }
1714 
1715 /*
1716  * A thread delete for a process.
1717  */
1718 static void
1719 pmc_process_thread_delete(struct thread *td)
1720 {
1721 	struct pmc_process *pmc;
1722 
1723 	pmc = pmc_find_process_descriptor(td->td_proc, PMC_FLAG_NONE);
1724 	if (pmc != NULL)
1725 		pmc_thread_descriptor_pool_free(pmc_find_thread_descriptor(pmc,
1726 		    td, PMC_FLAG_REMOVE));
1727 }
1728 
1729 /*
1730  * A userret() call for a thread.
1731  */
1732 static void
1733 pmc_process_thread_userret(struct thread *td)
1734 {
1735 	sched_pin();
1736 	pmc_capture_user_callchain(curcpu, PMC_UR, td->td_frame);
1737 	sched_unpin();
1738 }
1739 
1740 /*
1741  * A mapping change for a process.
1742  */
1743 
1744 static void
1745 pmc_process_mmap(struct thread *td, struct pmckern_map_in *pkm)
1746 {
1747 	int ri;
1748 	pid_t pid;
1749 	char *fullpath, *freepath;
1750 	const struct pmc *pm;
1751 	struct pmc_owner *po;
1752 	const struct pmc_process *pp;
1753 
1754 	freepath = fullpath = NULL;
1755 	MPASS(!in_epoch());
1756 	pmc_getfilename((struct vnode *) pkm->pm_file, &fullpath, &freepath);
1757 
1758 	pid = td->td_proc->p_pid;
1759 
1760 	epoch_enter_preempt(global_epoch_preempt);
1761 	/* Inform owners of all system-wide sampling PMCs. */
1762 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1763 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1764 			pmclog_process_map_in(po, pid, pkm->pm_address, fullpath);
1765 
1766 	if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL)
1767 		goto done;
1768 
1769 	/*
1770 	 * Inform sampling PMC owners tracking this process.
1771 	 */
1772 	for (ri = 0; ri < md->pmd_npmc; ri++)
1773 		if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL &&
1774 		    PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1775 			pmclog_process_map_in(pm->pm_owner,
1776 			    pid, pkm->pm_address, fullpath);
1777 
1778   done:
1779 	if (freepath)
1780 		free(freepath, M_TEMP);
1781 	epoch_exit_preempt(global_epoch_preempt);
1782 }
1783 
1784 
1785 /*
1786  * Log an munmap request.
1787  */
1788 
1789 static void
1790 pmc_process_munmap(struct thread *td, struct pmckern_map_out *pkm)
1791 {
1792 	int ri;
1793 	pid_t pid;
1794 	struct pmc_owner *po;
1795 	const struct pmc *pm;
1796 	const struct pmc_process *pp;
1797 
1798 	pid = td->td_proc->p_pid;
1799 
1800 	epoch_enter_preempt(global_epoch_preempt);
1801 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1802 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1803 		pmclog_process_map_out(po, pid, pkm->pm_address,
1804 		    pkm->pm_address + pkm->pm_size);
1805 	epoch_exit_preempt(global_epoch_preempt);
1806 
1807 	if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL)
1808 		return;
1809 
1810 	for (ri = 0; ri < md->pmd_npmc; ri++)
1811 		if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL &&
1812 		    PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1813 			pmclog_process_map_out(pm->pm_owner, pid,
1814 			    pkm->pm_address, pkm->pm_address + pkm->pm_size);
1815 }
1816 
1817 /*
1818  * Log mapping information about the kernel.
1819  */
1820 
1821 static void
1822 pmc_log_kernel_mappings(struct pmc *pm)
1823 {
1824 	struct pmc_owner *po;
1825 	struct pmckern_map_in *km, *kmbase;
1826 
1827 	MPASS(in_epoch() || sx_xlocked(&pmc_sx));
1828 	KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)),
1829 	    ("[pmc,%d] non-sampling PMC (%p) desires mapping information",
1830 		__LINE__, (void *) pm));
1831 
1832 	po = pm->pm_owner;
1833 
1834 	if (po->po_flags & PMC_PO_INITIAL_MAPPINGS_DONE)
1835 		return;
1836 	if (PMC_TO_MODE(pm) == PMC_MODE_SS)
1837 		pmc_process_allproc(pm);
1838 	/*
1839 	 * Log the current set of kernel modules.
1840 	 */
1841 	kmbase = linker_hwpmc_list_objects();
1842 	for (km = kmbase; km->pm_file != NULL; km++) {
1843 		PMCDBG2(LOG,REG,1,"%s %p", (char *) km->pm_file,
1844 		    (void *) km->pm_address);
1845 		pmclog_process_map_in(po, (pid_t) -1, km->pm_address,
1846 		    km->pm_file);
1847 	}
1848 	free(kmbase, M_LINKER);
1849 
1850 	po->po_flags |= PMC_PO_INITIAL_MAPPINGS_DONE;
1851 }
1852 
1853 /*
1854  * Log the mappings for a single process.
1855  */
1856 
1857 static void
1858 pmc_log_process_mappings(struct pmc_owner *po, struct proc *p)
1859 {
1860 	vm_map_t map;
1861 	struct vnode *vp;
1862 	struct vmspace *vm;
1863 	vm_map_entry_t entry;
1864 	vm_offset_t last_end;
1865 	u_int last_timestamp;
1866 	struct vnode *last_vp;
1867 	vm_offset_t start_addr;
1868 	vm_object_t obj, lobj, tobj;
1869 	char *fullpath, *freepath;
1870 
1871 	last_vp = NULL;
1872 	last_end = (vm_offset_t) 0;
1873 	fullpath = freepath = NULL;
1874 
1875 	if ((vm = vmspace_acquire_ref(p)) == NULL)
1876 		return;
1877 
1878 	map = &vm->vm_map;
1879 	vm_map_lock_read(map);
1880 
1881 	for (entry = map->header.next; entry != &map->header; entry = entry->next) {
1882 
1883 		if (entry == NULL) {
1884 			PMCDBG2(LOG,OPS,2, "hwpmc: vm_map entry unexpectedly "
1885 			    "NULL! pid=%d vm_map=%p\n", p->p_pid, map);
1886 			break;
1887 		}
1888 
1889 		/*
1890 		 * We only care about executable map entries.
1891 		 */
1892 		if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) ||
1893 		    !(entry->protection & VM_PROT_EXECUTE) ||
1894 		    (entry->object.vm_object == NULL)) {
1895 			continue;
1896 		}
1897 
1898 		obj = entry->object.vm_object;
1899 		VM_OBJECT_RLOCK(obj);
1900 
1901 		/*
1902 		 * Walk the backing_object list to find the base
1903 		 * (non-shadowed) vm_object.
1904 		 */
1905 		for (lobj = tobj = obj; tobj != NULL; tobj = tobj->backing_object) {
1906 			if (tobj != obj)
1907 				VM_OBJECT_RLOCK(tobj);
1908 			if (lobj != obj)
1909 				VM_OBJECT_RUNLOCK(lobj);
1910 			lobj = tobj;
1911 		}
1912 
1913 		/*
1914 		 * At this point lobj is the base vm_object and it is locked.
1915 		 */
1916 		if (lobj == NULL) {
1917 			PMCDBG3(LOG,OPS,2, "hwpmc: lobj unexpectedly NULL! pid=%d "
1918 			    "vm_map=%p vm_obj=%p\n", p->p_pid, map, obj);
1919 			VM_OBJECT_RUNLOCK(obj);
1920 			continue;
1921 		}
1922 
1923 		vp = vm_object_vnode(lobj);
1924 		if (vp == NULL) {
1925 			if (lobj != obj)
1926 				VM_OBJECT_RUNLOCK(lobj);
1927 			VM_OBJECT_RUNLOCK(obj);
1928 			continue;
1929 		}
1930 
1931 		/*
1932 		 * Skip contiguous regions that point to the same
1933 		 * vnode, so we don't emit redundant MAP-IN
1934 		 * directives.
1935 		 */
1936 		if (entry->start == last_end && vp == last_vp) {
1937 			last_end = entry->end;
1938 			if (lobj != obj)
1939 				VM_OBJECT_RUNLOCK(lobj);
1940 			VM_OBJECT_RUNLOCK(obj);
1941 			continue;
1942 		}
1943 
1944 		/*
1945 		 * We don't want to keep the proc's vm_map or this
1946 		 * vm_object locked while we walk the pathname, since
1947 		 * vn_fullpath() can sleep.  However, if we drop the
1948 		 * lock, it's possible for concurrent activity to
1949 		 * modify the vm_map list.  To protect against this,
1950 		 * we save the vm_map timestamp before we release the
1951 		 * lock, and check it after we reacquire the lock
1952 		 * below.
1953 		 */
1954 		start_addr = entry->start;
1955 		last_end = entry->end;
1956 		last_timestamp = map->timestamp;
1957 		vm_map_unlock_read(map);
1958 
1959 		vref(vp);
1960 		if (lobj != obj)
1961 			VM_OBJECT_RUNLOCK(lobj);
1962 
1963 		VM_OBJECT_RUNLOCK(obj);
1964 
1965 		freepath = NULL;
1966 		pmc_getfilename(vp, &fullpath, &freepath);
1967 		last_vp = vp;
1968 
1969 		vrele(vp);
1970 
1971 		vp = NULL;
1972 		pmclog_process_map_in(po, p->p_pid, start_addr, fullpath);
1973 		if (freepath)
1974 			free(freepath, M_TEMP);
1975 
1976 		vm_map_lock_read(map);
1977 
1978 		/*
1979 		 * If our saved timestamp doesn't match, this means
1980 		 * that the vm_map was modified out from under us and
1981 		 * we can't trust our current "entry" pointer.  Do a
1982 		 * new lookup for this entry.  If there is no entry
1983 		 * for this address range, vm_map_lookup_entry() will
1984 		 * return the previous one, so we always want to go to
1985 		 * entry->next on the next loop iteration.
1986 		 *
1987 		 * There is an edge condition here that can occur if
1988 		 * there is no entry at or before this address.  In
1989 		 * this situation, vm_map_lookup_entry returns
1990 		 * &map->header, which would cause our loop to abort
1991 		 * without processing the rest of the map.  However,
1992 		 * in practice this will never happen for process
1993 		 * vm_map.  This is because the executable's text
1994 		 * segment is the first mapping in the proc's address
1995 		 * space, and this mapping is never removed until the
1996 		 * process exits, so there will always be a non-header
1997 		 * entry at or before the requested address for
1998 		 * vm_map_lookup_entry to return.
1999 		 */
2000 		if (map->timestamp != last_timestamp)
2001 			vm_map_lookup_entry(map, last_end - 1, &entry);
2002 	}
2003 
2004 	vm_map_unlock_read(map);
2005 	vmspace_free(vm);
2006 	return;
2007 }
2008 
2009 /*
2010  * Log mappings for all processes in the system.
2011  */
2012 
2013 static void
2014 pmc_log_all_process_mappings(struct pmc_owner *po)
2015 {
2016 	struct proc *p, *top;
2017 
2018 	sx_assert(&pmc_sx, SX_XLOCKED);
2019 
2020 	if ((p = pfind(1)) == NULL)
2021 		panic("[pmc,%d] Cannot find init", __LINE__);
2022 
2023 	PROC_UNLOCK(p);
2024 
2025 	sx_slock(&proctree_lock);
2026 
2027 	top = p;
2028 
2029 	for (;;) {
2030 		pmc_log_process_mappings(po, p);
2031 		if (!LIST_EMPTY(&p->p_children))
2032 			p = LIST_FIRST(&p->p_children);
2033 		else for (;;) {
2034 			if (p == top)
2035 				goto done;
2036 			if (LIST_NEXT(p, p_sibling)) {
2037 				p = LIST_NEXT(p, p_sibling);
2038 				break;
2039 			}
2040 			p = p->p_pptr;
2041 		}
2042 	}
2043  done:
2044 	sx_sunlock(&proctree_lock);
2045 }
2046 
2047 /*
2048  * The 'hook' invoked from the kernel proper
2049  */
2050 
2051 
2052 #ifdef	HWPMC_DEBUG
2053 const char *pmc_hooknames[] = {
2054 	/* these strings correspond to PMC_FN_* in <sys/pmckern.h> */
2055 	"",
2056 	"EXEC",
2057 	"CSW-IN",
2058 	"CSW-OUT",
2059 	"SAMPLE",
2060 	"UNUSED1",
2061 	"UNUSED2",
2062 	"MMAP",
2063 	"MUNMAP",
2064 	"CALLCHAIN-NMI",
2065 	"CALLCHAIN-SOFT",
2066 	"SOFTSAMPLING",
2067 	"THR-CREATE",
2068 	"THR-EXIT",
2069 	"THR-USERRET",
2070 	"THR-CREATE-LOG",
2071 	"THR-EXIT-LOG",
2072 	"PROC-CREATE-LOG"
2073 };
2074 #endif
2075 
2076 static int
2077 pmc_hook_handler(struct thread *td, int function, void *arg)
2078 {
2079 	int cpu;
2080 
2081 	PMCDBG4(MOD,PMH,1, "hook td=%p func=%d \"%s\" arg=%p", td, function,
2082 	    pmc_hooknames[function], arg);
2083 
2084 	switch (function)
2085 	{
2086 
2087 	/*
2088 	 * Process exec()
2089 	 */
2090 
2091 	case PMC_FN_PROCESS_EXEC:
2092 	{
2093 		char *fullpath, *freepath;
2094 		unsigned int ri;
2095 		int is_using_hwpmcs;
2096 		struct pmc *pm;
2097 		struct proc *p;
2098 		struct pmc_owner *po;
2099 		struct pmc_process *pp;
2100 		struct pmckern_procexec *pk;
2101 
2102 		sx_assert(&pmc_sx, SX_XLOCKED);
2103 
2104 		p = td->td_proc;
2105 		pmc_getfilename(p->p_textvp, &fullpath, &freepath);
2106 
2107 		pk = (struct pmckern_procexec *) arg;
2108 
2109 		epoch_enter_preempt(global_epoch_preempt);
2110 		/* Inform owners of SS mode PMCs of the exec event. */
2111 		CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
2112 		    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
2113 			    pmclog_process_procexec(po, PMC_ID_INVALID,
2114 				p->p_pid, pk->pm_entryaddr, fullpath);
2115 		epoch_exit_preempt(global_epoch_preempt);
2116 
2117 		PROC_LOCK(p);
2118 		is_using_hwpmcs = p->p_flag & P_HWPMC;
2119 		PROC_UNLOCK(p);
2120 
2121 		if (!is_using_hwpmcs) {
2122 			if (freepath)
2123 				free(freepath, M_TEMP);
2124 			break;
2125 		}
2126 
2127 		/*
2128 		 * PMCs are not inherited across an exec():  remove any
2129 		 * PMCs that this process is the owner of.
2130 		 */
2131 
2132 		if ((po = pmc_find_owner_descriptor(p)) != NULL) {
2133 			pmc_remove_owner(po);
2134 			pmc_destroy_owner_descriptor(po);
2135 		}
2136 
2137 		/*
2138 		 * If the process being exec'ed is not the target of any
2139 		 * PMC, we are done.
2140 		 */
2141 		if ((pp = pmc_find_process_descriptor(p, 0)) == NULL) {
2142 			if (freepath)
2143 				free(freepath, M_TEMP);
2144 			break;
2145 		}
2146 
2147 		/*
2148 		 * Log the exec event to all monitoring owners.  Skip
2149 		 * owners who have already received the event because
2150 		 * they had system sampling PMCs active.
2151 		 */
2152 		for (ri = 0; ri < md->pmd_npmc; ri++)
2153 			if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) {
2154 				po = pm->pm_owner;
2155 				if (po->po_sscount == 0 &&
2156 				    po->po_flags & PMC_PO_OWNS_LOGFILE)
2157 					pmclog_process_procexec(po, pm->pm_id,
2158 					    p->p_pid, pk->pm_entryaddr,
2159 					    fullpath);
2160 			}
2161 
2162 		if (freepath)
2163 			free(freepath, M_TEMP);
2164 
2165 
2166 		PMCDBG4(PRC,EXC,1, "exec proc=%p (%d, %s) cred-changed=%d",
2167 		    p, p->p_pid, p->p_comm, pk->pm_credentialschanged);
2168 
2169 		if (pk->pm_credentialschanged == 0) /* no change */
2170 			break;
2171 
2172 		/*
2173 		 * If the newly exec()'ed process has a different credential
2174 		 * than before, allow it to be the target of a PMC only if
2175 		 * the PMC's owner has sufficient privilege.
2176 		 */
2177 
2178 		for (ri = 0; ri < md->pmd_npmc; ri++)
2179 			if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL)
2180 				if (pmc_can_attach(pm, td->td_proc) != 0)
2181 					pmc_detach_one_process(td->td_proc,
2182 					    pm, PMC_FLAG_NONE);
2183 
2184 		KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= (int) md->pmd_npmc,
2185 		    ("[pmc,%d] Illegal ref count %d on pp %p", __LINE__,
2186 			pp->pp_refcnt, pp));
2187 
2188 		/*
2189 		 * If this process is no longer the target of any
2190 		 * PMCs, we can remove the process entry and free
2191 		 * up space.
2192 		 */
2193 
2194 		if (pp->pp_refcnt == 0) {
2195 			pmc_remove_process_descriptor(pp);
2196 			pmc_destroy_process_descriptor(pp);
2197 			break;
2198 		}
2199 
2200 	}
2201 	break;
2202 
2203 	case PMC_FN_CSW_IN:
2204 		pmc_process_csw_in(td);
2205 		break;
2206 
2207 	case PMC_FN_CSW_OUT:
2208 		pmc_process_csw_out(td);
2209 		break;
2210 
2211 	/*
2212 	 * Process accumulated PC samples.
2213 	 *
2214 	 * This function is expected to be called by hardclock() for
2215 	 * each CPU that has accumulated PC samples.
2216 	 *
2217 	 * This function is to be executed on the CPU whose samples
2218 	 * are being processed.
2219 	 */
2220 	case PMC_FN_DO_SAMPLES:
2221 
2222 		/*
2223 		 * Clear the cpu specific bit in the CPU mask before
2224 		 * do the rest of the processing.  If the NMI handler
2225 		 * gets invoked after the "atomic_clear_int()" call
2226 		 * below but before "pmc_process_samples()" gets
2227 		 * around to processing the interrupt, then we will
2228 		 * come back here at the next hardclock() tick (and
2229 		 * may find nothing to do if "pmc_process_samples()"
2230 		 * had already processed the interrupt).  We don't
2231 		 * lose the interrupt sample.
2232 		 */
2233 		DPCPU_SET(pmc_sampled, 0);
2234 		cpu = PCPU_GET(cpuid);
2235 		pmc_process_samples(cpu, PMC_HR);
2236 		pmc_process_samples(cpu, PMC_SR);
2237 		pmc_process_samples(cpu, PMC_UR);
2238 		break;
2239 
2240 	case PMC_FN_MMAP:
2241 		pmc_process_mmap(td, (struct pmckern_map_in *) arg);
2242 		break;
2243 
2244 	case PMC_FN_MUNMAP:
2245 		MPASS(in_epoch() || sx_xlocked(&pmc_sx));
2246 		pmc_process_munmap(td, (struct pmckern_map_out *) arg);
2247 		break;
2248 
2249 	case PMC_FN_PROC_CREATE_LOG:
2250 		pmc_process_proccreate((struct proc *)arg);
2251 		break;
2252 
2253 	case PMC_FN_USER_CALLCHAIN:
2254 		/*
2255 		 * Record a call chain.
2256 		 */
2257 		KASSERT(td == curthread, ("[pmc,%d] td != curthread",
2258 		    __LINE__));
2259 
2260 		pmc_capture_user_callchain(PCPU_GET(cpuid), PMC_HR,
2261 		    (struct trapframe *) arg);
2262 
2263 		KASSERT(td->td_pinned == 1,
2264 			("[pmc,%d] invalid td_pinned value", __LINE__));
2265 		sched_unpin();  /* Can migrate safely now. */
2266 
2267 		td->td_pflags &= ~TDP_CALLCHAIN;
2268 		break;
2269 
2270 	case PMC_FN_USER_CALLCHAIN_SOFT:
2271 		/*
2272 		 * Record a call chain.
2273 		 */
2274 		KASSERT(td == curthread, ("[pmc,%d] td != curthread",
2275 		    __LINE__));
2276 
2277 		cpu = PCPU_GET(cpuid);
2278 		pmc_capture_user_callchain(cpu, PMC_SR,
2279 		    (struct trapframe *) arg);
2280 
2281 		KASSERT(td->td_pinned == 1,
2282 		    ("[pmc,%d] invalid td_pinned value", __LINE__));
2283 
2284 		sched_unpin();  /* Can migrate safely now. */
2285 
2286 		td->td_pflags &= ~TDP_CALLCHAIN;
2287 		break;
2288 
2289 	case PMC_FN_SOFT_SAMPLING:
2290 		/*
2291 		 * Call soft PMC sampling intr.
2292 		 */
2293 		pmc_soft_intr((struct pmckern_soft *) arg);
2294 		break;
2295 
2296 	case PMC_FN_THR_CREATE:
2297 		pmc_process_thread_add(td);
2298 		pmc_process_threadcreate(td);
2299 		break;
2300 
2301 	case PMC_FN_THR_CREATE_LOG:
2302 		pmc_process_threadcreate(td);
2303 		break;
2304 
2305 	case PMC_FN_THR_EXIT:
2306 		KASSERT(td == curthread, ("[pmc,%d] td != curthread",
2307 		    __LINE__));
2308 		pmc_process_thread_delete(td);
2309 		pmc_process_threadexit(td);
2310 		break;
2311 	case PMC_FN_THR_EXIT_LOG:
2312 		pmc_process_threadexit(td);
2313 		break;
2314 	case PMC_FN_THR_USERRET:
2315 		KASSERT(td == curthread, ("[pmc,%d] td != curthread",
2316 		    __LINE__));
2317 		pmc_process_thread_userret(td);
2318 		break;
2319 
2320 	default:
2321 #ifdef	HWPMC_DEBUG
2322 		KASSERT(0, ("[pmc,%d] unknown hook %d\n", __LINE__, function));
2323 #endif
2324 		break;
2325 
2326 	}
2327 
2328 	return 0;
2329 }
2330 
2331 /*
2332  * allocate a 'struct pmc_owner' descriptor in the owner hash table.
2333  */
2334 
2335 static struct pmc_owner *
2336 pmc_allocate_owner_descriptor(struct proc *p)
2337 {
2338 	uint32_t hindex;
2339 	struct pmc_owner *po;
2340 	struct pmc_ownerhash *poh;
2341 
2342 	hindex = PMC_HASH_PTR(p, pmc_ownerhashmask);
2343 	poh = &pmc_ownerhash[hindex];
2344 
2345 	/* allocate space for N pointers and one descriptor struct */
2346 	po = malloc(sizeof(struct pmc_owner), M_PMC, M_WAITOK|M_ZERO);
2347 	po->po_owner = p;
2348 	LIST_INSERT_HEAD(poh, po, po_next); /* insert into hash table */
2349 
2350 	TAILQ_INIT(&po->po_logbuffers);
2351 	mtx_init(&po->po_mtx, "pmc-owner-mtx", "pmc-per-proc", MTX_SPIN);
2352 
2353 	PMCDBG4(OWN,ALL,1, "allocate-owner proc=%p (%d, %s) pmc-owner=%p",
2354 	    p, p->p_pid, p->p_comm, po);
2355 
2356 	return po;
2357 }
2358 
2359 static void
2360 pmc_destroy_owner_descriptor(struct pmc_owner *po)
2361 {
2362 
2363 	PMCDBG4(OWN,REL,1, "destroy-owner po=%p proc=%p (%d, %s)",
2364 	    po, po->po_owner, po->po_owner->p_pid, po->po_owner->p_comm);
2365 
2366 	mtx_destroy(&po->po_mtx);
2367 	free(po, M_PMC);
2368 }
2369 
2370 /*
2371  * Allocate a thread descriptor from the free pool.
2372  *
2373  * NOTE: This *can* return NULL.
2374  */
2375 static struct pmc_thread *
2376 pmc_thread_descriptor_pool_alloc(void)
2377 {
2378 	struct pmc_thread *pt;
2379 
2380 	mtx_lock_spin(&pmc_threadfreelist_mtx);
2381 	if ((pt = LIST_FIRST(&pmc_threadfreelist)) != NULL) {
2382 		LIST_REMOVE(pt, pt_next);
2383 		pmc_threadfreelist_entries--;
2384 	}
2385 	mtx_unlock_spin(&pmc_threadfreelist_mtx);
2386 
2387 	return (pt);
2388 }
2389 
2390 /*
2391  * Add a thread descriptor to the free pool. We use this instead of free()
2392  * to maintain a cache of free entries. Additionally, we can safely call
2393  * this function when we cannot call free(), such as in a critical section.
2394  *
2395  */
2396 static void
2397 pmc_thread_descriptor_pool_free(struct pmc_thread *pt)
2398 {
2399 
2400 	if (pt == NULL)
2401 		return;
2402 
2403 	memset(pt, 0, THREADENTRY_SIZE);
2404 	mtx_lock_spin(&pmc_threadfreelist_mtx);
2405 	LIST_INSERT_HEAD(&pmc_threadfreelist, pt, pt_next);
2406 	pmc_threadfreelist_entries++;
2407 	if (pmc_threadfreelist_entries > pmc_threadfreelist_max)
2408 		GROUPTASK_ENQUEUE(&free_gtask);
2409 	mtx_unlock_spin(&pmc_threadfreelist_mtx);
2410 }
2411 
2412 /*
2413  * A callout to manage the free list.
2414  */
2415 static void
2416 pmc_thread_descriptor_pool_free_task(void *arg __unused)
2417 {
2418 	struct pmc_thread *pt;
2419 	LIST_HEAD(, pmc_thread) tmplist;
2420 	int delta;
2421 
2422 	LIST_INIT(&tmplist);
2423 	/* Determine what changes, if any, we need to make. */
2424 	mtx_lock_spin(&pmc_threadfreelist_mtx);
2425 	delta = pmc_threadfreelist_entries - pmc_threadfreelist_max;
2426 	while (delta > 0 &&
2427 		   (pt = LIST_FIRST(&pmc_threadfreelist)) != NULL) {
2428 		delta--;
2429 		LIST_REMOVE(pt, pt_next);
2430 		LIST_INSERT_HEAD(&tmplist, pt, pt_next);
2431 	}
2432 	mtx_unlock_spin(&pmc_threadfreelist_mtx);
2433 
2434 	/* If there are entries to free, free them. */
2435 	while (!LIST_EMPTY(&tmplist)) {
2436 		pt = LIST_FIRST(&tmplist);
2437 		LIST_REMOVE(pt, pt_next);
2438 		free(pt, M_PMC);
2439 	}
2440 }
2441 
2442 /*
2443  * Drain the thread free pool, freeing all allocations.
2444  */
2445 static void
2446 pmc_thread_descriptor_pool_drain()
2447 {
2448 	struct pmc_thread *pt, *next;
2449 
2450 	LIST_FOREACH_SAFE(pt, &pmc_threadfreelist, pt_next, next) {
2451 		LIST_REMOVE(pt, pt_next);
2452 		free(pt, M_PMC);
2453 	}
2454 }
2455 
2456 /*
2457  * find the descriptor corresponding to thread 'td', adding or removing it
2458  * as specified by 'mode'.
2459  *
2460  * Note that this supports additional mode flags in addition to those
2461  * supported by pmc_find_process_descriptor():
2462  * PMC_FLAG_NOWAIT: Causes the function to not wait for mallocs.
2463  *     This makes it safe to call while holding certain other locks.
2464  */
2465 
2466 static struct pmc_thread *
2467 pmc_find_thread_descriptor(struct pmc_process *pp, struct thread *td,
2468     uint32_t mode)
2469 {
2470 	struct pmc_thread *pt = NULL, *ptnew = NULL;
2471 	int wait_flag;
2472 
2473 	KASSERT(td != NULL, ("[pmc,%d] called to add NULL td", __LINE__));
2474 
2475 	/*
2476 	 * Pre-allocate memory in the PMC_FLAG_ALLOCATE case prior to
2477 	 * acquiring the lock.
2478 	 */
2479 	if (mode & PMC_FLAG_ALLOCATE) {
2480 		if ((ptnew = pmc_thread_descriptor_pool_alloc()) == NULL) {
2481 			wait_flag = M_WAITOK;
2482 			if ((mode & PMC_FLAG_NOWAIT) || in_epoch())
2483 				wait_flag = M_NOWAIT;
2484 
2485 			ptnew = malloc(THREADENTRY_SIZE, M_PMC,
2486 			    wait_flag|M_ZERO);
2487 		}
2488 	}
2489 
2490 	mtx_lock_spin(pp->pp_tdslock);
2491 
2492 	LIST_FOREACH(pt, &pp->pp_tds, pt_next)
2493 		if (pt->pt_td == td)
2494 			break;
2495 
2496 	if ((mode & PMC_FLAG_REMOVE) && pt != NULL)
2497 		LIST_REMOVE(pt, pt_next);
2498 
2499 	if ((mode & PMC_FLAG_ALLOCATE) && pt == NULL && ptnew != NULL) {
2500 		pt = ptnew;
2501 		ptnew = NULL;
2502 		pt->pt_td = td;
2503 		LIST_INSERT_HEAD(&pp->pp_tds, pt, pt_next);
2504 	}
2505 
2506 	mtx_unlock_spin(pp->pp_tdslock);
2507 
2508 	if (ptnew != NULL) {
2509 		free(ptnew, M_PMC);
2510 	}
2511 
2512 	return pt;
2513 }
2514 
2515 /*
2516  * Try to add thread descriptors for each thread in a process.
2517  */
2518 
2519 static void
2520 pmc_add_thread_descriptors_from_proc(struct proc *p, struct pmc_process *pp)
2521 {
2522 	struct thread *curtd;
2523 	struct pmc_thread **tdlist;
2524 	int i, tdcnt, tdlistsz;
2525 
2526 	KASSERT(!PROC_LOCKED(p), ("[pmc,%d] proc unexpectedly locked",
2527 	    __LINE__));
2528 	tdcnt = 32;
2529  restart:
2530 	tdlistsz = roundup2(tdcnt, 32);
2531 
2532 	tdcnt = 0;
2533 	tdlist = malloc(sizeof(struct pmc_thread*) * tdlistsz, M_TEMP, M_WAITOK);
2534 
2535 	PROC_LOCK(p);
2536 	FOREACH_THREAD_IN_PROC(p, curtd)
2537 		tdcnt++;
2538 	if (tdcnt >= tdlistsz) {
2539 		PROC_UNLOCK(p);
2540 		free(tdlist, M_TEMP);
2541 		goto restart;
2542 	}
2543 	/*
2544 	 * Try to add each thread to the list without sleeping. If unable,
2545 	 * add to a queue to retry after dropping the process lock.
2546 	 */
2547 	tdcnt = 0;
2548 	FOREACH_THREAD_IN_PROC(p, curtd) {
2549 		tdlist[tdcnt] = pmc_find_thread_descriptor(pp, curtd,
2550 						   PMC_FLAG_ALLOCATE|PMC_FLAG_NOWAIT);
2551 		if (tdlist[tdcnt] == NULL) {
2552 			PROC_UNLOCK(p);
2553 			for (i = 0; i <= tdcnt; i++)
2554 				pmc_thread_descriptor_pool_free(tdlist[i]);
2555 			free(tdlist, M_TEMP);
2556 			goto restart;
2557 		}
2558 		tdcnt++;
2559 	}
2560 	PROC_UNLOCK(p);
2561 	free(tdlist, M_TEMP);
2562 }
2563 
2564 /*
2565  * find the descriptor corresponding to process 'p', adding or removing it
2566  * as specified by 'mode'.
2567  */
2568 
2569 static struct pmc_process *
2570 pmc_find_process_descriptor(struct proc *p, uint32_t mode)
2571 {
2572 	uint32_t hindex;
2573 	struct pmc_process *pp, *ppnew;
2574 	struct pmc_processhash *pph;
2575 
2576 	hindex = PMC_HASH_PTR(p, pmc_processhashmask);
2577 	pph = &pmc_processhash[hindex];
2578 
2579 	ppnew = NULL;
2580 
2581 	/*
2582 	 * Pre-allocate memory in the PMC_FLAG_ALLOCATE case since we
2583 	 * cannot call malloc(9) once we hold a spin lock.
2584 	 */
2585 	if (mode & PMC_FLAG_ALLOCATE)
2586 		ppnew = malloc(sizeof(struct pmc_process) + md->pmd_npmc *
2587 		    sizeof(struct pmc_targetstate), M_PMC, M_WAITOK|M_ZERO);
2588 
2589 	mtx_lock_spin(&pmc_processhash_mtx);
2590 	LIST_FOREACH(pp, pph, pp_next)
2591 	    if (pp->pp_proc == p)
2592 		    break;
2593 
2594 	if ((mode & PMC_FLAG_REMOVE) && pp != NULL)
2595 		LIST_REMOVE(pp, pp_next);
2596 
2597 	if ((mode & PMC_FLAG_ALLOCATE) && pp == NULL &&
2598 	    ppnew != NULL) {
2599 		ppnew->pp_proc = p;
2600 		LIST_INIT(&ppnew->pp_tds);
2601 		ppnew->pp_tdslock = mtx_pool_find(pmc_mtxpool, ppnew);
2602 		LIST_INSERT_HEAD(pph, ppnew, pp_next);
2603 		mtx_unlock_spin(&pmc_processhash_mtx);
2604 		pp = ppnew;
2605 		ppnew = NULL;
2606 
2607 		/* Add thread descriptors for this process' current threads. */
2608 		pmc_add_thread_descriptors_from_proc(p, pp);
2609 	}
2610 	else
2611 		mtx_unlock_spin(&pmc_processhash_mtx);
2612 
2613 	if (ppnew != NULL)
2614 		free(ppnew, M_PMC);
2615 
2616 	return pp;
2617 }
2618 
2619 /*
2620  * remove a process descriptor from the process hash table.
2621  */
2622 
2623 static void
2624 pmc_remove_process_descriptor(struct pmc_process *pp)
2625 {
2626 	KASSERT(pp->pp_refcnt == 0,
2627 	    ("[pmc,%d] Removing process descriptor %p with count %d",
2628 		__LINE__, pp, pp->pp_refcnt));
2629 
2630 	mtx_lock_spin(&pmc_processhash_mtx);
2631 	LIST_REMOVE(pp, pp_next);
2632 	mtx_unlock_spin(&pmc_processhash_mtx);
2633 }
2634 
2635 /*
2636  * destroy a process descriptor.
2637  */
2638 
2639 static void
2640 pmc_destroy_process_descriptor(struct pmc_process *pp)
2641 {
2642 	struct pmc_thread *pmc_td;
2643 
2644 	while ((pmc_td = LIST_FIRST(&pp->pp_tds)) != NULL) {
2645 		LIST_REMOVE(pmc_td, pt_next);
2646 		pmc_thread_descriptor_pool_free(pmc_td);
2647 	}
2648 	free(pp, M_PMC);
2649 }
2650 
2651 
2652 /*
2653  * find an owner descriptor corresponding to proc 'p'
2654  */
2655 
2656 static struct pmc_owner *
2657 pmc_find_owner_descriptor(struct proc *p)
2658 {
2659 	uint32_t hindex;
2660 	struct pmc_owner *po;
2661 	struct pmc_ownerhash *poh;
2662 
2663 	hindex = PMC_HASH_PTR(p, pmc_ownerhashmask);
2664 	poh = &pmc_ownerhash[hindex];
2665 
2666 	po = NULL;
2667 	LIST_FOREACH(po, poh, po_next)
2668 	    if (po->po_owner == p)
2669 		    break;
2670 
2671 	PMCDBG5(OWN,FND,1, "find-owner proc=%p (%d, %s) hindex=0x%x -> "
2672 	    "pmc-owner=%p", p, p->p_pid, p->p_comm, hindex, po);
2673 
2674 	return po;
2675 }
2676 
2677 /*
2678  * pmc_allocate_pmc_descriptor
2679  *
2680  * Allocate a pmc descriptor and initialize its
2681  * fields.
2682  */
2683 
2684 static struct pmc *
2685 pmc_allocate_pmc_descriptor(void)
2686 {
2687 	struct pmc *pmc;
2688 
2689 	pmc = malloc(sizeof(struct pmc), M_PMC, M_WAITOK|M_ZERO);
2690 	pmc->pm_runcount = counter_u64_alloc(M_WAITOK);
2691 	pmc->pm_pcpu_state = malloc(sizeof(struct pmc_pcpu_state)*mp_ncpus, M_PMC, M_WAITOK|M_ZERO);
2692 	PMCDBG1(PMC,ALL,1, "allocate-pmc -> pmc=%p", pmc);
2693 
2694 	return pmc;
2695 }
2696 
2697 /*
2698  * Destroy a pmc descriptor.
2699  */
2700 
2701 static void
2702 pmc_destroy_pmc_descriptor(struct pmc *pm)
2703 {
2704 
2705 	KASSERT(pm->pm_state == PMC_STATE_DELETED ||
2706 	    pm->pm_state == PMC_STATE_FREE,
2707 	    ("[pmc,%d] destroying non-deleted PMC", __LINE__));
2708 	KASSERT(LIST_EMPTY(&pm->pm_targets),
2709 	    ("[pmc,%d] destroying pmc with targets", __LINE__));
2710 	KASSERT(pm->pm_owner == NULL,
2711 	    ("[pmc,%d] destroying pmc attached to an owner", __LINE__));
2712 	KASSERT(counter_u64_fetch(pm->pm_runcount) == 0,
2713 	    ("[pmc,%d] pmc has non-zero run count %ld", __LINE__,
2714 		 (unsigned long)counter_u64_fetch(pm->pm_runcount)));
2715 
2716 	counter_u64_free(pm->pm_runcount);
2717 	free(pm->pm_pcpu_state, M_PMC);
2718 	free(pm, M_PMC);
2719 }
2720 
2721 static void
2722 pmc_wait_for_pmc_idle(struct pmc *pm)
2723 {
2724 #ifdef HWPMC_DEBUG
2725 	volatile int maxloop;
2726 
2727 	maxloop = 100 * pmc_cpu_max();
2728 #endif
2729 	/*
2730 	 * Loop (with a forced context switch) till the PMC's runcount
2731 	 * comes down to zero.
2732 	 */
2733 	pmclog_flush(pm->pm_owner, 1);
2734 	while (counter_u64_fetch(pm->pm_runcount) > 0) {
2735 		pmclog_flush(pm->pm_owner, 1);
2736 #ifdef HWPMC_DEBUG
2737 		maxloop--;
2738 		KASSERT(maxloop > 0,
2739 		    ("[pmc,%d] (ri%d, rc%ld) waiting too long for "
2740 			"pmc to be free", __LINE__,
2741 			 PMC_TO_ROWINDEX(pm), (unsigned long)counter_u64_fetch(pm->pm_runcount)));
2742 #endif
2743 		pmc_force_context_switch();
2744 	}
2745 }
2746 
2747 /*
2748  * This function does the following things:
2749  *
2750  *  - detaches the PMC from hardware
2751  *  - unlinks all target threads that were attached to it
2752  *  - removes the PMC from its owner's list
2753  *  - destroys the PMC private mutex
2754  *
2755  * Once this function completes, the given pmc pointer can be freed by
2756  * calling pmc_destroy_pmc_descriptor().
2757  */
2758 
2759 static void
2760 pmc_release_pmc_descriptor(struct pmc *pm)
2761 {
2762 	enum pmc_mode mode;
2763 	struct pmc_hw *phw;
2764 	u_int adjri, ri, cpu;
2765 	struct pmc_owner *po;
2766 	struct pmc_binding pb;
2767 	struct pmc_process *pp;
2768 	struct pmc_classdep *pcd;
2769 	struct pmc_target *ptgt, *tmp;
2770 
2771 	sx_assert(&pmc_sx, SX_XLOCKED);
2772 
2773 	KASSERT(pm, ("[pmc,%d] null pmc", __LINE__));
2774 
2775 	ri   = PMC_TO_ROWINDEX(pm);
2776 	pcd  = pmc_ri_to_classdep(md, ri, &adjri);
2777 	mode = PMC_TO_MODE(pm);
2778 
2779 	PMCDBG3(PMC,REL,1, "release-pmc pmc=%p ri=%d mode=%d", pm, ri,
2780 	    mode);
2781 
2782 	/*
2783 	 * First, we take the PMC off hardware.
2784 	 */
2785 	cpu = 0;
2786 	if (PMC_IS_SYSTEM_MODE(mode)) {
2787 
2788 		/*
2789 		 * A system mode PMC runs on a specific CPU.  Switch
2790 		 * to this CPU and turn hardware off.
2791 		 */
2792 		pmc_save_cpu_binding(&pb);
2793 
2794 		cpu = PMC_TO_CPU(pm);
2795 
2796 		pmc_select_cpu(cpu);
2797 
2798 		/* switch off non-stalled CPUs */
2799 		pm->pm_pcpu_state[cpu].pps_cpustate = 0;
2800 		if (pm->pm_state == PMC_STATE_RUNNING &&
2801 			pm->pm_pcpu_state[cpu].pps_stalled == 0) {
2802 
2803 			phw = pmc_pcpu[cpu]->pc_hwpmcs[ri];
2804 
2805 			KASSERT(phw->phw_pmc == pm,
2806 			    ("[pmc, %d] pmc ptr ri(%d) hw(%p) pm(%p)",
2807 				__LINE__, ri, phw->phw_pmc, pm));
2808 			PMCDBG2(PMC,REL,2, "stopping cpu=%d ri=%d", cpu, ri);
2809 
2810 			critical_enter();
2811 			pcd->pcd_stop_pmc(cpu, adjri);
2812 			critical_exit();
2813 		}
2814 
2815 		PMCDBG2(PMC,REL,2, "decfg cpu=%d ri=%d", cpu, ri);
2816 
2817 		critical_enter();
2818 		pcd->pcd_config_pmc(cpu, adjri, NULL);
2819 		critical_exit();
2820 
2821 		/* adjust the global and process count of SS mode PMCs */
2822 		if (mode == PMC_MODE_SS && pm->pm_state == PMC_STATE_RUNNING) {
2823 			po = pm->pm_owner;
2824 			po->po_sscount--;
2825 			if (po->po_sscount == 0) {
2826 				atomic_subtract_rel_int(&pmc_ss_count, 1);
2827 				CK_LIST_REMOVE(po, po_ssnext);
2828 				epoch_wait_preempt(global_epoch_preempt);
2829 			}
2830 		}
2831 
2832 		pm->pm_state = PMC_STATE_DELETED;
2833 
2834 		pmc_restore_cpu_binding(&pb);
2835 
2836 		/*
2837 		 * We could have references to this PMC structure in
2838 		 * the per-cpu sample queues.  Wait for the queue to
2839 		 * drain.
2840 		 */
2841 		pmc_wait_for_pmc_idle(pm);
2842 
2843 	} else if (PMC_IS_VIRTUAL_MODE(mode)) {
2844 
2845 		/*
2846 		 * A virtual PMC could be running on multiple CPUs at
2847 		 * a given instant.
2848 		 *
2849 		 * By marking its state as DELETED, we ensure that
2850 		 * this PMC is never further scheduled on hardware.
2851 		 *
2852 		 * Then we wait till all CPUs are done with this PMC.
2853 		 */
2854 		pm->pm_state = PMC_STATE_DELETED;
2855 
2856 
2857 		/* Wait for the PMCs runcount to come to zero. */
2858 		pmc_wait_for_pmc_idle(pm);
2859 
2860 		/*
2861 		 * At this point the PMC is off all CPUs and cannot be
2862 		 * freshly scheduled onto a CPU.  It is now safe to
2863 		 * unlink all targets from this PMC.  If a
2864 		 * process-record's refcount falls to zero, we remove
2865 		 * it from the hash table.  The module-wide SX lock
2866 		 * protects us from races.
2867 		 */
2868 		LIST_FOREACH_SAFE(ptgt, &pm->pm_targets, pt_next, tmp) {
2869 			pp = ptgt->pt_process;
2870 			pmc_unlink_target_process(pm, pp); /* frees 'ptgt' */
2871 
2872 			PMCDBG1(PMC,REL,3, "pp->refcnt=%d", pp->pp_refcnt);
2873 
2874 			/*
2875 			 * If the target process record shows that no
2876 			 * PMCs are attached to it, reclaim its space.
2877 			 */
2878 
2879 			if (pp->pp_refcnt == 0) {
2880 				pmc_remove_process_descriptor(pp);
2881 				pmc_destroy_process_descriptor(pp);
2882 			}
2883 		}
2884 
2885 		cpu = curthread->td_oncpu; /* setup cpu for pmd_release() */
2886 
2887 	}
2888 
2889 	/*
2890 	 * Release any MD resources
2891 	 */
2892 	(void) pcd->pcd_release_pmc(cpu, adjri, pm);
2893 
2894 	/*
2895 	 * Update row disposition
2896 	 */
2897 
2898 	if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm)))
2899 		PMC_UNMARK_ROW_STANDALONE(ri);
2900 	else
2901 		PMC_UNMARK_ROW_THREAD(ri);
2902 
2903 	/* unlink from the owner's list */
2904 	if (pm->pm_owner) {
2905 		LIST_REMOVE(pm, pm_next);
2906 		pm->pm_owner = NULL;
2907 	}
2908 }
2909 
2910 /*
2911  * Register an owner and a pmc.
2912  */
2913 
2914 static int
2915 pmc_register_owner(struct proc *p, struct pmc *pmc)
2916 {
2917 	struct pmc_owner *po;
2918 
2919 	sx_assert(&pmc_sx, SX_XLOCKED);
2920 
2921 	if ((po = pmc_find_owner_descriptor(p)) == NULL)
2922 		if ((po = pmc_allocate_owner_descriptor(p)) == NULL)
2923 			return ENOMEM;
2924 
2925 	KASSERT(pmc->pm_owner == NULL,
2926 	    ("[pmc,%d] attempting to own an initialized PMC", __LINE__));
2927 	pmc->pm_owner  = po;
2928 
2929 	LIST_INSERT_HEAD(&po->po_pmcs, pmc, pm_next);
2930 
2931 	PROC_LOCK(p);
2932 	p->p_flag |= P_HWPMC;
2933 	PROC_UNLOCK(p);
2934 
2935 	if (po->po_flags & PMC_PO_OWNS_LOGFILE)
2936 		pmclog_process_pmcallocate(pmc);
2937 
2938 	PMCDBG2(PMC,REG,1, "register-owner pmc-owner=%p pmc=%p",
2939 	    po, pmc);
2940 
2941 	return 0;
2942 }
2943 
2944 /*
2945  * Return the current row disposition:
2946  * == 0 => FREE
2947  *  > 0 => PROCESS MODE
2948  *  < 0 => SYSTEM MODE
2949  */
2950 
2951 int
2952 pmc_getrowdisp(int ri)
2953 {
2954 	return pmc_pmcdisp[ri];
2955 }
2956 
2957 /*
2958  * Check if a PMC at row index 'ri' can be allocated to the current
2959  * process.
2960  *
2961  * Allocation can fail if:
2962  *   - the current process is already being profiled by a PMC at index 'ri',
2963  *     attached to it via OP_PMCATTACH.
2964  *   - the current process has already allocated a PMC at index 'ri'
2965  *     via OP_ALLOCATE.
2966  */
2967 
2968 static int
2969 pmc_can_allocate_rowindex(struct proc *p, unsigned int ri, int cpu)
2970 {
2971 	enum pmc_mode mode;
2972 	struct pmc *pm;
2973 	struct pmc_owner *po;
2974 	struct pmc_process *pp;
2975 
2976 	PMCDBG5(PMC,ALR,1, "can-allocate-rowindex proc=%p (%d, %s) ri=%d "
2977 	    "cpu=%d", p, p->p_pid, p->p_comm, ri, cpu);
2978 
2979 	/*
2980 	 * We shouldn't have already allocated a process-mode PMC at
2981 	 * row index 'ri'.
2982 	 *
2983 	 * We shouldn't have allocated a system-wide PMC on the same
2984 	 * CPU and same RI.
2985 	 */
2986 	if ((po = pmc_find_owner_descriptor(p)) != NULL)
2987 		LIST_FOREACH(pm, &po->po_pmcs, pm_next) {
2988 		    if (PMC_TO_ROWINDEX(pm) == ri) {
2989 			    mode = PMC_TO_MODE(pm);
2990 			    if (PMC_IS_VIRTUAL_MODE(mode))
2991 				    return EEXIST;
2992 			    if (PMC_IS_SYSTEM_MODE(mode) &&
2993 				(int) PMC_TO_CPU(pm) == cpu)
2994 				    return EEXIST;
2995 		    }
2996 	        }
2997 
2998 	/*
2999 	 * We also shouldn't be the target of any PMC at this index
3000 	 * since otherwise a PMC_ATTACH to ourselves will fail.
3001 	 */
3002 	if ((pp = pmc_find_process_descriptor(p, 0)) != NULL)
3003 		if (pp->pp_pmcs[ri].pp_pmc)
3004 			return EEXIST;
3005 
3006 	PMCDBG4(PMC,ALR,2, "can-allocate-rowindex proc=%p (%d, %s) ri=%d ok",
3007 	    p, p->p_pid, p->p_comm, ri);
3008 
3009 	return 0;
3010 }
3011 
3012 /*
3013  * Check if a given PMC at row index 'ri' can be currently used in
3014  * mode 'mode'.
3015  */
3016 
3017 static int
3018 pmc_can_allocate_row(int ri, enum pmc_mode mode)
3019 {
3020 	enum pmc_disp	disp;
3021 
3022 	sx_assert(&pmc_sx, SX_XLOCKED);
3023 
3024 	PMCDBG2(PMC,ALR,1, "can-allocate-row ri=%d mode=%d", ri, mode);
3025 
3026 	if (PMC_IS_SYSTEM_MODE(mode))
3027 		disp = PMC_DISP_STANDALONE;
3028 	else
3029 		disp = PMC_DISP_THREAD;
3030 
3031 	/*
3032 	 * check disposition for PMC row 'ri':
3033 	 *
3034 	 * Expected disposition		Row-disposition		Result
3035 	 *
3036 	 * STANDALONE			STANDALONE or FREE	proceed
3037 	 * STANDALONE			THREAD			fail
3038 	 * THREAD			THREAD or FREE		proceed
3039 	 * THREAD			STANDALONE		fail
3040 	 */
3041 
3042 	if (!PMC_ROW_DISP_IS_FREE(ri) &&
3043 	    !(disp == PMC_DISP_THREAD && PMC_ROW_DISP_IS_THREAD(ri)) &&
3044 	    !(disp == PMC_DISP_STANDALONE && PMC_ROW_DISP_IS_STANDALONE(ri)))
3045 		return EBUSY;
3046 
3047 	/*
3048 	 * All OK
3049 	 */
3050 
3051 	PMCDBG2(PMC,ALR,2, "can-allocate-row ri=%d mode=%d ok", ri, mode);
3052 
3053 	return 0;
3054 
3055 }
3056 
3057 /*
3058  * Find a PMC descriptor with user handle 'pmcid' for thread 'td'.
3059  */
3060 
3061 static struct pmc *
3062 pmc_find_pmc_descriptor_in_process(struct pmc_owner *po, pmc_id_t pmcid)
3063 {
3064 	struct pmc *pm;
3065 
3066 	KASSERT(PMC_ID_TO_ROWINDEX(pmcid) < md->pmd_npmc,
3067 	    ("[pmc,%d] Illegal pmc index %d (max %d)", __LINE__,
3068 		PMC_ID_TO_ROWINDEX(pmcid), md->pmd_npmc));
3069 
3070 	LIST_FOREACH(pm, &po->po_pmcs, pm_next)
3071 	    if (pm->pm_id == pmcid)
3072 		    return pm;
3073 
3074 	return NULL;
3075 }
3076 
3077 static int
3078 pmc_find_pmc(pmc_id_t pmcid, struct pmc **pmc)
3079 {
3080 
3081 	struct pmc *pm, *opm;
3082 	struct pmc_owner *po;
3083 	struct pmc_process *pp;
3084 
3085 	PMCDBG1(PMC,FND,1, "find-pmc id=%d", pmcid);
3086 	if (PMC_ID_TO_ROWINDEX(pmcid) >= md->pmd_npmc)
3087 		return (EINVAL);
3088 
3089 	if ((po = pmc_find_owner_descriptor(curthread->td_proc)) == NULL) {
3090 		/*
3091 		 * In case of PMC_F_DESCENDANTS child processes we will not find
3092 		 * the current process in the owners hash list.  Find the owner
3093 		 * process first and from there lookup the po.
3094 		 */
3095 		if ((pp = pmc_find_process_descriptor(curthread->td_proc,
3096 		    PMC_FLAG_NONE)) == NULL) {
3097 			return ESRCH;
3098 		} else {
3099 			opm = pp->pp_pmcs[PMC_ID_TO_ROWINDEX(pmcid)].pp_pmc;
3100 			if (opm == NULL)
3101 				return ESRCH;
3102 			if ((opm->pm_flags & (PMC_F_ATTACHED_TO_OWNER|
3103 			    PMC_F_DESCENDANTS)) != (PMC_F_ATTACHED_TO_OWNER|
3104 			    PMC_F_DESCENDANTS))
3105 				return ESRCH;
3106 			po = opm->pm_owner;
3107 		}
3108 	}
3109 
3110 	if ((pm = pmc_find_pmc_descriptor_in_process(po, pmcid)) == NULL)
3111 		return EINVAL;
3112 
3113 	PMCDBG2(PMC,FND,2, "find-pmc id=%d -> pmc=%p", pmcid, pm);
3114 
3115 	*pmc = pm;
3116 	return 0;
3117 }
3118 
3119 /*
3120  * Start a PMC.
3121  */
3122 
3123 static int
3124 pmc_start(struct pmc *pm)
3125 {
3126 	enum pmc_mode mode;
3127 	struct pmc_owner *po;
3128 	struct pmc_binding pb;
3129 	struct pmc_classdep *pcd;
3130 	int adjri, error, cpu, ri;
3131 
3132 	KASSERT(pm != NULL,
3133 	    ("[pmc,%d] null pm", __LINE__));
3134 
3135 	mode = PMC_TO_MODE(pm);
3136 	ri   = PMC_TO_ROWINDEX(pm);
3137 	pcd  = pmc_ri_to_classdep(md, ri, &adjri);
3138 
3139 	error = 0;
3140 
3141 	PMCDBG3(PMC,OPS,1, "start pmc=%p mode=%d ri=%d", pm, mode, ri);
3142 
3143 	po = pm->pm_owner;
3144 
3145 	/*
3146 	 * Disallow PMCSTART if a logfile is required but has not been
3147 	 * configured yet.
3148 	 */
3149 	if ((pm->pm_flags & PMC_F_NEEDS_LOGFILE) &&
3150 	    (po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
3151 		return (EDOOFUS);	/* programming error */
3152 
3153 	/*
3154 	 * If this is a sampling mode PMC, log mapping information for
3155 	 * the kernel modules that are currently loaded.
3156 	 */
3157 	if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
3158 	    pmc_log_kernel_mappings(pm);
3159 
3160 	if (PMC_IS_VIRTUAL_MODE(mode)) {
3161 
3162 		/*
3163 		 * If a PMCATTACH has never been done on this PMC,
3164 		 * attach it to its owner process.
3165 		 */
3166 
3167 		if (LIST_EMPTY(&pm->pm_targets))
3168 			error = (pm->pm_flags & PMC_F_ATTACH_DONE) ? ESRCH :
3169 			    pmc_attach_process(po->po_owner, pm);
3170 
3171 		/*
3172 		 * If the PMC is attached to its owner, then force a context
3173 		 * switch to ensure that the MD state gets set correctly.
3174 		 */
3175 
3176 		if (error == 0) {
3177 			pm->pm_state = PMC_STATE_RUNNING;
3178 			if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER)
3179 				pmc_force_context_switch();
3180 		}
3181 
3182 		return (error);
3183 	}
3184 
3185 
3186 	/*
3187 	 * A system-wide PMC.
3188 	 *
3189 	 * Add the owner to the global list if this is a system-wide
3190 	 * sampling PMC.
3191 	 */
3192 
3193 	if (mode == PMC_MODE_SS) {
3194 		/*
3195 		 * Log mapping information for all existing processes in the
3196 		 * system.  Subsequent mappings are logged as they happen;
3197 		 * see pmc_process_mmap().
3198 		 */
3199 		if (po->po_logprocmaps == 0) {
3200 			pmc_log_all_process_mappings(po);
3201 			po->po_logprocmaps = 1;
3202 		}
3203 		po->po_sscount++;
3204 		if (po->po_sscount == 1) {
3205 			atomic_add_rel_int(&pmc_ss_count, 1);
3206 			CK_LIST_INSERT_HEAD(&pmc_ss_owners, po, po_ssnext);
3207 			PMCDBG1(PMC,OPS,1, "po=%p in global list", po);
3208 		}
3209 	}
3210 
3211 	/*
3212 	 * Move to the CPU associated with this
3213 	 * PMC, and start the hardware.
3214 	 */
3215 
3216 	pmc_save_cpu_binding(&pb);
3217 
3218 	cpu = PMC_TO_CPU(pm);
3219 
3220 	if (!pmc_cpu_is_active(cpu))
3221 		return (ENXIO);
3222 
3223 	pmc_select_cpu(cpu);
3224 
3225 	/*
3226 	 * global PMCs are configured at allocation time
3227 	 * so write out the initial value and start the PMC.
3228 	 */
3229 
3230 	pm->pm_state = PMC_STATE_RUNNING;
3231 
3232 	critical_enter();
3233 	if ((error = pcd->pcd_write_pmc(cpu, adjri,
3234 		 PMC_IS_SAMPLING_MODE(mode) ?
3235 		 pm->pm_sc.pm_reloadcount :
3236 		 pm->pm_sc.pm_initial)) == 0) {
3237 		/* If a sampling mode PMC, reset stalled state. */
3238 		if (PMC_IS_SAMPLING_MODE(mode))
3239 			pm->pm_pcpu_state[cpu].pps_stalled = 0;
3240 
3241 		/* Indicate that we desire this to run. Start it. */
3242 		pm->pm_pcpu_state[cpu].pps_cpustate = 1;
3243 		error = pcd->pcd_start_pmc(cpu, adjri);
3244 	}
3245 	critical_exit();
3246 
3247 	pmc_restore_cpu_binding(&pb);
3248 
3249 	return (error);
3250 }
3251 
3252 /*
3253  * Stop a PMC.
3254  */
3255 
3256 static int
3257 pmc_stop(struct pmc *pm)
3258 {
3259 	struct pmc_owner *po;
3260 	struct pmc_binding pb;
3261 	struct pmc_classdep *pcd;
3262 	int adjri, cpu, error, ri;
3263 
3264 	KASSERT(pm != NULL, ("[pmc,%d] null pmc", __LINE__));
3265 
3266 	PMCDBG3(PMC,OPS,1, "stop pmc=%p mode=%d ri=%d", pm,
3267 	    PMC_TO_MODE(pm), PMC_TO_ROWINDEX(pm));
3268 
3269 	pm->pm_state = PMC_STATE_STOPPED;
3270 
3271 	/*
3272 	 * If the PMC is a virtual mode one, changing the state to
3273 	 * non-RUNNING is enough to ensure that the PMC never gets
3274 	 * scheduled.
3275 	 *
3276 	 * If this PMC is current running on a CPU, then it will
3277 	 * handled correctly at the time its target process is context
3278 	 * switched out.
3279 	 */
3280 
3281 	if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)))
3282 		return 0;
3283 
3284 	/*
3285 	 * A system-mode PMC.  Move to the CPU associated with
3286 	 * this PMC, and stop the hardware.  We update the
3287 	 * 'initial count' so that a subsequent PMCSTART will
3288 	 * resume counting from the current hardware count.
3289 	 */
3290 
3291 	pmc_save_cpu_binding(&pb);
3292 
3293 	cpu = PMC_TO_CPU(pm);
3294 
3295 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
3296 	    ("[pmc,%d] illegal cpu=%d", __LINE__, cpu));
3297 
3298 	if (!pmc_cpu_is_active(cpu))
3299 		return ENXIO;
3300 
3301 	pmc_select_cpu(cpu);
3302 
3303 	ri = PMC_TO_ROWINDEX(pm);
3304 	pcd = pmc_ri_to_classdep(md, ri, &adjri);
3305 
3306 	pm->pm_pcpu_state[cpu].pps_cpustate = 0;
3307 	critical_enter();
3308 	if ((error = pcd->pcd_stop_pmc(cpu, adjri)) == 0)
3309 		error = pcd->pcd_read_pmc(cpu, adjri, &pm->pm_sc.pm_initial);
3310 	critical_exit();
3311 
3312 	pmc_restore_cpu_binding(&pb);
3313 
3314 	po = pm->pm_owner;
3315 
3316 	/* remove this owner from the global list of SS PMC owners */
3317 	if (PMC_TO_MODE(pm) == PMC_MODE_SS) {
3318 		po->po_sscount--;
3319 		if (po->po_sscount == 0) {
3320 			atomic_subtract_rel_int(&pmc_ss_count, 1);
3321 			CK_LIST_REMOVE(po, po_ssnext);
3322 			epoch_wait_preempt(global_epoch_preempt);
3323 			PMCDBG1(PMC,OPS,2,"po=%p removed from global list", po);
3324 		}
3325 	}
3326 
3327 	return (error);
3328 }
3329 
3330 static struct pmc_classdep *
3331 pmc_class_to_classdep(enum pmc_class class)
3332 {
3333 	int n;
3334 
3335 	for (n = 0; n < md->pmd_nclass; n++)
3336 		if (md->pmd_classdep[n].pcd_class == class)
3337 			return (&md->pmd_classdep[n]);
3338 	return (NULL);
3339 }
3340 
3341 #if defined(HWPMC_DEBUG) && defined(KTR)
3342 static const char *pmc_op_to_name[] = {
3343 #undef	__PMC_OP
3344 #define	__PMC_OP(N, D)	#N ,
3345 	__PMC_OPS()
3346 	NULL
3347 };
3348 #endif
3349 
3350 /*
3351  * The syscall interface
3352  */
3353 
3354 #define	PMC_GET_SX_XLOCK(...) do {		\
3355 	sx_xlock(&pmc_sx);			\
3356 	if (pmc_hook == NULL) {			\
3357 		sx_xunlock(&pmc_sx);		\
3358 		return __VA_ARGS__;		\
3359 	}					\
3360 } while (0)
3361 
3362 #define	PMC_DOWNGRADE_SX() do {			\
3363 	sx_downgrade(&pmc_sx);			\
3364 	is_sx_downgraded = 1;			\
3365 } while (0)
3366 
3367 static int
3368 pmc_syscall_handler(struct thread *td, void *syscall_args)
3369 {
3370 	int error, is_sx_downgraded, op;
3371 	struct pmc_syscall_args *c;
3372 	void *pmclog_proc_handle;
3373 	void *arg;
3374 
3375 	c = (struct pmc_syscall_args *)syscall_args;
3376 	op = c->pmop_code;
3377 	arg = c->pmop_data;
3378 	/* PMC isn't set up yet */
3379 	if (pmc_hook == NULL)
3380 		return (EINVAL);
3381 	if (op == PMC_OP_CONFIGURELOG) {
3382 		/*
3383 		 * We cannot create the logging process inside
3384 		 * pmclog_configure_log() because there is a LOR
3385 		 * between pmc_sx and process structure locks.
3386 		 * Instead, pre-create the process and ignite the loop
3387 		 * if everything is fine, otherwise direct the process
3388 		 * to exit.
3389 		 */
3390 		error = pmclog_proc_create(td, &pmclog_proc_handle);
3391 		if (error != 0)
3392 			goto done_syscall;
3393 	}
3394 
3395 	PMC_GET_SX_XLOCK(ENOSYS);
3396 	is_sx_downgraded = 0;
3397 	PMCDBG3(MOD,PMS,1, "syscall op=%d \"%s\" arg=%p", op,
3398 	    pmc_op_to_name[op], arg);
3399 
3400 	error = 0;
3401 	counter_u64_add(pmc_stats.pm_syscalls, 1);
3402 
3403 	switch (op) {
3404 
3405 
3406 	/*
3407 	 * Configure a log file.
3408 	 *
3409 	 * XXX This OP will be reworked.
3410 	 */
3411 
3412 	case PMC_OP_CONFIGURELOG:
3413 	{
3414 		struct proc *p;
3415 		struct pmc *pm;
3416 		struct pmc_owner *po;
3417 		struct pmc_op_configurelog cl;
3418 
3419 		if ((error = copyin(arg, &cl, sizeof(cl))) != 0) {
3420 			pmclog_proc_ignite(pmclog_proc_handle, NULL);
3421 			break;
3422 		}
3423 
3424 		/* mark this process as owning a log file */
3425 		p = td->td_proc;
3426 		if ((po = pmc_find_owner_descriptor(p)) == NULL)
3427 			if ((po = pmc_allocate_owner_descriptor(p)) == NULL) {
3428 				pmclog_proc_ignite(pmclog_proc_handle, NULL);
3429 				error = ENOMEM;
3430 				break;
3431 			}
3432 
3433 		/*
3434 		 * If a valid fd was passed in, try to configure that,
3435 		 * otherwise if 'fd' was less than zero and there was
3436 		 * a log file configured, flush its buffers and
3437 		 * de-configure it.
3438 		 */
3439 		if (cl.pm_logfd >= 0) {
3440 			error = pmclog_configure_log(md, po, cl.pm_logfd);
3441 			pmclog_proc_ignite(pmclog_proc_handle, error == 0 ?
3442 			    po : NULL);
3443 		} else if (po->po_flags & PMC_PO_OWNS_LOGFILE) {
3444 			pmclog_proc_ignite(pmclog_proc_handle, NULL);
3445 			error = pmclog_close(po);
3446 			if (error == 0) {
3447 				LIST_FOREACH(pm, &po->po_pmcs, pm_next)
3448 				    if (pm->pm_flags & PMC_F_NEEDS_LOGFILE &&
3449 					pm->pm_state == PMC_STATE_RUNNING)
3450 					    pmc_stop(pm);
3451 				error = pmclog_deconfigure_log(po);
3452 			}
3453 		} else {
3454 			pmclog_proc_ignite(pmclog_proc_handle, NULL);
3455 			error = EINVAL;
3456 		}
3457 	}
3458 	break;
3459 
3460 	/*
3461 	 * Flush a log file.
3462 	 */
3463 
3464 	case PMC_OP_FLUSHLOG:
3465 	{
3466 		struct pmc_owner *po;
3467 
3468 		sx_assert(&pmc_sx, SX_XLOCKED);
3469 
3470 		if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) {
3471 			error = EINVAL;
3472 			break;
3473 		}
3474 
3475 		error = pmclog_flush(po, 0);
3476 	}
3477 	break;
3478 
3479 	/*
3480 	 * Close a log file.
3481 	 */
3482 
3483 	case PMC_OP_CLOSELOG:
3484 	{
3485 		struct pmc_owner *po;
3486 
3487 		sx_assert(&pmc_sx, SX_XLOCKED);
3488 
3489 		if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) {
3490 			error = EINVAL;
3491 			break;
3492 		}
3493 
3494 		error = pmclog_close(po);
3495 	}
3496 	break;
3497 
3498 	/*
3499 	 * Retrieve hardware configuration.
3500 	 */
3501 
3502 	case PMC_OP_GETCPUINFO:	/* CPU information */
3503 	{
3504 		struct pmc_op_getcpuinfo gci;
3505 		struct pmc_classinfo *pci;
3506 		struct pmc_classdep *pcd;
3507 		int cl;
3508 
3509 		gci.pm_cputype = md->pmd_cputype;
3510 		gci.pm_ncpu    = pmc_cpu_max();
3511 		gci.pm_npmc    = md->pmd_npmc;
3512 		gci.pm_nclass  = md->pmd_nclass;
3513 		pci = gci.pm_classes;
3514 		pcd = md->pmd_classdep;
3515 		for (cl = 0; cl < md->pmd_nclass; cl++, pci++, pcd++) {
3516 			pci->pm_caps  = pcd->pcd_caps;
3517 			pci->pm_class = pcd->pcd_class;
3518 			pci->pm_width = pcd->pcd_width;
3519 			pci->pm_num   = pcd->pcd_num;
3520 		}
3521 		error = copyout(&gci, arg, sizeof(gci));
3522 	}
3523 	break;
3524 
3525 	/*
3526 	 * Retrieve soft events list.
3527 	 */
3528 	case PMC_OP_GETDYNEVENTINFO:
3529 	{
3530 		enum pmc_class			cl;
3531 		enum pmc_event			ev;
3532 		struct pmc_op_getdyneventinfo	*gei;
3533 		struct pmc_dyn_event_descr	dev;
3534 		struct pmc_soft			*ps;
3535 		uint32_t			nevent;
3536 
3537 		sx_assert(&pmc_sx, SX_LOCKED);
3538 
3539 		gei = (struct pmc_op_getdyneventinfo *) arg;
3540 
3541 		if ((error = copyin(&gei->pm_class, &cl, sizeof(cl))) != 0)
3542 			break;
3543 
3544 		/* Only SOFT class is dynamic. */
3545 		if (cl != PMC_CLASS_SOFT) {
3546 			error = EINVAL;
3547 			break;
3548 		}
3549 
3550 		nevent = 0;
3551 		for (ev = PMC_EV_SOFT_FIRST; (int)ev <= PMC_EV_SOFT_LAST; ev++) {
3552 			ps = pmc_soft_ev_acquire(ev);
3553 			if (ps == NULL)
3554 				continue;
3555 			bcopy(&ps->ps_ev, &dev, sizeof(dev));
3556 			pmc_soft_ev_release(ps);
3557 
3558 			error = copyout(&dev,
3559 			    &gei->pm_events[nevent],
3560 			    sizeof(struct pmc_dyn_event_descr));
3561 			if (error != 0)
3562 				break;
3563 			nevent++;
3564 		}
3565 		if (error != 0)
3566 			break;
3567 
3568 		error = copyout(&nevent, &gei->pm_nevent,
3569 		    sizeof(nevent));
3570 	}
3571 	break;
3572 
3573 	/*
3574 	 * Get module statistics
3575 	 */
3576 
3577 	case PMC_OP_GETDRIVERSTATS:
3578 	{
3579 		struct pmc_op_getdriverstats gms;
3580 #define CFETCH(a, b, field) a.field = counter_u64_fetch(b.field)
3581 		CFETCH(gms, pmc_stats, pm_intr_ignored);
3582 		CFETCH(gms, pmc_stats, pm_intr_processed);
3583 		CFETCH(gms, pmc_stats, pm_intr_bufferfull);
3584 		CFETCH(gms, pmc_stats, pm_syscalls);
3585 		CFETCH(gms, pmc_stats, pm_syscall_errors);
3586 		CFETCH(gms, pmc_stats, pm_buffer_requests);
3587 		CFETCH(gms, pmc_stats, pm_buffer_requests_failed);
3588 		CFETCH(gms, pmc_stats, pm_log_sweeps);
3589 #undef CFETCH
3590 		error = copyout(&gms, arg, sizeof(gms));
3591 	}
3592 	break;
3593 
3594 
3595 	/*
3596 	 * Retrieve module version number
3597 	 */
3598 
3599 	case PMC_OP_GETMODULEVERSION:
3600 	{
3601 		uint32_t cv, modv;
3602 
3603 		/* retrieve the client's idea of the ABI version */
3604 		if ((error = copyin(arg, &cv, sizeof(uint32_t))) != 0)
3605 			break;
3606 		/* don't service clients newer than our driver */
3607 		modv = PMC_VERSION;
3608 		if ((cv & 0xFFFF0000) > (modv & 0xFFFF0000)) {
3609 			error = EPROGMISMATCH;
3610 			break;
3611 		}
3612 		error = copyout(&modv, arg, sizeof(int));
3613 	}
3614 	break;
3615 
3616 
3617 	/*
3618 	 * Retrieve the state of all the PMCs on a given
3619 	 * CPU.
3620 	 */
3621 
3622 	case PMC_OP_GETPMCINFO:
3623 	{
3624 		int ari;
3625 		struct pmc *pm;
3626 		size_t pmcinfo_size;
3627 		uint32_t cpu, n, npmc;
3628 		struct pmc_owner *po;
3629 		struct pmc_binding pb;
3630 		struct pmc_classdep *pcd;
3631 		struct pmc_info *p, *pmcinfo;
3632 		struct pmc_op_getpmcinfo *gpi;
3633 
3634 		PMC_DOWNGRADE_SX();
3635 
3636 		gpi = (struct pmc_op_getpmcinfo *) arg;
3637 
3638 		if ((error = copyin(&gpi->pm_cpu, &cpu, sizeof(cpu))) != 0)
3639 			break;
3640 
3641 		if (cpu >= pmc_cpu_max()) {
3642 			error = EINVAL;
3643 			break;
3644 		}
3645 
3646 		if (!pmc_cpu_is_active(cpu)) {
3647 			error = ENXIO;
3648 			break;
3649 		}
3650 
3651 		/* switch to CPU 'cpu' */
3652 		pmc_save_cpu_binding(&pb);
3653 		pmc_select_cpu(cpu);
3654 
3655 		npmc = md->pmd_npmc;
3656 
3657 		pmcinfo_size = npmc * sizeof(struct pmc_info);
3658 		pmcinfo = malloc(pmcinfo_size, M_PMC, M_WAITOK);
3659 
3660 		p = pmcinfo;
3661 
3662 		for (n = 0; n < md->pmd_npmc; n++, p++) {
3663 
3664 			pcd = pmc_ri_to_classdep(md, n, &ari);
3665 
3666 			KASSERT(pcd != NULL,
3667 			    ("[pmc,%d] null pcd ri=%d", __LINE__, n));
3668 
3669 			if ((error = pcd->pcd_describe(cpu, ari, p, &pm)) != 0)
3670 				break;
3671 
3672 			if (PMC_ROW_DISP_IS_STANDALONE(n))
3673 				p->pm_rowdisp = PMC_DISP_STANDALONE;
3674 			else if (PMC_ROW_DISP_IS_THREAD(n))
3675 				p->pm_rowdisp = PMC_DISP_THREAD;
3676 			else
3677 				p->pm_rowdisp = PMC_DISP_FREE;
3678 
3679 			p->pm_ownerpid = -1;
3680 
3681 			if (pm == NULL)	/* no PMC associated */
3682 				continue;
3683 
3684 			po = pm->pm_owner;
3685 
3686 			KASSERT(po->po_owner != NULL,
3687 			    ("[pmc,%d] pmc_owner had a null proc pointer",
3688 				__LINE__));
3689 
3690 			p->pm_ownerpid = po->po_owner->p_pid;
3691 			p->pm_mode     = PMC_TO_MODE(pm);
3692 			p->pm_event    = pm->pm_event;
3693 			p->pm_flags    = pm->pm_flags;
3694 
3695 			if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
3696 				p->pm_reloadcount =
3697 				    pm->pm_sc.pm_reloadcount;
3698 		}
3699 
3700 		pmc_restore_cpu_binding(&pb);
3701 
3702 		/* now copy out the PMC info collected */
3703 		if (error == 0)
3704 			error = copyout(pmcinfo, &gpi->pm_pmcs, pmcinfo_size);
3705 
3706 		free(pmcinfo, M_PMC);
3707 	}
3708 	break;
3709 
3710 
3711 	/*
3712 	 * Set the administrative state of a PMC.  I.e. whether
3713 	 * the PMC is to be used or not.
3714 	 */
3715 
3716 	case PMC_OP_PMCADMIN:
3717 	{
3718 		int cpu, ri;
3719 		enum pmc_state request;
3720 		struct pmc_cpu *pc;
3721 		struct pmc_hw *phw;
3722 		struct pmc_op_pmcadmin pma;
3723 		struct pmc_binding pb;
3724 
3725 		sx_assert(&pmc_sx, SX_XLOCKED);
3726 
3727 		KASSERT(td == curthread,
3728 		    ("[pmc,%d] td != curthread", __LINE__));
3729 
3730 		error = priv_check(td, PRIV_PMC_MANAGE);
3731 		if (error)
3732 			break;
3733 
3734 		if ((error = copyin(arg, &pma, sizeof(pma))) != 0)
3735 			break;
3736 
3737 		cpu = pma.pm_cpu;
3738 
3739 		if (cpu < 0 || cpu >= (int) pmc_cpu_max()) {
3740 			error = EINVAL;
3741 			break;
3742 		}
3743 
3744 		if (!pmc_cpu_is_active(cpu)) {
3745 			error = ENXIO;
3746 			break;
3747 		}
3748 
3749 		request = pma.pm_state;
3750 
3751 		if (request != PMC_STATE_DISABLED &&
3752 		    request != PMC_STATE_FREE) {
3753 			error = EINVAL;
3754 			break;
3755 		}
3756 
3757 		ri = pma.pm_pmc; /* pmc id == row index */
3758 		if (ri < 0 || ri >= (int) md->pmd_npmc) {
3759 			error = EINVAL;
3760 			break;
3761 		}
3762 
3763 		/*
3764 		 * We can't disable a PMC with a row-index allocated
3765 		 * for process virtual PMCs.
3766 		 */
3767 
3768 		if (PMC_ROW_DISP_IS_THREAD(ri) &&
3769 		    request == PMC_STATE_DISABLED) {
3770 			error = EBUSY;
3771 			break;
3772 		}
3773 
3774 		/*
3775 		 * otherwise, this PMC on this CPU is either free or
3776 		 * in system-wide mode.
3777 		 */
3778 
3779 		pmc_save_cpu_binding(&pb);
3780 		pmc_select_cpu(cpu);
3781 
3782 		pc  = pmc_pcpu[cpu];
3783 		phw = pc->pc_hwpmcs[ri];
3784 
3785 		/*
3786 		 * XXX do we need some kind of 'forced' disable?
3787 		 */
3788 
3789 		if (phw->phw_pmc == NULL) {
3790 			if (request == PMC_STATE_DISABLED &&
3791 			    (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED)) {
3792 				phw->phw_state &= ~PMC_PHW_FLAG_IS_ENABLED;
3793 				PMC_MARK_ROW_STANDALONE(ri);
3794 			} else if (request == PMC_STATE_FREE &&
3795 			    (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0) {
3796 				phw->phw_state |=  PMC_PHW_FLAG_IS_ENABLED;
3797 				PMC_UNMARK_ROW_STANDALONE(ri);
3798 			}
3799 			/* other cases are a no-op */
3800 		} else
3801 			error = EBUSY;
3802 
3803 		pmc_restore_cpu_binding(&pb);
3804 	}
3805 	break;
3806 
3807 
3808 	/*
3809 	 * Allocate a PMC.
3810 	 */
3811 
3812 	case PMC_OP_PMCALLOCATE:
3813 	{
3814 		int adjri, n;
3815 		u_int cpu;
3816 		uint32_t caps;
3817 		struct pmc *pmc;
3818 		enum pmc_mode mode;
3819 		struct pmc_hw *phw;
3820 		struct pmc_binding pb;
3821 		struct pmc_classdep *pcd;
3822 		struct pmc_op_pmcallocate pa;
3823 
3824 		if ((error = copyin(arg, &pa, sizeof(pa))) != 0)
3825 			break;
3826 
3827 		caps = pa.pm_caps;
3828 		mode = pa.pm_mode;
3829 		cpu  = pa.pm_cpu;
3830 
3831 		if ((mode != PMC_MODE_SS  &&  mode != PMC_MODE_SC  &&
3832 		     mode != PMC_MODE_TS  &&  mode != PMC_MODE_TC) ||
3833 		    (cpu != (u_int) PMC_CPU_ANY && cpu >= pmc_cpu_max())) {
3834 			error = EINVAL;
3835 			break;
3836 		}
3837 
3838 		/*
3839 		 * Virtual PMCs should only ask for a default CPU.
3840 		 * System mode PMCs need to specify a non-default CPU.
3841 		 */
3842 
3843 		if ((PMC_IS_VIRTUAL_MODE(mode) && cpu != (u_int) PMC_CPU_ANY) ||
3844 		    (PMC_IS_SYSTEM_MODE(mode) && cpu == (u_int) PMC_CPU_ANY)) {
3845 			error = EINVAL;
3846 			break;
3847 		}
3848 
3849 		/*
3850 		 * Check that an inactive CPU is not being asked for.
3851 		 */
3852 
3853 		if (PMC_IS_SYSTEM_MODE(mode) && !pmc_cpu_is_active(cpu)) {
3854 			error = ENXIO;
3855 			break;
3856 		}
3857 
3858 		/*
3859 		 * Refuse an allocation for a system-wide PMC if this
3860 		 * process has been jailed, or if this process lacks
3861 		 * super-user credentials and the sysctl tunable
3862 		 * 'security.bsd.unprivileged_syspmcs' is zero.
3863 		 */
3864 
3865 		if (PMC_IS_SYSTEM_MODE(mode)) {
3866 			if (jailed(curthread->td_ucred)) {
3867 				error = EPERM;
3868 				break;
3869 			}
3870 			if (!pmc_unprivileged_syspmcs) {
3871 				error = priv_check(curthread,
3872 				    PRIV_PMC_SYSTEM);
3873 				if (error)
3874 					break;
3875 			}
3876 		}
3877 
3878 		/*
3879 		 * Look for valid values for 'pm_flags'
3880 		 */
3881 
3882 		if ((pa.pm_flags & ~(PMC_F_DESCENDANTS | PMC_F_LOG_PROCCSW |
3883 		    PMC_F_LOG_PROCEXIT | PMC_F_CALLCHAIN |
3884 		    PMC_F_USERCALLCHAIN)) != 0) {
3885 			error = EINVAL;
3886 			break;
3887 		}
3888 
3889 		/* PMC_F_USERCALLCHAIN is only valid with PMC_F_CALLCHAIN */
3890 		if ((pa.pm_flags & (PMC_F_CALLCHAIN | PMC_F_USERCALLCHAIN)) ==
3891 		    PMC_F_USERCALLCHAIN) {
3892 			error = EINVAL;
3893 			break;
3894 		}
3895 
3896 		/* PMC_F_USERCALLCHAIN is only valid for sampling mode */
3897 		if (pa.pm_flags & PMC_F_USERCALLCHAIN &&
3898 			mode != PMC_MODE_TS && mode != PMC_MODE_SS) {
3899 			error = EINVAL;
3900 			break;
3901 		}
3902 
3903 		/* process logging options are not allowed for system PMCs */
3904 		if (PMC_IS_SYSTEM_MODE(mode) && (pa.pm_flags &
3905 		    (PMC_F_LOG_PROCCSW | PMC_F_LOG_PROCEXIT))) {
3906 			error = EINVAL;
3907 			break;
3908 		}
3909 
3910 		/*
3911 		 * All sampling mode PMCs need to be able to interrupt the
3912 		 * CPU.
3913 		 */
3914 		if (PMC_IS_SAMPLING_MODE(mode))
3915 			caps |= PMC_CAP_INTERRUPT;
3916 
3917 		/* A valid class specifier should have been passed in. */
3918 		pcd = pmc_class_to_classdep(pa.pm_class);
3919 		if (pcd == NULL) {
3920 			error = EINVAL;
3921 			break;
3922 		}
3923 
3924 		/* The requested PMC capabilities should be feasible. */
3925 		if ((pcd->pcd_caps & caps) != caps) {
3926 			error = EOPNOTSUPP;
3927 			break;
3928 		}
3929 
3930 		PMCDBG4(PMC,ALL,2, "event=%d caps=0x%x mode=%d cpu=%d",
3931 		    pa.pm_ev, caps, mode, cpu);
3932 
3933 		pmc = pmc_allocate_pmc_descriptor();
3934 		pmc->pm_id    = PMC_ID_MAKE_ID(cpu,pa.pm_mode,pa.pm_class,
3935 		    PMC_ID_INVALID);
3936 		pmc->pm_event = pa.pm_ev;
3937 		pmc->pm_state = PMC_STATE_FREE;
3938 		pmc->pm_caps  = caps;
3939 		pmc->pm_flags = pa.pm_flags;
3940 
3941 		/* XXX set lower bound on sampling for process counters */
3942 		if (PMC_IS_SAMPLING_MODE(mode))
3943 			pmc->pm_sc.pm_reloadcount = pa.pm_count;
3944 		else
3945 			pmc->pm_sc.pm_initial = pa.pm_count;
3946 
3947 		/* switch thread to CPU 'cpu' */
3948 		pmc_save_cpu_binding(&pb);
3949 
3950 #define	PMC_IS_SHAREABLE_PMC(cpu, n)				\
3951 	(pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_state &		\
3952 	 PMC_PHW_FLAG_IS_SHAREABLE)
3953 #define	PMC_IS_UNALLOCATED(cpu, n)				\
3954 	(pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_pmc == NULL)
3955 
3956 		if (PMC_IS_SYSTEM_MODE(mode)) {
3957 			pmc_select_cpu(cpu);
3958 			for (n = pcd->pcd_ri; n < (int) md->pmd_npmc; n++) {
3959 				pcd = pmc_ri_to_classdep(md, n, &adjri);
3960 				if (pmc_can_allocate_row(n, mode) == 0 &&
3961 				    pmc_can_allocate_rowindex(
3962 					    curthread->td_proc, n, cpu) == 0 &&
3963 				    (PMC_IS_UNALLOCATED(cpu, n) ||
3964 				     PMC_IS_SHAREABLE_PMC(cpu, n)) &&
3965 				    pcd->pcd_allocate_pmc(cpu, adjri, pmc,
3966 					&pa) == 0)
3967 					break;
3968 			}
3969 		} else {
3970 			/* Process virtual mode */
3971 			for (n = pcd->pcd_ri; n < (int) md->pmd_npmc; n++) {
3972 				pcd = pmc_ri_to_classdep(md, n, &adjri);
3973 				if (pmc_can_allocate_row(n, mode) == 0 &&
3974 				    pmc_can_allocate_rowindex(
3975 					    curthread->td_proc, n,
3976 					    PMC_CPU_ANY) == 0 &&
3977 				    pcd->pcd_allocate_pmc(curthread->td_oncpu,
3978 					adjri, pmc, &pa) == 0)
3979 					break;
3980 			}
3981 		}
3982 
3983 #undef	PMC_IS_UNALLOCATED
3984 #undef	PMC_IS_SHAREABLE_PMC
3985 
3986 		pmc_restore_cpu_binding(&pb);
3987 
3988 		if (n == (int) md->pmd_npmc) {
3989 			pmc_destroy_pmc_descriptor(pmc);
3990 			pmc = NULL;
3991 			error = EINVAL;
3992 			break;
3993 		}
3994 
3995 		/* Fill in the correct value in the ID field */
3996 		pmc->pm_id = PMC_ID_MAKE_ID(cpu,mode,pa.pm_class,n);
3997 
3998 		PMCDBG5(PMC,ALL,2, "ev=%d class=%d mode=%d n=%d -> pmcid=%x",
3999 		    pmc->pm_event, pa.pm_class, mode, n, pmc->pm_id);
4000 
4001 		/* Process mode PMCs with logging enabled need log files */
4002 		if (pmc->pm_flags & (PMC_F_LOG_PROCEXIT | PMC_F_LOG_PROCCSW))
4003 			pmc->pm_flags |= PMC_F_NEEDS_LOGFILE;
4004 
4005 		/* All system mode sampling PMCs require a log file */
4006 		if (PMC_IS_SAMPLING_MODE(mode) && PMC_IS_SYSTEM_MODE(mode))
4007 			pmc->pm_flags |= PMC_F_NEEDS_LOGFILE;
4008 
4009 		/*
4010 		 * Configure global pmc's immediately
4011 		 */
4012 
4013 		if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pmc))) {
4014 
4015 			pmc_save_cpu_binding(&pb);
4016 			pmc_select_cpu(cpu);
4017 
4018 			phw = pmc_pcpu[cpu]->pc_hwpmcs[n];
4019 			pcd = pmc_ri_to_classdep(md, n, &adjri);
4020 
4021 			if ((phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0 ||
4022 			    (error = pcd->pcd_config_pmc(cpu, adjri, pmc)) != 0) {
4023 				(void) pcd->pcd_release_pmc(cpu, adjri, pmc);
4024 				pmc_destroy_pmc_descriptor(pmc);
4025 				pmc = NULL;
4026 				pmc_restore_cpu_binding(&pb);
4027 				error = EPERM;
4028 				break;
4029 			}
4030 
4031 			pmc_restore_cpu_binding(&pb);
4032 		}
4033 
4034 		pmc->pm_state    = PMC_STATE_ALLOCATED;
4035 		pmc->pm_class	= pa.pm_class;
4036 
4037 		/*
4038 		 * mark row disposition
4039 		 */
4040 
4041 		if (PMC_IS_SYSTEM_MODE(mode))
4042 			PMC_MARK_ROW_STANDALONE(n);
4043 		else
4044 			PMC_MARK_ROW_THREAD(n);
4045 
4046 		/*
4047 		 * Register this PMC with the current thread as its owner.
4048 		 */
4049 
4050 		if ((error =
4051 		    pmc_register_owner(curthread->td_proc, pmc)) != 0) {
4052 			pmc_release_pmc_descriptor(pmc);
4053 			pmc_destroy_pmc_descriptor(pmc);
4054 			pmc = NULL;
4055 			break;
4056 		}
4057 
4058 
4059 		/*
4060 		 * Return the allocated index.
4061 		 */
4062 
4063 		pa.pm_pmcid = pmc->pm_id;
4064 
4065 		error = copyout(&pa, arg, sizeof(pa));
4066 	}
4067 	break;
4068 
4069 
4070 	/*
4071 	 * Attach a PMC to a process.
4072 	 */
4073 
4074 	case PMC_OP_PMCATTACH:
4075 	{
4076 		struct pmc *pm;
4077 		struct proc *p;
4078 		struct pmc_op_pmcattach a;
4079 
4080 		sx_assert(&pmc_sx, SX_XLOCKED);
4081 
4082 		if ((error = copyin(arg, &a, sizeof(a))) != 0)
4083 			break;
4084 
4085 		if (a.pm_pid < 0) {
4086 			error = EINVAL;
4087 			break;
4088 		} else if (a.pm_pid == 0)
4089 			a.pm_pid = td->td_proc->p_pid;
4090 
4091 		if ((error = pmc_find_pmc(a.pm_pmc, &pm)) != 0)
4092 			break;
4093 
4094 		if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm))) {
4095 			error = EINVAL;
4096 			break;
4097 		}
4098 
4099 		/* PMCs may be (re)attached only when allocated or stopped */
4100 		if (pm->pm_state == PMC_STATE_RUNNING) {
4101 			error = EBUSY;
4102 			break;
4103 		} else if (pm->pm_state != PMC_STATE_ALLOCATED &&
4104 		    pm->pm_state != PMC_STATE_STOPPED) {
4105 			error = EINVAL;
4106 			break;
4107 		}
4108 
4109 		/* lookup pid */
4110 		if ((p = pfind(a.pm_pid)) == NULL) {
4111 			error = ESRCH;
4112 			break;
4113 		}
4114 
4115 		/*
4116 		 * Ignore processes that are working on exiting.
4117 		 */
4118 		if (p->p_flag & P_WEXIT) {
4119 			error = ESRCH;
4120 			PROC_UNLOCK(p);	/* pfind() returns a locked process */
4121 			break;
4122 		}
4123 
4124 		/*
4125 		 * we are allowed to attach a PMC to a process if
4126 		 * we can debug it.
4127 		 */
4128 		error = p_candebug(curthread, p);
4129 
4130 		PROC_UNLOCK(p);
4131 
4132 		if (error == 0)
4133 			error = pmc_attach_process(p, pm);
4134 	}
4135 	break;
4136 
4137 
4138 	/*
4139 	 * Detach an attached PMC from a process.
4140 	 */
4141 
4142 	case PMC_OP_PMCDETACH:
4143 	{
4144 		struct pmc *pm;
4145 		struct proc *p;
4146 		struct pmc_op_pmcattach a;
4147 
4148 		if ((error = copyin(arg, &a, sizeof(a))) != 0)
4149 			break;
4150 
4151 		if (a.pm_pid < 0) {
4152 			error = EINVAL;
4153 			break;
4154 		} else if (a.pm_pid == 0)
4155 			a.pm_pid = td->td_proc->p_pid;
4156 
4157 		if ((error = pmc_find_pmc(a.pm_pmc, &pm)) != 0)
4158 			break;
4159 
4160 		if ((p = pfind(a.pm_pid)) == NULL) {
4161 			error = ESRCH;
4162 			break;
4163 		}
4164 
4165 		/*
4166 		 * Treat processes that are in the process of exiting
4167 		 * as if they were not present.
4168 		 */
4169 
4170 		if (p->p_flag & P_WEXIT)
4171 			error = ESRCH;
4172 
4173 		PROC_UNLOCK(p);	/* pfind() returns a locked process */
4174 
4175 		if (error == 0)
4176 			error = pmc_detach_process(p, pm);
4177 	}
4178 	break;
4179 
4180 
4181 	/*
4182 	 * Retrieve the MSR number associated with the counter
4183 	 * 'pmc_id'.  This allows processes to directly use RDPMC
4184 	 * instructions to read their PMCs, without the overhead of a
4185 	 * system call.
4186 	 */
4187 
4188 	case PMC_OP_PMCGETMSR:
4189 	{
4190 		int adjri, ri;
4191 		struct pmc *pm;
4192 		struct pmc_target *pt;
4193 		struct pmc_op_getmsr gm;
4194 		struct pmc_classdep *pcd;
4195 
4196 		PMC_DOWNGRADE_SX();
4197 
4198 		if ((error = copyin(arg, &gm, sizeof(gm))) != 0)
4199 			break;
4200 
4201 		if ((error = pmc_find_pmc(gm.pm_pmcid, &pm)) != 0)
4202 			break;
4203 
4204 		/*
4205 		 * The allocated PMC has to be a process virtual PMC,
4206 		 * i.e., of type MODE_T[CS].  Global PMCs can only be
4207 		 * read using the PMCREAD operation since they may be
4208 		 * allocated on a different CPU than the one we could
4209 		 * be running on at the time of the RDPMC instruction.
4210 		 *
4211 		 * The GETMSR operation is not allowed for PMCs that
4212 		 * are inherited across processes.
4213 		 */
4214 
4215 		if (!PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)) ||
4216 		    (pm->pm_flags & PMC_F_DESCENDANTS)) {
4217 			error = EINVAL;
4218 			break;
4219 		}
4220 
4221 		/*
4222 		 * It only makes sense to use a RDPMC (or its
4223 		 * equivalent instruction on non-x86 architectures) on
4224 		 * a process that has allocated and attached a PMC to
4225 		 * itself.  Conversely the PMC is only allowed to have
4226 		 * one process attached to it -- its owner.
4227 		 */
4228 
4229 		if ((pt = LIST_FIRST(&pm->pm_targets)) == NULL ||
4230 		    LIST_NEXT(pt, pt_next) != NULL ||
4231 		    pt->pt_process->pp_proc != pm->pm_owner->po_owner) {
4232 			error = EINVAL;
4233 			break;
4234 		}
4235 
4236 		ri = PMC_TO_ROWINDEX(pm);
4237 		pcd = pmc_ri_to_classdep(md, ri, &adjri);
4238 
4239 		/* PMC class has no 'GETMSR' support */
4240 		if (pcd->pcd_get_msr == NULL) {
4241 			error = ENOSYS;
4242 			break;
4243 		}
4244 
4245 		if ((error = (*pcd->pcd_get_msr)(adjri, &gm.pm_msr)) < 0)
4246 			break;
4247 
4248 		if ((error = copyout(&gm, arg, sizeof(gm))) < 0)
4249 			break;
4250 
4251 		/*
4252 		 * Mark our process as using MSRs.  Update machine
4253 		 * state using a forced context switch.
4254 		 */
4255 
4256 		pt->pt_process->pp_flags |= PMC_PP_ENABLE_MSR_ACCESS;
4257 		pmc_force_context_switch();
4258 
4259 	}
4260 	break;
4261 
4262 	/*
4263 	 * Release an allocated PMC
4264 	 */
4265 
4266 	case PMC_OP_PMCRELEASE:
4267 	{
4268 		pmc_id_t pmcid;
4269 		struct pmc *pm;
4270 		struct pmc_owner *po;
4271 		struct pmc_op_simple sp;
4272 
4273 		/*
4274 		 * Find PMC pointer for the named PMC.
4275 		 *
4276 		 * Use pmc_release_pmc_descriptor() to switch off the
4277 		 * PMC, remove all its target threads, and remove the
4278 		 * PMC from its owner's list.
4279 		 *
4280 		 * Remove the owner record if this is the last PMC
4281 		 * owned.
4282 		 *
4283 		 * Free up space.
4284 		 */
4285 
4286 		if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
4287 			break;
4288 
4289 		pmcid = sp.pm_pmcid;
4290 
4291 		if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
4292 			break;
4293 
4294 		po = pm->pm_owner;
4295 		pmc_release_pmc_descriptor(pm);
4296 		pmc_maybe_remove_owner(po);
4297 		pmc_destroy_pmc_descriptor(pm);
4298 	}
4299 	break;
4300 
4301 
4302 	/*
4303 	 * Read and/or write a PMC.
4304 	 */
4305 
4306 	case PMC_OP_PMCRW:
4307 	{
4308 		int adjri;
4309 		struct pmc *pm;
4310 		uint32_t cpu, ri;
4311 		pmc_value_t oldvalue;
4312 		struct pmc_binding pb;
4313 		struct pmc_op_pmcrw prw;
4314 		struct pmc_classdep *pcd;
4315 		struct pmc_op_pmcrw *pprw;
4316 
4317 		PMC_DOWNGRADE_SX();
4318 
4319 		if ((error = copyin(arg, &prw, sizeof(prw))) != 0)
4320 			break;
4321 
4322 		ri = 0;
4323 		PMCDBG2(PMC,OPS,1, "rw id=%d flags=0x%x", prw.pm_pmcid,
4324 		    prw.pm_flags);
4325 
4326 		/* must have at least one flag set */
4327 		if ((prw.pm_flags & (PMC_F_OLDVALUE|PMC_F_NEWVALUE)) == 0) {
4328 			error = EINVAL;
4329 			break;
4330 		}
4331 
4332 		/* locate pmc descriptor */
4333 		if ((error = pmc_find_pmc(prw.pm_pmcid, &pm)) != 0)
4334 			break;
4335 
4336 		/* Can't read a PMC that hasn't been started. */
4337 		if (pm->pm_state != PMC_STATE_ALLOCATED &&
4338 		    pm->pm_state != PMC_STATE_STOPPED &&
4339 		    pm->pm_state != PMC_STATE_RUNNING) {
4340 			error = EINVAL;
4341 			break;
4342 		}
4343 
4344 		/* writing a new value is allowed only for 'STOPPED' pmcs */
4345 		if (pm->pm_state == PMC_STATE_RUNNING &&
4346 		    (prw.pm_flags & PMC_F_NEWVALUE)) {
4347 			error = EBUSY;
4348 			break;
4349 		}
4350 
4351 		if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) {
4352 
4353 			/*
4354 			 * If this PMC is attached to its owner (i.e.,
4355 			 * the process requesting this operation) and
4356 			 * is running, then attempt to get an
4357 			 * upto-date reading from hardware for a READ.
4358 			 * Writes are only allowed when the PMC is
4359 			 * stopped, so only update the saved value
4360 			 * field.
4361 			 *
4362 			 * If the PMC is not running, or is not
4363 			 * attached to its owner, read/write to the
4364 			 * savedvalue field.
4365 			 */
4366 
4367 			ri = PMC_TO_ROWINDEX(pm);
4368 			pcd = pmc_ri_to_classdep(md, ri, &adjri);
4369 
4370 			mtx_pool_lock_spin(pmc_mtxpool, pm);
4371 			cpu = curthread->td_oncpu;
4372 
4373 			if (prw.pm_flags & PMC_F_OLDVALUE) {
4374 				if ((pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) &&
4375 				    (pm->pm_state == PMC_STATE_RUNNING))
4376 					error = (*pcd->pcd_read_pmc)(cpu, adjri,
4377 					    &oldvalue);
4378 				else
4379 					oldvalue = pm->pm_gv.pm_savedvalue;
4380 			}
4381 			if (prw.pm_flags & PMC_F_NEWVALUE)
4382 				pm->pm_gv.pm_savedvalue = prw.pm_value;
4383 
4384 			mtx_pool_unlock_spin(pmc_mtxpool, pm);
4385 
4386 		} else { /* System mode PMCs */
4387 			cpu = PMC_TO_CPU(pm);
4388 			ri  = PMC_TO_ROWINDEX(pm);
4389 			pcd = pmc_ri_to_classdep(md, ri, &adjri);
4390 
4391 			if (!pmc_cpu_is_active(cpu)) {
4392 				error = ENXIO;
4393 				break;
4394 			}
4395 
4396 			/* move this thread to CPU 'cpu' */
4397 			pmc_save_cpu_binding(&pb);
4398 			pmc_select_cpu(cpu);
4399 
4400 			critical_enter();
4401 			/* save old value */
4402 			if (prw.pm_flags & PMC_F_OLDVALUE)
4403 				if ((error = (*pcd->pcd_read_pmc)(cpu, adjri,
4404 					 &oldvalue)))
4405 					goto error;
4406 			/* write out new value */
4407 			if (prw.pm_flags & PMC_F_NEWVALUE)
4408 				error = (*pcd->pcd_write_pmc)(cpu, adjri,
4409 				    prw.pm_value);
4410 		error:
4411 			critical_exit();
4412 			pmc_restore_cpu_binding(&pb);
4413 			if (error)
4414 				break;
4415 		}
4416 
4417 		pprw = (struct pmc_op_pmcrw *) arg;
4418 
4419 #ifdef	HWPMC_DEBUG
4420 		if (prw.pm_flags & PMC_F_NEWVALUE)
4421 			PMCDBG3(PMC,OPS,2, "rw id=%d new %jx -> old %jx",
4422 			    ri, prw.pm_value, oldvalue);
4423 		else if (prw.pm_flags & PMC_F_OLDVALUE)
4424 			PMCDBG2(PMC,OPS,2, "rw id=%d -> old %jx", ri, oldvalue);
4425 #endif
4426 
4427 		/* return old value if requested */
4428 		if (prw.pm_flags & PMC_F_OLDVALUE)
4429 			if ((error = copyout(&oldvalue, &pprw->pm_value,
4430 				 sizeof(prw.pm_value))))
4431 				break;
4432 
4433 	}
4434 	break;
4435 
4436 
4437 	/*
4438 	 * Set the sampling rate for a sampling mode PMC and the
4439 	 * initial count for a counting mode PMC.
4440 	 */
4441 
4442 	case PMC_OP_PMCSETCOUNT:
4443 	{
4444 		struct pmc *pm;
4445 		struct pmc_op_pmcsetcount sc;
4446 
4447 		PMC_DOWNGRADE_SX();
4448 
4449 		if ((error = copyin(arg, &sc, sizeof(sc))) != 0)
4450 			break;
4451 
4452 		if ((error = pmc_find_pmc(sc.pm_pmcid, &pm)) != 0)
4453 			break;
4454 
4455 		if (pm->pm_state == PMC_STATE_RUNNING) {
4456 			error = EBUSY;
4457 			break;
4458 		}
4459 
4460 		if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
4461 			pm->pm_sc.pm_reloadcount = sc.pm_count;
4462 		else
4463 			pm->pm_sc.pm_initial = sc.pm_count;
4464 	}
4465 	break;
4466 
4467 
4468 	/*
4469 	 * Start a PMC.
4470 	 */
4471 
4472 	case PMC_OP_PMCSTART:
4473 	{
4474 		pmc_id_t pmcid;
4475 		struct pmc *pm;
4476 		struct pmc_op_simple sp;
4477 
4478 		sx_assert(&pmc_sx, SX_XLOCKED);
4479 
4480 		if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
4481 			break;
4482 
4483 		pmcid = sp.pm_pmcid;
4484 
4485 		if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
4486 			break;
4487 
4488 		KASSERT(pmcid == pm->pm_id,
4489 		    ("[pmc,%d] pmcid %x != id %x", __LINE__,
4490 			pm->pm_id, pmcid));
4491 
4492 		if (pm->pm_state == PMC_STATE_RUNNING) /* already running */
4493 			break;
4494 		else if (pm->pm_state != PMC_STATE_STOPPED &&
4495 		    pm->pm_state != PMC_STATE_ALLOCATED) {
4496 			error = EINVAL;
4497 			break;
4498 		}
4499 
4500 		error = pmc_start(pm);
4501 	}
4502 	break;
4503 
4504 
4505 	/*
4506 	 * Stop a PMC.
4507 	 */
4508 
4509 	case PMC_OP_PMCSTOP:
4510 	{
4511 		pmc_id_t pmcid;
4512 		struct pmc *pm;
4513 		struct pmc_op_simple sp;
4514 
4515 		PMC_DOWNGRADE_SX();
4516 
4517 		if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
4518 			break;
4519 
4520 		pmcid = sp.pm_pmcid;
4521 
4522 		/*
4523 		 * Mark the PMC as inactive and invoke the MD stop
4524 		 * routines if needed.
4525 		 */
4526 
4527 		if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
4528 			break;
4529 
4530 		KASSERT(pmcid == pm->pm_id,
4531 		    ("[pmc,%d] pmc id %x != pmcid %x", __LINE__,
4532 			pm->pm_id, pmcid));
4533 
4534 		if (pm->pm_state == PMC_STATE_STOPPED) /* already stopped */
4535 			break;
4536 		else if (pm->pm_state != PMC_STATE_RUNNING) {
4537 			error = EINVAL;
4538 			break;
4539 		}
4540 
4541 		error = pmc_stop(pm);
4542 	}
4543 	break;
4544 
4545 
4546 	/*
4547 	 * Write a user supplied value to the log file.
4548 	 */
4549 
4550 	case PMC_OP_WRITELOG:
4551 	{
4552 		struct pmc_op_writelog wl;
4553 		struct pmc_owner *po;
4554 
4555 		PMC_DOWNGRADE_SX();
4556 
4557 		if ((error = copyin(arg, &wl, sizeof(wl))) != 0)
4558 			break;
4559 
4560 		if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) {
4561 			error = EINVAL;
4562 			break;
4563 		}
4564 
4565 		if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) {
4566 			error = EINVAL;
4567 			break;
4568 		}
4569 
4570 		error = pmclog_process_userlog(po, &wl);
4571 	}
4572 	break;
4573 
4574 
4575 	default:
4576 		error = EINVAL;
4577 		break;
4578 	}
4579 
4580 	if (is_sx_downgraded)
4581 		sx_sunlock(&pmc_sx);
4582 	else
4583 		sx_xunlock(&pmc_sx);
4584 done_syscall:
4585 	if (error)
4586 		counter_u64_add(pmc_stats.pm_syscall_errors, 1);
4587 
4588 	return (error);
4589 }
4590 
4591 /*
4592  * Helper functions
4593  */
4594 
4595 
4596 /*
4597  * Mark the thread as needing callchain capture and post an AST.  The
4598  * actual callchain capture will be done in a context where it is safe
4599  * to take page faults.
4600  */
4601 
4602 static void
4603 pmc_post_callchain_callback(void)
4604 {
4605 	struct thread *td;
4606 
4607 	td = curthread;
4608 
4609 	/*
4610 	 * If there is multiple PMCs for the same interrupt ignore new post
4611 	 */
4612 	if (td->td_pflags & TDP_CALLCHAIN)
4613 		return;
4614 
4615 	/*
4616 	 * Mark this thread as needing callchain capture.
4617 	 * `td->td_pflags' will be safe to touch because this thread
4618 	 * was in user space when it was interrupted.
4619 	 */
4620 	td->td_pflags |= TDP_CALLCHAIN;
4621 
4622 	/*
4623 	 * Don't let this thread migrate between CPUs until callchain
4624 	 * capture completes.
4625 	 */
4626 	sched_pin();
4627 
4628 	return;
4629 }
4630 
4631 /*
4632  * Find a free slot in the per-cpu array of samples and capture the
4633  * current callchain there.  If a sample was successfully added, a bit
4634  * is set in mask 'pmc_cpumask' denoting that the DO_SAMPLES hook
4635  * needs to be invoked from the clock handler.
4636  *
4637  * This function is meant to be called from an NMI handler.  It cannot
4638  * use any of the locking primitives supplied by the OS.
4639  */
4640 
4641 static int
4642 pmc_add_sample(int ring, struct pmc *pm, struct trapframe *tf)
4643 {
4644 	int error, cpu, callchaindepth, inuserspace;
4645 	struct thread *td;
4646 	struct pmc_sample *ps;
4647 	struct pmc_samplebuffer *psb;
4648 
4649 	error = 0;
4650 
4651 	/*
4652 	 * Allocate space for a sample buffer.
4653 	 */
4654 	cpu = curcpu;
4655 	psb = pmc_pcpu[cpu]->pc_sb[ring];
4656 	inuserspace = TRAPF_USERMODE(tf);
4657 	ps = psb->ps_write;
4658 	if (ps->ps_nsamples == PMC_SAMPLE_INUSE) {
4659 		counter_u64_add(ps->ps_pmc->pm_runcount, -1);
4660 		counter_u64_add(pmc_stats.pm_overwrites, 1);
4661 		ps->ps_nsamples = 0;
4662 	} else if (ps->ps_nsamples) {	/* in use, reader hasn't caught up */
4663 		pm->pm_pcpu_state[cpu].pps_stalled = 1;
4664 		counter_u64_add(pmc_stats.pm_intr_bufferfull, 1);
4665 		PMCDBG6(SAM,INT,1,"(spc) cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d",
4666 		    cpu, pm, (void *) tf, inuserspace,
4667 		    (int) (psb->ps_write - psb->ps_samples),
4668 		    (int) (psb->ps_read - psb->ps_samples));
4669 		callchaindepth = 1;
4670 		error = ENOMEM;
4671 		goto done;
4672 	}
4673 
4674 	/* Fill in entry. */
4675 	PMCDBG6(SAM,INT,1,"cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d", cpu, pm,
4676 	    (void *) tf, inuserspace,
4677 	    (int) (psb->ps_write - psb->ps_samples),
4678 	    (int) (psb->ps_read - psb->ps_samples));
4679 
4680 	KASSERT(counter_u64_fetch(pm->pm_runcount) >= 0,
4681 	    ("[pmc,%d] pm=%p runcount %ld", __LINE__, (void *) pm,
4682 		 (unsigned long)counter_u64_fetch(pm->pm_runcount)));
4683 
4684 	counter_u64_add(pm->pm_runcount, 1);	/* hold onto PMC */
4685 
4686 	td = curthread;
4687 	ps->ps_pmc = pm;
4688 	ps->ps_td = td;
4689 	ps->ps_pid = td->td_proc->p_pid;
4690 	ps->ps_tid = td->td_tid;
4691 	ps->ps_tsc = pmc_rdtsc();
4692 
4693 	ps->ps_cpu = cpu;
4694 	ps->ps_flags = inuserspace ? PMC_CC_F_USERSPACE : 0;
4695 
4696 	callchaindepth = (pm->pm_flags & PMC_F_CALLCHAIN) ?
4697 	    pmc_callchaindepth : 1;
4698 
4699 	if (callchaindepth == 1)
4700 		ps->ps_pc[0] = PMC_TRAPFRAME_TO_PC(tf);
4701 	else {
4702 		/*
4703 		 * Kernel stack traversals can be done immediately,
4704 		 * while we defer to an AST for user space traversals.
4705 		 */
4706 		if (!inuserspace) {
4707 			callchaindepth =
4708 			    pmc_save_kernel_callchain(ps->ps_pc,
4709 				callchaindepth, tf);
4710 		} else {
4711 			pmc_post_callchain_callback();
4712 			callchaindepth = PMC_SAMPLE_INUSE;
4713 		}
4714 	}
4715 
4716 	ps->ps_nsamples = callchaindepth;	/* mark entry as in use */
4717 	if (ring == PMC_UR) {
4718 		ps->ps_nsamples_actual = callchaindepth;	/* mark entry as in use */
4719 		ps->ps_nsamples = PMC_SAMPLE_INUSE;
4720 	} else
4721 		ps->ps_nsamples = callchaindepth;	/* mark entry as in use */
4722 	/* increment write pointer, modulo ring buffer size */
4723 	ps++;
4724 	if (ps == psb->ps_fence)
4725 		psb->ps_write = psb->ps_samples;
4726 	else
4727 		psb->ps_write = ps;
4728 
4729  done:
4730 	/* mark CPU as needing processing */
4731 	if (callchaindepth != PMC_SAMPLE_INUSE)
4732 		DPCPU_SET(pmc_sampled, 1);
4733 
4734 	return (error);
4735 }
4736 
4737 /*
4738  * Interrupt processing.
4739  *
4740  * This function is meant to be called from an NMI handler.  It cannot
4741  * use any of the locking primitives supplied by the OS.
4742  */
4743 
4744 int
4745 pmc_process_interrupt(int ring, struct pmc *pm, struct trapframe *tf)
4746 {
4747 	struct thread *td;
4748 
4749 	td = curthread;
4750 	if ((pm->pm_flags & PMC_F_USERCALLCHAIN) &&
4751 	    (td->td_proc->p_flag & P_KPROC) == 0 &&
4752 	    !TRAPF_USERMODE(tf)) {
4753 		atomic_add_int(&td->td_pmcpend, 1);
4754 		return (pmc_add_sample(PMC_UR, pm, tf));
4755 	}
4756 	return (pmc_add_sample(ring, pm, tf));
4757 }
4758 
4759 /*
4760  * Capture a user call chain.  This function will be called from ast()
4761  * before control returns to userland and before the process gets
4762  * rescheduled.
4763  */
4764 
4765 static void
4766 pmc_capture_user_callchain(int cpu, int ring, struct trapframe *tf)
4767 {
4768 	struct pmc *pm;
4769 	struct thread *td;
4770 	struct pmc_sample *ps, *ps_end;
4771 	struct pmc_samplebuffer *psb;
4772 	int nsamples, nrecords, pass;
4773 #ifdef	INVARIANTS
4774 	int ncallchains;
4775 	int nfree;
4776 #endif
4777 
4778 	psb = pmc_pcpu[cpu]->pc_sb[ring];
4779 	td = curthread;
4780 
4781 	KASSERT(td->td_pflags & TDP_CALLCHAIN,
4782 	    ("[pmc,%d] Retrieving callchain for thread that doesn't want it",
4783 		__LINE__));
4784 
4785 #ifdef	INVARIANTS
4786 	ncallchains = 0;
4787 	nfree = 0;
4788 #endif
4789 	nrecords = INT_MAX;
4790 	pass = 0;
4791  restart:
4792 	if (ring == PMC_UR)
4793 		nrecords = atomic_readandclear_32(&td->td_pmcpend);
4794 
4795 	/*
4796 	 * Iterate through all deferred callchain requests.
4797 	 * Walk from the current read pointer to the current
4798 	 * write pointer.
4799 	 */
4800 
4801 	ps = psb->ps_read;
4802 	ps_end = psb->ps_write;
4803 	do {
4804 #ifdef	INVARIANTS
4805 		if (ps->ps_nsamples == PMC_SAMPLE_FREE) {
4806 			nfree++;
4807 			goto next;
4808 		}
4809 
4810 		if ((ps->ps_pmc == NULL) ||
4811 		    (ps->ps_pmc->pm_state != PMC_STATE_RUNNING))
4812 			nfree++;
4813 #endif
4814 		if (ps->ps_nsamples != PMC_SAMPLE_INUSE)
4815 			goto next;
4816 		if (ps->ps_td != td)
4817 			goto next;
4818 
4819 		KASSERT(ps->ps_cpu == cpu,
4820 		    ("[pmc,%d] cpu mismatch ps_cpu=%d pcpu=%d", __LINE__,
4821 			ps->ps_cpu, PCPU_GET(cpuid)));
4822 
4823 		pm = ps->ps_pmc;
4824 
4825 		KASSERT(pm->pm_flags & PMC_F_CALLCHAIN,
4826 		    ("[pmc,%d] Retrieving callchain for PMC that doesn't "
4827 			"want it", __LINE__));
4828 
4829 		KASSERT(counter_u64_fetch(pm->pm_runcount) > 0,
4830 		    ("[pmc,%d] runcount %ld", __LINE__, (unsigned long)counter_u64_fetch(pm->pm_runcount)));
4831 
4832 		if (ring == PMC_UR) {
4833 			nsamples = ps->ps_nsamples_actual;
4834 			counter_u64_add(pmc_stats.pm_merges, 1);
4835 		} else
4836 			nsamples = 0;
4837 
4838 		/*
4839 		 * Retrieve the callchain and mark the sample buffer
4840 		 * as 'processable' by the timer tick sweep code.
4841 		 */
4842 
4843 #ifdef INVARIANTS
4844 		ncallchains++;
4845 #endif
4846 
4847 		if (__predict_true(nsamples < pmc_callchaindepth - 1))
4848 			nsamples += pmc_save_user_callchain(ps->ps_pc + nsamples,
4849 		       pmc_callchaindepth - nsamples - 1, tf);
4850 		wmb();
4851 		ps->ps_nsamples = nsamples;
4852 		if (nrecords-- == 1)
4853 			break;
4854 next:
4855 		/* increment the pointer, modulo sample ring size */
4856 		if (++ps == psb->ps_fence)
4857 			ps = psb->ps_samples;
4858 	} while (ps != ps_end);
4859 	if (__predict_false(ring == PMC_UR && td->td_pmcpend)) {
4860 		if (pass == 0) {
4861 			pass = 1;
4862 			goto restart;
4863 		}
4864 		/* only collect samples for this part once */
4865 		td->td_pmcpend = 0;
4866 	}
4867 
4868 #ifdef INVARIANTS
4869 	if (ring == PMC_HR)
4870 		KASSERT(ncallchains > 0 || nfree > 0,
4871 		    ("[pmc,%d] cpu %d didn't find a sample to collect", __LINE__,
4872 			    cpu));
4873 #endif
4874 
4875 	/* mark CPU as needing processing */
4876 	DPCPU_SET(pmc_sampled, 1);
4877 }
4878 
4879 
4880 static void
4881 pmc_flush_ring(int cpu, int ring)
4882 {
4883 	struct pmc *pm;
4884 	struct pmc_sample *ps;
4885 	struct pmc_samplebuffer *psb;
4886 	int n;
4887 
4888 	psb = pmc_pcpu[cpu]->pc_sb[ring];
4889 
4890 	for (n = 0; n < pmc_nsamples; n++) { /* bound on #iterations */
4891 
4892 		ps = psb->ps_read;
4893 		if (ps->ps_nsamples == PMC_SAMPLE_FREE)
4894 			goto next;
4895 		pm = ps->ps_pmc;
4896 		counter_u64_add(pm->pm_runcount, -1);
4897 		ps->ps_nsamples = PMC_SAMPLE_FREE;
4898 		/* increment read pointer, modulo sample size */
4899 	next:
4900 		if (++ps == psb->ps_fence)
4901 			psb->ps_read = psb->ps_samples;
4902 		else
4903 			psb->ps_read = ps;
4904 	}
4905 }
4906 
4907 void
4908 pmc_flush_samples(int cpu)
4909 {
4910 	int n;
4911 
4912 	for (n = 0; n < PMC_NUM_SR; n++)
4913 		pmc_flush_ring(cpu, n);
4914 }
4915 
4916 
4917 /*
4918  * Process saved PC samples.
4919  */
4920 
4921 static void
4922 pmc_process_samples(int cpu, int ring)
4923 {
4924 	struct pmc *pm;
4925 	int adjri, n;
4926 	struct thread *td;
4927 	struct pmc_owner *po;
4928 	struct pmc_sample *ps;
4929 	struct pmc_classdep *pcd;
4930 	struct pmc_samplebuffer *psb;
4931 
4932 	KASSERT(PCPU_GET(cpuid) == cpu,
4933 	    ("[pmc,%d] not on the correct CPU pcpu=%d cpu=%d", __LINE__,
4934 		PCPU_GET(cpuid), cpu));
4935 
4936 	psb = pmc_pcpu[cpu]->pc_sb[ring];
4937 
4938 	for (n = 0; n < pmc_nsamples; n++) { /* bound on #iterations */
4939 
4940 		ps = psb->ps_read;
4941 		if (ps->ps_nsamples == PMC_SAMPLE_FREE)
4942 			break;
4943 
4944 		pm = ps->ps_pmc;
4945 
4946 		KASSERT(counter_u64_fetch(pm->pm_runcount) > 0,
4947 		    ("[pmc,%d] pm=%p runcount %ld", __LINE__, (void *) pm,
4948 			 (unsigned long)counter_u64_fetch(pm->pm_runcount)));
4949 
4950 		po = pm->pm_owner;
4951 
4952 		KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)),
4953 		    ("[pmc,%d] pmc=%p non-sampling mode=%d", __LINE__,
4954 			pm, PMC_TO_MODE(pm)));
4955 
4956 		/* Ignore PMCs that have been switched off */
4957 		if (pm->pm_state != PMC_STATE_RUNNING)
4958 			goto entrydone;
4959 
4960 		/* If there is a pending AST wait for completion */
4961 		if (ps->ps_nsamples == PMC_SAMPLE_INUSE) {
4962 			/* Need a rescan at a later time. */
4963 			DPCPU_SET(pmc_sampled, 1);
4964 			break;
4965 		}
4966 
4967 		PMCDBG6(SAM,OPS,1,"cpu=%d pm=%p n=%d fl=%x wr=%d rd=%d", cpu,
4968 		    pm, ps->ps_nsamples, ps->ps_flags,
4969 		    (int) (psb->ps_write - psb->ps_samples),
4970 		    (int) (psb->ps_read - psb->ps_samples));
4971 
4972 		/*
4973 		 * If this is a process-mode PMC that is attached to
4974 		 * its owner, and if the PC is in user mode, update
4975 		 * profiling statistics like timer-based profiling
4976 		 * would have done.
4977 		 *
4978 		 * Otherwise, this is either a sampling-mode PMC that
4979 		 * is attached to a different process than its owner,
4980 		 * or a system-wide sampling PMC. Dispatch a log
4981 		 * entry to the PMC's owner process.
4982 		 */
4983 		if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) {
4984 			if (ps->ps_flags & PMC_CC_F_USERSPACE) {
4985 				td = FIRST_THREAD_IN_PROC(po->po_owner);
4986 				addupc_intr(td, ps->ps_pc[0], 1);
4987 			}
4988 		} else
4989 			pmclog_process_callchain(pm, ps);
4990 
4991 	entrydone:
4992 		ps->ps_nsamples = 0; /* mark entry as free */
4993 		counter_u64_add(pm->pm_runcount, -1);
4994 
4995 		/* increment read pointer, modulo sample size */
4996 		if (++ps == psb->ps_fence)
4997 			psb->ps_read = psb->ps_samples;
4998 		else
4999 			psb->ps_read = ps;
5000 	}
5001 
5002 	counter_u64_add(pmc_stats.pm_log_sweeps, 1);
5003 
5004 	/* Do not re-enable stalled PMCs if we failed to process any samples */
5005 	if (n == 0)
5006 		return;
5007 
5008 	/*
5009 	 * Restart any stalled sampling PMCs on this CPU.
5010 	 *
5011 	 * If the NMI handler sets the pm_stalled field of a PMC after
5012 	 * the check below, we'll end up processing the stalled PMC at
5013 	 * the next hardclock tick.
5014 	 */
5015 	for (n = 0; n < md->pmd_npmc; n++) {
5016 		pcd = pmc_ri_to_classdep(md, n, &adjri);
5017 		KASSERT(pcd != NULL,
5018 		    ("[pmc,%d] null pcd ri=%d", __LINE__, n));
5019 		(void) (*pcd->pcd_get_config)(cpu,adjri,&pm);
5020 
5021 		if (pm == NULL ||			 /* !cfg'ed */
5022 		    pm->pm_state != PMC_STATE_RUNNING || /* !active */
5023 		    !PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) || /* !sampling */
5024 			!pm->pm_pcpu_state[cpu].pps_cpustate  || /* !desired */
5025 		    !pm->pm_pcpu_state[cpu].pps_stalled) /* !stalled */
5026 			continue;
5027 
5028 		pm->pm_pcpu_state[cpu].pps_stalled = 0;
5029 		(*pcd->pcd_start_pmc)(cpu, adjri);
5030 	}
5031 }
5032 
5033 /*
5034  * Event handlers.
5035  */
5036 
5037 /*
5038  * Handle a process exit.
5039  *
5040  * Remove this process from all hash tables.  If this process
5041  * owned any PMCs, turn off those PMCs and deallocate them,
5042  * removing any associations with target processes.
5043  *
5044  * This function will be called by the last 'thread' of a
5045  * process.
5046  *
5047  * XXX This eventhandler gets called early in the exit process.
5048  * Consider using a 'hook' invocation from thread_exit() or equivalent
5049  * spot.  Another negative is that kse_exit doesn't seem to call
5050  * exit1() [??].
5051  *
5052  */
5053 
5054 static void
5055 pmc_process_exit(void *arg __unused, struct proc *p)
5056 {
5057 	struct pmc *pm;
5058 	int adjri, cpu;
5059 	unsigned int ri;
5060 	int is_using_hwpmcs;
5061 	struct pmc_owner *po;
5062 	struct pmc_process *pp;
5063 	struct pmc_classdep *pcd;
5064 	pmc_value_t newvalue, tmp;
5065 
5066 	PROC_LOCK(p);
5067 	is_using_hwpmcs = p->p_flag & P_HWPMC;
5068 	PROC_UNLOCK(p);
5069 
5070 	/*
5071 	 * Log a sysexit event to all SS PMC owners.
5072 	 */
5073 	epoch_enter_preempt(global_epoch_preempt);
5074 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
5075 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
5076 		    pmclog_process_sysexit(po, p->p_pid);
5077 	epoch_exit_preempt(global_epoch_preempt);
5078 
5079 	if (!is_using_hwpmcs)
5080 		return;
5081 
5082 	PMC_GET_SX_XLOCK();
5083 	PMCDBG3(PRC,EXT,1,"process-exit proc=%p (%d, %s)", p, p->p_pid,
5084 	    p->p_comm);
5085 
5086 	/*
5087 	 * Since this code is invoked by the last thread in an exiting
5088 	 * process, we would have context switched IN at some prior
5089 	 * point.  However, with PREEMPTION, kernel mode context
5090 	 * switches may happen any time, so we want to disable a
5091 	 * context switch OUT till we get any PMCs targeting this
5092 	 * process off the hardware.
5093 	 *
5094 	 * We also need to atomically remove this process'
5095 	 * entry from our target process hash table, using
5096 	 * PMC_FLAG_REMOVE.
5097 	 */
5098 	PMCDBG3(PRC,EXT,1, "process-exit proc=%p (%d, %s)", p, p->p_pid,
5099 	    p->p_comm);
5100 
5101 	critical_enter(); /* no preemption */
5102 
5103 	cpu = curthread->td_oncpu;
5104 
5105 	if ((pp = pmc_find_process_descriptor(p,
5106 		 PMC_FLAG_REMOVE)) != NULL) {
5107 
5108 		PMCDBG2(PRC,EXT,2,
5109 		    "process-exit proc=%p pmc-process=%p", p, pp);
5110 
5111 		/*
5112 		 * The exiting process could the target of
5113 		 * some PMCs which will be running on
5114 		 * currently executing CPU.
5115 		 *
5116 		 * We need to turn these PMCs off like we
5117 		 * would do at context switch OUT time.
5118 		 */
5119 		for (ri = 0; ri < md->pmd_npmc; ri++) {
5120 
5121 			/*
5122 			 * Pick up the pmc pointer from hardware
5123 			 * state similar to the CSW_OUT code.
5124 			 */
5125 			pm = NULL;
5126 
5127 			pcd = pmc_ri_to_classdep(md, ri, &adjri);
5128 
5129 			(void) (*pcd->pcd_get_config)(cpu, adjri, &pm);
5130 
5131 			PMCDBG2(PRC,EXT,2, "ri=%d pm=%p", ri, pm);
5132 
5133 			if (pm == NULL ||
5134 			    !PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)))
5135 				continue;
5136 
5137 			PMCDBG4(PRC,EXT,2, "ppmcs[%d]=%p pm=%p "
5138 			    "state=%d", ri, pp->pp_pmcs[ri].pp_pmc,
5139 			    pm, pm->pm_state);
5140 
5141 			KASSERT(PMC_TO_ROWINDEX(pm) == ri,
5142 			    ("[pmc,%d] ri mismatch pmc(%d) ri(%d)",
5143 				__LINE__, PMC_TO_ROWINDEX(pm), ri));
5144 
5145 			KASSERT(pm == pp->pp_pmcs[ri].pp_pmc,
5146 			    ("[pmc,%d] pm %p != pp_pmcs[%d] %p",
5147 				__LINE__, pm, ri, pp->pp_pmcs[ri].pp_pmc));
5148 
5149 			KASSERT(counter_u64_fetch(pm->pm_runcount) > 0,
5150 			    ("[pmc,%d] bad runcount ri %d rc %ld",
5151 				 __LINE__, ri, (unsigned long)counter_u64_fetch(pm->pm_runcount)));
5152 
5153 			/*
5154 			 * Change desired state, and then stop if not
5155 			 * stalled. This two-step dance should avoid
5156 			 * race conditions where an interrupt re-enables
5157 			 * the PMC after this code has already checked
5158 			 * the pm_stalled flag.
5159 			 */
5160 			if (pm->pm_pcpu_state[cpu].pps_cpustate) {
5161 				pm->pm_pcpu_state[cpu].pps_cpustate = 0;
5162 				if (!pm->pm_pcpu_state[cpu].pps_stalled) {
5163 					(void) pcd->pcd_stop_pmc(cpu, adjri);
5164 
5165 					if (PMC_TO_MODE(pm) == PMC_MODE_TC) {
5166 						pcd->pcd_read_pmc(cpu, adjri,
5167 						    &newvalue);
5168 						tmp = newvalue -
5169 						    PMC_PCPU_SAVED(cpu,ri);
5170 
5171 						mtx_pool_lock_spin(pmc_mtxpool,
5172 						    pm);
5173 						pm->pm_gv.pm_savedvalue += tmp;
5174 						pp->pp_pmcs[ri].pp_pmcval +=
5175 						    tmp;
5176 						mtx_pool_unlock_spin(
5177 						    pmc_mtxpool, pm);
5178 					}
5179 				}
5180 			}
5181 
5182 			counter_u64_add(pm->pm_runcount, -1);
5183 
5184 			KASSERT((int) counter_u64_fetch(pm->pm_runcount) >= 0,
5185 			    ("[pmc,%d] runcount is %d", __LINE__, ri));
5186 
5187 			(void) pcd->pcd_config_pmc(cpu, adjri, NULL);
5188 		}
5189 
5190 		/*
5191 		 * Inform the MD layer of this pseudo "context switch
5192 		 * out"
5193 		 */
5194 		(void) md->pmd_switch_out(pmc_pcpu[cpu], pp);
5195 
5196 		critical_exit(); /* ok to be pre-empted now */
5197 
5198 		/*
5199 		 * Unlink this process from the PMCs that are
5200 		 * targeting it.  This will send a signal to
5201 		 * all PMC owner's whose PMCs are orphaned.
5202 		 *
5203 		 * Log PMC value at exit time if requested.
5204 		 */
5205 		for (ri = 0; ri < md->pmd_npmc; ri++)
5206 			if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) {
5207 				if (pm->pm_flags & PMC_F_NEEDS_LOGFILE &&
5208 				    PMC_IS_COUNTING_MODE(PMC_TO_MODE(pm)))
5209 					pmclog_process_procexit(pm, pp);
5210 				pmc_unlink_target_process(pm, pp);
5211 			}
5212 		free(pp, M_PMC);
5213 
5214 	} else
5215 		critical_exit(); /* pp == NULL */
5216 
5217 
5218 	/*
5219 	 * If the process owned PMCs, free them up and free up
5220 	 * memory.
5221 	 */
5222 	if ((po = pmc_find_owner_descriptor(p)) != NULL) {
5223 		pmc_remove_owner(po);
5224 		pmc_destroy_owner_descriptor(po);
5225 	}
5226 
5227 	sx_xunlock(&pmc_sx);
5228 }
5229 
5230 /*
5231  * Handle a process fork.
5232  *
5233  * If the parent process 'p1' is under HWPMC monitoring, then copy
5234  * over any attached PMCs that have 'do_descendants' semantics.
5235  */
5236 
5237 static void
5238 pmc_process_fork(void *arg __unused, struct proc *p1, struct proc *newproc,
5239     int flags)
5240 {
5241 	int is_using_hwpmcs;
5242 	unsigned int ri;
5243 	uint32_t do_descendants;
5244 	struct pmc *pm;
5245 	struct pmc_owner *po;
5246 	struct pmc_process *ppnew, *ppold;
5247 
5248 	(void) flags;		/* unused parameter */
5249 
5250 	PROC_LOCK(p1);
5251 	is_using_hwpmcs = p1->p_flag & P_HWPMC;
5252 	PROC_UNLOCK(p1);
5253 
5254 	/*
5255 	 * If there are system-wide sampling PMCs active, we need to
5256 	 * log all fork events to their owner's logs.
5257 	 */
5258 	epoch_enter_preempt(global_epoch_preempt);
5259 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
5260 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE) {
5261 		    pmclog_process_procfork(po, p1->p_pid, newproc->p_pid);
5262 			pmclog_process_proccreate(po, newproc, 1);
5263 		}
5264 	epoch_exit_preempt(global_epoch_preempt);
5265 
5266 	if (!is_using_hwpmcs)
5267 		return;
5268 
5269 	PMC_GET_SX_XLOCK();
5270 	PMCDBG4(PMC,FRK,1, "process-fork proc=%p (%d, %s) -> %p", p1,
5271 	    p1->p_pid, p1->p_comm, newproc);
5272 
5273 	/*
5274 	 * If the parent process (curthread->td_proc) is a
5275 	 * target of any PMCs, look for PMCs that are to be
5276 	 * inherited, and link these into the new process
5277 	 * descriptor.
5278 	 */
5279 	if ((ppold = pmc_find_process_descriptor(curthread->td_proc,
5280 		 PMC_FLAG_NONE)) == NULL)
5281 		goto done;		/* nothing to do */
5282 
5283 	do_descendants = 0;
5284 	for (ri = 0; ri < md->pmd_npmc; ri++)
5285 		if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL)
5286 			do_descendants |= pm->pm_flags & PMC_F_DESCENDANTS;
5287 	if (do_descendants == 0) /* nothing to do */
5288 		goto done;
5289 
5290 	/*
5291 	 * Now mark the new process as being tracked by this driver.
5292 	 */
5293 	PROC_LOCK(newproc);
5294 	newproc->p_flag |= P_HWPMC;
5295 	PROC_UNLOCK(newproc);
5296 
5297 	/* allocate a descriptor for the new process  */
5298 	if ((ppnew = pmc_find_process_descriptor(newproc,
5299 		 PMC_FLAG_ALLOCATE)) == NULL)
5300 		goto done;
5301 
5302 	/*
5303 	 * Run through all PMCs that were targeting the old process
5304 	 * and which specified F_DESCENDANTS and attach them to the
5305 	 * new process.
5306 	 *
5307 	 * Log the fork event to all owners of PMCs attached to this
5308 	 * process, if not already logged.
5309 	 */
5310 	for (ri = 0; ri < md->pmd_npmc; ri++)
5311 		if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL &&
5312 		    (pm->pm_flags & PMC_F_DESCENDANTS)) {
5313 			pmc_link_target_process(pm, ppnew);
5314 			po = pm->pm_owner;
5315 			if (po->po_sscount == 0 &&
5316 			    po->po_flags & PMC_PO_OWNS_LOGFILE)
5317 				pmclog_process_procfork(po, p1->p_pid,
5318 				    newproc->p_pid);
5319 		}
5320 
5321  done:
5322 	sx_xunlock(&pmc_sx);
5323 }
5324 
5325 static void
5326 pmc_process_threadcreate(struct thread *td)
5327 {
5328 	struct pmc_owner *po;
5329 
5330 	epoch_enter_preempt(global_epoch_preempt);
5331 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
5332 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
5333 			pmclog_process_threadcreate(po, td, 1);
5334 	epoch_exit_preempt(global_epoch_preempt);
5335 }
5336 
5337 static void
5338 pmc_process_threadexit(struct thread *td)
5339 {
5340 	struct pmc_owner *po;
5341 
5342 	epoch_enter_preempt(global_epoch_preempt);
5343 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
5344 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
5345 			pmclog_process_threadexit(po, td);
5346 	epoch_exit_preempt(global_epoch_preempt);
5347 }
5348 
5349 static void
5350 pmc_process_proccreate(struct proc *p)
5351 {
5352 	struct pmc_owner *po;
5353 
5354 	epoch_enter_preempt(global_epoch_preempt);
5355 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
5356 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
5357 			pmclog_process_proccreate(po, p, 1 /* sync */);
5358 	epoch_exit_preempt(global_epoch_preempt);
5359 }
5360 
5361 static void
5362 pmc_process_allproc(struct pmc *pm)
5363 {
5364 	struct pmc_owner *po;
5365 	struct thread *td;
5366 	struct proc *p;
5367 
5368 	po = pm->pm_owner;
5369 	if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
5370 		return;
5371 	sx_slock(&allproc_lock);
5372 	FOREACH_PROC_IN_SYSTEM(p) {
5373 		pmclog_process_proccreate(po, p, 0 /* sync */);
5374 		PROC_LOCK(p);
5375 		FOREACH_THREAD_IN_PROC(p, td)
5376 			pmclog_process_threadcreate(po, td, 0 /* sync */);
5377 		PROC_UNLOCK(p);
5378 	}
5379 	sx_sunlock(&allproc_lock);
5380 	pmclog_flush(po, 0);
5381 }
5382 
5383 static void
5384 pmc_kld_load(void *arg __unused, linker_file_t lf)
5385 {
5386 	struct pmc_owner *po;
5387 
5388 	/*
5389 	 * Notify owners of system sampling PMCs about KLD operations.
5390 	 */
5391 	epoch_enter_preempt(global_epoch_preempt);
5392 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
5393 		if (po->po_flags & PMC_PO_OWNS_LOGFILE)
5394 			pmclog_process_map_in(po, (pid_t) -1,
5395 			    (uintfptr_t) lf->address, lf->filename);
5396 	epoch_exit_preempt(global_epoch_preempt);
5397 
5398 	/*
5399 	 * TODO: Notify owners of (all) process-sampling PMCs too.
5400 	 */
5401 }
5402 
5403 static void
5404 pmc_kld_unload(void *arg __unused, const char *filename __unused,
5405     caddr_t address, size_t size)
5406 {
5407 	struct pmc_owner *po;
5408 
5409 	epoch_enter_preempt(global_epoch_preempt);
5410 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
5411 		if (po->po_flags & PMC_PO_OWNS_LOGFILE)
5412 			pmclog_process_map_out(po, (pid_t) -1,
5413 			    (uintfptr_t) address, (uintfptr_t) address + size);
5414 	epoch_exit_preempt(global_epoch_preempt);
5415 
5416 	/*
5417 	 * TODO: Notify owners of process-sampling PMCs.
5418 	 */
5419 }
5420 
5421 /*
5422  * initialization
5423  */
5424 static const char *
5425 pmc_name_of_pmcclass(enum pmc_class class)
5426 {
5427 
5428 	switch (class) {
5429 #undef	__PMC_CLASS
5430 #define	__PMC_CLASS(S,V,D)						\
5431 	case PMC_CLASS_##S:						\
5432 		return #S;
5433 	__PMC_CLASSES();
5434 	default:
5435 		return ("<unknown>");
5436 	}
5437 }
5438 
5439 /*
5440  * Base class initializer: allocate structure and set default classes.
5441  */
5442 struct pmc_mdep *
5443 pmc_mdep_alloc(int nclasses)
5444 {
5445 	struct pmc_mdep *md;
5446 	int	n;
5447 
5448 	/* SOFT + md classes */
5449 	n = 1 + nclasses;
5450 	md = malloc(sizeof(struct pmc_mdep) + n *
5451 	    sizeof(struct pmc_classdep), M_PMC, M_WAITOK|M_ZERO);
5452 	md->pmd_nclass = n;
5453 
5454 	/* Add base class. */
5455 	pmc_soft_initialize(md);
5456 	return md;
5457 }
5458 
5459 void
5460 pmc_mdep_free(struct pmc_mdep *md)
5461 {
5462 	pmc_soft_finalize(md);
5463 	free(md, M_PMC);
5464 }
5465 
5466 static int
5467 generic_switch_in(struct pmc_cpu *pc, struct pmc_process *pp)
5468 {
5469 	(void) pc; (void) pp;
5470 
5471 	return (0);
5472 }
5473 
5474 static int
5475 generic_switch_out(struct pmc_cpu *pc, struct pmc_process *pp)
5476 {
5477 	(void) pc; (void) pp;
5478 
5479 	return (0);
5480 }
5481 
5482 static struct pmc_mdep *
5483 pmc_generic_cpu_initialize(void)
5484 {
5485 	struct pmc_mdep *md;
5486 
5487 	md = pmc_mdep_alloc(0);
5488 
5489 	md->pmd_cputype    = PMC_CPU_GENERIC;
5490 
5491 	md->pmd_pcpu_init  = NULL;
5492 	md->pmd_pcpu_fini  = NULL;
5493 	md->pmd_switch_in  = generic_switch_in;
5494 	md->pmd_switch_out = generic_switch_out;
5495 
5496 	return (md);
5497 }
5498 
5499 static void
5500 pmc_generic_cpu_finalize(struct pmc_mdep *md)
5501 {
5502 	(void) md;
5503 }
5504 
5505 
5506 static int
5507 pmc_initialize(void)
5508 {
5509 	int c, cpu, error, n, ri;
5510 	unsigned int maxcpu, domain;
5511 	struct pcpu *pc;
5512 	struct pmc_binding pb;
5513 	struct pmc_sample *ps;
5514 	struct pmc_classdep *pcd;
5515 	struct pmc_samplebuffer *sb;
5516 
5517 	md = NULL;
5518 	error = 0;
5519 
5520 	pmc_stats.pm_intr_ignored = counter_u64_alloc(M_WAITOK);
5521 	pmc_stats.pm_intr_processed = counter_u64_alloc(M_WAITOK);
5522 	pmc_stats.pm_intr_bufferfull = counter_u64_alloc(M_WAITOK);
5523 	pmc_stats.pm_syscalls = counter_u64_alloc(M_WAITOK);
5524 	pmc_stats.pm_syscall_errors = counter_u64_alloc(M_WAITOK);
5525 	pmc_stats.pm_buffer_requests = counter_u64_alloc(M_WAITOK);
5526 	pmc_stats.pm_buffer_requests_failed = counter_u64_alloc(M_WAITOK);
5527 	pmc_stats.pm_log_sweeps = counter_u64_alloc(M_WAITOK);
5528 	pmc_stats.pm_merges = counter_u64_alloc(M_WAITOK);
5529 	pmc_stats.pm_overwrites = counter_u64_alloc(M_WAITOK);
5530 
5531 #ifdef	HWPMC_DEBUG
5532 	/* parse debug flags first */
5533 	if (TUNABLE_STR_FETCH(PMC_SYSCTL_NAME_PREFIX "debugflags",
5534 		pmc_debugstr, sizeof(pmc_debugstr)))
5535 		pmc_debugflags_parse(pmc_debugstr,
5536 		    pmc_debugstr+strlen(pmc_debugstr));
5537 #endif
5538 
5539 	PMCDBG1(MOD,INI,0, "PMC Initialize (version %x)", PMC_VERSION);
5540 
5541 	/* check kernel version */
5542 	if (pmc_kernel_version != PMC_VERSION) {
5543 		if (pmc_kernel_version == 0)
5544 			printf("hwpmc: this kernel has not been compiled with "
5545 			    "'options HWPMC_HOOKS'.\n");
5546 		else
5547 			printf("hwpmc: kernel version (0x%x) does not match "
5548 			    "module version (0x%x).\n", pmc_kernel_version,
5549 			    PMC_VERSION);
5550 		return EPROGMISMATCH;
5551 	}
5552 
5553 	/*
5554 	 * check sysctl parameters
5555 	 */
5556 
5557 	if (pmc_hashsize <= 0) {
5558 		(void) printf("hwpmc: tunable \"hashsize\"=%d must be "
5559 		    "greater than zero.\n", pmc_hashsize);
5560 		pmc_hashsize = PMC_HASH_SIZE;
5561 	}
5562 
5563 	if (pmc_nsamples <= 0 || pmc_nsamples > 65535) {
5564 		(void) printf("hwpmc: tunable \"nsamples\"=%d out of "
5565 		    "range.\n", pmc_nsamples);
5566 		pmc_nsamples = PMC_NSAMPLES;
5567 	}
5568 
5569 	if (pmc_callchaindepth <= 0 ||
5570 	    pmc_callchaindepth > PMC_CALLCHAIN_DEPTH_MAX) {
5571 		(void) printf("hwpmc: tunable \"callchaindepth\"=%d out of "
5572 		    "range - using %d.\n", pmc_callchaindepth,
5573 		    PMC_CALLCHAIN_DEPTH_MAX);
5574 		pmc_callchaindepth = PMC_CALLCHAIN_DEPTH_MAX;
5575 	}
5576 
5577 	md = pmc_md_initialize();
5578 	if (md == NULL) {
5579 		/* Default to generic CPU. */
5580 		md = pmc_generic_cpu_initialize();
5581 		if (md == NULL)
5582 			return (ENOSYS);
5583         }
5584 
5585 	KASSERT(md->pmd_nclass >= 1 && md->pmd_npmc >= 1,
5586 	    ("[pmc,%d] no classes or pmcs", __LINE__));
5587 
5588 	/* Compute the map from row-indices to classdep pointers. */
5589 	pmc_rowindex_to_classdep = malloc(sizeof(struct pmc_classdep *) *
5590 	    md->pmd_npmc, M_PMC, M_WAITOK|M_ZERO);
5591 
5592 	for (n = 0; n < md->pmd_npmc; n++)
5593 		pmc_rowindex_to_classdep[n] = NULL;
5594 	for (ri = c = 0; c < md->pmd_nclass; c++) {
5595 		pcd = &md->pmd_classdep[c];
5596 		for (n = 0; n < pcd->pcd_num; n++, ri++)
5597 			pmc_rowindex_to_classdep[ri] = pcd;
5598 	}
5599 
5600 	KASSERT(ri == md->pmd_npmc,
5601 	    ("[pmc,%d] npmc miscomputed: ri=%d, md->npmc=%d", __LINE__,
5602 	    ri, md->pmd_npmc));
5603 
5604 	maxcpu = pmc_cpu_max();
5605 
5606 	/* allocate space for the per-cpu array */
5607 	pmc_pcpu = malloc(maxcpu * sizeof(struct pmc_cpu *), M_PMC,
5608 	    M_WAITOK|M_ZERO);
5609 
5610 	/* per-cpu 'saved values' for managing process-mode PMCs */
5611 	pmc_pcpu_saved = malloc(sizeof(pmc_value_t) * maxcpu * md->pmd_npmc,
5612 	    M_PMC, M_WAITOK);
5613 
5614 	/* Perform CPU-dependent initialization. */
5615 	pmc_save_cpu_binding(&pb);
5616 	error = 0;
5617 	for (cpu = 0; error == 0 && cpu < maxcpu; cpu++) {
5618 		if (!pmc_cpu_is_active(cpu))
5619 			continue;
5620 		pmc_select_cpu(cpu);
5621 		pmc_pcpu[cpu] = malloc(sizeof(struct pmc_cpu) +
5622 		    md->pmd_npmc * sizeof(struct pmc_hw *), M_PMC,
5623 		    M_WAITOK|M_ZERO);
5624 		if (md->pmd_pcpu_init)
5625 			error = md->pmd_pcpu_init(md, cpu);
5626 		for (n = 0; error == 0 && n < md->pmd_nclass; n++)
5627 			error = md->pmd_classdep[n].pcd_pcpu_init(md, cpu);
5628 	}
5629 	pmc_restore_cpu_binding(&pb);
5630 
5631 	if (error)
5632 		return (error);
5633 
5634 	/* allocate space for the sample array */
5635 	for (cpu = 0; cpu < maxcpu; cpu++) {
5636 		if (!pmc_cpu_is_active(cpu))
5637 			continue;
5638 		pc = pcpu_find(cpu);
5639 		domain = pc->pc_domain;
5640 		sb = malloc_domain(sizeof(struct pmc_samplebuffer) +
5641 			pmc_nsamples * sizeof(struct pmc_sample), M_PMC, domain,
5642 		    M_WAITOK|M_ZERO);
5643 		sb->ps_read = sb->ps_write = sb->ps_samples;
5644 		sb->ps_fence = sb->ps_samples + pmc_nsamples;
5645 
5646 		KASSERT(pmc_pcpu[cpu] != NULL,
5647 		    ("[pmc,%d] cpu=%d Null per-cpu data", __LINE__, cpu));
5648 
5649 		sb->ps_callchains = malloc_domain(pmc_callchaindepth * pmc_nsamples *
5650 			sizeof(uintptr_t), M_PMC, domain, M_WAITOK|M_ZERO);
5651 
5652 		for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++)
5653 			ps->ps_pc = sb->ps_callchains +
5654 			    (n * pmc_callchaindepth);
5655 
5656 		pmc_pcpu[cpu]->pc_sb[PMC_HR] = sb;
5657 
5658 		sb = malloc_domain(sizeof(struct pmc_samplebuffer) +
5659 			pmc_nsamples * sizeof(struct pmc_sample), M_PMC, domain,
5660 		    M_WAITOK|M_ZERO);
5661 		sb->ps_read = sb->ps_write = sb->ps_samples;
5662 		sb->ps_fence = sb->ps_samples + pmc_nsamples;
5663 
5664 		KASSERT(pmc_pcpu[cpu] != NULL,
5665 		    ("[pmc,%d] cpu=%d Null per-cpu data", __LINE__, cpu));
5666 
5667 		sb->ps_callchains = malloc_domain(pmc_callchaindepth * pmc_nsamples *
5668 			sizeof(uintptr_t), M_PMC, domain, M_WAITOK|M_ZERO);
5669 
5670 		for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++)
5671 			ps->ps_pc = sb->ps_callchains +
5672 			    (n * pmc_callchaindepth);
5673 
5674 		pmc_pcpu[cpu]->pc_sb[PMC_SR] = sb;
5675 
5676 		sb = malloc_domain(sizeof(struct pmc_samplebuffer) +
5677 			pmc_nsamples * sizeof(struct pmc_sample), M_PMC, domain,
5678 		    M_WAITOK|M_ZERO);
5679 		sb->ps_read = sb->ps_write = sb->ps_samples;
5680 		sb->ps_fence = sb->ps_samples + pmc_nsamples;
5681 
5682 		KASSERT(pmc_pcpu[cpu] != NULL,
5683 		    ("[pmc,%d] cpu=%d Null per-cpu data", __LINE__, cpu));
5684 
5685 		sb->ps_callchains = malloc_domain(pmc_callchaindepth * pmc_nsamples *
5686 		    sizeof(uintptr_t), M_PMC, domain, M_WAITOK|M_ZERO);
5687 
5688 		for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++)
5689 			ps->ps_pc = sb->ps_callchains +
5690 			    (n * pmc_callchaindepth);
5691 
5692 		pmc_pcpu[cpu]->pc_sb[PMC_UR] = sb;
5693 	}
5694 
5695 	/* allocate space for the row disposition array */
5696 	pmc_pmcdisp = malloc(sizeof(enum pmc_mode) * md->pmd_npmc,
5697 	    M_PMC, M_WAITOK|M_ZERO);
5698 
5699 	/* mark all PMCs as available */
5700 	for (n = 0; n < (int) md->pmd_npmc; n++)
5701 		PMC_MARK_ROW_FREE(n);
5702 
5703 	/* allocate thread hash tables */
5704 	pmc_ownerhash = hashinit(pmc_hashsize, M_PMC,
5705 	    &pmc_ownerhashmask);
5706 
5707 	pmc_processhash = hashinit(pmc_hashsize, M_PMC,
5708 	    &pmc_processhashmask);
5709 	mtx_init(&pmc_processhash_mtx, "pmc-process-hash", "pmc-leaf",
5710 	    MTX_SPIN);
5711 
5712 	CK_LIST_INIT(&pmc_ss_owners);
5713 	pmc_ss_count = 0;
5714 
5715 	/* allocate a pool of spin mutexes */
5716 	pmc_mtxpool = mtx_pool_create("pmc-leaf", pmc_mtxpool_size,
5717 	    MTX_SPIN);
5718 
5719 	PMCDBG4(MOD,INI,1, "pmc_ownerhash=%p, mask=0x%lx "
5720 	    "targethash=%p mask=0x%lx", pmc_ownerhash, pmc_ownerhashmask,
5721 	    pmc_processhash, pmc_processhashmask);
5722 
5723 	/* Initialize a spin mutex for the thread free list. */
5724 	mtx_init(&pmc_threadfreelist_mtx, "pmc-threadfreelist", "pmc-leaf",
5725 	    MTX_SPIN);
5726 
5727 	/*
5728 	 * Initialize the callout to monitor the thread free list.
5729 	 * This callout will also handle the initial population of the list.
5730 	 */
5731 	taskqgroup_config_gtask_init(NULL, &free_gtask, pmc_thread_descriptor_pool_free_task, "thread descriptor pool free task");
5732 
5733 	/* register process {exit,fork,exec} handlers */
5734 	pmc_exit_tag = EVENTHANDLER_REGISTER(process_exit,
5735 	    pmc_process_exit, NULL, EVENTHANDLER_PRI_ANY);
5736 	pmc_fork_tag = EVENTHANDLER_REGISTER(process_fork,
5737 	    pmc_process_fork, NULL, EVENTHANDLER_PRI_ANY);
5738 
5739 	/* register kld event handlers */
5740 	pmc_kld_load_tag = EVENTHANDLER_REGISTER(kld_load, pmc_kld_load,
5741 	    NULL, EVENTHANDLER_PRI_ANY);
5742 	pmc_kld_unload_tag = EVENTHANDLER_REGISTER(kld_unload, pmc_kld_unload,
5743 	    NULL, EVENTHANDLER_PRI_ANY);
5744 
5745 	/* initialize logging */
5746 	pmclog_initialize();
5747 
5748 	/* set hook functions */
5749 	pmc_intr = md->pmd_intr;
5750 	wmb();
5751 	pmc_hook = pmc_hook_handler;
5752 
5753 	if (error == 0) {
5754 		printf(PMC_MODULE_NAME ":");
5755 		for (n = 0; n < (int) md->pmd_nclass; n++) {
5756 			pcd = &md->pmd_classdep[n];
5757 			printf(" %s/%d/%d/0x%b",
5758 			    pmc_name_of_pmcclass(pcd->pcd_class),
5759 			    pcd->pcd_num,
5760 			    pcd->pcd_width,
5761 			    pcd->pcd_caps,
5762 			    "\20"
5763 			    "\1INT\2USR\3SYS\4EDG\5THR"
5764 			    "\6REA\7WRI\10INV\11QUA\12PRC"
5765 			    "\13TAG\14CSC");
5766 		}
5767 		printf("\n");
5768 	}
5769 
5770 	return (error);
5771 }
5772 
5773 /* prepare to be unloaded */
5774 static void
5775 pmc_cleanup(void)
5776 {
5777 	int c, cpu;
5778 	unsigned int maxcpu;
5779 	struct pmc_ownerhash *ph;
5780 	struct pmc_owner *po, *tmp;
5781 	struct pmc_binding pb;
5782 #ifdef	HWPMC_DEBUG
5783 	struct pmc_processhash *prh;
5784 #endif
5785 
5786 	PMCDBG0(MOD,INI,0, "cleanup");
5787 
5788 	/* switch off sampling */
5789 	CPU_FOREACH(cpu)
5790 		DPCPU_ID_SET(cpu, pmc_sampled, 0);
5791 	pmc_intr = NULL;
5792 
5793 	sx_xlock(&pmc_sx);
5794 	if (pmc_hook == NULL) {	/* being unloaded already */
5795 		sx_xunlock(&pmc_sx);
5796 		return;
5797 	}
5798 
5799 	pmc_hook = NULL; /* prevent new threads from entering module */
5800 
5801 	/* deregister event handlers */
5802 	EVENTHANDLER_DEREGISTER(process_fork, pmc_fork_tag);
5803 	EVENTHANDLER_DEREGISTER(process_exit, pmc_exit_tag);
5804 	EVENTHANDLER_DEREGISTER(kld_load, pmc_kld_load_tag);
5805 	EVENTHANDLER_DEREGISTER(kld_unload, pmc_kld_unload_tag);
5806 
5807 	/* send SIGBUS to all owner threads, free up allocations */
5808 	if (pmc_ownerhash)
5809 		for (ph = pmc_ownerhash;
5810 		     ph <= &pmc_ownerhash[pmc_ownerhashmask];
5811 		     ph++) {
5812 			LIST_FOREACH_SAFE(po, ph, po_next, tmp) {
5813 				pmc_remove_owner(po);
5814 
5815 				/* send SIGBUS to owner processes */
5816 				PMCDBG3(MOD,INI,2, "cleanup signal proc=%p "
5817 				    "(%d, %s)", po->po_owner,
5818 				    po->po_owner->p_pid,
5819 				    po->po_owner->p_comm);
5820 
5821 				PROC_LOCK(po->po_owner);
5822 				kern_psignal(po->po_owner, SIGBUS);
5823 				PROC_UNLOCK(po->po_owner);
5824 
5825 				pmc_destroy_owner_descriptor(po);
5826 			}
5827 		}
5828 
5829 	/* reclaim allocated data structures */
5830 	mtx_destroy(&pmc_threadfreelist_mtx);
5831 	pmc_thread_descriptor_pool_drain();
5832 
5833 	if (pmc_mtxpool)
5834 		mtx_pool_destroy(&pmc_mtxpool);
5835 
5836 	mtx_destroy(&pmc_processhash_mtx);
5837 	taskqgroup_config_gtask_deinit(&free_gtask);
5838 	if (pmc_processhash) {
5839 #ifdef	HWPMC_DEBUG
5840 		struct pmc_process *pp;
5841 
5842 		PMCDBG0(MOD,INI,3, "destroy process hash");
5843 		for (prh = pmc_processhash;
5844 		     prh <= &pmc_processhash[pmc_processhashmask];
5845 		     prh++)
5846 			LIST_FOREACH(pp, prh, pp_next)
5847 			    PMCDBG1(MOD,INI,3, "pid=%d", pp->pp_proc->p_pid);
5848 #endif
5849 
5850 		hashdestroy(pmc_processhash, M_PMC, pmc_processhashmask);
5851 		pmc_processhash = NULL;
5852 	}
5853 
5854 	if (pmc_ownerhash) {
5855 		PMCDBG0(MOD,INI,3, "destroy owner hash");
5856 		hashdestroy(pmc_ownerhash, M_PMC, pmc_ownerhashmask);
5857 		pmc_ownerhash = NULL;
5858 	}
5859 
5860 	KASSERT(CK_LIST_EMPTY(&pmc_ss_owners),
5861 	    ("[pmc,%d] Global SS owner list not empty", __LINE__));
5862 	KASSERT(pmc_ss_count == 0,
5863 	    ("[pmc,%d] Global SS count not empty", __LINE__));
5864 
5865  	/* do processor and pmc-class dependent cleanup */
5866 	maxcpu = pmc_cpu_max();
5867 
5868 	PMCDBG0(MOD,INI,3, "md cleanup");
5869 	if (md) {
5870 		pmc_save_cpu_binding(&pb);
5871 		for (cpu = 0; cpu < maxcpu; cpu++) {
5872 			PMCDBG2(MOD,INI,1,"pmc-cleanup cpu=%d pcs=%p",
5873 			    cpu, pmc_pcpu[cpu]);
5874 			if (!pmc_cpu_is_active(cpu) || pmc_pcpu[cpu] == NULL)
5875 				continue;
5876 			pmc_select_cpu(cpu);
5877 			for (c = 0; c < md->pmd_nclass; c++)
5878 				md->pmd_classdep[c].pcd_pcpu_fini(md, cpu);
5879 			if (md->pmd_pcpu_fini)
5880 				md->pmd_pcpu_fini(md, cpu);
5881 		}
5882 
5883 		if (md->pmd_cputype == PMC_CPU_GENERIC)
5884 			pmc_generic_cpu_finalize(md);
5885 		else
5886 			pmc_md_finalize(md);
5887 
5888 		pmc_mdep_free(md);
5889 		md = NULL;
5890 		pmc_restore_cpu_binding(&pb);
5891 	}
5892 
5893 	/* Free per-cpu descriptors. */
5894 	for (cpu = 0; cpu < maxcpu; cpu++) {
5895 		if (!pmc_cpu_is_active(cpu))
5896 			continue;
5897 		KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_HR] != NULL,
5898 		    ("[pmc,%d] Null hw cpu sample buffer cpu=%d", __LINE__,
5899 			cpu));
5900 		KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_SR] != NULL,
5901 		    ("[pmc,%d] Null sw cpu sample buffer cpu=%d", __LINE__,
5902 			cpu));
5903 		KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_UR] != NULL,
5904 		    ("[pmc,%d] Null userret cpu sample buffer cpu=%d", __LINE__,
5905 			cpu));
5906 		free_domain(pmc_pcpu[cpu]->pc_sb[PMC_HR]->ps_callchains, M_PMC);
5907 		free_domain(pmc_pcpu[cpu]->pc_sb[PMC_HR], M_PMC);
5908 		free_domain(pmc_pcpu[cpu]->pc_sb[PMC_SR]->ps_callchains, M_PMC);
5909 		free_domain(pmc_pcpu[cpu]->pc_sb[PMC_SR], M_PMC);
5910 		free_domain(pmc_pcpu[cpu]->pc_sb[PMC_UR]->ps_callchains, M_PMC);
5911 		free_domain(pmc_pcpu[cpu]->pc_sb[PMC_UR], M_PMC);
5912 		free_domain(pmc_pcpu[cpu], M_PMC);
5913 	}
5914 
5915 	free(pmc_pcpu, M_PMC);
5916 	pmc_pcpu = NULL;
5917 
5918 	free(pmc_pcpu_saved, M_PMC);
5919 	pmc_pcpu_saved = NULL;
5920 
5921 	if (pmc_pmcdisp) {
5922 		free(pmc_pmcdisp, M_PMC);
5923 		pmc_pmcdisp = NULL;
5924 	}
5925 
5926 	if (pmc_rowindex_to_classdep) {
5927 		free(pmc_rowindex_to_classdep, M_PMC);
5928 		pmc_rowindex_to_classdep = NULL;
5929 	}
5930 
5931 	pmclog_shutdown();
5932 	counter_u64_free(pmc_stats.pm_intr_ignored);
5933 	counter_u64_free(pmc_stats.pm_intr_processed);
5934 	counter_u64_free(pmc_stats.pm_intr_bufferfull);
5935 	counter_u64_free(pmc_stats.pm_syscalls);
5936 	counter_u64_free(pmc_stats.pm_syscall_errors);
5937 	counter_u64_free(pmc_stats.pm_buffer_requests);
5938 	counter_u64_free(pmc_stats.pm_buffer_requests_failed);
5939 	counter_u64_free(pmc_stats.pm_log_sweeps);
5940 	counter_u64_free(pmc_stats.pm_merges);
5941 	counter_u64_free(pmc_stats.pm_overwrites);
5942 	sx_xunlock(&pmc_sx); 	/* we are done */
5943 }
5944 
5945 /*
5946  * The function called at load/unload.
5947  */
5948 
5949 static int
5950 load (struct module *module __unused, int cmd, void *arg __unused)
5951 {
5952 	int error;
5953 
5954 	error = 0;
5955 
5956 	switch (cmd) {
5957 	case MOD_LOAD :
5958 		/* initialize the subsystem */
5959 		error = pmc_initialize();
5960 		if (error != 0)
5961 			break;
5962 		PMCDBG2(MOD,INI,1, "syscall=%d maxcpu=%d",
5963 		    pmc_syscall_num, pmc_cpu_max());
5964 		break;
5965 
5966 
5967 	case MOD_UNLOAD :
5968 	case MOD_SHUTDOWN:
5969 		pmc_cleanup();
5970 		PMCDBG0(MOD,INI,1, "unloaded");
5971 		break;
5972 
5973 	default :
5974 		error = EINVAL;	/* XXX should panic(9) */
5975 		break;
5976 	}
5977 
5978 	return error;
5979 }
5980