xref: /freebsd/sys/dev/hwpmc/hwpmc_mod.c (revision e87ec409fa9b21abf79895837fe375ab3d7e408a)
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/systm.h>
40 #include <sys/domainset.h>
41 #include <sys/eventhandler.h>
42 #include <sys/jail.h>
43 #include <sys/kernel.h>
44 #include <sys/kthread.h>
45 #include <sys/limits.h>
46 #include <sys/lock.h>
47 #include <sys/malloc.h>
48 #include <sys/module.h>
49 #include <sys/mount.h>
50 #include <sys/mutex.h>
51 #include <sys/pmc.h>
52 #include <sys/pmckern.h>
53 #include <sys/pmclog.h>
54 #include <sys/priv.h>
55 #include <sys/proc.h>
56 #include <sys/queue.h>
57 #include <sys/resourcevar.h>
58 #include <sys/rwlock.h>
59 #include <sys/sched.h>
60 #include <sys/signalvar.h>
61 #include <sys/smp.h>
62 #include <sys/sx.h>
63 #include <sys/sysctl.h>
64 #include <sys/sysent.h>
65 #include <sys/syslog.h>
66 #include <sys/taskqueue.h>
67 #include <sys/vnode.h>
68 
69 #include <sys/linker.h>		/* needs to be after <sys/malloc.h> */
70 
71 #include <machine/atomic.h>
72 #include <machine/md_var.h>
73 
74 #include <vm/vm.h>
75 #include <vm/vm_extern.h>
76 #include <vm/pmap.h>
77 #include <vm/vm_map.h>
78 #include <vm/vm_object.h>
79 
80 #include "hwpmc_soft.h"
81 
82 #define PMC_EPOCH_ENTER() struct epoch_tracker pmc_et; epoch_enter_preempt(global_epoch_preempt, &pmc_et)
83 #define PMC_EPOCH_EXIT() epoch_exit_preempt(global_epoch_preempt, &pmc_et)
84 
85 /*
86  * Types
87  */
88 
89 enum pmc_flags {
90 	PMC_FLAG_NONE	  = 0x00, /* do nothing */
91 	PMC_FLAG_REMOVE   = 0x01, /* atomically remove entry from hash */
92 	PMC_FLAG_ALLOCATE = 0x02, /* add entry to hash if not found */
93 	PMC_FLAG_NOWAIT   = 0x04, /* do not wait for mallocs */
94 };
95 
96 /*
97  * The offset in sysent where the syscall is allocated.
98  */
99 
100 static int pmc_syscall_num = NO_SYSCALL;
101 struct pmc_cpu		**pmc_pcpu;	 /* per-cpu state */
102 pmc_value_t		*pmc_pcpu_saved; /* saved PMC values: CSW handling */
103 
104 #define	PMC_PCPU_SAVED(C,R)	pmc_pcpu_saved[(R) + md->pmd_npmc*(C)]
105 
106 struct mtx_pool		*pmc_mtxpool;
107 static int		*pmc_pmcdisp;	 /* PMC row dispositions */
108 
109 #define	PMC_ROW_DISP_IS_FREE(R)		(pmc_pmcdisp[(R)] == 0)
110 #define	PMC_ROW_DISP_IS_THREAD(R)	(pmc_pmcdisp[(R)] > 0)
111 #define	PMC_ROW_DISP_IS_STANDALONE(R)	(pmc_pmcdisp[(R)] < 0)
112 
113 #define	PMC_MARK_ROW_FREE(R) do {					  \
114 	pmc_pmcdisp[(R)] = 0;						  \
115 } while (0)
116 
117 #define	PMC_MARK_ROW_STANDALONE(R) do {					  \
118 	KASSERT(pmc_pmcdisp[(R)] <= 0, ("[pmc,%d] row disposition error", \
119 		    __LINE__));						  \
120 	atomic_add_int(&pmc_pmcdisp[(R)], -1);				  \
121 	KASSERT(pmc_pmcdisp[(R)] >= (-pmc_cpu_max_active()),		  \
122 		("[pmc,%d] row disposition error", __LINE__));		  \
123 } while (0)
124 
125 #define	PMC_UNMARK_ROW_STANDALONE(R) do { 				  \
126 	atomic_add_int(&pmc_pmcdisp[(R)], 1);				  \
127 	KASSERT(pmc_pmcdisp[(R)] <= 0, ("[pmc,%d] row disposition error", \
128 		    __LINE__));						  \
129 } while (0)
130 
131 #define	PMC_MARK_ROW_THREAD(R) do {					  \
132 	KASSERT(pmc_pmcdisp[(R)] >= 0, ("[pmc,%d] row disposition error", \
133 		    __LINE__));						  \
134 	atomic_add_int(&pmc_pmcdisp[(R)], 1);				  \
135 } while (0)
136 
137 #define	PMC_UNMARK_ROW_THREAD(R) do {					  \
138 	atomic_add_int(&pmc_pmcdisp[(R)], -1);				  \
139 	KASSERT(pmc_pmcdisp[(R)] >= 0, ("[pmc,%d] row disposition error", \
140 		    __LINE__));						  \
141 } while (0)
142 
143 
144 /* various event handlers */
145 static eventhandler_tag	pmc_exit_tag, pmc_fork_tag, pmc_kld_load_tag,
146     pmc_kld_unload_tag;
147 
148 /* Module statistics */
149 struct pmc_driverstats pmc_stats;
150 
151 
152 /* Machine/processor dependent operations */
153 static struct pmc_mdep  *md;
154 
155 /*
156  * Hash tables mapping owner processes and target threads to PMCs.
157  */
158 
159 struct mtx pmc_processhash_mtx;		/* spin mutex */
160 static u_long pmc_processhashmask;
161 static LIST_HEAD(pmc_processhash, pmc_process)	*pmc_processhash;
162 
163 /*
164  * Hash table of PMC owner descriptors.  This table is protected by
165  * the shared PMC "sx" lock.
166  */
167 
168 static u_long pmc_ownerhashmask;
169 static LIST_HEAD(pmc_ownerhash, pmc_owner)	*pmc_ownerhash;
170 
171 /*
172  * List of PMC owners with system-wide sampling PMCs.
173  */
174 
175 static CK_LIST_HEAD(, pmc_owner)			pmc_ss_owners;
176 
177 /*
178  * List of free thread entries. This is protected by the spin
179  * mutex.
180  */
181 static struct mtx pmc_threadfreelist_mtx;	/* spin mutex */
182 static LIST_HEAD(, pmc_thread)			pmc_threadfreelist;
183 static int pmc_threadfreelist_entries=0;
184 #define	THREADENTRY_SIZE						\
185 (sizeof(struct pmc_thread) + (md->pmd_npmc * sizeof(struct pmc_threadpmcstate)))
186 
187 /*
188  * Task to free thread descriptors
189  */
190 static struct task free_task;
191 
192 /*
193  * A map of row indices to classdep structures.
194  */
195 static struct pmc_classdep **pmc_rowindex_to_classdep;
196 
197 /*
198  * Prototypes
199  */
200 
201 #ifdef	HWPMC_DEBUG
202 static int	pmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS);
203 static int	pmc_debugflags_parse(char *newstr, char *fence);
204 #endif
205 
206 static int	load(struct module *module, int cmd, void *arg);
207 static int	pmc_add_sample(ring_type_t ring, struct pmc *pm, struct trapframe *tf);
208 static void	pmc_add_thread_descriptors_from_proc(struct proc *p,
209     struct pmc_process *pp);
210 static int	pmc_attach_process(struct proc *p, struct pmc *pm);
211 static struct pmc *pmc_allocate_pmc_descriptor(void);
212 static struct pmc_owner *pmc_allocate_owner_descriptor(struct proc *p);
213 static int	pmc_attach_one_process(struct proc *p, struct pmc *pm);
214 static int	pmc_can_allocate_rowindex(struct proc *p, unsigned int ri,
215     int cpu);
216 static int	pmc_can_attach(struct pmc *pm, struct proc *p);
217 static void	pmc_capture_user_callchain(int cpu, int soft, struct trapframe *tf);
218 static void	pmc_cleanup(void);
219 static int	pmc_detach_process(struct proc *p, struct pmc *pm);
220 static int	pmc_detach_one_process(struct proc *p, struct pmc *pm,
221     int flags);
222 static void	pmc_destroy_owner_descriptor(struct pmc_owner *po);
223 static void	pmc_destroy_pmc_descriptor(struct pmc *pm);
224 static void	pmc_destroy_process_descriptor(struct pmc_process *pp);
225 static struct pmc_owner *pmc_find_owner_descriptor(struct proc *p);
226 static int	pmc_find_pmc(pmc_id_t pmcid, struct pmc **pm);
227 static struct pmc *pmc_find_pmc_descriptor_in_process(struct pmc_owner *po,
228     pmc_id_t pmc);
229 static struct pmc_process *pmc_find_process_descriptor(struct proc *p,
230     uint32_t mode);
231 static struct pmc_thread *pmc_find_thread_descriptor(struct pmc_process *pp,
232     struct thread *td, uint32_t mode);
233 static void	pmc_force_context_switch(void);
234 static void	pmc_link_target_process(struct pmc *pm,
235     struct pmc_process *pp);
236 static void	pmc_log_all_process_mappings(struct pmc_owner *po);
237 static void	pmc_log_kernel_mappings(struct pmc *pm);
238 static void	pmc_log_process_mappings(struct pmc_owner *po, struct proc *p);
239 static void	pmc_maybe_remove_owner(struct pmc_owner *po);
240 static void	pmc_process_csw_in(struct thread *td);
241 static void	pmc_process_csw_out(struct thread *td);
242 static void	pmc_process_exit(void *arg, struct proc *p);
243 static void	pmc_process_fork(void *arg, struct proc *p1,
244     struct proc *p2, int n);
245 static void	pmc_process_samples(int cpu, ring_type_t soft);
246 static void	pmc_release_pmc_descriptor(struct pmc *pmc);
247 static void	pmc_process_thread_add(struct thread *td);
248 static void	pmc_process_thread_delete(struct thread *td);
249 static void	pmc_process_thread_userret(struct thread *td);
250 static void	pmc_remove_owner(struct pmc_owner *po);
251 static void	pmc_remove_process_descriptor(struct pmc_process *pp);
252 static void	pmc_restore_cpu_binding(struct pmc_binding *pb);
253 static void	pmc_save_cpu_binding(struct pmc_binding *pb);
254 static void	pmc_select_cpu(int cpu);
255 static int	pmc_start(struct pmc *pm);
256 static int	pmc_stop(struct pmc *pm);
257 static int	pmc_syscall_handler(struct thread *td, void *syscall_args);
258 static struct pmc_thread *pmc_thread_descriptor_pool_alloc(void);
259 static void	pmc_thread_descriptor_pool_drain(void);
260 static void	pmc_thread_descriptor_pool_free(struct pmc_thread *pt);
261 static void	pmc_unlink_target_process(struct pmc *pmc,
262     struct pmc_process *pp);
263 static int generic_switch_in(struct pmc_cpu *pc, struct pmc_process *pp);
264 static int generic_switch_out(struct pmc_cpu *pc, struct pmc_process *pp);
265 static struct pmc_mdep *pmc_generic_cpu_initialize(void);
266 static void pmc_generic_cpu_finalize(struct pmc_mdep *md);
267 static void pmc_post_callchain_callback(void);
268 static void pmc_process_threadcreate(struct thread *td);
269 static void pmc_process_threadexit(struct thread *td);
270 static void pmc_process_proccreate(struct proc *p);
271 static void pmc_process_allproc(struct pmc *pm);
272 
273 /*
274  * Kernel tunables and sysctl(8) interface.
275  */
276 
277 SYSCTL_DECL(_kern_hwpmc);
278 SYSCTL_NODE(_kern_hwpmc, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
279     "HWPMC stats");
280 
281 
282 /* Stats. */
283 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, intr_ignored, CTLFLAG_RW,
284 				   &pmc_stats.pm_intr_ignored, "# of interrupts ignored");
285 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, intr_processed, CTLFLAG_RW,
286 				   &pmc_stats.pm_intr_processed, "# of interrupts processed");
287 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, intr_bufferfull, CTLFLAG_RW,
288 				   &pmc_stats.pm_intr_bufferfull, "# of interrupts where buffer was full");
289 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, syscalls, CTLFLAG_RW,
290 				   &pmc_stats.pm_syscalls, "# of syscalls");
291 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, syscall_errors, CTLFLAG_RW,
292 				   &pmc_stats.pm_syscall_errors, "# of syscall_errors");
293 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, buffer_requests, CTLFLAG_RW,
294 				   &pmc_stats.pm_buffer_requests, "# of buffer requests");
295 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, buffer_requests_failed, CTLFLAG_RW,
296 				   &pmc_stats.pm_buffer_requests_failed, "# of buffer requests which failed");
297 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, log_sweeps, CTLFLAG_RW,
298 				   &pmc_stats.pm_log_sweeps, "# of ?");
299 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, merges, CTLFLAG_RW,
300 				   &pmc_stats.pm_merges, "# of times kernel stack was found for user trace");
301 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, overwrites, CTLFLAG_RW,
302 				   &pmc_stats.pm_overwrites, "# of times a sample was overwritten before being logged");
303 
304 static int pmc_callchaindepth = PMC_CALLCHAIN_DEPTH;
305 SYSCTL_INT(_kern_hwpmc, OID_AUTO, callchaindepth, CTLFLAG_RDTUN,
306     &pmc_callchaindepth, 0, "depth of call chain records");
307 
308 char pmc_cpuid[PMC_CPUID_LEN];
309 SYSCTL_STRING(_kern_hwpmc, OID_AUTO, cpuid, CTLFLAG_RD,
310 	pmc_cpuid, 0, "cpu version string");
311 #ifdef	HWPMC_DEBUG
312 struct pmc_debugflags pmc_debugflags = PMC_DEBUG_DEFAULT_FLAGS;
313 char	pmc_debugstr[PMC_DEBUG_STRSIZE];
314 TUNABLE_STR(PMC_SYSCTL_NAME_PREFIX "debugflags", pmc_debugstr,
315     sizeof(pmc_debugstr));
316 SYSCTL_PROC(_kern_hwpmc, OID_AUTO, debugflags,
317     CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE,
318     0, 0, pmc_debugflags_sysctl_handler, "A",
319     "debug flags");
320 #endif
321 
322 
323 /*
324  * kern.hwpmc.hashrows -- determines the number of rows in the
325  * of the hash table used to look up threads
326  */
327 
328 static int pmc_hashsize = PMC_HASH_SIZE;
329 SYSCTL_INT(_kern_hwpmc, OID_AUTO, hashsize, CTLFLAG_RDTUN,
330     &pmc_hashsize, 0, "rows in hash tables");
331 
332 /*
333  * kern.hwpmc.nsamples --- number of PC samples/callchain stacks per CPU
334  */
335 
336 static int pmc_nsamples = PMC_NSAMPLES;
337 SYSCTL_INT(_kern_hwpmc, OID_AUTO, nsamples, CTLFLAG_RDTUN,
338     &pmc_nsamples, 0, "number of PC samples per CPU");
339 
340 static uint64_t pmc_sample_mask = PMC_NSAMPLES-1;
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 available 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(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 __diagused;
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 		KASSERT(counter_u64_fetch(pm->pm_runcount) >= 0,
1402 	    ("[pmc,%d] pm=%p runcount %ld", __LINE__, (void *) pm,
1403 		 (unsigned long)counter_u64_fetch(pm->pm_runcount)));
1404 
1405 		/* increment PMC runcount */
1406 		counter_u64_add(pm->pm_runcount, 1);
1407 
1408 		/* configure the HWPMC we are going to use. */
1409 		pcd = pmc_ri_to_classdep(md, ri, &adjri);
1410 		pcd->pcd_config_pmc(cpu, adjri, pm);
1411 
1412 		phw = pc->pc_hwpmcs[ri];
1413 
1414 		KASSERT(phw != NULL,
1415 		    ("[pmc,%d] null hw pointer", __LINE__));
1416 
1417 		KASSERT(phw->phw_pmc == pm,
1418 		    ("[pmc,%d] hw->pmc %p != pmc %p", __LINE__,
1419 			phw->phw_pmc, pm));
1420 
1421 		/*
1422 		 * Write out saved value and start the PMC.
1423 		 *
1424 		 * Sampling PMCs use a per-thread value, while
1425 		 * counting mode PMCs use a per-pmc value that is
1426 		 * inherited across descendants.
1427 		 */
1428 		if (PMC_TO_MODE(pm) == PMC_MODE_TS) {
1429 			if (pt == NULL)
1430 				pt = pmc_find_thread_descriptor(pp, td,
1431 				    PMC_FLAG_NONE);
1432 
1433 			KASSERT(pt != NULL,
1434 			    ("[pmc,%d] No thread found for td=%p", __LINE__,
1435 			    td));
1436 
1437 			mtx_pool_lock_spin(pmc_mtxpool, pm);
1438 
1439 			/*
1440 			 * If we have a thread descriptor, use the per-thread
1441 			 * counter in the descriptor. If not, we will use
1442 			 * a per-process counter.
1443 			 *
1444 			 * TODO: Remove the per-process "safety net" once
1445 			 * we have thoroughly tested that we don't hit the
1446 			 * above assert.
1447 			 */
1448 			if (pt != NULL) {
1449 				if (pt->pt_pmcs[ri].pt_pmcval > 0)
1450 					newvalue = pt->pt_pmcs[ri].pt_pmcval;
1451 				else
1452 					newvalue = pm->pm_sc.pm_reloadcount;
1453 			} else {
1454 				/*
1455 				 * Use the saved value calculated after the most
1456 				 * recent time a thread using the shared counter
1457 				 * switched out. Reset the saved count in case
1458 				 * another thread from this process switches in
1459 				 * before any threads switch out.
1460 				 */
1461 
1462 				newvalue = pp->pp_pmcs[ri].pp_pmcval;
1463 				pp->pp_pmcs[ri].pp_pmcval =
1464 				    pm->pm_sc.pm_reloadcount;
1465 			}
1466 			mtx_pool_unlock_spin(pmc_mtxpool, pm);
1467 			KASSERT(newvalue > 0 && newvalue <=
1468 			    pm->pm_sc.pm_reloadcount,
1469 			    ("[pmc,%d] pmcval outside of expected range cpu=%d "
1470 			    "ri=%d pmcval=%jx pm_reloadcount=%jx", __LINE__,
1471 			    cpu, ri, newvalue, pm->pm_sc.pm_reloadcount));
1472 		} else {
1473 			KASSERT(PMC_TO_MODE(pm) == PMC_MODE_TC,
1474 			    ("[pmc,%d] illegal mode=%d", __LINE__,
1475 			    PMC_TO_MODE(pm)));
1476 			mtx_pool_lock_spin(pmc_mtxpool, pm);
1477 			newvalue = PMC_PCPU_SAVED(cpu, ri) =
1478 			    pm->pm_gv.pm_savedvalue;
1479 			mtx_pool_unlock_spin(pmc_mtxpool, pm);
1480 		}
1481 
1482 		PMCDBG3(CSW,SWI,1,"cpu=%d ri=%d new=%jd", cpu, ri, newvalue);
1483 
1484 		pcd->pcd_write_pmc(cpu, adjri, newvalue);
1485 
1486 		/* If a sampling mode PMC, reset stalled state. */
1487 		if (PMC_TO_MODE(pm) == PMC_MODE_TS)
1488 			pm->pm_pcpu_state[cpu].pps_stalled = 0;
1489 
1490 		/* Indicate that we desire this to run. */
1491 		pm->pm_pcpu_state[cpu].pps_cpustate = 1;
1492 
1493 		/* Start the PMC. */
1494 		pcd->pcd_start_pmc(cpu, adjri);
1495 	}
1496 
1497 	/*
1498 	 * perform any other architecture/cpu dependent thread
1499 	 * switch-in actions.
1500 	 */
1501 
1502 	(void) (*md->pmd_switch_in)(pc, pp);
1503 
1504 	critical_exit();
1505 
1506 }
1507 
1508 /*
1509  * Thread context switch OUT.
1510  */
1511 
1512 static void
1513 pmc_process_csw_out(struct thread *td)
1514 {
1515 	int cpu;
1516 	int64_t tmp;
1517 	struct pmc *pm;
1518 	struct proc *p;
1519 	enum pmc_mode mode;
1520 	struct pmc_cpu *pc;
1521 	pmc_value_t newvalue;
1522 	unsigned int adjri, ri;
1523 	struct pmc_process *pp;
1524 	struct pmc_thread *pt = NULL;
1525 	struct pmc_classdep *pcd;
1526 
1527 
1528 	/*
1529 	 * Locate our process descriptor; this may be NULL if
1530 	 * this process is exiting and we have already removed
1531 	 * the process from the target process table.
1532 	 *
1533 	 * Note that due to kernel preemption, multiple
1534 	 * context switches may happen while the process is
1535 	 * exiting.
1536 	 *
1537 	 * Note also that if the target process cannot be
1538 	 * found we still need to deconfigure any PMCs that
1539 	 * are currently running on hardware.
1540 	 */
1541 
1542 	p = td->td_proc;
1543 	pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE);
1544 
1545 	/*
1546 	 * save PMCs
1547 	 */
1548 
1549 	critical_enter();
1550 
1551 	cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */
1552 
1553 	PMCDBG5(CSW,SWO,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p,
1554 	    p->p_pid, p->p_comm, pp);
1555 
1556 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
1557 	    ("[pmc,%d weird CPU id %d", __LINE__, cpu));
1558 
1559 	pc = pmc_pcpu[cpu];
1560 
1561 	/*
1562 	 * When a PMC gets unlinked from a target PMC, it will
1563 	 * be removed from the target's pp_pmc[] array.
1564 	 *
1565 	 * However, on a MP system, the target could have been
1566 	 * executing on another CPU at the time of the unlink.
1567 	 * So, at context switch OUT time, we need to look at
1568 	 * the hardware to determine if a PMC is scheduled on
1569 	 * it.
1570 	 */
1571 
1572 	for (ri = 0; ri < md->pmd_npmc; ri++) {
1573 
1574 		pcd = pmc_ri_to_classdep(md, ri, &adjri);
1575 		pm  = NULL;
1576 		(void) (*pcd->pcd_get_config)(cpu, adjri, &pm);
1577 
1578 		if (pm == NULL)	/* nothing at this row index */
1579 			continue;
1580 
1581 		mode = PMC_TO_MODE(pm);
1582 		if (!PMC_IS_VIRTUAL_MODE(mode))
1583 			continue; /* not a process virtual PMC */
1584 
1585 		KASSERT(PMC_TO_ROWINDEX(pm) == ri,
1586 		    ("[pmc,%d] ri mismatch pmc(%d) ri(%d)",
1587 			__LINE__, PMC_TO_ROWINDEX(pm), ri));
1588 
1589 		/*
1590 		 * Change desired state, and then stop if not stalled.
1591 		 * This two-step dance should avoid race conditions where
1592 		 * an interrupt re-enables the PMC after this code has
1593 		 * already checked the pm_stalled flag.
1594 		 */
1595 		pm->pm_pcpu_state[cpu].pps_cpustate = 0;
1596 		if (pm->pm_pcpu_state[cpu].pps_stalled == 0)
1597 			pcd->pcd_stop_pmc(cpu, adjri);
1598 
1599 		KASSERT(counter_u64_fetch(pm->pm_runcount) > 0,
1600 			("[pmc,%d] pm=%p runcount %ld", __LINE__, (void *) pm,
1601 			 (unsigned long)counter_u64_fetch(pm->pm_runcount)));
1602 
1603 		/* reduce this PMC's runcount */
1604 		counter_u64_add(pm->pm_runcount, -1);
1605 
1606 		/*
1607 		 * If this PMC is associated with this process,
1608 		 * save the reading.
1609 		 */
1610 
1611 		if (pm->pm_state != PMC_STATE_DELETED && pp != NULL &&
1612 		    pp->pp_pmcs[ri].pp_pmc != NULL) {
1613 			KASSERT(pm == pp->pp_pmcs[ri].pp_pmc,
1614 			    ("[pmc,%d] pm %p != pp_pmcs[%d] %p", __LINE__,
1615 				pm, ri, pp->pp_pmcs[ri].pp_pmc));
1616 
1617 			KASSERT(pp->pp_refcnt > 0,
1618 			    ("[pmc,%d] pp refcnt = %d", __LINE__,
1619 				pp->pp_refcnt));
1620 
1621 			pcd->pcd_read_pmc(cpu, adjri, &newvalue);
1622 
1623 			if (mode == PMC_MODE_TS) {
1624 				PMCDBG3(CSW,SWO,1,"cpu=%d ri=%d val=%jd (samp)",
1625 				    cpu, ri, newvalue);
1626 
1627 				if (pt == NULL)
1628 					pt = pmc_find_thread_descriptor(pp, td,
1629 					    PMC_FLAG_NONE);
1630 
1631 				KASSERT(pt != NULL,
1632 				    ("[pmc,%d] No thread found for td=%p",
1633 				    __LINE__, td));
1634 
1635 				mtx_pool_lock_spin(pmc_mtxpool, pm);
1636 
1637 				/*
1638 				 * If we have a thread descriptor, save the
1639 				 * per-thread counter in the descriptor. If not,
1640 				 * we will update the per-process counter.
1641 				 *
1642 				 * TODO: Remove the per-process "safety net"
1643 				 * once we have thoroughly tested that we
1644 				 * don't hit the above assert.
1645 				 */
1646 				if (pt != NULL)
1647 					pt->pt_pmcs[ri].pt_pmcval = newvalue;
1648 				else {
1649 					/*
1650 					 * For sampling process-virtual PMCs,
1651 					 * newvalue is the number of events to
1652 					 * be seen until the next sampling
1653 					 * interrupt. We can just add the events
1654 					 * left from this invocation to the
1655 					 * counter, then adjust in case we
1656 					 * overflow our range.
1657 					 *
1658 					 * (Recall that we reload the counter
1659 					 * every time we use it.)
1660 					 */
1661 					pp->pp_pmcs[ri].pp_pmcval += newvalue;
1662 					if (pp->pp_pmcs[ri].pp_pmcval >
1663 					    pm->pm_sc.pm_reloadcount)
1664 						pp->pp_pmcs[ri].pp_pmcval -=
1665 						    pm->pm_sc.pm_reloadcount;
1666 				}
1667 				mtx_pool_unlock_spin(pmc_mtxpool, pm);
1668 			} else {
1669 				tmp = newvalue - PMC_PCPU_SAVED(cpu,ri);
1670 
1671 				PMCDBG3(CSW,SWO,1,"cpu=%d ri=%d tmp=%jd (count)",
1672 				    cpu, ri, tmp);
1673 
1674 				/*
1675 				 * For counting process-virtual PMCs,
1676 				 * we expect the count to be
1677 				 * increasing monotonically, modulo a 64
1678 				 * bit wraparound.
1679 				 */
1680 				KASSERT(tmp >= 0,
1681 				    ("[pmc,%d] negative increment cpu=%d "
1682 				     "ri=%d newvalue=%jx saved=%jx "
1683 				     "incr=%jx", __LINE__, cpu, ri,
1684 				     newvalue, PMC_PCPU_SAVED(cpu,ri), tmp));
1685 
1686 				mtx_pool_lock_spin(pmc_mtxpool, pm);
1687 				pm->pm_gv.pm_savedvalue += tmp;
1688 				pp->pp_pmcs[ri].pp_pmcval += tmp;
1689 				mtx_pool_unlock_spin(pmc_mtxpool, pm);
1690 
1691 				if (pm->pm_flags & PMC_F_LOG_PROCCSW)
1692 					pmclog_process_proccsw(pm, pp, tmp, td);
1693 			}
1694 		}
1695 
1696 		/* mark hardware as free */
1697 		pcd->pcd_config_pmc(cpu, adjri, NULL);
1698 	}
1699 
1700 	/*
1701 	 * perform any other architecture/cpu dependent thread
1702 	 * switch out functions.
1703 	 */
1704 
1705 	(void) (*md->pmd_switch_out)(pc, pp);
1706 
1707 	critical_exit();
1708 }
1709 
1710 /*
1711  * A new thread for a process.
1712  */
1713 static void
1714 pmc_process_thread_add(struct thread *td)
1715 {
1716 	struct pmc_process *pmc;
1717 
1718 	pmc = pmc_find_process_descriptor(td->td_proc, PMC_FLAG_NONE);
1719 	if (pmc != NULL)
1720 		pmc_find_thread_descriptor(pmc, td, PMC_FLAG_ALLOCATE);
1721 }
1722 
1723 /*
1724  * A thread delete for a process.
1725  */
1726 static void
1727 pmc_process_thread_delete(struct thread *td)
1728 {
1729 	struct pmc_process *pmc;
1730 
1731 	pmc = pmc_find_process_descriptor(td->td_proc, PMC_FLAG_NONE);
1732 	if (pmc != NULL)
1733 		pmc_thread_descriptor_pool_free(pmc_find_thread_descriptor(pmc,
1734 		    td, PMC_FLAG_REMOVE));
1735 }
1736 
1737 /*
1738  * A userret() call for a thread.
1739  */
1740 static void
1741 pmc_process_thread_userret(struct thread *td)
1742 {
1743 	sched_pin();
1744 	pmc_capture_user_callchain(curcpu, PMC_UR, td->td_frame);
1745 	sched_unpin();
1746 }
1747 
1748 /*
1749  * A mapping change for a process.
1750  */
1751 
1752 static void
1753 pmc_process_mmap(struct thread *td, struct pmckern_map_in *pkm)
1754 {
1755 	int ri;
1756 	pid_t pid;
1757 	char *fullpath, *freepath;
1758 	const struct pmc *pm;
1759 	struct pmc_owner *po;
1760 	const struct pmc_process *pp;
1761 
1762 	freepath = fullpath = NULL;
1763 	MPASS(!in_epoch(global_epoch_preempt));
1764 	pmc_getfilename((struct vnode *) pkm->pm_file, &fullpath, &freepath);
1765 
1766 	pid = td->td_proc->p_pid;
1767 
1768 	PMC_EPOCH_ENTER();
1769 	/* Inform owners of all system-wide sampling PMCs. */
1770 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1771 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1772 			pmclog_process_map_in(po, pid, pkm->pm_address, fullpath);
1773 
1774 	if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL)
1775 		goto done;
1776 
1777 	/*
1778 	 * Inform sampling PMC owners tracking this process.
1779 	 */
1780 	for (ri = 0; ri < md->pmd_npmc; ri++)
1781 		if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL &&
1782 		    PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1783 			pmclog_process_map_in(pm->pm_owner,
1784 			    pid, pkm->pm_address, fullpath);
1785 
1786   done:
1787 	if (freepath)
1788 		free(freepath, M_TEMP);
1789 	PMC_EPOCH_EXIT();
1790 }
1791 
1792 
1793 /*
1794  * Log an munmap request.
1795  */
1796 
1797 static void
1798 pmc_process_munmap(struct thread *td, struct pmckern_map_out *pkm)
1799 {
1800 	int ri;
1801 	pid_t pid;
1802 	struct pmc_owner *po;
1803 	const struct pmc *pm;
1804 	const struct pmc_process *pp;
1805 
1806 	pid = td->td_proc->p_pid;
1807 
1808 	PMC_EPOCH_ENTER();
1809 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1810 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1811 		pmclog_process_map_out(po, pid, pkm->pm_address,
1812 		    pkm->pm_address + pkm->pm_size);
1813 	PMC_EPOCH_EXIT();
1814 
1815 	if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL)
1816 		return;
1817 
1818 	for (ri = 0; ri < md->pmd_npmc; ri++)
1819 		if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL &&
1820 		    PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1821 			pmclog_process_map_out(pm->pm_owner, pid,
1822 			    pkm->pm_address, pkm->pm_address + pkm->pm_size);
1823 }
1824 
1825 /*
1826  * Log mapping information about the kernel.
1827  */
1828 
1829 static void
1830 pmc_log_kernel_mappings(struct pmc *pm)
1831 {
1832 	struct pmc_owner *po;
1833 	struct pmckern_map_in *km, *kmbase;
1834 
1835 	MPASS(in_epoch(global_epoch_preempt) || sx_xlocked(&pmc_sx));
1836 	KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)),
1837 	    ("[pmc,%d] non-sampling PMC (%p) desires mapping information",
1838 		__LINE__, (void *) pm));
1839 
1840 	po = pm->pm_owner;
1841 
1842 	if (po->po_flags & PMC_PO_INITIAL_MAPPINGS_DONE)
1843 		return;
1844 	if (PMC_TO_MODE(pm) == PMC_MODE_SS)
1845 		pmc_process_allproc(pm);
1846 	/*
1847 	 * Log the current set of kernel modules.
1848 	 */
1849 	kmbase = linker_hwpmc_list_objects();
1850 	for (km = kmbase; km->pm_file != NULL; km++) {
1851 		PMCDBG2(LOG,REG,1,"%s %p", (char *) km->pm_file,
1852 		    (void *) km->pm_address);
1853 		pmclog_process_map_in(po, (pid_t) -1, km->pm_address,
1854 		    km->pm_file);
1855 	}
1856 	free(kmbase, M_LINKER);
1857 
1858 	po->po_flags |= PMC_PO_INITIAL_MAPPINGS_DONE;
1859 }
1860 
1861 /*
1862  * Log the mappings for a single process.
1863  */
1864 
1865 static void
1866 pmc_log_process_mappings(struct pmc_owner *po, struct proc *p)
1867 {
1868 	vm_map_t map;
1869 	struct vnode *vp;
1870 	struct vmspace *vm;
1871 	vm_map_entry_t entry;
1872 	vm_offset_t last_end;
1873 	u_int last_timestamp;
1874 	struct vnode *last_vp;
1875 	vm_offset_t start_addr;
1876 	vm_object_t obj, lobj, tobj;
1877 	char *fullpath, *freepath;
1878 
1879 	last_vp = NULL;
1880 	last_end = (vm_offset_t) 0;
1881 	fullpath = freepath = NULL;
1882 
1883 	if ((vm = vmspace_acquire_ref(p)) == NULL)
1884 		return;
1885 
1886 	map = &vm->vm_map;
1887 	vm_map_lock_read(map);
1888 
1889 	VM_MAP_ENTRY_FOREACH(entry, map) {
1890 
1891 		if (entry == NULL) {
1892 			PMCDBG2(LOG,OPS,2, "hwpmc: vm_map entry unexpectedly "
1893 			    "NULL! pid=%d vm_map=%p\n", p->p_pid, map);
1894 			break;
1895 		}
1896 
1897 		/*
1898 		 * We only care about executable map entries.
1899 		 */
1900 		if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) ||
1901 		    !(entry->protection & VM_PROT_EXECUTE) ||
1902 		    (entry->object.vm_object == NULL)) {
1903 			continue;
1904 		}
1905 
1906 		obj = entry->object.vm_object;
1907 		VM_OBJECT_RLOCK(obj);
1908 
1909 		/*
1910 		 * Walk the backing_object list to find the base
1911 		 * (non-shadowed) vm_object.
1912 		 */
1913 		for (lobj = tobj = obj; tobj != NULL; tobj = tobj->backing_object) {
1914 			if (tobj != obj)
1915 				VM_OBJECT_RLOCK(tobj);
1916 			if (lobj != obj)
1917 				VM_OBJECT_RUNLOCK(lobj);
1918 			lobj = tobj;
1919 		}
1920 
1921 		/*
1922 		 * At this point lobj is the base vm_object and it is locked.
1923 		 */
1924 		if (lobj == NULL) {
1925 			PMCDBG3(LOG,OPS,2, "hwpmc: lobj unexpectedly NULL! pid=%d "
1926 			    "vm_map=%p vm_obj=%p\n", p->p_pid, map, obj);
1927 			VM_OBJECT_RUNLOCK(obj);
1928 			continue;
1929 		}
1930 
1931 		vp = vm_object_vnode(lobj);
1932 		if (vp == NULL) {
1933 			if (lobj != obj)
1934 				VM_OBJECT_RUNLOCK(lobj);
1935 			VM_OBJECT_RUNLOCK(obj);
1936 			continue;
1937 		}
1938 
1939 		/*
1940 		 * Skip contiguous regions that point to the same
1941 		 * vnode, so we don't emit redundant MAP-IN
1942 		 * directives.
1943 		 */
1944 		if (entry->start == last_end && vp == last_vp) {
1945 			last_end = entry->end;
1946 			if (lobj != obj)
1947 				VM_OBJECT_RUNLOCK(lobj);
1948 			VM_OBJECT_RUNLOCK(obj);
1949 			continue;
1950 		}
1951 
1952 		/*
1953 		 * We don't want to keep the proc's vm_map or this
1954 		 * vm_object locked while we walk the pathname, since
1955 		 * vn_fullpath() can sleep.  However, if we drop the
1956 		 * lock, it's possible for concurrent activity to
1957 		 * modify the vm_map list.  To protect against this,
1958 		 * we save the vm_map timestamp before we release the
1959 		 * lock, and check it after we reacquire the lock
1960 		 * below.
1961 		 */
1962 		start_addr = entry->start;
1963 		last_end = entry->end;
1964 		last_timestamp = map->timestamp;
1965 		vm_map_unlock_read(map);
1966 
1967 		vref(vp);
1968 		if (lobj != obj)
1969 			VM_OBJECT_RUNLOCK(lobj);
1970 
1971 		VM_OBJECT_RUNLOCK(obj);
1972 
1973 		freepath = NULL;
1974 		pmc_getfilename(vp, &fullpath, &freepath);
1975 		last_vp = vp;
1976 
1977 		vrele(vp);
1978 
1979 		vp = NULL;
1980 		pmclog_process_map_in(po, p->p_pid, start_addr, fullpath);
1981 		if (freepath)
1982 			free(freepath, M_TEMP);
1983 
1984 		vm_map_lock_read(map);
1985 
1986 		/*
1987 		 * If our saved timestamp doesn't match, this means
1988 		 * that the vm_map was modified out from under us and
1989 		 * we can't trust our current "entry" pointer.  Do a
1990 		 * new lookup for this entry.  If there is no entry
1991 		 * for this address range, vm_map_lookup_entry() will
1992 		 * return the previous one, so we always want to go to
1993 		 * the next entry on the next loop iteration.
1994 		 *
1995 		 * There is an edge condition here that can occur if
1996 		 * there is no entry at or before this address.  In
1997 		 * this situation, vm_map_lookup_entry returns
1998 		 * &map->header, which would cause our loop to abort
1999 		 * without processing the rest of the map.  However,
2000 		 * in practice this will never happen for process
2001 		 * vm_map.  This is because the executable's text
2002 		 * segment is the first mapping in the proc's address
2003 		 * space, and this mapping is never removed until the
2004 		 * process exits, so there will always be a non-header
2005 		 * entry at or before the requested address for
2006 		 * vm_map_lookup_entry to return.
2007 		 */
2008 		if (map->timestamp != last_timestamp)
2009 			vm_map_lookup_entry(map, last_end - 1, &entry);
2010 	}
2011 
2012 	vm_map_unlock_read(map);
2013 	vmspace_free(vm);
2014 	return;
2015 }
2016 
2017 /*
2018  * Log mappings for all processes in the system.
2019  */
2020 
2021 static void
2022 pmc_log_all_process_mappings(struct pmc_owner *po)
2023 {
2024 	struct proc *p, *top;
2025 
2026 	sx_assert(&pmc_sx, SX_XLOCKED);
2027 
2028 	if ((p = pfind(1)) == NULL)
2029 		panic("[pmc,%d] Cannot find init", __LINE__);
2030 
2031 	PROC_UNLOCK(p);
2032 
2033 	sx_slock(&proctree_lock);
2034 
2035 	top = p;
2036 
2037 	for (;;) {
2038 		pmc_log_process_mappings(po, p);
2039 		if (!LIST_EMPTY(&p->p_children))
2040 			p = LIST_FIRST(&p->p_children);
2041 		else for (;;) {
2042 			if (p == top)
2043 				goto done;
2044 			if (LIST_NEXT(p, p_sibling)) {
2045 				p = LIST_NEXT(p, p_sibling);
2046 				break;
2047 			}
2048 			p = p->p_pptr;
2049 		}
2050 	}
2051  done:
2052 	sx_sunlock(&proctree_lock);
2053 }
2054 
2055 /*
2056  * The 'hook' invoked from the kernel proper
2057  */
2058 
2059 
2060 #ifdef	HWPMC_DEBUG
2061 const char *pmc_hooknames[] = {
2062 	/* these strings correspond to PMC_FN_* in <sys/pmckern.h> */
2063 	"",
2064 	"EXEC",
2065 	"CSW-IN",
2066 	"CSW-OUT",
2067 	"SAMPLE",
2068 	"UNUSED1",
2069 	"UNUSED2",
2070 	"MMAP",
2071 	"MUNMAP",
2072 	"CALLCHAIN-NMI",
2073 	"CALLCHAIN-SOFT",
2074 	"SOFTSAMPLING",
2075 	"THR-CREATE",
2076 	"THR-EXIT",
2077 	"THR-USERRET",
2078 	"THR-CREATE-LOG",
2079 	"THR-EXIT-LOG",
2080 	"PROC-CREATE-LOG"
2081 };
2082 #endif
2083 
2084 static int
2085 pmc_hook_handler(struct thread *td, int function, void *arg)
2086 {
2087 	int cpu;
2088 
2089 	PMCDBG4(MOD,PMH,1, "hook td=%p func=%d \"%s\" arg=%p", td, function,
2090 	    pmc_hooknames[function], arg);
2091 
2092 	switch (function)
2093 	{
2094 
2095 	/*
2096 	 * Process exec()
2097 	 */
2098 
2099 	case PMC_FN_PROCESS_EXEC:
2100 	{
2101 		char *fullpath, *freepath;
2102 		unsigned int ri;
2103 		int is_using_hwpmcs;
2104 		struct pmc *pm;
2105 		struct proc *p;
2106 		struct pmc_owner *po;
2107 		struct pmc_process *pp;
2108 		struct pmckern_procexec *pk;
2109 
2110 		sx_assert(&pmc_sx, SX_XLOCKED);
2111 
2112 		p = td->td_proc;
2113 		pmc_getfilename(p->p_textvp, &fullpath, &freepath);
2114 
2115 		pk = (struct pmckern_procexec *) arg;
2116 
2117 		PMC_EPOCH_ENTER();
2118 		/* Inform owners of SS mode PMCs of the exec event. */
2119 		CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
2120 		    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
2121 			    pmclog_process_procexec(po, PMC_ID_INVALID,
2122 				p->p_pid, pk->pm_entryaddr, fullpath);
2123 		PMC_EPOCH_EXIT();
2124 
2125 		PROC_LOCK(p);
2126 		is_using_hwpmcs = p->p_flag & P_HWPMC;
2127 		PROC_UNLOCK(p);
2128 
2129 		if (!is_using_hwpmcs) {
2130 			if (freepath)
2131 				free(freepath, M_TEMP);
2132 			break;
2133 		}
2134 
2135 		/*
2136 		 * PMCs are not inherited across an exec():  remove any
2137 		 * PMCs that this process is the owner of.
2138 		 */
2139 
2140 		if ((po = pmc_find_owner_descriptor(p)) != NULL) {
2141 			pmc_remove_owner(po);
2142 			pmc_destroy_owner_descriptor(po);
2143 		}
2144 
2145 		/*
2146 		 * If the process being exec'ed is not the target of any
2147 		 * PMC, we are done.
2148 		 */
2149 		if ((pp = pmc_find_process_descriptor(p, 0)) == NULL) {
2150 			if (freepath)
2151 				free(freepath, M_TEMP);
2152 			break;
2153 		}
2154 
2155 		/*
2156 		 * Log the exec event to all monitoring owners.  Skip
2157 		 * owners who have already received the event because
2158 		 * they had system sampling PMCs active.
2159 		 */
2160 		for (ri = 0; ri < md->pmd_npmc; ri++)
2161 			if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) {
2162 				po = pm->pm_owner;
2163 				if (po->po_sscount == 0 &&
2164 				    po->po_flags & PMC_PO_OWNS_LOGFILE)
2165 					pmclog_process_procexec(po, pm->pm_id,
2166 					    p->p_pid, pk->pm_entryaddr,
2167 					    fullpath);
2168 			}
2169 
2170 		if (freepath)
2171 			free(freepath, M_TEMP);
2172 
2173 
2174 		PMCDBG4(PRC,EXC,1, "exec proc=%p (%d, %s) cred-changed=%d",
2175 		    p, p->p_pid, p->p_comm, pk->pm_credentialschanged);
2176 
2177 		if (pk->pm_credentialschanged == 0) /* no change */
2178 			break;
2179 
2180 		/*
2181 		 * If the newly exec()'ed process has a different credential
2182 		 * than before, allow it to be the target of a PMC only if
2183 		 * the PMC's owner has sufficient privilege.
2184 		 */
2185 
2186 		for (ri = 0; ri < md->pmd_npmc; ri++)
2187 			if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL)
2188 				if (pmc_can_attach(pm, td->td_proc) != 0)
2189 					pmc_detach_one_process(td->td_proc,
2190 					    pm, PMC_FLAG_NONE);
2191 
2192 		KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= (int) md->pmd_npmc,
2193 		    ("[pmc,%d] Illegal ref count %d on pp %p", __LINE__,
2194 			pp->pp_refcnt, pp));
2195 
2196 		/*
2197 		 * If this process is no longer the target of any
2198 		 * PMCs, we can remove the process entry and free
2199 		 * up space.
2200 		 */
2201 
2202 		if (pp->pp_refcnt == 0) {
2203 			pmc_remove_process_descriptor(pp);
2204 			pmc_destroy_process_descriptor(pp);
2205 			break;
2206 		}
2207 
2208 	}
2209 	break;
2210 
2211 	case PMC_FN_CSW_IN:
2212 		pmc_process_csw_in(td);
2213 		break;
2214 
2215 	case PMC_FN_CSW_OUT:
2216 		pmc_process_csw_out(td);
2217 		break;
2218 
2219 	/*
2220 	 * Process accumulated PC samples.
2221 	 *
2222 	 * This function is expected to be called by hardclock() for
2223 	 * each CPU that has accumulated PC samples.
2224 	 *
2225 	 * This function is to be executed on the CPU whose samples
2226 	 * are being processed.
2227 	 */
2228 	case PMC_FN_DO_SAMPLES:
2229 
2230 		/*
2231 		 * Clear the cpu specific bit in the CPU mask before
2232 		 * do the rest of the processing.  If the NMI handler
2233 		 * gets invoked after the "atomic_clear_int()" call
2234 		 * below but before "pmc_process_samples()" gets
2235 		 * around to processing the interrupt, then we will
2236 		 * come back here at the next hardclock() tick (and
2237 		 * may find nothing to do if "pmc_process_samples()"
2238 		 * had already processed the interrupt).  We don't
2239 		 * lose the interrupt sample.
2240 		 */
2241 		DPCPU_SET(pmc_sampled, 0);
2242 		cpu = PCPU_GET(cpuid);
2243 		pmc_process_samples(cpu, PMC_HR);
2244 		pmc_process_samples(cpu, PMC_SR);
2245 		pmc_process_samples(cpu, PMC_UR);
2246 		break;
2247 
2248 	case PMC_FN_MMAP:
2249 		pmc_process_mmap(td, (struct pmckern_map_in *) arg);
2250 		break;
2251 
2252 	case PMC_FN_MUNMAP:
2253 		MPASS(in_epoch(global_epoch_preempt) || sx_xlocked(&pmc_sx));
2254 		pmc_process_munmap(td, (struct pmckern_map_out *) arg);
2255 		break;
2256 
2257 	case PMC_FN_PROC_CREATE_LOG:
2258 		pmc_process_proccreate((struct proc *)arg);
2259 		break;
2260 
2261 	case PMC_FN_USER_CALLCHAIN:
2262 		/*
2263 		 * Record a call chain.
2264 		 */
2265 		KASSERT(td == curthread, ("[pmc,%d] td != curthread",
2266 		    __LINE__));
2267 
2268 		pmc_capture_user_callchain(PCPU_GET(cpuid), PMC_HR,
2269 		    (struct trapframe *) arg);
2270 
2271 		KASSERT(td->td_pinned == 1,
2272 			("[pmc,%d] invalid td_pinned value", __LINE__));
2273 		sched_unpin();  /* Can migrate safely now. */
2274 
2275 		td->td_pflags &= ~TDP_CALLCHAIN;
2276 		break;
2277 
2278 	case PMC_FN_USER_CALLCHAIN_SOFT:
2279 		/*
2280 		 * Record a call chain.
2281 		 */
2282 		KASSERT(td == curthread, ("[pmc,%d] td != curthread",
2283 		    __LINE__));
2284 
2285 		cpu = PCPU_GET(cpuid);
2286 		pmc_capture_user_callchain(cpu, PMC_SR,
2287 		    (struct trapframe *) arg);
2288 
2289 		KASSERT(td->td_pinned == 1,
2290 		    ("[pmc,%d] invalid td_pinned value", __LINE__));
2291 
2292 		sched_unpin();  /* Can migrate safely now. */
2293 
2294 		td->td_pflags &= ~TDP_CALLCHAIN;
2295 		break;
2296 
2297 	case PMC_FN_SOFT_SAMPLING:
2298 		/*
2299 		 * Call soft PMC sampling intr.
2300 		 */
2301 		pmc_soft_intr((struct pmckern_soft *) arg);
2302 		break;
2303 
2304 	case PMC_FN_THR_CREATE:
2305 		pmc_process_thread_add(td);
2306 		pmc_process_threadcreate(td);
2307 		break;
2308 
2309 	case PMC_FN_THR_CREATE_LOG:
2310 		pmc_process_threadcreate(td);
2311 		break;
2312 
2313 	case PMC_FN_THR_EXIT:
2314 		KASSERT(td == curthread, ("[pmc,%d] td != curthread",
2315 		    __LINE__));
2316 		pmc_process_thread_delete(td);
2317 		pmc_process_threadexit(td);
2318 		break;
2319 	case PMC_FN_THR_EXIT_LOG:
2320 		pmc_process_threadexit(td);
2321 		break;
2322 	case PMC_FN_THR_USERRET:
2323 		KASSERT(td == curthread, ("[pmc,%d] td != curthread",
2324 		    __LINE__));
2325 		pmc_process_thread_userret(td);
2326 		break;
2327 
2328 	default:
2329 #ifdef	HWPMC_DEBUG
2330 		KASSERT(0, ("[pmc,%d] unknown hook %d\n", __LINE__, function));
2331 #endif
2332 		break;
2333 
2334 	}
2335 
2336 	return 0;
2337 }
2338 
2339 /*
2340  * allocate a 'struct pmc_owner' descriptor in the owner hash table.
2341  */
2342 
2343 static struct pmc_owner *
2344 pmc_allocate_owner_descriptor(struct proc *p)
2345 {
2346 	uint32_t hindex;
2347 	struct pmc_owner *po;
2348 	struct pmc_ownerhash *poh;
2349 
2350 	hindex = PMC_HASH_PTR(p, pmc_ownerhashmask);
2351 	poh = &pmc_ownerhash[hindex];
2352 
2353 	/* allocate space for N pointers and one descriptor struct */
2354 	po = malloc(sizeof(struct pmc_owner), M_PMC, M_WAITOK|M_ZERO);
2355 	po->po_owner = p;
2356 	LIST_INSERT_HEAD(poh, po, po_next); /* insert into hash table */
2357 
2358 	TAILQ_INIT(&po->po_logbuffers);
2359 	mtx_init(&po->po_mtx, "pmc-owner-mtx", "pmc-per-proc", MTX_SPIN);
2360 
2361 	PMCDBG4(OWN,ALL,1, "allocate-owner proc=%p (%d, %s) pmc-owner=%p",
2362 	    p, p->p_pid, p->p_comm, po);
2363 
2364 	return po;
2365 }
2366 
2367 static void
2368 pmc_destroy_owner_descriptor(struct pmc_owner *po)
2369 {
2370 
2371 	PMCDBG4(OWN,REL,1, "destroy-owner po=%p proc=%p (%d, %s)",
2372 	    po, po->po_owner, po->po_owner->p_pid, po->po_owner->p_comm);
2373 
2374 	mtx_destroy(&po->po_mtx);
2375 	free(po, M_PMC);
2376 }
2377 
2378 /*
2379  * Allocate a thread descriptor from the free pool.
2380  *
2381  * NOTE: This *can* return NULL.
2382  */
2383 static struct pmc_thread *
2384 pmc_thread_descriptor_pool_alloc(void)
2385 {
2386 	struct pmc_thread *pt;
2387 
2388 	mtx_lock_spin(&pmc_threadfreelist_mtx);
2389 	if ((pt = LIST_FIRST(&pmc_threadfreelist)) != NULL) {
2390 		LIST_REMOVE(pt, pt_next);
2391 		pmc_threadfreelist_entries--;
2392 	}
2393 	mtx_unlock_spin(&pmc_threadfreelist_mtx);
2394 
2395 	return (pt);
2396 }
2397 
2398 /*
2399  * Add a thread descriptor to the free pool. We use this instead of free()
2400  * to maintain a cache of free entries. Additionally, we can safely call
2401  * this function when we cannot call free(), such as in a critical section.
2402  *
2403  */
2404 static void
2405 pmc_thread_descriptor_pool_free(struct pmc_thread *pt)
2406 {
2407 
2408 	if (pt == NULL)
2409 		return;
2410 
2411 	memset(pt, 0, THREADENTRY_SIZE);
2412 	mtx_lock_spin(&pmc_threadfreelist_mtx);
2413 	LIST_INSERT_HEAD(&pmc_threadfreelist, pt, pt_next);
2414 	pmc_threadfreelist_entries++;
2415 	if (pmc_threadfreelist_entries > pmc_threadfreelist_max)
2416 		taskqueue_enqueue(taskqueue_fast, &free_task);
2417 	mtx_unlock_spin(&pmc_threadfreelist_mtx);
2418 }
2419 
2420 /*
2421  * An asynchronous task to manage the free list.
2422  */
2423 static void
2424 pmc_thread_descriptor_pool_free_task(void *arg __unused, int pending __unused)
2425 {
2426 	struct pmc_thread *pt;
2427 	LIST_HEAD(, pmc_thread) tmplist;
2428 	int delta;
2429 
2430 	LIST_INIT(&tmplist);
2431 
2432 	/* Determine what changes, if any, we need to make. */
2433 	mtx_lock_spin(&pmc_threadfreelist_mtx);
2434 	delta = pmc_threadfreelist_entries - pmc_threadfreelist_max;
2435 	while (delta > 0 && (pt = LIST_FIRST(&pmc_threadfreelist)) != NULL) {
2436 		delta--;
2437 		pmc_threadfreelist_entries--;
2438 		LIST_REMOVE(pt, pt_next);
2439 		LIST_INSERT_HEAD(&tmplist, pt, pt_next);
2440 	}
2441 	mtx_unlock_spin(&pmc_threadfreelist_mtx);
2442 
2443 	/* If there are entries to free, free them. */
2444 	while (!LIST_EMPTY(&tmplist)) {
2445 		pt = LIST_FIRST(&tmplist);
2446 		LIST_REMOVE(pt, pt_next);
2447 		free(pt, M_PMC);
2448 	}
2449 }
2450 
2451 /*
2452  * Drain the thread free pool, freeing all allocations.
2453  */
2454 static void
2455 pmc_thread_descriptor_pool_drain()
2456 {
2457 	struct pmc_thread *pt, *next;
2458 
2459 	LIST_FOREACH_SAFE(pt, &pmc_threadfreelist, pt_next, next) {
2460 		LIST_REMOVE(pt, pt_next);
2461 		free(pt, M_PMC);
2462 	}
2463 }
2464 
2465 /*
2466  * find the descriptor corresponding to thread 'td', adding or removing it
2467  * as specified by 'mode'.
2468  *
2469  * Note that this supports additional mode flags in addition to those
2470  * supported by pmc_find_process_descriptor():
2471  * PMC_FLAG_NOWAIT: Causes the function to not wait for mallocs.
2472  *     This makes it safe to call while holding certain other locks.
2473  */
2474 
2475 static struct pmc_thread *
2476 pmc_find_thread_descriptor(struct pmc_process *pp, struct thread *td,
2477     uint32_t mode)
2478 {
2479 	struct pmc_thread *pt = NULL, *ptnew = NULL;
2480 	int wait_flag;
2481 
2482 	KASSERT(td != NULL, ("[pmc,%d] called to add NULL td", __LINE__));
2483 
2484 	/*
2485 	 * Pre-allocate memory in the PMC_FLAG_ALLOCATE case prior to
2486 	 * acquiring the lock.
2487 	 */
2488 	if (mode & PMC_FLAG_ALLOCATE) {
2489 		if ((ptnew = pmc_thread_descriptor_pool_alloc()) == NULL) {
2490 			wait_flag = M_WAITOK;
2491 			if ((mode & PMC_FLAG_NOWAIT) || in_epoch(global_epoch_preempt))
2492 				wait_flag = M_NOWAIT;
2493 
2494 			ptnew = malloc(THREADENTRY_SIZE, M_PMC,
2495 			    wait_flag|M_ZERO);
2496 		}
2497 	}
2498 
2499 	mtx_lock_spin(pp->pp_tdslock);
2500 
2501 	LIST_FOREACH(pt, &pp->pp_tds, pt_next)
2502 		if (pt->pt_td == td)
2503 			break;
2504 
2505 	if ((mode & PMC_FLAG_REMOVE) && pt != NULL)
2506 		LIST_REMOVE(pt, pt_next);
2507 
2508 	if ((mode & PMC_FLAG_ALLOCATE) && pt == NULL && ptnew != NULL) {
2509 		pt = ptnew;
2510 		ptnew = NULL;
2511 		pt->pt_td = td;
2512 		LIST_INSERT_HEAD(&pp->pp_tds, pt, pt_next);
2513 	}
2514 
2515 	mtx_unlock_spin(pp->pp_tdslock);
2516 
2517 	if (ptnew != NULL) {
2518 		free(ptnew, M_PMC);
2519 	}
2520 
2521 	return pt;
2522 }
2523 
2524 /*
2525  * Try to add thread descriptors for each thread in a process.
2526  */
2527 
2528 static void
2529 pmc_add_thread_descriptors_from_proc(struct proc *p, struct pmc_process *pp)
2530 {
2531 	struct thread *curtd;
2532 	struct pmc_thread **tdlist;
2533 	int i, tdcnt, tdlistsz;
2534 
2535 	KASSERT(!PROC_LOCKED(p), ("[pmc,%d] proc unexpectedly locked",
2536 	    __LINE__));
2537 	tdcnt = 32;
2538  restart:
2539 	tdlistsz = roundup2(tdcnt, 32);
2540 
2541 	tdcnt = 0;
2542 	tdlist = malloc(sizeof(struct pmc_thread*) * tdlistsz, M_TEMP, M_WAITOK);
2543 
2544 	PROC_LOCK(p);
2545 	FOREACH_THREAD_IN_PROC(p, curtd)
2546 		tdcnt++;
2547 	if (tdcnt >= tdlistsz) {
2548 		PROC_UNLOCK(p);
2549 		free(tdlist, M_TEMP);
2550 		goto restart;
2551 	}
2552 	/*
2553 	 * Try to add each thread to the list without sleeping. If unable,
2554 	 * add to a queue to retry after dropping the process lock.
2555 	 */
2556 	tdcnt = 0;
2557 	FOREACH_THREAD_IN_PROC(p, curtd) {
2558 		tdlist[tdcnt] = pmc_find_thread_descriptor(pp, curtd,
2559 						   PMC_FLAG_ALLOCATE|PMC_FLAG_NOWAIT);
2560 		if (tdlist[tdcnt] == NULL) {
2561 			PROC_UNLOCK(p);
2562 			for (i = 0; i <= tdcnt; i++)
2563 				pmc_thread_descriptor_pool_free(tdlist[i]);
2564 			free(tdlist, M_TEMP);
2565 			goto restart;
2566 		}
2567 		tdcnt++;
2568 	}
2569 	PROC_UNLOCK(p);
2570 	free(tdlist, M_TEMP);
2571 }
2572 
2573 /*
2574  * find the descriptor corresponding to process 'p', adding or removing it
2575  * as specified by 'mode'.
2576  */
2577 
2578 static struct pmc_process *
2579 pmc_find_process_descriptor(struct proc *p, uint32_t mode)
2580 {
2581 	uint32_t hindex;
2582 	struct pmc_process *pp, *ppnew;
2583 	struct pmc_processhash *pph;
2584 
2585 	hindex = PMC_HASH_PTR(p, pmc_processhashmask);
2586 	pph = &pmc_processhash[hindex];
2587 
2588 	ppnew = NULL;
2589 
2590 	/*
2591 	 * Pre-allocate memory in the PMC_FLAG_ALLOCATE case since we
2592 	 * cannot call malloc(9) once we hold a spin lock.
2593 	 */
2594 	if (mode & PMC_FLAG_ALLOCATE)
2595 		ppnew = malloc(sizeof(struct pmc_process) + md->pmd_npmc *
2596 		    sizeof(struct pmc_targetstate), M_PMC, M_WAITOK|M_ZERO);
2597 
2598 	mtx_lock_spin(&pmc_processhash_mtx);
2599 	LIST_FOREACH(pp, pph, pp_next)
2600 	    if (pp->pp_proc == p)
2601 		    break;
2602 
2603 	if ((mode & PMC_FLAG_REMOVE) && pp != NULL)
2604 		LIST_REMOVE(pp, pp_next);
2605 
2606 	if ((mode & PMC_FLAG_ALLOCATE) && pp == NULL &&
2607 	    ppnew != NULL) {
2608 		ppnew->pp_proc = p;
2609 		LIST_INIT(&ppnew->pp_tds);
2610 		ppnew->pp_tdslock = mtx_pool_find(pmc_mtxpool, ppnew);
2611 		LIST_INSERT_HEAD(pph, ppnew, pp_next);
2612 		mtx_unlock_spin(&pmc_processhash_mtx);
2613 		pp = ppnew;
2614 		ppnew = NULL;
2615 
2616 		/* Add thread descriptors for this process' current threads. */
2617 		pmc_add_thread_descriptors_from_proc(p, pp);
2618 	}
2619 	else
2620 		mtx_unlock_spin(&pmc_processhash_mtx);
2621 
2622 	if (ppnew != NULL)
2623 		free(ppnew, M_PMC);
2624 
2625 	return pp;
2626 }
2627 
2628 /*
2629  * remove a process descriptor from the process hash table.
2630  */
2631 
2632 static void
2633 pmc_remove_process_descriptor(struct pmc_process *pp)
2634 {
2635 	KASSERT(pp->pp_refcnt == 0,
2636 	    ("[pmc,%d] Removing process descriptor %p with count %d",
2637 		__LINE__, pp, pp->pp_refcnt));
2638 
2639 	mtx_lock_spin(&pmc_processhash_mtx);
2640 	LIST_REMOVE(pp, pp_next);
2641 	mtx_unlock_spin(&pmc_processhash_mtx);
2642 }
2643 
2644 /*
2645  * destroy a process descriptor.
2646  */
2647 
2648 static void
2649 pmc_destroy_process_descriptor(struct pmc_process *pp)
2650 {
2651 	struct pmc_thread *pmc_td;
2652 
2653 	while ((pmc_td = LIST_FIRST(&pp->pp_tds)) != NULL) {
2654 		LIST_REMOVE(pmc_td, pt_next);
2655 		pmc_thread_descriptor_pool_free(pmc_td);
2656 	}
2657 	free(pp, M_PMC);
2658 }
2659 
2660 
2661 /*
2662  * find an owner descriptor corresponding to proc 'p'
2663  */
2664 
2665 static struct pmc_owner *
2666 pmc_find_owner_descriptor(struct proc *p)
2667 {
2668 	uint32_t hindex;
2669 	struct pmc_owner *po;
2670 	struct pmc_ownerhash *poh;
2671 
2672 	hindex = PMC_HASH_PTR(p, pmc_ownerhashmask);
2673 	poh = &pmc_ownerhash[hindex];
2674 
2675 	po = NULL;
2676 	LIST_FOREACH(po, poh, po_next)
2677 	    if (po->po_owner == p)
2678 		    break;
2679 
2680 	PMCDBG5(OWN,FND,1, "find-owner proc=%p (%d, %s) hindex=0x%x -> "
2681 	    "pmc-owner=%p", p, p->p_pid, p->p_comm, hindex, po);
2682 
2683 	return po;
2684 }
2685 
2686 /*
2687  * pmc_allocate_pmc_descriptor
2688  *
2689  * Allocate a pmc descriptor and initialize its
2690  * fields.
2691  */
2692 
2693 static struct pmc *
2694 pmc_allocate_pmc_descriptor(void)
2695 {
2696 	struct pmc *pmc;
2697 
2698 	pmc = malloc(sizeof(struct pmc), M_PMC, M_WAITOK|M_ZERO);
2699 	pmc->pm_runcount = counter_u64_alloc(M_WAITOK);
2700 	pmc->pm_pcpu_state = malloc(sizeof(struct pmc_pcpu_state)*mp_ncpus, M_PMC, M_WAITOK|M_ZERO);
2701 	PMCDBG1(PMC,ALL,1, "allocate-pmc -> pmc=%p", pmc);
2702 
2703 	return pmc;
2704 }
2705 
2706 /*
2707  * Destroy a pmc descriptor.
2708  */
2709 
2710 static void
2711 pmc_destroy_pmc_descriptor(struct pmc *pm)
2712 {
2713 
2714 	KASSERT(pm->pm_state == PMC_STATE_DELETED ||
2715 	    pm->pm_state == PMC_STATE_FREE,
2716 	    ("[pmc,%d] destroying non-deleted PMC", __LINE__));
2717 	KASSERT(LIST_EMPTY(&pm->pm_targets),
2718 	    ("[pmc,%d] destroying pmc with targets", __LINE__));
2719 	KASSERT(pm->pm_owner == NULL,
2720 	    ("[pmc,%d] destroying pmc attached to an owner", __LINE__));
2721 	KASSERT(counter_u64_fetch(pm->pm_runcount) == 0,
2722 	    ("[pmc,%d] pmc has non-zero run count %ld", __LINE__,
2723 		 (unsigned long)counter_u64_fetch(pm->pm_runcount)));
2724 
2725 	counter_u64_free(pm->pm_runcount);
2726 	free(pm->pm_pcpu_state, M_PMC);
2727 	free(pm, M_PMC);
2728 }
2729 
2730 static void
2731 pmc_wait_for_pmc_idle(struct pmc *pm)
2732 {
2733 #ifdef INVARIANTS
2734 	volatile int maxloop;
2735 
2736 	maxloop = 100 * pmc_cpu_max();
2737 #endif
2738 	/*
2739 	 * Loop (with a forced context switch) till the PMC's runcount
2740 	 * comes down to zero.
2741 	 */
2742 	pmclog_flush(pm->pm_owner, 1);
2743 	while (counter_u64_fetch(pm->pm_runcount) > 0) {
2744 		pmclog_flush(pm->pm_owner, 1);
2745 #ifdef INVARIANTS
2746 		maxloop--;
2747 		KASSERT(maxloop > 0,
2748 		    ("[pmc,%d] (ri%d, rc%ld) waiting too long for "
2749 			"pmc to be free", __LINE__,
2750 			 PMC_TO_ROWINDEX(pm), (unsigned long)counter_u64_fetch(pm->pm_runcount)));
2751 #endif
2752 		pmc_force_context_switch();
2753 	}
2754 }
2755 
2756 /*
2757  * This function does the following things:
2758  *
2759  *  - detaches the PMC from hardware
2760  *  - unlinks all target threads that were attached to it
2761  *  - removes the PMC from its owner's list
2762  *  - destroys the PMC private mutex
2763  *
2764  * Once this function completes, the given pmc pointer can be freed by
2765  * calling pmc_destroy_pmc_descriptor().
2766  */
2767 
2768 static void
2769 pmc_release_pmc_descriptor(struct pmc *pm)
2770 {
2771 	enum pmc_mode mode;
2772 	struct pmc_hw *phw __diagused;
2773 	u_int adjri, ri, cpu;
2774 	struct pmc_owner *po;
2775 	struct pmc_binding pb;
2776 	struct pmc_process *pp;
2777 	struct pmc_classdep *pcd;
2778 	struct pmc_target *ptgt, *tmp;
2779 
2780 	sx_assert(&pmc_sx, SX_XLOCKED);
2781 
2782 	KASSERT(pm, ("[pmc,%d] null pmc", __LINE__));
2783 
2784 	ri   = PMC_TO_ROWINDEX(pm);
2785 	pcd  = pmc_ri_to_classdep(md, ri, &adjri);
2786 	mode = PMC_TO_MODE(pm);
2787 
2788 	PMCDBG3(PMC,REL,1, "release-pmc pmc=%p ri=%d mode=%d", pm, ri,
2789 	    mode);
2790 
2791 	/*
2792 	 * First, we take the PMC off hardware.
2793 	 */
2794 	cpu = 0;
2795 	if (PMC_IS_SYSTEM_MODE(mode)) {
2796 
2797 		/*
2798 		 * A system mode PMC runs on a specific CPU.  Switch
2799 		 * to this CPU and turn hardware off.
2800 		 */
2801 		pmc_save_cpu_binding(&pb);
2802 
2803 		cpu = PMC_TO_CPU(pm);
2804 
2805 		pmc_select_cpu(cpu);
2806 
2807 		/* switch off non-stalled CPUs */
2808 		pm->pm_pcpu_state[cpu].pps_cpustate = 0;
2809 		if (pm->pm_state == PMC_STATE_RUNNING &&
2810 			pm->pm_pcpu_state[cpu].pps_stalled == 0) {
2811 
2812 			phw = pmc_pcpu[cpu]->pc_hwpmcs[ri];
2813 
2814 			KASSERT(phw->phw_pmc == pm,
2815 			    ("[pmc, %d] pmc ptr ri(%d) hw(%p) pm(%p)",
2816 				__LINE__, ri, phw->phw_pmc, pm));
2817 			PMCDBG2(PMC,REL,2, "stopping cpu=%d ri=%d", cpu, ri);
2818 
2819 			critical_enter();
2820 			pcd->pcd_stop_pmc(cpu, adjri);
2821 			critical_exit();
2822 		}
2823 
2824 		PMCDBG2(PMC,REL,2, "decfg cpu=%d ri=%d", cpu, ri);
2825 
2826 		critical_enter();
2827 		pcd->pcd_config_pmc(cpu, adjri, NULL);
2828 		critical_exit();
2829 
2830 		/* adjust the global and process count of SS mode PMCs */
2831 		if (mode == PMC_MODE_SS && pm->pm_state == PMC_STATE_RUNNING) {
2832 			po = pm->pm_owner;
2833 			po->po_sscount--;
2834 			if (po->po_sscount == 0) {
2835 				atomic_subtract_rel_int(&pmc_ss_count, 1);
2836 				CK_LIST_REMOVE(po, po_ssnext);
2837 				epoch_wait_preempt(global_epoch_preempt);
2838 			}
2839 		}
2840 
2841 		pm->pm_state = PMC_STATE_DELETED;
2842 
2843 		pmc_restore_cpu_binding(&pb);
2844 
2845 		/*
2846 		 * We could have references to this PMC structure in
2847 		 * the per-cpu sample queues.  Wait for the queue to
2848 		 * drain.
2849 		 */
2850 		pmc_wait_for_pmc_idle(pm);
2851 
2852 	} else if (PMC_IS_VIRTUAL_MODE(mode)) {
2853 
2854 		/*
2855 		 * A virtual PMC could be running on multiple CPUs at
2856 		 * a given instant.
2857 		 *
2858 		 * By marking its state as DELETED, we ensure that
2859 		 * this PMC is never further scheduled on hardware.
2860 		 *
2861 		 * Then we wait till all CPUs are done with this PMC.
2862 		 */
2863 		pm->pm_state = PMC_STATE_DELETED;
2864 
2865 
2866 		/* Wait for the PMCs runcount to come to zero. */
2867 		pmc_wait_for_pmc_idle(pm);
2868 
2869 		/*
2870 		 * At this point the PMC is off all CPUs and cannot be
2871 		 * freshly scheduled onto a CPU.  It is now safe to
2872 		 * unlink all targets from this PMC.  If a
2873 		 * process-record's refcount falls to zero, we remove
2874 		 * it from the hash table.  The module-wide SX lock
2875 		 * protects us from races.
2876 		 */
2877 		LIST_FOREACH_SAFE(ptgt, &pm->pm_targets, pt_next, tmp) {
2878 			pp = ptgt->pt_process;
2879 			pmc_unlink_target_process(pm, pp); /* frees 'ptgt' */
2880 
2881 			PMCDBG1(PMC,REL,3, "pp->refcnt=%d", pp->pp_refcnt);
2882 
2883 			/*
2884 			 * If the target process record shows that no
2885 			 * PMCs are attached to it, reclaim its space.
2886 			 */
2887 
2888 			if (pp->pp_refcnt == 0) {
2889 				pmc_remove_process_descriptor(pp);
2890 				pmc_destroy_process_descriptor(pp);
2891 			}
2892 		}
2893 
2894 		cpu = curthread->td_oncpu; /* setup cpu for pmd_release() */
2895 
2896 	}
2897 
2898 	/*
2899 	 * Release any MD resources
2900 	 */
2901 	(void) pcd->pcd_release_pmc(cpu, adjri, pm);
2902 
2903 	/*
2904 	 * Update row disposition
2905 	 */
2906 
2907 	if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm)))
2908 		PMC_UNMARK_ROW_STANDALONE(ri);
2909 	else
2910 		PMC_UNMARK_ROW_THREAD(ri);
2911 
2912 	/* unlink from the owner's list */
2913 	if (pm->pm_owner) {
2914 		LIST_REMOVE(pm, pm_next);
2915 		pm->pm_owner = NULL;
2916 	}
2917 }
2918 
2919 /*
2920  * Register an owner and a pmc.
2921  */
2922 
2923 static int
2924 pmc_register_owner(struct proc *p, struct pmc *pmc)
2925 {
2926 	struct pmc_owner *po;
2927 
2928 	sx_assert(&pmc_sx, SX_XLOCKED);
2929 
2930 	if ((po = pmc_find_owner_descriptor(p)) == NULL)
2931 		if ((po = pmc_allocate_owner_descriptor(p)) == NULL)
2932 			return ENOMEM;
2933 
2934 	KASSERT(pmc->pm_owner == NULL,
2935 	    ("[pmc,%d] attempting to own an initialized PMC", __LINE__));
2936 	pmc->pm_owner  = po;
2937 
2938 	LIST_INSERT_HEAD(&po->po_pmcs, pmc, pm_next);
2939 
2940 	PROC_LOCK(p);
2941 	p->p_flag |= P_HWPMC;
2942 	PROC_UNLOCK(p);
2943 
2944 	if (po->po_flags & PMC_PO_OWNS_LOGFILE)
2945 		pmclog_process_pmcallocate(pmc);
2946 
2947 	PMCDBG2(PMC,REG,1, "register-owner pmc-owner=%p pmc=%p",
2948 	    po, pmc);
2949 
2950 	return 0;
2951 }
2952 
2953 /*
2954  * Return the current row disposition:
2955  * == 0 => FREE
2956  *  > 0 => PROCESS MODE
2957  *  < 0 => SYSTEM MODE
2958  */
2959 
2960 int
2961 pmc_getrowdisp(int ri)
2962 {
2963 	return pmc_pmcdisp[ri];
2964 }
2965 
2966 /*
2967  * Check if a PMC at row index 'ri' can be allocated to the current
2968  * process.
2969  *
2970  * Allocation can fail if:
2971  *   - the current process is already being profiled by a PMC at index 'ri',
2972  *     attached to it via OP_PMCATTACH.
2973  *   - the current process has already allocated a PMC at index 'ri'
2974  *     via OP_ALLOCATE.
2975  */
2976 
2977 static int
2978 pmc_can_allocate_rowindex(struct proc *p, unsigned int ri, int cpu)
2979 {
2980 	enum pmc_mode mode;
2981 	struct pmc *pm;
2982 	struct pmc_owner *po;
2983 	struct pmc_process *pp;
2984 
2985 	PMCDBG5(PMC,ALR,1, "can-allocate-rowindex proc=%p (%d, %s) ri=%d "
2986 	    "cpu=%d", p, p->p_pid, p->p_comm, ri, cpu);
2987 
2988 	/*
2989 	 * We shouldn't have already allocated a process-mode PMC at
2990 	 * row index 'ri'.
2991 	 *
2992 	 * We shouldn't have allocated a system-wide PMC on the same
2993 	 * CPU and same RI.
2994 	 */
2995 	if ((po = pmc_find_owner_descriptor(p)) != NULL)
2996 		LIST_FOREACH(pm, &po->po_pmcs, pm_next) {
2997 		    if (PMC_TO_ROWINDEX(pm) == ri) {
2998 			    mode = PMC_TO_MODE(pm);
2999 			    if (PMC_IS_VIRTUAL_MODE(mode))
3000 				    return EEXIST;
3001 			    if (PMC_IS_SYSTEM_MODE(mode) &&
3002 				(int) PMC_TO_CPU(pm) == cpu)
3003 				    return EEXIST;
3004 		    }
3005 	        }
3006 
3007 	/*
3008 	 * We also shouldn't be the target of any PMC at this index
3009 	 * since otherwise a PMC_ATTACH to ourselves will fail.
3010 	 */
3011 	if ((pp = pmc_find_process_descriptor(p, 0)) != NULL)
3012 		if (pp->pp_pmcs[ri].pp_pmc)
3013 			return EEXIST;
3014 
3015 	PMCDBG4(PMC,ALR,2, "can-allocate-rowindex proc=%p (%d, %s) ri=%d ok",
3016 	    p, p->p_pid, p->p_comm, ri);
3017 
3018 	return 0;
3019 }
3020 
3021 /*
3022  * Check if a given PMC at row index 'ri' can be currently used in
3023  * mode 'mode'.
3024  */
3025 
3026 static int
3027 pmc_can_allocate_row(int ri, enum pmc_mode mode)
3028 {
3029 	enum pmc_disp	disp;
3030 
3031 	sx_assert(&pmc_sx, SX_XLOCKED);
3032 
3033 	PMCDBG2(PMC,ALR,1, "can-allocate-row ri=%d mode=%d", ri, mode);
3034 
3035 	if (PMC_IS_SYSTEM_MODE(mode))
3036 		disp = PMC_DISP_STANDALONE;
3037 	else
3038 		disp = PMC_DISP_THREAD;
3039 
3040 	/*
3041 	 * check disposition for PMC row 'ri':
3042 	 *
3043 	 * Expected disposition		Row-disposition		Result
3044 	 *
3045 	 * STANDALONE			STANDALONE or FREE	proceed
3046 	 * STANDALONE			THREAD			fail
3047 	 * THREAD			THREAD or FREE		proceed
3048 	 * THREAD			STANDALONE		fail
3049 	 */
3050 
3051 	if (!PMC_ROW_DISP_IS_FREE(ri) &&
3052 	    !(disp == PMC_DISP_THREAD && PMC_ROW_DISP_IS_THREAD(ri)) &&
3053 	    !(disp == PMC_DISP_STANDALONE && PMC_ROW_DISP_IS_STANDALONE(ri)))
3054 		return EBUSY;
3055 
3056 	/*
3057 	 * All OK
3058 	 */
3059 
3060 	PMCDBG2(PMC,ALR,2, "can-allocate-row ri=%d mode=%d ok", ri, mode);
3061 
3062 	return 0;
3063 
3064 }
3065 
3066 /*
3067  * Find a PMC descriptor with user handle 'pmcid' for thread 'td'.
3068  */
3069 
3070 static struct pmc *
3071 pmc_find_pmc_descriptor_in_process(struct pmc_owner *po, pmc_id_t pmcid)
3072 {
3073 	struct pmc *pm;
3074 
3075 	KASSERT(PMC_ID_TO_ROWINDEX(pmcid) < md->pmd_npmc,
3076 	    ("[pmc,%d] Illegal pmc index %d (max %d)", __LINE__,
3077 		PMC_ID_TO_ROWINDEX(pmcid), md->pmd_npmc));
3078 
3079 	LIST_FOREACH(pm, &po->po_pmcs, pm_next)
3080 	    if (pm->pm_id == pmcid)
3081 		    return pm;
3082 
3083 	return NULL;
3084 }
3085 
3086 static int
3087 pmc_find_pmc(pmc_id_t pmcid, struct pmc **pmc)
3088 {
3089 
3090 	struct pmc *pm, *opm;
3091 	struct pmc_owner *po;
3092 	struct pmc_process *pp;
3093 
3094 	PMCDBG1(PMC,FND,1, "find-pmc id=%d", pmcid);
3095 	if (PMC_ID_TO_ROWINDEX(pmcid) >= md->pmd_npmc)
3096 		return (EINVAL);
3097 
3098 	if ((po = pmc_find_owner_descriptor(curthread->td_proc)) == NULL) {
3099 		/*
3100 		 * In case of PMC_F_DESCENDANTS child processes we will not find
3101 		 * the current process in the owners hash list.  Find the owner
3102 		 * process first and from there lookup the po.
3103 		 */
3104 		if ((pp = pmc_find_process_descriptor(curthread->td_proc,
3105 		    PMC_FLAG_NONE)) == NULL) {
3106 			return ESRCH;
3107 		} else {
3108 			opm = pp->pp_pmcs[PMC_ID_TO_ROWINDEX(pmcid)].pp_pmc;
3109 			if (opm == NULL)
3110 				return ESRCH;
3111 			if ((opm->pm_flags & (PMC_F_ATTACHED_TO_OWNER|
3112 			    PMC_F_DESCENDANTS)) != (PMC_F_ATTACHED_TO_OWNER|
3113 			    PMC_F_DESCENDANTS))
3114 				return ESRCH;
3115 			po = opm->pm_owner;
3116 		}
3117 	}
3118 
3119 	if ((pm = pmc_find_pmc_descriptor_in_process(po, pmcid)) == NULL)
3120 		return EINVAL;
3121 
3122 	PMCDBG2(PMC,FND,2, "find-pmc id=%d -> pmc=%p", pmcid, pm);
3123 
3124 	*pmc = pm;
3125 	return 0;
3126 }
3127 
3128 /*
3129  * Start a PMC.
3130  */
3131 
3132 static int
3133 pmc_start(struct pmc *pm)
3134 {
3135 	enum pmc_mode mode;
3136 	struct pmc_owner *po;
3137 	struct pmc_binding pb;
3138 	struct pmc_classdep *pcd;
3139 	int adjri, error, cpu, ri;
3140 
3141 	KASSERT(pm != NULL,
3142 	    ("[pmc,%d] null pm", __LINE__));
3143 
3144 	mode = PMC_TO_MODE(pm);
3145 	ri   = PMC_TO_ROWINDEX(pm);
3146 	pcd  = pmc_ri_to_classdep(md, ri, &adjri);
3147 
3148 	error = 0;
3149 
3150 	PMCDBG3(PMC,OPS,1, "start pmc=%p mode=%d ri=%d", pm, mode, ri);
3151 
3152 	po = pm->pm_owner;
3153 
3154 	/*
3155 	 * Disallow PMCSTART if a logfile is required but has not been
3156 	 * configured yet.
3157 	 */
3158 	if ((pm->pm_flags & PMC_F_NEEDS_LOGFILE) &&
3159 	    (po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
3160 		return (EDOOFUS);	/* programming error */
3161 
3162 	/*
3163 	 * If this is a sampling mode PMC, log mapping information for
3164 	 * the kernel modules that are currently loaded.
3165 	 */
3166 	if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
3167 	    pmc_log_kernel_mappings(pm);
3168 
3169 	if (PMC_IS_VIRTUAL_MODE(mode)) {
3170 
3171 		/*
3172 		 * If a PMCATTACH has never been done on this PMC,
3173 		 * attach it to its owner process.
3174 		 */
3175 
3176 		if (LIST_EMPTY(&pm->pm_targets))
3177 			error = (pm->pm_flags & PMC_F_ATTACH_DONE) ? ESRCH :
3178 			    pmc_attach_process(po->po_owner, pm);
3179 
3180 		/*
3181 		 * If the PMC is attached to its owner, then force a context
3182 		 * switch to ensure that the MD state gets set correctly.
3183 		 */
3184 
3185 		if (error == 0) {
3186 			pm->pm_state = PMC_STATE_RUNNING;
3187 			if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER)
3188 				pmc_force_context_switch();
3189 		}
3190 
3191 		return (error);
3192 	}
3193 
3194 
3195 	/*
3196 	 * A system-wide PMC.
3197 	 *
3198 	 * Add the owner to the global list if this is a system-wide
3199 	 * sampling PMC.
3200 	 */
3201 
3202 	if (mode == PMC_MODE_SS) {
3203 		/*
3204 		 * Log mapping information for all existing processes in the
3205 		 * system.  Subsequent mappings are logged as they happen;
3206 		 * see pmc_process_mmap().
3207 		 */
3208 		if (po->po_logprocmaps == 0) {
3209 			pmc_log_all_process_mappings(po);
3210 			po->po_logprocmaps = 1;
3211 		}
3212 		po->po_sscount++;
3213 		if (po->po_sscount == 1) {
3214 			atomic_add_rel_int(&pmc_ss_count, 1);
3215 			CK_LIST_INSERT_HEAD(&pmc_ss_owners, po, po_ssnext);
3216 			PMCDBG1(PMC,OPS,1, "po=%p in global list", po);
3217 		}
3218 	}
3219 
3220 	/*
3221 	 * Move to the CPU associated with this
3222 	 * PMC, and start the hardware.
3223 	 */
3224 
3225 	pmc_save_cpu_binding(&pb);
3226 
3227 	cpu = PMC_TO_CPU(pm);
3228 
3229 	if (!pmc_cpu_is_active(cpu))
3230 		return (ENXIO);
3231 
3232 	pmc_select_cpu(cpu);
3233 
3234 	/*
3235 	 * global PMCs are configured at allocation time
3236 	 * so write out the initial value and start the PMC.
3237 	 */
3238 
3239 	pm->pm_state = PMC_STATE_RUNNING;
3240 
3241 	critical_enter();
3242 	if ((error = pcd->pcd_write_pmc(cpu, adjri,
3243 		 PMC_IS_SAMPLING_MODE(mode) ?
3244 		 pm->pm_sc.pm_reloadcount :
3245 		 pm->pm_sc.pm_initial)) == 0) {
3246 		/* If a sampling mode PMC, reset stalled state. */
3247 		if (PMC_IS_SAMPLING_MODE(mode))
3248 			pm->pm_pcpu_state[cpu].pps_stalled = 0;
3249 
3250 		/* Indicate that we desire this to run. Start it. */
3251 		pm->pm_pcpu_state[cpu].pps_cpustate = 1;
3252 		error = pcd->pcd_start_pmc(cpu, adjri);
3253 	}
3254 	critical_exit();
3255 
3256 	pmc_restore_cpu_binding(&pb);
3257 
3258 	return (error);
3259 }
3260 
3261 /*
3262  * Stop a PMC.
3263  */
3264 
3265 static int
3266 pmc_stop(struct pmc *pm)
3267 {
3268 	struct pmc_owner *po;
3269 	struct pmc_binding pb;
3270 	struct pmc_classdep *pcd;
3271 	int adjri, cpu, error, ri;
3272 
3273 	KASSERT(pm != NULL, ("[pmc,%d] null pmc", __LINE__));
3274 
3275 	PMCDBG3(PMC,OPS,1, "stop pmc=%p mode=%d ri=%d", pm,
3276 	    PMC_TO_MODE(pm), PMC_TO_ROWINDEX(pm));
3277 
3278 	pm->pm_state = PMC_STATE_STOPPED;
3279 
3280 	/*
3281 	 * If the PMC is a virtual mode one, changing the state to
3282 	 * non-RUNNING is enough to ensure that the PMC never gets
3283 	 * scheduled.
3284 	 *
3285 	 * If this PMC is current running on a CPU, then it will
3286 	 * handled correctly at the time its target process is context
3287 	 * switched out.
3288 	 */
3289 
3290 	if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)))
3291 		return 0;
3292 
3293 	/*
3294 	 * A system-mode PMC.  Move to the CPU associated with
3295 	 * this PMC, and stop the hardware.  We update the
3296 	 * 'initial count' so that a subsequent PMCSTART will
3297 	 * resume counting from the current hardware count.
3298 	 */
3299 
3300 	pmc_save_cpu_binding(&pb);
3301 
3302 	cpu = PMC_TO_CPU(pm);
3303 
3304 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
3305 	    ("[pmc,%d] illegal cpu=%d", __LINE__, cpu));
3306 
3307 	if (!pmc_cpu_is_active(cpu))
3308 		return ENXIO;
3309 
3310 	pmc_select_cpu(cpu);
3311 
3312 	ri = PMC_TO_ROWINDEX(pm);
3313 	pcd = pmc_ri_to_classdep(md, ri, &adjri);
3314 
3315 	pm->pm_pcpu_state[cpu].pps_cpustate = 0;
3316 	critical_enter();
3317 	if ((error = pcd->pcd_stop_pmc(cpu, adjri)) == 0)
3318 		error = pcd->pcd_read_pmc(cpu, adjri, &pm->pm_sc.pm_initial);
3319 	critical_exit();
3320 
3321 	pmc_restore_cpu_binding(&pb);
3322 
3323 	po = pm->pm_owner;
3324 
3325 	/* remove this owner from the global list of SS PMC owners */
3326 	if (PMC_TO_MODE(pm) == PMC_MODE_SS) {
3327 		po->po_sscount--;
3328 		if (po->po_sscount == 0) {
3329 			atomic_subtract_rel_int(&pmc_ss_count, 1);
3330 			CK_LIST_REMOVE(po, po_ssnext);
3331 			epoch_wait_preempt(global_epoch_preempt);
3332 			PMCDBG1(PMC,OPS,2,"po=%p removed from global list", po);
3333 		}
3334 	}
3335 
3336 	return (error);
3337 }
3338 
3339 static struct pmc_classdep *
3340 pmc_class_to_classdep(enum pmc_class class)
3341 {
3342 	int n;
3343 
3344 	for (n = 0; n < md->pmd_nclass; n++)
3345 		if (md->pmd_classdep[n].pcd_class == class)
3346 			return (&md->pmd_classdep[n]);
3347 	return (NULL);
3348 }
3349 
3350 #if defined(HWPMC_DEBUG) && defined(KTR)
3351 static const char *pmc_op_to_name[] = {
3352 #undef	__PMC_OP
3353 #define	__PMC_OP(N, D)	#N ,
3354 	__PMC_OPS()
3355 	NULL
3356 };
3357 #endif
3358 
3359 /*
3360  * The syscall interface
3361  */
3362 
3363 #define	PMC_GET_SX_XLOCK(...) do {		\
3364 	sx_xlock(&pmc_sx);			\
3365 	if (pmc_hook == NULL) {			\
3366 		sx_xunlock(&pmc_sx);		\
3367 		return __VA_ARGS__;		\
3368 	}					\
3369 } while (0)
3370 
3371 #define	PMC_DOWNGRADE_SX() do {			\
3372 	sx_downgrade(&pmc_sx);			\
3373 	is_sx_downgraded = 1;			\
3374 } while (0)
3375 
3376 static int
3377 pmc_syscall_handler(struct thread *td, void *syscall_args)
3378 {
3379 	int error, is_sx_downgraded, op;
3380 	struct pmc_syscall_args *c;
3381 	void *pmclog_proc_handle;
3382 	void *arg;
3383 
3384 	c = (struct pmc_syscall_args *)syscall_args;
3385 	op = c->pmop_code;
3386 	arg = c->pmop_data;
3387 	/* PMC isn't set up yet */
3388 	if (pmc_hook == NULL)
3389 		return (EINVAL);
3390 	if (op == PMC_OP_CONFIGURELOG) {
3391 		/*
3392 		 * We cannot create the logging process inside
3393 		 * pmclog_configure_log() because there is a LOR
3394 		 * between pmc_sx and process structure locks.
3395 		 * Instead, pre-create the process and ignite the loop
3396 		 * if everything is fine, otherwise direct the process
3397 		 * to exit.
3398 		 */
3399 		error = pmclog_proc_create(td, &pmclog_proc_handle);
3400 		if (error != 0)
3401 			goto done_syscall;
3402 	}
3403 
3404 	PMC_GET_SX_XLOCK(ENOSYS);
3405 	is_sx_downgraded = 0;
3406 	PMCDBG3(MOD,PMS,1, "syscall op=%d \"%s\" arg=%p", op,
3407 	    pmc_op_to_name[op], arg);
3408 
3409 	error = 0;
3410 	counter_u64_add(pmc_stats.pm_syscalls, 1);
3411 
3412 	switch (op) {
3413 
3414 
3415 	/*
3416 	 * Configure a log file.
3417 	 *
3418 	 * XXX This OP will be reworked.
3419 	 */
3420 
3421 	case PMC_OP_CONFIGURELOG:
3422 	{
3423 		struct proc *p;
3424 		struct pmc *pm;
3425 		struct pmc_owner *po;
3426 		struct pmc_op_configurelog cl;
3427 
3428 		if ((error = copyin(arg, &cl, sizeof(cl))) != 0) {
3429 			pmclog_proc_ignite(pmclog_proc_handle, NULL);
3430 			break;
3431 		}
3432 
3433 		/* mark this process as owning a log file */
3434 		p = td->td_proc;
3435 		if ((po = pmc_find_owner_descriptor(p)) == NULL)
3436 			if ((po = pmc_allocate_owner_descriptor(p)) == NULL) {
3437 				pmclog_proc_ignite(pmclog_proc_handle, NULL);
3438 				error = ENOMEM;
3439 				break;
3440 			}
3441 
3442 		/*
3443 		 * If a valid fd was passed in, try to configure that,
3444 		 * otherwise if 'fd' was less than zero and there was
3445 		 * a log file configured, flush its buffers and
3446 		 * de-configure it.
3447 		 */
3448 		if (cl.pm_logfd >= 0) {
3449 			error = pmclog_configure_log(md, po, cl.pm_logfd);
3450 			pmclog_proc_ignite(pmclog_proc_handle, error == 0 ?
3451 			    po : NULL);
3452 		} else if (po->po_flags & PMC_PO_OWNS_LOGFILE) {
3453 			pmclog_proc_ignite(pmclog_proc_handle, NULL);
3454 			error = pmclog_close(po);
3455 			if (error == 0) {
3456 				LIST_FOREACH(pm, &po->po_pmcs, pm_next)
3457 				    if (pm->pm_flags & PMC_F_NEEDS_LOGFILE &&
3458 					pm->pm_state == PMC_STATE_RUNNING)
3459 					    pmc_stop(pm);
3460 				error = pmclog_deconfigure_log(po);
3461 			}
3462 		} else {
3463 			pmclog_proc_ignite(pmclog_proc_handle, NULL);
3464 			error = EINVAL;
3465 		}
3466 	}
3467 	break;
3468 
3469 	/*
3470 	 * Flush a log file.
3471 	 */
3472 
3473 	case PMC_OP_FLUSHLOG:
3474 	{
3475 		struct pmc_owner *po;
3476 
3477 		sx_assert(&pmc_sx, SX_XLOCKED);
3478 
3479 		if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) {
3480 			error = EINVAL;
3481 			break;
3482 		}
3483 
3484 		error = pmclog_flush(po, 0);
3485 	}
3486 	break;
3487 
3488 	/*
3489 	 * Close a log file.
3490 	 */
3491 
3492 	case PMC_OP_CLOSELOG:
3493 	{
3494 		struct pmc_owner *po;
3495 
3496 		sx_assert(&pmc_sx, SX_XLOCKED);
3497 
3498 		if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) {
3499 			error = EINVAL;
3500 			break;
3501 		}
3502 
3503 		error = pmclog_close(po);
3504 	}
3505 	break;
3506 
3507 	/*
3508 	 * Retrieve hardware configuration.
3509 	 */
3510 
3511 	case PMC_OP_GETCPUINFO:	/* CPU information */
3512 	{
3513 		struct pmc_op_getcpuinfo gci;
3514 		struct pmc_classinfo *pci;
3515 		struct pmc_classdep *pcd;
3516 		int cl;
3517 
3518 		memset(&gci, 0, sizeof(gci));
3519 		gci.pm_cputype = md->pmd_cputype;
3520 		gci.pm_ncpu    = pmc_cpu_max();
3521 		gci.pm_npmc    = md->pmd_npmc;
3522 		gci.pm_nclass  = md->pmd_nclass;
3523 		pci = gci.pm_classes;
3524 		pcd = md->pmd_classdep;
3525 		for (cl = 0; cl < md->pmd_nclass; cl++, pci++, pcd++) {
3526 			pci->pm_caps  = pcd->pcd_caps;
3527 			pci->pm_class = pcd->pcd_class;
3528 			pci->pm_width = pcd->pcd_width;
3529 			pci->pm_num   = pcd->pcd_num;
3530 		}
3531 		error = copyout(&gci, arg, sizeof(gci));
3532 	}
3533 	break;
3534 
3535 	/*
3536 	 * Retrieve soft events list.
3537 	 */
3538 	case PMC_OP_GETDYNEVENTINFO:
3539 	{
3540 		enum pmc_class			cl;
3541 		enum pmc_event			ev;
3542 		struct pmc_op_getdyneventinfo	*gei;
3543 		struct pmc_dyn_event_descr	dev;
3544 		struct pmc_soft			*ps;
3545 		uint32_t			nevent;
3546 
3547 		sx_assert(&pmc_sx, SX_LOCKED);
3548 
3549 		gei = (struct pmc_op_getdyneventinfo *) arg;
3550 
3551 		if ((error = copyin(&gei->pm_class, &cl, sizeof(cl))) != 0)
3552 			break;
3553 
3554 		/* Only SOFT class is dynamic. */
3555 		if (cl != PMC_CLASS_SOFT) {
3556 			error = EINVAL;
3557 			break;
3558 		}
3559 
3560 		nevent = 0;
3561 		for (ev = PMC_EV_SOFT_FIRST; (int)ev <= PMC_EV_SOFT_LAST; ev++) {
3562 			ps = pmc_soft_ev_acquire(ev);
3563 			if (ps == NULL)
3564 				continue;
3565 			bcopy(&ps->ps_ev, &dev, sizeof(dev));
3566 			pmc_soft_ev_release(ps);
3567 
3568 			error = copyout(&dev,
3569 			    &gei->pm_events[nevent],
3570 			    sizeof(struct pmc_dyn_event_descr));
3571 			if (error != 0)
3572 				break;
3573 			nevent++;
3574 		}
3575 		if (error != 0)
3576 			break;
3577 
3578 		error = copyout(&nevent, &gei->pm_nevent,
3579 		    sizeof(nevent));
3580 	}
3581 	break;
3582 
3583 	/*
3584 	 * Get module statistics
3585 	 */
3586 
3587 	case PMC_OP_GETDRIVERSTATS:
3588 	{
3589 		struct pmc_op_getdriverstats gms;
3590 #define CFETCH(a, b, field) a.field = counter_u64_fetch(b.field)
3591 		CFETCH(gms, pmc_stats, pm_intr_ignored);
3592 		CFETCH(gms, pmc_stats, pm_intr_processed);
3593 		CFETCH(gms, pmc_stats, pm_intr_bufferfull);
3594 		CFETCH(gms, pmc_stats, pm_syscalls);
3595 		CFETCH(gms, pmc_stats, pm_syscall_errors);
3596 		CFETCH(gms, pmc_stats, pm_buffer_requests);
3597 		CFETCH(gms, pmc_stats, pm_buffer_requests_failed);
3598 		CFETCH(gms, pmc_stats, pm_log_sweeps);
3599 #undef CFETCH
3600 		error = copyout(&gms, arg, sizeof(gms));
3601 	}
3602 	break;
3603 
3604 
3605 	/*
3606 	 * Retrieve module version number
3607 	 */
3608 
3609 	case PMC_OP_GETMODULEVERSION:
3610 	{
3611 		uint32_t cv, modv;
3612 
3613 		/* retrieve the client's idea of the ABI version */
3614 		if ((error = copyin(arg, &cv, sizeof(uint32_t))) != 0)
3615 			break;
3616 		/* don't service clients newer than our driver */
3617 		modv = PMC_VERSION;
3618 		if ((cv & 0xFFFF0000) > (modv & 0xFFFF0000)) {
3619 			error = EPROGMISMATCH;
3620 			break;
3621 		}
3622 		error = copyout(&modv, arg, sizeof(int));
3623 	}
3624 	break;
3625 
3626 
3627 	/*
3628 	 * Retrieve the state of all the PMCs on a given
3629 	 * CPU.
3630 	 */
3631 
3632 	case PMC_OP_GETPMCINFO:
3633 	{
3634 		int ari;
3635 		struct pmc *pm;
3636 		size_t pmcinfo_size;
3637 		uint32_t cpu, n, npmc;
3638 		struct pmc_owner *po;
3639 		struct pmc_binding pb;
3640 		struct pmc_classdep *pcd;
3641 		struct pmc_info *p, *pmcinfo;
3642 		struct pmc_op_getpmcinfo *gpi;
3643 
3644 		PMC_DOWNGRADE_SX();
3645 
3646 		gpi = (struct pmc_op_getpmcinfo *) arg;
3647 
3648 		if ((error = copyin(&gpi->pm_cpu, &cpu, sizeof(cpu))) != 0)
3649 			break;
3650 
3651 		if (cpu >= pmc_cpu_max()) {
3652 			error = EINVAL;
3653 			break;
3654 		}
3655 
3656 		if (!pmc_cpu_is_active(cpu)) {
3657 			error = ENXIO;
3658 			break;
3659 		}
3660 
3661 		/* switch to CPU 'cpu' */
3662 		pmc_save_cpu_binding(&pb);
3663 		pmc_select_cpu(cpu);
3664 
3665 		npmc = md->pmd_npmc;
3666 
3667 		pmcinfo_size = npmc * sizeof(struct pmc_info);
3668 		pmcinfo = malloc(pmcinfo_size, M_PMC, M_WAITOK | M_ZERO);
3669 
3670 		p = pmcinfo;
3671 
3672 		for (n = 0; n < md->pmd_npmc; n++, p++) {
3673 
3674 			pcd = pmc_ri_to_classdep(md, n, &ari);
3675 
3676 			KASSERT(pcd != NULL,
3677 			    ("[pmc,%d] null pcd ri=%d", __LINE__, n));
3678 
3679 			if ((error = pcd->pcd_describe(cpu, ari, p, &pm)) != 0)
3680 				break;
3681 
3682 			if (PMC_ROW_DISP_IS_STANDALONE(n))
3683 				p->pm_rowdisp = PMC_DISP_STANDALONE;
3684 			else if (PMC_ROW_DISP_IS_THREAD(n))
3685 				p->pm_rowdisp = PMC_DISP_THREAD;
3686 			else
3687 				p->pm_rowdisp = PMC_DISP_FREE;
3688 
3689 			p->pm_ownerpid = -1;
3690 
3691 			if (pm == NULL)	/* no PMC associated */
3692 				continue;
3693 
3694 			po = pm->pm_owner;
3695 
3696 			KASSERT(po->po_owner != NULL,
3697 			    ("[pmc,%d] pmc_owner had a null proc pointer",
3698 				__LINE__));
3699 
3700 			p->pm_ownerpid = po->po_owner->p_pid;
3701 			p->pm_mode     = PMC_TO_MODE(pm);
3702 			p->pm_event    = pm->pm_event;
3703 			p->pm_flags    = pm->pm_flags;
3704 
3705 			if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
3706 				p->pm_reloadcount =
3707 				    pm->pm_sc.pm_reloadcount;
3708 		}
3709 
3710 		pmc_restore_cpu_binding(&pb);
3711 
3712 		/* now copy out the PMC info collected */
3713 		if (error == 0)
3714 			error = copyout(pmcinfo, &gpi->pm_pmcs, pmcinfo_size);
3715 
3716 		free(pmcinfo, M_PMC);
3717 	}
3718 	break;
3719 
3720 
3721 	/*
3722 	 * Set the administrative state of a PMC.  I.e. whether
3723 	 * the PMC is to be used or not.
3724 	 */
3725 
3726 	case PMC_OP_PMCADMIN:
3727 	{
3728 		int cpu, ri;
3729 		enum pmc_state request;
3730 		struct pmc_cpu *pc;
3731 		struct pmc_hw *phw;
3732 		struct pmc_op_pmcadmin pma;
3733 		struct pmc_binding pb;
3734 
3735 		sx_assert(&pmc_sx, SX_XLOCKED);
3736 
3737 		KASSERT(td == curthread,
3738 		    ("[pmc,%d] td != curthread", __LINE__));
3739 
3740 		error = priv_check(td, PRIV_PMC_MANAGE);
3741 		if (error)
3742 			break;
3743 
3744 		if ((error = copyin(arg, &pma, sizeof(pma))) != 0)
3745 			break;
3746 
3747 		cpu = pma.pm_cpu;
3748 
3749 		if (cpu < 0 || cpu >= (int) pmc_cpu_max()) {
3750 			error = EINVAL;
3751 			break;
3752 		}
3753 
3754 		if (!pmc_cpu_is_active(cpu)) {
3755 			error = ENXIO;
3756 			break;
3757 		}
3758 
3759 		request = pma.pm_state;
3760 
3761 		if (request != PMC_STATE_DISABLED &&
3762 		    request != PMC_STATE_FREE) {
3763 			error = EINVAL;
3764 			break;
3765 		}
3766 
3767 		ri = pma.pm_pmc; /* pmc id == row index */
3768 		if (ri < 0 || ri >= (int) md->pmd_npmc) {
3769 			error = EINVAL;
3770 			break;
3771 		}
3772 
3773 		/*
3774 		 * We can't disable a PMC with a row-index allocated
3775 		 * for process virtual PMCs.
3776 		 */
3777 
3778 		if (PMC_ROW_DISP_IS_THREAD(ri) &&
3779 		    request == PMC_STATE_DISABLED) {
3780 			error = EBUSY;
3781 			break;
3782 		}
3783 
3784 		/*
3785 		 * otherwise, this PMC on this CPU is either free or
3786 		 * in system-wide mode.
3787 		 */
3788 
3789 		pmc_save_cpu_binding(&pb);
3790 		pmc_select_cpu(cpu);
3791 
3792 		pc  = pmc_pcpu[cpu];
3793 		phw = pc->pc_hwpmcs[ri];
3794 
3795 		/*
3796 		 * XXX do we need some kind of 'forced' disable?
3797 		 */
3798 
3799 		if (phw->phw_pmc == NULL) {
3800 			if (request == PMC_STATE_DISABLED &&
3801 			    (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED)) {
3802 				phw->phw_state &= ~PMC_PHW_FLAG_IS_ENABLED;
3803 				PMC_MARK_ROW_STANDALONE(ri);
3804 			} else if (request == PMC_STATE_FREE &&
3805 			    (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0) {
3806 				phw->phw_state |=  PMC_PHW_FLAG_IS_ENABLED;
3807 				PMC_UNMARK_ROW_STANDALONE(ri);
3808 			}
3809 			/* other cases are a no-op */
3810 		} else
3811 			error = EBUSY;
3812 
3813 		pmc_restore_cpu_binding(&pb);
3814 	}
3815 	break;
3816 
3817 
3818 	/*
3819 	 * Allocate a PMC.
3820 	 */
3821 
3822 	case PMC_OP_PMCALLOCATE:
3823 	{
3824 		int adjri, n;
3825 		u_int cpu;
3826 		uint32_t caps;
3827 		struct pmc *pmc;
3828 		enum pmc_mode mode;
3829 		struct pmc_hw *phw;
3830 		struct pmc_binding pb;
3831 		struct pmc_classdep *pcd;
3832 		struct pmc_op_pmcallocate pa;
3833 
3834 		if ((error = copyin(arg, &pa, sizeof(pa))) != 0)
3835 			break;
3836 
3837 		caps = pa.pm_caps;
3838 		mode = pa.pm_mode;
3839 		cpu  = pa.pm_cpu;
3840 
3841 		if ((mode != PMC_MODE_SS  &&  mode != PMC_MODE_SC  &&
3842 		     mode != PMC_MODE_TS  &&  mode != PMC_MODE_TC) ||
3843 		    (cpu != (u_int) PMC_CPU_ANY && cpu >= pmc_cpu_max())) {
3844 			error = EINVAL;
3845 			break;
3846 		}
3847 
3848 		/*
3849 		 * Virtual PMCs should only ask for a default CPU.
3850 		 * System mode PMCs need to specify a non-default CPU.
3851 		 */
3852 
3853 		if ((PMC_IS_VIRTUAL_MODE(mode) && cpu != (u_int) PMC_CPU_ANY) ||
3854 		    (PMC_IS_SYSTEM_MODE(mode) && cpu == (u_int) PMC_CPU_ANY)) {
3855 			error = EINVAL;
3856 			break;
3857 		}
3858 
3859 		/*
3860 		 * Check that an inactive CPU is not being asked for.
3861 		 */
3862 
3863 		if (PMC_IS_SYSTEM_MODE(mode) && !pmc_cpu_is_active(cpu)) {
3864 			error = ENXIO;
3865 			break;
3866 		}
3867 
3868 		/*
3869 		 * Refuse an allocation for a system-wide PMC if this
3870 		 * process has been jailed, or if this process lacks
3871 		 * super-user credentials and the sysctl tunable
3872 		 * 'security.bsd.unprivileged_syspmcs' is zero.
3873 		 */
3874 
3875 		if (PMC_IS_SYSTEM_MODE(mode)) {
3876 			if (jailed(curthread->td_ucred)) {
3877 				error = EPERM;
3878 				break;
3879 			}
3880 			if (!pmc_unprivileged_syspmcs) {
3881 				error = priv_check(curthread,
3882 				    PRIV_PMC_SYSTEM);
3883 				if (error)
3884 					break;
3885 			}
3886 		}
3887 
3888 		/*
3889 		 * Look for valid values for 'pm_flags'
3890 		 */
3891 
3892 		if ((pa.pm_flags & ~(PMC_F_DESCENDANTS | PMC_F_LOG_PROCCSW |
3893 		    PMC_F_LOG_PROCEXIT | PMC_F_CALLCHAIN |
3894 		    PMC_F_USERCALLCHAIN)) != 0) {
3895 			error = EINVAL;
3896 			break;
3897 		}
3898 
3899 		/* PMC_F_USERCALLCHAIN is only valid with PMC_F_CALLCHAIN */
3900 		if ((pa.pm_flags & (PMC_F_CALLCHAIN | PMC_F_USERCALLCHAIN)) ==
3901 		    PMC_F_USERCALLCHAIN) {
3902 			error = EINVAL;
3903 			break;
3904 		}
3905 
3906 		/* PMC_F_USERCALLCHAIN is only valid for sampling mode */
3907 		if (pa.pm_flags & PMC_F_USERCALLCHAIN &&
3908 			mode != PMC_MODE_TS && mode != PMC_MODE_SS) {
3909 			error = EINVAL;
3910 			break;
3911 		}
3912 
3913 		/* process logging options are not allowed for system PMCs */
3914 		if (PMC_IS_SYSTEM_MODE(mode) && (pa.pm_flags &
3915 		    (PMC_F_LOG_PROCCSW | PMC_F_LOG_PROCEXIT))) {
3916 			error = EINVAL;
3917 			break;
3918 		}
3919 
3920 		/*
3921 		 * All sampling mode PMCs need to be able to interrupt the
3922 		 * CPU.
3923 		 */
3924 		if (PMC_IS_SAMPLING_MODE(mode))
3925 			caps |= PMC_CAP_INTERRUPT;
3926 
3927 		/* A valid class specifier should have been passed in. */
3928 		pcd = pmc_class_to_classdep(pa.pm_class);
3929 		if (pcd == NULL) {
3930 			error = EINVAL;
3931 			break;
3932 		}
3933 
3934 		/* The requested PMC capabilities should be feasible. */
3935 		if ((pcd->pcd_caps & caps) != caps) {
3936 			error = EOPNOTSUPP;
3937 			break;
3938 		}
3939 
3940 		PMCDBG4(PMC,ALL,2, "event=%d caps=0x%x mode=%d cpu=%d",
3941 		    pa.pm_ev, caps, mode, cpu);
3942 
3943 		pmc = pmc_allocate_pmc_descriptor();
3944 		pmc->pm_id    = PMC_ID_MAKE_ID(cpu,pa.pm_mode,pa.pm_class,
3945 		    PMC_ID_INVALID);
3946 		pmc->pm_event = pa.pm_ev;
3947 		pmc->pm_state = PMC_STATE_FREE;
3948 		pmc->pm_caps  = caps;
3949 		pmc->pm_flags = pa.pm_flags;
3950 
3951 		/* XXX set lower bound on sampling for process counters */
3952 		if (PMC_IS_SAMPLING_MODE(mode)) {
3953 			/*
3954 			 * Don't permit requested sample rate to be less than 1000
3955 			 */
3956 			if (pa.pm_count < 1000)
3957 				log(LOG_WARNING,
3958 					"pmcallocate: passed sample rate %ju - setting to 1000\n",
3959 					(uintmax_t)pa.pm_count);
3960 			pmc->pm_sc.pm_reloadcount = MAX(1000, pa.pm_count);
3961 		} else
3962 			pmc->pm_sc.pm_initial = pa.pm_count;
3963 
3964 		/* switch thread to CPU 'cpu' */
3965 		pmc_save_cpu_binding(&pb);
3966 
3967 #define	PMC_IS_SHAREABLE_PMC(cpu, n)				\
3968 	(pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_state &		\
3969 	 PMC_PHW_FLAG_IS_SHAREABLE)
3970 #define	PMC_IS_UNALLOCATED(cpu, n)				\
3971 	(pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_pmc == NULL)
3972 
3973 		if (PMC_IS_SYSTEM_MODE(mode)) {
3974 			pmc_select_cpu(cpu);
3975 			for (n = pcd->pcd_ri; n < (int) md->pmd_npmc; n++) {
3976 				pcd = pmc_ri_to_classdep(md, n, &adjri);
3977 				if (pmc_can_allocate_row(n, mode) == 0 &&
3978 				    pmc_can_allocate_rowindex(
3979 					    curthread->td_proc, n, cpu) == 0 &&
3980 				    (PMC_IS_UNALLOCATED(cpu, n) ||
3981 				     PMC_IS_SHAREABLE_PMC(cpu, n)) &&
3982 				    pcd->pcd_allocate_pmc(cpu, adjri, pmc,
3983 					&pa) == 0)
3984 					break;
3985 			}
3986 		} else {
3987 			/* Process virtual mode */
3988 			for (n = pcd->pcd_ri; n < (int) md->pmd_npmc; n++) {
3989 				pcd = pmc_ri_to_classdep(md, n, &adjri);
3990 				if (pmc_can_allocate_row(n, mode) == 0 &&
3991 				    pmc_can_allocate_rowindex(
3992 					    curthread->td_proc, n,
3993 					    PMC_CPU_ANY) == 0 &&
3994 				    pcd->pcd_allocate_pmc(curthread->td_oncpu,
3995 					adjri, pmc, &pa) == 0)
3996 					break;
3997 			}
3998 		}
3999 
4000 #undef	PMC_IS_UNALLOCATED
4001 #undef	PMC_IS_SHAREABLE_PMC
4002 
4003 		pmc_restore_cpu_binding(&pb);
4004 
4005 		if (n == (int) md->pmd_npmc) {
4006 			pmc_destroy_pmc_descriptor(pmc);
4007 			pmc = NULL;
4008 			error = EINVAL;
4009 			break;
4010 		}
4011 
4012 		/* Fill in the correct value in the ID field */
4013 		pmc->pm_id = PMC_ID_MAKE_ID(cpu,mode,pa.pm_class,n);
4014 
4015 		PMCDBG5(PMC,ALL,2, "ev=%d class=%d mode=%d n=%d -> pmcid=%x",
4016 		    pmc->pm_event, pa.pm_class, mode, n, pmc->pm_id);
4017 
4018 		/* Process mode PMCs with logging enabled need log files */
4019 		if (pmc->pm_flags & (PMC_F_LOG_PROCEXIT | PMC_F_LOG_PROCCSW))
4020 			pmc->pm_flags |= PMC_F_NEEDS_LOGFILE;
4021 
4022 		/* All system mode sampling PMCs require a log file */
4023 		if (PMC_IS_SAMPLING_MODE(mode) && PMC_IS_SYSTEM_MODE(mode))
4024 			pmc->pm_flags |= PMC_F_NEEDS_LOGFILE;
4025 
4026 		/*
4027 		 * Configure global pmc's immediately
4028 		 */
4029 
4030 		if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pmc))) {
4031 
4032 			pmc_save_cpu_binding(&pb);
4033 			pmc_select_cpu(cpu);
4034 
4035 			phw = pmc_pcpu[cpu]->pc_hwpmcs[n];
4036 			pcd = pmc_ri_to_classdep(md, n, &adjri);
4037 
4038 			if ((phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0 ||
4039 			    (error = pcd->pcd_config_pmc(cpu, adjri, pmc)) != 0) {
4040 				(void) pcd->pcd_release_pmc(cpu, adjri, pmc);
4041 				pmc_destroy_pmc_descriptor(pmc);
4042 				pmc = NULL;
4043 				pmc_restore_cpu_binding(&pb);
4044 				error = EPERM;
4045 				break;
4046 			}
4047 
4048 			pmc_restore_cpu_binding(&pb);
4049 		}
4050 
4051 		pmc->pm_state    = PMC_STATE_ALLOCATED;
4052 		pmc->pm_class	= pa.pm_class;
4053 
4054 		/*
4055 		 * mark row disposition
4056 		 */
4057 
4058 		if (PMC_IS_SYSTEM_MODE(mode))
4059 			PMC_MARK_ROW_STANDALONE(n);
4060 		else
4061 			PMC_MARK_ROW_THREAD(n);
4062 
4063 		/*
4064 		 * Register this PMC with the current thread as its owner.
4065 		 */
4066 
4067 		if ((error =
4068 		    pmc_register_owner(curthread->td_proc, pmc)) != 0) {
4069 			pmc_release_pmc_descriptor(pmc);
4070 			pmc_destroy_pmc_descriptor(pmc);
4071 			pmc = NULL;
4072 			break;
4073 		}
4074 
4075 
4076 		/*
4077 		 * Return the allocated index.
4078 		 */
4079 
4080 		pa.pm_pmcid = pmc->pm_id;
4081 
4082 		error = copyout(&pa, arg, sizeof(pa));
4083 	}
4084 	break;
4085 
4086 
4087 	/*
4088 	 * Attach a PMC to a process.
4089 	 */
4090 
4091 	case PMC_OP_PMCATTACH:
4092 	{
4093 		struct pmc *pm;
4094 		struct proc *p;
4095 		struct pmc_op_pmcattach a;
4096 
4097 		sx_assert(&pmc_sx, SX_XLOCKED);
4098 
4099 		if ((error = copyin(arg, &a, sizeof(a))) != 0)
4100 			break;
4101 
4102 		if (a.pm_pid < 0) {
4103 			error = EINVAL;
4104 			break;
4105 		} else if (a.pm_pid == 0)
4106 			a.pm_pid = td->td_proc->p_pid;
4107 
4108 		if ((error = pmc_find_pmc(a.pm_pmc, &pm)) != 0)
4109 			break;
4110 
4111 		if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm))) {
4112 			error = EINVAL;
4113 			break;
4114 		}
4115 
4116 		/* PMCs may be (re)attached only when allocated or stopped */
4117 		if (pm->pm_state == PMC_STATE_RUNNING) {
4118 			error = EBUSY;
4119 			break;
4120 		} else if (pm->pm_state != PMC_STATE_ALLOCATED &&
4121 		    pm->pm_state != PMC_STATE_STOPPED) {
4122 			error = EINVAL;
4123 			break;
4124 		}
4125 
4126 		/* lookup pid */
4127 		if ((p = pfind(a.pm_pid)) == NULL) {
4128 			error = ESRCH;
4129 			break;
4130 		}
4131 
4132 		/*
4133 		 * Ignore processes that are working on exiting.
4134 		 */
4135 		if (p->p_flag & P_WEXIT) {
4136 			error = ESRCH;
4137 			PROC_UNLOCK(p);	/* pfind() returns a locked process */
4138 			break;
4139 		}
4140 
4141 		/*
4142 		 * we are allowed to attach a PMC to a process if
4143 		 * we can debug it.
4144 		 */
4145 		error = p_candebug(curthread, p);
4146 
4147 		PROC_UNLOCK(p);
4148 
4149 		if (error == 0)
4150 			error = pmc_attach_process(p, pm);
4151 	}
4152 	break;
4153 
4154 
4155 	/*
4156 	 * Detach an attached PMC from a process.
4157 	 */
4158 
4159 	case PMC_OP_PMCDETACH:
4160 	{
4161 		struct pmc *pm;
4162 		struct proc *p;
4163 		struct pmc_op_pmcattach a;
4164 
4165 		if ((error = copyin(arg, &a, sizeof(a))) != 0)
4166 			break;
4167 
4168 		if (a.pm_pid < 0) {
4169 			error = EINVAL;
4170 			break;
4171 		} else if (a.pm_pid == 0)
4172 			a.pm_pid = td->td_proc->p_pid;
4173 
4174 		if ((error = pmc_find_pmc(a.pm_pmc, &pm)) != 0)
4175 			break;
4176 
4177 		if ((p = pfind(a.pm_pid)) == NULL) {
4178 			error = ESRCH;
4179 			break;
4180 		}
4181 
4182 		/*
4183 		 * Treat processes that are in the process of exiting
4184 		 * as if they were not present.
4185 		 */
4186 
4187 		if (p->p_flag & P_WEXIT)
4188 			error = ESRCH;
4189 
4190 		PROC_UNLOCK(p);	/* pfind() returns a locked process */
4191 
4192 		if (error == 0)
4193 			error = pmc_detach_process(p, pm);
4194 	}
4195 	break;
4196 
4197 
4198 	/*
4199 	 * Retrieve the MSR number associated with the counter
4200 	 * 'pmc_id'.  This allows processes to directly use RDPMC
4201 	 * instructions to read their PMCs, without the overhead of a
4202 	 * system call.
4203 	 */
4204 
4205 	case PMC_OP_PMCGETMSR:
4206 	{
4207 		int adjri, ri;
4208 		struct pmc *pm;
4209 		struct pmc_target *pt;
4210 		struct pmc_op_getmsr gm;
4211 		struct pmc_classdep *pcd;
4212 
4213 		PMC_DOWNGRADE_SX();
4214 
4215 		if ((error = copyin(arg, &gm, sizeof(gm))) != 0)
4216 			break;
4217 
4218 		if ((error = pmc_find_pmc(gm.pm_pmcid, &pm)) != 0)
4219 			break;
4220 
4221 		/*
4222 		 * The allocated PMC has to be a process virtual PMC,
4223 		 * i.e., of type MODE_T[CS].  Global PMCs can only be
4224 		 * read using the PMCREAD operation since they may be
4225 		 * allocated on a different CPU than the one we could
4226 		 * be running on at the time of the RDPMC instruction.
4227 		 *
4228 		 * The GETMSR operation is not allowed for PMCs that
4229 		 * are inherited across processes.
4230 		 */
4231 
4232 		if (!PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)) ||
4233 		    (pm->pm_flags & PMC_F_DESCENDANTS)) {
4234 			error = EINVAL;
4235 			break;
4236 		}
4237 
4238 		/*
4239 		 * It only makes sense to use a RDPMC (or its
4240 		 * equivalent instruction on non-x86 architectures) on
4241 		 * a process that has allocated and attached a PMC to
4242 		 * itself.  Conversely the PMC is only allowed to have
4243 		 * one process attached to it -- its owner.
4244 		 */
4245 
4246 		if ((pt = LIST_FIRST(&pm->pm_targets)) == NULL ||
4247 		    LIST_NEXT(pt, pt_next) != NULL ||
4248 		    pt->pt_process->pp_proc != pm->pm_owner->po_owner) {
4249 			error = EINVAL;
4250 			break;
4251 		}
4252 
4253 		ri = PMC_TO_ROWINDEX(pm);
4254 		pcd = pmc_ri_to_classdep(md, ri, &adjri);
4255 
4256 		/* PMC class has no 'GETMSR' support */
4257 		if (pcd->pcd_get_msr == NULL) {
4258 			error = ENOSYS;
4259 			break;
4260 		}
4261 
4262 		if ((error = (*pcd->pcd_get_msr)(adjri, &gm.pm_msr)) < 0)
4263 			break;
4264 
4265 		if ((error = copyout(&gm, arg, sizeof(gm))) < 0)
4266 			break;
4267 
4268 		/*
4269 		 * Mark our process as using MSRs.  Update machine
4270 		 * state using a forced context switch.
4271 		 */
4272 
4273 		pt->pt_process->pp_flags |= PMC_PP_ENABLE_MSR_ACCESS;
4274 		pmc_force_context_switch();
4275 
4276 	}
4277 	break;
4278 
4279 	/*
4280 	 * Release an allocated PMC
4281 	 */
4282 
4283 	case PMC_OP_PMCRELEASE:
4284 	{
4285 		pmc_id_t pmcid;
4286 		struct pmc *pm;
4287 		struct pmc_owner *po;
4288 		struct pmc_op_simple sp;
4289 
4290 		/*
4291 		 * Find PMC pointer for the named PMC.
4292 		 *
4293 		 * Use pmc_release_pmc_descriptor() to switch off the
4294 		 * PMC, remove all its target threads, and remove the
4295 		 * PMC from its owner's list.
4296 		 *
4297 		 * Remove the owner record if this is the last PMC
4298 		 * owned.
4299 		 *
4300 		 * Free up space.
4301 		 */
4302 
4303 		if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
4304 			break;
4305 
4306 		pmcid = sp.pm_pmcid;
4307 
4308 		if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
4309 			break;
4310 
4311 		po = pm->pm_owner;
4312 		pmc_release_pmc_descriptor(pm);
4313 		pmc_maybe_remove_owner(po);
4314 		pmc_destroy_pmc_descriptor(pm);
4315 	}
4316 	break;
4317 
4318 
4319 	/*
4320 	 * Read and/or write a PMC.
4321 	 */
4322 
4323 	case PMC_OP_PMCRW:
4324 	{
4325 		int adjri;
4326 		struct pmc *pm;
4327 		uint32_t cpu, ri;
4328 		pmc_value_t oldvalue;
4329 		struct pmc_binding pb;
4330 		struct pmc_op_pmcrw prw;
4331 		struct pmc_classdep *pcd;
4332 		struct pmc_op_pmcrw *pprw;
4333 
4334 		PMC_DOWNGRADE_SX();
4335 
4336 		if ((error = copyin(arg, &prw, sizeof(prw))) != 0)
4337 			break;
4338 
4339 		PMCDBG2(PMC,OPS,1, "rw id=%d flags=0x%x", prw.pm_pmcid,
4340 		    prw.pm_flags);
4341 
4342 		/* must have at least one flag set */
4343 		if ((prw.pm_flags & (PMC_F_OLDVALUE|PMC_F_NEWVALUE)) == 0) {
4344 			error = EINVAL;
4345 			break;
4346 		}
4347 
4348 		/* locate pmc descriptor */
4349 		if ((error = pmc_find_pmc(prw.pm_pmcid, &pm)) != 0)
4350 			break;
4351 
4352 		/* Can't read a PMC that hasn't been started. */
4353 		if (pm->pm_state != PMC_STATE_ALLOCATED &&
4354 		    pm->pm_state != PMC_STATE_STOPPED &&
4355 		    pm->pm_state != PMC_STATE_RUNNING) {
4356 			error = EINVAL;
4357 			break;
4358 		}
4359 
4360 		/* writing a new value is allowed only for 'STOPPED' pmcs */
4361 		if (pm->pm_state == PMC_STATE_RUNNING &&
4362 		    (prw.pm_flags & PMC_F_NEWVALUE)) {
4363 			error = EBUSY;
4364 			break;
4365 		}
4366 
4367 		if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) {
4368 
4369 			/*
4370 			 * If this PMC is attached to its owner (i.e.,
4371 			 * the process requesting this operation) and
4372 			 * is running, then attempt to get an
4373 			 * upto-date reading from hardware for a READ.
4374 			 * Writes are only allowed when the PMC is
4375 			 * stopped, so only update the saved value
4376 			 * field.
4377 			 *
4378 			 * If the PMC is not running, or is not
4379 			 * attached to its owner, read/write to the
4380 			 * savedvalue field.
4381 			 */
4382 
4383 			ri = PMC_TO_ROWINDEX(pm);
4384 			pcd = pmc_ri_to_classdep(md, ri, &adjri);
4385 
4386 			mtx_pool_lock_spin(pmc_mtxpool, pm);
4387 			cpu = curthread->td_oncpu;
4388 
4389 			if (prw.pm_flags & PMC_F_OLDVALUE) {
4390 				if ((pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) &&
4391 				    (pm->pm_state == PMC_STATE_RUNNING))
4392 					error = (*pcd->pcd_read_pmc)(cpu, adjri,
4393 					    &oldvalue);
4394 				else
4395 					oldvalue = pm->pm_gv.pm_savedvalue;
4396 			}
4397 			if (prw.pm_flags & PMC_F_NEWVALUE)
4398 				pm->pm_gv.pm_savedvalue = prw.pm_value;
4399 
4400 			mtx_pool_unlock_spin(pmc_mtxpool, pm);
4401 
4402 		} else { /* System mode PMCs */
4403 			cpu = PMC_TO_CPU(pm);
4404 			ri  = PMC_TO_ROWINDEX(pm);
4405 			pcd = pmc_ri_to_classdep(md, ri, &adjri);
4406 
4407 			if (!pmc_cpu_is_active(cpu)) {
4408 				error = ENXIO;
4409 				break;
4410 			}
4411 
4412 			/* move this thread to CPU 'cpu' */
4413 			pmc_save_cpu_binding(&pb);
4414 			pmc_select_cpu(cpu);
4415 
4416 			critical_enter();
4417 			/* save old value */
4418 			if (prw.pm_flags & PMC_F_OLDVALUE)
4419 				if ((error = (*pcd->pcd_read_pmc)(cpu, adjri,
4420 					 &oldvalue)))
4421 					goto error;
4422 			/* write out new value */
4423 			if (prw.pm_flags & PMC_F_NEWVALUE)
4424 				error = (*pcd->pcd_write_pmc)(cpu, adjri,
4425 				    prw.pm_value);
4426 		error:
4427 			critical_exit();
4428 			pmc_restore_cpu_binding(&pb);
4429 			if (error)
4430 				break;
4431 		}
4432 
4433 		pprw = (struct pmc_op_pmcrw *) arg;
4434 
4435 #ifdef	HWPMC_DEBUG
4436 		if (prw.pm_flags & PMC_F_NEWVALUE)
4437 			PMCDBG3(PMC,OPS,2, "rw id=%d new %jx -> old %jx",
4438 			    ri, prw.pm_value, oldvalue);
4439 		else if (prw.pm_flags & PMC_F_OLDVALUE)
4440 			PMCDBG2(PMC,OPS,2, "rw id=%d -> old %jx", ri, oldvalue);
4441 #endif
4442 
4443 		/* return old value if requested */
4444 		if (prw.pm_flags & PMC_F_OLDVALUE)
4445 			if ((error = copyout(&oldvalue, &pprw->pm_value,
4446 				 sizeof(prw.pm_value))))
4447 				break;
4448 
4449 	}
4450 	break;
4451 
4452 
4453 	/*
4454 	 * Set the sampling rate for a sampling mode PMC and the
4455 	 * initial count for a counting mode PMC.
4456 	 */
4457 
4458 	case PMC_OP_PMCSETCOUNT:
4459 	{
4460 		struct pmc *pm;
4461 		struct pmc_op_pmcsetcount sc;
4462 
4463 		PMC_DOWNGRADE_SX();
4464 
4465 		if ((error = copyin(arg, &sc, sizeof(sc))) != 0)
4466 			break;
4467 
4468 		if ((error = pmc_find_pmc(sc.pm_pmcid, &pm)) != 0)
4469 			break;
4470 
4471 		if (pm->pm_state == PMC_STATE_RUNNING) {
4472 			error = EBUSY;
4473 			break;
4474 		}
4475 
4476 		if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) {
4477 			/*
4478 			 * Don't permit requested sample rate to be less than 1000
4479 			 */
4480 			if (sc.pm_count < 1000)
4481 				log(LOG_WARNING,
4482 					"pmcsetcount: passed sample rate %ju - setting to 1000\n",
4483 					(uintmax_t)sc.pm_count);
4484 			pm->pm_sc.pm_reloadcount = MAX(1000, sc.pm_count);
4485 		} else
4486 			pm->pm_sc.pm_initial = sc.pm_count;
4487 	}
4488 	break;
4489 
4490 
4491 	/*
4492 	 * Start a PMC.
4493 	 */
4494 
4495 	case PMC_OP_PMCSTART:
4496 	{
4497 		pmc_id_t pmcid;
4498 		struct pmc *pm;
4499 		struct pmc_op_simple sp;
4500 
4501 		sx_assert(&pmc_sx, SX_XLOCKED);
4502 
4503 		if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
4504 			break;
4505 
4506 		pmcid = sp.pm_pmcid;
4507 
4508 		if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
4509 			break;
4510 
4511 		KASSERT(pmcid == pm->pm_id,
4512 		    ("[pmc,%d] pmcid %x != id %x", __LINE__,
4513 			pm->pm_id, pmcid));
4514 
4515 		if (pm->pm_state == PMC_STATE_RUNNING) /* already running */
4516 			break;
4517 		else if (pm->pm_state != PMC_STATE_STOPPED &&
4518 		    pm->pm_state != PMC_STATE_ALLOCATED) {
4519 			error = EINVAL;
4520 			break;
4521 		}
4522 
4523 		error = pmc_start(pm);
4524 	}
4525 	break;
4526 
4527 
4528 	/*
4529 	 * Stop a PMC.
4530 	 */
4531 
4532 	case PMC_OP_PMCSTOP:
4533 	{
4534 		pmc_id_t pmcid;
4535 		struct pmc *pm;
4536 		struct pmc_op_simple sp;
4537 
4538 		PMC_DOWNGRADE_SX();
4539 
4540 		if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
4541 			break;
4542 
4543 		pmcid = sp.pm_pmcid;
4544 
4545 		/*
4546 		 * Mark the PMC as inactive and invoke the MD stop
4547 		 * routines if needed.
4548 		 */
4549 
4550 		if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
4551 			break;
4552 
4553 		KASSERT(pmcid == pm->pm_id,
4554 		    ("[pmc,%d] pmc id %x != pmcid %x", __LINE__,
4555 			pm->pm_id, pmcid));
4556 
4557 		if (pm->pm_state == PMC_STATE_STOPPED) /* already stopped */
4558 			break;
4559 		else if (pm->pm_state != PMC_STATE_RUNNING) {
4560 			error = EINVAL;
4561 			break;
4562 		}
4563 
4564 		error = pmc_stop(pm);
4565 	}
4566 	break;
4567 
4568 
4569 	/*
4570 	 * Write a user supplied value to the log file.
4571 	 */
4572 
4573 	case PMC_OP_WRITELOG:
4574 	{
4575 		struct pmc_op_writelog wl;
4576 		struct pmc_owner *po;
4577 
4578 		PMC_DOWNGRADE_SX();
4579 
4580 		if ((error = copyin(arg, &wl, sizeof(wl))) != 0)
4581 			break;
4582 
4583 		if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) {
4584 			error = EINVAL;
4585 			break;
4586 		}
4587 
4588 		if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) {
4589 			error = EINVAL;
4590 			break;
4591 		}
4592 
4593 		error = pmclog_process_userlog(po, &wl);
4594 	}
4595 	break;
4596 
4597 
4598 	default:
4599 		error = EINVAL;
4600 		break;
4601 	}
4602 
4603 	if (is_sx_downgraded)
4604 		sx_sunlock(&pmc_sx);
4605 	else
4606 		sx_xunlock(&pmc_sx);
4607 done_syscall:
4608 	if (error)
4609 		counter_u64_add(pmc_stats.pm_syscall_errors, 1);
4610 
4611 	return (error);
4612 }
4613 
4614 /*
4615  * Helper functions
4616  */
4617 
4618 
4619 /*
4620  * Mark the thread as needing callchain capture and post an AST.  The
4621  * actual callchain capture will be done in a context where it is safe
4622  * to take page faults.
4623  */
4624 
4625 static void
4626 pmc_post_callchain_callback(void)
4627 {
4628 	struct thread *td;
4629 
4630 	td = curthread;
4631 
4632 	/*
4633 	 * If there is multiple PMCs for the same interrupt ignore new post
4634 	 */
4635 	if (td->td_pflags & TDP_CALLCHAIN)
4636 		return;
4637 
4638 	/*
4639 	 * Mark this thread as needing callchain capture.
4640 	 * `td->td_pflags' will be safe to touch because this thread
4641 	 * was in user space when it was interrupted.
4642 	 */
4643 	td->td_pflags |= TDP_CALLCHAIN;
4644 
4645 	/*
4646 	 * Don't let this thread migrate between CPUs until callchain
4647 	 * capture completes.
4648 	 */
4649 	sched_pin();
4650 
4651 	return;
4652 }
4653 
4654 /*
4655  * Find a free slot in the per-cpu array of samples and capture the
4656  * current callchain there.  If a sample was successfully added, a bit
4657  * is set in mask 'pmc_cpumask' denoting that the DO_SAMPLES hook
4658  * needs to be invoked from the clock handler.
4659  *
4660  * This function is meant to be called from an NMI handler.  It cannot
4661  * use any of the locking primitives supplied by the OS.
4662  */
4663 
4664 static int
4665 pmc_add_sample(ring_type_t ring, struct pmc *pm, struct trapframe *tf)
4666 {
4667 	int error, cpu, callchaindepth, inuserspace;
4668 	struct thread *td;
4669 	struct pmc_sample *ps;
4670 	struct pmc_samplebuffer *psb;
4671 
4672 	error = 0;
4673 
4674 	/*
4675 	 * Allocate space for a sample buffer.
4676 	 */
4677 	cpu = curcpu;
4678 	psb = pmc_pcpu[cpu]->pc_sb[ring];
4679 	inuserspace = TRAPF_USERMODE(tf);
4680 	ps = PMC_PROD_SAMPLE(psb);
4681 	if (psb->ps_considx != psb->ps_prodidx &&
4682 		ps->ps_nsamples) {	/* in use, reader hasn't caught up */
4683 		pm->pm_pcpu_state[cpu].pps_stalled = 1;
4684 		counter_u64_add(pmc_stats.pm_intr_bufferfull, 1);
4685 		PMCDBG6(SAM,INT,1,"(spc) cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d",
4686 		    cpu, pm, (void *) tf, inuserspace,
4687 		    (int) (psb->ps_prodidx & pmc_sample_mask),
4688 		    (int) (psb->ps_considx & pmc_sample_mask));
4689 		callchaindepth = 1;
4690 		error = ENOMEM;
4691 		goto done;
4692 	}
4693 
4694 	/* Fill in entry. */
4695 	PMCDBG6(SAM,INT,1,"cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d", cpu, pm,
4696 	    (void *) tf, inuserspace,
4697 	    (int) (psb->ps_prodidx & pmc_sample_mask),
4698 	    (int) (psb->ps_considx & pmc_sample_mask));
4699 
4700 	td = curthread;
4701 	ps->ps_pmc = pm;
4702 	ps->ps_td = td;
4703 	ps->ps_pid = td->td_proc->p_pid;
4704 	ps->ps_tid = td->td_tid;
4705 	ps->ps_tsc = pmc_rdtsc();
4706 	ps->ps_ticks = ticks;
4707 	ps->ps_cpu = cpu;
4708 	ps->ps_flags = inuserspace ? PMC_CC_F_USERSPACE : 0;
4709 
4710 	callchaindepth = (pm->pm_flags & PMC_F_CALLCHAIN) ?
4711 	    pmc_callchaindepth : 1;
4712 
4713 	MPASS(ps->ps_pc != NULL);
4714 	if (callchaindepth == 1)
4715 		ps->ps_pc[0] = PMC_TRAPFRAME_TO_PC(tf);
4716 	else {
4717 		/*
4718 		 * Kernel stack traversals can be done immediately,
4719 		 * while we defer to an AST for user space traversals.
4720 		 */
4721 		if (!inuserspace) {
4722 			callchaindepth =
4723 			    pmc_save_kernel_callchain(ps->ps_pc,
4724 				callchaindepth, tf);
4725 		} else {
4726 			pmc_post_callchain_callback();
4727 			callchaindepth = PMC_USER_CALLCHAIN_PENDING;
4728 		}
4729 	}
4730 
4731 	ps->ps_nsamples = callchaindepth;	/* mark entry as in use */
4732 	if (ring == PMC_UR) {
4733 		ps->ps_nsamples_actual = callchaindepth;	/* mark entry as in use */
4734 		ps->ps_nsamples = PMC_USER_CALLCHAIN_PENDING;
4735 	} else
4736 		ps->ps_nsamples = callchaindepth;	/* mark entry as in use */
4737 
4738 	KASSERT(counter_u64_fetch(pm->pm_runcount) >= 0,
4739 	    ("[pmc,%d] pm=%p runcount %ld", __LINE__, (void *) pm,
4740 		 (unsigned long)counter_u64_fetch(pm->pm_runcount)));
4741 
4742 	counter_u64_add(pm->pm_runcount, 1);	/* hold onto PMC */
4743 	/* increment write pointer */
4744 	psb->ps_prodidx++;
4745  done:
4746 	/* mark CPU as needing processing */
4747 	if (callchaindepth != PMC_USER_CALLCHAIN_PENDING)
4748 		DPCPU_SET(pmc_sampled, 1);
4749 
4750 	return (error);
4751 }
4752 
4753 /*
4754  * Interrupt processing.
4755  *
4756  * This function is meant to be called from an NMI handler.  It cannot
4757  * use any of the locking primitives supplied by the OS.
4758  */
4759 
4760 int
4761 pmc_process_interrupt(int ring, struct pmc *pm, struct trapframe *tf)
4762 {
4763 	struct thread *td;
4764 
4765 	td = curthread;
4766 	if ((pm->pm_flags & PMC_F_USERCALLCHAIN) &&
4767 	    (td->td_proc->p_flag & P_KPROC) == 0 &&
4768 	    !TRAPF_USERMODE(tf)) {
4769 		atomic_add_int(&td->td_pmcpend, 1);
4770 		return (pmc_add_sample(PMC_UR, pm, tf));
4771 	}
4772 	return (pmc_add_sample(ring, pm, tf));
4773 }
4774 
4775 /*
4776  * Capture a user call chain.  This function will be called from ast()
4777  * before control returns to userland and before the process gets
4778  * rescheduled.
4779  */
4780 
4781 static void
4782 pmc_capture_user_callchain(int cpu, int ring, struct trapframe *tf)
4783 {
4784 	struct pmc *pm;
4785 	struct thread *td;
4786 	struct pmc_sample *ps;
4787 	struct pmc_samplebuffer *psb;
4788 	uint64_t considx, prodidx;
4789 	int nsamples, nrecords, pass, iter;
4790 #ifdef	INVARIANTS
4791 	int ncallchains;
4792 	int nfree;
4793 	int start_ticks = ticks;
4794 #endif
4795 	psb = pmc_pcpu[cpu]->pc_sb[ring];
4796 	td = curthread;
4797 
4798 	KASSERT(td->td_pflags & TDP_CALLCHAIN,
4799 	    ("[pmc,%d] Retrieving callchain for thread that doesn't want it",
4800 		__LINE__));
4801 
4802 #ifdef	INVARIANTS
4803 	ncallchains = 0;
4804 	nfree = 0;
4805 #endif
4806 	nrecords = INT_MAX;
4807 	pass = 0;
4808  restart:
4809 	if (ring == PMC_UR)
4810 		nrecords = atomic_readandclear_32(&td->td_pmcpend);
4811 
4812 	for (iter = 0, considx = psb->ps_considx, prodidx = psb->ps_prodidx;
4813 	    considx < prodidx && iter < pmc_nsamples; considx++, iter++) {
4814 		ps = PMC_CONS_SAMPLE_OFF(psb, considx);
4815 
4816 	/*
4817 	 * Iterate through all deferred callchain requests.
4818 	 * Walk from the current read pointer to the current
4819 	 * write pointer.
4820 	 */
4821 
4822 #ifdef	INVARIANTS
4823 		if (ps->ps_nsamples == PMC_SAMPLE_FREE) {
4824 			nfree++;
4825 			continue;
4826 		}
4827 
4828 		if ((ps->ps_pmc == NULL) ||
4829 		    (ps->ps_pmc->pm_state != PMC_STATE_RUNNING))
4830 			nfree++;
4831 #endif
4832 		if (ps->ps_td != td ||
4833 		   ps->ps_nsamples != PMC_USER_CALLCHAIN_PENDING ||
4834 		   ps->ps_pmc->pm_state != PMC_STATE_RUNNING)
4835 			continue;
4836 
4837 		KASSERT(ps->ps_cpu == cpu,
4838 		    ("[pmc,%d] cpu mismatch ps_cpu=%d pcpu=%d", __LINE__,
4839 			ps->ps_cpu, PCPU_GET(cpuid)));
4840 
4841 		pm = ps->ps_pmc;
4842 
4843 		KASSERT(pm->pm_flags & PMC_F_CALLCHAIN,
4844 		    ("[pmc,%d] Retrieving callchain for PMC that doesn't "
4845 			"want it", __LINE__));
4846 
4847 		KASSERT(counter_u64_fetch(pm->pm_runcount) > 0,
4848 		    ("[pmc,%d] runcount %ld", __LINE__, (unsigned long)counter_u64_fetch(pm->pm_runcount)));
4849 
4850 		if (ring == PMC_UR) {
4851 			nsamples = ps->ps_nsamples_actual;
4852 			counter_u64_add(pmc_stats.pm_merges, 1);
4853 		} else
4854 			nsamples = 0;
4855 
4856 		/*
4857 		 * Retrieve the callchain and mark the sample buffer
4858 		 * as 'processable' by the timer tick sweep code.
4859 		 */
4860 
4861 #ifdef INVARIANTS
4862 		ncallchains++;
4863 #endif
4864 
4865 		if (__predict_true(nsamples < pmc_callchaindepth - 1))
4866 			nsamples += pmc_save_user_callchain(ps->ps_pc + nsamples,
4867 		       pmc_callchaindepth - nsamples - 1, tf);
4868 
4869 		/*
4870 		 * We have to prevent hardclock from potentially overwriting
4871 		 * this sample between when we read the value and when we set
4872 		 * it
4873 		 */
4874 		spinlock_enter();
4875 		/*
4876 		 * Verify that the sample hasn't been dropped in the meantime
4877 		 */
4878 		if (ps->ps_nsamples == PMC_USER_CALLCHAIN_PENDING) {
4879 			ps->ps_nsamples = nsamples;
4880 			/*
4881 			 * If we couldn't get a sample, simply drop the reference
4882 			 */
4883 			if (nsamples == 0)
4884 				counter_u64_add(pm->pm_runcount, -1);
4885 		}
4886 		spinlock_exit();
4887 		if (nrecords-- == 1)
4888 			break;
4889 	}
4890 	if (__predict_false(ring == PMC_UR && td->td_pmcpend)) {
4891 		if (pass == 0) {
4892 			pass = 1;
4893 			goto restart;
4894 		}
4895 		/* only collect samples for this part once */
4896 		td->td_pmcpend = 0;
4897 	}
4898 
4899 #ifdef INVARIANTS
4900 	if ((ticks - start_ticks) > hz)
4901 		log(LOG_ERR, "%s took %d ticks\n", __func__, (ticks - start_ticks));
4902 #endif
4903 
4904 	/* mark CPU as needing processing */
4905 	DPCPU_SET(pmc_sampled, 1);
4906 }
4907 
4908 /*
4909  * Process saved PC samples.
4910  */
4911 
4912 static void
4913 pmc_process_samples(int cpu, ring_type_t ring)
4914 {
4915 	struct pmc *pm;
4916 	int adjri, n;
4917 	struct thread *td;
4918 	struct pmc_owner *po;
4919 	struct pmc_sample *ps;
4920 	struct pmc_classdep *pcd;
4921 	struct pmc_samplebuffer *psb;
4922 	uint64_t delta __diagused;
4923 
4924 	KASSERT(PCPU_GET(cpuid) == cpu,
4925 	    ("[pmc,%d] not on the correct CPU pcpu=%d cpu=%d", __LINE__,
4926 		PCPU_GET(cpuid), cpu));
4927 
4928 	psb = pmc_pcpu[cpu]->pc_sb[ring];
4929 	delta = psb->ps_prodidx - psb->ps_considx;
4930 	MPASS(delta <= pmc_nsamples);
4931 	MPASS(psb->ps_considx <= psb->ps_prodidx);
4932 	for (n = 0; psb->ps_considx < psb->ps_prodidx; psb->ps_considx++, n++) {
4933 		ps = PMC_CONS_SAMPLE(psb);
4934 
4935 		if (__predict_false(ps->ps_nsamples == PMC_SAMPLE_FREE))
4936 			continue;
4937 		pm = ps->ps_pmc;
4938 		/* skip non-running samples */
4939 		if (pm->pm_state != PMC_STATE_RUNNING)
4940 			goto entrydone;
4941 
4942 		KASSERT(counter_u64_fetch(pm->pm_runcount) > 0,
4943 		    ("[pmc,%d] pm=%p runcount %ld", __LINE__, (void *) pm,
4944 			 (unsigned long)counter_u64_fetch(pm->pm_runcount)));
4945 
4946 		po = pm->pm_owner;
4947 
4948 		KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)),
4949 		    ("[pmc,%d] pmc=%p non-sampling mode=%d", __LINE__,
4950 			pm, PMC_TO_MODE(pm)));
4951 
4952 
4953 		/* If there is a pending AST wait for completion */
4954 		if (ps->ps_nsamples == PMC_USER_CALLCHAIN_PENDING) {
4955 			/* if we've been waiting more than 1 tick to
4956 			 * collect a callchain for this record then
4957 			 * drop it and move on.
4958 			 */
4959 			if (ticks - ps->ps_ticks > 1) {
4960 				/*
4961 				 * track how often we hit this as it will
4962 				 * preferentially lose user samples
4963 				 * for long running system calls
4964 				 */
4965 				counter_u64_add(pmc_stats.pm_overwrites, 1);
4966 				goto entrydone;
4967 			}
4968 			/* Need a rescan at a later time. */
4969 			DPCPU_SET(pmc_sampled, 1);
4970 			break;
4971 		}
4972 
4973 		PMCDBG6(SAM,OPS,1,"cpu=%d pm=%p n=%d fl=%x wr=%d rd=%d", cpu,
4974 		    pm, ps->ps_nsamples, ps->ps_flags,
4975 		    (int) (psb->ps_prodidx & pmc_sample_mask),
4976 		    (int) (psb->ps_considx & pmc_sample_mask));
4977 
4978 		/*
4979 		 * If this is a process-mode PMC that is attached to
4980 		 * its owner, and if the PC is in user mode, update
4981 		 * profiling statistics like timer-based profiling
4982 		 * would have done.
4983 		 *
4984 		 * Otherwise, this is either a sampling-mode PMC that
4985 		 * is attached to a different process than its owner,
4986 		 * or a system-wide sampling PMC. Dispatch a log
4987 		 * entry to the PMC's owner process.
4988 		 */
4989 		if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) {
4990 			if (ps->ps_flags & PMC_CC_F_USERSPACE) {
4991 				td = FIRST_THREAD_IN_PROC(po->po_owner);
4992 				addupc_intr(td, ps->ps_pc[0], 1);
4993 			}
4994 		} else
4995 			pmclog_process_callchain(pm, ps);
4996 
4997 	entrydone:
4998 		ps->ps_nsamples = 0; /* mark entry as free */
4999 		KASSERT(counter_u64_fetch(pm->pm_runcount) > 0,
5000 				("[pmc,%d] pm=%p runcount %ld", __LINE__, (void *) pm,
5001 				 (unsigned long)counter_u64_fetch(pm->pm_runcount)));
5002 
5003 		counter_u64_add(pm->pm_runcount, -1);
5004 	}
5005 
5006 	counter_u64_add(pmc_stats.pm_log_sweeps, 1);
5007 
5008 	/* Do not re-enable stalled PMCs if we failed to process any samples */
5009 	if (n == 0)
5010 		return;
5011 
5012 	/*
5013 	 * Restart any stalled sampling PMCs on this CPU.
5014 	 *
5015 	 * If the NMI handler sets the pm_stalled field of a PMC after
5016 	 * the check below, we'll end up processing the stalled PMC at
5017 	 * the next hardclock tick.
5018 	 */
5019 	for (n = 0; n < md->pmd_npmc; n++) {
5020 		pcd = pmc_ri_to_classdep(md, n, &adjri);
5021 		KASSERT(pcd != NULL,
5022 		    ("[pmc,%d] null pcd ri=%d", __LINE__, n));
5023 		(void) (*pcd->pcd_get_config)(cpu,adjri,&pm);
5024 
5025 		if (pm == NULL ||			 /* !cfg'ed */
5026 		    pm->pm_state != PMC_STATE_RUNNING || /* !active */
5027 		    !PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) || /* !sampling */
5028 			!pm->pm_pcpu_state[cpu].pps_cpustate  || /* !desired */
5029 		    !pm->pm_pcpu_state[cpu].pps_stalled) /* !stalled */
5030 			continue;
5031 
5032 		pm->pm_pcpu_state[cpu].pps_stalled = 0;
5033 		(*pcd->pcd_start_pmc)(cpu, adjri);
5034 	}
5035 }
5036 
5037 /*
5038  * Event handlers.
5039  */
5040 
5041 /*
5042  * Handle a process exit.
5043  *
5044  * Remove this process from all hash tables.  If this process
5045  * owned any PMCs, turn off those PMCs and deallocate them,
5046  * removing any associations with target processes.
5047  *
5048  * This function will be called by the last 'thread' of a
5049  * process.
5050  *
5051  * XXX This eventhandler gets called early in the exit process.
5052  * Consider using a 'hook' invocation from thread_exit() or equivalent
5053  * spot.  Another negative is that kse_exit doesn't seem to call
5054  * exit1() [??].
5055  *
5056  */
5057 
5058 static void
5059 pmc_process_exit(void *arg __unused, struct proc *p)
5060 {
5061 	struct pmc *pm;
5062 	int adjri, cpu;
5063 	unsigned int ri;
5064 	int is_using_hwpmcs;
5065 	struct pmc_owner *po;
5066 	struct pmc_process *pp;
5067 	struct pmc_classdep *pcd;
5068 	pmc_value_t newvalue, tmp;
5069 
5070 	PROC_LOCK(p);
5071 	is_using_hwpmcs = p->p_flag & P_HWPMC;
5072 	PROC_UNLOCK(p);
5073 
5074 	/*
5075 	 * Log a sysexit event to all SS PMC owners.
5076 	 */
5077 	PMC_EPOCH_ENTER();
5078 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
5079 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
5080 		    pmclog_process_sysexit(po, p->p_pid);
5081 	PMC_EPOCH_EXIT();
5082 
5083 	if (!is_using_hwpmcs)
5084 		return;
5085 
5086 	PMC_GET_SX_XLOCK();
5087 	PMCDBG3(PRC,EXT,1,"process-exit proc=%p (%d, %s)", p, p->p_pid,
5088 	    p->p_comm);
5089 
5090 	/*
5091 	 * Since this code is invoked by the last thread in an exiting
5092 	 * process, we would have context switched IN at some prior
5093 	 * point.  However, with PREEMPTION, kernel mode context
5094 	 * switches may happen any time, so we want to disable a
5095 	 * context switch OUT till we get any PMCs targeting this
5096 	 * process off the hardware.
5097 	 *
5098 	 * We also need to atomically remove this process'
5099 	 * entry from our target process hash table, using
5100 	 * PMC_FLAG_REMOVE.
5101 	 */
5102 	PMCDBG3(PRC,EXT,1, "process-exit proc=%p (%d, %s)", p, p->p_pid,
5103 	    p->p_comm);
5104 
5105 	critical_enter(); /* no preemption */
5106 
5107 	cpu = curthread->td_oncpu;
5108 
5109 	if ((pp = pmc_find_process_descriptor(p,
5110 		 PMC_FLAG_REMOVE)) != NULL) {
5111 
5112 		PMCDBG2(PRC,EXT,2,
5113 		    "process-exit proc=%p pmc-process=%p", p, pp);
5114 
5115 		/*
5116 		 * The exiting process could the target of
5117 		 * some PMCs which will be running on
5118 		 * currently executing CPU.
5119 		 *
5120 		 * We need to turn these PMCs off like we
5121 		 * would do at context switch OUT time.
5122 		 */
5123 		for (ri = 0; ri < md->pmd_npmc; ri++) {
5124 
5125 			/*
5126 			 * Pick up the pmc pointer from hardware
5127 			 * state similar to the CSW_OUT code.
5128 			 */
5129 			pm = NULL;
5130 
5131 			pcd = pmc_ri_to_classdep(md, ri, &adjri);
5132 
5133 			(void) (*pcd->pcd_get_config)(cpu, adjri, &pm);
5134 
5135 			PMCDBG2(PRC,EXT,2, "ri=%d pm=%p", ri, pm);
5136 
5137 			if (pm == NULL ||
5138 			    !PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)))
5139 				continue;
5140 
5141 			PMCDBG4(PRC,EXT,2, "ppmcs[%d]=%p pm=%p "
5142 			    "state=%d", ri, pp->pp_pmcs[ri].pp_pmc,
5143 			    pm, pm->pm_state);
5144 
5145 			KASSERT(PMC_TO_ROWINDEX(pm) == ri,
5146 			    ("[pmc,%d] ri mismatch pmc(%d) ri(%d)",
5147 				__LINE__, PMC_TO_ROWINDEX(pm), ri));
5148 
5149 			KASSERT(pm == pp->pp_pmcs[ri].pp_pmc,
5150 			    ("[pmc,%d] pm %p != pp_pmcs[%d] %p",
5151 				__LINE__, pm, ri, pp->pp_pmcs[ri].pp_pmc));
5152 
5153 			KASSERT(counter_u64_fetch(pm->pm_runcount) > 0,
5154 			    ("[pmc,%d] bad runcount ri %d rc %ld",
5155 				 __LINE__, ri, (unsigned long)counter_u64_fetch(pm->pm_runcount)));
5156 
5157 			/*
5158 			 * Change desired state, and then stop if not
5159 			 * stalled. This two-step dance should avoid
5160 			 * race conditions where an interrupt re-enables
5161 			 * the PMC after this code has already checked
5162 			 * the pm_stalled flag.
5163 			 */
5164 			if (pm->pm_pcpu_state[cpu].pps_cpustate) {
5165 				pm->pm_pcpu_state[cpu].pps_cpustate = 0;
5166 				if (!pm->pm_pcpu_state[cpu].pps_stalled) {
5167 					(void) pcd->pcd_stop_pmc(cpu, adjri);
5168 
5169 					if (PMC_TO_MODE(pm) == PMC_MODE_TC) {
5170 						pcd->pcd_read_pmc(cpu, adjri,
5171 						    &newvalue);
5172 						tmp = newvalue -
5173 						    PMC_PCPU_SAVED(cpu,ri);
5174 
5175 						mtx_pool_lock_spin(pmc_mtxpool,
5176 						    pm);
5177 						pm->pm_gv.pm_savedvalue += tmp;
5178 						pp->pp_pmcs[ri].pp_pmcval +=
5179 						    tmp;
5180 						mtx_pool_unlock_spin(
5181 						    pmc_mtxpool, pm);
5182 					}
5183 				}
5184 			}
5185 
5186 			KASSERT((int64_t) counter_u64_fetch(pm->pm_runcount) > 0,
5187 			    ("[pmc,%d] runcount is %d", __LINE__, ri));
5188 
5189 			counter_u64_add(pm->pm_runcount, -1);
5190 
5191 			(void) pcd->pcd_config_pmc(cpu, adjri, NULL);
5192 		}
5193 
5194 		/*
5195 		 * Inform the MD layer of this pseudo "context switch
5196 		 * out"
5197 		 */
5198 		(void) md->pmd_switch_out(pmc_pcpu[cpu], pp);
5199 
5200 		critical_exit(); /* ok to be pre-empted now */
5201 
5202 		/*
5203 		 * Unlink this process from the PMCs that are
5204 		 * targeting it.  This will send a signal to
5205 		 * all PMC owner's whose PMCs are orphaned.
5206 		 *
5207 		 * Log PMC value at exit time if requested.
5208 		 */
5209 		for (ri = 0; ri < md->pmd_npmc; ri++)
5210 			if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) {
5211 				if (pm->pm_flags & PMC_F_NEEDS_LOGFILE &&
5212 				    PMC_IS_COUNTING_MODE(PMC_TO_MODE(pm)))
5213 					pmclog_process_procexit(pm, pp);
5214 				pmc_unlink_target_process(pm, pp);
5215 			}
5216 		free(pp, M_PMC);
5217 
5218 	} else
5219 		critical_exit(); /* pp == NULL */
5220 
5221 
5222 	/*
5223 	 * If the process owned PMCs, free them up and free up
5224 	 * memory.
5225 	 */
5226 	if ((po = pmc_find_owner_descriptor(p)) != NULL) {
5227 		pmc_remove_owner(po);
5228 		pmc_destroy_owner_descriptor(po);
5229 	}
5230 
5231 	sx_xunlock(&pmc_sx);
5232 }
5233 
5234 /*
5235  * Handle a process fork.
5236  *
5237  * If the parent process 'p1' is under HWPMC monitoring, then copy
5238  * over any attached PMCs that have 'do_descendants' semantics.
5239  */
5240 
5241 static void
5242 pmc_process_fork(void *arg __unused, struct proc *p1, struct proc *newproc,
5243     int flags)
5244 {
5245 	int is_using_hwpmcs;
5246 	unsigned int ri;
5247 	uint32_t do_descendants;
5248 	struct pmc *pm;
5249 	struct pmc_owner *po;
5250 	struct pmc_process *ppnew, *ppold;
5251 
5252 	(void) flags;		/* unused parameter */
5253 
5254 	PROC_LOCK(p1);
5255 	is_using_hwpmcs = p1->p_flag & P_HWPMC;
5256 	PROC_UNLOCK(p1);
5257 
5258 	/*
5259 	 * If there are system-wide sampling PMCs active, we need to
5260 	 * log all fork events to their owner's logs.
5261 	 */
5262 	PMC_EPOCH_ENTER();
5263 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
5264 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE) {
5265 		    pmclog_process_procfork(po, p1->p_pid, newproc->p_pid);
5266 			pmclog_process_proccreate(po, newproc, 1);
5267 		}
5268 	PMC_EPOCH_EXIT();
5269 
5270 	if (!is_using_hwpmcs)
5271 		return;
5272 
5273 	PMC_GET_SX_XLOCK();
5274 	PMCDBG4(PMC,FRK,1, "process-fork proc=%p (%d, %s) -> %p", p1,
5275 	    p1->p_pid, p1->p_comm, newproc);
5276 
5277 	/*
5278 	 * If the parent process (curthread->td_proc) is a
5279 	 * target of any PMCs, look for PMCs that are to be
5280 	 * inherited, and link these into the new process
5281 	 * descriptor.
5282 	 */
5283 	if ((ppold = pmc_find_process_descriptor(curthread->td_proc,
5284 		 PMC_FLAG_NONE)) == NULL)
5285 		goto done;		/* nothing to do */
5286 
5287 	do_descendants = 0;
5288 	for (ri = 0; ri < md->pmd_npmc; ri++)
5289 		if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL)
5290 			do_descendants |= pm->pm_flags & PMC_F_DESCENDANTS;
5291 	if (do_descendants == 0) /* nothing to do */
5292 		goto done;
5293 
5294 	/*
5295 	 * Now mark the new process as being tracked by this driver.
5296 	 */
5297 	PROC_LOCK(newproc);
5298 	newproc->p_flag |= P_HWPMC;
5299 	PROC_UNLOCK(newproc);
5300 
5301 	/* allocate a descriptor for the new process  */
5302 	if ((ppnew = pmc_find_process_descriptor(newproc,
5303 		 PMC_FLAG_ALLOCATE)) == NULL)
5304 		goto done;
5305 
5306 	/*
5307 	 * Run through all PMCs that were targeting the old process
5308 	 * and which specified F_DESCENDANTS and attach them to the
5309 	 * new process.
5310 	 *
5311 	 * Log the fork event to all owners of PMCs attached to this
5312 	 * process, if not already logged.
5313 	 */
5314 	for (ri = 0; ri < md->pmd_npmc; ri++)
5315 		if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL &&
5316 		    (pm->pm_flags & PMC_F_DESCENDANTS)) {
5317 			pmc_link_target_process(pm, ppnew);
5318 			po = pm->pm_owner;
5319 			if (po->po_sscount == 0 &&
5320 			    po->po_flags & PMC_PO_OWNS_LOGFILE)
5321 				pmclog_process_procfork(po, p1->p_pid,
5322 				    newproc->p_pid);
5323 		}
5324 
5325  done:
5326 	sx_xunlock(&pmc_sx);
5327 }
5328 
5329 static void
5330 pmc_process_threadcreate(struct thread *td)
5331 {
5332 	struct pmc_owner *po;
5333 
5334 	PMC_EPOCH_ENTER();
5335 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
5336 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
5337 			pmclog_process_threadcreate(po, td, 1);
5338 	PMC_EPOCH_EXIT();
5339 }
5340 
5341 static void
5342 pmc_process_threadexit(struct thread *td)
5343 {
5344 	struct pmc_owner *po;
5345 
5346 	PMC_EPOCH_ENTER();
5347 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
5348 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
5349 			pmclog_process_threadexit(po, td);
5350 	PMC_EPOCH_EXIT();
5351 }
5352 
5353 static void
5354 pmc_process_proccreate(struct proc *p)
5355 {
5356 	struct pmc_owner *po;
5357 
5358 	PMC_EPOCH_ENTER();
5359 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
5360 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
5361 			pmclog_process_proccreate(po, p, 1 /* sync */);
5362 	PMC_EPOCH_EXIT();
5363 }
5364 
5365 static void
5366 pmc_process_allproc(struct pmc *pm)
5367 {
5368 	struct pmc_owner *po;
5369 	struct thread *td;
5370 	struct proc *p;
5371 
5372 	po = pm->pm_owner;
5373 	if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
5374 		return;
5375 	sx_slock(&allproc_lock);
5376 	FOREACH_PROC_IN_SYSTEM(p) {
5377 		pmclog_process_proccreate(po, p, 0 /* sync */);
5378 		PROC_LOCK(p);
5379 		FOREACH_THREAD_IN_PROC(p, td)
5380 			pmclog_process_threadcreate(po, td, 0 /* sync */);
5381 		PROC_UNLOCK(p);
5382 	}
5383 	sx_sunlock(&allproc_lock);
5384 	pmclog_flush(po, 0);
5385 }
5386 
5387 static void
5388 pmc_kld_load(void *arg __unused, linker_file_t lf)
5389 {
5390 	struct pmc_owner *po;
5391 
5392 	/*
5393 	 * Notify owners of system sampling PMCs about KLD operations.
5394 	 */
5395 	PMC_EPOCH_ENTER();
5396 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
5397 		if (po->po_flags & PMC_PO_OWNS_LOGFILE)
5398 			pmclog_process_map_in(po, (pid_t) -1,
5399 			    (uintfptr_t) lf->address, lf->filename);
5400 	PMC_EPOCH_EXIT();
5401 
5402 	/*
5403 	 * TODO: Notify owners of (all) process-sampling PMCs too.
5404 	 */
5405 }
5406 
5407 static void
5408 pmc_kld_unload(void *arg __unused, const char *filename __unused,
5409     caddr_t address, size_t size)
5410 {
5411 	struct pmc_owner *po;
5412 
5413 	PMC_EPOCH_ENTER();
5414 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
5415 		if (po->po_flags & PMC_PO_OWNS_LOGFILE)
5416 			pmclog_process_map_out(po, (pid_t) -1,
5417 			    (uintfptr_t) address, (uintfptr_t) address + size);
5418 	PMC_EPOCH_EXIT();
5419 
5420 	/*
5421 	 * TODO: Notify owners of process-sampling PMCs.
5422 	 */
5423 }
5424 
5425 /*
5426  * initialization
5427  */
5428 static const char *
5429 pmc_name_of_pmcclass(enum pmc_class class)
5430 {
5431 
5432 	switch (class) {
5433 #undef	__PMC_CLASS
5434 #define	__PMC_CLASS(S,V,D)						\
5435 	case PMC_CLASS_##S:						\
5436 		return #S;
5437 	__PMC_CLASSES();
5438 	default:
5439 		return ("<unknown>");
5440 	}
5441 }
5442 
5443 /*
5444  * Base class initializer: allocate structure and set default classes.
5445  */
5446 struct pmc_mdep *
5447 pmc_mdep_alloc(int nclasses)
5448 {
5449 	struct pmc_mdep *md;
5450 	int	n;
5451 
5452 	/* SOFT + md classes */
5453 	n = 1 + nclasses;
5454 	md = malloc(sizeof(struct pmc_mdep) + n *
5455 	    sizeof(struct pmc_classdep), M_PMC, M_WAITOK|M_ZERO);
5456 	md->pmd_nclass = n;
5457 
5458 	/* Add base class. */
5459 	pmc_soft_initialize(md);
5460 	return md;
5461 }
5462 
5463 void
5464 pmc_mdep_free(struct pmc_mdep *md)
5465 {
5466 	pmc_soft_finalize(md);
5467 	free(md, M_PMC);
5468 }
5469 
5470 static int
5471 generic_switch_in(struct pmc_cpu *pc, struct pmc_process *pp)
5472 {
5473 	(void) pc; (void) pp;
5474 
5475 	return (0);
5476 }
5477 
5478 static int
5479 generic_switch_out(struct pmc_cpu *pc, struct pmc_process *pp)
5480 {
5481 	(void) pc; (void) pp;
5482 
5483 	return (0);
5484 }
5485 
5486 static struct pmc_mdep *
5487 pmc_generic_cpu_initialize(void)
5488 {
5489 	struct pmc_mdep *md;
5490 
5491 	md = pmc_mdep_alloc(0);
5492 
5493 	md->pmd_cputype    = PMC_CPU_GENERIC;
5494 
5495 	md->pmd_pcpu_init  = NULL;
5496 	md->pmd_pcpu_fini  = NULL;
5497 	md->pmd_switch_in  = generic_switch_in;
5498 	md->pmd_switch_out = generic_switch_out;
5499 
5500 	return (md);
5501 }
5502 
5503 static void
5504 pmc_generic_cpu_finalize(struct pmc_mdep *md)
5505 {
5506 	(void) md;
5507 }
5508 
5509 
5510 static int
5511 pmc_initialize(void)
5512 {
5513 	int c, cpu, error, n, ri;
5514 	unsigned int maxcpu, domain;
5515 	struct pcpu *pc;
5516 	struct pmc_binding pb;
5517 	struct pmc_sample *ps;
5518 	struct pmc_classdep *pcd;
5519 	struct pmc_samplebuffer *sb;
5520 
5521 	md = NULL;
5522 	error = 0;
5523 
5524 	pmc_stats.pm_intr_ignored = counter_u64_alloc(M_WAITOK);
5525 	pmc_stats.pm_intr_processed = counter_u64_alloc(M_WAITOK);
5526 	pmc_stats.pm_intr_bufferfull = counter_u64_alloc(M_WAITOK);
5527 	pmc_stats.pm_syscalls = counter_u64_alloc(M_WAITOK);
5528 	pmc_stats.pm_syscall_errors = counter_u64_alloc(M_WAITOK);
5529 	pmc_stats.pm_buffer_requests = counter_u64_alloc(M_WAITOK);
5530 	pmc_stats.pm_buffer_requests_failed = counter_u64_alloc(M_WAITOK);
5531 	pmc_stats.pm_log_sweeps = counter_u64_alloc(M_WAITOK);
5532 	pmc_stats.pm_merges = counter_u64_alloc(M_WAITOK);
5533 	pmc_stats.pm_overwrites = counter_u64_alloc(M_WAITOK);
5534 
5535 #ifdef	HWPMC_DEBUG
5536 	/* parse debug flags first */
5537 	if (TUNABLE_STR_FETCH(PMC_SYSCTL_NAME_PREFIX "debugflags",
5538 		pmc_debugstr, sizeof(pmc_debugstr)))
5539 		pmc_debugflags_parse(pmc_debugstr,
5540 		    pmc_debugstr+strlen(pmc_debugstr));
5541 #endif
5542 
5543 	PMCDBG1(MOD,INI,0, "PMC Initialize (version %x)", PMC_VERSION);
5544 
5545 	/* check kernel version */
5546 	if (pmc_kernel_version != PMC_VERSION) {
5547 		if (pmc_kernel_version == 0)
5548 			printf("hwpmc: this kernel has not been compiled with "
5549 			    "'options HWPMC_HOOKS'.\n");
5550 		else
5551 			printf("hwpmc: kernel version (0x%x) does not match "
5552 			    "module version (0x%x).\n", pmc_kernel_version,
5553 			    PMC_VERSION);
5554 		return EPROGMISMATCH;
5555 	}
5556 
5557 	/*
5558 	 * check sysctl parameters
5559 	 */
5560 
5561 	if (pmc_hashsize <= 0) {
5562 		(void) printf("hwpmc: tunable \"hashsize\"=%d must be "
5563 		    "greater than zero.\n", pmc_hashsize);
5564 		pmc_hashsize = PMC_HASH_SIZE;
5565 	}
5566 
5567 	if (pmc_nsamples <= 0 || pmc_nsamples > 65535) {
5568 		(void) printf("hwpmc: tunable \"nsamples\"=%d out of "
5569 		    "range.\n", pmc_nsamples);
5570 		pmc_nsamples = PMC_NSAMPLES;
5571 	}
5572 	pmc_sample_mask = pmc_nsamples-1;
5573 
5574 	if (pmc_callchaindepth <= 0 ||
5575 	    pmc_callchaindepth > PMC_CALLCHAIN_DEPTH_MAX) {
5576 		(void) printf("hwpmc: tunable \"callchaindepth\"=%d out of "
5577 		    "range - using %d.\n", pmc_callchaindepth,
5578 		    PMC_CALLCHAIN_DEPTH_MAX);
5579 		pmc_callchaindepth = PMC_CALLCHAIN_DEPTH_MAX;
5580 	}
5581 
5582 	md = pmc_md_initialize();
5583 	if (md == NULL) {
5584 		/* Default to generic CPU. */
5585 		md = pmc_generic_cpu_initialize();
5586 		if (md == NULL)
5587 			return (ENOSYS);
5588         }
5589 
5590 	KASSERT(md->pmd_nclass >= 1 && md->pmd_npmc >= 1,
5591 	    ("[pmc,%d] no classes or pmcs", __LINE__));
5592 
5593 	/* Compute the map from row-indices to classdep pointers. */
5594 	pmc_rowindex_to_classdep = malloc(sizeof(struct pmc_classdep *) *
5595 	    md->pmd_npmc, M_PMC, M_WAITOK|M_ZERO);
5596 
5597 	for (n = 0; n < md->pmd_npmc; n++)
5598 		pmc_rowindex_to_classdep[n] = NULL;
5599 	for (ri = c = 0; c < md->pmd_nclass; c++) {
5600 		pcd = &md->pmd_classdep[c];
5601 		for (n = 0; n < pcd->pcd_num; n++, ri++)
5602 			pmc_rowindex_to_classdep[ri] = pcd;
5603 	}
5604 
5605 	KASSERT(ri == md->pmd_npmc,
5606 	    ("[pmc,%d] npmc miscomputed: ri=%d, md->npmc=%d", __LINE__,
5607 	    ri, md->pmd_npmc));
5608 
5609 	maxcpu = pmc_cpu_max();
5610 
5611 	/* allocate space for the per-cpu array */
5612 	pmc_pcpu = malloc(maxcpu * sizeof(struct pmc_cpu *), M_PMC,
5613 	    M_WAITOK|M_ZERO);
5614 
5615 	/* per-cpu 'saved values' for managing process-mode PMCs */
5616 	pmc_pcpu_saved = malloc(sizeof(pmc_value_t) * maxcpu * md->pmd_npmc,
5617 	    M_PMC, M_WAITOK);
5618 
5619 	/* Perform CPU-dependent initialization. */
5620 	pmc_save_cpu_binding(&pb);
5621 	error = 0;
5622 	for (cpu = 0; error == 0 && cpu < maxcpu; cpu++) {
5623 		if (!pmc_cpu_is_active(cpu))
5624 			continue;
5625 		pmc_select_cpu(cpu);
5626 		pmc_pcpu[cpu] = malloc(sizeof(struct pmc_cpu) +
5627 		    md->pmd_npmc * sizeof(struct pmc_hw *), M_PMC,
5628 		    M_WAITOK|M_ZERO);
5629 		if (md->pmd_pcpu_init)
5630 			error = md->pmd_pcpu_init(md, cpu);
5631 		for (n = 0; error == 0 && n < md->pmd_nclass; n++)
5632 			error = md->pmd_classdep[n].pcd_pcpu_init(md, cpu);
5633 	}
5634 	pmc_restore_cpu_binding(&pb);
5635 
5636 	if (error)
5637 		return (error);
5638 
5639 	/* allocate space for the sample array */
5640 	for (cpu = 0; cpu < maxcpu; cpu++) {
5641 		if (!pmc_cpu_is_active(cpu))
5642 			continue;
5643 		pc = pcpu_find(cpu);
5644 		domain = pc->pc_domain;
5645 		sb = malloc_domainset(sizeof(struct pmc_samplebuffer) +
5646 		    pmc_nsamples * sizeof(struct pmc_sample), M_PMC,
5647 		    DOMAINSET_PREF(domain), M_WAITOK | M_ZERO);
5648 
5649 		KASSERT(pmc_pcpu[cpu] != NULL,
5650 		    ("[pmc,%d] cpu=%d Null per-cpu data", __LINE__, cpu));
5651 
5652 		sb->ps_callchains = malloc_domainset(pmc_callchaindepth *
5653 		    pmc_nsamples * sizeof(uintptr_t), M_PMC,
5654 		    DOMAINSET_PREF(domain), M_WAITOK | M_ZERO);
5655 
5656 		for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++)
5657 			ps->ps_pc = sb->ps_callchains +
5658 			    (n * pmc_callchaindepth);
5659 
5660 		pmc_pcpu[cpu]->pc_sb[PMC_HR] = sb;
5661 
5662 		sb = malloc_domainset(sizeof(struct pmc_samplebuffer) +
5663 		    pmc_nsamples * sizeof(struct pmc_sample), M_PMC,
5664 		    DOMAINSET_PREF(domain), M_WAITOK | M_ZERO);
5665 
5666 		sb->ps_callchains = malloc_domainset(pmc_callchaindepth *
5667 		    pmc_nsamples * sizeof(uintptr_t), M_PMC,
5668 		    DOMAINSET_PREF(domain), M_WAITOK | M_ZERO);
5669 		for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++)
5670 			ps->ps_pc = sb->ps_callchains +
5671 			    (n * pmc_callchaindepth);
5672 
5673 		pmc_pcpu[cpu]->pc_sb[PMC_SR] = sb;
5674 
5675 		sb = malloc_domainset(sizeof(struct pmc_samplebuffer) +
5676 		    pmc_nsamples * sizeof(struct pmc_sample), M_PMC,
5677 		    DOMAINSET_PREF(domain), M_WAITOK | M_ZERO);
5678 		sb->ps_callchains = malloc_domainset(pmc_callchaindepth *
5679 		    pmc_nsamples * sizeof(uintptr_t), M_PMC,
5680 		    DOMAINSET_PREF(domain), M_WAITOK | M_ZERO);
5681 		for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++)
5682 			ps->ps_pc = sb->ps_callchains + n * pmc_callchaindepth;
5683 
5684 		pmc_pcpu[cpu]->pc_sb[PMC_UR] = sb;
5685 	}
5686 
5687 	/* allocate space for the row disposition array */
5688 	pmc_pmcdisp = malloc(sizeof(enum pmc_mode) * md->pmd_npmc,
5689 	    M_PMC, M_WAITOK|M_ZERO);
5690 
5691 	/* mark all PMCs as available */
5692 	for (n = 0; n < (int) md->pmd_npmc; n++)
5693 		PMC_MARK_ROW_FREE(n);
5694 
5695 	/* allocate thread hash tables */
5696 	pmc_ownerhash = hashinit(pmc_hashsize, M_PMC,
5697 	    &pmc_ownerhashmask);
5698 
5699 	pmc_processhash = hashinit(pmc_hashsize, M_PMC,
5700 	    &pmc_processhashmask);
5701 	mtx_init(&pmc_processhash_mtx, "pmc-process-hash", "pmc-leaf",
5702 	    MTX_SPIN);
5703 
5704 	CK_LIST_INIT(&pmc_ss_owners);
5705 	pmc_ss_count = 0;
5706 
5707 	/* allocate a pool of spin mutexes */
5708 	pmc_mtxpool = mtx_pool_create("pmc-leaf", pmc_mtxpool_size,
5709 	    MTX_SPIN);
5710 
5711 	PMCDBG4(MOD,INI,1, "pmc_ownerhash=%p, mask=0x%lx "
5712 	    "targethash=%p mask=0x%lx", pmc_ownerhash, pmc_ownerhashmask,
5713 	    pmc_processhash, pmc_processhashmask);
5714 
5715 	/* Initialize a spin mutex for the thread free list. */
5716 	mtx_init(&pmc_threadfreelist_mtx, "pmc-threadfreelist", "pmc-leaf",
5717 	    MTX_SPIN);
5718 
5719 	/* Initialize the task to prune the thread free list. */
5720 	TASK_INIT(&free_task, 0, pmc_thread_descriptor_pool_free_task, NULL);
5721 
5722 	/* register process {exit,fork,exec} handlers */
5723 	pmc_exit_tag = EVENTHANDLER_REGISTER(process_exit,
5724 	    pmc_process_exit, NULL, EVENTHANDLER_PRI_ANY);
5725 	pmc_fork_tag = EVENTHANDLER_REGISTER(process_fork,
5726 	    pmc_process_fork, NULL, EVENTHANDLER_PRI_ANY);
5727 
5728 	/* register kld event handlers */
5729 	pmc_kld_load_tag = EVENTHANDLER_REGISTER(kld_load, pmc_kld_load,
5730 	    NULL, EVENTHANDLER_PRI_ANY);
5731 	pmc_kld_unload_tag = EVENTHANDLER_REGISTER(kld_unload, pmc_kld_unload,
5732 	    NULL, EVENTHANDLER_PRI_ANY);
5733 
5734 	/* initialize logging */
5735 	pmclog_initialize();
5736 
5737 	/* set hook functions */
5738 	pmc_intr = md->pmd_intr;
5739 	wmb();
5740 	pmc_hook = pmc_hook_handler;
5741 
5742 	if (error == 0) {
5743 		printf(PMC_MODULE_NAME ":");
5744 		for (n = 0; n < (int) md->pmd_nclass; n++) {
5745 			pcd = &md->pmd_classdep[n];
5746 			printf(" %s/%d/%d/0x%b",
5747 			    pmc_name_of_pmcclass(pcd->pcd_class),
5748 			    pcd->pcd_num,
5749 			    pcd->pcd_width,
5750 			    pcd->pcd_caps,
5751 			    "\20"
5752 			    "\1INT\2USR\3SYS\4EDG\5THR"
5753 			    "\6REA\7WRI\10INV\11QUA\12PRC"
5754 			    "\13TAG\14CSC");
5755 		}
5756 		printf("\n");
5757 	}
5758 
5759 	return (error);
5760 }
5761 
5762 /* prepare to be unloaded */
5763 static void
5764 pmc_cleanup(void)
5765 {
5766 	int c, cpu;
5767 	unsigned int maxcpu;
5768 	struct pmc_ownerhash *ph;
5769 	struct pmc_owner *po, *tmp;
5770 	struct pmc_binding pb;
5771 #ifdef	HWPMC_DEBUG
5772 	struct pmc_processhash *prh;
5773 #endif
5774 
5775 	PMCDBG0(MOD,INI,0, "cleanup");
5776 
5777 	/* switch off sampling */
5778 	CPU_FOREACH(cpu)
5779 		DPCPU_ID_SET(cpu, pmc_sampled, 0);
5780 	pmc_intr = NULL;
5781 
5782 	sx_xlock(&pmc_sx);
5783 	if (pmc_hook == NULL) {	/* being unloaded already */
5784 		sx_xunlock(&pmc_sx);
5785 		return;
5786 	}
5787 
5788 	pmc_hook = NULL; /* prevent new threads from entering module */
5789 
5790 	/* deregister event handlers */
5791 	EVENTHANDLER_DEREGISTER(process_fork, pmc_fork_tag);
5792 	EVENTHANDLER_DEREGISTER(process_exit, pmc_exit_tag);
5793 	EVENTHANDLER_DEREGISTER(kld_load, pmc_kld_load_tag);
5794 	EVENTHANDLER_DEREGISTER(kld_unload, pmc_kld_unload_tag);
5795 
5796 	/* send SIGBUS to all owner threads, free up allocations */
5797 	if (pmc_ownerhash)
5798 		for (ph = pmc_ownerhash;
5799 		     ph <= &pmc_ownerhash[pmc_ownerhashmask];
5800 		     ph++) {
5801 			LIST_FOREACH_SAFE(po, ph, po_next, tmp) {
5802 				pmc_remove_owner(po);
5803 
5804 				/* send SIGBUS to owner processes */
5805 				PMCDBG3(MOD,INI,2, "cleanup signal proc=%p "
5806 				    "(%d, %s)", po->po_owner,
5807 				    po->po_owner->p_pid,
5808 				    po->po_owner->p_comm);
5809 
5810 				PROC_LOCK(po->po_owner);
5811 				kern_psignal(po->po_owner, SIGBUS);
5812 				PROC_UNLOCK(po->po_owner);
5813 
5814 				pmc_destroy_owner_descriptor(po);
5815 			}
5816 		}
5817 
5818 	/* reclaim allocated data structures */
5819 	taskqueue_drain(taskqueue_fast, &free_task);
5820 	mtx_destroy(&pmc_threadfreelist_mtx);
5821 	pmc_thread_descriptor_pool_drain();
5822 
5823 	if (pmc_mtxpool)
5824 		mtx_pool_destroy(&pmc_mtxpool);
5825 
5826 	mtx_destroy(&pmc_processhash_mtx);
5827 	if (pmc_processhash) {
5828 #ifdef	HWPMC_DEBUG
5829 		struct pmc_process *pp;
5830 
5831 		PMCDBG0(MOD,INI,3, "destroy process hash");
5832 		for (prh = pmc_processhash;
5833 		     prh <= &pmc_processhash[pmc_processhashmask];
5834 		     prh++)
5835 			LIST_FOREACH(pp, prh, pp_next)
5836 			    PMCDBG1(MOD,INI,3, "pid=%d", pp->pp_proc->p_pid);
5837 #endif
5838 
5839 		hashdestroy(pmc_processhash, M_PMC, pmc_processhashmask);
5840 		pmc_processhash = NULL;
5841 	}
5842 
5843 	if (pmc_ownerhash) {
5844 		PMCDBG0(MOD,INI,3, "destroy owner hash");
5845 		hashdestroy(pmc_ownerhash, M_PMC, pmc_ownerhashmask);
5846 		pmc_ownerhash = NULL;
5847 	}
5848 
5849 	KASSERT(CK_LIST_EMPTY(&pmc_ss_owners),
5850 	    ("[pmc,%d] Global SS owner list not empty", __LINE__));
5851 	KASSERT(pmc_ss_count == 0,
5852 	    ("[pmc,%d] Global SS count not empty", __LINE__));
5853 
5854  	/* do processor and pmc-class dependent cleanup */
5855 	maxcpu = pmc_cpu_max();
5856 
5857 	PMCDBG0(MOD,INI,3, "md cleanup");
5858 	if (md) {
5859 		pmc_save_cpu_binding(&pb);
5860 		for (cpu = 0; cpu < maxcpu; cpu++) {
5861 			PMCDBG2(MOD,INI,1,"pmc-cleanup cpu=%d pcs=%p",
5862 			    cpu, pmc_pcpu[cpu]);
5863 			if (!pmc_cpu_is_active(cpu) || pmc_pcpu[cpu] == NULL)
5864 				continue;
5865 			pmc_select_cpu(cpu);
5866 			for (c = 0; c < md->pmd_nclass; c++)
5867 				md->pmd_classdep[c].pcd_pcpu_fini(md, cpu);
5868 			if (md->pmd_pcpu_fini)
5869 				md->pmd_pcpu_fini(md, cpu);
5870 		}
5871 
5872 		if (md->pmd_cputype == PMC_CPU_GENERIC)
5873 			pmc_generic_cpu_finalize(md);
5874 		else
5875 			pmc_md_finalize(md);
5876 
5877 		pmc_mdep_free(md);
5878 		md = NULL;
5879 		pmc_restore_cpu_binding(&pb);
5880 	}
5881 
5882 	/* Free per-cpu descriptors. */
5883 	for (cpu = 0; cpu < maxcpu; cpu++) {
5884 		if (!pmc_cpu_is_active(cpu))
5885 			continue;
5886 		KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_HR] != NULL,
5887 		    ("[pmc,%d] Null hw cpu sample buffer cpu=%d", __LINE__,
5888 			cpu));
5889 		KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_SR] != NULL,
5890 		    ("[pmc,%d] Null sw cpu sample buffer cpu=%d", __LINE__,
5891 			cpu));
5892 		KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_UR] != NULL,
5893 		    ("[pmc,%d] Null userret cpu sample buffer cpu=%d", __LINE__,
5894 			cpu));
5895 		free(pmc_pcpu[cpu]->pc_sb[PMC_HR]->ps_callchains, M_PMC);
5896 		free(pmc_pcpu[cpu]->pc_sb[PMC_HR], M_PMC);
5897 		free(pmc_pcpu[cpu]->pc_sb[PMC_SR]->ps_callchains, M_PMC);
5898 		free(pmc_pcpu[cpu]->pc_sb[PMC_SR], M_PMC);
5899 		free(pmc_pcpu[cpu]->pc_sb[PMC_UR]->ps_callchains, M_PMC);
5900 		free(pmc_pcpu[cpu]->pc_sb[PMC_UR], M_PMC);
5901 		free(pmc_pcpu[cpu], M_PMC);
5902 	}
5903 
5904 	free(pmc_pcpu, M_PMC);
5905 	pmc_pcpu = NULL;
5906 
5907 	free(pmc_pcpu_saved, M_PMC);
5908 	pmc_pcpu_saved = NULL;
5909 
5910 	if (pmc_pmcdisp) {
5911 		free(pmc_pmcdisp, M_PMC);
5912 		pmc_pmcdisp = NULL;
5913 	}
5914 
5915 	if (pmc_rowindex_to_classdep) {
5916 		free(pmc_rowindex_to_classdep, M_PMC);
5917 		pmc_rowindex_to_classdep = NULL;
5918 	}
5919 
5920 	pmclog_shutdown();
5921 	counter_u64_free(pmc_stats.pm_intr_ignored);
5922 	counter_u64_free(pmc_stats.pm_intr_processed);
5923 	counter_u64_free(pmc_stats.pm_intr_bufferfull);
5924 	counter_u64_free(pmc_stats.pm_syscalls);
5925 	counter_u64_free(pmc_stats.pm_syscall_errors);
5926 	counter_u64_free(pmc_stats.pm_buffer_requests);
5927 	counter_u64_free(pmc_stats.pm_buffer_requests_failed);
5928 	counter_u64_free(pmc_stats.pm_log_sweeps);
5929 	counter_u64_free(pmc_stats.pm_merges);
5930 	counter_u64_free(pmc_stats.pm_overwrites);
5931 	sx_xunlock(&pmc_sx); 	/* we are done */
5932 }
5933 
5934 /*
5935  * The function called at load/unload.
5936  */
5937 
5938 static int
5939 load (struct module *module __unused, int cmd, void *arg __unused)
5940 {
5941 	int error;
5942 
5943 	error = 0;
5944 
5945 	switch (cmd) {
5946 	case MOD_LOAD :
5947 		/* initialize the subsystem */
5948 		error = pmc_initialize();
5949 		if (error != 0)
5950 			break;
5951 		PMCDBG2(MOD,INI,1, "syscall=%d maxcpu=%d",
5952 		    pmc_syscall_num, pmc_cpu_max());
5953 		break;
5954 
5955 
5956 	case MOD_UNLOAD :
5957 	case MOD_SHUTDOWN:
5958 		pmc_cleanup();
5959 		PMCDBG0(MOD,INI,1, "unloaded");
5960 		break;
5961 
5962 	default :
5963 		error = EINVAL;	/* XXX should panic(9) */
5964 		break;
5965 	}
5966 
5967 	return error;
5968 }
5969