xref: /freebsd/sys/dev/hwpmc/hwpmc_mod.c (revision 0b57cec536236d46e3dba9bd041533462f33dbb7)
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/domainset.h>
40 #include <sys/eventhandler.h>
41 #include <sys/gtaskqueue.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/systm.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 grouptask free_gtask;
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, 0, "HWPMC stats");
279 
280 
281 /* Stats. */
282 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, intr_ignored, CTLFLAG_RW,
283 				   &pmc_stats.pm_intr_ignored, "# of interrupts ignored");
284 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, intr_processed, CTLFLAG_RW,
285 				   &pmc_stats.pm_intr_processed, "# of interrupts processed");
286 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, intr_bufferfull, CTLFLAG_RW,
287 				   &pmc_stats.pm_intr_bufferfull, "# of interrupts where buffer was full");
288 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, syscalls, CTLFLAG_RW,
289 				   &pmc_stats.pm_syscalls, "# of syscalls");
290 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, syscall_errors, CTLFLAG_RW,
291 				   &pmc_stats.pm_syscall_errors, "# of syscall_errors");
292 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, buffer_requests, CTLFLAG_RW,
293 				   &pmc_stats.pm_buffer_requests, "# of buffer requests");
294 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, buffer_requests_failed, CTLFLAG_RW,
295 				   &pmc_stats.pm_buffer_requests_failed, "# of buffer requests which failed");
296 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, log_sweeps, CTLFLAG_RW,
297 				   &pmc_stats.pm_log_sweeps, "# of ?");
298 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, merges, CTLFLAG_RW,
299 				   &pmc_stats.pm_merges, "# of times kernel stack was found for user trace");
300 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, overwrites, CTLFLAG_RW,
301 				   &pmc_stats.pm_overwrites, "# of times a sample was overwritten before being logged");
302 
303 static int pmc_callchaindepth = PMC_CALLCHAIN_DEPTH;
304 SYSCTL_INT(_kern_hwpmc, OID_AUTO, callchaindepth, CTLFLAG_RDTUN,
305     &pmc_callchaindepth, 0, "depth of call chain records");
306 
307 char pmc_cpuid[64];
308 SYSCTL_STRING(_kern_hwpmc, OID_AUTO, cpuid, CTLFLAG_RD,
309 	pmc_cpuid, 0, "cpu version string");
310 #ifdef	HWPMC_DEBUG
311 struct pmc_debugflags pmc_debugflags = PMC_DEBUG_DEFAULT_FLAGS;
312 char	pmc_debugstr[PMC_DEBUG_STRSIZE];
313 TUNABLE_STR(PMC_SYSCTL_NAME_PREFIX "debugflags", pmc_debugstr,
314     sizeof(pmc_debugstr));
315 SYSCTL_PROC(_kern_hwpmc, OID_AUTO, debugflags,
316     CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NOFETCH,
317     0, 0, pmc_debugflags_sysctl_handler, "A", "debug flags");
318 #endif
319 
320 
321 /*
322  * kern.hwpmc.hashrows -- determines the number of rows in the
323  * of the hash table used to look up threads
324  */
325 
326 static int pmc_hashsize = PMC_HASH_SIZE;
327 SYSCTL_INT(_kern_hwpmc, OID_AUTO, hashsize, CTLFLAG_RDTUN,
328     &pmc_hashsize, 0, "rows in hash tables");
329 
330 /*
331  * kern.hwpmc.nsamples --- number of PC samples/callchain stacks per CPU
332  */
333 
334 static int pmc_nsamples = PMC_NSAMPLES;
335 SYSCTL_INT(_kern_hwpmc, OID_AUTO, nsamples, CTLFLAG_RDTUN,
336     &pmc_nsamples, 0, "number of PC samples per CPU");
337 
338 static uint64_t pmc_sample_mask = PMC_NSAMPLES-1;
339 
340 /*
341  * kern.hwpmc.mtxpoolsize -- number of mutexes in the mutex pool.
342  */
343 
344 static int pmc_mtxpool_size = PMC_MTXPOOL_SIZE;
345 SYSCTL_INT(_kern_hwpmc, OID_AUTO, mtxpoolsize, CTLFLAG_RDTUN,
346     &pmc_mtxpool_size, 0, "size of spin mutex pool");
347 
348 
349 /*
350  * kern.hwpmc.threadfreelist_entries -- number of free entries
351  */
352 
353 SYSCTL_INT(_kern_hwpmc, OID_AUTO, threadfreelist_entries, CTLFLAG_RD,
354     &pmc_threadfreelist_entries, 0, "number of avalable thread entries");
355 
356 
357 /*
358  * kern.hwpmc.threadfreelist_max -- maximum number of free entries
359  */
360 
361 static int pmc_threadfreelist_max = PMC_THREADLIST_MAX;
362 SYSCTL_INT(_kern_hwpmc, OID_AUTO, threadfreelist_max, CTLFLAG_RW,
363     &pmc_threadfreelist_max, 0,
364     "maximum number of available thread entries before freeing some");
365 
366 
367 /*
368  * security.bsd.unprivileged_syspmcs -- allow non-root processes to
369  * allocate system-wide PMCs.
370  *
371  * Allowing unprivileged processes to allocate system PMCs is convenient
372  * if system-wide measurements need to be taken concurrently with other
373  * per-process measurements.  This feature is turned off by default.
374  */
375 
376 static int pmc_unprivileged_syspmcs = 0;
377 SYSCTL_INT(_security_bsd, OID_AUTO, unprivileged_syspmcs, CTLFLAG_RWTUN,
378     &pmc_unprivileged_syspmcs, 0,
379     "allow unprivileged process to allocate system PMCs");
380 
381 /*
382  * Hash function.  Discard the lower 2 bits of the pointer since
383  * these are always zero for our uses.  The hash multiplier is
384  * round((2^LONG_BIT) * ((sqrt(5)-1)/2)).
385  */
386 
387 #if	LONG_BIT == 64
388 #define	_PMC_HM		11400714819323198486u
389 #elif	LONG_BIT == 32
390 #define	_PMC_HM		2654435769u
391 #else
392 #error 	Must know the size of 'long' to compile
393 #endif
394 
395 #define	PMC_HASH_PTR(P,M)	((((unsigned long) (P) >> 2) * _PMC_HM) & (M))
396 
397 /*
398  * Syscall structures
399  */
400 
401 /* The `sysent' for the new syscall */
402 static struct sysent pmc_sysent = {
403 	.sy_narg =	2,
404 	.sy_call =	pmc_syscall_handler,
405 };
406 
407 static struct syscall_module_data pmc_syscall_mod = {
408 	.chainevh =	load,
409 	.chainarg =	NULL,
410 	.offset =	&pmc_syscall_num,
411 	.new_sysent =	&pmc_sysent,
412 	.old_sysent =	{ .sy_narg = 0, .sy_call = NULL },
413 	.flags =	SY_THR_STATIC_KLD,
414 };
415 
416 static moduledata_t pmc_mod = {
417 	.name =		PMC_MODULE_NAME,
418 	.evhand =	syscall_module_handler,
419 	.priv =		&pmc_syscall_mod,
420 };
421 
422 #ifdef EARLY_AP_STARTUP
423 DECLARE_MODULE(pmc, pmc_mod, SI_SUB_SYSCALLS, SI_ORDER_ANY);
424 #else
425 DECLARE_MODULE(pmc, pmc_mod, SI_SUB_SMP, SI_ORDER_ANY);
426 #endif
427 MODULE_VERSION(pmc, PMC_VERSION);
428 
429 #ifdef	HWPMC_DEBUG
430 enum pmc_dbgparse_state {
431 	PMCDS_WS,		/* in whitespace */
432 	PMCDS_MAJOR,		/* seen a major keyword */
433 	PMCDS_MINOR
434 };
435 
436 static int
437 pmc_debugflags_parse(char *newstr, char *fence)
438 {
439 	char c, *p, *q;
440 	struct pmc_debugflags *tmpflags;
441 	int error, found, *newbits, tmp;
442 	size_t kwlen;
443 
444 	tmpflags = malloc(sizeof(*tmpflags), M_PMC, M_WAITOK|M_ZERO);
445 
446 	p = newstr;
447 	error = 0;
448 
449 	for (; p < fence && (c = *p); p++) {
450 
451 		/* skip white space */
452 		if (c == ' ' || c == '\t')
453 			continue;
454 
455 		/* look for a keyword followed by "=" */
456 		for (q = p; p < fence && (c = *p) && c != '='; p++)
457 			;
458 		if (c != '=') {
459 			error = EINVAL;
460 			goto done;
461 		}
462 
463 		kwlen = p - q;
464 		newbits = NULL;
465 
466 		/* lookup flag group name */
467 #define	DBG_SET_FLAG_MAJ(S,F)						\
468 		if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0)	\
469 			newbits = &tmpflags->pdb_ ## F;
470 
471 		DBG_SET_FLAG_MAJ("cpu",		CPU);
472 		DBG_SET_FLAG_MAJ("csw",		CSW);
473 		DBG_SET_FLAG_MAJ("logging",	LOG);
474 		DBG_SET_FLAG_MAJ("module",	MOD);
475 		DBG_SET_FLAG_MAJ("md", 		MDP);
476 		DBG_SET_FLAG_MAJ("owner",	OWN);
477 		DBG_SET_FLAG_MAJ("pmc",		PMC);
478 		DBG_SET_FLAG_MAJ("process",	PRC);
479 		DBG_SET_FLAG_MAJ("sampling", 	SAM);
480 
481 		if (newbits == NULL) {
482 			error = EINVAL;
483 			goto done;
484 		}
485 
486 		p++;		/* skip the '=' */
487 
488 		/* Now parse the individual flags */
489 		tmp = 0;
490 	newflag:
491 		for (q = p; p < fence && (c = *p); p++)
492 			if (c == ' ' || c == '\t' || c == ',')
493 				break;
494 
495 		/* p == fence or c == ws or c == "," or c == 0 */
496 
497 		if ((kwlen = p - q) == 0) {
498 			*newbits = tmp;
499 			continue;
500 		}
501 
502 		found = 0;
503 #define	DBG_SET_FLAG_MIN(S,F)						\
504 		if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0)	\
505 			tmp |= found = (1 << PMC_DEBUG_MIN_ ## F)
506 
507 		/* a '*' denotes all possible flags in the group */
508 		if (kwlen == 1 && *q == '*')
509 			tmp = found = ~0;
510 		/* look for individual flag names */
511 		DBG_SET_FLAG_MIN("allocaterow", ALR);
512 		DBG_SET_FLAG_MIN("allocate",	ALL);
513 		DBG_SET_FLAG_MIN("attach",	ATT);
514 		DBG_SET_FLAG_MIN("bind",	BND);
515 		DBG_SET_FLAG_MIN("config",	CFG);
516 		DBG_SET_FLAG_MIN("exec",	EXC);
517 		DBG_SET_FLAG_MIN("exit",	EXT);
518 		DBG_SET_FLAG_MIN("find",	FND);
519 		DBG_SET_FLAG_MIN("flush",	FLS);
520 		DBG_SET_FLAG_MIN("fork",	FRK);
521 		DBG_SET_FLAG_MIN("getbuf",	GTB);
522 		DBG_SET_FLAG_MIN("hook",	PMH);
523 		DBG_SET_FLAG_MIN("init",	INI);
524 		DBG_SET_FLAG_MIN("intr",	INT);
525 		DBG_SET_FLAG_MIN("linktarget",	TLK);
526 		DBG_SET_FLAG_MIN("mayberemove", OMR);
527 		DBG_SET_FLAG_MIN("ops",		OPS);
528 		DBG_SET_FLAG_MIN("read",	REA);
529 		DBG_SET_FLAG_MIN("register",	REG);
530 		DBG_SET_FLAG_MIN("release",	REL);
531 		DBG_SET_FLAG_MIN("remove",	ORM);
532 		DBG_SET_FLAG_MIN("sample",	SAM);
533 		DBG_SET_FLAG_MIN("scheduleio",	SIO);
534 		DBG_SET_FLAG_MIN("select",	SEL);
535 		DBG_SET_FLAG_MIN("signal",	SIG);
536 		DBG_SET_FLAG_MIN("swi",		SWI);
537 		DBG_SET_FLAG_MIN("swo",		SWO);
538 		DBG_SET_FLAG_MIN("start",	STA);
539 		DBG_SET_FLAG_MIN("stop",	STO);
540 		DBG_SET_FLAG_MIN("syscall",	PMS);
541 		DBG_SET_FLAG_MIN("unlinktarget", TUL);
542 		DBG_SET_FLAG_MIN("write",	WRI);
543 		if (found == 0) {
544 			/* unrecognized flag name */
545 			error = EINVAL;
546 			goto done;
547 		}
548 
549 		if (c == 0 || c == ' ' || c == '\t') {	/* end of flag group */
550 			*newbits = tmp;
551 			continue;
552 		}
553 
554 		p++;
555 		goto newflag;
556 	}
557 
558 	/* save the new flag set */
559 	bcopy(tmpflags, &pmc_debugflags, sizeof(pmc_debugflags));
560 
561  done:
562 	free(tmpflags, M_PMC);
563 	return error;
564 }
565 
566 static int
567 pmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS)
568 {
569 	char *fence, *newstr;
570 	int error;
571 	unsigned int n;
572 
573 	(void) arg1; (void) arg2; /* unused parameters */
574 
575 	n = sizeof(pmc_debugstr);
576 	newstr = malloc(n, M_PMC, M_WAITOK|M_ZERO);
577 	(void) strlcpy(newstr, pmc_debugstr, n);
578 
579 	error = sysctl_handle_string(oidp, newstr, n, req);
580 
581 	/* if there is a new string, parse and copy it */
582 	if (error == 0 && req->newptr != NULL) {
583 		fence = newstr + (n < req->newlen ? n : req->newlen + 1);
584 		if ((error = pmc_debugflags_parse(newstr, fence)) == 0)
585 			(void) strlcpy(pmc_debugstr, newstr,
586 			    sizeof(pmc_debugstr));
587 	}
588 
589 	free(newstr, M_PMC);
590 
591 	return error;
592 }
593 #endif
594 
595 /*
596  * Map a row index to a classdep structure and return the adjusted row
597  * index for the PMC class index.
598  */
599 static struct pmc_classdep *
600 pmc_ri_to_classdep(struct pmc_mdep *md, int ri, int *adjri)
601 {
602 	struct pmc_classdep *pcd;
603 
604 	(void) md;
605 
606 	KASSERT(ri >= 0 && ri < md->pmd_npmc,
607 	    ("[pmc,%d] illegal row-index %d", __LINE__, ri));
608 
609 	pcd = pmc_rowindex_to_classdep[ri];
610 
611 	KASSERT(pcd != NULL,
612 	    ("[pmc,%d] ri %d null pcd", __LINE__, ri));
613 
614 	*adjri = ri - pcd->pcd_ri;
615 
616 	KASSERT(*adjri >= 0 && *adjri < pcd->pcd_num,
617 	    ("[pmc,%d] adjusted row-index %d", __LINE__, *adjri));
618 
619 	return (pcd);
620 }
621 
622 /*
623  * Concurrency Control
624  *
625  * The driver manages the following data structures:
626  *
627  *   - target process descriptors, one per target process
628  *   - owner process descriptors (and attached lists), one per owner process
629  *   - lookup hash tables for owner and target processes
630  *   - PMC descriptors (and attached lists)
631  *   - per-cpu hardware state
632  *   - the 'hook' variable through which the kernel calls into
633  *     this module
634  *   - the machine hardware state (managed by the MD layer)
635  *
636  * These data structures are accessed from:
637  *
638  * - thread context-switch code
639  * - interrupt handlers (possibly on multiple cpus)
640  * - kernel threads on multiple cpus running on behalf of user
641  *   processes doing system calls
642  * - this driver's private kernel threads
643  *
644  * = Locks and Locking strategy =
645  *
646  * The driver uses four locking strategies for its operation:
647  *
648  * - The global SX lock "pmc_sx" is used to protect internal
649  *   data structures.
650  *
651  *   Calls into the module by syscall() start with this lock being
652  *   held in exclusive mode.  Depending on the requested operation,
653  *   the lock may be downgraded to 'shared' mode to allow more
654  *   concurrent readers into the module.  Calls into the module from
655  *   other parts of the kernel acquire the lock in shared mode.
656  *
657  *   This SX lock is held in exclusive mode for any operations that
658  *   modify the linkages between the driver's internal data structures.
659  *
660  *   The 'pmc_hook' function pointer is also protected by this lock.
661  *   It is only examined with the sx lock held in exclusive mode.  The
662  *   kernel module is allowed to be unloaded only with the sx lock held
663  *   in exclusive mode.  In normal syscall handling, after acquiring the
664  *   pmc_sx lock we first check that 'pmc_hook' is non-null before
665  *   proceeding.  This prevents races between the thread unloading the module
666  *   and other threads seeking to use the module.
667  *
668  * - Lookups of target process structures and owner process structures
669  *   cannot use the global "pmc_sx" SX lock because these lookups need
670  *   to happen during context switches and in other critical sections
671  *   where sleeping is not allowed.  We protect these lookup tables
672  *   with their own private spin-mutexes, "pmc_processhash_mtx" and
673  *   "pmc_ownerhash_mtx".
674  *
675  * - Interrupt handlers work in a lock free manner.  At interrupt
676  *   time, handlers look at the PMC pointer (phw->phw_pmc) configured
677  *   when the PMC was started.  If this pointer is NULL, the interrupt
678  *   is ignored after updating driver statistics.  We ensure that this
679  *   pointer is set (using an atomic operation if necessary) before the
680  *   PMC hardware is started.  Conversely, this pointer is unset atomically
681  *   only after the PMC hardware is stopped.
682  *
683  *   We ensure that everything needed for the operation of an
684  *   interrupt handler is available without it needing to acquire any
685  *   locks.  We also ensure that a PMC's software state is destroyed only
686  *   after the PMC is taken off hardware (on all CPUs).
687  *
688  * - Context-switch handling with process-private PMCs needs more
689  *   care.
690  *
691  *   A given process may be the target of multiple PMCs.  For example,
692  *   PMCATTACH and PMCDETACH may be requested by a process on one CPU
693  *   while the target process is running on another.  A PMC could also
694  *   be getting released because its owner is exiting.  We tackle
695  *   these situations in the following manner:
696  *
697  *   - each target process structure 'pmc_process' has an array
698  *     of 'struct pmc *' pointers, one for each hardware PMC.
699  *
700  *   - At context switch IN time, each "target" PMC in RUNNING state
701  *     gets started on hardware and a pointer to each PMC is copied into
702  *     the per-cpu phw array.  The 'runcount' for the PMC is
703  *     incremented.
704  *
705  *   - At context switch OUT time, all process-virtual PMCs are stopped
706  *     on hardware.  The saved value is added to the PMCs value field
707  *     only if the PMC is in a non-deleted state (the PMCs state could
708  *     have changed during the current time slice).
709  *
710  *     Note that since in-between a switch IN on a processor and a switch
711  *     OUT, the PMC could have been released on another CPU.  Therefore
712  *     context switch OUT always looks at the hardware state to turn
713  *     OFF PMCs and will update a PMC's saved value only if reachable
714  *     from the target process record.
715  *
716  *   - OP PMCRELEASE could be called on a PMC at any time (the PMC could
717  *     be attached to many processes at the time of the call and could
718  *     be active on multiple CPUs).
719  *
720  *     We prevent further scheduling of the PMC by marking it as in
721  *     state 'DELETED'.  If the runcount of the PMC is non-zero then
722  *     this PMC is currently running on a CPU somewhere.  The thread
723  *     doing the PMCRELEASE operation waits by repeatedly doing a
724  *     pause() till the runcount comes to zero.
725  *
726  * The contents of a PMC descriptor (struct pmc) are protected using
727  * a spin-mutex.  In order to save space, we use a mutex pool.
728  *
729  * In terms of lock types used by witness(4), we use:
730  * - Type "pmc-sx", used by the global SX lock.
731  * - Type "pmc-sleep", for sleep mutexes used by logger threads.
732  * - Type "pmc-per-proc", for protecting PMC owner descriptors.
733  * - Type "pmc-leaf", used for all other spin mutexes.
734  */
735 
736 /*
737  * save the cpu binding of the current kthread
738  */
739 
740 static void
741 pmc_save_cpu_binding(struct pmc_binding *pb)
742 {
743 	PMCDBG0(CPU,BND,2, "save-cpu");
744 	thread_lock(curthread);
745 	pb->pb_bound = sched_is_bound(curthread);
746 	pb->pb_cpu   = curthread->td_oncpu;
747 	thread_unlock(curthread);
748 	PMCDBG1(CPU,BND,2, "save-cpu cpu=%d", pb->pb_cpu);
749 }
750 
751 /*
752  * restore the cpu binding of the current thread
753  */
754 
755 static void
756 pmc_restore_cpu_binding(struct pmc_binding *pb)
757 {
758 	PMCDBG2(CPU,BND,2, "restore-cpu curcpu=%d restore=%d",
759 	    curthread->td_oncpu, pb->pb_cpu);
760 	thread_lock(curthread);
761 	if (pb->pb_bound)
762 		sched_bind(curthread, pb->pb_cpu);
763 	else
764 		sched_unbind(curthread);
765 	thread_unlock(curthread);
766 	PMCDBG0(CPU,BND,2, "restore-cpu done");
767 }
768 
769 /*
770  * move execution over the specified cpu and bind it there.
771  */
772 
773 static void
774 pmc_select_cpu(int cpu)
775 {
776 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
777 	    ("[pmc,%d] bad cpu number %d", __LINE__, cpu));
778 
779 	/* Never move to an inactive CPU. */
780 	KASSERT(pmc_cpu_is_active(cpu), ("[pmc,%d] selecting inactive "
781 	    "CPU %d", __LINE__, cpu));
782 
783 	PMCDBG1(CPU,SEL,2, "select-cpu cpu=%d", cpu);
784 	thread_lock(curthread);
785 	sched_bind(curthread, cpu);
786 	thread_unlock(curthread);
787 
788 	KASSERT(curthread->td_oncpu == cpu,
789 	    ("[pmc,%d] CPU not bound [cpu=%d, curr=%d]", __LINE__,
790 		cpu, curthread->td_oncpu));
791 
792 	PMCDBG1(CPU,SEL,2, "select-cpu cpu=%d ok", cpu);
793 }
794 
795 /*
796  * Force a context switch.
797  *
798  * We do this by pause'ing for 1 tick -- invoking mi_switch() is not
799  * guaranteed to force a context switch.
800  */
801 
802 static void
803 pmc_force_context_switch(void)
804 {
805 
806 	pause("pmcctx", 1);
807 }
808 
809 uint64_t
810 pmc_rdtsc(void)
811 {
812 #if defined(__i386__) || defined(__amd64__)
813 	if (__predict_true(amd_feature & AMDID_RDTSCP))
814 		return rdtscp();
815 	else
816 		return rdtsc();
817 #else
818 	return get_cyclecount();
819 #endif
820 }
821 
822 /*
823  * Get the file name for an executable.  This is a simple wrapper
824  * around vn_fullpath(9).
825  */
826 
827 static void
828 pmc_getfilename(struct vnode *v, char **fullpath, char **freepath)
829 {
830 
831 	*fullpath = "unknown";
832 	*freepath = NULL;
833 	vn_fullpath(curthread, v, fullpath, freepath);
834 }
835 
836 /*
837  * remove an process owning PMCs
838  */
839 
840 void
841 pmc_remove_owner(struct pmc_owner *po)
842 {
843 	struct pmc *pm, *tmp;
844 
845 	sx_assert(&pmc_sx, SX_XLOCKED);
846 
847 	PMCDBG1(OWN,ORM,1, "remove-owner po=%p", po);
848 
849 	/* Remove descriptor from the owner hash table */
850 	LIST_REMOVE(po, po_next);
851 
852 	/* release all owned PMC descriptors */
853 	LIST_FOREACH_SAFE(pm, &po->po_pmcs, pm_next, tmp) {
854 		PMCDBG1(OWN,ORM,2, "pmc=%p", pm);
855 		KASSERT(pm->pm_owner == po,
856 		    ("[pmc,%d] owner %p != po %p", __LINE__, pm->pm_owner, po));
857 
858 		pmc_release_pmc_descriptor(pm);	/* will unlink from the list */
859 		pmc_destroy_pmc_descriptor(pm);
860 	}
861 
862 	KASSERT(po->po_sscount == 0,
863 	    ("[pmc,%d] SS count not zero", __LINE__));
864 	KASSERT(LIST_EMPTY(&po->po_pmcs),
865 	    ("[pmc,%d] PMC list not empty", __LINE__));
866 
867 	/* de-configure the log file if present */
868 	if (po->po_flags & PMC_PO_OWNS_LOGFILE)
869 		pmclog_deconfigure_log(po);
870 }
871 
872 /*
873  * remove an owner process record if all conditions are met.
874  */
875 
876 static void
877 pmc_maybe_remove_owner(struct pmc_owner *po)
878 {
879 
880 	PMCDBG1(OWN,OMR,1, "maybe-remove-owner po=%p", po);
881 
882 	/*
883 	 * Remove owner record if
884 	 * - this process does not own any PMCs
885 	 * - this process has not allocated a system-wide sampling buffer
886 	 */
887 
888 	if (LIST_EMPTY(&po->po_pmcs) &&
889 	    ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)) {
890 		pmc_remove_owner(po);
891 		pmc_destroy_owner_descriptor(po);
892 	}
893 }
894 
895 /*
896  * Add an association between a target process and a PMC.
897  */
898 
899 static void
900 pmc_link_target_process(struct pmc *pm, struct pmc_process *pp)
901 {
902 	int ri;
903 	struct pmc_target *pt;
904 #ifdef INVARIANTS
905 	struct pmc_thread *pt_td;
906 #endif
907 
908 	sx_assert(&pmc_sx, SX_XLOCKED);
909 
910 	KASSERT(pm != NULL && pp != NULL,
911 	    ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp));
912 	KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)),
913 	    ("[pmc,%d] Attaching a non-process-virtual pmc=%p to pid=%d",
914 		__LINE__, pm, pp->pp_proc->p_pid));
915 	KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= ((int) md->pmd_npmc - 1),
916 	    ("[pmc,%d] Illegal reference count %d for process record %p",
917 		__LINE__, pp->pp_refcnt, (void *) pp));
918 
919 	ri = PMC_TO_ROWINDEX(pm);
920 
921 	PMCDBG3(PRC,TLK,1, "link-target pmc=%p ri=%d pmc-process=%p",
922 	    pm, ri, pp);
923 
924 #ifdef	HWPMC_DEBUG
925 	LIST_FOREACH(pt, &pm->pm_targets, pt_next)
926 	    if (pt->pt_process == pp)
927 		    KASSERT(0, ("[pmc,%d] pp %p already in pmc %p targets",
928 				__LINE__, pp, pm));
929 #endif
930 
931 	pt = malloc(sizeof(struct pmc_target), M_PMC, M_WAITOK|M_ZERO);
932 	pt->pt_process = pp;
933 
934 	LIST_INSERT_HEAD(&pm->pm_targets, pt, pt_next);
935 
936 	atomic_store_rel_ptr((uintptr_t *)&pp->pp_pmcs[ri].pp_pmc,
937 	    (uintptr_t)pm);
938 
939 	if (pm->pm_owner->po_owner == pp->pp_proc)
940 		pm->pm_flags |= PMC_F_ATTACHED_TO_OWNER;
941 
942 	/*
943 	 * Initialize the per-process values at this row index.
944 	 */
945 	pp->pp_pmcs[ri].pp_pmcval = PMC_TO_MODE(pm) == PMC_MODE_TS ?
946 	    pm->pm_sc.pm_reloadcount : 0;
947 
948 	pp->pp_refcnt++;
949 
950 #ifdef INVARIANTS
951 	/* Confirm that the per-thread values at this row index are cleared. */
952 	if (PMC_TO_MODE(pm) == PMC_MODE_TS) {
953 		mtx_lock_spin(pp->pp_tdslock);
954 		LIST_FOREACH(pt_td, &pp->pp_tds, pt_next) {
955 			KASSERT(pt_td->pt_pmcs[ri].pt_pmcval == (pmc_value_t) 0,
956 			    ("[pmc,%d] pt_pmcval not cleared for pid=%d at "
957 			    "ri=%d", __LINE__, pp->pp_proc->p_pid, ri));
958 		}
959 		mtx_unlock_spin(pp->pp_tdslock);
960 	}
961 #endif
962 }
963 
964 /*
965  * Removes the association between a target process and a PMC.
966  */
967 
968 static void
969 pmc_unlink_target_process(struct pmc *pm, struct pmc_process *pp)
970 {
971 	int ri;
972 	struct proc *p;
973 	struct pmc_target *ptgt;
974 	struct pmc_thread *pt;
975 
976 	sx_assert(&pmc_sx, SX_XLOCKED);
977 
978 	KASSERT(pm != NULL && pp != NULL,
979 	    ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp));
980 
981 	KASSERT(pp->pp_refcnt >= 1 && pp->pp_refcnt <= (int) md->pmd_npmc,
982 	    ("[pmc,%d] Illegal ref count %d on process record %p",
983 		__LINE__, pp->pp_refcnt, (void *) pp));
984 
985 	ri = PMC_TO_ROWINDEX(pm);
986 
987 	PMCDBG3(PRC,TUL,1, "unlink-target pmc=%p ri=%d pmc-process=%p",
988 	    pm, ri, pp);
989 
990 	KASSERT(pp->pp_pmcs[ri].pp_pmc == pm,
991 	    ("[pmc,%d] PMC ri %d mismatch pmc %p pp->[ri] %p", __LINE__,
992 		ri, pm, pp->pp_pmcs[ri].pp_pmc));
993 
994 	pp->pp_pmcs[ri].pp_pmc = NULL;
995 	pp->pp_pmcs[ri].pp_pmcval = (pmc_value_t) 0;
996 
997 	/* Clear the per-thread values at this row index. */
998 	if (PMC_TO_MODE(pm) == PMC_MODE_TS) {
999 		mtx_lock_spin(pp->pp_tdslock);
1000 		LIST_FOREACH(pt, &pp->pp_tds, pt_next)
1001 			pt->pt_pmcs[ri].pt_pmcval = (pmc_value_t) 0;
1002 		mtx_unlock_spin(pp->pp_tdslock);
1003 	}
1004 
1005 	/* Remove owner-specific flags */
1006 	if (pm->pm_owner->po_owner == pp->pp_proc) {
1007 		pp->pp_flags &= ~PMC_PP_ENABLE_MSR_ACCESS;
1008 		pm->pm_flags &= ~PMC_F_ATTACHED_TO_OWNER;
1009 	}
1010 
1011 	pp->pp_refcnt--;
1012 
1013 	/* Remove the target process from the PMC structure */
1014 	LIST_FOREACH(ptgt, &pm->pm_targets, pt_next)
1015 		if (ptgt->pt_process == pp)
1016 			break;
1017 
1018 	KASSERT(ptgt != NULL, ("[pmc,%d] process %p (pp: %p) not found "
1019 		    "in pmc %p", __LINE__, pp->pp_proc, pp, pm));
1020 
1021 	LIST_REMOVE(ptgt, pt_next);
1022 	free(ptgt, M_PMC);
1023 
1024 	/* if the PMC now lacks targets, send the owner a SIGIO */
1025 	if (LIST_EMPTY(&pm->pm_targets)) {
1026 		p = pm->pm_owner->po_owner;
1027 		PROC_LOCK(p);
1028 		kern_psignal(p, SIGIO);
1029 		PROC_UNLOCK(p);
1030 
1031 		PMCDBG2(PRC,SIG,2, "signalling proc=%p signal=%d", p,
1032 		    SIGIO);
1033 	}
1034 }
1035 
1036 /*
1037  * Check if PMC 'pm' may be attached to target process 't'.
1038  */
1039 
1040 static int
1041 pmc_can_attach(struct pmc *pm, struct proc *t)
1042 {
1043 	struct proc *o;		/* pmc owner */
1044 	struct ucred *oc, *tc;	/* owner, target credentials */
1045 	int decline_attach, i;
1046 
1047 	/*
1048 	 * A PMC's owner can always attach that PMC to itself.
1049 	 */
1050 
1051 	if ((o = pm->pm_owner->po_owner) == t)
1052 		return 0;
1053 
1054 	PROC_LOCK(o);
1055 	oc = o->p_ucred;
1056 	crhold(oc);
1057 	PROC_UNLOCK(o);
1058 
1059 	PROC_LOCK(t);
1060 	tc = t->p_ucred;
1061 	crhold(tc);
1062 	PROC_UNLOCK(t);
1063 
1064 	/*
1065 	 * The effective uid of the PMC owner should match at least one
1066 	 * of the {effective,real,saved} uids of the target process.
1067 	 */
1068 
1069 	decline_attach = oc->cr_uid != tc->cr_uid &&
1070 	    oc->cr_uid != tc->cr_svuid &&
1071 	    oc->cr_uid != tc->cr_ruid;
1072 
1073 	/*
1074 	 * Every one of the target's group ids, must be in the owner's
1075 	 * group list.
1076 	 */
1077 	for (i = 0; !decline_attach && i < tc->cr_ngroups; i++)
1078 		decline_attach = !groupmember(tc->cr_groups[i], oc);
1079 
1080 	/* check the read and saved gids too */
1081 	if (decline_attach == 0)
1082 		decline_attach = !groupmember(tc->cr_rgid, oc) ||
1083 		    !groupmember(tc->cr_svgid, oc);
1084 
1085 	crfree(tc);
1086 	crfree(oc);
1087 
1088 	return !decline_attach;
1089 }
1090 
1091 /*
1092  * Attach a process to a PMC.
1093  */
1094 
1095 static int
1096 pmc_attach_one_process(struct proc *p, struct pmc *pm)
1097 {
1098 	int ri, error;
1099 	char *fullpath, *freepath;
1100 	struct pmc_process	*pp;
1101 
1102 	sx_assert(&pmc_sx, SX_XLOCKED);
1103 
1104 	PMCDBG5(PRC,ATT,2, "attach-one pm=%p ri=%d proc=%p (%d, %s)", pm,
1105 	    PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm);
1106 
1107 	/*
1108 	 * Locate the process descriptor corresponding to process 'p',
1109 	 * allocating space as needed.
1110 	 *
1111 	 * Verify that rowindex 'pm_rowindex' is free in the process
1112 	 * descriptor.
1113 	 *
1114 	 * If not, allocate space for a descriptor and link the
1115 	 * process descriptor and PMC.
1116 	 */
1117 	ri = PMC_TO_ROWINDEX(pm);
1118 
1119 	/* mark process as using HWPMCs */
1120 	PROC_LOCK(p);
1121 	p->p_flag |= P_HWPMC;
1122 	PROC_UNLOCK(p);
1123 
1124 	if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_ALLOCATE)) == NULL) {
1125 		error = ENOMEM;
1126 		goto fail;
1127 	}
1128 
1129 	if (pp->pp_pmcs[ri].pp_pmc == pm) {/* already present at slot [ri] */
1130 		error = EEXIST;
1131 		goto fail;
1132 	}
1133 
1134 	if (pp->pp_pmcs[ri].pp_pmc != NULL) {
1135 		error = EBUSY;
1136 		goto fail;
1137 	}
1138 
1139 	pmc_link_target_process(pm, pp);
1140 
1141 	if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) &&
1142 	    (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) == 0)
1143 		pm->pm_flags |= PMC_F_NEEDS_LOGFILE;
1144 
1145 	pm->pm_flags |= PMC_F_ATTACH_DONE; /* mark as attached */
1146 
1147 	/* issue an attach event to a configured log file */
1148 	if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE) {
1149 		if (p->p_flag & P_KPROC) {
1150 			fullpath = kernelname;
1151 			freepath = NULL;
1152 		} else {
1153 			pmc_getfilename(p->p_textvp, &fullpath, &freepath);
1154 			pmclog_process_pmcattach(pm, p->p_pid, fullpath);
1155 		}
1156 		free(freepath, M_TEMP);
1157 		if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1158 			pmc_log_process_mappings(pm->pm_owner, p);
1159 	}
1160 
1161 	return (0);
1162  fail:
1163 	PROC_LOCK(p);
1164 	p->p_flag &= ~P_HWPMC;
1165 	PROC_UNLOCK(p);
1166 	return (error);
1167 }
1168 
1169 /*
1170  * Attach a process and optionally its children
1171  */
1172 
1173 static int
1174 pmc_attach_process(struct proc *p, struct pmc *pm)
1175 {
1176 	int error;
1177 	struct proc *top;
1178 
1179 	sx_assert(&pmc_sx, SX_XLOCKED);
1180 
1181 	PMCDBG5(PRC,ATT,1, "attach pm=%p ri=%d proc=%p (%d, %s)", pm,
1182 	    PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm);
1183 
1184 
1185 	/*
1186 	 * If this PMC successfully allowed a GETMSR operation
1187 	 * in the past, disallow further ATTACHes.
1188 	 */
1189 
1190 	if ((pm->pm_flags & PMC_PP_ENABLE_MSR_ACCESS) != 0)
1191 		return EPERM;
1192 
1193 	if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0)
1194 		return pmc_attach_one_process(p, pm);
1195 
1196 	/*
1197 	 * Traverse all child processes, attaching them to
1198 	 * this PMC.
1199 	 */
1200 
1201 	sx_slock(&proctree_lock);
1202 
1203 	top = p;
1204 
1205 	for (;;) {
1206 		if ((error = pmc_attach_one_process(p, pm)) != 0)
1207 			break;
1208 		if (!LIST_EMPTY(&p->p_children))
1209 			p = LIST_FIRST(&p->p_children);
1210 		else for (;;) {
1211 			if (p == top)
1212 				goto done;
1213 			if (LIST_NEXT(p, p_sibling)) {
1214 				p = LIST_NEXT(p, p_sibling);
1215 				break;
1216 			}
1217 			p = p->p_pptr;
1218 		}
1219 	}
1220 
1221 	if (error)
1222 		(void) pmc_detach_process(top, pm);
1223 
1224  done:
1225 	sx_sunlock(&proctree_lock);
1226 	return error;
1227 }
1228 
1229 /*
1230  * Detach a process from a PMC.  If there are no other PMCs tracking
1231  * this process, remove the process structure from its hash table.  If
1232  * 'flags' contains PMC_FLAG_REMOVE, then free the process structure.
1233  */
1234 
1235 static int
1236 pmc_detach_one_process(struct proc *p, struct pmc *pm, int flags)
1237 {
1238 	int ri;
1239 	struct pmc_process *pp;
1240 
1241 	sx_assert(&pmc_sx, SX_XLOCKED);
1242 
1243 	KASSERT(pm != NULL,
1244 	    ("[pmc,%d] null pm pointer", __LINE__));
1245 
1246 	ri = PMC_TO_ROWINDEX(pm);
1247 
1248 	PMCDBG6(PRC,ATT,2, "detach-one pm=%p ri=%d proc=%p (%d, %s) flags=0x%x",
1249 	    pm, ri, p, p->p_pid, p->p_comm, flags);
1250 
1251 	if ((pp = pmc_find_process_descriptor(p, 0)) == NULL)
1252 		return ESRCH;
1253 
1254 	if (pp->pp_pmcs[ri].pp_pmc != pm)
1255 		return EINVAL;
1256 
1257 	pmc_unlink_target_process(pm, pp);
1258 
1259 	/* Issue a detach entry if a log file is configured */
1260 	if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE)
1261 		pmclog_process_pmcdetach(pm, p->p_pid);
1262 
1263 	/*
1264 	 * If there are no PMCs targeting this process, we remove its
1265 	 * descriptor from the target hash table and unset the P_HWPMC
1266 	 * flag in the struct proc.
1267 	 */
1268 	KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= (int) md->pmd_npmc,
1269 	    ("[pmc,%d] Illegal refcnt %d for process struct %p",
1270 		__LINE__, pp->pp_refcnt, pp));
1271 
1272 	if (pp->pp_refcnt != 0)	/* still a target of some PMC */
1273 		return 0;
1274 
1275 	pmc_remove_process_descriptor(pp);
1276 
1277 	if (flags & PMC_FLAG_REMOVE)
1278 		pmc_destroy_process_descriptor(pp);
1279 
1280 	PROC_LOCK(p);
1281 	p->p_flag &= ~P_HWPMC;
1282 	PROC_UNLOCK(p);
1283 
1284 	return 0;
1285 }
1286 
1287 /*
1288  * Detach a process and optionally its descendants from a PMC.
1289  */
1290 
1291 static int
1292 pmc_detach_process(struct proc *p, struct pmc *pm)
1293 {
1294 	struct proc *top;
1295 
1296 	sx_assert(&pmc_sx, SX_XLOCKED);
1297 
1298 	PMCDBG5(PRC,ATT,1, "detach pm=%p ri=%d proc=%p (%d, %s)", pm,
1299 	    PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm);
1300 
1301 	if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0)
1302 		return pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE);
1303 
1304 	/*
1305 	 * Traverse all children, detaching them from this PMC.  We
1306 	 * ignore errors since we could be detaching a PMC from a
1307 	 * partially attached proc tree.
1308 	 */
1309 
1310 	sx_slock(&proctree_lock);
1311 
1312 	top = p;
1313 
1314 	for (;;) {
1315 		(void) pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE);
1316 
1317 		if (!LIST_EMPTY(&p->p_children))
1318 			p = LIST_FIRST(&p->p_children);
1319 		else for (;;) {
1320 			if (p == top)
1321 				goto done;
1322 			if (LIST_NEXT(p, p_sibling)) {
1323 				p = LIST_NEXT(p, p_sibling);
1324 				break;
1325 			}
1326 			p = p->p_pptr;
1327 		}
1328 	}
1329 
1330  done:
1331 	sx_sunlock(&proctree_lock);
1332 
1333 	if (LIST_EMPTY(&pm->pm_targets))
1334 		pm->pm_flags &= ~PMC_F_ATTACH_DONE;
1335 
1336 	return 0;
1337 }
1338 
1339 
1340 /*
1341  * Thread context switch IN
1342  */
1343 
1344 static void
1345 pmc_process_csw_in(struct thread *td)
1346 {
1347 	int cpu;
1348 	unsigned int adjri, ri;
1349 	struct pmc *pm;
1350 	struct proc *p;
1351 	struct pmc_cpu *pc;
1352 	struct pmc_hw *phw;
1353 	pmc_value_t newvalue;
1354 	struct pmc_process *pp;
1355 	struct pmc_thread *pt;
1356 	struct pmc_classdep *pcd;
1357 
1358 	p = td->td_proc;
1359 	pt = NULL;
1360 	if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE)) == NULL)
1361 		return;
1362 
1363 	KASSERT(pp->pp_proc == td->td_proc,
1364 	    ("[pmc,%d] not my thread state", __LINE__));
1365 
1366 	critical_enter(); /* no preemption from this point */
1367 
1368 	cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */
1369 
1370 	PMCDBG5(CSW,SWI,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p,
1371 	    p->p_pid, p->p_comm, pp);
1372 
1373 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
1374 	    ("[pmc,%d] weird CPU id %d", __LINE__, cpu));
1375 
1376 	pc = pmc_pcpu[cpu];
1377 
1378 	for (ri = 0; ri < md->pmd_npmc; ri++) {
1379 
1380 		if ((pm = pp->pp_pmcs[ri].pp_pmc) == NULL)
1381 			continue;
1382 
1383 		KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)),
1384 		    ("[pmc,%d] Target PMC in non-virtual mode (%d)",
1385 			__LINE__, PMC_TO_MODE(pm)));
1386 
1387 		KASSERT(PMC_TO_ROWINDEX(pm) == ri,
1388 		    ("[pmc,%d] Row index mismatch pmc %d != ri %d",
1389 			__LINE__, PMC_TO_ROWINDEX(pm), ri));
1390 
1391 		/*
1392 		 * Only PMCs that are marked as 'RUNNING' need
1393 		 * be placed on hardware.
1394 		 */
1395 
1396 		if (pm->pm_state != PMC_STATE_RUNNING)
1397 			continue;
1398 
1399 		KASSERT(counter_u64_fetch(pm->pm_runcount) >= 0,
1400 	    ("[pmc,%d] pm=%p runcount %ld", __LINE__, (void *) pm,
1401 		 (unsigned long)counter_u64_fetch(pm->pm_runcount)));
1402 
1403 		/* increment PMC runcount */
1404 		counter_u64_add(pm->pm_runcount, 1);
1405 
1406 		/* configure the HWPMC we are going to use. */
1407 		pcd = pmc_ri_to_classdep(md, ri, &adjri);
1408 		pcd->pcd_config_pmc(cpu, adjri, pm);
1409 
1410 		phw = pc->pc_hwpmcs[ri];
1411 
1412 		KASSERT(phw != NULL,
1413 		    ("[pmc,%d] null hw pointer", __LINE__));
1414 
1415 		KASSERT(phw->phw_pmc == pm,
1416 		    ("[pmc,%d] hw->pmc %p != pmc %p", __LINE__,
1417 			phw->phw_pmc, pm));
1418 
1419 		/*
1420 		 * Write out saved value and start the PMC.
1421 		 *
1422 		 * Sampling PMCs use a per-thread value, while
1423 		 * counting mode PMCs use a per-pmc value that is
1424 		 * inherited across descendants.
1425 		 */
1426 		if (PMC_TO_MODE(pm) == PMC_MODE_TS) {
1427 			if (pt == NULL)
1428 				pt = pmc_find_thread_descriptor(pp, td,
1429 				    PMC_FLAG_NONE);
1430 
1431 			KASSERT(pt != NULL,
1432 			    ("[pmc,%d] No thread found for td=%p", __LINE__,
1433 			    td));
1434 
1435 			mtx_pool_lock_spin(pmc_mtxpool, pm);
1436 
1437 			/*
1438 			 * If we have a thread descriptor, use the per-thread
1439 			 * counter in the descriptor. If not, we will use
1440 			 * a per-process counter.
1441 			 *
1442 			 * TODO: Remove the per-process "safety net" once
1443 			 * we have thoroughly tested that we don't hit the
1444 			 * above assert.
1445 			 */
1446 			if (pt != NULL) {
1447 				if (pt->pt_pmcs[ri].pt_pmcval > 0)
1448 					newvalue = pt->pt_pmcs[ri].pt_pmcval;
1449 				else
1450 					newvalue = pm->pm_sc.pm_reloadcount;
1451 			} else {
1452 				/*
1453 				 * Use the saved value calculated after the most
1454 				 * recent time a thread using the shared counter
1455 				 * switched out. Reset the saved count in case
1456 				 * another thread from this process switches in
1457 				 * before any threads switch out.
1458 				 */
1459 
1460 				newvalue = pp->pp_pmcs[ri].pp_pmcval;
1461 				pp->pp_pmcs[ri].pp_pmcval =
1462 				    pm->pm_sc.pm_reloadcount;
1463 			}
1464 			mtx_pool_unlock_spin(pmc_mtxpool, pm);
1465 			KASSERT(newvalue > 0 && newvalue <=
1466 			    pm->pm_sc.pm_reloadcount,
1467 			    ("[pmc,%d] pmcval outside of expected range cpu=%d "
1468 			    "ri=%d pmcval=%jx pm_reloadcount=%jx", __LINE__,
1469 			    cpu, ri, newvalue, pm->pm_sc.pm_reloadcount));
1470 		} else {
1471 			KASSERT(PMC_TO_MODE(pm) == PMC_MODE_TC,
1472 			    ("[pmc,%d] illegal mode=%d", __LINE__,
1473 			    PMC_TO_MODE(pm)));
1474 			mtx_pool_lock_spin(pmc_mtxpool, pm);
1475 			newvalue = PMC_PCPU_SAVED(cpu, ri) =
1476 			    pm->pm_gv.pm_savedvalue;
1477 			mtx_pool_unlock_spin(pmc_mtxpool, pm);
1478 		}
1479 
1480 		PMCDBG3(CSW,SWI,1,"cpu=%d ri=%d new=%jd", cpu, ri, newvalue);
1481 
1482 		pcd->pcd_write_pmc(cpu, adjri, newvalue);
1483 
1484 		/* If a sampling mode PMC, reset stalled state. */
1485 		if (PMC_TO_MODE(pm) == PMC_MODE_TS)
1486 			pm->pm_pcpu_state[cpu].pps_stalled = 0;
1487 
1488 		/* Indicate that we desire this to run. */
1489 		pm->pm_pcpu_state[cpu].pps_cpustate = 1;
1490 
1491 		/* Start the PMC. */
1492 		pcd->pcd_start_pmc(cpu, adjri);
1493 	}
1494 
1495 	/*
1496 	 * perform any other architecture/cpu dependent thread
1497 	 * switch-in actions.
1498 	 */
1499 
1500 	(void) (*md->pmd_switch_in)(pc, pp);
1501 
1502 	critical_exit();
1503 
1504 }
1505 
1506 /*
1507  * Thread context switch OUT.
1508  */
1509 
1510 static void
1511 pmc_process_csw_out(struct thread *td)
1512 {
1513 	int cpu;
1514 	int64_t tmp;
1515 	struct pmc *pm;
1516 	struct proc *p;
1517 	enum pmc_mode mode;
1518 	struct pmc_cpu *pc;
1519 	pmc_value_t newvalue;
1520 	unsigned int adjri, ri;
1521 	struct pmc_process *pp;
1522 	struct pmc_thread *pt = NULL;
1523 	struct pmc_classdep *pcd;
1524 
1525 
1526 	/*
1527 	 * Locate our process descriptor; this may be NULL if
1528 	 * this process is exiting and we have already removed
1529 	 * the process from the target process table.
1530 	 *
1531 	 * Note that due to kernel preemption, multiple
1532 	 * context switches may happen while the process is
1533 	 * exiting.
1534 	 *
1535 	 * Note also that if the target process cannot be
1536 	 * found we still need to deconfigure any PMCs that
1537 	 * are currently running on hardware.
1538 	 */
1539 
1540 	p = td->td_proc;
1541 	pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE);
1542 
1543 	/*
1544 	 * save PMCs
1545 	 */
1546 
1547 	critical_enter();
1548 
1549 	cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */
1550 
1551 	PMCDBG5(CSW,SWO,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p,
1552 	    p->p_pid, p->p_comm, pp);
1553 
1554 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
1555 	    ("[pmc,%d weird CPU id %d", __LINE__, cpu));
1556 
1557 	pc = pmc_pcpu[cpu];
1558 
1559 	/*
1560 	 * When a PMC gets unlinked from a target PMC, it will
1561 	 * be removed from the target's pp_pmc[] array.
1562 	 *
1563 	 * However, on a MP system, the target could have been
1564 	 * executing on another CPU at the time of the unlink.
1565 	 * So, at context switch OUT time, we need to look at
1566 	 * the hardware to determine if a PMC is scheduled on
1567 	 * it.
1568 	 */
1569 
1570 	for (ri = 0; ri < md->pmd_npmc; ri++) {
1571 
1572 		pcd = pmc_ri_to_classdep(md, ri, &adjri);
1573 		pm  = NULL;
1574 		(void) (*pcd->pcd_get_config)(cpu, adjri, &pm);
1575 
1576 		if (pm == NULL)	/* nothing at this row index */
1577 			continue;
1578 
1579 		mode = PMC_TO_MODE(pm);
1580 		if (!PMC_IS_VIRTUAL_MODE(mode))
1581 			continue; /* not a process virtual PMC */
1582 
1583 		KASSERT(PMC_TO_ROWINDEX(pm) == ri,
1584 		    ("[pmc,%d] ri mismatch pmc(%d) ri(%d)",
1585 			__LINE__, PMC_TO_ROWINDEX(pm), ri));
1586 
1587 		/*
1588 		 * Change desired state, and then stop if not stalled.
1589 		 * This two-step dance should avoid race conditions where
1590 		 * an interrupt re-enables the PMC after this code has
1591 		 * already checked the pm_stalled flag.
1592 		 */
1593 		pm->pm_pcpu_state[cpu].pps_cpustate = 0;
1594 		if (pm->pm_pcpu_state[cpu].pps_stalled == 0)
1595 			pcd->pcd_stop_pmc(cpu, adjri);
1596 
1597 		KASSERT(counter_u64_fetch(pm->pm_runcount) > 0,
1598 			("[pmc,%d] pm=%p runcount %ld", __LINE__, (void *) pm,
1599 			 (unsigned long)counter_u64_fetch(pm->pm_runcount)));
1600 
1601 		/* reduce this PMC's runcount */
1602 		counter_u64_add(pm->pm_runcount, -1);
1603 
1604 		/*
1605 		 * If this PMC is associated with this process,
1606 		 * save the reading.
1607 		 */
1608 
1609 		if (pm->pm_state != PMC_STATE_DELETED && pp != NULL &&
1610 		    pp->pp_pmcs[ri].pp_pmc != NULL) {
1611 			KASSERT(pm == pp->pp_pmcs[ri].pp_pmc,
1612 			    ("[pmc,%d] pm %p != pp_pmcs[%d] %p", __LINE__,
1613 				pm, ri, pp->pp_pmcs[ri].pp_pmc));
1614 
1615 			KASSERT(pp->pp_refcnt > 0,
1616 			    ("[pmc,%d] pp refcnt = %d", __LINE__,
1617 				pp->pp_refcnt));
1618 
1619 			pcd->pcd_read_pmc(cpu, adjri, &newvalue);
1620 
1621 			if (mode == PMC_MODE_TS) {
1622 				PMCDBG3(CSW,SWO,1,"cpu=%d ri=%d val=%jd (samp)",
1623 				    cpu, ri, newvalue);
1624 
1625 				if (pt == NULL)
1626 					pt = pmc_find_thread_descriptor(pp, td,
1627 					    PMC_FLAG_NONE);
1628 
1629 				KASSERT(pt != NULL,
1630 				    ("[pmc,%d] No thread found for td=%p",
1631 				    __LINE__, td));
1632 
1633 				mtx_pool_lock_spin(pmc_mtxpool, pm);
1634 
1635 				/*
1636 				 * If we have a thread descriptor, save the
1637 				 * per-thread counter in the descriptor. If not,
1638 				 * we will update the per-process counter.
1639 				 *
1640 				 * TODO: Remove the per-process "safety net"
1641 				 * once we have thoroughly tested that we
1642 				 * don't hit the above assert.
1643 				 */
1644 				if (pt != NULL)
1645 					pt->pt_pmcs[ri].pt_pmcval = newvalue;
1646 				else {
1647 					/*
1648 					 * For sampling process-virtual PMCs,
1649 					 * newvalue is the number of events to
1650 					 * be seen until the next sampling
1651 					 * interrupt. We can just add the events
1652 					 * left from this invocation to the
1653 					 * counter, then adjust in case we
1654 					 * overflow our range.
1655 					 *
1656 					 * (Recall that we reload the counter
1657 					 * every time we use it.)
1658 					 */
1659 					pp->pp_pmcs[ri].pp_pmcval += newvalue;
1660 					if (pp->pp_pmcs[ri].pp_pmcval >
1661 					    pm->pm_sc.pm_reloadcount)
1662 						pp->pp_pmcs[ri].pp_pmcval -=
1663 						    pm->pm_sc.pm_reloadcount;
1664 				}
1665 				mtx_pool_unlock_spin(pmc_mtxpool, pm);
1666 			} else {
1667 				tmp = newvalue - PMC_PCPU_SAVED(cpu,ri);
1668 
1669 				PMCDBG3(CSW,SWO,1,"cpu=%d ri=%d tmp=%jd (count)",
1670 				    cpu, ri, tmp);
1671 
1672 				/*
1673 				 * For counting process-virtual PMCs,
1674 				 * we expect the count to be
1675 				 * increasing monotonically, modulo a 64
1676 				 * bit wraparound.
1677 				 */
1678 				KASSERT(tmp >= 0,
1679 				    ("[pmc,%d] negative increment cpu=%d "
1680 				     "ri=%d newvalue=%jx saved=%jx "
1681 				     "incr=%jx", __LINE__, cpu, ri,
1682 				     newvalue, PMC_PCPU_SAVED(cpu,ri), tmp));
1683 
1684 				mtx_pool_lock_spin(pmc_mtxpool, pm);
1685 				pm->pm_gv.pm_savedvalue += tmp;
1686 				pp->pp_pmcs[ri].pp_pmcval += tmp;
1687 				mtx_pool_unlock_spin(pmc_mtxpool, pm);
1688 
1689 				if (pm->pm_flags & PMC_F_LOG_PROCCSW)
1690 					pmclog_process_proccsw(pm, pp, tmp, td);
1691 			}
1692 		}
1693 
1694 		/* mark hardware as free */
1695 		pcd->pcd_config_pmc(cpu, adjri, NULL);
1696 	}
1697 
1698 	/*
1699 	 * perform any other architecture/cpu dependent thread
1700 	 * switch out functions.
1701 	 */
1702 
1703 	(void) (*md->pmd_switch_out)(pc, pp);
1704 
1705 	critical_exit();
1706 }
1707 
1708 /*
1709  * A new thread for a process.
1710  */
1711 static void
1712 pmc_process_thread_add(struct thread *td)
1713 {
1714 	struct pmc_process *pmc;
1715 
1716 	pmc = pmc_find_process_descriptor(td->td_proc, PMC_FLAG_NONE);
1717 	if (pmc != NULL)
1718 		pmc_find_thread_descriptor(pmc, td, PMC_FLAG_ALLOCATE);
1719 }
1720 
1721 /*
1722  * A thread delete for a process.
1723  */
1724 static void
1725 pmc_process_thread_delete(struct thread *td)
1726 {
1727 	struct pmc_process *pmc;
1728 
1729 	pmc = pmc_find_process_descriptor(td->td_proc, PMC_FLAG_NONE);
1730 	if (pmc != NULL)
1731 		pmc_thread_descriptor_pool_free(pmc_find_thread_descriptor(pmc,
1732 		    td, PMC_FLAG_REMOVE));
1733 }
1734 
1735 /*
1736  * A userret() call for a thread.
1737  */
1738 static void
1739 pmc_process_thread_userret(struct thread *td)
1740 {
1741 	sched_pin();
1742 	pmc_capture_user_callchain(curcpu, PMC_UR, td->td_frame);
1743 	sched_unpin();
1744 }
1745 
1746 /*
1747  * A mapping change for a process.
1748  */
1749 
1750 static void
1751 pmc_process_mmap(struct thread *td, struct pmckern_map_in *pkm)
1752 {
1753 	int ri;
1754 	pid_t pid;
1755 	char *fullpath, *freepath;
1756 	const struct pmc *pm;
1757 	struct pmc_owner *po;
1758 	const struct pmc_process *pp;
1759 
1760 	freepath = fullpath = NULL;
1761 	MPASS(!in_epoch(global_epoch_preempt));
1762 	pmc_getfilename((struct vnode *) pkm->pm_file, &fullpath, &freepath);
1763 
1764 	pid = td->td_proc->p_pid;
1765 
1766 	PMC_EPOCH_ENTER();
1767 	/* Inform owners of all system-wide sampling PMCs. */
1768 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1769 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1770 			pmclog_process_map_in(po, pid, pkm->pm_address, fullpath);
1771 
1772 	if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL)
1773 		goto done;
1774 
1775 	/*
1776 	 * Inform sampling PMC owners tracking this process.
1777 	 */
1778 	for (ri = 0; ri < md->pmd_npmc; ri++)
1779 		if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL &&
1780 		    PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1781 			pmclog_process_map_in(pm->pm_owner,
1782 			    pid, pkm->pm_address, fullpath);
1783 
1784   done:
1785 	if (freepath)
1786 		free(freepath, M_TEMP);
1787 	PMC_EPOCH_EXIT();
1788 }
1789 
1790 
1791 /*
1792  * Log an munmap request.
1793  */
1794 
1795 static void
1796 pmc_process_munmap(struct thread *td, struct pmckern_map_out *pkm)
1797 {
1798 	int ri;
1799 	pid_t pid;
1800 	struct pmc_owner *po;
1801 	const struct pmc *pm;
1802 	const struct pmc_process *pp;
1803 
1804 	pid = td->td_proc->p_pid;
1805 
1806 	PMC_EPOCH_ENTER();
1807 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1808 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1809 		pmclog_process_map_out(po, pid, pkm->pm_address,
1810 		    pkm->pm_address + pkm->pm_size);
1811 	PMC_EPOCH_EXIT();
1812 
1813 	if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL)
1814 		return;
1815 
1816 	for (ri = 0; ri < md->pmd_npmc; ri++)
1817 		if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL &&
1818 		    PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1819 			pmclog_process_map_out(pm->pm_owner, pid,
1820 			    pkm->pm_address, pkm->pm_address + pkm->pm_size);
1821 }
1822 
1823 /*
1824  * Log mapping information about the kernel.
1825  */
1826 
1827 static void
1828 pmc_log_kernel_mappings(struct pmc *pm)
1829 {
1830 	struct pmc_owner *po;
1831 	struct pmckern_map_in *km, *kmbase;
1832 
1833 	MPASS(in_epoch(global_epoch_preempt) || sx_xlocked(&pmc_sx));
1834 	KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)),
1835 	    ("[pmc,%d] non-sampling PMC (%p) desires mapping information",
1836 		__LINE__, (void *) pm));
1837 
1838 	po = pm->pm_owner;
1839 
1840 	if (po->po_flags & PMC_PO_INITIAL_MAPPINGS_DONE)
1841 		return;
1842 	if (PMC_TO_MODE(pm) == PMC_MODE_SS)
1843 		pmc_process_allproc(pm);
1844 	/*
1845 	 * Log the current set of kernel modules.
1846 	 */
1847 	kmbase = linker_hwpmc_list_objects();
1848 	for (km = kmbase; km->pm_file != NULL; km++) {
1849 		PMCDBG2(LOG,REG,1,"%s %p", (char *) km->pm_file,
1850 		    (void *) km->pm_address);
1851 		pmclog_process_map_in(po, (pid_t) -1, km->pm_address,
1852 		    km->pm_file);
1853 	}
1854 	free(kmbase, M_LINKER);
1855 
1856 	po->po_flags |= PMC_PO_INITIAL_MAPPINGS_DONE;
1857 }
1858 
1859 /*
1860  * Log the mappings for a single process.
1861  */
1862 
1863 static void
1864 pmc_log_process_mappings(struct pmc_owner *po, struct proc *p)
1865 {
1866 	vm_map_t map;
1867 	struct vnode *vp;
1868 	struct vmspace *vm;
1869 	vm_map_entry_t entry;
1870 	vm_offset_t last_end;
1871 	u_int last_timestamp;
1872 	struct vnode *last_vp;
1873 	vm_offset_t start_addr;
1874 	vm_object_t obj, lobj, tobj;
1875 	char *fullpath, *freepath;
1876 
1877 	last_vp = NULL;
1878 	last_end = (vm_offset_t) 0;
1879 	fullpath = freepath = NULL;
1880 
1881 	if ((vm = vmspace_acquire_ref(p)) == NULL)
1882 		return;
1883 
1884 	map = &vm->vm_map;
1885 	vm_map_lock_read(map);
1886 
1887 	VM_MAP_ENTRY_FOREACH(entry, map) {
1888 
1889 		if (entry == NULL) {
1890 			PMCDBG2(LOG,OPS,2, "hwpmc: vm_map entry unexpectedly "
1891 			    "NULL! pid=%d vm_map=%p\n", p->p_pid, map);
1892 			break;
1893 		}
1894 
1895 		/*
1896 		 * We only care about executable map entries.
1897 		 */
1898 		if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) ||
1899 		    !(entry->protection & VM_PROT_EXECUTE) ||
1900 		    (entry->object.vm_object == NULL)) {
1901 			continue;
1902 		}
1903 
1904 		obj = entry->object.vm_object;
1905 		VM_OBJECT_RLOCK(obj);
1906 
1907 		/*
1908 		 * Walk the backing_object list to find the base
1909 		 * (non-shadowed) vm_object.
1910 		 */
1911 		for (lobj = tobj = obj; tobj != NULL; tobj = tobj->backing_object) {
1912 			if (tobj != obj)
1913 				VM_OBJECT_RLOCK(tobj);
1914 			if (lobj != obj)
1915 				VM_OBJECT_RUNLOCK(lobj);
1916 			lobj = tobj;
1917 		}
1918 
1919 		/*
1920 		 * At this point lobj is the base vm_object and it is locked.
1921 		 */
1922 		if (lobj == NULL) {
1923 			PMCDBG3(LOG,OPS,2, "hwpmc: lobj unexpectedly NULL! pid=%d "
1924 			    "vm_map=%p vm_obj=%p\n", p->p_pid, map, obj);
1925 			VM_OBJECT_RUNLOCK(obj);
1926 			continue;
1927 		}
1928 
1929 		vp = vm_object_vnode(lobj);
1930 		if (vp == NULL) {
1931 			if (lobj != obj)
1932 				VM_OBJECT_RUNLOCK(lobj);
1933 			VM_OBJECT_RUNLOCK(obj);
1934 			continue;
1935 		}
1936 
1937 		/*
1938 		 * Skip contiguous regions that point to the same
1939 		 * vnode, so we don't emit redundant MAP-IN
1940 		 * directives.
1941 		 */
1942 		if (entry->start == last_end && vp == last_vp) {
1943 			last_end = entry->end;
1944 			if (lobj != obj)
1945 				VM_OBJECT_RUNLOCK(lobj);
1946 			VM_OBJECT_RUNLOCK(obj);
1947 			continue;
1948 		}
1949 
1950 		/*
1951 		 * We don't want to keep the proc's vm_map or this
1952 		 * vm_object locked while we walk the pathname, since
1953 		 * vn_fullpath() can sleep.  However, if we drop the
1954 		 * lock, it's possible for concurrent activity to
1955 		 * modify the vm_map list.  To protect against this,
1956 		 * we save the vm_map timestamp before we release the
1957 		 * lock, and check it after we reacquire the lock
1958 		 * below.
1959 		 */
1960 		start_addr = entry->start;
1961 		last_end = entry->end;
1962 		last_timestamp = map->timestamp;
1963 		vm_map_unlock_read(map);
1964 
1965 		vref(vp);
1966 		if (lobj != obj)
1967 			VM_OBJECT_RUNLOCK(lobj);
1968 
1969 		VM_OBJECT_RUNLOCK(obj);
1970 
1971 		freepath = NULL;
1972 		pmc_getfilename(vp, &fullpath, &freepath);
1973 		last_vp = vp;
1974 
1975 		vrele(vp);
1976 
1977 		vp = NULL;
1978 		pmclog_process_map_in(po, p->p_pid, start_addr, fullpath);
1979 		if (freepath)
1980 			free(freepath, M_TEMP);
1981 
1982 		vm_map_lock_read(map);
1983 
1984 		/*
1985 		 * If our saved timestamp doesn't match, this means
1986 		 * that the vm_map was modified out from under us and
1987 		 * we can't trust our current "entry" pointer.  Do a
1988 		 * new lookup for this entry.  If there is no entry
1989 		 * for this address range, vm_map_lookup_entry() will
1990 		 * return the previous one, so we always want to go to
1991 		 * the next entry on the next loop iteration.
1992 		 *
1993 		 * There is an edge condition here that can occur if
1994 		 * there is no entry at or before this address.  In
1995 		 * this situation, vm_map_lookup_entry returns
1996 		 * &map->header, which would cause our loop to abort
1997 		 * without processing the rest of the map.  However,
1998 		 * in practice this will never happen for process
1999 		 * vm_map.  This is because the executable's text
2000 		 * segment is the first mapping in the proc's address
2001 		 * space, and this mapping is never removed until the
2002 		 * process exits, so there will always be a non-header
2003 		 * entry at or before the requested address for
2004 		 * vm_map_lookup_entry to return.
2005 		 */
2006 		if (map->timestamp != last_timestamp)
2007 			vm_map_lookup_entry(map, last_end - 1, &entry);
2008 	}
2009 
2010 	vm_map_unlock_read(map);
2011 	vmspace_free(vm);
2012 	return;
2013 }
2014 
2015 /*
2016  * Log mappings for all processes in the system.
2017  */
2018 
2019 static void
2020 pmc_log_all_process_mappings(struct pmc_owner *po)
2021 {
2022 	struct proc *p, *top;
2023 
2024 	sx_assert(&pmc_sx, SX_XLOCKED);
2025 
2026 	if ((p = pfind(1)) == NULL)
2027 		panic("[pmc,%d] Cannot find init", __LINE__);
2028 
2029 	PROC_UNLOCK(p);
2030 
2031 	sx_slock(&proctree_lock);
2032 
2033 	top = p;
2034 
2035 	for (;;) {
2036 		pmc_log_process_mappings(po, p);
2037 		if (!LIST_EMPTY(&p->p_children))
2038 			p = LIST_FIRST(&p->p_children);
2039 		else for (;;) {
2040 			if (p == top)
2041 				goto done;
2042 			if (LIST_NEXT(p, p_sibling)) {
2043 				p = LIST_NEXT(p, p_sibling);
2044 				break;
2045 			}
2046 			p = p->p_pptr;
2047 		}
2048 	}
2049  done:
2050 	sx_sunlock(&proctree_lock);
2051 }
2052 
2053 /*
2054  * The 'hook' invoked from the kernel proper
2055  */
2056 
2057 
2058 #ifdef	HWPMC_DEBUG
2059 const char *pmc_hooknames[] = {
2060 	/* these strings correspond to PMC_FN_* in <sys/pmckern.h> */
2061 	"",
2062 	"EXEC",
2063 	"CSW-IN",
2064 	"CSW-OUT",
2065 	"SAMPLE",
2066 	"UNUSED1",
2067 	"UNUSED2",
2068 	"MMAP",
2069 	"MUNMAP",
2070 	"CALLCHAIN-NMI",
2071 	"CALLCHAIN-SOFT",
2072 	"SOFTSAMPLING",
2073 	"THR-CREATE",
2074 	"THR-EXIT",
2075 	"THR-USERRET",
2076 	"THR-CREATE-LOG",
2077 	"THR-EXIT-LOG",
2078 	"PROC-CREATE-LOG"
2079 };
2080 #endif
2081 
2082 static int
2083 pmc_hook_handler(struct thread *td, int function, void *arg)
2084 {
2085 	int cpu;
2086 
2087 	PMCDBG4(MOD,PMH,1, "hook td=%p func=%d \"%s\" arg=%p", td, function,
2088 	    pmc_hooknames[function], arg);
2089 
2090 	switch (function)
2091 	{
2092 
2093 	/*
2094 	 * Process exec()
2095 	 */
2096 
2097 	case PMC_FN_PROCESS_EXEC:
2098 	{
2099 		char *fullpath, *freepath;
2100 		unsigned int ri;
2101 		int is_using_hwpmcs;
2102 		struct pmc *pm;
2103 		struct proc *p;
2104 		struct pmc_owner *po;
2105 		struct pmc_process *pp;
2106 		struct pmckern_procexec *pk;
2107 
2108 		sx_assert(&pmc_sx, SX_XLOCKED);
2109 
2110 		p = td->td_proc;
2111 		pmc_getfilename(p->p_textvp, &fullpath, &freepath);
2112 
2113 		pk = (struct pmckern_procexec *) arg;
2114 
2115 		PMC_EPOCH_ENTER();
2116 		/* Inform owners of SS mode PMCs of the exec event. */
2117 		CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
2118 		    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
2119 			    pmclog_process_procexec(po, PMC_ID_INVALID,
2120 				p->p_pid, pk->pm_entryaddr, fullpath);
2121 		PMC_EPOCH_EXIT();
2122 
2123 		PROC_LOCK(p);
2124 		is_using_hwpmcs = p->p_flag & P_HWPMC;
2125 		PROC_UNLOCK(p);
2126 
2127 		if (!is_using_hwpmcs) {
2128 			if (freepath)
2129 				free(freepath, M_TEMP);
2130 			break;
2131 		}
2132 
2133 		/*
2134 		 * PMCs are not inherited across an exec():  remove any
2135 		 * PMCs that this process is the owner of.
2136 		 */
2137 
2138 		if ((po = pmc_find_owner_descriptor(p)) != NULL) {
2139 			pmc_remove_owner(po);
2140 			pmc_destroy_owner_descriptor(po);
2141 		}
2142 
2143 		/*
2144 		 * If the process being exec'ed is not the target of any
2145 		 * PMC, we are done.
2146 		 */
2147 		if ((pp = pmc_find_process_descriptor(p, 0)) == NULL) {
2148 			if (freepath)
2149 				free(freepath, M_TEMP);
2150 			break;
2151 		}
2152 
2153 		/*
2154 		 * Log the exec event to all monitoring owners.  Skip
2155 		 * owners who have already received the event because
2156 		 * they had system sampling PMCs active.
2157 		 */
2158 		for (ri = 0; ri < md->pmd_npmc; ri++)
2159 			if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) {
2160 				po = pm->pm_owner;
2161 				if (po->po_sscount == 0 &&
2162 				    po->po_flags & PMC_PO_OWNS_LOGFILE)
2163 					pmclog_process_procexec(po, pm->pm_id,
2164 					    p->p_pid, pk->pm_entryaddr,
2165 					    fullpath);
2166 			}
2167 
2168 		if (freepath)
2169 			free(freepath, M_TEMP);
2170 
2171 
2172 		PMCDBG4(PRC,EXC,1, "exec proc=%p (%d, %s) cred-changed=%d",
2173 		    p, p->p_pid, p->p_comm, pk->pm_credentialschanged);
2174 
2175 		if (pk->pm_credentialschanged == 0) /* no change */
2176 			break;
2177 
2178 		/*
2179 		 * If the newly exec()'ed process has a different credential
2180 		 * than before, allow it to be the target of a PMC only if
2181 		 * the PMC's owner has sufficient privilege.
2182 		 */
2183 
2184 		for (ri = 0; ri < md->pmd_npmc; ri++)
2185 			if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL)
2186 				if (pmc_can_attach(pm, td->td_proc) != 0)
2187 					pmc_detach_one_process(td->td_proc,
2188 					    pm, PMC_FLAG_NONE);
2189 
2190 		KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= (int) md->pmd_npmc,
2191 		    ("[pmc,%d] Illegal ref count %d on pp %p", __LINE__,
2192 			pp->pp_refcnt, pp));
2193 
2194 		/*
2195 		 * If this process is no longer the target of any
2196 		 * PMCs, we can remove the process entry and free
2197 		 * up space.
2198 		 */
2199 
2200 		if (pp->pp_refcnt == 0) {
2201 			pmc_remove_process_descriptor(pp);
2202 			pmc_destroy_process_descriptor(pp);
2203 			break;
2204 		}
2205 
2206 	}
2207 	break;
2208 
2209 	case PMC_FN_CSW_IN:
2210 		pmc_process_csw_in(td);
2211 		break;
2212 
2213 	case PMC_FN_CSW_OUT:
2214 		pmc_process_csw_out(td);
2215 		break;
2216 
2217 	/*
2218 	 * Process accumulated PC samples.
2219 	 *
2220 	 * This function is expected to be called by hardclock() for
2221 	 * each CPU that has accumulated PC samples.
2222 	 *
2223 	 * This function is to be executed on the CPU whose samples
2224 	 * are being processed.
2225 	 */
2226 	case PMC_FN_DO_SAMPLES:
2227 
2228 		/*
2229 		 * Clear the cpu specific bit in the CPU mask before
2230 		 * do the rest of the processing.  If the NMI handler
2231 		 * gets invoked after the "atomic_clear_int()" call
2232 		 * below but before "pmc_process_samples()" gets
2233 		 * around to processing the interrupt, then we will
2234 		 * come back here at the next hardclock() tick (and
2235 		 * may find nothing to do if "pmc_process_samples()"
2236 		 * had already processed the interrupt).  We don't
2237 		 * lose the interrupt sample.
2238 		 */
2239 		DPCPU_SET(pmc_sampled, 0);
2240 		cpu = PCPU_GET(cpuid);
2241 		pmc_process_samples(cpu, PMC_HR);
2242 		pmc_process_samples(cpu, PMC_SR);
2243 		pmc_process_samples(cpu, PMC_UR);
2244 		break;
2245 
2246 	case PMC_FN_MMAP:
2247 		pmc_process_mmap(td, (struct pmckern_map_in *) arg);
2248 		break;
2249 
2250 	case PMC_FN_MUNMAP:
2251 		MPASS(in_epoch(global_epoch_preempt) || sx_xlocked(&pmc_sx));
2252 		pmc_process_munmap(td, (struct pmckern_map_out *) arg);
2253 		break;
2254 
2255 	case PMC_FN_PROC_CREATE_LOG:
2256 		pmc_process_proccreate((struct proc *)arg);
2257 		break;
2258 
2259 	case PMC_FN_USER_CALLCHAIN:
2260 		/*
2261 		 * Record a call chain.
2262 		 */
2263 		KASSERT(td == curthread, ("[pmc,%d] td != curthread",
2264 		    __LINE__));
2265 
2266 		pmc_capture_user_callchain(PCPU_GET(cpuid), PMC_HR,
2267 		    (struct trapframe *) arg);
2268 
2269 		KASSERT(td->td_pinned == 1,
2270 			("[pmc,%d] invalid td_pinned value", __LINE__));
2271 		sched_unpin();  /* Can migrate safely now. */
2272 
2273 		td->td_pflags &= ~TDP_CALLCHAIN;
2274 		break;
2275 
2276 	case PMC_FN_USER_CALLCHAIN_SOFT:
2277 		/*
2278 		 * Record a call chain.
2279 		 */
2280 		KASSERT(td == curthread, ("[pmc,%d] td != curthread",
2281 		    __LINE__));
2282 
2283 		cpu = PCPU_GET(cpuid);
2284 		pmc_capture_user_callchain(cpu, PMC_SR,
2285 		    (struct trapframe *) arg);
2286 
2287 		KASSERT(td->td_pinned == 1,
2288 		    ("[pmc,%d] invalid td_pinned value", __LINE__));
2289 
2290 		sched_unpin();  /* Can migrate safely now. */
2291 
2292 		td->td_pflags &= ~TDP_CALLCHAIN;
2293 		break;
2294 
2295 	case PMC_FN_SOFT_SAMPLING:
2296 		/*
2297 		 * Call soft PMC sampling intr.
2298 		 */
2299 		pmc_soft_intr((struct pmckern_soft *) arg);
2300 		break;
2301 
2302 	case PMC_FN_THR_CREATE:
2303 		pmc_process_thread_add(td);
2304 		pmc_process_threadcreate(td);
2305 		break;
2306 
2307 	case PMC_FN_THR_CREATE_LOG:
2308 		pmc_process_threadcreate(td);
2309 		break;
2310 
2311 	case PMC_FN_THR_EXIT:
2312 		KASSERT(td == curthread, ("[pmc,%d] td != curthread",
2313 		    __LINE__));
2314 		pmc_process_thread_delete(td);
2315 		pmc_process_threadexit(td);
2316 		break;
2317 	case PMC_FN_THR_EXIT_LOG:
2318 		pmc_process_threadexit(td);
2319 		break;
2320 	case PMC_FN_THR_USERRET:
2321 		KASSERT(td == curthread, ("[pmc,%d] td != curthread",
2322 		    __LINE__));
2323 		pmc_process_thread_userret(td);
2324 		break;
2325 
2326 	default:
2327 #ifdef	HWPMC_DEBUG
2328 		KASSERT(0, ("[pmc,%d] unknown hook %d\n", __LINE__, function));
2329 #endif
2330 		break;
2331 
2332 	}
2333 
2334 	return 0;
2335 }
2336 
2337 /*
2338  * allocate a 'struct pmc_owner' descriptor in the owner hash table.
2339  */
2340 
2341 static struct pmc_owner *
2342 pmc_allocate_owner_descriptor(struct proc *p)
2343 {
2344 	uint32_t hindex;
2345 	struct pmc_owner *po;
2346 	struct pmc_ownerhash *poh;
2347 
2348 	hindex = PMC_HASH_PTR(p, pmc_ownerhashmask);
2349 	poh = &pmc_ownerhash[hindex];
2350 
2351 	/* allocate space for N pointers and one descriptor struct */
2352 	po = malloc(sizeof(struct pmc_owner), M_PMC, M_WAITOK|M_ZERO);
2353 	po->po_owner = p;
2354 	LIST_INSERT_HEAD(poh, po, po_next); /* insert into hash table */
2355 
2356 	TAILQ_INIT(&po->po_logbuffers);
2357 	mtx_init(&po->po_mtx, "pmc-owner-mtx", "pmc-per-proc", MTX_SPIN);
2358 
2359 	PMCDBG4(OWN,ALL,1, "allocate-owner proc=%p (%d, %s) pmc-owner=%p",
2360 	    p, p->p_pid, p->p_comm, po);
2361 
2362 	return po;
2363 }
2364 
2365 static void
2366 pmc_destroy_owner_descriptor(struct pmc_owner *po)
2367 {
2368 
2369 	PMCDBG4(OWN,REL,1, "destroy-owner po=%p proc=%p (%d, %s)",
2370 	    po, po->po_owner, po->po_owner->p_pid, po->po_owner->p_comm);
2371 
2372 	mtx_destroy(&po->po_mtx);
2373 	free(po, M_PMC);
2374 }
2375 
2376 /*
2377  * Allocate a thread descriptor from the free pool.
2378  *
2379  * NOTE: This *can* return NULL.
2380  */
2381 static struct pmc_thread *
2382 pmc_thread_descriptor_pool_alloc(void)
2383 {
2384 	struct pmc_thread *pt;
2385 
2386 	mtx_lock_spin(&pmc_threadfreelist_mtx);
2387 	if ((pt = LIST_FIRST(&pmc_threadfreelist)) != NULL) {
2388 		LIST_REMOVE(pt, pt_next);
2389 		pmc_threadfreelist_entries--;
2390 	}
2391 	mtx_unlock_spin(&pmc_threadfreelist_mtx);
2392 
2393 	return (pt);
2394 }
2395 
2396 /*
2397  * Add a thread descriptor to the free pool. We use this instead of free()
2398  * to maintain a cache of free entries. Additionally, we can safely call
2399  * this function when we cannot call free(), such as in a critical section.
2400  *
2401  */
2402 static void
2403 pmc_thread_descriptor_pool_free(struct pmc_thread *pt)
2404 {
2405 
2406 	if (pt == NULL)
2407 		return;
2408 
2409 	memset(pt, 0, THREADENTRY_SIZE);
2410 	mtx_lock_spin(&pmc_threadfreelist_mtx);
2411 	LIST_INSERT_HEAD(&pmc_threadfreelist, pt, pt_next);
2412 	pmc_threadfreelist_entries++;
2413 	if (pmc_threadfreelist_entries > pmc_threadfreelist_max)
2414 		GROUPTASK_ENQUEUE(&free_gtask);
2415 	mtx_unlock_spin(&pmc_threadfreelist_mtx);
2416 }
2417 
2418 /*
2419  * A callout to manage the free list.
2420  */
2421 static void
2422 pmc_thread_descriptor_pool_free_task(void *arg __unused)
2423 {
2424 	struct pmc_thread *pt;
2425 	LIST_HEAD(, pmc_thread) tmplist;
2426 	int delta;
2427 
2428 	LIST_INIT(&tmplist);
2429 	/* Determine what changes, if any, we need to make. */
2430 	mtx_lock_spin(&pmc_threadfreelist_mtx);
2431 	delta = pmc_threadfreelist_entries - pmc_threadfreelist_max;
2432 	while (delta > 0 &&
2433 		   (pt = LIST_FIRST(&pmc_threadfreelist)) != NULL) {
2434 		delta--;
2435 		LIST_REMOVE(pt, pt_next);
2436 		LIST_INSERT_HEAD(&tmplist, pt, pt_next);
2437 	}
2438 	mtx_unlock_spin(&pmc_threadfreelist_mtx);
2439 
2440 	/* If there are entries to free, free them. */
2441 	while (!LIST_EMPTY(&tmplist)) {
2442 		pt = LIST_FIRST(&tmplist);
2443 		LIST_REMOVE(pt, pt_next);
2444 		free(pt, M_PMC);
2445 	}
2446 }
2447 
2448 /*
2449  * Drain the thread free pool, freeing all allocations.
2450  */
2451 static void
2452 pmc_thread_descriptor_pool_drain()
2453 {
2454 	struct pmc_thread *pt, *next;
2455 
2456 	LIST_FOREACH_SAFE(pt, &pmc_threadfreelist, pt_next, next) {
2457 		LIST_REMOVE(pt, pt_next);
2458 		free(pt, M_PMC);
2459 	}
2460 }
2461 
2462 /*
2463  * find the descriptor corresponding to thread 'td', adding or removing it
2464  * as specified by 'mode'.
2465  *
2466  * Note that this supports additional mode flags in addition to those
2467  * supported by pmc_find_process_descriptor():
2468  * PMC_FLAG_NOWAIT: Causes the function to not wait for mallocs.
2469  *     This makes it safe to call while holding certain other locks.
2470  */
2471 
2472 static struct pmc_thread *
2473 pmc_find_thread_descriptor(struct pmc_process *pp, struct thread *td,
2474     uint32_t mode)
2475 {
2476 	struct pmc_thread *pt = NULL, *ptnew = NULL;
2477 	int wait_flag;
2478 
2479 	KASSERT(td != NULL, ("[pmc,%d] called to add NULL td", __LINE__));
2480 
2481 	/*
2482 	 * Pre-allocate memory in the PMC_FLAG_ALLOCATE case prior to
2483 	 * acquiring the lock.
2484 	 */
2485 	if (mode & PMC_FLAG_ALLOCATE) {
2486 		if ((ptnew = pmc_thread_descriptor_pool_alloc()) == NULL) {
2487 			wait_flag = M_WAITOK;
2488 			if ((mode & PMC_FLAG_NOWAIT) || in_epoch(global_epoch_preempt))
2489 				wait_flag = M_NOWAIT;
2490 
2491 			ptnew = malloc(THREADENTRY_SIZE, M_PMC,
2492 			    wait_flag|M_ZERO);
2493 		}
2494 	}
2495 
2496 	mtx_lock_spin(pp->pp_tdslock);
2497 
2498 	LIST_FOREACH(pt, &pp->pp_tds, pt_next)
2499 		if (pt->pt_td == td)
2500 			break;
2501 
2502 	if ((mode & PMC_FLAG_REMOVE) && pt != NULL)
2503 		LIST_REMOVE(pt, pt_next);
2504 
2505 	if ((mode & PMC_FLAG_ALLOCATE) && pt == NULL && ptnew != NULL) {
2506 		pt = ptnew;
2507 		ptnew = NULL;
2508 		pt->pt_td = td;
2509 		LIST_INSERT_HEAD(&pp->pp_tds, pt, pt_next);
2510 	}
2511 
2512 	mtx_unlock_spin(pp->pp_tdslock);
2513 
2514 	if (ptnew != NULL) {
2515 		free(ptnew, M_PMC);
2516 	}
2517 
2518 	return pt;
2519 }
2520 
2521 /*
2522  * Try to add thread descriptors for each thread in a process.
2523  */
2524 
2525 static void
2526 pmc_add_thread_descriptors_from_proc(struct proc *p, struct pmc_process *pp)
2527 {
2528 	struct thread *curtd;
2529 	struct pmc_thread **tdlist;
2530 	int i, tdcnt, tdlistsz;
2531 
2532 	KASSERT(!PROC_LOCKED(p), ("[pmc,%d] proc unexpectedly locked",
2533 	    __LINE__));
2534 	tdcnt = 32;
2535  restart:
2536 	tdlistsz = roundup2(tdcnt, 32);
2537 
2538 	tdcnt = 0;
2539 	tdlist = malloc(sizeof(struct pmc_thread*) * tdlistsz, M_TEMP, M_WAITOK);
2540 
2541 	PROC_LOCK(p);
2542 	FOREACH_THREAD_IN_PROC(p, curtd)
2543 		tdcnt++;
2544 	if (tdcnt >= tdlistsz) {
2545 		PROC_UNLOCK(p);
2546 		free(tdlist, M_TEMP);
2547 		goto restart;
2548 	}
2549 	/*
2550 	 * Try to add each thread to the list without sleeping. If unable,
2551 	 * add to a queue to retry after dropping the process lock.
2552 	 */
2553 	tdcnt = 0;
2554 	FOREACH_THREAD_IN_PROC(p, curtd) {
2555 		tdlist[tdcnt] = pmc_find_thread_descriptor(pp, curtd,
2556 						   PMC_FLAG_ALLOCATE|PMC_FLAG_NOWAIT);
2557 		if (tdlist[tdcnt] == NULL) {
2558 			PROC_UNLOCK(p);
2559 			for (i = 0; i <= tdcnt; i++)
2560 				pmc_thread_descriptor_pool_free(tdlist[i]);
2561 			free(tdlist, M_TEMP);
2562 			goto restart;
2563 		}
2564 		tdcnt++;
2565 	}
2566 	PROC_UNLOCK(p);
2567 	free(tdlist, M_TEMP);
2568 }
2569 
2570 /*
2571  * find the descriptor corresponding to process 'p', adding or removing it
2572  * as specified by 'mode'.
2573  */
2574 
2575 static struct pmc_process *
2576 pmc_find_process_descriptor(struct proc *p, uint32_t mode)
2577 {
2578 	uint32_t hindex;
2579 	struct pmc_process *pp, *ppnew;
2580 	struct pmc_processhash *pph;
2581 
2582 	hindex = PMC_HASH_PTR(p, pmc_processhashmask);
2583 	pph = &pmc_processhash[hindex];
2584 
2585 	ppnew = NULL;
2586 
2587 	/*
2588 	 * Pre-allocate memory in the PMC_FLAG_ALLOCATE case since we
2589 	 * cannot call malloc(9) once we hold a spin lock.
2590 	 */
2591 	if (mode & PMC_FLAG_ALLOCATE)
2592 		ppnew = malloc(sizeof(struct pmc_process) + md->pmd_npmc *
2593 		    sizeof(struct pmc_targetstate), M_PMC, M_WAITOK|M_ZERO);
2594 
2595 	mtx_lock_spin(&pmc_processhash_mtx);
2596 	LIST_FOREACH(pp, pph, pp_next)
2597 	    if (pp->pp_proc == p)
2598 		    break;
2599 
2600 	if ((mode & PMC_FLAG_REMOVE) && pp != NULL)
2601 		LIST_REMOVE(pp, pp_next);
2602 
2603 	if ((mode & PMC_FLAG_ALLOCATE) && pp == NULL &&
2604 	    ppnew != NULL) {
2605 		ppnew->pp_proc = p;
2606 		LIST_INIT(&ppnew->pp_tds);
2607 		ppnew->pp_tdslock = mtx_pool_find(pmc_mtxpool, ppnew);
2608 		LIST_INSERT_HEAD(pph, ppnew, pp_next);
2609 		mtx_unlock_spin(&pmc_processhash_mtx);
2610 		pp = ppnew;
2611 		ppnew = NULL;
2612 
2613 		/* Add thread descriptors for this process' current threads. */
2614 		pmc_add_thread_descriptors_from_proc(p, pp);
2615 	}
2616 	else
2617 		mtx_unlock_spin(&pmc_processhash_mtx);
2618 
2619 	if (ppnew != NULL)
2620 		free(ppnew, M_PMC);
2621 
2622 	return pp;
2623 }
2624 
2625 /*
2626  * remove a process descriptor from the process hash table.
2627  */
2628 
2629 static void
2630 pmc_remove_process_descriptor(struct pmc_process *pp)
2631 {
2632 	KASSERT(pp->pp_refcnt == 0,
2633 	    ("[pmc,%d] Removing process descriptor %p with count %d",
2634 		__LINE__, pp, pp->pp_refcnt));
2635 
2636 	mtx_lock_spin(&pmc_processhash_mtx);
2637 	LIST_REMOVE(pp, pp_next);
2638 	mtx_unlock_spin(&pmc_processhash_mtx);
2639 }
2640 
2641 /*
2642  * destroy a process descriptor.
2643  */
2644 
2645 static void
2646 pmc_destroy_process_descriptor(struct pmc_process *pp)
2647 {
2648 	struct pmc_thread *pmc_td;
2649 
2650 	while ((pmc_td = LIST_FIRST(&pp->pp_tds)) != NULL) {
2651 		LIST_REMOVE(pmc_td, pt_next);
2652 		pmc_thread_descriptor_pool_free(pmc_td);
2653 	}
2654 	free(pp, M_PMC);
2655 }
2656 
2657 
2658 /*
2659  * find an owner descriptor corresponding to proc 'p'
2660  */
2661 
2662 static struct pmc_owner *
2663 pmc_find_owner_descriptor(struct proc *p)
2664 {
2665 	uint32_t hindex;
2666 	struct pmc_owner *po;
2667 	struct pmc_ownerhash *poh;
2668 
2669 	hindex = PMC_HASH_PTR(p, pmc_ownerhashmask);
2670 	poh = &pmc_ownerhash[hindex];
2671 
2672 	po = NULL;
2673 	LIST_FOREACH(po, poh, po_next)
2674 	    if (po->po_owner == p)
2675 		    break;
2676 
2677 	PMCDBG5(OWN,FND,1, "find-owner proc=%p (%d, %s) hindex=0x%x -> "
2678 	    "pmc-owner=%p", p, p->p_pid, p->p_comm, hindex, po);
2679 
2680 	return po;
2681 }
2682 
2683 /*
2684  * pmc_allocate_pmc_descriptor
2685  *
2686  * Allocate a pmc descriptor and initialize its
2687  * fields.
2688  */
2689 
2690 static struct pmc *
2691 pmc_allocate_pmc_descriptor(void)
2692 {
2693 	struct pmc *pmc;
2694 
2695 	pmc = malloc(sizeof(struct pmc), M_PMC, M_WAITOK|M_ZERO);
2696 	pmc->pm_runcount = counter_u64_alloc(M_WAITOK);
2697 	pmc->pm_pcpu_state = malloc(sizeof(struct pmc_pcpu_state)*mp_ncpus, M_PMC, M_WAITOK|M_ZERO);
2698 	PMCDBG1(PMC,ALL,1, "allocate-pmc -> pmc=%p", pmc);
2699 
2700 	return pmc;
2701 }
2702 
2703 /*
2704  * Destroy a pmc descriptor.
2705  */
2706 
2707 static void
2708 pmc_destroy_pmc_descriptor(struct pmc *pm)
2709 {
2710 
2711 	KASSERT(pm->pm_state == PMC_STATE_DELETED ||
2712 	    pm->pm_state == PMC_STATE_FREE,
2713 	    ("[pmc,%d] destroying non-deleted PMC", __LINE__));
2714 	KASSERT(LIST_EMPTY(&pm->pm_targets),
2715 	    ("[pmc,%d] destroying pmc with targets", __LINE__));
2716 	KASSERT(pm->pm_owner == NULL,
2717 	    ("[pmc,%d] destroying pmc attached to an owner", __LINE__));
2718 	KASSERT(counter_u64_fetch(pm->pm_runcount) == 0,
2719 	    ("[pmc,%d] pmc has non-zero run count %ld", __LINE__,
2720 		 (unsigned long)counter_u64_fetch(pm->pm_runcount)));
2721 
2722 	counter_u64_free(pm->pm_runcount);
2723 	free(pm->pm_pcpu_state, M_PMC);
2724 	free(pm, M_PMC);
2725 }
2726 
2727 static void
2728 pmc_wait_for_pmc_idle(struct pmc *pm)
2729 {
2730 #ifdef INVARIANTS
2731 	volatile int maxloop;
2732 
2733 	maxloop = 100 * pmc_cpu_max();
2734 #endif
2735 	/*
2736 	 * Loop (with a forced context switch) till the PMC's runcount
2737 	 * comes down to zero.
2738 	 */
2739 	pmclog_flush(pm->pm_owner, 1);
2740 	while (counter_u64_fetch(pm->pm_runcount) > 0) {
2741 		pmclog_flush(pm->pm_owner, 1);
2742 #ifdef INVARIANTS
2743 		maxloop--;
2744 		KASSERT(maxloop > 0,
2745 		    ("[pmc,%d] (ri%d, rc%ld) waiting too long for "
2746 			"pmc to be free", __LINE__,
2747 			 PMC_TO_ROWINDEX(pm), (unsigned long)counter_u64_fetch(pm->pm_runcount)));
2748 #endif
2749 		pmc_force_context_switch();
2750 	}
2751 }
2752 
2753 /*
2754  * This function does the following things:
2755  *
2756  *  - detaches the PMC from hardware
2757  *  - unlinks all target threads that were attached to it
2758  *  - removes the PMC from its owner's list
2759  *  - destroys the PMC private mutex
2760  *
2761  * Once this function completes, the given pmc pointer can be freed by
2762  * calling pmc_destroy_pmc_descriptor().
2763  */
2764 
2765 static void
2766 pmc_release_pmc_descriptor(struct pmc *pm)
2767 {
2768 	enum pmc_mode mode;
2769 	struct pmc_hw *phw;
2770 	u_int adjri, ri, cpu;
2771 	struct pmc_owner *po;
2772 	struct pmc_binding pb;
2773 	struct pmc_process *pp;
2774 	struct pmc_classdep *pcd;
2775 	struct pmc_target *ptgt, *tmp;
2776 
2777 	sx_assert(&pmc_sx, SX_XLOCKED);
2778 
2779 	KASSERT(pm, ("[pmc,%d] null pmc", __LINE__));
2780 
2781 	ri   = PMC_TO_ROWINDEX(pm);
2782 	pcd  = pmc_ri_to_classdep(md, ri, &adjri);
2783 	mode = PMC_TO_MODE(pm);
2784 
2785 	PMCDBG3(PMC,REL,1, "release-pmc pmc=%p ri=%d mode=%d", pm, ri,
2786 	    mode);
2787 
2788 	/*
2789 	 * First, we take the PMC off hardware.
2790 	 */
2791 	cpu = 0;
2792 	if (PMC_IS_SYSTEM_MODE(mode)) {
2793 
2794 		/*
2795 		 * A system mode PMC runs on a specific CPU.  Switch
2796 		 * to this CPU and turn hardware off.
2797 		 */
2798 		pmc_save_cpu_binding(&pb);
2799 
2800 		cpu = PMC_TO_CPU(pm);
2801 
2802 		pmc_select_cpu(cpu);
2803 
2804 		/* switch off non-stalled CPUs */
2805 		pm->pm_pcpu_state[cpu].pps_cpustate = 0;
2806 		if (pm->pm_state == PMC_STATE_RUNNING &&
2807 			pm->pm_pcpu_state[cpu].pps_stalled == 0) {
2808 
2809 			phw = pmc_pcpu[cpu]->pc_hwpmcs[ri];
2810 
2811 			KASSERT(phw->phw_pmc == pm,
2812 			    ("[pmc, %d] pmc ptr ri(%d) hw(%p) pm(%p)",
2813 				__LINE__, ri, phw->phw_pmc, pm));
2814 			PMCDBG2(PMC,REL,2, "stopping cpu=%d ri=%d", cpu, ri);
2815 
2816 			critical_enter();
2817 			pcd->pcd_stop_pmc(cpu, adjri);
2818 			critical_exit();
2819 		}
2820 
2821 		PMCDBG2(PMC,REL,2, "decfg cpu=%d ri=%d", cpu, ri);
2822 
2823 		critical_enter();
2824 		pcd->pcd_config_pmc(cpu, adjri, NULL);
2825 		critical_exit();
2826 
2827 		/* adjust the global and process count of SS mode PMCs */
2828 		if (mode == PMC_MODE_SS && pm->pm_state == PMC_STATE_RUNNING) {
2829 			po = pm->pm_owner;
2830 			po->po_sscount--;
2831 			if (po->po_sscount == 0) {
2832 				atomic_subtract_rel_int(&pmc_ss_count, 1);
2833 				CK_LIST_REMOVE(po, po_ssnext);
2834 				epoch_wait_preempt(global_epoch_preempt);
2835 			}
2836 		}
2837 
2838 		pm->pm_state = PMC_STATE_DELETED;
2839 
2840 		pmc_restore_cpu_binding(&pb);
2841 
2842 		/*
2843 		 * We could have references to this PMC structure in
2844 		 * the per-cpu sample queues.  Wait for the queue to
2845 		 * drain.
2846 		 */
2847 		pmc_wait_for_pmc_idle(pm);
2848 
2849 	} else if (PMC_IS_VIRTUAL_MODE(mode)) {
2850 
2851 		/*
2852 		 * A virtual PMC could be running on multiple CPUs at
2853 		 * a given instant.
2854 		 *
2855 		 * By marking its state as DELETED, we ensure that
2856 		 * this PMC is never further scheduled on hardware.
2857 		 *
2858 		 * Then we wait till all CPUs are done with this PMC.
2859 		 */
2860 		pm->pm_state = PMC_STATE_DELETED;
2861 
2862 
2863 		/* Wait for the PMCs runcount to come to zero. */
2864 		pmc_wait_for_pmc_idle(pm);
2865 
2866 		/*
2867 		 * At this point the PMC is off all CPUs and cannot be
2868 		 * freshly scheduled onto a CPU.  It is now safe to
2869 		 * unlink all targets from this PMC.  If a
2870 		 * process-record's refcount falls to zero, we remove
2871 		 * it from the hash table.  The module-wide SX lock
2872 		 * protects us from races.
2873 		 */
2874 		LIST_FOREACH_SAFE(ptgt, &pm->pm_targets, pt_next, tmp) {
2875 			pp = ptgt->pt_process;
2876 			pmc_unlink_target_process(pm, pp); /* frees 'ptgt' */
2877 
2878 			PMCDBG1(PMC,REL,3, "pp->refcnt=%d", pp->pp_refcnt);
2879 
2880 			/*
2881 			 * If the target process record shows that no
2882 			 * PMCs are attached to it, reclaim its space.
2883 			 */
2884 
2885 			if (pp->pp_refcnt == 0) {
2886 				pmc_remove_process_descriptor(pp);
2887 				pmc_destroy_process_descriptor(pp);
2888 			}
2889 		}
2890 
2891 		cpu = curthread->td_oncpu; /* setup cpu for pmd_release() */
2892 
2893 	}
2894 
2895 	/*
2896 	 * Release any MD resources
2897 	 */
2898 	(void) pcd->pcd_release_pmc(cpu, adjri, pm);
2899 
2900 	/*
2901 	 * Update row disposition
2902 	 */
2903 
2904 	if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm)))
2905 		PMC_UNMARK_ROW_STANDALONE(ri);
2906 	else
2907 		PMC_UNMARK_ROW_THREAD(ri);
2908 
2909 	/* unlink from the owner's list */
2910 	if (pm->pm_owner) {
2911 		LIST_REMOVE(pm, pm_next);
2912 		pm->pm_owner = NULL;
2913 	}
2914 }
2915 
2916 /*
2917  * Register an owner and a pmc.
2918  */
2919 
2920 static int
2921 pmc_register_owner(struct proc *p, struct pmc *pmc)
2922 {
2923 	struct pmc_owner *po;
2924 
2925 	sx_assert(&pmc_sx, SX_XLOCKED);
2926 
2927 	if ((po = pmc_find_owner_descriptor(p)) == NULL)
2928 		if ((po = pmc_allocate_owner_descriptor(p)) == NULL)
2929 			return ENOMEM;
2930 
2931 	KASSERT(pmc->pm_owner == NULL,
2932 	    ("[pmc,%d] attempting to own an initialized PMC", __LINE__));
2933 	pmc->pm_owner  = po;
2934 
2935 	LIST_INSERT_HEAD(&po->po_pmcs, pmc, pm_next);
2936 
2937 	PROC_LOCK(p);
2938 	p->p_flag |= P_HWPMC;
2939 	PROC_UNLOCK(p);
2940 
2941 	if (po->po_flags & PMC_PO_OWNS_LOGFILE)
2942 		pmclog_process_pmcallocate(pmc);
2943 
2944 	PMCDBG2(PMC,REG,1, "register-owner pmc-owner=%p pmc=%p",
2945 	    po, pmc);
2946 
2947 	return 0;
2948 }
2949 
2950 /*
2951  * Return the current row disposition:
2952  * == 0 => FREE
2953  *  > 0 => PROCESS MODE
2954  *  < 0 => SYSTEM MODE
2955  */
2956 
2957 int
2958 pmc_getrowdisp(int ri)
2959 {
2960 	return pmc_pmcdisp[ri];
2961 }
2962 
2963 /*
2964  * Check if a PMC at row index 'ri' can be allocated to the current
2965  * process.
2966  *
2967  * Allocation can fail if:
2968  *   - the current process is already being profiled by a PMC at index 'ri',
2969  *     attached to it via OP_PMCATTACH.
2970  *   - the current process has already allocated a PMC at index 'ri'
2971  *     via OP_ALLOCATE.
2972  */
2973 
2974 static int
2975 pmc_can_allocate_rowindex(struct proc *p, unsigned int ri, int cpu)
2976 {
2977 	enum pmc_mode mode;
2978 	struct pmc *pm;
2979 	struct pmc_owner *po;
2980 	struct pmc_process *pp;
2981 
2982 	PMCDBG5(PMC,ALR,1, "can-allocate-rowindex proc=%p (%d, %s) ri=%d "
2983 	    "cpu=%d", p, p->p_pid, p->p_comm, ri, cpu);
2984 
2985 	/*
2986 	 * We shouldn't have already allocated a process-mode PMC at
2987 	 * row index 'ri'.
2988 	 *
2989 	 * We shouldn't have allocated a system-wide PMC on the same
2990 	 * CPU and same RI.
2991 	 */
2992 	if ((po = pmc_find_owner_descriptor(p)) != NULL)
2993 		LIST_FOREACH(pm, &po->po_pmcs, pm_next) {
2994 		    if (PMC_TO_ROWINDEX(pm) == ri) {
2995 			    mode = PMC_TO_MODE(pm);
2996 			    if (PMC_IS_VIRTUAL_MODE(mode))
2997 				    return EEXIST;
2998 			    if (PMC_IS_SYSTEM_MODE(mode) &&
2999 				(int) PMC_TO_CPU(pm) == cpu)
3000 				    return EEXIST;
3001 		    }
3002 	        }
3003 
3004 	/*
3005 	 * We also shouldn't be the target of any PMC at this index
3006 	 * since otherwise a PMC_ATTACH to ourselves will fail.
3007 	 */
3008 	if ((pp = pmc_find_process_descriptor(p, 0)) != NULL)
3009 		if (pp->pp_pmcs[ri].pp_pmc)
3010 			return EEXIST;
3011 
3012 	PMCDBG4(PMC,ALR,2, "can-allocate-rowindex proc=%p (%d, %s) ri=%d ok",
3013 	    p, p->p_pid, p->p_comm, ri);
3014 
3015 	return 0;
3016 }
3017 
3018 /*
3019  * Check if a given PMC at row index 'ri' can be currently used in
3020  * mode 'mode'.
3021  */
3022 
3023 static int
3024 pmc_can_allocate_row(int ri, enum pmc_mode mode)
3025 {
3026 	enum pmc_disp	disp;
3027 
3028 	sx_assert(&pmc_sx, SX_XLOCKED);
3029 
3030 	PMCDBG2(PMC,ALR,1, "can-allocate-row ri=%d mode=%d", ri, mode);
3031 
3032 	if (PMC_IS_SYSTEM_MODE(mode))
3033 		disp = PMC_DISP_STANDALONE;
3034 	else
3035 		disp = PMC_DISP_THREAD;
3036 
3037 	/*
3038 	 * check disposition for PMC row 'ri':
3039 	 *
3040 	 * Expected disposition		Row-disposition		Result
3041 	 *
3042 	 * STANDALONE			STANDALONE or FREE	proceed
3043 	 * STANDALONE			THREAD			fail
3044 	 * THREAD			THREAD or FREE		proceed
3045 	 * THREAD			STANDALONE		fail
3046 	 */
3047 
3048 	if (!PMC_ROW_DISP_IS_FREE(ri) &&
3049 	    !(disp == PMC_DISP_THREAD && PMC_ROW_DISP_IS_THREAD(ri)) &&
3050 	    !(disp == PMC_DISP_STANDALONE && PMC_ROW_DISP_IS_STANDALONE(ri)))
3051 		return EBUSY;
3052 
3053 	/*
3054 	 * All OK
3055 	 */
3056 
3057 	PMCDBG2(PMC,ALR,2, "can-allocate-row ri=%d mode=%d ok", ri, mode);
3058 
3059 	return 0;
3060 
3061 }
3062 
3063 /*
3064  * Find a PMC descriptor with user handle 'pmcid' for thread 'td'.
3065  */
3066 
3067 static struct pmc *
3068 pmc_find_pmc_descriptor_in_process(struct pmc_owner *po, pmc_id_t pmcid)
3069 {
3070 	struct pmc *pm;
3071 
3072 	KASSERT(PMC_ID_TO_ROWINDEX(pmcid) < md->pmd_npmc,
3073 	    ("[pmc,%d] Illegal pmc index %d (max %d)", __LINE__,
3074 		PMC_ID_TO_ROWINDEX(pmcid), md->pmd_npmc));
3075 
3076 	LIST_FOREACH(pm, &po->po_pmcs, pm_next)
3077 	    if (pm->pm_id == pmcid)
3078 		    return pm;
3079 
3080 	return NULL;
3081 }
3082 
3083 static int
3084 pmc_find_pmc(pmc_id_t pmcid, struct pmc **pmc)
3085 {
3086 
3087 	struct pmc *pm, *opm;
3088 	struct pmc_owner *po;
3089 	struct pmc_process *pp;
3090 
3091 	PMCDBG1(PMC,FND,1, "find-pmc id=%d", pmcid);
3092 	if (PMC_ID_TO_ROWINDEX(pmcid) >= md->pmd_npmc)
3093 		return (EINVAL);
3094 
3095 	if ((po = pmc_find_owner_descriptor(curthread->td_proc)) == NULL) {
3096 		/*
3097 		 * In case of PMC_F_DESCENDANTS child processes we will not find
3098 		 * the current process in the owners hash list.  Find the owner
3099 		 * process first and from there lookup the po.
3100 		 */
3101 		if ((pp = pmc_find_process_descriptor(curthread->td_proc,
3102 		    PMC_FLAG_NONE)) == NULL) {
3103 			return ESRCH;
3104 		} else {
3105 			opm = pp->pp_pmcs[PMC_ID_TO_ROWINDEX(pmcid)].pp_pmc;
3106 			if (opm == NULL)
3107 				return ESRCH;
3108 			if ((opm->pm_flags & (PMC_F_ATTACHED_TO_OWNER|
3109 			    PMC_F_DESCENDANTS)) != (PMC_F_ATTACHED_TO_OWNER|
3110 			    PMC_F_DESCENDANTS))
3111 				return ESRCH;
3112 			po = opm->pm_owner;
3113 		}
3114 	}
3115 
3116 	if ((pm = pmc_find_pmc_descriptor_in_process(po, pmcid)) == NULL)
3117 		return EINVAL;
3118 
3119 	PMCDBG2(PMC,FND,2, "find-pmc id=%d -> pmc=%p", pmcid, pm);
3120 
3121 	*pmc = pm;
3122 	return 0;
3123 }
3124 
3125 /*
3126  * Start a PMC.
3127  */
3128 
3129 static int
3130 pmc_start(struct pmc *pm)
3131 {
3132 	enum pmc_mode mode;
3133 	struct pmc_owner *po;
3134 	struct pmc_binding pb;
3135 	struct pmc_classdep *pcd;
3136 	int adjri, error, cpu, ri;
3137 
3138 	KASSERT(pm != NULL,
3139 	    ("[pmc,%d] null pm", __LINE__));
3140 
3141 	mode = PMC_TO_MODE(pm);
3142 	ri   = PMC_TO_ROWINDEX(pm);
3143 	pcd  = pmc_ri_to_classdep(md, ri, &adjri);
3144 
3145 	error = 0;
3146 
3147 	PMCDBG3(PMC,OPS,1, "start pmc=%p mode=%d ri=%d", pm, mode, ri);
3148 
3149 	po = pm->pm_owner;
3150 
3151 	/*
3152 	 * Disallow PMCSTART if a logfile is required but has not been
3153 	 * configured yet.
3154 	 */
3155 	if ((pm->pm_flags & PMC_F_NEEDS_LOGFILE) &&
3156 	    (po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
3157 		return (EDOOFUS);	/* programming error */
3158 
3159 	/*
3160 	 * If this is a sampling mode PMC, log mapping information for
3161 	 * the kernel modules that are currently loaded.
3162 	 */
3163 	if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
3164 	    pmc_log_kernel_mappings(pm);
3165 
3166 	if (PMC_IS_VIRTUAL_MODE(mode)) {
3167 
3168 		/*
3169 		 * If a PMCATTACH has never been done on this PMC,
3170 		 * attach it to its owner process.
3171 		 */
3172 
3173 		if (LIST_EMPTY(&pm->pm_targets))
3174 			error = (pm->pm_flags & PMC_F_ATTACH_DONE) ? ESRCH :
3175 			    pmc_attach_process(po->po_owner, pm);
3176 
3177 		/*
3178 		 * If the PMC is attached to its owner, then force a context
3179 		 * switch to ensure that the MD state gets set correctly.
3180 		 */
3181 
3182 		if (error == 0) {
3183 			pm->pm_state = PMC_STATE_RUNNING;
3184 			if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER)
3185 				pmc_force_context_switch();
3186 		}
3187 
3188 		return (error);
3189 	}
3190 
3191 
3192 	/*
3193 	 * A system-wide PMC.
3194 	 *
3195 	 * Add the owner to the global list if this is a system-wide
3196 	 * sampling PMC.
3197 	 */
3198 
3199 	if (mode == PMC_MODE_SS) {
3200 		/*
3201 		 * Log mapping information for all existing processes in the
3202 		 * system.  Subsequent mappings are logged as they happen;
3203 		 * see pmc_process_mmap().
3204 		 */
3205 		if (po->po_logprocmaps == 0) {
3206 			pmc_log_all_process_mappings(po);
3207 			po->po_logprocmaps = 1;
3208 		}
3209 		po->po_sscount++;
3210 		if (po->po_sscount == 1) {
3211 			atomic_add_rel_int(&pmc_ss_count, 1);
3212 			CK_LIST_INSERT_HEAD(&pmc_ss_owners, po, po_ssnext);
3213 			PMCDBG1(PMC,OPS,1, "po=%p in global list", po);
3214 		}
3215 	}
3216 
3217 	/*
3218 	 * Move to the CPU associated with this
3219 	 * PMC, and start the hardware.
3220 	 */
3221 
3222 	pmc_save_cpu_binding(&pb);
3223 
3224 	cpu = PMC_TO_CPU(pm);
3225 
3226 	if (!pmc_cpu_is_active(cpu))
3227 		return (ENXIO);
3228 
3229 	pmc_select_cpu(cpu);
3230 
3231 	/*
3232 	 * global PMCs are configured at allocation time
3233 	 * so write out the initial value and start the PMC.
3234 	 */
3235 
3236 	pm->pm_state = PMC_STATE_RUNNING;
3237 
3238 	critical_enter();
3239 	if ((error = pcd->pcd_write_pmc(cpu, adjri,
3240 		 PMC_IS_SAMPLING_MODE(mode) ?
3241 		 pm->pm_sc.pm_reloadcount :
3242 		 pm->pm_sc.pm_initial)) == 0) {
3243 		/* If a sampling mode PMC, reset stalled state. */
3244 		if (PMC_IS_SAMPLING_MODE(mode))
3245 			pm->pm_pcpu_state[cpu].pps_stalled = 0;
3246 
3247 		/* Indicate that we desire this to run. Start it. */
3248 		pm->pm_pcpu_state[cpu].pps_cpustate = 1;
3249 		error = pcd->pcd_start_pmc(cpu, adjri);
3250 	}
3251 	critical_exit();
3252 
3253 	pmc_restore_cpu_binding(&pb);
3254 
3255 	return (error);
3256 }
3257 
3258 /*
3259  * Stop a PMC.
3260  */
3261 
3262 static int
3263 pmc_stop(struct pmc *pm)
3264 {
3265 	struct pmc_owner *po;
3266 	struct pmc_binding pb;
3267 	struct pmc_classdep *pcd;
3268 	int adjri, cpu, error, ri;
3269 
3270 	KASSERT(pm != NULL, ("[pmc,%d] null pmc", __LINE__));
3271 
3272 	PMCDBG3(PMC,OPS,1, "stop pmc=%p mode=%d ri=%d", pm,
3273 	    PMC_TO_MODE(pm), PMC_TO_ROWINDEX(pm));
3274 
3275 	pm->pm_state = PMC_STATE_STOPPED;
3276 
3277 	/*
3278 	 * If the PMC is a virtual mode one, changing the state to
3279 	 * non-RUNNING is enough to ensure that the PMC never gets
3280 	 * scheduled.
3281 	 *
3282 	 * If this PMC is current running on a CPU, then it will
3283 	 * handled correctly at the time its target process is context
3284 	 * switched out.
3285 	 */
3286 
3287 	if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)))
3288 		return 0;
3289 
3290 	/*
3291 	 * A system-mode PMC.  Move to the CPU associated with
3292 	 * this PMC, and stop the hardware.  We update the
3293 	 * 'initial count' so that a subsequent PMCSTART will
3294 	 * resume counting from the current hardware count.
3295 	 */
3296 
3297 	pmc_save_cpu_binding(&pb);
3298 
3299 	cpu = PMC_TO_CPU(pm);
3300 
3301 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
3302 	    ("[pmc,%d] illegal cpu=%d", __LINE__, cpu));
3303 
3304 	if (!pmc_cpu_is_active(cpu))
3305 		return ENXIO;
3306 
3307 	pmc_select_cpu(cpu);
3308 
3309 	ri = PMC_TO_ROWINDEX(pm);
3310 	pcd = pmc_ri_to_classdep(md, ri, &adjri);
3311 
3312 	pm->pm_pcpu_state[cpu].pps_cpustate = 0;
3313 	critical_enter();
3314 	if ((error = pcd->pcd_stop_pmc(cpu, adjri)) == 0)
3315 		error = pcd->pcd_read_pmc(cpu, adjri, &pm->pm_sc.pm_initial);
3316 	critical_exit();
3317 
3318 	pmc_restore_cpu_binding(&pb);
3319 
3320 	po = pm->pm_owner;
3321 
3322 	/* remove this owner from the global list of SS PMC owners */
3323 	if (PMC_TO_MODE(pm) == PMC_MODE_SS) {
3324 		po->po_sscount--;
3325 		if (po->po_sscount == 0) {
3326 			atomic_subtract_rel_int(&pmc_ss_count, 1);
3327 			CK_LIST_REMOVE(po, po_ssnext);
3328 			epoch_wait_preempt(global_epoch_preempt);
3329 			PMCDBG1(PMC,OPS,2,"po=%p removed from global list", po);
3330 		}
3331 	}
3332 
3333 	return (error);
3334 }
3335 
3336 static struct pmc_classdep *
3337 pmc_class_to_classdep(enum pmc_class class)
3338 {
3339 	int n;
3340 
3341 	for (n = 0; n < md->pmd_nclass; n++)
3342 		if (md->pmd_classdep[n].pcd_class == class)
3343 			return (&md->pmd_classdep[n]);
3344 	return (NULL);
3345 }
3346 
3347 #if defined(HWPMC_DEBUG) && defined(KTR)
3348 static const char *pmc_op_to_name[] = {
3349 #undef	__PMC_OP
3350 #define	__PMC_OP(N, D)	#N ,
3351 	__PMC_OPS()
3352 	NULL
3353 };
3354 #endif
3355 
3356 /*
3357  * The syscall interface
3358  */
3359 
3360 #define	PMC_GET_SX_XLOCK(...) do {		\
3361 	sx_xlock(&pmc_sx);			\
3362 	if (pmc_hook == NULL) {			\
3363 		sx_xunlock(&pmc_sx);		\
3364 		return __VA_ARGS__;		\
3365 	}					\
3366 } while (0)
3367 
3368 #define	PMC_DOWNGRADE_SX() do {			\
3369 	sx_downgrade(&pmc_sx);			\
3370 	is_sx_downgraded = 1;			\
3371 } while (0)
3372 
3373 static int
3374 pmc_syscall_handler(struct thread *td, void *syscall_args)
3375 {
3376 	int error, is_sx_downgraded, op;
3377 	struct pmc_syscall_args *c;
3378 	void *pmclog_proc_handle;
3379 	void *arg;
3380 
3381 	c = (struct pmc_syscall_args *)syscall_args;
3382 	op = c->pmop_code;
3383 	arg = c->pmop_data;
3384 	/* PMC isn't set up yet */
3385 	if (pmc_hook == NULL)
3386 		return (EINVAL);
3387 	if (op == PMC_OP_CONFIGURELOG) {
3388 		/*
3389 		 * We cannot create the logging process inside
3390 		 * pmclog_configure_log() because there is a LOR
3391 		 * between pmc_sx and process structure locks.
3392 		 * Instead, pre-create the process and ignite the loop
3393 		 * if everything is fine, otherwise direct the process
3394 		 * to exit.
3395 		 */
3396 		error = pmclog_proc_create(td, &pmclog_proc_handle);
3397 		if (error != 0)
3398 			goto done_syscall;
3399 	}
3400 
3401 	PMC_GET_SX_XLOCK(ENOSYS);
3402 	is_sx_downgraded = 0;
3403 	PMCDBG3(MOD,PMS,1, "syscall op=%d \"%s\" arg=%p", op,
3404 	    pmc_op_to_name[op], arg);
3405 
3406 	error = 0;
3407 	counter_u64_add(pmc_stats.pm_syscalls, 1);
3408 
3409 	switch (op) {
3410 
3411 
3412 	/*
3413 	 * Configure a log file.
3414 	 *
3415 	 * XXX This OP will be reworked.
3416 	 */
3417 
3418 	case PMC_OP_CONFIGURELOG:
3419 	{
3420 		struct proc *p;
3421 		struct pmc *pm;
3422 		struct pmc_owner *po;
3423 		struct pmc_op_configurelog cl;
3424 
3425 		if ((error = copyin(arg, &cl, sizeof(cl))) != 0) {
3426 			pmclog_proc_ignite(pmclog_proc_handle, NULL);
3427 			break;
3428 		}
3429 
3430 		/* mark this process as owning a log file */
3431 		p = td->td_proc;
3432 		if ((po = pmc_find_owner_descriptor(p)) == NULL)
3433 			if ((po = pmc_allocate_owner_descriptor(p)) == NULL) {
3434 				pmclog_proc_ignite(pmclog_proc_handle, NULL);
3435 				error = ENOMEM;
3436 				break;
3437 			}
3438 
3439 		/*
3440 		 * If a valid fd was passed in, try to configure that,
3441 		 * otherwise if 'fd' was less than zero and there was
3442 		 * a log file configured, flush its buffers and
3443 		 * de-configure it.
3444 		 */
3445 		if (cl.pm_logfd >= 0) {
3446 			error = pmclog_configure_log(md, po, cl.pm_logfd);
3447 			pmclog_proc_ignite(pmclog_proc_handle, error == 0 ?
3448 			    po : NULL);
3449 		} else if (po->po_flags & PMC_PO_OWNS_LOGFILE) {
3450 			pmclog_proc_ignite(pmclog_proc_handle, NULL);
3451 			error = pmclog_close(po);
3452 			if (error == 0) {
3453 				LIST_FOREACH(pm, &po->po_pmcs, pm_next)
3454 				    if (pm->pm_flags & PMC_F_NEEDS_LOGFILE &&
3455 					pm->pm_state == PMC_STATE_RUNNING)
3456 					    pmc_stop(pm);
3457 				error = pmclog_deconfigure_log(po);
3458 			}
3459 		} else {
3460 			pmclog_proc_ignite(pmclog_proc_handle, NULL);
3461 			error = EINVAL;
3462 		}
3463 	}
3464 	break;
3465 
3466 	/*
3467 	 * Flush a log file.
3468 	 */
3469 
3470 	case PMC_OP_FLUSHLOG:
3471 	{
3472 		struct pmc_owner *po;
3473 
3474 		sx_assert(&pmc_sx, SX_XLOCKED);
3475 
3476 		if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) {
3477 			error = EINVAL;
3478 			break;
3479 		}
3480 
3481 		error = pmclog_flush(po, 0);
3482 	}
3483 	break;
3484 
3485 	/*
3486 	 * Close a log file.
3487 	 */
3488 
3489 	case PMC_OP_CLOSELOG:
3490 	{
3491 		struct pmc_owner *po;
3492 
3493 		sx_assert(&pmc_sx, SX_XLOCKED);
3494 
3495 		if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) {
3496 			error = EINVAL;
3497 			break;
3498 		}
3499 
3500 		error = pmclog_close(po);
3501 	}
3502 	break;
3503 
3504 	/*
3505 	 * Retrieve hardware configuration.
3506 	 */
3507 
3508 	case PMC_OP_GETCPUINFO:	/* CPU information */
3509 	{
3510 		struct pmc_op_getcpuinfo gci;
3511 		struct pmc_classinfo *pci;
3512 		struct pmc_classdep *pcd;
3513 		int cl;
3514 
3515 		memset(&gci, 0, sizeof(gci));
3516 		gci.pm_cputype = md->pmd_cputype;
3517 		gci.pm_ncpu    = pmc_cpu_max();
3518 		gci.pm_npmc    = md->pmd_npmc;
3519 		gci.pm_nclass  = md->pmd_nclass;
3520 		pci = gci.pm_classes;
3521 		pcd = md->pmd_classdep;
3522 		for (cl = 0; cl < md->pmd_nclass; cl++, pci++, pcd++) {
3523 			pci->pm_caps  = pcd->pcd_caps;
3524 			pci->pm_class = pcd->pcd_class;
3525 			pci->pm_width = pcd->pcd_width;
3526 			pci->pm_num   = pcd->pcd_num;
3527 		}
3528 		error = copyout(&gci, arg, sizeof(gci));
3529 	}
3530 	break;
3531 
3532 	/*
3533 	 * Retrieve soft events list.
3534 	 */
3535 	case PMC_OP_GETDYNEVENTINFO:
3536 	{
3537 		enum pmc_class			cl;
3538 		enum pmc_event			ev;
3539 		struct pmc_op_getdyneventinfo	*gei;
3540 		struct pmc_dyn_event_descr	dev;
3541 		struct pmc_soft			*ps;
3542 		uint32_t			nevent;
3543 
3544 		sx_assert(&pmc_sx, SX_LOCKED);
3545 
3546 		gei = (struct pmc_op_getdyneventinfo *) arg;
3547 
3548 		if ((error = copyin(&gei->pm_class, &cl, sizeof(cl))) != 0)
3549 			break;
3550 
3551 		/* Only SOFT class is dynamic. */
3552 		if (cl != PMC_CLASS_SOFT) {
3553 			error = EINVAL;
3554 			break;
3555 		}
3556 
3557 		nevent = 0;
3558 		for (ev = PMC_EV_SOFT_FIRST; (int)ev <= PMC_EV_SOFT_LAST; ev++) {
3559 			ps = pmc_soft_ev_acquire(ev);
3560 			if (ps == NULL)
3561 				continue;
3562 			bcopy(&ps->ps_ev, &dev, sizeof(dev));
3563 			pmc_soft_ev_release(ps);
3564 
3565 			error = copyout(&dev,
3566 			    &gei->pm_events[nevent],
3567 			    sizeof(struct pmc_dyn_event_descr));
3568 			if (error != 0)
3569 				break;
3570 			nevent++;
3571 		}
3572 		if (error != 0)
3573 			break;
3574 
3575 		error = copyout(&nevent, &gei->pm_nevent,
3576 		    sizeof(nevent));
3577 	}
3578 	break;
3579 
3580 	/*
3581 	 * Get module statistics
3582 	 */
3583 
3584 	case PMC_OP_GETDRIVERSTATS:
3585 	{
3586 		struct pmc_op_getdriverstats gms;
3587 #define CFETCH(a, b, field) a.field = counter_u64_fetch(b.field)
3588 		CFETCH(gms, pmc_stats, pm_intr_ignored);
3589 		CFETCH(gms, pmc_stats, pm_intr_processed);
3590 		CFETCH(gms, pmc_stats, pm_intr_bufferfull);
3591 		CFETCH(gms, pmc_stats, pm_syscalls);
3592 		CFETCH(gms, pmc_stats, pm_syscall_errors);
3593 		CFETCH(gms, pmc_stats, pm_buffer_requests);
3594 		CFETCH(gms, pmc_stats, pm_buffer_requests_failed);
3595 		CFETCH(gms, pmc_stats, pm_log_sweeps);
3596 #undef CFETCH
3597 		error = copyout(&gms, arg, sizeof(gms));
3598 	}
3599 	break;
3600 
3601 
3602 	/*
3603 	 * Retrieve module version number
3604 	 */
3605 
3606 	case PMC_OP_GETMODULEVERSION:
3607 	{
3608 		uint32_t cv, modv;
3609 
3610 		/* retrieve the client's idea of the ABI version */
3611 		if ((error = copyin(arg, &cv, sizeof(uint32_t))) != 0)
3612 			break;
3613 		/* don't service clients newer than our driver */
3614 		modv = PMC_VERSION;
3615 		if ((cv & 0xFFFF0000) > (modv & 0xFFFF0000)) {
3616 			error = EPROGMISMATCH;
3617 			break;
3618 		}
3619 		error = copyout(&modv, arg, sizeof(int));
3620 	}
3621 	break;
3622 
3623 
3624 	/*
3625 	 * Retrieve the state of all the PMCs on a given
3626 	 * CPU.
3627 	 */
3628 
3629 	case PMC_OP_GETPMCINFO:
3630 	{
3631 		int ari;
3632 		struct pmc *pm;
3633 		size_t pmcinfo_size;
3634 		uint32_t cpu, n, npmc;
3635 		struct pmc_owner *po;
3636 		struct pmc_binding pb;
3637 		struct pmc_classdep *pcd;
3638 		struct pmc_info *p, *pmcinfo;
3639 		struct pmc_op_getpmcinfo *gpi;
3640 
3641 		PMC_DOWNGRADE_SX();
3642 
3643 		gpi = (struct pmc_op_getpmcinfo *) arg;
3644 
3645 		if ((error = copyin(&gpi->pm_cpu, &cpu, sizeof(cpu))) != 0)
3646 			break;
3647 
3648 		if (cpu >= pmc_cpu_max()) {
3649 			error = EINVAL;
3650 			break;
3651 		}
3652 
3653 		if (!pmc_cpu_is_active(cpu)) {
3654 			error = ENXIO;
3655 			break;
3656 		}
3657 
3658 		/* switch to CPU 'cpu' */
3659 		pmc_save_cpu_binding(&pb);
3660 		pmc_select_cpu(cpu);
3661 
3662 		npmc = md->pmd_npmc;
3663 
3664 		pmcinfo_size = npmc * sizeof(struct pmc_info);
3665 		pmcinfo = malloc(pmcinfo_size, M_PMC, M_WAITOK | M_ZERO);
3666 
3667 		p = pmcinfo;
3668 
3669 		for (n = 0; n < md->pmd_npmc; n++, p++) {
3670 
3671 			pcd = pmc_ri_to_classdep(md, n, &ari);
3672 
3673 			KASSERT(pcd != NULL,
3674 			    ("[pmc,%d] null pcd ri=%d", __LINE__, n));
3675 
3676 			if ((error = pcd->pcd_describe(cpu, ari, p, &pm)) != 0)
3677 				break;
3678 
3679 			if (PMC_ROW_DISP_IS_STANDALONE(n))
3680 				p->pm_rowdisp = PMC_DISP_STANDALONE;
3681 			else if (PMC_ROW_DISP_IS_THREAD(n))
3682 				p->pm_rowdisp = PMC_DISP_THREAD;
3683 			else
3684 				p->pm_rowdisp = PMC_DISP_FREE;
3685 
3686 			p->pm_ownerpid = -1;
3687 
3688 			if (pm == NULL)	/* no PMC associated */
3689 				continue;
3690 
3691 			po = pm->pm_owner;
3692 
3693 			KASSERT(po->po_owner != NULL,
3694 			    ("[pmc,%d] pmc_owner had a null proc pointer",
3695 				__LINE__));
3696 
3697 			p->pm_ownerpid = po->po_owner->p_pid;
3698 			p->pm_mode     = PMC_TO_MODE(pm);
3699 			p->pm_event    = pm->pm_event;
3700 			p->pm_flags    = pm->pm_flags;
3701 
3702 			if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
3703 				p->pm_reloadcount =
3704 				    pm->pm_sc.pm_reloadcount;
3705 		}
3706 
3707 		pmc_restore_cpu_binding(&pb);
3708 
3709 		/* now copy out the PMC info collected */
3710 		if (error == 0)
3711 			error = copyout(pmcinfo, &gpi->pm_pmcs, pmcinfo_size);
3712 
3713 		free(pmcinfo, M_PMC);
3714 	}
3715 	break;
3716 
3717 
3718 	/*
3719 	 * Set the administrative state of a PMC.  I.e. whether
3720 	 * the PMC is to be used or not.
3721 	 */
3722 
3723 	case PMC_OP_PMCADMIN:
3724 	{
3725 		int cpu, ri;
3726 		enum pmc_state request;
3727 		struct pmc_cpu *pc;
3728 		struct pmc_hw *phw;
3729 		struct pmc_op_pmcadmin pma;
3730 		struct pmc_binding pb;
3731 
3732 		sx_assert(&pmc_sx, SX_XLOCKED);
3733 
3734 		KASSERT(td == curthread,
3735 		    ("[pmc,%d] td != curthread", __LINE__));
3736 
3737 		error = priv_check(td, PRIV_PMC_MANAGE);
3738 		if (error)
3739 			break;
3740 
3741 		if ((error = copyin(arg, &pma, sizeof(pma))) != 0)
3742 			break;
3743 
3744 		cpu = pma.pm_cpu;
3745 
3746 		if (cpu < 0 || cpu >= (int) pmc_cpu_max()) {
3747 			error = EINVAL;
3748 			break;
3749 		}
3750 
3751 		if (!pmc_cpu_is_active(cpu)) {
3752 			error = ENXIO;
3753 			break;
3754 		}
3755 
3756 		request = pma.pm_state;
3757 
3758 		if (request != PMC_STATE_DISABLED &&
3759 		    request != PMC_STATE_FREE) {
3760 			error = EINVAL;
3761 			break;
3762 		}
3763 
3764 		ri = pma.pm_pmc; /* pmc id == row index */
3765 		if (ri < 0 || ri >= (int) md->pmd_npmc) {
3766 			error = EINVAL;
3767 			break;
3768 		}
3769 
3770 		/*
3771 		 * We can't disable a PMC with a row-index allocated
3772 		 * for process virtual PMCs.
3773 		 */
3774 
3775 		if (PMC_ROW_DISP_IS_THREAD(ri) &&
3776 		    request == PMC_STATE_DISABLED) {
3777 			error = EBUSY;
3778 			break;
3779 		}
3780 
3781 		/*
3782 		 * otherwise, this PMC on this CPU is either free or
3783 		 * in system-wide mode.
3784 		 */
3785 
3786 		pmc_save_cpu_binding(&pb);
3787 		pmc_select_cpu(cpu);
3788 
3789 		pc  = pmc_pcpu[cpu];
3790 		phw = pc->pc_hwpmcs[ri];
3791 
3792 		/*
3793 		 * XXX do we need some kind of 'forced' disable?
3794 		 */
3795 
3796 		if (phw->phw_pmc == NULL) {
3797 			if (request == PMC_STATE_DISABLED &&
3798 			    (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED)) {
3799 				phw->phw_state &= ~PMC_PHW_FLAG_IS_ENABLED;
3800 				PMC_MARK_ROW_STANDALONE(ri);
3801 			} else if (request == PMC_STATE_FREE &&
3802 			    (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0) {
3803 				phw->phw_state |=  PMC_PHW_FLAG_IS_ENABLED;
3804 				PMC_UNMARK_ROW_STANDALONE(ri);
3805 			}
3806 			/* other cases are a no-op */
3807 		} else
3808 			error = EBUSY;
3809 
3810 		pmc_restore_cpu_binding(&pb);
3811 	}
3812 	break;
3813 
3814 
3815 	/*
3816 	 * Allocate a PMC.
3817 	 */
3818 
3819 	case PMC_OP_PMCALLOCATE:
3820 	{
3821 		int adjri, n;
3822 		u_int cpu;
3823 		uint32_t caps;
3824 		struct pmc *pmc;
3825 		enum pmc_mode mode;
3826 		struct pmc_hw *phw;
3827 		struct pmc_binding pb;
3828 		struct pmc_classdep *pcd;
3829 		struct pmc_op_pmcallocate pa;
3830 
3831 		if ((error = copyin(arg, &pa, sizeof(pa))) != 0)
3832 			break;
3833 
3834 		caps = pa.pm_caps;
3835 		mode = pa.pm_mode;
3836 		cpu  = pa.pm_cpu;
3837 
3838 		if ((mode != PMC_MODE_SS  &&  mode != PMC_MODE_SC  &&
3839 		     mode != PMC_MODE_TS  &&  mode != PMC_MODE_TC) ||
3840 		    (cpu != (u_int) PMC_CPU_ANY && cpu >= pmc_cpu_max())) {
3841 			error = EINVAL;
3842 			break;
3843 		}
3844 
3845 		/*
3846 		 * Virtual PMCs should only ask for a default CPU.
3847 		 * System mode PMCs need to specify a non-default CPU.
3848 		 */
3849 
3850 		if ((PMC_IS_VIRTUAL_MODE(mode) && cpu != (u_int) PMC_CPU_ANY) ||
3851 		    (PMC_IS_SYSTEM_MODE(mode) && cpu == (u_int) PMC_CPU_ANY)) {
3852 			error = EINVAL;
3853 			break;
3854 		}
3855 
3856 		/*
3857 		 * Check that an inactive CPU is not being asked for.
3858 		 */
3859 
3860 		if (PMC_IS_SYSTEM_MODE(mode) && !pmc_cpu_is_active(cpu)) {
3861 			error = ENXIO;
3862 			break;
3863 		}
3864 
3865 		/*
3866 		 * Refuse an allocation for a system-wide PMC if this
3867 		 * process has been jailed, or if this process lacks
3868 		 * super-user credentials and the sysctl tunable
3869 		 * 'security.bsd.unprivileged_syspmcs' is zero.
3870 		 */
3871 
3872 		if (PMC_IS_SYSTEM_MODE(mode)) {
3873 			if (jailed(curthread->td_ucred)) {
3874 				error = EPERM;
3875 				break;
3876 			}
3877 			if (!pmc_unprivileged_syspmcs) {
3878 				error = priv_check(curthread,
3879 				    PRIV_PMC_SYSTEM);
3880 				if (error)
3881 					break;
3882 			}
3883 		}
3884 
3885 		/*
3886 		 * Look for valid values for 'pm_flags'
3887 		 */
3888 
3889 		if ((pa.pm_flags & ~(PMC_F_DESCENDANTS | PMC_F_LOG_PROCCSW |
3890 		    PMC_F_LOG_PROCEXIT | PMC_F_CALLCHAIN |
3891 		    PMC_F_USERCALLCHAIN)) != 0) {
3892 			error = EINVAL;
3893 			break;
3894 		}
3895 
3896 		/* PMC_F_USERCALLCHAIN is only valid with PMC_F_CALLCHAIN */
3897 		if ((pa.pm_flags & (PMC_F_CALLCHAIN | PMC_F_USERCALLCHAIN)) ==
3898 		    PMC_F_USERCALLCHAIN) {
3899 			error = EINVAL;
3900 			break;
3901 		}
3902 
3903 		/* PMC_F_USERCALLCHAIN is only valid for sampling mode */
3904 		if (pa.pm_flags & PMC_F_USERCALLCHAIN &&
3905 			mode != PMC_MODE_TS && mode != PMC_MODE_SS) {
3906 			error = EINVAL;
3907 			break;
3908 		}
3909 
3910 		/* process logging options are not allowed for system PMCs */
3911 		if (PMC_IS_SYSTEM_MODE(mode) && (pa.pm_flags &
3912 		    (PMC_F_LOG_PROCCSW | PMC_F_LOG_PROCEXIT))) {
3913 			error = EINVAL;
3914 			break;
3915 		}
3916 
3917 		/*
3918 		 * All sampling mode PMCs need to be able to interrupt the
3919 		 * CPU.
3920 		 */
3921 		if (PMC_IS_SAMPLING_MODE(mode))
3922 			caps |= PMC_CAP_INTERRUPT;
3923 
3924 		/* A valid class specifier should have been passed in. */
3925 		pcd = pmc_class_to_classdep(pa.pm_class);
3926 		if (pcd == NULL) {
3927 			error = EINVAL;
3928 			break;
3929 		}
3930 
3931 		/* The requested PMC capabilities should be feasible. */
3932 		if ((pcd->pcd_caps & caps) != caps) {
3933 			error = EOPNOTSUPP;
3934 			break;
3935 		}
3936 
3937 		PMCDBG4(PMC,ALL,2, "event=%d caps=0x%x mode=%d cpu=%d",
3938 		    pa.pm_ev, caps, mode, cpu);
3939 
3940 		pmc = pmc_allocate_pmc_descriptor();
3941 		pmc->pm_id    = PMC_ID_MAKE_ID(cpu,pa.pm_mode,pa.pm_class,
3942 		    PMC_ID_INVALID);
3943 		pmc->pm_event = pa.pm_ev;
3944 		pmc->pm_state = PMC_STATE_FREE;
3945 		pmc->pm_caps  = caps;
3946 		pmc->pm_flags = pa.pm_flags;
3947 
3948 		/* XXX set lower bound on sampling for process counters */
3949 		if (PMC_IS_SAMPLING_MODE(mode)) {
3950 			/*
3951 			 * Don't permit requested sample rate to be less than 1000
3952 			 */
3953 			if (pa.pm_count < 1000)
3954 				log(LOG_WARNING,
3955 					"pmcallocate: passed sample rate %ju - setting to 1000\n",
3956 					(uintmax_t)pa.pm_count);
3957 			pmc->pm_sc.pm_reloadcount = MAX(1000, pa.pm_count);
3958 		} else
3959 			pmc->pm_sc.pm_initial = pa.pm_count;
3960 
3961 		/* switch thread to CPU 'cpu' */
3962 		pmc_save_cpu_binding(&pb);
3963 
3964 #define	PMC_IS_SHAREABLE_PMC(cpu, n)				\
3965 	(pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_state &		\
3966 	 PMC_PHW_FLAG_IS_SHAREABLE)
3967 #define	PMC_IS_UNALLOCATED(cpu, n)				\
3968 	(pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_pmc == NULL)
3969 
3970 		if (PMC_IS_SYSTEM_MODE(mode)) {
3971 			pmc_select_cpu(cpu);
3972 			for (n = pcd->pcd_ri; n < (int) md->pmd_npmc; n++) {
3973 				pcd = pmc_ri_to_classdep(md, n, &adjri);
3974 				if (pmc_can_allocate_row(n, mode) == 0 &&
3975 				    pmc_can_allocate_rowindex(
3976 					    curthread->td_proc, n, cpu) == 0 &&
3977 				    (PMC_IS_UNALLOCATED(cpu, n) ||
3978 				     PMC_IS_SHAREABLE_PMC(cpu, n)) &&
3979 				    pcd->pcd_allocate_pmc(cpu, adjri, pmc,
3980 					&pa) == 0)
3981 					break;
3982 			}
3983 		} else {
3984 			/* Process virtual mode */
3985 			for (n = pcd->pcd_ri; n < (int) md->pmd_npmc; n++) {
3986 				pcd = pmc_ri_to_classdep(md, n, &adjri);
3987 				if (pmc_can_allocate_row(n, mode) == 0 &&
3988 				    pmc_can_allocate_rowindex(
3989 					    curthread->td_proc, n,
3990 					    PMC_CPU_ANY) == 0 &&
3991 				    pcd->pcd_allocate_pmc(curthread->td_oncpu,
3992 					adjri, pmc, &pa) == 0)
3993 					break;
3994 			}
3995 		}
3996 
3997 #undef	PMC_IS_UNALLOCATED
3998 #undef	PMC_IS_SHAREABLE_PMC
3999 
4000 		pmc_restore_cpu_binding(&pb);
4001 
4002 		if (n == (int) md->pmd_npmc) {
4003 			pmc_destroy_pmc_descriptor(pmc);
4004 			pmc = NULL;
4005 			error = EINVAL;
4006 			break;
4007 		}
4008 
4009 		/* Fill in the correct value in the ID field */
4010 		pmc->pm_id = PMC_ID_MAKE_ID(cpu,mode,pa.pm_class,n);
4011 
4012 		PMCDBG5(PMC,ALL,2, "ev=%d class=%d mode=%d n=%d -> pmcid=%x",
4013 		    pmc->pm_event, pa.pm_class, mode, n, pmc->pm_id);
4014 
4015 		/* Process mode PMCs with logging enabled need log files */
4016 		if (pmc->pm_flags & (PMC_F_LOG_PROCEXIT | PMC_F_LOG_PROCCSW))
4017 			pmc->pm_flags |= PMC_F_NEEDS_LOGFILE;
4018 
4019 		/* All system mode sampling PMCs require a log file */
4020 		if (PMC_IS_SAMPLING_MODE(mode) && PMC_IS_SYSTEM_MODE(mode))
4021 			pmc->pm_flags |= PMC_F_NEEDS_LOGFILE;
4022 
4023 		/*
4024 		 * Configure global pmc's immediately
4025 		 */
4026 
4027 		if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pmc))) {
4028 
4029 			pmc_save_cpu_binding(&pb);
4030 			pmc_select_cpu(cpu);
4031 
4032 			phw = pmc_pcpu[cpu]->pc_hwpmcs[n];
4033 			pcd = pmc_ri_to_classdep(md, n, &adjri);
4034 
4035 			if ((phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0 ||
4036 			    (error = pcd->pcd_config_pmc(cpu, adjri, pmc)) != 0) {
4037 				(void) pcd->pcd_release_pmc(cpu, adjri, pmc);
4038 				pmc_destroy_pmc_descriptor(pmc);
4039 				pmc = NULL;
4040 				pmc_restore_cpu_binding(&pb);
4041 				error = EPERM;
4042 				break;
4043 			}
4044 
4045 			pmc_restore_cpu_binding(&pb);
4046 		}
4047 
4048 		pmc->pm_state    = PMC_STATE_ALLOCATED;
4049 		pmc->pm_class	= pa.pm_class;
4050 
4051 		/*
4052 		 * mark row disposition
4053 		 */
4054 
4055 		if (PMC_IS_SYSTEM_MODE(mode))
4056 			PMC_MARK_ROW_STANDALONE(n);
4057 		else
4058 			PMC_MARK_ROW_THREAD(n);
4059 
4060 		/*
4061 		 * Register this PMC with the current thread as its owner.
4062 		 */
4063 
4064 		if ((error =
4065 		    pmc_register_owner(curthread->td_proc, pmc)) != 0) {
4066 			pmc_release_pmc_descriptor(pmc);
4067 			pmc_destroy_pmc_descriptor(pmc);
4068 			pmc = NULL;
4069 			break;
4070 		}
4071 
4072 
4073 		/*
4074 		 * Return the allocated index.
4075 		 */
4076 
4077 		pa.pm_pmcid = pmc->pm_id;
4078 
4079 		error = copyout(&pa, arg, sizeof(pa));
4080 	}
4081 	break;
4082 
4083 
4084 	/*
4085 	 * Attach a PMC to a process.
4086 	 */
4087 
4088 	case PMC_OP_PMCATTACH:
4089 	{
4090 		struct pmc *pm;
4091 		struct proc *p;
4092 		struct pmc_op_pmcattach a;
4093 
4094 		sx_assert(&pmc_sx, SX_XLOCKED);
4095 
4096 		if ((error = copyin(arg, &a, sizeof(a))) != 0)
4097 			break;
4098 
4099 		if (a.pm_pid < 0) {
4100 			error = EINVAL;
4101 			break;
4102 		} else if (a.pm_pid == 0)
4103 			a.pm_pid = td->td_proc->p_pid;
4104 
4105 		if ((error = pmc_find_pmc(a.pm_pmc, &pm)) != 0)
4106 			break;
4107 
4108 		if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm))) {
4109 			error = EINVAL;
4110 			break;
4111 		}
4112 
4113 		/* PMCs may be (re)attached only when allocated or stopped */
4114 		if (pm->pm_state == PMC_STATE_RUNNING) {
4115 			error = EBUSY;
4116 			break;
4117 		} else if (pm->pm_state != PMC_STATE_ALLOCATED &&
4118 		    pm->pm_state != PMC_STATE_STOPPED) {
4119 			error = EINVAL;
4120 			break;
4121 		}
4122 
4123 		/* lookup pid */
4124 		if ((p = pfind(a.pm_pid)) == NULL) {
4125 			error = ESRCH;
4126 			break;
4127 		}
4128 
4129 		/*
4130 		 * Ignore processes that are working on exiting.
4131 		 */
4132 		if (p->p_flag & P_WEXIT) {
4133 			error = ESRCH;
4134 			PROC_UNLOCK(p);	/* pfind() returns a locked process */
4135 			break;
4136 		}
4137 
4138 		/*
4139 		 * we are allowed to attach a PMC to a process if
4140 		 * we can debug it.
4141 		 */
4142 		error = p_candebug(curthread, p);
4143 
4144 		PROC_UNLOCK(p);
4145 
4146 		if (error == 0)
4147 			error = pmc_attach_process(p, pm);
4148 	}
4149 	break;
4150 
4151 
4152 	/*
4153 	 * Detach an attached PMC from a process.
4154 	 */
4155 
4156 	case PMC_OP_PMCDETACH:
4157 	{
4158 		struct pmc *pm;
4159 		struct proc *p;
4160 		struct pmc_op_pmcattach a;
4161 
4162 		if ((error = copyin(arg, &a, sizeof(a))) != 0)
4163 			break;
4164 
4165 		if (a.pm_pid < 0) {
4166 			error = EINVAL;
4167 			break;
4168 		} else if (a.pm_pid == 0)
4169 			a.pm_pid = td->td_proc->p_pid;
4170 
4171 		if ((error = pmc_find_pmc(a.pm_pmc, &pm)) != 0)
4172 			break;
4173 
4174 		if ((p = pfind(a.pm_pid)) == NULL) {
4175 			error = ESRCH;
4176 			break;
4177 		}
4178 
4179 		/*
4180 		 * Treat processes that are in the process of exiting
4181 		 * as if they were not present.
4182 		 */
4183 
4184 		if (p->p_flag & P_WEXIT)
4185 			error = ESRCH;
4186 
4187 		PROC_UNLOCK(p);	/* pfind() returns a locked process */
4188 
4189 		if (error == 0)
4190 			error = pmc_detach_process(p, pm);
4191 	}
4192 	break;
4193 
4194 
4195 	/*
4196 	 * Retrieve the MSR number associated with the counter
4197 	 * 'pmc_id'.  This allows processes to directly use RDPMC
4198 	 * instructions to read their PMCs, without the overhead of a
4199 	 * system call.
4200 	 */
4201 
4202 	case PMC_OP_PMCGETMSR:
4203 	{
4204 		int adjri, ri;
4205 		struct pmc *pm;
4206 		struct pmc_target *pt;
4207 		struct pmc_op_getmsr gm;
4208 		struct pmc_classdep *pcd;
4209 
4210 		PMC_DOWNGRADE_SX();
4211 
4212 		if ((error = copyin(arg, &gm, sizeof(gm))) != 0)
4213 			break;
4214 
4215 		if ((error = pmc_find_pmc(gm.pm_pmcid, &pm)) != 0)
4216 			break;
4217 
4218 		/*
4219 		 * The allocated PMC has to be a process virtual PMC,
4220 		 * i.e., of type MODE_T[CS].  Global PMCs can only be
4221 		 * read using the PMCREAD operation since they may be
4222 		 * allocated on a different CPU than the one we could
4223 		 * be running on at the time of the RDPMC instruction.
4224 		 *
4225 		 * The GETMSR operation is not allowed for PMCs that
4226 		 * are inherited across processes.
4227 		 */
4228 
4229 		if (!PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)) ||
4230 		    (pm->pm_flags & PMC_F_DESCENDANTS)) {
4231 			error = EINVAL;
4232 			break;
4233 		}
4234 
4235 		/*
4236 		 * It only makes sense to use a RDPMC (or its
4237 		 * equivalent instruction on non-x86 architectures) on
4238 		 * a process that has allocated and attached a PMC to
4239 		 * itself.  Conversely the PMC is only allowed to have
4240 		 * one process attached to it -- its owner.
4241 		 */
4242 
4243 		if ((pt = LIST_FIRST(&pm->pm_targets)) == NULL ||
4244 		    LIST_NEXT(pt, pt_next) != NULL ||
4245 		    pt->pt_process->pp_proc != pm->pm_owner->po_owner) {
4246 			error = EINVAL;
4247 			break;
4248 		}
4249 
4250 		ri = PMC_TO_ROWINDEX(pm);
4251 		pcd = pmc_ri_to_classdep(md, ri, &adjri);
4252 
4253 		/* PMC class has no 'GETMSR' support */
4254 		if (pcd->pcd_get_msr == NULL) {
4255 			error = ENOSYS;
4256 			break;
4257 		}
4258 
4259 		if ((error = (*pcd->pcd_get_msr)(adjri, &gm.pm_msr)) < 0)
4260 			break;
4261 
4262 		if ((error = copyout(&gm, arg, sizeof(gm))) < 0)
4263 			break;
4264 
4265 		/*
4266 		 * Mark our process as using MSRs.  Update machine
4267 		 * state using a forced context switch.
4268 		 */
4269 
4270 		pt->pt_process->pp_flags |= PMC_PP_ENABLE_MSR_ACCESS;
4271 		pmc_force_context_switch();
4272 
4273 	}
4274 	break;
4275 
4276 	/*
4277 	 * Release an allocated PMC
4278 	 */
4279 
4280 	case PMC_OP_PMCRELEASE:
4281 	{
4282 		pmc_id_t pmcid;
4283 		struct pmc *pm;
4284 		struct pmc_owner *po;
4285 		struct pmc_op_simple sp;
4286 
4287 		/*
4288 		 * Find PMC pointer for the named PMC.
4289 		 *
4290 		 * Use pmc_release_pmc_descriptor() to switch off the
4291 		 * PMC, remove all its target threads, and remove the
4292 		 * PMC from its owner's list.
4293 		 *
4294 		 * Remove the owner record if this is the last PMC
4295 		 * owned.
4296 		 *
4297 		 * Free up space.
4298 		 */
4299 
4300 		if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
4301 			break;
4302 
4303 		pmcid = sp.pm_pmcid;
4304 
4305 		if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
4306 			break;
4307 
4308 		po = pm->pm_owner;
4309 		pmc_release_pmc_descriptor(pm);
4310 		pmc_maybe_remove_owner(po);
4311 		pmc_destroy_pmc_descriptor(pm);
4312 	}
4313 	break;
4314 
4315 
4316 	/*
4317 	 * Read and/or write a PMC.
4318 	 */
4319 
4320 	case PMC_OP_PMCRW:
4321 	{
4322 		int adjri;
4323 		struct pmc *pm;
4324 		uint32_t cpu, ri;
4325 		pmc_value_t oldvalue;
4326 		struct pmc_binding pb;
4327 		struct pmc_op_pmcrw prw;
4328 		struct pmc_classdep *pcd;
4329 		struct pmc_op_pmcrw *pprw;
4330 
4331 		PMC_DOWNGRADE_SX();
4332 
4333 		if ((error = copyin(arg, &prw, sizeof(prw))) != 0)
4334 			break;
4335 
4336 		ri = 0;
4337 		PMCDBG2(PMC,OPS,1, "rw id=%d flags=0x%x", prw.pm_pmcid,
4338 		    prw.pm_flags);
4339 
4340 		/* must have at least one flag set */
4341 		if ((prw.pm_flags & (PMC_F_OLDVALUE|PMC_F_NEWVALUE)) == 0) {
4342 			error = EINVAL;
4343 			break;
4344 		}
4345 
4346 		/* locate pmc descriptor */
4347 		if ((error = pmc_find_pmc(prw.pm_pmcid, &pm)) != 0)
4348 			break;
4349 
4350 		/* Can't read a PMC that hasn't been started. */
4351 		if (pm->pm_state != PMC_STATE_ALLOCATED &&
4352 		    pm->pm_state != PMC_STATE_STOPPED &&
4353 		    pm->pm_state != PMC_STATE_RUNNING) {
4354 			error = EINVAL;
4355 			break;
4356 		}
4357 
4358 		/* writing a new value is allowed only for 'STOPPED' pmcs */
4359 		if (pm->pm_state == PMC_STATE_RUNNING &&
4360 		    (prw.pm_flags & PMC_F_NEWVALUE)) {
4361 			error = EBUSY;
4362 			break;
4363 		}
4364 
4365 		if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) {
4366 
4367 			/*
4368 			 * If this PMC is attached to its owner (i.e.,
4369 			 * the process requesting this operation) and
4370 			 * is running, then attempt to get an
4371 			 * upto-date reading from hardware for a READ.
4372 			 * Writes are only allowed when the PMC is
4373 			 * stopped, so only update the saved value
4374 			 * field.
4375 			 *
4376 			 * If the PMC is not running, or is not
4377 			 * attached to its owner, read/write to the
4378 			 * savedvalue field.
4379 			 */
4380 
4381 			ri = PMC_TO_ROWINDEX(pm);
4382 			pcd = pmc_ri_to_classdep(md, ri, &adjri);
4383 
4384 			mtx_pool_lock_spin(pmc_mtxpool, pm);
4385 			cpu = curthread->td_oncpu;
4386 
4387 			if (prw.pm_flags & PMC_F_OLDVALUE) {
4388 				if ((pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) &&
4389 				    (pm->pm_state == PMC_STATE_RUNNING))
4390 					error = (*pcd->pcd_read_pmc)(cpu, adjri,
4391 					    &oldvalue);
4392 				else
4393 					oldvalue = pm->pm_gv.pm_savedvalue;
4394 			}
4395 			if (prw.pm_flags & PMC_F_NEWVALUE)
4396 				pm->pm_gv.pm_savedvalue = prw.pm_value;
4397 
4398 			mtx_pool_unlock_spin(pmc_mtxpool, pm);
4399 
4400 		} else { /* System mode PMCs */
4401 			cpu = PMC_TO_CPU(pm);
4402 			ri  = PMC_TO_ROWINDEX(pm);
4403 			pcd = pmc_ri_to_classdep(md, ri, &adjri);
4404 
4405 			if (!pmc_cpu_is_active(cpu)) {
4406 				error = ENXIO;
4407 				break;
4408 			}
4409 
4410 			/* move this thread to CPU 'cpu' */
4411 			pmc_save_cpu_binding(&pb);
4412 			pmc_select_cpu(cpu);
4413 
4414 			critical_enter();
4415 			/* save old value */
4416 			if (prw.pm_flags & PMC_F_OLDVALUE)
4417 				if ((error = (*pcd->pcd_read_pmc)(cpu, adjri,
4418 					 &oldvalue)))
4419 					goto error;
4420 			/* write out new value */
4421 			if (prw.pm_flags & PMC_F_NEWVALUE)
4422 				error = (*pcd->pcd_write_pmc)(cpu, adjri,
4423 				    prw.pm_value);
4424 		error:
4425 			critical_exit();
4426 			pmc_restore_cpu_binding(&pb);
4427 			if (error)
4428 				break;
4429 		}
4430 
4431 		pprw = (struct pmc_op_pmcrw *) arg;
4432 
4433 #ifdef	HWPMC_DEBUG
4434 		if (prw.pm_flags & PMC_F_NEWVALUE)
4435 			PMCDBG3(PMC,OPS,2, "rw id=%d new %jx -> old %jx",
4436 			    ri, prw.pm_value, oldvalue);
4437 		else if (prw.pm_flags & PMC_F_OLDVALUE)
4438 			PMCDBG2(PMC,OPS,2, "rw id=%d -> old %jx", ri, oldvalue);
4439 #endif
4440 
4441 		/* return old value if requested */
4442 		if (prw.pm_flags & PMC_F_OLDVALUE)
4443 			if ((error = copyout(&oldvalue, &pprw->pm_value,
4444 				 sizeof(prw.pm_value))))
4445 				break;
4446 
4447 	}
4448 	break;
4449 
4450 
4451 	/*
4452 	 * Set the sampling rate for a sampling mode PMC and the
4453 	 * initial count for a counting mode PMC.
4454 	 */
4455 
4456 	case PMC_OP_PMCSETCOUNT:
4457 	{
4458 		struct pmc *pm;
4459 		struct pmc_op_pmcsetcount sc;
4460 
4461 		PMC_DOWNGRADE_SX();
4462 
4463 		if ((error = copyin(arg, &sc, sizeof(sc))) != 0)
4464 			break;
4465 
4466 		if ((error = pmc_find_pmc(sc.pm_pmcid, &pm)) != 0)
4467 			break;
4468 
4469 		if (pm->pm_state == PMC_STATE_RUNNING) {
4470 			error = EBUSY;
4471 			break;
4472 		}
4473 
4474 		if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) {
4475 			/*
4476 			 * Don't permit requested sample rate to be less than 1000
4477 			 */
4478 			if (sc.pm_count < 1000)
4479 				log(LOG_WARNING,
4480 					"pmcsetcount: passed sample rate %ju - setting to 1000\n",
4481 					(uintmax_t)sc.pm_count);
4482 			pm->pm_sc.pm_reloadcount = MAX(1000, sc.pm_count);
4483 		} else
4484 			pm->pm_sc.pm_initial = sc.pm_count;
4485 	}
4486 	break;
4487 
4488 
4489 	/*
4490 	 * Start a PMC.
4491 	 */
4492 
4493 	case PMC_OP_PMCSTART:
4494 	{
4495 		pmc_id_t pmcid;
4496 		struct pmc *pm;
4497 		struct pmc_op_simple sp;
4498 
4499 		sx_assert(&pmc_sx, SX_XLOCKED);
4500 
4501 		if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
4502 			break;
4503 
4504 		pmcid = sp.pm_pmcid;
4505 
4506 		if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
4507 			break;
4508 
4509 		KASSERT(pmcid == pm->pm_id,
4510 		    ("[pmc,%d] pmcid %x != id %x", __LINE__,
4511 			pm->pm_id, pmcid));
4512 
4513 		if (pm->pm_state == PMC_STATE_RUNNING) /* already running */
4514 			break;
4515 		else if (pm->pm_state != PMC_STATE_STOPPED &&
4516 		    pm->pm_state != PMC_STATE_ALLOCATED) {
4517 			error = EINVAL;
4518 			break;
4519 		}
4520 
4521 		error = pmc_start(pm);
4522 	}
4523 	break;
4524 
4525 
4526 	/*
4527 	 * Stop a PMC.
4528 	 */
4529 
4530 	case PMC_OP_PMCSTOP:
4531 	{
4532 		pmc_id_t pmcid;
4533 		struct pmc *pm;
4534 		struct pmc_op_simple sp;
4535 
4536 		PMC_DOWNGRADE_SX();
4537 
4538 		if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
4539 			break;
4540 
4541 		pmcid = sp.pm_pmcid;
4542 
4543 		/*
4544 		 * Mark the PMC as inactive and invoke the MD stop
4545 		 * routines if needed.
4546 		 */
4547 
4548 		if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
4549 			break;
4550 
4551 		KASSERT(pmcid == pm->pm_id,
4552 		    ("[pmc,%d] pmc id %x != pmcid %x", __LINE__,
4553 			pm->pm_id, pmcid));
4554 
4555 		if (pm->pm_state == PMC_STATE_STOPPED) /* already stopped */
4556 			break;
4557 		else if (pm->pm_state != PMC_STATE_RUNNING) {
4558 			error = EINVAL;
4559 			break;
4560 		}
4561 
4562 		error = pmc_stop(pm);
4563 	}
4564 	break;
4565 
4566 
4567 	/*
4568 	 * Write a user supplied value to the log file.
4569 	 */
4570 
4571 	case PMC_OP_WRITELOG:
4572 	{
4573 		struct pmc_op_writelog wl;
4574 		struct pmc_owner *po;
4575 
4576 		PMC_DOWNGRADE_SX();
4577 
4578 		if ((error = copyin(arg, &wl, sizeof(wl))) != 0)
4579 			break;
4580 
4581 		if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) {
4582 			error = EINVAL;
4583 			break;
4584 		}
4585 
4586 		if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) {
4587 			error = EINVAL;
4588 			break;
4589 		}
4590 
4591 		error = pmclog_process_userlog(po, &wl);
4592 	}
4593 	break;
4594 
4595 
4596 	default:
4597 		error = EINVAL;
4598 		break;
4599 	}
4600 
4601 	if (is_sx_downgraded)
4602 		sx_sunlock(&pmc_sx);
4603 	else
4604 		sx_xunlock(&pmc_sx);
4605 done_syscall:
4606 	if (error)
4607 		counter_u64_add(pmc_stats.pm_syscall_errors, 1);
4608 
4609 	return (error);
4610 }
4611 
4612 /*
4613  * Helper functions
4614  */
4615 
4616 
4617 /*
4618  * Mark the thread as needing callchain capture and post an AST.  The
4619  * actual callchain capture will be done in a context where it is safe
4620  * to take page faults.
4621  */
4622 
4623 static void
4624 pmc_post_callchain_callback(void)
4625 {
4626 	struct thread *td;
4627 
4628 	td = curthread;
4629 
4630 	/*
4631 	 * If there is multiple PMCs for the same interrupt ignore new post
4632 	 */
4633 	if (td->td_pflags & TDP_CALLCHAIN)
4634 		return;
4635 
4636 	/*
4637 	 * Mark this thread as needing callchain capture.
4638 	 * `td->td_pflags' will be safe to touch because this thread
4639 	 * was in user space when it was interrupted.
4640 	 */
4641 	td->td_pflags |= TDP_CALLCHAIN;
4642 
4643 	/*
4644 	 * Don't let this thread migrate between CPUs until callchain
4645 	 * capture completes.
4646 	 */
4647 	sched_pin();
4648 
4649 	return;
4650 }
4651 
4652 /*
4653  * Find a free slot in the per-cpu array of samples and capture the
4654  * current callchain there.  If a sample was successfully added, a bit
4655  * is set in mask 'pmc_cpumask' denoting that the DO_SAMPLES hook
4656  * needs to be invoked from the clock handler.
4657  *
4658  * This function is meant to be called from an NMI handler.  It cannot
4659  * use any of the locking primitives supplied by the OS.
4660  */
4661 
4662 static int
4663 pmc_add_sample(ring_type_t ring, struct pmc *pm, struct trapframe *tf)
4664 {
4665 	int error, cpu, callchaindepth, inuserspace;
4666 	struct thread *td;
4667 	struct pmc_sample *ps;
4668 	struct pmc_samplebuffer *psb;
4669 
4670 	error = 0;
4671 
4672 	/*
4673 	 * Allocate space for a sample buffer.
4674 	 */
4675 	cpu = curcpu;
4676 	psb = pmc_pcpu[cpu]->pc_sb[ring];
4677 	inuserspace = TRAPF_USERMODE(tf);
4678 	ps = PMC_PROD_SAMPLE(psb);
4679 	if (psb->ps_considx != psb->ps_prodidx &&
4680 		ps->ps_nsamples) {	/* in use, reader hasn't caught up */
4681 		pm->pm_pcpu_state[cpu].pps_stalled = 1;
4682 		counter_u64_add(pmc_stats.pm_intr_bufferfull, 1);
4683 		PMCDBG6(SAM,INT,1,"(spc) cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d",
4684 		    cpu, pm, (void *) tf, inuserspace,
4685 		    (int) (psb->ps_prodidx & pmc_sample_mask),
4686 		    (int) (psb->ps_considx & pmc_sample_mask));
4687 		callchaindepth = 1;
4688 		error = ENOMEM;
4689 		goto done;
4690 	}
4691 
4692 	/* Fill in entry. */
4693 	PMCDBG6(SAM,INT,1,"cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d", cpu, pm,
4694 	    (void *) tf, inuserspace,
4695 	    (int) (psb->ps_prodidx & pmc_sample_mask),
4696 	    (int) (psb->ps_considx & pmc_sample_mask));
4697 
4698 	td = curthread;
4699 	ps->ps_pmc = pm;
4700 	ps->ps_td = td;
4701 	ps->ps_pid = td->td_proc->p_pid;
4702 	ps->ps_tid = td->td_tid;
4703 	ps->ps_tsc = pmc_rdtsc();
4704 	ps->ps_ticks = ticks;
4705 	ps->ps_cpu = cpu;
4706 	ps->ps_flags = inuserspace ? PMC_CC_F_USERSPACE : 0;
4707 
4708 	callchaindepth = (pm->pm_flags & PMC_F_CALLCHAIN) ?
4709 	    pmc_callchaindepth : 1;
4710 
4711 	MPASS(ps->ps_pc != NULL);
4712 	if (callchaindepth == 1)
4713 		ps->ps_pc[0] = PMC_TRAPFRAME_TO_PC(tf);
4714 	else {
4715 		/*
4716 		 * Kernel stack traversals can be done immediately,
4717 		 * while we defer to an AST for user space traversals.
4718 		 */
4719 		if (!inuserspace) {
4720 			callchaindepth =
4721 			    pmc_save_kernel_callchain(ps->ps_pc,
4722 				callchaindepth, tf);
4723 		} else {
4724 			pmc_post_callchain_callback();
4725 			callchaindepth = PMC_USER_CALLCHAIN_PENDING;
4726 		}
4727 	}
4728 
4729 	ps->ps_nsamples = callchaindepth;	/* mark entry as in use */
4730 	if (ring == PMC_UR) {
4731 		ps->ps_nsamples_actual = callchaindepth;	/* mark entry as in use */
4732 		ps->ps_nsamples = PMC_USER_CALLCHAIN_PENDING;
4733 	} else
4734 		ps->ps_nsamples = callchaindepth;	/* mark entry as in use */
4735 
4736 	KASSERT(counter_u64_fetch(pm->pm_runcount) >= 0,
4737 	    ("[pmc,%d] pm=%p runcount %ld", __LINE__, (void *) pm,
4738 		 (unsigned long)counter_u64_fetch(pm->pm_runcount)));
4739 
4740 	counter_u64_add(pm->pm_runcount, 1);	/* hold onto PMC */
4741 	/* increment write pointer */
4742 	psb->ps_prodidx++;
4743  done:
4744 	/* mark CPU as needing processing */
4745 	if (callchaindepth != PMC_USER_CALLCHAIN_PENDING)
4746 		DPCPU_SET(pmc_sampled, 1);
4747 
4748 	return (error);
4749 }
4750 
4751 /*
4752  * Interrupt processing.
4753  *
4754  * This function is meant to be called from an NMI handler.  It cannot
4755  * use any of the locking primitives supplied by the OS.
4756  */
4757 
4758 int
4759 pmc_process_interrupt(int ring, struct pmc *pm, struct trapframe *tf)
4760 {
4761 	struct thread *td;
4762 
4763 	td = curthread;
4764 	if ((pm->pm_flags & PMC_F_USERCALLCHAIN) &&
4765 	    (td->td_proc->p_flag & P_KPROC) == 0 &&
4766 	    !TRAPF_USERMODE(tf)) {
4767 		atomic_add_int(&td->td_pmcpend, 1);
4768 		return (pmc_add_sample(PMC_UR, pm, tf));
4769 	}
4770 	return (pmc_add_sample(ring, pm, tf));
4771 }
4772 
4773 /*
4774  * Capture a user call chain.  This function will be called from ast()
4775  * before control returns to userland and before the process gets
4776  * rescheduled.
4777  */
4778 
4779 static void
4780 pmc_capture_user_callchain(int cpu, int ring, struct trapframe *tf)
4781 {
4782 	struct pmc *pm;
4783 	struct thread *td;
4784 	struct pmc_sample *ps;
4785 	struct pmc_samplebuffer *psb;
4786 	uint64_t considx, prodidx;
4787 	int nsamples, nrecords, pass, iter;
4788 #ifdef	INVARIANTS
4789 	int ncallchains;
4790 	int nfree;
4791 	int start_ticks = ticks;
4792 #endif
4793 	psb = pmc_pcpu[cpu]->pc_sb[ring];
4794 	td = curthread;
4795 
4796 	KASSERT(td->td_pflags & TDP_CALLCHAIN,
4797 	    ("[pmc,%d] Retrieving callchain for thread that doesn't want it",
4798 		__LINE__));
4799 
4800 #ifdef	INVARIANTS
4801 	ncallchains = 0;
4802 	nfree = 0;
4803 #endif
4804 	nrecords = INT_MAX;
4805 	pass = 0;
4806  restart:
4807 	if (ring == PMC_UR)
4808 		nrecords = atomic_readandclear_32(&td->td_pmcpend);
4809 
4810 	for (iter = 0, considx = psb->ps_considx, prodidx = psb->ps_prodidx;
4811 	    considx < prodidx && iter < pmc_nsamples; considx++, iter++) {
4812 		ps = PMC_CONS_SAMPLE_OFF(psb, considx);
4813 
4814 	/*
4815 	 * Iterate through all deferred callchain requests.
4816 	 * Walk from the current read pointer to the current
4817 	 * write pointer.
4818 	 */
4819 
4820 #ifdef	INVARIANTS
4821 		if (ps->ps_nsamples == PMC_SAMPLE_FREE) {
4822 			nfree++;
4823 			continue;
4824 		}
4825 
4826 		if ((ps->ps_pmc == NULL) ||
4827 		    (ps->ps_pmc->pm_state != PMC_STATE_RUNNING))
4828 			nfree++;
4829 #endif
4830 		if (ps->ps_td != td ||
4831 		   ps->ps_nsamples != PMC_USER_CALLCHAIN_PENDING ||
4832 		   ps->ps_pmc->pm_state != PMC_STATE_RUNNING)
4833 			continue;
4834 
4835 		KASSERT(ps->ps_cpu == cpu,
4836 		    ("[pmc,%d] cpu mismatch ps_cpu=%d pcpu=%d", __LINE__,
4837 			ps->ps_cpu, PCPU_GET(cpuid)));
4838 
4839 		pm = ps->ps_pmc;
4840 
4841 		KASSERT(pm->pm_flags & PMC_F_CALLCHAIN,
4842 		    ("[pmc,%d] Retrieving callchain for PMC that doesn't "
4843 			"want it", __LINE__));
4844 
4845 		KASSERT(counter_u64_fetch(pm->pm_runcount) > 0,
4846 		    ("[pmc,%d] runcount %ld", __LINE__, (unsigned long)counter_u64_fetch(pm->pm_runcount)));
4847 
4848 		if (ring == PMC_UR) {
4849 			nsamples = ps->ps_nsamples_actual;
4850 			counter_u64_add(pmc_stats.pm_merges, 1);
4851 		} else
4852 			nsamples = 0;
4853 
4854 		/*
4855 		 * Retrieve the callchain and mark the sample buffer
4856 		 * as 'processable' by the timer tick sweep code.
4857 		 */
4858 
4859 #ifdef INVARIANTS
4860 		ncallchains++;
4861 #endif
4862 
4863 		if (__predict_true(nsamples < pmc_callchaindepth - 1))
4864 			nsamples += pmc_save_user_callchain(ps->ps_pc + nsamples,
4865 		       pmc_callchaindepth - nsamples - 1, tf);
4866 
4867 		/*
4868 		 * We have to prevent hardclock from potentially overwriting
4869 		 * this sample between when we read the value and when we set
4870 		 * it
4871 		 */
4872 		spinlock_enter();
4873 		/*
4874 		 * Verify that the sample hasn't been dropped in the meantime
4875 		 */
4876 		if (ps->ps_nsamples == PMC_USER_CALLCHAIN_PENDING) {
4877 			ps->ps_nsamples = nsamples;
4878 			/*
4879 			 * If we couldn't get a sample, simply drop the reference
4880 			 */
4881 			if (nsamples == 0)
4882 				counter_u64_add(pm->pm_runcount, -1);
4883 		}
4884 		spinlock_exit();
4885 		if (nrecords-- == 1)
4886 			break;
4887 	}
4888 	if (__predict_false(ring == PMC_UR && td->td_pmcpend)) {
4889 		if (pass == 0) {
4890 			pass = 1;
4891 			goto restart;
4892 		}
4893 		/* only collect samples for this part once */
4894 		td->td_pmcpend = 0;
4895 	}
4896 
4897 #ifdef INVARIANTS
4898 	if ((ticks - start_ticks) > hz)
4899 		log(LOG_ERR, "%s took %d ticks\n", __func__, (ticks - start_ticks));
4900 #endif
4901 
4902 	/* mark CPU as needing processing */
4903 	DPCPU_SET(pmc_sampled, 1);
4904 }
4905 
4906 /*
4907  * Process saved PC samples.
4908  */
4909 
4910 static void
4911 pmc_process_samples(int cpu, ring_type_t ring)
4912 {
4913 	struct pmc *pm;
4914 	int adjri, n;
4915 	struct thread *td;
4916 	struct pmc_owner *po;
4917 	struct pmc_sample *ps;
4918 	struct pmc_classdep *pcd;
4919 	struct pmc_samplebuffer *psb;
4920 	uint64_t delta;
4921 
4922 	KASSERT(PCPU_GET(cpuid) == cpu,
4923 	    ("[pmc,%d] not on the correct CPU pcpu=%d cpu=%d", __LINE__,
4924 		PCPU_GET(cpuid), cpu));
4925 
4926 	psb = pmc_pcpu[cpu]->pc_sb[ring];
4927 	delta = psb->ps_prodidx - psb->ps_considx;
4928 	MPASS(delta <= pmc_nsamples);
4929 	MPASS(psb->ps_considx <= psb->ps_prodidx);
4930 	for (n = 0; psb->ps_considx < psb->ps_prodidx; psb->ps_considx++, n++) {
4931 		ps = PMC_CONS_SAMPLE(psb);
4932 
4933 		if (__predict_false(ps->ps_nsamples == PMC_SAMPLE_FREE))
4934 			continue;
4935 		pm = ps->ps_pmc;
4936 		/* skip non-running samples */
4937 		if (pm->pm_state != PMC_STATE_RUNNING)
4938 			goto entrydone;
4939 
4940 		KASSERT(counter_u64_fetch(pm->pm_runcount) > 0,
4941 		    ("[pmc,%d] pm=%p runcount %ld", __LINE__, (void *) pm,
4942 			 (unsigned long)counter_u64_fetch(pm->pm_runcount)));
4943 
4944 		po = pm->pm_owner;
4945 
4946 		KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)),
4947 		    ("[pmc,%d] pmc=%p non-sampling mode=%d", __LINE__,
4948 			pm, PMC_TO_MODE(pm)));
4949 
4950 
4951 		/* If there is a pending AST wait for completion */
4952 		if (ps->ps_nsamples == PMC_USER_CALLCHAIN_PENDING) {
4953 			/* if we've been waiting more than 1 tick to
4954 			 * collect a callchain for this record then
4955 			 * drop it and move on.
4956 			 */
4957 			if (ticks - ps->ps_ticks > 1) {
4958 				/*
4959 				 * track how often we hit this as it will
4960 				 * preferentially lose user samples
4961 				 * for long running system calls
4962 				 */
4963 				counter_u64_add(pmc_stats.pm_overwrites, 1);
4964 				goto entrydone;
4965 			}
4966 			/* Need a rescan at a later time. */
4967 			DPCPU_SET(pmc_sampled, 1);
4968 			break;
4969 		}
4970 
4971 		PMCDBG6(SAM,OPS,1,"cpu=%d pm=%p n=%d fl=%x wr=%d rd=%d", cpu,
4972 		    pm, ps->ps_nsamples, ps->ps_flags,
4973 		    (int) (psb->ps_prodidx & pmc_sample_mask),
4974 		    (int) (psb->ps_considx & pmc_sample_mask));
4975 
4976 		/*
4977 		 * If this is a process-mode PMC that is attached to
4978 		 * its owner, and if the PC is in user mode, update
4979 		 * profiling statistics like timer-based profiling
4980 		 * would have done.
4981 		 *
4982 		 * Otherwise, this is either a sampling-mode PMC that
4983 		 * is attached to a different process than its owner,
4984 		 * or a system-wide sampling PMC. Dispatch a log
4985 		 * entry to the PMC's owner process.
4986 		 */
4987 		if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) {
4988 			if (ps->ps_flags & PMC_CC_F_USERSPACE) {
4989 				td = FIRST_THREAD_IN_PROC(po->po_owner);
4990 				addupc_intr(td, ps->ps_pc[0], 1);
4991 			}
4992 		} else
4993 			pmclog_process_callchain(pm, ps);
4994 
4995 	entrydone:
4996 		ps->ps_nsamples = 0; /* mark entry as free */
4997 		KASSERT(counter_u64_fetch(pm->pm_runcount) > 0,
4998 				("[pmc,%d] pm=%p runcount %ld", __LINE__, (void *) pm,
4999 				 (unsigned long)counter_u64_fetch(pm->pm_runcount)));
5000 
5001 		counter_u64_add(pm->pm_runcount, -1);
5002 	}
5003 
5004 	counter_u64_add(pmc_stats.pm_log_sweeps, 1);
5005 
5006 	/* Do not re-enable stalled PMCs if we failed to process any samples */
5007 	if (n == 0)
5008 		return;
5009 
5010 	/*
5011 	 * Restart any stalled sampling PMCs on this CPU.
5012 	 *
5013 	 * If the NMI handler sets the pm_stalled field of a PMC after
5014 	 * the check below, we'll end up processing the stalled PMC at
5015 	 * the next hardclock tick.
5016 	 */
5017 	for (n = 0; n < md->pmd_npmc; n++) {
5018 		pcd = pmc_ri_to_classdep(md, n, &adjri);
5019 		KASSERT(pcd != NULL,
5020 		    ("[pmc,%d] null pcd ri=%d", __LINE__, n));
5021 		(void) (*pcd->pcd_get_config)(cpu,adjri,&pm);
5022 
5023 		if (pm == NULL ||			 /* !cfg'ed */
5024 		    pm->pm_state != PMC_STATE_RUNNING || /* !active */
5025 		    !PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) || /* !sampling */
5026 			!pm->pm_pcpu_state[cpu].pps_cpustate  || /* !desired */
5027 		    !pm->pm_pcpu_state[cpu].pps_stalled) /* !stalled */
5028 			continue;
5029 
5030 		pm->pm_pcpu_state[cpu].pps_stalled = 0;
5031 		(*pcd->pcd_start_pmc)(cpu, adjri);
5032 	}
5033 }
5034 
5035 /*
5036  * Event handlers.
5037  */
5038 
5039 /*
5040  * Handle a process exit.
5041  *
5042  * Remove this process from all hash tables.  If this process
5043  * owned any PMCs, turn off those PMCs and deallocate them,
5044  * removing any associations with target processes.
5045  *
5046  * This function will be called by the last 'thread' of a
5047  * process.
5048  *
5049  * XXX This eventhandler gets called early in the exit process.
5050  * Consider using a 'hook' invocation from thread_exit() or equivalent
5051  * spot.  Another negative is that kse_exit doesn't seem to call
5052  * exit1() [??].
5053  *
5054  */
5055 
5056 static void
5057 pmc_process_exit(void *arg __unused, struct proc *p)
5058 {
5059 	struct pmc *pm;
5060 	int adjri, cpu;
5061 	unsigned int ri;
5062 	int is_using_hwpmcs;
5063 	struct pmc_owner *po;
5064 	struct pmc_process *pp;
5065 	struct pmc_classdep *pcd;
5066 	pmc_value_t newvalue, tmp;
5067 
5068 	PROC_LOCK(p);
5069 	is_using_hwpmcs = p->p_flag & P_HWPMC;
5070 	PROC_UNLOCK(p);
5071 
5072 	/*
5073 	 * Log a sysexit event to all SS PMC owners.
5074 	 */
5075 	PMC_EPOCH_ENTER();
5076 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
5077 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
5078 		    pmclog_process_sysexit(po, p->p_pid);
5079 	PMC_EPOCH_EXIT();
5080 
5081 	if (!is_using_hwpmcs)
5082 		return;
5083 
5084 	PMC_GET_SX_XLOCK();
5085 	PMCDBG3(PRC,EXT,1,"process-exit proc=%p (%d, %s)", p, p->p_pid,
5086 	    p->p_comm);
5087 
5088 	/*
5089 	 * Since this code is invoked by the last thread in an exiting
5090 	 * process, we would have context switched IN at some prior
5091 	 * point.  However, with PREEMPTION, kernel mode context
5092 	 * switches may happen any time, so we want to disable a
5093 	 * context switch OUT till we get any PMCs targeting this
5094 	 * process off the hardware.
5095 	 *
5096 	 * We also need to atomically remove this process'
5097 	 * entry from our target process hash table, using
5098 	 * PMC_FLAG_REMOVE.
5099 	 */
5100 	PMCDBG3(PRC,EXT,1, "process-exit proc=%p (%d, %s)", p, p->p_pid,
5101 	    p->p_comm);
5102 
5103 	critical_enter(); /* no preemption */
5104 
5105 	cpu = curthread->td_oncpu;
5106 
5107 	if ((pp = pmc_find_process_descriptor(p,
5108 		 PMC_FLAG_REMOVE)) != NULL) {
5109 
5110 		PMCDBG2(PRC,EXT,2,
5111 		    "process-exit proc=%p pmc-process=%p", p, pp);
5112 
5113 		/*
5114 		 * The exiting process could the target of
5115 		 * some PMCs which will be running on
5116 		 * currently executing CPU.
5117 		 *
5118 		 * We need to turn these PMCs off like we
5119 		 * would do at context switch OUT time.
5120 		 */
5121 		for (ri = 0; ri < md->pmd_npmc; ri++) {
5122 
5123 			/*
5124 			 * Pick up the pmc pointer from hardware
5125 			 * state similar to the CSW_OUT code.
5126 			 */
5127 			pm = NULL;
5128 
5129 			pcd = pmc_ri_to_classdep(md, ri, &adjri);
5130 
5131 			(void) (*pcd->pcd_get_config)(cpu, adjri, &pm);
5132 
5133 			PMCDBG2(PRC,EXT,2, "ri=%d pm=%p", ri, pm);
5134 
5135 			if (pm == NULL ||
5136 			    !PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)))
5137 				continue;
5138 
5139 			PMCDBG4(PRC,EXT,2, "ppmcs[%d]=%p pm=%p "
5140 			    "state=%d", ri, pp->pp_pmcs[ri].pp_pmc,
5141 			    pm, pm->pm_state);
5142 
5143 			KASSERT(PMC_TO_ROWINDEX(pm) == ri,
5144 			    ("[pmc,%d] ri mismatch pmc(%d) ri(%d)",
5145 				__LINE__, PMC_TO_ROWINDEX(pm), ri));
5146 
5147 			KASSERT(pm == pp->pp_pmcs[ri].pp_pmc,
5148 			    ("[pmc,%d] pm %p != pp_pmcs[%d] %p",
5149 				__LINE__, pm, ri, pp->pp_pmcs[ri].pp_pmc));
5150 
5151 			KASSERT(counter_u64_fetch(pm->pm_runcount) > 0,
5152 			    ("[pmc,%d] bad runcount ri %d rc %ld",
5153 				 __LINE__, ri, (unsigned long)counter_u64_fetch(pm->pm_runcount)));
5154 
5155 			/*
5156 			 * Change desired state, and then stop if not
5157 			 * stalled. This two-step dance should avoid
5158 			 * race conditions where an interrupt re-enables
5159 			 * the PMC after this code has already checked
5160 			 * the pm_stalled flag.
5161 			 */
5162 			if (pm->pm_pcpu_state[cpu].pps_cpustate) {
5163 				pm->pm_pcpu_state[cpu].pps_cpustate = 0;
5164 				if (!pm->pm_pcpu_state[cpu].pps_stalled) {
5165 					(void) pcd->pcd_stop_pmc(cpu, adjri);
5166 
5167 					if (PMC_TO_MODE(pm) == PMC_MODE_TC) {
5168 						pcd->pcd_read_pmc(cpu, adjri,
5169 						    &newvalue);
5170 						tmp = newvalue -
5171 						    PMC_PCPU_SAVED(cpu,ri);
5172 
5173 						mtx_pool_lock_spin(pmc_mtxpool,
5174 						    pm);
5175 						pm->pm_gv.pm_savedvalue += tmp;
5176 						pp->pp_pmcs[ri].pp_pmcval +=
5177 						    tmp;
5178 						mtx_pool_unlock_spin(
5179 						    pmc_mtxpool, pm);
5180 					}
5181 				}
5182 			}
5183 
5184 			KASSERT((int64_t) counter_u64_fetch(pm->pm_runcount) > 0,
5185 			    ("[pmc,%d] runcount is %d", __LINE__, ri));
5186 
5187 			counter_u64_add(pm->pm_runcount, -1);
5188 
5189 			(void) pcd->pcd_config_pmc(cpu, adjri, NULL);
5190 		}
5191 
5192 		/*
5193 		 * Inform the MD layer of this pseudo "context switch
5194 		 * out"
5195 		 */
5196 		(void) md->pmd_switch_out(pmc_pcpu[cpu], pp);
5197 
5198 		critical_exit(); /* ok to be pre-empted now */
5199 
5200 		/*
5201 		 * Unlink this process from the PMCs that are
5202 		 * targeting it.  This will send a signal to
5203 		 * all PMC owner's whose PMCs are orphaned.
5204 		 *
5205 		 * Log PMC value at exit time if requested.
5206 		 */
5207 		for (ri = 0; ri < md->pmd_npmc; ri++)
5208 			if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) {
5209 				if (pm->pm_flags & PMC_F_NEEDS_LOGFILE &&
5210 				    PMC_IS_COUNTING_MODE(PMC_TO_MODE(pm)))
5211 					pmclog_process_procexit(pm, pp);
5212 				pmc_unlink_target_process(pm, pp);
5213 			}
5214 		free(pp, M_PMC);
5215 
5216 	} else
5217 		critical_exit(); /* pp == NULL */
5218 
5219 
5220 	/*
5221 	 * If the process owned PMCs, free them up and free up
5222 	 * memory.
5223 	 */
5224 	if ((po = pmc_find_owner_descriptor(p)) != NULL) {
5225 		pmc_remove_owner(po);
5226 		pmc_destroy_owner_descriptor(po);
5227 	}
5228 
5229 	sx_xunlock(&pmc_sx);
5230 }
5231 
5232 /*
5233  * Handle a process fork.
5234  *
5235  * If the parent process 'p1' is under HWPMC monitoring, then copy
5236  * over any attached PMCs that have 'do_descendants' semantics.
5237  */
5238 
5239 static void
5240 pmc_process_fork(void *arg __unused, struct proc *p1, struct proc *newproc,
5241     int flags)
5242 {
5243 	int is_using_hwpmcs;
5244 	unsigned int ri;
5245 	uint32_t do_descendants;
5246 	struct pmc *pm;
5247 	struct pmc_owner *po;
5248 	struct pmc_process *ppnew, *ppold;
5249 
5250 	(void) flags;		/* unused parameter */
5251 
5252 	PROC_LOCK(p1);
5253 	is_using_hwpmcs = p1->p_flag & P_HWPMC;
5254 	PROC_UNLOCK(p1);
5255 
5256 	/*
5257 	 * If there are system-wide sampling PMCs active, we need to
5258 	 * log all fork events to their owner's logs.
5259 	 */
5260 	PMC_EPOCH_ENTER();
5261 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
5262 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE) {
5263 		    pmclog_process_procfork(po, p1->p_pid, newproc->p_pid);
5264 			pmclog_process_proccreate(po, newproc, 1);
5265 		}
5266 	PMC_EPOCH_EXIT();
5267 
5268 	if (!is_using_hwpmcs)
5269 		return;
5270 
5271 	PMC_GET_SX_XLOCK();
5272 	PMCDBG4(PMC,FRK,1, "process-fork proc=%p (%d, %s) -> %p", p1,
5273 	    p1->p_pid, p1->p_comm, newproc);
5274 
5275 	/*
5276 	 * If the parent process (curthread->td_proc) is a
5277 	 * target of any PMCs, look for PMCs that are to be
5278 	 * inherited, and link these into the new process
5279 	 * descriptor.
5280 	 */
5281 	if ((ppold = pmc_find_process_descriptor(curthread->td_proc,
5282 		 PMC_FLAG_NONE)) == NULL)
5283 		goto done;		/* nothing to do */
5284 
5285 	do_descendants = 0;
5286 	for (ri = 0; ri < md->pmd_npmc; ri++)
5287 		if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL)
5288 			do_descendants |= pm->pm_flags & PMC_F_DESCENDANTS;
5289 	if (do_descendants == 0) /* nothing to do */
5290 		goto done;
5291 
5292 	/*
5293 	 * Now mark the new process as being tracked by this driver.
5294 	 */
5295 	PROC_LOCK(newproc);
5296 	newproc->p_flag |= P_HWPMC;
5297 	PROC_UNLOCK(newproc);
5298 
5299 	/* allocate a descriptor for the new process  */
5300 	if ((ppnew = pmc_find_process_descriptor(newproc,
5301 		 PMC_FLAG_ALLOCATE)) == NULL)
5302 		goto done;
5303 
5304 	/*
5305 	 * Run through all PMCs that were targeting the old process
5306 	 * and which specified F_DESCENDANTS and attach them to the
5307 	 * new process.
5308 	 *
5309 	 * Log the fork event to all owners of PMCs attached to this
5310 	 * process, if not already logged.
5311 	 */
5312 	for (ri = 0; ri < md->pmd_npmc; ri++)
5313 		if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL &&
5314 		    (pm->pm_flags & PMC_F_DESCENDANTS)) {
5315 			pmc_link_target_process(pm, ppnew);
5316 			po = pm->pm_owner;
5317 			if (po->po_sscount == 0 &&
5318 			    po->po_flags & PMC_PO_OWNS_LOGFILE)
5319 				pmclog_process_procfork(po, p1->p_pid,
5320 				    newproc->p_pid);
5321 		}
5322 
5323  done:
5324 	sx_xunlock(&pmc_sx);
5325 }
5326 
5327 static void
5328 pmc_process_threadcreate(struct thread *td)
5329 {
5330 	struct pmc_owner *po;
5331 
5332 	PMC_EPOCH_ENTER();
5333 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
5334 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
5335 			pmclog_process_threadcreate(po, td, 1);
5336 	PMC_EPOCH_EXIT();
5337 }
5338 
5339 static void
5340 pmc_process_threadexit(struct thread *td)
5341 {
5342 	struct pmc_owner *po;
5343 
5344 	PMC_EPOCH_ENTER();
5345 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
5346 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
5347 			pmclog_process_threadexit(po, td);
5348 	PMC_EPOCH_EXIT();
5349 }
5350 
5351 static void
5352 pmc_process_proccreate(struct proc *p)
5353 {
5354 	struct pmc_owner *po;
5355 
5356 	PMC_EPOCH_ENTER();
5357 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
5358 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
5359 			pmclog_process_proccreate(po, p, 1 /* sync */);
5360 	PMC_EPOCH_EXIT();
5361 }
5362 
5363 static void
5364 pmc_process_allproc(struct pmc *pm)
5365 {
5366 	struct pmc_owner *po;
5367 	struct thread *td;
5368 	struct proc *p;
5369 
5370 	po = pm->pm_owner;
5371 	if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
5372 		return;
5373 	sx_slock(&allproc_lock);
5374 	FOREACH_PROC_IN_SYSTEM(p) {
5375 		pmclog_process_proccreate(po, p, 0 /* sync */);
5376 		PROC_LOCK(p);
5377 		FOREACH_THREAD_IN_PROC(p, td)
5378 			pmclog_process_threadcreate(po, td, 0 /* sync */);
5379 		PROC_UNLOCK(p);
5380 	}
5381 	sx_sunlock(&allproc_lock);
5382 	pmclog_flush(po, 0);
5383 }
5384 
5385 static void
5386 pmc_kld_load(void *arg __unused, linker_file_t lf)
5387 {
5388 	struct pmc_owner *po;
5389 
5390 	/*
5391 	 * Notify owners of system sampling PMCs about KLD operations.
5392 	 */
5393 	PMC_EPOCH_ENTER();
5394 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
5395 		if (po->po_flags & PMC_PO_OWNS_LOGFILE)
5396 			pmclog_process_map_in(po, (pid_t) -1,
5397 			    (uintfptr_t) lf->address, lf->filename);
5398 	PMC_EPOCH_EXIT();
5399 
5400 	/*
5401 	 * TODO: Notify owners of (all) process-sampling PMCs too.
5402 	 */
5403 }
5404 
5405 static void
5406 pmc_kld_unload(void *arg __unused, const char *filename __unused,
5407     caddr_t address, size_t size)
5408 {
5409 	struct pmc_owner *po;
5410 
5411 	PMC_EPOCH_ENTER();
5412 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
5413 		if (po->po_flags & PMC_PO_OWNS_LOGFILE)
5414 			pmclog_process_map_out(po, (pid_t) -1,
5415 			    (uintfptr_t) address, (uintfptr_t) address + size);
5416 	PMC_EPOCH_EXIT();
5417 
5418 	/*
5419 	 * TODO: Notify owners of process-sampling PMCs.
5420 	 */
5421 }
5422 
5423 /*
5424  * initialization
5425  */
5426 static const char *
5427 pmc_name_of_pmcclass(enum pmc_class class)
5428 {
5429 
5430 	switch (class) {
5431 #undef	__PMC_CLASS
5432 #define	__PMC_CLASS(S,V,D)						\
5433 	case PMC_CLASS_##S:						\
5434 		return #S;
5435 	__PMC_CLASSES();
5436 	default:
5437 		return ("<unknown>");
5438 	}
5439 }
5440 
5441 /*
5442  * Base class initializer: allocate structure and set default classes.
5443  */
5444 struct pmc_mdep *
5445 pmc_mdep_alloc(int nclasses)
5446 {
5447 	struct pmc_mdep *md;
5448 	int	n;
5449 
5450 	/* SOFT + md classes */
5451 	n = 1 + nclasses;
5452 	md = malloc(sizeof(struct pmc_mdep) + n *
5453 	    sizeof(struct pmc_classdep), M_PMC, M_WAITOK|M_ZERO);
5454 	md->pmd_nclass = n;
5455 
5456 	/* Add base class. */
5457 	pmc_soft_initialize(md);
5458 	return md;
5459 }
5460 
5461 void
5462 pmc_mdep_free(struct pmc_mdep *md)
5463 {
5464 	pmc_soft_finalize(md);
5465 	free(md, M_PMC);
5466 }
5467 
5468 static int
5469 generic_switch_in(struct pmc_cpu *pc, struct pmc_process *pp)
5470 {
5471 	(void) pc; (void) pp;
5472 
5473 	return (0);
5474 }
5475 
5476 static int
5477 generic_switch_out(struct pmc_cpu *pc, struct pmc_process *pp)
5478 {
5479 	(void) pc; (void) pp;
5480 
5481 	return (0);
5482 }
5483 
5484 static struct pmc_mdep *
5485 pmc_generic_cpu_initialize(void)
5486 {
5487 	struct pmc_mdep *md;
5488 
5489 	md = pmc_mdep_alloc(0);
5490 
5491 	md->pmd_cputype    = PMC_CPU_GENERIC;
5492 
5493 	md->pmd_pcpu_init  = NULL;
5494 	md->pmd_pcpu_fini  = NULL;
5495 	md->pmd_switch_in  = generic_switch_in;
5496 	md->pmd_switch_out = generic_switch_out;
5497 
5498 	return (md);
5499 }
5500 
5501 static void
5502 pmc_generic_cpu_finalize(struct pmc_mdep *md)
5503 {
5504 	(void) md;
5505 }
5506 
5507 
5508 static int
5509 pmc_initialize(void)
5510 {
5511 	int c, cpu, error, n, ri;
5512 	unsigned int maxcpu, domain;
5513 	struct pcpu *pc;
5514 	struct pmc_binding pb;
5515 	struct pmc_sample *ps;
5516 	struct pmc_classdep *pcd;
5517 	struct pmc_samplebuffer *sb;
5518 
5519 	md = NULL;
5520 	error = 0;
5521 
5522 	pmc_stats.pm_intr_ignored = counter_u64_alloc(M_WAITOK);
5523 	pmc_stats.pm_intr_processed = counter_u64_alloc(M_WAITOK);
5524 	pmc_stats.pm_intr_bufferfull = counter_u64_alloc(M_WAITOK);
5525 	pmc_stats.pm_syscalls = counter_u64_alloc(M_WAITOK);
5526 	pmc_stats.pm_syscall_errors = counter_u64_alloc(M_WAITOK);
5527 	pmc_stats.pm_buffer_requests = counter_u64_alloc(M_WAITOK);
5528 	pmc_stats.pm_buffer_requests_failed = counter_u64_alloc(M_WAITOK);
5529 	pmc_stats.pm_log_sweeps = counter_u64_alloc(M_WAITOK);
5530 	pmc_stats.pm_merges = counter_u64_alloc(M_WAITOK);
5531 	pmc_stats.pm_overwrites = counter_u64_alloc(M_WAITOK);
5532 
5533 #ifdef	HWPMC_DEBUG
5534 	/* parse debug flags first */
5535 	if (TUNABLE_STR_FETCH(PMC_SYSCTL_NAME_PREFIX "debugflags",
5536 		pmc_debugstr, sizeof(pmc_debugstr)))
5537 		pmc_debugflags_parse(pmc_debugstr,
5538 		    pmc_debugstr+strlen(pmc_debugstr));
5539 #endif
5540 
5541 	PMCDBG1(MOD,INI,0, "PMC Initialize (version %x)", PMC_VERSION);
5542 
5543 	/* check kernel version */
5544 	if (pmc_kernel_version != PMC_VERSION) {
5545 		if (pmc_kernel_version == 0)
5546 			printf("hwpmc: this kernel has not been compiled with "
5547 			    "'options HWPMC_HOOKS'.\n");
5548 		else
5549 			printf("hwpmc: kernel version (0x%x) does not match "
5550 			    "module version (0x%x).\n", pmc_kernel_version,
5551 			    PMC_VERSION);
5552 		return EPROGMISMATCH;
5553 	}
5554 
5555 	/*
5556 	 * check sysctl parameters
5557 	 */
5558 
5559 	if (pmc_hashsize <= 0) {
5560 		(void) printf("hwpmc: tunable \"hashsize\"=%d must be "
5561 		    "greater than zero.\n", pmc_hashsize);
5562 		pmc_hashsize = PMC_HASH_SIZE;
5563 	}
5564 
5565 	if (pmc_nsamples <= 0 || pmc_nsamples > 65535) {
5566 		(void) printf("hwpmc: tunable \"nsamples\"=%d out of "
5567 		    "range.\n", pmc_nsamples);
5568 		pmc_nsamples = PMC_NSAMPLES;
5569 	}
5570 	pmc_sample_mask = pmc_nsamples-1;
5571 
5572 	if (pmc_callchaindepth <= 0 ||
5573 	    pmc_callchaindepth > PMC_CALLCHAIN_DEPTH_MAX) {
5574 		(void) printf("hwpmc: tunable \"callchaindepth\"=%d out of "
5575 		    "range - using %d.\n", pmc_callchaindepth,
5576 		    PMC_CALLCHAIN_DEPTH_MAX);
5577 		pmc_callchaindepth = PMC_CALLCHAIN_DEPTH_MAX;
5578 	}
5579 
5580 	md = pmc_md_initialize();
5581 	if (md == NULL) {
5582 		/* Default to generic CPU. */
5583 		md = pmc_generic_cpu_initialize();
5584 		if (md == NULL)
5585 			return (ENOSYS);
5586         }
5587 
5588 	KASSERT(md->pmd_nclass >= 1 && md->pmd_npmc >= 1,
5589 	    ("[pmc,%d] no classes or pmcs", __LINE__));
5590 
5591 	/* Compute the map from row-indices to classdep pointers. */
5592 	pmc_rowindex_to_classdep = malloc(sizeof(struct pmc_classdep *) *
5593 	    md->pmd_npmc, M_PMC, M_WAITOK|M_ZERO);
5594 
5595 	for (n = 0; n < md->pmd_npmc; n++)
5596 		pmc_rowindex_to_classdep[n] = NULL;
5597 	for (ri = c = 0; c < md->pmd_nclass; c++) {
5598 		pcd = &md->pmd_classdep[c];
5599 		for (n = 0; n < pcd->pcd_num; n++, ri++)
5600 			pmc_rowindex_to_classdep[ri] = pcd;
5601 	}
5602 
5603 	KASSERT(ri == md->pmd_npmc,
5604 	    ("[pmc,%d] npmc miscomputed: ri=%d, md->npmc=%d", __LINE__,
5605 	    ri, md->pmd_npmc));
5606 
5607 	maxcpu = pmc_cpu_max();
5608 
5609 	/* allocate space for the per-cpu array */
5610 	pmc_pcpu = malloc(maxcpu * sizeof(struct pmc_cpu *), M_PMC,
5611 	    M_WAITOK|M_ZERO);
5612 
5613 	/* per-cpu 'saved values' for managing process-mode PMCs */
5614 	pmc_pcpu_saved = malloc(sizeof(pmc_value_t) * maxcpu * md->pmd_npmc,
5615 	    M_PMC, M_WAITOK);
5616 
5617 	/* Perform CPU-dependent initialization. */
5618 	pmc_save_cpu_binding(&pb);
5619 	error = 0;
5620 	for (cpu = 0; error == 0 && cpu < maxcpu; cpu++) {
5621 		if (!pmc_cpu_is_active(cpu))
5622 			continue;
5623 		pmc_select_cpu(cpu);
5624 		pmc_pcpu[cpu] = malloc(sizeof(struct pmc_cpu) +
5625 		    md->pmd_npmc * sizeof(struct pmc_hw *), M_PMC,
5626 		    M_WAITOK|M_ZERO);
5627 		if (md->pmd_pcpu_init)
5628 			error = md->pmd_pcpu_init(md, cpu);
5629 		for (n = 0; error == 0 && n < md->pmd_nclass; n++)
5630 			error = md->pmd_classdep[n].pcd_pcpu_init(md, cpu);
5631 	}
5632 	pmc_restore_cpu_binding(&pb);
5633 
5634 	if (error)
5635 		return (error);
5636 
5637 	/* allocate space for the sample array */
5638 	for (cpu = 0; cpu < maxcpu; cpu++) {
5639 		if (!pmc_cpu_is_active(cpu))
5640 			continue;
5641 		pc = pcpu_find(cpu);
5642 		domain = pc->pc_domain;
5643 		sb = malloc_domainset(sizeof(struct pmc_samplebuffer) +
5644 		    pmc_nsamples * sizeof(struct pmc_sample), M_PMC,
5645 		    DOMAINSET_PREF(domain), M_WAITOK | M_ZERO);
5646 
5647 		KASSERT(pmc_pcpu[cpu] != NULL,
5648 		    ("[pmc,%d] cpu=%d Null per-cpu data", __LINE__, cpu));
5649 
5650 		sb->ps_callchains = malloc_domainset(pmc_callchaindepth *
5651 		    pmc_nsamples * sizeof(uintptr_t), M_PMC,
5652 		    DOMAINSET_PREF(domain), M_WAITOK | M_ZERO);
5653 
5654 		for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++)
5655 			ps->ps_pc = sb->ps_callchains +
5656 			    (n * pmc_callchaindepth);
5657 
5658 		pmc_pcpu[cpu]->pc_sb[PMC_HR] = sb;
5659 
5660 		sb = malloc_domainset(sizeof(struct pmc_samplebuffer) +
5661 		    pmc_nsamples * sizeof(struct pmc_sample), M_PMC,
5662 		    DOMAINSET_PREF(domain), M_WAITOK | M_ZERO);
5663 
5664 		sb->ps_callchains = malloc_domainset(pmc_callchaindepth *
5665 		    pmc_nsamples * sizeof(uintptr_t), M_PMC,
5666 		    DOMAINSET_PREF(domain), M_WAITOK | M_ZERO);
5667 		for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++)
5668 			ps->ps_pc = sb->ps_callchains +
5669 			    (n * pmc_callchaindepth);
5670 
5671 		pmc_pcpu[cpu]->pc_sb[PMC_SR] = sb;
5672 
5673 		sb = malloc_domainset(sizeof(struct pmc_samplebuffer) +
5674 		    pmc_nsamples * sizeof(struct pmc_sample), M_PMC,
5675 		    DOMAINSET_PREF(domain), M_WAITOK | M_ZERO);
5676 		sb->ps_callchains = malloc_domainset(pmc_callchaindepth *
5677 		    pmc_nsamples * sizeof(uintptr_t), M_PMC,
5678 		    DOMAINSET_PREF(domain), M_WAITOK | M_ZERO);
5679 		for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++)
5680 			ps->ps_pc = sb->ps_callchains + n * pmc_callchaindepth;
5681 
5682 		pmc_pcpu[cpu]->pc_sb[PMC_UR] = sb;
5683 	}
5684 
5685 	/* allocate space for the row disposition array */
5686 	pmc_pmcdisp = malloc(sizeof(enum pmc_mode) * md->pmd_npmc,
5687 	    M_PMC, M_WAITOK|M_ZERO);
5688 
5689 	/* mark all PMCs as available */
5690 	for (n = 0; n < (int) md->pmd_npmc; n++)
5691 		PMC_MARK_ROW_FREE(n);
5692 
5693 	/* allocate thread hash tables */
5694 	pmc_ownerhash = hashinit(pmc_hashsize, M_PMC,
5695 	    &pmc_ownerhashmask);
5696 
5697 	pmc_processhash = hashinit(pmc_hashsize, M_PMC,
5698 	    &pmc_processhashmask);
5699 	mtx_init(&pmc_processhash_mtx, "pmc-process-hash", "pmc-leaf",
5700 	    MTX_SPIN);
5701 
5702 	CK_LIST_INIT(&pmc_ss_owners);
5703 	pmc_ss_count = 0;
5704 
5705 	/* allocate a pool of spin mutexes */
5706 	pmc_mtxpool = mtx_pool_create("pmc-leaf", pmc_mtxpool_size,
5707 	    MTX_SPIN);
5708 
5709 	PMCDBG4(MOD,INI,1, "pmc_ownerhash=%p, mask=0x%lx "
5710 	    "targethash=%p mask=0x%lx", pmc_ownerhash, pmc_ownerhashmask,
5711 	    pmc_processhash, pmc_processhashmask);
5712 
5713 	/* Initialize a spin mutex for the thread free list. */
5714 	mtx_init(&pmc_threadfreelist_mtx, "pmc-threadfreelist", "pmc-leaf",
5715 	    MTX_SPIN);
5716 
5717 	/*
5718 	 * Initialize the callout to monitor the thread free list.
5719 	 * This callout will also handle the initial population of the list.
5720 	 */
5721 	taskqgroup_config_gtask_init(NULL, &free_gtask, pmc_thread_descriptor_pool_free_task, "thread descriptor pool free task");
5722 
5723 	/* register process {exit,fork,exec} handlers */
5724 	pmc_exit_tag = EVENTHANDLER_REGISTER(process_exit,
5725 	    pmc_process_exit, NULL, EVENTHANDLER_PRI_ANY);
5726 	pmc_fork_tag = EVENTHANDLER_REGISTER(process_fork,
5727 	    pmc_process_fork, NULL, EVENTHANDLER_PRI_ANY);
5728 
5729 	/* register kld event handlers */
5730 	pmc_kld_load_tag = EVENTHANDLER_REGISTER(kld_load, pmc_kld_load,
5731 	    NULL, EVENTHANDLER_PRI_ANY);
5732 	pmc_kld_unload_tag = EVENTHANDLER_REGISTER(kld_unload, pmc_kld_unload,
5733 	    NULL, EVENTHANDLER_PRI_ANY);
5734 
5735 	/* initialize logging */
5736 	pmclog_initialize();
5737 
5738 	/* set hook functions */
5739 	pmc_intr = md->pmd_intr;
5740 	wmb();
5741 	pmc_hook = pmc_hook_handler;
5742 
5743 	if (error == 0) {
5744 		printf(PMC_MODULE_NAME ":");
5745 		for (n = 0; n < (int) md->pmd_nclass; n++) {
5746 			pcd = &md->pmd_classdep[n];
5747 			printf(" %s/%d/%d/0x%b",
5748 			    pmc_name_of_pmcclass(pcd->pcd_class),
5749 			    pcd->pcd_num,
5750 			    pcd->pcd_width,
5751 			    pcd->pcd_caps,
5752 			    "\20"
5753 			    "\1INT\2USR\3SYS\4EDG\5THR"
5754 			    "\6REA\7WRI\10INV\11QUA\12PRC"
5755 			    "\13TAG\14CSC");
5756 		}
5757 		printf("\n");
5758 	}
5759 
5760 	return (error);
5761 }
5762 
5763 /* prepare to be unloaded */
5764 static void
5765 pmc_cleanup(void)
5766 {
5767 	int c, cpu;
5768 	unsigned int maxcpu;
5769 	struct pmc_ownerhash *ph;
5770 	struct pmc_owner *po, *tmp;
5771 	struct pmc_binding pb;
5772 #ifdef	HWPMC_DEBUG
5773 	struct pmc_processhash *prh;
5774 #endif
5775 
5776 	PMCDBG0(MOD,INI,0, "cleanup");
5777 
5778 	/* switch off sampling */
5779 	CPU_FOREACH(cpu)
5780 		DPCPU_ID_SET(cpu, pmc_sampled, 0);
5781 	pmc_intr = NULL;
5782 
5783 	sx_xlock(&pmc_sx);
5784 	if (pmc_hook == NULL) {	/* being unloaded already */
5785 		sx_xunlock(&pmc_sx);
5786 		return;
5787 	}
5788 
5789 	pmc_hook = NULL; /* prevent new threads from entering module */
5790 
5791 	/* deregister event handlers */
5792 	EVENTHANDLER_DEREGISTER(process_fork, pmc_fork_tag);
5793 	EVENTHANDLER_DEREGISTER(process_exit, pmc_exit_tag);
5794 	EVENTHANDLER_DEREGISTER(kld_load, pmc_kld_load_tag);
5795 	EVENTHANDLER_DEREGISTER(kld_unload, pmc_kld_unload_tag);
5796 
5797 	/* send SIGBUS to all owner threads, free up allocations */
5798 	if (pmc_ownerhash)
5799 		for (ph = pmc_ownerhash;
5800 		     ph <= &pmc_ownerhash[pmc_ownerhashmask];
5801 		     ph++) {
5802 			LIST_FOREACH_SAFE(po, ph, po_next, tmp) {
5803 				pmc_remove_owner(po);
5804 
5805 				/* send SIGBUS to owner processes */
5806 				PMCDBG3(MOD,INI,2, "cleanup signal proc=%p "
5807 				    "(%d, %s)", po->po_owner,
5808 				    po->po_owner->p_pid,
5809 				    po->po_owner->p_comm);
5810 
5811 				PROC_LOCK(po->po_owner);
5812 				kern_psignal(po->po_owner, SIGBUS);
5813 				PROC_UNLOCK(po->po_owner);
5814 
5815 				pmc_destroy_owner_descriptor(po);
5816 			}
5817 		}
5818 
5819 	/* reclaim allocated data structures */
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 	taskqgroup_config_gtask_deinit(&free_gtask);
5828 	if (pmc_processhash) {
5829 #ifdef	HWPMC_DEBUG
5830 		struct pmc_process *pp;
5831 
5832 		PMCDBG0(MOD,INI,3, "destroy process hash");
5833 		for (prh = pmc_processhash;
5834 		     prh <= &pmc_processhash[pmc_processhashmask];
5835 		     prh++)
5836 			LIST_FOREACH(pp, prh, pp_next)
5837 			    PMCDBG1(MOD,INI,3, "pid=%d", pp->pp_proc->p_pid);
5838 #endif
5839 
5840 		hashdestroy(pmc_processhash, M_PMC, pmc_processhashmask);
5841 		pmc_processhash = NULL;
5842 	}
5843 
5844 	if (pmc_ownerhash) {
5845 		PMCDBG0(MOD,INI,3, "destroy owner hash");
5846 		hashdestroy(pmc_ownerhash, M_PMC, pmc_ownerhashmask);
5847 		pmc_ownerhash = NULL;
5848 	}
5849 
5850 	KASSERT(CK_LIST_EMPTY(&pmc_ss_owners),
5851 	    ("[pmc,%d] Global SS owner list not empty", __LINE__));
5852 	KASSERT(pmc_ss_count == 0,
5853 	    ("[pmc,%d] Global SS count not empty", __LINE__));
5854 
5855  	/* do processor and pmc-class dependent cleanup */
5856 	maxcpu = pmc_cpu_max();
5857 
5858 	PMCDBG0(MOD,INI,3, "md cleanup");
5859 	if (md) {
5860 		pmc_save_cpu_binding(&pb);
5861 		for (cpu = 0; cpu < maxcpu; cpu++) {
5862 			PMCDBG2(MOD,INI,1,"pmc-cleanup cpu=%d pcs=%p",
5863 			    cpu, pmc_pcpu[cpu]);
5864 			if (!pmc_cpu_is_active(cpu) || pmc_pcpu[cpu] == NULL)
5865 				continue;
5866 			pmc_select_cpu(cpu);
5867 			for (c = 0; c < md->pmd_nclass; c++)
5868 				md->pmd_classdep[c].pcd_pcpu_fini(md, cpu);
5869 			if (md->pmd_pcpu_fini)
5870 				md->pmd_pcpu_fini(md, cpu);
5871 		}
5872 
5873 		if (md->pmd_cputype == PMC_CPU_GENERIC)
5874 			pmc_generic_cpu_finalize(md);
5875 		else
5876 			pmc_md_finalize(md);
5877 
5878 		pmc_mdep_free(md);
5879 		md = NULL;
5880 		pmc_restore_cpu_binding(&pb);
5881 	}
5882 
5883 	/* Free per-cpu descriptors. */
5884 	for (cpu = 0; cpu < maxcpu; cpu++) {
5885 		if (!pmc_cpu_is_active(cpu))
5886 			continue;
5887 		KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_HR] != NULL,
5888 		    ("[pmc,%d] Null hw cpu sample buffer cpu=%d", __LINE__,
5889 			cpu));
5890 		KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_SR] != NULL,
5891 		    ("[pmc,%d] Null sw cpu sample buffer cpu=%d", __LINE__,
5892 			cpu));
5893 		KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_UR] != NULL,
5894 		    ("[pmc,%d] Null userret cpu sample buffer cpu=%d", __LINE__,
5895 			cpu));
5896 		free_domain(pmc_pcpu[cpu]->pc_sb[PMC_HR]->ps_callchains, M_PMC);
5897 		free_domain(pmc_pcpu[cpu]->pc_sb[PMC_HR], M_PMC);
5898 		free_domain(pmc_pcpu[cpu]->pc_sb[PMC_SR]->ps_callchains, M_PMC);
5899 		free_domain(pmc_pcpu[cpu]->pc_sb[PMC_SR], M_PMC);
5900 		free_domain(pmc_pcpu[cpu]->pc_sb[PMC_UR]->ps_callchains, M_PMC);
5901 		free_domain(pmc_pcpu[cpu]->pc_sb[PMC_UR], M_PMC);
5902 		free_domain(pmc_pcpu[cpu], M_PMC);
5903 	}
5904 
5905 	free(pmc_pcpu, M_PMC);
5906 	pmc_pcpu = NULL;
5907 
5908 	free(pmc_pcpu_saved, M_PMC);
5909 	pmc_pcpu_saved = NULL;
5910 
5911 	if (pmc_pmcdisp) {
5912 		free(pmc_pmcdisp, M_PMC);
5913 		pmc_pmcdisp = NULL;
5914 	}
5915 
5916 	if (pmc_rowindex_to_classdep) {
5917 		free(pmc_rowindex_to_classdep, M_PMC);
5918 		pmc_rowindex_to_classdep = NULL;
5919 	}
5920 
5921 	pmclog_shutdown();
5922 	counter_u64_free(pmc_stats.pm_intr_ignored);
5923 	counter_u64_free(pmc_stats.pm_intr_processed);
5924 	counter_u64_free(pmc_stats.pm_intr_bufferfull);
5925 	counter_u64_free(pmc_stats.pm_syscalls);
5926 	counter_u64_free(pmc_stats.pm_syscall_errors);
5927 	counter_u64_free(pmc_stats.pm_buffer_requests);
5928 	counter_u64_free(pmc_stats.pm_buffer_requests_failed);
5929 	counter_u64_free(pmc_stats.pm_log_sweeps);
5930 	counter_u64_free(pmc_stats.pm_merges);
5931 	counter_u64_free(pmc_stats.pm_overwrites);
5932 	sx_xunlock(&pmc_sx); 	/* we are done */
5933 }
5934 
5935 /*
5936  * The function called at load/unload.
5937  */
5938 
5939 static int
5940 load (struct module *module __unused, int cmd, void *arg __unused)
5941 {
5942 	int error;
5943 
5944 	error = 0;
5945 
5946 	switch (cmd) {
5947 	case MOD_LOAD :
5948 		/* initialize the subsystem */
5949 		error = pmc_initialize();
5950 		if (error != 0)
5951 			break;
5952 		PMCDBG2(MOD,INI,1, "syscall=%d maxcpu=%d",
5953 		    pmc_syscall_num, pmc_cpu_max());
5954 		break;
5955 
5956 
5957 	case MOD_UNLOAD :
5958 	case MOD_SHUTDOWN:
5959 		pmc_cleanup();
5960 		PMCDBG0(MOD,INI,1, "unloaded");
5961 		break;
5962 
5963 	default :
5964 		error = EINVAL;	/* XXX should panic(9) */
5965 		break;
5966 	}
5967 
5968 	return error;
5969 }
5970