xref: /freebsd/sys/dev/hwpmc/hwpmc_mod.c (revision e9b1dc32c9bd2ebae5f9e140bfa0e0321bc366b5)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2003-2008 Joseph Koshy
5  * Copyright (c) 2007 The FreeBSD Foundation
6  * Copyright (c) 2018 Matthew Macy
7  * All rights reserved.
8  *
9  * Portions of this software were developed by A. Joseph Koshy under
10  * sponsorship from the FreeBSD Foundation and Google, Inc.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  *
33  */
34 
35 #include <sys/cdefs.h>
36 __FBSDID("$FreeBSD$");
37 
38 #include <sys/param.h>
39 #include <sys/eventhandler.h>
40 #include <sys/gtaskqueue.h>
41 #include <sys/jail.h>
42 #include <sys/kernel.h>
43 #include <sys/kthread.h>
44 #include <sys/limits.h>
45 #include <sys/lock.h>
46 #include <sys/malloc.h>
47 #include <sys/module.h>
48 #include <sys/mount.h>
49 #include <sys/mutex.h>
50 #include <sys/pmc.h>
51 #include <sys/pmckern.h>
52 #include <sys/pmclog.h>
53 #include <sys/priv.h>
54 #include <sys/proc.h>
55 #include <sys/queue.h>
56 #include <sys/resourcevar.h>
57 #include <sys/rwlock.h>
58 #include <sys/sched.h>
59 #include <sys/signalvar.h>
60 #include <sys/smp.h>
61 #include <sys/sx.h>
62 #include <sys/sysctl.h>
63 #include <sys/sysent.h>
64 #include <sys/syslog.h>
65 #include <sys/systm.h>
66 #include <sys/vnode.h>
67 
68 #include <sys/linker.h>		/* needs to be after <sys/malloc.h> */
69 
70 #include <machine/atomic.h>
71 #include <machine/md_var.h>
72 
73 #include <vm/vm.h>
74 #include <vm/vm_extern.h>
75 #include <vm/pmap.h>
76 #include <vm/vm_map.h>
77 #include <vm/vm_object.h>
78 
79 #include "hwpmc_soft.h"
80 
81 #ifdef NUMA
82 #define NDOMAINS vm_ndomains
83 #else
84 #define NDOMAINS 1
85 #define malloc_domain(size, type, domain, flags) malloc((size), (type), (flags))
86 #define free_domain(addr, type) free(addr, type)
87 #endif
88 
89 #define PMC_EPOCH_ENTER() struct epoch_tracker pmc_et; epoch_enter_preempt(global_epoch_preempt, &pmc_et)
90 #define PMC_EPOCH_EXIT() epoch_exit_preempt(global_epoch_preempt, &pmc_et)
91 
92 /*
93  * Types
94  */
95 
96 enum pmc_flags {
97 	PMC_FLAG_NONE	  = 0x00, /* do nothing */
98 	PMC_FLAG_REMOVE   = 0x01, /* atomically remove entry from hash */
99 	PMC_FLAG_ALLOCATE = 0x02, /* add entry to hash if not found */
100 	PMC_FLAG_NOWAIT   = 0x04, /* do not wait for mallocs */
101 };
102 
103 /*
104  * The offset in sysent where the syscall is allocated.
105  */
106 
107 static int pmc_syscall_num = NO_SYSCALL;
108 struct pmc_cpu		**pmc_pcpu;	 /* per-cpu state */
109 pmc_value_t		*pmc_pcpu_saved; /* saved PMC values: CSW handling */
110 
111 #define	PMC_PCPU_SAVED(C,R)	pmc_pcpu_saved[(R) + md->pmd_npmc*(C)]
112 
113 struct mtx_pool		*pmc_mtxpool;
114 static int		*pmc_pmcdisp;	 /* PMC row dispositions */
115 
116 #define	PMC_ROW_DISP_IS_FREE(R)		(pmc_pmcdisp[(R)] == 0)
117 #define	PMC_ROW_DISP_IS_THREAD(R)	(pmc_pmcdisp[(R)] > 0)
118 #define	PMC_ROW_DISP_IS_STANDALONE(R)	(pmc_pmcdisp[(R)] < 0)
119 
120 #define	PMC_MARK_ROW_FREE(R) do {					  \
121 	pmc_pmcdisp[(R)] = 0;						  \
122 } while (0)
123 
124 #define	PMC_MARK_ROW_STANDALONE(R) do {					  \
125 	KASSERT(pmc_pmcdisp[(R)] <= 0, ("[pmc,%d] row disposition error", \
126 		    __LINE__));						  \
127 	atomic_add_int(&pmc_pmcdisp[(R)], -1);				  \
128 	KASSERT(pmc_pmcdisp[(R)] >= (-pmc_cpu_max_active()),		  \
129 		("[pmc,%d] row disposition error", __LINE__));		  \
130 } while (0)
131 
132 #define	PMC_UNMARK_ROW_STANDALONE(R) do { 				  \
133 	atomic_add_int(&pmc_pmcdisp[(R)], 1);				  \
134 	KASSERT(pmc_pmcdisp[(R)] <= 0, ("[pmc,%d] row disposition error", \
135 		    __LINE__));						  \
136 } while (0)
137 
138 #define	PMC_MARK_ROW_THREAD(R) do {					  \
139 	KASSERT(pmc_pmcdisp[(R)] >= 0, ("[pmc,%d] row disposition error", \
140 		    __LINE__));						  \
141 	atomic_add_int(&pmc_pmcdisp[(R)], 1);				  \
142 } while (0)
143 
144 #define	PMC_UNMARK_ROW_THREAD(R) do {					  \
145 	atomic_add_int(&pmc_pmcdisp[(R)], -1);				  \
146 	KASSERT(pmc_pmcdisp[(R)] >= 0, ("[pmc,%d] row disposition error", \
147 		    __LINE__));						  \
148 } while (0)
149 
150 
151 /* various event handlers */
152 static eventhandler_tag	pmc_exit_tag, pmc_fork_tag, pmc_kld_load_tag,
153     pmc_kld_unload_tag;
154 
155 /* Module statistics */
156 struct pmc_driverstats pmc_stats;
157 
158 
159 /* Machine/processor dependent operations */
160 static struct pmc_mdep  *md;
161 
162 /*
163  * Hash tables mapping owner processes and target threads to PMCs.
164  */
165 
166 struct mtx pmc_processhash_mtx;		/* spin mutex */
167 static u_long pmc_processhashmask;
168 static LIST_HEAD(pmc_processhash, pmc_process)	*pmc_processhash;
169 
170 /*
171  * Hash table of PMC owner descriptors.  This table is protected by
172  * the shared PMC "sx" lock.
173  */
174 
175 static u_long pmc_ownerhashmask;
176 static LIST_HEAD(pmc_ownerhash, pmc_owner)	*pmc_ownerhash;
177 
178 /*
179  * List of PMC owners with system-wide sampling PMCs.
180  */
181 
182 static CK_LIST_HEAD(, pmc_owner)			pmc_ss_owners;
183 
184 /*
185  * List of free thread entries. This is protected by the spin
186  * mutex.
187  */
188 static struct mtx pmc_threadfreelist_mtx;	/* spin mutex */
189 static LIST_HEAD(, pmc_thread)			pmc_threadfreelist;
190 static int pmc_threadfreelist_entries=0;
191 #define	THREADENTRY_SIZE						\
192 (sizeof(struct pmc_thread) + (md->pmd_npmc * sizeof(struct pmc_threadpmcstate)))
193 
194 /*
195  * Task to free thread descriptors
196  */
197 static struct grouptask free_gtask;
198 
199 /*
200  * A map of row indices to classdep structures.
201  */
202 static struct pmc_classdep **pmc_rowindex_to_classdep;
203 
204 /*
205  * Prototypes
206  */
207 
208 #ifdef	HWPMC_DEBUG
209 static int	pmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS);
210 static int	pmc_debugflags_parse(char *newstr, char *fence);
211 #endif
212 
213 static int	load(struct module *module, int cmd, void *arg);
214 static int	pmc_add_sample(ring_type_t ring, struct pmc *pm, struct trapframe *tf);
215 static void	pmc_add_thread_descriptors_from_proc(struct proc *p,
216     struct pmc_process *pp);
217 static int	pmc_attach_process(struct proc *p, struct pmc *pm);
218 static struct pmc *pmc_allocate_pmc_descriptor(void);
219 static struct pmc_owner *pmc_allocate_owner_descriptor(struct proc *p);
220 static int	pmc_attach_one_process(struct proc *p, struct pmc *pm);
221 static int	pmc_can_allocate_rowindex(struct proc *p, unsigned int ri,
222     int cpu);
223 static int	pmc_can_attach(struct pmc *pm, struct proc *p);
224 static void	pmc_capture_user_callchain(int cpu, int soft, struct trapframe *tf);
225 static void	pmc_cleanup(void);
226 static int	pmc_detach_process(struct proc *p, struct pmc *pm);
227 static int	pmc_detach_one_process(struct proc *p, struct pmc *pm,
228     int flags);
229 static void	pmc_destroy_owner_descriptor(struct pmc_owner *po);
230 static void	pmc_destroy_pmc_descriptor(struct pmc *pm);
231 static void	pmc_destroy_process_descriptor(struct pmc_process *pp);
232 static struct pmc_owner *pmc_find_owner_descriptor(struct proc *p);
233 static int	pmc_find_pmc(pmc_id_t pmcid, struct pmc **pm);
234 static struct pmc *pmc_find_pmc_descriptor_in_process(struct pmc_owner *po,
235     pmc_id_t pmc);
236 static struct pmc_process *pmc_find_process_descriptor(struct proc *p,
237     uint32_t mode);
238 static struct pmc_thread *pmc_find_thread_descriptor(struct pmc_process *pp,
239     struct thread *td, uint32_t mode);
240 static void	pmc_force_context_switch(void);
241 static void	pmc_link_target_process(struct pmc *pm,
242     struct pmc_process *pp);
243 static void	pmc_log_all_process_mappings(struct pmc_owner *po);
244 static void	pmc_log_kernel_mappings(struct pmc *pm);
245 static void	pmc_log_process_mappings(struct pmc_owner *po, struct proc *p);
246 static void	pmc_maybe_remove_owner(struct pmc_owner *po);
247 static void	pmc_process_csw_in(struct thread *td);
248 static void	pmc_process_csw_out(struct thread *td);
249 static void	pmc_process_exit(void *arg, struct proc *p);
250 static void	pmc_process_fork(void *arg, struct proc *p1,
251     struct proc *p2, int n);
252 static void	pmc_process_samples(int cpu, ring_type_t soft);
253 static void	pmc_release_pmc_descriptor(struct pmc *pmc);
254 static void	pmc_process_thread_add(struct thread *td);
255 static void	pmc_process_thread_delete(struct thread *td);
256 static void	pmc_process_thread_userret(struct thread *td);
257 static void	pmc_remove_owner(struct pmc_owner *po);
258 static void	pmc_remove_process_descriptor(struct pmc_process *pp);
259 static void	pmc_restore_cpu_binding(struct pmc_binding *pb);
260 static void	pmc_save_cpu_binding(struct pmc_binding *pb);
261 static void	pmc_select_cpu(int cpu);
262 static int	pmc_start(struct pmc *pm);
263 static int	pmc_stop(struct pmc *pm);
264 static int	pmc_syscall_handler(struct thread *td, void *syscall_args);
265 static struct pmc_thread *pmc_thread_descriptor_pool_alloc(void);
266 static void	pmc_thread_descriptor_pool_drain(void);
267 static void	pmc_thread_descriptor_pool_free(struct pmc_thread *pt);
268 static void	pmc_unlink_target_process(struct pmc *pmc,
269     struct pmc_process *pp);
270 static int generic_switch_in(struct pmc_cpu *pc, struct pmc_process *pp);
271 static int generic_switch_out(struct pmc_cpu *pc, struct pmc_process *pp);
272 static struct pmc_mdep *pmc_generic_cpu_initialize(void);
273 static void pmc_generic_cpu_finalize(struct pmc_mdep *md);
274 static void pmc_post_callchain_callback(void);
275 static void pmc_process_threadcreate(struct thread *td);
276 static void pmc_process_threadexit(struct thread *td);
277 static void pmc_process_proccreate(struct proc *p);
278 static void pmc_process_allproc(struct pmc *pm);
279 
280 /*
281  * Kernel tunables and sysctl(8) interface.
282  */
283 
284 SYSCTL_DECL(_kern_hwpmc);
285 SYSCTL_NODE(_kern_hwpmc, OID_AUTO, stats, CTLFLAG_RW, 0, "HWPMC stats");
286 
287 
288 /* Stats. */
289 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, intr_ignored, CTLFLAG_RW,
290 				   &pmc_stats.pm_intr_ignored, "# of interrupts ignored");
291 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, intr_processed, CTLFLAG_RW,
292 				   &pmc_stats.pm_intr_processed, "# of interrupts processed");
293 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, intr_bufferfull, CTLFLAG_RW,
294 				   &pmc_stats.pm_intr_bufferfull, "# of interrupts where buffer was full");
295 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, syscalls, CTLFLAG_RW,
296 				   &pmc_stats.pm_syscalls, "# of syscalls");
297 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, syscall_errors, CTLFLAG_RW,
298 				   &pmc_stats.pm_syscall_errors, "# of syscall_errors");
299 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, buffer_requests, CTLFLAG_RW,
300 				   &pmc_stats.pm_buffer_requests, "# of buffer requests");
301 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, buffer_requests_failed, CTLFLAG_RW,
302 				   &pmc_stats.pm_buffer_requests_failed, "# of buffer requests which failed");
303 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, log_sweeps, CTLFLAG_RW,
304 				   &pmc_stats.pm_log_sweeps, "# of ?");
305 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, merges, CTLFLAG_RW,
306 				   &pmc_stats.pm_merges, "# of times kernel stack was found for user trace");
307 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, overwrites, CTLFLAG_RW,
308 				   &pmc_stats.pm_overwrites, "# of times a sample was overwritten before being logged");
309 
310 static int pmc_callchaindepth = PMC_CALLCHAIN_DEPTH;
311 SYSCTL_INT(_kern_hwpmc, OID_AUTO, callchaindepth, CTLFLAG_RDTUN,
312     &pmc_callchaindepth, 0, "depth of call chain records");
313 
314 char pmc_cpuid[64];
315 SYSCTL_STRING(_kern_hwpmc, OID_AUTO, cpuid, CTLFLAG_RD,
316 	pmc_cpuid, 0, "cpu version string");
317 #ifdef	HWPMC_DEBUG
318 struct pmc_debugflags pmc_debugflags = PMC_DEBUG_DEFAULT_FLAGS;
319 char	pmc_debugstr[PMC_DEBUG_STRSIZE];
320 TUNABLE_STR(PMC_SYSCTL_NAME_PREFIX "debugflags", pmc_debugstr,
321     sizeof(pmc_debugstr));
322 SYSCTL_PROC(_kern_hwpmc, OID_AUTO, debugflags,
323     CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NOFETCH,
324     0, 0, pmc_debugflags_sysctl_handler, "A", "debug flags");
325 #endif
326 
327 
328 /*
329  * kern.hwpmc.hashrows -- determines the number of rows in the
330  * of the hash table used to look up threads
331  */
332 
333 static int pmc_hashsize = PMC_HASH_SIZE;
334 SYSCTL_INT(_kern_hwpmc, OID_AUTO, hashsize, CTLFLAG_RDTUN,
335     &pmc_hashsize, 0, "rows in hash tables");
336 
337 /*
338  * kern.hwpmc.nsamples --- number of PC samples/callchain stacks per CPU
339  */
340 
341 static int pmc_nsamples = PMC_NSAMPLES;
342 SYSCTL_INT(_kern_hwpmc, OID_AUTO, nsamples, CTLFLAG_RDTUN,
343     &pmc_nsamples, 0, "number of PC samples per CPU");
344 
345 static uint64_t pmc_sample_mask = PMC_NSAMPLES-1;
346 
347 /*
348  * kern.hwpmc.mtxpoolsize -- number of mutexes in the mutex pool.
349  */
350 
351 static int pmc_mtxpool_size = PMC_MTXPOOL_SIZE;
352 SYSCTL_INT(_kern_hwpmc, OID_AUTO, mtxpoolsize, CTLFLAG_RDTUN,
353     &pmc_mtxpool_size, 0, "size of spin mutex pool");
354 
355 
356 /*
357  * kern.hwpmc.threadfreelist_entries -- number of free entries
358  */
359 
360 SYSCTL_INT(_kern_hwpmc, OID_AUTO, threadfreelist_entries, CTLFLAG_RD,
361     &pmc_threadfreelist_entries, 0, "number of avalable thread entries");
362 
363 
364 /*
365  * kern.hwpmc.threadfreelist_max -- maximum number of free entries
366  */
367 
368 static int pmc_threadfreelist_max = PMC_THREADLIST_MAX;
369 SYSCTL_INT(_kern_hwpmc, OID_AUTO, threadfreelist_max, CTLFLAG_RW,
370     &pmc_threadfreelist_max, 0,
371     "maximum number of available thread entries before freeing some");
372 
373 
374 /*
375  * security.bsd.unprivileged_syspmcs -- allow non-root processes to
376  * allocate system-wide PMCs.
377  *
378  * Allowing unprivileged processes to allocate system PMCs is convenient
379  * if system-wide measurements need to be taken concurrently with other
380  * per-process measurements.  This feature is turned off by default.
381  */
382 
383 static int pmc_unprivileged_syspmcs = 0;
384 SYSCTL_INT(_security_bsd, OID_AUTO, unprivileged_syspmcs, CTLFLAG_RWTUN,
385     &pmc_unprivileged_syspmcs, 0,
386     "allow unprivileged process to allocate system PMCs");
387 
388 /*
389  * Hash function.  Discard the lower 2 bits of the pointer since
390  * these are always zero for our uses.  The hash multiplier is
391  * round((2^LONG_BIT) * ((sqrt(5)-1)/2)).
392  */
393 
394 #if	LONG_BIT == 64
395 #define	_PMC_HM		11400714819323198486u
396 #elif	LONG_BIT == 32
397 #define	_PMC_HM		2654435769u
398 #else
399 #error 	Must know the size of 'long' to compile
400 #endif
401 
402 #define	PMC_HASH_PTR(P,M)	((((unsigned long) (P) >> 2) * _PMC_HM) & (M))
403 
404 /*
405  * Syscall structures
406  */
407 
408 /* The `sysent' for the new syscall */
409 static struct sysent pmc_sysent = {
410 	.sy_narg =	2,
411 	.sy_call =	pmc_syscall_handler,
412 };
413 
414 static struct syscall_module_data pmc_syscall_mod = {
415 	.chainevh =	load,
416 	.chainarg =	NULL,
417 	.offset =	&pmc_syscall_num,
418 	.new_sysent =	&pmc_sysent,
419 	.old_sysent =	{ .sy_narg = 0, .sy_call = NULL },
420 	.flags =	SY_THR_STATIC_KLD,
421 };
422 
423 static moduledata_t pmc_mod = {
424 	.name =		PMC_MODULE_NAME,
425 	.evhand =	syscall_module_handler,
426 	.priv =		&pmc_syscall_mod,
427 };
428 
429 #ifdef EARLY_AP_STARTUP
430 DECLARE_MODULE(pmc, pmc_mod, SI_SUB_SYSCALLS, SI_ORDER_ANY);
431 #else
432 DECLARE_MODULE(pmc, pmc_mod, SI_SUB_SMP, SI_ORDER_ANY);
433 #endif
434 MODULE_VERSION(pmc, PMC_VERSION);
435 
436 #ifdef	HWPMC_DEBUG
437 enum pmc_dbgparse_state {
438 	PMCDS_WS,		/* in whitespace */
439 	PMCDS_MAJOR,		/* seen a major keyword */
440 	PMCDS_MINOR
441 };
442 
443 static int
444 pmc_debugflags_parse(char *newstr, char *fence)
445 {
446 	char c, *p, *q;
447 	struct pmc_debugflags *tmpflags;
448 	int error, found, *newbits, tmp;
449 	size_t kwlen;
450 
451 	tmpflags = malloc(sizeof(*tmpflags), M_PMC, M_WAITOK|M_ZERO);
452 
453 	p = newstr;
454 	error = 0;
455 
456 	for (; p < fence && (c = *p); p++) {
457 
458 		/* skip white space */
459 		if (c == ' ' || c == '\t')
460 			continue;
461 
462 		/* look for a keyword followed by "=" */
463 		for (q = p; p < fence && (c = *p) && c != '='; p++)
464 			;
465 		if (c != '=') {
466 			error = EINVAL;
467 			goto done;
468 		}
469 
470 		kwlen = p - q;
471 		newbits = NULL;
472 
473 		/* lookup flag group name */
474 #define	DBG_SET_FLAG_MAJ(S,F)						\
475 		if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0)	\
476 			newbits = &tmpflags->pdb_ ## F;
477 
478 		DBG_SET_FLAG_MAJ("cpu",		CPU);
479 		DBG_SET_FLAG_MAJ("csw",		CSW);
480 		DBG_SET_FLAG_MAJ("logging",	LOG);
481 		DBG_SET_FLAG_MAJ("module",	MOD);
482 		DBG_SET_FLAG_MAJ("md", 		MDP);
483 		DBG_SET_FLAG_MAJ("owner",	OWN);
484 		DBG_SET_FLAG_MAJ("pmc",		PMC);
485 		DBG_SET_FLAG_MAJ("process",	PRC);
486 		DBG_SET_FLAG_MAJ("sampling", 	SAM);
487 
488 		if (newbits == NULL) {
489 			error = EINVAL;
490 			goto done;
491 		}
492 
493 		p++;		/* skip the '=' */
494 
495 		/* Now parse the individual flags */
496 		tmp = 0;
497 	newflag:
498 		for (q = p; p < fence && (c = *p); p++)
499 			if (c == ' ' || c == '\t' || c == ',')
500 				break;
501 
502 		/* p == fence or c == ws or c == "," or c == 0 */
503 
504 		if ((kwlen = p - q) == 0) {
505 			*newbits = tmp;
506 			continue;
507 		}
508 
509 		found = 0;
510 #define	DBG_SET_FLAG_MIN(S,F)						\
511 		if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0)	\
512 			tmp |= found = (1 << PMC_DEBUG_MIN_ ## F)
513 
514 		/* a '*' denotes all possible flags in the group */
515 		if (kwlen == 1 && *q == '*')
516 			tmp = found = ~0;
517 		/* look for individual flag names */
518 		DBG_SET_FLAG_MIN("allocaterow", ALR);
519 		DBG_SET_FLAG_MIN("allocate",	ALL);
520 		DBG_SET_FLAG_MIN("attach",	ATT);
521 		DBG_SET_FLAG_MIN("bind",	BND);
522 		DBG_SET_FLAG_MIN("config",	CFG);
523 		DBG_SET_FLAG_MIN("exec",	EXC);
524 		DBG_SET_FLAG_MIN("exit",	EXT);
525 		DBG_SET_FLAG_MIN("find",	FND);
526 		DBG_SET_FLAG_MIN("flush",	FLS);
527 		DBG_SET_FLAG_MIN("fork",	FRK);
528 		DBG_SET_FLAG_MIN("getbuf",	GTB);
529 		DBG_SET_FLAG_MIN("hook",	PMH);
530 		DBG_SET_FLAG_MIN("init",	INI);
531 		DBG_SET_FLAG_MIN("intr",	INT);
532 		DBG_SET_FLAG_MIN("linktarget",	TLK);
533 		DBG_SET_FLAG_MIN("mayberemove", OMR);
534 		DBG_SET_FLAG_MIN("ops",		OPS);
535 		DBG_SET_FLAG_MIN("read",	REA);
536 		DBG_SET_FLAG_MIN("register",	REG);
537 		DBG_SET_FLAG_MIN("release",	REL);
538 		DBG_SET_FLAG_MIN("remove",	ORM);
539 		DBG_SET_FLAG_MIN("sample",	SAM);
540 		DBG_SET_FLAG_MIN("scheduleio",	SIO);
541 		DBG_SET_FLAG_MIN("select",	SEL);
542 		DBG_SET_FLAG_MIN("signal",	SIG);
543 		DBG_SET_FLAG_MIN("swi",		SWI);
544 		DBG_SET_FLAG_MIN("swo",		SWO);
545 		DBG_SET_FLAG_MIN("start",	STA);
546 		DBG_SET_FLAG_MIN("stop",	STO);
547 		DBG_SET_FLAG_MIN("syscall",	PMS);
548 		DBG_SET_FLAG_MIN("unlinktarget", TUL);
549 		DBG_SET_FLAG_MIN("write",	WRI);
550 		if (found == 0) {
551 			/* unrecognized flag name */
552 			error = EINVAL;
553 			goto done;
554 		}
555 
556 		if (c == 0 || c == ' ' || c == '\t') {	/* end of flag group */
557 			*newbits = tmp;
558 			continue;
559 		}
560 
561 		p++;
562 		goto newflag;
563 	}
564 
565 	/* save the new flag set */
566 	bcopy(tmpflags, &pmc_debugflags, sizeof(pmc_debugflags));
567 
568  done:
569 	free(tmpflags, M_PMC);
570 	return error;
571 }
572 
573 static int
574 pmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS)
575 {
576 	char *fence, *newstr;
577 	int error;
578 	unsigned int n;
579 
580 	(void) arg1; (void) arg2; /* unused parameters */
581 
582 	n = sizeof(pmc_debugstr);
583 	newstr = malloc(n, M_PMC, M_WAITOK|M_ZERO);
584 	(void) strlcpy(newstr, pmc_debugstr, n);
585 
586 	error = sysctl_handle_string(oidp, newstr, n, req);
587 
588 	/* if there is a new string, parse and copy it */
589 	if (error == 0 && req->newptr != NULL) {
590 		fence = newstr + (n < req->newlen ? n : req->newlen + 1);
591 		if ((error = pmc_debugflags_parse(newstr, fence)) == 0)
592 			(void) strlcpy(pmc_debugstr, newstr,
593 			    sizeof(pmc_debugstr));
594 	}
595 
596 	free(newstr, M_PMC);
597 
598 	return error;
599 }
600 #endif
601 
602 /*
603  * Map a row index to a classdep structure and return the adjusted row
604  * index for the PMC class index.
605  */
606 static struct pmc_classdep *
607 pmc_ri_to_classdep(struct pmc_mdep *md, int ri, int *adjri)
608 {
609 	struct pmc_classdep *pcd;
610 
611 	(void) md;
612 
613 	KASSERT(ri >= 0 && ri < md->pmd_npmc,
614 	    ("[pmc,%d] illegal row-index %d", __LINE__, ri));
615 
616 	pcd = pmc_rowindex_to_classdep[ri];
617 
618 	KASSERT(pcd != NULL,
619 	    ("[pmc,%d] ri %d null pcd", __LINE__, ri));
620 
621 	*adjri = ri - pcd->pcd_ri;
622 
623 	KASSERT(*adjri >= 0 && *adjri < pcd->pcd_num,
624 	    ("[pmc,%d] adjusted row-index %d", __LINE__, *adjri));
625 
626 	return (pcd);
627 }
628 
629 /*
630  * Concurrency Control
631  *
632  * The driver manages the following data structures:
633  *
634  *   - target process descriptors, one per target process
635  *   - owner process descriptors (and attached lists), one per owner process
636  *   - lookup hash tables for owner and target processes
637  *   - PMC descriptors (and attached lists)
638  *   - per-cpu hardware state
639  *   - the 'hook' variable through which the kernel calls into
640  *     this module
641  *   - the machine hardware state (managed by the MD layer)
642  *
643  * These data structures are accessed from:
644  *
645  * - thread context-switch code
646  * - interrupt handlers (possibly on multiple cpus)
647  * - kernel threads on multiple cpus running on behalf of user
648  *   processes doing system calls
649  * - this driver's private kernel threads
650  *
651  * = Locks and Locking strategy =
652  *
653  * The driver uses four locking strategies for its operation:
654  *
655  * - The global SX lock "pmc_sx" is used to protect internal
656  *   data structures.
657  *
658  *   Calls into the module by syscall() start with this lock being
659  *   held in exclusive mode.  Depending on the requested operation,
660  *   the lock may be downgraded to 'shared' mode to allow more
661  *   concurrent readers into the module.  Calls into the module from
662  *   other parts of the kernel acquire the lock in shared mode.
663  *
664  *   This SX lock is held in exclusive mode for any operations that
665  *   modify the linkages between the driver's internal data structures.
666  *
667  *   The 'pmc_hook' function pointer is also protected by this lock.
668  *   It is only examined with the sx lock held in exclusive mode.  The
669  *   kernel module is allowed to be unloaded only with the sx lock held
670  *   in exclusive mode.  In normal syscall handling, after acquiring the
671  *   pmc_sx lock we first check that 'pmc_hook' is non-null before
672  *   proceeding.  This prevents races between the thread unloading the module
673  *   and other threads seeking to use the module.
674  *
675  * - Lookups of target process structures and owner process structures
676  *   cannot use the global "pmc_sx" SX lock because these lookups need
677  *   to happen during context switches and in other critical sections
678  *   where sleeping is not allowed.  We protect these lookup tables
679  *   with their own private spin-mutexes, "pmc_processhash_mtx" and
680  *   "pmc_ownerhash_mtx".
681  *
682  * - Interrupt handlers work in a lock free manner.  At interrupt
683  *   time, handlers look at the PMC pointer (phw->phw_pmc) configured
684  *   when the PMC was started.  If this pointer is NULL, the interrupt
685  *   is ignored after updating driver statistics.  We ensure that this
686  *   pointer is set (using an atomic operation if necessary) before the
687  *   PMC hardware is started.  Conversely, this pointer is unset atomically
688  *   only after the PMC hardware is stopped.
689  *
690  *   We ensure that everything needed for the operation of an
691  *   interrupt handler is available without it needing to acquire any
692  *   locks.  We also ensure that a PMC's software state is destroyed only
693  *   after the PMC is taken off hardware (on all CPUs).
694  *
695  * - Context-switch handling with process-private PMCs needs more
696  *   care.
697  *
698  *   A given process may be the target of multiple PMCs.  For example,
699  *   PMCATTACH and PMCDETACH may be requested by a process on one CPU
700  *   while the target process is running on another.  A PMC could also
701  *   be getting released because its owner is exiting.  We tackle
702  *   these situations in the following manner:
703  *
704  *   - each target process structure 'pmc_process' has an array
705  *     of 'struct pmc *' pointers, one for each hardware PMC.
706  *
707  *   - At context switch IN time, each "target" PMC in RUNNING state
708  *     gets started on hardware and a pointer to each PMC is copied into
709  *     the per-cpu phw array.  The 'runcount' for the PMC is
710  *     incremented.
711  *
712  *   - At context switch OUT time, all process-virtual PMCs are stopped
713  *     on hardware.  The saved value is added to the PMCs value field
714  *     only if the PMC is in a non-deleted state (the PMCs state could
715  *     have changed during the current time slice).
716  *
717  *     Note that since in-between a switch IN on a processor and a switch
718  *     OUT, the PMC could have been released on another CPU.  Therefore
719  *     context switch OUT always looks at the hardware state to turn
720  *     OFF PMCs and will update a PMC's saved value only if reachable
721  *     from the target process record.
722  *
723  *   - OP PMCRELEASE could be called on a PMC at any time (the PMC could
724  *     be attached to many processes at the time of the call and could
725  *     be active on multiple CPUs).
726  *
727  *     We prevent further scheduling of the PMC by marking it as in
728  *     state 'DELETED'.  If the runcount of the PMC is non-zero then
729  *     this PMC is currently running on a CPU somewhere.  The thread
730  *     doing the PMCRELEASE operation waits by repeatedly doing a
731  *     pause() till the runcount comes to zero.
732  *
733  * The contents of a PMC descriptor (struct pmc) are protected using
734  * a spin-mutex.  In order to save space, we use a mutex pool.
735  *
736  * In terms of lock types used by witness(4), we use:
737  * - Type "pmc-sx", used by the global SX lock.
738  * - Type "pmc-sleep", for sleep mutexes used by logger threads.
739  * - Type "pmc-per-proc", for protecting PMC owner descriptors.
740  * - Type "pmc-leaf", used for all other spin mutexes.
741  */
742 
743 /*
744  * save the cpu binding of the current kthread
745  */
746 
747 static void
748 pmc_save_cpu_binding(struct pmc_binding *pb)
749 {
750 	PMCDBG0(CPU,BND,2, "save-cpu");
751 	thread_lock(curthread);
752 	pb->pb_bound = sched_is_bound(curthread);
753 	pb->pb_cpu   = curthread->td_oncpu;
754 	thread_unlock(curthread);
755 	PMCDBG1(CPU,BND,2, "save-cpu cpu=%d", pb->pb_cpu);
756 }
757 
758 /*
759  * restore the cpu binding of the current thread
760  */
761 
762 static void
763 pmc_restore_cpu_binding(struct pmc_binding *pb)
764 {
765 	PMCDBG2(CPU,BND,2, "restore-cpu curcpu=%d restore=%d",
766 	    curthread->td_oncpu, pb->pb_cpu);
767 	thread_lock(curthread);
768 	if (pb->pb_bound)
769 		sched_bind(curthread, pb->pb_cpu);
770 	else
771 		sched_unbind(curthread);
772 	thread_unlock(curthread);
773 	PMCDBG0(CPU,BND,2, "restore-cpu done");
774 }
775 
776 /*
777  * move execution over the specified cpu and bind it there.
778  */
779 
780 static void
781 pmc_select_cpu(int cpu)
782 {
783 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
784 	    ("[pmc,%d] bad cpu number %d", __LINE__, cpu));
785 
786 	/* Never move to an inactive CPU. */
787 	KASSERT(pmc_cpu_is_active(cpu), ("[pmc,%d] selecting inactive "
788 	    "CPU %d", __LINE__, cpu));
789 
790 	PMCDBG1(CPU,SEL,2, "select-cpu cpu=%d", cpu);
791 	thread_lock(curthread);
792 	sched_bind(curthread, cpu);
793 	thread_unlock(curthread);
794 
795 	KASSERT(curthread->td_oncpu == cpu,
796 	    ("[pmc,%d] CPU not bound [cpu=%d, curr=%d]", __LINE__,
797 		cpu, curthread->td_oncpu));
798 
799 	PMCDBG1(CPU,SEL,2, "select-cpu cpu=%d ok", cpu);
800 }
801 
802 /*
803  * Force a context switch.
804  *
805  * We do this by pause'ing for 1 tick -- invoking mi_switch() is not
806  * guaranteed to force a context switch.
807  */
808 
809 static void
810 pmc_force_context_switch(void)
811 {
812 
813 	pause("pmcctx", 1);
814 }
815 
816 uint64_t
817 pmc_rdtsc(void)
818 {
819 #if defined(__i386__) || defined(__amd64__)
820 	if (__predict_true(amd_feature & AMDID_RDTSCP))
821 		return rdtscp();
822 	else
823 		return rdtsc();
824 #else
825 	return get_cyclecount();
826 #endif
827 }
828 
829 /*
830  * Get the file name for an executable.  This is a simple wrapper
831  * around vn_fullpath(9).
832  */
833 
834 static void
835 pmc_getfilename(struct vnode *v, char **fullpath, char **freepath)
836 {
837 
838 	*fullpath = "unknown";
839 	*freepath = NULL;
840 	vn_fullpath(curthread, v, fullpath, freepath);
841 }
842 
843 /*
844  * remove an process owning PMCs
845  */
846 
847 void
848 pmc_remove_owner(struct pmc_owner *po)
849 {
850 	struct pmc *pm, *tmp;
851 
852 	sx_assert(&pmc_sx, SX_XLOCKED);
853 
854 	PMCDBG1(OWN,ORM,1, "remove-owner po=%p", po);
855 
856 	/* Remove descriptor from the owner hash table */
857 	LIST_REMOVE(po, po_next);
858 
859 	/* release all owned PMC descriptors */
860 	LIST_FOREACH_SAFE(pm, &po->po_pmcs, pm_next, tmp) {
861 		PMCDBG1(OWN,ORM,2, "pmc=%p", pm);
862 		KASSERT(pm->pm_owner == po,
863 		    ("[pmc,%d] owner %p != po %p", __LINE__, pm->pm_owner, po));
864 
865 		pmc_release_pmc_descriptor(pm);	/* will unlink from the list */
866 		pmc_destroy_pmc_descriptor(pm);
867 	}
868 
869 	KASSERT(po->po_sscount == 0,
870 	    ("[pmc,%d] SS count not zero", __LINE__));
871 	KASSERT(LIST_EMPTY(&po->po_pmcs),
872 	    ("[pmc,%d] PMC list not empty", __LINE__));
873 
874 	/* de-configure the log file if present */
875 	if (po->po_flags & PMC_PO_OWNS_LOGFILE)
876 		pmclog_deconfigure_log(po);
877 }
878 
879 /*
880  * remove an owner process record if all conditions are met.
881  */
882 
883 static void
884 pmc_maybe_remove_owner(struct pmc_owner *po)
885 {
886 
887 	PMCDBG1(OWN,OMR,1, "maybe-remove-owner po=%p", po);
888 
889 	/*
890 	 * Remove owner record if
891 	 * - this process does not own any PMCs
892 	 * - this process has not allocated a system-wide sampling buffer
893 	 */
894 
895 	if (LIST_EMPTY(&po->po_pmcs) &&
896 	    ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)) {
897 		pmc_remove_owner(po);
898 		pmc_destroy_owner_descriptor(po);
899 	}
900 }
901 
902 /*
903  * Add an association between a target process and a PMC.
904  */
905 
906 static void
907 pmc_link_target_process(struct pmc *pm, struct pmc_process *pp)
908 {
909 	int ri;
910 	struct pmc_target *pt;
911 #ifdef INVARIANTS
912 	struct pmc_thread *pt_td;
913 #endif
914 
915 	sx_assert(&pmc_sx, SX_XLOCKED);
916 
917 	KASSERT(pm != NULL && pp != NULL,
918 	    ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp));
919 	KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)),
920 	    ("[pmc,%d] Attaching a non-process-virtual pmc=%p to pid=%d",
921 		__LINE__, pm, pp->pp_proc->p_pid));
922 	KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= ((int) md->pmd_npmc - 1),
923 	    ("[pmc,%d] Illegal reference count %d for process record %p",
924 		__LINE__, pp->pp_refcnt, (void *) pp));
925 
926 	ri = PMC_TO_ROWINDEX(pm);
927 
928 	PMCDBG3(PRC,TLK,1, "link-target pmc=%p ri=%d pmc-process=%p",
929 	    pm, ri, pp);
930 
931 #ifdef	HWPMC_DEBUG
932 	LIST_FOREACH(pt, &pm->pm_targets, pt_next)
933 	    if (pt->pt_process == pp)
934 		    KASSERT(0, ("[pmc,%d] pp %p already in pmc %p targets",
935 				__LINE__, pp, pm));
936 #endif
937 
938 	pt = malloc(sizeof(struct pmc_target), M_PMC, M_WAITOK|M_ZERO);
939 	pt->pt_process = pp;
940 
941 	LIST_INSERT_HEAD(&pm->pm_targets, pt, pt_next);
942 
943 	atomic_store_rel_ptr((uintptr_t *)&pp->pp_pmcs[ri].pp_pmc,
944 	    (uintptr_t)pm);
945 
946 	if (pm->pm_owner->po_owner == pp->pp_proc)
947 		pm->pm_flags |= PMC_F_ATTACHED_TO_OWNER;
948 
949 	/*
950 	 * Initialize the per-process values at this row index.
951 	 */
952 	pp->pp_pmcs[ri].pp_pmcval = PMC_TO_MODE(pm) == PMC_MODE_TS ?
953 	    pm->pm_sc.pm_reloadcount : 0;
954 
955 	pp->pp_refcnt++;
956 
957 #ifdef INVARIANTS
958 	/* Confirm that the per-thread values at this row index are cleared. */
959 	if (PMC_TO_MODE(pm) == PMC_MODE_TS) {
960 		mtx_lock_spin(pp->pp_tdslock);
961 		LIST_FOREACH(pt_td, &pp->pp_tds, pt_next) {
962 			KASSERT(pt_td->pt_pmcs[ri].pt_pmcval == (pmc_value_t) 0,
963 			    ("[pmc,%d] pt_pmcval not cleared for pid=%d at "
964 			    "ri=%d", __LINE__, pp->pp_proc->p_pid, ri));
965 		}
966 		mtx_unlock_spin(pp->pp_tdslock);
967 	}
968 #endif
969 }
970 
971 /*
972  * Removes the association between a target process and a PMC.
973  */
974 
975 static void
976 pmc_unlink_target_process(struct pmc *pm, struct pmc_process *pp)
977 {
978 	int ri;
979 	struct proc *p;
980 	struct pmc_target *ptgt;
981 	struct pmc_thread *pt;
982 
983 	sx_assert(&pmc_sx, SX_XLOCKED);
984 
985 	KASSERT(pm != NULL && pp != NULL,
986 	    ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp));
987 
988 	KASSERT(pp->pp_refcnt >= 1 && pp->pp_refcnt <= (int) md->pmd_npmc,
989 	    ("[pmc,%d] Illegal ref count %d on process record %p",
990 		__LINE__, pp->pp_refcnt, (void *) pp));
991 
992 	ri = PMC_TO_ROWINDEX(pm);
993 
994 	PMCDBG3(PRC,TUL,1, "unlink-target pmc=%p ri=%d pmc-process=%p",
995 	    pm, ri, pp);
996 
997 	KASSERT(pp->pp_pmcs[ri].pp_pmc == pm,
998 	    ("[pmc,%d] PMC ri %d mismatch pmc %p pp->[ri] %p", __LINE__,
999 		ri, pm, pp->pp_pmcs[ri].pp_pmc));
1000 
1001 	pp->pp_pmcs[ri].pp_pmc = NULL;
1002 	pp->pp_pmcs[ri].pp_pmcval = (pmc_value_t) 0;
1003 
1004 	/* Clear the per-thread values at this row index. */
1005 	if (PMC_TO_MODE(pm) == PMC_MODE_TS) {
1006 		mtx_lock_spin(pp->pp_tdslock);
1007 		LIST_FOREACH(pt, &pp->pp_tds, pt_next)
1008 			pt->pt_pmcs[ri].pt_pmcval = (pmc_value_t) 0;
1009 		mtx_unlock_spin(pp->pp_tdslock);
1010 	}
1011 
1012 	/* Remove owner-specific flags */
1013 	if (pm->pm_owner->po_owner == pp->pp_proc) {
1014 		pp->pp_flags &= ~PMC_PP_ENABLE_MSR_ACCESS;
1015 		pm->pm_flags &= ~PMC_F_ATTACHED_TO_OWNER;
1016 	}
1017 
1018 	pp->pp_refcnt--;
1019 
1020 	/* Remove the target process from the PMC structure */
1021 	LIST_FOREACH(ptgt, &pm->pm_targets, pt_next)
1022 		if (ptgt->pt_process == pp)
1023 			break;
1024 
1025 	KASSERT(ptgt != NULL, ("[pmc,%d] process %p (pp: %p) not found "
1026 		    "in pmc %p", __LINE__, pp->pp_proc, pp, pm));
1027 
1028 	LIST_REMOVE(ptgt, pt_next);
1029 	free(ptgt, M_PMC);
1030 
1031 	/* if the PMC now lacks targets, send the owner a SIGIO */
1032 	if (LIST_EMPTY(&pm->pm_targets)) {
1033 		p = pm->pm_owner->po_owner;
1034 		PROC_LOCK(p);
1035 		kern_psignal(p, SIGIO);
1036 		PROC_UNLOCK(p);
1037 
1038 		PMCDBG2(PRC,SIG,2, "signalling proc=%p signal=%d", p,
1039 		    SIGIO);
1040 	}
1041 }
1042 
1043 /*
1044  * Check if PMC 'pm' may be attached to target process 't'.
1045  */
1046 
1047 static int
1048 pmc_can_attach(struct pmc *pm, struct proc *t)
1049 {
1050 	struct proc *o;		/* pmc owner */
1051 	struct ucred *oc, *tc;	/* owner, target credentials */
1052 	int decline_attach, i;
1053 
1054 	/*
1055 	 * A PMC's owner can always attach that PMC to itself.
1056 	 */
1057 
1058 	if ((o = pm->pm_owner->po_owner) == t)
1059 		return 0;
1060 
1061 	PROC_LOCK(o);
1062 	oc = o->p_ucred;
1063 	crhold(oc);
1064 	PROC_UNLOCK(o);
1065 
1066 	PROC_LOCK(t);
1067 	tc = t->p_ucred;
1068 	crhold(tc);
1069 	PROC_UNLOCK(t);
1070 
1071 	/*
1072 	 * The effective uid of the PMC owner should match at least one
1073 	 * of the {effective,real,saved} uids of the target process.
1074 	 */
1075 
1076 	decline_attach = oc->cr_uid != tc->cr_uid &&
1077 	    oc->cr_uid != tc->cr_svuid &&
1078 	    oc->cr_uid != tc->cr_ruid;
1079 
1080 	/*
1081 	 * Every one of the target's group ids, must be in the owner's
1082 	 * group list.
1083 	 */
1084 	for (i = 0; !decline_attach && i < tc->cr_ngroups; i++)
1085 		decline_attach = !groupmember(tc->cr_groups[i], oc);
1086 
1087 	/* check the read and saved gids too */
1088 	if (decline_attach == 0)
1089 		decline_attach = !groupmember(tc->cr_rgid, oc) ||
1090 		    !groupmember(tc->cr_svgid, oc);
1091 
1092 	crfree(tc);
1093 	crfree(oc);
1094 
1095 	return !decline_attach;
1096 }
1097 
1098 /*
1099  * Attach a process to a PMC.
1100  */
1101 
1102 static int
1103 pmc_attach_one_process(struct proc *p, struct pmc *pm)
1104 {
1105 	int ri, error;
1106 	char *fullpath, *freepath;
1107 	struct pmc_process	*pp;
1108 
1109 	sx_assert(&pmc_sx, SX_XLOCKED);
1110 
1111 	PMCDBG5(PRC,ATT,2, "attach-one pm=%p ri=%d proc=%p (%d, %s)", pm,
1112 	    PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm);
1113 
1114 	/*
1115 	 * Locate the process descriptor corresponding to process 'p',
1116 	 * allocating space as needed.
1117 	 *
1118 	 * Verify that rowindex 'pm_rowindex' is free in the process
1119 	 * descriptor.
1120 	 *
1121 	 * If not, allocate space for a descriptor and link the
1122 	 * process descriptor and PMC.
1123 	 */
1124 	ri = PMC_TO_ROWINDEX(pm);
1125 
1126 	/* mark process as using HWPMCs */
1127 	PROC_LOCK(p);
1128 	p->p_flag |= P_HWPMC;
1129 	PROC_UNLOCK(p);
1130 
1131 	if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_ALLOCATE)) == NULL) {
1132 		error = ENOMEM;
1133 		goto fail;
1134 	}
1135 
1136 	if (pp->pp_pmcs[ri].pp_pmc == pm) {/* already present at slot [ri] */
1137 		error = EEXIST;
1138 		goto fail;
1139 	}
1140 
1141 	if (pp->pp_pmcs[ri].pp_pmc != NULL) {
1142 		error = EBUSY;
1143 		goto fail;
1144 	}
1145 
1146 	pmc_link_target_process(pm, pp);
1147 
1148 	if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) &&
1149 	    (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) == 0)
1150 		pm->pm_flags |= PMC_F_NEEDS_LOGFILE;
1151 
1152 	pm->pm_flags |= PMC_F_ATTACH_DONE; /* mark as attached */
1153 
1154 	/* issue an attach event to a configured log file */
1155 	if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE) {
1156 		if (p->p_flag & P_KPROC) {
1157 			fullpath = kernelname;
1158 			freepath = NULL;
1159 		} else {
1160 			pmc_getfilename(p->p_textvp, &fullpath, &freepath);
1161 			pmclog_process_pmcattach(pm, p->p_pid, fullpath);
1162 		}
1163 		free(freepath, M_TEMP);
1164 		if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1165 			pmc_log_process_mappings(pm->pm_owner, p);
1166 	}
1167 
1168 	return (0);
1169  fail:
1170 	PROC_LOCK(p);
1171 	p->p_flag &= ~P_HWPMC;
1172 	PROC_UNLOCK(p);
1173 	return (error);
1174 }
1175 
1176 /*
1177  * Attach a process and optionally its children
1178  */
1179 
1180 static int
1181 pmc_attach_process(struct proc *p, struct pmc *pm)
1182 {
1183 	int error;
1184 	struct proc *top;
1185 
1186 	sx_assert(&pmc_sx, SX_XLOCKED);
1187 
1188 	PMCDBG5(PRC,ATT,1, "attach pm=%p ri=%d proc=%p (%d, %s)", pm,
1189 	    PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm);
1190 
1191 
1192 	/*
1193 	 * If this PMC successfully allowed a GETMSR operation
1194 	 * in the past, disallow further ATTACHes.
1195 	 */
1196 
1197 	if ((pm->pm_flags & PMC_PP_ENABLE_MSR_ACCESS) != 0)
1198 		return EPERM;
1199 
1200 	if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0)
1201 		return pmc_attach_one_process(p, pm);
1202 
1203 	/*
1204 	 * Traverse all child processes, attaching them to
1205 	 * this PMC.
1206 	 */
1207 
1208 	sx_slock(&proctree_lock);
1209 
1210 	top = p;
1211 
1212 	for (;;) {
1213 		if ((error = pmc_attach_one_process(p, pm)) != 0)
1214 			break;
1215 		if (!LIST_EMPTY(&p->p_children))
1216 			p = LIST_FIRST(&p->p_children);
1217 		else for (;;) {
1218 			if (p == top)
1219 				goto done;
1220 			if (LIST_NEXT(p, p_sibling)) {
1221 				p = LIST_NEXT(p, p_sibling);
1222 				break;
1223 			}
1224 			p = p->p_pptr;
1225 		}
1226 	}
1227 
1228 	if (error)
1229 		(void) pmc_detach_process(top, pm);
1230 
1231  done:
1232 	sx_sunlock(&proctree_lock);
1233 	return error;
1234 }
1235 
1236 /*
1237  * Detach a process from a PMC.  If there are no other PMCs tracking
1238  * this process, remove the process structure from its hash table.  If
1239  * 'flags' contains PMC_FLAG_REMOVE, then free the process structure.
1240  */
1241 
1242 static int
1243 pmc_detach_one_process(struct proc *p, struct pmc *pm, int flags)
1244 {
1245 	int ri;
1246 	struct pmc_process *pp;
1247 
1248 	sx_assert(&pmc_sx, SX_XLOCKED);
1249 
1250 	KASSERT(pm != NULL,
1251 	    ("[pmc,%d] null pm pointer", __LINE__));
1252 
1253 	ri = PMC_TO_ROWINDEX(pm);
1254 
1255 	PMCDBG6(PRC,ATT,2, "detach-one pm=%p ri=%d proc=%p (%d, %s) flags=0x%x",
1256 	    pm, ri, p, p->p_pid, p->p_comm, flags);
1257 
1258 	if ((pp = pmc_find_process_descriptor(p, 0)) == NULL)
1259 		return ESRCH;
1260 
1261 	if (pp->pp_pmcs[ri].pp_pmc != pm)
1262 		return EINVAL;
1263 
1264 	pmc_unlink_target_process(pm, pp);
1265 
1266 	/* Issue a detach entry if a log file is configured */
1267 	if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE)
1268 		pmclog_process_pmcdetach(pm, p->p_pid);
1269 
1270 	/*
1271 	 * If there are no PMCs targeting this process, we remove its
1272 	 * descriptor from the target hash table and unset the P_HWPMC
1273 	 * flag in the struct proc.
1274 	 */
1275 	KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= (int) md->pmd_npmc,
1276 	    ("[pmc,%d] Illegal refcnt %d for process struct %p",
1277 		__LINE__, pp->pp_refcnt, pp));
1278 
1279 	if (pp->pp_refcnt != 0)	/* still a target of some PMC */
1280 		return 0;
1281 
1282 	pmc_remove_process_descriptor(pp);
1283 
1284 	if (flags & PMC_FLAG_REMOVE)
1285 		pmc_destroy_process_descriptor(pp);
1286 
1287 	PROC_LOCK(p);
1288 	p->p_flag &= ~P_HWPMC;
1289 	PROC_UNLOCK(p);
1290 
1291 	return 0;
1292 }
1293 
1294 /*
1295  * Detach a process and optionally its descendants from a PMC.
1296  */
1297 
1298 static int
1299 pmc_detach_process(struct proc *p, struct pmc *pm)
1300 {
1301 	struct proc *top;
1302 
1303 	sx_assert(&pmc_sx, SX_XLOCKED);
1304 
1305 	PMCDBG5(PRC,ATT,1, "detach pm=%p ri=%d proc=%p (%d, %s)", pm,
1306 	    PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm);
1307 
1308 	if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0)
1309 		return pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE);
1310 
1311 	/*
1312 	 * Traverse all children, detaching them from this PMC.  We
1313 	 * ignore errors since we could be detaching a PMC from a
1314 	 * partially attached proc tree.
1315 	 */
1316 
1317 	sx_slock(&proctree_lock);
1318 
1319 	top = p;
1320 
1321 	for (;;) {
1322 		(void) pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE);
1323 
1324 		if (!LIST_EMPTY(&p->p_children))
1325 			p = LIST_FIRST(&p->p_children);
1326 		else for (;;) {
1327 			if (p == top)
1328 				goto done;
1329 			if (LIST_NEXT(p, p_sibling)) {
1330 				p = LIST_NEXT(p, p_sibling);
1331 				break;
1332 			}
1333 			p = p->p_pptr;
1334 		}
1335 	}
1336 
1337  done:
1338 	sx_sunlock(&proctree_lock);
1339 
1340 	if (LIST_EMPTY(&pm->pm_targets))
1341 		pm->pm_flags &= ~PMC_F_ATTACH_DONE;
1342 
1343 	return 0;
1344 }
1345 
1346 
1347 /*
1348  * Thread context switch IN
1349  */
1350 
1351 static void
1352 pmc_process_csw_in(struct thread *td)
1353 {
1354 	int cpu;
1355 	unsigned int adjri, ri;
1356 	struct pmc *pm;
1357 	struct proc *p;
1358 	struct pmc_cpu *pc;
1359 	struct pmc_hw *phw;
1360 	pmc_value_t newvalue;
1361 	struct pmc_process *pp;
1362 	struct pmc_thread *pt;
1363 	struct pmc_classdep *pcd;
1364 
1365 	p = td->td_proc;
1366 	pt = NULL;
1367 	if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE)) == NULL)
1368 		return;
1369 
1370 	KASSERT(pp->pp_proc == td->td_proc,
1371 	    ("[pmc,%d] not my thread state", __LINE__));
1372 
1373 	critical_enter(); /* no preemption from this point */
1374 
1375 	cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */
1376 
1377 	PMCDBG5(CSW,SWI,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p,
1378 	    p->p_pid, p->p_comm, pp);
1379 
1380 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
1381 	    ("[pmc,%d] weird CPU id %d", __LINE__, cpu));
1382 
1383 	pc = pmc_pcpu[cpu];
1384 
1385 	for (ri = 0; ri < md->pmd_npmc; ri++) {
1386 
1387 		if ((pm = pp->pp_pmcs[ri].pp_pmc) == NULL)
1388 			continue;
1389 
1390 		KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)),
1391 		    ("[pmc,%d] Target PMC in non-virtual mode (%d)",
1392 			__LINE__, PMC_TO_MODE(pm)));
1393 
1394 		KASSERT(PMC_TO_ROWINDEX(pm) == ri,
1395 		    ("[pmc,%d] Row index mismatch pmc %d != ri %d",
1396 			__LINE__, PMC_TO_ROWINDEX(pm), ri));
1397 
1398 		/*
1399 		 * Only PMCs that are marked as 'RUNNING' need
1400 		 * be placed on hardware.
1401 		 */
1402 
1403 		if (pm->pm_state != PMC_STATE_RUNNING)
1404 			continue;
1405 
1406 		KASSERT(counter_u64_fetch(pm->pm_runcount) >= 0,
1407 	    ("[pmc,%d] pm=%p runcount %ld", __LINE__, (void *) pm,
1408 		 (unsigned long)counter_u64_fetch(pm->pm_runcount)));
1409 
1410 		/* increment PMC runcount */
1411 		counter_u64_add(pm->pm_runcount, 1);
1412 
1413 		/* configure the HWPMC we are going to use. */
1414 		pcd = pmc_ri_to_classdep(md, ri, &adjri);
1415 		pcd->pcd_config_pmc(cpu, adjri, pm);
1416 
1417 		phw = pc->pc_hwpmcs[ri];
1418 
1419 		KASSERT(phw != NULL,
1420 		    ("[pmc,%d] null hw pointer", __LINE__));
1421 
1422 		KASSERT(phw->phw_pmc == pm,
1423 		    ("[pmc,%d] hw->pmc %p != pmc %p", __LINE__,
1424 			phw->phw_pmc, pm));
1425 
1426 		/*
1427 		 * Write out saved value and start the PMC.
1428 		 *
1429 		 * Sampling PMCs use a per-thread value, while
1430 		 * counting mode PMCs use a per-pmc value that is
1431 		 * inherited across descendants.
1432 		 */
1433 		if (PMC_TO_MODE(pm) == PMC_MODE_TS) {
1434 			if (pt == NULL)
1435 				pt = pmc_find_thread_descriptor(pp, td,
1436 				    PMC_FLAG_NONE);
1437 
1438 			KASSERT(pt != NULL,
1439 			    ("[pmc,%d] No thread found for td=%p", __LINE__,
1440 			    td));
1441 
1442 			mtx_pool_lock_spin(pmc_mtxpool, pm);
1443 
1444 			/*
1445 			 * If we have a thread descriptor, use the per-thread
1446 			 * counter in the descriptor. If not, we will use
1447 			 * a per-process counter.
1448 			 *
1449 			 * TODO: Remove the per-process "safety net" once
1450 			 * we have thoroughly tested that we don't hit the
1451 			 * above assert.
1452 			 */
1453 			if (pt != NULL) {
1454 				if (pt->pt_pmcs[ri].pt_pmcval > 0)
1455 					newvalue = pt->pt_pmcs[ri].pt_pmcval;
1456 				else
1457 					newvalue = pm->pm_sc.pm_reloadcount;
1458 			} else {
1459 				/*
1460 				 * Use the saved value calculated after the most
1461 				 * recent time a thread using the shared counter
1462 				 * switched out. Reset the saved count in case
1463 				 * another thread from this process switches in
1464 				 * before any threads switch out.
1465 				 */
1466 
1467 				newvalue = pp->pp_pmcs[ri].pp_pmcval;
1468 				pp->pp_pmcs[ri].pp_pmcval =
1469 				    pm->pm_sc.pm_reloadcount;
1470 			}
1471 			mtx_pool_unlock_spin(pmc_mtxpool, pm);
1472 			KASSERT(newvalue > 0 && newvalue <=
1473 			    pm->pm_sc.pm_reloadcount,
1474 			    ("[pmc,%d] pmcval outside of expected range cpu=%d "
1475 			    "ri=%d pmcval=%jx pm_reloadcount=%jx", __LINE__,
1476 			    cpu, ri, newvalue, pm->pm_sc.pm_reloadcount));
1477 		} else {
1478 			KASSERT(PMC_TO_MODE(pm) == PMC_MODE_TC,
1479 			    ("[pmc,%d] illegal mode=%d", __LINE__,
1480 			    PMC_TO_MODE(pm)));
1481 			mtx_pool_lock_spin(pmc_mtxpool, pm);
1482 			newvalue = PMC_PCPU_SAVED(cpu, ri) =
1483 			    pm->pm_gv.pm_savedvalue;
1484 			mtx_pool_unlock_spin(pmc_mtxpool, pm);
1485 		}
1486 
1487 		PMCDBG3(CSW,SWI,1,"cpu=%d ri=%d new=%jd", cpu, ri, newvalue);
1488 
1489 		pcd->pcd_write_pmc(cpu, adjri, newvalue);
1490 
1491 		/* If a sampling mode PMC, reset stalled state. */
1492 		if (PMC_TO_MODE(pm) == PMC_MODE_TS)
1493 			pm->pm_pcpu_state[cpu].pps_stalled = 0;
1494 
1495 		/* Indicate that we desire this to run. */
1496 		pm->pm_pcpu_state[cpu].pps_cpustate = 1;
1497 
1498 		/* Start the PMC. */
1499 		pcd->pcd_start_pmc(cpu, adjri);
1500 	}
1501 
1502 	/*
1503 	 * perform any other architecture/cpu dependent thread
1504 	 * switch-in actions.
1505 	 */
1506 
1507 	(void) (*md->pmd_switch_in)(pc, pp);
1508 
1509 	critical_exit();
1510 
1511 }
1512 
1513 /*
1514  * Thread context switch OUT.
1515  */
1516 
1517 static void
1518 pmc_process_csw_out(struct thread *td)
1519 {
1520 	int cpu;
1521 	int64_t tmp;
1522 	struct pmc *pm;
1523 	struct proc *p;
1524 	enum pmc_mode mode;
1525 	struct pmc_cpu *pc;
1526 	pmc_value_t newvalue;
1527 	unsigned int adjri, ri;
1528 	struct pmc_process *pp;
1529 	struct pmc_thread *pt = NULL;
1530 	struct pmc_classdep *pcd;
1531 
1532 
1533 	/*
1534 	 * Locate our process descriptor; this may be NULL if
1535 	 * this process is exiting and we have already removed
1536 	 * the process from the target process table.
1537 	 *
1538 	 * Note that due to kernel preemption, multiple
1539 	 * context switches may happen while the process is
1540 	 * exiting.
1541 	 *
1542 	 * Note also that if the target process cannot be
1543 	 * found we still need to deconfigure any PMCs that
1544 	 * are currently running on hardware.
1545 	 */
1546 
1547 	p = td->td_proc;
1548 	pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE);
1549 
1550 	/*
1551 	 * save PMCs
1552 	 */
1553 
1554 	critical_enter();
1555 
1556 	cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */
1557 
1558 	PMCDBG5(CSW,SWO,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p,
1559 	    p->p_pid, p->p_comm, pp);
1560 
1561 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
1562 	    ("[pmc,%d weird CPU id %d", __LINE__, cpu));
1563 
1564 	pc = pmc_pcpu[cpu];
1565 
1566 	/*
1567 	 * When a PMC gets unlinked from a target PMC, it will
1568 	 * be removed from the target's pp_pmc[] array.
1569 	 *
1570 	 * However, on a MP system, the target could have been
1571 	 * executing on another CPU at the time of the unlink.
1572 	 * So, at context switch OUT time, we need to look at
1573 	 * the hardware to determine if a PMC is scheduled on
1574 	 * it.
1575 	 */
1576 
1577 	for (ri = 0; ri < md->pmd_npmc; ri++) {
1578 
1579 		pcd = pmc_ri_to_classdep(md, ri, &adjri);
1580 		pm  = NULL;
1581 		(void) (*pcd->pcd_get_config)(cpu, adjri, &pm);
1582 
1583 		if (pm == NULL)	/* nothing at this row index */
1584 			continue;
1585 
1586 		mode = PMC_TO_MODE(pm);
1587 		if (!PMC_IS_VIRTUAL_MODE(mode))
1588 			continue; /* not a process virtual PMC */
1589 
1590 		KASSERT(PMC_TO_ROWINDEX(pm) == ri,
1591 		    ("[pmc,%d] ri mismatch pmc(%d) ri(%d)",
1592 			__LINE__, PMC_TO_ROWINDEX(pm), ri));
1593 
1594 		/*
1595 		 * Change desired state, and then stop if not stalled.
1596 		 * This two-step dance should avoid race conditions where
1597 		 * an interrupt re-enables the PMC after this code has
1598 		 * already checked the pm_stalled flag.
1599 		 */
1600 		pm->pm_pcpu_state[cpu].pps_cpustate = 0;
1601 		if (pm->pm_pcpu_state[cpu].pps_stalled == 0)
1602 			pcd->pcd_stop_pmc(cpu, adjri);
1603 
1604 		KASSERT(counter_u64_fetch(pm->pm_runcount) > 0,
1605 			("[pmc,%d] pm=%p runcount %ld", __LINE__, (void *) pm,
1606 			 (unsigned long)counter_u64_fetch(pm->pm_runcount)));
1607 
1608 		/* reduce this PMC's runcount */
1609 		counter_u64_add(pm->pm_runcount, -1);
1610 
1611 		/*
1612 		 * If this PMC is associated with this process,
1613 		 * save the reading.
1614 		 */
1615 
1616 		if (pm->pm_state != PMC_STATE_DELETED && pp != NULL &&
1617 		    pp->pp_pmcs[ri].pp_pmc != NULL) {
1618 			KASSERT(pm == pp->pp_pmcs[ri].pp_pmc,
1619 			    ("[pmc,%d] pm %p != pp_pmcs[%d] %p", __LINE__,
1620 				pm, ri, pp->pp_pmcs[ri].pp_pmc));
1621 
1622 			KASSERT(pp->pp_refcnt > 0,
1623 			    ("[pmc,%d] pp refcnt = %d", __LINE__,
1624 				pp->pp_refcnt));
1625 
1626 			pcd->pcd_read_pmc(cpu, adjri, &newvalue);
1627 
1628 			if (mode == PMC_MODE_TS) {
1629 				PMCDBG3(CSW,SWO,1,"cpu=%d ri=%d val=%jd (samp)",
1630 				    cpu, ri, newvalue);
1631 
1632 				if (pt == NULL)
1633 					pt = pmc_find_thread_descriptor(pp, td,
1634 					    PMC_FLAG_NONE);
1635 
1636 				KASSERT(pt != NULL,
1637 				    ("[pmc,%d] No thread found for td=%p",
1638 				    __LINE__, td));
1639 
1640 				mtx_pool_lock_spin(pmc_mtxpool, pm);
1641 
1642 				/*
1643 				 * If we have a thread descriptor, save the
1644 				 * per-thread counter in the descriptor. If not,
1645 				 * we will update the per-process counter.
1646 				 *
1647 				 * TODO: Remove the per-process "safety net"
1648 				 * once we have thoroughly tested that we
1649 				 * don't hit the above assert.
1650 				 */
1651 				if (pt != NULL)
1652 					pt->pt_pmcs[ri].pt_pmcval = newvalue;
1653 				else {
1654 					/*
1655 					 * For sampling process-virtual PMCs,
1656 					 * newvalue is the number of events to
1657 					 * be seen until the next sampling
1658 					 * interrupt. We can just add the events
1659 					 * left from this invocation to the
1660 					 * counter, then adjust in case we
1661 					 * overflow our range.
1662 					 *
1663 					 * (Recall that we reload the counter
1664 					 * every time we use it.)
1665 					 */
1666 					pp->pp_pmcs[ri].pp_pmcval += newvalue;
1667 					if (pp->pp_pmcs[ri].pp_pmcval >
1668 					    pm->pm_sc.pm_reloadcount)
1669 						pp->pp_pmcs[ri].pp_pmcval -=
1670 						    pm->pm_sc.pm_reloadcount;
1671 				}
1672 				mtx_pool_unlock_spin(pmc_mtxpool, pm);
1673 			} else {
1674 				tmp = newvalue - PMC_PCPU_SAVED(cpu,ri);
1675 
1676 				PMCDBG3(CSW,SWO,1,"cpu=%d ri=%d tmp=%jd (count)",
1677 				    cpu, ri, tmp);
1678 
1679 				/*
1680 				 * For counting process-virtual PMCs,
1681 				 * we expect the count to be
1682 				 * increasing monotonically, modulo a 64
1683 				 * bit wraparound.
1684 				 */
1685 				KASSERT(tmp >= 0,
1686 				    ("[pmc,%d] negative increment cpu=%d "
1687 				     "ri=%d newvalue=%jx saved=%jx "
1688 				     "incr=%jx", __LINE__, cpu, ri,
1689 				     newvalue, PMC_PCPU_SAVED(cpu,ri), tmp));
1690 
1691 				mtx_pool_lock_spin(pmc_mtxpool, pm);
1692 				pm->pm_gv.pm_savedvalue += tmp;
1693 				pp->pp_pmcs[ri].pp_pmcval += tmp;
1694 				mtx_pool_unlock_spin(pmc_mtxpool, pm);
1695 
1696 				if (pm->pm_flags & PMC_F_LOG_PROCCSW)
1697 					pmclog_process_proccsw(pm, pp, tmp, td);
1698 			}
1699 		}
1700 
1701 		/* mark hardware as free */
1702 		pcd->pcd_config_pmc(cpu, adjri, NULL);
1703 	}
1704 
1705 	/*
1706 	 * perform any other architecture/cpu dependent thread
1707 	 * switch out functions.
1708 	 */
1709 
1710 	(void) (*md->pmd_switch_out)(pc, pp);
1711 
1712 	critical_exit();
1713 }
1714 
1715 /*
1716  * A new thread for a process.
1717  */
1718 static void
1719 pmc_process_thread_add(struct thread *td)
1720 {
1721 	struct pmc_process *pmc;
1722 
1723 	pmc = pmc_find_process_descriptor(td->td_proc, PMC_FLAG_NONE);
1724 	if (pmc != NULL)
1725 		pmc_find_thread_descriptor(pmc, td, PMC_FLAG_ALLOCATE);
1726 }
1727 
1728 /*
1729  * A thread delete for a process.
1730  */
1731 static void
1732 pmc_process_thread_delete(struct thread *td)
1733 {
1734 	struct pmc_process *pmc;
1735 
1736 	pmc = pmc_find_process_descriptor(td->td_proc, PMC_FLAG_NONE);
1737 	if (pmc != NULL)
1738 		pmc_thread_descriptor_pool_free(pmc_find_thread_descriptor(pmc,
1739 		    td, PMC_FLAG_REMOVE));
1740 }
1741 
1742 /*
1743  * A userret() call for a thread.
1744  */
1745 static void
1746 pmc_process_thread_userret(struct thread *td)
1747 {
1748 	sched_pin();
1749 	pmc_capture_user_callchain(curcpu, PMC_UR, td->td_frame);
1750 	sched_unpin();
1751 }
1752 
1753 /*
1754  * A mapping change for a process.
1755  */
1756 
1757 static void
1758 pmc_process_mmap(struct thread *td, struct pmckern_map_in *pkm)
1759 {
1760 	int ri;
1761 	pid_t pid;
1762 	char *fullpath, *freepath;
1763 	const struct pmc *pm;
1764 	struct pmc_owner *po;
1765 	const struct pmc_process *pp;
1766 
1767 	freepath = fullpath = NULL;
1768 	MPASS(!in_epoch(global_epoch_preempt));
1769 	pmc_getfilename((struct vnode *) pkm->pm_file, &fullpath, &freepath);
1770 
1771 	pid = td->td_proc->p_pid;
1772 
1773 	PMC_EPOCH_ENTER();
1774 	/* Inform owners of all system-wide sampling PMCs. */
1775 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1776 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1777 			pmclog_process_map_in(po, pid, pkm->pm_address, fullpath);
1778 
1779 	if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL)
1780 		goto done;
1781 
1782 	/*
1783 	 * Inform sampling PMC owners tracking this process.
1784 	 */
1785 	for (ri = 0; ri < md->pmd_npmc; ri++)
1786 		if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL &&
1787 		    PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1788 			pmclog_process_map_in(pm->pm_owner,
1789 			    pid, pkm->pm_address, fullpath);
1790 
1791   done:
1792 	if (freepath)
1793 		free(freepath, M_TEMP);
1794 	PMC_EPOCH_EXIT();
1795 }
1796 
1797 
1798 /*
1799  * Log an munmap request.
1800  */
1801 
1802 static void
1803 pmc_process_munmap(struct thread *td, struct pmckern_map_out *pkm)
1804 {
1805 	int ri;
1806 	pid_t pid;
1807 	struct pmc_owner *po;
1808 	const struct pmc *pm;
1809 	const struct pmc_process *pp;
1810 
1811 	pid = td->td_proc->p_pid;
1812 
1813 	PMC_EPOCH_ENTER();
1814 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1815 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1816 		pmclog_process_map_out(po, pid, pkm->pm_address,
1817 		    pkm->pm_address + pkm->pm_size);
1818 	PMC_EPOCH_EXIT();
1819 
1820 	if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL)
1821 		return;
1822 
1823 	for (ri = 0; ri < md->pmd_npmc; ri++)
1824 		if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL &&
1825 		    PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1826 			pmclog_process_map_out(pm->pm_owner, pid,
1827 			    pkm->pm_address, pkm->pm_address + pkm->pm_size);
1828 }
1829 
1830 /*
1831  * Log mapping information about the kernel.
1832  */
1833 
1834 static void
1835 pmc_log_kernel_mappings(struct pmc *pm)
1836 {
1837 	struct pmc_owner *po;
1838 	struct pmckern_map_in *km, *kmbase;
1839 
1840 	MPASS(in_epoch(global_epoch_preempt) || sx_xlocked(&pmc_sx));
1841 	KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)),
1842 	    ("[pmc,%d] non-sampling PMC (%p) desires mapping information",
1843 		__LINE__, (void *) pm));
1844 
1845 	po = pm->pm_owner;
1846 
1847 	if (po->po_flags & PMC_PO_INITIAL_MAPPINGS_DONE)
1848 		return;
1849 	if (PMC_TO_MODE(pm) == PMC_MODE_SS)
1850 		pmc_process_allproc(pm);
1851 	/*
1852 	 * Log the current set of kernel modules.
1853 	 */
1854 	kmbase = linker_hwpmc_list_objects();
1855 	for (km = kmbase; km->pm_file != NULL; km++) {
1856 		PMCDBG2(LOG,REG,1,"%s %p", (char *) km->pm_file,
1857 		    (void *) km->pm_address);
1858 		pmclog_process_map_in(po, (pid_t) -1, km->pm_address,
1859 		    km->pm_file);
1860 	}
1861 	free(kmbase, M_LINKER);
1862 
1863 	po->po_flags |= PMC_PO_INITIAL_MAPPINGS_DONE;
1864 }
1865 
1866 /*
1867  * Log the mappings for a single process.
1868  */
1869 
1870 static void
1871 pmc_log_process_mappings(struct pmc_owner *po, struct proc *p)
1872 {
1873 	vm_map_t map;
1874 	struct vnode *vp;
1875 	struct vmspace *vm;
1876 	vm_map_entry_t entry;
1877 	vm_offset_t last_end;
1878 	u_int last_timestamp;
1879 	struct vnode *last_vp;
1880 	vm_offset_t start_addr;
1881 	vm_object_t obj, lobj, tobj;
1882 	char *fullpath, *freepath;
1883 
1884 	last_vp = NULL;
1885 	last_end = (vm_offset_t) 0;
1886 	fullpath = freepath = NULL;
1887 
1888 	if ((vm = vmspace_acquire_ref(p)) == NULL)
1889 		return;
1890 
1891 	map = &vm->vm_map;
1892 	vm_map_lock_read(map);
1893 
1894 	for (entry = map->header.next; entry != &map->header; entry = entry->next) {
1895 
1896 		if (entry == NULL) {
1897 			PMCDBG2(LOG,OPS,2, "hwpmc: vm_map entry unexpectedly "
1898 			    "NULL! pid=%d vm_map=%p\n", p->p_pid, map);
1899 			break;
1900 		}
1901 
1902 		/*
1903 		 * We only care about executable map entries.
1904 		 */
1905 		if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) ||
1906 		    !(entry->protection & VM_PROT_EXECUTE) ||
1907 		    (entry->object.vm_object == NULL)) {
1908 			continue;
1909 		}
1910 
1911 		obj = entry->object.vm_object;
1912 		VM_OBJECT_RLOCK(obj);
1913 
1914 		/*
1915 		 * Walk the backing_object list to find the base
1916 		 * (non-shadowed) vm_object.
1917 		 */
1918 		for (lobj = tobj = obj; tobj != NULL; tobj = tobj->backing_object) {
1919 			if (tobj != obj)
1920 				VM_OBJECT_RLOCK(tobj);
1921 			if (lobj != obj)
1922 				VM_OBJECT_RUNLOCK(lobj);
1923 			lobj = tobj;
1924 		}
1925 
1926 		/*
1927 		 * At this point lobj is the base vm_object and it is locked.
1928 		 */
1929 		if (lobj == NULL) {
1930 			PMCDBG3(LOG,OPS,2, "hwpmc: lobj unexpectedly NULL! pid=%d "
1931 			    "vm_map=%p vm_obj=%p\n", p->p_pid, map, obj);
1932 			VM_OBJECT_RUNLOCK(obj);
1933 			continue;
1934 		}
1935 
1936 		vp = vm_object_vnode(lobj);
1937 		if (vp == NULL) {
1938 			if (lobj != obj)
1939 				VM_OBJECT_RUNLOCK(lobj);
1940 			VM_OBJECT_RUNLOCK(obj);
1941 			continue;
1942 		}
1943 
1944 		/*
1945 		 * Skip contiguous regions that point to the same
1946 		 * vnode, so we don't emit redundant MAP-IN
1947 		 * directives.
1948 		 */
1949 		if (entry->start == last_end && vp == last_vp) {
1950 			last_end = entry->end;
1951 			if (lobj != obj)
1952 				VM_OBJECT_RUNLOCK(lobj);
1953 			VM_OBJECT_RUNLOCK(obj);
1954 			continue;
1955 		}
1956 
1957 		/*
1958 		 * We don't want to keep the proc's vm_map or this
1959 		 * vm_object locked while we walk the pathname, since
1960 		 * vn_fullpath() can sleep.  However, if we drop the
1961 		 * lock, it's possible for concurrent activity to
1962 		 * modify the vm_map list.  To protect against this,
1963 		 * we save the vm_map timestamp before we release the
1964 		 * lock, and check it after we reacquire the lock
1965 		 * below.
1966 		 */
1967 		start_addr = entry->start;
1968 		last_end = entry->end;
1969 		last_timestamp = map->timestamp;
1970 		vm_map_unlock_read(map);
1971 
1972 		vref(vp);
1973 		if (lobj != obj)
1974 			VM_OBJECT_RUNLOCK(lobj);
1975 
1976 		VM_OBJECT_RUNLOCK(obj);
1977 
1978 		freepath = NULL;
1979 		pmc_getfilename(vp, &fullpath, &freepath);
1980 		last_vp = vp;
1981 
1982 		vrele(vp);
1983 
1984 		vp = NULL;
1985 		pmclog_process_map_in(po, p->p_pid, start_addr, fullpath);
1986 		if (freepath)
1987 			free(freepath, M_TEMP);
1988 
1989 		vm_map_lock_read(map);
1990 
1991 		/*
1992 		 * If our saved timestamp doesn't match, this means
1993 		 * that the vm_map was modified out from under us and
1994 		 * we can't trust our current "entry" pointer.  Do a
1995 		 * new lookup for this entry.  If there is no entry
1996 		 * for this address range, vm_map_lookup_entry() will
1997 		 * return the previous one, so we always want to go to
1998 		 * entry->next on the next loop iteration.
1999 		 *
2000 		 * There is an edge condition here that can occur if
2001 		 * there is no entry at or before this address.  In
2002 		 * this situation, vm_map_lookup_entry returns
2003 		 * &map->header, which would cause our loop to abort
2004 		 * without processing the rest of the map.  However,
2005 		 * in practice this will never happen for process
2006 		 * vm_map.  This is because the executable's text
2007 		 * segment is the first mapping in the proc's address
2008 		 * space, and this mapping is never removed until the
2009 		 * process exits, so there will always be a non-header
2010 		 * entry at or before the requested address for
2011 		 * vm_map_lookup_entry to return.
2012 		 */
2013 		if (map->timestamp != last_timestamp)
2014 			vm_map_lookup_entry(map, last_end - 1, &entry);
2015 	}
2016 
2017 	vm_map_unlock_read(map);
2018 	vmspace_free(vm);
2019 	return;
2020 }
2021 
2022 /*
2023  * Log mappings for all processes in the system.
2024  */
2025 
2026 static void
2027 pmc_log_all_process_mappings(struct pmc_owner *po)
2028 {
2029 	struct proc *p, *top;
2030 
2031 	sx_assert(&pmc_sx, SX_XLOCKED);
2032 
2033 	if ((p = pfind(1)) == NULL)
2034 		panic("[pmc,%d] Cannot find init", __LINE__);
2035 
2036 	PROC_UNLOCK(p);
2037 
2038 	sx_slock(&proctree_lock);
2039 
2040 	top = p;
2041 
2042 	for (;;) {
2043 		pmc_log_process_mappings(po, p);
2044 		if (!LIST_EMPTY(&p->p_children))
2045 			p = LIST_FIRST(&p->p_children);
2046 		else for (;;) {
2047 			if (p == top)
2048 				goto done;
2049 			if (LIST_NEXT(p, p_sibling)) {
2050 				p = LIST_NEXT(p, p_sibling);
2051 				break;
2052 			}
2053 			p = p->p_pptr;
2054 		}
2055 	}
2056  done:
2057 	sx_sunlock(&proctree_lock);
2058 }
2059 
2060 /*
2061  * The 'hook' invoked from the kernel proper
2062  */
2063 
2064 
2065 #ifdef	HWPMC_DEBUG
2066 const char *pmc_hooknames[] = {
2067 	/* these strings correspond to PMC_FN_* in <sys/pmckern.h> */
2068 	"",
2069 	"EXEC",
2070 	"CSW-IN",
2071 	"CSW-OUT",
2072 	"SAMPLE",
2073 	"UNUSED1",
2074 	"UNUSED2",
2075 	"MMAP",
2076 	"MUNMAP",
2077 	"CALLCHAIN-NMI",
2078 	"CALLCHAIN-SOFT",
2079 	"SOFTSAMPLING",
2080 	"THR-CREATE",
2081 	"THR-EXIT",
2082 	"THR-USERRET",
2083 	"THR-CREATE-LOG",
2084 	"THR-EXIT-LOG",
2085 	"PROC-CREATE-LOG"
2086 };
2087 #endif
2088 
2089 static int
2090 pmc_hook_handler(struct thread *td, int function, void *arg)
2091 {
2092 	int cpu;
2093 
2094 	PMCDBG4(MOD,PMH,1, "hook td=%p func=%d \"%s\" arg=%p", td, function,
2095 	    pmc_hooknames[function], arg);
2096 
2097 	switch (function)
2098 	{
2099 
2100 	/*
2101 	 * Process exec()
2102 	 */
2103 
2104 	case PMC_FN_PROCESS_EXEC:
2105 	{
2106 		char *fullpath, *freepath;
2107 		unsigned int ri;
2108 		int is_using_hwpmcs;
2109 		struct pmc *pm;
2110 		struct proc *p;
2111 		struct pmc_owner *po;
2112 		struct pmc_process *pp;
2113 		struct pmckern_procexec *pk;
2114 
2115 		sx_assert(&pmc_sx, SX_XLOCKED);
2116 
2117 		p = td->td_proc;
2118 		pmc_getfilename(p->p_textvp, &fullpath, &freepath);
2119 
2120 		pk = (struct pmckern_procexec *) arg;
2121 
2122 		PMC_EPOCH_ENTER();
2123 		/* Inform owners of SS mode PMCs of the exec event. */
2124 		CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
2125 		    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
2126 			    pmclog_process_procexec(po, PMC_ID_INVALID,
2127 				p->p_pid, pk->pm_entryaddr, fullpath);
2128 		PMC_EPOCH_EXIT();
2129 
2130 		PROC_LOCK(p);
2131 		is_using_hwpmcs = p->p_flag & P_HWPMC;
2132 		PROC_UNLOCK(p);
2133 
2134 		if (!is_using_hwpmcs) {
2135 			if (freepath)
2136 				free(freepath, M_TEMP);
2137 			break;
2138 		}
2139 
2140 		/*
2141 		 * PMCs are not inherited across an exec():  remove any
2142 		 * PMCs that this process is the owner of.
2143 		 */
2144 
2145 		if ((po = pmc_find_owner_descriptor(p)) != NULL) {
2146 			pmc_remove_owner(po);
2147 			pmc_destroy_owner_descriptor(po);
2148 		}
2149 
2150 		/*
2151 		 * If the process being exec'ed is not the target of any
2152 		 * PMC, we are done.
2153 		 */
2154 		if ((pp = pmc_find_process_descriptor(p, 0)) == NULL) {
2155 			if (freepath)
2156 				free(freepath, M_TEMP);
2157 			break;
2158 		}
2159 
2160 		/*
2161 		 * Log the exec event to all monitoring owners.  Skip
2162 		 * owners who have already received the event because
2163 		 * they had system sampling PMCs active.
2164 		 */
2165 		for (ri = 0; ri < md->pmd_npmc; ri++)
2166 			if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) {
2167 				po = pm->pm_owner;
2168 				if (po->po_sscount == 0 &&
2169 				    po->po_flags & PMC_PO_OWNS_LOGFILE)
2170 					pmclog_process_procexec(po, pm->pm_id,
2171 					    p->p_pid, pk->pm_entryaddr,
2172 					    fullpath);
2173 			}
2174 
2175 		if (freepath)
2176 			free(freepath, M_TEMP);
2177 
2178 
2179 		PMCDBG4(PRC,EXC,1, "exec proc=%p (%d, %s) cred-changed=%d",
2180 		    p, p->p_pid, p->p_comm, pk->pm_credentialschanged);
2181 
2182 		if (pk->pm_credentialschanged == 0) /* no change */
2183 			break;
2184 
2185 		/*
2186 		 * If the newly exec()'ed process has a different credential
2187 		 * than before, allow it to be the target of a PMC only if
2188 		 * the PMC's owner has sufficient privilege.
2189 		 */
2190 
2191 		for (ri = 0; ri < md->pmd_npmc; ri++)
2192 			if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL)
2193 				if (pmc_can_attach(pm, td->td_proc) != 0)
2194 					pmc_detach_one_process(td->td_proc,
2195 					    pm, PMC_FLAG_NONE);
2196 
2197 		KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= (int) md->pmd_npmc,
2198 		    ("[pmc,%d] Illegal ref count %d on pp %p", __LINE__,
2199 			pp->pp_refcnt, pp));
2200 
2201 		/*
2202 		 * If this process is no longer the target of any
2203 		 * PMCs, we can remove the process entry and free
2204 		 * up space.
2205 		 */
2206 
2207 		if (pp->pp_refcnt == 0) {
2208 			pmc_remove_process_descriptor(pp);
2209 			pmc_destroy_process_descriptor(pp);
2210 			break;
2211 		}
2212 
2213 	}
2214 	break;
2215 
2216 	case PMC_FN_CSW_IN:
2217 		pmc_process_csw_in(td);
2218 		break;
2219 
2220 	case PMC_FN_CSW_OUT:
2221 		pmc_process_csw_out(td);
2222 		break;
2223 
2224 	/*
2225 	 * Process accumulated PC samples.
2226 	 *
2227 	 * This function is expected to be called by hardclock() for
2228 	 * each CPU that has accumulated PC samples.
2229 	 *
2230 	 * This function is to be executed on the CPU whose samples
2231 	 * are being processed.
2232 	 */
2233 	case PMC_FN_DO_SAMPLES:
2234 
2235 		/*
2236 		 * Clear the cpu specific bit in the CPU mask before
2237 		 * do the rest of the processing.  If the NMI handler
2238 		 * gets invoked after the "atomic_clear_int()" call
2239 		 * below but before "pmc_process_samples()" gets
2240 		 * around to processing the interrupt, then we will
2241 		 * come back here at the next hardclock() tick (and
2242 		 * may find nothing to do if "pmc_process_samples()"
2243 		 * had already processed the interrupt).  We don't
2244 		 * lose the interrupt sample.
2245 		 */
2246 		DPCPU_SET(pmc_sampled, 0);
2247 		cpu = PCPU_GET(cpuid);
2248 		pmc_process_samples(cpu, PMC_HR);
2249 		pmc_process_samples(cpu, PMC_SR);
2250 		pmc_process_samples(cpu, PMC_UR);
2251 		break;
2252 
2253 	case PMC_FN_MMAP:
2254 		pmc_process_mmap(td, (struct pmckern_map_in *) arg);
2255 		break;
2256 
2257 	case PMC_FN_MUNMAP:
2258 		MPASS(in_epoch(global_epoch_preempt) || sx_xlocked(&pmc_sx));
2259 		pmc_process_munmap(td, (struct pmckern_map_out *) arg);
2260 		break;
2261 
2262 	case PMC_FN_PROC_CREATE_LOG:
2263 		pmc_process_proccreate((struct proc *)arg);
2264 		break;
2265 
2266 	case PMC_FN_USER_CALLCHAIN:
2267 		/*
2268 		 * Record a call chain.
2269 		 */
2270 		KASSERT(td == curthread, ("[pmc,%d] td != curthread",
2271 		    __LINE__));
2272 
2273 		pmc_capture_user_callchain(PCPU_GET(cpuid), PMC_HR,
2274 		    (struct trapframe *) arg);
2275 
2276 		KASSERT(td->td_pinned == 1,
2277 			("[pmc,%d] invalid td_pinned value", __LINE__));
2278 		sched_unpin();  /* Can migrate safely now. */
2279 
2280 		td->td_pflags &= ~TDP_CALLCHAIN;
2281 		break;
2282 
2283 	case PMC_FN_USER_CALLCHAIN_SOFT:
2284 		/*
2285 		 * Record a call chain.
2286 		 */
2287 		KASSERT(td == curthread, ("[pmc,%d] td != curthread",
2288 		    __LINE__));
2289 
2290 		cpu = PCPU_GET(cpuid);
2291 		pmc_capture_user_callchain(cpu, PMC_SR,
2292 		    (struct trapframe *) arg);
2293 
2294 		KASSERT(td->td_pinned == 1,
2295 		    ("[pmc,%d] invalid td_pinned value", __LINE__));
2296 
2297 		sched_unpin();  /* Can migrate safely now. */
2298 
2299 		td->td_pflags &= ~TDP_CALLCHAIN;
2300 		break;
2301 
2302 	case PMC_FN_SOFT_SAMPLING:
2303 		/*
2304 		 * Call soft PMC sampling intr.
2305 		 */
2306 		pmc_soft_intr((struct pmckern_soft *) arg);
2307 		break;
2308 
2309 	case PMC_FN_THR_CREATE:
2310 		pmc_process_thread_add(td);
2311 		pmc_process_threadcreate(td);
2312 		break;
2313 
2314 	case PMC_FN_THR_CREATE_LOG:
2315 		pmc_process_threadcreate(td);
2316 		break;
2317 
2318 	case PMC_FN_THR_EXIT:
2319 		KASSERT(td == curthread, ("[pmc,%d] td != curthread",
2320 		    __LINE__));
2321 		pmc_process_thread_delete(td);
2322 		pmc_process_threadexit(td);
2323 		break;
2324 	case PMC_FN_THR_EXIT_LOG:
2325 		pmc_process_threadexit(td);
2326 		break;
2327 	case PMC_FN_THR_USERRET:
2328 		KASSERT(td == curthread, ("[pmc,%d] td != curthread",
2329 		    __LINE__));
2330 		pmc_process_thread_userret(td);
2331 		break;
2332 
2333 	default:
2334 #ifdef	HWPMC_DEBUG
2335 		KASSERT(0, ("[pmc,%d] unknown hook %d\n", __LINE__, function));
2336 #endif
2337 		break;
2338 
2339 	}
2340 
2341 	return 0;
2342 }
2343 
2344 /*
2345  * allocate a 'struct pmc_owner' descriptor in the owner hash table.
2346  */
2347 
2348 static struct pmc_owner *
2349 pmc_allocate_owner_descriptor(struct proc *p)
2350 {
2351 	uint32_t hindex;
2352 	struct pmc_owner *po;
2353 	struct pmc_ownerhash *poh;
2354 
2355 	hindex = PMC_HASH_PTR(p, pmc_ownerhashmask);
2356 	poh = &pmc_ownerhash[hindex];
2357 
2358 	/* allocate space for N pointers and one descriptor struct */
2359 	po = malloc(sizeof(struct pmc_owner), M_PMC, M_WAITOK|M_ZERO);
2360 	po->po_owner = p;
2361 	LIST_INSERT_HEAD(poh, po, po_next); /* insert into hash table */
2362 
2363 	TAILQ_INIT(&po->po_logbuffers);
2364 	mtx_init(&po->po_mtx, "pmc-owner-mtx", "pmc-per-proc", MTX_SPIN);
2365 
2366 	PMCDBG4(OWN,ALL,1, "allocate-owner proc=%p (%d, %s) pmc-owner=%p",
2367 	    p, p->p_pid, p->p_comm, po);
2368 
2369 	return po;
2370 }
2371 
2372 static void
2373 pmc_destroy_owner_descriptor(struct pmc_owner *po)
2374 {
2375 
2376 	PMCDBG4(OWN,REL,1, "destroy-owner po=%p proc=%p (%d, %s)",
2377 	    po, po->po_owner, po->po_owner->p_pid, po->po_owner->p_comm);
2378 
2379 	mtx_destroy(&po->po_mtx);
2380 	free(po, M_PMC);
2381 }
2382 
2383 /*
2384  * Allocate a thread descriptor from the free pool.
2385  *
2386  * NOTE: This *can* return NULL.
2387  */
2388 static struct pmc_thread *
2389 pmc_thread_descriptor_pool_alloc(void)
2390 {
2391 	struct pmc_thread *pt;
2392 
2393 	mtx_lock_spin(&pmc_threadfreelist_mtx);
2394 	if ((pt = LIST_FIRST(&pmc_threadfreelist)) != NULL) {
2395 		LIST_REMOVE(pt, pt_next);
2396 		pmc_threadfreelist_entries--;
2397 	}
2398 	mtx_unlock_spin(&pmc_threadfreelist_mtx);
2399 
2400 	return (pt);
2401 }
2402 
2403 /*
2404  * Add a thread descriptor to the free pool. We use this instead of free()
2405  * to maintain a cache of free entries. Additionally, we can safely call
2406  * this function when we cannot call free(), such as in a critical section.
2407  *
2408  */
2409 static void
2410 pmc_thread_descriptor_pool_free(struct pmc_thread *pt)
2411 {
2412 
2413 	if (pt == NULL)
2414 		return;
2415 
2416 	memset(pt, 0, THREADENTRY_SIZE);
2417 	mtx_lock_spin(&pmc_threadfreelist_mtx);
2418 	LIST_INSERT_HEAD(&pmc_threadfreelist, pt, pt_next);
2419 	pmc_threadfreelist_entries++;
2420 	if (pmc_threadfreelist_entries > pmc_threadfreelist_max)
2421 		GROUPTASK_ENQUEUE(&free_gtask);
2422 	mtx_unlock_spin(&pmc_threadfreelist_mtx);
2423 }
2424 
2425 /*
2426  * A callout to manage the free list.
2427  */
2428 static void
2429 pmc_thread_descriptor_pool_free_task(void *arg __unused)
2430 {
2431 	struct pmc_thread *pt;
2432 	LIST_HEAD(, pmc_thread) tmplist;
2433 	int delta;
2434 
2435 	LIST_INIT(&tmplist);
2436 	/* Determine what changes, if any, we need to make. */
2437 	mtx_lock_spin(&pmc_threadfreelist_mtx);
2438 	delta = pmc_threadfreelist_entries - pmc_threadfreelist_max;
2439 	while (delta > 0 &&
2440 		   (pt = LIST_FIRST(&pmc_threadfreelist)) != NULL) {
2441 		delta--;
2442 		LIST_REMOVE(pt, pt_next);
2443 		LIST_INSERT_HEAD(&tmplist, pt, pt_next);
2444 	}
2445 	mtx_unlock_spin(&pmc_threadfreelist_mtx);
2446 
2447 	/* If there are entries to free, free them. */
2448 	while (!LIST_EMPTY(&tmplist)) {
2449 		pt = LIST_FIRST(&tmplist);
2450 		LIST_REMOVE(pt, pt_next);
2451 		free(pt, M_PMC);
2452 	}
2453 }
2454 
2455 /*
2456  * Drain the thread free pool, freeing all allocations.
2457  */
2458 static void
2459 pmc_thread_descriptor_pool_drain()
2460 {
2461 	struct pmc_thread *pt, *next;
2462 
2463 	LIST_FOREACH_SAFE(pt, &pmc_threadfreelist, pt_next, next) {
2464 		LIST_REMOVE(pt, pt_next);
2465 		free(pt, M_PMC);
2466 	}
2467 }
2468 
2469 /*
2470  * find the descriptor corresponding to thread 'td', adding or removing it
2471  * as specified by 'mode'.
2472  *
2473  * Note that this supports additional mode flags in addition to those
2474  * supported by pmc_find_process_descriptor():
2475  * PMC_FLAG_NOWAIT: Causes the function to not wait for mallocs.
2476  *     This makes it safe to call while holding certain other locks.
2477  */
2478 
2479 static struct pmc_thread *
2480 pmc_find_thread_descriptor(struct pmc_process *pp, struct thread *td,
2481     uint32_t mode)
2482 {
2483 	struct pmc_thread *pt = NULL, *ptnew = NULL;
2484 	int wait_flag;
2485 
2486 	KASSERT(td != NULL, ("[pmc,%d] called to add NULL td", __LINE__));
2487 
2488 	/*
2489 	 * Pre-allocate memory in the PMC_FLAG_ALLOCATE case prior to
2490 	 * acquiring the lock.
2491 	 */
2492 	if (mode & PMC_FLAG_ALLOCATE) {
2493 		if ((ptnew = pmc_thread_descriptor_pool_alloc()) == NULL) {
2494 			wait_flag = M_WAITOK;
2495 			if ((mode & PMC_FLAG_NOWAIT) || in_epoch(global_epoch_preempt))
2496 				wait_flag = M_NOWAIT;
2497 
2498 			ptnew = malloc(THREADENTRY_SIZE, M_PMC,
2499 			    wait_flag|M_ZERO);
2500 		}
2501 	}
2502 
2503 	mtx_lock_spin(pp->pp_tdslock);
2504 
2505 	LIST_FOREACH(pt, &pp->pp_tds, pt_next)
2506 		if (pt->pt_td == td)
2507 			break;
2508 
2509 	if ((mode & PMC_FLAG_REMOVE) && pt != NULL)
2510 		LIST_REMOVE(pt, pt_next);
2511 
2512 	if ((mode & PMC_FLAG_ALLOCATE) && pt == NULL && ptnew != NULL) {
2513 		pt = ptnew;
2514 		ptnew = NULL;
2515 		pt->pt_td = td;
2516 		LIST_INSERT_HEAD(&pp->pp_tds, pt, pt_next);
2517 	}
2518 
2519 	mtx_unlock_spin(pp->pp_tdslock);
2520 
2521 	if (ptnew != NULL) {
2522 		free(ptnew, M_PMC);
2523 	}
2524 
2525 	return pt;
2526 }
2527 
2528 /*
2529  * Try to add thread descriptors for each thread in a process.
2530  */
2531 
2532 static void
2533 pmc_add_thread_descriptors_from_proc(struct proc *p, struct pmc_process *pp)
2534 {
2535 	struct thread *curtd;
2536 	struct pmc_thread **tdlist;
2537 	int i, tdcnt, tdlistsz;
2538 
2539 	KASSERT(!PROC_LOCKED(p), ("[pmc,%d] proc unexpectedly locked",
2540 	    __LINE__));
2541 	tdcnt = 32;
2542  restart:
2543 	tdlistsz = roundup2(tdcnt, 32);
2544 
2545 	tdcnt = 0;
2546 	tdlist = malloc(sizeof(struct pmc_thread*) * tdlistsz, M_TEMP, M_WAITOK);
2547 
2548 	PROC_LOCK(p);
2549 	FOREACH_THREAD_IN_PROC(p, curtd)
2550 		tdcnt++;
2551 	if (tdcnt >= tdlistsz) {
2552 		PROC_UNLOCK(p);
2553 		free(tdlist, M_TEMP);
2554 		goto restart;
2555 	}
2556 	/*
2557 	 * Try to add each thread to the list without sleeping. If unable,
2558 	 * add to a queue to retry after dropping the process lock.
2559 	 */
2560 	tdcnt = 0;
2561 	FOREACH_THREAD_IN_PROC(p, curtd) {
2562 		tdlist[tdcnt] = pmc_find_thread_descriptor(pp, curtd,
2563 						   PMC_FLAG_ALLOCATE|PMC_FLAG_NOWAIT);
2564 		if (tdlist[tdcnt] == NULL) {
2565 			PROC_UNLOCK(p);
2566 			for (i = 0; i <= tdcnt; i++)
2567 				pmc_thread_descriptor_pool_free(tdlist[i]);
2568 			free(tdlist, M_TEMP);
2569 			goto restart;
2570 		}
2571 		tdcnt++;
2572 	}
2573 	PROC_UNLOCK(p);
2574 	free(tdlist, M_TEMP);
2575 }
2576 
2577 /*
2578  * find the descriptor corresponding to process 'p', adding or removing it
2579  * as specified by 'mode'.
2580  */
2581 
2582 static struct pmc_process *
2583 pmc_find_process_descriptor(struct proc *p, uint32_t mode)
2584 {
2585 	uint32_t hindex;
2586 	struct pmc_process *pp, *ppnew;
2587 	struct pmc_processhash *pph;
2588 
2589 	hindex = PMC_HASH_PTR(p, pmc_processhashmask);
2590 	pph = &pmc_processhash[hindex];
2591 
2592 	ppnew = NULL;
2593 
2594 	/*
2595 	 * Pre-allocate memory in the PMC_FLAG_ALLOCATE case since we
2596 	 * cannot call malloc(9) once we hold a spin lock.
2597 	 */
2598 	if (mode & PMC_FLAG_ALLOCATE)
2599 		ppnew = malloc(sizeof(struct pmc_process) + md->pmd_npmc *
2600 		    sizeof(struct pmc_targetstate), M_PMC, M_WAITOK|M_ZERO);
2601 
2602 	mtx_lock_spin(&pmc_processhash_mtx);
2603 	LIST_FOREACH(pp, pph, pp_next)
2604 	    if (pp->pp_proc == p)
2605 		    break;
2606 
2607 	if ((mode & PMC_FLAG_REMOVE) && pp != NULL)
2608 		LIST_REMOVE(pp, pp_next);
2609 
2610 	if ((mode & PMC_FLAG_ALLOCATE) && pp == NULL &&
2611 	    ppnew != NULL) {
2612 		ppnew->pp_proc = p;
2613 		LIST_INIT(&ppnew->pp_tds);
2614 		ppnew->pp_tdslock = mtx_pool_find(pmc_mtxpool, ppnew);
2615 		LIST_INSERT_HEAD(pph, ppnew, pp_next);
2616 		mtx_unlock_spin(&pmc_processhash_mtx);
2617 		pp = ppnew;
2618 		ppnew = NULL;
2619 
2620 		/* Add thread descriptors for this process' current threads. */
2621 		pmc_add_thread_descriptors_from_proc(p, pp);
2622 	}
2623 	else
2624 		mtx_unlock_spin(&pmc_processhash_mtx);
2625 
2626 	if (ppnew != NULL)
2627 		free(ppnew, M_PMC);
2628 
2629 	return pp;
2630 }
2631 
2632 /*
2633  * remove a process descriptor from the process hash table.
2634  */
2635 
2636 static void
2637 pmc_remove_process_descriptor(struct pmc_process *pp)
2638 {
2639 	KASSERT(pp->pp_refcnt == 0,
2640 	    ("[pmc,%d] Removing process descriptor %p with count %d",
2641 		__LINE__, pp, pp->pp_refcnt));
2642 
2643 	mtx_lock_spin(&pmc_processhash_mtx);
2644 	LIST_REMOVE(pp, pp_next);
2645 	mtx_unlock_spin(&pmc_processhash_mtx);
2646 }
2647 
2648 /*
2649  * destroy a process descriptor.
2650  */
2651 
2652 static void
2653 pmc_destroy_process_descriptor(struct pmc_process *pp)
2654 {
2655 	struct pmc_thread *pmc_td;
2656 
2657 	while ((pmc_td = LIST_FIRST(&pp->pp_tds)) != NULL) {
2658 		LIST_REMOVE(pmc_td, pt_next);
2659 		pmc_thread_descriptor_pool_free(pmc_td);
2660 	}
2661 	free(pp, M_PMC);
2662 }
2663 
2664 
2665 /*
2666  * find an owner descriptor corresponding to proc 'p'
2667  */
2668 
2669 static struct pmc_owner *
2670 pmc_find_owner_descriptor(struct proc *p)
2671 {
2672 	uint32_t hindex;
2673 	struct pmc_owner *po;
2674 	struct pmc_ownerhash *poh;
2675 
2676 	hindex = PMC_HASH_PTR(p, pmc_ownerhashmask);
2677 	poh = &pmc_ownerhash[hindex];
2678 
2679 	po = NULL;
2680 	LIST_FOREACH(po, poh, po_next)
2681 	    if (po->po_owner == p)
2682 		    break;
2683 
2684 	PMCDBG5(OWN,FND,1, "find-owner proc=%p (%d, %s) hindex=0x%x -> "
2685 	    "pmc-owner=%p", p, p->p_pid, p->p_comm, hindex, po);
2686 
2687 	return po;
2688 }
2689 
2690 /*
2691  * pmc_allocate_pmc_descriptor
2692  *
2693  * Allocate a pmc descriptor and initialize its
2694  * fields.
2695  */
2696 
2697 static struct pmc *
2698 pmc_allocate_pmc_descriptor(void)
2699 {
2700 	struct pmc *pmc;
2701 
2702 	pmc = malloc(sizeof(struct pmc), M_PMC, M_WAITOK|M_ZERO);
2703 	pmc->pm_runcount = counter_u64_alloc(M_WAITOK);
2704 	pmc->pm_pcpu_state = malloc(sizeof(struct pmc_pcpu_state)*mp_ncpus, M_PMC, M_WAITOK|M_ZERO);
2705 	PMCDBG1(PMC,ALL,1, "allocate-pmc -> pmc=%p", pmc);
2706 
2707 	return pmc;
2708 }
2709 
2710 /*
2711  * Destroy a pmc descriptor.
2712  */
2713 
2714 static void
2715 pmc_destroy_pmc_descriptor(struct pmc *pm)
2716 {
2717 
2718 	KASSERT(pm->pm_state == PMC_STATE_DELETED ||
2719 	    pm->pm_state == PMC_STATE_FREE,
2720 	    ("[pmc,%d] destroying non-deleted PMC", __LINE__));
2721 	KASSERT(LIST_EMPTY(&pm->pm_targets),
2722 	    ("[pmc,%d] destroying pmc with targets", __LINE__));
2723 	KASSERT(pm->pm_owner == NULL,
2724 	    ("[pmc,%d] destroying pmc attached to an owner", __LINE__));
2725 	KASSERT(counter_u64_fetch(pm->pm_runcount) == 0,
2726 	    ("[pmc,%d] pmc has non-zero run count %ld", __LINE__,
2727 		 (unsigned long)counter_u64_fetch(pm->pm_runcount)));
2728 
2729 	counter_u64_free(pm->pm_runcount);
2730 	free(pm->pm_pcpu_state, M_PMC);
2731 	free(pm, M_PMC);
2732 }
2733 
2734 static void
2735 pmc_wait_for_pmc_idle(struct pmc *pm)
2736 {
2737 #ifdef INVARIANTS
2738 	volatile int maxloop;
2739 
2740 	maxloop = 100 * pmc_cpu_max();
2741 #endif
2742 	/*
2743 	 * Loop (with a forced context switch) till the PMC's runcount
2744 	 * comes down to zero.
2745 	 */
2746 	pmclog_flush(pm->pm_owner, 1);
2747 	while (counter_u64_fetch(pm->pm_runcount) > 0) {
2748 		pmclog_flush(pm->pm_owner, 1);
2749 #ifdef INVARIANTS
2750 		maxloop--;
2751 		KASSERT(maxloop > 0,
2752 		    ("[pmc,%d] (ri%d, rc%ld) waiting too long for "
2753 			"pmc to be free", __LINE__,
2754 			 PMC_TO_ROWINDEX(pm), (unsigned long)counter_u64_fetch(pm->pm_runcount)));
2755 #endif
2756 		pmc_force_context_switch();
2757 	}
2758 }
2759 
2760 /*
2761  * This function does the following things:
2762  *
2763  *  - detaches the PMC from hardware
2764  *  - unlinks all target threads that were attached to it
2765  *  - removes the PMC from its owner's list
2766  *  - destroys the PMC private mutex
2767  *
2768  * Once this function completes, the given pmc pointer can be freed by
2769  * calling pmc_destroy_pmc_descriptor().
2770  */
2771 
2772 static void
2773 pmc_release_pmc_descriptor(struct pmc *pm)
2774 {
2775 	enum pmc_mode mode;
2776 	struct pmc_hw *phw;
2777 	u_int adjri, ri, cpu;
2778 	struct pmc_owner *po;
2779 	struct pmc_binding pb;
2780 	struct pmc_process *pp;
2781 	struct pmc_classdep *pcd;
2782 	struct pmc_target *ptgt, *tmp;
2783 
2784 	sx_assert(&pmc_sx, SX_XLOCKED);
2785 
2786 	KASSERT(pm, ("[pmc,%d] null pmc", __LINE__));
2787 
2788 	ri   = PMC_TO_ROWINDEX(pm);
2789 	pcd  = pmc_ri_to_classdep(md, ri, &adjri);
2790 	mode = PMC_TO_MODE(pm);
2791 
2792 	PMCDBG3(PMC,REL,1, "release-pmc pmc=%p ri=%d mode=%d", pm, ri,
2793 	    mode);
2794 
2795 	/*
2796 	 * First, we take the PMC off hardware.
2797 	 */
2798 	cpu = 0;
2799 	if (PMC_IS_SYSTEM_MODE(mode)) {
2800 
2801 		/*
2802 		 * A system mode PMC runs on a specific CPU.  Switch
2803 		 * to this CPU and turn hardware off.
2804 		 */
2805 		pmc_save_cpu_binding(&pb);
2806 
2807 		cpu = PMC_TO_CPU(pm);
2808 
2809 		pmc_select_cpu(cpu);
2810 
2811 		/* switch off non-stalled CPUs */
2812 		pm->pm_pcpu_state[cpu].pps_cpustate = 0;
2813 		if (pm->pm_state == PMC_STATE_RUNNING &&
2814 			pm->pm_pcpu_state[cpu].pps_stalled == 0) {
2815 
2816 			phw = pmc_pcpu[cpu]->pc_hwpmcs[ri];
2817 
2818 			KASSERT(phw->phw_pmc == pm,
2819 			    ("[pmc, %d] pmc ptr ri(%d) hw(%p) pm(%p)",
2820 				__LINE__, ri, phw->phw_pmc, pm));
2821 			PMCDBG2(PMC,REL,2, "stopping cpu=%d ri=%d", cpu, ri);
2822 
2823 			critical_enter();
2824 			pcd->pcd_stop_pmc(cpu, adjri);
2825 			critical_exit();
2826 		}
2827 
2828 		PMCDBG2(PMC,REL,2, "decfg cpu=%d ri=%d", cpu, ri);
2829 
2830 		critical_enter();
2831 		pcd->pcd_config_pmc(cpu, adjri, NULL);
2832 		critical_exit();
2833 
2834 		/* adjust the global and process count of SS mode PMCs */
2835 		if (mode == PMC_MODE_SS && pm->pm_state == PMC_STATE_RUNNING) {
2836 			po = pm->pm_owner;
2837 			po->po_sscount--;
2838 			if (po->po_sscount == 0) {
2839 				atomic_subtract_rel_int(&pmc_ss_count, 1);
2840 				CK_LIST_REMOVE(po, po_ssnext);
2841 				epoch_wait_preempt(global_epoch_preempt);
2842 			}
2843 		}
2844 
2845 		pm->pm_state = PMC_STATE_DELETED;
2846 
2847 		pmc_restore_cpu_binding(&pb);
2848 
2849 		/*
2850 		 * We could have references to this PMC structure in
2851 		 * the per-cpu sample queues.  Wait for the queue to
2852 		 * drain.
2853 		 */
2854 		pmc_wait_for_pmc_idle(pm);
2855 
2856 	} else if (PMC_IS_VIRTUAL_MODE(mode)) {
2857 
2858 		/*
2859 		 * A virtual PMC could be running on multiple CPUs at
2860 		 * a given instant.
2861 		 *
2862 		 * By marking its state as DELETED, we ensure that
2863 		 * this PMC is never further scheduled on hardware.
2864 		 *
2865 		 * Then we wait till all CPUs are done with this PMC.
2866 		 */
2867 		pm->pm_state = PMC_STATE_DELETED;
2868 
2869 
2870 		/* Wait for the PMCs runcount to come to zero. */
2871 		pmc_wait_for_pmc_idle(pm);
2872 
2873 		/*
2874 		 * At this point the PMC is off all CPUs and cannot be
2875 		 * freshly scheduled onto a CPU.  It is now safe to
2876 		 * unlink all targets from this PMC.  If a
2877 		 * process-record's refcount falls to zero, we remove
2878 		 * it from the hash table.  The module-wide SX lock
2879 		 * protects us from races.
2880 		 */
2881 		LIST_FOREACH_SAFE(ptgt, &pm->pm_targets, pt_next, tmp) {
2882 			pp = ptgt->pt_process;
2883 			pmc_unlink_target_process(pm, pp); /* frees 'ptgt' */
2884 
2885 			PMCDBG1(PMC,REL,3, "pp->refcnt=%d", pp->pp_refcnt);
2886 
2887 			/*
2888 			 * If the target process record shows that no
2889 			 * PMCs are attached to it, reclaim its space.
2890 			 */
2891 
2892 			if (pp->pp_refcnt == 0) {
2893 				pmc_remove_process_descriptor(pp);
2894 				pmc_destroy_process_descriptor(pp);
2895 			}
2896 		}
2897 
2898 		cpu = curthread->td_oncpu; /* setup cpu for pmd_release() */
2899 
2900 	}
2901 
2902 	/*
2903 	 * Release any MD resources
2904 	 */
2905 	(void) pcd->pcd_release_pmc(cpu, adjri, pm);
2906 
2907 	/*
2908 	 * Update row disposition
2909 	 */
2910 
2911 	if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm)))
2912 		PMC_UNMARK_ROW_STANDALONE(ri);
2913 	else
2914 		PMC_UNMARK_ROW_THREAD(ri);
2915 
2916 	/* unlink from the owner's list */
2917 	if (pm->pm_owner) {
2918 		LIST_REMOVE(pm, pm_next);
2919 		pm->pm_owner = NULL;
2920 	}
2921 }
2922 
2923 /*
2924  * Register an owner and a pmc.
2925  */
2926 
2927 static int
2928 pmc_register_owner(struct proc *p, struct pmc *pmc)
2929 {
2930 	struct pmc_owner *po;
2931 
2932 	sx_assert(&pmc_sx, SX_XLOCKED);
2933 
2934 	if ((po = pmc_find_owner_descriptor(p)) == NULL)
2935 		if ((po = pmc_allocate_owner_descriptor(p)) == NULL)
2936 			return ENOMEM;
2937 
2938 	KASSERT(pmc->pm_owner == NULL,
2939 	    ("[pmc,%d] attempting to own an initialized PMC", __LINE__));
2940 	pmc->pm_owner  = po;
2941 
2942 	LIST_INSERT_HEAD(&po->po_pmcs, pmc, pm_next);
2943 
2944 	PROC_LOCK(p);
2945 	p->p_flag |= P_HWPMC;
2946 	PROC_UNLOCK(p);
2947 
2948 	if (po->po_flags & PMC_PO_OWNS_LOGFILE)
2949 		pmclog_process_pmcallocate(pmc);
2950 
2951 	PMCDBG2(PMC,REG,1, "register-owner pmc-owner=%p pmc=%p",
2952 	    po, pmc);
2953 
2954 	return 0;
2955 }
2956 
2957 /*
2958  * Return the current row disposition:
2959  * == 0 => FREE
2960  *  > 0 => PROCESS MODE
2961  *  < 0 => SYSTEM MODE
2962  */
2963 
2964 int
2965 pmc_getrowdisp(int ri)
2966 {
2967 	return pmc_pmcdisp[ri];
2968 }
2969 
2970 /*
2971  * Check if a PMC at row index 'ri' can be allocated to the current
2972  * process.
2973  *
2974  * Allocation can fail if:
2975  *   - the current process is already being profiled by a PMC at index 'ri',
2976  *     attached to it via OP_PMCATTACH.
2977  *   - the current process has already allocated a PMC at index 'ri'
2978  *     via OP_ALLOCATE.
2979  */
2980 
2981 static int
2982 pmc_can_allocate_rowindex(struct proc *p, unsigned int ri, int cpu)
2983 {
2984 	enum pmc_mode mode;
2985 	struct pmc *pm;
2986 	struct pmc_owner *po;
2987 	struct pmc_process *pp;
2988 
2989 	PMCDBG5(PMC,ALR,1, "can-allocate-rowindex proc=%p (%d, %s) ri=%d "
2990 	    "cpu=%d", p, p->p_pid, p->p_comm, ri, cpu);
2991 
2992 	/*
2993 	 * We shouldn't have already allocated a process-mode PMC at
2994 	 * row index 'ri'.
2995 	 *
2996 	 * We shouldn't have allocated a system-wide PMC on the same
2997 	 * CPU and same RI.
2998 	 */
2999 	if ((po = pmc_find_owner_descriptor(p)) != NULL)
3000 		LIST_FOREACH(pm, &po->po_pmcs, pm_next) {
3001 		    if (PMC_TO_ROWINDEX(pm) == ri) {
3002 			    mode = PMC_TO_MODE(pm);
3003 			    if (PMC_IS_VIRTUAL_MODE(mode))
3004 				    return EEXIST;
3005 			    if (PMC_IS_SYSTEM_MODE(mode) &&
3006 				(int) PMC_TO_CPU(pm) == cpu)
3007 				    return EEXIST;
3008 		    }
3009 	        }
3010 
3011 	/*
3012 	 * We also shouldn't be the target of any PMC at this index
3013 	 * since otherwise a PMC_ATTACH to ourselves will fail.
3014 	 */
3015 	if ((pp = pmc_find_process_descriptor(p, 0)) != NULL)
3016 		if (pp->pp_pmcs[ri].pp_pmc)
3017 			return EEXIST;
3018 
3019 	PMCDBG4(PMC,ALR,2, "can-allocate-rowindex proc=%p (%d, %s) ri=%d ok",
3020 	    p, p->p_pid, p->p_comm, ri);
3021 
3022 	return 0;
3023 }
3024 
3025 /*
3026  * Check if a given PMC at row index 'ri' can be currently used in
3027  * mode 'mode'.
3028  */
3029 
3030 static int
3031 pmc_can_allocate_row(int ri, enum pmc_mode mode)
3032 {
3033 	enum pmc_disp	disp;
3034 
3035 	sx_assert(&pmc_sx, SX_XLOCKED);
3036 
3037 	PMCDBG2(PMC,ALR,1, "can-allocate-row ri=%d mode=%d", ri, mode);
3038 
3039 	if (PMC_IS_SYSTEM_MODE(mode))
3040 		disp = PMC_DISP_STANDALONE;
3041 	else
3042 		disp = PMC_DISP_THREAD;
3043 
3044 	/*
3045 	 * check disposition for PMC row 'ri':
3046 	 *
3047 	 * Expected disposition		Row-disposition		Result
3048 	 *
3049 	 * STANDALONE			STANDALONE or FREE	proceed
3050 	 * STANDALONE			THREAD			fail
3051 	 * THREAD			THREAD or FREE		proceed
3052 	 * THREAD			STANDALONE		fail
3053 	 */
3054 
3055 	if (!PMC_ROW_DISP_IS_FREE(ri) &&
3056 	    !(disp == PMC_DISP_THREAD && PMC_ROW_DISP_IS_THREAD(ri)) &&
3057 	    !(disp == PMC_DISP_STANDALONE && PMC_ROW_DISP_IS_STANDALONE(ri)))
3058 		return EBUSY;
3059 
3060 	/*
3061 	 * All OK
3062 	 */
3063 
3064 	PMCDBG2(PMC,ALR,2, "can-allocate-row ri=%d mode=%d ok", ri, mode);
3065 
3066 	return 0;
3067 
3068 }
3069 
3070 /*
3071  * Find a PMC descriptor with user handle 'pmcid' for thread 'td'.
3072  */
3073 
3074 static struct pmc *
3075 pmc_find_pmc_descriptor_in_process(struct pmc_owner *po, pmc_id_t pmcid)
3076 {
3077 	struct pmc *pm;
3078 
3079 	KASSERT(PMC_ID_TO_ROWINDEX(pmcid) < md->pmd_npmc,
3080 	    ("[pmc,%d] Illegal pmc index %d (max %d)", __LINE__,
3081 		PMC_ID_TO_ROWINDEX(pmcid), md->pmd_npmc));
3082 
3083 	LIST_FOREACH(pm, &po->po_pmcs, pm_next)
3084 	    if (pm->pm_id == pmcid)
3085 		    return pm;
3086 
3087 	return NULL;
3088 }
3089 
3090 static int
3091 pmc_find_pmc(pmc_id_t pmcid, struct pmc **pmc)
3092 {
3093 
3094 	struct pmc *pm, *opm;
3095 	struct pmc_owner *po;
3096 	struct pmc_process *pp;
3097 
3098 	PMCDBG1(PMC,FND,1, "find-pmc id=%d", pmcid);
3099 	if (PMC_ID_TO_ROWINDEX(pmcid) >= md->pmd_npmc)
3100 		return (EINVAL);
3101 
3102 	if ((po = pmc_find_owner_descriptor(curthread->td_proc)) == NULL) {
3103 		/*
3104 		 * In case of PMC_F_DESCENDANTS child processes we will not find
3105 		 * the current process in the owners hash list.  Find the owner
3106 		 * process first and from there lookup the po.
3107 		 */
3108 		if ((pp = pmc_find_process_descriptor(curthread->td_proc,
3109 		    PMC_FLAG_NONE)) == NULL) {
3110 			return ESRCH;
3111 		} else {
3112 			opm = pp->pp_pmcs[PMC_ID_TO_ROWINDEX(pmcid)].pp_pmc;
3113 			if (opm == NULL)
3114 				return ESRCH;
3115 			if ((opm->pm_flags & (PMC_F_ATTACHED_TO_OWNER|
3116 			    PMC_F_DESCENDANTS)) != (PMC_F_ATTACHED_TO_OWNER|
3117 			    PMC_F_DESCENDANTS))
3118 				return ESRCH;
3119 			po = opm->pm_owner;
3120 		}
3121 	}
3122 
3123 	if ((pm = pmc_find_pmc_descriptor_in_process(po, pmcid)) == NULL)
3124 		return EINVAL;
3125 
3126 	PMCDBG2(PMC,FND,2, "find-pmc id=%d -> pmc=%p", pmcid, pm);
3127 
3128 	*pmc = pm;
3129 	return 0;
3130 }
3131 
3132 /*
3133  * Start a PMC.
3134  */
3135 
3136 static int
3137 pmc_start(struct pmc *pm)
3138 {
3139 	enum pmc_mode mode;
3140 	struct pmc_owner *po;
3141 	struct pmc_binding pb;
3142 	struct pmc_classdep *pcd;
3143 	int adjri, error, cpu, ri;
3144 
3145 	KASSERT(pm != NULL,
3146 	    ("[pmc,%d] null pm", __LINE__));
3147 
3148 	mode = PMC_TO_MODE(pm);
3149 	ri   = PMC_TO_ROWINDEX(pm);
3150 	pcd  = pmc_ri_to_classdep(md, ri, &adjri);
3151 
3152 	error = 0;
3153 
3154 	PMCDBG3(PMC,OPS,1, "start pmc=%p mode=%d ri=%d", pm, mode, ri);
3155 
3156 	po = pm->pm_owner;
3157 
3158 	/*
3159 	 * Disallow PMCSTART if a logfile is required but has not been
3160 	 * configured yet.
3161 	 */
3162 	if ((pm->pm_flags & PMC_F_NEEDS_LOGFILE) &&
3163 	    (po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
3164 		return (EDOOFUS);	/* programming error */
3165 
3166 	/*
3167 	 * If this is a sampling mode PMC, log mapping information for
3168 	 * the kernel modules that are currently loaded.
3169 	 */
3170 	if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
3171 	    pmc_log_kernel_mappings(pm);
3172 
3173 	if (PMC_IS_VIRTUAL_MODE(mode)) {
3174 
3175 		/*
3176 		 * If a PMCATTACH has never been done on this PMC,
3177 		 * attach it to its owner process.
3178 		 */
3179 
3180 		if (LIST_EMPTY(&pm->pm_targets))
3181 			error = (pm->pm_flags & PMC_F_ATTACH_DONE) ? ESRCH :
3182 			    pmc_attach_process(po->po_owner, pm);
3183 
3184 		/*
3185 		 * If the PMC is attached to its owner, then force a context
3186 		 * switch to ensure that the MD state gets set correctly.
3187 		 */
3188 
3189 		if (error == 0) {
3190 			pm->pm_state = PMC_STATE_RUNNING;
3191 			if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER)
3192 				pmc_force_context_switch();
3193 		}
3194 
3195 		return (error);
3196 	}
3197 
3198 
3199 	/*
3200 	 * A system-wide PMC.
3201 	 *
3202 	 * Add the owner to the global list if this is a system-wide
3203 	 * sampling PMC.
3204 	 */
3205 
3206 	if (mode == PMC_MODE_SS) {
3207 		/*
3208 		 * Log mapping information for all existing processes in the
3209 		 * system.  Subsequent mappings are logged as they happen;
3210 		 * see pmc_process_mmap().
3211 		 */
3212 		if (po->po_logprocmaps == 0) {
3213 			pmc_log_all_process_mappings(po);
3214 			po->po_logprocmaps = 1;
3215 		}
3216 		po->po_sscount++;
3217 		if (po->po_sscount == 1) {
3218 			atomic_add_rel_int(&pmc_ss_count, 1);
3219 			CK_LIST_INSERT_HEAD(&pmc_ss_owners, po, po_ssnext);
3220 			PMCDBG1(PMC,OPS,1, "po=%p in global list", po);
3221 		}
3222 	}
3223 
3224 	/*
3225 	 * Move to the CPU associated with this
3226 	 * PMC, and start the hardware.
3227 	 */
3228 
3229 	pmc_save_cpu_binding(&pb);
3230 
3231 	cpu = PMC_TO_CPU(pm);
3232 
3233 	if (!pmc_cpu_is_active(cpu))
3234 		return (ENXIO);
3235 
3236 	pmc_select_cpu(cpu);
3237 
3238 	/*
3239 	 * global PMCs are configured at allocation time
3240 	 * so write out the initial value and start the PMC.
3241 	 */
3242 
3243 	pm->pm_state = PMC_STATE_RUNNING;
3244 
3245 	critical_enter();
3246 	if ((error = pcd->pcd_write_pmc(cpu, adjri,
3247 		 PMC_IS_SAMPLING_MODE(mode) ?
3248 		 pm->pm_sc.pm_reloadcount :
3249 		 pm->pm_sc.pm_initial)) == 0) {
3250 		/* If a sampling mode PMC, reset stalled state. */
3251 		if (PMC_IS_SAMPLING_MODE(mode))
3252 			pm->pm_pcpu_state[cpu].pps_stalled = 0;
3253 
3254 		/* Indicate that we desire this to run. Start it. */
3255 		pm->pm_pcpu_state[cpu].pps_cpustate = 1;
3256 		error = pcd->pcd_start_pmc(cpu, adjri);
3257 	}
3258 	critical_exit();
3259 
3260 	pmc_restore_cpu_binding(&pb);
3261 
3262 	return (error);
3263 }
3264 
3265 /*
3266  * Stop a PMC.
3267  */
3268 
3269 static int
3270 pmc_stop(struct pmc *pm)
3271 {
3272 	struct pmc_owner *po;
3273 	struct pmc_binding pb;
3274 	struct pmc_classdep *pcd;
3275 	int adjri, cpu, error, ri;
3276 
3277 	KASSERT(pm != NULL, ("[pmc,%d] null pmc", __LINE__));
3278 
3279 	PMCDBG3(PMC,OPS,1, "stop pmc=%p mode=%d ri=%d", pm,
3280 	    PMC_TO_MODE(pm), PMC_TO_ROWINDEX(pm));
3281 
3282 	pm->pm_state = PMC_STATE_STOPPED;
3283 
3284 	/*
3285 	 * If the PMC is a virtual mode one, changing the state to
3286 	 * non-RUNNING is enough to ensure that the PMC never gets
3287 	 * scheduled.
3288 	 *
3289 	 * If this PMC is current running on a CPU, then it will
3290 	 * handled correctly at the time its target process is context
3291 	 * switched out.
3292 	 */
3293 
3294 	if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)))
3295 		return 0;
3296 
3297 	/*
3298 	 * A system-mode PMC.  Move to the CPU associated with
3299 	 * this PMC, and stop the hardware.  We update the
3300 	 * 'initial count' so that a subsequent PMCSTART will
3301 	 * resume counting from the current hardware count.
3302 	 */
3303 
3304 	pmc_save_cpu_binding(&pb);
3305 
3306 	cpu = PMC_TO_CPU(pm);
3307 
3308 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
3309 	    ("[pmc,%d] illegal cpu=%d", __LINE__, cpu));
3310 
3311 	if (!pmc_cpu_is_active(cpu))
3312 		return ENXIO;
3313 
3314 	pmc_select_cpu(cpu);
3315 
3316 	ri = PMC_TO_ROWINDEX(pm);
3317 	pcd = pmc_ri_to_classdep(md, ri, &adjri);
3318 
3319 	pm->pm_pcpu_state[cpu].pps_cpustate = 0;
3320 	critical_enter();
3321 	if ((error = pcd->pcd_stop_pmc(cpu, adjri)) == 0)
3322 		error = pcd->pcd_read_pmc(cpu, adjri, &pm->pm_sc.pm_initial);
3323 	critical_exit();
3324 
3325 	pmc_restore_cpu_binding(&pb);
3326 
3327 	po = pm->pm_owner;
3328 
3329 	/* remove this owner from the global list of SS PMC owners */
3330 	if (PMC_TO_MODE(pm) == PMC_MODE_SS) {
3331 		po->po_sscount--;
3332 		if (po->po_sscount == 0) {
3333 			atomic_subtract_rel_int(&pmc_ss_count, 1);
3334 			CK_LIST_REMOVE(po, po_ssnext);
3335 			epoch_wait_preempt(global_epoch_preempt);
3336 			PMCDBG1(PMC,OPS,2,"po=%p removed from global list", po);
3337 		}
3338 	}
3339 
3340 	return (error);
3341 }
3342 
3343 static struct pmc_classdep *
3344 pmc_class_to_classdep(enum pmc_class class)
3345 {
3346 	int n;
3347 
3348 	for (n = 0; n < md->pmd_nclass; n++)
3349 		if (md->pmd_classdep[n].pcd_class == class)
3350 			return (&md->pmd_classdep[n]);
3351 	return (NULL);
3352 }
3353 
3354 #if defined(HWPMC_DEBUG) && defined(KTR)
3355 static const char *pmc_op_to_name[] = {
3356 #undef	__PMC_OP
3357 #define	__PMC_OP(N, D)	#N ,
3358 	__PMC_OPS()
3359 	NULL
3360 };
3361 #endif
3362 
3363 /*
3364  * The syscall interface
3365  */
3366 
3367 #define	PMC_GET_SX_XLOCK(...) do {		\
3368 	sx_xlock(&pmc_sx);			\
3369 	if (pmc_hook == NULL) {			\
3370 		sx_xunlock(&pmc_sx);		\
3371 		return __VA_ARGS__;		\
3372 	}					\
3373 } while (0)
3374 
3375 #define	PMC_DOWNGRADE_SX() do {			\
3376 	sx_downgrade(&pmc_sx);			\
3377 	is_sx_downgraded = 1;			\
3378 } while (0)
3379 
3380 static int
3381 pmc_syscall_handler(struct thread *td, void *syscall_args)
3382 {
3383 	int error, is_sx_downgraded, op;
3384 	struct pmc_syscall_args *c;
3385 	void *pmclog_proc_handle;
3386 	void *arg;
3387 
3388 	c = (struct pmc_syscall_args *)syscall_args;
3389 	op = c->pmop_code;
3390 	arg = c->pmop_data;
3391 	/* PMC isn't set up yet */
3392 	if (pmc_hook == NULL)
3393 		return (EINVAL);
3394 	if (op == PMC_OP_CONFIGURELOG) {
3395 		/*
3396 		 * We cannot create the logging process inside
3397 		 * pmclog_configure_log() because there is a LOR
3398 		 * between pmc_sx and process structure locks.
3399 		 * Instead, pre-create the process and ignite the loop
3400 		 * if everything is fine, otherwise direct the process
3401 		 * to exit.
3402 		 */
3403 		error = pmclog_proc_create(td, &pmclog_proc_handle);
3404 		if (error != 0)
3405 			goto done_syscall;
3406 	}
3407 
3408 	PMC_GET_SX_XLOCK(ENOSYS);
3409 	is_sx_downgraded = 0;
3410 	PMCDBG3(MOD,PMS,1, "syscall op=%d \"%s\" arg=%p", op,
3411 	    pmc_op_to_name[op], arg);
3412 
3413 	error = 0;
3414 	counter_u64_add(pmc_stats.pm_syscalls, 1);
3415 
3416 	switch (op) {
3417 
3418 
3419 	/*
3420 	 * Configure a log file.
3421 	 *
3422 	 * XXX This OP will be reworked.
3423 	 */
3424 
3425 	case PMC_OP_CONFIGURELOG:
3426 	{
3427 		struct proc *p;
3428 		struct pmc *pm;
3429 		struct pmc_owner *po;
3430 		struct pmc_op_configurelog cl;
3431 
3432 		if ((error = copyin(arg, &cl, sizeof(cl))) != 0) {
3433 			pmclog_proc_ignite(pmclog_proc_handle, NULL);
3434 			break;
3435 		}
3436 
3437 		/* mark this process as owning a log file */
3438 		p = td->td_proc;
3439 		if ((po = pmc_find_owner_descriptor(p)) == NULL)
3440 			if ((po = pmc_allocate_owner_descriptor(p)) == NULL) {
3441 				pmclog_proc_ignite(pmclog_proc_handle, NULL);
3442 				error = ENOMEM;
3443 				break;
3444 			}
3445 
3446 		/*
3447 		 * If a valid fd was passed in, try to configure that,
3448 		 * otherwise if 'fd' was less than zero and there was
3449 		 * a log file configured, flush its buffers and
3450 		 * de-configure it.
3451 		 */
3452 		if (cl.pm_logfd >= 0) {
3453 			error = pmclog_configure_log(md, po, cl.pm_logfd);
3454 			pmclog_proc_ignite(pmclog_proc_handle, error == 0 ?
3455 			    po : NULL);
3456 		} else if (po->po_flags & PMC_PO_OWNS_LOGFILE) {
3457 			pmclog_proc_ignite(pmclog_proc_handle, NULL);
3458 			error = pmclog_close(po);
3459 			if (error == 0) {
3460 				LIST_FOREACH(pm, &po->po_pmcs, pm_next)
3461 				    if (pm->pm_flags & PMC_F_NEEDS_LOGFILE &&
3462 					pm->pm_state == PMC_STATE_RUNNING)
3463 					    pmc_stop(pm);
3464 				error = pmclog_deconfigure_log(po);
3465 			}
3466 		} else {
3467 			pmclog_proc_ignite(pmclog_proc_handle, NULL);
3468 			error = EINVAL;
3469 		}
3470 	}
3471 	break;
3472 
3473 	/*
3474 	 * Flush a log file.
3475 	 */
3476 
3477 	case PMC_OP_FLUSHLOG:
3478 	{
3479 		struct pmc_owner *po;
3480 
3481 		sx_assert(&pmc_sx, SX_XLOCKED);
3482 
3483 		if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) {
3484 			error = EINVAL;
3485 			break;
3486 		}
3487 
3488 		error = pmclog_flush(po, 0);
3489 	}
3490 	break;
3491 
3492 	/*
3493 	 * Close a log file.
3494 	 */
3495 
3496 	case PMC_OP_CLOSELOG:
3497 	{
3498 		struct pmc_owner *po;
3499 
3500 		sx_assert(&pmc_sx, SX_XLOCKED);
3501 
3502 		if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) {
3503 			error = EINVAL;
3504 			break;
3505 		}
3506 
3507 		error = pmclog_close(po);
3508 	}
3509 	break;
3510 
3511 	/*
3512 	 * Retrieve hardware configuration.
3513 	 */
3514 
3515 	case PMC_OP_GETCPUINFO:	/* CPU information */
3516 	{
3517 		struct pmc_op_getcpuinfo gci;
3518 		struct pmc_classinfo *pci;
3519 		struct pmc_classdep *pcd;
3520 		int cl;
3521 
3522 		gci.pm_cputype = md->pmd_cputype;
3523 		gci.pm_ncpu    = pmc_cpu_max();
3524 		gci.pm_npmc    = md->pmd_npmc;
3525 		gci.pm_nclass  = md->pmd_nclass;
3526 		pci = gci.pm_classes;
3527 		pcd = md->pmd_classdep;
3528 		for (cl = 0; cl < md->pmd_nclass; cl++, pci++, pcd++) {
3529 			pci->pm_caps  = pcd->pcd_caps;
3530 			pci->pm_class = pcd->pcd_class;
3531 			pci->pm_width = pcd->pcd_width;
3532 			pci->pm_num   = pcd->pcd_num;
3533 		}
3534 		error = copyout(&gci, arg, sizeof(gci));
3535 	}
3536 	break;
3537 
3538 	/*
3539 	 * Retrieve soft events list.
3540 	 */
3541 	case PMC_OP_GETDYNEVENTINFO:
3542 	{
3543 		enum pmc_class			cl;
3544 		enum pmc_event			ev;
3545 		struct pmc_op_getdyneventinfo	*gei;
3546 		struct pmc_dyn_event_descr	dev;
3547 		struct pmc_soft			*ps;
3548 		uint32_t			nevent;
3549 
3550 		sx_assert(&pmc_sx, SX_LOCKED);
3551 
3552 		gei = (struct pmc_op_getdyneventinfo *) arg;
3553 
3554 		if ((error = copyin(&gei->pm_class, &cl, sizeof(cl))) != 0)
3555 			break;
3556 
3557 		/* Only SOFT class is dynamic. */
3558 		if (cl != PMC_CLASS_SOFT) {
3559 			error = EINVAL;
3560 			break;
3561 		}
3562 
3563 		nevent = 0;
3564 		for (ev = PMC_EV_SOFT_FIRST; (int)ev <= PMC_EV_SOFT_LAST; ev++) {
3565 			ps = pmc_soft_ev_acquire(ev);
3566 			if (ps == NULL)
3567 				continue;
3568 			bcopy(&ps->ps_ev, &dev, sizeof(dev));
3569 			pmc_soft_ev_release(ps);
3570 
3571 			error = copyout(&dev,
3572 			    &gei->pm_events[nevent],
3573 			    sizeof(struct pmc_dyn_event_descr));
3574 			if (error != 0)
3575 				break;
3576 			nevent++;
3577 		}
3578 		if (error != 0)
3579 			break;
3580 
3581 		error = copyout(&nevent, &gei->pm_nevent,
3582 		    sizeof(nevent));
3583 	}
3584 	break;
3585 
3586 	/*
3587 	 * Get module statistics
3588 	 */
3589 
3590 	case PMC_OP_GETDRIVERSTATS:
3591 	{
3592 		struct pmc_op_getdriverstats gms;
3593 #define CFETCH(a, b, field) a.field = counter_u64_fetch(b.field)
3594 		CFETCH(gms, pmc_stats, pm_intr_ignored);
3595 		CFETCH(gms, pmc_stats, pm_intr_processed);
3596 		CFETCH(gms, pmc_stats, pm_intr_bufferfull);
3597 		CFETCH(gms, pmc_stats, pm_syscalls);
3598 		CFETCH(gms, pmc_stats, pm_syscall_errors);
3599 		CFETCH(gms, pmc_stats, pm_buffer_requests);
3600 		CFETCH(gms, pmc_stats, pm_buffer_requests_failed);
3601 		CFETCH(gms, pmc_stats, pm_log_sweeps);
3602 #undef CFETCH
3603 		error = copyout(&gms, arg, sizeof(gms));
3604 	}
3605 	break;
3606 
3607 
3608 	/*
3609 	 * Retrieve module version number
3610 	 */
3611 
3612 	case PMC_OP_GETMODULEVERSION:
3613 	{
3614 		uint32_t cv, modv;
3615 
3616 		/* retrieve the client's idea of the ABI version */
3617 		if ((error = copyin(arg, &cv, sizeof(uint32_t))) != 0)
3618 			break;
3619 		/* don't service clients newer than our driver */
3620 		modv = PMC_VERSION;
3621 		if ((cv & 0xFFFF0000) > (modv & 0xFFFF0000)) {
3622 			error = EPROGMISMATCH;
3623 			break;
3624 		}
3625 		error = copyout(&modv, arg, sizeof(int));
3626 	}
3627 	break;
3628 
3629 
3630 	/*
3631 	 * Retrieve the state of all the PMCs on a given
3632 	 * CPU.
3633 	 */
3634 
3635 	case PMC_OP_GETPMCINFO:
3636 	{
3637 		int ari;
3638 		struct pmc *pm;
3639 		size_t pmcinfo_size;
3640 		uint32_t cpu, n, npmc;
3641 		struct pmc_owner *po;
3642 		struct pmc_binding pb;
3643 		struct pmc_classdep *pcd;
3644 		struct pmc_info *p, *pmcinfo;
3645 		struct pmc_op_getpmcinfo *gpi;
3646 
3647 		PMC_DOWNGRADE_SX();
3648 
3649 		gpi = (struct pmc_op_getpmcinfo *) arg;
3650 
3651 		if ((error = copyin(&gpi->pm_cpu, &cpu, sizeof(cpu))) != 0)
3652 			break;
3653 
3654 		if (cpu >= pmc_cpu_max()) {
3655 			error = EINVAL;
3656 			break;
3657 		}
3658 
3659 		if (!pmc_cpu_is_active(cpu)) {
3660 			error = ENXIO;
3661 			break;
3662 		}
3663 
3664 		/* switch to CPU 'cpu' */
3665 		pmc_save_cpu_binding(&pb);
3666 		pmc_select_cpu(cpu);
3667 
3668 		npmc = md->pmd_npmc;
3669 
3670 		pmcinfo_size = npmc * sizeof(struct pmc_info);
3671 		pmcinfo = malloc(pmcinfo_size, M_PMC, M_WAITOK);
3672 
3673 		p = pmcinfo;
3674 
3675 		for (n = 0; n < md->pmd_npmc; n++, p++) {
3676 
3677 			pcd = pmc_ri_to_classdep(md, n, &ari);
3678 
3679 			KASSERT(pcd != NULL,
3680 			    ("[pmc,%d] null pcd ri=%d", __LINE__, n));
3681 
3682 			if ((error = pcd->pcd_describe(cpu, ari, p, &pm)) != 0)
3683 				break;
3684 
3685 			if (PMC_ROW_DISP_IS_STANDALONE(n))
3686 				p->pm_rowdisp = PMC_DISP_STANDALONE;
3687 			else if (PMC_ROW_DISP_IS_THREAD(n))
3688 				p->pm_rowdisp = PMC_DISP_THREAD;
3689 			else
3690 				p->pm_rowdisp = PMC_DISP_FREE;
3691 
3692 			p->pm_ownerpid = -1;
3693 
3694 			if (pm == NULL)	/* no PMC associated */
3695 				continue;
3696 
3697 			po = pm->pm_owner;
3698 
3699 			KASSERT(po->po_owner != NULL,
3700 			    ("[pmc,%d] pmc_owner had a null proc pointer",
3701 				__LINE__));
3702 
3703 			p->pm_ownerpid = po->po_owner->p_pid;
3704 			p->pm_mode     = PMC_TO_MODE(pm);
3705 			p->pm_event    = pm->pm_event;
3706 			p->pm_flags    = pm->pm_flags;
3707 
3708 			if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
3709 				p->pm_reloadcount =
3710 				    pm->pm_sc.pm_reloadcount;
3711 		}
3712 
3713 		pmc_restore_cpu_binding(&pb);
3714 
3715 		/* now copy out the PMC info collected */
3716 		if (error == 0)
3717 			error = copyout(pmcinfo, &gpi->pm_pmcs, pmcinfo_size);
3718 
3719 		free(pmcinfo, M_PMC);
3720 	}
3721 	break;
3722 
3723 
3724 	/*
3725 	 * Set the administrative state of a PMC.  I.e. whether
3726 	 * the PMC is to be used or not.
3727 	 */
3728 
3729 	case PMC_OP_PMCADMIN:
3730 	{
3731 		int cpu, ri;
3732 		enum pmc_state request;
3733 		struct pmc_cpu *pc;
3734 		struct pmc_hw *phw;
3735 		struct pmc_op_pmcadmin pma;
3736 		struct pmc_binding pb;
3737 
3738 		sx_assert(&pmc_sx, SX_XLOCKED);
3739 
3740 		KASSERT(td == curthread,
3741 		    ("[pmc,%d] td != curthread", __LINE__));
3742 
3743 		error = priv_check(td, PRIV_PMC_MANAGE);
3744 		if (error)
3745 			break;
3746 
3747 		if ((error = copyin(arg, &pma, sizeof(pma))) != 0)
3748 			break;
3749 
3750 		cpu = pma.pm_cpu;
3751 
3752 		if (cpu < 0 || cpu >= (int) pmc_cpu_max()) {
3753 			error = EINVAL;
3754 			break;
3755 		}
3756 
3757 		if (!pmc_cpu_is_active(cpu)) {
3758 			error = ENXIO;
3759 			break;
3760 		}
3761 
3762 		request = pma.pm_state;
3763 
3764 		if (request != PMC_STATE_DISABLED &&
3765 		    request != PMC_STATE_FREE) {
3766 			error = EINVAL;
3767 			break;
3768 		}
3769 
3770 		ri = pma.pm_pmc; /* pmc id == row index */
3771 		if (ri < 0 || ri >= (int) md->pmd_npmc) {
3772 			error = EINVAL;
3773 			break;
3774 		}
3775 
3776 		/*
3777 		 * We can't disable a PMC with a row-index allocated
3778 		 * for process virtual PMCs.
3779 		 */
3780 
3781 		if (PMC_ROW_DISP_IS_THREAD(ri) &&
3782 		    request == PMC_STATE_DISABLED) {
3783 			error = EBUSY;
3784 			break;
3785 		}
3786 
3787 		/*
3788 		 * otherwise, this PMC on this CPU is either free or
3789 		 * in system-wide mode.
3790 		 */
3791 
3792 		pmc_save_cpu_binding(&pb);
3793 		pmc_select_cpu(cpu);
3794 
3795 		pc  = pmc_pcpu[cpu];
3796 		phw = pc->pc_hwpmcs[ri];
3797 
3798 		/*
3799 		 * XXX do we need some kind of 'forced' disable?
3800 		 */
3801 
3802 		if (phw->phw_pmc == NULL) {
3803 			if (request == PMC_STATE_DISABLED &&
3804 			    (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED)) {
3805 				phw->phw_state &= ~PMC_PHW_FLAG_IS_ENABLED;
3806 				PMC_MARK_ROW_STANDALONE(ri);
3807 			} else if (request == PMC_STATE_FREE &&
3808 			    (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0) {
3809 				phw->phw_state |=  PMC_PHW_FLAG_IS_ENABLED;
3810 				PMC_UNMARK_ROW_STANDALONE(ri);
3811 			}
3812 			/* other cases are a no-op */
3813 		} else
3814 			error = EBUSY;
3815 
3816 		pmc_restore_cpu_binding(&pb);
3817 	}
3818 	break;
3819 
3820 
3821 	/*
3822 	 * Allocate a PMC.
3823 	 */
3824 
3825 	case PMC_OP_PMCALLOCATE:
3826 	{
3827 		int adjri, n;
3828 		u_int cpu;
3829 		uint32_t caps;
3830 		struct pmc *pmc;
3831 		enum pmc_mode mode;
3832 		struct pmc_hw *phw;
3833 		struct pmc_binding pb;
3834 		struct pmc_classdep *pcd;
3835 		struct pmc_op_pmcallocate pa;
3836 
3837 		if ((error = copyin(arg, &pa, sizeof(pa))) != 0)
3838 			break;
3839 
3840 		caps = pa.pm_caps;
3841 		mode = pa.pm_mode;
3842 		cpu  = pa.pm_cpu;
3843 
3844 		if ((mode != PMC_MODE_SS  &&  mode != PMC_MODE_SC  &&
3845 		     mode != PMC_MODE_TS  &&  mode != PMC_MODE_TC) ||
3846 		    (cpu != (u_int) PMC_CPU_ANY && cpu >= pmc_cpu_max())) {
3847 			error = EINVAL;
3848 			break;
3849 		}
3850 
3851 		/*
3852 		 * Virtual PMCs should only ask for a default CPU.
3853 		 * System mode PMCs need to specify a non-default CPU.
3854 		 */
3855 
3856 		if ((PMC_IS_VIRTUAL_MODE(mode) && cpu != (u_int) PMC_CPU_ANY) ||
3857 		    (PMC_IS_SYSTEM_MODE(mode) && cpu == (u_int) PMC_CPU_ANY)) {
3858 			error = EINVAL;
3859 			break;
3860 		}
3861 
3862 		/*
3863 		 * Check that an inactive CPU is not being asked for.
3864 		 */
3865 
3866 		if (PMC_IS_SYSTEM_MODE(mode) && !pmc_cpu_is_active(cpu)) {
3867 			error = ENXIO;
3868 			break;
3869 		}
3870 
3871 		/*
3872 		 * Refuse an allocation for a system-wide PMC if this
3873 		 * process has been jailed, or if this process lacks
3874 		 * super-user credentials and the sysctl tunable
3875 		 * 'security.bsd.unprivileged_syspmcs' is zero.
3876 		 */
3877 
3878 		if (PMC_IS_SYSTEM_MODE(mode)) {
3879 			if (jailed(curthread->td_ucred)) {
3880 				error = EPERM;
3881 				break;
3882 			}
3883 			if (!pmc_unprivileged_syspmcs) {
3884 				error = priv_check(curthread,
3885 				    PRIV_PMC_SYSTEM);
3886 				if (error)
3887 					break;
3888 			}
3889 		}
3890 
3891 		/*
3892 		 * Look for valid values for 'pm_flags'
3893 		 */
3894 
3895 		if ((pa.pm_flags & ~(PMC_F_DESCENDANTS | PMC_F_LOG_PROCCSW |
3896 		    PMC_F_LOG_PROCEXIT | PMC_F_CALLCHAIN |
3897 		    PMC_F_USERCALLCHAIN)) != 0) {
3898 			error = EINVAL;
3899 			break;
3900 		}
3901 
3902 		/* PMC_F_USERCALLCHAIN is only valid with PMC_F_CALLCHAIN */
3903 		if ((pa.pm_flags & (PMC_F_CALLCHAIN | PMC_F_USERCALLCHAIN)) ==
3904 		    PMC_F_USERCALLCHAIN) {
3905 			error = EINVAL;
3906 			break;
3907 		}
3908 
3909 		/* PMC_F_USERCALLCHAIN is only valid for sampling mode */
3910 		if (pa.pm_flags & PMC_F_USERCALLCHAIN &&
3911 			mode != PMC_MODE_TS && mode != PMC_MODE_SS) {
3912 			error = EINVAL;
3913 			break;
3914 		}
3915 
3916 		/* process logging options are not allowed for system PMCs */
3917 		if (PMC_IS_SYSTEM_MODE(mode) && (pa.pm_flags &
3918 		    (PMC_F_LOG_PROCCSW | PMC_F_LOG_PROCEXIT))) {
3919 			error = EINVAL;
3920 			break;
3921 		}
3922 
3923 		/*
3924 		 * All sampling mode PMCs need to be able to interrupt the
3925 		 * CPU.
3926 		 */
3927 		if (PMC_IS_SAMPLING_MODE(mode))
3928 			caps |= PMC_CAP_INTERRUPT;
3929 
3930 		/* A valid class specifier should have been passed in. */
3931 		pcd = pmc_class_to_classdep(pa.pm_class);
3932 		if (pcd == NULL) {
3933 			error = EINVAL;
3934 			break;
3935 		}
3936 
3937 		/* The requested PMC capabilities should be feasible. */
3938 		if ((pcd->pcd_caps & caps) != caps) {
3939 			error = EOPNOTSUPP;
3940 			break;
3941 		}
3942 
3943 		PMCDBG4(PMC,ALL,2, "event=%d caps=0x%x mode=%d cpu=%d",
3944 		    pa.pm_ev, caps, mode, cpu);
3945 
3946 		pmc = pmc_allocate_pmc_descriptor();
3947 		pmc->pm_id    = PMC_ID_MAKE_ID(cpu,pa.pm_mode,pa.pm_class,
3948 		    PMC_ID_INVALID);
3949 		pmc->pm_event = pa.pm_ev;
3950 		pmc->pm_state = PMC_STATE_FREE;
3951 		pmc->pm_caps  = caps;
3952 		pmc->pm_flags = pa.pm_flags;
3953 
3954 		/* XXX set lower bound on sampling for process counters */
3955 		if (PMC_IS_SAMPLING_MODE(mode)) {
3956 			/*
3957 			 * Don't permit requested sample rate to be less than 1000
3958 			 */
3959 			if (pa.pm_count < 1000)
3960 				log(LOG_WARNING,
3961 					"pmcallocate: passed sample rate %ju - setting to 1000\n",
3962 					(uintmax_t)pa.pm_count);
3963 			pmc->pm_sc.pm_reloadcount = MAX(1000, pa.pm_count);
3964 		} else
3965 			pmc->pm_sc.pm_initial = pa.pm_count;
3966 
3967 		/* switch thread to CPU 'cpu' */
3968 		pmc_save_cpu_binding(&pb);
3969 
3970 #define	PMC_IS_SHAREABLE_PMC(cpu, n)				\
3971 	(pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_state &		\
3972 	 PMC_PHW_FLAG_IS_SHAREABLE)
3973 #define	PMC_IS_UNALLOCATED(cpu, n)				\
3974 	(pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_pmc == NULL)
3975 
3976 		if (PMC_IS_SYSTEM_MODE(mode)) {
3977 			pmc_select_cpu(cpu);
3978 			for (n = pcd->pcd_ri; n < (int) md->pmd_npmc; n++) {
3979 				pcd = pmc_ri_to_classdep(md, n, &adjri);
3980 				if (pmc_can_allocate_row(n, mode) == 0 &&
3981 				    pmc_can_allocate_rowindex(
3982 					    curthread->td_proc, n, cpu) == 0 &&
3983 				    (PMC_IS_UNALLOCATED(cpu, n) ||
3984 				     PMC_IS_SHAREABLE_PMC(cpu, n)) &&
3985 				    pcd->pcd_allocate_pmc(cpu, adjri, pmc,
3986 					&pa) == 0)
3987 					break;
3988 			}
3989 		} else {
3990 			/* Process virtual mode */
3991 			for (n = pcd->pcd_ri; n < (int) md->pmd_npmc; n++) {
3992 				pcd = pmc_ri_to_classdep(md, n, &adjri);
3993 				if (pmc_can_allocate_row(n, mode) == 0 &&
3994 				    pmc_can_allocate_rowindex(
3995 					    curthread->td_proc, n,
3996 					    PMC_CPU_ANY) == 0 &&
3997 				    pcd->pcd_allocate_pmc(curthread->td_oncpu,
3998 					adjri, pmc, &pa) == 0)
3999 					break;
4000 			}
4001 		}
4002 
4003 #undef	PMC_IS_UNALLOCATED
4004 #undef	PMC_IS_SHAREABLE_PMC
4005 
4006 		pmc_restore_cpu_binding(&pb);
4007 
4008 		if (n == (int) md->pmd_npmc) {
4009 			pmc_destroy_pmc_descriptor(pmc);
4010 			pmc = NULL;
4011 			error = EINVAL;
4012 			break;
4013 		}
4014 
4015 		/* Fill in the correct value in the ID field */
4016 		pmc->pm_id = PMC_ID_MAKE_ID(cpu,mode,pa.pm_class,n);
4017 
4018 		PMCDBG5(PMC,ALL,2, "ev=%d class=%d mode=%d n=%d -> pmcid=%x",
4019 		    pmc->pm_event, pa.pm_class, mode, n, pmc->pm_id);
4020 
4021 		/* Process mode PMCs with logging enabled need log files */
4022 		if (pmc->pm_flags & (PMC_F_LOG_PROCEXIT | PMC_F_LOG_PROCCSW))
4023 			pmc->pm_flags |= PMC_F_NEEDS_LOGFILE;
4024 
4025 		/* All system mode sampling PMCs require a log file */
4026 		if (PMC_IS_SAMPLING_MODE(mode) && PMC_IS_SYSTEM_MODE(mode))
4027 			pmc->pm_flags |= PMC_F_NEEDS_LOGFILE;
4028 
4029 		/*
4030 		 * Configure global pmc's immediately
4031 		 */
4032 
4033 		if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pmc))) {
4034 
4035 			pmc_save_cpu_binding(&pb);
4036 			pmc_select_cpu(cpu);
4037 
4038 			phw = pmc_pcpu[cpu]->pc_hwpmcs[n];
4039 			pcd = pmc_ri_to_classdep(md, n, &adjri);
4040 
4041 			if ((phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0 ||
4042 			    (error = pcd->pcd_config_pmc(cpu, adjri, pmc)) != 0) {
4043 				(void) pcd->pcd_release_pmc(cpu, adjri, pmc);
4044 				pmc_destroy_pmc_descriptor(pmc);
4045 				pmc = NULL;
4046 				pmc_restore_cpu_binding(&pb);
4047 				error = EPERM;
4048 				break;
4049 			}
4050 
4051 			pmc_restore_cpu_binding(&pb);
4052 		}
4053 
4054 		pmc->pm_state    = PMC_STATE_ALLOCATED;
4055 		pmc->pm_class	= pa.pm_class;
4056 
4057 		/*
4058 		 * mark row disposition
4059 		 */
4060 
4061 		if (PMC_IS_SYSTEM_MODE(mode))
4062 			PMC_MARK_ROW_STANDALONE(n);
4063 		else
4064 			PMC_MARK_ROW_THREAD(n);
4065 
4066 		/*
4067 		 * Register this PMC with the current thread as its owner.
4068 		 */
4069 
4070 		if ((error =
4071 		    pmc_register_owner(curthread->td_proc, pmc)) != 0) {
4072 			pmc_release_pmc_descriptor(pmc);
4073 			pmc_destroy_pmc_descriptor(pmc);
4074 			pmc = NULL;
4075 			break;
4076 		}
4077 
4078 
4079 		/*
4080 		 * Return the allocated index.
4081 		 */
4082 
4083 		pa.pm_pmcid = pmc->pm_id;
4084 
4085 		error = copyout(&pa, arg, sizeof(pa));
4086 	}
4087 	break;
4088 
4089 
4090 	/*
4091 	 * Attach a PMC to a process.
4092 	 */
4093 
4094 	case PMC_OP_PMCATTACH:
4095 	{
4096 		struct pmc *pm;
4097 		struct proc *p;
4098 		struct pmc_op_pmcattach a;
4099 
4100 		sx_assert(&pmc_sx, SX_XLOCKED);
4101 
4102 		if ((error = copyin(arg, &a, sizeof(a))) != 0)
4103 			break;
4104 
4105 		if (a.pm_pid < 0) {
4106 			error = EINVAL;
4107 			break;
4108 		} else if (a.pm_pid == 0)
4109 			a.pm_pid = td->td_proc->p_pid;
4110 
4111 		if ((error = pmc_find_pmc(a.pm_pmc, &pm)) != 0)
4112 			break;
4113 
4114 		if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm))) {
4115 			error = EINVAL;
4116 			break;
4117 		}
4118 
4119 		/* PMCs may be (re)attached only when allocated or stopped */
4120 		if (pm->pm_state == PMC_STATE_RUNNING) {
4121 			error = EBUSY;
4122 			break;
4123 		} else if (pm->pm_state != PMC_STATE_ALLOCATED &&
4124 		    pm->pm_state != PMC_STATE_STOPPED) {
4125 			error = EINVAL;
4126 			break;
4127 		}
4128 
4129 		/* lookup pid */
4130 		if ((p = pfind(a.pm_pid)) == NULL) {
4131 			error = ESRCH;
4132 			break;
4133 		}
4134 
4135 		/*
4136 		 * Ignore processes that are working on exiting.
4137 		 */
4138 		if (p->p_flag & P_WEXIT) {
4139 			error = ESRCH;
4140 			PROC_UNLOCK(p);	/* pfind() returns a locked process */
4141 			break;
4142 		}
4143 
4144 		/*
4145 		 * we are allowed to attach a PMC to a process if
4146 		 * we can debug it.
4147 		 */
4148 		error = p_candebug(curthread, p);
4149 
4150 		PROC_UNLOCK(p);
4151 
4152 		if (error == 0)
4153 			error = pmc_attach_process(p, pm);
4154 	}
4155 	break;
4156 
4157 
4158 	/*
4159 	 * Detach an attached PMC from a process.
4160 	 */
4161 
4162 	case PMC_OP_PMCDETACH:
4163 	{
4164 		struct pmc *pm;
4165 		struct proc *p;
4166 		struct pmc_op_pmcattach a;
4167 
4168 		if ((error = copyin(arg, &a, sizeof(a))) != 0)
4169 			break;
4170 
4171 		if (a.pm_pid < 0) {
4172 			error = EINVAL;
4173 			break;
4174 		} else if (a.pm_pid == 0)
4175 			a.pm_pid = td->td_proc->p_pid;
4176 
4177 		if ((error = pmc_find_pmc(a.pm_pmc, &pm)) != 0)
4178 			break;
4179 
4180 		if ((p = pfind(a.pm_pid)) == NULL) {
4181 			error = ESRCH;
4182 			break;
4183 		}
4184 
4185 		/*
4186 		 * Treat processes that are in the process of exiting
4187 		 * as if they were not present.
4188 		 */
4189 
4190 		if (p->p_flag & P_WEXIT)
4191 			error = ESRCH;
4192 
4193 		PROC_UNLOCK(p);	/* pfind() returns a locked process */
4194 
4195 		if (error == 0)
4196 			error = pmc_detach_process(p, pm);
4197 	}
4198 	break;
4199 
4200 
4201 	/*
4202 	 * Retrieve the MSR number associated with the counter
4203 	 * 'pmc_id'.  This allows processes to directly use RDPMC
4204 	 * instructions to read their PMCs, without the overhead of a
4205 	 * system call.
4206 	 */
4207 
4208 	case PMC_OP_PMCGETMSR:
4209 	{
4210 		int adjri, ri;
4211 		struct pmc *pm;
4212 		struct pmc_target *pt;
4213 		struct pmc_op_getmsr gm;
4214 		struct pmc_classdep *pcd;
4215 
4216 		PMC_DOWNGRADE_SX();
4217 
4218 		if ((error = copyin(arg, &gm, sizeof(gm))) != 0)
4219 			break;
4220 
4221 		if ((error = pmc_find_pmc(gm.pm_pmcid, &pm)) != 0)
4222 			break;
4223 
4224 		/*
4225 		 * The allocated PMC has to be a process virtual PMC,
4226 		 * i.e., of type MODE_T[CS].  Global PMCs can only be
4227 		 * read using the PMCREAD operation since they may be
4228 		 * allocated on a different CPU than the one we could
4229 		 * be running on at the time of the RDPMC instruction.
4230 		 *
4231 		 * The GETMSR operation is not allowed for PMCs that
4232 		 * are inherited across processes.
4233 		 */
4234 
4235 		if (!PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)) ||
4236 		    (pm->pm_flags & PMC_F_DESCENDANTS)) {
4237 			error = EINVAL;
4238 			break;
4239 		}
4240 
4241 		/*
4242 		 * It only makes sense to use a RDPMC (or its
4243 		 * equivalent instruction on non-x86 architectures) on
4244 		 * a process that has allocated and attached a PMC to
4245 		 * itself.  Conversely the PMC is only allowed to have
4246 		 * one process attached to it -- its owner.
4247 		 */
4248 
4249 		if ((pt = LIST_FIRST(&pm->pm_targets)) == NULL ||
4250 		    LIST_NEXT(pt, pt_next) != NULL ||
4251 		    pt->pt_process->pp_proc != pm->pm_owner->po_owner) {
4252 			error = EINVAL;
4253 			break;
4254 		}
4255 
4256 		ri = PMC_TO_ROWINDEX(pm);
4257 		pcd = pmc_ri_to_classdep(md, ri, &adjri);
4258 
4259 		/* PMC class has no 'GETMSR' support */
4260 		if (pcd->pcd_get_msr == NULL) {
4261 			error = ENOSYS;
4262 			break;
4263 		}
4264 
4265 		if ((error = (*pcd->pcd_get_msr)(adjri, &gm.pm_msr)) < 0)
4266 			break;
4267 
4268 		if ((error = copyout(&gm, arg, sizeof(gm))) < 0)
4269 			break;
4270 
4271 		/*
4272 		 * Mark our process as using MSRs.  Update machine
4273 		 * state using a forced context switch.
4274 		 */
4275 
4276 		pt->pt_process->pp_flags |= PMC_PP_ENABLE_MSR_ACCESS;
4277 		pmc_force_context_switch();
4278 
4279 	}
4280 	break;
4281 
4282 	/*
4283 	 * Release an allocated PMC
4284 	 */
4285 
4286 	case PMC_OP_PMCRELEASE:
4287 	{
4288 		pmc_id_t pmcid;
4289 		struct pmc *pm;
4290 		struct pmc_owner *po;
4291 		struct pmc_op_simple sp;
4292 
4293 		/*
4294 		 * Find PMC pointer for the named PMC.
4295 		 *
4296 		 * Use pmc_release_pmc_descriptor() to switch off the
4297 		 * PMC, remove all its target threads, and remove the
4298 		 * PMC from its owner's list.
4299 		 *
4300 		 * Remove the owner record if this is the last PMC
4301 		 * owned.
4302 		 *
4303 		 * Free up space.
4304 		 */
4305 
4306 		if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
4307 			break;
4308 
4309 		pmcid = sp.pm_pmcid;
4310 
4311 		if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
4312 			break;
4313 
4314 		po = pm->pm_owner;
4315 		pmc_release_pmc_descriptor(pm);
4316 		pmc_maybe_remove_owner(po);
4317 		pmc_destroy_pmc_descriptor(pm);
4318 	}
4319 	break;
4320 
4321 
4322 	/*
4323 	 * Read and/or write a PMC.
4324 	 */
4325 
4326 	case PMC_OP_PMCRW:
4327 	{
4328 		int adjri;
4329 		struct pmc *pm;
4330 		uint32_t cpu, ri;
4331 		pmc_value_t oldvalue;
4332 		struct pmc_binding pb;
4333 		struct pmc_op_pmcrw prw;
4334 		struct pmc_classdep *pcd;
4335 		struct pmc_op_pmcrw *pprw;
4336 
4337 		PMC_DOWNGRADE_SX();
4338 
4339 		if ((error = copyin(arg, &prw, sizeof(prw))) != 0)
4340 			break;
4341 
4342 		ri = 0;
4343 		PMCDBG2(PMC,OPS,1, "rw id=%d flags=0x%x", prw.pm_pmcid,
4344 		    prw.pm_flags);
4345 
4346 		/* must have at least one flag set */
4347 		if ((prw.pm_flags & (PMC_F_OLDVALUE|PMC_F_NEWVALUE)) == 0) {
4348 			error = EINVAL;
4349 			break;
4350 		}
4351 
4352 		/* locate pmc descriptor */
4353 		if ((error = pmc_find_pmc(prw.pm_pmcid, &pm)) != 0)
4354 			break;
4355 
4356 		/* Can't read a PMC that hasn't been started. */
4357 		if (pm->pm_state != PMC_STATE_ALLOCATED &&
4358 		    pm->pm_state != PMC_STATE_STOPPED &&
4359 		    pm->pm_state != PMC_STATE_RUNNING) {
4360 			error = EINVAL;
4361 			break;
4362 		}
4363 
4364 		/* writing a new value is allowed only for 'STOPPED' pmcs */
4365 		if (pm->pm_state == PMC_STATE_RUNNING &&
4366 		    (prw.pm_flags & PMC_F_NEWVALUE)) {
4367 			error = EBUSY;
4368 			break;
4369 		}
4370 
4371 		if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) {
4372 
4373 			/*
4374 			 * If this PMC is attached to its owner (i.e.,
4375 			 * the process requesting this operation) and
4376 			 * is running, then attempt to get an
4377 			 * upto-date reading from hardware for a READ.
4378 			 * Writes are only allowed when the PMC is
4379 			 * stopped, so only update the saved value
4380 			 * field.
4381 			 *
4382 			 * If the PMC is not running, or is not
4383 			 * attached to its owner, read/write to the
4384 			 * savedvalue field.
4385 			 */
4386 
4387 			ri = PMC_TO_ROWINDEX(pm);
4388 			pcd = pmc_ri_to_classdep(md, ri, &adjri);
4389 
4390 			mtx_pool_lock_spin(pmc_mtxpool, pm);
4391 			cpu = curthread->td_oncpu;
4392 
4393 			if (prw.pm_flags & PMC_F_OLDVALUE) {
4394 				if ((pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) &&
4395 				    (pm->pm_state == PMC_STATE_RUNNING))
4396 					error = (*pcd->pcd_read_pmc)(cpu, adjri,
4397 					    &oldvalue);
4398 				else
4399 					oldvalue = pm->pm_gv.pm_savedvalue;
4400 			}
4401 			if (prw.pm_flags & PMC_F_NEWVALUE)
4402 				pm->pm_gv.pm_savedvalue = prw.pm_value;
4403 
4404 			mtx_pool_unlock_spin(pmc_mtxpool, pm);
4405 
4406 		} else { /* System mode PMCs */
4407 			cpu = PMC_TO_CPU(pm);
4408 			ri  = PMC_TO_ROWINDEX(pm);
4409 			pcd = pmc_ri_to_classdep(md, ri, &adjri);
4410 
4411 			if (!pmc_cpu_is_active(cpu)) {
4412 				error = ENXIO;
4413 				break;
4414 			}
4415 
4416 			/* move this thread to CPU 'cpu' */
4417 			pmc_save_cpu_binding(&pb);
4418 			pmc_select_cpu(cpu);
4419 
4420 			critical_enter();
4421 			/* save old value */
4422 			if (prw.pm_flags & PMC_F_OLDVALUE)
4423 				if ((error = (*pcd->pcd_read_pmc)(cpu, adjri,
4424 					 &oldvalue)))
4425 					goto error;
4426 			/* write out new value */
4427 			if (prw.pm_flags & PMC_F_NEWVALUE)
4428 				error = (*pcd->pcd_write_pmc)(cpu, adjri,
4429 				    prw.pm_value);
4430 		error:
4431 			critical_exit();
4432 			pmc_restore_cpu_binding(&pb);
4433 			if (error)
4434 				break;
4435 		}
4436 
4437 		pprw = (struct pmc_op_pmcrw *) arg;
4438 
4439 #ifdef	HWPMC_DEBUG
4440 		if (prw.pm_flags & PMC_F_NEWVALUE)
4441 			PMCDBG3(PMC,OPS,2, "rw id=%d new %jx -> old %jx",
4442 			    ri, prw.pm_value, oldvalue);
4443 		else if (prw.pm_flags & PMC_F_OLDVALUE)
4444 			PMCDBG2(PMC,OPS,2, "rw id=%d -> old %jx", ri, oldvalue);
4445 #endif
4446 
4447 		/* return old value if requested */
4448 		if (prw.pm_flags & PMC_F_OLDVALUE)
4449 			if ((error = copyout(&oldvalue, &pprw->pm_value,
4450 				 sizeof(prw.pm_value))))
4451 				break;
4452 
4453 	}
4454 	break;
4455 
4456 
4457 	/*
4458 	 * Set the sampling rate for a sampling mode PMC and the
4459 	 * initial count for a counting mode PMC.
4460 	 */
4461 
4462 	case PMC_OP_PMCSETCOUNT:
4463 	{
4464 		struct pmc *pm;
4465 		struct pmc_op_pmcsetcount sc;
4466 
4467 		PMC_DOWNGRADE_SX();
4468 
4469 		if ((error = copyin(arg, &sc, sizeof(sc))) != 0)
4470 			break;
4471 
4472 		if ((error = pmc_find_pmc(sc.pm_pmcid, &pm)) != 0)
4473 			break;
4474 
4475 		if (pm->pm_state == PMC_STATE_RUNNING) {
4476 			error = EBUSY;
4477 			break;
4478 		}
4479 
4480 		if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) {
4481 			/*
4482 			 * Don't permit requested sample rate to be less than 1000
4483 			 */
4484 			if (sc.pm_count < 1000)
4485 				log(LOG_WARNING,
4486 					"pmcsetcount: passed sample rate %ju - setting to 1000\n",
4487 					(uintmax_t)sc.pm_count);
4488 			pm->pm_sc.pm_reloadcount = MAX(1000, sc.pm_count);
4489 		} else
4490 			pm->pm_sc.pm_initial = sc.pm_count;
4491 	}
4492 	break;
4493 
4494 
4495 	/*
4496 	 * Start a PMC.
4497 	 */
4498 
4499 	case PMC_OP_PMCSTART:
4500 	{
4501 		pmc_id_t pmcid;
4502 		struct pmc *pm;
4503 		struct pmc_op_simple sp;
4504 
4505 		sx_assert(&pmc_sx, SX_XLOCKED);
4506 
4507 		if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
4508 			break;
4509 
4510 		pmcid = sp.pm_pmcid;
4511 
4512 		if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
4513 			break;
4514 
4515 		KASSERT(pmcid == pm->pm_id,
4516 		    ("[pmc,%d] pmcid %x != id %x", __LINE__,
4517 			pm->pm_id, pmcid));
4518 
4519 		if (pm->pm_state == PMC_STATE_RUNNING) /* already running */
4520 			break;
4521 		else if (pm->pm_state != PMC_STATE_STOPPED &&
4522 		    pm->pm_state != PMC_STATE_ALLOCATED) {
4523 			error = EINVAL;
4524 			break;
4525 		}
4526 
4527 		error = pmc_start(pm);
4528 	}
4529 	break;
4530 
4531 
4532 	/*
4533 	 * Stop a PMC.
4534 	 */
4535 
4536 	case PMC_OP_PMCSTOP:
4537 	{
4538 		pmc_id_t pmcid;
4539 		struct pmc *pm;
4540 		struct pmc_op_simple sp;
4541 
4542 		PMC_DOWNGRADE_SX();
4543 
4544 		if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
4545 			break;
4546 
4547 		pmcid = sp.pm_pmcid;
4548 
4549 		/*
4550 		 * Mark the PMC as inactive and invoke the MD stop
4551 		 * routines if needed.
4552 		 */
4553 
4554 		if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
4555 			break;
4556 
4557 		KASSERT(pmcid == pm->pm_id,
4558 		    ("[pmc,%d] pmc id %x != pmcid %x", __LINE__,
4559 			pm->pm_id, pmcid));
4560 
4561 		if (pm->pm_state == PMC_STATE_STOPPED) /* already stopped */
4562 			break;
4563 		else if (pm->pm_state != PMC_STATE_RUNNING) {
4564 			error = EINVAL;
4565 			break;
4566 		}
4567 
4568 		error = pmc_stop(pm);
4569 	}
4570 	break;
4571 
4572 
4573 	/*
4574 	 * Write a user supplied value to the log file.
4575 	 */
4576 
4577 	case PMC_OP_WRITELOG:
4578 	{
4579 		struct pmc_op_writelog wl;
4580 		struct pmc_owner *po;
4581 
4582 		PMC_DOWNGRADE_SX();
4583 
4584 		if ((error = copyin(arg, &wl, sizeof(wl))) != 0)
4585 			break;
4586 
4587 		if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) {
4588 			error = EINVAL;
4589 			break;
4590 		}
4591 
4592 		if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) {
4593 			error = EINVAL;
4594 			break;
4595 		}
4596 
4597 		error = pmclog_process_userlog(po, &wl);
4598 	}
4599 	break;
4600 
4601 
4602 	default:
4603 		error = EINVAL;
4604 		break;
4605 	}
4606 
4607 	if (is_sx_downgraded)
4608 		sx_sunlock(&pmc_sx);
4609 	else
4610 		sx_xunlock(&pmc_sx);
4611 done_syscall:
4612 	if (error)
4613 		counter_u64_add(pmc_stats.pm_syscall_errors, 1);
4614 
4615 	return (error);
4616 }
4617 
4618 /*
4619  * Helper functions
4620  */
4621 
4622 
4623 /*
4624  * Mark the thread as needing callchain capture and post an AST.  The
4625  * actual callchain capture will be done in a context where it is safe
4626  * to take page faults.
4627  */
4628 
4629 static void
4630 pmc_post_callchain_callback(void)
4631 {
4632 	struct thread *td;
4633 
4634 	td = curthread;
4635 
4636 	/*
4637 	 * If there is multiple PMCs for the same interrupt ignore new post
4638 	 */
4639 	if (td->td_pflags & TDP_CALLCHAIN)
4640 		return;
4641 
4642 	/*
4643 	 * Mark this thread as needing callchain capture.
4644 	 * `td->td_pflags' will be safe to touch because this thread
4645 	 * was in user space when it was interrupted.
4646 	 */
4647 	td->td_pflags |= TDP_CALLCHAIN;
4648 
4649 	/*
4650 	 * Don't let this thread migrate between CPUs until callchain
4651 	 * capture completes.
4652 	 */
4653 	sched_pin();
4654 
4655 	return;
4656 }
4657 
4658 /*
4659  * Find a free slot in the per-cpu array of samples and capture the
4660  * current callchain there.  If a sample was successfully added, a bit
4661  * is set in mask 'pmc_cpumask' denoting that the DO_SAMPLES hook
4662  * needs to be invoked from the clock handler.
4663  *
4664  * This function is meant to be called from an NMI handler.  It cannot
4665  * use any of the locking primitives supplied by the OS.
4666  */
4667 
4668 static int
4669 pmc_add_sample(ring_type_t ring, struct pmc *pm, struct trapframe *tf)
4670 {
4671 	int error, cpu, callchaindepth, inuserspace;
4672 	struct thread *td;
4673 	struct pmc_sample *ps;
4674 	struct pmc_samplebuffer *psb;
4675 
4676 	error = 0;
4677 
4678 	/*
4679 	 * Allocate space for a sample buffer.
4680 	 */
4681 	cpu = curcpu;
4682 	psb = pmc_pcpu[cpu]->pc_sb[ring];
4683 	inuserspace = TRAPF_USERMODE(tf);
4684 	ps = PMC_PROD_SAMPLE(psb);
4685 	if (psb->ps_considx != psb->ps_prodidx &&
4686 		ps->ps_nsamples) {	/* in use, reader hasn't caught up */
4687 		pm->pm_pcpu_state[cpu].pps_stalled = 1;
4688 		counter_u64_add(pmc_stats.pm_intr_bufferfull, 1);
4689 		PMCDBG6(SAM,INT,1,"(spc) cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d",
4690 		    cpu, pm, (void *) tf, inuserspace,
4691 		    (int) (psb->ps_prodidx & pmc_sample_mask),
4692 		    (int) (psb->ps_considx & pmc_sample_mask));
4693 		callchaindepth = 1;
4694 		error = ENOMEM;
4695 		goto done;
4696 	}
4697 
4698 	/* Fill in entry. */
4699 	PMCDBG6(SAM,INT,1,"cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d", cpu, pm,
4700 	    (void *) tf, inuserspace,
4701 	    (int) (psb->ps_prodidx & pmc_sample_mask),
4702 	    (int) (psb->ps_considx & pmc_sample_mask));
4703 
4704 	td = curthread;
4705 	ps->ps_pmc = pm;
4706 	ps->ps_td = td;
4707 	ps->ps_pid = td->td_proc->p_pid;
4708 	ps->ps_tid = td->td_tid;
4709 	ps->ps_tsc = pmc_rdtsc();
4710 	ps->ps_ticks = ticks;
4711 	ps->ps_cpu = cpu;
4712 	ps->ps_flags = inuserspace ? PMC_CC_F_USERSPACE : 0;
4713 
4714 	callchaindepth = (pm->pm_flags & PMC_F_CALLCHAIN) ?
4715 	    pmc_callchaindepth : 1;
4716 
4717 	MPASS(ps->ps_pc != NULL);
4718 	if (callchaindepth == 1)
4719 		ps->ps_pc[0] = PMC_TRAPFRAME_TO_PC(tf);
4720 	else {
4721 		/*
4722 		 * Kernel stack traversals can be done immediately,
4723 		 * while we defer to an AST for user space traversals.
4724 		 */
4725 		if (!inuserspace) {
4726 			callchaindepth =
4727 			    pmc_save_kernel_callchain(ps->ps_pc,
4728 				callchaindepth, tf);
4729 		} else {
4730 			pmc_post_callchain_callback();
4731 			callchaindepth = PMC_USER_CALLCHAIN_PENDING;
4732 		}
4733 	}
4734 
4735 	ps->ps_nsamples = callchaindepth;	/* mark entry as in use */
4736 	if (ring == PMC_UR) {
4737 		ps->ps_nsamples_actual = callchaindepth;	/* mark entry as in use */
4738 		ps->ps_nsamples = PMC_USER_CALLCHAIN_PENDING;
4739 	} else
4740 		ps->ps_nsamples = callchaindepth;	/* mark entry as in use */
4741 
4742 	KASSERT(counter_u64_fetch(pm->pm_runcount) >= 0,
4743 	    ("[pmc,%d] pm=%p runcount %ld", __LINE__, (void *) pm,
4744 		 (unsigned long)counter_u64_fetch(pm->pm_runcount)));
4745 
4746 	counter_u64_add(pm->pm_runcount, 1);	/* hold onto PMC */
4747 	/* increment write pointer */
4748 	psb->ps_prodidx++;
4749  done:
4750 	/* mark CPU as needing processing */
4751 	if (callchaindepth != PMC_USER_CALLCHAIN_PENDING)
4752 		DPCPU_SET(pmc_sampled, 1);
4753 
4754 	return (error);
4755 }
4756 
4757 /*
4758  * Interrupt processing.
4759  *
4760  * This function is meant to be called from an NMI handler.  It cannot
4761  * use any of the locking primitives supplied by the OS.
4762  */
4763 
4764 int
4765 pmc_process_interrupt(int ring, struct pmc *pm, struct trapframe *tf)
4766 {
4767 	struct thread *td;
4768 
4769 	td = curthread;
4770 	if ((pm->pm_flags & PMC_F_USERCALLCHAIN) &&
4771 	    (td->td_proc->p_flag & P_KPROC) == 0 &&
4772 	    !TRAPF_USERMODE(tf)) {
4773 		atomic_add_int(&td->td_pmcpend, 1);
4774 		return (pmc_add_sample(PMC_UR, pm, tf));
4775 	}
4776 	return (pmc_add_sample(ring, pm, tf));
4777 }
4778 
4779 /*
4780  * Capture a user call chain.  This function will be called from ast()
4781  * before control returns to userland and before the process gets
4782  * rescheduled.
4783  */
4784 
4785 static void
4786 pmc_capture_user_callchain(int cpu, int ring, struct trapframe *tf)
4787 {
4788 	struct pmc *pm;
4789 	struct thread *td;
4790 	struct pmc_sample *ps;
4791 	struct pmc_samplebuffer *psb;
4792 	uint64_t considx, prodidx;
4793 	int nsamples, nrecords, pass, iter;
4794 #ifdef	INVARIANTS
4795 	int ncallchains;
4796 	int nfree;
4797 	int start_ticks = ticks;
4798 #endif
4799 	psb = pmc_pcpu[cpu]->pc_sb[ring];
4800 	td = curthread;
4801 
4802 	KASSERT(td->td_pflags & TDP_CALLCHAIN,
4803 	    ("[pmc,%d] Retrieving callchain for thread that doesn't want it",
4804 		__LINE__));
4805 
4806 #ifdef	INVARIANTS
4807 	ncallchains = 0;
4808 	nfree = 0;
4809 #endif
4810 	nrecords = INT_MAX;
4811 	pass = 0;
4812  restart:
4813 	if (ring == PMC_UR)
4814 		nrecords = atomic_readandclear_32(&td->td_pmcpend);
4815 
4816 	for (iter = 0, considx = psb->ps_considx, prodidx = psb->ps_prodidx;
4817 	    considx < prodidx && iter < pmc_nsamples; considx++, iter++) {
4818 		ps = PMC_CONS_SAMPLE_OFF(psb, considx);
4819 
4820 	/*
4821 	 * Iterate through all deferred callchain requests.
4822 	 * Walk from the current read pointer to the current
4823 	 * write pointer.
4824 	 */
4825 
4826 #ifdef	INVARIANTS
4827 		if (ps->ps_nsamples == PMC_SAMPLE_FREE) {
4828 			nfree++;
4829 			continue;
4830 		}
4831 
4832 		if ((ps->ps_pmc == NULL) ||
4833 		    (ps->ps_pmc->pm_state != PMC_STATE_RUNNING))
4834 			nfree++;
4835 #endif
4836 		if (ps->ps_td != td ||
4837 		   ps->ps_nsamples == PMC_USER_CALLCHAIN_PENDING ||
4838 		   ps->ps_pmc->pm_state != PMC_STATE_RUNNING)
4839 			continue;
4840 
4841 		KASSERT(ps->ps_cpu == cpu,
4842 		    ("[pmc,%d] cpu mismatch ps_cpu=%d pcpu=%d", __LINE__,
4843 			ps->ps_cpu, PCPU_GET(cpuid)));
4844 
4845 		pm = ps->ps_pmc;
4846 
4847 		KASSERT(pm->pm_flags & PMC_F_CALLCHAIN,
4848 		    ("[pmc,%d] Retrieving callchain for PMC that doesn't "
4849 			"want it", __LINE__));
4850 
4851 		KASSERT(counter_u64_fetch(pm->pm_runcount) > 0,
4852 		    ("[pmc,%d] runcount %ld", __LINE__, (unsigned long)counter_u64_fetch(pm->pm_runcount)));
4853 
4854 		if (ring == PMC_UR) {
4855 			nsamples = ps->ps_nsamples_actual;
4856 			counter_u64_add(pmc_stats.pm_merges, 1);
4857 		} else
4858 			nsamples = 0;
4859 
4860 		/*
4861 		 * Retrieve the callchain and mark the sample buffer
4862 		 * as 'processable' by the timer tick sweep code.
4863 		 */
4864 
4865 #ifdef INVARIANTS
4866 		ncallchains++;
4867 #endif
4868 
4869 		if (__predict_true(nsamples < pmc_callchaindepth - 1))
4870 			nsamples += pmc_save_user_callchain(ps->ps_pc + nsamples,
4871 		       pmc_callchaindepth - nsamples - 1, tf);
4872 
4873 		/*
4874 		 * We have to prevent hardclock from potentially overwriting
4875 		 * this sample between when we read the value and when we set
4876 		 * it
4877 		 */
4878 		spinlock_enter();
4879 		/*
4880 		 * Verify that the sample hasn't been dropped in the meantime
4881 		 */
4882 		if (ps->ps_nsamples == PMC_USER_CALLCHAIN_PENDING) {
4883 			ps->ps_nsamples = nsamples;
4884 			/*
4885 			 * If we couldn't get a sample, simply drop the reference
4886 			 */
4887 			if (nsamples == 0)
4888 				counter_u64_add(pm->pm_runcount, -1);
4889 		}
4890 		spinlock_exit();
4891 		if (nrecords-- == 1)
4892 			break;
4893 	}
4894 	if (__predict_false(ring == PMC_UR && td->td_pmcpend)) {
4895 		if (pass == 0) {
4896 			pass = 1;
4897 			goto restart;
4898 		}
4899 		/* only collect samples for this part once */
4900 		td->td_pmcpend = 0;
4901 	}
4902 
4903 #ifdef INVARIANTS
4904 	if ((ticks - start_ticks) > hz)
4905 		log(LOG_ERR, "%s took %d ticks\n", __func__, (ticks - start_ticks));
4906 #endif
4907 
4908 	/* mark CPU as needing processing */
4909 	DPCPU_SET(pmc_sampled, 1);
4910 }
4911 
4912 /*
4913  * Process saved PC samples.
4914  */
4915 
4916 static void
4917 pmc_process_samples(int cpu, ring_type_t ring)
4918 {
4919 	struct pmc *pm;
4920 	int adjri, n;
4921 	struct thread *td;
4922 	struct pmc_owner *po;
4923 	struct pmc_sample *ps;
4924 	struct pmc_classdep *pcd;
4925 	struct pmc_samplebuffer *psb;
4926 	uint64_t delta;
4927 
4928 	KASSERT(PCPU_GET(cpuid) == cpu,
4929 	    ("[pmc,%d] not on the correct CPU pcpu=%d cpu=%d", __LINE__,
4930 		PCPU_GET(cpuid), cpu));
4931 
4932 	psb = pmc_pcpu[cpu]->pc_sb[ring];
4933 	delta = psb->ps_prodidx - psb->ps_considx;
4934 	MPASS(delta <= pmc_nsamples);
4935 	MPASS(psb->ps_considx <= psb->ps_prodidx);
4936 	for (n = 0; psb->ps_considx < psb->ps_prodidx; psb->ps_considx++, n++) {
4937 		ps = PMC_CONS_SAMPLE(psb);
4938 
4939 		if (__predict_false(ps->ps_nsamples == PMC_SAMPLE_FREE))
4940 			continue;
4941 		pm = ps->ps_pmc;
4942 		/* skip non-running samples */
4943 		if (pm->pm_state != PMC_STATE_RUNNING)
4944 			goto entrydone;
4945 
4946 		KASSERT(counter_u64_fetch(pm->pm_runcount) > 0,
4947 		    ("[pmc,%d] pm=%p runcount %ld", __LINE__, (void *) pm,
4948 			 (unsigned long)counter_u64_fetch(pm->pm_runcount)));
4949 
4950 		po = pm->pm_owner;
4951 
4952 		KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)),
4953 		    ("[pmc,%d] pmc=%p non-sampling mode=%d", __LINE__,
4954 			pm, PMC_TO_MODE(pm)));
4955 
4956 
4957 		/* If there is a pending AST wait for completion */
4958 		if (ps->ps_nsamples == PMC_USER_CALLCHAIN_PENDING) {
4959 			/* if sample is more than 65 ms old, drop it */
4960 			if (ticks - ps->ps_ticks > (hz >> 4)) {
4961 				/*
4962 				 * track how often we hit this as it will
4963 				 * preferentially lose user samples
4964 				 * for long running system calls
4965 				 */
4966 				counter_u64_add(pmc_stats.pm_overwrites, 1);
4967 				goto entrydone;
4968 			}
4969 			/* Need a rescan at a later time. */
4970 			DPCPU_SET(pmc_sampled, 1);
4971 			break;
4972 		}
4973 
4974 		PMCDBG6(SAM,OPS,1,"cpu=%d pm=%p n=%d fl=%x wr=%d rd=%d", cpu,
4975 		    pm, ps->ps_nsamples, ps->ps_flags,
4976 		    (int) (psb->ps_prodidx & pmc_sample_mask),
4977 		    (int) (psb->ps_considx & pmc_sample_mask));
4978 
4979 		/*
4980 		 * If this is a process-mode PMC that is attached to
4981 		 * its owner, and if the PC is in user mode, update
4982 		 * profiling statistics like timer-based profiling
4983 		 * would have done.
4984 		 *
4985 		 * Otherwise, this is either a sampling-mode PMC that
4986 		 * is attached to a different process than its owner,
4987 		 * or a system-wide sampling PMC. Dispatch a log
4988 		 * entry to the PMC's owner process.
4989 		 */
4990 		if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) {
4991 			if (ps->ps_flags & PMC_CC_F_USERSPACE) {
4992 				td = FIRST_THREAD_IN_PROC(po->po_owner);
4993 				addupc_intr(td, ps->ps_pc[0], 1);
4994 			}
4995 		} else
4996 			pmclog_process_callchain(pm, ps);
4997 
4998 	entrydone:
4999 		ps->ps_nsamples = 0; /* mark entry as free */
5000 		KASSERT(counter_u64_fetch(pm->pm_runcount) > 0,
5001 				("[pmc,%d] pm=%p runcount %ld", __LINE__, (void *) pm,
5002 				 (unsigned long)counter_u64_fetch(pm->pm_runcount)));
5003 
5004 		counter_u64_add(pm->pm_runcount, -1);
5005 	}
5006 
5007 	counter_u64_add(pmc_stats.pm_log_sweeps, 1);
5008 
5009 	/* Do not re-enable stalled PMCs if we failed to process any samples */
5010 	if (n == 0)
5011 		return;
5012 
5013 	/*
5014 	 * Restart any stalled sampling PMCs on this CPU.
5015 	 *
5016 	 * If the NMI handler sets the pm_stalled field of a PMC after
5017 	 * the check below, we'll end up processing the stalled PMC at
5018 	 * the next hardclock tick.
5019 	 */
5020 	for (n = 0; n < md->pmd_npmc; n++) {
5021 		pcd = pmc_ri_to_classdep(md, n, &adjri);
5022 		KASSERT(pcd != NULL,
5023 		    ("[pmc,%d] null pcd ri=%d", __LINE__, n));
5024 		(void) (*pcd->pcd_get_config)(cpu,adjri,&pm);
5025 
5026 		if (pm == NULL ||			 /* !cfg'ed */
5027 		    pm->pm_state != PMC_STATE_RUNNING || /* !active */
5028 		    !PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) || /* !sampling */
5029 			!pm->pm_pcpu_state[cpu].pps_cpustate  || /* !desired */
5030 		    !pm->pm_pcpu_state[cpu].pps_stalled) /* !stalled */
5031 			continue;
5032 
5033 		pm->pm_pcpu_state[cpu].pps_stalled = 0;
5034 		(*pcd->pcd_start_pmc)(cpu, adjri);
5035 	}
5036 }
5037 
5038 /*
5039  * Event handlers.
5040  */
5041 
5042 /*
5043  * Handle a process exit.
5044  *
5045  * Remove this process from all hash tables.  If this process
5046  * owned any PMCs, turn off those PMCs and deallocate them,
5047  * removing any associations with target processes.
5048  *
5049  * This function will be called by the last 'thread' of a
5050  * process.
5051  *
5052  * XXX This eventhandler gets called early in the exit process.
5053  * Consider using a 'hook' invocation from thread_exit() or equivalent
5054  * spot.  Another negative is that kse_exit doesn't seem to call
5055  * exit1() [??].
5056  *
5057  */
5058 
5059 static void
5060 pmc_process_exit(void *arg __unused, struct proc *p)
5061 {
5062 	struct pmc *pm;
5063 	int adjri, cpu;
5064 	unsigned int ri;
5065 	int is_using_hwpmcs;
5066 	struct pmc_owner *po;
5067 	struct pmc_process *pp;
5068 	struct pmc_classdep *pcd;
5069 	pmc_value_t newvalue, tmp;
5070 
5071 	PROC_LOCK(p);
5072 	is_using_hwpmcs = p->p_flag & P_HWPMC;
5073 	PROC_UNLOCK(p);
5074 
5075 	/*
5076 	 * Log a sysexit event to all SS PMC owners.
5077 	 */
5078 	PMC_EPOCH_ENTER();
5079 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
5080 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
5081 		    pmclog_process_sysexit(po, p->p_pid);
5082 	PMC_EPOCH_EXIT();
5083 
5084 	if (!is_using_hwpmcs)
5085 		return;
5086 
5087 	PMC_GET_SX_XLOCK();
5088 	PMCDBG3(PRC,EXT,1,"process-exit proc=%p (%d, %s)", p, p->p_pid,
5089 	    p->p_comm);
5090 
5091 	/*
5092 	 * Since this code is invoked by the last thread in an exiting
5093 	 * process, we would have context switched IN at some prior
5094 	 * point.  However, with PREEMPTION, kernel mode context
5095 	 * switches may happen any time, so we want to disable a
5096 	 * context switch OUT till we get any PMCs targeting this
5097 	 * process off the hardware.
5098 	 *
5099 	 * We also need to atomically remove this process'
5100 	 * entry from our target process hash table, using
5101 	 * PMC_FLAG_REMOVE.
5102 	 */
5103 	PMCDBG3(PRC,EXT,1, "process-exit proc=%p (%d, %s)", p, p->p_pid,
5104 	    p->p_comm);
5105 
5106 	critical_enter(); /* no preemption */
5107 
5108 	cpu = curthread->td_oncpu;
5109 
5110 	if ((pp = pmc_find_process_descriptor(p,
5111 		 PMC_FLAG_REMOVE)) != NULL) {
5112 
5113 		PMCDBG2(PRC,EXT,2,
5114 		    "process-exit proc=%p pmc-process=%p", p, pp);
5115 
5116 		/*
5117 		 * The exiting process could the target of
5118 		 * some PMCs which will be running on
5119 		 * currently executing CPU.
5120 		 *
5121 		 * We need to turn these PMCs off like we
5122 		 * would do at context switch OUT time.
5123 		 */
5124 		for (ri = 0; ri < md->pmd_npmc; ri++) {
5125 
5126 			/*
5127 			 * Pick up the pmc pointer from hardware
5128 			 * state similar to the CSW_OUT code.
5129 			 */
5130 			pm = NULL;
5131 
5132 			pcd = pmc_ri_to_classdep(md, ri, &adjri);
5133 
5134 			(void) (*pcd->pcd_get_config)(cpu, adjri, &pm);
5135 
5136 			PMCDBG2(PRC,EXT,2, "ri=%d pm=%p", ri, pm);
5137 
5138 			if (pm == NULL ||
5139 			    !PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)))
5140 				continue;
5141 
5142 			PMCDBG4(PRC,EXT,2, "ppmcs[%d]=%p pm=%p "
5143 			    "state=%d", ri, pp->pp_pmcs[ri].pp_pmc,
5144 			    pm, pm->pm_state);
5145 
5146 			KASSERT(PMC_TO_ROWINDEX(pm) == ri,
5147 			    ("[pmc,%d] ri mismatch pmc(%d) ri(%d)",
5148 				__LINE__, PMC_TO_ROWINDEX(pm), ri));
5149 
5150 			KASSERT(pm == pp->pp_pmcs[ri].pp_pmc,
5151 			    ("[pmc,%d] pm %p != pp_pmcs[%d] %p",
5152 				__LINE__, pm, ri, pp->pp_pmcs[ri].pp_pmc));
5153 
5154 			KASSERT(counter_u64_fetch(pm->pm_runcount) > 0,
5155 			    ("[pmc,%d] bad runcount ri %d rc %ld",
5156 				 __LINE__, ri, (unsigned long)counter_u64_fetch(pm->pm_runcount)));
5157 
5158 			/*
5159 			 * Change desired state, and then stop if not
5160 			 * stalled. This two-step dance should avoid
5161 			 * race conditions where an interrupt re-enables
5162 			 * the PMC after this code has already checked
5163 			 * the pm_stalled flag.
5164 			 */
5165 			if (pm->pm_pcpu_state[cpu].pps_cpustate) {
5166 				pm->pm_pcpu_state[cpu].pps_cpustate = 0;
5167 				if (!pm->pm_pcpu_state[cpu].pps_stalled) {
5168 					(void) pcd->pcd_stop_pmc(cpu, adjri);
5169 
5170 					if (PMC_TO_MODE(pm) == PMC_MODE_TC) {
5171 						pcd->pcd_read_pmc(cpu, adjri,
5172 						    &newvalue);
5173 						tmp = newvalue -
5174 						    PMC_PCPU_SAVED(cpu,ri);
5175 
5176 						mtx_pool_lock_spin(pmc_mtxpool,
5177 						    pm);
5178 						pm->pm_gv.pm_savedvalue += tmp;
5179 						pp->pp_pmcs[ri].pp_pmcval +=
5180 						    tmp;
5181 						mtx_pool_unlock_spin(
5182 						    pmc_mtxpool, pm);
5183 					}
5184 				}
5185 			}
5186 
5187 			KASSERT((int64_t) counter_u64_fetch(pm->pm_runcount) > 0,
5188 			    ("[pmc,%d] runcount is %d", __LINE__, ri));
5189 
5190 			counter_u64_add(pm->pm_runcount, -1);
5191 
5192 			(void) pcd->pcd_config_pmc(cpu, adjri, NULL);
5193 		}
5194 
5195 		/*
5196 		 * Inform the MD layer of this pseudo "context switch
5197 		 * out"
5198 		 */
5199 		(void) md->pmd_switch_out(pmc_pcpu[cpu], pp);
5200 
5201 		critical_exit(); /* ok to be pre-empted now */
5202 
5203 		/*
5204 		 * Unlink this process from the PMCs that are
5205 		 * targeting it.  This will send a signal to
5206 		 * all PMC owner's whose PMCs are orphaned.
5207 		 *
5208 		 * Log PMC value at exit time if requested.
5209 		 */
5210 		for (ri = 0; ri < md->pmd_npmc; ri++)
5211 			if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) {
5212 				if (pm->pm_flags & PMC_F_NEEDS_LOGFILE &&
5213 				    PMC_IS_COUNTING_MODE(PMC_TO_MODE(pm)))
5214 					pmclog_process_procexit(pm, pp);
5215 				pmc_unlink_target_process(pm, pp);
5216 			}
5217 		free(pp, M_PMC);
5218 
5219 	} else
5220 		critical_exit(); /* pp == NULL */
5221 
5222 
5223 	/*
5224 	 * If the process owned PMCs, free them up and free up
5225 	 * memory.
5226 	 */
5227 	if ((po = pmc_find_owner_descriptor(p)) != NULL) {
5228 		pmc_remove_owner(po);
5229 		pmc_destroy_owner_descriptor(po);
5230 	}
5231 
5232 	sx_xunlock(&pmc_sx);
5233 }
5234 
5235 /*
5236  * Handle a process fork.
5237  *
5238  * If the parent process 'p1' is under HWPMC monitoring, then copy
5239  * over any attached PMCs that have 'do_descendants' semantics.
5240  */
5241 
5242 static void
5243 pmc_process_fork(void *arg __unused, struct proc *p1, struct proc *newproc,
5244     int flags)
5245 {
5246 	int is_using_hwpmcs;
5247 	unsigned int ri;
5248 	uint32_t do_descendants;
5249 	struct pmc *pm;
5250 	struct pmc_owner *po;
5251 	struct pmc_process *ppnew, *ppold;
5252 
5253 	(void) flags;		/* unused parameter */
5254 
5255 	PROC_LOCK(p1);
5256 	is_using_hwpmcs = p1->p_flag & P_HWPMC;
5257 	PROC_UNLOCK(p1);
5258 
5259 	/*
5260 	 * If there are system-wide sampling PMCs active, we need to
5261 	 * log all fork events to their owner's logs.
5262 	 */
5263 	PMC_EPOCH_ENTER();
5264 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
5265 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE) {
5266 		    pmclog_process_procfork(po, p1->p_pid, newproc->p_pid);
5267 			pmclog_process_proccreate(po, newproc, 1);
5268 		}
5269 	PMC_EPOCH_EXIT();
5270 
5271 	if (!is_using_hwpmcs)
5272 		return;
5273 
5274 	PMC_GET_SX_XLOCK();
5275 	PMCDBG4(PMC,FRK,1, "process-fork proc=%p (%d, %s) -> %p", p1,
5276 	    p1->p_pid, p1->p_comm, newproc);
5277 
5278 	/*
5279 	 * If the parent process (curthread->td_proc) is a
5280 	 * target of any PMCs, look for PMCs that are to be
5281 	 * inherited, and link these into the new process
5282 	 * descriptor.
5283 	 */
5284 	if ((ppold = pmc_find_process_descriptor(curthread->td_proc,
5285 		 PMC_FLAG_NONE)) == NULL)
5286 		goto done;		/* nothing to do */
5287 
5288 	do_descendants = 0;
5289 	for (ri = 0; ri < md->pmd_npmc; ri++)
5290 		if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL)
5291 			do_descendants |= pm->pm_flags & PMC_F_DESCENDANTS;
5292 	if (do_descendants == 0) /* nothing to do */
5293 		goto done;
5294 
5295 	/*
5296 	 * Now mark the new process as being tracked by this driver.
5297 	 */
5298 	PROC_LOCK(newproc);
5299 	newproc->p_flag |= P_HWPMC;
5300 	PROC_UNLOCK(newproc);
5301 
5302 	/* allocate a descriptor for the new process  */
5303 	if ((ppnew = pmc_find_process_descriptor(newproc,
5304 		 PMC_FLAG_ALLOCATE)) == NULL)
5305 		goto done;
5306 
5307 	/*
5308 	 * Run through all PMCs that were targeting the old process
5309 	 * and which specified F_DESCENDANTS and attach them to the
5310 	 * new process.
5311 	 *
5312 	 * Log the fork event to all owners of PMCs attached to this
5313 	 * process, if not already logged.
5314 	 */
5315 	for (ri = 0; ri < md->pmd_npmc; ri++)
5316 		if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL &&
5317 		    (pm->pm_flags & PMC_F_DESCENDANTS)) {
5318 			pmc_link_target_process(pm, ppnew);
5319 			po = pm->pm_owner;
5320 			if (po->po_sscount == 0 &&
5321 			    po->po_flags & PMC_PO_OWNS_LOGFILE)
5322 				pmclog_process_procfork(po, p1->p_pid,
5323 				    newproc->p_pid);
5324 		}
5325 
5326  done:
5327 	sx_xunlock(&pmc_sx);
5328 }
5329 
5330 static void
5331 pmc_process_threadcreate(struct thread *td)
5332 {
5333 	struct pmc_owner *po;
5334 
5335 	PMC_EPOCH_ENTER();
5336 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
5337 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
5338 			pmclog_process_threadcreate(po, td, 1);
5339 	PMC_EPOCH_EXIT();
5340 }
5341 
5342 static void
5343 pmc_process_threadexit(struct thread *td)
5344 {
5345 	struct pmc_owner *po;
5346 
5347 	PMC_EPOCH_ENTER();
5348 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
5349 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
5350 			pmclog_process_threadexit(po, td);
5351 	PMC_EPOCH_EXIT();
5352 }
5353 
5354 static void
5355 pmc_process_proccreate(struct proc *p)
5356 {
5357 	struct pmc_owner *po;
5358 
5359 	PMC_EPOCH_ENTER();
5360 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
5361 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
5362 			pmclog_process_proccreate(po, p, 1 /* sync */);
5363 	PMC_EPOCH_EXIT();
5364 }
5365 
5366 static void
5367 pmc_process_allproc(struct pmc *pm)
5368 {
5369 	struct pmc_owner *po;
5370 	struct thread *td;
5371 	struct proc *p;
5372 
5373 	po = pm->pm_owner;
5374 	if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
5375 		return;
5376 	sx_slock(&allproc_lock);
5377 	FOREACH_PROC_IN_SYSTEM(p) {
5378 		pmclog_process_proccreate(po, p, 0 /* sync */);
5379 		PROC_LOCK(p);
5380 		FOREACH_THREAD_IN_PROC(p, td)
5381 			pmclog_process_threadcreate(po, td, 0 /* sync */);
5382 		PROC_UNLOCK(p);
5383 	}
5384 	sx_sunlock(&allproc_lock);
5385 	pmclog_flush(po, 0);
5386 }
5387 
5388 static void
5389 pmc_kld_load(void *arg __unused, linker_file_t lf)
5390 {
5391 	struct pmc_owner *po;
5392 
5393 	/*
5394 	 * Notify owners of system sampling PMCs about KLD operations.
5395 	 */
5396 	PMC_EPOCH_ENTER();
5397 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
5398 		if (po->po_flags & PMC_PO_OWNS_LOGFILE)
5399 			pmclog_process_map_in(po, (pid_t) -1,
5400 			    (uintfptr_t) lf->address, lf->filename);
5401 	PMC_EPOCH_EXIT();
5402 
5403 	/*
5404 	 * TODO: Notify owners of (all) process-sampling PMCs too.
5405 	 */
5406 }
5407 
5408 static void
5409 pmc_kld_unload(void *arg __unused, const char *filename __unused,
5410     caddr_t address, size_t size)
5411 {
5412 	struct pmc_owner *po;
5413 
5414 	PMC_EPOCH_ENTER();
5415 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
5416 		if (po->po_flags & PMC_PO_OWNS_LOGFILE)
5417 			pmclog_process_map_out(po, (pid_t) -1,
5418 			    (uintfptr_t) address, (uintfptr_t) address + size);
5419 	PMC_EPOCH_EXIT();
5420 
5421 	/*
5422 	 * TODO: Notify owners of process-sampling PMCs.
5423 	 */
5424 }
5425 
5426 /*
5427  * initialization
5428  */
5429 static const char *
5430 pmc_name_of_pmcclass(enum pmc_class class)
5431 {
5432 
5433 	switch (class) {
5434 #undef	__PMC_CLASS
5435 #define	__PMC_CLASS(S,V,D)						\
5436 	case PMC_CLASS_##S:						\
5437 		return #S;
5438 	__PMC_CLASSES();
5439 	default:
5440 		return ("<unknown>");
5441 	}
5442 }
5443 
5444 /*
5445  * Base class initializer: allocate structure and set default classes.
5446  */
5447 struct pmc_mdep *
5448 pmc_mdep_alloc(int nclasses)
5449 {
5450 	struct pmc_mdep *md;
5451 	int	n;
5452 
5453 	/* SOFT + md classes */
5454 	n = 1 + nclasses;
5455 	md = malloc(sizeof(struct pmc_mdep) + n *
5456 	    sizeof(struct pmc_classdep), M_PMC, M_WAITOK|M_ZERO);
5457 	md->pmd_nclass = n;
5458 
5459 	/* Add base class. */
5460 	pmc_soft_initialize(md);
5461 	return md;
5462 }
5463 
5464 void
5465 pmc_mdep_free(struct pmc_mdep *md)
5466 {
5467 	pmc_soft_finalize(md);
5468 	free(md, M_PMC);
5469 }
5470 
5471 static int
5472 generic_switch_in(struct pmc_cpu *pc, struct pmc_process *pp)
5473 {
5474 	(void) pc; (void) pp;
5475 
5476 	return (0);
5477 }
5478 
5479 static int
5480 generic_switch_out(struct pmc_cpu *pc, struct pmc_process *pp)
5481 {
5482 	(void) pc; (void) pp;
5483 
5484 	return (0);
5485 }
5486 
5487 static struct pmc_mdep *
5488 pmc_generic_cpu_initialize(void)
5489 {
5490 	struct pmc_mdep *md;
5491 
5492 	md = pmc_mdep_alloc(0);
5493 
5494 	md->pmd_cputype    = PMC_CPU_GENERIC;
5495 
5496 	md->pmd_pcpu_init  = NULL;
5497 	md->pmd_pcpu_fini  = NULL;
5498 	md->pmd_switch_in  = generic_switch_in;
5499 	md->pmd_switch_out = generic_switch_out;
5500 
5501 	return (md);
5502 }
5503 
5504 static void
5505 pmc_generic_cpu_finalize(struct pmc_mdep *md)
5506 {
5507 	(void) md;
5508 }
5509 
5510 
5511 static int
5512 pmc_initialize(void)
5513 {
5514 	int c, cpu, error, n, ri;
5515 	unsigned int maxcpu, domain;
5516 	struct pcpu *pc;
5517 	struct pmc_binding pb;
5518 	struct pmc_sample *ps;
5519 	struct pmc_classdep *pcd;
5520 	struct pmc_samplebuffer *sb;
5521 
5522 	md = NULL;
5523 	error = 0;
5524 
5525 	pmc_stats.pm_intr_ignored = counter_u64_alloc(M_WAITOK);
5526 	pmc_stats.pm_intr_processed = counter_u64_alloc(M_WAITOK);
5527 	pmc_stats.pm_intr_bufferfull = counter_u64_alloc(M_WAITOK);
5528 	pmc_stats.pm_syscalls = counter_u64_alloc(M_WAITOK);
5529 	pmc_stats.pm_syscall_errors = counter_u64_alloc(M_WAITOK);
5530 	pmc_stats.pm_buffer_requests = counter_u64_alloc(M_WAITOK);
5531 	pmc_stats.pm_buffer_requests_failed = counter_u64_alloc(M_WAITOK);
5532 	pmc_stats.pm_log_sweeps = counter_u64_alloc(M_WAITOK);
5533 	pmc_stats.pm_merges = counter_u64_alloc(M_WAITOK);
5534 	pmc_stats.pm_overwrites = counter_u64_alloc(M_WAITOK);
5535 
5536 #ifdef	HWPMC_DEBUG
5537 	/* parse debug flags first */
5538 	if (TUNABLE_STR_FETCH(PMC_SYSCTL_NAME_PREFIX "debugflags",
5539 		pmc_debugstr, sizeof(pmc_debugstr)))
5540 		pmc_debugflags_parse(pmc_debugstr,
5541 		    pmc_debugstr+strlen(pmc_debugstr));
5542 #endif
5543 
5544 	PMCDBG1(MOD,INI,0, "PMC Initialize (version %x)", PMC_VERSION);
5545 
5546 	/* check kernel version */
5547 	if (pmc_kernel_version != PMC_VERSION) {
5548 		if (pmc_kernel_version == 0)
5549 			printf("hwpmc: this kernel has not been compiled with "
5550 			    "'options HWPMC_HOOKS'.\n");
5551 		else
5552 			printf("hwpmc: kernel version (0x%x) does not match "
5553 			    "module version (0x%x).\n", pmc_kernel_version,
5554 			    PMC_VERSION);
5555 		return EPROGMISMATCH;
5556 	}
5557 
5558 	/*
5559 	 * check sysctl parameters
5560 	 */
5561 
5562 	if (pmc_hashsize <= 0) {
5563 		(void) printf("hwpmc: tunable \"hashsize\"=%d must be "
5564 		    "greater than zero.\n", pmc_hashsize);
5565 		pmc_hashsize = PMC_HASH_SIZE;
5566 	}
5567 
5568 	if (pmc_nsamples <= 0 || pmc_nsamples > 65535) {
5569 		(void) printf("hwpmc: tunable \"nsamples\"=%d out of "
5570 		    "range.\n", pmc_nsamples);
5571 		pmc_nsamples = PMC_NSAMPLES;
5572 	}
5573 	pmc_sample_mask = pmc_nsamples-1;
5574 
5575 	if (pmc_callchaindepth <= 0 ||
5576 	    pmc_callchaindepth > PMC_CALLCHAIN_DEPTH_MAX) {
5577 		(void) printf("hwpmc: tunable \"callchaindepth\"=%d out of "
5578 		    "range - using %d.\n", pmc_callchaindepth,
5579 		    PMC_CALLCHAIN_DEPTH_MAX);
5580 		pmc_callchaindepth = PMC_CALLCHAIN_DEPTH_MAX;
5581 	}
5582 
5583 	md = pmc_md_initialize();
5584 	if (md == NULL) {
5585 		/* Default to generic CPU. */
5586 		md = pmc_generic_cpu_initialize();
5587 		if (md == NULL)
5588 			return (ENOSYS);
5589         }
5590 
5591 	KASSERT(md->pmd_nclass >= 1 && md->pmd_npmc >= 1,
5592 	    ("[pmc,%d] no classes or pmcs", __LINE__));
5593 
5594 	/* Compute the map from row-indices to classdep pointers. */
5595 	pmc_rowindex_to_classdep = malloc(sizeof(struct pmc_classdep *) *
5596 	    md->pmd_npmc, M_PMC, M_WAITOK|M_ZERO);
5597 
5598 	for (n = 0; n < md->pmd_npmc; n++)
5599 		pmc_rowindex_to_classdep[n] = NULL;
5600 	for (ri = c = 0; c < md->pmd_nclass; c++) {
5601 		pcd = &md->pmd_classdep[c];
5602 		for (n = 0; n < pcd->pcd_num; n++, ri++)
5603 			pmc_rowindex_to_classdep[ri] = pcd;
5604 	}
5605 
5606 	KASSERT(ri == md->pmd_npmc,
5607 	    ("[pmc,%d] npmc miscomputed: ri=%d, md->npmc=%d", __LINE__,
5608 	    ri, md->pmd_npmc));
5609 
5610 	maxcpu = pmc_cpu_max();
5611 
5612 	/* allocate space for the per-cpu array */
5613 	pmc_pcpu = malloc(maxcpu * sizeof(struct pmc_cpu *), M_PMC,
5614 	    M_WAITOK|M_ZERO);
5615 
5616 	/* per-cpu 'saved values' for managing process-mode PMCs */
5617 	pmc_pcpu_saved = malloc(sizeof(pmc_value_t) * maxcpu * md->pmd_npmc,
5618 	    M_PMC, M_WAITOK);
5619 
5620 	/* Perform CPU-dependent initialization. */
5621 	pmc_save_cpu_binding(&pb);
5622 	error = 0;
5623 	for (cpu = 0; error == 0 && cpu < maxcpu; cpu++) {
5624 		if (!pmc_cpu_is_active(cpu))
5625 			continue;
5626 		pmc_select_cpu(cpu);
5627 		pmc_pcpu[cpu] = malloc(sizeof(struct pmc_cpu) +
5628 		    md->pmd_npmc * sizeof(struct pmc_hw *), M_PMC,
5629 		    M_WAITOK|M_ZERO);
5630 		if (md->pmd_pcpu_init)
5631 			error = md->pmd_pcpu_init(md, cpu);
5632 		for (n = 0; error == 0 && n < md->pmd_nclass; n++)
5633 			error = md->pmd_classdep[n].pcd_pcpu_init(md, cpu);
5634 	}
5635 	pmc_restore_cpu_binding(&pb);
5636 
5637 	if (error)
5638 		return (error);
5639 
5640 	/* allocate space for the sample array */
5641 	for (cpu = 0; cpu < maxcpu; cpu++) {
5642 		if (!pmc_cpu_is_active(cpu))
5643 			continue;
5644 		pc = pcpu_find(cpu);
5645 		domain = pc->pc_domain;
5646 		sb = malloc_domain(sizeof(struct pmc_samplebuffer) +
5647 			pmc_nsamples * sizeof(struct pmc_sample), M_PMC, domain,
5648 		    M_WAITOK|M_ZERO);
5649 
5650 		KASSERT(pmc_pcpu[cpu] != NULL,
5651 		    ("[pmc,%d] cpu=%d Null per-cpu data", __LINE__, cpu));
5652 
5653 		sb->ps_callchains = malloc_domain(pmc_callchaindepth * pmc_nsamples *
5654 			sizeof(uintptr_t), M_PMC, domain, M_WAITOK|M_ZERO);
5655 
5656 		for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++)
5657 			ps->ps_pc = sb->ps_callchains +
5658 			    (n * pmc_callchaindepth);
5659 
5660 		pmc_pcpu[cpu]->pc_sb[PMC_HR] = sb;
5661 
5662 		sb = malloc_domain(sizeof(struct pmc_samplebuffer) +
5663 			pmc_nsamples * sizeof(struct pmc_sample), M_PMC, domain,
5664 		    M_WAITOK|M_ZERO);
5665 
5666 		KASSERT(pmc_pcpu[cpu] != NULL,
5667 		    ("[pmc,%d] cpu=%d Null per-cpu data", __LINE__, cpu));
5668 
5669 		sb->ps_callchains = malloc_domain(pmc_callchaindepth * pmc_nsamples *
5670 			sizeof(uintptr_t), M_PMC, domain, M_WAITOK|M_ZERO);
5671 
5672 		for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++)
5673 			ps->ps_pc = sb->ps_callchains +
5674 			    (n * pmc_callchaindepth);
5675 
5676 		pmc_pcpu[cpu]->pc_sb[PMC_SR] = sb;
5677 
5678 		sb = malloc_domain(sizeof(struct pmc_samplebuffer) +
5679 			pmc_nsamples * sizeof(struct pmc_sample), M_PMC, domain,
5680 		    M_WAITOK|M_ZERO);
5681 
5682 		KASSERT(pmc_pcpu[cpu] != NULL,
5683 		    ("[pmc,%d] cpu=%d Null per-cpu data", __LINE__, cpu));
5684 
5685 		sb->ps_callchains = malloc_domain(pmc_callchaindepth * pmc_nsamples *
5686 		    sizeof(uintptr_t), M_PMC, domain, M_WAITOK|M_ZERO);
5687 
5688 		for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++)
5689 			ps->ps_pc = sb->ps_callchains +
5690 			    (n * pmc_callchaindepth);
5691 
5692 		pmc_pcpu[cpu]->pc_sb[PMC_UR] = sb;
5693 	}
5694 
5695 	/* allocate space for the row disposition array */
5696 	pmc_pmcdisp = malloc(sizeof(enum pmc_mode) * md->pmd_npmc,
5697 	    M_PMC, M_WAITOK|M_ZERO);
5698 
5699 	/* mark all PMCs as available */
5700 	for (n = 0; n < (int) md->pmd_npmc; n++)
5701 		PMC_MARK_ROW_FREE(n);
5702 
5703 	/* allocate thread hash tables */
5704 	pmc_ownerhash = hashinit(pmc_hashsize, M_PMC,
5705 	    &pmc_ownerhashmask);
5706 
5707 	pmc_processhash = hashinit(pmc_hashsize, M_PMC,
5708 	    &pmc_processhashmask);
5709 	mtx_init(&pmc_processhash_mtx, "pmc-process-hash", "pmc-leaf",
5710 	    MTX_SPIN);
5711 
5712 	CK_LIST_INIT(&pmc_ss_owners);
5713 	pmc_ss_count = 0;
5714 
5715 	/* allocate a pool of spin mutexes */
5716 	pmc_mtxpool = mtx_pool_create("pmc-leaf", pmc_mtxpool_size,
5717 	    MTX_SPIN);
5718 
5719 	PMCDBG4(MOD,INI,1, "pmc_ownerhash=%p, mask=0x%lx "
5720 	    "targethash=%p mask=0x%lx", pmc_ownerhash, pmc_ownerhashmask,
5721 	    pmc_processhash, pmc_processhashmask);
5722 
5723 	/* Initialize a spin mutex for the thread free list. */
5724 	mtx_init(&pmc_threadfreelist_mtx, "pmc-threadfreelist", "pmc-leaf",
5725 	    MTX_SPIN);
5726 
5727 	/*
5728 	 * Initialize the callout to monitor the thread free list.
5729 	 * This callout will also handle the initial population of the list.
5730 	 */
5731 	taskqgroup_config_gtask_init(NULL, &free_gtask, pmc_thread_descriptor_pool_free_task, "thread descriptor pool free task");
5732 
5733 	/* register process {exit,fork,exec} handlers */
5734 	pmc_exit_tag = EVENTHANDLER_REGISTER(process_exit,
5735 	    pmc_process_exit, NULL, EVENTHANDLER_PRI_ANY);
5736 	pmc_fork_tag = EVENTHANDLER_REGISTER(process_fork,
5737 	    pmc_process_fork, NULL, EVENTHANDLER_PRI_ANY);
5738 
5739 	/* register kld event handlers */
5740 	pmc_kld_load_tag = EVENTHANDLER_REGISTER(kld_load, pmc_kld_load,
5741 	    NULL, EVENTHANDLER_PRI_ANY);
5742 	pmc_kld_unload_tag = EVENTHANDLER_REGISTER(kld_unload, pmc_kld_unload,
5743 	    NULL, EVENTHANDLER_PRI_ANY);
5744 
5745 	/* initialize logging */
5746 	pmclog_initialize();
5747 
5748 	/* set hook functions */
5749 	pmc_intr = md->pmd_intr;
5750 	wmb();
5751 	pmc_hook = pmc_hook_handler;
5752 
5753 	if (error == 0) {
5754 		printf(PMC_MODULE_NAME ":");
5755 		for (n = 0; n < (int) md->pmd_nclass; n++) {
5756 			pcd = &md->pmd_classdep[n];
5757 			printf(" %s/%d/%d/0x%b",
5758 			    pmc_name_of_pmcclass(pcd->pcd_class),
5759 			    pcd->pcd_num,
5760 			    pcd->pcd_width,
5761 			    pcd->pcd_caps,
5762 			    "\20"
5763 			    "\1INT\2USR\3SYS\4EDG\5THR"
5764 			    "\6REA\7WRI\10INV\11QUA\12PRC"
5765 			    "\13TAG\14CSC");
5766 		}
5767 		printf("\n");
5768 	}
5769 
5770 	return (error);
5771 }
5772 
5773 /* prepare to be unloaded */
5774 static void
5775 pmc_cleanup(void)
5776 {
5777 	int c, cpu;
5778 	unsigned int maxcpu;
5779 	struct pmc_ownerhash *ph;
5780 	struct pmc_owner *po, *tmp;
5781 	struct pmc_binding pb;
5782 #ifdef	HWPMC_DEBUG
5783 	struct pmc_processhash *prh;
5784 #endif
5785 
5786 	PMCDBG0(MOD,INI,0, "cleanup");
5787 
5788 	/* switch off sampling */
5789 	CPU_FOREACH(cpu)
5790 		DPCPU_ID_SET(cpu, pmc_sampled, 0);
5791 	pmc_intr = NULL;
5792 
5793 	sx_xlock(&pmc_sx);
5794 	if (pmc_hook == NULL) {	/* being unloaded already */
5795 		sx_xunlock(&pmc_sx);
5796 		return;
5797 	}
5798 
5799 	pmc_hook = NULL; /* prevent new threads from entering module */
5800 
5801 	/* deregister event handlers */
5802 	EVENTHANDLER_DEREGISTER(process_fork, pmc_fork_tag);
5803 	EVENTHANDLER_DEREGISTER(process_exit, pmc_exit_tag);
5804 	EVENTHANDLER_DEREGISTER(kld_load, pmc_kld_load_tag);
5805 	EVENTHANDLER_DEREGISTER(kld_unload, pmc_kld_unload_tag);
5806 
5807 	/* send SIGBUS to all owner threads, free up allocations */
5808 	if (pmc_ownerhash)
5809 		for (ph = pmc_ownerhash;
5810 		     ph <= &pmc_ownerhash[pmc_ownerhashmask];
5811 		     ph++) {
5812 			LIST_FOREACH_SAFE(po, ph, po_next, tmp) {
5813 				pmc_remove_owner(po);
5814 
5815 				/* send SIGBUS to owner processes */
5816 				PMCDBG3(MOD,INI,2, "cleanup signal proc=%p "
5817 				    "(%d, %s)", po->po_owner,
5818 				    po->po_owner->p_pid,
5819 				    po->po_owner->p_comm);
5820 
5821 				PROC_LOCK(po->po_owner);
5822 				kern_psignal(po->po_owner, SIGBUS);
5823 				PROC_UNLOCK(po->po_owner);
5824 
5825 				pmc_destroy_owner_descriptor(po);
5826 			}
5827 		}
5828 
5829 	/* reclaim allocated data structures */
5830 	mtx_destroy(&pmc_threadfreelist_mtx);
5831 	pmc_thread_descriptor_pool_drain();
5832 
5833 	if (pmc_mtxpool)
5834 		mtx_pool_destroy(&pmc_mtxpool);
5835 
5836 	mtx_destroy(&pmc_processhash_mtx);
5837 	taskqgroup_config_gtask_deinit(&free_gtask);
5838 	if (pmc_processhash) {
5839 #ifdef	HWPMC_DEBUG
5840 		struct pmc_process *pp;
5841 
5842 		PMCDBG0(MOD,INI,3, "destroy process hash");
5843 		for (prh = pmc_processhash;
5844 		     prh <= &pmc_processhash[pmc_processhashmask];
5845 		     prh++)
5846 			LIST_FOREACH(pp, prh, pp_next)
5847 			    PMCDBG1(MOD,INI,3, "pid=%d", pp->pp_proc->p_pid);
5848 #endif
5849 
5850 		hashdestroy(pmc_processhash, M_PMC, pmc_processhashmask);
5851 		pmc_processhash = NULL;
5852 	}
5853 
5854 	if (pmc_ownerhash) {
5855 		PMCDBG0(MOD,INI,3, "destroy owner hash");
5856 		hashdestroy(pmc_ownerhash, M_PMC, pmc_ownerhashmask);
5857 		pmc_ownerhash = NULL;
5858 	}
5859 
5860 	KASSERT(CK_LIST_EMPTY(&pmc_ss_owners),
5861 	    ("[pmc,%d] Global SS owner list not empty", __LINE__));
5862 	KASSERT(pmc_ss_count == 0,
5863 	    ("[pmc,%d] Global SS count not empty", __LINE__));
5864 
5865  	/* do processor and pmc-class dependent cleanup */
5866 	maxcpu = pmc_cpu_max();
5867 
5868 	PMCDBG0(MOD,INI,3, "md cleanup");
5869 	if (md) {
5870 		pmc_save_cpu_binding(&pb);
5871 		for (cpu = 0; cpu < maxcpu; cpu++) {
5872 			PMCDBG2(MOD,INI,1,"pmc-cleanup cpu=%d pcs=%p",
5873 			    cpu, pmc_pcpu[cpu]);
5874 			if (!pmc_cpu_is_active(cpu) || pmc_pcpu[cpu] == NULL)
5875 				continue;
5876 			pmc_select_cpu(cpu);
5877 			for (c = 0; c < md->pmd_nclass; c++)
5878 				md->pmd_classdep[c].pcd_pcpu_fini(md, cpu);
5879 			if (md->pmd_pcpu_fini)
5880 				md->pmd_pcpu_fini(md, cpu);
5881 		}
5882 
5883 		if (md->pmd_cputype == PMC_CPU_GENERIC)
5884 			pmc_generic_cpu_finalize(md);
5885 		else
5886 			pmc_md_finalize(md);
5887 
5888 		pmc_mdep_free(md);
5889 		md = NULL;
5890 		pmc_restore_cpu_binding(&pb);
5891 	}
5892 
5893 	/* Free per-cpu descriptors. */
5894 	for (cpu = 0; cpu < maxcpu; cpu++) {
5895 		if (!pmc_cpu_is_active(cpu))
5896 			continue;
5897 		KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_HR] != NULL,
5898 		    ("[pmc,%d] Null hw cpu sample buffer cpu=%d", __LINE__,
5899 			cpu));
5900 		KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_SR] != NULL,
5901 		    ("[pmc,%d] Null sw cpu sample buffer cpu=%d", __LINE__,
5902 			cpu));
5903 		KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_UR] != NULL,
5904 		    ("[pmc,%d] Null userret cpu sample buffer cpu=%d", __LINE__,
5905 			cpu));
5906 		free_domain(pmc_pcpu[cpu]->pc_sb[PMC_HR]->ps_callchains, M_PMC);
5907 		free_domain(pmc_pcpu[cpu]->pc_sb[PMC_HR], M_PMC);
5908 		free_domain(pmc_pcpu[cpu]->pc_sb[PMC_SR]->ps_callchains, M_PMC);
5909 		free_domain(pmc_pcpu[cpu]->pc_sb[PMC_SR], M_PMC);
5910 		free_domain(pmc_pcpu[cpu]->pc_sb[PMC_UR]->ps_callchains, M_PMC);
5911 		free_domain(pmc_pcpu[cpu]->pc_sb[PMC_UR], M_PMC);
5912 		free_domain(pmc_pcpu[cpu], M_PMC);
5913 	}
5914 
5915 	free(pmc_pcpu, M_PMC);
5916 	pmc_pcpu = NULL;
5917 
5918 	free(pmc_pcpu_saved, M_PMC);
5919 	pmc_pcpu_saved = NULL;
5920 
5921 	if (pmc_pmcdisp) {
5922 		free(pmc_pmcdisp, M_PMC);
5923 		pmc_pmcdisp = NULL;
5924 	}
5925 
5926 	if (pmc_rowindex_to_classdep) {
5927 		free(pmc_rowindex_to_classdep, M_PMC);
5928 		pmc_rowindex_to_classdep = NULL;
5929 	}
5930 
5931 	pmclog_shutdown();
5932 	counter_u64_free(pmc_stats.pm_intr_ignored);
5933 	counter_u64_free(pmc_stats.pm_intr_processed);
5934 	counter_u64_free(pmc_stats.pm_intr_bufferfull);
5935 	counter_u64_free(pmc_stats.pm_syscalls);
5936 	counter_u64_free(pmc_stats.pm_syscall_errors);
5937 	counter_u64_free(pmc_stats.pm_buffer_requests);
5938 	counter_u64_free(pmc_stats.pm_buffer_requests_failed);
5939 	counter_u64_free(pmc_stats.pm_log_sweeps);
5940 	counter_u64_free(pmc_stats.pm_merges);
5941 	counter_u64_free(pmc_stats.pm_overwrites);
5942 	sx_xunlock(&pmc_sx); 	/* we are done */
5943 }
5944 
5945 /*
5946  * The function called at load/unload.
5947  */
5948 
5949 static int
5950 load (struct module *module __unused, int cmd, void *arg __unused)
5951 {
5952 	int error;
5953 
5954 	error = 0;
5955 
5956 	switch (cmd) {
5957 	case MOD_LOAD :
5958 		/* initialize the subsystem */
5959 		error = pmc_initialize();
5960 		if (error != 0)
5961 			break;
5962 		PMCDBG2(MOD,INI,1, "syscall=%d maxcpu=%d",
5963 		    pmc_syscall_num, pmc_cpu_max());
5964 		break;
5965 
5966 
5967 	case MOD_UNLOAD :
5968 	case MOD_SHUTDOWN:
5969 		pmc_cleanup();
5970 		PMCDBG0(MOD,INI,1, "unloaded");
5971 		break;
5972 
5973 	default :
5974 		error = EINVAL;	/* XXX should panic(9) */
5975 		break;
5976 	}
5977 
5978 	return error;
5979 }
5980