xref: /freebsd/sys/dev/hwpmc/hwpmc_mod.c (revision d3d381b2b194b4d24853e92eecef55f262688d1a)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2003-2008 Joseph Koshy
5  * Copyright (c) 2007 The FreeBSD Foundation
6  * Copyright (c) 2018 Matthew Macy
7  * All rights reserved.
8  *
9  * Portions of this software were developed by A. Joseph Koshy under
10  * sponsorship from the FreeBSD Foundation and Google, Inc.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  *
33  */
34 
35 #include <sys/cdefs.h>
36 __FBSDID("$FreeBSD$");
37 
38 #include <sys/param.h>
39 #include <sys/eventhandler.h>
40 #include <sys/gtaskqueue.h>
41 #include <sys/jail.h>
42 #include <sys/kernel.h>
43 #include <sys/kthread.h>
44 #include <sys/limits.h>
45 #include <sys/lock.h>
46 #include <sys/malloc.h>
47 #include <sys/module.h>
48 #include <sys/mount.h>
49 #include <sys/mutex.h>
50 #include <sys/pmc.h>
51 #include <sys/pmckern.h>
52 #include <sys/pmclog.h>
53 #include <sys/priv.h>
54 #include <sys/proc.h>
55 #include <sys/queue.h>
56 #include <sys/resourcevar.h>
57 #include <sys/rwlock.h>
58 #include <sys/sched.h>
59 #include <sys/signalvar.h>
60 #include <sys/smp.h>
61 #include <sys/sx.h>
62 #include <sys/sysctl.h>
63 #include <sys/sysent.h>
64 #include <sys/systm.h>
65 #include <sys/vnode.h>
66 
67 #include <sys/linker.h>		/* needs to be after <sys/malloc.h> */
68 
69 #include <machine/atomic.h>
70 #include <machine/md_var.h>
71 
72 #include <vm/vm.h>
73 #include <vm/vm_extern.h>
74 #include <vm/pmap.h>
75 #include <vm/vm_map.h>
76 #include <vm/vm_object.h>
77 
78 #include "hwpmc_soft.h"
79 
80 #ifdef NUMA
81 #define NDOMAINS vm_ndomains
82 #else
83 #define NDOMAINS 1
84 #define malloc_domain(size, type, domain, flags) malloc((size), (type), (flags))
85 #define free_domain(addr, type) free(addr, type)
86 #endif
87 
88 #define PMC_EPOCH_ENTER() struct epoch_tracker pmc_et; epoch_enter_preempt(global_epoch_preempt, &pmc_et)
89 #define PMC_EPOCH_EXIT() epoch_exit_preempt(global_epoch_preempt, &pmc_et)
90 
91 /*
92  * Types
93  */
94 
95 enum pmc_flags {
96 	PMC_FLAG_NONE	  = 0x00, /* do nothing */
97 	PMC_FLAG_REMOVE   = 0x01, /* atomically remove entry from hash */
98 	PMC_FLAG_ALLOCATE = 0x02, /* add entry to hash if not found */
99 	PMC_FLAG_NOWAIT   = 0x04, /* do not wait for mallocs */
100 };
101 
102 /*
103  * The offset in sysent where the syscall is allocated.
104  */
105 
106 static int pmc_syscall_num = NO_SYSCALL;
107 struct pmc_cpu		**pmc_pcpu;	 /* per-cpu state */
108 pmc_value_t		*pmc_pcpu_saved; /* saved PMC values: CSW handling */
109 
110 #define	PMC_PCPU_SAVED(C,R)	pmc_pcpu_saved[(R) + md->pmd_npmc*(C)]
111 
112 struct mtx_pool		*pmc_mtxpool;
113 static int		*pmc_pmcdisp;	 /* PMC row dispositions */
114 
115 #define	PMC_ROW_DISP_IS_FREE(R)		(pmc_pmcdisp[(R)] == 0)
116 #define	PMC_ROW_DISP_IS_THREAD(R)	(pmc_pmcdisp[(R)] > 0)
117 #define	PMC_ROW_DISP_IS_STANDALONE(R)	(pmc_pmcdisp[(R)] < 0)
118 
119 #define	PMC_MARK_ROW_FREE(R) do {					  \
120 	pmc_pmcdisp[(R)] = 0;						  \
121 } while (0)
122 
123 #define	PMC_MARK_ROW_STANDALONE(R) do {					  \
124 	KASSERT(pmc_pmcdisp[(R)] <= 0, ("[pmc,%d] row disposition error", \
125 		    __LINE__));						  \
126 	atomic_add_int(&pmc_pmcdisp[(R)], -1);				  \
127 	KASSERT(pmc_pmcdisp[(R)] >= (-pmc_cpu_max_active()),		  \
128 		("[pmc,%d] row disposition error", __LINE__));		  \
129 } while (0)
130 
131 #define	PMC_UNMARK_ROW_STANDALONE(R) do { 				  \
132 	atomic_add_int(&pmc_pmcdisp[(R)], 1);				  \
133 	KASSERT(pmc_pmcdisp[(R)] <= 0, ("[pmc,%d] row disposition error", \
134 		    __LINE__));						  \
135 } while (0)
136 
137 #define	PMC_MARK_ROW_THREAD(R) do {					  \
138 	KASSERT(pmc_pmcdisp[(R)] >= 0, ("[pmc,%d] row disposition error", \
139 		    __LINE__));						  \
140 	atomic_add_int(&pmc_pmcdisp[(R)], 1);				  \
141 } while (0)
142 
143 #define	PMC_UNMARK_ROW_THREAD(R) do {					  \
144 	atomic_add_int(&pmc_pmcdisp[(R)], -1);				  \
145 	KASSERT(pmc_pmcdisp[(R)] >= 0, ("[pmc,%d] row disposition error", \
146 		    __LINE__));						  \
147 } while (0)
148 
149 
150 /* various event handlers */
151 static eventhandler_tag	pmc_exit_tag, pmc_fork_tag, pmc_kld_load_tag,
152     pmc_kld_unload_tag;
153 
154 /* Module statistics */
155 struct pmc_driverstats pmc_stats;
156 
157 
158 /* Machine/processor dependent operations */
159 static struct pmc_mdep  *md;
160 
161 /*
162  * Hash tables mapping owner processes and target threads to PMCs.
163  */
164 
165 struct mtx pmc_processhash_mtx;		/* spin mutex */
166 static u_long pmc_processhashmask;
167 static LIST_HEAD(pmc_processhash, pmc_process)	*pmc_processhash;
168 
169 /*
170  * Hash table of PMC owner descriptors.  This table is protected by
171  * the shared PMC "sx" lock.
172  */
173 
174 static u_long pmc_ownerhashmask;
175 static LIST_HEAD(pmc_ownerhash, pmc_owner)	*pmc_ownerhash;
176 
177 /*
178  * List of PMC owners with system-wide sampling PMCs.
179  */
180 
181 static CK_LIST_HEAD(, pmc_owner)			pmc_ss_owners;
182 
183 /*
184  * List of free thread entries. This is protected by the spin
185  * mutex.
186  */
187 static struct mtx pmc_threadfreelist_mtx;	/* spin mutex */
188 static LIST_HEAD(, pmc_thread)			pmc_threadfreelist;
189 static int pmc_threadfreelist_entries=0;
190 #define	THREADENTRY_SIZE						\
191 (sizeof(struct pmc_thread) + (md->pmd_npmc * sizeof(struct pmc_threadpmcstate)))
192 
193 /*
194  * Task to free thread descriptors
195  */
196 static struct grouptask free_gtask;
197 
198 /*
199  * A map of row indices to classdep structures.
200  */
201 static struct pmc_classdep **pmc_rowindex_to_classdep;
202 
203 /*
204  * Prototypes
205  */
206 
207 #ifdef	HWPMC_DEBUG
208 static int	pmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS);
209 static int	pmc_debugflags_parse(char *newstr, char *fence);
210 #endif
211 
212 static int	load(struct module *module, int cmd, void *arg);
213 static int	pmc_add_sample(int ring, struct pmc *pm, struct trapframe *tf);
214 static void	pmc_add_thread_descriptors_from_proc(struct proc *p,
215     struct pmc_process *pp);
216 static int	pmc_attach_process(struct proc *p, struct pmc *pm);
217 static struct pmc *pmc_allocate_pmc_descriptor(void);
218 static struct pmc_owner *pmc_allocate_owner_descriptor(struct proc *p);
219 static int	pmc_attach_one_process(struct proc *p, struct pmc *pm);
220 static int	pmc_can_allocate_rowindex(struct proc *p, unsigned int ri,
221     int cpu);
222 static int	pmc_can_attach(struct pmc *pm, struct proc *p);
223 static void	pmc_capture_user_callchain(int cpu, int soft, struct trapframe *tf);
224 static void	pmc_cleanup(void);
225 static int	pmc_detach_process(struct proc *p, struct pmc *pm);
226 static int	pmc_detach_one_process(struct proc *p, struct pmc *pm,
227     int flags);
228 static void	pmc_destroy_owner_descriptor(struct pmc_owner *po);
229 static void	pmc_destroy_pmc_descriptor(struct pmc *pm);
230 static void	pmc_destroy_process_descriptor(struct pmc_process *pp);
231 static struct pmc_owner *pmc_find_owner_descriptor(struct proc *p);
232 static int	pmc_find_pmc(pmc_id_t pmcid, struct pmc **pm);
233 static struct pmc *pmc_find_pmc_descriptor_in_process(struct pmc_owner *po,
234     pmc_id_t pmc);
235 static struct pmc_process *pmc_find_process_descriptor(struct proc *p,
236     uint32_t mode);
237 static struct pmc_thread *pmc_find_thread_descriptor(struct pmc_process *pp,
238     struct thread *td, uint32_t mode);
239 static void	pmc_force_context_switch(void);
240 static void	pmc_link_target_process(struct pmc *pm,
241     struct pmc_process *pp);
242 static void	pmc_log_all_process_mappings(struct pmc_owner *po);
243 static void	pmc_log_kernel_mappings(struct pmc *pm);
244 static void	pmc_log_process_mappings(struct pmc_owner *po, struct proc *p);
245 static void	pmc_maybe_remove_owner(struct pmc_owner *po);
246 static void	pmc_process_csw_in(struct thread *td);
247 static void	pmc_process_csw_out(struct thread *td);
248 static void	pmc_process_exit(void *arg, struct proc *p);
249 static void	pmc_process_fork(void *arg, struct proc *p1,
250     struct proc *p2, int n);
251 static void	pmc_process_samples(int cpu, int soft);
252 static void	pmc_release_pmc_descriptor(struct pmc *pmc);
253 static void	pmc_process_thread_add(struct thread *td);
254 static void	pmc_process_thread_delete(struct thread *td);
255 static void	pmc_process_thread_userret(struct thread *td);
256 static void	pmc_remove_owner(struct pmc_owner *po);
257 static void	pmc_remove_process_descriptor(struct pmc_process *pp);
258 static void	pmc_restore_cpu_binding(struct pmc_binding *pb);
259 static void	pmc_save_cpu_binding(struct pmc_binding *pb);
260 static void	pmc_select_cpu(int cpu);
261 static int	pmc_start(struct pmc *pm);
262 static int	pmc_stop(struct pmc *pm);
263 static int	pmc_syscall_handler(struct thread *td, void *syscall_args);
264 static struct pmc_thread *pmc_thread_descriptor_pool_alloc(void);
265 static void	pmc_thread_descriptor_pool_drain(void);
266 static void	pmc_thread_descriptor_pool_free(struct pmc_thread *pt);
267 static void	pmc_unlink_target_process(struct pmc *pmc,
268     struct pmc_process *pp);
269 static int generic_switch_in(struct pmc_cpu *pc, struct pmc_process *pp);
270 static int generic_switch_out(struct pmc_cpu *pc, struct pmc_process *pp);
271 static struct pmc_mdep *pmc_generic_cpu_initialize(void);
272 static void pmc_generic_cpu_finalize(struct pmc_mdep *md);
273 static void pmc_post_callchain_callback(void);
274 static void pmc_process_threadcreate(struct thread *td);
275 static void pmc_process_threadexit(struct thread *td);
276 static void pmc_process_proccreate(struct proc *p);
277 static void pmc_process_allproc(struct pmc *pm);
278 
279 /*
280  * Kernel tunables and sysctl(8) interface.
281  */
282 
283 SYSCTL_DECL(_kern_hwpmc);
284 SYSCTL_NODE(_kern_hwpmc, OID_AUTO, stats, CTLFLAG_RW, 0, "HWPMC stats");
285 
286 
287 /* Stats. */
288 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, intr_ignored, CTLFLAG_RW,
289 				   &pmc_stats.pm_intr_ignored, "# of interrupts ignored");
290 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, intr_processed, CTLFLAG_RW,
291 				   &pmc_stats.pm_intr_processed, "# of interrupts processed");
292 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, intr_bufferfull, CTLFLAG_RW,
293 				   &pmc_stats.pm_intr_bufferfull, "# of interrupts where buffer was full");
294 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, syscalls, CTLFLAG_RW,
295 				   &pmc_stats.pm_syscalls, "# of syscalls");
296 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, syscall_errors, CTLFLAG_RW,
297 				   &pmc_stats.pm_syscall_errors, "# of syscall_errors");
298 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, buffer_requests, CTLFLAG_RW,
299 				   &pmc_stats.pm_buffer_requests, "# of buffer requests");
300 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, buffer_requests_failed, CTLFLAG_RW,
301 				   &pmc_stats.pm_buffer_requests_failed, "# of buffer requests which failed");
302 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, log_sweeps, CTLFLAG_RW,
303 				   &pmc_stats.pm_log_sweeps, "# of ?");
304 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, merges, CTLFLAG_RW,
305 				   &pmc_stats.pm_merges, "# of times kernel stack was found for user trace");
306 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, overwrites, CTLFLAG_RW,
307 				   &pmc_stats.pm_overwrites, "# of times a sample was overwritten before being logged");
308 
309 static int pmc_callchaindepth = PMC_CALLCHAIN_DEPTH;
310 SYSCTL_INT(_kern_hwpmc, OID_AUTO, callchaindepth, CTLFLAG_RDTUN,
311     &pmc_callchaindepth, 0, "depth of call chain records");
312 
313 char pmc_cpuid[64];
314 SYSCTL_STRING(_kern_hwpmc, OID_AUTO, cpuid, CTLFLAG_RD,
315 	pmc_cpuid, 0, "cpu version string");
316 #ifdef	HWPMC_DEBUG
317 struct pmc_debugflags pmc_debugflags = PMC_DEBUG_DEFAULT_FLAGS;
318 char	pmc_debugstr[PMC_DEBUG_STRSIZE];
319 TUNABLE_STR(PMC_SYSCTL_NAME_PREFIX "debugflags", pmc_debugstr,
320     sizeof(pmc_debugstr));
321 SYSCTL_PROC(_kern_hwpmc, OID_AUTO, debugflags,
322     CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NOFETCH,
323     0, 0, pmc_debugflags_sysctl_handler, "A", "debug flags");
324 #endif
325 
326 
327 /*
328  * kern.hwpmc.hashrows -- determines the number of rows in the
329  * of the hash table used to look up threads
330  */
331 
332 static int pmc_hashsize = PMC_HASH_SIZE;
333 SYSCTL_INT(_kern_hwpmc, OID_AUTO, hashsize, CTLFLAG_RDTUN,
334     &pmc_hashsize, 0, "rows in hash tables");
335 
336 /*
337  * kern.hwpmc.nsamples --- number of PC samples/callchain stacks per CPU
338  */
339 
340 static int pmc_nsamples = PMC_NSAMPLES;
341 SYSCTL_INT(_kern_hwpmc, OID_AUTO, nsamples, CTLFLAG_RDTUN,
342     &pmc_nsamples, 0, "number of PC samples per CPU");
343 
344 
345 /*
346  * kern.hwpmc.mtxpoolsize -- number of mutexes in the mutex pool.
347  */
348 
349 static int pmc_mtxpool_size = PMC_MTXPOOL_SIZE;
350 SYSCTL_INT(_kern_hwpmc, OID_AUTO, mtxpoolsize, CTLFLAG_RDTUN,
351     &pmc_mtxpool_size, 0, "size of spin mutex pool");
352 
353 
354 /*
355  * kern.hwpmc.threadfreelist_entries -- number of free entries
356  */
357 
358 SYSCTL_INT(_kern_hwpmc, OID_AUTO, threadfreelist_entries, CTLFLAG_RD,
359     &pmc_threadfreelist_entries, 0, "number of avalable thread entries");
360 
361 
362 /*
363  * kern.hwpmc.threadfreelist_max -- maximum number of free entries
364  */
365 
366 static int pmc_threadfreelist_max = PMC_THREADLIST_MAX;
367 SYSCTL_INT(_kern_hwpmc, OID_AUTO, threadfreelist_max, CTLFLAG_RW,
368     &pmc_threadfreelist_max, 0,
369     "maximum number of available thread entries before freeing some");
370 
371 
372 /*
373  * security.bsd.unprivileged_syspmcs -- allow non-root processes to
374  * allocate system-wide PMCs.
375  *
376  * Allowing unprivileged processes to allocate system PMCs is convenient
377  * if system-wide measurements need to be taken concurrently with other
378  * per-process measurements.  This feature is turned off by default.
379  */
380 
381 static int pmc_unprivileged_syspmcs = 0;
382 SYSCTL_INT(_security_bsd, OID_AUTO, unprivileged_syspmcs, CTLFLAG_RWTUN,
383     &pmc_unprivileged_syspmcs, 0,
384     "allow unprivileged process to allocate system PMCs");
385 
386 /*
387  * Hash function.  Discard the lower 2 bits of the pointer since
388  * these are always zero for our uses.  The hash multiplier is
389  * round((2^LONG_BIT) * ((sqrt(5)-1)/2)).
390  */
391 
392 #if	LONG_BIT == 64
393 #define	_PMC_HM		11400714819323198486u
394 #elif	LONG_BIT == 32
395 #define	_PMC_HM		2654435769u
396 #else
397 #error 	Must know the size of 'long' to compile
398 #endif
399 
400 #define	PMC_HASH_PTR(P,M)	((((unsigned long) (P) >> 2) * _PMC_HM) & (M))
401 
402 /*
403  * Syscall structures
404  */
405 
406 /* The `sysent' for the new syscall */
407 static struct sysent pmc_sysent = {
408 	.sy_narg =	2,
409 	.sy_call =	pmc_syscall_handler,
410 };
411 
412 static struct syscall_module_data pmc_syscall_mod = {
413 	.chainevh =	load,
414 	.chainarg =	NULL,
415 	.offset =	&pmc_syscall_num,
416 	.new_sysent =	&pmc_sysent,
417 	.old_sysent =	{ .sy_narg = 0, .sy_call = NULL },
418 	.flags =	SY_THR_STATIC_KLD,
419 };
420 
421 static moduledata_t pmc_mod = {
422 	.name =		PMC_MODULE_NAME,
423 	.evhand =	syscall_module_handler,
424 	.priv =		&pmc_syscall_mod,
425 };
426 
427 #ifdef EARLY_AP_STARTUP
428 DECLARE_MODULE(pmc, pmc_mod, SI_SUB_SYSCALLS, SI_ORDER_ANY);
429 #else
430 DECLARE_MODULE(pmc, pmc_mod, SI_SUB_SMP, SI_ORDER_ANY);
431 #endif
432 MODULE_VERSION(pmc, PMC_VERSION);
433 
434 #ifdef	HWPMC_DEBUG
435 enum pmc_dbgparse_state {
436 	PMCDS_WS,		/* in whitespace */
437 	PMCDS_MAJOR,		/* seen a major keyword */
438 	PMCDS_MINOR
439 };
440 
441 static int
442 pmc_debugflags_parse(char *newstr, char *fence)
443 {
444 	char c, *p, *q;
445 	struct pmc_debugflags *tmpflags;
446 	int error, found, *newbits, tmp;
447 	size_t kwlen;
448 
449 	tmpflags = malloc(sizeof(*tmpflags), M_PMC, M_WAITOK|M_ZERO);
450 
451 	p = newstr;
452 	error = 0;
453 
454 	for (; p < fence && (c = *p); p++) {
455 
456 		/* skip white space */
457 		if (c == ' ' || c == '\t')
458 			continue;
459 
460 		/* look for a keyword followed by "=" */
461 		for (q = p; p < fence && (c = *p) && c != '='; p++)
462 			;
463 		if (c != '=') {
464 			error = EINVAL;
465 			goto done;
466 		}
467 
468 		kwlen = p - q;
469 		newbits = NULL;
470 
471 		/* lookup flag group name */
472 #define	DBG_SET_FLAG_MAJ(S,F)						\
473 		if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0)	\
474 			newbits = &tmpflags->pdb_ ## F;
475 
476 		DBG_SET_FLAG_MAJ("cpu",		CPU);
477 		DBG_SET_FLAG_MAJ("csw",		CSW);
478 		DBG_SET_FLAG_MAJ("logging",	LOG);
479 		DBG_SET_FLAG_MAJ("module",	MOD);
480 		DBG_SET_FLAG_MAJ("md", 		MDP);
481 		DBG_SET_FLAG_MAJ("owner",	OWN);
482 		DBG_SET_FLAG_MAJ("pmc",		PMC);
483 		DBG_SET_FLAG_MAJ("process",	PRC);
484 		DBG_SET_FLAG_MAJ("sampling", 	SAM);
485 
486 		if (newbits == NULL) {
487 			error = EINVAL;
488 			goto done;
489 		}
490 
491 		p++;		/* skip the '=' */
492 
493 		/* Now parse the individual flags */
494 		tmp = 0;
495 	newflag:
496 		for (q = p; p < fence && (c = *p); p++)
497 			if (c == ' ' || c == '\t' || c == ',')
498 				break;
499 
500 		/* p == fence or c == ws or c == "," or c == 0 */
501 
502 		if ((kwlen = p - q) == 0) {
503 			*newbits = tmp;
504 			continue;
505 		}
506 
507 		found = 0;
508 #define	DBG_SET_FLAG_MIN(S,F)						\
509 		if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0)	\
510 			tmp |= found = (1 << PMC_DEBUG_MIN_ ## F)
511 
512 		/* a '*' denotes all possible flags in the group */
513 		if (kwlen == 1 && *q == '*')
514 			tmp = found = ~0;
515 		/* look for individual flag names */
516 		DBG_SET_FLAG_MIN("allocaterow", ALR);
517 		DBG_SET_FLAG_MIN("allocate",	ALL);
518 		DBG_SET_FLAG_MIN("attach",	ATT);
519 		DBG_SET_FLAG_MIN("bind",	BND);
520 		DBG_SET_FLAG_MIN("config",	CFG);
521 		DBG_SET_FLAG_MIN("exec",	EXC);
522 		DBG_SET_FLAG_MIN("exit",	EXT);
523 		DBG_SET_FLAG_MIN("find",	FND);
524 		DBG_SET_FLAG_MIN("flush",	FLS);
525 		DBG_SET_FLAG_MIN("fork",	FRK);
526 		DBG_SET_FLAG_MIN("getbuf",	GTB);
527 		DBG_SET_FLAG_MIN("hook",	PMH);
528 		DBG_SET_FLAG_MIN("init",	INI);
529 		DBG_SET_FLAG_MIN("intr",	INT);
530 		DBG_SET_FLAG_MIN("linktarget",	TLK);
531 		DBG_SET_FLAG_MIN("mayberemove", OMR);
532 		DBG_SET_FLAG_MIN("ops",		OPS);
533 		DBG_SET_FLAG_MIN("read",	REA);
534 		DBG_SET_FLAG_MIN("register",	REG);
535 		DBG_SET_FLAG_MIN("release",	REL);
536 		DBG_SET_FLAG_MIN("remove",	ORM);
537 		DBG_SET_FLAG_MIN("sample",	SAM);
538 		DBG_SET_FLAG_MIN("scheduleio",	SIO);
539 		DBG_SET_FLAG_MIN("select",	SEL);
540 		DBG_SET_FLAG_MIN("signal",	SIG);
541 		DBG_SET_FLAG_MIN("swi",		SWI);
542 		DBG_SET_FLAG_MIN("swo",		SWO);
543 		DBG_SET_FLAG_MIN("start",	STA);
544 		DBG_SET_FLAG_MIN("stop",	STO);
545 		DBG_SET_FLAG_MIN("syscall",	PMS);
546 		DBG_SET_FLAG_MIN("unlinktarget", TUL);
547 		DBG_SET_FLAG_MIN("write",	WRI);
548 		if (found == 0) {
549 			/* unrecognized flag name */
550 			error = EINVAL;
551 			goto done;
552 		}
553 
554 		if (c == 0 || c == ' ' || c == '\t') {	/* end of flag group */
555 			*newbits = tmp;
556 			continue;
557 		}
558 
559 		p++;
560 		goto newflag;
561 	}
562 
563 	/* save the new flag set */
564 	bcopy(tmpflags, &pmc_debugflags, sizeof(pmc_debugflags));
565 
566  done:
567 	free(tmpflags, M_PMC);
568 	return error;
569 }
570 
571 static int
572 pmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS)
573 {
574 	char *fence, *newstr;
575 	int error;
576 	unsigned int n;
577 
578 	(void) arg1; (void) arg2; /* unused parameters */
579 
580 	n = sizeof(pmc_debugstr);
581 	newstr = malloc(n, M_PMC, M_WAITOK|M_ZERO);
582 	(void) strlcpy(newstr, pmc_debugstr, n);
583 
584 	error = sysctl_handle_string(oidp, newstr, n, req);
585 
586 	/* if there is a new string, parse and copy it */
587 	if (error == 0 && req->newptr != NULL) {
588 		fence = newstr + (n < req->newlen ? n : req->newlen + 1);
589 		if ((error = pmc_debugflags_parse(newstr, fence)) == 0)
590 			(void) strlcpy(pmc_debugstr, newstr,
591 			    sizeof(pmc_debugstr));
592 	}
593 
594 	free(newstr, M_PMC);
595 
596 	return error;
597 }
598 #endif
599 
600 /*
601  * Map a row index to a classdep structure and return the adjusted row
602  * index for the PMC class index.
603  */
604 static struct pmc_classdep *
605 pmc_ri_to_classdep(struct pmc_mdep *md, int ri, int *adjri)
606 {
607 	struct pmc_classdep *pcd;
608 
609 	(void) md;
610 
611 	KASSERT(ri >= 0 && ri < md->pmd_npmc,
612 	    ("[pmc,%d] illegal row-index %d", __LINE__, ri));
613 
614 	pcd = pmc_rowindex_to_classdep[ri];
615 
616 	KASSERT(pcd != NULL,
617 	    ("[pmc,%d] ri %d null pcd", __LINE__, ri));
618 
619 	*adjri = ri - pcd->pcd_ri;
620 
621 	KASSERT(*adjri >= 0 && *adjri < pcd->pcd_num,
622 	    ("[pmc,%d] adjusted row-index %d", __LINE__, *adjri));
623 
624 	return (pcd);
625 }
626 
627 /*
628  * Concurrency Control
629  *
630  * The driver manages the following data structures:
631  *
632  *   - target process descriptors, one per target process
633  *   - owner process descriptors (and attached lists), one per owner process
634  *   - lookup hash tables for owner and target processes
635  *   - PMC descriptors (and attached lists)
636  *   - per-cpu hardware state
637  *   - the 'hook' variable through which the kernel calls into
638  *     this module
639  *   - the machine hardware state (managed by the MD layer)
640  *
641  * These data structures are accessed from:
642  *
643  * - thread context-switch code
644  * - interrupt handlers (possibly on multiple cpus)
645  * - kernel threads on multiple cpus running on behalf of user
646  *   processes doing system calls
647  * - this driver's private kernel threads
648  *
649  * = Locks and Locking strategy =
650  *
651  * The driver uses four locking strategies for its operation:
652  *
653  * - The global SX lock "pmc_sx" is used to protect internal
654  *   data structures.
655  *
656  *   Calls into the module by syscall() start with this lock being
657  *   held in exclusive mode.  Depending on the requested operation,
658  *   the lock may be downgraded to 'shared' mode to allow more
659  *   concurrent readers into the module.  Calls into the module from
660  *   other parts of the kernel acquire the lock in shared mode.
661  *
662  *   This SX lock is held in exclusive mode for any operations that
663  *   modify the linkages between the driver's internal data structures.
664  *
665  *   The 'pmc_hook' function pointer is also protected by this lock.
666  *   It is only examined with the sx lock held in exclusive mode.  The
667  *   kernel module is allowed to be unloaded only with the sx lock held
668  *   in exclusive mode.  In normal syscall handling, after acquiring the
669  *   pmc_sx lock we first check that 'pmc_hook' is non-null before
670  *   proceeding.  This prevents races between the thread unloading the module
671  *   and other threads seeking to use the module.
672  *
673  * - Lookups of target process structures and owner process structures
674  *   cannot use the global "pmc_sx" SX lock because these lookups need
675  *   to happen during context switches and in other critical sections
676  *   where sleeping is not allowed.  We protect these lookup tables
677  *   with their own private spin-mutexes, "pmc_processhash_mtx" and
678  *   "pmc_ownerhash_mtx".
679  *
680  * - Interrupt handlers work in a lock free manner.  At interrupt
681  *   time, handlers look at the PMC pointer (phw->phw_pmc) configured
682  *   when the PMC was started.  If this pointer is NULL, the interrupt
683  *   is ignored after updating driver statistics.  We ensure that this
684  *   pointer is set (using an atomic operation if necessary) before the
685  *   PMC hardware is started.  Conversely, this pointer is unset atomically
686  *   only after the PMC hardware is stopped.
687  *
688  *   We ensure that everything needed for the operation of an
689  *   interrupt handler is available without it needing to acquire any
690  *   locks.  We also ensure that a PMC's software state is destroyed only
691  *   after the PMC is taken off hardware (on all CPUs).
692  *
693  * - Context-switch handling with process-private PMCs needs more
694  *   care.
695  *
696  *   A given process may be the target of multiple PMCs.  For example,
697  *   PMCATTACH and PMCDETACH may be requested by a process on one CPU
698  *   while the target process is running on another.  A PMC could also
699  *   be getting released because its owner is exiting.  We tackle
700  *   these situations in the following manner:
701  *
702  *   - each target process structure 'pmc_process' has an array
703  *     of 'struct pmc *' pointers, one for each hardware PMC.
704  *
705  *   - At context switch IN time, each "target" PMC in RUNNING state
706  *     gets started on hardware and a pointer to each PMC is copied into
707  *     the per-cpu phw array.  The 'runcount' for the PMC is
708  *     incremented.
709  *
710  *   - At context switch OUT time, all process-virtual PMCs are stopped
711  *     on hardware.  The saved value is added to the PMCs value field
712  *     only if the PMC is in a non-deleted state (the PMCs state could
713  *     have changed during the current time slice).
714  *
715  *     Note that since in-between a switch IN on a processor and a switch
716  *     OUT, the PMC could have been released on another CPU.  Therefore
717  *     context switch OUT always looks at the hardware state to turn
718  *     OFF PMCs and will update a PMC's saved value only if reachable
719  *     from the target process record.
720  *
721  *   - OP PMCRELEASE could be called on a PMC at any time (the PMC could
722  *     be attached to many processes at the time of the call and could
723  *     be active on multiple CPUs).
724  *
725  *     We prevent further scheduling of the PMC by marking it as in
726  *     state 'DELETED'.  If the runcount of the PMC is non-zero then
727  *     this PMC is currently running on a CPU somewhere.  The thread
728  *     doing the PMCRELEASE operation waits by repeatedly doing a
729  *     pause() till the runcount comes to zero.
730  *
731  * The contents of a PMC descriptor (struct pmc) are protected using
732  * a spin-mutex.  In order to save space, we use a mutex pool.
733  *
734  * In terms of lock types used by witness(4), we use:
735  * - Type "pmc-sx", used by the global SX lock.
736  * - Type "pmc-sleep", for sleep mutexes used by logger threads.
737  * - Type "pmc-per-proc", for protecting PMC owner descriptors.
738  * - Type "pmc-leaf", used for all other spin mutexes.
739  */
740 
741 /*
742  * save the cpu binding of the current kthread
743  */
744 
745 static void
746 pmc_save_cpu_binding(struct pmc_binding *pb)
747 {
748 	PMCDBG0(CPU,BND,2, "save-cpu");
749 	thread_lock(curthread);
750 	pb->pb_bound = sched_is_bound(curthread);
751 	pb->pb_cpu   = curthread->td_oncpu;
752 	thread_unlock(curthread);
753 	PMCDBG1(CPU,BND,2, "save-cpu cpu=%d", pb->pb_cpu);
754 }
755 
756 /*
757  * restore the cpu binding of the current thread
758  */
759 
760 static void
761 pmc_restore_cpu_binding(struct pmc_binding *pb)
762 {
763 	PMCDBG2(CPU,BND,2, "restore-cpu curcpu=%d restore=%d",
764 	    curthread->td_oncpu, pb->pb_cpu);
765 	thread_lock(curthread);
766 	if (pb->pb_bound)
767 		sched_bind(curthread, pb->pb_cpu);
768 	else
769 		sched_unbind(curthread);
770 	thread_unlock(curthread);
771 	PMCDBG0(CPU,BND,2, "restore-cpu done");
772 }
773 
774 /*
775  * move execution over the specified cpu and bind it there.
776  */
777 
778 static void
779 pmc_select_cpu(int cpu)
780 {
781 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
782 	    ("[pmc,%d] bad cpu number %d", __LINE__, cpu));
783 
784 	/* Never move to an inactive CPU. */
785 	KASSERT(pmc_cpu_is_active(cpu), ("[pmc,%d] selecting inactive "
786 	    "CPU %d", __LINE__, cpu));
787 
788 	PMCDBG1(CPU,SEL,2, "select-cpu cpu=%d", cpu);
789 	thread_lock(curthread);
790 	sched_bind(curthread, cpu);
791 	thread_unlock(curthread);
792 
793 	KASSERT(curthread->td_oncpu == cpu,
794 	    ("[pmc,%d] CPU not bound [cpu=%d, curr=%d]", __LINE__,
795 		cpu, curthread->td_oncpu));
796 
797 	PMCDBG1(CPU,SEL,2, "select-cpu cpu=%d ok", cpu);
798 }
799 
800 /*
801  * Force a context switch.
802  *
803  * We do this by pause'ing for 1 tick -- invoking mi_switch() is not
804  * guaranteed to force a context switch.
805  */
806 
807 static void
808 pmc_force_context_switch(void)
809 {
810 
811 	pause("pmcctx", 1);
812 }
813 
814 uint64_t
815 pmc_rdtsc(void)
816 {
817 #if defined(__i386__) || defined(__amd64__)
818 	if (__predict_true(amd_feature & AMDID_RDTSCP))
819 		return rdtscp();
820 	else
821 		return rdtsc();
822 #else
823 	return get_cyclecount();
824 #endif
825 }
826 
827 /*
828  * Get the file name for an executable.  This is a simple wrapper
829  * around vn_fullpath(9).
830  */
831 
832 static void
833 pmc_getfilename(struct vnode *v, char **fullpath, char **freepath)
834 {
835 
836 	*fullpath = "unknown";
837 	*freepath = NULL;
838 	vn_fullpath(curthread, v, fullpath, freepath);
839 }
840 
841 /*
842  * remove an process owning PMCs
843  */
844 
845 void
846 pmc_remove_owner(struct pmc_owner *po)
847 {
848 	struct pmc *pm, *tmp;
849 
850 	sx_assert(&pmc_sx, SX_XLOCKED);
851 
852 	PMCDBG1(OWN,ORM,1, "remove-owner po=%p", po);
853 
854 	/* Remove descriptor from the owner hash table */
855 	LIST_REMOVE(po, po_next);
856 
857 	/* release all owned PMC descriptors */
858 	LIST_FOREACH_SAFE(pm, &po->po_pmcs, pm_next, tmp) {
859 		PMCDBG1(OWN,ORM,2, "pmc=%p", pm);
860 		KASSERT(pm->pm_owner == po,
861 		    ("[pmc,%d] owner %p != po %p", __LINE__, pm->pm_owner, po));
862 
863 		pmc_release_pmc_descriptor(pm);	/* will unlink from the list */
864 		pmc_destroy_pmc_descriptor(pm);
865 	}
866 
867 	KASSERT(po->po_sscount == 0,
868 	    ("[pmc,%d] SS count not zero", __LINE__));
869 	KASSERT(LIST_EMPTY(&po->po_pmcs),
870 	    ("[pmc,%d] PMC list not empty", __LINE__));
871 
872 	/* de-configure the log file if present */
873 	if (po->po_flags & PMC_PO_OWNS_LOGFILE)
874 		pmclog_deconfigure_log(po);
875 }
876 
877 /*
878  * remove an owner process record if all conditions are met.
879  */
880 
881 static void
882 pmc_maybe_remove_owner(struct pmc_owner *po)
883 {
884 
885 	PMCDBG1(OWN,OMR,1, "maybe-remove-owner po=%p", po);
886 
887 	/*
888 	 * Remove owner record if
889 	 * - this process does not own any PMCs
890 	 * - this process has not allocated a system-wide sampling buffer
891 	 */
892 
893 	if (LIST_EMPTY(&po->po_pmcs) &&
894 	    ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)) {
895 		pmc_remove_owner(po);
896 		pmc_destroy_owner_descriptor(po);
897 	}
898 }
899 
900 /*
901  * Add an association between a target process and a PMC.
902  */
903 
904 static void
905 pmc_link_target_process(struct pmc *pm, struct pmc_process *pp)
906 {
907 	int ri;
908 	struct pmc_target *pt;
909 #ifdef INVARIANTS
910 	struct pmc_thread *pt_td;
911 #endif
912 
913 	sx_assert(&pmc_sx, SX_XLOCKED);
914 
915 	KASSERT(pm != NULL && pp != NULL,
916 	    ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp));
917 	KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)),
918 	    ("[pmc,%d] Attaching a non-process-virtual pmc=%p to pid=%d",
919 		__LINE__, pm, pp->pp_proc->p_pid));
920 	KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= ((int) md->pmd_npmc - 1),
921 	    ("[pmc,%d] Illegal reference count %d for process record %p",
922 		__LINE__, pp->pp_refcnt, (void *) pp));
923 
924 	ri = PMC_TO_ROWINDEX(pm);
925 
926 	PMCDBG3(PRC,TLK,1, "link-target pmc=%p ri=%d pmc-process=%p",
927 	    pm, ri, pp);
928 
929 #ifdef	HWPMC_DEBUG
930 	LIST_FOREACH(pt, &pm->pm_targets, pt_next)
931 	    if (pt->pt_process == pp)
932 		    KASSERT(0, ("[pmc,%d] pp %p already in pmc %p targets",
933 				__LINE__, pp, pm));
934 #endif
935 
936 	pt = malloc(sizeof(struct pmc_target), M_PMC, M_WAITOK|M_ZERO);
937 	pt->pt_process = pp;
938 
939 	LIST_INSERT_HEAD(&pm->pm_targets, pt, pt_next);
940 
941 	atomic_store_rel_ptr((uintptr_t *)&pp->pp_pmcs[ri].pp_pmc,
942 	    (uintptr_t)pm);
943 
944 	if (pm->pm_owner->po_owner == pp->pp_proc)
945 		pm->pm_flags |= PMC_F_ATTACHED_TO_OWNER;
946 
947 	/*
948 	 * Initialize the per-process values at this row index.
949 	 */
950 	pp->pp_pmcs[ri].pp_pmcval = PMC_TO_MODE(pm) == PMC_MODE_TS ?
951 	    pm->pm_sc.pm_reloadcount : 0;
952 
953 	pp->pp_refcnt++;
954 
955 #ifdef INVARIANTS
956 	/* Confirm that the per-thread values at this row index are cleared. */
957 	if (PMC_TO_MODE(pm) == PMC_MODE_TS) {
958 		mtx_lock_spin(pp->pp_tdslock);
959 		LIST_FOREACH(pt_td, &pp->pp_tds, pt_next) {
960 			KASSERT(pt_td->pt_pmcs[ri].pt_pmcval == (pmc_value_t) 0,
961 			    ("[pmc,%d] pt_pmcval not cleared for pid=%d at "
962 			    "ri=%d", __LINE__, pp->pp_proc->p_pid, ri));
963 		}
964 		mtx_unlock_spin(pp->pp_tdslock);
965 	}
966 #endif
967 }
968 
969 /*
970  * Removes the association between a target process and a PMC.
971  */
972 
973 static void
974 pmc_unlink_target_process(struct pmc *pm, struct pmc_process *pp)
975 {
976 	int ri;
977 	struct proc *p;
978 	struct pmc_target *ptgt;
979 	struct pmc_thread *pt;
980 
981 	sx_assert(&pmc_sx, SX_XLOCKED);
982 
983 	KASSERT(pm != NULL && pp != NULL,
984 	    ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp));
985 
986 	KASSERT(pp->pp_refcnt >= 1 && pp->pp_refcnt <= (int) md->pmd_npmc,
987 	    ("[pmc,%d] Illegal ref count %d on process record %p",
988 		__LINE__, pp->pp_refcnt, (void *) pp));
989 
990 	ri = PMC_TO_ROWINDEX(pm);
991 
992 	PMCDBG3(PRC,TUL,1, "unlink-target pmc=%p ri=%d pmc-process=%p",
993 	    pm, ri, pp);
994 
995 	KASSERT(pp->pp_pmcs[ri].pp_pmc == pm,
996 	    ("[pmc,%d] PMC ri %d mismatch pmc %p pp->[ri] %p", __LINE__,
997 		ri, pm, pp->pp_pmcs[ri].pp_pmc));
998 
999 	pp->pp_pmcs[ri].pp_pmc = NULL;
1000 	pp->pp_pmcs[ri].pp_pmcval = (pmc_value_t) 0;
1001 
1002 	/* Clear the per-thread values at this row index. */
1003 	if (PMC_TO_MODE(pm) == PMC_MODE_TS) {
1004 		mtx_lock_spin(pp->pp_tdslock);
1005 		LIST_FOREACH(pt, &pp->pp_tds, pt_next)
1006 			pt->pt_pmcs[ri].pt_pmcval = (pmc_value_t) 0;
1007 		mtx_unlock_spin(pp->pp_tdslock);
1008 	}
1009 
1010 	/* Remove owner-specific flags */
1011 	if (pm->pm_owner->po_owner == pp->pp_proc) {
1012 		pp->pp_flags &= ~PMC_PP_ENABLE_MSR_ACCESS;
1013 		pm->pm_flags &= ~PMC_F_ATTACHED_TO_OWNER;
1014 	}
1015 
1016 	pp->pp_refcnt--;
1017 
1018 	/* Remove the target process from the PMC structure */
1019 	LIST_FOREACH(ptgt, &pm->pm_targets, pt_next)
1020 		if (ptgt->pt_process == pp)
1021 			break;
1022 
1023 	KASSERT(ptgt != NULL, ("[pmc,%d] process %p (pp: %p) not found "
1024 		    "in pmc %p", __LINE__, pp->pp_proc, pp, pm));
1025 
1026 	LIST_REMOVE(ptgt, pt_next);
1027 	free(ptgt, M_PMC);
1028 
1029 	/* if the PMC now lacks targets, send the owner a SIGIO */
1030 	if (LIST_EMPTY(&pm->pm_targets)) {
1031 		p = pm->pm_owner->po_owner;
1032 		PROC_LOCK(p);
1033 		kern_psignal(p, SIGIO);
1034 		PROC_UNLOCK(p);
1035 
1036 		PMCDBG2(PRC,SIG,2, "signalling proc=%p signal=%d", p,
1037 		    SIGIO);
1038 	}
1039 }
1040 
1041 /*
1042  * Check if PMC 'pm' may be attached to target process 't'.
1043  */
1044 
1045 static int
1046 pmc_can_attach(struct pmc *pm, struct proc *t)
1047 {
1048 	struct proc *o;		/* pmc owner */
1049 	struct ucred *oc, *tc;	/* owner, target credentials */
1050 	int decline_attach, i;
1051 
1052 	/*
1053 	 * A PMC's owner can always attach that PMC to itself.
1054 	 */
1055 
1056 	if ((o = pm->pm_owner->po_owner) == t)
1057 		return 0;
1058 
1059 	PROC_LOCK(o);
1060 	oc = o->p_ucred;
1061 	crhold(oc);
1062 	PROC_UNLOCK(o);
1063 
1064 	PROC_LOCK(t);
1065 	tc = t->p_ucred;
1066 	crhold(tc);
1067 	PROC_UNLOCK(t);
1068 
1069 	/*
1070 	 * The effective uid of the PMC owner should match at least one
1071 	 * of the {effective,real,saved} uids of the target process.
1072 	 */
1073 
1074 	decline_attach = oc->cr_uid != tc->cr_uid &&
1075 	    oc->cr_uid != tc->cr_svuid &&
1076 	    oc->cr_uid != tc->cr_ruid;
1077 
1078 	/*
1079 	 * Every one of the target's group ids, must be in the owner's
1080 	 * group list.
1081 	 */
1082 	for (i = 0; !decline_attach && i < tc->cr_ngroups; i++)
1083 		decline_attach = !groupmember(tc->cr_groups[i], oc);
1084 
1085 	/* check the read and saved gids too */
1086 	if (decline_attach == 0)
1087 		decline_attach = !groupmember(tc->cr_rgid, oc) ||
1088 		    !groupmember(tc->cr_svgid, oc);
1089 
1090 	crfree(tc);
1091 	crfree(oc);
1092 
1093 	return !decline_attach;
1094 }
1095 
1096 /*
1097  * Attach a process to a PMC.
1098  */
1099 
1100 static int
1101 pmc_attach_one_process(struct proc *p, struct pmc *pm)
1102 {
1103 	int ri, error;
1104 	char *fullpath, *freepath;
1105 	struct pmc_process	*pp;
1106 
1107 	sx_assert(&pmc_sx, SX_XLOCKED);
1108 
1109 	PMCDBG5(PRC,ATT,2, "attach-one pm=%p ri=%d proc=%p (%d, %s)", pm,
1110 	    PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm);
1111 
1112 	/*
1113 	 * Locate the process descriptor corresponding to process 'p',
1114 	 * allocating space as needed.
1115 	 *
1116 	 * Verify that rowindex 'pm_rowindex' is free in the process
1117 	 * descriptor.
1118 	 *
1119 	 * If not, allocate space for a descriptor and link the
1120 	 * process descriptor and PMC.
1121 	 */
1122 	ri = PMC_TO_ROWINDEX(pm);
1123 
1124 	/* mark process as using HWPMCs */
1125 	PROC_LOCK(p);
1126 	p->p_flag |= P_HWPMC;
1127 	PROC_UNLOCK(p);
1128 
1129 	if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_ALLOCATE)) == NULL) {
1130 		error = ENOMEM;
1131 		goto fail;
1132 	}
1133 
1134 	if (pp->pp_pmcs[ri].pp_pmc == pm) {/* already present at slot [ri] */
1135 		error = EEXIST;
1136 		goto fail;
1137 	}
1138 
1139 	if (pp->pp_pmcs[ri].pp_pmc != NULL) {
1140 		error = EBUSY;
1141 		goto fail;
1142 	}
1143 
1144 	pmc_link_target_process(pm, pp);
1145 
1146 	if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) &&
1147 	    (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) == 0)
1148 		pm->pm_flags |= PMC_F_NEEDS_LOGFILE;
1149 
1150 	pm->pm_flags |= PMC_F_ATTACH_DONE; /* mark as attached */
1151 
1152 	/* issue an attach event to a configured log file */
1153 	if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE) {
1154 		if (p->p_flag & P_KPROC) {
1155 			fullpath = kernelname;
1156 			freepath = NULL;
1157 		} else {
1158 			pmc_getfilename(p->p_textvp, &fullpath, &freepath);
1159 			pmclog_process_pmcattach(pm, p->p_pid, fullpath);
1160 		}
1161 		free(freepath, M_TEMP);
1162 		if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1163 			pmc_log_process_mappings(pm->pm_owner, p);
1164 	}
1165 
1166 	return (0);
1167  fail:
1168 	PROC_LOCK(p);
1169 	p->p_flag &= ~P_HWPMC;
1170 	PROC_UNLOCK(p);
1171 	return (error);
1172 }
1173 
1174 /*
1175  * Attach a process and optionally its children
1176  */
1177 
1178 static int
1179 pmc_attach_process(struct proc *p, struct pmc *pm)
1180 {
1181 	int error;
1182 	struct proc *top;
1183 
1184 	sx_assert(&pmc_sx, SX_XLOCKED);
1185 
1186 	PMCDBG5(PRC,ATT,1, "attach pm=%p ri=%d proc=%p (%d, %s)", pm,
1187 	    PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm);
1188 
1189 
1190 	/*
1191 	 * If this PMC successfully allowed a GETMSR operation
1192 	 * in the past, disallow further ATTACHes.
1193 	 */
1194 
1195 	if ((pm->pm_flags & PMC_PP_ENABLE_MSR_ACCESS) != 0)
1196 		return EPERM;
1197 
1198 	if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0)
1199 		return pmc_attach_one_process(p, pm);
1200 
1201 	/*
1202 	 * Traverse all child processes, attaching them to
1203 	 * this PMC.
1204 	 */
1205 
1206 	sx_slock(&proctree_lock);
1207 
1208 	top = p;
1209 
1210 	for (;;) {
1211 		if ((error = pmc_attach_one_process(p, pm)) != 0)
1212 			break;
1213 		if (!LIST_EMPTY(&p->p_children))
1214 			p = LIST_FIRST(&p->p_children);
1215 		else for (;;) {
1216 			if (p == top)
1217 				goto done;
1218 			if (LIST_NEXT(p, p_sibling)) {
1219 				p = LIST_NEXT(p, p_sibling);
1220 				break;
1221 			}
1222 			p = p->p_pptr;
1223 		}
1224 	}
1225 
1226 	if (error)
1227 		(void) pmc_detach_process(top, pm);
1228 
1229  done:
1230 	sx_sunlock(&proctree_lock);
1231 	return error;
1232 }
1233 
1234 /*
1235  * Detach a process from a PMC.  If there are no other PMCs tracking
1236  * this process, remove the process structure from its hash table.  If
1237  * 'flags' contains PMC_FLAG_REMOVE, then free the process structure.
1238  */
1239 
1240 static int
1241 pmc_detach_one_process(struct proc *p, struct pmc *pm, int flags)
1242 {
1243 	int ri;
1244 	struct pmc_process *pp;
1245 
1246 	sx_assert(&pmc_sx, SX_XLOCKED);
1247 
1248 	KASSERT(pm != NULL,
1249 	    ("[pmc,%d] null pm pointer", __LINE__));
1250 
1251 	ri = PMC_TO_ROWINDEX(pm);
1252 
1253 	PMCDBG6(PRC,ATT,2, "detach-one pm=%p ri=%d proc=%p (%d, %s) flags=0x%x",
1254 	    pm, ri, p, p->p_pid, p->p_comm, flags);
1255 
1256 	if ((pp = pmc_find_process_descriptor(p, 0)) == NULL)
1257 		return ESRCH;
1258 
1259 	if (pp->pp_pmcs[ri].pp_pmc != pm)
1260 		return EINVAL;
1261 
1262 	pmc_unlink_target_process(pm, pp);
1263 
1264 	/* Issue a detach entry if a log file is configured */
1265 	if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE)
1266 		pmclog_process_pmcdetach(pm, p->p_pid);
1267 
1268 	/*
1269 	 * If there are no PMCs targeting this process, we remove its
1270 	 * descriptor from the target hash table and unset the P_HWPMC
1271 	 * flag in the struct proc.
1272 	 */
1273 	KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= (int) md->pmd_npmc,
1274 	    ("[pmc,%d] Illegal refcnt %d for process struct %p",
1275 		__LINE__, pp->pp_refcnt, pp));
1276 
1277 	if (pp->pp_refcnt != 0)	/* still a target of some PMC */
1278 		return 0;
1279 
1280 	pmc_remove_process_descriptor(pp);
1281 
1282 	if (flags & PMC_FLAG_REMOVE)
1283 		pmc_destroy_process_descriptor(pp);
1284 
1285 	PROC_LOCK(p);
1286 	p->p_flag &= ~P_HWPMC;
1287 	PROC_UNLOCK(p);
1288 
1289 	return 0;
1290 }
1291 
1292 /*
1293  * Detach a process and optionally its descendants from a PMC.
1294  */
1295 
1296 static int
1297 pmc_detach_process(struct proc *p, struct pmc *pm)
1298 {
1299 	struct proc *top;
1300 
1301 	sx_assert(&pmc_sx, SX_XLOCKED);
1302 
1303 	PMCDBG5(PRC,ATT,1, "detach pm=%p ri=%d proc=%p (%d, %s)", pm,
1304 	    PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm);
1305 
1306 	if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0)
1307 		return pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE);
1308 
1309 	/*
1310 	 * Traverse all children, detaching them from this PMC.  We
1311 	 * ignore errors since we could be detaching a PMC from a
1312 	 * partially attached proc tree.
1313 	 */
1314 
1315 	sx_slock(&proctree_lock);
1316 
1317 	top = p;
1318 
1319 	for (;;) {
1320 		(void) pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE);
1321 
1322 		if (!LIST_EMPTY(&p->p_children))
1323 			p = LIST_FIRST(&p->p_children);
1324 		else for (;;) {
1325 			if (p == top)
1326 				goto done;
1327 			if (LIST_NEXT(p, p_sibling)) {
1328 				p = LIST_NEXT(p, p_sibling);
1329 				break;
1330 			}
1331 			p = p->p_pptr;
1332 		}
1333 	}
1334 
1335  done:
1336 	sx_sunlock(&proctree_lock);
1337 
1338 	if (LIST_EMPTY(&pm->pm_targets))
1339 		pm->pm_flags &= ~PMC_F_ATTACH_DONE;
1340 
1341 	return 0;
1342 }
1343 
1344 
1345 /*
1346  * Thread context switch IN
1347  */
1348 
1349 static void
1350 pmc_process_csw_in(struct thread *td)
1351 {
1352 	int cpu;
1353 	unsigned int adjri, ri;
1354 	struct pmc *pm;
1355 	struct proc *p;
1356 	struct pmc_cpu *pc;
1357 	struct pmc_hw *phw;
1358 	pmc_value_t newvalue;
1359 	struct pmc_process *pp;
1360 	struct pmc_thread *pt;
1361 	struct pmc_classdep *pcd;
1362 
1363 	p = td->td_proc;
1364 	pt = NULL;
1365 	if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE)) == NULL)
1366 		return;
1367 
1368 	KASSERT(pp->pp_proc == td->td_proc,
1369 	    ("[pmc,%d] not my thread state", __LINE__));
1370 
1371 	critical_enter(); /* no preemption from this point */
1372 
1373 	cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */
1374 
1375 	PMCDBG5(CSW,SWI,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p,
1376 	    p->p_pid, p->p_comm, pp);
1377 
1378 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
1379 	    ("[pmc,%d] weird CPU id %d", __LINE__, cpu));
1380 
1381 	pc = pmc_pcpu[cpu];
1382 
1383 	for (ri = 0; ri < md->pmd_npmc; ri++) {
1384 
1385 		if ((pm = pp->pp_pmcs[ri].pp_pmc) == NULL)
1386 			continue;
1387 
1388 		KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)),
1389 		    ("[pmc,%d] Target PMC in non-virtual mode (%d)",
1390 			__LINE__, PMC_TO_MODE(pm)));
1391 
1392 		KASSERT(PMC_TO_ROWINDEX(pm) == ri,
1393 		    ("[pmc,%d] Row index mismatch pmc %d != ri %d",
1394 			__LINE__, PMC_TO_ROWINDEX(pm), ri));
1395 
1396 		/*
1397 		 * Only PMCs that are marked as 'RUNNING' need
1398 		 * be placed on hardware.
1399 		 */
1400 
1401 		if (pm->pm_state != PMC_STATE_RUNNING)
1402 			continue;
1403 
1404 		/* increment PMC runcount */
1405 		counter_u64_add(pm->pm_runcount, 1);
1406 
1407 		/* configure the HWPMC we are going to use. */
1408 		pcd = pmc_ri_to_classdep(md, ri, &adjri);
1409 		pcd->pcd_config_pmc(cpu, adjri, pm);
1410 
1411 		phw = pc->pc_hwpmcs[ri];
1412 
1413 		KASSERT(phw != NULL,
1414 		    ("[pmc,%d] null hw pointer", __LINE__));
1415 
1416 		KASSERT(phw->phw_pmc == pm,
1417 		    ("[pmc,%d] hw->pmc %p != pmc %p", __LINE__,
1418 			phw->phw_pmc, pm));
1419 
1420 		/*
1421 		 * Write out saved value and start the PMC.
1422 		 *
1423 		 * Sampling PMCs use a per-thread value, while
1424 		 * counting mode PMCs use a per-pmc value that is
1425 		 * inherited across descendants.
1426 		 */
1427 		if (PMC_TO_MODE(pm) == PMC_MODE_TS) {
1428 			if (pt == NULL)
1429 				pt = pmc_find_thread_descriptor(pp, td,
1430 				    PMC_FLAG_NONE);
1431 
1432 			KASSERT(pt != NULL,
1433 			    ("[pmc,%d] No thread found for td=%p", __LINE__,
1434 			    td));
1435 
1436 			mtx_pool_lock_spin(pmc_mtxpool, pm);
1437 
1438 			/*
1439 			 * If we have a thread descriptor, use the per-thread
1440 			 * counter in the descriptor. If not, we will use
1441 			 * a per-process counter.
1442 			 *
1443 			 * TODO: Remove the per-process "safety net" once
1444 			 * we have thoroughly tested that we don't hit the
1445 			 * above assert.
1446 			 */
1447 			if (pt != NULL) {
1448 				if (pt->pt_pmcs[ri].pt_pmcval > 0)
1449 					newvalue = pt->pt_pmcs[ri].pt_pmcval;
1450 				else
1451 					newvalue = pm->pm_sc.pm_reloadcount;
1452 			} else {
1453 				/*
1454 				 * Use the saved value calculated after the most
1455 				 * recent time a thread using the shared counter
1456 				 * switched out. Reset the saved count in case
1457 				 * another thread from this process switches in
1458 				 * before any threads switch out.
1459 				 */
1460 
1461 				newvalue = pp->pp_pmcs[ri].pp_pmcval;
1462 				pp->pp_pmcs[ri].pp_pmcval =
1463 				    pm->pm_sc.pm_reloadcount;
1464 			}
1465 			mtx_pool_unlock_spin(pmc_mtxpool, pm);
1466 			KASSERT(newvalue > 0 && newvalue <=
1467 			    pm->pm_sc.pm_reloadcount,
1468 			    ("[pmc,%d] pmcval outside of expected range cpu=%d "
1469 			    "ri=%d pmcval=%jx pm_reloadcount=%jx", __LINE__,
1470 			    cpu, ri, newvalue, pm->pm_sc.pm_reloadcount));
1471 		} else {
1472 			KASSERT(PMC_TO_MODE(pm) == PMC_MODE_TC,
1473 			    ("[pmc,%d] illegal mode=%d", __LINE__,
1474 			    PMC_TO_MODE(pm)));
1475 			mtx_pool_lock_spin(pmc_mtxpool, pm);
1476 			newvalue = PMC_PCPU_SAVED(cpu, ri) =
1477 			    pm->pm_gv.pm_savedvalue;
1478 			mtx_pool_unlock_spin(pmc_mtxpool, pm);
1479 		}
1480 
1481 		PMCDBG3(CSW,SWI,1,"cpu=%d ri=%d new=%jd", cpu, ri, newvalue);
1482 
1483 		pcd->pcd_write_pmc(cpu, adjri, newvalue);
1484 
1485 		/* If a sampling mode PMC, reset stalled state. */
1486 		if (PMC_TO_MODE(pm) == PMC_MODE_TS)
1487 			pm->pm_pcpu_state[cpu].pps_stalled = 0;
1488 
1489 		/* Indicate that we desire this to run. */
1490 		pm->pm_pcpu_state[cpu].pps_cpustate = 1;
1491 
1492 		/* Start the PMC. */
1493 		pcd->pcd_start_pmc(cpu, adjri);
1494 	}
1495 
1496 	/*
1497 	 * perform any other architecture/cpu dependent thread
1498 	 * switch-in actions.
1499 	 */
1500 
1501 	(void) (*md->pmd_switch_in)(pc, pp);
1502 
1503 	critical_exit();
1504 
1505 }
1506 
1507 /*
1508  * Thread context switch OUT.
1509  */
1510 
1511 static void
1512 pmc_process_csw_out(struct thread *td)
1513 {
1514 	int cpu;
1515 	int64_t tmp;
1516 	struct pmc *pm;
1517 	struct proc *p;
1518 	enum pmc_mode mode;
1519 	struct pmc_cpu *pc;
1520 	pmc_value_t newvalue;
1521 	unsigned int adjri, ri;
1522 	struct pmc_process *pp;
1523 	struct pmc_thread *pt = NULL;
1524 	struct pmc_classdep *pcd;
1525 
1526 
1527 	/*
1528 	 * Locate our process descriptor; this may be NULL if
1529 	 * this process is exiting and we have already removed
1530 	 * the process from the target process table.
1531 	 *
1532 	 * Note that due to kernel preemption, multiple
1533 	 * context switches may happen while the process is
1534 	 * exiting.
1535 	 *
1536 	 * Note also that if the target process cannot be
1537 	 * found we still need to deconfigure any PMCs that
1538 	 * are currently running on hardware.
1539 	 */
1540 
1541 	p = td->td_proc;
1542 	pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE);
1543 
1544 	/*
1545 	 * save PMCs
1546 	 */
1547 
1548 	critical_enter();
1549 
1550 	cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */
1551 
1552 	PMCDBG5(CSW,SWO,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p,
1553 	    p->p_pid, p->p_comm, pp);
1554 
1555 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
1556 	    ("[pmc,%d weird CPU id %d", __LINE__, cpu));
1557 
1558 	pc = pmc_pcpu[cpu];
1559 
1560 	/*
1561 	 * When a PMC gets unlinked from a target PMC, it will
1562 	 * be removed from the target's pp_pmc[] array.
1563 	 *
1564 	 * However, on a MP system, the target could have been
1565 	 * executing on another CPU at the time of the unlink.
1566 	 * So, at context switch OUT time, we need to look at
1567 	 * the hardware to determine if a PMC is scheduled on
1568 	 * it.
1569 	 */
1570 
1571 	for (ri = 0; ri < md->pmd_npmc; ri++) {
1572 
1573 		pcd = pmc_ri_to_classdep(md, ri, &adjri);
1574 		pm  = NULL;
1575 		(void) (*pcd->pcd_get_config)(cpu, adjri, &pm);
1576 
1577 		if (pm == NULL)	/* nothing at this row index */
1578 			continue;
1579 
1580 		mode = PMC_TO_MODE(pm);
1581 		if (!PMC_IS_VIRTUAL_MODE(mode))
1582 			continue; /* not a process virtual PMC */
1583 
1584 		KASSERT(PMC_TO_ROWINDEX(pm) == ri,
1585 		    ("[pmc,%d] ri mismatch pmc(%d) ri(%d)",
1586 			__LINE__, PMC_TO_ROWINDEX(pm), ri));
1587 
1588 		/*
1589 		 * Change desired state, and then stop if not stalled.
1590 		 * This two-step dance should avoid race conditions where
1591 		 * an interrupt re-enables the PMC after this code has
1592 		 * already checked the pm_stalled flag.
1593 		 */
1594 		pm->pm_pcpu_state[cpu].pps_cpustate = 0;
1595 		if (pm->pm_pcpu_state[cpu].pps_stalled == 0)
1596 			pcd->pcd_stop_pmc(cpu, adjri);
1597 
1598 		/* reduce this PMC's runcount */
1599 		counter_u64_add(pm->pm_runcount, -1);
1600 
1601 		/*
1602 		 * If this PMC is associated with this process,
1603 		 * save the reading.
1604 		 */
1605 
1606 		if (pm->pm_state != PMC_STATE_DELETED && pp != NULL &&
1607 		    pp->pp_pmcs[ri].pp_pmc != NULL) {
1608 			KASSERT(pm == pp->pp_pmcs[ri].pp_pmc,
1609 			    ("[pmc,%d] pm %p != pp_pmcs[%d] %p", __LINE__,
1610 				pm, ri, pp->pp_pmcs[ri].pp_pmc));
1611 
1612 			KASSERT(pp->pp_refcnt > 0,
1613 			    ("[pmc,%d] pp refcnt = %d", __LINE__,
1614 				pp->pp_refcnt));
1615 
1616 			pcd->pcd_read_pmc(cpu, adjri, &newvalue);
1617 
1618 			if (mode == PMC_MODE_TS) {
1619 				PMCDBG3(CSW,SWO,1,"cpu=%d ri=%d val=%jd (samp)",
1620 				    cpu, ri, newvalue);
1621 
1622 				if (pt == NULL)
1623 					pt = pmc_find_thread_descriptor(pp, td,
1624 					    PMC_FLAG_NONE);
1625 
1626 				KASSERT(pt != NULL,
1627 				    ("[pmc,%d] No thread found for td=%p",
1628 				    __LINE__, td));
1629 
1630 				mtx_pool_lock_spin(pmc_mtxpool, pm);
1631 
1632 				/*
1633 				 * If we have a thread descriptor, save the
1634 				 * per-thread counter in the descriptor. If not,
1635 				 * we will update the per-process counter.
1636 				 *
1637 				 * TODO: Remove the per-process "safety net"
1638 				 * once we have thoroughly tested that we
1639 				 * don't hit the above assert.
1640 				 */
1641 				if (pt != NULL)
1642 					pt->pt_pmcs[ri].pt_pmcval = newvalue;
1643 				else {
1644 					/*
1645 					 * For sampling process-virtual PMCs,
1646 					 * newvalue is the number of events to
1647 					 * be seen until the next sampling
1648 					 * interrupt. We can just add the events
1649 					 * left from this invocation to the
1650 					 * counter, then adjust in case we
1651 					 * overflow our range.
1652 					 *
1653 					 * (Recall that we reload the counter
1654 					 * every time we use it.)
1655 					 */
1656 					pp->pp_pmcs[ri].pp_pmcval += newvalue;
1657 					if (pp->pp_pmcs[ri].pp_pmcval >
1658 					    pm->pm_sc.pm_reloadcount)
1659 						pp->pp_pmcs[ri].pp_pmcval -=
1660 						    pm->pm_sc.pm_reloadcount;
1661 				}
1662 				mtx_pool_unlock_spin(pmc_mtxpool, pm);
1663 			} else {
1664 				tmp = newvalue - PMC_PCPU_SAVED(cpu,ri);
1665 
1666 				PMCDBG3(CSW,SWO,1,"cpu=%d ri=%d tmp=%jd (count)",
1667 				    cpu, ri, tmp);
1668 
1669 				/*
1670 				 * For counting process-virtual PMCs,
1671 				 * we expect the count to be
1672 				 * increasing monotonically, modulo a 64
1673 				 * bit wraparound.
1674 				 */
1675 				KASSERT(tmp >= 0,
1676 				    ("[pmc,%d] negative increment cpu=%d "
1677 				     "ri=%d newvalue=%jx saved=%jx "
1678 				     "incr=%jx", __LINE__, cpu, ri,
1679 				     newvalue, PMC_PCPU_SAVED(cpu,ri), tmp));
1680 
1681 				mtx_pool_lock_spin(pmc_mtxpool, pm);
1682 				pm->pm_gv.pm_savedvalue += tmp;
1683 				pp->pp_pmcs[ri].pp_pmcval += tmp;
1684 				mtx_pool_unlock_spin(pmc_mtxpool, pm);
1685 
1686 				if (pm->pm_flags & PMC_F_LOG_PROCCSW)
1687 					pmclog_process_proccsw(pm, pp, tmp, td);
1688 			}
1689 		}
1690 
1691 		/* mark hardware as free */
1692 		pcd->pcd_config_pmc(cpu, adjri, NULL);
1693 	}
1694 
1695 	/*
1696 	 * perform any other architecture/cpu dependent thread
1697 	 * switch out functions.
1698 	 */
1699 
1700 	(void) (*md->pmd_switch_out)(pc, pp);
1701 
1702 	critical_exit();
1703 }
1704 
1705 /*
1706  * A new thread for a process.
1707  */
1708 static void
1709 pmc_process_thread_add(struct thread *td)
1710 {
1711 	struct pmc_process *pmc;
1712 
1713 	pmc = pmc_find_process_descriptor(td->td_proc, PMC_FLAG_NONE);
1714 	if (pmc != NULL)
1715 		pmc_find_thread_descriptor(pmc, td, PMC_FLAG_ALLOCATE);
1716 }
1717 
1718 /*
1719  * A thread delete for a process.
1720  */
1721 static void
1722 pmc_process_thread_delete(struct thread *td)
1723 {
1724 	struct pmc_process *pmc;
1725 
1726 	pmc = pmc_find_process_descriptor(td->td_proc, PMC_FLAG_NONE);
1727 	if (pmc != NULL)
1728 		pmc_thread_descriptor_pool_free(pmc_find_thread_descriptor(pmc,
1729 		    td, PMC_FLAG_REMOVE));
1730 }
1731 
1732 /*
1733  * A userret() call for a thread.
1734  */
1735 static void
1736 pmc_process_thread_userret(struct thread *td)
1737 {
1738 	sched_pin();
1739 	pmc_capture_user_callchain(curcpu, PMC_UR, td->td_frame);
1740 	sched_unpin();
1741 }
1742 
1743 /*
1744  * A mapping change for a process.
1745  */
1746 
1747 static void
1748 pmc_process_mmap(struct thread *td, struct pmckern_map_in *pkm)
1749 {
1750 	int ri;
1751 	pid_t pid;
1752 	char *fullpath, *freepath;
1753 	const struct pmc *pm;
1754 	struct pmc_owner *po;
1755 	const struct pmc_process *pp;
1756 
1757 	freepath = fullpath = NULL;
1758 	MPASS(!in_epoch(global_epoch_preempt));
1759 	pmc_getfilename((struct vnode *) pkm->pm_file, &fullpath, &freepath);
1760 
1761 	pid = td->td_proc->p_pid;
1762 
1763 	PMC_EPOCH_ENTER();
1764 	/* Inform owners of all system-wide sampling PMCs. */
1765 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1766 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1767 			pmclog_process_map_in(po, pid, pkm->pm_address, fullpath);
1768 
1769 	if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL)
1770 		goto done;
1771 
1772 	/*
1773 	 * Inform sampling PMC owners tracking this process.
1774 	 */
1775 	for (ri = 0; ri < md->pmd_npmc; ri++)
1776 		if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL &&
1777 		    PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1778 			pmclog_process_map_in(pm->pm_owner,
1779 			    pid, pkm->pm_address, fullpath);
1780 
1781   done:
1782 	if (freepath)
1783 		free(freepath, M_TEMP);
1784 	PMC_EPOCH_EXIT();
1785 }
1786 
1787 
1788 /*
1789  * Log an munmap request.
1790  */
1791 
1792 static void
1793 pmc_process_munmap(struct thread *td, struct pmckern_map_out *pkm)
1794 {
1795 	int ri;
1796 	pid_t pid;
1797 	struct pmc_owner *po;
1798 	const struct pmc *pm;
1799 	const struct pmc_process *pp;
1800 
1801 	pid = td->td_proc->p_pid;
1802 
1803 	PMC_EPOCH_ENTER();
1804 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1805 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1806 		pmclog_process_map_out(po, pid, pkm->pm_address,
1807 		    pkm->pm_address + pkm->pm_size);
1808 	PMC_EPOCH_EXIT();
1809 
1810 	if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL)
1811 		return;
1812 
1813 	for (ri = 0; ri < md->pmd_npmc; ri++)
1814 		if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL &&
1815 		    PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1816 			pmclog_process_map_out(pm->pm_owner, pid,
1817 			    pkm->pm_address, pkm->pm_address + pkm->pm_size);
1818 }
1819 
1820 /*
1821  * Log mapping information about the kernel.
1822  */
1823 
1824 static void
1825 pmc_log_kernel_mappings(struct pmc *pm)
1826 {
1827 	struct pmc_owner *po;
1828 	struct pmckern_map_in *km, *kmbase;
1829 
1830 	MPASS(in_epoch(global_epoch_preempt) || sx_xlocked(&pmc_sx));
1831 	KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)),
1832 	    ("[pmc,%d] non-sampling PMC (%p) desires mapping information",
1833 		__LINE__, (void *) pm));
1834 
1835 	po = pm->pm_owner;
1836 
1837 	if (po->po_flags & PMC_PO_INITIAL_MAPPINGS_DONE)
1838 		return;
1839 	if (PMC_TO_MODE(pm) == PMC_MODE_SS)
1840 		pmc_process_allproc(pm);
1841 	/*
1842 	 * Log the current set of kernel modules.
1843 	 */
1844 	kmbase = linker_hwpmc_list_objects();
1845 	for (km = kmbase; km->pm_file != NULL; km++) {
1846 		PMCDBG2(LOG,REG,1,"%s %p", (char *) km->pm_file,
1847 		    (void *) km->pm_address);
1848 		pmclog_process_map_in(po, (pid_t) -1, km->pm_address,
1849 		    km->pm_file);
1850 	}
1851 	free(kmbase, M_LINKER);
1852 
1853 	po->po_flags |= PMC_PO_INITIAL_MAPPINGS_DONE;
1854 }
1855 
1856 /*
1857  * Log the mappings for a single process.
1858  */
1859 
1860 static void
1861 pmc_log_process_mappings(struct pmc_owner *po, struct proc *p)
1862 {
1863 	vm_map_t map;
1864 	struct vnode *vp;
1865 	struct vmspace *vm;
1866 	vm_map_entry_t entry;
1867 	vm_offset_t last_end;
1868 	u_int last_timestamp;
1869 	struct vnode *last_vp;
1870 	vm_offset_t start_addr;
1871 	vm_object_t obj, lobj, tobj;
1872 	char *fullpath, *freepath;
1873 
1874 	last_vp = NULL;
1875 	last_end = (vm_offset_t) 0;
1876 	fullpath = freepath = NULL;
1877 
1878 	if ((vm = vmspace_acquire_ref(p)) == NULL)
1879 		return;
1880 
1881 	map = &vm->vm_map;
1882 	vm_map_lock_read(map);
1883 
1884 	for (entry = map->header.next; entry != &map->header; entry = entry->next) {
1885 
1886 		if (entry == NULL) {
1887 			PMCDBG2(LOG,OPS,2, "hwpmc: vm_map entry unexpectedly "
1888 			    "NULL! pid=%d vm_map=%p\n", p->p_pid, map);
1889 			break;
1890 		}
1891 
1892 		/*
1893 		 * We only care about executable map entries.
1894 		 */
1895 		if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) ||
1896 		    !(entry->protection & VM_PROT_EXECUTE) ||
1897 		    (entry->object.vm_object == NULL)) {
1898 			continue;
1899 		}
1900 
1901 		obj = entry->object.vm_object;
1902 		VM_OBJECT_RLOCK(obj);
1903 
1904 		/*
1905 		 * Walk the backing_object list to find the base
1906 		 * (non-shadowed) vm_object.
1907 		 */
1908 		for (lobj = tobj = obj; tobj != NULL; tobj = tobj->backing_object) {
1909 			if (tobj != obj)
1910 				VM_OBJECT_RLOCK(tobj);
1911 			if (lobj != obj)
1912 				VM_OBJECT_RUNLOCK(lobj);
1913 			lobj = tobj;
1914 		}
1915 
1916 		/*
1917 		 * At this point lobj is the base vm_object and it is locked.
1918 		 */
1919 		if (lobj == NULL) {
1920 			PMCDBG3(LOG,OPS,2, "hwpmc: lobj unexpectedly NULL! pid=%d "
1921 			    "vm_map=%p vm_obj=%p\n", p->p_pid, map, obj);
1922 			VM_OBJECT_RUNLOCK(obj);
1923 			continue;
1924 		}
1925 
1926 		vp = vm_object_vnode(lobj);
1927 		if (vp == NULL) {
1928 			if (lobj != obj)
1929 				VM_OBJECT_RUNLOCK(lobj);
1930 			VM_OBJECT_RUNLOCK(obj);
1931 			continue;
1932 		}
1933 
1934 		/*
1935 		 * Skip contiguous regions that point to the same
1936 		 * vnode, so we don't emit redundant MAP-IN
1937 		 * directives.
1938 		 */
1939 		if (entry->start == last_end && vp == last_vp) {
1940 			last_end = entry->end;
1941 			if (lobj != obj)
1942 				VM_OBJECT_RUNLOCK(lobj);
1943 			VM_OBJECT_RUNLOCK(obj);
1944 			continue;
1945 		}
1946 
1947 		/*
1948 		 * We don't want to keep the proc's vm_map or this
1949 		 * vm_object locked while we walk the pathname, since
1950 		 * vn_fullpath() can sleep.  However, if we drop the
1951 		 * lock, it's possible for concurrent activity to
1952 		 * modify the vm_map list.  To protect against this,
1953 		 * we save the vm_map timestamp before we release the
1954 		 * lock, and check it after we reacquire the lock
1955 		 * below.
1956 		 */
1957 		start_addr = entry->start;
1958 		last_end = entry->end;
1959 		last_timestamp = map->timestamp;
1960 		vm_map_unlock_read(map);
1961 
1962 		vref(vp);
1963 		if (lobj != obj)
1964 			VM_OBJECT_RUNLOCK(lobj);
1965 
1966 		VM_OBJECT_RUNLOCK(obj);
1967 
1968 		freepath = NULL;
1969 		pmc_getfilename(vp, &fullpath, &freepath);
1970 		last_vp = vp;
1971 
1972 		vrele(vp);
1973 
1974 		vp = NULL;
1975 		pmclog_process_map_in(po, p->p_pid, start_addr, fullpath);
1976 		if (freepath)
1977 			free(freepath, M_TEMP);
1978 
1979 		vm_map_lock_read(map);
1980 
1981 		/*
1982 		 * If our saved timestamp doesn't match, this means
1983 		 * that the vm_map was modified out from under us and
1984 		 * we can't trust our current "entry" pointer.  Do a
1985 		 * new lookup for this entry.  If there is no entry
1986 		 * for this address range, vm_map_lookup_entry() will
1987 		 * return the previous one, so we always want to go to
1988 		 * entry->next on the next loop iteration.
1989 		 *
1990 		 * There is an edge condition here that can occur if
1991 		 * there is no entry at or before this address.  In
1992 		 * this situation, vm_map_lookup_entry returns
1993 		 * &map->header, which would cause our loop to abort
1994 		 * without processing the rest of the map.  However,
1995 		 * in practice this will never happen for process
1996 		 * vm_map.  This is because the executable's text
1997 		 * segment is the first mapping in the proc's address
1998 		 * space, and this mapping is never removed until the
1999 		 * process exits, so there will always be a non-header
2000 		 * entry at or before the requested address for
2001 		 * vm_map_lookup_entry to return.
2002 		 */
2003 		if (map->timestamp != last_timestamp)
2004 			vm_map_lookup_entry(map, last_end - 1, &entry);
2005 	}
2006 
2007 	vm_map_unlock_read(map);
2008 	vmspace_free(vm);
2009 	return;
2010 }
2011 
2012 /*
2013  * Log mappings for all processes in the system.
2014  */
2015 
2016 static void
2017 pmc_log_all_process_mappings(struct pmc_owner *po)
2018 {
2019 	struct proc *p, *top;
2020 
2021 	sx_assert(&pmc_sx, SX_XLOCKED);
2022 
2023 	if ((p = pfind(1)) == NULL)
2024 		panic("[pmc,%d] Cannot find init", __LINE__);
2025 
2026 	PROC_UNLOCK(p);
2027 
2028 	sx_slock(&proctree_lock);
2029 
2030 	top = p;
2031 
2032 	for (;;) {
2033 		pmc_log_process_mappings(po, p);
2034 		if (!LIST_EMPTY(&p->p_children))
2035 			p = LIST_FIRST(&p->p_children);
2036 		else for (;;) {
2037 			if (p == top)
2038 				goto done;
2039 			if (LIST_NEXT(p, p_sibling)) {
2040 				p = LIST_NEXT(p, p_sibling);
2041 				break;
2042 			}
2043 			p = p->p_pptr;
2044 		}
2045 	}
2046  done:
2047 	sx_sunlock(&proctree_lock);
2048 }
2049 
2050 /*
2051  * The 'hook' invoked from the kernel proper
2052  */
2053 
2054 
2055 #ifdef	HWPMC_DEBUG
2056 const char *pmc_hooknames[] = {
2057 	/* these strings correspond to PMC_FN_* in <sys/pmckern.h> */
2058 	"",
2059 	"EXEC",
2060 	"CSW-IN",
2061 	"CSW-OUT",
2062 	"SAMPLE",
2063 	"UNUSED1",
2064 	"UNUSED2",
2065 	"MMAP",
2066 	"MUNMAP",
2067 	"CALLCHAIN-NMI",
2068 	"CALLCHAIN-SOFT",
2069 	"SOFTSAMPLING",
2070 	"THR-CREATE",
2071 	"THR-EXIT",
2072 	"THR-USERRET",
2073 	"THR-CREATE-LOG",
2074 	"THR-EXIT-LOG",
2075 	"PROC-CREATE-LOG"
2076 };
2077 #endif
2078 
2079 static int
2080 pmc_hook_handler(struct thread *td, int function, void *arg)
2081 {
2082 	int cpu;
2083 
2084 	PMCDBG4(MOD,PMH,1, "hook td=%p func=%d \"%s\" arg=%p", td, function,
2085 	    pmc_hooknames[function], arg);
2086 
2087 	switch (function)
2088 	{
2089 
2090 	/*
2091 	 * Process exec()
2092 	 */
2093 
2094 	case PMC_FN_PROCESS_EXEC:
2095 	{
2096 		char *fullpath, *freepath;
2097 		unsigned int ri;
2098 		int is_using_hwpmcs;
2099 		struct pmc *pm;
2100 		struct proc *p;
2101 		struct pmc_owner *po;
2102 		struct pmc_process *pp;
2103 		struct pmckern_procexec *pk;
2104 
2105 		sx_assert(&pmc_sx, SX_XLOCKED);
2106 
2107 		p = td->td_proc;
2108 		pmc_getfilename(p->p_textvp, &fullpath, &freepath);
2109 
2110 		pk = (struct pmckern_procexec *) arg;
2111 
2112 		PMC_EPOCH_ENTER();
2113 		/* Inform owners of SS mode PMCs of the exec event. */
2114 		CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
2115 		    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
2116 			    pmclog_process_procexec(po, PMC_ID_INVALID,
2117 				p->p_pid, pk->pm_entryaddr, fullpath);
2118 		PMC_EPOCH_EXIT();
2119 
2120 		PROC_LOCK(p);
2121 		is_using_hwpmcs = p->p_flag & P_HWPMC;
2122 		PROC_UNLOCK(p);
2123 
2124 		if (!is_using_hwpmcs) {
2125 			if (freepath)
2126 				free(freepath, M_TEMP);
2127 			break;
2128 		}
2129 
2130 		/*
2131 		 * PMCs are not inherited across an exec():  remove any
2132 		 * PMCs that this process is the owner of.
2133 		 */
2134 
2135 		if ((po = pmc_find_owner_descriptor(p)) != NULL) {
2136 			pmc_remove_owner(po);
2137 			pmc_destroy_owner_descriptor(po);
2138 		}
2139 
2140 		/*
2141 		 * If the process being exec'ed is not the target of any
2142 		 * PMC, we are done.
2143 		 */
2144 		if ((pp = pmc_find_process_descriptor(p, 0)) == NULL) {
2145 			if (freepath)
2146 				free(freepath, M_TEMP);
2147 			break;
2148 		}
2149 
2150 		/*
2151 		 * Log the exec event to all monitoring owners.  Skip
2152 		 * owners who have already received the event because
2153 		 * they had system sampling PMCs active.
2154 		 */
2155 		for (ri = 0; ri < md->pmd_npmc; ri++)
2156 			if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) {
2157 				po = pm->pm_owner;
2158 				if (po->po_sscount == 0 &&
2159 				    po->po_flags & PMC_PO_OWNS_LOGFILE)
2160 					pmclog_process_procexec(po, pm->pm_id,
2161 					    p->p_pid, pk->pm_entryaddr,
2162 					    fullpath);
2163 			}
2164 
2165 		if (freepath)
2166 			free(freepath, M_TEMP);
2167 
2168 
2169 		PMCDBG4(PRC,EXC,1, "exec proc=%p (%d, %s) cred-changed=%d",
2170 		    p, p->p_pid, p->p_comm, pk->pm_credentialschanged);
2171 
2172 		if (pk->pm_credentialschanged == 0) /* no change */
2173 			break;
2174 
2175 		/*
2176 		 * If the newly exec()'ed process has a different credential
2177 		 * than before, allow it to be the target of a PMC only if
2178 		 * the PMC's owner has sufficient privilege.
2179 		 */
2180 
2181 		for (ri = 0; ri < md->pmd_npmc; ri++)
2182 			if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL)
2183 				if (pmc_can_attach(pm, td->td_proc) != 0)
2184 					pmc_detach_one_process(td->td_proc,
2185 					    pm, PMC_FLAG_NONE);
2186 
2187 		KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= (int) md->pmd_npmc,
2188 		    ("[pmc,%d] Illegal ref count %d on pp %p", __LINE__,
2189 			pp->pp_refcnt, pp));
2190 
2191 		/*
2192 		 * If this process is no longer the target of any
2193 		 * PMCs, we can remove the process entry and free
2194 		 * up space.
2195 		 */
2196 
2197 		if (pp->pp_refcnt == 0) {
2198 			pmc_remove_process_descriptor(pp);
2199 			pmc_destroy_process_descriptor(pp);
2200 			break;
2201 		}
2202 
2203 	}
2204 	break;
2205 
2206 	case PMC_FN_CSW_IN:
2207 		pmc_process_csw_in(td);
2208 		break;
2209 
2210 	case PMC_FN_CSW_OUT:
2211 		pmc_process_csw_out(td);
2212 		break;
2213 
2214 	/*
2215 	 * Process accumulated PC samples.
2216 	 *
2217 	 * This function is expected to be called by hardclock() for
2218 	 * each CPU that has accumulated PC samples.
2219 	 *
2220 	 * This function is to be executed on the CPU whose samples
2221 	 * are being processed.
2222 	 */
2223 	case PMC_FN_DO_SAMPLES:
2224 
2225 		/*
2226 		 * Clear the cpu specific bit in the CPU mask before
2227 		 * do the rest of the processing.  If the NMI handler
2228 		 * gets invoked after the "atomic_clear_int()" call
2229 		 * below but before "pmc_process_samples()" gets
2230 		 * around to processing the interrupt, then we will
2231 		 * come back here at the next hardclock() tick (and
2232 		 * may find nothing to do if "pmc_process_samples()"
2233 		 * had already processed the interrupt).  We don't
2234 		 * lose the interrupt sample.
2235 		 */
2236 		DPCPU_SET(pmc_sampled, 0);
2237 		cpu = PCPU_GET(cpuid);
2238 		pmc_process_samples(cpu, PMC_HR);
2239 		pmc_process_samples(cpu, PMC_SR);
2240 		pmc_process_samples(cpu, PMC_UR);
2241 		break;
2242 
2243 	case PMC_FN_MMAP:
2244 		pmc_process_mmap(td, (struct pmckern_map_in *) arg);
2245 		break;
2246 
2247 	case PMC_FN_MUNMAP:
2248 		MPASS(in_epoch(global_epoch_preempt) || sx_xlocked(&pmc_sx));
2249 		pmc_process_munmap(td, (struct pmckern_map_out *) arg);
2250 		break;
2251 
2252 	case PMC_FN_PROC_CREATE_LOG:
2253 		pmc_process_proccreate((struct proc *)arg);
2254 		break;
2255 
2256 	case PMC_FN_USER_CALLCHAIN:
2257 		/*
2258 		 * Record a call chain.
2259 		 */
2260 		KASSERT(td == curthread, ("[pmc,%d] td != curthread",
2261 		    __LINE__));
2262 
2263 		pmc_capture_user_callchain(PCPU_GET(cpuid), PMC_HR,
2264 		    (struct trapframe *) arg);
2265 
2266 		KASSERT(td->td_pinned == 1,
2267 			("[pmc,%d] invalid td_pinned value", __LINE__));
2268 		sched_unpin();  /* Can migrate safely now. */
2269 
2270 		td->td_pflags &= ~TDP_CALLCHAIN;
2271 		break;
2272 
2273 	case PMC_FN_USER_CALLCHAIN_SOFT:
2274 		/*
2275 		 * Record a call chain.
2276 		 */
2277 		KASSERT(td == curthread, ("[pmc,%d] td != curthread",
2278 		    __LINE__));
2279 
2280 		cpu = PCPU_GET(cpuid);
2281 		pmc_capture_user_callchain(cpu, PMC_SR,
2282 		    (struct trapframe *) arg);
2283 
2284 		KASSERT(td->td_pinned == 1,
2285 		    ("[pmc,%d] invalid td_pinned value", __LINE__));
2286 
2287 		sched_unpin();  /* Can migrate safely now. */
2288 
2289 		td->td_pflags &= ~TDP_CALLCHAIN;
2290 		break;
2291 
2292 	case PMC_FN_SOFT_SAMPLING:
2293 		/*
2294 		 * Call soft PMC sampling intr.
2295 		 */
2296 		pmc_soft_intr((struct pmckern_soft *) arg);
2297 		break;
2298 
2299 	case PMC_FN_THR_CREATE:
2300 		pmc_process_thread_add(td);
2301 		pmc_process_threadcreate(td);
2302 		break;
2303 
2304 	case PMC_FN_THR_CREATE_LOG:
2305 		pmc_process_threadcreate(td);
2306 		break;
2307 
2308 	case PMC_FN_THR_EXIT:
2309 		KASSERT(td == curthread, ("[pmc,%d] td != curthread",
2310 		    __LINE__));
2311 		pmc_process_thread_delete(td);
2312 		pmc_process_threadexit(td);
2313 		break;
2314 	case PMC_FN_THR_EXIT_LOG:
2315 		pmc_process_threadexit(td);
2316 		break;
2317 	case PMC_FN_THR_USERRET:
2318 		KASSERT(td == curthread, ("[pmc,%d] td != curthread",
2319 		    __LINE__));
2320 		pmc_process_thread_userret(td);
2321 		break;
2322 
2323 	default:
2324 #ifdef	HWPMC_DEBUG
2325 		KASSERT(0, ("[pmc,%d] unknown hook %d\n", __LINE__, function));
2326 #endif
2327 		break;
2328 
2329 	}
2330 
2331 	return 0;
2332 }
2333 
2334 /*
2335  * allocate a 'struct pmc_owner' descriptor in the owner hash table.
2336  */
2337 
2338 static struct pmc_owner *
2339 pmc_allocate_owner_descriptor(struct proc *p)
2340 {
2341 	uint32_t hindex;
2342 	struct pmc_owner *po;
2343 	struct pmc_ownerhash *poh;
2344 
2345 	hindex = PMC_HASH_PTR(p, pmc_ownerhashmask);
2346 	poh = &pmc_ownerhash[hindex];
2347 
2348 	/* allocate space for N pointers and one descriptor struct */
2349 	po = malloc(sizeof(struct pmc_owner), M_PMC, M_WAITOK|M_ZERO);
2350 	po->po_owner = p;
2351 	LIST_INSERT_HEAD(poh, po, po_next); /* insert into hash table */
2352 
2353 	TAILQ_INIT(&po->po_logbuffers);
2354 	mtx_init(&po->po_mtx, "pmc-owner-mtx", "pmc-per-proc", MTX_SPIN);
2355 
2356 	PMCDBG4(OWN,ALL,1, "allocate-owner proc=%p (%d, %s) pmc-owner=%p",
2357 	    p, p->p_pid, p->p_comm, po);
2358 
2359 	return po;
2360 }
2361 
2362 static void
2363 pmc_destroy_owner_descriptor(struct pmc_owner *po)
2364 {
2365 
2366 	PMCDBG4(OWN,REL,1, "destroy-owner po=%p proc=%p (%d, %s)",
2367 	    po, po->po_owner, po->po_owner->p_pid, po->po_owner->p_comm);
2368 
2369 	mtx_destroy(&po->po_mtx);
2370 	free(po, M_PMC);
2371 }
2372 
2373 /*
2374  * Allocate a thread descriptor from the free pool.
2375  *
2376  * NOTE: This *can* return NULL.
2377  */
2378 static struct pmc_thread *
2379 pmc_thread_descriptor_pool_alloc(void)
2380 {
2381 	struct pmc_thread *pt;
2382 
2383 	mtx_lock_spin(&pmc_threadfreelist_mtx);
2384 	if ((pt = LIST_FIRST(&pmc_threadfreelist)) != NULL) {
2385 		LIST_REMOVE(pt, pt_next);
2386 		pmc_threadfreelist_entries--;
2387 	}
2388 	mtx_unlock_spin(&pmc_threadfreelist_mtx);
2389 
2390 	return (pt);
2391 }
2392 
2393 /*
2394  * Add a thread descriptor to the free pool. We use this instead of free()
2395  * to maintain a cache of free entries. Additionally, we can safely call
2396  * this function when we cannot call free(), such as in a critical section.
2397  *
2398  */
2399 static void
2400 pmc_thread_descriptor_pool_free(struct pmc_thread *pt)
2401 {
2402 
2403 	if (pt == NULL)
2404 		return;
2405 
2406 	memset(pt, 0, THREADENTRY_SIZE);
2407 	mtx_lock_spin(&pmc_threadfreelist_mtx);
2408 	LIST_INSERT_HEAD(&pmc_threadfreelist, pt, pt_next);
2409 	pmc_threadfreelist_entries++;
2410 	if (pmc_threadfreelist_entries > pmc_threadfreelist_max)
2411 		GROUPTASK_ENQUEUE(&free_gtask);
2412 	mtx_unlock_spin(&pmc_threadfreelist_mtx);
2413 }
2414 
2415 /*
2416  * A callout to manage the free list.
2417  */
2418 static void
2419 pmc_thread_descriptor_pool_free_task(void *arg __unused)
2420 {
2421 	struct pmc_thread *pt;
2422 	LIST_HEAD(, pmc_thread) tmplist;
2423 	int delta;
2424 
2425 	LIST_INIT(&tmplist);
2426 	/* Determine what changes, if any, we need to make. */
2427 	mtx_lock_spin(&pmc_threadfreelist_mtx);
2428 	delta = pmc_threadfreelist_entries - pmc_threadfreelist_max;
2429 	while (delta > 0 &&
2430 		   (pt = LIST_FIRST(&pmc_threadfreelist)) != NULL) {
2431 		delta--;
2432 		LIST_REMOVE(pt, pt_next);
2433 		LIST_INSERT_HEAD(&tmplist, pt, pt_next);
2434 	}
2435 	mtx_unlock_spin(&pmc_threadfreelist_mtx);
2436 
2437 	/* If there are entries to free, free them. */
2438 	while (!LIST_EMPTY(&tmplist)) {
2439 		pt = LIST_FIRST(&tmplist);
2440 		LIST_REMOVE(pt, pt_next);
2441 		free(pt, M_PMC);
2442 	}
2443 }
2444 
2445 /*
2446  * Drain the thread free pool, freeing all allocations.
2447  */
2448 static void
2449 pmc_thread_descriptor_pool_drain()
2450 {
2451 	struct pmc_thread *pt, *next;
2452 
2453 	LIST_FOREACH_SAFE(pt, &pmc_threadfreelist, pt_next, next) {
2454 		LIST_REMOVE(pt, pt_next);
2455 		free(pt, M_PMC);
2456 	}
2457 }
2458 
2459 /*
2460  * find the descriptor corresponding to thread 'td', adding or removing it
2461  * as specified by 'mode'.
2462  *
2463  * Note that this supports additional mode flags in addition to those
2464  * supported by pmc_find_process_descriptor():
2465  * PMC_FLAG_NOWAIT: Causes the function to not wait for mallocs.
2466  *     This makes it safe to call while holding certain other locks.
2467  */
2468 
2469 static struct pmc_thread *
2470 pmc_find_thread_descriptor(struct pmc_process *pp, struct thread *td,
2471     uint32_t mode)
2472 {
2473 	struct pmc_thread *pt = NULL, *ptnew = NULL;
2474 	int wait_flag;
2475 
2476 	KASSERT(td != NULL, ("[pmc,%d] called to add NULL td", __LINE__));
2477 
2478 	/*
2479 	 * Pre-allocate memory in the PMC_FLAG_ALLOCATE case prior to
2480 	 * acquiring the lock.
2481 	 */
2482 	if (mode & PMC_FLAG_ALLOCATE) {
2483 		if ((ptnew = pmc_thread_descriptor_pool_alloc()) == NULL) {
2484 			wait_flag = M_WAITOK;
2485 			if ((mode & PMC_FLAG_NOWAIT) || in_epoch(global_epoch_preempt))
2486 				wait_flag = M_NOWAIT;
2487 
2488 			ptnew = malloc(THREADENTRY_SIZE, M_PMC,
2489 			    wait_flag|M_ZERO);
2490 		}
2491 	}
2492 
2493 	mtx_lock_spin(pp->pp_tdslock);
2494 
2495 	LIST_FOREACH(pt, &pp->pp_tds, pt_next)
2496 		if (pt->pt_td == td)
2497 			break;
2498 
2499 	if ((mode & PMC_FLAG_REMOVE) && pt != NULL)
2500 		LIST_REMOVE(pt, pt_next);
2501 
2502 	if ((mode & PMC_FLAG_ALLOCATE) && pt == NULL && ptnew != NULL) {
2503 		pt = ptnew;
2504 		ptnew = NULL;
2505 		pt->pt_td = td;
2506 		LIST_INSERT_HEAD(&pp->pp_tds, pt, pt_next);
2507 	}
2508 
2509 	mtx_unlock_spin(pp->pp_tdslock);
2510 
2511 	if (ptnew != NULL) {
2512 		free(ptnew, M_PMC);
2513 	}
2514 
2515 	return pt;
2516 }
2517 
2518 /*
2519  * Try to add thread descriptors for each thread in a process.
2520  */
2521 
2522 static void
2523 pmc_add_thread_descriptors_from_proc(struct proc *p, struct pmc_process *pp)
2524 {
2525 	struct thread *curtd;
2526 	struct pmc_thread **tdlist;
2527 	int i, tdcnt, tdlistsz;
2528 
2529 	KASSERT(!PROC_LOCKED(p), ("[pmc,%d] proc unexpectedly locked",
2530 	    __LINE__));
2531 	tdcnt = 32;
2532  restart:
2533 	tdlistsz = roundup2(tdcnt, 32);
2534 
2535 	tdcnt = 0;
2536 	tdlist = malloc(sizeof(struct pmc_thread*) * tdlistsz, M_TEMP, M_WAITOK);
2537 
2538 	PROC_LOCK(p);
2539 	FOREACH_THREAD_IN_PROC(p, curtd)
2540 		tdcnt++;
2541 	if (tdcnt >= tdlistsz) {
2542 		PROC_UNLOCK(p);
2543 		free(tdlist, M_TEMP);
2544 		goto restart;
2545 	}
2546 	/*
2547 	 * Try to add each thread to the list without sleeping. If unable,
2548 	 * add to a queue to retry after dropping the process lock.
2549 	 */
2550 	tdcnt = 0;
2551 	FOREACH_THREAD_IN_PROC(p, curtd) {
2552 		tdlist[tdcnt] = pmc_find_thread_descriptor(pp, curtd,
2553 						   PMC_FLAG_ALLOCATE|PMC_FLAG_NOWAIT);
2554 		if (tdlist[tdcnt] == NULL) {
2555 			PROC_UNLOCK(p);
2556 			for (i = 0; i <= tdcnt; i++)
2557 				pmc_thread_descriptor_pool_free(tdlist[i]);
2558 			free(tdlist, M_TEMP);
2559 			goto restart;
2560 		}
2561 		tdcnt++;
2562 	}
2563 	PROC_UNLOCK(p);
2564 	free(tdlist, M_TEMP);
2565 }
2566 
2567 /*
2568  * find the descriptor corresponding to process 'p', adding or removing it
2569  * as specified by 'mode'.
2570  */
2571 
2572 static struct pmc_process *
2573 pmc_find_process_descriptor(struct proc *p, uint32_t mode)
2574 {
2575 	uint32_t hindex;
2576 	struct pmc_process *pp, *ppnew;
2577 	struct pmc_processhash *pph;
2578 
2579 	hindex = PMC_HASH_PTR(p, pmc_processhashmask);
2580 	pph = &pmc_processhash[hindex];
2581 
2582 	ppnew = NULL;
2583 
2584 	/*
2585 	 * Pre-allocate memory in the PMC_FLAG_ALLOCATE case since we
2586 	 * cannot call malloc(9) once we hold a spin lock.
2587 	 */
2588 	if (mode & PMC_FLAG_ALLOCATE)
2589 		ppnew = malloc(sizeof(struct pmc_process) + md->pmd_npmc *
2590 		    sizeof(struct pmc_targetstate), M_PMC, M_WAITOK|M_ZERO);
2591 
2592 	mtx_lock_spin(&pmc_processhash_mtx);
2593 	LIST_FOREACH(pp, pph, pp_next)
2594 	    if (pp->pp_proc == p)
2595 		    break;
2596 
2597 	if ((mode & PMC_FLAG_REMOVE) && pp != NULL)
2598 		LIST_REMOVE(pp, pp_next);
2599 
2600 	if ((mode & PMC_FLAG_ALLOCATE) && pp == NULL &&
2601 	    ppnew != NULL) {
2602 		ppnew->pp_proc = p;
2603 		LIST_INIT(&ppnew->pp_tds);
2604 		ppnew->pp_tdslock = mtx_pool_find(pmc_mtxpool, ppnew);
2605 		LIST_INSERT_HEAD(pph, ppnew, pp_next);
2606 		mtx_unlock_spin(&pmc_processhash_mtx);
2607 		pp = ppnew;
2608 		ppnew = NULL;
2609 
2610 		/* Add thread descriptors for this process' current threads. */
2611 		pmc_add_thread_descriptors_from_proc(p, pp);
2612 	}
2613 	else
2614 		mtx_unlock_spin(&pmc_processhash_mtx);
2615 
2616 	if (ppnew != NULL)
2617 		free(ppnew, M_PMC);
2618 
2619 	return pp;
2620 }
2621 
2622 /*
2623  * remove a process descriptor from the process hash table.
2624  */
2625 
2626 static void
2627 pmc_remove_process_descriptor(struct pmc_process *pp)
2628 {
2629 	KASSERT(pp->pp_refcnt == 0,
2630 	    ("[pmc,%d] Removing process descriptor %p with count %d",
2631 		__LINE__, pp, pp->pp_refcnt));
2632 
2633 	mtx_lock_spin(&pmc_processhash_mtx);
2634 	LIST_REMOVE(pp, pp_next);
2635 	mtx_unlock_spin(&pmc_processhash_mtx);
2636 }
2637 
2638 /*
2639  * destroy a process descriptor.
2640  */
2641 
2642 static void
2643 pmc_destroy_process_descriptor(struct pmc_process *pp)
2644 {
2645 	struct pmc_thread *pmc_td;
2646 
2647 	while ((pmc_td = LIST_FIRST(&pp->pp_tds)) != NULL) {
2648 		LIST_REMOVE(pmc_td, pt_next);
2649 		pmc_thread_descriptor_pool_free(pmc_td);
2650 	}
2651 	free(pp, M_PMC);
2652 }
2653 
2654 
2655 /*
2656  * find an owner descriptor corresponding to proc 'p'
2657  */
2658 
2659 static struct pmc_owner *
2660 pmc_find_owner_descriptor(struct proc *p)
2661 {
2662 	uint32_t hindex;
2663 	struct pmc_owner *po;
2664 	struct pmc_ownerhash *poh;
2665 
2666 	hindex = PMC_HASH_PTR(p, pmc_ownerhashmask);
2667 	poh = &pmc_ownerhash[hindex];
2668 
2669 	po = NULL;
2670 	LIST_FOREACH(po, poh, po_next)
2671 	    if (po->po_owner == p)
2672 		    break;
2673 
2674 	PMCDBG5(OWN,FND,1, "find-owner proc=%p (%d, %s) hindex=0x%x -> "
2675 	    "pmc-owner=%p", p, p->p_pid, p->p_comm, hindex, po);
2676 
2677 	return po;
2678 }
2679 
2680 /*
2681  * pmc_allocate_pmc_descriptor
2682  *
2683  * Allocate a pmc descriptor and initialize its
2684  * fields.
2685  */
2686 
2687 static struct pmc *
2688 pmc_allocate_pmc_descriptor(void)
2689 {
2690 	struct pmc *pmc;
2691 
2692 	pmc = malloc(sizeof(struct pmc), M_PMC, M_WAITOK|M_ZERO);
2693 	pmc->pm_runcount = counter_u64_alloc(M_WAITOK);
2694 	pmc->pm_pcpu_state = malloc(sizeof(struct pmc_pcpu_state)*mp_ncpus, M_PMC, M_WAITOK|M_ZERO);
2695 	PMCDBG1(PMC,ALL,1, "allocate-pmc -> pmc=%p", pmc);
2696 
2697 	return pmc;
2698 }
2699 
2700 /*
2701  * Destroy a pmc descriptor.
2702  */
2703 
2704 static void
2705 pmc_destroy_pmc_descriptor(struct pmc *pm)
2706 {
2707 
2708 	KASSERT(pm->pm_state == PMC_STATE_DELETED ||
2709 	    pm->pm_state == PMC_STATE_FREE,
2710 	    ("[pmc,%d] destroying non-deleted PMC", __LINE__));
2711 	KASSERT(LIST_EMPTY(&pm->pm_targets),
2712 	    ("[pmc,%d] destroying pmc with targets", __LINE__));
2713 	KASSERT(pm->pm_owner == NULL,
2714 	    ("[pmc,%d] destroying pmc attached to an owner", __LINE__));
2715 	KASSERT(counter_u64_fetch(pm->pm_runcount) == 0,
2716 	    ("[pmc,%d] pmc has non-zero run count %ld", __LINE__,
2717 		 (unsigned long)counter_u64_fetch(pm->pm_runcount)));
2718 
2719 	counter_u64_free(pm->pm_runcount);
2720 	free(pm->pm_pcpu_state, M_PMC);
2721 	free(pm, M_PMC);
2722 }
2723 
2724 static void
2725 pmc_wait_for_pmc_idle(struct pmc *pm)
2726 {
2727 #ifdef HWPMC_DEBUG
2728 	volatile int maxloop;
2729 
2730 	maxloop = 100 * pmc_cpu_max();
2731 #endif
2732 	/*
2733 	 * Loop (with a forced context switch) till the PMC's runcount
2734 	 * comes down to zero.
2735 	 */
2736 	pmclog_flush(pm->pm_owner, 1);
2737 	while (counter_u64_fetch(pm->pm_runcount) > 0) {
2738 		pmclog_flush(pm->pm_owner, 1);
2739 #ifdef HWPMC_DEBUG
2740 		maxloop--;
2741 		KASSERT(maxloop > 0,
2742 		    ("[pmc,%d] (ri%d, rc%ld) waiting too long for "
2743 			"pmc to be free", __LINE__,
2744 			 PMC_TO_ROWINDEX(pm), (unsigned long)counter_u64_fetch(pm->pm_runcount)));
2745 #endif
2746 		pmc_force_context_switch();
2747 	}
2748 }
2749 
2750 /*
2751  * This function does the following things:
2752  *
2753  *  - detaches the PMC from hardware
2754  *  - unlinks all target threads that were attached to it
2755  *  - removes the PMC from its owner's list
2756  *  - destroys the PMC private mutex
2757  *
2758  * Once this function completes, the given pmc pointer can be freed by
2759  * calling pmc_destroy_pmc_descriptor().
2760  */
2761 
2762 static void
2763 pmc_release_pmc_descriptor(struct pmc *pm)
2764 {
2765 	enum pmc_mode mode;
2766 	struct pmc_hw *phw;
2767 	u_int adjri, ri, cpu;
2768 	struct pmc_owner *po;
2769 	struct pmc_binding pb;
2770 	struct pmc_process *pp;
2771 	struct pmc_classdep *pcd;
2772 	struct pmc_target *ptgt, *tmp;
2773 
2774 	sx_assert(&pmc_sx, SX_XLOCKED);
2775 
2776 	KASSERT(pm, ("[pmc,%d] null pmc", __LINE__));
2777 
2778 	ri   = PMC_TO_ROWINDEX(pm);
2779 	pcd  = pmc_ri_to_classdep(md, ri, &adjri);
2780 	mode = PMC_TO_MODE(pm);
2781 
2782 	PMCDBG3(PMC,REL,1, "release-pmc pmc=%p ri=%d mode=%d", pm, ri,
2783 	    mode);
2784 
2785 	/*
2786 	 * First, we take the PMC off hardware.
2787 	 */
2788 	cpu = 0;
2789 	if (PMC_IS_SYSTEM_MODE(mode)) {
2790 
2791 		/*
2792 		 * A system mode PMC runs on a specific CPU.  Switch
2793 		 * to this CPU and turn hardware off.
2794 		 */
2795 		pmc_save_cpu_binding(&pb);
2796 
2797 		cpu = PMC_TO_CPU(pm);
2798 
2799 		pmc_select_cpu(cpu);
2800 
2801 		/* switch off non-stalled CPUs */
2802 		pm->pm_pcpu_state[cpu].pps_cpustate = 0;
2803 		if (pm->pm_state == PMC_STATE_RUNNING &&
2804 			pm->pm_pcpu_state[cpu].pps_stalled == 0) {
2805 
2806 			phw = pmc_pcpu[cpu]->pc_hwpmcs[ri];
2807 
2808 			KASSERT(phw->phw_pmc == pm,
2809 			    ("[pmc, %d] pmc ptr ri(%d) hw(%p) pm(%p)",
2810 				__LINE__, ri, phw->phw_pmc, pm));
2811 			PMCDBG2(PMC,REL,2, "stopping cpu=%d ri=%d", cpu, ri);
2812 
2813 			critical_enter();
2814 			pcd->pcd_stop_pmc(cpu, adjri);
2815 			critical_exit();
2816 		}
2817 
2818 		PMCDBG2(PMC,REL,2, "decfg cpu=%d ri=%d", cpu, ri);
2819 
2820 		critical_enter();
2821 		pcd->pcd_config_pmc(cpu, adjri, NULL);
2822 		critical_exit();
2823 
2824 		/* adjust the global and process count of SS mode PMCs */
2825 		if (mode == PMC_MODE_SS && pm->pm_state == PMC_STATE_RUNNING) {
2826 			po = pm->pm_owner;
2827 			po->po_sscount--;
2828 			if (po->po_sscount == 0) {
2829 				atomic_subtract_rel_int(&pmc_ss_count, 1);
2830 				CK_LIST_REMOVE(po, po_ssnext);
2831 				epoch_wait_preempt(global_epoch_preempt);
2832 			}
2833 		}
2834 
2835 		pm->pm_state = PMC_STATE_DELETED;
2836 
2837 		pmc_restore_cpu_binding(&pb);
2838 
2839 		/*
2840 		 * We could have references to this PMC structure in
2841 		 * the per-cpu sample queues.  Wait for the queue to
2842 		 * drain.
2843 		 */
2844 		pmc_wait_for_pmc_idle(pm);
2845 
2846 	} else if (PMC_IS_VIRTUAL_MODE(mode)) {
2847 
2848 		/*
2849 		 * A virtual PMC could be running on multiple CPUs at
2850 		 * a given instant.
2851 		 *
2852 		 * By marking its state as DELETED, we ensure that
2853 		 * this PMC is never further scheduled on hardware.
2854 		 *
2855 		 * Then we wait till all CPUs are done with this PMC.
2856 		 */
2857 		pm->pm_state = PMC_STATE_DELETED;
2858 
2859 
2860 		/* Wait for the PMCs runcount to come to zero. */
2861 		pmc_wait_for_pmc_idle(pm);
2862 
2863 		/*
2864 		 * At this point the PMC is off all CPUs and cannot be
2865 		 * freshly scheduled onto a CPU.  It is now safe to
2866 		 * unlink all targets from this PMC.  If a
2867 		 * process-record's refcount falls to zero, we remove
2868 		 * it from the hash table.  The module-wide SX lock
2869 		 * protects us from races.
2870 		 */
2871 		LIST_FOREACH_SAFE(ptgt, &pm->pm_targets, pt_next, tmp) {
2872 			pp = ptgt->pt_process;
2873 			pmc_unlink_target_process(pm, pp); /* frees 'ptgt' */
2874 
2875 			PMCDBG1(PMC,REL,3, "pp->refcnt=%d", pp->pp_refcnt);
2876 
2877 			/*
2878 			 * If the target process record shows that no
2879 			 * PMCs are attached to it, reclaim its space.
2880 			 */
2881 
2882 			if (pp->pp_refcnt == 0) {
2883 				pmc_remove_process_descriptor(pp);
2884 				pmc_destroy_process_descriptor(pp);
2885 			}
2886 		}
2887 
2888 		cpu = curthread->td_oncpu; /* setup cpu for pmd_release() */
2889 
2890 	}
2891 
2892 	/*
2893 	 * Release any MD resources
2894 	 */
2895 	(void) pcd->pcd_release_pmc(cpu, adjri, pm);
2896 
2897 	/*
2898 	 * Update row disposition
2899 	 */
2900 
2901 	if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm)))
2902 		PMC_UNMARK_ROW_STANDALONE(ri);
2903 	else
2904 		PMC_UNMARK_ROW_THREAD(ri);
2905 
2906 	/* unlink from the owner's list */
2907 	if (pm->pm_owner) {
2908 		LIST_REMOVE(pm, pm_next);
2909 		pm->pm_owner = NULL;
2910 	}
2911 }
2912 
2913 /*
2914  * Register an owner and a pmc.
2915  */
2916 
2917 static int
2918 pmc_register_owner(struct proc *p, struct pmc *pmc)
2919 {
2920 	struct pmc_owner *po;
2921 
2922 	sx_assert(&pmc_sx, SX_XLOCKED);
2923 
2924 	if ((po = pmc_find_owner_descriptor(p)) == NULL)
2925 		if ((po = pmc_allocate_owner_descriptor(p)) == NULL)
2926 			return ENOMEM;
2927 
2928 	KASSERT(pmc->pm_owner == NULL,
2929 	    ("[pmc,%d] attempting to own an initialized PMC", __LINE__));
2930 	pmc->pm_owner  = po;
2931 
2932 	LIST_INSERT_HEAD(&po->po_pmcs, pmc, pm_next);
2933 
2934 	PROC_LOCK(p);
2935 	p->p_flag |= P_HWPMC;
2936 	PROC_UNLOCK(p);
2937 
2938 	if (po->po_flags & PMC_PO_OWNS_LOGFILE)
2939 		pmclog_process_pmcallocate(pmc);
2940 
2941 	PMCDBG2(PMC,REG,1, "register-owner pmc-owner=%p pmc=%p",
2942 	    po, pmc);
2943 
2944 	return 0;
2945 }
2946 
2947 /*
2948  * Return the current row disposition:
2949  * == 0 => FREE
2950  *  > 0 => PROCESS MODE
2951  *  < 0 => SYSTEM MODE
2952  */
2953 
2954 int
2955 pmc_getrowdisp(int ri)
2956 {
2957 	return pmc_pmcdisp[ri];
2958 }
2959 
2960 /*
2961  * Check if a PMC at row index 'ri' can be allocated to the current
2962  * process.
2963  *
2964  * Allocation can fail if:
2965  *   - the current process is already being profiled by a PMC at index 'ri',
2966  *     attached to it via OP_PMCATTACH.
2967  *   - the current process has already allocated a PMC at index 'ri'
2968  *     via OP_ALLOCATE.
2969  */
2970 
2971 static int
2972 pmc_can_allocate_rowindex(struct proc *p, unsigned int ri, int cpu)
2973 {
2974 	enum pmc_mode mode;
2975 	struct pmc *pm;
2976 	struct pmc_owner *po;
2977 	struct pmc_process *pp;
2978 
2979 	PMCDBG5(PMC,ALR,1, "can-allocate-rowindex proc=%p (%d, %s) ri=%d "
2980 	    "cpu=%d", p, p->p_pid, p->p_comm, ri, cpu);
2981 
2982 	/*
2983 	 * We shouldn't have already allocated a process-mode PMC at
2984 	 * row index 'ri'.
2985 	 *
2986 	 * We shouldn't have allocated a system-wide PMC on the same
2987 	 * CPU and same RI.
2988 	 */
2989 	if ((po = pmc_find_owner_descriptor(p)) != NULL)
2990 		LIST_FOREACH(pm, &po->po_pmcs, pm_next) {
2991 		    if (PMC_TO_ROWINDEX(pm) == ri) {
2992 			    mode = PMC_TO_MODE(pm);
2993 			    if (PMC_IS_VIRTUAL_MODE(mode))
2994 				    return EEXIST;
2995 			    if (PMC_IS_SYSTEM_MODE(mode) &&
2996 				(int) PMC_TO_CPU(pm) == cpu)
2997 				    return EEXIST;
2998 		    }
2999 	        }
3000 
3001 	/*
3002 	 * We also shouldn't be the target of any PMC at this index
3003 	 * since otherwise a PMC_ATTACH to ourselves will fail.
3004 	 */
3005 	if ((pp = pmc_find_process_descriptor(p, 0)) != NULL)
3006 		if (pp->pp_pmcs[ri].pp_pmc)
3007 			return EEXIST;
3008 
3009 	PMCDBG4(PMC,ALR,2, "can-allocate-rowindex proc=%p (%d, %s) ri=%d ok",
3010 	    p, p->p_pid, p->p_comm, ri);
3011 
3012 	return 0;
3013 }
3014 
3015 /*
3016  * Check if a given PMC at row index 'ri' can be currently used in
3017  * mode 'mode'.
3018  */
3019 
3020 static int
3021 pmc_can_allocate_row(int ri, enum pmc_mode mode)
3022 {
3023 	enum pmc_disp	disp;
3024 
3025 	sx_assert(&pmc_sx, SX_XLOCKED);
3026 
3027 	PMCDBG2(PMC,ALR,1, "can-allocate-row ri=%d mode=%d", ri, mode);
3028 
3029 	if (PMC_IS_SYSTEM_MODE(mode))
3030 		disp = PMC_DISP_STANDALONE;
3031 	else
3032 		disp = PMC_DISP_THREAD;
3033 
3034 	/*
3035 	 * check disposition for PMC row 'ri':
3036 	 *
3037 	 * Expected disposition		Row-disposition		Result
3038 	 *
3039 	 * STANDALONE			STANDALONE or FREE	proceed
3040 	 * STANDALONE			THREAD			fail
3041 	 * THREAD			THREAD or FREE		proceed
3042 	 * THREAD			STANDALONE		fail
3043 	 */
3044 
3045 	if (!PMC_ROW_DISP_IS_FREE(ri) &&
3046 	    !(disp == PMC_DISP_THREAD && PMC_ROW_DISP_IS_THREAD(ri)) &&
3047 	    !(disp == PMC_DISP_STANDALONE && PMC_ROW_DISP_IS_STANDALONE(ri)))
3048 		return EBUSY;
3049 
3050 	/*
3051 	 * All OK
3052 	 */
3053 
3054 	PMCDBG2(PMC,ALR,2, "can-allocate-row ri=%d mode=%d ok", ri, mode);
3055 
3056 	return 0;
3057 
3058 }
3059 
3060 /*
3061  * Find a PMC descriptor with user handle 'pmcid' for thread 'td'.
3062  */
3063 
3064 static struct pmc *
3065 pmc_find_pmc_descriptor_in_process(struct pmc_owner *po, pmc_id_t pmcid)
3066 {
3067 	struct pmc *pm;
3068 
3069 	KASSERT(PMC_ID_TO_ROWINDEX(pmcid) < md->pmd_npmc,
3070 	    ("[pmc,%d] Illegal pmc index %d (max %d)", __LINE__,
3071 		PMC_ID_TO_ROWINDEX(pmcid), md->pmd_npmc));
3072 
3073 	LIST_FOREACH(pm, &po->po_pmcs, pm_next)
3074 	    if (pm->pm_id == pmcid)
3075 		    return pm;
3076 
3077 	return NULL;
3078 }
3079 
3080 static int
3081 pmc_find_pmc(pmc_id_t pmcid, struct pmc **pmc)
3082 {
3083 
3084 	struct pmc *pm, *opm;
3085 	struct pmc_owner *po;
3086 	struct pmc_process *pp;
3087 
3088 	PMCDBG1(PMC,FND,1, "find-pmc id=%d", pmcid);
3089 	if (PMC_ID_TO_ROWINDEX(pmcid) >= md->pmd_npmc)
3090 		return (EINVAL);
3091 
3092 	if ((po = pmc_find_owner_descriptor(curthread->td_proc)) == NULL) {
3093 		/*
3094 		 * In case of PMC_F_DESCENDANTS child processes we will not find
3095 		 * the current process in the owners hash list.  Find the owner
3096 		 * process first and from there lookup the po.
3097 		 */
3098 		if ((pp = pmc_find_process_descriptor(curthread->td_proc,
3099 		    PMC_FLAG_NONE)) == NULL) {
3100 			return ESRCH;
3101 		} else {
3102 			opm = pp->pp_pmcs[PMC_ID_TO_ROWINDEX(pmcid)].pp_pmc;
3103 			if (opm == NULL)
3104 				return ESRCH;
3105 			if ((opm->pm_flags & (PMC_F_ATTACHED_TO_OWNER|
3106 			    PMC_F_DESCENDANTS)) != (PMC_F_ATTACHED_TO_OWNER|
3107 			    PMC_F_DESCENDANTS))
3108 				return ESRCH;
3109 			po = opm->pm_owner;
3110 		}
3111 	}
3112 
3113 	if ((pm = pmc_find_pmc_descriptor_in_process(po, pmcid)) == NULL)
3114 		return EINVAL;
3115 
3116 	PMCDBG2(PMC,FND,2, "find-pmc id=%d -> pmc=%p", pmcid, pm);
3117 
3118 	*pmc = pm;
3119 	return 0;
3120 }
3121 
3122 /*
3123  * Start a PMC.
3124  */
3125 
3126 static int
3127 pmc_start(struct pmc *pm)
3128 {
3129 	enum pmc_mode mode;
3130 	struct pmc_owner *po;
3131 	struct pmc_binding pb;
3132 	struct pmc_classdep *pcd;
3133 	int adjri, error, cpu, ri;
3134 
3135 	KASSERT(pm != NULL,
3136 	    ("[pmc,%d] null pm", __LINE__));
3137 
3138 	mode = PMC_TO_MODE(pm);
3139 	ri   = PMC_TO_ROWINDEX(pm);
3140 	pcd  = pmc_ri_to_classdep(md, ri, &adjri);
3141 
3142 	error = 0;
3143 
3144 	PMCDBG3(PMC,OPS,1, "start pmc=%p mode=%d ri=%d", pm, mode, ri);
3145 
3146 	po = pm->pm_owner;
3147 
3148 	/*
3149 	 * Disallow PMCSTART if a logfile is required but has not been
3150 	 * configured yet.
3151 	 */
3152 	if ((pm->pm_flags & PMC_F_NEEDS_LOGFILE) &&
3153 	    (po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
3154 		return (EDOOFUS);	/* programming error */
3155 
3156 	/*
3157 	 * If this is a sampling mode PMC, log mapping information for
3158 	 * the kernel modules that are currently loaded.
3159 	 */
3160 	if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
3161 	    pmc_log_kernel_mappings(pm);
3162 
3163 	if (PMC_IS_VIRTUAL_MODE(mode)) {
3164 
3165 		/*
3166 		 * If a PMCATTACH has never been done on this PMC,
3167 		 * attach it to its owner process.
3168 		 */
3169 
3170 		if (LIST_EMPTY(&pm->pm_targets))
3171 			error = (pm->pm_flags & PMC_F_ATTACH_DONE) ? ESRCH :
3172 			    pmc_attach_process(po->po_owner, pm);
3173 
3174 		/*
3175 		 * If the PMC is attached to its owner, then force a context
3176 		 * switch to ensure that the MD state gets set correctly.
3177 		 */
3178 
3179 		if (error == 0) {
3180 			pm->pm_state = PMC_STATE_RUNNING;
3181 			if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER)
3182 				pmc_force_context_switch();
3183 		}
3184 
3185 		return (error);
3186 	}
3187 
3188 
3189 	/*
3190 	 * A system-wide PMC.
3191 	 *
3192 	 * Add the owner to the global list if this is a system-wide
3193 	 * sampling PMC.
3194 	 */
3195 
3196 	if (mode == PMC_MODE_SS) {
3197 		/*
3198 		 * Log mapping information for all existing processes in the
3199 		 * system.  Subsequent mappings are logged as they happen;
3200 		 * see pmc_process_mmap().
3201 		 */
3202 		if (po->po_logprocmaps == 0) {
3203 			pmc_log_all_process_mappings(po);
3204 			po->po_logprocmaps = 1;
3205 		}
3206 		po->po_sscount++;
3207 		if (po->po_sscount == 1) {
3208 			atomic_add_rel_int(&pmc_ss_count, 1);
3209 			CK_LIST_INSERT_HEAD(&pmc_ss_owners, po, po_ssnext);
3210 			PMCDBG1(PMC,OPS,1, "po=%p in global list", po);
3211 		}
3212 	}
3213 
3214 	/*
3215 	 * Move to the CPU associated with this
3216 	 * PMC, and start the hardware.
3217 	 */
3218 
3219 	pmc_save_cpu_binding(&pb);
3220 
3221 	cpu = PMC_TO_CPU(pm);
3222 
3223 	if (!pmc_cpu_is_active(cpu))
3224 		return (ENXIO);
3225 
3226 	pmc_select_cpu(cpu);
3227 
3228 	/*
3229 	 * global PMCs are configured at allocation time
3230 	 * so write out the initial value and start the PMC.
3231 	 */
3232 
3233 	pm->pm_state = PMC_STATE_RUNNING;
3234 
3235 	critical_enter();
3236 	if ((error = pcd->pcd_write_pmc(cpu, adjri,
3237 		 PMC_IS_SAMPLING_MODE(mode) ?
3238 		 pm->pm_sc.pm_reloadcount :
3239 		 pm->pm_sc.pm_initial)) == 0) {
3240 		/* If a sampling mode PMC, reset stalled state. */
3241 		if (PMC_IS_SAMPLING_MODE(mode))
3242 			pm->pm_pcpu_state[cpu].pps_stalled = 0;
3243 
3244 		/* Indicate that we desire this to run. Start it. */
3245 		pm->pm_pcpu_state[cpu].pps_cpustate = 1;
3246 		error = pcd->pcd_start_pmc(cpu, adjri);
3247 	}
3248 	critical_exit();
3249 
3250 	pmc_restore_cpu_binding(&pb);
3251 
3252 	return (error);
3253 }
3254 
3255 /*
3256  * Stop a PMC.
3257  */
3258 
3259 static int
3260 pmc_stop(struct pmc *pm)
3261 {
3262 	struct pmc_owner *po;
3263 	struct pmc_binding pb;
3264 	struct pmc_classdep *pcd;
3265 	int adjri, cpu, error, ri;
3266 
3267 	KASSERT(pm != NULL, ("[pmc,%d] null pmc", __LINE__));
3268 
3269 	PMCDBG3(PMC,OPS,1, "stop pmc=%p mode=%d ri=%d", pm,
3270 	    PMC_TO_MODE(pm), PMC_TO_ROWINDEX(pm));
3271 
3272 	pm->pm_state = PMC_STATE_STOPPED;
3273 
3274 	/*
3275 	 * If the PMC is a virtual mode one, changing the state to
3276 	 * non-RUNNING is enough to ensure that the PMC never gets
3277 	 * scheduled.
3278 	 *
3279 	 * If this PMC is current running on a CPU, then it will
3280 	 * handled correctly at the time its target process is context
3281 	 * switched out.
3282 	 */
3283 
3284 	if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)))
3285 		return 0;
3286 
3287 	/*
3288 	 * A system-mode PMC.  Move to the CPU associated with
3289 	 * this PMC, and stop the hardware.  We update the
3290 	 * 'initial count' so that a subsequent PMCSTART will
3291 	 * resume counting from the current hardware count.
3292 	 */
3293 
3294 	pmc_save_cpu_binding(&pb);
3295 
3296 	cpu = PMC_TO_CPU(pm);
3297 
3298 	KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
3299 	    ("[pmc,%d] illegal cpu=%d", __LINE__, cpu));
3300 
3301 	if (!pmc_cpu_is_active(cpu))
3302 		return ENXIO;
3303 
3304 	pmc_select_cpu(cpu);
3305 
3306 	ri = PMC_TO_ROWINDEX(pm);
3307 	pcd = pmc_ri_to_classdep(md, ri, &adjri);
3308 
3309 	pm->pm_pcpu_state[cpu].pps_cpustate = 0;
3310 	critical_enter();
3311 	if ((error = pcd->pcd_stop_pmc(cpu, adjri)) == 0)
3312 		error = pcd->pcd_read_pmc(cpu, adjri, &pm->pm_sc.pm_initial);
3313 	critical_exit();
3314 
3315 	pmc_restore_cpu_binding(&pb);
3316 
3317 	po = pm->pm_owner;
3318 
3319 	/* remove this owner from the global list of SS PMC owners */
3320 	if (PMC_TO_MODE(pm) == PMC_MODE_SS) {
3321 		po->po_sscount--;
3322 		if (po->po_sscount == 0) {
3323 			atomic_subtract_rel_int(&pmc_ss_count, 1);
3324 			CK_LIST_REMOVE(po, po_ssnext);
3325 			epoch_wait_preempt(global_epoch_preempt);
3326 			PMCDBG1(PMC,OPS,2,"po=%p removed from global list", po);
3327 		}
3328 	}
3329 
3330 	return (error);
3331 }
3332 
3333 static struct pmc_classdep *
3334 pmc_class_to_classdep(enum pmc_class class)
3335 {
3336 	int n;
3337 
3338 	for (n = 0; n < md->pmd_nclass; n++)
3339 		if (md->pmd_classdep[n].pcd_class == class)
3340 			return (&md->pmd_classdep[n]);
3341 	return (NULL);
3342 }
3343 
3344 #if defined(HWPMC_DEBUG) && defined(KTR)
3345 static const char *pmc_op_to_name[] = {
3346 #undef	__PMC_OP
3347 #define	__PMC_OP(N, D)	#N ,
3348 	__PMC_OPS()
3349 	NULL
3350 };
3351 #endif
3352 
3353 /*
3354  * The syscall interface
3355  */
3356 
3357 #define	PMC_GET_SX_XLOCK(...) do {		\
3358 	sx_xlock(&pmc_sx);			\
3359 	if (pmc_hook == NULL) {			\
3360 		sx_xunlock(&pmc_sx);		\
3361 		return __VA_ARGS__;		\
3362 	}					\
3363 } while (0)
3364 
3365 #define	PMC_DOWNGRADE_SX() do {			\
3366 	sx_downgrade(&pmc_sx);			\
3367 	is_sx_downgraded = 1;			\
3368 } while (0)
3369 
3370 static int
3371 pmc_syscall_handler(struct thread *td, void *syscall_args)
3372 {
3373 	int error, is_sx_downgraded, op;
3374 	struct pmc_syscall_args *c;
3375 	void *pmclog_proc_handle;
3376 	void *arg;
3377 
3378 	c = (struct pmc_syscall_args *)syscall_args;
3379 	op = c->pmop_code;
3380 	arg = c->pmop_data;
3381 	/* PMC isn't set up yet */
3382 	if (pmc_hook == NULL)
3383 		return (EINVAL);
3384 	if (op == PMC_OP_CONFIGURELOG) {
3385 		/*
3386 		 * We cannot create the logging process inside
3387 		 * pmclog_configure_log() because there is a LOR
3388 		 * between pmc_sx and process structure locks.
3389 		 * Instead, pre-create the process and ignite the loop
3390 		 * if everything is fine, otherwise direct the process
3391 		 * to exit.
3392 		 */
3393 		error = pmclog_proc_create(td, &pmclog_proc_handle);
3394 		if (error != 0)
3395 			goto done_syscall;
3396 	}
3397 
3398 	PMC_GET_SX_XLOCK(ENOSYS);
3399 	is_sx_downgraded = 0;
3400 	PMCDBG3(MOD,PMS,1, "syscall op=%d \"%s\" arg=%p", op,
3401 	    pmc_op_to_name[op], arg);
3402 
3403 	error = 0;
3404 	counter_u64_add(pmc_stats.pm_syscalls, 1);
3405 
3406 	switch (op) {
3407 
3408 
3409 	/*
3410 	 * Configure a log file.
3411 	 *
3412 	 * XXX This OP will be reworked.
3413 	 */
3414 
3415 	case PMC_OP_CONFIGURELOG:
3416 	{
3417 		struct proc *p;
3418 		struct pmc *pm;
3419 		struct pmc_owner *po;
3420 		struct pmc_op_configurelog cl;
3421 
3422 		if ((error = copyin(arg, &cl, sizeof(cl))) != 0) {
3423 			pmclog_proc_ignite(pmclog_proc_handle, NULL);
3424 			break;
3425 		}
3426 
3427 		/* mark this process as owning a log file */
3428 		p = td->td_proc;
3429 		if ((po = pmc_find_owner_descriptor(p)) == NULL)
3430 			if ((po = pmc_allocate_owner_descriptor(p)) == NULL) {
3431 				pmclog_proc_ignite(pmclog_proc_handle, NULL);
3432 				error = ENOMEM;
3433 				break;
3434 			}
3435 
3436 		/*
3437 		 * If a valid fd was passed in, try to configure that,
3438 		 * otherwise if 'fd' was less than zero and there was
3439 		 * a log file configured, flush its buffers and
3440 		 * de-configure it.
3441 		 */
3442 		if (cl.pm_logfd >= 0) {
3443 			error = pmclog_configure_log(md, po, cl.pm_logfd);
3444 			pmclog_proc_ignite(pmclog_proc_handle, error == 0 ?
3445 			    po : NULL);
3446 		} else if (po->po_flags & PMC_PO_OWNS_LOGFILE) {
3447 			pmclog_proc_ignite(pmclog_proc_handle, NULL);
3448 			error = pmclog_close(po);
3449 			if (error == 0) {
3450 				LIST_FOREACH(pm, &po->po_pmcs, pm_next)
3451 				    if (pm->pm_flags & PMC_F_NEEDS_LOGFILE &&
3452 					pm->pm_state == PMC_STATE_RUNNING)
3453 					    pmc_stop(pm);
3454 				error = pmclog_deconfigure_log(po);
3455 			}
3456 		} else {
3457 			pmclog_proc_ignite(pmclog_proc_handle, NULL);
3458 			error = EINVAL;
3459 		}
3460 	}
3461 	break;
3462 
3463 	/*
3464 	 * Flush a log file.
3465 	 */
3466 
3467 	case PMC_OP_FLUSHLOG:
3468 	{
3469 		struct pmc_owner *po;
3470 
3471 		sx_assert(&pmc_sx, SX_XLOCKED);
3472 
3473 		if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) {
3474 			error = EINVAL;
3475 			break;
3476 		}
3477 
3478 		error = pmclog_flush(po, 0);
3479 	}
3480 	break;
3481 
3482 	/*
3483 	 * Close a log file.
3484 	 */
3485 
3486 	case PMC_OP_CLOSELOG:
3487 	{
3488 		struct pmc_owner *po;
3489 
3490 		sx_assert(&pmc_sx, SX_XLOCKED);
3491 
3492 		if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) {
3493 			error = EINVAL;
3494 			break;
3495 		}
3496 
3497 		error = pmclog_close(po);
3498 	}
3499 	break;
3500 
3501 	/*
3502 	 * Retrieve hardware configuration.
3503 	 */
3504 
3505 	case PMC_OP_GETCPUINFO:	/* CPU information */
3506 	{
3507 		struct pmc_op_getcpuinfo gci;
3508 		struct pmc_classinfo *pci;
3509 		struct pmc_classdep *pcd;
3510 		int cl;
3511 
3512 		gci.pm_cputype = md->pmd_cputype;
3513 		gci.pm_ncpu    = pmc_cpu_max();
3514 		gci.pm_npmc    = md->pmd_npmc;
3515 		gci.pm_nclass  = md->pmd_nclass;
3516 		pci = gci.pm_classes;
3517 		pcd = md->pmd_classdep;
3518 		for (cl = 0; cl < md->pmd_nclass; cl++, pci++, pcd++) {
3519 			pci->pm_caps  = pcd->pcd_caps;
3520 			pci->pm_class = pcd->pcd_class;
3521 			pci->pm_width = pcd->pcd_width;
3522 			pci->pm_num   = pcd->pcd_num;
3523 		}
3524 		error = copyout(&gci, arg, sizeof(gci));
3525 	}
3526 	break;
3527 
3528 	/*
3529 	 * Retrieve soft events list.
3530 	 */
3531 	case PMC_OP_GETDYNEVENTINFO:
3532 	{
3533 		enum pmc_class			cl;
3534 		enum pmc_event			ev;
3535 		struct pmc_op_getdyneventinfo	*gei;
3536 		struct pmc_dyn_event_descr	dev;
3537 		struct pmc_soft			*ps;
3538 		uint32_t			nevent;
3539 
3540 		sx_assert(&pmc_sx, SX_LOCKED);
3541 
3542 		gei = (struct pmc_op_getdyneventinfo *) arg;
3543 
3544 		if ((error = copyin(&gei->pm_class, &cl, sizeof(cl))) != 0)
3545 			break;
3546 
3547 		/* Only SOFT class is dynamic. */
3548 		if (cl != PMC_CLASS_SOFT) {
3549 			error = EINVAL;
3550 			break;
3551 		}
3552 
3553 		nevent = 0;
3554 		for (ev = PMC_EV_SOFT_FIRST; (int)ev <= PMC_EV_SOFT_LAST; ev++) {
3555 			ps = pmc_soft_ev_acquire(ev);
3556 			if (ps == NULL)
3557 				continue;
3558 			bcopy(&ps->ps_ev, &dev, sizeof(dev));
3559 			pmc_soft_ev_release(ps);
3560 
3561 			error = copyout(&dev,
3562 			    &gei->pm_events[nevent],
3563 			    sizeof(struct pmc_dyn_event_descr));
3564 			if (error != 0)
3565 				break;
3566 			nevent++;
3567 		}
3568 		if (error != 0)
3569 			break;
3570 
3571 		error = copyout(&nevent, &gei->pm_nevent,
3572 		    sizeof(nevent));
3573 	}
3574 	break;
3575 
3576 	/*
3577 	 * Get module statistics
3578 	 */
3579 
3580 	case PMC_OP_GETDRIVERSTATS:
3581 	{
3582 		struct pmc_op_getdriverstats gms;
3583 #define CFETCH(a, b, field) a.field = counter_u64_fetch(b.field)
3584 		CFETCH(gms, pmc_stats, pm_intr_ignored);
3585 		CFETCH(gms, pmc_stats, pm_intr_processed);
3586 		CFETCH(gms, pmc_stats, pm_intr_bufferfull);
3587 		CFETCH(gms, pmc_stats, pm_syscalls);
3588 		CFETCH(gms, pmc_stats, pm_syscall_errors);
3589 		CFETCH(gms, pmc_stats, pm_buffer_requests);
3590 		CFETCH(gms, pmc_stats, pm_buffer_requests_failed);
3591 		CFETCH(gms, pmc_stats, pm_log_sweeps);
3592 #undef CFETCH
3593 		error = copyout(&gms, arg, sizeof(gms));
3594 	}
3595 	break;
3596 
3597 
3598 	/*
3599 	 * Retrieve module version number
3600 	 */
3601 
3602 	case PMC_OP_GETMODULEVERSION:
3603 	{
3604 		uint32_t cv, modv;
3605 
3606 		/* retrieve the client's idea of the ABI version */
3607 		if ((error = copyin(arg, &cv, sizeof(uint32_t))) != 0)
3608 			break;
3609 		/* don't service clients newer than our driver */
3610 		modv = PMC_VERSION;
3611 		if ((cv & 0xFFFF0000) > (modv & 0xFFFF0000)) {
3612 			error = EPROGMISMATCH;
3613 			break;
3614 		}
3615 		error = copyout(&modv, arg, sizeof(int));
3616 	}
3617 	break;
3618 
3619 
3620 	/*
3621 	 * Retrieve the state of all the PMCs on a given
3622 	 * CPU.
3623 	 */
3624 
3625 	case PMC_OP_GETPMCINFO:
3626 	{
3627 		int ari;
3628 		struct pmc *pm;
3629 		size_t pmcinfo_size;
3630 		uint32_t cpu, n, npmc;
3631 		struct pmc_owner *po;
3632 		struct pmc_binding pb;
3633 		struct pmc_classdep *pcd;
3634 		struct pmc_info *p, *pmcinfo;
3635 		struct pmc_op_getpmcinfo *gpi;
3636 
3637 		PMC_DOWNGRADE_SX();
3638 
3639 		gpi = (struct pmc_op_getpmcinfo *) arg;
3640 
3641 		if ((error = copyin(&gpi->pm_cpu, &cpu, sizeof(cpu))) != 0)
3642 			break;
3643 
3644 		if (cpu >= pmc_cpu_max()) {
3645 			error = EINVAL;
3646 			break;
3647 		}
3648 
3649 		if (!pmc_cpu_is_active(cpu)) {
3650 			error = ENXIO;
3651 			break;
3652 		}
3653 
3654 		/* switch to CPU 'cpu' */
3655 		pmc_save_cpu_binding(&pb);
3656 		pmc_select_cpu(cpu);
3657 
3658 		npmc = md->pmd_npmc;
3659 
3660 		pmcinfo_size = npmc * sizeof(struct pmc_info);
3661 		pmcinfo = malloc(pmcinfo_size, M_PMC, M_WAITOK);
3662 
3663 		p = pmcinfo;
3664 
3665 		for (n = 0; n < md->pmd_npmc; n++, p++) {
3666 
3667 			pcd = pmc_ri_to_classdep(md, n, &ari);
3668 
3669 			KASSERT(pcd != NULL,
3670 			    ("[pmc,%d] null pcd ri=%d", __LINE__, n));
3671 
3672 			if ((error = pcd->pcd_describe(cpu, ari, p, &pm)) != 0)
3673 				break;
3674 
3675 			if (PMC_ROW_DISP_IS_STANDALONE(n))
3676 				p->pm_rowdisp = PMC_DISP_STANDALONE;
3677 			else if (PMC_ROW_DISP_IS_THREAD(n))
3678 				p->pm_rowdisp = PMC_DISP_THREAD;
3679 			else
3680 				p->pm_rowdisp = PMC_DISP_FREE;
3681 
3682 			p->pm_ownerpid = -1;
3683 
3684 			if (pm == NULL)	/* no PMC associated */
3685 				continue;
3686 
3687 			po = pm->pm_owner;
3688 
3689 			KASSERT(po->po_owner != NULL,
3690 			    ("[pmc,%d] pmc_owner had a null proc pointer",
3691 				__LINE__));
3692 
3693 			p->pm_ownerpid = po->po_owner->p_pid;
3694 			p->pm_mode     = PMC_TO_MODE(pm);
3695 			p->pm_event    = pm->pm_event;
3696 			p->pm_flags    = pm->pm_flags;
3697 
3698 			if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
3699 				p->pm_reloadcount =
3700 				    pm->pm_sc.pm_reloadcount;
3701 		}
3702 
3703 		pmc_restore_cpu_binding(&pb);
3704 
3705 		/* now copy out the PMC info collected */
3706 		if (error == 0)
3707 			error = copyout(pmcinfo, &gpi->pm_pmcs, pmcinfo_size);
3708 
3709 		free(pmcinfo, M_PMC);
3710 	}
3711 	break;
3712 
3713 
3714 	/*
3715 	 * Set the administrative state of a PMC.  I.e. whether
3716 	 * the PMC is to be used or not.
3717 	 */
3718 
3719 	case PMC_OP_PMCADMIN:
3720 	{
3721 		int cpu, ri;
3722 		enum pmc_state request;
3723 		struct pmc_cpu *pc;
3724 		struct pmc_hw *phw;
3725 		struct pmc_op_pmcadmin pma;
3726 		struct pmc_binding pb;
3727 
3728 		sx_assert(&pmc_sx, SX_XLOCKED);
3729 
3730 		KASSERT(td == curthread,
3731 		    ("[pmc,%d] td != curthread", __LINE__));
3732 
3733 		error = priv_check(td, PRIV_PMC_MANAGE);
3734 		if (error)
3735 			break;
3736 
3737 		if ((error = copyin(arg, &pma, sizeof(pma))) != 0)
3738 			break;
3739 
3740 		cpu = pma.pm_cpu;
3741 
3742 		if (cpu < 0 || cpu >= (int) pmc_cpu_max()) {
3743 			error = EINVAL;
3744 			break;
3745 		}
3746 
3747 		if (!pmc_cpu_is_active(cpu)) {
3748 			error = ENXIO;
3749 			break;
3750 		}
3751 
3752 		request = pma.pm_state;
3753 
3754 		if (request != PMC_STATE_DISABLED &&
3755 		    request != PMC_STATE_FREE) {
3756 			error = EINVAL;
3757 			break;
3758 		}
3759 
3760 		ri = pma.pm_pmc; /* pmc id == row index */
3761 		if (ri < 0 || ri >= (int) md->pmd_npmc) {
3762 			error = EINVAL;
3763 			break;
3764 		}
3765 
3766 		/*
3767 		 * We can't disable a PMC with a row-index allocated
3768 		 * for process virtual PMCs.
3769 		 */
3770 
3771 		if (PMC_ROW_DISP_IS_THREAD(ri) &&
3772 		    request == PMC_STATE_DISABLED) {
3773 			error = EBUSY;
3774 			break;
3775 		}
3776 
3777 		/*
3778 		 * otherwise, this PMC on this CPU is either free or
3779 		 * in system-wide mode.
3780 		 */
3781 
3782 		pmc_save_cpu_binding(&pb);
3783 		pmc_select_cpu(cpu);
3784 
3785 		pc  = pmc_pcpu[cpu];
3786 		phw = pc->pc_hwpmcs[ri];
3787 
3788 		/*
3789 		 * XXX do we need some kind of 'forced' disable?
3790 		 */
3791 
3792 		if (phw->phw_pmc == NULL) {
3793 			if (request == PMC_STATE_DISABLED &&
3794 			    (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED)) {
3795 				phw->phw_state &= ~PMC_PHW_FLAG_IS_ENABLED;
3796 				PMC_MARK_ROW_STANDALONE(ri);
3797 			} else if (request == PMC_STATE_FREE &&
3798 			    (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0) {
3799 				phw->phw_state |=  PMC_PHW_FLAG_IS_ENABLED;
3800 				PMC_UNMARK_ROW_STANDALONE(ri);
3801 			}
3802 			/* other cases are a no-op */
3803 		} else
3804 			error = EBUSY;
3805 
3806 		pmc_restore_cpu_binding(&pb);
3807 	}
3808 	break;
3809 
3810 
3811 	/*
3812 	 * Allocate a PMC.
3813 	 */
3814 
3815 	case PMC_OP_PMCALLOCATE:
3816 	{
3817 		int adjri, n;
3818 		u_int cpu;
3819 		uint32_t caps;
3820 		struct pmc *pmc;
3821 		enum pmc_mode mode;
3822 		struct pmc_hw *phw;
3823 		struct pmc_binding pb;
3824 		struct pmc_classdep *pcd;
3825 		struct pmc_op_pmcallocate pa;
3826 
3827 		if ((error = copyin(arg, &pa, sizeof(pa))) != 0)
3828 			break;
3829 
3830 		caps = pa.pm_caps;
3831 		mode = pa.pm_mode;
3832 		cpu  = pa.pm_cpu;
3833 
3834 		if ((mode != PMC_MODE_SS  &&  mode != PMC_MODE_SC  &&
3835 		     mode != PMC_MODE_TS  &&  mode != PMC_MODE_TC) ||
3836 		    (cpu != (u_int) PMC_CPU_ANY && cpu >= pmc_cpu_max())) {
3837 			error = EINVAL;
3838 			break;
3839 		}
3840 
3841 		/*
3842 		 * Virtual PMCs should only ask for a default CPU.
3843 		 * System mode PMCs need to specify a non-default CPU.
3844 		 */
3845 
3846 		if ((PMC_IS_VIRTUAL_MODE(mode) && cpu != (u_int) PMC_CPU_ANY) ||
3847 		    (PMC_IS_SYSTEM_MODE(mode) && cpu == (u_int) PMC_CPU_ANY)) {
3848 			error = EINVAL;
3849 			break;
3850 		}
3851 
3852 		/*
3853 		 * Check that an inactive CPU is not being asked for.
3854 		 */
3855 
3856 		if (PMC_IS_SYSTEM_MODE(mode) && !pmc_cpu_is_active(cpu)) {
3857 			error = ENXIO;
3858 			break;
3859 		}
3860 
3861 		/*
3862 		 * Refuse an allocation for a system-wide PMC if this
3863 		 * process has been jailed, or if this process lacks
3864 		 * super-user credentials and the sysctl tunable
3865 		 * 'security.bsd.unprivileged_syspmcs' is zero.
3866 		 */
3867 
3868 		if (PMC_IS_SYSTEM_MODE(mode)) {
3869 			if (jailed(curthread->td_ucred)) {
3870 				error = EPERM;
3871 				break;
3872 			}
3873 			if (!pmc_unprivileged_syspmcs) {
3874 				error = priv_check(curthread,
3875 				    PRIV_PMC_SYSTEM);
3876 				if (error)
3877 					break;
3878 			}
3879 		}
3880 
3881 		/*
3882 		 * Look for valid values for 'pm_flags'
3883 		 */
3884 
3885 		if ((pa.pm_flags & ~(PMC_F_DESCENDANTS | PMC_F_LOG_PROCCSW |
3886 		    PMC_F_LOG_PROCEXIT | PMC_F_CALLCHAIN |
3887 		    PMC_F_USERCALLCHAIN)) != 0) {
3888 			error = EINVAL;
3889 			break;
3890 		}
3891 
3892 		/* PMC_F_USERCALLCHAIN is only valid with PMC_F_CALLCHAIN */
3893 		if ((pa.pm_flags & (PMC_F_CALLCHAIN | PMC_F_USERCALLCHAIN)) ==
3894 		    PMC_F_USERCALLCHAIN) {
3895 			error = EINVAL;
3896 			break;
3897 		}
3898 
3899 		/* PMC_F_USERCALLCHAIN is only valid for sampling mode */
3900 		if (pa.pm_flags & PMC_F_USERCALLCHAIN &&
3901 			mode != PMC_MODE_TS && mode != PMC_MODE_SS) {
3902 			error = EINVAL;
3903 			break;
3904 		}
3905 
3906 		/* process logging options are not allowed for system PMCs */
3907 		if (PMC_IS_SYSTEM_MODE(mode) && (pa.pm_flags &
3908 		    (PMC_F_LOG_PROCCSW | PMC_F_LOG_PROCEXIT))) {
3909 			error = EINVAL;
3910 			break;
3911 		}
3912 
3913 		/*
3914 		 * All sampling mode PMCs need to be able to interrupt the
3915 		 * CPU.
3916 		 */
3917 		if (PMC_IS_SAMPLING_MODE(mode))
3918 			caps |= PMC_CAP_INTERRUPT;
3919 
3920 		/* A valid class specifier should have been passed in. */
3921 		pcd = pmc_class_to_classdep(pa.pm_class);
3922 		if (pcd == NULL) {
3923 			error = EINVAL;
3924 			break;
3925 		}
3926 
3927 		/* The requested PMC capabilities should be feasible. */
3928 		if ((pcd->pcd_caps & caps) != caps) {
3929 			error = EOPNOTSUPP;
3930 			break;
3931 		}
3932 
3933 		PMCDBG4(PMC,ALL,2, "event=%d caps=0x%x mode=%d cpu=%d",
3934 		    pa.pm_ev, caps, mode, cpu);
3935 
3936 		pmc = pmc_allocate_pmc_descriptor();
3937 		pmc->pm_id    = PMC_ID_MAKE_ID(cpu,pa.pm_mode,pa.pm_class,
3938 		    PMC_ID_INVALID);
3939 		pmc->pm_event = pa.pm_ev;
3940 		pmc->pm_state = PMC_STATE_FREE;
3941 		pmc->pm_caps  = caps;
3942 		pmc->pm_flags = pa.pm_flags;
3943 
3944 		/* XXX set lower bound on sampling for process counters */
3945 		if (PMC_IS_SAMPLING_MODE(mode))
3946 			pmc->pm_sc.pm_reloadcount = pa.pm_count;
3947 		else
3948 			pmc->pm_sc.pm_initial = pa.pm_count;
3949 
3950 		/* switch thread to CPU 'cpu' */
3951 		pmc_save_cpu_binding(&pb);
3952 
3953 #define	PMC_IS_SHAREABLE_PMC(cpu, n)				\
3954 	(pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_state &		\
3955 	 PMC_PHW_FLAG_IS_SHAREABLE)
3956 #define	PMC_IS_UNALLOCATED(cpu, n)				\
3957 	(pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_pmc == NULL)
3958 
3959 		if (PMC_IS_SYSTEM_MODE(mode)) {
3960 			pmc_select_cpu(cpu);
3961 			for (n = pcd->pcd_ri; n < (int) md->pmd_npmc; n++) {
3962 				pcd = pmc_ri_to_classdep(md, n, &adjri);
3963 				if (pmc_can_allocate_row(n, mode) == 0 &&
3964 				    pmc_can_allocate_rowindex(
3965 					    curthread->td_proc, n, cpu) == 0 &&
3966 				    (PMC_IS_UNALLOCATED(cpu, n) ||
3967 				     PMC_IS_SHAREABLE_PMC(cpu, n)) &&
3968 				    pcd->pcd_allocate_pmc(cpu, adjri, pmc,
3969 					&pa) == 0)
3970 					break;
3971 			}
3972 		} else {
3973 			/* Process virtual mode */
3974 			for (n = pcd->pcd_ri; n < (int) md->pmd_npmc; n++) {
3975 				pcd = pmc_ri_to_classdep(md, n, &adjri);
3976 				if (pmc_can_allocate_row(n, mode) == 0 &&
3977 				    pmc_can_allocate_rowindex(
3978 					    curthread->td_proc, n,
3979 					    PMC_CPU_ANY) == 0 &&
3980 				    pcd->pcd_allocate_pmc(curthread->td_oncpu,
3981 					adjri, pmc, &pa) == 0)
3982 					break;
3983 			}
3984 		}
3985 
3986 #undef	PMC_IS_UNALLOCATED
3987 #undef	PMC_IS_SHAREABLE_PMC
3988 
3989 		pmc_restore_cpu_binding(&pb);
3990 
3991 		if (n == (int) md->pmd_npmc) {
3992 			pmc_destroy_pmc_descriptor(pmc);
3993 			pmc = NULL;
3994 			error = EINVAL;
3995 			break;
3996 		}
3997 
3998 		/* Fill in the correct value in the ID field */
3999 		pmc->pm_id = PMC_ID_MAKE_ID(cpu,mode,pa.pm_class,n);
4000 
4001 		PMCDBG5(PMC,ALL,2, "ev=%d class=%d mode=%d n=%d -> pmcid=%x",
4002 		    pmc->pm_event, pa.pm_class, mode, n, pmc->pm_id);
4003 
4004 		/* Process mode PMCs with logging enabled need log files */
4005 		if (pmc->pm_flags & (PMC_F_LOG_PROCEXIT | PMC_F_LOG_PROCCSW))
4006 			pmc->pm_flags |= PMC_F_NEEDS_LOGFILE;
4007 
4008 		/* All system mode sampling PMCs require a log file */
4009 		if (PMC_IS_SAMPLING_MODE(mode) && PMC_IS_SYSTEM_MODE(mode))
4010 			pmc->pm_flags |= PMC_F_NEEDS_LOGFILE;
4011 
4012 		/*
4013 		 * Configure global pmc's immediately
4014 		 */
4015 
4016 		if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pmc))) {
4017 
4018 			pmc_save_cpu_binding(&pb);
4019 			pmc_select_cpu(cpu);
4020 
4021 			phw = pmc_pcpu[cpu]->pc_hwpmcs[n];
4022 			pcd = pmc_ri_to_classdep(md, n, &adjri);
4023 
4024 			if ((phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0 ||
4025 			    (error = pcd->pcd_config_pmc(cpu, adjri, pmc)) != 0) {
4026 				(void) pcd->pcd_release_pmc(cpu, adjri, pmc);
4027 				pmc_destroy_pmc_descriptor(pmc);
4028 				pmc = NULL;
4029 				pmc_restore_cpu_binding(&pb);
4030 				error = EPERM;
4031 				break;
4032 			}
4033 
4034 			pmc_restore_cpu_binding(&pb);
4035 		}
4036 
4037 		pmc->pm_state    = PMC_STATE_ALLOCATED;
4038 		pmc->pm_class	= pa.pm_class;
4039 
4040 		/*
4041 		 * mark row disposition
4042 		 */
4043 
4044 		if (PMC_IS_SYSTEM_MODE(mode))
4045 			PMC_MARK_ROW_STANDALONE(n);
4046 		else
4047 			PMC_MARK_ROW_THREAD(n);
4048 
4049 		/*
4050 		 * Register this PMC with the current thread as its owner.
4051 		 */
4052 
4053 		if ((error =
4054 		    pmc_register_owner(curthread->td_proc, pmc)) != 0) {
4055 			pmc_release_pmc_descriptor(pmc);
4056 			pmc_destroy_pmc_descriptor(pmc);
4057 			pmc = NULL;
4058 			break;
4059 		}
4060 
4061 
4062 		/*
4063 		 * Return the allocated index.
4064 		 */
4065 
4066 		pa.pm_pmcid = pmc->pm_id;
4067 
4068 		error = copyout(&pa, arg, sizeof(pa));
4069 	}
4070 	break;
4071 
4072 
4073 	/*
4074 	 * Attach a PMC to a process.
4075 	 */
4076 
4077 	case PMC_OP_PMCATTACH:
4078 	{
4079 		struct pmc *pm;
4080 		struct proc *p;
4081 		struct pmc_op_pmcattach a;
4082 
4083 		sx_assert(&pmc_sx, SX_XLOCKED);
4084 
4085 		if ((error = copyin(arg, &a, sizeof(a))) != 0)
4086 			break;
4087 
4088 		if (a.pm_pid < 0) {
4089 			error = EINVAL;
4090 			break;
4091 		} else if (a.pm_pid == 0)
4092 			a.pm_pid = td->td_proc->p_pid;
4093 
4094 		if ((error = pmc_find_pmc(a.pm_pmc, &pm)) != 0)
4095 			break;
4096 
4097 		if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm))) {
4098 			error = EINVAL;
4099 			break;
4100 		}
4101 
4102 		/* PMCs may be (re)attached only when allocated or stopped */
4103 		if (pm->pm_state == PMC_STATE_RUNNING) {
4104 			error = EBUSY;
4105 			break;
4106 		} else if (pm->pm_state != PMC_STATE_ALLOCATED &&
4107 		    pm->pm_state != PMC_STATE_STOPPED) {
4108 			error = EINVAL;
4109 			break;
4110 		}
4111 
4112 		/* lookup pid */
4113 		if ((p = pfind(a.pm_pid)) == NULL) {
4114 			error = ESRCH;
4115 			break;
4116 		}
4117 
4118 		/*
4119 		 * Ignore processes that are working on exiting.
4120 		 */
4121 		if (p->p_flag & P_WEXIT) {
4122 			error = ESRCH;
4123 			PROC_UNLOCK(p);	/* pfind() returns a locked process */
4124 			break;
4125 		}
4126 
4127 		/*
4128 		 * we are allowed to attach a PMC to a process if
4129 		 * we can debug it.
4130 		 */
4131 		error = p_candebug(curthread, p);
4132 
4133 		PROC_UNLOCK(p);
4134 
4135 		if (error == 0)
4136 			error = pmc_attach_process(p, pm);
4137 	}
4138 	break;
4139 
4140 
4141 	/*
4142 	 * Detach an attached PMC from a process.
4143 	 */
4144 
4145 	case PMC_OP_PMCDETACH:
4146 	{
4147 		struct pmc *pm;
4148 		struct proc *p;
4149 		struct pmc_op_pmcattach a;
4150 
4151 		if ((error = copyin(arg, &a, sizeof(a))) != 0)
4152 			break;
4153 
4154 		if (a.pm_pid < 0) {
4155 			error = EINVAL;
4156 			break;
4157 		} else if (a.pm_pid == 0)
4158 			a.pm_pid = td->td_proc->p_pid;
4159 
4160 		if ((error = pmc_find_pmc(a.pm_pmc, &pm)) != 0)
4161 			break;
4162 
4163 		if ((p = pfind(a.pm_pid)) == NULL) {
4164 			error = ESRCH;
4165 			break;
4166 		}
4167 
4168 		/*
4169 		 * Treat processes that are in the process of exiting
4170 		 * as if they were not present.
4171 		 */
4172 
4173 		if (p->p_flag & P_WEXIT)
4174 			error = ESRCH;
4175 
4176 		PROC_UNLOCK(p);	/* pfind() returns a locked process */
4177 
4178 		if (error == 0)
4179 			error = pmc_detach_process(p, pm);
4180 	}
4181 	break;
4182 
4183 
4184 	/*
4185 	 * Retrieve the MSR number associated with the counter
4186 	 * 'pmc_id'.  This allows processes to directly use RDPMC
4187 	 * instructions to read their PMCs, without the overhead of a
4188 	 * system call.
4189 	 */
4190 
4191 	case PMC_OP_PMCGETMSR:
4192 	{
4193 		int adjri, ri;
4194 		struct pmc *pm;
4195 		struct pmc_target *pt;
4196 		struct pmc_op_getmsr gm;
4197 		struct pmc_classdep *pcd;
4198 
4199 		PMC_DOWNGRADE_SX();
4200 
4201 		if ((error = copyin(arg, &gm, sizeof(gm))) != 0)
4202 			break;
4203 
4204 		if ((error = pmc_find_pmc(gm.pm_pmcid, &pm)) != 0)
4205 			break;
4206 
4207 		/*
4208 		 * The allocated PMC has to be a process virtual PMC,
4209 		 * i.e., of type MODE_T[CS].  Global PMCs can only be
4210 		 * read using the PMCREAD operation since they may be
4211 		 * allocated on a different CPU than the one we could
4212 		 * be running on at the time of the RDPMC instruction.
4213 		 *
4214 		 * The GETMSR operation is not allowed for PMCs that
4215 		 * are inherited across processes.
4216 		 */
4217 
4218 		if (!PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)) ||
4219 		    (pm->pm_flags & PMC_F_DESCENDANTS)) {
4220 			error = EINVAL;
4221 			break;
4222 		}
4223 
4224 		/*
4225 		 * It only makes sense to use a RDPMC (or its
4226 		 * equivalent instruction on non-x86 architectures) on
4227 		 * a process that has allocated and attached a PMC to
4228 		 * itself.  Conversely the PMC is only allowed to have
4229 		 * one process attached to it -- its owner.
4230 		 */
4231 
4232 		if ((pt = LIST_FIRST(&pm->pm_targets)) == NULL ||
4233 		    LIST_NEXT(pt, pt_next) != NULL ||
4234 		    pt->pt_process->pp_proc != pm->pm_owner->po_owner) {
4235 			error = EINVAL;
4236 			break;
4237 		}
4238 
4239 		ri = PMC_TO_ROWINDEX(pm);
4240 		pcd = pmc_ri_to_classdep(md, ri, &adjri);
4241 
4242 		/* PMC class has no 'GETMSR' support */
4243 		if (pcd->pcd_get_msr == NULL) {
4244 			error = ENOSYS;
4245 			break;
4246 		}
4247 
4248 		if ((error = (*pcd->pcd_get_msr)(adjri, &gm.pm_msr)) < 0)
4249 			break;
4250 
4251 		if ((error = copyout(&gm, arg, sizeof(gm))) < 0)
4252 			break;
4253 
4254 		/*
4255 		 * Mark our process as using MSRs.  Update machine
4256 		 * state using a forced context switch.
4257 		 */
4258 
4259 		pt->pt_process->pp_flags |= PMC_PP_ENABLE_MSR_ACCESS;
4260 		pmc_force_context_switch();
4261 
4262 	}
4263 	break;
4264 
4265 	/*
4266 	 * Release an allocated PMC
4267 	 */
4268 
4269 	case PMC_OP_PMCRELEASE:
4270 	{
4271 		pmc_id_t pmcid;
4272 		struct pmc *pm;
4273 		struct pmc_owner *po;
4274 		struct pmc_op_simple sp;
4275 
4276 		/*
4277 		 * Find PMC pointer for the named PMC.
4278 		 *
4279 		 * Use pmc_release_pmc_descriptor() to switch off the
4280 		 * PMC, remove all its target threads, and remove the
4281 		 * PMC from its owner's list.
4282 		 *
4283 		 * Remove the owner record if this is the last PMC
4284 		 * owned.
4285 		 *
4286 		 * Free up space.
4287 		 */
4288 
4289 		if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
4290 			break;
4291 
4292 		pmcid = sp.pm_pmcid;
4293 
4294 		if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
4295 			break;
4296 
4297 		po = pm->pm_owner;
4298 		pmc_release_pmc_descriptor(pm);
4299 		pmc_maybe_remove_owner(po);
4300 		pmc_destroy_pmc_descriptor(pm);
4301 	}
4302 	break;
4303 
4304 
4305 	/*
4306 	 * Read and/or write a PMC.
4307 	 */
4308 
4309 	case PMC_OP_PMCRW:
4310 	{
4311 		int adjri;
4312 		struct pmc *pm;
4313 		uint32_t cpu, ri;
4314 		pmc_value_t oldvalue;
4315 		struct pmc_binding pb;
4316 		struct pmc_op_pmcrw prw;
4317 		struct pmc_classdep *pcd;
4318 		struct pmc_op_pmcrw *pprw;
4319 
4320 		PMC_DOWNGRADE_SX();
4321 
4322 		if ((error = copyin(arg, &prw, sizeof(prw))) != 0)
4323 			break;
4324 
4325 		ri = 0;
4326 		PMCDBG2(PMC,OPS,1, "rw id=%d flags=0x%x", prw.pm_pmcid,
4327 		    prw.pm_flags);
4328 
4329 		/* must have at least one flag set */
4330 		if ((prw.pm_flags & (PMC_F_OLDVALUE|PMC_F_NEWVALUE)) == 0) {
4331 			error = EINVAL;
4332 			break;
4333 		}
4334 
4335 		/* locate pmc descriptor */
4336 		if ((error = pmc_find_pmc(prw.pm_pmcid, &pm)) != 0)
4337 			break;
4338 
4339 		/* Can't read a PMC that hasn't been started. */
4340 		if (pm->pm_state != PMC_STATE_ALLOCATED &&
4341 		    pm->pm_state != PMC_STATE_STOPPED &&
4342 		    pm->pm_state != PMC_STATE_RUNNING) {
4343 			error = EINVAL;
4344 			break;
4345 		}
4346 
4347 		/* writing a new value is allowed only for 'STOPPED' pmcs */
4348 		if (pm->pm_state == PMC_STATE_RUNNING &&
4349 		    (prw.pm_flags & PMC_F_NEWVALUE)) {
4350 			error = EBUSY;
4351 			break;
4352 		}
4353 
4354 		if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) {
4355 
4356 			/*
4357 			 * If this PMC is attached to its owner (i.e.,
4358 			 * the process requesting this operation) and
4359 			 * is running, then attempt to get an
4360 			 * upto-date reading from hardware for a READ.
4361 			 * Writes are only allowed when the PMC is
4362 			 * stopped, so only update the saved value
4363 			 * field.
4364 			 *
4365 			 * If the PMC is not running, or is not
4366 			 * attached to its owner, read/write to the
4367 			 * savedvalue field.
4368 			 */
4369 
4370 			ri = PMC_TO_ROWINDEX(pm);
4371 			pcd = pmc_ri_to_classdep(md, ri, &adjri);
4372 
4373 			mtx_pool_lock_spin(pmc_mtxpool, pm);
4374 			cpu = curthread->td_oncpu;
4375 
4376 			if (prw.pm_flags & PMC_F_OLDVALUE) {
4377 				if ((pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) &&
4378 				    (pm->pm_state == PMC_STATE_RUNNING))
4379 					error = (*pcd->pcd_read_pmc)(cpu, adjri,
4380 					    &oldvalue);
4381 				else
4382 					oldvalue = pm->pm_gv.pm_savedvalue;
4383 			}
4384 			if (prw.pm_flags & PMC_F_NEWVALUE)
4385 				pm->pm_gv.pm_savedvalue = prw.pm_value;
4386 
4387 			mtx_pool_unlock_spin(pmc_mtxpool, pm);
4388 
4389 		} else { /* System mode PMCs */
4390 			cpu = PMC_TO_CPU(pm);
4391 			ri  = PMC_TO_ROWINDEX(pm);
4392 			pcd = pmc_ri_to_classdep(md, ri, &adjri);
4393 
4394 			if (!pmc_cpu_is_active(cpu)) {
4395 				error = ENXIO;
4396 				break;
4397 			}
4398 
4399 			/* move this thread to CPU 'cpu' */
4400 			pmc_save_cpu_binding(&pb);
4401 			pmc_select_cpu(cpu);
4402 
4403 			critical_enter();
4404 			/* save old value */
4405 			if (prw.pm_flags & PMC_F_OLDVALUE)
4406 				if ((error = (*pcd->pcd_read_pmc)(cpu, adjri,
4407 					 &oldvalue)))
4408 					goto error;
4409 			/* write out new value */
4410 			if (prw.pm_flags & PMC_F_NEWVALUE)
4411 				error = (*pcd->pcd_write_pmc)(cpu, adjri,
4412 				    prw.pm_value);
4413 		error:
4414 			critical_exit();
4415 			pmc_restore_cpu_binding(&pb);
4416 			if (error)
4417 				break;
4418 		}
4419 
4420 		pprw = (struct pmc_op_pmcrw *) arg;
4421 
4422 #ifdef	HWPMC_DEBUG
4423 		if (prw.pm_flags & PMC_F_NEWVALUE)
4424 			PMCDBG3(PMC,OPS,2, "rw id=%d new %jx -> old %jx",
4425 			    ri, prw.pm_value, oldvalue);
4426 		else if (prw.pm_flags & PMC_F_OLDVALUE)
4427 			PMCDBG2(PMC,OPS,2, "rw id=%d -> old %jx", ri, oldvalue);
4428 #endif
4429 
4430 		/* return old value if requested */
4431 		if (prw.pm_flags & PMC_F_OLDVALUE)
4432 			if ((error = copyout(&oldvalue, &pprw->pm_value,
4433 				 sizeof(prw.pm_value))))
4434 				break;
4435 
4436 	}
4437 	break;
4438 
4439 
4440 	/*
4441 	 * Set the sampling rate for a sampling mode PMC and the
4442 	 * initial count for a counting mode PMC.
4443 	 */
4444 
4445 	case PMC_OP_PMCSETCOUNT:
4446 	{
4447 		struct pmc *pm;
4448 		struct pmc_op_pmcsetcount sc;
4449 
4450 		PMC_DOWNGRADE_SX();
4451 
4452 		if ((error = copyin(arg, &sc, sizeof(sc))) != 0)
4453 			break;
4454 
4455 		if ((error = pmc_find_pmc(sc.pm_pmcid, &pm)) != 0)
4456 			break;
4457 
4458 		if (pm->pm_state == PMC_STATE_RUNNING) {
4459 			error = EBUSY;
4460 			break;
4461 		}
4462 
4463 		if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
4464 			pm->pm_sc.pm_reloadcount = sc.pm_count;
4465 		else
4466 			pm->pm_sc.pm_initial = sc.pm_count;
4467 	}
4468 	break;
4469 
4470 
4471 	/*
4472 	 * Start a PMC.
4473 	 */
4474 
4475 	case PMC_OP_PMCSTART:
4476 	{
4477 		pmc_id_t pmcid;
4478 		struct pmc *pm;
4479 		struct pmc_op_simple sp;
4480 
4481 		sx_assert(&pmc_sx, SX_XLOCKED);
4482 
4483 		if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
4484 			break;
4485 
4486 		pmcid = sp.pm_pmcid;
4487 
4488 		if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
4489 			break;
4490 
4491 		KASSERT(pmcid == pm->pm_id,
4492 		    ("[pmc,%d] pmcid %x != id %x", __LINE__,
4493 			pm->pm_id, pmcid));
4494 
4495 		if (pm->pm_state == PMC_STATE_RUNNING) /* already running */
4496 			break;
4497 		else if (pm->pm_state != PMC_STATE_STOPPED &&
4498 		    pm->pm_state != PMC_STATE_ALLOCATED) {
4499 			error = EINVAL;
4500 			break;
4501 		}
4502 
4503 		error = pmc_start(pm);
4504 	}
4505 	break;
4506 
4507 
4508 	/*
4509 	 * Stop a PMC.
4510 	 */
4511 
4512 	case PMC_OP_PMCSTOP:
4513 	{
4514 		pmc_id_t pmcid;
4515 		struct pmc *pm;
4516 		struct pmc_op_simple sp;
4517 
4518 		PMC_DOWNGRADE_SX();
4519 
4520 		if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
4521 			break;
4522 
4523 		pmcid = sp.pm_pmcid;
4524 
4525 		/*
4526 		 * Mark the PMC as inactive and invoke the MD stop
4527 		 * routines if needed.
4528 		 */
4529 
4530 		if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
4531 			break;
4532 
4533 		KASSERT(pmcid == pm->pm_id,
4534 		    ("[pmc,%d] pmc id %x != pmcid %x", __LINE__,
4535 			pm->pm_id, pmcid));
4536 
4537 		if (pm->pm_state == PMC_STATE_STOPPED) /* already stopped */
4538 			break;
4539 		else if (pm->pm_state != PMC_STATE_RUNNING) {
4540 			error = EINVAL;
4541 			break;
4542 		}
4543 
4544 		error = pmc_stop(pm);
4545 	}
4546 	break;
4547 
4548 
4549 	/*
4550 	 * Write a user supplied value to the log file.
4551 	 */
4552 
4553 	case PMC_OP_WRITELOG:
4554 	{
4555 		struct pmc_op_writelog wl;
4556 		struct pmc_owner *po;
4557 
4558 		PMC_DOWNGRADE_SX();
4559 
4560 		if ((error = copyin(arg, &wl, sizeof(wl))) != 0)
4561 			break;
4562 
4563 		if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) {
4564 			error = EINVAL;
4565 			break;
4566 		}
4567 
4568 		if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) {
4569 			error = EINVAL;
4570 			break;
4571 		}
4572 
4573 		error = pmclog_process_userlog(po, &wl);
4574 	}
4575 	break;
4576 
4577 
4578 	default:
4579 		error = EINVAL;
4580 		break;
4581 	}
4582 
4583 	if (is_sx_downgraded)
4584 		sx_sunlock(&pmc_sx);
4585 	else
4586 		sx_xunlock(&pmc_sx);
4587 done_syscall:
4588 	if (error)
4589 		counter_u64_add(pmc_stats.pm_syscall_errors, 1);
4590 
4591 	return (error);
4592 }
4593 
4594 /*
4595  * Helper functions
4596  */
4597 
4598 
4599 /*
4600  * Mark the thread as needing callchain capture and post an AST.  The
4601  * actual callchain capture will be done in a context where it is safe
4602  * to take page faults.
4603  */
4604 
4605 static void
4606 pmc_post_callchain_callback(void)
4607 {
4608 	struct thread *td;
4609 
4610 	td = curthread;
4611 
4612 	/*
4613 	 * If there is multiple PMCs for the same interrupt ignore new post
4614 	 */
4615 	if (td->td_pflags & TDP_CALLCHAIN)
4616 		return;
4617 
4618 	/*
4619 	 * Mark this thread as needing callchain capture.
4620 	 * `td->td_pflags' will be safe to touch because this thread
4621 	 * was in user space when it was interrupted.
4622 	 */
4623 	td->td_pflags |= TDP_CALLCHAIN;
4624 
4625 	/*
4626 	 * Don't let this thread migrate between CPUs until callchain
4627 	 * capture completes.
4628 	 */
4629 	sched_pin();
4630 
4631 	return;
4632 }
4633 
4634 /*
4635  * Find a free slot in the per-cpu array of samples and capture the
4636  * current callchain there.  If a sample was successfully added, a bit
4637  * is set in mask 'pmc_cpumask' denoting that the DO_SAMPLES hook
4638  * needs to be invoked from the clock handler.
4639  *
4640  * This function is meant to be called from an NMI handler.  It cannot
4641  * use any of the locking primitives supplied by the OS.
4642  */
4643 
4644 static int
4645 pmc_add_sample(int ring, struct pmc *pm, struct trapframe *tf)
4646 {
4647 	int error, cpu, callchaindepth, inuserspace;
4648 	struct thread *td;
4649 	struct pmc_sample *ps;
4650 	struct pmc_samplebuffer *psb;
4651 
4652 	error = 0;
4653 
4654 	/*
4655 	 * Allocate space for a sample buffer.
4656 	 */
4657 	cpu = curcpu;
4658 	psb = pmc_pcpu[cpu]->pc_sb[ring];
4659 	inuserspace = TRAPF_USERMODE(tf);
4660 	ps = psb->ps_write;
4661 	if (ps->ps_nsamples == PMC_SAMPLE_INUSE) {
4662 		counter_u64_add(ps->ps_pmc->pm_runcount, -1);
4663 		counter_u64_add(pmc_stats.pm_overwrites, 1);
4664 		ps->ps_nsamples = 0;
4665 	} else if (ps->ps_nsamples) {	/* in use, reader hasn't caught up */
4666 		pm->pm_pcpu_state[cpu].pps_stalled = 1;
4667 		counter_u64_add(pmc_stats.pm_intr_bufferfull, 1);
4668 		PMCDBG6(SAM,INT,1,"(spc) cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d",
4669 		    cpu, pm, (void *) tf, inuserspace,
4670 		    (int) (psb->ps_write - psb->ps_samples),
4671 		    (int) (psb->ps_read - psb->ps_samples));
4672 		callchaindepth = 1;
4673 		error = ENOMEM;
4674 		goto done;
4675 	}
4676 
4677 	/* Fill in entry. */
4678 	PMCDBG6(SAM,INT,1,"cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d", cpu, pm,
4679 	    (void *) tf, inuserspace,
4680 	    (int) (psb->ps_write - psb->ps_samples),
4681 	    (int) (psb->ps_read - psb->ps_samples));
4682 
4683 	KASSERT(counter_u64_fetch(pm->pm_runcount) >= 0,
4684 	    ("[pmc,%d] pm=%p runcount %ld", __LINE__, (void *) pm,
4685 		 (unsigned long)counter_u64_fetch(pm->pm_runcount)));
4686 
4687 	counter_u64_add(pm->pm_runcount, 1);	/* hold onto PMC */
4688 
4689 	td = curthread;
4690 	ps->ps_pmc = pm;
4691 	ps->ps_td = td;
4692 	ps->ps_pid = td->td_proc->p_pid;
4693 	ps->ps_tid = td->td_tid;
4694 	ps->ps_tsc = pmc_rdtsc();
4695 
4696 	ps->ps_cpu = cpu;
4697 	ps->ps_flags = inuserspace ? PMC_CC_F_USERSPACE : 0;
4698 
4699 	callchaindepth = (pm->pm_flags & PMC_F_CALLCHAIN) ?
4700 	    pmc_callchaindepth : 1;
4701 
4702 	if (callchaindepth == 1)
4703 		ps->ps_pc[0] = PMC_TRAPFRAME_TO_PC(tf);
4704 	else {
4705 		/*
4706 		 * Kernel stack traversals can be done immediately,
4707 		 * while we defer to an AST for user space traversals.
4708 		 */
4709 		if (!inuserspace) {
4710 			callchaindepth =
4711 			    pmc_save_kernel_callchain(ps->ps_pc,
4712 				callchaindepth, tf);
4713 		} else {
4714 			pmc_post_callchain_callback();
4715 			callchaindepth = PMC_SAMPLE_INUSE;
4716 		}
4717 	}
4718 
4719 	ps->ps_nsamples = callchaindepth;	/* mark entry as in use */
4720 	if (ring == PMC_UR) {
4721 		ps->ps_nsamples_actual = callchaindepth;	/* mark entry as in use */
4722 		ps->ps_nsamples = PMC_SAMPLE_INUSE;
4723 	} else
4724 		ps->ps_nsamples = callchaindepth;	/* mark entry as in use */
4725 	/* increment write pointer, modulo ring buffer size */
4726 	ps++;
4727 	if (ps == psb->ps_fence)
4728 		psb->ps_write = psb->ps_samples;
4729 	else
4730 		psb->ps_write = ps;
4731 
4732  done:
4733 	/* mark CPU as needing processing */
4734 	if (callchaindepth != PMC_SAMPLE_INUSE)
4735 		DPCPU_SET(pmc_sampled, 1);
4736 
4737 	return (error);
4738 }
4739 
4740 /*
4741  * Interrupt processing.
4742  *
4743  * This function is meant to be called from an NMI handler.  It cannot
4744  * use any of the locking primitives supplied by the OS.
4745  */
4746 
4747 int
4748 pmc_process_interrupt(int ring, struct pmc *pm, struct trapframe *tf)
4749 {
4750 	struct thread *td;
4751 
4752 	td = curthread;
4753 	if ((pm->pm_flags & PMC_F_USERCALLCHAIN) &&
4754 	    (td->td_proc->p_flag & P_KPROC) == 0 &&
4755 	    !TRAPF_USERMODE(tf)) {
4756 		atomic_add_int(&td->td_pmcpend, 1);
4757 		return (pmc_add_sample(PMC_UR, pm, tf));
4758 	}
4759 	return (pmc_add_sample(ring, pm, tf));
4760 }
4761 
4762 /*
4763  * Capture a user call chain.  This function will be called from ast()
4764  * before control returns to userland and before the process gets
4765  * rescheduled.
4766  */
4767 
4768 static void
4769 pmc_capture_user_callchain(int cpu, int ring, struct trapframe *tf)
4770 {
4771 	struct pmc *pm;
4772 	struct thread *td;
4773 	struct pmc_sample *ps, *ps_end;
4774 	struct pmc_samplebuffer *psb;
4775 	int nsamples, nrecords, pass;
4776 #ifdef	INVARIANTS
4777 	int ncallchains;
4778 	int nfree;
4779 #endif
4780 
4781 	psb = pmc_pcpu[cpu]->pc_sb[ring];
4782 	td = curthread;
4783 
4784 	KASSERT(td->td_pflags & TDP_CALLCHAIN,
4785 	    ("[pmc,%d] Retrieving callchain for thread that doesn't want it",
4786 		__LINE__));
4787 
4788 #ifdef	INVARIANTS
4789 	ncallchains = 0;
4790 	nfree = 0;
4791 #endif
4792 	nrecords = INT_MAX;
4793 	pass = 0;
4794  restart:
4795 	if (ring == PMC_UR)
4796 		nrecords = atomic_readandclear_32(&td->td_pmcpend);
4797 
4798 	/*
4799 	 * Iterate through all deferred callchain requests.
4800 	 * Walk from the current read pointer to the current
4801 	 * write pointer.
4802 	 */
4803 
4804 	ps = psb->ps_read;
4805 	ps_end = psb->ps_write;
4806 	do {
4807 #ifdef	INVARIANTS
4808 		if (ps->ps_nsamples == PMC_SAMPLE_FREE) {
4809 			nfree++;
4810 			goto next;
4811 		}
4812 
4813 		if ((ps->ps_pmc == NULL) ||
4814 		    (ps->ps_pmc->pm_state != PMC_STATE_RUNNING))
4815 			nfree++;
4816 #endif
4817 		if (ps->ps_nsamples != PMC_SAMPLE_INUSE)
4818 			goto next;
4819 		if (ps->ps_td != td)
4820 			goto next;
4821 
4822 		KASSERT(ps->ps_cpu == cpu,
4823 		    ("[pmc,%d] cpu mismatch ps_cpu=%d pcpu=%d", __LINE__,
4824 			ps->ps_cpu, PCPU_GET(cpuid)));
4825 
4826 		pm = ps->ps_pmc;
4827 
4828 		KASSERT(pm->pm_flags & PMC_F_CALLCHAIN,
4829 		    ("[pmc,%d] Retrieving callchain for PMC that doesn't "
4830 			"want it", __LINE__));
4831 
4832 		KASSERT(counter_u64_fetch(pm->pm_runcount) > 0,
4833 		    ("[pmc,%d] runcount %ld", __LINE__, (unsigned long)counter_u64_fetch(pm->pm_runcount)));
4834 
4835 		if (ring == PMC_UR) {
4836 			nsamples = ps->ps_nsamples_actual;
4837 			counter_u64_add(pmc_stats.pm_merges, 1);
4838 		} else
4839 			nsamples = 0;
4840 
4841 		/*
4842 		 * Retrieve the callchain and mark the sample buffer
4843 		 * as 'processable' by the timer tick sweep code.
4844 		 */
4845 
4846 #ifdef INVARIANTS
4847 		ncallchains++;
4848 #endif
4849 
4850 		if (__predict_true(nsamples < pmc_callchaindepth - 1))
4851 			nsamples += pmc_save_user_callchain(ps->ps_pc + nsamples,
4852 		       pmc_callchaindepth - nsamples - 1, tf);
4853 		wmb();
4854 		ps->ps_nsamples = nsamples;
4855 		if (nrecords-- == 1)
4856 			break;
4857 next:
4858 		/* increment the pointer, modulo sample ring size */
4859 		if (++ps == psb->ps_fence)
4860 			ps = psb->ps_samples;
4861 	} while (ps != ps_end);
4862 	if (__predict_false(ring == PMC_UR && td->td_pmcpend)) {
4863 		if (pass == 0) {
4864 			pass = 1;
4865 			goto restart;
4866 		}
4867 		/* only collect samples for this part once */
4868 		td->td_pmcpend = 0;
4869 	}
4870 
4871 #ifdef INVARIANTS
4872 	if (ring == PMC_HR)
4873 		KASSERT(ncallchains > 0 || nfree > 0,
4874 		    ("[pmc,%d] cpu %d didn't find a sample to collect", __LINE__,
4875 			    cpu));
4876 #endif
4877 
4878 	/* mark CPU as needing processing */
4879 	DPCPU_SET(pmc_sampled, 1);
4880 }
4881 
4882 
4883 static void
4884 pmc_flush_ring(int cpu, int ring)
4885 {
4886 	struct pmc *pm;
4887 	struct pmc_sample *ps;
4888 	struct pmc_samplebuffer *psb;
4889 	int n;
4890 
4891 	psb = pmc_pcpu[cpu]->pc_sb[ring];
4892 
4893 	for (n = 0; n < pmc_nsamples; n++) { /* bound on #iterations */
4894 
4895 		ps = psb->ps_read;
4896 		if (ps->ps_nsamples == PMC_SAMPLE_FREE)
4897 			goto next;
4898 		pm = ps->ps_pmc;
4899 		counter_u64_add(pm->pm_runcount, -1);
4900 		ps->ps_nsamples = PMC_SAMPLE_FREE;
4901 		/* increment read pointer, modulo sample size */
4902 	next:
4903 		if (++ps == psb->ps_fence)
4904 			psb->ps_read = psb->ps_samples;
4905 		else
4906 			psb->ps_read = ps;
4907 	}
4908 }
4909 
4910 void
4911 pmc_flush_samples(int cpu)
4912 {
4913 	int n;
4914 
4915 	for (n = 0; n < PMC_NUM_SR; n++)
4916 		pmc_flush_ring(cpu, n);
4917 }
4918 
4919 
4920 /*
4921  * Process saved PC samples.
4922  */
4923 
4924 static void
4925 pmc_process_samples(int cpu, int ring)
4926 {
4927 	struct pmc *pm;
4928 	int adjri, n;
4929 	struct thread *td;
4930 	struct pmc_owner *po;
4931 	struct pmc_sample *ps;
4932 	struct pmc_classdep *pcd;
4933 	struct pmc_samplebuffer *psb;
4934 
4935 	KASSERT(PCPU_GET(cpuid) == cpu,
4936 	    ("[pmc,%d] not on the correct CPU pcpu=%d cpu=%d", __LINE__,
4937 		PCPU_GET(cpuid), cpu));
4938 
4939 	psb = pmc_pcpu[cpu]->pc_sb[ring];
4940 
4941 	for (n = 0; n < pmc_nsamples; n++) { /* bound on #iterations */
4942 
4943 		ps = psb->ps_read;
4944 		if (ps->ps_nsamples == PMC_SAMPLE_FREE)
4945 			break;
4946 
4947 		pm = ps->ps_pmc;
4948 
4949 		KASSERT(counter_u64_fetch(pm->pm_runcount) > 0,
4950 		    ("[pmc,%d] pm=%p runcount %ld", __LINE__, (void *) pm,
4951 			 (unsigned long)counter_u64_fetch(pm->pm_runcount)));
4952 
4953 		po = pm->pm_owner;
4954 
4955 		KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)),
4956 		    ("[pmc,%d] pmc=%p non-sampling mode=%d", __LINE__,
4957 			pm, PMC_TO_MODE(pm)));
4958 
4959 		/* Ignore PMCs that have been switched off */
4960 		if (pm->pm_state != PMC_STATE_RUNNING)
4961 			goto entrydone;
4962 
4963 		/* If there is a pending AST wait for completion */
4964 		if (ps->ps_nsamples == PMC_SAMPLE_INUSE) {
4965 			/* Need a rescan at a later time. */
4966 			DPCPU_SET(pmc_sampled, 1);
4967 			break;
4968 		}
4969 
4970 		PMCDBG6(SAM,OPS,1,"cpu=%d pm=%p n=%d fl=%x wr=%d rd=%d", cpu,
4971 		    pm, ps->ps_nsamples, ps->ps_flags,
4972 		    (int) (psb->ps_write - psb->ps_samples),
4973 		    (int) (psb->ps_read - psb->ps_samples));
4974 
4975 		/*
4976 		 * If this is a process-mode PMC that is attached to
4977 		 * its owner, and if the PC is in user mode, update
4978 		 * profiling statistics like timer-based profiling
4979 		 * would have done.
4980 		 *
4981 		 * Otherwise, this is either a sampling-mode PMC that
4982 		 * is attached to a different process than its owner,
4983 		 * or a system-wide sampling PMC. Dispatch a log
4984 		 * entry to the PMC's owner process.
4985 		 */
4986 		if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) {
4987 			if (ps->ps_flags & PMC_CC_F_USERSPACE) {
4988 				td = FIRST_THREAD_IN_PROC(po->po_owner);
4989 				addupc_intr(td, ps->ps_pc[0], 1);
4990 			}
4991 		} else
4992 			pmclog_process_callchain(pm, ps);
4993 
4994 	entrydone:
4995 		ps->ps_nsamples = 0; /* mark entry as free */
4996 		counter_u64_add(pm->pm_runcount, -1);
4997 
4998 		/* increment read pointer, modulo sample size */
4999 		if (++ps == psb->ps_fence)
5000 			psb->ps_read = psb->ps_samples;
5001 		else
5002 			psb->ps_read = ps;
5003 	}
5004 
5005 	counter_u64_add(pmc_stats.pm_log_sweeps, 1);
5006 
5007 	/* Do not re-enable stalled PMCs if we failed to process any samples */
5008 	if (n == 0)
5009 		return;
5010 
5011 	/*
5012 	 * Restart any stalled sampling PMCs on this CPU.
5013 	 *
5014 	 * If the NMI handler sets the pm_stalled field of a PMC after
5015 	 * the check below, we'll end up processing the stalled PMC at
5016 	 * the next hardclock tick.
5017 	 */
5018 	for (n = 0; n < md->pmd_npmc; n++) {
5019 		pcd = pmc_ri_to_classdep(md, n, &adjri);
5020 		KASSERT(pcd != NULL,
5021 		    ("[pmc,%d] null pcd ri=%d", __LINE__, n));
5022 		(void) (*pcd->pcd_get_config)(cpu,adjri,&pm);
5023 
5024 		if (pm == NULL ||			 /* !cfg'ed */
5025 		    pm->pm_state != PMC_STATE_RUNNING || /* !active */
5026 		    !PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) || /* !sampling */
5027 			!pm->pm_pcpu_state[cpu].pps_cpustate  || /* !desired */
5028 		    !pm->pm_pcpu_state[cpu].pps_stalled) /* !stalled */
5029 			continue;
5030 
5031 		pm->pm_pcpu_state[cpu].pps_stalled = 0;
5032 		(*pcd->pcd_start_pmc)(cpu, adjri);
5033 	}
5034 }
5035 
5036 /*
5037  * Event handlers.
5038  */
5039 
5040 /*
5041  * Handle a process exit.
5042  *
5043  * Remove this process from all hash tables.  If this process
5044  * owned any PMCs, turn off those PMCs and deallocate them,
5045  * removing any associations with target processes.
5046  *
5047  * This function will be called by the last 'thread' of a
5048  * process.
5049  *
5050  * XXX This eventhandler gets called early in the exit process.
5051  * Consider using a 'hook' invocation from thread_exit() or equivalent
5052  * spot.  Another negative is that kse_exit doesn't seem to call
5053  * exit1() [??].
5054  *
5055  */
5056 
5057 static void
5058 pmc_process_exit(void *arg __unused, struct proc *p)
5059 {
5060 	struct pmc *pm;
5061 	int adjri, cpu;
5062 	unsigned int ri;
5063 	int is_using_hwpmcs;
5064 	struct pmc_owner *po;
5065 	struct pmc_process *pp;
5066 	struct pmc_classdep *pcd;
5067 	pmc_value_t newvalue, tmp;
5068 
5069 	PROC_LOCK(p);
5070 	is_using_hwpmcs = p->p_flag & P_HWPMC;
5071 	PROC_UNLOCK(p);
5072 
5073 	/*
5074 	 * Log a sysexit event to all SS PMC owners.
5075 	 */
5076 	PMC_EPOCH_ENTER();
5077 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
5078 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
5079 		    pmclog_process_sysexit(po, p->p_pid);
5080 	PMC_EPOCH_EXIT();
5081 
5082 	if (!is_using_hwpmcs)
5083 		return;
5084 
5085 	PMC_GET_SX_XLOCK();
5086 	PMCDBG3(PRC,EXT,1,"process-exit proc=%p (%d, %s)", p, p->p_pid,
5087 	    p->p_comm);
5088 
5089 	/*
5090 	 * Since this code is invoked by the last thread in an exiting
5091 	 * process, we would have context switched IN at some prior
5092 	 * point.  However, with PREEMPTION, kernel mode context
5093 	 * switches may happen any time, so we want to disable a
5094 	 * context switch OUT till we get any PMCs targeting this
5095 	 * process off the hardware.
5096 	 *
5097 	 * We also need to atomically remove this process'
5098 	 * entry from our target process hash table, using
5099 	 * PMC_FLAG_REMOVE.
5100 	 */
5101 	PMCDBG3(PRC,EXT,1, "process-exit proc=%p (%d, %s)", p, p->p_pid,
5102 	    p->p_comm);
5103 
5104 	critical_enter(); /* no preemption */
5105 
5106 	cpu = curthread->td_oncpu;
5107 
5108 	if ((pp = pmc_find_process_descriptor(p,
5109 		 PMC_FLAG_REMOVE)) != NULL) {
5110 
5111 		PMCDBG2(PRC,EXT,2,
5112 		    "process-exit proc=%p pmc-process=%p", p, pp);
5113 
5114 		/*
5115 		 * The exiting process could the target of
5116 		 * some PMCs which will be running on
5117 		 * currently executing CPU.
5118 		 *
5119 		 * We need to turn these PMCs off like we
5120 		 * would do at context switch OUT time.
5121 		 */
5122 		for (ri = 0; ri < md->pmd_npmc; ri++) {
5123 
5124 			/*
5125 			 * Pick up the pmc pointer from hardware
5126 			 * state similar to the CSW_OUT code.
5127 			 */
5128 			pm = NULL;
5129 
5130 			pcd = pmc_ri_to_classdep(md, ri, &adjri);
5131 
5132 			(void) (*pcd->pcd_get_config)(cpu, adjri, &pm);
5133 
5134 			PMCDBG2(PRC,EXT,2, "ri=%d pm=%p", ri, pm);
5135 
5136 			if (pm == NULL ||
5137 			    !PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)))
5138 				continue;
5139 
5140 			PMCDBG4(PRC,EXT,2, "ppmcs[%d]=%p pm=%p "
5141 			    "state=%d", ri, pp->pp_pmcs[ri].pp_pmc,
5142 			    pm, pm->pm_state);
5143 
5144 			KASSERT(PMC_TO_ROWINDEX(pm) == ri,
5145 			    ("[pmc,%d] ri mismatch pmc(%d) ri(%d)",
5146 				__LINE__, PMC_TO_ROWINDEX(pm), ri));
5147 
5148 			KASSERT(pm == pp->pp_pmcs[ri].pp_pmc,
5149 			    ("[pmc,%d] pm %p != pp_pmcs[%d] %p",
5150 				__LINE__, pm, ri, pp->pp_pmcs[ri].pp_pmc));
5151 
5152 			KASSERT(counter_u64_fetch(pm->pm_runcount) > 0,
5153 			    ("[pmc,%d] bad runcount ri %d rc %ld",
5154 				 __LINE__, ri, (unsigned long)counter_u64_fetch(pm->pm_runcount)));
5155 
5156 			/*
5157 			 * Change desired state, and then stop if not
5158 			 * stalled. This two-step dance should avoid
5159 			 * race conditions where an interrupt re-enables
5160 			 * the PMC after this code has already checked
5161 			 * the pm_stalled flag.
5162 			 */
5163 			if (pm->pm_pcpu_state[cpu].pps_cpustate) {
5164 				pm->pm_pcpu_state[cpu].pps_cpustate = 0;
5165 				if (!pm->pm_pcpu_state[cpu].pps_stalled) {
5166 					(void) pcd->pcd_stop_pmc(cpu, adjri);
5167 
5168 					if (PMC_TO_MODE(pm) == PMC_MODE_TC) {
5169 						pcd->pcd_read_pmc(cpu, adjri,
5170 						    &newvalue);
5171 						tmp = newvalue -
5172 						    PMC_PCPU_SAVED(cpu,ri);
5173 
5174 						mtx_pool_lock_spin(pmc_mtxpool,
5175 						    pm);
5176 						pm->pm_gv.pm_savedvalue += tmp;
5177 						pp->pp_pmcs[ri].pp_pmcval +=
5178 						    tmp;
5179 						mtx_pool_unlock_spin(
5180 						    pmc_mtxpool, pm);
5181 					}
5182 				}
5183 			}
5184 
5185 			counter_u64_add(pm->pm_runcount, -1);
5186 
5187 			KASSERT((int) counter_u64_fetch(pm->pm_runcount) >= 0,
5188 			    ("[pmc,%d] runcount is %d", __LINE__, ri));
5189 
5190 			(void) pcd->pcd_config_pmc(cpu, adjri, NULL);
5191 		}
5192 
5193 		/*
5194 		 * Inform the MD layer of this pseudo "context switch
5195 		 * out"
5196 		 */
5197 		(void) md->pmd_switch_out(pmc_pcpu[cpu], pp);
5198 
5199 		critical_exit(); /* ok to be pre-empted now */
5200 
5201 		/*
5202 		 * Unlink this process from the PMCs that are
5203 		 * targeting it.  This will send a signal to
5204 		 * all PMC owner's whose PMCs are orphaned.
5205 		 *
5206 		 * Log PMC value at exit time if requested.
5207 		 */
5208 		for (ri = 0; ri < md->pmd_npmc; ri++)
5209 			if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) {
5210 				if (pm->pm_flags & PMC_F_NEEDS_LOGFILE &&
5211 				    PMC_IS_COUNTING_MODE(PMC_TO_MODE(pm)))
5212 					pmclog_process_procexit(pm, pp);
5213 				pmc_unlink_target_process(pm, pp);
5214 			}
5215 		free(pp, M_PMC);
5216 
5217 	} else
5218 		critical_exit(); /* pp == NULL */
5219 
5220 
5221 	/*
5222 	 * If the process owned PMCs, free them up and free up
5223 	 * memory.
5224 	 */
5225 	if ((po = pmc_find_owner_descriptor(p)) != NULL) {
5226 		pmc_remove_owner(po);
5227 		pmc_destroy_owner_descriptor(po);
5228 	}
5229 
5230 	sx_xunlock(&pmc_sx);
5231 }
5232 
5233 /*
5234  * Handle a process fork.
5235  *
5236  * If the parent process 'p1' is under HWPMC monitoring, then copy
5237  * over any attached PMCs that have 'do_descendants' semantics.
5238  */
5239 
5240 static void
5241 pmc_process_fork(void *arg __unused, struct proc *p1, struct proc *newproc,
5242     int flags)
5243 {
5244 	int is_using_hwpmcs;
5245 	unsigned int ri;
5246 	uint32_t do_descendants;
5247 	struct pmc *pm;
5248 	struct pmc_owner *po;
5249 	struct pmc_process *ppnew, *ppold;
5250 
5251 	(void) flags;		/* unused parameter */
5252 
5253 	PROC_LOCK(p1);
5254 	is_using_hwpmcs = p1->p_flag & P_HWPMC;
5255 	PROC_UNLOCK(p1);
5256 
5257 	/*
5258 	 * If there are system-wide sampling PMCs active, we need to
5259 	 * log all fork events to their owner's logs.
5260 	 */
5261 	PMC_EPOCH_ENTER();
5262 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
5263 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE) {
5264 		    pmclog_process_procfork(po, p1->p_pid, newproc->p_pid);
5265 			pmclog_process_proccreate(po, newproc, 1);
5266 		}
5267 	PMC_EPOCH_EXIT();
5268 
5269 	if (!is_using_hwpmcs)
5270 		return;
5271 
5272 	PMC_GET_SX_XLOCK();
5273 	PMCDBG4(PMC,FRK,1, "process-fork proc=%p (%d, %s) -> %p", p1,
5274 	    p1->p_pid, p1->p_comm, newproc);
5275 
5276 	/*
5277 	 * If the parent process (curthread->td_proc) is a
5278 	 * target of any PMCs, look for PMCs that are to be
5279 	 * inherited, and link these into the new process
5280 	 * descriptor.
5281 	 */
5282 	if ((ppold = pmc_find_process_descriptor(curthread->td_proc,
5283 		 PMC_FLAG_NONE)) == NULL)
5284 		goto done;		/* nothing to do */
5285 
5286 	do_descendants = 0;
5287 	for (ri = 0; ri < md->pmd_npmc; ri++)
5288 		if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL)
5289 			do_descendants |= pm->pm_flags & PMC_F_DESCENDANTS;
5290 	if (do_descendants == 0) /* nothing to do */
5291 		goto done;
5292 
5293 	/*
5294 	 * Now mark the new process as being tracked by this driver.
5295 	 */
5296 	PROC_LOCK(newproc);
5297 	newproc->p_flag |= P_HWPMC;
5298 	PROC_UNLOCK(newproc);
5299 
5300 	/* allocate a descriptor for the new process  */
5301 	if ((ppnew = pmc_find_process_descriptor(newproc,
5302 		 PMC_FLAG_ALLOCATE)) == NULL)
5303 		goto done;
5304 
5305 	/*
5306 	 * Run through all PMCs that were targeting the old process
5307 	 * and which specified F_DESCENDANTS and attach them to the
5308 	 * new process.
5309 	 *
5310 	 * Log the fork event to all owners of PMCs attached to this
5311 	 * process, if not already logged.
5312 	 */
5313 	for (ri = 0; ri < md->pmd_npmc; ri++)
5314 		if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL &&
5315 		    (pm->pm_flags & PMC_F_DESCENDANTS)) {
5316 			pmc_link_target_process(pm, ppnew);
5317 			po = pm->pm_owner;
5318 			if (po->po_sscount == 0 &&
5319 			    po->po_flags & PMC_PO_OWNS_LOGFILE)
5320 				pmclog_process_procfork(po, p1->p_pid,
5321 				    newproc->p_pid);
5322 		}
5323 
5324  done:
5325 	sx_xunlock(&pmc_sx);
5326 }
5327 
5328 static void
5329 pmc_process_threadcreate(struct thread *td)
5330 {
5331 	struct pmc_owner *po;
5332 
5333 	PMC_EPOCH_ENTER();
5334 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
5335 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
5336 			pmclog_process_threadcreate(po, td, 1);
5337 	PMC_EPOCH_EXIT();
5338 }
5339 
5340 static void
5341 pmc_process_threadexit(struct thread *td)
5342 {
5343 	struct pmc_owner *po;
5344 
5345 	PMC_EPOCH_ENTER();
5346 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
5347 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
5348 			pmclog_process_threadexit(po, td);
5349 	PMC_EPOCH_EXIT();
5350 }
5351 
5352 static void
5353 pmc_process_proccreate(struct proc *p)
5354 {
5355 	struct pmc_owner *po;
5356 
5357 	PMC_EPOCH_ENTER();
5358 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
5359 	    if (po->po_flags & PMC_PO_OWNS_LOGFILE)
5360 			pmclog_process_proccreate(po, p, 1 /* sync */);
5361 	PMC_EPOCH_EXIT();
5362 }
5363 
5364 static void
5365 pmc_process_allproc(struct pmc *pm)
5366 {
5367 	struct pmc_owner *po;
5368 	struct thread *td;
5369 	struct proc *p;
5370 
5371 	po = pm->pm_owner;
5372 	if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
5373 		return;
5374 	sx_slock(&allproc_lock);
5375 	FOREACH_PROC_IN_SYSTEM(p) {
5376 		pmclog_process_proccreate(po, p, 0 /* sync */);
5377 		PROC_LOCK(p);
5378 		FOREACH_THREAD_IN_PROC(p, td)
5379 			pmclog_process_threadcreate(po, td, 0 /* sync */);
5380 		PROC_UNLOCK(p);
5381 	}
5382 	sx_sunlock(&allproc_lock);
5383 	pmclog_flush(po, 0);
5384 }
5385 
5386 static void
5387 pmc_kld_load(void *arg __unused, linker_file_t lf)
5388 {
5389 	struct pmc_owner *po;
5390 
5391 	/*
5392 	 * Notify owners of system sampling PMCs about KLD operations.
5393 	 */
5394 	PMC_EPOCH_ENTER();
5395 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
5396 		if (po->po_flags & PMC_PO_OWNS_LOGFILE)
5397 			pmclog_process_map_in(po, (pid_t) -1,
5398 			    (uintfptr_t) lf->address, lf->filename);
5399 	PMC_EPOCH_EXIT();
5400 
5401 	/*
5402 	 * TODO: Notify owners of (all) process-sampling PMCs too.
5403 	 */
5404 }
5405 
5406 static void
5407 pmc_kld_unload(void *arg __unused, const char *filename __unused,
5408     caddr_t address, size_t size)
5409 {
5410 	struct pmc_owner *po;
5411 
5412 	PMC_EPOCH_ENTER();
5413 	CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
5414 		if (po->po_flags & PMC_PO_OWNS_LOGFILE)
5415 			pmclog_process_map_out(po, (pid_t) -1,
5416 			    (uintfptr_t) address, (uintfptr_t) address + size);
5417 	PMC_EPOCH_EXIT();
5418 
5419 	/*
5420 	 * TODO: Notify owners of process-sampling PMCs.
5421 	 */
5422 }
5423 
5424 /*
5425  * initialization
5426  */
5427 static const char *
5428 pmc_name_of_pmcclass(enum pmc_class class)
5429 {
5430 
5431 	switch (class) {
5432 #undef	__PMC_CLASS
5433 #define	__PMC_CLASS(S,V,D)						\
5434 	case PMC_CLASS_##S:						\
5435 		return #S;
5436 	__PMC_CLASSES();
5437 	default:
5438 		return ("<unknown>");
5439 	}
5440 }
5441 
5442 /*
5443  * Base class initializer: allocate structure and set default classes.
5444  */
5445 struct pmc_mdep *
5446 pmc_mdep_alloc(int nclasses)
5447 {
5448 	struct pmc_mdep *md;
5449 	int	n;
5450 
5451 	/* SOFT + md classes */
5452 	n = 1 + nclasses;
5453 	md = malloc(sizeof(struct pmc_mdep) + n *
5454 	    sizeof(struct pmc_classdep), M_PMC, M_WAITOK|M_ZERO);
5455 	md->pmd_nclass = n;
5456 
5457 	/* Add base class. */
5458 	pmc_soft_initialize(md);
5459 	return md;
5460 }
5461 
5462 void
5463 pmc_mdep_free(struct pmc_mdep *md)
5464 {
5465 	pmc_soft_finalize(md);
5466 	free(md, M_PMC);
5467 }
5468 
5469 static int
5470 generic_switch_in(struct pmc_cpu *pc, struct pmc_process *pp)
5471 {
5472 	(void) pc; (void) pp;
5473 
5474 	return (0);
5475 }
5476 
5477 static int
5478 generic_switch_out(struct pmc_cpu *pc, struct pmc_process *pp)
5479 {
5480 	(void) pc; (void) pp;
5481 
5482 	return (0);
5483 }
5484 
5485 static struct pmc_mdep *
5486 pmc_generic_cpu_initialize(void)
5487 {
5488 	struct pmc_mdep *md;
5489 
5490 	md = pmc_mdep_alloc(0);
5491 
5492 	md->pmd_cputype    = PMC_CPU_GENERIC;
5493 
5494 	md->pmd_pcpu_init  = NULL;
5495 	md->pmd_pcpu_fini  = NULL;
5496 	md->pmd_switch_in  = generic_switch_in;
5497 	md->pmd_switch_out = generic_switch_out;
5498 
5499 	return (md);
5500 }
5501 
5502 static void
5503 pmc_generic_cpu_finalize(struct pmc_mdep *md)
5504 {
5505 	(void) md;
5506 }
5507 
5508 
5509 static int
5510 pmc_initialize(void)
5511 {
5512 	int c, cpu, error, n, ri;
5513 	unsigned int maxcpu, domain;
5514 	struct pcpu *pc;
5515 	struct pmc_binding pb;
5516 	struct pmc_sample *ps;
5517 	struct pmc_classdep *pcd;
5518 	struct pmc_samplebuffer *sb;
5519 
5520 	md = NULL;
5521 	error = 0;
5522 
5523 	pmc_stats.pm_intr_ignored = counter_u64_alloc(M_WAITOK);
5524 	pmc_stats.pm_intr_processed = counter_u64_alloc(M_WAITOK);
5525 	pmc_stats.pm_intr_bufferfull = counter_u64_alloc(M_WAITOK);
5526 	pmc_stats.pm_syscalls = counter_u64_alloc(M_WAITOK);
5527 	pmc_stats.pm_syscall_errors = counter_u64_alloc(M_WAITOK);
5528 	pmc_stats.pm_buffer_requests = counter_u64_alloc(M_WAITOK);
5529 	pmc_stats.pm_buffer_requests_failed = counter_u64_alloc(M_WAITOK);
5530 	pmc_stats.pm_log_sweeps = counter_u64_alloc(M_WAITOK);
5531 	pmc_stats.pm_merges = counter_u64_alloc(M_WAITOK);
5532 	pmc_stats.pm_overwrites = counter_u64_alloc(M_WAITOK);
5533 
5534 #ifdef	HWPMC_DEBUG
5535 	/* parse debug flags first */
5536 	if (TUNABLE_STR_FETCH(PMC_SYSCTL_NAME_PREFIX "debugflags",
5537 		pmc_debugstr, sizeof(pmc_debugstr)))
5538 		pmc_debugflags_parse(pmc_debugstr,
5539 		    pmc_debugstr+strlen(pmc_debugstr));
5540 #endif
5541 
5542 	PMCDBG1(MOD,INI,0, "PMC Initialize (version %x)", PMC_VERSION);
5543 
5544 	/* check kernel version */
5545 	if (pmc_kernel_version != PMC_VERSION) {
5546 		if (pmc_kernel_version == 0)
5547 			printf("hwpmc: this kernel has not been compiled with "
5548 			    "'options HWPMC_HOOKS'.\n");
5549 		else
5550 			printf("hwpmc: kernel version (0x%x) does not match "
5551 			    "module version (0x%x).\n", pmc_kernel_version,
5552 			    PMC_VERSION);
5553 		return EPROGMISMATCH;
5554 	}
5555 
5556 	/*
5557 	 * check sysctl parameters
5558 	 */
5559 
5560 	if (pmc_hashsize <= 0) {
5561 		(void) printf("hwpmc: tunable \"hashsize\"=%d must be "
5562 		    "greater than zero.\n", pmc_hashsize);
5563 		pmc_hashsize = PMC_HASH_SIZE;
5564 	}
5565 
5566 	if (pmc_nsamples <= 0 || pmc_nsamples > 65535) {
5567 		(void) printf("hwpmc: tunable \"nsamples\"=%d out of "
5568 		    "range.\n", pmc_nsamples);
5569 		pmc_nsamples = PMC_NSAMPLES;
5570 	}
5571 
5572 	if (pmc_callchaindepth <= 0 ||
5573 	    pmc_callchaindepth > PMC_CALLCHAIN_DEPTH_MAX) {
5574 		(void) printf("hwpmc: tunable \"callchaindepth\"=%d out of "
5575 		    "range - using %d.\n", pmc_callchaindepth,
5576 		    PMC_CALLCHAIN_DEPTH_MAX);
5577 		pmc_callchaindepth = PMC_CALLCHAIN_DEPTH_MAX;
5578 	}
5579 
5580 	md = pmc_md_initialize();
5581 	if (md == NULL) {
5582 		/* Default to generic CPU. */
5583 		md = pmc_generic_cpu_initialize();
5584 		if (md == NULL)
5585 			return (ENOSYS);
5586         }
5587 
5588 	KASSERT(md->pmd_nclass >= 1 && md->pmd_npmc >= 1,
5589 	    ("[pmc,%d] no classes or pmcs", __LINE__));
5590 
5591 	/* Compute the map from row-indices to classdep pointers. */
5592 	pmc_rowindex_to_classdep = malloc(sizeof(struct pmc_classdep *) *
5593 	    md->pmd_npmc, M_PMC, M_WAITOK|M_ZERO);
5594 
5595 	for (n = 0; n < md->pmd_npmc; n++)
5596 		pmc_rowindex_to_classdep[n] = NULL;
5597 	for (ri = c = 0; c < md->pmd_nclass; c++) {
5598 		pcd = &md->pmd_classdep[c];
5599 		for (n = 0; n < pcd->pcd_num; n++, ri++)
5600 			pmc_rowindex_to_classdep[ri] = pcd;
5601 	}
5602 
5603 	KASSERT(ri == md->pmd_npmc,
5604 	    ("[pmc,%d] npmc miscomputed: ri=%d, md->npmc=%d", __LINE__,
5605 	    ri, md->pmd_npmc));
5606 
5607 	maxcpu = pmc_cpu_max();
5608 
5609 	/* allocate space for the per-cpu array */
5610 	pmc_pcpu = malloc(maxcpu * sizeof(struct pmc_cpu *), M_PMC,
5611 	    M_WAITOK|M_ZERO);
5612 
5613 	/* per-cpu 'saved values' for managing process-mode PMCs */
5614 	pmc_pcpu_saved = malloc(sizeof(pmc_value_t) * maxcpu * md->pmd_npmc,
5615 	    M_PMC, M_WAITOK);
5616 
5617 	/* Perform CPU-dependent initialization. */
5618 	pmc_save_cpu_binding(&pb);
5619 	error = 0;
5620 	for (cpu = 0; error == 0 && cpu < maxcpu; cpu++) {
5621 		if (!pmc_cpu_is_active(cpu))
5622 			continue;
5623 		pmc_select_cpu(cpu);
5624 		pmc_pcpu[cpu] = malloc(sizeof(struct pmc_cpu) +
5625 		    md->pmd_npmc * sizeof(struct pmc_hw *), M_PMC,
5626 		    M_WAITOK|M_ZERO);
5627 		if (md->pmd_pcpu_init)
5628 			error = md->pmd_pcpu_init(md, cpu);
5629 		for (n = 0; error == 0 && n < md->pmd_nclass; n++)
5630 			error = md->pmd_classdep[n].pcd_pcpu_init(md, cpu);
5631 	}
5632 	pmc_restore_cpu_binding(&pb);
5633 
5634 	if (error)
5635 		return (error);
5636 
5637 	/* allocate space for the sample array */
5638 	for (cpu = 0; cpu < maxcpu; cpu++) {
5639 		if (!pmc_cpu_is_active(cpu))
5640 			continue;
5641 		pc = pcpu_find(cpu);
5642 		domain = pc->pc_domain;
5643 		sb = malloc_domain(sizeof(struct pmc_samplebuffer) +
5644 			pmc_nsamples * sizeof(struct pmc_sample), M_PMC, domain,
5645 		    M_WAITOK|M_ZERO);
5646 		sb->ps_read = sb->ps_write = sb->ps_samples;
5647 		sb->ps_fence = sb->ps_samples + pmc_nsamples;
5648 
5649 		KASSERT(pmc_pcpu[cpu] != NULL,
5650 		    ("[pmc,%d] cpu=%d Null per-cpu data", __LINE__, cpu));
5651 
5652 		sb->ps_callchains = malloc_domain(pmc_callchaindepth * pmc_nsamples *
5653 			sizeof(uintptr_t), M_PMC, domain, M_WAITOK|M_ZERO);
5654 
5655 		for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++)
5656 			ps->ps_pc = sb->ps_callchains +
5657 			    (n * pmc_callchaindepth);
5658 
5659 		pmc_pcpu[cpu]->pc_sb[PMC_HR] = sb;
5660 
5661 		sb = malloc_domain(sizeof(struct pmc_samplebuffer) +
5662 			pmc_nsamples * sizeof(struct pmc_sample), M_PMC, domain,
5663 		    M_WAITOK|M_ZERO);
5664 		sb->ps_read = sb->ps_write = sb->ps_samples;
5665 		sb->ps_fence = sb->ps_samples + pmc_nsamples;
5666 
5667 		KASSERT(pmc_pcpu[cpu] != NULL,
5668 		    ("[pmc,%d] cpu=%d Null per-cpu data", __LINE__, cpu));
5669 
5670 		sb->ps_callchains = malloc_domain(pmc_callchaindepth * pmc_nsamples *
5671 			sizeof(uintptr_t), M_PMC, domain, M_WAITOK|M_ZERO);
5672 
5673 		for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++)
5674 			ps->ps_pc = sb->ps_callchains +
5675 			    (n * pmc_callchaindepth);
5676 
5677 		pmc_pcpu[cpu]->pc_sb[PMC_SR] = sb;
5678 
5679 		sb = malloc_domain(sizeof(struct pmc_samplebuffer) +
5680 			pmc_nsamples * sizeof(struct pmc_sample), M_PMC, domain,
5681 		    M_WAITOK|M_ZERO);
5682 		sb->ps_read = sb->ps_write = sb->ps_samples;
5683 		sb->ps_fence = sb->ps_samples + pmc_nsamples;
5684 
5685 		KASSERT(pmc_pcpu[cpu] != NULL,
5686 		    ("[pmc,%d] cpu=%d Null per-cpu data", __LINE__, cpu));
5687 
5688 		sb->ps_callchains = malloc_domain(pmc_callchaindepth * pmc_nsamples *
5689 		    sizeof(uintptr_t), M_PMC, domain, M_WAITOK|M_ZERO);
5690 
5691 		for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++)
5692 			ps->ps_pc = sb->ps_callchains +
5693 			    (n * pmc_callchaindepth);
5694 
5695 		pmc_pcpu[cpu]->pc_sb[PMC_UR] = sb;
5696 	}
5697 
5698 	/* allocate space for the row disposition array */
5699 	pmc_pmcdisp = malloc(sizeof(enum pmc_mode) * md->pmd_npmc,
5700 	    M_PMC, M_WAITOK|M_ZERO);
5701 
5702 	/* mark all PMCs as available */
5703 	for (n = 0; n < (int) md->pmd_npmc; n++)
5704 		PMC_MARK_ROW_FREE(n);
5705 
5706 	/* allocate thread hash tables */
5707 	pmc_ownerhash = hashinit(pmc_hashsize, M_PMC,
5708 	    &pmc_ownerhashmask);
5709 
5710 	pmc_processhash = hashinit(pmc_hashsize, M_PMC,
5711 	    &pmc_processhashmask);
5712 	mtx_init(&pmc_processhash_mtx, "pmc-process-hash", "pmc-leaf",
5713 	    MTX_SPIN);
5714 
5715 	CK_LIST_INIT(&pmc_ss_owners);
5716 	pmc_ss_count = 0;
5717 
5718 	/* allocate a pool of spin mutexes */
5719 	pmc_mtxpool = mtx_pool_create("pmc-leaf", pmc_mtxpool_size,
5720 	    MTX_SPIN);
5721 
5722 	PMCDBG4(MOD,INI,1, "pmc_ownerhash=%p, mask=0x%lx "
5723 	    "targethash=%p mask=0x%lx", pmc_ownerhash, pmc_ownerhashmask,
5724 	    pmc_processhash, pmc_processhashmask);
5725 
5726 	/* Initialize a spin mutex for the thread free list. */
5727 	mtx_init(&pmc_threadfreelist_mtx, "pmc-threadfreelist", "pmc-leaf",
5728 	    MTX_SPIN);
5729 
5730 	/*
5731 	 * Initialize the callout to monitor the thread free list.
5732 	 * This callout will also handle the initial population of the list.
5733 	 */
5734 	taskqgroup_config_gtask_init(NULL, &free_gtask, pmc_thread_descriptor_pool_free_task, "thread descriptor pool free task");
5735 
5736 	/* register process {exit,fork,exec} handlers */
5737 	pmc_exit_tag = EVENTHANDLER_REGISTER(process_exit,
5738 	    pmc_process_exit, NULL, EVENTHANDLER_PRI_ANY);
5739 	pmc_fork_tag = EVENTHANDLER_REGISTER(process_fork,
5740 	    pmc_process_fork, NULL, EVENTHANDLER_PRI_ANY);
5741 
5742 	/* register kld event handlers */
5743 	pmc_kld_load_tag = EVENTHANDLER_REGISTER(kld_load, pmc_kld_load,
5744 	    NULL, EVENTHANDLER_PRI_ANY);
5745 	pmc_kld_unload_tag = EVENTHANDLER_REGISTER(kld_unload, pmc_kld_unload,
5746 	    NULL, EVENTHANDLER_PRI_ANY);
5747 
5748 	/* initialize logging */
5749 	pmclog_initialize();
5750 
5751 	/* set hook functions */
5752 	pmc_intr = md->pmd_intr;
5753 	wmb();
5754 	pmc_hook = pmc_hook_handler;
5755 
5756 	if (error == 0) {
5757 		printf(PMC_MODULE_NAME ":");
5758 		for (n = 0; n < (int) md->pmd_nclass; n++) {
5759 			pcd = &md->pmd_classdep[n];
5760 			printf(" %s/%d/%d/0x%b",
5761 			    pmc_name_of_pmcclass(pcd->pcd_class),
5762 			    pcd->pcd_num,
5763 			    pcd->pcd_width,
5764 			    pcd->pcd_caps,
5765 			    "\20"
5766 			    "\1INT\2USR\3SYS\4EDG\5THR"
5767 			    "\6REA\7WRI\10INV\11QUA\12PRC"
5768 			    "\13TAG\14CSC");
5769 		}
5770 		printf("\n");
5771 	}
5772 
5773 	return (error);
5774 }
5775 
5776 /* prepare to be unloaded */
5777 static void
5778 pmc_cleanup(void)
5779 {
5780 	int c, cpu;
5781 	unsigned int maxcpu;
5782 	struct pmc_ownerhash *ph;
5783 	struct pmc_owner *po, *tmp;
5784 	struct pmc_binding pb;
5785 #ifdef	HWPMC_DEBUG
5786 	struct pmc_processhash *prh;
5787 #endif
5788 
5789 	PMCDBG0(MOD,INI,0, "cleanup");
5790 
5791 	/* switch off sampling */
5792 	CPU_FOREACH(cpu)
5793 		DPCPU_ID_SET(cpu, pmc_sampled, 0);
5794 	pmc_intr = NULL;
5795 
5796 	sx_xlock(&pmc_sx);
5797 	if (pmc_hook == NULL) {	/* being unloaded already */
5798 		sx_xunlock(&pmc_sx);
5799 		return;
5800 	}
5801 
5802 	pmc_hook = NULL; /* prevent new threads from entering module */
5803 
5804 	/* deregister event handlers */
5805 	EVENTHANDLER_DEREGISTER(process_fork, pmc_fork_tag);
5806 	EVENTHANDLER_DEREGISTER(process_exit, pmc_exit_tag);
5807 	EVENTHANDLER_DEREGISTER(kld_load, pmc_kld_load_tag);
5808 	EVENTHANDLER_DEREGISTER(kld_unload, pmc_kld_unload_tag);
5809 
5810 	/* send SIGBUS to all owner threads, free up allocations */
5811 	if (pmc_ownerhash)
5812 		for (ph = pmc_ownerhash;
5813 		     ph <= &pmc_ownerhash[pmc_ownerhashmask];
5814 		     ph++) {
5815 			LIST_FOREACH_SAFE(po, ph, po_next, tmp) {
5816 				pmc_remove_owner(po);
5817 
5818 				/* send SIGBUS to owner processes */
5819 				PMCDBG3(MOD,INI,2, "cleanup signal proc=%p "
5820 				    "(%d, %s)", po->po_owner,
5821 				    po->po_owner->p_pid,
5822 				    po->po_owner->p_comm);
5823 
5824 				PROC_LOCK(po->po_owner);
5825 				kern_psignal(po->po_owner, SIGBUS);
5826 				PROC_UNLOCK(po->po_owner);
5827 
5828 				pmc_destroy_owner_descriptor(po);
5829 			}
5830 		}
5831 
5832 	/* reclaim allocated data structures */
5833 	mtx_destroy(&pmc_threadfreelist_mtx);
5834 	pmc_thread_descriptor_pool_drain();
5835 
5836 	if (pmc_mtxpool)
5837 		mtx_pool_destroy(&pmc_mtxpool);
5838 
5839 	mtx_destroy(&pmc_processhash_mtx);
5840 	taskqgroup_config_gtask_deinit(&free_gtask);
5841 	if (pmc_processhash) {
5842 #ifdef	HWPMC_DEBUG
5843 		struct pmc_process *pp;
5844 
5845 		PMCDBG0(MOD,INI,3, "destroy process hash");
5846 		for (prh = pmc_processhash;
5847 		     prh <= &pmc_processhash[pmc_processhashmask];
5848 		     prh++)
5849 			LIST_FOREACH(pp, prh, pp_next)
5850 			    PMCDBG1(MOD,INI,3, "pid=%d", pp->pp_proc->p_pid);
5851 #endif
5852 
5853 		hashdestroy(pmc_processhash, M_PMC, pmc_processhashmask);
5854 		pmc_processhash = NULL;
5855 	}
5856 
5857 	if (pmc_ownerhash) {
5858 		PMCDBG0(MOD,INI,3, "destroy owner hash");
5859 		hashdestroy(pmc_ownerhash, M_PMC, pmc_ownerhashmask);
5860 		pmc_ownerhash = NULL;
5861 	}
5862 
5863 	KASSERT(CK_LIST_EMPTY(&pmc_ss_owners),
5864 	    ("[pmc,%d] Global SS owner list not empty", __LINE__));
5865 	KASSERT(pmc_ss_count == 0,
5866 	    ("[pmc,%d] Global SS count not empty", __LINE__));
5867 
5868  	/* do processor and pmc-class dependent cleanup */
5869 	maxcpu = pmc_cpu_max();
5870 
5871 	PMCDBG0(MOD,INI,3, "md cleanup");
5872 	if (md) {
5873 		pmc_save_cpu_binding(&pb);
5874 		for (cpu = 0; cpu < maxcpu; cpu++) {
5875 			PMCDBG2(MOD,INI,1,"pmc-cleanup cpu=%d pcs=%p",
5876 			    cpu, pmc_pcpu[cpu]);
5877 			if (!pmc_cpu_is_active(cpu) || pmc_pcpu[cpu] == NULL)
5878 				continue;
5879 			pmc_select_cpu(cpu);
5880 			for (c = 0; c < md->pmd_nclass; c++)
5881 				md->pmd_classdep[c].pcd_pcpu_fini(md, cpu);
5882 			if (md->pmd_pcpu_fini)
5883 				md->pmd_pcpu_fini(md, cpu);
5884 		}
5885 
5886 		if (md->pmd_cputype == PMC_CPU_GENERIC)
5887 			pmc_generic_cpu_finalize(md);
5888 		else
5889 			pmc_md_finalize(md);
5890 
5891 		pmc_mdep_free(md);
5892 		md = NULL;
5893 		pmc_restore_cpu_binding(&pb);
5894 	}
5895 
5896 	/* Free per-cpu descriptors. */
5897 	for (cpu = 0; cpu < maxcpu; cpu++) {
5898 		if (!pmc_cpu_is_active(cpu))
5899 			continue;
5900 		KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_HR] != NULL,
5901 		    ("[pmc,%d] Null hw cpu sample buffer cpu=%d", __LINE__,
5902 			cpu));
5903 		KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_SR] != NULL,
5904 		    ("[pmc,%d] Null sw cpu sample buffer cpu=%d", __LINE__,
5905 			cpu));
5906 		KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_UR] != NULL,
5907 		    ("[pmc,%d] Null userret cpu sample buffer cpu=%d", __LINE__,
5908 			cpu));
5909 		free_domain(pmc_pcpu[cpu]->pc_sb[PMC_HR]->ps_callchains, M_PMC);
5910 		free_domain(pmc_pcpu[cpu]->pc_sb[PMC_HR], M_PMC);
5911 		free_domain(pmc_pcpu[cpu]->pc_sb[PMC_SR]->ps_callchains, M_PMC);
5912 		free_domain(pmc_pcpu[cpu]->pc_sb[PMC_SR], M_PMC);
5913 		free_domain(pmc_pcpu[cpu]->pc_sb[PMC_UR]->ps_callchains, M_PMC);
5914 		free_domain(pmc_pcpu[cpu]->pc_sb[PMC_UR], M_PMC);
5915 		free_domain(pmc_pcpu[cpu], M_PMC);
5916 	}
5917 
5918 	free(pmc_pcpu, M_PMC);
5919 	pmc_pcpu = NULL;
5920 
5921 	free(pmc_pcpu_saved, M_PMC);
5922 	pmc_pcpu_saved = NULL;
5923 
5924 	if (pmc_pmcdisp) {
5925 		free(pmc_pmcdisp, M_PMC);
5926 		pmc_pmcdisp = NULL;
5927 	}
5928 
5929 	if (pmc_rowindex_to_classdep) {
5930 		free(pmc_rowindex_to_classdep, M_PMC);
5931 		pmc_rowindex_to_classdep = NULL;
5932 	}
5933 
5934 	pmclog_shutdown();
5935 	counter_u64_free(pmc_stats.pm_intr_ignored);
5936 	counter_u64_free(pmc_stats.pm_intr_processed);
5937 	counter_u64_free(pmc_stats.pm_intr_bufferfull);
5938 	counter_u64_free(pmc_stats.pm_syscalls);
5939 	counter_u64_free(pmc_stats.pm_syscall_errors);
5940 	counter_u64_free(pmc_stats.pm_buffer_requests);
5941 	counter_u64_free(pmc_stats.pm_buffer_requests_failed);
5942 	counter_u64_free(pmc_stats.pm_log_sweeps);
5943 	counter_u64_free(pmc_stats.pm_merges);
5944 	counter_u64_free(pmc_stats.pm_overwrites);
5945 	sx_xunlock(&pmc_sx); 	/* we are done */
5946 }
5947 
5948 /*
5949  * The function called at load/unload.
5950  */
5951 
5952 static int
5953 load (struct module *module __unused, int cmd, void *arg __unused)
5954 {
5955 	int error;
5956 
5957 	error = 0;
5958 
5959 	switch (cmd) {
5960 	case MOD_LOAD :
5961 		/* initialize the subsystem */
5962 		error = pmc_initialize();
5963 		if (error != 0)
5964 			break;
5965 		PMCDBG2(MOD,INI,1, "syscall=%d maxcpu=%d",
5966 		    pmc_syscall_num, pmc_cpu_max());
5967 		break;
5968 
5969 
5970 	case MOD_UNLOAD :
5971 	case MOD_SHUTDOWN:
5972 		pmc_cleanup();
5973 		PMCDBG0(MOD,INI,1, "unloaded");
5974 		break;
5975 
5976 	default :
5977 		error = EINVAL;	/* XXX should panic(9) */
5978 		break;
5979 	}
5980 
5981 	return error;
5982 }
5983