xref: /freebsd/sys/dev/hwpmc/hwpmc_logging.c (revision dd41de95a84d979615a2ef11df6850622bf6184e)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3  *
4  * Copyright (c) 2005-2007 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 /*
36  * Logging code for hwpmc(4)
37  */
38 
39 #include <sys/cdefs.h>
40 __FBSDID("$FreeBSD$");
41 
42 #include <sys/param.h>
43 #include <sys/capsicum.h>
44 #include <sys/domainset.h>
45 #include <sys/file.h>
46 #include <sys/kernel.h>
47 #include <sys/kthread.h>
48 #include <sys/lock.h>
49 #include <sys/module.h>
50 #include <sys/mutex.h>
51 #include <sys/pmc.h>
52 #include <sys/pmckern.h>
53 #include <sys/pmclog.h>
54 #include <sys/proc.h>
55 #include <sys/sched.h>
56 #include <sys/signalvar.h>
57 #include <sys/smp.h>
58 #include <sys/syscallsubr.h>
59 #include <sys/sysctl.h>
60 #include <sys/systm.h>
61 #include <sys/uio.h>
62 #include <sys/unistd.h>
63 #include <sys/vnode.h>
64 
65 #if defined(__i386__) || defined(__amd64__)
66 #include <machine/clock.h>
67 #endif
68 
69 #define curdomain PCPU_GET(domain)
70 
71 /*
72  * Sysctl tunables
73  */
74 
75 SYSCTL_DECL(_kern_hwpmc);
76 
77 /*
78  * kern.hwpmc.logbuffersize -- size of the per-cpu owner buffers.
79  */
80 
81 static int pmclog_buffer_size = PMC_LOG_BUFFER_SIZE;
82 #if (__FreeBSD_version < 1100000)
83 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "logbuffersize", &pmclog_buffer_size);
84 #endif
85 SYSCTL_INT(_kern_hwpmc, OID_AUTO, logbuffersize, CTLFLAG_RDTUN,
86     &pmclog_buffer_size, 0, "size of log buffers in kilobytes");
87 
88 /*
89  * kern.hwpmc.nbuffer -- number of global log buffers
90  */
91 
92 static int pmc_nlogbuffers_pcpu = PMC_NLOGBUFFERS_PCPU;
93 #if (__FreeBSD_version < 1100000)
94 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "nbuffers", &pmc_nlogbuffers_pcpu);
95 #endif
96 SYSCTL_INT(_kern_hwpmc, OID_AUTO, nbuffers_pcpu, CTLFLAG_RDTUN,
97     &pmc_nlogbuffers_pcpu, 0, "number of log buffers per cpu");
98 
99 /*
100  * Global log buffer list and associated spin lock.
101  */
102 
103 static struct mtx pmc_kthread_mtx;	/* sleep lock */
104 
105 #define	PMCLOG_INIT_BUFFER_DESCRIPTOR(D, buf, domain) do {						\
106 		(D)->plb_fence = ((char *) (buf)) +	1024*pmclog_buffer_size;			\
107 		(D)->plb_base  = (D)->plb_ptr = ((char *) (buf));				\
108 		(D)->plb_domain = domain; \
109 	} while (0)
110 
111 #define	PMCLOG_RESET_BUFFER_DESCRIPTOR(D) do {			\
112 		(D)->plb_ptr  = (D)->plb_base; \
113 	} while (0)
114 
115 /*
116  * Log file record constructors.
117  */
118 #define	_PMCLOG_TO_HEADER(T,L)						\
119 	((PMCLOG_HEADER_MAGIC << 24) |					\
120 	 (PMCLOG_TYPE_ ## T << 16)   |					\
121 	 ((L) & 0xFFFF))
122 
123 /* reserve LEN bytes of space and initialize the entry header */
124 #define	_PMCLOG_RESERVE_SAFE(PO,TYPE,LEN,ACTION, TSC) do {	\
125 		uint32_t *_le;						\
126 		int _len = roundup((LEN), sizeof(uint32_t));	\
127 		struct pmclog_header *ph;							\
128 		if ((_le = pmclog_reserve((PO), _len)) == NULL) {	\
129 			ACTION;											\
130 		}													\
131 		ph = (struct pmclog_header *)_le;					\
132 		ph->pl_header =_PMCLOG_TO_HEADER(TYPE,_len);	\
133 		ph->pl_tsc = (TSC);									\
134 		_le += sizeof(*ph)/4	/* skip over timestamp */
135 
136 /* reserve LEN bytes of space and initialize the entry header */
137 #define	_PMCLOG_RESERVE(PO,TYPE,LEN,ACTION) do {			\
138 		uint32_t *_le;						\
139 		int _len = roundup((LEN), sizeof(uint32_t));	\
140 		uint64_t tsc;										\
141 		struct pmclog_header *ph;							\
142 		tsc = pmc_rdtsc();									\
143 		spinlock_enter();									\
144 		if ((_le = pmclog_reserve((PO), _len)) == NULL) {	\
145 			spinlock_exit();								\
146 			ACTION;											\
147 		}												\
148 		ph = (struct pmclog_header *)_le;					\
149 		ph->pl_header =_PMCLOG_TO_HEADER(TYPE,_len);	\
150 		ph->pl_tsc = tsc;									\
151 		_le += sizeof(*ph)/4	/* skip over timestamp */
152 
153 
154 
155 #define	PMCLOG_RESERVE_SAFE(P,T,L,TSC)		_PMCLOG_RESERVE_SAFE(P,T,L,return,TSC)
156 #define	PMCLOG_RESERVE(P,T,L)		_PMCLOG_RESERVE(P,T,L,return)
157 #define	PMCLOG_RESERVE_WITH_ERROR(P,T,L) _PMCLOG_RESERVE(P,T,L,		\
158 	error=ENOMEM;goto error)
159 
160 #define	PMCLOG_EMIT32(V)	do { *_le++ = (V); } while (0)
161 #define	PMCLOG_EMIT64(V)	do { 					\
162 		*_le++ = (uint32_t) ((V) & 0xFFFFFFFF);			\
163 		*_le++ = (uint32_t) (((V) >> 32) & 0xFFFFFFFF);		\
164 	} while (0)
165 
166 
167 /* Emit a string.  Caution: does NOT update _le, so needs to be last */
168 #define	PMCLOG_EMITSTRING(S,L)	do { bcopy((S), _le, (L)); } while (0)
169 #define	PMCLOG_EMITNULLSTRING(L) do { bzero(_le, (L)); } while (0)
170 
171 #define	PMCLOG_DESPATCH_SAFE(PO)						\
172 	    pmclog_release((PO));						\
173 	} while (0)
174 
175 #define	PMCLOG_DESPATCH_SCHED_LOCK(PO)						\
176 	     pmclog_release_flags((PO), 0);							\
177 	} while (0)
178 
179 #define	PMCLOG_DESPATCH(PO)							\
180 	    pmclog_release((PO));						\
181 		spinlock_exit();							\
182 	} while (0)
183 
184 #define	PMCLOG_DESPATCH_SYNC(PO)						\
185 	    pmclog_schedule_io((PO), 1);						\
186 		spinlock_exit();								\
187 		} while (0)
188 
189 
190 #define TSDELTA 4
191 /*
192  * Assertions about the log file format.
193  */
194 CTASSERT(sizeof(struct pmclog_callchain) == 7*4 + TSDELTA +
195     PMC_CALLCHAIN_DEPTH_MAX*sizeof(uintfptr_t));
196 CTASSERT(sizeof(struct pmclog_closelog) == 3*4 + TSDELTA);
197 CTASSERT(sizeof(struct pmclog_dropnotify) == 3*4 + TSDELTA);
198 CTASSERT(sizeof(struct pmclog_map_in) == PATH_MAX + TSDELTA +
199     5*4 + sizeof(uintfptr_t));
200 CTASSERT(offsetof(struct pmclog_map_in,pl_pathname) ==
201     5*4 + TSDELTA + sizeof(uintfptr_t));
202 CTASSERT(sizeof(struct pmclog_map_out) == 5*4 + 2*sizeof(uintfptr_t) + TSDELTA);
203 CTASSERT(sizeof(struct pmclog_pmcallocate) == 9*4 + TSDELTA);
204 CTASSERT(sizeof(struct pmclog_pmcattach) == 5*4 + PATH_MAX + TSDELTA);
205 CTASSERT(offsetof(struct pmclog_pmcattach,pl_pathname) == 5*4 + TSDELTA);
206 CTASSERT(sizeof(struct pmclog_pmcdetach) == 5*4 + TSDELTA);
207 CTASSERT(sizeof(struct pmclog_proccsw) == 7*4 + 8 + TSDELTA);
208 CTASSERT(sizeof(struct pmclog_procexec) == 5*4 + PATH_MAX +
209     sizeof(uintfptr_t) + TSDELTA);
210 CTASSERT(offsetof(struct pmclog_procexec,pl_pathname) == 5*4 + TSDELTA +
211     sizeof(uintfptr_t));
212 CTASSERT(sizeof(struct pmclog_procexit) == 5*4 + 8 + TSDELTA);
213 CTASSERT(sizeof(struct pmclog_procfork) == 5*4 + TSDELTA);
214 CTASSERT(sizeof(struct pmclog_sysexit) == 6*4);
215 CTASSERT(sizeof(struct pmclog_userdata) == 6*4);
216 
217 /*
218  * Log buffer structure
219  */
220 
221 struct pmclog_buffer {
222 	TAILQ_ENTRY(pmclog_buffer) plb_next;
223 	char 		*plb_base;
224 	char		*plb_ptr;
225 	char 		*plb_fence;
226 	uint16_t	 plb_domain;
227 } __aligned(CACHE_LINE_SIZE);
228 
229 /*
230  * Prototypes
231  */
232 
233 static int pmclog_get_buffer(struct pmc_owner *po);
234 static void pmclog_loop(void *arg);
235 static void pmclog_release(struct pmc_owner *po);
236 static uint32_t *pmclog_reserve(struct pmc_owner *po, int length);
237 static void pmclog_schedule_io(struct pmc_owner *po, int wakeup);
238 static void pmclog_schedule_all(struct pmc_owner *po);
239 static void pmclog_stop_kthread(struct pmc_owner *po);
240 
241 /*
242  * Helper functions
243  */
244 
245 static inline void
246 pmc_plb_rele_unlocked(struct pmclog_buffer *plb)
247 {
248 	TAILQ_INSERT_HEAD(&pmc_dom_hdrs[plb->plb_domain]->pdbh_head, plb, plb_next);
249 }
250 
251 static inline void
252 pmc_plb_rele(struct pmclog_buffer *plb)
253 {
254 	mtx_lock_spin(&pmc_dom_hdrs[plb->plb_domain]->pdbh_mtx);
255 	pmc_plb_rele_unlocked(plb);
256 	mtx_unlock_spin(&pmc_dom_hdrs[plb->plb_domain]->pdbh_mtx);
257 }
258 
259 /*
260  * Get a log buffer
261  */
262 static int
263 pmclog_get_buffer(struct pmc_owner *po)
264 {
265 	struct pmclog_buffer *plb;
266 	int domain;
267 
268 	KASSERT(po->po_curbuf[curcpu] == NULL,
269 	    ("[pmclog,%d] po=%p current buffer still valid", __LINE__, po));
270 
271 	domain = curdomain;
272 	MPASS(pmc_dom_hdrs[domain]);
273 	mtx_lock_spin(&pmc_dom_hdrs[domain]->pdbh_mtx);
274 	if ((plb = TAILQ_FIRST(&pmc_dom_hdrs[domain]->pdbh_head)) != NULL)
275 		TAILQ_REMOVE(&pmc_dom_hdrs[domain]->pdbh_head, plb, plb_next);
276 	mtx_unlock_spin(&pmc_dom_hdrs[domain]->pdbh_mtx);
277 
278 	PMCDBG2(LOG,GTB,1, "po=%p plb=%p", po, plb);
279 
280 #ifdef	HWPMC_DEBUG
281 	if (plb)
282 		KASSERT(plb->plb_ptr == plb->plb_base &&
283 		    plb->plb_base < plb->plb_fence,
284 		    ("[pmclog,%d] po=%p buffer invariants: ptr=%p "
285 		    "base=%p fence=%p", __LINE__, po, plb->plb_ptr,
286 		    plb->plb_base, plb->plb_fence));
287 #endif
288 
289 	po->po_curbuf[curcpu] = plb;
290 
291 	/* update stats */
292 	counter_u64_add(pmc_stats.pm_buffer_requests, 1);
293 	if (plb == NULL)
294 		counter_u64_add(pmc_stats.pm_buffer_requests_failed, 1);
295 
296 	return (plb ? 0 : ENOMEM);
297 }
298 
299 struct pmclog_proc_init_args {
300 	struct proc *kthr;
301 	struct pmc_owner *po;
302 	bool exit;
303 	bool acted;
304 };
305 
306 int
307 pmclog_proc_create(struct thread *td, void **handlep)
308 {
309 	struct pmclog_proc_init_args *ia;
310 	int error;
311 
312 	ia = malloc(sizeof(*ia), M_TEMP, M_WAITOK | M_ZERO);
313 	error = kproc_create(pmclog_loop, ia, &ia->kthr,
314 	    RFHIGHPID, 0, "hwpmc: proc(%d)", td->td_proc->p_pid);
315 	if (error == 0)
316 		*handlep = ia;
317 	return (error);
318 }
319 
320 void
321 pmclog_proc_ignite(void *handle, struct pmc_owner *po)
322 {
323 	struct pmclog_proc_init_args *ia;
324 
325 	ia = handle;
326 	mtx_lock(&pmc_kthread_mtx);
327 	MPASS(!ia->acted);
328 	MPASS(ia->po == NULL);
329 	MPASS(!ia->exit);
330 	MPASS(ia->kthr != NULL);
331 	if (po == NULL) {
332 		ia->exit = true;
333 	} else {
334 		ia->po = po;
335 		KASSERT(po->po_kthread == NULL,
336 		    ("[pmclog,%d] po=%p kthread (%p) already present",
337 		    __LINE__, po, po->po_kthread));
338 		po->po_kthread = ia->kthr;
339 	}
340 	wakeup(ia);
341 	while (!ia->acted)
342 		msleep(ia, &pmc_kthread_mtx, PWAIT, "pmclogw", 0);
343 	mtx_unlock(&pmc_kthread_mtx);
344 	free(ia, M_TEMP);
345 }
346 
347 /*
348  * Log handler loop.
349  *
350  * This function is executed by each pmc owner's helper thread.
351  */
352 static void
353 pmclog_loop(void *arg)
354 {
355 	struct pmclog_proc_init_args *ia;
356 	struct pmc_owner *po;
357 	struct pmclog_buffer *lb;
358 	struct proc *p;
359 	struct ucred *ownercred;
360 	struct ucred *mycred;
361 	struct thread *td;
362 	sigset_t unb;
363 	struct uio auio;
364 	struct iovec aiov;
365 	size_t nbytes;
366 	int error;
367 
368 	td = curthread;
369 
370 	SIGEMPTYSET(unb);
371 	SIGADDSET(unb, SIGHUP);
372 	(void)kern_sigprocmask(td, SIG_UNBLOCK, &unb, NULL, 0);
373 
374 	ia = arg;
375 	MPASS(ia->kthr == curproc);
376 	MPASS(!ia->acted);
377 	mtx_lock(&pmc_kthread_mtx);
378 	while (ia->po == NULL && !ia->exit)
379 		msleep(ia, &pmc_kthread_mtx, PWAIT, "pmclogi", 0);
380 	if (ia->exit) {
381 		ia->acted = true;
382 		wakeup(ia);
383 		mtx_unlock(&pmc_kthread_mtx);
384 		kproc_exit(0);
385 	}
386 	MPASS(ia->po != NULL);
387 	po = ia->po;
388 	ia->acted = true;
389 	wakeup(ia);
390 	mtx_unlock(&pmc_kthread_mtx);
391 	ia = NULL;
392 
393 	p = po->po_owner;
394 	mycred = td->td_ucred;
395 
396 	PROC_LOCK(p);
397 	ownercred = crhold(p->p_ucred);
398 	PROC_UNLOCK(p);
399 
400 	PMCDBG2(LOG,INI,1, "po=%p kt=%p", po, po->po_kthread);
401 	KASSERT(po->po_kthread == curthread->td_proc,
402 	    ("[pmclog,%d] proc mismatch po=%p po/kt=%p curproc=%p", __LINE__,
403 		po, po->po_kthread, curthread->td_proc));
404 
405 	lb = NULL;
406 
407 
408 	/*
409 	 * Loop waiting for I/O requests to be added to the owner
410 	 * struct's queue.  The loop is exited when the log file
411 	 * is deconfigured.
412 	 */
413 
414 	mtx_lock(&pmc_kthread_mtx);
415 
416 	for (;;) {
417 
418 		/* check if we've been asked to exit */
419 		if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
420 			break;
421 
422 		if (lb == NULL) { /* look for a fresh buffer to write */
423 			mtx_lock_spin(&po->po_mtx);
424 			if ((lb = TAILQ_FIRST(&po->po_logbuffers)) == NULL) {
425 				mtx_unlock_spin(&po->po_mtx);
426 
427 				/* No more buffers and shutdown required. */
428 				if (po->po_flags & PMC_PO_SHUTDOWN)
429 					break;
430 
431 				(void) msleep(po, &pmc_kthread_mtx, PWAIT,
432 				    "pmcloop", 250);
433 				continue;
434 			}
435 
436 			TAILQ_REMOVE(&po->po_logbuffers, lb, plb_next);
437 			mtx_unlock_spin(&po->po_mtx);
438 		}
439 
440 		mtx_unlock(&pmc_kthread_mtx);
441 
442 		/* process the request */
443 		PMCDBG3(LOG,WRI,2, "po=%p base=%p ptr=%p", po,
444 		    lb->plb_base, lb->plb_ptr);
445 		/* change our thread's credentials before issuing the I/O */
446 
447 		aiov.iov_base = lb->plb_base;
448 		aiov.iov_len  = nbytes = lb->plb_ptr - lb->plb_base;
449 
450 		auio.uio_iov    = &aiov;
451 		auio.uio_iovcnt = 1;
452 		auio.uio_offset = -1;
453 		auio.uio_resid  = nbytes;
454 		auio.uio_rw     = UIO_WRITE;
455 		auio.uio_segflg = UIO_SYSSPACE;
456 		auio.uio_td     = td;
457 
458 		/* switch thread credentials -- see kern_ktrace.c */
459 		td->td_ucred = ownercred;
460 		error = fo_write(po->po_file, &auio, ownercred, 0, td);
461 		td->td_ucred = mycred;
462 
463 		if (error) {
464 			/* XXX some errors are recoverable */
465 			/* send a SIGIO to the owner and exit */
466 			PROC_LOCK(p);
467 			kern_psignal(p, SIGIO);
468 			PROC_UNLOCK(p);
469 
470 			mtx_lock(&pmc_kthread_mtx);
471 
472 			po->po_error = error; /* save for flush log */
473 
474 			PMCDBG2(LOG,WRI,2, "po=%p error=%d", po, error);
475 
476 			break;
477 		}
478 
479 		mtx_lock(&pmc_kthread_mtx);
480 
481 		/* put the used buffer back into the global pool */
482 		PMCLOG_RESET_BUFFER_DESCRIPTOR(lb);
483 
484 		pmc_plb_rele(lb);
485 		lb = NULL;
486 	}
487 
488 	wakeup_one(po->po_kthread);
489 	po->po_kthread = NULL;
490 
491 	mtx_unlock(&pmc_kthread_mtx);
492 
493 	/* return the current I/O buffer to the global pool */
494 	if (lb) {
495 		PMCLOG_RESET_BUFFER_DESCRIPTOR(lb);
496 
497 		pmc_plb_rele(lb);
498 	}
499 
500 	/*
501 	 * Exit this thread, signalling the waiter
502 	 */
503 
504 	crfree(ownercred);
505 
506 	kproc_exit(0);
507 }
508 
509 /*
510  * Release and log entry and schedule an I/O if needed.
511  */
512 
513 static void
514 pmclog_release_flags(struct pmc_owner *po, int wakeup)
515 {
516 	struct pmclog_buffer *plb;
517 
518 	plb = po->po_curbuf[curcpu];
519 	KASSERT(plb->plb_ptr >= plb->plb_base,
520 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p base=%p", __LINE__,
521 		po, plb->plb_ptr, plb->plb_base));
522 	KASSERT(plb->plb_ptr <= plb->plb_fence,
523 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p fenc=%p", __LINE__,
524 		po, plb->plb_ptr, plb->plb_fence));
525 
526 	/* schedule an I/O if we've filled a buffer */
527 	if (plb->plb_ptr >= plb->plb_fence)
528 		pmclog_schedule_io(po, wakeup);
529 
530 	PMCDBG1(LOG,REL,1, "po=%p", po);
531 }
532 
533 static void
534 pmclog_release(struct pmc_owner *po)
535 {
536 
537 	pmclog_release_flags(po, 1);
538 }
539 
540 
541 /*
542  * Attempt to reserve 'length' bytes of space in an owner's log
543  * buffer.  The function returns a pointer to 'length' bytes of space
544  * if there was enough space or returns NULL if no space was
545  * available.  Non-null returns do so with the po mutex locked.  The
546  * caller must invoke pmclog_release() on the pmc owner structure
547  * when done.
548  */
549 
550 static uint32_t *
551 pmclog_reserve(struct pmc_owner *po, int length)
552 {
553 	uintptr_t newptr, oldptr;
554 	struct pmclog_buffer *plb, **pplb;
555 
556 	PMCDBG2(LOG,ALL,1, "po=%p len=%d", po, length);
557 
558 	KASSERT(length % sizeof(uint32_t) == 0,
559 	    ("[pmclog,%d] length not a multiple of word size", __LINE__));
560 
561 	/* No more data when shutdown in progress. */
562 	if (po->po_flags & PMC_PO_SHUTDOWN)
563 		return (NULL);
564 
565 	pplb = &po->po_curbuf[curcpu];
566 	if (*pplb == NULL && pmclog_get_buffer(po) != 0)
567 		goto fail;
568 
569 	KASSERT(*pplb != NULL,
570 	    ("[pmclog,%d] po=%p no current buffer", __LINE__, po));
571 
572 	plb = *pplb;
573 	KASSERT(plb->plb_ptr >= plb->plb_base &&
574 	    plb->plb_ptr <= plb->plb_fence,
575 	    ("[pmclog,%d] po=%p buffer invariants: ptr=%p base=%p fence=%p",
576 		__LINE__, po, plb->plb_ptr, plb->plb_base,
577 		plb->plb_fence));
578 
579 	oldptr = (uintptr_t) plb->plb_ptr;
580 	newptr = oldptr + length;
581 
582 	KASSERT(oldptr != (uintptr_t) NULL,
583 	    ("[pmclog,%d] po=%p Null log buffer pointer", __LINE__, po));
584 
585 	/*
586 	 * If we have space in the current buffer, return a pointer to
587 	 * available space with the PO structure locked.
588 	 */
589 	if (newptr <= (uintptr_t) plb->plb_fence) {
590 		plb->plb_ptr = (char *) newptr;
591 		goto done;
592 	}
593 
594 	/*
595 	 * Otherwise, schedule the current buffer for output and get a
596 	 * fresh buffer.
597 	 */
598 	pmclog_schedule_io(po, 0);
599 
600 	if (pmclog_get_buffer(po) != 0)
601 		goto fail;
602 
603 	plb = *pplb;
604 	KASSERT(plb != NULL,
605 	    ("[pmclog,%d] po=%p no current buffer", __LINE__, po));
606 
607 	KASSERT(plb->plb_ptr != NULL,
608 	    ("[pmclog,%d] null return from pmc_get_log_buffer", __LINE__));
609 
610 	KASSERT(plb->plb_ptr == plb->plb_base &&
611 	    plb->plb_ptr <= plb->plb_fence,
612 	    ("[pmclog,%d] po=%p buffer invariants: ptr=%p base=%p fence=%p",
613 		__LINE__, po, plb->plb_ptr, plb->plb_base,
614 		plb->plb_fence));
615 
616 	oldptr = (uintptr_t) plb->plb_ptr;
617 
618  done:
619 	return ((uint32_t *) oldptr);
620  fail:
621 	return (NULL);
622 }
623 
624 /*
625  * Schedule an I/O.
626  *
627  * Transfer the current buffer to the helper kthread.
628  */
629 
630 static void
631 pmclog_schedule_io(struct pmc_owner *po, int wakeup)
632 {
633 	struct pmclog_buffer *plb;
634 
635 	plb = po->po_curbuf[curcpu];
636 	po->po_curbuf[curcpu] = NULL;
637 	KASSERT(plb != NULL,
638 	    ("[pmclog,%d] schedule_io with null buffer po=%p", __LINE__, po));
639 	KASSERT(plb->plb_ptr >= plb->plb_base,
640 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p base=%p", __LINE__,
641 		po, plb->plb_ptr, plb->plb_base));
642 	KASSERT(plb->plb_ptr <= plb->plb_fence,
643 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p fenc=%p", __LINE__,
644 		po, plb->plb_ptr, plb->plb_fence));
645 
646 	PMCDBG1(LOG,SIO, 1, "po=%p", po);
647 
648 	/*
649 	 * Add the current buffer to the tail of the buffer list and
650 	 * wakeup the helper.
651 	 */
652 	mtx_lock_spin(&po->po_mtx);
653 	TAILQ_INSERT_TAIL(&po->po_logbuffers, plb, plb_next);
654 	mtx_unlock_spin(&po->po_mtx);
655 	if (wakeup)
656 		wakeup_one(po);
657 }
658 
659 /*
660  * Stop the helper kthread.
661  */
662 
663 static void
664 pmclog_stop_kthread(struct pmc_owner *po)
665 {
666 
667 	mtx_lock(&pmc_kthread_mtx);
668 	po->po_flags &= ~PMC_PO_OWNS_LOGFILE;
669 	if (po->po_kthread != NULL) {
670 		PROC_LOCK(po->po_kthread);
671 		kern_psignal(po->po_kthread, SIGHUP);
672 		PROC_UNLOCK(po->po_kthread);
673 	}
674 	wakeup_one(po);
675 	while (po->po_kthread)
676 		msleep(po->po_kthread, &pmc_kthread_mtx, PPAUSE, "pmckstp", 0);
677 	mtx_unlock(&pmc_kthread_mtx);
678 }
679 
680 /*
681  * Public functions
682  */
683 
684 /*
685  * Configure a log file for pmc owner 'po'.
686  *
687  * Parameter 'logfd' is a file handle referencing an open file in the
688  * owner process.  This file needs to have been opened for writing.
689  */
690 
691 int
692 pmclog_configure_log(struct pmc_mdep *md, struct pmc_owner *po, int logfd)
693 {
694 	struct proc *p;
695 	struct timespec ts;
696 	uint64_t tsc;
697 	int error;
698 
699 	sx_assert(&pmc_sx, SA_XLOCKED);
700 	PMCDBG2(LOG,CFG,1, "config po=%p logfd=%d", po, logfd);
701 
702 	p = po->po_owner;
703 
704 	/* return EBUSY if a log file was already present */
705 	if (po->po_flags & PMC_PO_OWNS_LOGFILE)
706 		return (EBUSY);
707 
708 	KASSERT(po->po_file == NULL,
709 	    ("[pmclog,%d] po=%p file (%p) already present", __LINE__, po,
710 		po->po_file));
711 
712 	/* get a reference to the file state */
713 	error = fget_write(curthread, logfd, &cap_write_rights, &po->po_file);
714 	if (error)
715 		goto error;
716 
717 	/* mark process as owning a log file */
718 	po->po_flags |= PMC_PO_OWNS_LOGFILE;
719 
720 	/* mark process as using HWPMCs */
721 	PROC_LOCK(p);
722 	p->p_flag |= P_HWPMC;
723 	PROC_UNLOCK(p);
724 	nanotime(&ts);
725 	tsc = pmc_rdtsc();
726 	/* create a log initialization entry */
727 	PMCLOG_RESERVE_WITH_ERROR(po, INITIALIZE,
728 	    sizeof(struct pmclog_initialize));
729 	PMCLOG_EMIT32(PMC_VERSION);
730 	PMCLOG_EMIT32(md->pmd_cputype);
731 #if defined(__i386__) || defined(__amd64__)
732 	PMCLOG_EMIT64(tsc_freq);
733 #else
734 	/* other architectures will need to fill this in */
735 	PMCLOG_EMIT32(0);
736 	PMCLOG_EMIT32(0);
737 #endif
738 	memcpy(_le, &ts, sizeof(ts));
739 	_le += sizeof(ts)/4;
740 	PMCLOG_EMITSTRING(pmc_cpuid, PMC_CPUID_LEN);
741 	PMCLOG_DESPATCH_SYNC(po);
742 
743 	return (0);
744 
745  error:
746 	KASSERT(po->po_kthread == NULL, ("[pmclog,%d] po=%p kthread not "
747 	    "stopped", __LINE__, po));
748 
749 	if (po->po_file)
750 		(void) fdrop(po->po_file, curthread);
751 	po->po_file  = NULL;	/* clear file and error state */
752 	po->po_error = 0;
753 	po->po_flags &= ~PMC_PO_OWNS_LOGFILE;
754 
755 	return (error);
756 }
757 
758 
759 /*
760  * De-configure a log file.  This will throw away any buffers queued
761  * for this owner process.
762  */
763 
764 int
765 pmclog_deconfigure_log(struct pmc_owner *po)
766 {
767 	int error;
768 	struct pmclog_buffer *lb;
769 
770 	PMCDBG1(LOG,CFG,1, "de-config po=%p", po);
771 
772 	if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
773 		return (EINVAL);
774 
775 	KASSERT(po->po_sscount == 0,
776 	    ("[pmclog,%d] po=%p still owning SS PMCs", __LINE__, po));
777 	KASSERT(po->po_file != NULL,
778 	    ("[pmclog,%d] po=%p no log file", __LINE__, po));
779 
780 	/* stop the kthread, this will reset the 'OWNS_LOGFILE' flag */
781 	pmclog_stop_kthread(po);
782 
783 	KASSERT(po->po_kthread == NULL,
784 	    ("[pmclog,%d] po=%p kthread not stopped", __LINE__, po));
785 
786 	/* return all queued log buffers to the global pool */
787 	while ((lb = TAILQ_FIRST(&po->po_logbuffers)) != NULL) {
788 		TAILQ_REMOVE(&po->po_logbuffers, lb, plb_next);
789 		PMCLOG_RESET_BUFFER_DESCRIPTOR(lb);
790 		pmc_plb_rele(lb);
791 	}
792 	for (int i = 0; i < mp_ncpus; i++) {
793 		thread_lock(curthread);
794 		sched_bind(curthread, i);
795 		thread_unlock(curthread);
796 		/* return the 'current' buffer to the global pool */
797 		if ((lb = po->po_curbuf[curcpu]) != NULL) {
798 			PMCLOG_RESET_BUFFER_DESCRIPTOR(lb);
799 			pmc_plb_rele(lb);
800 		}
801 	}
802 	thread_lock(curthread);
803 	sched_unbind(curthread);
804 	thread_unlock(curthread);
805 
806 	/* drop a reference to the fd */
807 	if (po->po_file != NULL) {
808 		error = fdrop(po->po_file, curthread);
809 		po->po_file = NULL;
810 	} else
811 		error = 0;
812 	po->po_error = 0;
813 
814 	return (error);
815 }
816 
817 /*
818  * Flush a process' log buffer.
819  */
820 
821 int
822 pmclog_flush(struct pmc_owner *po, int force)
823 {
824 	int error;
825 
826 	PMCDBG1(LOG,FLS,1, "po=%p", po);
827 
828 	/*
829 	 * If there is a pending error recorded by the logger thread,
830 	 * return that.
831 	 */
832 	if (po->po_error)
833 		return (po->po_error);
834 
835 	error = 0;
836 
837 	/*
838 	 * Check that we do have an active log file.
839 	 */
840 	mtx_lock(&pmc_kthread_mtx);
841 	if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) {
842 		error = EINVAL;
843 		goto error;
844 	}
845 
846 	pmclog_schedule_all(po);
847  error:
848 	mtx_unlock(&pmc_kthread_mtx);
849 
850 	return (error);
851 }
852 
853 static void
854 pmclog_schedule_one_cond(struct pmc_owner *po)
855 {
856 	struct pmclog_buffer *plb;
857 	int cpu;
858 
859 	spinlock_enter();
860 	cpu = curcpu;
861 	/* tell hardclock not to run again */
862 	if (PMC_CPU_HAS_SAMPLES(cpu))
863 		PMC_CALL_HOOK_UNLOCKED(curthread, PMC_FN_DO_SAMPLES, NULL);
864 
865 	plb = po->po_curbuf[cpu];
866 	if (plb && plb->plb_ptr != plb->plb_base)
867 		pmclog_schedule_io(po, 1);
868 	spinlock_exit();
869 }
870 
871 static void
872 pmclog_schedule_all(struct pmc_owner *po)
873 {
874 	/*
875 	 * Schedule the current buffer if any and not empty.
876 	 */
877 	for (int i = 0; i < mp_ncpus; i++) {
878 		thread_lock(curthread);
879 		sched_bind(curthread, i);
880 		thread_unlock(curthread);
881 		pmclog_schedule_one_cond(po);
882 	}
883 	thread_lock(curthread);
884 	sched_unbind(curthread);
885 	thread_unlock(curthread);
886 }
887 
888 int
889 pmclog_close(struct pmc_owner *po)
890 {
891 
892 	PMCDBG1(LOG,CLO,1, "po=%p", po);
893 
894 	pmclog_process_closelog(po);
895 
896 	mtx_lock(&pmc_kthread_mtx);
897 	/*
898 	 * Initiate shutdown: no new data queued,
899 	 * thread will close file on last block.
900 	 */
901 	po->po_flags |= PMC_PO_SHUTDOWN;
902 	/* give time for all to see */
903 	DELAY(50);
904 
905 	/*
906 	 * Schedule the current buffer.
907 	 */
908 	pmclog_schedule_all(po);
909 	wakeup_one(po);
910 
911 	mtx_unlock(&pmc_kthread_mtx);
912 
913 	return (0);
914 }
915 
916 void
917 pmclog_process_callchain(struct pmc *pm, struct pmc_sample *ps)
918 {
919 	int n, recordlen;
920 	uint32_t flags;
921 	struct pmc_owner *po;
922 
923 	PMCDBG3(LOG,SAM,1,"pm=%p pid=%d n=%d", pm, ps->ps_pid,
924 	    ps->ps_nsamples);
925 
926 	recordlen = offsetof(struct pmclog_callchain, pl_pc) +
927 	    ps->ps_nsamples * sizeof(uintfptr_t);
928 	po = pm->pm_owner;
929 	flags = PMC_CALLCHAIN_TO_CPUFLAGS(ps->ps_cpu,ps->ps_flags);
930 	PMCLOG_RESERVE_SAFE(po, CALLCHAIN, recordlen, ps->ps_tsc);
931 	PMCLOG_EMIT32(ps->ps_pid);
932 	PMCLOG_EMIT32(ps->ps_tid);
933 	PMCLOG_EMIT32(pm->pm_id);
934 	PMCLOG_EMIT32(flags);
935 	for (n = 0; n < ps->ps_nsamples; n++)
936 		PMCLOG_EMITADDR(ps->ps_pc[n]);
937 	PMCLOG_DESPATCH_SAFE(po);
938 }
939 
940 void
941 pmclog_process_closelog(struct pmc_owner *po)
942 {
943 	PMCLOG_RESERVE(po,CLOSELOG,sizeof(struct pmclog_closelog));
944 	PMCLOG_DESPATCH_SYNC(po);
945 }
946 
947 void
948 pmclog_process_dropnotify(struct pmc_owner *po)
949 {
950 	PMCLOG_RESERVE(po,DROPNOTIFY,sizeof(struct pmclog_dropnotify));
951 	PMCLOG_DESPATCH(po);
952 }
953 
954 void
955 pmclog_process_map_in(struct pmc_owner *po, pid_t pid, uintfptr_t start,
956     const char *path)
957 {
958 	int pathlen, recordlen;
959 
960 	KASSERT(path != NULL, ("[pmclog,%d] map-in, null path", __LINE__));
961 
962 	pathlen = strlen(path) + 1;	/* #bytes for path name */
963 	recordlen = offsetof(struct pmclog_map_in, pl_pathname) +
964 	    pathlen;
965 
966 	PMCLOG_RESERVE(po, MAP_IN, recordlen);
967 	PMCLOG_EMIT32(pid);
968 	PMCLOG_EMIT32(0);
969 	PMCLOG_EMITADDR(start);
970 	PMCLOG_EMITSTRING(path,pathlen);
971 	PMCLOG_DESPATCH_SYNC(po);
972 }
973 
974 void
975 pmclog_process_map_out(struct pmc_owner *po, pid_t pid, uintfptr_t start,
976     uintfptr_t end)
977 {
978 	KASSERT(start <= end, ("[pmclog,%d] start > end", __LINE__));
979 
980 	PMCLOG_RESERVE(po, MAP_OUT, sizeof(struct pmclog_map_out));
981 	PMCLOG_EMIT32(pid);
982 	PMCLOG_EMIT32(0);
983 	PMCLOG_EMITADDR(start);
984 	PMCLOG_EMITADDR(end);
985 	PMCLOG_DESPATCH(po);
986 }
987 
988 void
989 pmclog_process_pmcallocate(struct pmc *pm)
990 {
991 	struct pmc_owner *po;
992 	struct pmc_soft *ps;
993 
994 	po = pm->pm_owner;
995 
996 	PMCDBG1(LOG,ALL,1, "pm=%p", pm);
997 
998 	if (PMC_TO_CLASS(pm) == PMC_CLASS_SOFT) {
999 		PMCLOG_RESERVE(po, PMCALLOCATEDYN,
1000 		    sizeof(struct pmclog_pmcallocatedyn));
1001 		PMCLOG_EMIT32(pm->pm_id);
1002 		PMCLOG_EMIT32(pm->pm_event);
1003 		PMCLOG_EMIT32(pm->pm_flags);
1004 		PMCLOG_EMIT32(0);
1005 		PMCLOG_EMIT64(pm->pm_sc.pm_reloadcount);
1006 		ps = pmc_soft_ev_acquire(pm->pm_event);
1007 		if (ps != NULL)
1008 			PMCLOG_EMITSTRING(ps->ps_ev.pm_ev_name,PMC_NAME_MAX);
1009 		else
1010 			PMCLOG_EMITNULLSTRING(PMC_NAME_MAX);
1011 		pmc_soft_ev_release(ps);
1012 		PMCLOG_DESPATCH_SYNC(po);
1013 	} else {
1014 		PMCLOG_RESERVE(po, PMCALLOCATE,
1015 		    sizeof(struct pmclog_pmcallocate));
1016 		PMCLOG_EMIT32(pm->pm_id);
1017 		PMCLOG_EMIT32(pm->pm_event);
1018 		PMCLOG_EMIT32(pm->pm_flags);
1019 		PMCLOG_EMIT32(0);
1020 		PMCLOG_EMIT64(pm->pm_sc.pm_reloadcount);
1021 		PMCLOG_DESPATCH_SYNC(po);
1022 	}
1023 }
1024 
1025 void
1026 pmclog_process_pmcattach(struct pmc *pm, pid_t pid, char *path)
1027 {
1028 	int pathlen, recordlen;
1029 	struct pmc_owner *po;
1030 
1031 	PMCDBG2(LOG,ATT,1,"pm=%p pid=%d", pm, pid);
1032 
1033 	po = pm->pm_owner;
1034 
1035 	pathlen = strlen(path) + 1;	/* #bytes for the string */
1036 	recordlen = offsetof(struct pmclog_pmcattach, pl_pathname) + pathlen;
1037 
1038 	PMCLOG_RESERVE(po, PMCATTACH, recordlen);
1039 	PMCLOG_EMIT32(pm->pm_id);
1040 	PMCLOG_EMIT32(pid);
1041 	PMCLOG_EMITSTRING(path, pathlen);
1042 	PMCLOG_DESPATCH_SYNC(po);
1043 }
1044 
1045 void
1046 pmclog_process_pmcdetach(struct pmc *pm, pid_t pid)
1047 {
1048 	struct pmc_owner *po;
1049 
1050 	PMCDBG2(LOG,ATT,1,"!pm=%p pid=%d", pm, pid);
1051 
1052 	po = pm->pm_owner;
1053 
1054 	PMCLOG_RESERVE(po, PMCDETACH, sizeof(struct pmclog_pmcdetach));
1055 	PMCLOG_EMIT32(pm->pm_id);
1056 	PMCLOG_EMIT32(pid);
1057 	PMCLOG_DESPATCH_SYNC(po);
1058 }
1059 
1060 void
1061 pmclog_process_proccreate(struct pmc_owner *po, struct proc *p, int sync)
1062 {
1063 	if (sync) {
1064 		PMCLOG_RESERVE(po, PROC_CREATE, sizeof(struct pmclog_proccreate));
1065 		PMCLOG_EMIT32(p->p_pid);
1066 		PMCLOG_EMIT32(p->p_flag);
1067 		PMCLOG_EMITSTRING(p->p_comm, MAXCOMLEN+1);
1068 		PMCLOG_DESPATCH_SYNC(po);
1069 	} else {
1070 		PMCLOG_RESERVE(po, PROC_CREATE, sizeof(struct pmclog_proccreate));
1071 		PMCLOG_EMIT32(p->p_pid);
1072 		PMCLOG_EMIT32(p->p_flag);
1073 		PMCLOG_EMITSTRING(p->p_comm, MAXCOMLEN+1);
1074 		PMCLOG_DESPATCH(po);
1075 	}
1076 }
1077 
1078 /*
1079  * Log a context switch event to the log file.
1080  */
1081 
1082 void
1083 pmclog_process_proccsw(struct pmc *pm, struct pmc_process *pp, pmc_value_t v, struct thread *td)
1084 {
1085 	struct pmc_owner *po;
1086 
1087 	KASSERT(pm->pm_flags & PMC_F_LOG_PROCCSW,
1088 	    ("[pmclog,%d] log-process-csw called gratuitously", __LINE__));
1089 
1090 	PMCDBG3(LOG,SWO,1,"pm=%p pid=%d v=%jx", pm, pp->pp_proc->p_pid,
1091 	    v);
1092 
1093 	po = pm->pm_owner;
1094 
1095 	PMCLOG_RESERVE_SAFE(po, PROCCSW, sizeof(struct pmclog_proccsw), pmc_rdtsc());
1096 	PMCLOG_EMIT64(v);
1097 	PMCLOG_EMIT32(pm->pm_id);
1098 	PMCLOG_EMIT32(pp->pp_proc->p_pid);
1099 	PMCLOG_EMIT32(td->td_tid);
1100 	PMCLOG_EMIT32(0);
1101 	PMCLOG_DESPATCH_SCHED_LOCK(po);
1102 }
1103 
1104 void
1105 pmclog_process_procexec(struct pmc_owner *po, pmc_id_t pmid, pid_t pid,
1106     uintfptr_t startaddr, char *path)
1107 {
1108 	int pathlen, recordlen;
1109 
1110 	PMCDBG3(LOG,EXC,1,"po=%p pid=%d path=\"%s\"", po, pid, path);
1111 
1112 	pathlen   = strlen(path) + 1;	/* #bytes for the path */
1113 	recordlen = offsetof(struct pmclog_procexec, pl_pathname) + pathlen;
1114 	PMCLOG_RESERVE(po, PROCEXEC, recordlen);
1115 	PMCLOG_EMIT32(pid);
1116 	PMCLOG_EMIT32(pmid);
1117 	PMCLOG_EMITADDR(startaddr);
1118 	PMCLOG_EMITSTRING(path,pathlen);
1119 	PMCLOG_DESPATCH_SYNC(po);
1120 }
1121 
1122 /*
1123  * Log a process exit event (and accumulated pmc value) to the log file.
1124  */
1125 
1126 void
1127 pmclog_process_procexit(struct pmc *pm, struct pmc_process *pp)
1128 {
1129 	int ri;
1130 	struct pmc_owner *po;
1131 
1132 	ri = PMC_TO_ROWINDEX(pm);
1133 	PMCDBG3(LOG,EXT,1,"pm=%p pid=%d v=%jx", pm, pp->pp_proc->p_pid,
1134 	    pp->pp_pmcs[ri].pp_pmcval);
1135 
1136 	po = pm->pm_owner;
1137 
1138 	PMCLOG_RESERVE(po, PROCEXIT, sizeof(struct pmclog_procexit));
1139 	PMCLOG_EMIT32(pm->pm_id);
1140 	PMCLOG_EMIT32(pp->pp_proc->p_pid);
1141 	PMCLOG_EMIT64(pp->pp_pmcs[ri].pp_pmcval);
1142 	PMCLOG_DESPATCH(po);
1143 }
1144 
1145 /*
1146  * Log a fork event.
1147  */
1148 
1149 void
1150 pmclog_process_procfork(struct pmc_owner *po, pid_t oldpid, pid_t newpid)
1151 {
1152 	PMCLOG_RESERVE(po, PROCFORK, sizeof(struct pmclog_procfork));
1153 	PMCLOG_EMIT32(oldpid);
1154 	PMCLOG_EMIT32(newpid);
1155 	PMCLOG_DESPATCH(po);
1156 }
1157 
1158 /*
1159  * Log a process exit event of the form suitable for system-wide PMCs.
1160  */
1161 
1162 void
1163 pmclog_process_sysexit(struct pmc_owner *po, pid_t pid)
1164 {
1165 	PMCLOG_RESERVE(po, SYSEXIT, sizeof(struct pmclog_sysexit));
1166 	PMCLOG_EMIT32(pid);
1167 	PMCLOG_DESPATCH(po);
1168 }
1169 
1170 void
1171 pmclog_process_threadcreate(struct pmc_owner *po, struct thread *td, int sync)
1172 {
1173 	struct proc *p;
1174 
1175 	p = td->td_proc;
1176 	if (sync) {
1177 		PMCLOG_RESERVE(po, THR_CREATE, sizeof(struct pmclog_threadcreate));
1178 		PMCLOG_EMIT32(td->td_tid);
1179 		PMCLOG_EMIT32(p->p_pid);
1180 		PMCLOG_EMIT32(p->p_flag);
1181 		PMCLOG_EMIT32(0);
1182 		PMCLOG_EMITSTRING(td->td_name, MAXCOMLEN+1);
1183 		PMCLOG_DESPATCH_SYNC(po);
1184 	} else {
1185 		PMCLOG_RESERVE(po, THR_CREATE, sizeof(struct pmclog_threadcreate));
1186 		PMCLOG_EMIT32(td->td_tid);
1187 		PMCLOG_EMIT32(p->p_pid);
1188 		PMCLOG_EMIT32(p->p_flag);
1189 		PMCLOG_EMIT32(0);
1190 		PMCLOG_EMITSTRING(td->td_name, MAXCOMLEN+1);
1191 		PMCLOG_DESPATCH(po);
1192 	}
1193 }
1194 
1195 void
1196 pmclog_process_threadexit(struct pmc_owner *po, struct thread *td)
1197 {
1198 
1199 	PMCLOG_RESERVE(po, THR_EXIT, sizeof(struct pmclog_threadexit));
1200 	PMCLOG_EMIT32(td->td_tid);
1201 	PMCLOG_DESPATCH(po);
1202 }
1203 
1204 /*
1205  * Write a user log entry.
1206  */
1207 
1208 int
1209 pmclog_process_userlog(struct pmc_owner *po, struct pmc_op_writelog *wl)
1210 {
1211 	int error;
1212 
1213 	PMCDBG2(LOG,WRI,1, "writelog po=%p ud=0x%x", po, wl->pm_userdata);
1214 
1215 	error = 0;
1216 
1217 	PMCLOG_RESERVE_WITH_ERROR(po, USERDATA,
1218 	    sizeof(struct pmclog_userdata));
1219 	PMCLOG_EMIT32(wl->pm_userdata);
1220 	PMCLOG_DESPATCH(po);
1221 
1222  error:
1223 	return (error);
1224 }
1225 
1226 /*
1227  * Initialization.
1228  *
1229  * Create a pool of log buffers and initialize mutexes.
1230  */
1231 
1232 void
1233 pmclog_initialize()
1234 {
1235 	struct pmclog_buffer *plb;
1236 	int domain, ncpus, total;
1237 
1238 	if (pmclog_buffer_size <= 0 || pmclog_buffer_size > 16*1024) {
1239 		(void) printf("hwpmc: tunable logbuffersize=%d must be "
1240 					  "greater than zero and less than or equal to 16MB.\n",
1241 					  pmclog_buffer_size);
1242 		pmclog_buffer_size = PMC_LOG_BUFFER_SIZE;
1243 	}
1244 
1245 	if (pmc_nlogbuffers_pcpu <= 0) {
1246 		(void) printf("hwpmc: tunable nlogbuffers=%d must be greater "
1247 					  "than zero.\n", pmc_nlogbuffers_pcpu);
1248 		pmc_nlogbuffers_pcpu = PMC_NLOGBUFFERS_PCPU;
1249 	}
1250 	if (pmc_nlogbuffers_pcpu*pmclog_buffer_size > 32*1024) {
1251 		(void) printf("hwpmc: memory allocated pcpu must be less than 32MB (is %dK).\n",
1252 					  pmc_nlogbuffers_pcpu*pmclog_buffer_size);
1253 		pmc_nlogbuffers_pcpu = PMC_NLOGBUFFERS_PCPU;
1254 		pmclog_buffer_size = PMC_LOG_BUFFER_SIZE;
1255 	}
1256 	for (domain = 0; domain < vm_ndomains; domain++) {
1257 		ncpus = pmc_dom_hdrs[domain]->pdbh_ncpus;
1258 		total = ncpus * pmc_nlogbuffers_pcpu;
1259 
1260 		plb = malloc_domainset(sizeof(struct pmclog_buffer) * total,
1261 		    M_PMC, DOMAINSET_PREF(domain), M_WAITOK | M_ZERO);
1262 		pmc_dom_hdrs[domain]->pdbh_plbs = plb;
1263 		for (; total > 0; total--, plb++) {
1264 			void *buf;
1265 
1266 			buf = malloc_domainset(1024 * pmclog_buffer_size, M_PMC,
1267 			    DOMAINSET_PREF(domain), M_WAITOK | M_ZERO);
1268 			PMCLOG_INIT_BUFFER_DESCRIPTOR(plb, buf, domain);
1269 			pmc_plb_rele_unlocked(plb);
1270 		}
1271 	}
1272 	mtx_init(&pmc_kthread_mtx, "pmc-kthread", "pmc-sleep", MTX_DEF);
1273 }
1274 
1275 /*
1276  * Shutdown logging.
1277  *
1278  * Destroy mutexes and release memory back the to free pool.
1279  */
1280 
1281 void
1282 pmclog_shutdown()
1283 {
1284 	struct pmclog_buffer *plb;
1285 	int domain;
1286 
1287 	mtx_destroy(&pmc_kthread_mtx);
1288 
1289 	for (domain = 0; domain < vm_ndomains; domain++) {
1290 		while ((plb = TAILQ_FIRST(&pmc_dom_hdrs[domain]->pdbh_head)) != NULL) {
1291 			TAILQ_REMOVE(&pmc_dom_hdrs[domain]->pdbh_head, plb, plb_next);
1292 			free(plb->plb_base, M_PMC);
1293 		}
1294 		free(pmc_dom_hdrs[domain]->pdbh_plbs, M_PMC);
1295 	}
1296 }
1297