xref: /freebsd/sys/dev/hwpmc/hwpmc_logging.c (revision 22d7dd834bc5cd189810e414701e3ad1e98102e4)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
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 #include <sys/param.h>
41 #include <sys/capsicum.h>
42 #include <sys/domainset.h>
43 #include <sys/file.h>
44 #include <sys/kernel.h>
45 #include <sys/kthread.h>
46 #include <sys/lock.h>
47 #include <sys/module.h>
48 #include <sys/mutex.h>
49 #include <sys/pmc.h>
50 #include <sys/pmckern.h>
51 #include <sys/pmclog.h>
52 #include <sys/proc.h>
53 #include <sys/sched.h>
54 #include <sys/signalvar.h>
55 #include <sys/smp.h>
56 #include <sys/syscallsubr.h>
57 #include <sys/sysctl.h>
58 #include <sys/systm.h>
59 #include <sys/uio.h>
60 #include <sys/unistd.h>
61 #include <sys/vnode.h>
62 
63 #if defined(__i386__) || defined(__amd64__)
64 #include <machine/clock.h>
65 #endif
66 
67 #define curdomain PCPU_GET(domain)
68 
69 /*
70  * Sysctl tunables
71  */
72 
73 SYSCTL_DECL(_kern_hwpmc);
74 
75 /*
76  * kern.hwpmc.logbuffersize -- size of the per-cpu owner buffers.
77  */
78 
79 static int pmclog_buffer_size = PMC_LOG_BUFFER_SIZE;
80 SYSCTL_INT(_kern_hwpmc, OID_AUTO, logbuffersize, CTLFLAG_RDTUN,
81     &pmclog_buffer_size, 0, "size of log buffers in kilobytes");
82 
83 /*
84  * kern.hwpmc.nbuffers_pcpu -- number of global log buffers
85  */
86 
87 static int pmc_nlogbuffers_pcpu = PMC_NLOGBUFFERS_PCPU;
88 SYSCTL_INT(_kern_hwpmc, OID_AUTO, nbuffers_pcpu, CTLFLAG_RDTUN,
89     &pmc_nlogbuffers_pcpu, 0, "number of log buffers per cpu");
90 
91 /*
92  * Global log buffer list and associated spin lock.
93  */
94 
95 static struct mtx pmc_kthread_mtx;	/* sleep lock */
96 
97 #define	PMCLOG_INIT_BUFFER_DESCRIPTOR(D, buf, domain) do {						\
98 		(D)->plb_fence = ((char *) (buf)) +	1024*pmclog_buffer_size;			\
99 		(D)->plb_base  = (D)->plb_ptr = ((char *) (buf));				\
100 		(D)->plb_domain = domain; \
101 	} while (0)
102 
103 #define	PMCLOG_RESET_BUFFER_DESCRIPTOR(D) do {			\
104 		(D)->plb_ptr  = (D)->plb_base; \
105 	} while (0)
106 
107 /*
108  * Log file record constructors.
109  */
110 #define	_PMCLOG_TO_HEADER(T,L)						\
111 	((PMCLOG_HEADER_MAGIC << 24) | (T << 16) | ((L) & 0xFFFF))
112 
113 /* reserve LEN bytes of space and initialize the entry header */
114 #define	_PMCLOG_RESERVE_SAFE(PO,TYPE,LEN,ACTION, TSC) do {	\
115 		uint32_t *_le;						\
116 		int _len = roundup((LEN), sizeof(uint32_t));	\
117 		struct pmclog_header *ph;							\
118 		if ((_le = pmclog_reserve((PO), _len)) == NULL) {	\
119 			ACTION;											\
120 		}													\
121 		ph = (struct pmclog_header *)_le;					\
122 		ph->pl_header =_PMCLOG_TO_HEADER(TYPE,_len);	\
123 		ph->pl_tsc = (TSC);									\
124 		_le += sizeof(*ph)/4	/* skip over timestamp */
125 
126 /* reserve LEN bytes of space and initialize the entry header */
127 #define	_PMCLOG_RESERVE(PO,TYPE,LEN,ACTION) do {			\
128 		uint32_t *_le;						\
129 		int _len = roundup((LEN), sizeof(uint32_t));	\
130 		uint64_t tsc;										\
131 		struct pmclog_header *ph;							\
132 		tsc = pmc_rdtsc();									\
133 		spinlock_enter();									\
134 		if ((_le = pmclog_reserve((PO), _len)) == NULL) {	\
135 			spinlock_exit();								\
136 			ACTION;											\
137 		}												\
138 		ph = (struct pmclog_header *)_le;					\
139 		ph->pl_header =_PMCLOG_TO_HEADER(TYPE,_len);	\
140 		ph->pl_tsc = tsc;									\
141 		_le += sizeof(*ph)/4	/* skip over timestamp */
142 
143 
144 
145 #define	PMCLOG_RESERVE_SAFE(P,T,L,TSC)		_PMCLOG_RESERVE_SAFE(P,T,L,return,TSC)
146 #define	PMCLOG_RESERVE(P,T,L)		_PMCLOG_RESERVE(P,T,L,return)
147 #define	PMCLOG_RESERVE_WITH_ERROR(P,T,L) _PMCLOG_RESERVE(P,T,L,		\
148 	error=ENOMEM;goto error)
149 
150 #define	PMCLOG_EMIT32(V)	do { *_le++ = (V); } while (0)
151 #define	PMCLOG_EMIT64(V)	do { 					\
152 		*_le++ = (uint32_t) ((V) & 0xFFFFFFFF);			\
153 		*_le++ = (uint32_t) (((V) >> 32) & 0xFFFFFFFF);		\
154 	} while (0)
155 
156 
157 /* Emit a string.  Caution: does NOT update _le, so needs to be last */
158 #define	PMCLOG_EMITSTRING(S,L)	do { bcopy((S), _le, (L)); } while (0)
159 #define	PMCLOG_EMITNULLSTRING(L) do { bzero(_le, (L)); } while (0)
160 
161 #define	PMCLOG_DESPATCH_SAFE(PO)						\
162 	    pmclog_release((PO));						\
163 	} while (0)
164 
165 #define	PMCLOG_DESPATCH_SCHED_LOCK(PO)						\
166 	     pmclog_release_flags((PO), 0);							\
167 	} while (0)
168 
169 #define	PMCLOG_DESPATCH(PO)							\
170 	    pmclog_release((PO));						\
171 		spinlock_exit();							\
172 	} while (0)
173 
174 #define	PMCLOG_DESPATCH_SYNC(PO)						\
175 	    pmclog_schedule_io((PO), 1);						\
176 		spinlock_exit();								\
177 		} while (0)
178 
179 
180 #define TSDELTA 4
181 /*
182  * Assertions about the log file format.
183  */
184 CTASSERT(sizeof(struct pmclog_callchain) == 7*4 + TSDELTA +
185     PMC_CALLCHAIN_DEPTH_MAX*sizeof(uintfptr_t));
186 CTASSERT(sizeof(struct pmclog_closelog) == 3*4 + TSDELTA);
187 CTASSERT(sizeof(struct pmclog_dropnotify) == 3*4 + TSDELTA);
188 CTASSERT(sizeof(struct pmclog_map_in) == PATH_MAX + TSDELTA +
189     5*4 + sizeof(uintfptr_t));
190 CTASSERT(offsetof(struct pmclog_map_in,pl_pathname) ==
191     5*4 + TSDELTA + sizeof(uintfptr_t));
192 CTASSERT(sizeof(struct pmclog_map_out) == 5*4 + 2*sizeof(uintfptr_t) + TSDELTA);
193 CTASSERT(sizeof(struct pmclog_pmcallocate) == 9*4 + TSDELTA);
194 CTASSERT(sizeof(struct pmclog_pmcattach) == 5*4 + PATH_MAX + TSDELTA);
195 CTASSERT(offsetof(struct pmclog_pmcattach,pl_pathname) == 5*4 + TSDELTA);
196 CTASSERT(sizeof(struct pmclog_pmcdetach) == 5*4 + TSDELTA);
197 CTASSERT(sizeof(struct pmclog_proccsw) == 7*4 + 8 + TSDELTA);
198 CTASSERT(sizeof(struct pmclog_procexec) == 5*4 + PATH_MAX +
199     2*sizeof(uintptr_t) + TSDELTA);
200 CTASSERT(offsetof(struct pmclog_procexec,pl_pathname) == 5*4 + TSDELTA +
201     2*sizeof(uintptr_t));
202 CTASSERT(sizeof(struct pmclog_procexit) == 5*4 + 8 + TSDELTA);
203 CTASSERT(sizeof(struct pmclog_procfork) == 5*4 + TSDELTA);
204 CTASSERT(sizeof(struct pmclog_sysexit) == 6*4);
205 CTASSERT(sizeof(struct pmclog_userdata) == 6*4);
206 
207 /*
208  * Log buffer structure
209  */
210 
211 struct pmclog_buffer {
212 	TAILQ_ENTRY(pmclog_buffer) plb_next;
213 	char 		*plb_base;
214 	char		*plb_ptr;
215 	char 		*plb_fence;
216 	uint16_t	 plb_domain;
217 } __aligned(CACHE_LINE_SIZE);
218 
219 /*
220  * Prototypes
221  */
222 
223 static int pmclog_get_buffer(struct pmc_owner *po);
224 static void pmclog_loop(void *arg);
225 static void pmclog_release(struct pmc_owner *po);
226 static uint32_t *pmclog_reserve(struct pmc_owner *po, int length);
227 static void pmclog_schedule_io(struct pmc_owner *po, int wakeup);
228 static void pmclog_schedule_all(struct pmc_owner *po);
229 static void pmclog_stop_kthread(struct pmc_owner *po);
230 
231 /*
232  * Helper functions
233  */
234 
235 static inline void
236 pmc_plb_rele_unlocked(struct pmclog_buffer *plb)
237 {
238 	TAILQ_INSERT_HEAD(&pmc_dom_hdrs[plb->plb_domain]->pdbh_head, plb, plb_next);
239 }
240 
241 static inline void
242 pmc_plb_rele(struct pmclog_buffer *plb)
243 {
244 	mtx_lock_spin(&pmc_dom_hdrs[plb->plb_domain]->pdbh_mtx);
245 	pmc_plb_rele_unlocked(plb);
246 	mtx_unlock_spin(&pmc_dom_hdrs[plb->plb_domain]->pdbh_mtx);
247 }
248 
249 /*
250  * Get a log buffer
251  */
252 static int
253 pmclog_get_buffer(struct pmc_owner *po)
254 {
255 	struct pmclog_buffer *plb;
256 	int domain;
257 
258 	KASSERT(po->po_curbuf[curcpu] == NULL,
259 	    ("[pmclog,%d] po=%p current buffer still valid", __LINE__, po));
260 
261 	domain = curdomain;
262 	MPASS(pmc_dom_hdrs[domain]);
263 	mtx_lock_spin(&pmc_dom_hdrs[domain]->pdbh_mtx);
264 	if ((plb = TAILQ_FIRST(&pmc_dom_hdrs[domain]->pdbh_head)) != NULL)
265 		TAILQ_REMOVE(&pmc_dom_hdrs[domain]->pdbh_head, plb, plb_next);
266 	mtx_unlock_spin(&pmc_dom_hdrs[domain]->pdbh_mtx);
267 
268 	PMCDBG2(LOG,GTB,1, "po=%p plb=%p", po, plb);
269 
270 #ifdef	HWPMC_DEBUG
271 	if (plb)
272 		KASSERT(plb->plb_ptr == plb->plb_base &&
273 		    plb->plb_base < plb->plb_fence,
274 		    ("[pmclog,%d] po=%p buffer invariants: ptr=%p "
275 		    "base=%p fence=%p", __LINE__, po, plb->plb_ptr,
276 		    plb->plb_base, plb->plb_fence));
277 #endif
278 
279 	po->po_curbuf[curcpu] = plb;
280 
281 	/* update stats */
282 	counter_u64_add(pmc_stats.pm_buffer_requests, 1);
283 	if (plb == NULL)
284 		counter_u64_add(pmc_stats.pm_buffer_requests_failed, 1);
285 
286 	return (plb ? 0 : ENOMEM);
287 }
288 
289 struct pmclog_proc_init_args {
290 	struct proc *kthr;
291 	struct pmc_owner *po;
292 	bool exit;
293 	bool acted;
294 };
295 
296 int
297 pmclog_proc_create(struct thread *td, void **handlep)
298 {
299 	struct pmclog_proc_init_args *ia;
300 	int error;
301 
302 	ia = malloc(sizeof(*ia), M_TEMP, M_WAITOK | M_ZERO);
303 	error = kproc_create(pmclog_loop, ia, &ia->kthr,
304 	    RFHIGHPID, 0, "hwpmc: proc(%d)", td->td_proc->p_pid);
305 	if (error == 0)
306 		*handlep = ia;
307 	return (error);
308 }
309 
310 void
311 pmclog_proc_ignite(void *handle, struct pmc_owner *po)
312 {
313 	struct pmclog_proc_init_args *ia;
314 
315 	ia = handle;
316 	mtx_lock(&pmc_kthread_mtx);
317 	MPASS(!ia->acted);
318 	MPASS(ia->po == NULL);
319 	MPASS(!ia->exit);
320 	MPASS(ia->kthr != NULL);
321 	if (po == NULL) {
322 		ia->exit = true;
323 	} else {
324 		ia->po = po;
325 		KASSERT(po->po_kthread == NULL,
326 		    ("[pmclog,%d] po=%p kthread (%p) already present",
327 		    __LINE__, po, po->po_kthread));
328 		po->po_kthread = ia->kthr;
329 	}
330 	wakeup(ia);
331 	while (!ia->acted)
332 		msleep(ia, &pmc_kthread_mtx, PWAIT, "pmclogw", 0);
333 	mtx_unlock(&pmc_kthread_mtx);
334 	free(ia, M_TEMP);
335 }
336 
337 /*
338  * Log handler loop.
339  *
340  * This function is executed by each pmc owner's helper thread.
341  */
342 static void
343 pmclog_loop(void *arg)
344 {
345 	struct pmclog_proc_init_args *ia;
346 	struct pmc_owner *po;
347 	struct pmclog_buffer *lb;
348 	struct proc *p;
349 	struct ucred *ownercred;
350 	struct ucred *mycred;
351 	struct thread *td;
352 	sigset_t unb;
353 	struct uio auio;
354 	struct iovec aiov;
355 	size_t nbytes;
356 	int error;
357 
358 	td = curthread;
359 
360 	SIGEMPTYSET(unb);
361 	SIGADDSET(unb, SIGHUP);
362 	(void)kern_sigprocmask(td, SIG_UNBLOCK, &unb, NULL, 0);
363 
364 	ia = arg;
365 	MPASS(ia->kthr == curproc);
366 	MPASS(!ia->acted);
367 	mtx_lock(&pmc_kthread_mtx);
368 	while (ia->po == NULL && !ia->exit)
369 		msleep(ia, &pmc_kthread_mtx, PWAIT, "pmclogi", 0);
370 	if (ia->exit) {
371 		ia->acted = true;
372 		wakeup(ia);
373 		mtx_unlock(&pmc_kthread_mtx);
374 		kproc_exit(0);
375 	}
376 	MPASS(ia->po != NULL);
377 	po = ia->po;
378 	ia->acted = true;
379 	wakeup(ia);
380 	mtx_unlock(&pmc_kthread_mtx);
381 	ia = NULL;
382 
383 	p = po->po_owner;
384 	mycred = td->td_ucred;
385 
386 	PROC_LOCK(p);
387 	ownercred = crhold(p->p_ucred);
388 	PROC_UNLOCK(p);
389 
390 	PMCDBG2(LOG,INI,1, "po=%p kt=%p", po, po->po_kthread);
391 	KASSERT(po->po_kthread == curthread->td_proc,
392 	    ("[pmclog,%d] proc mismatch po=%p po/kt=%p curproc=%p", __LINE__,
393 		po, po->po_kthread, curthread->td_proc));
394 
395 	lb = NULL;
396 
397 
398 	/*
399 	 * Loop waiting for I/O requests to be added to the owner
400 	 * struct's queue.  The loop is exited when the log file
401 	 * is deconfigured.
402 	 */
403 
404 	mtx_lock(&pmc_kthread_mtx);
405 
406 	for (;;) {
407 
408 		/* check if we've been asked to exit */
409 		if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
410 			break;
411 
412 		if (lb == NULL) { /* look for a fresh buffer to write */
413 			mtx_lock_spin(&po->po_mtx);
414 			if ((lb = TAILQ_FIRST(&po->po_logbuffers)) == NULL) {
415 				mtx_unlock_spin(&po->po_mtx);
416 
417 				/* No more buffers and shutdown required. */
418 				if (po->po_flags & PMC_PO_SHUTDOWN)
419 					break;
420 
421 				(void) msleep(po, &pmc_kthread_mtx, PWAIT,
422 				    "pmcloop", 250);
423 				continue;
424 			}
425 
426 			TAILQ_REMOVE(&po->po_logbuffers, lb, plb_next);
427 			mtx_unlock_spin(&po->po_mtx);
428 		}
429 
430 		mtx_unlock(&pmc_kthread_mtx);
431 
432 		/* process the request */
433 		PMCDBG3(LOG,WRI,2, "po=%p base=%p ptr=%p", po,
434 		    lb->plb_base, lb->plb_ptr);
435 		/* change our thread's credentials before issuing the I/O */
436 
437 		aiov.iov_base = lb->plb_base;
438 		aiov.iov_len  = nbytes = lb->plb_ptr - lb->plb_base;
439 
440 		auio.uio_iov    = &aiov;
441 		auio.uio_iovcnt = 1;
442 		auio.uio_offset = -1;
443 		auio.uio_resid  = nbytes;
444 		auio.uio_rw     = UIO_WRITE;
445 		auio.uio_segflg = UIO_SYSSPACE;
446 		auio.uio_td     = td;
447 
448 		/* switch thread credentials -- see kern_ktrace.c */
449 		td->td_ucred = ownercred;
450 		error = fo_write(po->po_file, &auio, ownercred, 0, td);
451 		td->td_ucred = mycred;
452 
453 		if (error) {
454 			/* XXX some errors are recoverable */
455 			/* send a SIGIO to the owner and exit */
456 			PROC_LOCK(p);
457 			kern_psignal(p, SIGIO);
458 			PROC_UNLOCK(p);
459 
460 			mtx_lock(&pmc_kthread_mtx);
461 
462 			po->po_error = error; /* save for flush log */
463 
464 			PMCDBG2(LOG,WRI,2, "po=%p error=%d", po, error);
465 
466 			break;
467 		}
468 
469 		mtx_lock(&pmc_kthread_mtx);
470 
471 		/* put the used buffer back into the global pool */
472 		PMCLOG_RESET_BUFFER_DESCRIPTOR(lb);
473 
474 		pmc_plb_rele(lb);
475 		lb = NULL;
476 	}
477 
478 	wakeup_one(po->po_kthread);
479 	po->po_kthread = NULL;
480 
481 	mtx_unlock(&pmc_kthread_mtx);
482 
483 	/* return the current I/O buffer to the global pool */
484 	if (lb) {
485 		PMCLOG_RESET_BUFFER_DESCRIPTOR(lb);
486 
487 		pmc_plb_rele(lb);
488 	}
489 
490 	/*
491 	 * Exit this thread, signalling the waiter
492 	 */
493 
494 	crfree(ownercred);
495 
496 	kproc_exit(0);
497 }
498 
499 /*
500  * Release and log entry and schedule an I/O if needed.
501  */
502 
503 static void
504 pmclog_release_flags(struct pmc_owner *po, int wakeup)
505 {
506 	struct pmclog_buffer *plb;
507 
508 	plb = po->po_curbuf[curcpu];
509 	KASSERT(plb->plb_ptr >= plb->plb_base,
510 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p base=%p", __LINE__,
511 		po, plb->plb_ptr, plb->plb_base));
512 	KASSERT(plb->plb_ptr <= plb->plb_fence,
513 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p fenc=%p", __LINE__,
514 		po, plb->plb_ptr, plb->plb_fence));
515 
516 	/* schedule an I/O if we've filled a buffer */
517 	if (plb->plb_ptr >= plb->plb_fence)
518 		pmclog_schedule_io(po, wakeup);
519 
520 	PMCDBG1(LOG,REL,1, "po=%p", po);
521 }
522 
523 static void
524 pmclog_release(struct pmc_owner *po)
525 {
526 
527 	pmclog_release_flags(po, 1);
528 }
529 
530 
531 /*
532  * Attempt to reserve 'length' bytes of space in an owner's log
533  * buffer.  The function returns a pointer to 'length' bytes of space
534  * if there was enough space or returns NULL if no space was
535  * available.  Non-null returns do so with the po mutex locked.  The
536  * caller must invoke pmclog_release() on the pmc owner structure
537  * when done.
538  */
539 
540 static uint32_t *
541 pmclog_reserve(struct pmc_owner *po, int length)
542 {
543 	uintptr_t newptr, oldptr __diagused;
544 	struct pmclog_buffer *plb, **pplb;
545 
546 	PMCDBG2(LOG,ALL,1, "po=%p len=%d", po, length);
547 
548 	KASSERT(length % sizeof(uint32_t) == 0,
549 	    ("[pmclog,%d] length not a multiple of word size", __LINE__));
550 
551 	/* No more data when shutdown in progress. */
552 	if (po->po_flags & PMC_PO_SHUTDOWN)
553 		return (NULL);
554 
555 	pplb = &po->po_curbuf[curcpu];
556 	if (*pplb == NULL && pmclog_get_buffer(po) != 0)
557 		goto fail;
558 
559 	KASSERT(*pplb != NULL,
560 	    ("[pmclog,%d] po=%p no current buffer", __LINE__, po));
561 
562 	plb = *pplb;
563 	KASSERT(plb->plb_ptr >= plb->plb_base &&
564 	    plb->plb_ptr <= plb->plb_fence,
565 	    ("[pmclog,%d] po=%p buffer invariants: ptr=%p base=%p fence=%p",
566 		__LINE__, po, plb->plb_ptr, plb->plb_base,
567 		plb->plb_fence));
568 
569 	oldptr = (uintptr_t) plb->plb_ptr;
570 	newptr = oldptr + length;
571 
572 	KASSERT(oldptr != (uintptr_t) NULL,
573 	    ("[pmclog,%d] po=%p Null log buffer pointer", __LINE__, po));
574 
575 	/*
576 	 * If we have space in the current buffer, return a pointer to
577 	 * available space with the PO structure locked.
578 	 */
579 	if (newptr <= (uintptr_t) plb->plb_fence) {
580 		plb->plb_ptr = (char *) newptr;
581 		goto done;
582 	}
583 
584 	/*
585 	 * Otherwise, schedule the current buffer for output and get a
586 	 * fresh buffer.
587 	 */
588 	pmclog_schedule_io(po, 0);
589 
590 	if (pmclog_get_buffer(po) != 0)
591 		goto fail;
592 
593 	plb = *pplb;
594 	KASSERT(plb != NULL,
595 	    ("[pmclog,%d] po=%p no current buffer", __LINE__, po));
596 
597 	KASSERT(plb->plb_ptr != NULL,
598 	    ("[pmclog,%d] null return from pmc_get_log_buffer", __LINE__));
599 
600 	KASSERT(plb->plb_ptr == plb->plb_base &&
601 	    plb->plb_ptr <= plb->plb_fence,
602 	    ("[pmclog,%d] po=%p buffer invariants: ptr=%p base=%p fence=%p",
603 		__LINE__, po, plb->plb_ptr, plb->plb_base,
604 		plb->plb_fence));
605 
606 	oldptr = (uintptr_t) plb->plb_ptr;
607 
608  done:
609 	return ((uint32_t *) oldptr);
610  fail:
611 	return (NULL);
612 }
613 
614 /*
615  * Schedule an I/O.
616  *
617  * Transfer the current buffer to the helper kthread.
618  */
619 
620 static void
621 pmclog_schedule_io(struct pmc_owner *po, int wakeup)
622 {
623 	struct pmclog_buffer *plb;
624 
625 	plb = po->po_curbuf[curcpu];
626 	po->po_curbuf[curcpu] = NULL;
627 	KASSERT(plb != NULL,
628 	    ("[pmclog,%d] schedule_io with null buffer po=%p", __LINE__, po));
629 	KASSERT(plb->plb_ptr >= plb->plb_base,
630 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p base=%p", __LINE__,
631 		po, plb->plb_ptr, plb->plb_base));
632 	KASSERT(plb->plb_ptr <= plb->plb_fence,
633 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p fenc=%p", __LINE__,
634 		po, plb->plb_ptr, plb->plb_fence));
635 
636 	PMCDBG1(LOG,SIO, 1, "po=%p", po);
637 
638 	/*
639 	 * Add the current buffer to the tail of the buffer list and
640 	 * wakeup the helper.
641 	 */
642 	mtx_lock_spin(&po->po_mtx);
643 	TAILQ_INSERT_TAIL(&po->po_logbuffers, plb, plb_next);
644 	mtx_unlock_spin(&po->po_mtx);
645 	if (wakeup)
646 		wakeup_one(po);
647 }
648 
649 /*
650  * Stop the helper kthread.
651  */
652 
653 static void
654 pmclog_stop_kthread(struct pmc_owner *po)
655 {
656 
657 	mtx_lock(&pmc_kthread_mtx);
658 	po->po_flags &= ~PMC_PO_OWNS_LOGFILE;
659 	if (po->po_kthread != NULL) {
660 		PROC_LOCK(po->po_kthread);
661 		kern_psignal(po->po_kthread, SIGHUP);
662 		PROC_UNLOCK(po->po_kthread);
663 	}
664 	wakeup_one(po);
665 	while (po->po_kthread)
666 		msleep(po->po_kthread, &pmc_kthread_mtx, PPAUSE, "pmckstp", 0);
667 	mtx_unlock(&pmc_kthread_mtx);
668 }
669 
670 /*
671  * Public functions
672  */
673 
674 /*
675  * Configure a log file for pmc owner 'po'.
676  *
677  * Parameter 'logfd' is a file handle referencing an open file in the
678  * owner process.  This file needs to have been opened for writing.
679  */
680 
681 int
682 pmclog_configure_log(struct pmc_mdep *md, struct pmc_owner *po, int logfd)
683 {
684 	struct proc *p;
685 	struct timespec ts;
686 	int error;
687 
688 	sx_assert(&pmc_sx, SA_XLOCKED);
689 	PMCDBG2(LOG,CFG,1, "config po=%p logfd=%d", po, logfd);
690 
691 	p = po->po_owner;
692 
693 	/* return EBUSY if a log file was already present */
694 	if (po->po_flags & PMC_PO_OWNS_LOGFILE)
695 		return (EBUSY);
696 
697 	KASSERT(po->po_file == NULL,
698 	    ("[pmclog,%d] po=%p file (%p) already present", __LINE__, po,
699 		po->po_file));
700 
701 	/* get a reference to the file state */
702 	error = fget_write(curthread, logfd, &cap_write_rights, &po->po_file);
703 	if (error)
704 		goto error;
705 
706 	/* mark process as owning a log file */
707 	po->po_flags |= PMC_PO_OWNS_LOGFILE;
708 
709 	/* mark process as using HWPMCs */
710 	PROC_LOCK(p);
711 	p->p_flag |= P_HWPMC;
712 	PROC_UNLOCK(p);
713 	nanotime(&ts);
714 	/* create a log initialization entry */
715 	PMCLOG_RESERVE_WITH_ERROR(po, PMCLOG_TYPE_INITIALIZE,
716 	    sizeof(struct pmclog_initialize));
717 	PMCLOG_EMIT32(PMC_VERSION);
718 	PMCLOG_EMIT32(md->pmd_cputype);
719 #if defined(__i386__) || defined(__amd64__)
720 	PMCLOG_EMIT64(tsc_freq);
721 #else
722 	/* other architectures will need to fill this in */
723 	PMCLOG_EMIT32(0);
724 	PMCLOG_EMIT32(0);
725 #endif
726 	memcpy(_le, &ts, sizeof(ts));
727 	_le += sizeof(ts)/4;
728 	PMCLOG_EMITSTRING(pmc_cpuid, PMC_CPUID_LEN);
729 	PMCLOG_DESPATCH_SYNC(po);
730 
731 	return (0);
732 
733  error:
734 	KASSERT(po->po_kthread == NULL, ("[pmclog,%d] po=%p kthread not "
735 	    "stopped", __LINE__, po));
736 
737 	if (po->po_file)
738 		(void) fdrop(po->po_file, curthread);
739 	po->po_file  = NULL;	/* clear file and error state */
740 	po->po_error = 0;
741 	po->po_flags &= ~PMC_PO_OWNS_LOGFILE;
742 
743 	return (error);
744 }
745 
746 
747 /*
748  * De-configure a log file.  This will throw away any buffers queued
749  * for this owner process.
750  */
751 
752 int
753 pmclog_deconfigure_log(struct pmc_owner *po)
754 {
755 	int error;
756 	struct pmclog_buffer *lb;
757 	struct pmc_binding pb;
758 
759 	PMCDBG1(LOG,CFG,1, "de-config po=%p", po);
760 
761 	if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
762 		return (EINVAL);
763 
764 	KASSERT(po->po_sscount == 0,
765 	    ("[pmclog,%d] po=%p still owning SS PMCs", __LINE__, po));
766 	KASSERT(po->po_file != NULL,
767 	    ("[pmclog,%d] po=%p no log file", __LINE__, po));
768 
769 	/* stop the kthread, this will reset the 'OWNS_LOGFILE' flag */
770 	pmclog_stop_kthread(po);
771 
772 	KASSERT(po->po_kthread == NULL,
773 	    ("[pmclog,%d] po=%p kthread not stopped", __LINE__, po));
774 
775 	/* return all queued log buffers to the global pool */
776 	while ((lb = TAILQ_FIRST(&po->po_logbuffers)) != NULL) {
777 		TAILQ_REMOVE(&po->po_logbuffers, lb, plb_next);
778 		PMCLOG_RESET_BUFFER_DESCRIPTOR(lb);
779 		pmc_plb_rele(lb);
780 	}
781 	pmc_save_cpu_binding(&pb);
782 	for (int i = 0; i < mp_ncpus; i++) {
783 		pmc_select_cpu(i);
784 		/* return the 'current' buffer to the global pool */
785 		if ((lb = po->po_curbuf[curcpu]) != NULL) {
786 			PMCLOG_RESET_BUFFER_DESCRIPTOR(lb);
787 			pmc_plb_rele(lb);
788 		}
789 	}
790 	pmc_restore_cpu_binding(&pb);
791 
792 	/* drop a reference to the fd */
793 	if (po->po_file != NULL) {
794 		error = fdrop(po->po_file, curthread);
795 		po->po_file = NULL;
796 	} else
797 		error = 0;
798 	po->po_error = 0;
799 
800 	return (error);
801 }
802 
803 /*
804  * Flush a process' log buffer.
805  */
806 
807 int
808 pmclog_flush(struct pmc_owner *po, int force)
809 {
810 	int error;
811 
812 	PMCDBG1(LOG,FLS,1, "po=%p", po);
813 
814 	/*
815 	 * If there is a pending error recorded by the logger thread,
816 	 * return that.
817 	 */
818 	if (po->po_error)
819 		return (po->po_error);
820 
821 	error = 0;
822 
823 	/*
824 	 * Check that we do have an active log file.
825 	 */
826 	mtx_lock(&pmc_kthread_mtx);
827 	if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) {
828 		error = EINVAL;
829 		goto error;
830 	}
831 
832 	pmclog_schedule_all(po);
833  error:
834 	mtx_unlock(&pmc_kthread_mtx);
835 
836 	return (error);
837 }
838 
839 static void
840 pmclog_schedule_one_cond(struct pmc_owner *po)
841 {
842 	struct pmclog_buffer *plb;
843 	int cpu;
844 
845 	spinlock_enter();
846 	cpu = curcpu;
847 	/* tell hardclock not to run again */
848 	if (PMC_CPU_HAS_SAMPLES(cpu))
849 		PMC_CALL_HOOK_UNLOCKED(curthread, PMC_FN_DO_SAMPLES, NULL);
850 
851 	plb = po->po_curbuf[cpu];
852 	if (plb && plb->plb_ptr != plb->plb_base)
853 		pmclog_schedule_io(po, 1);
854 	spinlock_exit();
855 }
856 
857 static void
858 pmclog_schedule_all(struct pmc_owner *po)
859 {
860 	struct pmc_binding pb;
861 
862 	/*
863 	 * Schedule the current buffer if any and not empty.
864 	 */
865 	pmc_save_cpu_binding(&pb);
866 	for (int i = 0; i < mp_ncpus; i++) {
867 		pmc_select_cpu(i);
868 		pmclog_schedule_one_cond(po);
869 	}
870 	pmc_restore_cpu_binding(&pb);
871 }
872 
873 int
874 pmclog_close(struct pmc_owner *po)
875 {
876 
877 	PMCDBG1(LOG,CLO,1, "po=%p", po);
878 
879 	pmclog_process_closelog(po);
880 
881 	mtx_lock(&pmc_kthread_mtx);
882 	/*
883 	 * Initiate shutdown: no new data queued,
884 	 * thread will close file on last block.
885 	 */
886 	po->po_flags |= PMC_PO_SHUTDOWN;
887 	/* give time for all to see */
888 	DELAY(50);
889 
890 	/*
891 	 * Schedule the current buffer.
892 	 */
893 	pmclog_schedule_all(po);
894 	wakeup_one(po);
895 
896 	mtx_unlock(&pmc_kthread_mtx);
897 
898 	return (0);
899 }
900 
901 void
902 pmclog_process_callchain(struct pmc *pm, struct pmc_sample *ps)
903 {
904 	int n, recordlen;
905 	uint32_t flags;
906 	struct pmc_owner *po;
907 
908 	PMCDBG3(LOG,SAM,1,"pm=%p pid=%d n=%d", pm, ps->ps_pid,
909 	    ps->ps_nsamples);
910 
911 	recordlen = offsetof(struct pmclog_callchain, pl_pc) +
912 	    ps->ps_nsamples * sizeof(uintfptr_t);
913 	po = pm->pm_owner;
914 	flags = PMC_CALLCHAIN_TO_CPUFLAGS(ps->ps_cpu,ps->ps_flags);
915 	PMCLOG_RESERVE_SAFE(po, PMCLOG_TYPE_CALLCHAIN, recordlen, ps->ps_tsc);
916 	PMCLOG_EMIT32(ps->ps_pid);
917 	PMCLOG_EMIT32(ps->ps_tid);
918 	PMCLOG_EMIT32(pm->pm_id);
919 	PMCLOG_EMIT32(flags);
920 	for (n = 0; n < ps->ps_nsamples; n++)
921 		PMCLOG_EMITADDR(ps->ps_pc[n]);
922 	PMCLOG_DESPATCH_SAFE(po);
923 }
924 
925 void
926 pmclog_process_closelog(struct pmc_owner *po)
927 {
928 	PMCLOG_RESERVE(po, PMCLOG_TYPE_CLOSELOG,
929 	    sizeof(struct pmclog_closelog));
930 	PMCLOG_DESPATCH_SYNC(po);
931 }
932 
933 void
934 pmclog_process_dropnotify(struct pmc_owner *po)
935 {
936 	PMCLOG_RESERVE(po, PMCLOG_TYPE_DROPNOTIFY,
937 	    sizeof(struct pmclog_dropnotify));
938 	PMCLOG_DESPATCH(po);
939 }
940 
941 void
942 pmclog_process_map_in(struct pmc_owner *po, pid_t pid, uintfptr_t start,
943     const char *path)
944 {
945 	int pathlen, recordlen;
946 
947 	KASSERT(path != NULL, ("[pmclog,%d] map-in, null path", __LINE__));
948 
949 	pathlen = strlen(path) + 1;	/* #bytes for path name */
950 	recordlen = offsetof(struct pmclog_map_in, pl_pathname) +
951 	    pathlen;
952 
953 	PMCLOG_RESERVE(po, PMCLOG_TYPE_MAP_IN, recordlen);
954 	PMCLOG_EMIT32(pid);
955 	PMCLOG_EMIT32(0);
956 	PMCLOG_EMITADDR(start);
957 	PMCLOG_EMITSTRING(path,pathlen);
958 	PMCLOG_DESPATCH_SYNC(po);
959 }
960 
961 void
962 pmclog_process_map_out(struct pmc_owner *po, pid_t pid, uintfptr_t start,
963     uintfptr_t end)
964 {
965 	KASSERT(start <= end, ("[pmclog,%d] start > end", __LINE__));
966 
967 	PMCLOG_RESERVE(po, PMCLOG_TYPE_MAP_OUT, sizeof(struct pmclog_map_out));
968 	PMCLOG_EMIT32(pid);
969 	PMCLOG_EMIT32(0);
970 	PMCLOG_EMITADDR(start);
971 	PMCLOG_EMITADDR(end);
972 	PMCLOG_DESPATCH(po);
973 }
974 
975 void
976 pmclog_process_pmcallocate(struct pmc *pm)
977 {
978 	struct pmc_owner *po;
979 	struct pmc_soft *ps;
980 
981 	po = pm->pm_owner;
982 
983 	PMCDBG1(LOG,ALL,1, "pm=%p", pm);
984 
985 	if (PMC_TO_CLASS(pm) == PMC_CLASS_SOFT) {
986 		PMCLOG_RESERVE(po, PMCLOG_TYPE_PMCALLOCATEDYN,
987 		    sizeof(struct pmclog_pmcallocatedyn));
988 		PMCLOG_EMIT32(pm->pm_id);
989 		PMCLOG_EMIT32(pm->pm_event);
990 		PMCLOG_EMIT32(pm->pm_flags);
991 		PMCLOG_EMIT32(0);
992 		PMCLOG_EMIT64(pm->pm_sc.pm_reloadcount);
993 		ps = pmc_soft_ev_acquire(pm->pm_event);
994 		if (ps != NULL)
995 			PMCLOG_EMITSTRING(ps->ps_ev.pm_ev_name,PMC_NAME_MAX);
996 		else
997 			PMCLOG_EMITNULLSTRING(PMC_NAME_MAX);
998 		pmc_soft_ev_release(ps);
999 		PMCLOG_DESPATCH_SYNC(po);
1000 	} else {
1001 		PMCLOG_RESERVE(po, PMCLOG_TYPE_PMCALLOCATE,
1002 		    sizeof(struct pmclog_pmcallocate));
1003 		PMCLOG_EMIT32(pm->pm_id);
1004 		PMCLOG_EMIT32(pm->pm_event);
1005 		PMCLOG_EMIT32(pm->pm_flags);
1006 		PMCLOG_EMIT32(0);
1007 		PMCLOG_EMIT64(pm->pm_sc.pm_reloadcount);
1008 		PMCLOG_DESPATCH_SYNC(po);
1009 	}
1010 }
1011 
1012 void
1013 pmclog_process_pmcattach(struct pmc *pm, pid_t pid, char *path)
1014 {
1015 	int pathlen, recordlen;
1016 	struct pmc_owner *po;
1017 
1018 	PMCDBG2(LOG,ATT,1,"pm=%p pid=%d", pm, pid);
1019 
1020 	po = pm->pm_owner;
1021 
1022 	pathlen = strlen(path) + 1;	/* #bytes for the string */
1023 	recordlen = offsetof(struct pmclog_pmcattach, pl_pathname) + pathlen;
1024 
1025 	PMCLOG_RESERVE(po, PMCLOG_TYPE_PMCATTACH, recordlen);
1026 	PMCLOG_EMIT32(pm->pm_id);
1027 	PMCLOG_EMIT32(pid);
1028 	PMCLOG_EMITSTRING(path, pathlen);
1029 	PMCLOG_DESPATCH_SYNC(po);
1030 }
1031 
1032 void
1033 pmclog_process_pmcdetach(struct pmc *pm, pid_t pid)
1034 {
1035 	struct pmc_owner *po;
1036 
1037 	PMCDBG2(LOG,ATT,1,"!pm=%p pid=%d", pm, pid);
1038 
1039 	po = pm->pm_owner;
1040 
1041 	PMCLOG_RESERVE(po, PMCLOG_TYPE_PMCDETACH,
1042 	    sizeof(struct pmclog_pmcdetach));
1043 	PMCLOG_EMIT32(pm->pm_id);
1044 	PMCLOG_EMIT32(pid);
1045 	PMCLOG_DESPATCH_SYNC(po);
1046 }
1047 
1048 void
1049 pmclog_process_proccreate(struct pmc_owner *po, struct proc *p, int sync)
1050 {
1051 	if (sync) {
1052 		PMCLOG_RESERVE(po, PMCLOG_TYPE_PROC_CREATE,
1053 		    sizeof(struct pmclog_proccreate));
1054 		PMCLOG_EMIT32(p->p_pid);
1055 		PMCLOG_EMIT32(p->p_flag);
1056 		PMCLOG_EMITSTRING(p->p_comm, MAXCOMLEN+1);
1057 		PMCLOG_DESPATCH_SYNC(po);
1058 	} else {
1059 		PMCLOG_RESERVE(po, PMCLOG_TYPE_PROC_CREATE,
1060 		    sizeof(struct pmclog_proccreate));
1061 		PMCLOG_EMIT32(p->p_pid);
1062 		PMCLOG_EMIT32(p->p_flag);
1063 		PMCLOG_EMITSTRING(p->p_comm, MAXCOMLEN+1);
1064 		PMCLOG_DESPATCH(po);
1065 	}
1066 }
1067 
1068 /*
1069  * Log a context switch event to the log file.
1070  */
1071 
1072 void
1073 pmclog_process_proccsw(struct pmc *pm, struct pmc_process *pp, pmc_value_t v, struct thread *td)
1074 {
1075 	struct pmc_owner *po;
1076 
1077 	KASSERT(pm->pm_flags & PMC_F_LOG_PROCCSW,
1078 	    ("[pmclog,%d] log-process-csw called gratuitously", __LINE__));
1079 
1080 	PMCDBG3(LOG,SWO,1,"pm=%p pid=%d v=%jx", pm, pp->pp_proc->p_pid,
1081 	    v);
1082 
1083 	po = pm->pm_owner;
1084 
1085 	PMCLOG_RESERVE_SAFE(po, PMCLOG_TYPE_PROCCSW,
1086 	    sizeof(struct pmclog_proccsw), pmc_rdtsc());
1087 	PMCLOG_EMIT64(v);
1088 	PMCLOG_EMIT32(pm->pm_id);
1089 	PMCLOG_EMIT32(pp->pp_proc->p_pid);
1090 	PMCLOG_EMIT32(td->td_tid);
1091 	PMCLOG_EMIT32(0);
1092 	PMCLOG_DESPATCH_SCHED_LOCK(po);
1093 }
1094 
1095 void
1096 pmclog_process_procexec(struct pmc_owner *po, pmc_id_t pmid, pid_t pid,
1097     uintptr_t baseaddr, uintptr_t dynaddr, char *path)
1098 {
1099 	int pathlen, recordlen;
1100 
1101 	PMCDBG3(LOG,EXC,1,"po=%p pid=%d path=\"%s\"", po, pid, path);
1102 
1103 	pathlen   = strlen(path) + 1;	/* #bytes for the path */
1104 	recordlen = offsetof(struct pmclog_procexec, pl_pathname) + pathlen;
1105 	PMCLOG_RESERVE(po, PMCLOG_TYPE_PROCEXEC, recordlen);
1106 	PMCLOG_EMIT32(pid);
1107 	PMCLOG_EMIT32(pmid);
1108 	PMCLOG_EMITADDR(baseaddr);
1109 	PMCLOG_EMITADDR(dynaddr);
1110 	PMCLOG_EMITSTRING(path,pathlen);
1111 	PMCLOG_DESPATCH_SYNC(po);
1112 }
1113 
1114 /*
1115  * Log a process exit event (and accumulated pmc value) to the log file.
1116  */
1117 
1118 void
1119 pmclog_process_procexit(struct pmc *pm, struct pmc_process *pp)
1120 {
1121 	int ri;
1122 	struct pmc_owner *po;
1123 
1124 	ri = PMC_TO_ROWINDEX(pm);
1125 	PMCDBG3(LOG,EXT,1,"pm=%p pid=%d v=%jx", pm, pp->pp_proc->p_pid,
1126 	    pp->pp_pmcs[ri].pp_pmcval);
1127 
1128 	po = pm->pm_owner;
1129 
1130 	PMCLOG_RESERVE(po, PMCLOG_TYPE_PROCEXIT,
1131 	    sizeof(struct pmclog_procexit));
1132 	PMCLOG_EMIT32(pm->pm_id);
1133 	PMCLOG_EMIT32(pp->pp_proc->p_pid);
1134 	PMCLOG_EMIT64(pp->pp_pmcs[ri].pp_pmcval);
1135 	PMCLOG_DESPATCH(po);
1136 }
1137 
1138 /*
1139  * Log a fork event.
1140  */
1141 
1142 void
1143 pmclog_process_procfork(struct pmc_owner *po, pid_t oldpid, pid_t newpid)
1144 {
1145 	PMCLOG_RESERVE(po, PMCLOG_TYPE_PROCFORK,
1146 	    sizeof(struct pmclog_procfork));
1147 	PMCLOG_EMIT32(oldpid);
1148 	PMCLOG_EMIT32(newpid);
1149 	PMCLOG_DESPATCH(po);
1150 }
1151 
1152 /*
1153  * Log a process exit event of the form suitable for system-wide PMCs.
1154  */
1155 
1156 void
1157 pmclog_process_sysexit(struct pmc_owner *po, pid_t pid)
1158 {
1159 	PMCLOG_RESERVE(po, PMCLOG_TYPE_SYSEXIT, sizeof(struct pmclog_sysexit));
1160 	PMCLOG_EMIT32(pid);
1161 	PMCLOG_DESPATCH(po);
1162 }
1163 
1164 void
1165 pmclog_process_threadcreate(struct pmc_owner *po, struct thread *td, int sync)
1166 {
1167 	struct proc *p;
1168 
1169 	p = td->td_proc;
1170 	if (sync) {
1171 		PMCLOG_RESERVE(po, PMCLOG_TYPE_THR_CREATE,
1172 		    sizeof(struct pmclog_threadcreate));
1173 		PMCLOG_EMIT32(td->td_tid);
1174 		PMCLOG_EMIT32(p->p_pid);
1175 		PMCLOG_EMIT32(p->p_flag);
1176 		PMCLOG_EMIT32(0);
1177 		PMCLOG_EMITSTRING(td->td_name, MAXCOMLEN+1);
1178 		PMCLOG_DESPATCH_SYNC(po);
1179 	} else {
1180 		PMCLOG_RESERVE(po, PMCLOG_TYPE_THR_CREATE,
1181 		    sizeof(struct pmclog_threadcreate));
1182 		PMCLOG_EMIT32(td->td_tid);
1183 		PMCLOG_EMIT32(p->p_pid);
1184 		PMCLOG_EMIT32(p->p_flag);
1185 		PMCLOG_EMIT32(0);
1186 		PMCLOG_EMITSTRING(td->td_name, MAXCOMLEN+1);
1187 		PMCLOG_DESPATCH(po);
1188 	}
1189 }
1190 
1191 void
1192 pmclog_process_threadexit(struct pmc_owner *po, struct thread *td)
1193 {
1194 
1195 	PMCLOG_RESERVE(po, PMCLOG_TYPE_THR_EXIT,
1196 	    sizeof(struct pmclog_threadexit));
1197 	PMCLOG_EMIT32(td->td_tid);
1198 	PMCLOG_DESPATCH(po);
1199 }
1200 
1201 /*
1202  * Write a user log entry.
1203  */
1204 
1205 int
1206 pmclog_process_userlog(struct pmc_owner *po, struct pmc_op_writelog *wl)
1207 {
1208 	int error;
1209 
1210 	PMCDBG2(LOG,WRI,1, "writelog po=%p ud=0x%x", po, wl->pm_userdata);
1211 
1212 	error = 0;
1213 
1214 	PMCLOG_RESERVE_WITH_ERROR(po, PMCLOG_TYPE_USERDATA,
1215 	    sizeof(struct pmclog_userdata));
1216 	PMCLOG_EMIT32(wl->pm_userdata);
1217 	PMCLOG_DESPATCH(po);
1218 
1219  error:
1220 	return (error);
1221 }
1222 
1223 /*
1224  * Initialization.
1225  *
1226  * Create a pool of log buffers and initialize mutexes.
1227  */
1228 
1229 void
1230 pmclog_initialize(void)
1231 {
1232 	struct pmclog_buffer *plb;
1233 	int domain, ncpus, total;
1234 
1235 	if (pmclog_buffer_size <= 0 || pmclog_buffer_size > 16*1024) {
1236 		(void) printf("hwpmc: tunable logbuffersize=%d must be "
1237 					  "greater than zero and less than or equal to 16MB.\n",
1238 					  pmclog_buffer_size);
1239 		pmclog_buffer_size = PMC_LOG_BUFFER_SIZE;
1240 	}
1241 
1242 	if (pmc_nlogbuffers_pcpu <= 0) {
1243 		(void) printf("hwpmc: tunable nlogbuffers=%d must be greater "
1244 					  "than zero.\n", pmc_nlogbuffers_pcpu);
1245 		pmc_nlogbuffers_pcpu = PMC_NLOGBUFFERS_PCPU;
1246 	}
1247 	if (pmc_nlogbuffers_pcpu*pmclog_buffer_size > 32*1024) {
1248 		(void) printf("hwpmc: memory allocated pcpu must be less than 32MB (is %dK).\n",
1249 					  pmc_nlogbuffers_pcpu*pmclog_buffer_size);
1250 		pmc_nlogbuffers_pcpu = PMC_NLOGBUFFERS_PCPU;
1251 		pmclog_buffer_size = PMC_LOG_BUFFER_SIZE;
1252 	}
1253 	for (domain = 0; domain < vm_ndomains; domain++) {
1254 		ncpus = pmc_dom_hdrs[domain]->pdbh_ncpus;
1255 		total = ncpus * pmc_nlogbuffers_pcpu;
1256 
1257 		plb = malloc_domainset(sizeof(struct pmclog_buffer) * total,
1258 		    M_PMC, DOMAINSET_PREF(domain), M_WAITOK | M_ZERO);
1259 		pmc_dom_hdrs[domain]->pdbh_plbs = plb;
1260 		for (; total > 0; total--, plb++) {
1261 			void *buf;
1262 
1263 			buf = malloc_domainset(1024 * pmclog_buffer_size, M_PMC,
1264 			    DOMAINSET_PREF(domain), M_WAITOK | M_ZERO);
1265 			PMCLOG_INIT_BUFFER_DESCRIPTOR(plb, buf, domain);
1266 			pmc_plb_rele_unlocked(plb);
1267 		}
1268 	}
1269 	mtx_init(&pmc_kthread_mtx, "pmc-kthread", "pmc-sleep", MTX_DEF);
1270 }
1271 
1272 /*
1273  * Shutdown logging.
1274  *
1275  * Destroy mutexes and release memory back the to free pool.
1276  */
1277 
1278 void
1279 pmclog_shutdown(void)
1280 {
1281 	struct pmclog_buffer *plb;
1282 	int domain;
1283 
1284 	mtx_destroy(&pmc_kthread_mtx);
1285 
1286 	for (domain = 0; domain < vm_ndomains; domain++) {
1287 		while ((plb = TAILQ_FIRST(&pmc_dom_hdrs[domain]->pdbh_head)) != NULL) {
1288 			TAILQ_REMOVE(&pmc_dom_hdrs[domain]->pdbh_head, plb, plb_next);
1289 			free(plb->plb_base, M_PMC);
1290 		}
1291 		free(pmc_dom_hdrs[domain]->pdbh_plbs, M_PMC);
1292 	}
1293 }
1294