xref: /freebsd/sys/dev/hwpmc/hwpmc_logging.c (revision f39bffc62c1395bde25d152c7f68fdf7cbaab414)
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/file.h>
45 #include <sys/kernel.h>
46 #include <sys/kthread.h>
47 #include <sys/lock.h>
48 #include <sys/module.h>
49 #include <sys/mutex.h>
50 #include <sys/pmc.h>
51 #include <sys/pmckern.h>
52 #include <sys/pmclog.h>
53 #include <sys/proc.h>
54 #include <sys/sched.h>
55 #include <sys/signalvar.h>
56 #include <sys/smp.h>
57 #include <sys/syscallsubr.h>
58 #include <sys/sysctl.h>
59 #include <sys/systm.h>
60 #include <sys/uio.h>
61 #include <sys/unistd.h>
62 #include <sys/vnode.h>
63 
64 #if defined(__i386__) || defined(__amd64__)
65 #include <machine/clock.h>
66 #endif
67 
68 #ifdef NUMA
69 #define NDOMAINS vm_ndomains
70 #define curdomain PCPU_GET(domain)
71 #else
72 #define NDOMAINS 1
73 #define curdomain 0
74 #define malloc_domain(size, type, domain, flags) malloc((size), (type), (flags))
75 #define free_domain(addr, type) free(addr, type)
76 #endif
77 
78 /*
79  * Sysctl tunables
80  */
81 
82 SYSCTL_DECL(_kern_hwpmc);
83 
84 /*
85  * kern.hwpmc.logbuffersize -- size of the per-cpu owner buffers.
86  */
87 
88 static int pmclog_buffer_size = PMC_LOG_BUFFER_SIZE;
89 #if (__FreeBSD_version < 1100000)
90 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "logbuffersize", &pmclog_buffer_size);
91 #endif
92 SYSCTL_INT(_kern_hwpmc, OID_AUTO, logbuffersize, CTLFLAG_RDTUN,
93     &pmclog_buffer_size, 0, "size of log buffers in kilobytes");
94 
95 /*
96  * kern.hwpmc.nbuffer -- number of global log buffers
97  */
98 
99 static int pmc_nlogbuffers_pcpu = PMC_NLOGBUFFERS_PCPU;
100 #if (__FreeBSD_version < 1100000)
101 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "nbuffers", &pmc_nlogbuffers_pcpu);
102 #endif
103 SYSCTL_INT(_kern_hwpmc, OID_AUTO, nbuffers_pcpu, CTLFLAG_RDTUN,
104     &pmc_nlogbuffers_pcpu, 0, "number of log buffers per cpu");
105 
106 /*
107  * Global log buffer list and associated spin lock.
108  */
109 
110 static struct mtx pmc_kthread_mtx;	/* sleep lock */
111 
112 #define	PMCLOG_INIT_BUFFER_DESCRIPTOR(D, buf, domain) do {						\
113 		(D)->plb_fence = ((char *) (buf)) +	1024*pmclog_buffer_size;			\
114 		(D)->plb_base  = (D)->plb_ptr = ((char *) (buf));				\
115 		(D)->plb_domain = domain; \
116 	} while (0)
117 
118 #define	PMCLOG_RESET_BUFFER_DESCRIPTOR(D) do {			\
119 		(D)->plb_ptr  = (D)->plb_base; \
120 	} while (0)
121 
122 /*
123  * Log file record constructors.
124  */
125 #define	_PMCLOG_TO_HEADER(T,L)						\
126 	((PMCLOG_HEADER_MAGIC << 24) |					\
127 	 (PMCLOG_TYPE_ ## T << 16)   |					\
128 	 ((L) & 0xFFFF))
129 
130 /* reserve LEN bytes of space and initialize the entry header */
131 #define	_PMCLOG_RESERVE_SAFE(PO,TYPE,LEN,ACTION, TSC) do {	\
132 		uint32_t *_le;						\
133 		int _len = roundup((LEN), sizeof(uint32_t));	\
134 		struct pmclog_header *ph;							\
135 		if ((_le = pmclog_reserve((PO), _len)) == NULL) {	\
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 /* reserve LEN bytes of space and initialize the entry header */
144 #define	_PMCLOG_RESERVE(PO,TYPE,LEN,ACTION) do {			\
145 		uint32_t *_le;						\
146 		int _len = roundup((LEN), sizeof(uint32_t));	\
147 		uint64_t tsc;										\
148 		struct pmclog_header *ph;							\
149 		tsc = pmc_rdtsc();									\
150 		spinlock_enter();									\
151 		if ((_le = pmclog_reserve((PO), _len)) == NULL) {	\
152 			spinlock_exit();								\
153 			ACTION;											\
154 		}												\
155 		ph = (struct pmclog_header *)_le;					\
156 		ph->pl_header =_PMCLOG_TO_HEADER(TYPE,_len);	\
157 		ph->pl_tsc = tsc;									\
158 		_le += sizeof(*ph)/4	/* skip over timestamp */
159 
160 
161 
162 #define	PMCLOG_RESERVE_SAFE(P,T,L,TSC)		_PMCLOG_RESERVE_SAFE(P,T,L,return,TSC)
163 #define	PMCLOG_RESERVE(P,T,L)		_PMCLOG_RESERVE(P,T,L,return)
164 #define	PMCLOG_RESERVE_WITH_ERROR(P,T,L) _PMCLOG_RESERVE(P,T,L,		\
165 	error=ENOMEM;goto error)
166 
167 #define	PMCLOG_EMIT32(V)	do { *_le++ = (V); } while (0)
168 #define	PMCLOG_EMIT64(V)	do { 					\
169 		*_le++ = (uint32_t) ((V) & 0xFFFFFFFF);			\
170 		*_le++ = (uint32_t) (((V) >> 32) & 0xFFFFFFFF);		\
171 	} while (0)
172 
173 
174 /* Emit a string.  Caution: does NOT update _le, so needs to be last */
175 #define	PMCLOG_EMITSTRING(S,L)	do { bcopy((S), _le, (L)); } while (0)
176 #define	PMCLOG_EMITNULLSTRING(L) do { bzero(_le, (L)); } while (0)
177 
178 #define	PMCLOG_DESPATCH_SAFE(PO)						\
179 	    pmclog_release((PO));						\
180 	} while (0)
181 
182 #define	PMCLOG_DESPATCH_SCHED_LOCK(PO)						\
183 	     pmclog_release_flags((PO), 0);							\
184 	} while (0)
185 
186 #define	PMCLOG_DESPATCH(PO)							\
187 	    pmclog_release((PO));						\
188 		spinlock_exit();							\
189 	} while (0)
190 
191 #define	PMCLOG_DESPATCH_SYNC(PO)						\
192 	    pmclog_schedule_io((PO), 1);						\
193 		spinlock_exit();								\
194 		} while (0)
195 
196 
197 #define TSDELTA 4
198 /*
199  * Assertions about the log file format.
200  */
201 CTASSERT(sizeof(struct pmclog_callchain) == 7*4 + TSDELTA +
202     PMC_CALLCHAIN_DEPTH_MAX*sizeof(uintfptr_t));
203 CTASSERT(sizeof(struct pmclog_closelog) == 3*4 + TSDELTA);
204 CTASSERT(sizeof(struct pmclog_dropnotify) == 3*4 + TSDELTA);
205 CTASSERT(sizeof(struct pmclog_map_in) == PATH_MAX + TSDELTA +
206     5*4 + sizeof(uintfptr_t));
207 CTASSERT(offsetof(struct pmclog_map_in,pl_pathname) ==
208     5*4 + TSDELTA + sizeof(uintfptr_t));
209 CTASSERT(sizeof(struct pmclog_map_out) == 5*4 + 2*sizeof(uintfptr_t) + TSDELTA);
210 CTASSERT(sizeof(struct pmclog_pmcallocate) == 9*4 + TSDELTA);
211 CTASSERT(sizeof(struct pmclog_pmcattach) == 5*4 + PATH_MAX + TSDELTA);
212 CTASSERT(offsetof(struct pmclog_pmcattach,pl_pathname) == 5*4 + TSDELTA);
213 CTASSERT(sizeof(struct pmclog_pmcdetach) == 5*4 + TSDELTA);
214 CTASSERT(sizeof(struct pmclog_proccsw) == 7*4 + 8 + TSDELTA);
215 CTASSERT(sizeof(struct pmclog_procexec) == 5*4 + PATH_MAX +
216     sizeof(uintfptr_t) + TSDELTA);
217 CTASSERT(offsetof(struct pmclog_procexec,pl_pathname) == 5*4 + TSDELTA +
218     sizeof(uintfptr_t));
219 CTASSERT(sizeof(struct pmclog_procexit) == 5*4 + 8 + TSDELTA);
220 CTASSERT(sizeof(struct pmclog_procfork) == 5*4 + TSDELTA);
221 CTASSERT(sizeof(struct pmclog_sysexit) == 6*4);
222 CTASSERT(sizeof(struct pmclog_userdata) == 6*4);
223 
224 /*
225  * Log buffer structure
226  */
227 
228 struct pmclog_buffer {
229 	TAILQ_ENTRY(pmclog_buffer) plb_next;
230 	char 		*plb_base;
231 	char		*plb_ptr;
232 	char 		*plb_fence;
233 	uint16_t	 plb_domain;
234 } __aligned(CACHE_LINE_SIZE);
235 
236 /*
237  * Prototypes
238  */
239 
240 static int pmclog_get_buffer(struct pmc_owner *po);
241 static void pmclog_loop(void *arg);
242 static void pmclog_release(struct pmc_owner *po);
243 static uint32_t *pmclog_reserve(struct pmc_owner *po, int length);
244 static void pmclog_schedule_io(struct pmc_owner *po, int wakeup);
245 static void pmclog_schedule_all(struct pmc_owner *po, int force);
246 static void pmclog_stop_kthread(struct pmc_owner *po);
247 
248 /*
249  * Helper functions
250  */
251 
252 static inline void
253 pmc_plb_rele_unlocked(struct pmclog_buffer *plb)
254 {
255 	TAILQ_INSERT_HEAD(&pmc_dom_hdrs[plb->plb_domain]->pdbh_head, plb, plb_next);
256 }
257 
258 static inline void
259 pmc_plb_rele(struct pmclog_buffer *plb)
260 {
261 	mtx_lock_spin(&pmc_dom_hdrs[plb->plb_domain]->pdbh_mtx);
262 	pmc_plb_rele_unlocked(plb);
263 	mtx_unlock_spin(&pmc_dom_hdrs[plb->plb_domain]->pdbh_mtx);
264 }
265 
266 /*
267  * Get a log buffer
268  */
269 static int
270 pmclog_get_buffer(struct pmc_owner *po)
271 {
272 	struct pmclog_buffer *plb;
273 	int domain;
274 
275 	KASSERT(po->po_curbuf[curcpu] == NULL,
276 	    ("[pmclog,%d] po=%p current buffer still valid", __LINE__, po));
277 
278 	domain = curdomain;
279 	MPASS(pmc_dom_hdrs[domain]);
280 	mtx_lock_spin(&pmc_dom_hdrs[domain]->pdbh_mtx);
281 	if ((plb = TAILQ_FIRST(&pmc_dom_hdrs[domain]->pdbh_head)) != NULL)
282 		TAILQ_REMOVE(&pmc_dom_hdrs[domain]->pdbh_head, plb, plb_next);
283 	mtx_unlock_spin(&pmc_dom_hdrs[domain]->pdbh_mtx);
284 
285 	PMCDBG2(LOG,GTB,1, "po=%p plb=%p", po, plb);
286 
287 #ifdef	HWPMC_DEBUG
288 	if (plb)
289 		KASSERT(plb->plb_ptr == plb->plb_base &&
290 		    plb->plb_base < plb->plb_fence,
291 		    ("[pmclog,%d] po=%p buffer invariants: ptr=%p "
292 		    "base=%p fence=%p", __LINE__, po, plb->plb_ptr,
293 		    plb->plb_base, plb->plb_fence));
294 #endif
295 
296 	po->po_curbuf[curcpu] = plb;
297 
298 	/* update stats */
299 	counter_u64_add(pmc_stats.pm_buffer_requests, 1);
300 	if (plb == NULL)
301 		counter_u64_add(pmc_stats.pm_buffer_requests_failed, 1);
302 
303 	return (plb ? 0 : ENOMEM);
304 }
305 
306 struct pmclog_proc_init_args {
307 	struct proc *kthr;
308 	struct pmc_owner *po;
309 	bool exit;
310 	bool acted;
311 };
312 
313 int
314 pmclog_proc_create(struct thread *td, void **handlep)
315 {
316 	struct pmclog_proc_init_args *ia;
317 	int error;
318 
319 	ia = malloc(sizeof(*ia), M_TEMP, M_WAITOK | M_ZERO);
320 	error = kproc_create(pmclog_loop, ia, &ia->kthr,
321 	    RFHIGHPID, 0, "hwpmc: proc(%d)", td->td_proc->p_pid);
322 	if (error == 0)
323 		*handlep = ia;
324 	return (error);
325 }
326 
327 void
328 pmclog_proc_ignite(void *handle, struct pmc_owner *po)
329 {
330 	struct pmclog_proc_init_args *ia;
331 
332 	ia = handle;
333 	mtx_lock(&pmc_kthread_mtx);
334 	MPASS(!ia->acted);
335 	MPASS(ia->po == NULL);
336 	MPASS(!ia->exit);
337 	MPASS(ia->kthr != NULL);
338 	if (po == NULL) {
339 		ia->exit = true;
340 	} else {
341 		ia->po = po;
342 		KASSERT(po->po_kthread == NULL,
343 		    ("[pmclog,%d] po=%p kthread (%p) already present",
344 		    __LINE__, po, po->po_kthread));
345 		po->po_kthread = ia->kthr;
346 	}
347 	wakeup(ia);
348 	while (!ia->acted)
349 		msleep(ia, &pmc_kthread_mtx, PWAIT, "pmclogw", 0);
350 	mtx_unlock(&pmc_kthread_mtx);
351 	free(ia, M_TEMP);
352 }
353 
354 /*
355  * Log handler loop.
356  *
357  * This function is executed by each pmc owner's helper thread.
358  */
359 static void
360 pmclog_loop(void *arg)
361 {
362 	struct pmclog_proc_init_args *ia;
363 	struct pmc_owner *po;
364 	struct pmclog_buffer *lb;
365 	struct proc *p;
366 	struct ucred *ownercred;
367 	struct ucred *mycred;
368 	struct thread *td;
369 	sigset_t unb;
370 	struct uio auio;
371 	struct iovec aiov;
372 	size_t nbytes;
373 	int error;
374 
375 	td = curthread;
376 
377 	SIGEMPTYSET(unb);
378 	SIGADDSET(unb, SIGHUP);
379 	(void)kern_sigprocmask(td, SIG_UNBLOCK, &unb, NULL, 0);
380 
381 	ia = arg;
382 	MPASS(ia->kthr == curproc);
383 	MPASS(!ia->acted);
384 	mtx_lock(&pmc_kthread_mtx);
385 	while (ia->po == NULL && !ia->exit)
386 		msleep(ia, &pmc_kthread_mtx, PWAIT, "pmclogi", 0);
387 	if (ia->exit) {
388 		ia->acted = true;
389 		wakeup(ia);
390 		mtx_unlock(&pmc_kthread_mtx);
391 		kproc_exit(0);
392 	}
393 	MPASS(ia->po != NULL);
394 	po = ia->po;
395 	ia->acted = true;
396 	wakeup(ia);
397 	mtx_unlock(&pmc_kthread_mtx);
398 	ia = NULL;
399 
400 	p = po->po_owner;
401 	mycred = td->td_ucred;
402 
403 	PROC_LOCK(p);
404 	ownercred = crhold(p->p_ucred);
405 	PROC_UNLOCK(p);
406 
407 	PMCDBG2(LOG,INI,1, "po=%p kt=%p", po, po->po_kthread);
408 	KASSERT(po->po_kthread == curthread->td_proc,
409 	    ("[pmclog,%d] proc mismatch po=%p po/kt=%p curproc=%p", __LINE__,
410 		po, po->po_kthread, curthread->td_proc));
411 
412 	lb = NULL;
413 
414 
415 	/*
416 	 * Loop waiting for I/O requests to be added to the owner
417 	 * struct's queue.  The loop is exited when the log file
418 	 * is deconfigured.
419 	 */
420 
421 	mtx_lock(&pmc_kthread_mtx);
422 
423 	for (;;) {
424 
425 		/* check if we've been asked to exit */
426 		if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
427 			break;
428 
429 		if (lb == NULL) { /* look for a fresh buffer to write */
430 			mtx_lock_spin(&po->po_mtx);
431 			if ((lb = TAILQ_FIRST(&po->po_logbuffers)) == NULL) {
432 				mtx_unlock_spin(&po->po_mtx);
433 
434 				/* No more buffers and shutdown required. */
435 				if (po->po_flags & PMC_PO_SHUTDOWN)
436 					break;
437 
438 				(void) msleep(po, &pmc_kthread_mtx, PWAIT,
439 				    "pmcloop", 250);
440 				continue;
441 			}
442 
443 			TAILQ_REMOVE(&po->po_logbuffers, lb, plb_next);
444 			mtx_unlock_spin(&po->po_mtx);
445 		}
446 
447 		mtx_unlock(&pmc_kthread_mtx);
448 
449 		/* process the request */
450 		PMCDBG3(LOG,WRI,2, "po=%p base=%p ptr=%p", po,
451 		    lb->plb_base, lb->plb_ptr);
452 		/* change our thread's credentials before issuing the I/O */
453 
454 		aiov.iov_base = lb->plb_base;
455 		aiov.iov_len  = nbytes = lb->plb_ptr - lb->plb_base;
456 
457 		auio.uio_iov    = &aiov;
458 		auio.uio_iovcnt = 1;
459 		auio.uio_offset = -1;
460 		auio.uio_resid  = nbytes;
461 		auio.uio_rw     = UIO_WRITE;
462 		auio.uio_segflg = UIO_SYSSPACE;
463 		auio.uio_td     = td;
464 
465 		/* switch thread credentials -- see kern_ktrace.c */
466 		td->td_ucred = ownercred;
467 		error = fo_write(po->po_file, &auio, ownercred, 0, td);
468 		td->td_ucred = mycred;
469 
470 		if (error) {
471 			/* XXX some errors are recoverable */
472 			/* send a SIGIO to the owner and exit */
473 			PROC_LOCK(p);
474 			kern_psignal(p, SIGIO);
475 			PROC_UNLOCK(p);
476 
477 			mtx_lock(&pmc_kthread_mtx);
478 
479 			po->po_error = error; /* save for flush log */
480 
481 			PMCDBG2(LOG,WRI,2, "po=%p error=%d", po, error);
482 
483 			break;
484 		}
485 
486 		mtx_lock(&pmc_kthread_mtx);
487 
488 		/* put the used buffer back into the global pool */
489 		PMCLOG_RESET_BUFFER_DESCRIPTOR(lb);
490 
491 		pmc_plb_rele(lb);
492 		lb = NULL;
493 	}
494 
495 	wakeup_one(po->po_kthread);
496 	po->po_kthread = NULL;
497 
498 	mtx_unlock(&pmc_kthread_mtx);
499 
500 	/* return the current I/O buffer to the global pool */
501 	if (lb) {
502 		PMCLOG_RESET_BUFFER_DESCRIPTOR(lb);
503 
504 		pmc_plb_rele(lb);
505 	}
506 
507 	/*
508 	 * Exit this thread, signalling the waiter
509 	 */
510 
511 	crfree(ownercred);
512 
513 	kproc_exit(0);
514 }
515 
516 /*
517  * Release and log entry and schedule an I/O if needed.
518  */
519 
520 static void
521 pmclog_release_flags(struct pmc_owner *po, int wakeup)
522 {
523 	struct pmclog_buffer *plb;
524 
525 	plb = po->po_curbuf[curcpu];
526 	KASSERT(plb->plb_ptr >= plb->plb_base,
527 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p base=%p", __LINE__,
528 		po, plb->plb_ptr, plb->plb_base));
529 	KASSERT(plb->plb_ptr <= plb->plb_fence,
530 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p fenc=%p", __LINE__,
531 		po, plb->plb_ptr, plb->plb_fence));
532 
533 	/* schedule an I/O if we've filled a buffer */
534 	if (plb->plb_ptr >= plb->plb_fence)
535 		pmclog_schedule_io(po, wakeup);
536 
537 	PMCDBG1(LOG,REL,1, "po=%p", po);
538 }
539 
540 static void
541 pmclog_release(struct pmc_owner *po)
542 {
543 
544 	pmclog_release_flags(po, 1);
545 }
546 
547 
548 /*
549  * Attempt to reserve 'length' bytes of space in an owner's log
550  * buffer.  The function returns a pointer to 'length' bytes of space
551  * if there was enough space or returns NULL if no space was
552  * available.  Non-null returns do so with the po mutex locked.  The
553  * caller must invoke pmclog_release() on the pmc owner structure
554  * when done.
555  */
556 
557 static uint32_t *
558 pmclog_reserve(struct pmc_owner *po, int length)
559 {
560 	uintptr_t newptr, oldptr;
561 	struct pmclog_buffer *plb, **pplb;
562 
563 	PMCDBG2(LOG,ALL,1, "po=%p len=%d", po, length);
564 
565 	KASSERT(length % sizeof(uint32_t) == 0,
566 	    ("[pmclog,%d] length not a multiple of word size", __LINE__));
567 
568 	/* No more data when shutdown in progress. */
569 	if (po->po_flags & PMC_PO_SHUTDOWN)
570 		return (NULL);
571 
572 	pplb = &po->po_curbuf[curcpu];
573 	if (*pplb == NULL && pmclog_get_buffer(po) != 0)
574 		goto fail;
575 
576 	KASSERT(*pplb != NULL,
577 	    ("[pmclog,%d] po=%p no current buffer", __LINE__, po));
578 
579 	plb = *pplb;
580 	KASSERT(plb->plb_ptr >= plb->plb_base &&
581 	    plb->plb_ptr <= plb->plb_fence,
582 	    ("[pmclog,%d] po=%p buffer invariants: ptr=%p base=%p fence=%p",
583 		__LINE__, po, plb->plb_ptr, plb->plb_base,
584 		plb->plb_fence));
585 
586 	oldptr = (uintptr_t) plb->plb_ptr;
587 	newptr = oldptr + length;
588 
589 	KASSERT(oldptr != (uintptr_t) NULL,
590 	    ("[pmclog,%d] po=%p Null log buffer pointer", __LINE__, po));
591 
592 	/*
593 	 * If we have space in the current buffer, return a pointer to
594 	 * available space with the PO structure locked.
595 	 */
596 	if (newptr <= (uintptr_t) plb->plb_fence) {
597 		plb->plb_ptr = (char *) newptr;
598 		goto done;
599 	}
600 
601 	/*
602 	 * Otherwise, schedule the current buffer for output and get a
603 	 * fresh buffer.
604 	 */
605 	pmclog_schedule_io(po, 0);
606 
607 	if (pmclog_get_buffer(po) != 0)
608 		goto fail;
609 
610 	plb = *pplb;
611 	KASSERT(plb != NULL,
612 	    ("[pmclog,%d] po=%p no current buffer", __LINE__, po));
613 
614 	KASSERT(plb->plb_ptr != NULL,
615 	    ("[pmclog,%d] null return from pmc_get_log_buffer", __LINE__));
616 
617 	KASSERT(plb->plb_ptr == plb->plb_base &&
618 	    plb->plb_ptr <= plb->plb_fence,
619 	    ("[pmclog,%d] po=%p buffer invariants: ptr=%p base=%p fence=%p",
620 		__LINE__, po, plb->plb_ptr, plb->plb_base,
621 		plb->plb_fence));
622 
623 	oldptr = (uintptr_t) plb->plb_ptr;
624 
625  done:
626 	return ((uint32_t *) oldptr);
627  fail:
628 	return (NULL);
629 }
630 
631 /*
632  * Schedule an I/O.
633  *
634  * Transfer the current buffer to the helper kthread.
635  */
636 
637 static void
638 pmclog_schedule_io(struct pmc_owner *po, int wakeup)
639 {
640 	struct pmclog_buffer *plb;
641 
642 	plb = po->po_curbuf[curcpu];
643 	po->po_curbuf[curcpu] = NULL;
644 	KASSERT(plb != NULL,
645 	    ("[pmclog,%d] schedule_io with null buffer po=%p", __LINE__, po));
646 	KASSERT(plb->plb_ptr >= plb->plb_base,
647 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p base=%p", __LINE__,
648 		po, plb->plb_ptr, plb->plb_base));
649 	KASSERT(plb->plb_ptr <= plb->plb_fence,
650 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p fenc=%p", __LINE__,
651 		po, plb->plb_ptr, plb->plb_fence));
652 
653 	PMCDBG1(LOG,SIO, 1, "po=%p", po);
654 
655 	/*
656 	 * Add the current buffer to the tail of the buffer list and
657 	 * wakeup the helper.
658 	 */
659 	mtx_lock_spin(&po->po_mtx);
660 	TAILQ_INSERT_TAIL(&po->po_logbuffers, plb, plb_next);
661 	mtx_unlock_spin(&po->po_mtx);
662 	if (wakeup)
663 		wakeup_one(po);
664 }
665 
666 /*
667  * Stop the helper kthread.
668  */
669 
670 static void
671 pmclog_stop_kthread(struct pmc_owner *po)
672 {
673 
674 	mtx_lock(&pmc_kthread_mtx);
675 	po->po_flags &= ~PMC_PO_OWNS_LOGFILE;
676 	if (po->po_kthread != NULL) {
677 		PROC_LOCK(po->po_kthread);
678 		kern_psignal(po->po_kthread, SIGHUP);
679 		PROC_UNLOCK(po->po_kthread);
680 	}
681 	wakeup_one(po);
682 	while (po->po_kthread)
683 		msleep(po->po_kthread, &pmc_kthread_mtx, PPAUSE, "pmckstp", 0);
684 	mtx_unlock(&pmc_kthread_mtx);
685 }
686 
687 /*
688  * Public functions
689  */
690 
691 /*
692  * Configure a log file for pmc owner 'po'.
693  *
694  * Parameter 'logfd' is a file handle referencing an open file in the
695  * owner process.  This file needs to have been opened for writing.
696  */
697 
698 int
699 pmclog_configure_log(struct pmc_mdep *md, struct pmc_owner *po, int logfd)
700 {
701 	struct proc *p;
702 	struct timespec ts;
703 	uint64_t tsc;
704 	int error;
705 
706 	sx_assert(&pmc_sx, SA_XLOCKED);
707 	PMCDBG2(LOG,CFG,1, "config po=%p logfd=%d", po, logfd);
708 
709 	p = po->po_owner;
710 
711 	/* return EBUSY if a log file was already present */
712 	if (po->po_flags & PMC_PO_OWNS_LOGFILE)
713 		return (EBUSY);
714 
715 	KASSERT(po->po_file == NULL,
716 	    ("[pmclog,%d] po=%p file (%p) already present", __LINE__, po,
717 		po->po_file));
718 
719 	/* get a reference to the file state */
720 	error = fget_write(curthread, logfd, &cap_write_rights, &po->po_file);
721 	if (error)
722 		goto error;
723 
724 	/* mark process as owning a log file */
725 	po->po_flags |= PMC_PO_OWNS_LOGFILE;
726 
727 	/* mark process as using HWPMCs */
728 	PROC_LOCK(p);
729 	p->p_flag |= P_HWPMC;
730 	PROC_UNLOCK(p);
731 	nanotime(&ts);
732 	tsc = pmc_rdtsc();
733 	/* create a log initialization entry */
734 	PMCLOG_RESERVE_WITH_ERROR(po, INITIALIZE,
735 	    sizeof(struct pmclog_initialize));
736 	PMCLOG_EMIT32(PMC_VERSION);
737 	PMCLOG_EMIT32(md->pmd_cputype);
738 #if defined(__i386__) || defined(__amd64__)
739 	PMCLOG_EMIT64(tsc_freq);
740 #else
741 	/* other architectures will need to fill this in */
742 	PMCLOG_EMIT32(0);
743 	PMCLOG_EMIT32(0);
744 #endif
745 	memcpy(_le, &ts, sizeof(ts));
746 	_le += sizeof(ts)/4;
747 	PMCLOG_EMITSTRING(pmc_cpuid, PMC_CPUID_LEN);
748 	PMCLOG_DESPATCH_SYNC(po);
749 
750 	return (0);
751 
752  error:
753 	KASSERT(po->po_kthread == NULL, ("[pmclog,%d] po=%p kthread not "
754 	    "stopped", __LINE__, po));
755 
756 	if (po->po_file)
757 		(void) fdrop(po->po_file, curthread);
758 	po->po_file  = NULL;	/* clear file and error state */
759 	po->po_error = 0;
760 	po->po_flags &= ~PMC_PO_OWNS_LOGFILE;
761 
762 	return (error);
763 }
764 
765 
766 /*
767  * De-configure a log file.  This will throw away any buffers queued
768  * for this owner process.
769  */
770 
771 int
772 pmclog_deconfigure_log(struct pmc_owner *po)
773 {
774 	int error;
775 	struct pmclog_buffer *lb;
776 
777 	PMCDBG1(LOG,CFG,1, "de-config po=%p", po);
778 
779 	if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
780 		return (EINVAL);
781 
782 	KASSERT(po->po_sscount == 0,
783 	    ("[pmclog,%d] po=%p still owning SS PMCs", __LINE__, po));
784 	KASSERT(po->po_file != NULL,
785 	    ("[pmclog,%d] po=%p no log file", __LINE__, po));
786 
787 	/* stop the kthread, this will reset the 'OWNS_LOGFILE' flag */
788 	pmclog_stop_kthread(po);
789 
790 	KASSERT(po->po_kthread == NULL,
791 	    ("[pmclog,%d] po=%p kthread not stopped", __LINE__, po));
792 
793 	/* return all queued log buffers to the global pool */
794 	while ((lb = TAILQ_FIRST(&po->po_logbuffers)) != NULL) {
795 		TAILQ_REMOVE(&po->po_logbuffers, lb, plb_next);
796 		PMCLOG_RESET_BUFFER_DESCRIPTOR(lb);
797 		pmc_plb_rele(lb);
798 	}
799 	for (int i = 0; i < mp_ncpus; i++) {
800 		thread_lock(curthread);
801 		sched_bind(curthread, i);
802 		thread_unlock(curthread);
803 		/* return the 'current' buffer to the global pool */
804 		if ((lb = po->po_curbuf[curcpu]) != NULL) {
805 			PMCLOG_RESET_BUFFER_DESCRIPTOR(lb);
806 			pmc_plb_rele(lb);
807 		}
808 	}
809 	thread_lock(curthread);
810 	sched_unbind(curthread);
811 	thread_unlock(curthread);
812 
813 	/* drop a reference to the fd */
814 	if (po->po_file != NULL) {
815 		error = fdrop(po->po_file, curthread);
816 		po->po_file = NULL;
817 	} else
818 		error = 0;
819 	po->po_error = 0;
820 
821 	return (error);
822 }
823 
824 /*
825  * Flush a process' log buffer.
826  */
827 
828 int
829 pmclog_flush(struct pmc_owner *po, int force)
830 {
831 	int error;
832 
833 	PMCDBG1(LOG,FLS,1, "po=%p", po);
834 
835 	/*
836 	 * If there is a pending error recorded by the logger thread,
837 	 * return that.
838 	 */
839 	if (po->po_error)
840 		return (po->po_error);
841 
842 	error = 0;
843 
844 	/*
845 	 * Check that we do have an active log file.
846 	 */
847 	mtx_lock(&pmc_kthread_mtx);
848 	if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) {
849 		error = EINVAL;
850 		goto error;
851 	}
852 
853 	pmclog_schedule_all(po, force);
854  error:
855 	mtx_unlock(&pmc_kthread_mtx);
856 
857 	return (error);
858 }
859 
860 static void
861 pmclog_schedule_one_cond(struct pmc_owner *po, int force)
862 {
863 	struct pmclog_buffer *plb;
864 	int cpu;
865 
866 	spinlock_enter();
867 	cpu = curcpu;
868 	/* tell hardclock not to run again */
869 	if (PMC_CPU_HAS_SAMPLES(cpu))
870 		PMC_CALL_HOOK_UNLOCKED(curthread, PMC_FN_DO_SAMPLES, NULL);
871 	if (force)
872 		pmc_flush_samples(cpu);
873 	plb = po->po_curbuf[cpu];
874 	if (plb && plb->plb_ptr != plb->plb_base)
875 		pmclog_schedule_io(po, 1);
876 	spinlock_exit();
877 }
878 
879 static void
880 pmclog_schedule_all(struct pmc_owner *po, int force)
881 {
882 	/*
883 	 * Schedule the current buffer if any and not empty.
884 	 */
885 	for (int i = 0; i < mp_ncpus; i++) {
886 		thread_lock(curthread);
887 		sched_bind(curthread, i);
888 		thread_unlock(curthread);
889 		pmclog_schedule_one_cond(po, force);
890 	}
891 	thread_lock(curthread);
892 	sched_unbind(curthread);
893 	thread_unlock(curthread);
894 }
895 
896 int
897 pmclog_close(struct pmc_owner *po)
898 {
899 
900 	PMCDBG1(LOG,CLO,1, "po=%p", po);
901 
902 	pmclog_process_closelog(po);
903 
904 	mtx_lock(&pmc_kthread_mtx);
905 	/*
906 	 * Initiate shutdown: no new data queued,
907 	 * thread will close file on last block.
908 	 */
909 	po->po_flags |= PMC_PO_SHUTDOWN;
910 	/* give time for all to see */
911 	DELAY(50);
912 
913 	/*
914 	 * Schedule the current buffer.
915 	 */
916 	pmclog_schedule_all(po, 0);
917 	wakeup_one(po);
918 
919 	mtx_unlock(&pmc_kthread_mtx);
920 
921 	return (0);
922 }
923 
924 void
925 pmclog_process_callchain(struct pmc *pm, struct pmc_sample *ps)
926 {
927 	int n, recordlen;
928 	uint32_t flags;
929 	struct pmc_owner *po;
930 
931 	PMCDBG3(LOG,SAM,1,"pm=%p pid=%d n=%d", pm, ps->ps_pid,
932 	    ps->ps_nsamples);
933 
934 	recordlen = offsetof(struct pmclog_callchain, pl_pc) +
935 	    ps->ps_nsamples * sizeof(uintfptr_t);
936 	po = pm->pm_owner;
937 	flags = PMC_CALLCHAIN_TO_CPUFLAGS(ps->ps_cpu,ps->ps_flags);
938 	PMCLOG_RESERVE_SAFE(po, CALLCHAIN, recordlen, ps->ps_tsc);
939 	PMCLOG_EMIT32(ps->ps_pid);
940 	PMCLOG_EMIT32(ps->ps_tid);
941 	PMCLOG_EMIT32(pm->pm_id);
942 	PMCLOG_EMIT32(flags);
943 	for (n = 0; n < ps->ps_nsamples; n++)
944 		PMCLOG_EMITADDR(ps->ps_pc[n]);
945 	PMCLOG_DESPATCH_SAFE(po);
946 }
947 
948 void
949 pmclog_process_closelog(struct pmc_owner *po)
950 {
951 	PMCLOG_RESERVE(po,CLOSELOG,sizeof(struct pmclog_closelog));
952 	PMCLOG_DESPATCH_SYNC(po);
953 }
954 
955 void
956 pmclog_process_dropnotify(struct pmc_owner *po)
957 {
958 	PMCLOG_RESERVE(po,DROPNOTIFY,sizeof(struct pmclog_dropnotify));
959 	PMCLOG_DESPATCH(po);
960 }
961 
962 void
963 pmclog_process_map_in(struct pmc_owner *po, pid_t pid, uintfptr_t start,
964     const char *path)
965 {
966 	int pathlen, recordlen;
967 
968 	KASSERT(path != NULL, ("[pmclog,%d] map-in, null path", __LINE__));
969 
970 	pathlen = strlen(path) + 1;	/* #bytes for path name */
971 	recordlen = offsetof(struct pmclog_map_in, pl_pathname) +
972 	    pathlen;
973 
974 	PMCLOG_RESERVE(po, MAP_IN, recordlen);
975 	PMCLOG_EMIT32(pid);
976 	PMCLOG_EMIT32(0);
977 	PMCLOG_EMITADDR(start);
978 	PMCLOG_EMITSTRING(path,pathlen);
979 	PMCLOG_DESPATCH_SYNC(po);
980 }
981 
982 void
983 pmclog_process_map_out(struct pmc_owner *po, pid_t pid, uintfptr_t start,
984     uintfptr_t end)
985 {
986 	KASSERT(start <= end, ("[pmclog,%d] start > end", __LINE__));
987 
988 	PMCLOG_RESERVE(po, MAP_OUT, sizeof(struct pmclog_map_out));
989 	PMCLOG_EMIT32(pid);
990 	PMCLOG_EMIT32(0);
991 	PMCLOG_EMITADDR(start);
992 	PMCLOG_EMITADDR(end);
993 	PMCLOG_DESPATCH(po);
994 }
995 
996 void
997 pmclog_process_pmcallocate(struct pmc *pm)
998 {
999 	struct pmc_owner *po;
1000 	struct pmc_soft *ps;
1001 
1002 	po = pm->pm_owner;
1003 
1004 	PMCDBG1(LOG,ALL,1, "pm=%p", pm);
1005 
1006 	if (PMC_TO_CLASS(pm) == PMC_CLASS_SOFT) {
1007 		PMCLOG_RESERVE(po, PMCALLOCATEDYN,
1008 		    sizeof(struct pmclog_pmcallocatedyn));
1009 		PMCLOG_EMIT32(pm->pm_id);
1010 		PMCLOG_EMIT32(pm->pm_event);
1011 		PMCLOG_EMIT32(pm->pm_flags);
1012 		PMCLOG_EMIT32(0);
1013 		PMCLOG_EMIT64(pm->pm_sc.pm_reloadcount);
1014 		ps = pmc_soft_ev_acquire(pm->pm_event);
1015 		if (ps != NULL)
1016 			PMCLOG_EMITSTRING(ps->ps_ev.pm_ev_name,PMC_NAME_MAX);
1017 		else
1018 			PMCLOG_EMITNULLSTRING(PMC_NAME_MAX);
1019 		pmc_soft_ev_release(ps);
1020 		PMCLOG_DESPATCH_SYNC(po);
1021 	} else {
1022 		PMCLOG_RESERVE(po, PMCALLOCATE,
1023 		    sizeof(struct pmclog_pmcallocate));
1024 		PMCLOG_EMIT32(pm->pm_id);
1025 		PMCLOG_EMIT32(pm->pm_event);
1026 		PMCLOG_EMIT32(pm->pm_flags);
1027 		PMCLOG_EMIT32(0);
1028 		PMCLOG_EMIT64(pm->pm_sc.pm_reloadcount);
1029 		PMCLOG_DESPATCH_SYNC(po);
1030 	}
1031 }
1032 
1033 void
1034 pmclog_process_pmcattach(struct pmc *pm, pid_t pid, char *path)
1035 {
1036 	int pathlen, recordlen;
1037 	struct pmc_owner *po;
1038 
1039 	PMCDBG2(LOG,ATT,1,"pm=%p pid=%d", pm, pid);
1040 
1041 	po = pm->pm_owner;
1042 
1043 	pathlen = strlen(path) + 1;	/* #bytes for the string */
1044 	recordlen = offsetof(struct pmclog_pmcattach, pl_pathname) + pathlen;
1045 
1046 	PMCLOG_RESERVE(po, PMCATTACH, recordlen);
1047 	PMCLOG_EMIT32(pm->pm_id);
1048 	PMCLOG_EMIT32(pid);
1049 	PMCLOG_EMITSTRING(path, pathlen);
1050 	PMCLOG_DESPATCH_SYNC(po);
1051 }
1052 
1053 void
1054 pmclog_process_pmcdetach(struct pmc *pm, pid_t pid)
1055 {
1056 	struct pmc_owner *po;
1057 
1058 	PMCDBG2(LOG,ATT,1,"!pm=%p pid=%d", pm, pid);
1059 
1060 	po = pm->pm_owner;
1061 
1062 	PMCLOG_RESERVE(po, PMCDETACH, sizeof(struct pmclog_pmcdetach));
1063 	PMCLOG_EMIT32(pm->pm_id);
1064 	PMCLOG_EMIT32(pid);
1065 	PMCLOG_DESPATCH_SYNC(po);
1066 }
1067 
1068 void
1069 pmclog_process_proccreate(struct pmc_owner *po, struct proc *p, int sync)
1070 {
1071 	if (sync) {
1072 		PMCLOG_RESERVE(po, PROC_CREATE, sizeof(struct pmclog_proccreate));
1073 		PMCLOG_EMIT32(p->p_pid);
1074 		PMCLOG_EMIT32(p->p_flag);
1075 		PMCLOG_EMITSTRING(p->p_comm, MAXCOMLEN+1);
1076 		PMCLOG_DESPATCH_SYNC(po);
1077 	} else {
1078 		PMCLOG_RESERVE(po, PROC_CREATE, sizeof(struct pmclog_proccreate));
1079 		PMCLOG_EMIT32(p->p_pid);
1080 		PMCLOG_EMIT32(p->p_flag);
1081 		PMCLOG_EMITSTRING(p->p_comm, MAXCOMLEN+1);
1082 		PMCLOG_DESPATCH(po);
1083 	}
1084 }
1085 
1086 /*
1087  * Log a context switch event to the log file.
1088  */
1089 
1090 void
1091 pmclog_process_proccsw(struct pmc *pm, struct pmc_process *pp, pmc_value_t v, struct thread *td)
1092 {
1093 	struct pmc_owner *po;
1094 
1095 	KASSERT(pm->pm_flags & PMC_F_LOG_PROCCSW,
1096 	    ("[pmclog,%d] log-process-csw called gratuitously", __LINE__));
1097 
1098 	PMCDBG3(LOG,SWO,1,"pm=%p pid=%d v=%jx", pm, pp->pp_proc->p_pid,
1099 	    v);
1100 
1101 	po = pm->pm_owner;
1102 
1103 	PMCLOG_RESERVE_SAFE(po, PROCCSW, sizeof(struct pmclog_proccsw), pmc_rdtsc());
1104 	PMCLOG_EMIT64(v);
1105 	PMCLOG_EMIT32(pm->pm_id);
1106 	PMCLOG_EMIT32(pp->pp_proc->p_pid);
1107 	PMCLOG_EMIT32(td->td_tid);
1108 	PMCLOG_EMIT32(0);
1109 	PMCLOG_DESPATCH_SCHED_LOCK(po);
1110 }
1111 
1112 void
1113 pmclog_process_procexec(struct pmc_owner *po, pmc_id_t pmid, pid_t pid,
1114     uintfptr_t startaddr, char *path)
1115 {
1116 	int pathlen, recordlen;
1117 
1118 	PMCDBG3(LOG,EXC,1,"po=%p pid=%d path=\"%s\"", po, pid, path);
1119 
1120 	pathlen   = strlen(path) + 1;	/* #bytes for the path */
1121 	recordlen = offsetof(struct pmclog_procexec, pl_pathname) + pathlen;
1122 	PMCLOG_RESERVE(po, PROCEXEC, recordlen);
1123 	PMCLOG_EMIT32(pid);
1124 	PMCLOG_EMIT32(pmid);
1125 	PMCLOG_EMITADDR(startaddr);
1126 	PMCLOG_EMITSTRING(path,pathlen);
1127 	PMCLOG_DESPATCH_SYNC(po);
1128 }
1129 
1130 /*
1131  * Log a process exit event (and accumulated pmc value) to the log file.
1132  */
1133 
1134 void
1135 pmclog_process_procexit(struct pmc *pm, struct pmc_process *pp)
1136 {
1137 	int ri;
1138 	struct pmc_owner *po;
1139 
1140 	ri = PMC_TO_ROWINDEX(pm);
1141 	PMCDBG3(LOG,EXT,1,"pm=%p pid=%d v=%jx", pm, pp->pp_proc->p_pid,
1142 	    pp->pp_pmcs[ri].pp_pmcval);
1143 
1144 	po = pm->pm_owner;
1145 
1146 	PMCLOG_RESERVE(po, PROCEXIT, sizeof(struct pmclog_procexit));
1147 	PMCLOG_EMIT32(pm->pm_id);
1148 	PMCLOG_EMIT32(pp->pp_proc->p_pid);
1149 	PMCLOG_EMIT64(pp->pp_pmcs[ri].pp_pmcval);
1150 	PMCLOG_DESPATCH(po);
1151 }
1152 
1153 /*
1154  * Log a fork event.
1155  */
1156 
1157 void
1158 pmclog_process_procfork(struct pmc_owner *po, pid_t oldpid, pid_t newpid)
1159 {
1160 	PMCLOG_RESERVE(po, PROCFORK, sizeof(struct pmclog_procfork));
1161 	PMCLOG_EMIT32(oldpid);
1162 	PMCLOG_EMIT32(newpid);
1163 	PMCLOG_DESPATCH(po);
1164 }
1165 
1166 /*
1167  * Log a process exit event of the form suitable for system-wide PMCs.
1168  */
1169 
1170 void
1171 pmclog_process_sysexit(struct pmc_owner *po, pid_t pid)
1172 {
1173 	PMCLOG_RESERVE(po, SYSEXIT, sizeof(struct pmclog_sysexit));
1174 	PMCLOG_EMIT32(pid);
1175 	PMCLOG_DESPATCH(po);
1176 }
1177 
1178 void
1179 pmclog_process_threadcreate(struct pmc_owner *po, struct thread *td, int sync)
1180 {
1181 	struct proc *p;
1182 
1183 	p = td->td_proc;
1184 	if (sync) {
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_SYNC(po);
1192 	} else {
1193 		PMCLOG_RESERVE(po, THR_CREATE, sizeof(struct pmclog_threadcreate));
1194 		PMCLOG_EMIT32(td->td_tid);
1195 		PMCLOG_EMIT32(p->p_pid);
1196 		PMCLOG_EMIT32(p->p_flag);
1197 		PMCLOG_EMIT32(0);
1198 		PMCLOG_EMITSTRING(td->td_name, MAXCOMLEN+1);
1199 		PMCLOG_DESPATCH(po);
1200 	}
1201 }
1202 
1203 void
1204 pmclog_process_threadexit(struct pmc_owner *po, struct thread *td)
1205 {
1206 
1207 	PMCLOG_RESERVE(po, THR_EXIT, sizeof(struct pmclog_threadexit));
1208 	PMCLOG_EMIT32(td->td_tid);
1209 	PMCLOG_DESPATCH(po);
1210 }
1211 
1212 /*
1213  * Write a user log entry.
1214  */
1215 
1216 int
1217 pmclog_process_userlog(struct pmc_owner *po, struct pmc_op_writelog *wl)
1218 {
1219 	int error;
1220 
1221 	PMCDBG2(LOG,WRI,1, "writelog po=%p ud=0x%x", po, wl->pm_userdata);
1222 
1223 	error = 0;
1224 
1225 	PMCLOG_RESERVE_WITH_ERROR(po, USERDATA,
1226 	    sizeof(struct pmclog_userdata));
1227 	PMCLOG_EMIT32(wl->pm_userdata);
1228 	PMCLOG_DESPATCH(po);
1229 
1230  error:
1231 	return (error);
1232 }
1233 
1234 /*
1235  * Initialization.
1236  *
1237  * Create a pool of log buffers and initialize mutexes.
1238  */
1239 
1240 void
1241 pmclog_initialize()
1242 {
1243 	int domain;
1244 	struct pmclog_buffer *plb;
1245 
1246 	if (pmclog_buffer_size <= 0 || pmclog_buffer_size > 16*1024) {
1247 		(void) printf("hwpmc: tunable logbuffersize=%d must be "
1248 					  "greater than zero and less than or equal to 16MB.\n",
1249 					  pmclog_buffer_size);
1250 		pmclog_buffer_size = PMC_LOG_BUFFER_SIZE;
1251 	}
1252 
1253 	if (pmc_nlogbuffers_pcpu <= 0) {
1254 		(void) printf("hwpmc: tunable nlogbuffers=%d must be greater "
1255 					  "than zero.\n", pmc_nlogbuffers_pcpu);
1256 		pmc_nlogbuffers_pcpu = PMC_NLOGBUFFERS_PCPU;
1257 	}
1258 	if (pmc_nlogbuffers_pcpu*pmclog_buffer_size > 32*1024) {
1259 		(void) printf("hwpmc: memory allocated pcpu must be less than 32MB (is %dK).\n",
1260 					  pmc_nlogbuffers_pcpu*pmclog_buffer_size);
1261 		pmc_nlogbuffers_pcpu = PMC_NLOGBUFFERS_PCPU;
1262 		pmclog_buffer_size = PMC_LOG_BUFFER_SIZE;
1263 	}
1264 	for (domain = 0; domain < NDOMAINS; domain++) {
1265 		int ncpus = pmc_dom_hdrs[domain]->pdbh_ncpus;
1266 		int total = ncpus*pmc_nlogbuffers_pcpu;
1267 
1268 		plb = malloc_domain(sizeof(struct pmclog_buffer)*total, M_PMC, domain, M_WAITOK|M_ZERO);
1269 		pmc_dom_hdrs[domain]->pdbh_plbs = plb;
1270 		for (int i = 0; i < total; i++, plb++) {
1271 			void *buf;
1272 
1273 			buf = malloc_domain(1024 * pmclog_buffer_size, M_PMC, domain,
1274 								M_WAITOK|M_ZERO);
1275 			PMCLOG_INIT_BUFFER_DESCRIPTOR(plb, buf, domain);
1276 			pmc_plb_rele_unlocked(plb);
1277 		}
1278 	}
1279 	mtx_init(&pmc_kthread_mtx, "pmc-kthread", "pmc-sleep", MTX_DEF);
1280 }
1281 
1282 /*
1283  * Shutdown logging.
1284  *
1285  * Destroy mutexes and release memory back the to free pool.
1286  */
1287 
1288 void
1289 pmclog_shutdown()
1290 {
1291 	struct pmclog_buffer *plb;
1292 	int domain;
1293 
1294 	mtx_destroy(&pmc_kthread_mtx);
1295 
1296 	for (domain = 0; domain < NDOMAINS; domain++) {
1297 		while ((plb = TAILQ_FIRST(&pmc_dom_hdrs[domain]->pdbh_head)) != NULL) {
1298 			TAILQ_REMOVE(&pmc_dom_hdrs[domain]->pdbh_head, plb, plb_next);
1299 			free(plb->plb_base, M_PMC);
1300 		}
1301 		free(pmc_dom_hdrs[domain]->pdbh_plbs, M_PMC);
1302 	}
1303 }
1304