xref: /freebsd/sys/dev/hwpmc/hwpmc_logging.c (revision 0dbdecfd4f23486da21345fe446a28a1c27db3f2)
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 #define curdomain PCPU_GET(domain)
69 
70 /*
71  * Sysctl tunables
72  */
73 
74 SYSCTL_DECL(_kern_hwpmc);
75 
76 /*
77  * kern.hwpmc.logbuffersize -- size of the per-cpu owner buffers.
78  */
79 
80 static int pmclog_buffer_size = PMC_LOG_BUFFER_SIZE;
81 #if (__FreeBSD_version < 1100000)
82 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "logbuffersize", &pmclog_buffer_size);
83 #endif
84 SYSCTL_INT(_kern_hwpmc, OID_AUTO, logbuffersize, CTLFLAG_RDTUN,
85     &pmclog_buffer_size, 0, "size of log buffers in kilobytes");
86 
87 /*
88  * kern.hwpmc.nbuffer -- number of global log buffers
89  */
90 
91 static int pmc_nlogbuffers_pcpu = PMC_NLOGBUFFERS_PCPU;
92 #if (__FreeBSD_version < 1100000)
93 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "nbuffers", &pmc_nlogbuffers_pcpu);
94 #endif
95 SYSCTL_INT(_kern_hwpmc, OID_AUTO, nbuffers_pcpu, CTLFLAG_RDTUN,
96     &pmc_nlogbuffers_pcpu, 0, "number of log buffers per cpu");
97 
98 /*
99  * Global log buffer list and associated spin lock.
100  */
101 
102 static struct mtx pmc_kthread_mtx;	/* sleep lock */
103 
104 #define	PMCLOG_INIT_BUFFER_DESCRIPTOR(D, buf, domain) do {						\
105 		(D)->plb_fence = ((char *) (buf)) +	1024*pmclog_buffer_size;			\
106 		(D)->plb_base  = (D)->plb_ptr = ((char *) (buf));				\
107 		(D)->plb_domain = domain; \
108 	} while (0)
109 
110 #define	PMCLOG_RESET_BUFFER_DESCRIPTOR(D) do {			\
111 		(D)->plb_ptr  = (D)->plb_base; \
112 	} while (0)
113 
114 /*
115  * Log file record constructors.
116  */
117 #define	_PMCLOG_TO_HEADER(T,L)						\
118 	((PMCLOG_HEADER_MAGIC << 24) |					\
119 	 (PMCLOG_TYPE_ ## T << 16)   |					\
120 	 ((L) & 0xFFFF))
121 
122 /* reserve LEN bytes of space and initialize the entry header */
123 #define	_PMCLOG_RESERVE_SAFE(PO,TYPE,LEN,ACTION, TSC) do {	\
124 		uint32_t *_le;						\
125 		int _len = roundup((LEN), sizeof(uint32_t));	\
126 		struct pmclog_header *ph;							\
127 		if ((_le = pmclog_reserve((PO), _len)) == NULL) {	\
128 			ACTION;											\
129 		}													\
130 		ph = (struct pmclog_header *)_le;					\
131 		ph->pl_header =_PMCLOG_TO_HEADER(TYPE,_len);	\
132 		ph->pl_tsc = (TSC);									\
133 		_le += sizeof(*ph)/4	/* skip over timestamp */
134 
135 /* reserve LEN bytes of space and initialize the entry header */
136 #define	_PMCLOG_RESERVE(PO,TYPE,LEN,ACTION) do {			\
137 		uint32_t *_le;						\
138 		int _len = roundup((LEN), sizeof(uint32_t));	\
139 		uint64_t tsc;										\
140 		struct pmclog_header *ph;							\
141 		tsc = pmc_rdtsc();									\
142 		spinlock_enter();									\
143 		if ((_le = pmclog_reserve((PO), _len)) == NULL) {	\
144 			spinlock_exit();								\
145 			ACTION;											\
146 		}												\
147 		ph = (struct pmclog_header *)_le;					\
148 		ph->pl_header =_PMCLOG_TO_HEADER(TYPE,_len);	\
149 		ph->pl_tsc = tsc;									\
150 		_le += sizeof(*ph)/4	/* skip over timestamp */
151 
152 
153 
154 #define	PMCLOG_RESERVE_SAFE(P,T,L,TSC)		_PMCLOG_RESERVE_SAFE(P,T,L,return,TSC)
155 #define	PMCLOG_RESERVE(P,T,L)		_PMCLOG_RESERVE(P,T,L,return)
156 #define	PMCLOG_RESERVE_WITH_ERROR(P,T,L) _PMCLOG_RESERVE(P,T,L,		\
157 	error=ENOMEM;goto error)
158 
159 #define	PMCLOG_EMIT32(V)	do { *_le++ = (V); } while (0)
160 #define	PMCLOG_EMIT64(V)	do { 					\
161 		*_le++ = (uint32_t) ((V) & 0xFFFFFFFF);			\
162 		*_le++ = (uint32_t) (((V) >> 32) & 0xFFFFFFFF);		\
163 	} while (0)
164 
165 
166 /* Emit a string.  Caution: does NOT update _le, so needs to be last */
167 #define	PMCLOG_EMITSTRING(S,L)	do { bcopy((S), _le, (L)); } while (0)
168 #define	PMCLOG_EMITNULLSTRING(L) do { bzero(_le, (L)); } while (0)
169 
170 #define	PMCLOG_DESPATCH_SAFE(PO)						\
171 	    pmclog_release((PO));						\
172 	} while (0)
173 
174 #define	PMCLOG_DESPATCH_SCHED_LOCK(PO)						\
175 	     pmclog_release_flags((PO), 0);							\
176 	} while (0)
177 
178 #define	PMCLOG_DESPATCH(PO)							\
179 	    pmclog_release((PO));						\
180 		spinlock_exit();							\
181 	} while (0)
182 
183 #define	PMCLOG_DESPATCH_SYNC(PO)						\
184 	    pmclog_schedule_io((PO), 1);						\
185 		spinlock_exit();								\
186 		} while (0)
187 
188 
189 #define TSDELTA 4
190 /*
191  * Assertions about the log file format.
192  */
193 CTASSERT(sizeof(struct pmclog_callchain) == 7*4 + TSDELTA +
194     PMC_CALLCHAIN_DEPTH_MAX*sizeof(uintfptr_t));
195 CTASSERT(sizeof(struct pmclog_closelog) == 3*4 + TSDELTA);
196 CTASSERT(sizeof(struct pmclog_dropnotify) == 3*4 + TSDELTA);
197 CTASSERT(sizeof(struct pmclog_map_in) == PATH_MAX + TSDELTA +
198     5*4 + sizeof(uintfptr_t));
199 CTASSERT(offsetof(struct pmclog_map_in,pl_pathname) ==
200     5*4 + TSDELTA + sizeof(uintfptr_t));
201 CTASSERT(sizeof(struct pmclog_map_out) == 5*4 + 2*sizeof(uintfptr_t) + TSDELTA);
202 CTASSERT(sizeof(struct pmclog_pmcallocate) == 9*4 + TSDELTA);
203 CTASSERT(sizeof(struct pmclog_pmcattach) == 5*4 + PATH_MAX + TSDELTA);
204 CTASSERT(offsetof(struct pmclog_pmcattach,pl_pathname) == 5*4 + TSDELTA);
205 CTASSERT(sizeof(struct pmclog_pmcdetach) == 5*4 + TSDELTA);
206 CTASSERT(sizeof(struct pmclog_proccsw) == 7*4 + 8 + TSDELTA);
207 CTASSERT(sizeof(struct pmclog_procexec) == 5*4 + PATH_MAX +
208     sizeof(uintfptr_t) + TSDELTA);
209 CTASSERT(offsetof(struct pmclog_procexec,pl_pathname) == 5*4 + TSDELTA +
210     sizeof(uintfptr_t));
211 CTASSERT(sizeof(struct pmclog_procexit) == 5*4 + 8 + TSDELTA);
212 CTASSERT(sizeof(struct pmclog_procfork) == 5*4 + TSDELTA);
213 CTASSERT(sizeof(struct pmclog_sysexit) == 6*4);
214 CTASSERT(sizeof(struct pmclog_userdata) == 6*4);
215 
216 /*
217  * Log buffer structure
218  */
219 
220 struct pmclog_buffer {
221 	TAILQ_ENTRY(pmclog_buffer) plb_next;
222 	char 		*plb_base;
223 	char		*plb_ptr;
224 	char 		*plb_fence;
225 	uint16_t	 plb_domain;
226 } __aligned(CACHE_LINE_SIZE);
227 
228 /*
229  * Prototypes
230  */
231 
232 static int pmclog_get_buffer(struct pmc_owner *po);
233 static void pmclog_loop(void *arg);
234 static void pmclog_release(struct pmc_owner *po);
235 static uint32_t *pmclog_reserve(struct pmc_owner *po, int length);
236 static void pmclog_schedule_io(struct pmc_owner *po, int wakeup);
237 static void pmclog_schedule_all(struct pmc_owner *po, int force);
238 static void pmclog_stop_kthread(struct pmc_owner *po);
239 
240 /*
241  * Helper functions
242  */
243 
244 static inline void
245 pmc_plb_rele_unlocked(struct pmclog_buffer *plb)
246 {
247 	TAILQ_INSERT_HEAD(&pmc_dom_hdrs[plb->plb_domain]->pdbh_head, plb, plb_next);
248 }
249 
250 static inline void
251 pmc_plb_rele(struct pmclog_buffer *plb)
252 {
253 	mtx_lock_spin(&pmc_dom_hdrs[plb->plb_domain]->pdbh_mtx);
254 	pmc_plb_rele_unlocked(plb);
255 	mtx_unlock_spin(&pmc_dom_hdrs[plb->plb_domain]->pdbh_mtx);
256 }
257 
258 /*
259  * Get a log buffer
260  */
261 static int
262 pmclog_get_buffer(struct pmc_owner *po)
263 {
264 	struct pmclog_buffer *plb;
265 	int domain;
266 
267 	KASSERT(po->po_curbuf[curcpu] == NULL,
268 	    ("[pmclog,%d] po=%p current buffer still valid", __LINE__, po));
269 
270 	domain = curdomain;
271 	MPASS(pmc_dom_hdrs[domain]);
272 	mtx_lock_spin(&pmc_dom_hdrs[domain]->pdbh_mtx);
273 	if ((plb = TAILQ_FIRST(&pmc_dom_hdrs[domain]->pdbh_head)) != NULL)
274 		TAILQ_REMOVE(&pmc_dom_hdrs[domain]->pdbh_head, plb, plb_next);
275 	mtx_unlock_spin(&pmc_dom_hdrs[domain]->pdbh_mtx);
276 
277 	PMCDBG2(LOG,GTB,1, "po=%p plb=%p", po, plb);
278 
279 #ifdef	HWPMC_DEBUG
280 	if (plb)
281 		KASSERT(plb->plb_ptr == plb->plb_base &&
282 		    plb->plb_base < plb->plb_fence,
283 		    ("[pmclog,%d] po=%p buffer invariants: ptr=%p "
284 		    "base=%p fence=%p", __LINE__, po, plb->plb_ptr,
285 		    plb->plb_base, plb->plb_fence));
286 #endif
287 
288 	po->po_curbuf[curcpu] = plb;
289 
290 	/* update stats */
291 	counter_u64_add(pmc_stats.pm_buffer_requests, 1);
292 	if (plb == NULL)
293 		counter_u64_add(pmc_stats.pm_buffer_requests_failed, 1);
294 
295 	return (plb ? 0 : ENOMEM);
296 }
297 
298 struct pmclog_proc_init_args {
299 	struct proc *kthr;
300 	struct pmc_owner *po;
301 	bool exit;
302 	bool acted;
303 };
304 
305 int
306 pmclog_proc_create(struct thread *td, void **handlep)
307 {
308 	struct pmclog_proc_init_args *ia;
309 	int error;
310 
311 	ia = malloc(sizeof(*ia), M_TEMP, M_WAITOK | M_ZERO);
312 	error = kproc_create(pmclog_loop, ia, &ia->kthr,
313 	    RFHIGHPID, 0, "hwpmc: proc(%d)", td->td_proc->p_pid);
314 	if (error == 0)
315 		*handlep = ia;
316 	return (error);
317 }
318 
319 void
320 pmclog_proc_ignite(void *handle, struct pmc_owner *po)
321 {
322 	struct pmclog_proc_init_args *ia;
323 
324 	ia = handle;
325 	mtx_lock(&pmc_kthread_mtx);
326 	MPASS(!ia->acted);
327 	MPASS(ia->po == NULL);
328 	MPASS(!ia->exit);
329 	MPASS(ia->kthr != NULL);
330 	if (po == NULL) {
331 		ia->exit = true;
332 	} else {
333 		ia->po = po;
334 		KASSERT(po->po_kthread == NULL,
335 		    ("[pmclog,%d] po=%p kthread (%p) already present",
336 		    __LINE__, po, po->po_kthread));
337 		po->po_kthread = ia->kthr;
338 	}
339 	wakeup(ia);
340 	while (!ia->acted)
341 		msleep(ia, &pmc_kthread_mtx, PWAIT, "pmclogw", 0);
342 	mtx_unlock(&pmc_kthread_mtx);
343 	free(ia, M_TEMP);
344 }
345 
346 /*
347  * Log handler loop.
348  *
349  * This function is executed by each pmc owner's helper thread.
350  */
351 static void
352 pmclog_loop(void *arg)
353 {
354 	struct pmclog_proc_init_args *ia;
355 	struct pmc_owner *po;
356 	struct pmclog_buffer *lb;
357 	struct proc *p;
358 	struct ucred *ownercred;
359 	struct ucred *mycred;
360 	struct thread *td;
361 	sigset_t unb;
362 	struct uio auio;
363 	struct iovec aiov;
364 	size_t nbytes;
365 	int error;
366 
367 	td = curthread;
368 
369 	SIGEMPTYSET(unb);
370 	SIGADDSET(unb, SIGHUP);
371 	(void)kern_sigprocmask(td, SIG_UNBLOCK, &unb, NULL, 0);
372 
373 	ia = arg;
374 	MPASS(ia->kthr == curproc);
375 	MPASS(!ia->acted);
376 	mtx_lock(&pmc_kthread_mtx);
377 	while (ia->po == NULL && !ia->exit)
378 		msleep(ia, &pmc_kthread_mtx, PWAIT, "pmclogi", 0);
379 	if (ia->exit) {
380 		ia->acted = true;
381 		wakeup(ia);
382 		mtx_unlock(&pmc_kthread_mtx);
383 		kproc_exit(0);
384 	}
385 	MPASS(ia->po != NULL);
386 	po = ia->po;
387 	ia->acted = true;
388 	wakeup(ia);
389 	mtx_unlock(&pmc_kthread_mtx);
390 	ia = NULL;
391 
392 	p = po->po_owner;
393 	mycred = td->td_ucred;
394 
395 	PROC_LOCK(p);
396 	ownercred = crhold(p->p_ucred);
397 	PROC_UNLOCK(p);
398 
399 	PMCDBG2(LOG,INI,1, "po=%p kt=%p", po, po->po_kthread);
400 	KASSERT(po->po_kthread == curthread->td_proc,
401 	    ("[pmclog,%d] proc mismatch po=%p po/kt=%p curproc=%p", __LINE__,
402 		po, po->po_kthread, curthread->td_proc));
403 
404 	lb = NULL;
405 
406 
407 	/*
408 	 * Loop waiting for I/O requests to be added to the owner
409 	 * struct's queue.  The loop is exited when the log file
410 	 * is deconfigured.
411 	 */
412 
413 	mtx_lock(&pmc_kthread_mtx);
414 
415 	for (;;) {
416 
417 		/* check if we've been asked to exit */
418 		if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
419 			break;
420 
421 		if (lb == NULL) { /* look for a fresh buffer to write */
422 			mtx_lock_spin(&po->po_mtx);
423 			if ((lb = TAILQ_FIRST(&po->po_logbuffers)) == NULL) {
424 				mtx_unlock_spin(&po->po_mtx);
425 
426 				/* No more buffers and shutdown required. */
427 				if (po->po_flags & PMC_PO_SHUTDOWN)
428 					break;
429 
430 				(void) msleep(po, &pmc_kthread_mtx, PWAIT,
431 				    "pmcloop", 250);
432 				continue;
433 			}
434 
435 			TAILQ_REMOVE(&po->po_logbuffers, lb, plb_next);
436 			mtx_unlock_spin(&po->po_mtx);
437 		}
438 
439 		mtx_unlock(&pmc_kthread_mtx);
440 
441 		/* process the request */
442 		PMCDBG3(LOG,WRI,2, "po=%p base=%p ptr=%p", po,
443 		    lb->plb_base, lb->plb_ptr);
444 		/* change our thread's credentials before issuing the I/O */
445 
446 		aiov.iov_base = lb->plb_base;
447 		aiov.iov_len  = nbytes = lb->plb_ptr - lb->plb_base;
448 
449 		auio.uio_iov    = &aiov;
450 		auio.uio_iovcnt = 1;
451 		auio.uio_offset = -1;
452 		auio.uio_resid  = nbytes;
453 		auio.uio_rw     = UIO_WRITE;
454 		auio.uio_segflg = UIO_SYSSPACE;
455 		auio.uio_td     = td;
456 
457 		/* switch thread credentials -- see kern_ktrace.c */
458 		td->td_ucred = ownercred;
459 		error = fo_write(po->po_file, &auio, ownercred, 0, td);
460 		td->td_ucred = mycred;
461 
462 		if (error) {
463 			/* XXX some errors are recoverable */
464 			/* send a SIGIO to the owner and exit */
465 			PROC_LOCK(p);
466 			kern_psignal(p, SIGIO);
467 			PROC_UNLOCK(p);
468 
469 			mtx_lock(&pmc_kthread_mtx);
470 
471 			po->po_error = error; /* save for flush log */
472 
473 			PMCDBG2(LOG,WRI,2, "po=%p error=%d", po, error);
474 
475 			break;
476 		}
477 
478 		mtx_lock(&pmc_kthread_mtx);
479 
480 		/* put the used buffer back into the global pool */
481 		PMCLOG_RESET_BUFFER_DESCRIPTOR(lb);
482 
483 		pmc_plb_rele(lb);
484 		lb = NULL;
485 	}
486 
487 	wakeup_one(po->po_kthread);
488 	po->po_kthread = NULL;
489 
490 	mtx_unlock(&pmc_kthread_mtx);
491 
492 	/* return the current I/O buffer to the global pool */
493 	if (lb) {
494 		PMCLOG_RESET_BUFFER_DESCRIPTOR(lb);
495 
496 		pmc_plb_rele(lb);
497 	}
498 
499 	/*
500 	 * Exit this thread, signalling the waiter
501 	 */
502 
503 	crfree(ownercred);
504 
505 	kproc_exit(0);
506 }
507 
508 /*
509  * Release and log entry and schedule an I/O if needed.
510  */
511 
512 static void
513 pmclog_release_flags(struct pmc_owner *po, int wakeup)
514 {
515 	struct pmclog_buffer *plb;
516 
517 	plb = po->po_curbuf[curcpu];
518 	KASSERT(plb->plb_ptr >= plb->plb_base,
519 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p base=%p", __LINE__,
520 		po, plb->plb_ptr, plb->plb_base));
521 	KASSERT(plb->plb_ptr <= plb->plb_fence,
522 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p fenc=%p", __LINE__,
523 		po, plb->plb_ptr, plb->plb_fence));
524 
525 	/* schedule an I/O if we've filled a buffer */
526 	if (plb->plb_ptr >= plb->plb_fence)
527 		pmclog_schedule_io(po, wakeup);
528 
529 	PMCDBG1(LOG,REL,1, "po=%p", po);
530 }
531 
532 static void
533 pmclog_release(struct pmc_owner *po)
534 {
535 
536 	pmclog_release_flags(po, 1);
537 }
538 
539 
540 /*
541  * Attempt to reserve 'length' bytes of space in an owner's log
542  * buffer.  The function returns a pointer to 'length' bytes of space
543  * if there was enough space or returns NULL if no space was
544  * available.  Non-null returns do so with the po mutex locked.  The
545  * caller must invoke pmclog_release() on the pmc owner structure
546  * when done.
547  */
548 
549 static uint32_t *
550 pmclog_reserve(struct pmc_owner *po, int length)
551 {
552 	uintptr_t newptr, oldptr;
553 	struct pmclog_buffer *plb, **pplb;
554 
555 	PMCDBG2(LOG,ALL,1, "po=%p len=%d", po, length);
556 
557 	KASSERT(length % sizeof(uint32_t) == 0,
558 	    ("[pmclog,%d] length not a multiple of word size", __LINE__));
559 
560 	/* No more data when shutdown in progress. */
561 	if (po->po_flags & PMC_PO_SHUTDOWN)
562 		return (NULL);
563 
564 	pplb = &po->po_curbuf[curcpu];
565 	if (*pplb == NULL && pmclog_get_buffer(po) != 0)
566 		goto fail;
567 
568 	KASSERT(*pplb != NULL,
569 	    ("[pmclog,%d] po=%p no current buffer", __LINE__, po));
570 
571 	plb = *pplb;
572 	KASSERT(plb->plb_ptr >= plb->plb_base &&
573 	    plb->plb_ptr <= plb->plb_fence,
574 	    ("[pmclog,%d] po=%p buffer invariants: ptr=%p base=%p fence=%p",
575 		__LINE__, po, plb->plb_ptr, plb->plb_base,
576 		plb->plb_fence));
577 
578 	oldptr = (uintptr_t) plb->plb_ptr;
579 	newptr = oldptr + length;
580 
581 	KASSERT(oldptr != (uintptr_t) NULL,
582 	    ("[pmclog,%d] po=%p Null log buffer pointer", __LINE__, po));
583 
584 	/*
585 	 * If we have space in the current buffer, return a pointer to
586 	 * available space with the PO structure locked.
587 	 */
588 	if (newptr <= (uintptr_t) plb->plb_fence) {
589 		plb->plb_ptr = (char *) newptr;
590 		goto done;
591 	}
592 
593 	/*
594 	 * Otherwise, schedule the current buffer for output and get a
595 	 * fresh buffer.
596 	 */
597 	pmclog_schedule_io(po, 0);
598 
599 	if (pmclog_get_buffer(po) != 0)
600 		goto fail;
601 
602 	plb = *pplb;
603 	KASSERT(plb != NULL,
604 	    ("[pmclog,%d] po=%p no current buffer", __LINE__, po));
605 
606 	KASSERT(plb->plb_ptr != NULL,
607 	    ("[pmclog,%d] null return from pmc_get_log_buffer", __LINE__));
608 
609 	KASSERT(plb->plb_ptr == plb->plb_base &&
610 	    plb->plb_ptr <= plb->plb_fence,
611 	    ("[pmclog,%d] po=%p buffer invariants: ptr=%p base=%p fence=%p",
612 		__LINE__, po, plb->plb_ptr, plb->plb_base,
613 		plb->plb_fence));
614 
615 	oldptr = (uintptr_t) plb->plb_ptr;
616 
617  done:
618 	return ((uint32_t *) oldptr);
619  fail:
620 	return (NULL);
621 }
622 
623 /*
624  * Schedule an I/O.
625  *
626  * Transfer the current buffer to the helper kthread.
627  */
628 
629 static void
630 pmclog_schedule_io(struct pmc_owner *po, int wakeup)
631 {
632 	struct pmclog_buffer *plb;
633 
634 	plb = po->po_curbuf[curcpu];
635 	po->po_curbuf[curcpu] = NULL;
636 	KASSERT(plb != NULL,
637 	    ("[pmclog,%d] schedule_io with null buffer po=%p", __LINE__, po));
638 	KASSERT(plb->plb_ptr >= plb->plb_base,
639 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p base=%p", __LINE__,
640 		po, plb->plb_ptr, plb->plb_base));
641 	KASSERT(plb->plb_ptr <= plb->plb_fence,
642 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p fenc=%p", __LINE__,
643 		po, plb->plb_ptr, plb->plb_fence));
644 
645 	PMCDBG1(LOG,SIO, 1, "po=%p", po);
646 
647 	/*
648 	 * Add the current buffer to the tail of the buffer list and
649 	 * wakeup the helper.
650 	 */
651 	mtx_lock_spin(&po->po_mtx);
652 	TAILQ_INSERT_TAIL(&po->po_logbuffers, plb, plb_next);
653 	mtx_unlock_spin(&po->po_mtx);
654 	if (wakeup)
655 		wakeup_one(po);
656 }
657 
658 /*
659  * Stop the helper kthread.
660  */
661 
662 static void
663 pmclog_stop_kthread(struct pmc_owner *po)
664 {
665 
666 	mtx_lock(&pmc_kthread_mtx);
667 	po->po_flags &= ~PMC_PO_OWNS_LOGFILE;
668 	if (po->po_kthread != NULL) {
669 		PROC_LOCK(po->po_kthread);
670 		kern_psignal(po->po_kthread, SIGHUP);
671 		PROC_UNLOCK(po->po_kthread);
672 	}
673 	wakeup_one(po);
674 	while (po->po_kthread)
675 		msleep(po->po_kthread, &pmc_kthread_mtx, PPAUSE, "pmckstp", 0);
676 	mtx_unlock(&pmc_kthread_mtx);
677 }
678 
679 /*
680  * Public functions
681  */
682 
683 /*
684  * Configure a log file for pmc owner 'po'.
685  *
686  * Parameter 'logfd' is a file handle referencing an open file in the
687  * owner process.  This file needs to have been opened for writing.
688  */
689 
690 int
691 pmclog_configure_log(struct pmc_mdep *md, struct pmc_owner *po, int logfd)
692 {
693 	struct proc *p;
694 	struct timespec ts;
695 	uint64_t tsc;
696 	int error;
697 
698 	sx_assert(&pmc_sx, SA_XLOCKED);
699 	PMCDBG2(LOG,CFG,1, "config po=%p logfd=%d", po, logfd);
700 
701 	p = po->po_owner;
702 
703 	/* return EBUSY if a log file was already present */
704 	if (po->po_flags & PMC_PO_OWNS_LOGFILE)
705 		return (EBUSY);
706 
707 	KASSERT(po->po_file == NULL,
708 	    ("[pmclog,%d] po=%p file (%p) already present", __LINE__, po,
709 		po->po_file));
710 
711 	/* get a reference to the file state */
712 	error = fget_write(curthread, logfd, &cap_write_rights, &po->po_file);
713 	if (error)
714 		goto error;
715 
716 	/* mark process as owning a log file */
717 	po->po_flags |= PMC_PO_OWNS_LOGFILE;
718 
719 	/* mark process as using HWPMCs */
720 	PROC_LOCK(p);
721 	p->p_flag |= P_HWPMC;
722 	PROC_UNLOCK(p);
723 	nanotime(&ts);
724 	tsc = pmc_rdtsc();
725 	/* create a log initialization entry */
726 	PMCLOG_RESERVE_WITH_ERROR(po, INITIALIZE,
727 	    sizeof(struct pmclog_initialize));
728 	PMCLOG_EMIT32(PMC_VERSION);
729 	PMCLOG_EMIT32(md->pmd_cputype);
730 #if defined(__i386__) || defined(__amd64__)
731 	PMCLOG_EMIT64(tsc_freq);
732 #else
733 	/* other architectures will need to fill this in */
734 	PMCLOG_EMIT32(0);
735 	PMCLOG_EMIT32(0);
736 #endif
737 	memcpy(_le, &ts, sizeof(ts));
738 	_le += sizeof(ts)/4;
739 	PMCLOG_EMITSTRING(pmc_cpuid, PMC_CPUID_LEN);
740 	PMCLOG_DESPATCH_SYNC(po);
741 
742 	return (0);
743 
744  error:
745 	KASSERT(po->po_kthread == NULL, ("[pmclog,%d] po=%p kthread not "
746 	    "stopped", __LINE__, po));
747 
748 	if (po->po_file)
749 		(void) fdrop(po->po_file, curthread);
750 	po->po_file  = NULL;	/* clear file and error state */
751 	po->po_error = 0;
752 	po->po_flags &= ~PMC_PO_OWNS_LOGFILE;
753 
754 	return (error);
755 }
756 
757 
758 /*
759  * De-configure a log file.  This will throw away any buffers queued
760  * for this owner process.
761  */
762 
763 int
764 pmclog_deconfigure_log(struct pmc_owner *po)
765 {
766 	int error;
767 	struct pmclog_buffer *lb;
768 
769 	PMCDBG1(LOG,CFG,1, "de-config po=%p", po);
770 
771 	if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
772 		return (EINVAL);
773 
774 	KASSERT(po->po_sscount == 0,
775 	    ("[pmclog,%d] po=%p still owning SS PMCs", __LINE__, po));
776 	KASSERT(po->po_file != NULL,
777 	    ("[pmclog,%d] po=%p no log file", __LINE__, po));
778 
779 	/* stop the kthread, this will reset the 'OWNS_LOGFILE' flag */
780 	pmclog_stop_kthread(po);
781 
782 	KASSERT(po->po_kthread == NULL,
783 	    ("[pmclog,%d] po=%p kthread not stopped", __LINE__, po));
784 
785 	/* return all queued log buffers to the global pool */
786 	while ((lb = TAILQ_FIRST(&po->po_logbuffers)) != NULL) {
787 		TAILQ_REMOVE(&po->po_logbuffers, lb, plb_next);
788 		PMCLOG_RESET_BUFFER_DESCRIPTOR(lb);
789 		pmc_plb_rele(lb);
790 	}
791 	for (int i = 0; i < mp_ncpus; i++) {
792 		thread_lock(curthread);
793 		sched_bind(curthread, i);
794 		thread_unlock(curthread);
795 		/* return the 'current' buffer to the global pool */
796 		if ((lb = po->po_curbuf[curcpu]) != NULL) {
797 			PMCLOG_RESET_BUFFER_DESCRIPTOR(lb);
798 			pmc_plb_rele(lb);
799 		}
800 	}
801 	thread_lock(curthread);
802 	sched_unbind(curthread);
803 	thread_unlock(curthread);
804 
805 	/* drop a reference to the fd */
806 	if (po->po_file != NULL) {
807 		error = fdrop(po->po_file, curthread);
808 		po->po_file = NULL;
809 	} else
810 		error = 0;
811 	po->po_error = 0;
812 
813 	return (error);
814 }
815 
816 /*
817  * Flush a process' log buffer.
818  */
819 
820 int
821 pmclog_flush(struct pmc_owner *po, int force)
822 {
823 	int error;
824 
825 	PMCDBG1(LOG,FLS,1, "po=%p", po);
826 
827 	/*
828 	 * If there is a pending error recorded by the logger thread,
829 	 * return that.
830 	 */
831 	if (po->po_error)
832 		return (po->po_error);
833 
834 	error = 0;
835 
836 	/*
837 	 * Check that we do have an active log file.
838 	 */
839 	mtx_lock(&pmc_kthread_mtx);
840 	if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) {
841 		error = EINVAL;
842 		goto error;
843 	}
844 
845 	pmclog_schedule_all(po, force);
846  error:
847 	mtx_unlock(&pmc_kthread_mtx);
848 
849 	return (error);
850 }
851 
852 static void
853 pmclog_schedule_one_cond(struct pmc_owner *po, int force)
854 {
855 	struct pmclog_buffer *plb;
856 	int cpu;
857 
858 	spinlock_enter();
859 	cpu = curcpu;
860 	/* tell hardclock not to run again */
861 	if (PMC_CPU_HAS_SAMPLES(cpu))
862 		PMC_CALL_HOOK_UNLOCKED(curthread, PMC_FN_DO_SAMPLES, NULL);
863 	if (force)
864 		pmc_flush_samples(cpu);
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, int force)
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, force);
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, 0);
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 	int domain;
1236 	struct pmclog_buffer *plb;
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 		int ncpus = pmc_dom_hdrs[domain]->pdbh_ncpus;
1258 		int total = ncpus*pmc_nlogbuffers_pcpu;
1259 
1260 		plb = malloc_domain(sizeof(struct pmclog_buffer)*total, M_PMC, domain, M_WAITOK|M_ZERO);
1261 		pmc_dom_hdrs[domain]->pdbh_plbs = plb;
1262 		for (int i = 0; i < total; i++, plb++) {
1263 			void *buf;
1264 
1265 			buf = malloc_domain(1024 * pmclog_buffer_size, M_PMC, domain,
1266 								M_WAITOK|M_ZERO);
1267 			PMCLOG_INIT_BUFFER_DESCRIPTOR(plb, buf, domain);
1268 			pmc_plb_rele_unlocked(plb);
1269 		}
1270 	}
1271 	mtx_init(&pmc_kthread_mtx, "pmc-kthread", "pmc-sleep", MTX_DEF);
1272 }
1273 
1274 /*
1275  * Shutdown logging.
1276  *
1277  * Destroy mutexes and release memory back the to free pool.
1278  */
1279 
1280 void
1281 pmclog_shutdown()
1282 {
1283 	struct pmclog_buffer *plb;
1284 	int domain;
1285 
1286 	mtx_destroy(&pmc_kthread_mtx);
1287 
1288 	for (domain = 0; domain < vm_ndomains; domain++) {
1289 		while ((plb = TAILQ_FIRST(&pmc_dom_hdrs[domain]->pdbh_head)) != NULL) {
1290 			TAILQ_REMOVE(&pmc_dom_hdrs[domain]->pdbh_head, plb, plb_next);
1291 			free(plb->plb_base, M_PMC);
1292 		}
1293 		free(pmc_dom_hdrs[domain]->pdbh_plbs, M_PMC);
1294 	}
1295 }
1296