xref: /freebsd/sys/dev/hwpmc/hwpmc_logging.c (revision b1879975794772ee51f0b4865753364c7d7626c3)
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/param.h>
40 #include <sys/capsicum.h>
41 #include <sys/domainset.h>
42 #include <sys/file.h>
43 #include <sys/kernel.h>
44 #include <sys/kthread.h>
45 #include <sys/lock.h>
46 #include <sys/module.h>
47 #include <sys/mutex.h>
48 #include <sys/pmc.h>
49 #include <sys/pmckern.h>
50 #include <sys/pmclog.h>
51 #include <sys/proc.h>
52 #include <sys/sched.h>
53 #include <sys/signalvar.h>
54 #include <sys/smp.h>
55 #include <sys/syscallsubr.h>
56 #include <sys/sysctl.h>
57 #include <sys/systm.h>
58 #include <sys/uio.h>
59 #include <sys/unistd.h>
60 #include <sys/vnode.h>
61 
62 #if defined(__i386__) || defined(__amd64__)
63 #include <machine/clock.h>
64 #endif
65 
66 #define curdomain PCPU_GET(domain)
67 
68 /*
69  * Sysctl tunables
70  */
71 
72 SYSCTL_DECL(_kern_hwpmc);
73 
74 /*
75  * kern.hwpmc.logbuffersize -- size of the per-cpu owner buffers.
76  */
77 
78 static int pmclog_buffer_size = PMC_LOG_BUFFER_SIZE;
79 SYSCTL_INT(_kern_hwpmc, OID_AUTO, logbuffersize, CTLFLAG_RDTUN,
80     &pmclog_buffer_size, 0, "size of log buffers in kilobytes");
81 
82 /*
83  * kern.hwpmc.nbuffers_pcpu -- number of global log buffers
84  */
85 
86 static int pmc_nlogbuffers_pcpu = PMC_NLOGBUFFERS_PCPU;
87 SYSCTL_INT(_kern_hwpmc, OID_AUTO, nbuffers_pcpu, CTLFLAG_RDTUN,
88     &pmc_nlogbuffers_pcpu, 0, "number of log buffers per cpu");
89 
90 /*
91  * Global log buffer list and associated spin lock.
92  */
93 
94 static struct mtx pmc_kthread_mtx;	/* sleep lock */
95 
96 #define	PMCLOG_INIT_BUFFER_DESCRIPTOR(D, buf, domain) do {		\
97 	(D)->plb_fence = ((char *)(buf)) + 1024 * pmclog_buffer_size;	\
98 	(D)->plb_base = (D)->plb_ptr = ((char *)(buf));			\
99 	(D)->plb_domain = domain;					\
100 } while (0)
101 
102 #define	PMCLOG_RESET_BUFFER_DESCRIPTOR(D) do {				\
103 	(D)->plb_ptr = (D)->plb_base;					\
104 } while (0)
105 
106 /*
107  * Log file record constructors.
108  */
109 #define	_PMCLOG_TO_HEADER(T, L)						\
110 	((PMCLOG_HEADER_MAGIC << 24) | (T << 16) | ((L) & 0xFFFF))
111 
112 /* reserve LEN bytes of space and initialize the entry header */
113 #define	_PMCLOG_RESERVE_SAFE(PO, TYPE, LEN, ACTION, TSC) do {		\
114 	uint32_t *_le;							\
115 	int _len = roundup((LEN), sizeof(uint32_t));			\
116 	struct pmclog_header *ph;					\
117 									\
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 									\
133 	tsc = pmc_rdtsc();						\
134 	spinlock_enter();						\
135 	if ((_le = pmclog_reserve((PO), _len)) == NULL) {		\
136 		spinlock_exit();					\
137 		ACTION;							\
138 	}								\
139 	ph = (struct pmclog_header *)_le;				\
140 	ph->pl_header =_PMCLOG_TO_HEADER(TYPE,_len);			\
141 	ph->pl_tsc = tsc;						\
142 	_le += sizeof(*ph) / 4	/* skip over timestamp */
143 
144 #define	PMCLOG_RESERVE_SAFE(P, T, L, TSC)				\
145 	_PMCLOG_RESERVE_SAFE(P, T, L, return, TSC)
146 #define	PMCLOG_RESERVE(P,T,L)						\
147 	_PMCLOG_RESERVE(P, T, L, return)
148 #define	PMCLOG_RESERVE_WITH_ERROR(P, T, L)				\
149 	_PMCLOG_RESERVE(P, T, L, error = ENOMEM; goto error)
150 
151 #define	PMCLOG_EMIT32(V)	do { *_le++ = (V); } while (0)
152 #define	PMCLOG_EMIT64(V)	do { 					\
153 	*_le++ = (uint32_t) ((V) & 0xFFFFFFFF);				\
154 	*_le++ = (uint32_t) (((V) >> 32) & 0xFFFFFFFF);			\
155 } while (0)
156 
157 
158 /* Emit a string.  Caution: does NOT update _le, so needs to be last */
159 #define	PMCLOG_EMITSTRING(S,L)	do {					\
160 	bcopy((S), _le, (L));						\
161 } while (0)
162 #define	PMCLOG_EMITNULLSTRING(L) do {					\
163 	bzero(_le, (L));						\
164 } while (0)
165 
166 #define	PMCLOG_DESPATCH_SAFE(PO)					\
167 	pmclog_release((PO));						\
168 } while (0)
169 
170 #define	PMCLOG_DESPATCH_SCHED_LOCK(PO)					\
171 	pmclog_release_flags((PO), 0);					\
172 } while (0)
173 
174 #define	PMCLOG_DESPATCH(PO)						\
175 	pmclog_release((PO));						\
176 	spinlock_exit();						\
177 } while (0)
178 
179 #define	PMCLOG_DESPATCH_SYNC(PO)					\
180 	pmclog_schedule_io((PO), 1);					\
181 	spinlock_exit();						\
182 } while (0)
183 
184 #define TSDELTA 4
185 /*
186  * Assertions about the log file format.
187  */
188 CTASSERT(sizeof(struct pmclog_callchain) == 7*4 + TSDELTA +
189     PMC_CALLCHAIN_DEPTH_MAX*sizeof(uintfptr_t));
190 CTASSERT(sizeof(struct pmclog_closelog) == 3*4 + TSDELTA);
191 CTASSERT(sizeof(struct pmclog_dropnotify) == 3*4 + TSDELTA);
192 CTASSERT(sizeof(struct pmclog_map_in) == PATH_MAX + TSDELTA +
193     5*4 + sizeof(uintfptr_t));
194 CTASSERT(offsetof(struct pmclog_map_in,pl_pathname) ==
195     5*4 + TSDELTA + sizeof(uintfptr_t));
196 CTASSERT(sizeof(struct pmclog_map_out) == 5*4 + 2*sizeof(uintfptr_t) + TSDELTA);
197 CTASSERT(sizeof(struct pmclog_pmcallocate) == 9*4 + TSDELTA);
198 CTASSERT(sizeof(struct pmclog_pmcattach) == 5*4 + PATH_MAX + TSDELTA);
199 CTASSERT(offsetof(struct pmclog_pmcattach,pl_pathname) == 5*4 + TSDELTA);
200 CTASSERT(sizeof(struct pmclog_pmcdetach) == 5*4 + TSDELTA);
201 CTASSERT(sizeof(struct pmclog_proccsw) == 7*4 + 8 + TSDELTA);
202 CTASSERT(sizeof(struct pmclog_procexec) == 5*4 + PATH_MAX +
203     2*sizeof(uintptr_t) + TSDELTA);
204 CTASSERT(offsetof(struct pmclog_procexec,pl_pathname) == 5*4 + TSDELTA +
205     2*sizeof(uintptr_t));
206 CTASSERT(sizeof(struct pmclog_procexit) == 5*4 + 8 + TSDELTA);
207 CTASSERT(sizeof(struct pmclog_procfork) == 5*4 + TSDELTA);
208 CTASSERT(sizeof(struct pmclog_sysexit) == 6*4);
209 CTASSERT(sizeof(struct pmclog_userdata) == 6*4);
210 
211 /*
212  * Log buffer structure
213  */
214 
215 struct pmclog_buffer {
216 	TAILQ_ENTRY(pmclog_buffer) plb_next;
217 	char 		*plb_base;
218 	char		*plb_ptr;
219 	char 		*plb_fence;
220 	uint16_t	 plb_domain;
221 } __aligned(CACHE_LINE_SIZE);
222 
223 /*
224  * Prototypes
225  */
226 
227 static int pmclog_get_buffer(struct pmc_owner *po);
228 static void pmclog_loop(void *arg);
229 static void pmclog_release(struct pmc_owner *po);
230 static uint32_t *pmclog_reserve(struct pmc_owner *po, int length);
231 static void pmclog_schedule_io(struct pmc_owner *po, int wakeup);
232 static void pmclog_schedule_all(struct pmc_owner *po);
233 static void pmclog_stop_kthread(struct pmc_owner *po);
234 
235 /*
236  * Helper functions
237  */
238 
239 static inline void
240 pmc_plb_rele_unlocked(struct pmclog_buffer *plb)
241 {
242 	TAILQ_INSERT_HEAD(&pmc_dom_hdrs[plb->plb_domain]->pdbh_head, plb, plb_next);
243 }
244 
245 static inline void
246 pmc_plb_rele(struct pmclog_buffer *plb)
247 {
248 	mtx_lock_spin(&pmc_dom_hdrs[plb->plb_domain]->pdbh_mtx);
249 	pmc_plb_rele_unlocked(plb);
250 	mtx_unlock_spin(&pmc_dom_hdrs[plb->plb_domain]->pdbh_mtx);
251 }
252 
253 /*
254  * Get a log buffer
255  */
256 static int
257 pmclog_get_buffer(struct pmc_owner *po)
258 {
259 	struct pmclog_buffer *plb;
260 	int domain;
261 
262 	KASSERT(po->po_curbuf[curcpu] == NULL,
263 	    ("[pmclog,%d] po=%p current buffer still valid", __LINE__, po));
264 
265 	domain = curdomain;
266 	MPASS(pmc_dom_hdrs[domain]);
267 	mtx_lock_spin(&pmc_dom_hdrs[domain]->pdbh_mtx);
268 	if ((plb = TAILQ_FIRST(&pmc_dom_hdrs[domain]->pdbh_head)) != NULL)
269 		TAILQ_REMOVE(&pmc_dom_hdrs[domain]->pdbh_head, plb, plb_next);
270 	mtx_unlock_spin(&pmc_dom_hdrs[domain]->pdbh_mtx);
271 
272 	PMCDBG2(LOG,GTB,1, "po=%p plb=%p", po, plb);
273 
274 #ifdef	HWPMC_DEBUG
275 	if (plb)
276 		KASSERT(plb->plb_ptr == plb->plb_base &&
277 		    plb->plb_base < plb->plb_fence,
278 		    ("[pmclog,%d] po=%p buffer invariants: ptr=%p "
279 		    "base=%p fence=%p", __LINE__, po, plb->plb_ptr,
280 		    plb->plb_base, plb->plb_fence));
281 #endif
282 
283 	po->po_curbuf[curcpu] = plb;
284 
285 	/* update stats */
286 	counter_u64_add(pmc_stats.pm_buffer_requests, 1);
287 	if (plb == NULL)
288 		counter_u64_add(pmc_stats.pm_buffer_requests_failed, 1);
289 
290 	return (plb ? 0 : ENOMEM);
291 }
292 
293 struct pmclog_proc_init_args {
294 	struct proc *kthr;
295 	struct pmc_owner *po;
296 	bool exit;
297 	bool acted;
298 };
299 
300 int
301 pmclog_proc_create(struct thread *td, void **handlep)
302 {
303 	struct pmclog_proc_init_args *ia;
304 	int error;
305 
306 	ia = malloc(sizeof(*ia), M_TEMP, M_WAITOK | M_ZERO);
307 	error = kproc_create(pmclog_loop, ia, &ia->kthr,
308 	    RFHIGHPID, 0, "hwpmc: proc(%d)", td->td_proc->p_pid);
309 	if (error == 0)
310 		*handlep = ia;
311 	return (error);
312 }
313 
314 void
315 pmclog_proc_ignite(void *handle, struct pmc_owner *po)
316 {
317 	struct pmclog_proc_init_args *ia;
318 
319 	ia = handle;
320 	mtx_lock(&pmc_kthread_mtx);
321 	MPASS(!ia->acted);
322 	MPASS(ia->po == NULL);
323 	MPASS(!ia->exit);
324 	MPASS(ia->kthr != NULL);
325 	if (po == NULL) {
326 		ia->exit = true;
327 	} else {
328 		ia->po = po;
329 		KASSERT(po->po_kthread == NULL,
330 		    ("[pmclog,%d] po=%p kthread (%p) already present",
331 		    __LINE__, po, po->po_kthread));
332 		po->po_kthread = ia->kthr;
333 	}
334 	wakeup(ia);
335 	while (!ia->acted)
336 		msleep(ia, &pmc_kthread_mtx, PWAIT, "pmclogw", 0);
337 	mtx_unlock(&pmc_kthread_mtx);
338 	free(ia, M_TEMP);
339 }
340 
341 /*
342  * Log handler loop.
343  *
344  * This function is executed by each pmc owner's helper thread.
345  */
346 static void
347 pmclog_loop(void *arg)
348 {
349 	struct pmclog_proc_init_args *ia;
350 	struct pmc_owner *po;
351 	struct pmclog_buffer *lb;
352 	struct proc *p;
353 	struct ucred *ownercred;
354 	struct ucred *mycred;
355 	struct thread *td;
356 	sigset_t unb;
357 	struct uio auio;
358 	struct iovec aiov;
359 	size_t nbytes;
360 	int error;
361 
362 	td = curthread;
363 
364 	SIGEMPTYSET(unb);
365 	SIGADDSET(unb, SIGHUP);
366 	(void)kern_sigprocmask(td, SIG_UNBLOCK, &unb, NULL, 0);
367 
368 	ia = arg;
369 	MPASS(ia->kthr == curproc);
370 	MPASS(!ia->acted);
371 	mtx_lock(&pmc_kthread_mtx);
372 	while (ia->po == NULL && !ia->exit)
373 		msleep(ia, &pmc_kthread_mtx, PWAIT, "pmclogi", 0);
374 	if (ia->exit) {
375 		ia->acted = true;
376 		wakeup(ia);
377 		mtx_unlock(&pmc_kthread_mtx);
378 		kproc_exit(0);
379 	}
380 	MPASS(ia->po != NULL);
381 	po = ia->po;
382 	ia->acted = true;
383 	wakeup(ia);
384 	mtx_unlock(&pmc_kthread_mtx);
385 	ia = NULL;
386 
387 	p = po->po_owner;
388 	mycred = td->td_ucred;
389 
390 	PROC_LOCK(p);
391 	ownercred = crhold(p->p_ucred);
392 	PROC_UNLOCK(p);
393 
394 	PMCDBG2(LOG,INI,1, "po=%p kt=%p", po, po->po_kthread);
395 	KASSERT(po->po_kthread == curthread->td_proc,
396 	    ("[pmclog,%d] proc mismatch po=%p po/kt=%p curproc=%p", __LINE__,
397 		po, po->po_kthread, curthread->td_proc));
398 
399 	lb = NULL;
400 
401 
402 	/*
403 	 * Loop waiting for I/O requests to be added to the owner
404 	 * struct's queue.  The loop is exited when the log file
405 	 * is deconfigured.
406 	 */
407 
408 	mtx_lock(&pmc_kthread_mtx);
409 
410 	for (;;) {
411 
412 		/* check if we've been asked to exit */
413 		if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
414 			break;
415 
416 		if (lb == NULL) { /* look for a fresh buffer to write */
417 			mtx_lock_spin(&po->po_mtx);
418 			if ((lb = TAILQ_FIRST(&po->po_logbuffers)) == NULL) {
419 				mtx_unlock_spin(&po->po_mtx);
420 
421 				/* No more buffers and shutdown required. */
422 				if (po->po_flags & PMC_PO_SHUTDOWN)
423 					break;
424 
425 				(void) msleep(po, &pmc_kthread_mtx, PWAIT,
426 				    "pmcloop", 250);
427 				continue;
428 			}
429 
430 			TAILQ_REMOVE(&po->po_logbuffers, lb, plb_next);
431 			mtx_unlock_spin(&po->po_mtx);
432 		}
433 
434 		mtx_unlock(&pmc_kthread_mtx);
435 
436 		/* process the request */
437 		PMCDBG3(LOG,WRI,2, "po=%p base=%p ptr=%p", po,
438 		    lb->plb_base, lb->plb_ptr);
439 		/* change our thread's credentials before issuing the I/O */
440 
441 		aiov.iov_base = lb->plb_base;
442 		aiov.iov_len  = nbytes = lb->plb_ptr - lb->plb_base;
443 
444 		auio.uio_iov    = &aiov;
445 		auio.uio_iovcnt = 1;
446 		auio.uio_offset = -1;
447 		auio.uio_resid  = nbytes;
448 		auio.uio_rw     = UIO_WRITE;
449 		auio.uio_segflg = UIO_SYSSPACE;
450 		auio.uio_td     = td;
451 
452 		/* switch thread credentials -- see kern_ktrace.c */
453 		td->td_ucred = ownercred;
454 		error = fo_write(po->po_file, &auio, ownercred, 0, td);
455 		td->td_ucred = mycred;
456 
457 		if (error) {
458 			/* XXX some errors are recoverable */
459 			/* send a SIGIO to the owner and exit */
460 			PROC_LOCK(p);
461 			kern_psignal(p, SIGIO);
462 			PROC_UNLOCK(p);
463 
464 			mtx_lock(&pmc_kthread_mtx);
465 
466 			po->po_error = error; /* save for flush log */
467 
468 			PMCDBG2(LOG,WRI,2, "po=%p error=%d", po, error);
469 
470 			break;
471 		}
472 
473 		mtx_lock(&pmc_kthread_mtx);
474 
475 		/* put the used buffer back into the global pool */
476 		PMCLOG_RESET_BUFFER_DESCRIPTOR(lb);
477 
478 		pmc_plb_rele(lb);
479 		lb = NULL;
480 	}
481 
482 	wakeup_one(po->po_kthread);
483 	po->po_kthread = NULL;
484 
485 	mtx_unlock(&pmc_kthread_mtx);
486 
487 	/* return the current I/O buffer to the global pool */
488 	if (lb) {
489 		PMCLOG_RESET_BUFFER_DESCRIPTOR(lb);
490 
491 		pmc_plb_rele(lb);
492 	}
493 
494 	/*
495 	 * Exit this thread, signalling the waiter
496 	 */
497 
498 	crfree(ownercred);
499 
500 	kproc_exit(0);
501 }
502 
503 /*
504  * Release and log entry and schedule an I/O if needed.
505  */
506 
507 static void
508 pmclog_release_flags(struct pmc_owner *po, int wakeup)
509 {
510 	struct pmclog_buffer *plb;
511 
512 	plb = po->po_curbuf[curcpu];
513 	KASSERT(plb->plb_ptr >= plb->plb_base,
514 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p base=%p", __LINE__,
515 		po, plb->plb_ptr, plb->plb_base));
516 	KASSERT(plb->plb_ptr <= plb->plb_fence,
517 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p fenc=%p", __LINE__,
518 		po, plb->plb_ptr, plb->plb_fence));
519 
520 	/* schedule an I/O if we've filled a buffer */
521 	if (plb->plb_ptr >= plb->plb_fence)
522 		pmclog_schedule_io(po, wakeup);
523 
524 	PMCDBG1(LOG,REL,1, "po=%p", po);
525 }
526 
527 static void
528 pmclog_release(struct pmc_owner *po)
529 {
530 
531 	pmclog_release_flags(po, 1);
532 }
533 
534 
535 /*
536  * Attempt to reserve 'length' bytes of space in an owner's log
537  * buffer.  The function returns a pointer to 'length' bytes of space
538  * if there was enough space or returns NULL if no space was
539  * available.  Non-null returns do so with the po mutex locked.  The
540  * caller must invoke pmclog_release() on the pmc owner structure
541  * when done.
542  */
543 
544 static uint32_t *
545 pmclog_reserve(struct pmc_owner *po, int length)
546 {
547 	uintptr_t newptr, oldptr __diagused;
548 	struct pmclog_buffer *plb, **pplb;
549 
550 	PMCDBG2(LOG,ALL,1, "po=%p len=%d", po, length);
551 
552 	KASSERT(length % sizeof(uint32_t) == 0,
553 	    ("[pmclog,%d] length not a multiple of word size", __LINE__));
554 
555 	/* No more data when shutdown in progress. */
556 	if (po->po_flags & PMC_PO_SHUTDOWN)
557 		return (NULL);
558 
559 	pplb = &po->po_curbuf[curcpu];
560 	if (*pplb == NULL && pmclog_get_buffer(po) != 0)
561 		goto fail;
562 
563 	KASSERT(*pplb != NULL,
564 	    ("[pmclog,%d] po=%p no current buffer", __LINE__, po));
565 
566 	plb = *pplb;
567 	KASSERT(plb->plb_ptr >= plb->plb_base &&
568 	    plb->plb_ptr <= plb->plb_fence,
569 	    ("[pmclog,%d] po=%p buffer invariants: ptr=%p base=%p fence=%p",
570 		__LINE__, po, plb->plb_ptr, plb->plb_base,
571 		plb->plb_fence));
572 
573 	oldptr = (uintptr_t) plb->plb_ptr;
574 	newptr = oldptr + length;
575 
576 	KASSERT(oldptr != (uintptr_t) NULL,
577 	    ("[pmclog,%d] po=%p Null log buffer pointer", __LINE__, po));
578 
579 	/*
580 	 * If we have space in the current buffer, return a pointer to
581 	 * available space with the PO structure locked.
582 	 */
583 	if (newptr <= (uintptr_t) plb->plb_fence) {
584 		plb->plb_ptr = (char *) newptr;
585 		goto done;
586 	}
587 
588 	/*
589 	 * Otherwise, schedule the current buffer for output and get a
590 	 * fresh buffer.
591 	 */
592 	pmclog_schedule_io(po, 0);
593 
594 	if (pmclog_get_buffer(po) != 0)
595 		goto fail;
596 
597 	plb = *pplb;
598 	KASSERT(plb != NULL,
599 	    ("[pmclog,%d] po=%p no current buffer", __LINE__, po));
600 
601 	KASSERT(plb->plb_ptr != NULL,
602 	    ("[pmclog,%d] null return from pmc_get_log_buffer", __LINE__));
603 
604 	KASSERT(plb->plb_ptr == plb->plb_base &&
605 	    plb->plb_ptr <= plb->plb_fence,
606 	    ("[pmclog,%d] po=%p buffer invariants: ptr=%p base=%p fence=%p",
607 		__LINE__, po, plb->plb_ptr, plb->plb_base,
608 		plb->plb_fence));
609 
610 	oldptr = (uintptr_t) plb->plb_ptr;
611 
612  done:
613 	return ((uint32_t *) oldptr);
614  fail:
615 	return (NULL);
616 }
617 
618 /*
619  * Schedule an I/O.
620  *
621  * Transfer the current buffer to the helper kthread.
622  */
623 
624 static void
625 pmclog_schedule_io(struct pmc_owner *po, int wakeup)
626 {
627 	struct pmclog_buffer *plb;
628 
629 	plb = po->po_curbuf[curcpu];
630 	po->po_curbuf[curcpu] = NULL;
631 	KASSERT(plb != NULL,
632 	    ("[pmclog,%d] schedule_io with null buffer po=%p", __LINE__, po));
633 	KASSERT(plb->plb_ptr >= plb->plb_base,
634 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p base=%p", __LINE__,
635 		po, plb->plb_ptr, plb->plb_base));
636 	KASSERT(plb->plb_ptr <= plb->plb_fence,
637 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p fenc=%p", __LINE__,
638 		po, plb->plb_ptr, plb->plb_fence));
639 
640 	PMCDBG1(LOG,SIO, 1, "po=%p", po);
641 
642 	/*
643 	 * Add the current buffer to the tail of the buffer list and
644 	 * wakeup the helper.
645 	 */
646 	mtx_lock_spin(&po->po_mtx);
647 	TAILQ_INSERT_TAIL(&po->po_logbuffers, plb, plb_next);
648 	mtx_unlock_spin(&po->po_mtx);
649 	if (wakeup)
650 		wakeup_one(po);
651 }
652 
653 /*
654  * Stop the helper kthread.
655  */
656 
657 static void
658 pmclog_stop_kthread(struct pmc_owner *po)
659 {
660 
661 	mtx_lock(&pmc_kthread_mtx);
662 	po->po_flags &= ~PMC_PO_OWNS_LOGFILE;
663 	if (po->po_kthread != NULL) {
664 		PROC_LOCK(po->po_kthread);
665 		kern_psignal(po->po_kthread, SIGHUP);
666 		PROC_UNLOCK(po->po_kthread);
667 	}
668 	wakeup_one(po);
669 	while (po->po_kthread)
670 		msleep(po->po_kthread, &pmc_kthread_mtx, PPAUSE, "pmckstp", 0);
671 	mtx_unlock(&pmc_kthread_mtx);
672 }
673 
674 /*
675  * Public functions
676  */
677 
678 /*
679  * Configure a log file for pmc owner 'po'.
680  *
681  * Parameter 'logfd' is a file handle referencing an open file in the
682  * owner process.  This file needs to have been opened for writing.
683  */
684 
685 int
686 pmclog_configure_log(struct pmc_mdep *md, struct pmc_owner *po, int logfd)
687 {
688 	struct proc *p;
689 	struct timespec ts;
690 	int error;
691 
692 	sx_assert(&pmc_sx, SA_XLOCKED);
693 	PMCDBG2(LOG,CFG,1, "config po=%p logfd=%d", po, logfd);
694 
695 	p = po->po_owner;
696 
697 	/* return EBUSY if a log file was already present */
698 	if (po->po_flags & PMC_PO_OWNS_LOGFILE)
699 		return (EBUSY);
700 
701 	KASSERT(po->po_file == NULL,
702 	    ("[pmclog,%d] po=%p file (%p) already present", __LINE__, po,
703 		po->po_file));
704 
705 	/* get a reference to the file state */
706 	error = fget_write(curthread, logfd, &cap_write_rights, &po->po_file);
707 	if (error)
708 		goto error;
709 
710 	/* mark process as owning a log file */
711 	po->po_flags |= PMC_PO_OWNS_LOGFILE;
712 
713 	/* mark process as using HWPMCs */
714 	PROC_LOCK(p);
715 	p->p_flag |= P_HWPMC;
716 	PROC_UNLOCK(p);
717 	nanotime(&ts);
718 	/* create a log initialization entry */
719 	PMCLOG_RESERVE_WITH_ERROR(po, PMCLOG_TYPE_INITIALIZE,
720 	    sizeof(struct pmclog_initialize));
721 	PMCLOG_EMIT32(PMC_VERSION);
722 	PMCLOG_EMIT32(md->pmd_cputype);
723 #if defined(__i386__) || defined(__amd64__)
724 	PMCLOG_EMIT64(tsc_freq);
725 #else
726 	/* other architectures will need to fill this in */
727 	PMCLOG_EMIT32(0);
728 	PMCLOG_EMIT32(0);
729 #endif
730 	memcpy(_le, &ts, sizeof(ts));
731 	_le += sizeof(ts)/4;
732 	PMCLOG_EMITSTRING(pmc_cpuid, PMC_CPUID_LEN);
733 	PMCLOG_DESPATCH_SYNC(po);
734 
735 	return (0);
736 
737  error:
738 	KASSERT(po->po_kthread == NULL, ("[pmclog,%d] po=%p kthread not "
739 	    "stopped", __LINE__, po));
740 
741 	if (po->po_file)
742 		(void) fdrop(po->po_file, curthread);
743 	po->po_file  = NULL;	/* clear file and error state */
744 	po->po_error = 0;
745 	po->po_flags &= ~PMC_PO_OWNS_LOGFILE;
746 
747 	return (error);
748 }
749 
750 
751 /*
752  * De-configure a log file.  This will throw away any buffers queued
753  * for this owner process.
754  */
755 
756 int
757 pmclog_deconfigure_log(struct pmc_owner *po)
758 {
759 	int error;
760 	struct pmclog_buffer *lb;
761 	struct pmc_binding pb;
762 
763 	PMCDBG1(LOG,CFG,1, "de-config po=%p", po);
764 
765 	if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
766 		return (EINVAL);
767 
768 	KASSERT(po->po_sscount == 0,
769 	    ("[pmclog,%d] po=%p still owning SS PMCs", __LINE__, po));
770 	KASSERT(po->po_file != NULL,
771 	    ("[pmclog,%d] po=%p no log file", __LINE__, po));
772 
773 	/* stop the kthread, this will reset the 'OWNS_LOGFILE' flag */
774 	pmclog_stop_kthread(po);
775 
776 	KASSERT(po->po_kthread == NULL,
777 	    ("[pmclog,%d] po=%p kthread not stopped", __LINE__, po));
778 
779 	/* return all queued log buffers to the global pool */
780 	while ((lb = TAILQ_FIRST(&po->po_logbuffers)) != NULL) {
781 		TAILQ_REMOVE(&po->po_logbuffers, lb, plb_next);
782 		PMCLOG_RESET_BUFFER_DESCRIPTOR(lb);
783 		pmc_plb_rele(lb);
784 	}
785 	pmc_save_cpu_binding(&pb);
786 	for (int i = 0; i < mp_ncpus; i++) {
787 		pmc_select_cpu(i);
788 		/* return the 'current' buffer to the global pool */
789 		if ((lb = po->po_curbuf[curcpu]) != NULL) {
790 			PMCLOG_RESET_BUFFER_DESCRIPTOR(lb);
791 			pmc_plb_rele(lb);
792 		}
793 	}
794 	pmc_restore_cpu_binding(&pb);
795 
796 	/* drop a reference to the fd */
797 	if (po->po_file != NULL) {
798 		error = fdrop(po->po_file, curthread);
799 		po->po_file = NULL;
800 	} else
801 		error = 0;
802 	po->po_error = 0;
803 
804 	return (error);
805 }
806 
807 /*
808  * Flush a process' log buffer.
809  */
810 
811 int
812 pmclog_flush(struct pmc_owner *po, int force)
813 {
814 	int error;
815 
816 	PMCDBG1(LOG,FLS,1, "po=%p", po);
817 
818 	/*
819 	 * If there is a pending error recorded by the logger thread,
820 	 * return that.
821 	 */
822 	if (po->po_error)
823 		return (po->po_error);
824 
825 	error = 0;
826 
827 	/*
828 	 * Check that we do have an active log file.
829 	 */
830 	mtx_lock(&pmc_kthread_mtx);
831 	if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) {
832 		error = EINVAL;
833 		goto error;
834 	}
835 
836 	pmclog_schedule_all(po);
837  error:
838 	mtx_unlock(&pmc_kthread_mtx);
839 
840 	return (error);
841 }
842 
843 static void
844 pmclog_schedule_one_cond(struct pmc_owner *po)
845 {
846 	struct pmclog_buffer *plb;
847 	int cpu;
848 
849 	spinlock_enter();
850 	cpu = curcpu;
851 	/* tell hardclock not to run again */
852 	if (PMC_CPU_HAS_SAMPLES(cpu))
853 		PMC_CALL_HOOK_UNLOCKED(curthread, PMC_FN_DO_SAMPLES, NULL);
854 
855 	plb = po->po_curbuf[cpu];
856 	if (plb && plb->plb_ptr != plb->plb_base)
857 		pmclog_schedule_io(po, 1);
858 	spinlock_exit();
859 }
860 
861 static void
862 pmclog_schedule_all(struct pmc_owner *po)
863 {
864 	struct pmc_binding pb;
865 
866 	/*
867 	 * Schedule the current buffer if any and not empty.
868 	 */
869 	pmc_save_cpu_binding(&pb);
870 	for (int i = 0; i < mp_ncpus; i++) {
871 		pmc_select_cpu(i);
872 		pmclog_schedule_one_cond(po);
873 	}
874 	pmc_restore_cpu_binding(&pb);
875 }
876 
877 int
878 pmclog_close(struct pmc_owner *po)
879 {
880 
881 	PMCDBG1(LOG,CLO,1, "po=%p", po);
882 
883 	pmclog_process_closelog(po);
884 
885 	mtx_lock(&pmc_kthread_mtx);
886 	/*
887 	 * Initiate shutdown: no new data queued,
888 	 * thread will close file on last block.
889 	 */
890 	po->po_flags |= PMC_PO_SHUTDOWN;
891 	/* give time for all to see */
892 	DELAY(50);
893 
894 	/*
895 	 * Schedule the current buffer.
896 	 */
897 	pmclog_schedule_all(po);
898 	wakeup_one(po);
899 
900 	mtx_unlock(&pmc_kthread_mtx);
901 
902 	return (0);
903 }
904 
905 void
906 pmclog_process_callchain(struct pmc *pm, struct pmc_sample *ps)
907 {
908 	int n, recordlen;
909 	uint32_t flags;
910 	struct pmc_owner *po;
911 
912 	PMCDBG3(LOG,SAM,1,"pm=%p pid=%d n=%d", pm, ps->ps_pid,
913 	    ps->ps_nsamples);
914 
915 	recordlen = offsetof(struct pmclog_callchain, pl_pc) +
916 	    ps->ps_nsamples * sizeof(uintfptr_t);
917 	po = pm->pm_owner;
918 	flags = PMC_CALLCHAIN_TO_CPUFLAGS(ps->ps_cpu,ps->ps_flags);
919 	PMCLOG_RESERVE_SAFE(po, PMCLOG_TYPE_CALLCHAIN, recordlen, ps->ps_tsc);
920 	PMCLOG_EMIT32(ps->ps_pid);
921 	PMCLOG_EMIT32(ps->ps_tid);
922 	PMCLOG_EMIT32(pm->pm_id);
923 	PMCLOG_EMIT32(flags);
924 	for (n = 0; n < ps->ps_nsamples; n++)
925 		PMCLOG_EMITADDR(ps->ps_pc[n]);
926 	PMCLOG_DESPATCH_SAFE(po);
927 }
928 
929 void
930 pmclog_process_closelog(struct pmc_owner *po)
931 {
932 	PMCLOG_RESERVE(po, PMCLOG_TYPE_CLOSELOG,
933 	    sizeof(struct pmclog_closelog));
934 	PMCLOG_DESPATCH_SYNC(po);
935 }
936 
937 void
938 pmclog_process_dropnotify(struct pmc_owner *po)
939 {
940 	PMCLOG_RESERVE(po, PMCLOG_TYPE_DROPNOTIFY,
941 	    sizeof(struct pmclog_dropnotify));
942 	PMCLOG_DESPATCH(po);
943 }
944 
945 void
946 pmclog_process_map_in(struct pmc_owner *po, pid_t pid, uintfptr_t start,
947     const char *path)
948 {
949 	int pathlen, recordlen;
950 
951 	KASSERT(path != NULL, ("[pmclog,%d] map-in, null path", __LINE__));
952 
953 	pathlen = strlen(path) + 1;	/* #bytes for path name */
954 	recordlen = offsetof(struct pmclog_map_in, pl_pathname) +
955 	    pathlen;
956 
957 	PMCLOG_RESERVE(po, PMCLOG_TYPE_MAP_IN, recordlen);
958 	PMCLOG_EMIT32(pid);
959 	PMCLOG_EMIT32(0);
960 	PMCLOG_EMITADDR(start);
961 	PMCLOG_EMITSTRING(path,pathlen);
962 	PMCLOG_DESPATCH_SYNC(po);
963 }
964 
965 void
966 pmclog_process_map_out(struct pmc_owner *po, pid_t pid, uintfptr_t start,
967     uintfptr_t end)
968 {
969 	KASSERT(start <= end, ("[pmclog,%d] start > end", __LINE__));
970 
971 	PMCLOG_RESERVE(po, PMCLOG_TYPE_MAP_OUT, sizeof(struct pmclog_map_out));
972 	PMCLOG_EMIT32(pid);
973 	PMCLOG_EMIT32(0);
974 	PMCLOG_EMITADDR(start);
975 	PMCLOG_EMITADDR(end);
976 	PMCLOG_DESPATCH(po);
977 }
978 
979 void
980 pmclog_process_pmcallocate(struct pmc *pm)
981 {
982 	struct pmc_owner *po;
983 	struct pmc_soft *ps;
984 
985 	po = pm->pm_owner;
986 
987 	PMCDBG1(LOG,ALL,1, "pm=%p", pm);
988 
989 	if (PMC_TO_CLASS(pm) == PMC_CLASS_SOFT) {
990 		PMCLOG_RESERVE(po, PMCLOG_TYPE_PMCALLOCATEDYN,
991 		    sizeof(struct pmclog_pmcallocatedyn));
992 		PMCLOG_EMIT32(pm->pm_id);
993 		PMCLOG_EMIT32(pm->pm_event);
994 		PMCLOG_EMIT32(pm->pm_flags);
995 		PMCLOG_EMIT32(0);
996 		PMCLOG_EMIT64(pm->pm_sc.pm_reloadcount);
997 		ps = pmc_soft_ev_acquire(pm->pm_event);
998 		if (ps != NULL)
999 			PMCLOG_EMITSTRING(ps->ps_ev.pm_ev_name,PMC_NAME_MAX);
1000 		else
1001 			PMCLOG_EMITNULLSTRING(PMC_NAME_MAX);
1002 		pmc_soft_ev_release(ps);
1003 		PMCLOG_DESPATCH_SYNC(po);
1004 	} else {
1005 		PMCLOG_RESERVE(po, PMCLOG_TYPE_PMCALLOCATE,
1006 		    sizeof(struct pmclog_pmcallocate));
1007 		PMCLOG_EMIT32(pm->pm_id);
1008 		PMCLOG_EMIT32(pm->pm_event);
1009 		PMCLOG_EMIT32(pm->pm_flags);
1010 		PMCLOG_EMIT32(0);
1011 		PMCLOG_EMIT64(pm->pm_sc.pm_reloadcount);
1012 		PMCLOG_DESPATCH_SYNC(po);
1013 	}
1014 }
1015 
1016 void
1017 pmclog_process_pmcattach(struct pmc *pm, pid_t pid, char *path)
1018 {
1019 	int pathlen, recordlen;
1020 	struct pmc_owner *po;
1021 
1022 	PMCDBG2(LOG,ATT,1,"pm=%p pid=%d", pm, pid);
1023 
1024 	po = pm->pm_owner;
1025 
1026 	pathlen = strlen(path) + 1;	/* #bytes for the string */
1027 	recordlen = offsetof(struct pmclog_pmcattach, pl_pathname) + pathlen;
1028 
1029 	PMCLOG_RESERVE(po, PMCLOG_TYPE_PMCATTACH, recordlen);
1030 	PMCLOG_EMIT32(pm->pm_id);
1031 	PMCLOG_EMIT32(pid);
1032 	PMCLOG_EMITSTRING(path, pathlen);
1033 	PMCLOG_DESPATCH_SYNC(po);
1034 }
1035 
1036 void
1037 pmclog_process_pmcdetach(struct pmc *pm, pid_t pid)
1038 {
1039 	struct pmc_owner *po;
1040 
1041 	PMCDBG2(LOG,ATT,1,"!pm=%p pid=%d", pm, pid);
1042 
1043 	po = pm->pm_owner;
1044 
1045 	PMCLOG_RESERVE(po, PMCLOG_TYPE_PMCDETACH,
1046 	    sizeof(struct pmclog_pmcdetach));
1047 	PMCLOG_EMIT32(pm->pm_id);
1048 	PMCLOG_EMIT32(pid);
1049 	PMCLOG_DESPATCH_SYNC(po);
1050 }
1051 
1052 void
1053 pmclog_process_proccreate(struct pmc_owner *po, struct proc *p, int sync)
1054 {
1055 	if (sync) {
1056 		PMCLOG_RESERVE(po, PMCLOG_TYPE_PROC_CREATE,
1057 		    sizeof(struct pmclog_proccreate));
1058 		PMCLOG_EMIT32(p->p_pid);
1059 		PMCLOG_EMIT32(p->p_flag);
1060 		PMCLOG_EMITSTRING(p->p_comm, MAXCOMLEN+1);
1061 		PMCLOG_DESPATCH_SYNC(po);
1062 	} else {
1063 		PMCLOG_RESERVE(po, PMCLOG_TYPE_PROC_CREATE,
1064 		    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(po);
1069 	}
1070 }
1071 
1072 /*
1073  * Log a context switch event to the log file.
1074  */
1075 
1076 void
1077 pmclog_process_proccsw(struct pmc *pm, struct pmc_process *pp, pmc_value_t v, struct thread *td)
1078 {
1079 	struct pmc_owner *po;
1080 
1081 	KASSERT(pm->pm_flags & PMC_F_LOG_PROCCSW,
1082 	    ("[pmclog,%d] log-process-csw called gratuitously", __LINE__));
1083 
1084 	PMCDBG3(LOG,SWO,1,"pm=%p pid=%d v=%jx", pm, pp->pp_proc->p_pid,
1085 	    v);
1086 
1087 	po = pm->pm_owner;
1088 
1089 	PMCLOG_RESERVE_SAFE(po, PMCLOG_TYPE_PROCCSW,
1090 	    sizeof(struct pmclog_proccsw), pmc_rdtsc());
1091 	PMCLOG_EMIT64(v);
1092 	PMCLOG_EMIT32(pm->pm_id);
1093 	PMCLOG_EMIT32(pp->pp_proc->p_pid);
1094 	PMCLOG_EMIT32(td->td_tid);
1095 	PMCLOG_EMIT32(0);
1096 	PMCLOG_DESPATCH_SCHED_LOCK(po);
1097 }
1098 
1099 void
1100 pmclog_process_procexec(struct pmc_owner *po, pmc_id_t pmid, pid_t pid,
1101     uintptr_t baseaddr, uintptr_t dynaddr, char *path)
1102 {
1103 	int pathlen, recordlen;
1104 
1105 	PMCDBG3(LOG,EXC,1,"po=%p pid=%d path=\"%s\"", po, pid, path);
1106 
1107 	pathlen   = strlen(path) + 1;	/* #bytes for the path */
1108 	recordlen = offsetof(struct pmclog_procexec, pl_pathname) + pathlen;
1109 	PMCLOG_RESERVE(po, PMCLOG_TYPE_PROCEXEC, recordlen);
1110 	PMCLOG_EMIT32(pid);
1111 	PMCLOG_EMIT32(pmid);
1112 	PMCLOG_EMITADDR(baseaddr);
1113 	PMCLOG_EMITADDR(dynaddr);
1114 	PMCLOG_EMITSTRING(path,pathlen);
1115 	PMCLOG_DESPATCH_SYNC(po);
1116 }
1117 
1118 /*
1119  * Log a process exit event (and accumulated pmc value) to the log file.
1120  */
1121 
1122 void
1123 pmclog_process_procexit(struct pmc *pm, struct pmc_process *pp)
1124 {
1125 	int ri;
1126 	struct pmc_owner *po;
1127 
1128 	ri = PMC_TO_ROWINDEX(pm);
1129 	PMCDBG3(LOG,EXT,1,"pm=%p pid=%d v=%jx", pm, pp->pp_proc->p_pid,
1130 	    pp->pp_pmcs[ri].pp_pmcval);
1131 
1132 	po = pm->pm_owner;
1133 
1134 	PMCLOG_RESERVE(po, PMCLOG_TYPE_PROCEXIT,
1135 	    sizeof(struct pmclog_procexit));
1136 	PMCLOG_EMIT32(pm->pm_id);
1137 	PMCLOG_EMIT32(pp->pp_proc->p_pid);
1138 	PMCLOG_EMIT64(pp->pp_pmcs[ri].pp_pmcval);
1139 	PMCLOG_DESPATCH(po);
1140 }
1141 
1142 /*
1143  * Log a fork event.
1144  */
1145 
1146 void
1147 pmclog_process_procfork(struct pmc_owner *po, pid_t oldpid, pid_t newpid)
1148 {
1149 	PMCLOG_RESERVE(po, PMCLOG_TYPE_PROCFORK,
1150 	    sizeof(struct pmclog_procfork));
1151 	PMCLOG_EMIT32(oldpid);
1152 	PMCLOG_EMIT32(newpid);
1153 	PMCLOG_DESPATCH(po);
1154 }
1155 
1156 /*
1157  * Log a process exit event of the form suitable for system-wide PMCs.
1158  */
1159 
1160 void
1161 pmclog_process_sysexit(struct pmc_owner *po, pid_t pid)
1162 {
1163 	PMCLOG_RESERVE(po, PMCLOG_TYPE_SYSEXIT, sizeof(struct pmclog_sysexit));
1164 	PMCLOG_EMIT32(pid);
1165 	PMCLOG_DESPATCH(po);
1166 }
1167 
1168 void
1169 pmclog_process_threadcreate(struct pmc_owner *po, struct thread *td, int sync)
1170 {
1171 	struct proc *p;
1172 
1173 	p = td->td_proc;
1174 	if (sync) {
1175 		PMCLOG_RESERVE(po, PMCLOG_TYPE_THR_CREATE,
1176 		    sizeof(struct pmclog_threadcreate));
1177 		PMCLOG_EMIT32(td->td_tid);
1178 		PMCLOG_EMIT32(p->p_pid);
1179 		PMCLOG_EMIT32(p->p_flag);
1180 		PMCLOG_EMIT32(0);
1181 		PMCLOG_EMITSTRING(td->td_name, MAXCOMLEN+1);
1182 		PMCLOG_DESPATCH_SYNC(po);
1183 	} else {
1184 		PMCLOG_RESERVE(po, PMCLOG_TYPE_THR_CREATE,
1185 		    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, PMCLOG_TYPE_THR_EXIT,
1200 	    sizeof(struct pmclog_threadexit));
1201 	PMCLOG_EMIT32(td->td_tid);
1202 	PMCLOG_DESPATCH(po);
1203 }
1204 
1205 /*
1206  * Write a user log entry.
1207  */
1208 
1209 int
1210 pmclog_process_userlog(struct pmc_owner *po, struct pmc_op_writelog *wl)
1211 {
1212 	int error;
1213 
1214 	PMCDBG2(LOG,WRI,1, "writelog po=%p ud=0x%x", po, wl->pm_userdata);
1215 
1216 	error = 0;
1217 
1218 	PMCLOG_RESERVE_WITH_ERROR(po, PMCLOG_TYPE_USERDATA,
1219 	    sizeof(struct pmclog_userdata));
1220 	PMCLOG_EMIT32(wl->pm_userdata);
1221 	PMCLOG_DESPATCH(po);
1222 
1223  error:
1224 	return (error);
1225 }
1226 
1227 /*
1228  * Initialization.
1229  *
1230  * Create a pool of log buffers and initialize mutexes.
1231  */
1232 
1233 void
1234 pmclog_initialize(void)
1235 {
1236 	struct pmclog_buffer *plb;
1237 	int domain, ncpus, total;
1238 
1239 	if (pmclog_buffer_size <= 0 || pmclog_buffer_size > 16*1024) {
1240 		(void) printf("hwpmc: tunable logbuffersize=%d must be "
1241 					  "greater than zero and less than or equal to 16MB.\n",
1242 					  pmclog_buffer_size);
1243 		pmclog_buffer_size = PMC_LOG_BUFFER_SIZE;
1244 	}
1245 
1246 	if (pmc_nlogbuffers_pcpu <= 0) {
1247 		(void) printf("hwpmc: tunable nlogbuffers=%d must be greater "
1248 					  "than zero.\n", pmc_nlogbuffers_pcpu);
1249 		pmc_nlogbuffers_pcpu = PMC_NLOGBUFFERS_PCPU;
1250 	}
1251 	if (pmc_nlogbuffers_pcpu*pmclog_buffer_size > 32*1024) {
1252 		(void) printf("hwpmc: memory allocated pcpu must be less than 32MB (is %dK).\n",
1253 					  pmc_nlogbuffers_pcpu*pmclog_buffer_size);
1254 		pmc_nlogbuffers_pcpu = PMC_NLOGBUFFERS_PCPU;
1255 		pmclog_buffer_size = PMC_LOG_BUFFER_SIZE;
1256 	}
1257 	for (domain = 0; domain < vm_ndomains; domain++) {
1258 		ncpus = pmc_dom_hdrs[domain]->pdbh_ncpus;
1259 		total = ncpus * pmc_nlogbuffers_pcpu;
1260 
1261 		plb = malloc_domainset(sizeof(struct pmclog_buffer) * total,
1262 		    M_PMC, DOMAINSET_PREF(domain), M_WAITOK | M_ZERO);
1263 		pmc_dom_hdrs[domain]->pdbh_plbs = plb;
1264 		for (; total > 0; total--, plb++) {
1265 			void *buf;
1266 
1267 			buf = malloc_domainset(1024 * pmclog_buffer_size, M_PMC,
1268 			    DOMAINSET_PREF(domain), M_WAITOK | M_ZERO);
1269 			PMCLOG_INIT_BUFFER_DESCRIPTOR(plb, buf, domain);
1270 			pmc_plb_rele_unlocked(plb);
1271 		}
1272 	}
1273 	mtx_init(&pmc_kthread_mtx, "pmc-kthread", "pmc-sleep", MTX_DEF);
1274 }
1275 
1276 /*
1277  * Shutdown logging.
1278  *
1279  * Destroy mutexes and release memory back the to free pool.
1280  */
1281 
1282 void
1283 pmclog_shutdown(void)
1284 {
1285 	struct pmclog_buffer *plb;
1286 	int domain;
1287 
1288 	mtx_destroy(&pmc_kthread_mtx);
1289 
1290 	for (domain = 0; domain < vm_ndomains; domain++) {
1291 		while ((plb = TAILQ_FIRST(&pmc_dom_hdrs[domain]->pdbh_head)) != NULL) {
1292 			TAILQ_REMOVE(&pmc_dom_hdrs[domain]->pdbh_head, plb, plb_next);
1293 			free(plb->plb_base, M_PMC);
1294 		}
1295 		free(pmc_dom_hdrs[domain]->pdbh_plbs, M_PMC);
1296 	}
1297 }
1298