xref: /freebsd/sys/dev/hwpmc/hwpmc_logging.c (revision eb6d21b4ca6d668cf89afd99eef7baeafa712197)
1 /*-
2  * Copyright (c) 2005-2007 Joseph Koshy
3  * Copyright (c) 2007 The FreeBSD Foundation
4  * All rights reserved.
5  *
6  * Portions of this software were developed by A. Joseph Koshy under
7  * sponsorship from the FreeBSD Foundation and Google, Inc.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions and the following disclaimer.
14  * 2. Redistributions in binary form must reproduce the above copyright
15  *    notice, this list of conditions and the following disclaimer in the
16  *    documentation and/or other materials provided with the distribution.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
22  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28  * SUCH DAMAGE.
29  *
30  */
31 
32 /*
33  * Logging code for hwpmc(4)
34  */
35 
36 #include <sys/cdefs.h>
37 __FBSDID("$FreeBSD$");
38 
39 #include <sys/param.h>
40 #include <sys/file.h>
41 #include <sys/kernel.h>
42 #include <sys/kthread.h>
43 #include <sys/lock.h>
44 #include <sys/module.h>
45 #include <sys/mutex.h>
46 #include <sys/pmc.h>
47 #include <sys/pmckern.h>
48 #include <sys/pmclog.h>
49 #include <sys/proc.h>
50 #include <sys/signalvar.h>
51 #include <sys/sysctl.h>
52 #include <sys/systm.h>
53 #include <sys/uio.h>
54 #include <sys/unistd.h>
55 #include <sys/vnode.h>
56 
57 /*
58  * Sysctl tunables
59  */
60 
61 SYSCTL_DECL(_kern_hwpmc);
62 
63 /*
64  * kern.hwpmc.logbuffersize -- size of the per-cpu owner buffers.
65  */
66 
67 static int pmclog_buffer_size = PMC_LOG_BUFFER_SIZE;
68 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "logbuffersize", &pmclog_buffer_size);
69 SYSCTL_INT(_kern_hwpmc, OID_AUTO, logbuffersize, CTLFLAG_TUN|CTLFLAG_RD,
70     &pmclog_buffer_size, 0, "size of log buffers in kilobytes");
71 
72 /*
73  * kern.hwpmc.nbuffer -- number of global log buffers
74  */
75 
76 static int pmc_nlogbuffers = PMC_NLOGBUFFERS;
77 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "nbuffers", &pmc_nlogbuffers);
78 SYSCTL_INT(_kern_hwpmc, OID_AUTO, nbuffers, CTLFLAG_TUN|CTLFLAG_RD,
79     &pmc_nlogbuffers, 0, "number of global log buffers");
80 
81 /*
82  * Global log buffer list and associated spin lock.
83  */
84 
85 TAILQ_HEAD(, pmclog_buffer) pmc_bufferlist =
86 	TAILQ_HEAD_INITIALIZER(pmc_bufferlist);
87 static struct mtx pmc_bufferlist_mtx;	/* spin lock */
88 static struct mtx pmc_kthread_mtx;	/* sleep lock */
89 
90 #define	PMCLOG_INIT_BUFFER_DESCRIPTOR(D) do {				\
91 		const int __roundup = roundup(sizeof(*D),		\
92 			sizeof(uint32_t));				\
93 		(D)->plb_fence = ((char *) (D)) +			\
94 			 1024*pmclog_buffer_size;			\
95 		(D)->plb_base  = (D)->plb_ptr = ((char *) (D)) +	\
96 			__roundup;					\
97 	} while (0)
98 
99 
100 /*
101  * Log file record constructors.
102  */
103 #define	_PMCLOG_TO_HEADER(T,L)						\
104 	((PMCLOG_HEADER_MAGIC << 24) |					\
105 	 (PMCLOG_TYPE_ ## T << 16)   |					\
106 	 ((L) & 0xFFFF))
107 
108 /* reserve LEN bytes of space and initialize the entry header */
109 #define	_PMCLOG_RESERVE(PO,TYPE,LEN,ACTION) do {			\
110 		uint32_t *_le;						\
111 		int _len = roundup((LEN), sizeof(uint32_t));		\
112 		if ((_le = pmclog_reserve((PO), _len)) == NULL) {	\
113 			ACTION;						\
114 		}							\
115 		*_le = _PMCLOG_TO_HEADER(TYPE,_len);			\
116 		_le += 3	/* skip over timestamp */
117 
118 #define	PMCLOG_RESERVE(P,T,L)		_PMCLOG_RESERVE(P,T,L,return)
119 #define	PMCLOG_RESERVE_WITH_ERROR(P,T,L) _PMCLOG_RESERVE(P,T,L,		\
120 	error=ENOMEM;goto error)
121 
122 #define	PMCLOG_EMIT32(V)	do { *_le++ = (V); } while (0)
123 #define	PMCLOG_EMIT64(V)	do { 					\
124 		*_le++ = (uint32_t) ((V) & 0xFFFFFFFF);			\
125 		*_le++ = (uint32_t) (((V) >> 32) & 0xFFFFFFFF);		\
126 	} while (0)
127 
128 
129 /* Emit a string.  Caution: does NOT update _le, so needs to be last */
130 #define	PMCLOG_EMITSTRING(S,L)	do { bcopy((S), _le, (L)); } while (0)
131 
132 #define	PMCLOG_DESPATCH(PO)						\
133 		pmclog_release((PO));					\
134 	} while (0)
135 
136 
137 /*
138  * Assertions about the log file format.
139  */
140 
141 CTASSERT(sizeof(struct pmclog_callchain) == 6*4 +
142     PMC_CALLCHAIN_DEPTH_MAX*sizeof(uintfptr_t));
143 CTASSERT(sizeof(struct pmclog_closelog) == 3*4);
144 CTASSERT(sizeof(struct pmclog_dropnotify) == 3*4);
145 CTASSERT(sizeof(struct pmclog_map_in) == PATH_MAX +
146     4*4 + sizeof(uintfptr_t));
147 CTASSERT(offsetof(struct pmclog_map_in,pl_pathname) ==
148     4*4 + sizeof(uintfptr_t));
149 CTASSERT(sizeof(struct pmclog_map_out) == 4*4 + 2*sizeof(uintfptr_t));
150 CTASSERT(sizeof(struct pmclog_pcsample) == 6*4 + sizeof(uintfptr_t));
151 CTASSERT(sizeof(struct pmclog_pmcallocate) == 6*4);
152 CTASSERT(sizeof(struct pmclog_pmcattach) == 5*4 + PATH_MAX);
153 CTASSERT(offsetof(struct pmclog_pmcattach,pl_pathname) == 5*4);
154 CTASSERT(sizeof(struct pmclog_pmcdetach) == 5*4);
155 CTASSERT(sizeof(struct pmclog_proccsw) == 5*4 + 8);
156 CTASSERT(sizeof(struct pmclog_procexec) == 5*4 + PATH_MAX +
157     sizeof(uintfptr_t));
158 CTASSERT(offsetof(struct pmclog_procexec,pl_pathname) == 5*4 +
159     sizeof(uintfptr_t));
160 CTASSERT(sizeof(struct pmclog_procexit) == 5*4 + 8);
161 CTASSERT(sizeof(struct pmclog_procfork) == 5*4);
162 CTASSERT(sizeof(struct pmclog_sysexit) == 4*4);
163 CTASSERT(sizeof(struct pmclog_userdata) == 4*4);
164 
165 /*
166  * Log buffer structure
167  */
168 
169 struct pmclog_buffer {
170 	TAILQ_ENTRY(pmclog_buffer) plb_next;
171 	char 		*plb_base;
172 	char		*plb_ptr;
173 	char 		*plb_fence;
174 };
175 
176 /*
177  * Prototypes
178  */
179 
180 static int pmclog_get_buffer(struct pmc_owner *po);
181 static void pmclog_loop(void *arg);
182 static void pmclog_release(struct pmc_owner *po);
183 static uint32_t *pmclog_reserve(struct pmc_owner *po, int length);
184 static void pmclog_schedule_io(struct pmc_owner *po);
185 static void pmclog_stop_kthread(struct pmc_owner *po);
186 
187 /*
188  * Helper functions
189  */
190 
191 /*
192  * Get a log buffer
193  */
194 
195 static int
196 pmclog_get_buffer(struct pmc_owner *po)
197 {
198 	struct pmclog_buffer *plb;
199 
200 	mtx_assert(&po->po_mtx, MA_OWNED);
201 
202 	KASSERT(po->po_curbuf == NULL,
203 	    ("[pmclog,%d] po=%p current buffer still valid", __LINE__, po));
204 
205 	mtx_lock_spin(&pmc_bufferlist_mtx);
206 	if ((plb = TAILQ_FIRST(&pmc_bufferlist)) != NULL)
207 		TAILQ_REMOVE(&pmc_bufferlist, plb, plb_next);
208 	mtx_unlock_spin(&pmc_bufferlist_mtx);
209 
210 	PMCDBG(LOG,GTB,1, "po=%p plb=%p", po, plb);
211 
212 #ifdef	DEBUG
213 	if (plb)
214 		KASSERT(plb->plb_ptr == plb->plb_base &&
215 		    plb->plb_base < plb->plb_fence,
216 		    ("[pmclog,%d] po=%p buffer invariants: ptr=%p "
217 		    "base=%p fence=%p", __LINE__, po, plb->plb_ptr,
218 		    plb->plb_base, plb->plb_fence));
219 #endif
220 
221 	po->po_curbuf = plb;
222 
223 	/* update stats */
224 	atomic_add_int(&pmc_stats.pm_buffer_requests, 1);
225 	if (plb == NULL)
226 		atomic_add_int(&pmc_stats.pm_buffer_requests_failed, 1);
227 
228 	return (plb ? 0 : ENOMEM);
229 }
230 
231 /*
232  * Log handler loop.
233  *
234  * This function is executed by each pmc owner's helper thread.
235  */
236 
237 static void
238 pmclog_loop(void *arg)
239 {
240 	int error;
241 	struct pmc_owner *po;
242 	struct pmclog_buffer *lb;
243 	struct proc *p;
244 	struct ucred *ownercred;
245 	struct ucred *mycred;
246 	struct thread *td;
247 	struct uio auio;
248 	struct iovec aiov;
249 	size_t nbytes;
250 
251 	po = (struct pmc_owner *) arg;
252 	p = po->po_owner;
253 	td = curthread;
254 	mycred = td->td_ucred;
255 
256 	PROC_LOCK(p);
257 	ownercred = crhold(p->p_ucred);
258 	PROC_UNLOCK(p);
259 
260 	PMCDBG(LOG,INI,1, "po=%p kt=%p", po, po->po_kthread);
261 	KASSERT(po->po_kthread == curthread->td_proc,
262 	    ("[pmclog,%d] proc mismatch po=%p po/kt=%p curproc=%p", __LINE__,
263 		po, po->po_kthread, curthread->td_proc));
264 
265 	lb = NULL;
266 
267 
268 	/*
269 	 * Loop waiting for I/O requests to be added to the owner
270 	 * struct's queue.  The loop is exited when the log file
271 	 * is deconfigured.
272 	 */
273 
274 	mtx_lock(&pmc_kthread_mtx);
275 
276 	for (;;) {
277 
278 		/* check if we've been asked to exit */
279 		if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
280 			break;
281 
282 		if (lb == NULL) { /* look for a fresh buffer to write */
283 			mtx_lock_spin(&po->po_mtx);
284 			if ((lb = TAILQ_FIRST(&po->po_logbuffers)) == NULL) {
285 				mtx_unlock_spin(&po->po_mtx);
286 
287 				/*
288 				 * Wakeup the thread waiting for the
289 				 * PMC_OP_FLUSHLOG request to
290 				 * complete.
291 				 */
292 				if (po->po_flags & PMC_PO_IN_FLUSH) {
293 					po->po_flags &= ~PMC_PO_IN_FLUSH;
294 					wakeup_one(po->po_kthread);
295 				}
296 
297 				(void) msleep(po, &pmc_kthread_mtx, PWAIT,
298 				    "pmcloop", 0);
299 				continue;
300 			}
301 
302 			TAILQ_REMOVE(&po->po_logbuffers, lb, plb_next);
303 			mtx_unlock_spin(&po->po_mtx);
304 		}
305 
306 		mtx_unlock(&pmc_kthread_mtx);
307 
308 		/* process the request */
309 		PMCDBG(LOG,WRI,2, "po=%p base=%p ptr=%p", po,
310 		    lb->plb_base, lb->plb_ptr);
311 		/* change our thread's credentials before issuing the I/O */
312 
313 		aiov.iov_base = lb->plb_base;
314 		aiov.iov_len  = nbytes = lb->plb_ptr - lb->plb_base;
315 
316 		auio.uio_iov    = &aiov;
317 		auio.uio_iovcnt = 1;
318 		auio.uio_offset = -1;
319 		auio.uio_resid  = nbytes;
320 		auio.uio_rw     = UIO_WRITE;
321 		auio.uio_segflg = UIO_SYSSPACE;
322 		auio.uio_td     = td;
323 
324 		/* switch thread credentials -- see kern_ktrace.c */
325 		td->td_ucred = ownercred;
326 		error = fo_write(po->po_file, &auio, ownercred, 0, td);
327 		td->td_ucred = mycred;
328 
329 		if (error) {
330 			/* XXX some errors are recoverable */
331 			/* XXX also check for SIGPIPE if a socket */
332 
333 			/* send a SIGIO to the owner and exit */
334 			PROC_LOCK(p);
335 			psignal(p, SIGIO);
336 			PROC_UNLOCK(p);
337 
338 			mtx_lock(&pmc_kthread_mtx);
339 
340 			po->po_error = error; /* save for flush log */
341 
342 			PMCDBG(LOG,WRI,2, "po=%p error=%d", po, error);
343 
344 			break;
345 		}
346 
347 		mtx_lock(&pmc_kthread_mtx);
348 
349 		/* put the used buffer back into the global pool */
350 		PMCLOG_INIT_BUFFER_DESCRIPTOR(lb);
351 
352 		mtx_lock_spin(&pmc_bufferlist_mtx);
353 		TAILQ_INSERT_HEAD(&pmc_bufferlist, lb, plb_next);
354 		mtx_unlock_spin(&pmc_bufferlist_mtx);
355 
356 		lb = NULL;
357 	}
358 
359 	po->po_kthread = NULL;
360 
361 	mtx_unlock(&pmc_kthread_mtx);
362 
363 	/* return the current I/O buffer to the global pool */
364 	if (lb) {
365 		PMCLOG_INIT_BUFFER_DESCRIPTOR(lb);
366 
367 		mtx_lock_spin(&pmc_bufferlist_mtx);
368 		TAILQ_INSERT_HEAD(&pmc_bufferlist, lb, plb_next);
369 		mtx_unlock_spin(&pmc_bufferlist_mtx);
370 	}
371 
372 	/*
373 	 * Exit this thread, signalling the waiter
374 	 */
375 
376 	crfree(ownercred);
377 
378 	kproc_exit(0);
379 }
380 
381 /*
382  * Release and log entry and schedule an I/O if needed.
383  */
384 
385 static void
386 pmclog_release(struct pmc_owner *po)
387 {
388 	KASSERT(po->po_curbuf->plb_ptr >= po->po_curbuf->plb_base,
389 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p base=%p", __LINE__,
390 		po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_base));
391 	KASSERT(po->po_curbuf->plb_ptr <= po->po_curbuf->plb_fence,
392 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p fenc=%p", __LINE__,
393 		po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_fence));
394 
395 	/* schedule an I/O if we've filled a buffer */
396 	if (po->po_curbuf->plb_ptr >= po->po_curbuf->plb_fence)
397 		pmclog_schedule_io(po);
398 
399 	mtx_unlock_spin(&po->po_mtx);
400 
401 	PMCDBG(LOG,REL,1, "po=%p", po);
402 }
403 
404 
405 /*
406  * Attempt to reserve 'length' bytes of space in an owner's log
407  * buffer.  The function returns a pointer to 'length' bytes of space
408  * if there was enough space or returns NULL if no space was
409  * available.  Non-null returns do so with the po mutex locked.  The
410  * caller must invoke pmclog_release() on the pmc owner structure
411  * when done.
412  */
413 
414 static uint32_t *
415 pmclog_reserve(struct pmc_owner *po, int length)
416 {
417 	uintptr_t newptr, oldptr;
418 	uint32_t *lh;
419 	struct timespec ts;
420 
421 	PMCDBG(LOG,ALL,1, "po=%p len=%d", po, length);
422 
423 	KASSERT(length % sizeof(uint32_t) == 0,
424 	    ("[pmclog,%d] length not a multiple of word size", __LINE__));
425 
426 	mtx_lock_spin(&po->po_mtx);
427 
428 	if (po->po_curbuf == NULL)
429 		if (pmclog_get_buffer(po) != 0) {
430 			mtx_unlock_spin(&po->po_mtx);
431 			return (NULL);
432 		}
433 
434 	KASSERT(po->po_curbuf != NULL,
435 	    ("[pmclog,%d] po=%p no current buffer", __LINE__, po));
436 
437 	KASSERT(po->po_curbuf->plb_ptr >= po->po_curbuf->plb_base &&
438 	    po->po_curbuf->plb_ptr <= po->po_curbuf->plb_fence,
439 	    ("[pmclog,%d] po=%p buffer invariants: ptr=%p base=%p fence=%p",
440 		__LINE__, po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_base,
441 		po->po_curbuf->plb_fence));
442 
443 	oldptr = (uintptr_t) po->po_curbuf->plb_ptr;
444 	newptr = oldptr + length;
445 
446 	KASSERT(oldptr != (uintptr_t) NULL,
447 	    ("[pmclog,%d] po=%p Null log buffer pointer", __LINE__, po));
448 
449 	/*
450 	 * If we have space in the current buffer, return a pointer to
451 	 * available space with the PO structure locked.
452 	 */
453 	if (newptr <= (uintptr_t) po->po_curbuf->plb_fence) {
454 		po->po_curbuf->plb_ptr = (char *) newptr;
455 		goto done;
456 	}
457 
458 	/*
459 	 * Otherwise, schedule the current buffer for output and get a
460 	 * fresh buffer.
461 	 */
462 	pmclog_schedule_io(po);
463 
464 	if (pmclog_get_buffer(po) != 0) {
465 		mtx_unlock_spin(&po->po_mtx);
466 		return (NULL);
467 	}
468 
469 	KASSERT(po->po_curbuf != NULL,
470 	    ("[pmclog,%d] po=%p no current buffer", __LINE__, po));
471 
472 	KASSERT(po->po_curbuf->plb_ptr != NULL,
473 	    ("[pmclog,%d] null return from pmc_get_log_buffer", __LINE__));
474 
475 	KASSERT(po->po_curbuf->plb_ptr == po->po_curbuf->plb_base &&
476 	    po->po_curbuf->plb_ptr <= po->po_curbuf->plb_fence,
477 	    ("[pmclog,%d] po=%p buffer invariants: ptr=%p base=%p fence=%p",
478 		__LINE__, po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_base,
479 		po->po_curbuf->plb_fence));
480 
481 	oldptr = (uintptr_t) po->po_curbuf->plb_ptr;
482 
483  done:
484 	lh = (uint32_t *) oldptr;
485 	lh++;				/* skip header */
486 	getnanotime(&ts);		/* fill in the timestamp */
487 	*lh++ = ts.tv_sec & 0xFFFFFFFF;
488 	*lh++ = ts.tv_nsec & 0xFFFFFFF;
489 	return ((uint32_t *) oldptr);
490 }
491 
492 /*
493  * Schedule an I/O.
494  *
495  * Transfer the current buffer to the helper kthread.
496  */
497 
498 static void
499 pmclog_schedule_io(struct pmc_owner *po)
500 {
501 	KASSERT(po->po_curbuf != NULL,
502 	    ("[pmclog,%d] schedule_io with null buffer po=%p", __LINE__, po));
503 
504 	KASSERT(po->po_curbuf->plb_ptr >= po->po_curbuf->plb_base,
505 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p base=%p", __LINE__,
506 		po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_base));
507 	KASSERT(po->po_curbuf->plb_ptr <= po->po_curbuf->plb_fence,
508 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p fenc=%p", __LINE__,
509 		po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_fence));
510 
511 	PMCDBG(LOG,SIO, 1, "po=%p", po);
512 
513 	mtx_assert(&po->po_mtx, MA_OWNED);
514 
515 	/*
516 	 * Add the current buffer to the tail of the buffer list and
517 	 * wakeup the helper.
518 	 */
519 	TAILQ_INSERT_TAIL(&po->po_logbuffers, po->po_curbuf, plb_next);
520 	po->po_curbuf = NULL;
521 	wakeup_one(po);
522 }
523 
524 /*
525  * Stop the helper kthread.
526  */
527 
528 static void
529 pmclog_stop_kthread(struct pmc_owner *po)
530 {
531 	/*
532 	 * Close the file to force the thread out of fo_write,
533 	 * unset flag, wakeup the helper thread,
534 	 * wait for it to exit
535 	 */
536 
537 	if (po->po_file != NULL)
538 		fo_close(po->po_file, curthread);
539 
540 	mtx_lock(&pmc_kthread_mtx);
541 	po->po_flags &= ~PMC_PO_OWNS_LOGFILE;
542 	wakeup_one(po);
543 	if (po->po_kthread)
544 		msleep(po->po_kthread, &pmc_kthread_mtx, PPAUSE, "pmckstp", 0);
545 	mtx_unlock(&pmc_kthread_mtx);
546 }
547 
548 /*
549  * Public functions
550  */
551 
552 /*
553  * Configure a log file for pmc owner 'po'.
554  *
555  * Parameter 'logfd' is a file handle referencing an open file in the
556  * owner process.  This file needs to have been opened for writing.
557  */
558 
559 int
560 pmclog_configure_log(struct pmc_mdep *md, struct pmc_owner *po, int logfd)
561 {
562 	int error;
563 	struct proc *p;
564 
565 	/*
566 	 * As long as it is possible to get a LOR between pmc_sx lock and
567 	 * proctree/allproc sx locks used for adding a new process, assure
568 	 * the former is not held here.
569 	 */
570 	sx_assert(&pmc_sx, SA_UNLOCKED);
571 	PMCDBG(LOG,CFG,1, "config po=%p logfd=%d", po, logfd);
572 
573 	p = po->po_owner;
574 
575 	/* return EBUSY if a log file was already present */
576 	if (po->po_flags & PMC_PO_OWNS_LOGFILE)
577 		return (EBUSY);
578 
579 	KASSERT(po->po_kthread == NULL,
580 	    ("[pmclog,%d] po=%p kthread (%p) already present", __LINE__, po,
581 		po->po_kthread));
582 	KASSERT(po->po_file == NULL,
583 	    ("[pmclog,%d] po=%p file (%p) already present", __LINE__, po,
584 		po->po_file));
585 
586 	/* get a reference to the file state */
587 	error = fget_write(curthread, logfd, &po->po_file);
588 	if (error)
589 		goto error;
590 
591 	/* mark process as owning a log file */
592 	po->po_flags |= PMC_PO_OWNS_LOGFILE;
593 	error = kproc_create(pmclog_loop, po, &po->po_kthread,
594 	    RFHIGHPID, 0, "hwpmc: proc(%d)", p->p_pid);
595 	if (error)
596 		goto error;
597 
598 	/* mark process as using HWPMCs */
599 	PROC_LOCK(p);
600 	p->p_flag |= P_HWPMC;
601 	PROC_UNLOCK(p);
602 
603 	/* create a log initialization entry */
604 	PMCLOG_RESERVE_WITH_ERROR(po, INITIALIZE,
605 	    sizeof(struct pmclog_initialize));
606 	PMCLOG_EMIT32(PMC_VERSION);
607 	PMCLOG_EMIT32(md->pmd_cputype);
608 	PMCLOG_DESPATCH(po);
609 
610 	return (0);
611 
612  error:
613 	/* shutdown the thread */
614 	if (po->po_kthread)
615 		pmclog_stop_kthread(po);
616 
617 	KASSERT(po->po_kthread == NULL, ("[pmclog,%d] po=%p kthread not "
618 	    "stopped", __LINE__, po));
619 
620 	if (po->po_file)
621 		(void) fdrop(po->po_file, curthread);
622 	po->po_file  = NULL;	/* clear file and error state */
623 	po->po_error = 0;
624 
625 	return (error);
626 }
627 
628 
629 /*
630  * De-configure a log file.  This will throw away any buffers queued
631  * for this owner process.
632  */
633 
634 int
635 pmclog_deconfigure_log(struct pmc_owner *po)
636 {
637 	int error;
638 	struct pmclog_buffer *lb;
639 
640 	PMCDBG(LOG,CFG,1, "de-config po=%p", po);
641 
642 	if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
643 		return (EINVAL);
644 
645 	KASSERT(po->po_sscount == 0,
646 	    ("[pmclog,%d] po=%p still owning SS PMCs", __LINE__, po));
647 	KASSERT(po->po_file != NULL,
648 	    ("[pmclog,%d] po=%p no log file", __LINE__, po));
649 
650 	/* stop the kthread, this will reset the 'OWNS_LOGFILE' flag */
651 	if (po->po_kthread)
652 		pmclog_stop_kthread(po);
653 
654 	KASSERT(po->po_kthread == NULL,
655 	    ("[pmclog,%d] po=%p kthread not stopped", __LINE__, po));
656 
657 	/* return all queued log buffers to the global pool */
658 	while ((lb = TAILQ_FIRST(&po->po_logbuffers)) != NULL) {
659 		TAILQ_REMOVE(&po->po_logbuffers, lb, plb_next);
660 		PMCLOG_INIT_BUFFER_DESCRIPTOR(lb);
661 		mtx_lock_spin(&pmc_bufferlist_mtx);
662 		TAILQ_INSERT_HEAD(&pmc_bufferlist, lb, plb_next);
663 		mtx_unlock_spin(&pmc_bufferlist_mtx);
664 	}
665 
666 	/* return the 'current' buffer to the global pool */
667 	if ((lb = po->po_curbuf) != NULL) {
668 		PMCLOG_INIT_BUFFER_DESCRIPTOR(lb);
669 		mtx_lock_spin(&pmc_bufferlist_mtx);
670 		TAILQ_INSERT_HEAD(&pmc_bufferlist, lb, plb_next);
671 		mtx_unlock_spin(&pmc_bufferlist_mtx);
672 	}
673 
674 	/* drop a reference to the fd */
675 	error = fdrop(po->po_file, curthread);
676 	po->po_file  = NULL;
677 	po->po_error = 0;
678 
679 	return (error);
680 }
681 
682 /*
683  * Flush a process' log buffer.
684  */
685 
686 int
687 pmclog_flush(struct pmc_owner *po)
688 {
689 	int error, has_pending_buffers;
690 
691 	PMCDBG(LOG,FLS,1, "po=%p", po);
692 
693 	/*
694 	 * If there is a pending error recorded by the logger thread,
695 	 * return that.
696 	 */
697 	if (po->po_error)
698 		return (po->po_error);
699 
700 	error = 0;
701 
702 	/*
703 	 * Check that we do have an active log file.
704 	 */
705 	mtx_lock(&pmc_kthread_mtx);
706 	if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) {
707 		error = EINVAL;
708 		goto error;
709 	}
710 
711 	/*
712 	 * Schedule the current buffer if any.
713 	 */
714 	mtx_lock_spin(&po->po_mtx);
715 	if (po->po_curbuf)
716 		pmclog_schedule_io(po);
717 	has_pending_buffers = !TAILQ_EMPTY(&po->po_logbuffers);
718 	mtx_unlock_spin(&po->po_mtx);
719 
720 	if (has_pending_buffers) {
721 		po->po_flags |= PMC_PO_IN_FLUSH; /* ask for a wakeup */
722 		error = msleep(po->po_kthread, &pmc_kthread_mtx, PWAIT,
723 		    "pmcflush", 0);
724 		if (error == 0)
725 			error = po->po_error;
726 	}
727 
728  error:
729 	mtx_unlock(&pmc_kthread_mtx);
730 
731 	return (error);
732 }
733 
734 
735 void
736 pmclog_process_callchain(struct pmc *pm, struct pmc_sample *ps)
737 {
738 	int n, recordlen;
739 	uint32_t flags;
740 	struct pmc_owner *po;
741 
742 	PMCDBG(LOG,SAM,1,"pm=%p pid=%d n=%d", pm, ps->ps_pid,
743 	    ps->ps_nsamples);
744 
745 	recordlen = offsetof(struct pmclog_callchain, pl_pc) +
746 	    ps->ps_nsamples * sizeof(uintfptr_t);
747 	po = pm->pm_owner;
748 	flags = PMC_CALLCHAIN_TO_CPUFLAGS(ps->ps_cpu,ps->ps_flags);
749 	PMCLOG_RESERVE(po, CALLCHAIN, recordlen);
750 	PMCLOG_EMIT32(ps->ps_pid);
751 	PMCLOG_EMIT32(pm->pm_id);
752 	PMCLOG_EMIT32(flags);
753 	for (n = 0; n < ps->ps_nsamples; n++)
754 		PMCLOG_EMITADDR(ps->ps_pc[n]);
755 	PMCLOG_DESPATCH(po);
756 }
757 
758 void
759 pmclog_process_closelog(struct pmc_owner *po)
760 {
761 	PMCLOG_RESERVE(po,CLOSELOG,sizeof(struct pmclog_closelog));
762 	PMCLOG_DESPATCH(po);
763 }
764 
765 void
766 pmclog_process_dropnotify(struct pmc_owner *po)
767 {
768 	PMCLOG_RESERVE(po,DROPNOTIFY,sizeof(struct pmclog_dropnotify));
769 	PMCLOG_DESPATCH(po);
770 }
771 
772 void
773 pmclog_process_map_in(struct pmc_owner *po, pid_t pid, uintfptr_t start,
774     const char *path)
775 {
776 	int pathlen, recordlen;
777 
778 	KASSERT(path != NULL, ("[pmclog,%d] map-in, null path", __LINE__));
779 
780 	pathlen = strlen(path) + 1;	/* #bytes for path name */
781 	recordlen = offsetof(struct pmclog_map_in, pl_pathname) +
782 	    pathlen;
783 
784 	PMCLOG_RESERVE(po, MAP_IN, recordlen);
785 	PMCLOG_EMIT32(pid);
786 	PMCLOG_EMITADDR(start);
787 	PMCLOG_EMITSTRING(path,pathlen);
788 	PMCLOG_DESPATCH(po);
789 }
790 
791 void
792 pmclog_process_map_out(struct pmc_owner *po, pid_t pid, uintfptr_t start,
793     uintfptr_t end)
794 {
795 	KASSERT(start <= end, ("[pmclog,%d] start > end", __LINE__));
796 
797 	PMCLOG_RESERVE(po, MAP_OUT, sizeof(struct pmclog_map_out));
798 	PMCLOG_EMIT32(pid);
799 	PMCLOG_EMITADDR(start);
800 	PMCLOG_EMITADDR(end);
801 	PMCLOG_DESPATCH(po);
802 }
803 
804 void
805 pmclog_process_pmcallocate(struct pmc *pm)
806 {
807 	struct pmc_owner *po;
808 
809 	po = pm->pm_owner;
810 
811 	PMCDBG(LOG,ALL,1, "pm=%p", pm);
812 
813 	PMCLOG_RESERVE(po, PMCALLOCATE, sizeof(struct pmclog_pmcallocate));
814 	PMCLOG_EMIT32(pm->pm_id);
815 	PMCLOG_EMIT32(pm->pm_event);
816 	PMCLOG_EMIT32(pm->pm_flags);
817 	PMCLOG_DESPATCH(po);
818 }
819 
820 void
821 pmclog_process_pmcattach(struct pmc *pm, pid_t pid, char *path)
822 {
823 	int pathlen, recordlen;
824 	struct pmc_owner *po;
825 
826 	PMCDBG(LOG,ATT,1,"pm=%p pid=%d", pm, pid);
827 
828 	po = pm->pm_owner;
829 
830 	pathlen = strlen(path) + 1;	/* #bytes for the string */
831 	recordlen = offsetof(struct pmclog_pmcattach, pl_pathname) + pathlen;
832 
833 	PMCLOG_RESERVE(po, PMCATTACH, recordlen);
834 	PMCLOG_EMIT32(pm->pm_id);
835 	PMCLOG_EMIT32(pid);
836 	PMCLOG_EMITSTRING(path, pathlen);
837 	PMCLOG_DESPATCH(po);
838 }
839 
840 void
841 pmclog_process_pmcdetach(struct pmc *pm, pid_t pid)
842 {
843 	struct pmc_owner *po;
844 
845 	PMCDBG(LOG,ATT,1,"!pm=%p pid=%d", pm, pid);
846 
847 	po = pm->pm_owner;
848 
849 	PMCLOG_RESERVE(po, PMCDETACH, sizeof(struct pmclog_pmcdetach));
850 	PMCLOG_EMIT32(pm->pm_id);
851 	PMCLOG_EMIT32(pid);
852 	PMCLOG_DESPATCH(po);
853 }
854 
855 /*
856  * Log a context switch event to the log file.
857  */
858 
859 void
860 pmclog_process_proccsw(struct pmc *pm, struct pmc_process *pp, pmc_value_t v)
861 {
862 	struct pmc_owner *po;
863 
864 	KASSERT(pm->pm_flags & PMC_F_LOG_PROCCSW,
865 	    ("[pmclog,%d] log-process-csw called gratuitously", __LINE__));
866 
867 	PMCDBG(LOG,SWO,1,"pm=%p pid=%d v=%jx", pm, pp->pp_proc->p_pid,
868 	    v);
869 
870 	po = pm->pm_owner;
871 
872 	PMCLOG_RESERVE(po, PROCCSW, sizeof(struct pmclog_proccsw));
873 	PMCLOG_EMIT32(pm->pm_id);
874 	PMCLOG_EMIT64(v);
875 	PMCLOG_EMIT32(pp->pp_proc->p_pid);
876 	PMCLOG_DESPATCH(po);
877 }
878 
879 void
880 pmclog_process_procexec(struct pmc_owner *po, pmc_id_t pmid, pid_t pid,
881     uintfptr_t startaddr, char *path)
882 {
883 	int pathlen, recordlen;
884 
885 	PMCDBG(LOG,EXC,1,"po=%p pid=%d path=\"%s\"", po, pid, path);
886 
887 	pathlen   = strlen(path) + 1;	/* #bytes for the path */
888 	recordlen = offsetof(struct pmclog_procexec, pl_pathname) + pathlen;
889 
890 	PMCLOG_RESERVE(po, PROCEXEC, recordlen);
891 	PMCLOG_EMIT32(pid);
892 	PMCLOG_EMITADDR(startaddr);
893 	PMCLOG_EMIT32(pmid);
894 	PMCLOG_EMITSTRING(path,pathlen);
895 	PMCLOG_DESPATCH(po);
896 }
897 
898 /*
899  * Log a process exit event (and accumulated pmc value) to the log file.
900  */
901 
902 void
903 pmclog_process_procexit(struct pmc *pm, struct pmc_process *pp)
904 {
905 	int ri;
906 	struct pmc_owner *po;
907 
908 	ri = PMC_TO_ROWINDEX(pm);
909 	PMCDBG(LOG,EXT,1,"pm=%p pid=%d v=%jx", pm, pp->pp_proc->p_pid,
910 	    pp->pp_pmcs[ri].pp_pmcval);
911 
912 	po = pm->pm_owner;
913 
914 	PMCLOG_RESERVE(po, PROCEXIT, sizeof(struct pmclog_procexit));
915 	PMCLOG_EMIT32(pm->pm_id);
916 	PMCLOG_EMIT64(pp->pp_pmcs[ri].pp_pmcval);
917 	PMCLOG_EMIT32(pp->pp_proc->p_pid);
918 	PMCLOG_DESPATCH(po);
919 }
920 
921 /*
922  * Log a fork event.
923  */
924 
925 void
926 pmclog_process_procfork(struct pmc_owner *po, pid_t oldpid, pid_t newpid)
927 {
928 	PMCLOG_RESERVE(po, PROCFORK, sizeof(struct pmclog_procfork));
929 	PMCLOG_EMIT32(oldpid);
930 	PMCLOG_EMIT32(newpid);
931 	PMCLOG_DESPATCH(po);
932 }
933 
934 /*
935  * Log a process exit event of the form suitable for system-wide PMCs.
936  */
937 
938 void
939 pmclog_process_sysexit(struct pmc_owner *po, pid_t pid)
940 {
941 	PMCLOG_RESERVE(po, SYSEXIT, sizeof(struct pmclog_sysexit));
942 	PMCLOG_EMIT32(pid);
943 	PMCLOG_DESPATCH(po);
944 }
945 
946 /*
947  * Write a user log entry.
948  */
949 
950 int
951 pmclog_process_userlog(struct pmc_owner *po, struct pmc_op_writelog *wl)
952 {
953 	int error;
954 
955 	PMCDBG(LOG,WRI,1, "writelog po=%p ud=0x%x", po, wl->pm_userdata);
956 
957 	error = 0;
958 
959 	PMCLOG_RESERVE_WITH_ERROR(po, USERDATA,
960 	    sizeof(struct pmclog_userdata));
961 	PMCLOG_EMIT32(wl->pm_userdata);
962 	PMCLOG_DESPATCH(po);
963 
964  error:
965 	return (error);
966 }
967 
968 /*
969  * Initialization.
970  *
971  * Create a pool of log buffers and initialize mutexes.
972  */
973 
974 void
975 pmclog_initialize()
976 {
977 	int n;
978 	struct pmclog_buffer *plb;
979 
980 	if (pmclog_buffer_size <= 0) {
981 		(void) printf("hwpmc: tunable logbuffersize=%d must be "
982 		    "greater than zero.\n", pmclog_buffer_size);
983 		pmclog_buffer_size = PMC_LOG_BUFFER_SIZE;
984 	}
985 
986 	if (pmc_nlogbuffers <= 0) {
987 		(void) printf("hwpmc: tunable nlogbuffers=%d must be greater "
988 		    "than zero.\n", pmc_nlogbuffers);
989 		pmc_nlogbuffers = PMC_NLOGBUFFERS;
990 	}
991 
992 	/* create global pool of log buffers */
993 	for (n = 0; n < pmc_nlogbuffers; n++) {
994 		plb = malloc(1024 * pmclog_buffer_size, M_PMC,
995 		    M_WAITOK|M_ZERO);
996 		PMCLOG_INIT_BUFFER_DESCRIPTOR(plb);
997 		TAILQ_INSERT_HEAD(&pmc_bufferlist, plb, plb_next);
998 	}
999 	mtx_init(&pmc_bufferlist_mtx, "pmc-buffer-list", "pmc-leaf",
1000 	    MTX_SPIN);
1001 	mtx_init(&pmc_kthread_mtx, "pmc-kthread", "pmc-sleep", MTX_DEF);
1002 }
1003 
1004 /*
1005  * Shutdown logging.
1006  *
1007  * Destroy mutexes and release memory back the to free pool.
1008  */
1009 
1010 void
1011 pmclog_shutdown()
1012 {
1013 	struct pmclog_buffer *plb;
1014 
1015 	mtx_destroy(&pmc_kthread_mtx);
1016 	mtx_destroy(&pmc_bufferlist_mtx);
1017 
1018 	while ((plb = TAILQ_FIRST(&pmc_bufferlist)) != NULL) {
1019 		TAILQ_REMOVE(&pmc_bufferlist, plb, plb_next);
1020 		free(plb, M_PMC);
1021 	}
1022 }
1023