xref: /freebsd/sys/dev/hwpmc/hwpmc_logging.c (revision 6186fd1857626de0f7cb1a9e4dff19082f9ebb11)
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/capsicum.h>
41 #include <sys/file.h>
42 #include <sys/kernel.h>
43 #include <sys/kthread.h>
44 #include <sys/lock.h>
45 #include <sys/module.h>
46 #include <sys/mutex.h>
47 #include <sys/pmc.h>
48 #include <sys/pmckern.h>
49 #include <sys/pmclog.h>
50 #include <sys/proc.h>
51 #include <sys/signalvar.h>
52 #include <sys/sysctl.h>
53 #include <sys/systm.h>
54 #include <sys/uio.h>
55 #include <sys/unistd.h>
56 #include <sys/vnode.h>
57 
58 /*
59  * Sysctl tunables
60  */
61 
62 SYSCTL_DECL(_kern_hwpmc);
63 
64 /*
65  * kern.hwpmc.logbuffersize -- size of the per-cpu owner buffers.
66  */
67 
68 static int pmclog_buffer_size = PMC_LOG_BUFFER_SIZE;
69 SYSCTL_INT(_kern_hwpmc, OID_AUTO, logbuffersize, CTLFLAG_RDTUN,
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 SYSCTL_INT(_kern_hwpmc, OID_AUTO, nbuffers, CTLFLAG_RDTUN,
78     &pmc_nlogbuffers, 0, "number of global log buffers");
79 
80 /*
81  * Global log buffer list and associated spin lock.
82  */
83 
84 TAILQ_HEAD(, pmclog_buffer) pmc_bufferlist =
85 	TAILQ_HEAD_INITIALIZER(pmc_bufferlist);
86 static struct mtx pmc_bufferlist_mtx;	/* spin lock */
87 static struct mtx pmc_kthread_mtx;	/* sleep lock */
88 
89 #define	PMCLOG_INIT_BUFFER_DESCRIPTOR(D) do {				\
90 		const int __roundup = roundup(sizeof(*D),		\
91 			sizeof(uint32_t));				\
92 		(D)->plb_fence = ((char *) (D)) +			\
93 			 1024*pmclog_buffer_size;			\
94 		(D)->plb_base  = (D)->plb_ptr = ((char *) (D)) +	\
95 			__roundup;					\
96 	} while (0)
97 
98 
99 /*
100  * Log file record constructors.
101  */
102 #define	_PMCLOG_TO_HEADER(T,L)						\
103 	((PMCLOG_HEADER_MAGIC << 24) |					\
104 	 (PMCLOG_TYPE_ ## T << 16)   |					\
105 	 ((L) & 0xFFFF))
106 
107 /* reserve LEN bytes of space and initialize the entry header */
108 #define	_PMCLOG_RESERVE(PO,TYPE,LEN,ACTION) do {			\
109 		uint32_t *_le;						\
110 		int _len = roundup((LEN), sizeof(uint32_t));		\
111 		if ((_le = pmclog_reserve((PO), _len)) == NULL) {	\
112 			ACTION;						\
113 		}							\
114 		*_le = _PMCLOG_TO_HEADER(TYPE,_len);			\
115 		_le += 3	/* skip over timestamp */
116 
117 #define	PMCLOG_RESERVE(P,T,L)		_PMCLOG_RESERVE(P,T,L,return)
118 #define	PMCLOG_RESERVE_WITH_ERROR(P,T,L) _PMCLOG_RESERVE(P,T,L,		\
119 	error=ENOMEM;goto error)
120 
121 #define	PMCLOG_EMIT32(V)	do { *_le++ = (V); } while (0)
122 #define	PMCLOG_EMIT64(V)	do { 					\
123 		*_le++ = (uint32_t) ((V) & 0xFFFFFFFF);			\
124 		*_le++ = (uint32_t) (((V) >> 32) & 0xFFFFFFFF);		\
125 	} while (0)
126 
127 
128 /* Emit a string.  Caution: does NOT update _le, so needs to be last */
129 #define	PMCLOG_EMITSTRING(S,L)	do { bcopy((S), _le, (L)); } while (0)
130 #define	PMCLOG_EMITNULLSTRING(L) do { bzero(_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 				/* No more buffers and shutdown required. */
288 				if (po->po_flags & PMC_PO_SHUTDOWN) {
289 					mtx_unlock(&pmc_kthread_mtx);
290 					/*
291 			 		 * Close the file to get PMCLOG_EOF
292 					 * error in pmclog(3).
293 					 */
294 					fo_close(po->po_file, curthread);
295 					mtx_lock(&pmc_kthread_mtx);
296 					break;
297 				}
298 
299 				(void) msleep(po, &pmc_kthread_mtx, PWAIT,
300 				    "pmcloop", 0);
301 				continue;
302 			}
303 
304 			TAILQ_REMOVE(&po->po_logbuffers, lb, plb_next);
305 			mtx_unlock_spin(&po->po_mtx);
306 		}
307 
308 		mtx_unlock(&pmc_kthread_mtx);
309 
310 		/* process the request */
311 		PMCDBG(LOG,WRI,2, "po=%p base=%p ptr=%p", po,
312 		    lb->plb_base, lb->plb_ptr);
313 		/* change our thread's credentials before issuing the I/O */
314 
315 		aiov.iov_base = lb->plb_base;
316 		aiov.iov_len  = nbytes = lb->plb_ptr - lb->plb_base;
317 
318 		auio.uio_iov    = &aiov;
319 		auio.uio_iovcnt = 1;
320 		auio.uio_offset = -1;
321 		auio.uio_resid  = nbytes;
322 		auio.uio_rw     = UIO_WRITE;
323 		auio.uio_segflg = UIO_SYSSPACE;
324 		auio.uio_td     = td;
325 
326 		/* switch thread credentials -- see kern_ktrace.c */
327 		td->td_ucred = ownercred;
328 		error = fo_write(po->po_file, &auio, ownercred, 0, td);
329 		td->td_ucred = mycred;
330 
331 		if (error) {
332 			/* XXX some errors are recoverable */
333 			/* send a SIGIO to the owner and exit */
334 			PROC_LOCK(p);
335 			kern_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 	wakeup_one(po->po_kthread);
360 	po->po_kthread = NULL;
361 
362 	mtx_unlock(&pmc_kthread_mtx);
363 
364 	/* return the current I/O buffer to the global pool */
365 	if (lb) {
366 		PMCLOG_INIT_BUFFER_DESCRIPTOR(lb);
367 
368 		mtx_lock_spin(&pmc_bufferlist_mtx);
369 		TAILQ_INSERT_HEAD(&pmc_bufferlist, lb, plb_next);
370 		mtx_unlock_spin(&pmc_bufferlist_mtx);
371 	}
372 
373 	/*
374 	 * Exit this thread, signalling the waiter
375 	 */
376 
377 	crfree(ownercred);
378 
379 	kproc_exit(0);
380 }
381 
382 /*
383  * Release and log entry and schedule an I/O if needed.
384  */
385 
386 static void
387 pmclog_release(struct pmc_owner *po)
388 {
389 	KASSERT(po->po_curbuf->plb_ptr >= po->po_curbuf->plb_base,
390 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p base=%p", __LINE__,
391 		po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_base));
392 	KASSERT(po->po_curbuf->plb_ptr <= po->po_curbuf->plb_fence,
393 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p fenc=%p", __LINE__,
394 		po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_fence));
395 
396 	/* schedule an I/O if we've filled a buffer */
397 	if (po->po_curbuf->plb_ptr >= po->po_curbuf->plb_fence)
398 		pmclog_schedule_io(po);
399 
400 	mtx_unlock_spin(&po->po_mtx);
401 
402 	PMCDBG(LOG,REL,1, "po=%p", po);
403 }
404 
405 
406 /*
407  * Attempt to reserve 'length' bytes of space in an owner's log
408  * buffer.  The function returns a pointer to 'length' bytes of space
409  * if there was enough space or returns NULL if no space was
410  * available.  Non-null returns do so with the po mutex locked.  The
411  * caller must invoke pmclog_release() on the pmc owner structure
412  * when done.
413  */
414 
415 static uint32_t *
416 pmclog_reserve(struct pmc_owner *po, int length)
417 {
418 	uintptr_t newptr, oldptr;
419 	uint32_t *lh;
420 	struct timespec ts;
421 
422 	PMCDBG(LOG,ALL,1, "po=%p len=%d", po, length);
423 
424 	KASSERT(length % sizeof(uint32_t) == 0,
425 	    ("[pmclog,%d] length not a multiple of word size", __LINE__));
426 
427 	mtx_lock_spin(&po->po_mtx);
428 
429 	/* No more data when shutdown in progress. */
430 	if (po->po_flags & PMC_PO_SHUTDOWN) {
431 		mtx_unlock_spin(&po->po_mtx);
432 		return (NULL);
433 	}
434 
435 	if (po->po_curbuf == NULL)
436 		if (pmclog_get_buffer(po) != 0) {
437 			mtx_unlock_spin(&po->po_mtx);
438 			return (NULL);
439 		}
440 
441 	KASSERT(po->po_curbuf != NULL,
442 	    ("[pmclog,%d] po=%p no current buffer", __LINE__, po));
443 
444 	KASSERT(po->po_curbuf->plb_ptr >= po->po_curbuf->plb_base &&
445 	    po->po_curbuf->plb_ptr <= po->po_curbuf->plb_fence,
446 	    ("[pmclog,%d] po=%p buffer invariants: ptr=%p base=%p fence=%p",
447 		__LINE__, po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_base,
448 		po->po_curbuf->plb_fence));
449 
450 	oldptr = (uintptr_t) po->po_curbuf->plb_ptr;
451 	newptr = oldptr + length;
452 
453 	KASSERT(oldptr != (uintptr_t) NULL,
454 	    ("[pmclog,%d] po=%p Null log buffer pointer", __LINE__, po));
455 
456 	/*
457 	 * If we have space in the current buffer, return a pointer to
458 	 * available space with the PO structure locked.
459 	 */
460 	if (newptr <= (uintptr_t) po->po_curbuf->plb_fence) {
461 		po->po_curbuf->plb_ptr = (char *) newptr;
462 		goto done;
463 	}
464 
465 	/*
466 	 * Otherwise, schedule the current buffer for output and get a
467 	 * fresh buffer.
468 	 */
469 	pmclog_schedule_io(po);
470 
471 	if (pmclog_get_buffer(po) != 0) {
472 		mtx_unlock_spin(&po->po_mtx);
473 		return (NULL);
474 	}
475 
476 	KASSERT(po->po_curbuf != NULL,
477 	    ("[pmclog,%d] po=%p no current buffer", __LINE__, po));
478 
479 	KASSERT(po->po_curbuf->plb_ptr != NULL,
480 	    ("[pmclog,%d] null return from pmc_get_log_buffer", __LINE__));
481 
482 	KASSERT(po->po_curbuf->plb_ptr == po->po_curbuf->plb_base &&
483 	    po->po_curbuf->plb_ptr <= po->po_curbuf->plb_fence,
484 	    ("[pmclog,%d] po=%p buffer invariants: ptr=%p base=%p fence=%p",
485 		__LINE__, po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_base,
486 		po->po_curbuf->plb_fence));
487 
488 	oldptr = (uintptr_t) po->po_curbuf->plb_ptr;
489 
490  done:
491 	lh = (uint32_t *) oldptr;
492 	lh++;				/* skip header */
493 	getnanotime(&ts);		/* fill in the timestamp */
494 	*lh++ = ts.tv_sec & 0xFFFFFFFF;
495 	*lh++ = ts.tv_nsec & 0xFFFFFFF;
496 	return ((uint32_t *) oldptr);
497 }
498 
499 /*
500  * Schedule an I/O.
501  *
502  * Transfer the current buffer to the helper kthread.
503  */
504 
505 static void
506 pmclog_schedule_io(struct pmc_owner *po)
507 {
508 	KASSERT(po->po_curbuf != NULL,
509 	    ("[pmclog,%d] schedule_io with null buffer po=%p", __LINE__, po));
510 
511 	KASSERT(po->po_curbuf->plb_ptr >= po->po_curbuf->plb_base,
512 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p base=%p", __LINE__,
513 		po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_base));
514 	KASSERT(po->po_curbuf->plb_ptr <= po->po_curbuf->plb_fence,
515 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p fenc=%p", __LINE__,
516 		po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_fence));
517 
518 	PMCDBG(LOG,SIO, 1, "po=%p", po);
519 
520 	mtx_assert(&po->po_mtx, MA_OWNED);
521 
522 	/*
523 	 * Add the current buffer to the tail of the buffer list and
524 	 * wakeup the helper.
525 	 */
526 	TAILQ_INSERT_TAIL(&po->po_logbuffers, po->po_curbuf, plb_next);
527 	po->po_curbuf = NULL;
528 	wakeup_one(po);
529 }
530 
531 /*
532  * Stop the helper kthread.
533  */
534 
535 static void
536 pmclog_stop_kthread(struct pmc_owner *po)
537 {
538 	/*
539 	 * Close the file to force the thread out of fo_write,
540 	 * unset flag, wakeup the helper thread,
541 	 * wait for it to exit
542 	 */
543 
544 	if (po->po_file != NULL)
545 		fo_close(po->po_file, curthread);
546 
547 	mtx_lock(&pmc_kthread_mtx);
548 	po->po_flags &= ~PMC_PO_OWNS_LOGFILE;
549 	wakeup_one(po);
550 	if (po->po_kthread)
551 		msleep(po->po_kthread, &pmc_kthread_mtx, PPAUSE, "pmckstp", 0);
552 	mtx_unlock(&pmc_kthread_mtx);
553 }
554 
555 /*
556  * Public functions
557  */
558 
559 /*
560  * Configure a log file for pmc owner 'po'.
561  *
562  * Parameter 'logfd' is a file handle referencing an open file in the
563  * owner process.  This file needs to have been opened for writing.
564  */
565 
566 int
567 pmclog_configure_log(struct pmc_mdep *md, struct pmc_owner *po, int logfd)
568 {
569 	int error;
570 	struct proc *p;
571 	cap_rights_t rights;
572 
573 	/*
574 	 * As long as it is possible to get a LOR between pmc_sx lock and
575 	 * proctree/allproc sx locks used for adding a new process, assure
576 	 * the former is not held here.
577 	 */
578 	sx_assert(&pmc_sx, SA_UNLOCKED);
579 	PMCDBG(LOG,CFG,1, "config po=%p logfd=%d", po, logfd);
580 
581 	p = po->po_owner;
582 
583 	/* return EBUSY if a log file was already present */
584 	if (po->po_flags & PMC_PO_OWNS_LOGFILE)
585 		return (EBUSY);
586 
587 	KASSERT(po->po_kthread == NULL,
588 	    ("[pmclog,%d] po=%p kthread (%p) already present", __LINE__, po,
589 		po->po_kthread));
590 	KASSERT(po->po_file == NULL,
591 	    ("[pmclog,%d] po=%p file (%p) already present", __LINE__, po,
592 		po->po_file));
593 
594 	/* get a reference to the file state */
595 	error = fget_write(curthread, logfd,
596 	    cap_rights_init(&rights, CAP_WRITE), &po->po_file);
597 	if (error)
598 		goto error;
599 
600 	/* mark process as owning a log file */
601 	po->po_flags |= PMC_PO_OWNS_LOGFILE;
602 	error = kproc_create(pmclog_loop, po, &po->po_kthread,
603 	    RFHIGHPID, 0, "hwpmc: proc(%d)", p->p_pid);
604 	if (error)
605 		goto error;
606 
607 	/* mark process as using HWPMCs */
608 	PROC_LOCK(p);
609 	p->p_flag |= P_HWPMC;
610 	PROC_UNLOCK(p);
611 
612 	/* create a log initialization entry */
613 	PMCLOG_RESERVE_WITH_ERROR(po, INITIALIZE,
614 	    sizeof(struct pmclog_initialize));
615 	PMCLOG_EMIT32(PMC_VERSION);
616 	PMCLOG_EMIT32(md->pmd_cputype);
617 	PMCLOG_DESPATCH(po);
618 
619 	return (0);
620 
621  error:
622 	/* shutdown the thread */
623 	if (po->po_kthread)
624 		pmclog_stop_kthread(po);
625 
626 	KASSERT(po->po_kthread == NULL, ("[pmclog,%d] po=%p kthread not "
627 	    "stopped", __LINE__, po));
628 
629 	if (po->po_file)
630 		(void) fdrop(po->po_file, curthread);
631 	po->po_file  = NULL;	/* clear file and error state */
632 	po->po_error = 0;
633 
634 	return (error);
635 }
636 
637 
638 /*
639  * De-configure a log file.  This will throw away any buffers queued
640  * for this owner process.
641  */
642 
643 int
644 pmclog_deconfigure_log(struct pmc_owner *po)
645 {
646 	int error;
647 	struct pmclog_buffer *lb;
648 
649 	PMCDBG(LOG,CFG,1, "de-config po=%p", po);
650 
651 	if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
652 		return (EINVAL);
653 
654 	KASSERT(po->po_sscount == 0,
655 	    ("[pmclog,%d] po=%p still owning SS PMCs", __LINE__, po));
656 	KASSERT(po->po_file != NULL,
657 	    ("[pmclog,%d] po=%p no log file", __LINE__, po));
658 
659 	/* stop the kthread, this will reset the 'OWNS_LOGFILE' flag */
660 	pmclog_stop_kthread(po);
661 
662 	KASSERT(po->po_kthread == NULL,
663 	    ("[pmclog,%d] po=%p kthread not stopped", __LINE__, po));
664 
665 	/* return all queued log buffers to the global pool */
666 	while ((lb = TAILQ_FIRST(&po->po_logbuffers)) != NULL) {
667 		TAILQ_REMOVE(&po->po_logbuffers, lb, plb_next);
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 	/* return the 'current' buffer to the global pool */
675 	if ((lb = po->po_curbuf) != NULL) {
676 		PMCLOG_INIT_BUFFER_DESCRIPTOR(lb);
677 		mtx_lock_spin(&pmc_bufferlist_mtx);
678 		TAILQ_INSERT_HEAD(&pmc_bufferlist, lb, plb_next);
679 		mtx_unlock_spin(&pmc_bufferlist_mtx);
680 	}
681 
682 	/* drop a reference to the fd */
683 	error = fdrop(po->po_file, curthread);
684 	po->po_file  = NULL;
685 	po->po_error = 0;
686 
687 	return (error);
688 }
689 
690 /*
691  * Flush a process' log buffer.
692  */
693 
694 int
695 pmclog_flush(struct pmc_owner *po)
696 {
697 	int error;
698 	struct pmclog_buffer *lb;
699 
700 	PMCDBG(LOG,FLS,1, "po=%p", po);
701 
702 	/*
703 	 * If there is a pending error recorded by the logger thread,
704 	 * return that.
705 	 */
706 	if (po->po_error)
707 		return (po->po_error);
708 
709 	error = 0;
710 
711 	/*
712 	 * Check that we do have an active log file.
713 	 */
714 	mtx_lock(&pmc_kthread_mtx);
715 	if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) {
716 		error = EINVAL;
717 		goto error;
718 	}
719 
720 	/*
721 	 * Schedule the current buffer if any and not empty.
722 	 */
723 	mtx_lock_spin(&po->po_mtx);
724 	lb = po->po_curbuf;
725 	if (lb && lb->plb_ptr != lb->plb_base) {
726 		pmclog_schedule_io(po);
727 	} else
728 		error = ENOBUFS;
729 	mtx_unlock_spin(&po->po_mtx);
730 
731  error:
732 	mtx_unlock(&pmc_kthread_mtx);
733 
734 	return (error);
735 }
736 
737 int
738 pmclog_close(struct pmc_owner *po)
739 {
740 
741 	PMCDBG(LOG,CLO,1, "po=%p", po);
742 
743 	mtx_lock(&pmc_kthread_mtx);
744 
745 	/*
746 	 * Schedule the current buffer.
747 	 */
748 	mtx_lock_spin(&po->po_mtx);
749 	if (po->po_curbuf)
750 		pmclog_schedule_io(po);
751 	else
752 		wakeup_one(po);
753 	mtx_unlock_spin(&po->po_mtx);
754 
755 	/*
756 	 * Initiate shutdown: no new data queued,
757 	 * thread will close file on last block.
758 	 */
759 	po->po_flags |= PMC_PO_SHUTDOWN;
760 
761 	mtx_unlock(&pmc_kthread_mtx);
762 
763 	return (0);
764 }
765 
766 void
767 pmclog_process_callchain(struct pmc *pm, struct pmc_sample *ps)
768 {
769 	int n, recordlen;
770 	uint32_t flags;
771 	struct pmc_owner *po;
772 
773 	PMCDBG(LOG,SAM,1,"pm=%p pid=%d n=%d", pm, ps->ps_pid,
774 	    ps->ps_nsamples);
775 
776 	recordlen = offsetof(struct pmclog_callchain, pl_pc) +
777 	    ps->ps_nsamples * sizeof(uintfptr_t);
778 	po = pm->pm_owner;
779 	flags = PMC_CALLCHAIN_TO_CPUFLAGS(ps->ps_cpu,ps->ps_flags);
780 	PMCLOG_RESERVE(po, CALLCHAIN, recordlen);
781 	PMCLOG_EMIT32(ps->ps_pid);
782 	PMCLOG_EMIT32(pm->pm_id);
783 	PMCLOG_EMIT32(flags);
784 	for (n = 0; n < ps->ps_nsamples; n++)
785 		PMCLOG_EMITADDR(ps->ps_pc[n]);
786 	PMCLOG_DESPATCH(po);
787 }
788 
789 void
790 pmclog_process_closelog(struct pmc_owner *po)
791 {
792 	PMCLOG_RESERVE(po,CLOSELOG,sizeof(struct pmclog_closelog));
793 	PMCLOG_DESPATCH(po);
794 }
795 
796 void
797 pmclog_process_dropnotify(struct pmc_owner *po)
798 {
799 	PMCLOG_RESERVE(po,DROPNOTIFY,sizeof(struct pmclog_dropnotify));
800 	PMCLOG_DESPATCH(po);
801 }
802 
803 void
804 pmclog_process_map_in(struct pmc_owner *po, pid_t pid, uintfptr_t start,
805     const char *path)
806 {
807 	int pathlen, recordlen;
808 
809 	KASSERT(path != NULL, ("[pmclog,%d] map-in, null path", __LINE__));
810 
811 	pathlen = strlen(path) + 1;	/* #bytes for path name */
812 	recordlen = offsetof(struct pmclog_map_in, pl_pathname) +
813 	    pathlen;
814 
815 	PMCLOG_RESERVE(po, MAP_IN, recordlen);
816 	PMCLOG_EMIT32(pid);
817 	PMCLOG_EMITADDR(start);
818 	PMCLOG_EMITSTRING(path,pathlen);
819 	PMCLOG_DESPATCH(po);
820 }
821 
822 void
823 pmclog_process_map_out(struct pmc_owner *po, pid_t pid, uintfptr_t start,
824     uintfptr_t end)
825 {
826 	KASSERT(start <= end, ("[pmclog,%d] start > end", __LINE__));
827 
828 	PMCLOG_RESERVE(po, MAP_OUT, sizeof(struct pmclog_map_out));
829 	PMCLOG_EMIT32(pid);
830 	PMCLOG_EMITADDR(start);
831 	PMCLOG_EMITADDR(end);
832 	PMCLOG_DESPATCH(po);
833 }
834 
835 void
836 pmclog_process_pmcallocate(struct pmc *pm)
837 {
838 	struct pmc_owner *po;
839 	struct pmc_soft *ps;
840 
841 	po = pm->pm_owner;
842 
843 	PMCDBG(LOG,ALL,1, "pm=%p", pm);
844 
845 	if (PMC_TO_CLASS(pm) == PMC_CLASS_SOFT) {
846 		PMCLOG_RESERVE(po, PMCALLOCATEDYN,
847 		    sizeof(struct pmclog_pmcallocatedyn));
848 		PMCLOG_EMIT32(pm->pm_id);
849 		PMCLOG_EMIT32(pm->pm_event);
850 		PMCLOG_EMIT32(pm->pm_flags);
851 		ps = pmc_soft_ev_acquire(pm->pm_event);
852 		if (ps != NULL)
853 			PMCLOG_EMITSTRING(ps->ps_ev.pm_ev_name,PMC_NAME_MAX);
854 		else
855 			PMCLOG_EMITNULLSTRING(PMC_NAME_MAX);
856 		pmc_soft_ev_release(ps);
857 		PMCLOG_DESPATCH(po);
858 	} else {
859 		PMCLOG_RESERVE(po, PMCALLOCATE,
860 		    sizeof(struct pmclog_pmcallocate));
861 		PMCLOG_EMIT32(pm->pm_id);
862 		PMCLOG_EMIT32(pm->pm_event);
863 		PMCLOG_EMIT32(pm->pm_flags);
864 		PMCLOG_DESPATCH(po);
865 	}
866 }
867 
868 void
869 pmclog_process_pmcattach(struct pmc *pm, pid_t pid, char *path)
870 {
871 	int pathlen, recordlen;
872 	struct pmc_owner *po;
873 
874 	PMCDBG(LOG,ATT,1,"pm=%p pid=%d", pm, pid);
875 
876 	po = pm->pm_owner;
877 
878 	pathlen = strlen(path) + 1;	/* #bytes for the string */
879 	recordlen = offsetof(struct pmclog_pmcattach, pl_pathname) + pathlen;
880 
881 	PMCLOG_RESERVE(po, PMCATTACH, recordlen);
882 	PMCLOG_EMIT32(pm->pm_id);
883 	PMCLOG_EMIT32(pid);
884 	PMCLOG_EMITSTRING(path, pathlen);
885 	PMCLOG_DESPATCH(po);
886 }
887 
888 void
889 pmclog_process_pmcdetach(struct pmc *pm, pid_t pid)
890 {
891 	struct pmc_owner *po;
892 
893 	PMCDBG(LOG,ATT,1,"!pm=%p pid=%d", pm, pid);
894 
895 	po = pm->pm_owner;
896 
897 	PMCLOG_RESERVE(po, PMCDETACH, sizeof(struct pmclog_pmcdetach));
898 	PMCLOG_EMIT32(pm->pm_id);
899 	PMCLOG_EMIT32(pid);
900 	PMCLOG_DESPATCH(po);
901 }
902 
903 /*
904  * Log a context switch event to the log file.
905  */
906 
907 void
908 pmclog_process_proccsw(struct pmc *pm, struct pmc_process *pp, pmc_value_t v)
909 {
910 	struct pmc_owner *po;
911 
912 	KASSERT(pm->pm_flags & PMC_F_LOG_PROCCSW,
913 	    ("[pmclog,%d] log-process-csw called gratuitously", __LINE__));
914 
915 	PMCDBG(LOG,SWO,1,"pm=%p pid=%d v=%jx", pm, pp->pp_proc->p_pid,
916 	    v);
917 
918 	po = pm->pm_owner;
919 
920 	PMCLOG_RESERVE(po, PROCCSW, sizeof(struct pmclog_proccsw));
921 	PMCLOG_EMIT32(pm->pm_id);
922 	PMCLOG_EMIT64(v);
923 	PMCLOG_EMIT32(pp->pp_proc->p_pid);
924 	PMCLOG_DESPATCH(po);
925 }
926 
927 void
928 pmclog_process_procexec(struct pmc_owner *po, pmc_id_t pmid, pid_t pid,
929     uintfptr_t startaddr, char *path)
930 {
931 	int pathlen, recordlen;
932 
933 	PMCDBG(LOG,EXC,1,"po=%p pid=%d path=\"%s\"", po, pid, path);
934 
935 	pathlen   = strlen(path) + 1;	/* #bytes for the path */
936 	recordlen = offsetof(struct pmclog_procexec, pl_pathname) + pathlen;
937 
938 	PMCLOG_RESERVE(po, PROCEXEC, recordlen);
939 	PMCLOG_EMIT32(pid);
940 	PMCLOG_EMITADDR(startaddr);
941 	PMCLOG_EMIT32(pmid);
942 	PMCLOG_EMITSTRING(path,pathlen);
943 	PMCLOG_DESPATCH(po);
944 }
945 
946 /*
947  * Log a process exit event (and accumulated pmc value) to the log file.
948  */
949 
950 void
951 pmclog_process_procexit(struct pmc *pm, struct pmc_process *pp)
952 {
953 	int ri;
954 	struct pmc_owner *po;
955 
956 	ri = PMC_TO_ROWINDEX(pm);
957 	PMCDBG(LOG,EXT,1,"pm=%p pid=%d v=%jx", pm, pp->pp_proc->p_pid,
958 	    pp->pp_pmcs[ri].pp_pmcval);
959 
960 	po = pm->pm_owner;
961 
962 	PMCLOG_RESERVE(po, PROCEXIT, sizeof(struct pmclog_procexit));
963 	PMCLOG_EMIT32(pm->pm_id);
964 	PMCLOG_EMIT64(pp->pp_pmcs[ri].pp_pmcval);
965 	PMCLOG_EMIT32(pp->pp_proc->p_pid);
966 	PMCLOG_DESPATCH(po);
967 }
968 
969 /*
970  * Log a fork event.
971  */
972 
973 void
974 pmclog_process_procfork(struct pmc_owner *po, pid_t oldpid, pid_t newpid)
975 {
976 	PMCLOG_RESERVE(po, PROCFORK, sizeof(struct pmclog_procfork));
977 	PMCLOG_EMIT32(oldpid);
978 	PMCLOG_EMIT32(newpid);
979 	PMCLOG_DESPATCH(po);
980 }
981 
982 /*
983  * Log a process exit event of the form suitable for system-wide PMCs.
984  */
985 
986 void
987 pmclog_process_sysexit(struct pmc_owner *po, pid_t pid)
988 {
989 	PMCLOG_RESERVE(po, SYSEXIT, sizeof(struct pmclog_sysexit));
990 	PMCLOG_EMIT32(pid);
991 	PMCLOG_DESPATCH(po);
992 }
993 
994 /*
995  * Write a user log entry.
996  */
997 
998 int
999 pmclog_process_userlog(struct pmc_owner *po, struct pmc_op_writelog *wl)
1000 {
1001 	int error;
1002 
1003 	PMCDBG(LOG,WRI,1, "writelog po=%p ud=0x%x", po, wl->pm_userdata);
1004 
1005 	error = 0;
1006 
1007 	PMCLOG_RESERVE_WITH_ERROR(po, USERDATA,
1008 	    sizeof(struct pmclog_userdata));
1009 	PMCLOG_EMIT32(wl->pm_userdata);
1010 	PMCLOG_DESPATCH(po);
1011 
1012  error:
1013 	return (error);
1014 }
1015 
1016 /*
1017  * Initialization.
1018  *
1019  * Create a pool of log buffers and initialize mutexes.
1020  */
1021 
1022 void
1023 pmclog_initialize()
1024 {
1025 	int n;
1026 	struct pmclog_buffer *plb;
1027 
1028 	if (pmclog_buffer_size <= 0) {
1029 		(void) printf("hwpmc: tunable logbuffersize=%d must be "
1030 		    "greater than zero.\n", pmclog_buffer_size);
1031 		pmclog_buffer_size = PMC_LOG_BUFFER_SIZE;
1032 	}
1033 
1034 	if (pmc_nlogbuffers <= 0) {
1035 		(void) printf("hwpmc: tunable nlogbuffers=%d must be greater "
1036 		    "than zero.\n", pmc_nlogbuffers);
1037 		pmc_nlogbuffers = PMC_NLOGBUFFERS;
1038 	}
1039 
1040 	/* create global pool of log buffers */
1041 	for (n = 0; n < pmc_nlogbuffers; n++) {
1042 		plb = malloc(1024 * pmclog_buffer_size, M_PMC,
1043 		    M_WAITOK|M_ZERO);
1044 		PMCLOG_INIT_BUFFER_DESCRIPTOR(plb);
1045 		TAILQ_INSERT_HEAD(&pmc_bufferlist, plb, plb_next);
1046 	}
1047 	mtx_init(&pmc_bufferlist_mtx, "pmc-buffer-list", "pmc-leaf",
1048 	    MTX_SPIN);
1049 	mtx_init(&pmc_kthread_mtx, "pmc-kthread", "pmc-sleep", MTX_DEF);
1050 }
1051 
1052 /*
1053  * Shutdown logging.
1054  *
1055  * Destroy mutexes and release memory back the to free pool.
1056  */
1057 
1058 void
1059 pmclog_shutdown()
1060 {
1061 	struct pmclog_buffer *plb;
1062 
1063 	mtx_destroy(&pmc_kthread_mtx);
1064 	mtx_destroy(&pmc_bufferlist_mtx);
1065 
1066 	while ((plb = TAILQ_FIRST(&pmc_bufferlist)) != NULL) {
1067 		TAILQ_REMOVE(&pmc_bufferlist, plb, plb_next);
1068 		free(plb, M_PMC);
1069 	}
1070 }
1071