xref: /freebsd/sys/dev/hwpmc/hwpmc_logging.c (revision 907b59d76938e654f0d040a888e8dfca3de1e222)
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 #if (__FreeBSD_version >= 1100000)
41 #include <sys/capsicum.h>
42 #else
43 #include <sys/capability.h>
44 #endif
45 #include <sys/file.h>
46 #include <sys/kernel.h>
47 #include <sys/kthread.h>
48 #include <sys/lock.h>
49 #include <sys/module.h>
50 #include <sys/mutex.h>
51 #include <sys/pmc.h>
52 #include <sys/pmckern.h>
53 #include <sys/pmclog.h>
54 #include <sys/proc.h>
55 #include <sys/signalvar.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 /*
63  * Sysctl tunables
64  */
65 
66 SYSCTL_DECL(_kern_hwpmc);
67 
68 /*
69  * kern.hwpmc.logbuffersize -- size of the per-cpu owner buffers.
70  */
71 
72 static int pmclog_buffer_size = PMC_LOG_BUFFER_SIZE;
73 #if (__FreeBSD_version < 1100000)
74 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "logbuffersize", &pmclog_buffer_size);
75 #endif
76 SYSCTL_INT(_kern_hwpmc, OID_AUTO, logbuffersize, CTLFLAG_RDTUN,
77     &pmclog_buffer_size, 0, "size of log buffers in kilobytes");
78 
79 /*
80  * kern.hwpmc.nbuffer -- number of global log buffers
81  */
82 
83 static int pmc_nlogbuffers = PMC_NLOGBUFFERS;
84 #if (__FreeBSD_version < 1100000)
85 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "nbuffers", &pmc_nlogbuffers);
86 #endif
87 SYSCTL_INT(_kern_hwpmc, OID_AUTO, nbuffers, CTLFLAG_RDTUN,
88     &pmc_nlogbuffers, 0, "number of global log buffers");
89 
90 /*
91  * Global log buffer list and associated spin lock.
92  */
93 
94 TAILQ_HEAD(, pmclog_buffer) pmc_bufferlist =
95 	TAILQ_HEAD_INITIALIZER(pmc_bufferlist);
96 static struct mtx pmc_bufferlist_mtx;	/* spin lock */
97 static struct mtx pmc_kthread_mtx;	/* sleep lock */
98 
99 #define	PMCLOG_INIT_BUFFER_DESCRIPTOR(D) do {				\
100 		const int __roundup = roundup(sizeof(*D),		\
101 			sizeof(uint32_t));				\
102 		(D)->plb_fence = ((char *) (D)) +			\
103 			 1024*pmclog_buffer_size;			\
104 		(D)->plb_base  = (D)->plb_ptr = ((char *) (D)) +	\
105 			__roundup;					\
106 	} while (0)
107 
108 
109 /*
110  * Log file record constructors.
111  */
112 #define	_PMCLOG_TO_HEADER(T,L)						\
113 	((PMCLOG_HEADER_MAGIC << 24) |					\
114 	 (PMCLOG_TYPE_ ## T << 16)   |					\
115 	 ((L) & 0xFFFF))
116 
117 /* reserve LEN bytes of space and initialize the entry header */
118 #define	_PMCLOG_RESERVE(PO,TYPE,LEN,ACTION) do {			\
119 		uint32_t *_le;						\
120 		int _len = roundup((LEN), sizeof(uint32_t));		\
121 		if ((_le = pmclog_reserve((PO), _len)) == NULL) {	\
122 			ACTION;						\
123 		}							\
124 		*_le = _PMCLOG_TO_HEADER(TYPE,_len);			\
125 		_le += 3	/* skip over timestamp */
126 
127 #define	PMCLOG_RESERVE(P,T,L)		_PMCLOG_RESERVE(P,T,L,return)
128 #define	PMCLOG_RESERVE_WITH_ERROR(P,T,L) _PMCLOG_RESERVE(P,T,L,		\
129 	error=ENOMEM;goto error)
130 
131 #define	PMCLOG_EMIT32(V)	do { *_le++ = (V); } while (0)
132 #define	PMCLOG_EMIT64(V)	do { 					\
133 		*_le++ = (uint32_t) ((V) & 0xFFFFFFFF);			\
134 		*_le++ = (uint32_t) (((V) >> 32) & 0xFFFFFFFF);		\
135 	} while (0)
136 
137 
138 /* Emit a string.  Caution: does NOT update _le, so needs to be last */
139 #define	PMCLOG_EMITSTRING(S,L)	do { bcopy((S), _le, (L)); } while (0)
140 #define	PMCLOG_EMITNULLSTRING(L) do { bzero(_le, (L)); } while (0)
141 
142 #define	PMCLOG_DESPATCH(PO)						\
143 		pmclog_release((PO));					\
144 	} while (0)
145 
146 
147 /*
148  * Assertions about the log file format.
149  */
150 
151 CTASSERT(sizeof(struct pmclog_callchain) == 6*4 +
152     PMC_CALLCHAIN_DEPTH_MAX*sizeof(uintfptr_t));
153 CTASSERT(sizeof(struct pmclog_closelog) == 3*4);
154 CTASSERT(sizeof(struct pmclog_dropnotify) == 3*4);
155 CTASSERT(sizeof(struct pmclog_map_in) == PATH_MAX +
156     4*4 + sizeof(uintfptr_t));
157 CTASSERT(offsetof(struct pmclog_map_in,pl_pathname) ==
158     4*4 + sizeof(uintfptr_t));
159 CTASSERT(sizeof(struct pmclog_map_out) == 4*4 + 2*sizeof(uintfptr_t));
160 CTASSERT(sizeof(struct pmclog_pcsample) == 6*4 + sizeof(uintfptr_t));
161 CTASSERT(sizeof(struct pmclog_pmcallocate) == 6*4);
162 CTASSERT(sizeof(struct pmclog_pmcattach) == 5*4 + PATH_MAX);
163 CTASSERT(offsetof(struct pmclog_pmcattach,pl_pathname) == 5*4);
164 CTASSERT(sizeof(struct pmclog_pmcdetach) == 5*4);
165 CTASSERT(sizeof(struct pmclog_proccsw) == 5*4 + 8);
166 CTASSERT(sizeof(struct pmclog_procexec) == 5*4 + PATH_MAX +
167     sizeof(uintfptr_t));
168 CTASSERT(offsetof(struct pmclog_procexec,pl_pathname) == 5*4 +
169     sizeof(uintfptr_t));
170 CTASSERT(sizeof(struct pmclog_procexit) == 5*4 + 8);
171 CTASSERT(sizeof(struct pmclog_procfork) == 5*4);
172 CTASSERT(sizeof(struct pmclog_sysexit) == 4*4);
173 CTASSERT(sizeof(struct pmclog_userdata) == 4*4);
174 
175 /*
176  * Log buffer structure
177  */
178 
179 struct pmclog_buffer {
180 	TAILQ_ENTRY(pmclog_buffer) plb_next;
181 	char 		*plb_base;
182 	char		*plb_ptr;
183 	char 		*plb_fence;
184 };
185 
186 /*
187  * Prototypes
188  */
189 
190 static int pmclog_get_buffer(struct pmc_owner *po);
191 static void pmclog_loop(void *arg);
192 static void pmclog_release(struct pmc_owner *po);
193 static uint32_t *pmclog_reserve(struct pmc_owner *po, int length);
194 static void pmclog_schedule_io(struct pmc_owner *po);
195 static void pmclog_stop_kthread(struct pmc_owner *po);
196 
197 /*
198  * Helper functions
199  */
200 
201 /*
202  * Get a log buffer
203  */
204 
205 static int
206 pmclog_get_buffer(struct pmc_owner *po)
207 {
208 	struct pmclog_buffer *plb;
209 
210 	mtx_assert(&po->po_mtx, MA_OWNED);
211 
212 	KASSERT(po->po_curbuf == NULL,
213 	    ("[pmclog,%d] po=%p current buffer still valid", __LINE__, po));
214 
215 	mtx_lock_spin(&pmc_bufferlist_mtx);
216 	if ((plb = TAILQ_FIRST(&pmc_bufferlist)) != NULL)
217 		TAILQ_REMOVE(&pmc_bufferlist, plb, plb_next);
218 	mtx_unlock_spin(&pmc_bufferlist_mtx);
219 
220 	PMCDBG2(LOG,GTB,1, "po=%p plb=%p", po, plb);
221 
222 #ifdef	HWPMC_DEBUG
223 	if (plb)
224 		KASSERT(plb->plb_ptr == plb->plb_base &&
225 		    plb->plb_base < plb->plb_fence,
226 		    ("[pmclog,%d] po=%p buffer invariants: ptr=%p "
227 		    "base=%p fence=%p", __LINE__, po, plb->plb_ptr,
228 		    plb->plb_base, plb->plb_fence));
229 #endif
230 
231 	po->po_curbuf = plb;
232 
233 	/* update stats */
234 	atomic_add_int(&pmc_stats.pm_buffer_requests, 1);
235 	if (plb == NULL)
236 		atomic_add_int(&pmc_stats.pm_buffer_requests_failed, 1);
237 
238 	return (plb ? 0 : ENOMEM);
239 }
240 
241 /*
242  * Log handler loop.
243  *
244  * This function is executed by each pmc owner's helper thread.
245  */
246 
247 static void
248 pmclog_loop(void *arg)
249 {
250 	int error;
251 	struct pmc_owner *po;
252 	struct pmclog_buffer *lb;
253 	struct proc *p;
254 	struct ucred *ownercred;
255 	struct ucred *mycred;
256 	struct thread *td;
257 	struct uio auio;
258 	struct iovec aiov;
259 	size_t nbytes;
260 
261 	po = (struct pmc_owner *) arg;
262 	p = po->po_owner;
263 	td = curthread;
264 	mycred = td->td_ucred;
265 
266 	PROC_LOCK(p);
267 	ownercred = crhold(p->p_ucred);
268 	PROC_UNLOCK(p);
269 
270 	PMCDBG2(LOG,INI,1, "po=%p kt=%p", po, po->po_kthread);
271 	KASSERT(po->po_kthread == curthread->td_proc,
272 	    ("[pmclog,%d] proc mismatch po=%p po/kt=%p curproc=%p", __LINE__,
273 		po, po->po_kthread, curthread->td_proc));
274 
275 	lb = NULL;
276 
277 
278 	/*
279 	 * Loop waiting for I/O requests to be added to the owner
280 	 * struct's queue.  The loop is exited when the log file
281 	 * is deconfigured.
282 	 */
283 
284 	mtx_lock(&pmc_kthread_mtx);
285 
286 	for (;;) {
287 
288 		/* check if we've been asked to exit */
289 		if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
290 			break;
291 
292 		if (lb == NULL) { /* look for a fresh buffer to write */
293 			mtx_lock_spin(&po->po_mtx);
294 			if ((lb = TAILQ_FIRST(&po->po_logbuffers)) == NULL) {
295 				mtx_unlock_spin(&po->po_mtx);
296 
297 				/* No more buffers and shutdown required. */
298 				if (po->po_flags & PMC_PO_SHUTDOWN) {
299 					mtx_unlock(&pmc_kthread_mtx);
300 					/*
301 			 		 * Close the file to get PMCLOG_EOF
302 					 * error in pmclog(3).
303 					 */
304 					fo_close(po->po_file, curthread);
305 					mtx_lock(&pmc_kthread_mtx);
306 					break;
307 				}
308 
309 				(void) msleep(po, &pmc_kthread_mtx, PWAIT,
310 				    "pmcloop", 0);
311 				continue;
312 			}
313 
314 			TAILQ_REMOVE(&po->po_logbuffers, lb, plb_next);
315 			mtx_unlock_spin(&po->po_mtx);
316 		}
317 
318 		mtx_unlock(&pmc_kthread_mtx);
319 
320 		/* process the request */
321 		PMCDBG3(LOG,WRI,2, "po=%p base=%p ptr=%p", po,
322 		    lb->plb_base, lb->plb_ptr);
323 		/* change our thread's credentials before issuing the I/O */
324 
325 		aiov.iov_base = lb->plb_base;
326 		aiov.iov_len  = nbytes = lb->plb_ptr - lb->plb_base;
327 
328 		auio.uio_iov    = &aiov;
329 		auio.uio_iovcnt = 1;
330 		auio.uio_offset = -1;
331 		auio.uio_resid  = nbytes;
332 		auio.uio_rw     = UIO_WRITE;
333 		auio.uio_segflg = UIO_SYSSPACE;
334 		auio.uio_td     = td;
335 
336 		/* switch thread credentials -- see kern_ktrace.c */
337 		td->td_ucred = ownercred;
338 		error = fo_write(po->po_file, &auio, ownercred, 0, td);
339 		td->td_ucred = mycred;
340 
341 		if (error) {
342 			/* XXX some errors are recoverable */
343 			/* send a SIGIO to the owner and exit */
344 			PROC_LOCK(p);
345 			kern_psignal(p, SIGIO);
346 			PROC_UNLOCK(p);
347 
348 			mtx_lock(&pmc_kthread_mtx);
349 
350 			po->po_error = error; /* save for flush log */
351 
352 			PMCDBG2(LOG,WRI,2, "po=%p error=%d", po, error);
353 
354 			break;
355 		}
356 
357 		mtx_lock(&pmc_kthread_mtx);
358 
359 		/* put the used buffer back into the global pool */
360 		PMCLOG_INIT_BUFFER_DESCRIPTOR(lb);
361 
362 		mtx_lock_spin(&pmc_bufferlist_mtx);
363 		TAILQ_INSERT_HEAD(&pmc_bufferlist, lb, plb_next);
364 		mtx_unlock_spin(&pmc_bufferlist_mtx);
365 
366 		lb = NULL;
367 	}
368 
369 	wakeup_one(po->po_kthread);
370 	po->po_kthread = NULL;
371 
372 	mtx_unlock(&pmc_kthread_mtx);
373 
374 	/* return the current I/O buffer to the global pool */
375 	if (lb) {
376 		PMCLOG_INIT_BUFFER_DESCRIPTOR(lb);
377 
378 		mtx_lock_spin(&pmc_bufferlist_mtx);
379 		TAILQ_INSERT_HEAD(&pmc_bufferlist, lb, plb_next);
380 		mtx_unlock_spin(&pmc_bufferlist_mtx);
381 	}
382 
383 	/*
384 	 * Exit this thread, signalling the waiter
385 	 */
386 
387 	crfree(ownercred);
388 
389 	kproc_exit(0);
390 }
391 
392 /*
393  * Release and log entry and schedule an I/O if needed.
394  */
395 
396 static void
397 pmclog_release(struct pmc_owner *po)
398 {
399 	KASSERT(po->po_curbuf->plb_ptr >= po->po_curbuf->plb_base,
400 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p base=%p", __LINE__,
401 		po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_base));
402 	KASSERT(po->po_curbuf->plb_ptr <= po->po_curbuf->plb_fence,
403 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p fenc=%p", __LINE__,
404 		po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_fence));
405 
406 	/* schedule an I/O if we've filled a buffer */
407 	if (po->po_curbuf->plb_ptr >= po->po_curbuf->plb_fence)
408 		pmclog_schedule_io(po);
409 
410 	mtx_unlock_spin(&po->po_mtx);
411 
412 	PMCDBG1(LOG,REL,1, "po=%p", po);
413 }
414 
415 
416 /*
417  * Attempt to reserve 'length' bytes of space in an owner's log
418  * buffer.  The function returns a pointer to 'length' bytes of space
419  * if there was enough space or returns NULL if no space was
420  * available.  Non-null returns do so with the po mutex locked.  The
421  * caller must invoke pmclog_release() on the pmc owner structure
422  * when done.
423  */
424 
425 static uint32_t *
426 pmclog_reserve(struct pmc_owner *po, int length)
427 {
428 	uintptr_t newptr, oldptr;
429 	uint32_t *lh;
430 	struct timespec ts;
431 
432 	PMCDBG2(LOG,ALL,1, "po=%p len=%d", po, length);
433 
434 	KASSERT(length % sizeof(uint32_t) == 0,
435 	    ("[pmclog,%d] length not a multiple of word size", __LINE__));
436 
437 	mtx_lock_spin(&po->po_mtx);
438 
439 	/* No more data when shutdown in progress. */
440 	if (po->po_flags & PMC_PO_SHUTDOWN) {
441 		mtx_unlock_spin(&po->po_mtx);
442 		return (NULL);
443 	}
444 
445 	if (po->po_curbuf == NULL)
446 		if (pmclog_get_buffer(po) != 0) {
447 			mtx_unlock_spin(&po->po_mtx);
448 			return (NULL);
449 		}
450 
451 	KASSERT(po->po_curbuf != NULL,
452 	    ("[pmclog,%d] po=%p no current buffer", __LINE__, po));
453 
454 	KASSERT(po->po_curbuf->plb_ptr >= po->po_curbuf->plb_base &&
455 	    po->po_curbuf->plb_ptr <= po->po_curbuf->plb_fence,
456 	    ("[pmclog,%d] po=%p buffer invariants: ptr=%p base=%p fence=%p",
457 		__LINE__, po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_base,
458 		po->po_curbuf->plb_fence));
459 
460 	oldptr = (uintptr_t) po->po_curbuf->plb_ptr;
461 	newptr = oldptr + length;
462 
463 	KASSERT(oldptr != (uintptr_t) NULL,
464 	    ("[pmclog,%d] po=%p Null log buffer pointer", __LINE__, po));
465 
466 	/*
467 	 * If we have space in the current buffer, return a pointer to
468 	 * available space with the PO structure locked.
469 	 */
470 	if (newptr <= (uintptr_t) po->po_curbuf->plb_fence) {
471 		po->po_curbuf->plb_ptr = (char *) newptr;
472 		goto done;
473 	}
474 
475 	/*
476 	 * Otherwise, schedule the current buffer for output and get a
477 	 * fresh buffer.
478 	 */
479 	pmclog_schedule_io(po);
480 
481 	if (pmclog_get_buffer(po) != 0) {
482 		mtx_unlock_spin(&po->po_mtx);
483 		return (NULL);
484 	}
485 
486 	KASSERT(po->po_curbuf != NULL,
487 	    ("[pmclog,%d] po=%p no current buffer", __LINE__, po));
488 
489 	KASSERT(po->po_curbuf->plb_ptr != NULL,
490 	    ("[pmclog,%d] null return from pmc_get_log_buffer", __LINE__));
491 
492 	KASSERT(po->po_curbuf->plb_ptr == po->po_curbuf->plb_base &&
493 	    po->po_curbuf->plb_ptr <= po->po_curbuf->plb_fence,
494 	    ("[pmclog,%d] po=%p buffer invariants: ptr=%p base=%p fence=%p",
495 		__LINE__, po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_base,
496 		po->po_curbuf->plb_fence));
497 
498 	oldptr = (uintptr_t) po->po_curbuf->plb_ptr;
499 
500  done:
501 	lh = (uint32_t *) oldptr;
502 	lh++;				/* skip header */
503 	getnanotime(&ts);		/* fill in the timestamp */
504 	*lh++ = ts.tv_sec & 0xFFFFFFFF;
505 	*lh++ = ts.tv_nsec & 0xFFFFFFF;
506 	return ((uint32_t *) oldptr);
507 }
508 
509 /*
510  * Schedule an I/O.
511  *
512  * Transfer the current buffer to the helper kthread.
513  */
514 
515 static void
516 pmclog_schedule_io(struct pmc_owner *po)
517 {
518 	KASSERT(po->po_curbuf != NULL,
519 	    ("[pmclog,%d] schedule_io with null buffer po=%p", __LINE__, po));
520 
521 	KASSERT(po->po_curbuf->plb_ptr >= po->po_curbuf->plb_base,
522 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p base=%p", __LINE__,
523 		po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_base));
524 	KASSERT(po->po_curbuf->plb_ptr <= po->po_curbuf->plb_fence,
525 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p fenc=%p", __LINE__,
526 		po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_fence));
527 
528 	PMCDBG1(LOG,SIO, 1, "po=%p", po);
529 
530 	mtx_assert(&po->po_mtx, MA_OWNED);
531 
532 	/*
533 	 * Add the current buffer to the tail of the buffer list and
534 	 * wakeup the helper.
535 	 */
536 	TAILQ_INSERT_TAIL(&po->po_logbuffers, po->po_curbuf, plb_next);
537 	po->po_curbuf = NULL;
538 	wakeup_one(po);
539 }
540 
541 /*
542  * Stop the helper kthread.
543  */
544 
545 static void
546 pmclog_stop_kthread(struct pmc_owner *po)
547 {
548 	/*
549 	 * Close the file to force the thread out of fo_write,
550 	 * unset flag, wakeup the helper thread,
551 	 * wait for it to exit
552 	 */
553 
554 	if (po->po_file != NULL)
555 		fo_close(po->po_file, curthread);
556 
557 	mtx_lock(&pmc_kthread_mtx);
558 	po->po_flags &= ~PMC_PO_OWNS_LOGFILE;
559 	wakeup_one(po);
560 	if (po->po_kthread)
561 		msleep(po->po_kthread, &pmc_kthread_mtx, PPAUSE, "pmckstp", 0);
562 	mtx_unlock(&pmc_kthread_mtx);
563 }
564 
565 /*
566  * Public functions
567  */
568 
569 /*
570  * Configure a log file for pmc owner 'po'.
571  *
572  * Parameter 'logfd' is a file handle referencing an open file in the
573  * owner process.  This file needs to have been opened for writing.
574  */
575 
576 int
577 pmclog_configure_log(struct pmc_mdep *md, struct pmc_owner *po, int logfd)
578 {
579 	int error;
580 	struct proc *p;
581 	cap_rights_t rights;
582 	/*
583 	 * As long as it is possible to get a LOR between pmc_sx lock and
584 	 * proctree/allproc sx locks used for adding a new process, assure
585 	 * the former is not held here.
586 	 */
587 	sx_assert(&pmc_sx, SA_UNLOCKED);
588 	PMCDBG2(LOG,CFG,1, "config po=%p logfd=%d", po, logfd);
589 
590 	p = po->po_owner;
591 
592 	/* return EBUSY if a log file was already present */
593 	if (po->po_flags & PMC_PO_OWNS_LOGFILE)
594 		return (EBUSY);
595 
596 	KASSERT(po->po_kthread == NULL,
597 	    ("[pmclog,%d] po=%p kthread (%p) already present", __LINE__, po,
598 		po->po_kthread));
599 	KASSERT(po->po_file == NULL,
600 	    ("[pmclog,%d] po=%p file (%p) already present", __LINE__, po,
601 		po->po_file));
602 
603 	/* get a reference to the file state */
604 	error = fget_write(curthread, logfd,
605 	    cap_rights_init(&rights, CAP_WRITE), &po->po_file);
606 	if (error)
607 		goto error;
608 
609 	/* mark process as owning a log file */
610 	po->po_flags |= PMC_PO_OWNS_LOGFILE;
611 	error = kproc_create(pmclog_loop, po, &po->po_kthread,
612 	    RFHIGHPID, 0, "hwpmc: proc(%d)", p->p_pid);
613 	if (error)
614 		goto error;
615 
616 	/* mark process as using HWPMCs */
617 	PROC_LOCK(p);
618 	p->p_flag |= P_HWPMC;
619 	PROC_UNLOCK(p);
620 
621 	/* create a log initialization entry */
622 	PMCLOG_RESERVE_WITH_ERROR(po, INITIALIZE,
623 	    sizeof(struct pmclog_initialize));
624 	PMCLOG_EMIT32(PMC_VERSION);
625 	PMCLOG_EMIT32(md->pmd_cputype);
626 	PMCLOG_DESPATCH(po);
627 
628 	return (0);
629 
630  error:
631 	/* shutdown the thread */
632 	if (po->po_kthread)
633 		pmclog_stop_kthread(po);
634 
635 	KASSERT(po->po_kthread == NULL, ("[pmclog,%d] po=%p kthread not "
636 	    "stopped", __LINE__, po));
637 
638 	if (po->po_file)
639 		(void) fdrop(po->po_file, curthread);
640 	po->po_file  = NULL;	/* clear file and error state */
641 	po->po_error = 0;
642 
643 	return (error);
644 }
645 
646 
647 /*
648  * De-configure a log file.  This will throw away any buffers queued
649  * for this owner process.
650  */
651 
652 int
653 pmclog_deconfigure_log(struct pmc_owner *po)
654 {
655 	int error;
656 	struct pmclog_buffer *lb;
657 
658 	PMCDBG1(LOG,CFG,1, "de-config po=%p", po);
659 
660 	if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
661 		return (EINVAL);
662 
663 	KASSERT(po->po_sscount == 0,
664 	    ("[pmclog,%d] po=%p still owning SS PMCs", __LINE__, po));
665 	KASSERT(po->po_file != NULL,
666 	    ("[pmclog,%d] po=%p no log file", __LINE__, po));
667 
668 	/* stop the kthread, this will reset the 'OWNS_LOGFILE' flag */
669 	pmclog_stop_kthread(po);
670 
671 	KASSERT(po->po_kthread == NULL,
672 	    ("[pmclog,%d] po=%p kthread not stopped", __LINE__, po));
673 
674 	/* return all queued log buffers to the global pool */
675 	while ((lb = TAILQ_FIRST(&po->po_logbuffers)) != NULL) {
676 		TAILQ_REMOVE(&po->po_logbuffers, lb, plb_next);
677 		PMCLOG_INIT_BUFFER_DESCRIPTOR(lb);
678 		mtx_lock_spin(&pmc_bufferlist_mtx);
679 		TAILQ_INSERT_HEAD(&pmc_bufferlist, lb, plb_next);
680 		mtx_unlock_spin(&pmc_bufferlist_mtx);
681 	}
682 
683 	/* return the 'current' buffer to the global pool */
684 	if ((lb = po->po_curbuf) != NULL) {
685 		PMCLOG_INIT_BUFFER_DESCRIPTOR(lb);
686 		mtx_lock_spin(&pmc_bufferlist_mtx);
687 		TAILQ_INSERT_HEAD(&pmc_bufferlist, lb, plb_next);
688 		mtx_unlock_spin(&pmc_bufferlist_mtx);
689 	}
690 
691 	/* drop a reference to the fd */
692 	error = fdrop(po->po_file, curthread);
693 	po->po_file  = NULL;
694 	po->po_error = 0;
695 
696 	return (error);
697 }
698 
699 /*
700  * Flush a process' log buffer.
701  */
702 
703 int
704 pmclog_flush(struct pmc_owner *po)
705 {
706 	int error;
707 	struct pmclog_buffer *lb;
708 
709 	PMCDBG1(LOG,FLS,1, "po=%p", po);
710 
711 	/*
712 	 * If there is a pending error recorded by the logger thread,
713 	 * return that.
714 	 */
715 	if (po->po_error)
716 		return (po->po_error);
717 
718 	error = 0;
719 
720 	/*
721 	 * Check that we do have an active log file.
722 	 */
723 	mtx_lock(&pmc_kthread_mtx);
724 	if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) {
725 		error = EINVAL;
726 		goto error;
727 	}
728 
729 	/*
730 	 * Schedule the current buffer if any and not empty.
731 	 */
732 	mtx_lock_spin(&po->po_mtx);
733 	lb = po->po_curbuf;
734 	if (lb && lb->plb_ptr != lb->plb_base) {
735 		pmclog_schedule_io(po);
736 	} else
737 		error = ENOBUFS;
738 	mtx_unlock_spin(&po->po_mtx);
739 
740  error:
741 	mtx_unlock(&pmc_kthread_mtx);
742 
743 	return (error);
744 }
745 
746 int
747 pmclog_close(struct pmc_owner *po)
748 {
749 
750 	PMCDBG1(LOG,CLO,1, "po=%p", po);
751 
752 	mtx_lock(&pmc_kthread_mtx);
753 
754 	/*
755 	 * Schedule the current buffer.
756 	 */
757 	mtx_lock_spin(&po->po_mtx);
758 	if (po->po_curbuf)
759 		pmclog_schedule_io(po);
760 	else
761 		wakeup_one(po);
762 	mtx_unlock_spin(&po->po_mtx);
763 
764 	/*
765 	 * Initiate shutdown: no new data queued,
766 	 * thread will close file on last block.
767 	 */
768 	po->po_flags |= PMC_PO_SHUTDOWN;
769 
770 	mtx_unlock(&pmc_kthread_mtx);
771 
772 	return (0);
773 }
774 
775 void
776 pmclog_process_callchain(struct pmc *pm, struct pmc_sample *ps)
777 {
778 	int n, recordlen;
779 	uint32_t flags;
780 	struct pmc_owner *po;
781 
782 	PMCDBG3(LOG,SAM,1,"pm=%p pid=%d n=%d", pm, ps->ps_pid,
783 	    ps->ps_nsamples);
784 
785 	recordlen = offsetof(struct pmclog_callchain, pl_pc) +
786 	    ps->ps_nsamples * sizeof(uintfptr_t);
787 	po = pm->pm_owner;
788 	flags = PMC_CALLCHAIN_TO_CPUFLAGS(ps->ps_cpu,ps->ps_flags);
789 	PMCLOG_RESERVE(po, CALLCHAIN, recordlen);
790 	PMCLOG_EMIT32(ps->ps_pid);
791 	PMCLOG_EMIT32(pm->pm_id);
792 	PMCLOG_EMIT32(flags);
793 	for (n = 0; n < ps->ps_nsamples; n++)
794 		PMCLOG_EMITADDR(ps->ps_pc[n]);
795 	PMCLOG_DESPATCH(po);
796 }
797 
798 void
799 pmclog_process_closelog(struct pmc_owner *po)
800 {
801 	PMCLOG_RESERVE(po,CLOSELOG,sizeof(struct pmclog_closelog));
802 	PMCLOG_DESPATCH(po);
803 }
804 
805 void
806 pmclog_process_dropnotify(struct pmc_owner *po)
807 {
808 	PMCLOG_RESERVE(po,DROPNOTIFY,sizeof(struct pmclog_dropnotify));
809 	PMCLOG_DESPATCH(po);
810 }
811 
812 void
813 pmclog_process_map_in(struct pmc_owner *po, pid_t pid, uintfptr_t start,
814     const char *path)
815 {
816 	int pathlen, recordlen;
817 
818 	KASSERT(path != NULL, ("[pmclog,%d] map-in, null path", __LINE__));
819 
820 	pathlen = strlen(path) + 1;	/* #bytes for path name */
821 	recordlen = offsetof(struct pmclog_map_in, pl_pathname) +
822 	    pathlen;
823 
824 	PMCLOG_RESERVE(po, MAP_IN, recordlen);
825 	PMCLOG_EMIT32(pid);
826 	PMCLOG_EMITADDR(start);
827 	PMCLOG_EMITSTRING(path,pathlen);
828 	PMCLOG_DESPATCH(po);
829 }
830 
831 void
832 pmclog_process_map_out(struct pmc_owner *po, pid_t pid, uintfptr_t start,
833     uintfptr_t end)
834 {
835 	KASSERT(start <= end, ("[pmclog,%d] start > end", __LINE__));
836 
837 	PMCLOG_RESERVE(po, MAP_OUT, sizeof(struct pmclog_map_out));
838 	PMCLOG_EMIT32(pid);
839 	PMCLOG_EMITADDR(start);
840 	PMCLOG_EMITADDR(end);
841 	PMCLOG_DESPATCH(po);
842 }
843 
844 void
845 pmclog_process_pmcallocate(struct pmc *pm)
846 {
847 	struct pmc_owner *po;
848 	struct pmc_soft *ps;
849 
850 	po = pm->pm_owner;
851 
852 	PMCDBG1(LOG,ALL,1, "pm=%p", pm);
853 
854 	if (PMC_TO_CLASS(pm) == PMC_CLASS_SOFT) {
855 		PMCLOG_RESERVE(po, PMCALLOCATEDYN,
856 		    sizeof(struct pmclog_pmcallocatedyn));
857 		PMCLOG_EMIT32(pm->pm_id);
858 		PMCLOG_EMIT32(pm->pm_event);
859 		PMCLOG_EMIT32(pm->pm_flags);
860 		ps = pmc_soft_ev_acquire(pm->pm_event);
861 		if (ps != NULL)
862 			PMCLOG_EMITSTRING(ps->ps_ev.pm_ev_name,PMC_NAME_MAX);
863 		else
864 			PMCLOG_EMITNULLSTRING(PMC_NAME_MAX);
865 		pmc_soft_ev_release(ps);
866 		PMCLOG_DESPATCH(po);
867 	} else {
868 		PMCLOG_RESERVE(po, PMCALLOCATE,
869 		    sizeof(struct pmclog_pmcallocate));
870 		PMCLOG_EMIT32(pm->pm_id);
871 		PMCLOG_EMIT32(pm->pm_event);
872 		PMCLOG_EMIT32(pm->pm_flags);
873 		PMCLOG_DESPATCH(po);
874 	}
875 }
876 
877 void
878 pmclog_process_pmcattach(struct pmc *pm, pid_t pid, char *path)
879 {
880 	int pathlen, recordlen;
881 	struct pmc_owner *po;
882 
883 	PMCDBG2(LOG,ATT,1,"pm=%p pid=%d", pm, pid);
884 
885 	po = pm->pm_owner;
886 
887 	pathlen = strlen(path) + 1;	/* #bytes for the string */
888 	recordlen = offsetof(struct pmclog_pmcattach, pl_pathname) + pathlen;
889 
890 	PMCLOG_RESERVE(po, PMCATTACH, recordlen);
891 	PMCLOG_EMIT32(pm->pm_id);
892 	PMCLOG_EMIT32(pid);
893 	PMCLOG_EMITSTRING(path, pathlen);
894 	PMCLOG_DESPATCH(po);
895 }
896 
897 void
898 pmclog_process_pmcdetach(struct pmc *pm, pid_t pid)
899 {
900 	struct pmc_owner *po;
901 
902 	PMCDBG2(LOG,ATT,1,"!pm=%p pid=%d", pm, pid);
903 
904 	po = pm->pm_owner;
905 
906 	PMCLOG_RESERVE(po, PMCDETACH, sizeof(struct pmclog_pmcdetach));
907 	PMCLOG_EMIT32(pm->pm_id);
908 	PMCLOG_EMIT32(pid);
909 	PMCLOG_DESPATCH(po);
910 }
911 
912 /*
913  * Log a context switch event to the log file.
914  */
915 
916 void
917 pmclog_process_proccsw(struct pmc *pm, struct pmc_process *pp, pmc_value_t v)
918 {
919 	struct pmc_owner *po;
920 
921 	KASSERT(pm->pm_flags & PMC_F_LOG_PROCCSW,
922 	    ("[pmclog,%d] log-process-csw called gratuitously", __LINE__));
923 
924 	PMCDBG3(LOG,SWO,1,"pm=%p pid=%d v=%jx", pm, pp->pp_proc->p_pid,
925 	    v);
926 
927 	po = pm->pm_owner;
928 
929 	PMCLOG_RESERVE(po, PROCCSW, sizeof(struct pmclog_proccsw));
930 	PMCLOG_EMIT32(pm->pm_id);
931 	PMCLOG_EMIT64(v);
932 	PMCLOG_EMIT32(pp->pp_proc->p_pid);
933 	PMCLOG_DESPATCH(po);
934 }
935 
936 void
937 pmclog_process_procexec(struct pmc_owner *po, pmc_id_t pmid, pid_t pid,
938     uintfptr_t startaddr, char *path)
939 {
940 	int pathlen, recordlen;
941 
942 	PMCDBG3(LOG,EXC,1,"po=%p pid=%d path=\"%s\"", po, pid, path);
943 
944 	pathlen   = strlen(path) + 1;	/* #bytes for the path */
945 	recordlen = offsetof(struct pmclog_procexec, pl_pathname) + pathlen;
946 
947 	PMCLOG_RESERVE(po, PROCEXEC, recordlen);
948 	PMCLOG_EMIT32(pid);
949 	PMCLOG_EMITADDR(startaddr);
950 	PMCLOG_EMIT32(pmid);
951 	PMCLOG_EMITSTRING(path,pathlen);
952 	PMCLOG_DESPATCH(po);
953 }
954 
955 /*
956  * Log a process exit event (and accumulated pmc value) to the log file.
957  */
958 
959 void
960 pmclog_process_procexit(struct pmc *pm, struct pmc_process *pp)
961 {
962 	int ri;
963 	struct pmc_owner *po;
964 
965 	ri = PMC_TO_ROWINDEX(pm);
966 	PMCDBG3(LOG,EXT,1,"pm=%p pid=%d v=%jx", pm, pp->pp_proc->p_pid,
967 	    pp->pp_pmcs[ri].pp_pmcval);
968 
969 	po = pm->pm_owner;
970 
971 	PMCLOG_RESERVE(po, PROCEXIT, sizeof(struct pmclog_procexit));
972 	PMCLOG_EMIT32(pm->pm_id);
973 	PMCLOG_EMIT64(pp->pp_pmcs[ri].pp_pmcval);
974 	PMCLOG_EMIT32(pp->pp_proc->p_pid);
975 	PMCLOG_DESPATCH(po);
976 }
977 
978 /*
979  * Log a fork event.
980  */
981 
982 void
983 pmclog_process_procfork(struct pmc_owner *po, pid_t oldpid, pid_t newpid)
984 {
985 	PMCLOG_RESERVE(po, PROCFORK, sizeof(struct pmclog_procfork));
986 	PMCLOG_EMIT32(oldpid);
987 	PMCLOG_EMIT32(newpid);
988 	PMCLOG_DESPATCH(po);
989 }
990 
991 /*
992  * Log a process exit event of the form suitable for system-wide PMCs.
993  */
994 
995 void
996 pmclog_process_sysexit(struct pmc_owner *po, pid_t pid)
997 {
998 	PMCLOG_RESERVE(po, SYSEXIT, sizeof(struct pmclog_sysexit));
999 	PMCLOG_EMIT32(pid);
1000 	PMCLOG_DESPATCH(po);
1001 }
1002 
1003 /*
1004  * Write a user log entry.
1005  */
1006 
1007 int
1008 pmclog_process_userlog(struct pmc_owner *po, struct pmc_op_writelog *wl)
1009 {
1010 	int error;
1011 
1012 	PMCDBG2(LOG,WRI,1, "writelog po=%p ud=0x%x", po, wl->pm_userdata);
1013 
1014 	error = 0;
1015 
1016 	PMCLOG_RESERVE_WITH_ERROR(po, USERDATA,
1017 	    sizeof(struct pmclog_userdata));
1018 	PMCLOG_EMIT32(wl->pm_userdata);
1019 	PMCLOG_DESPATCH(po);
1020 
1021  error:
1022 	return (error);
1023 }
1024 
1025 /*
1026  * Initialization.
1027  *
1028  * Create a pool of log buffers and initialize mutexes.
1029  */
1030 
1031 void
1032 pmclog_initialize()
1033 {
1034 	int n;
1035 	struct pmclog_buffer *plb;
1036 
1037 	if (pmclog_buffer_size <= 0) {
1038 		(void) printf("hwpmc: tunable logbuffersize=%d must be "
1039 		    "greater than zero.\n", pmclog_buffer_size);
1040 		pmclog_buffer_size = PMC_LOG_BUFFER_SIZE;
1041 	}
1042 
1043 	if (pmc_nlogbuffers <= 0) {
1044 		(void) printf("hwpmc: tunable nlogbuffers=%d must be greater "
1045 		    "than zero.\n", pmc_nlogbuffers);
1046 		pmc_nlogbuffers = PMC_NLOGBUFFERS;
1047 	}
1048 
1049 	/* create global pool of log buffers */
1050 	for (n = 0; n < pmc_nlogbuffers; n++) {
1051 		plb = malloc(1024 * pmclog_buffer_size, M_PMC,
1052 		    M_WAITOK|M_ZERO);
1053 		PMCLOG_INIT_BUFFER_DESCRIPTOR(plb);
1054 		TAILQ_INSERT_HEAD(&pmc_bufferlist, plb, plb_next);
1055 	}
1056 	mtx_init(&pmc_bufferlist_mtx, "pmc-buffer-list", "pmc-leaf",
1057 	    MTX_SPIN);
1058 	mtx_init(&pmc_kthread_mtx, "pmc-kthread", "pmc-sleep", MTX_DEF);
1059 }
1060 
1061 /*
1062  * Shutdown logging.
1063  *
1064  * Destroy mutexes and release memory back the to free pool.
1065  */
1066 
1067 void
1068 pmclog_shutdown()
1069 {
1070 	struct pmclog_buffer *plb;
1071 
1072 	mtx_destroy(&pmc_kthread_mtx);
1073 	mtx_destroy(&pmc_bufferlist_mtx);
1074 
1075 	while ((plb = TAILQ_FIRST(&pmc_bufferlist)) != NULL) {
1076 		TAILQ_REMOVE(&pmc_bufferlist, plb, plb_next);
1077 		free(plb, M_PMC);
1078 	}
1079 }
1080