xref: /freebsd-11-stable/sys/dev/hwpmc/hwpmc_logging.c (revision 4ab2e064d7950be84256d671a7ae93f87cc6aa36)
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/syscallsubr.h>
57 #include <sys/sysctl.h>
58 #include <sys/systm.h>
59 #include <sys/uio.h>
60 #include <sys/unistd.h>
61 #include <sys/vnode.h>
62 
63 /*
64  * Sysctl tunables
65  */
66 
67 SYSCTL_DECL(_kern_hwpmc);
68 
69 /*
70  * kern.hwpmc.logbuffersize -- size of the per-cpu owner buffers.
71  */
72 
73 static int pmclog_buffer_size = PMC_LOG_BUFFER_SIZE;
74 #if (__FreeBSD_version < 1100000)
75 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "logbuffersize", &pmclog_buffer_size);
76 #endif
77 SYSCTL_INT(_kern_hwpmc, OID_AUTO, logbuffersize, CTLFLAG_RDTUN,
78     &pmclog_buffer_size, 0, "size of log buffers in kilobytes");
79 
80 /*
81  * kern.hwpmc.nbuffer -- number of global log buffers
82  */
83 
84 static int pmc_nlogbuffers = PMC_NLOGBUFFERS;
85 #if (__FreeBSD_version < 1100000)
86 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "nbuffers", &pmc_nlogbuffers);
87 #endif
88 SYSCTL_INT(_kern_hwpmc, OID_AUTO, nbuffers, CTLFLAG_RDTUN,
89     &pmc_nlogbuffers, 0, "number of global log buffers");
90 
91 /*
92  * Global log buffer list and associated spin lock.
93  */
94 
95 TAILQ_HEAD(, pmclog_buffer) pmc_bufferlist =
96 	TAILQ_HEAD_INITIALIZER(pmc_bufferlist);
97 static struct mtx pmc_bufferlist_mtx;	/* spin lock */
98 static struct mtx pmc_kthread_mtx;	/* sleep lock */
99 
100 #define	PMCLOG_INIT_BUFFER_DESCRIPTOR(D) do {				\
101 		const int __roundup = roundup(sizeof(*D),		\
102 			sizeof(uint32_t));				\
103 		(D)->plb_fence = ((char *) (D)) +			\
104 			 1024*pmclog_buffer_size;			\
105 		(D)->plb_base  = (D)->plb_ptr = ((char *) (D)) +	\
106 			__roundup;					\
107 	} while (0)
108 
109 
110 /*
111  * Log file record constructors.
112  */
113 #define	_PMCLOG_TO_HEADER(T,L)						\
114 	((PMCLOG_HEADER_MAGIC << 24) |					\
115 	 (PMCLOG_TYPE_ ## T << 16)   |					\
116 	 ((L) & 0xFFFF))
117 
118 /* reserve LEN bytes of space and initialize the entry header */
119 #define	_PMCLOG_RESERVE(PO,TYPE,LEN,ACTION) do {			\
120 		uint32_t *_le;						\
121 		int _len = roundup((LEN), sizeof(uint32_t));		\
122 		if ((_le = pmclog_reserve((PO), _len)) == NULL) {	\
123 			ACTION;						\
124 		}							\
125 		*_le = _PMCLOG_TO_HEADER(TYPE,_len);			\
126 		_le += 3	/* skip over timestamp */
127 
128 #define	PMCLOG_RESERVE(P,T,L)		_PMCLOG_RESERVE(P,T,L,return)
129 #define	PMCLOG_RESERVE_WITH_ERROR(P,T,L) _PMCLOG_RESERVE(P,T,L,		\
130 	error=ENOMEM;goto error)
131 
132 #define	PMCLOG_EMIT32(V)	do { *_le++ = (V); } while (0)
133 #define	PMCLOG_EMIT64(V)	do { 					\
134 		*_le++ = (uint32_t) ((V) & 0xFFFFFFFF);			\
135 		*_le++ = (uint32_t) (((V) >> 32) & 0xFFFFFFFF);		\
136 	} while (0)
137 
138 
139 /* Emit a string.  Caution: does NOT update _le, so needs to be last */
140 #define	PMCLOG_EMITSTRING(S,L)	do { bcopy((S), _le, (L)); } while (0)
141 #define	PMCLOG_EMITNULLSTRING(L) do { bzero(_le, (L)); } while (0)
142 
143 #define	PMCLOG_DESPATCH(PO)						\
144 		pmclog_release((PO));					\
145 	} while (0)
146 
147 
148 /*
149  * Assertions about the log file format.
150  */
151 
152 CTASSERT(sizeof(struct pmclog_callchain) == 6*4 +
153     PMC_CALLCHAIN_DEPTH_MAX*sizeof(uintfptr_t));
154 CTASSERT(sizeof(struct pmclog_closelog) == 3*4);
155 CTASSERT(sizeof(struct pmclog_dropnotify) == 3*4);
156 CTASSERT(sizeof(struct pmclog_map_in) == PATH_MAX +
157     4*4 + sizeof(uintfptr_t));
158 CTASSERT(offsetof(struct pmclog_map_in,pl_pathname) ==
159     4*4 + sizeof(uintfptr_t));
160 CTASSERT(sizeof(struct pmclog_map_out) == 4*4 + 2*sizeof(uintfptr_t));
161 CTASSERT(sizeof(struct pmclog_pcsample) == 6*4 + sizeof(uintfptr_t));
162 CTASSERT(sizeof(struct pmclog_pmcallocate) == 6*4);
163 CTASSERT(sizeof(struct pmclog_pmcattach) == 5*4 + PATH_MAX);
164 CTASSERT(offsetof(struct pmclog_pmcattach,pl_pathname) == 5*4);
165 CTASSERT(sizeof(struct pmclog_pmcdetach) == 5*4);
166 CTASSERT(sizeof(struct pmclog_proccsw) == 5*4 + 8);
167 CTASSERT(sizeof(struct pmclog_procexec) == 5*4 + PATH_MAX +
168     sizeof(uintfptr_t));
169 CTASSERT(offsetof(struct pmclog_procexec,pl_pathname) == 5*4 +
170     sizeof(uintfptr_t));
171 CTASSERT(sizeof(struct pmclog_procexit) == 5*4 + 8);
172 CTASSERT(sizeof(struct pmclog_procfork) == 5*4);
173 CTASSERT(sizeof(struct pmclog_sysexit) == 4*4);
174 CTASSERT(sizeof(struct pmclog_userdata) == 4*4);
175 
176 /*
177  * Log buffer structure
178  */
179 
180 struct pmclog_buffer {
181 	TAILQ_ENTRY(pmclog_buffer) plb_next;
182 	char 		*plb_base;
183 	char		*plb_ptr;
184 	char 		*plb_fence;
185 };
186 
187 /*
188  * Prototypes
189  */
190 
191 static int pmclog_get_buffer(struct pmc_owner *po);
192 static void pmclog_loop(void *arg);
193 static void pmclog_release(struct pmc_owner *po);
194 static uint32_t *pmclog_reserve(struct pmc_owner *po, int length);
195 static void pmclog_schedule_io(struct pmc_owner *po);
196 static void pmclog_stop_kthread(struct pmc_owner *po);
197 
198 /*
199  * Helper functions
200  */
201 
202 /*
203  * Get a log buffer
204  */
205 
206 static int
pmclog_get_buffer(struct pmc_owner * po)207 pmclog_get_buffer(struct pmc_owner *po)
208 {
209 	struct pmclog_buffer *plb;
210 
211 	mtx_assert(&po->po_mtx, MA_OWNED);
212 
213 	KASSERT(po->po_curbuf == NULL,
214 	    ("[pmclog,%d] po=%p current buffer still valid", __LINE__, po));
215 
216 	mtx_lock_spin(&pmc_bufferlist_mtx);
217 	if ((plb = TAILQ_FIRST(&pmc_bufferlist)) != NULL)
218 		TAILQ_REMOVE(&pmc_bufferlist, plb, plb_next);
219 	mtx_unlock_spin(&pmc_bufferlist_mtx);
220 
221 	PMCDBG2(LOG,GTB,1, "po=%p plb=%p", po, plb);
222 
223 #ifdef	HWPMC_DEBUG
224 	if (plb)
225 		KASSERT(plb->plb_ptr == plb->plb_base &&
226 		    plb->plb_base < plb->plb_fence,
227 		    ("[pmclog,%d] po=%p buffer invariants: ptr=%p "
228 		    "base=%p fence=%p", __LINE__, po, plb->plb_ptr,
229 		    plb->plb_base, plb->plb_fence));
230 #endif
231 
232 	po->po_curbuf = plb;
233 
234 	/* update stats */
235 	atomic_add_int(&pmc_stats.pm_buffer_requests, 1);
236 	if (plb == NULL)
237 		atomic_add_int(&pmc_stats.pm_buffer_requests_failed, 1);
238 
239 	return (plb ? 0 : ENOMEM);
240 }
241 
242 struct pmclog_proc_init_args {
243 	struct proc *kthr;
244 	struct pmc_owner *po;
245 	bool exit;
246 	bool acted;
247 };
248 
249 int
pmclog_proc_create(struct thread * td,void ** handlep)250 pmclog_proc_create(struct thread *td, void **handlep)
251 {
252 	struct pmclog_proc_init_args *ia;
253 	int error;
254 
255 	ia = malloc(sizeof(*ia), M_TEMP, M_WAITOK | M_ZERO);
256 	error = kproc_create(pmclog_loop, ia, &ia->kthr,
257 	    RFHIGHPID, 0, "hwpmc: proc(%d)", td->td_proc->p_pid);
258 	if (error == 0)
259 		*handlep = ia;
260 	return (error);
261 }
262 
263 void
pmclog_proc_ignite(void * handle,struct pmc_owner * po)264 pmclog_proc_ignite(void *handle, struct pmc_owner *po)
265 {
266 	struct pmclog_proc_init_args *ia;
267 
268 	ia = handle;
269 	mtx_lock(&pmc_kthread_mtx);
270 	MPASS(!ia->acted);
271 	MPASS(ia->po == NULL);
272 	MPASS(!ia->exit);
273 	MPASS(ia->kthr != NULL);
274 	if (po == NULL) {
275 		ia->exit = true;
276 	} else {
277 		ia->po = po;
278 		KASSERT(po->po_kthread == NULL,
279 		    ("[pmclog,%d] po=%p kthread (%p) already present",
280 		    __LINE__, po, po->po_kthread));
281 		po->po_kthread = ia->kthr;
282 	}
283 	wakeup(ia);
284 	while (!ia->acted)
285 		msleep(ia, &pmc_kthread_mtx, PWAIT, "pmclogw", 0);
286 	mtx_unlock(&pmc_kthread_mtx);
287 	free(ia, M_TEMP);
288 }
289 
290 /*
291  * Log handler loop.
292  *
293  * This function is executed by each pmc owner's helper thread.
294  */
295 
296 static void
pmclog_loop(void * arg)297 pmclog_loop(void *arg)
298 {
299 	struct pmclog_proc_init_args *ia;
300 	struct pmc_owner *po;
301 	struct pmclog_buffer *lb;
302 	struct proc *p;
303 	struct ucred *ownercred;
304 	struct ucred *mycred;
305 	struct thread *td;
306 	sigset_t unb;
307 	struct uio auio;
308 	struct iovec aiov;
309 	size_t nbytes;
310 	int error;
311 
312 	td = curthread;
313 
314 	SIGEMPTYSET(unb);
315 	SIGADDSET(unb, SIGHUP);
316 	(void)kern_sigprocmask(td, SIG_UNBLOCK, &unb, NULL, 0);
317 
318 	ia = arg;
319 	MPASS(ia->kthr == curproc);
320 	MPASS(!ia->acted);
321 	mtx_lock(&pmc_kthread_mtx);
322 	while (ia->po == NULL && !ia->exit)
323 		msleep(ia, &pmc_kthread_mtx, PWAIT, "pmclogi", 0);
324 	if (ia->exit) {
325 		ia->acted = true;
326 		wakeup(ia);
327 		mtx_unlock(&pmc_kthread_mtx);
328 		kproc_exit(0);
329 	}
330 	MPASS(ia->po != NULL);
331 	po = ia->po;
332 	ia->acted = true;
333 	wakeup(ia);
334 	mtx_unlock(&pmc_kthread_mtx);
335 	ia = NULL;
336 
337 	p = po->po_owner;
338 	mycred = td->td_ucred;
339 
340 	PROC_LOCK(p);
341 	ownercred = crhold(p->p_ucred);
342 	PROC_UNLOCK(p);
343 
344 	PMCDBG2(LOG,INI,1, "po=%p kt=%p", po, po->po_kthread);
345 	KASSERT(po->po_kthread == curthread->td_proc,
346 	    ("[pmclog,%d] proc mismatch po=%p po/kt=%p curproc=%p", __LINE__,
347 		po, po->po_kthread, curthread->td_proc));
348 
349 	lb = NULL;
350 
351 
352 	/*
353 	 * Loop waiting for I/O requests to be added to the owner
354 	 * struct's queue.  The loop is exited when the log file
355 	 * is deconfigured.
356 	 */
357 
358 	mtx_lock(&pmc_kthread_mtx);
359 
360 	for (;;) {
361 
362 		/* check if we've been asked to exit */
363 		if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
364 			break;
365 
366 		if (lb == NULL) { /* look for a fresh buffer to write */
367 			mtx_lock_spin(&po->po_mtx);
368 			if ((lb = TAILQ_FIRST(&po->po_logbuffers)) == NULL) {
369 				mtx_unlock_spin(&po->po_mtx);
370 
371 				/* No more buffers and shutdown required. */
372 				if (po->po_flags & PMC_PO_SHUTDOWN)
373 					break;
374 
375 				(void) msleep(po, &pmc_kthread_mtx, PWAIT,
376 				    "pmcloop", 0);
377 				continue;
378 			}
379 
380 			TAILQ_REMOVE(&po->po_logbuffers, lb, plb_next);
381 			mtx_unlock_spin(&po->po_mtx);
382 		}
383 
384 		mtx_unlock(&pmc_kthread_mtx);
385 
386 		/* process the request */
387 		PMCDBG3(LOG,WRI,2, "po=%p base=%p ptr=%p", po,
388 		    lb->plb_base, lb->plb_ptr);
389 		/* change our thread's credentials before issuing the I/O */
390 
391 		aiov.iov_base = lb->plb_base;
392 		aiov.iov_len  = nbytes = lb->plb_ptr - lb->plb_base;
393 
394 		auio.uio_iov    = &aiov;
395 		auio.uio_iovcnt = 1;
396 		auio.uio_offset = -1;
397 		auio.uio_resid  = nbytes;
398 		auio.uio_rw     = UIO_WRITE;
399 		auio.uio_segflg = UIO_SYSSPACE;
400 		auio.uio_td     = td;
401 
402 		/* switch thread credentials -- see kern_ktrace.c */
403 		td->td_ucred = ownercred;
404 		error = fo_write(po->po_file, &auio, ownercred, 0, td);
405 		td->td_ucred = mycred;
406 
407 		if (error) {
408 			/* XXX some errors are recoverable */
409 			/* send a SIGIO to the owner and exit */
410 			PROC_LOCK(p);
411 			kern_psignal(p, SIGIO);
412 			PROC_UNLOCK(p);
413 
414 			mtx_lock(&pmc_kthread_mtx);
415 
416 			po->po_error = error; /* save for flush log */
417 
418 			PMCDBG2(LOG,WRI,2, "po=%p error=%d", po, error);
419 
420 			break;
421 		}
422 
423 		mtx_lock(&pmc_kthread_mtx);
424 
425 		/* put the used buffer back into the global pool */
426 		PMCLOG_INIT_BUFFER_DESCRIPTOR(lb);
427 
428 		mtx_lock_spin(&pmc_bufferlist_mtx);
429 		TAILQ_INSERT_HEAD(&pmc_bufferlist, lb, plb_next);
430 		mtx_unlock_spin(&pmc_bufferlist_mtx);
431 
432 		lb = NULL;
433 	}
434 
435 	wakeup_one(po->po_kthread);
436 	po->po_kthread = NULL;
437 
438 	mtx_unlock(&pmc_kthread_mtx);
439 
440 	/* return the current I/O buffer to the global pool */
441 	if (lb) {
442 		PMCLOG_INIT_BUFFER_DESCRIPTOR(lb);
443 
444 		mtx_lock_spin(&pmc_bufferlist_mtx);
445 		TAILQ_INSERT_HEAD(&pmc_bufferlist, lb, plb_next);
446 		mtx_unlock_spin(&pmc_bufferlist_mtx);
447 	}
448 
449 	/*
450 	 * Exit this thread, signalling the waiter
451 	 */
452 
453 	crfree(ownercred);
454 
455 	kproc_exit(0);
456 }
457 
458 /*
459  * Release and log entry and schedule an I/O if needed.
460  */
461 
462 static void
pmclog_release(struct pmc_owner * po)463 pmclog_release(struct pmc_owner *po)
464 {
465 	KASSERT(po->po_curbuf->plb_ptr >= po->po_curbuf->plb_base,
466 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p base=%p", __LINE__,
467 		po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_base));
468 	KASSERT(po->po_curbuf->plb_ptr <= po->po_curbuf->plb_fence,
469 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p fenc=%p", __LINE__,
470 		po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_fence));
471 
472 	/* schedule an I/O if we've filled a buffer */
473 	if (po->po_curbuf->plb_ptr >= po->po_curbuf->plb_fence)
474 		pmclog_schedule_io(po);
475 
476 	mtx_unlock_spin(&po->po_mtx);
477 
478 	PMCDBG1(LOG,REL,1, "po=%p", po);
479 }
480 
481 
482 /*
483  * Attempt to reserve 'length' bytes of space in an owner's log
484  * buffer.  The function returns a pointer to 'length' bytes of space
485  * if there was enough space or returns NULL if no space was
486  * available.  Non-null returns do so with the po mutex locked.  The
487  * caller must invoke pmclog_release() on the pmc owner structure
488  * when done.
489  */
490 
491 static uint32_t *
pmclog_reserve(struct pmc_owner * po,int length)492 pmclog_reserve(struct pmc_owner *po, int length)
493 {
494 	uintptr_t newptr, oldptr;
495 	uint32_t *lh;
496 	struct timespec ts;
497 
498 	PMCDBG2(LOG,ALL,1, "po=%p len=%d", po, length);
499 
500 	KASSERT(length % sizeof(uint32_t) == 0,
501 	    ("[pmclog,%d] length not a multiple of word size", __LINE__));
502 
503 	mtx_lock_spin(&po->po_mtx);
504 
505 	/* No more data when shutdown in progress. */
506 	if (po->po_flags & PMC_PO_SHUTDOWN) {
507 		mtx_unlock_spin(&po->po_mtx);
508 		return (NULL);
509 	}
510 
511 	if (po->po_curbuf == NULL)
512 		if (pmclog_get_buffer(po) != 0) {
513 			mtx_unlock_spin(&po->po_mtx);
514 			return (NULL);
515 		}
516 
517 	KASSERT(po->po_curbuf != NULL,
518 	    ("[pmclog,%d] po=%p no current buffer", __LINE__, po));
519 
520 	KASSERT(po->po_curbuf->plb_ptr >= po->po_curbuf->plb_base &&
521 	    po->po_curbuf->plb_ptr <= po->po_curbuf->plb_fence,
522 	    ("[pmclog,%d] po=%p buffer invariants: ptr=%p base=%p fence=%p",
523 		__LINE__, po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_base,
524 		po->po_curbuf->plb_fence));
525 
526 	oldptr = (uintptr_t) po->po_curbuf->plb_ptr;
527 	newptr = oldptr + length;
528 
529 	KASSERT(oldptr != (uintptr_t) NULL,
530 	    ("[pmclog,%d] po=%p Null log buffer pointer", __LINE__, po));
531 
532 	/*
533 	 * If we have space in the current buffer, return a pointer to
534 	 * available space with the PO structure locked.
535 	 */
536 	if (newptr <= (uintptr_t) po->po_curbuf->plb_fence) {
537 		po->po_curbuf->plb_ptr = (char *) newptr;
538 		goto done;
539 	}
540 
541 	/*
542 	 * Otherwise, schedule the current buffer for output and get a
543 	 * fresh buffer.
544 	 */
545 	pmclog_schedule_io(po);
546 
547 	if (pmclog_get_buffer(po) != 0) {
548 		mtx_unlock_spin(&po->po_mtx);
549 		return (NULL);
550 	}
551 
552 	KASSERT(po->po_curbuf != NULL,
553 	    ("[pmclog,%d] po=%p no current buffer", __LINE__, po));
554 
555 	KASSERT(po->po_curbuf->plb_ptr != NULL,
556 	    ("[pmclog,%d] null return from pmc_get_log_buffer", __LINE__));
557 
558 	KASSERT(po->po_curbuf->plb_ptr == po->po_curbuf->plb_base &&
559 	    po->po_curbuf->plb_ptr <= po->po_curbuf->plb_fence,
560 	    ("[pmclog,%d] po=%p buffer invariants: ptr=%p base=%p fence=%p",
561 		__LINE__, po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_base,
562 		po->po_curbuf->plb_fence));
563 
564 	oldptr = (uintptr_t) po->po_curbuf->plb_ptr;
565 
566  done:
567 	lh = (uint32_t *) oldptr;
568 	lh++;				/* skip header */
569 	getnanotime(&ts);		/* fill in the timestamp */
570 	*lh++ = ts.tv_sec & 0xFFFFFFFF;
571 	*lh++ = ts.tv_nsec & 0xFFFFFFF;
572 	return ((uint32_t *) oldptr);
573 }
574 
575 /*
576  * Schedule an I/O.
577  *
578  * Transfer the current buffer to the helper kthread.
579  */
580 
581 static void
pmclog_schedule_io(struct pmc_owner * po)582 pmclog_schedule_io(struct pmc_owner *po)
583 {
584 	KASSERT(po->po_curbuf != NULL,
585 	    ("[pmclog,%d] schedule_io with null buffer po=%p", __LINE__, po));
586 
587 	KASSERT(po->po_curbuf->plb_ptr >= po->po_curbuf->plb_base,
588 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p base=%p", __LINE__,
589 		po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_base));
590 	KASSERT(po->po_curbuf->plb_ptr <= po->po_curbuf->plb_fence,
591 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p fenc=%p", __LINE__,
592 		po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_fence));
593 
594 	PMCDBG1(LOG,SIO, 1, "po=%p", po);
595 
596 	mtx_assert(&po->po_mtx, MA_OWNED);
597 
598 	/*
599 	 * Add the current buffer to the tail of the buffer list and
600 	 * wakeup the helper.
601 	 */
602 	TAILQ_INSERT_TAIL(&po->po_logbuffers, po->po_curbuf, plb_next);
603 	po->po_curbuf = NULL;
604 	wakeup_one(po);
605 }
606 
607 /*
608  * Stop the helper kthread.
609  */
610 
611 static void
pmclog_stop_kthread(struct pmc_owner * po)612 pmclog_stop_kthread(struct pmc_owner *po)
613 {
614 
615 	mtx_lock(&pmc_kthread_mtx);
616 	po->po_flags &= ~PMC_PO_OWNS_LOGFILE;
617 	if (po->po_kthread != NULL) {
618 		PROC_LOCK(po->po_kthread);
619 		kern_psignal(po->po_kthread, SIGHUP);
620 		PROC_UNLOCK(po->po_kthread);
621 	}
622 	wakeup_one(po);
623 	while (po->po_kthread)
624 		msleep(po->po_kthread, &pmc_kthread_mtx, PPAUSE, "pmckstp", 0);
625 	mtx_unlock(&pmc_kthread_mtx);
626 }
627 
628 /*
629  * Public functions
630  */
631 
632 /*
633  * Configure a log file for pmc owner 'po'.
634  *
635  * Parameter 'logfd' is a file handle referencing an open file in the
636  * owner process.  This file needs to have been opened for writing.
637  */
638 
639 int
pmclog_configure_log(struct pmc_mdep * md,struct pmc_owner * po,int logfd)640 pmclog_configure_log(struct pmc_mdep *md, struct pmc_owner *po, int logfd)
641 {
642 	struct proc *p;
643 	cap_rights_t rights;
644 	int error;
645 
646 	sx_assert(&pmc_sx, SA_XLOCKED);
647 	PMCDBG2(LOG,CFG,1, "config po=%p logfd=%d", po, logfd);
648 
649 	p = po->po_owner;
650 
651 	/* return EBUSY if a log file was already present */
652 	if (po->po_flags & PMC_PO_OWNS_LOGFILE)
653 		return (EBUSY);
654 
655 	KASSERT(po->po_file == NULL,
656 	    ("[pmclog,%d] po=%p file (%p) already present", __LINE__, po,
657 		po->po_file));
658 
659 	/* get a reference to the file state */
660 	error = fget_write(curthread, logfd,
661 	    cap_rights_init(&rights, CAP_WRITE), &po->po_file);
662 	if (error)
663 		goto error;
664 
665 	/* mark process as owning a log file */
666 	po->po_flags |= PMC_PO_OWNS_LOGFILE;
667 
668 	/* mark process as using HWPMCs */
669 	PROC_LOCK(p);
670 	p->p_flag |= P_HWPMC;
671 	PROC_UNLOCK(p);
672 
673 	/* create a log initialization entry */
674 	PMCLOG_RESERVE_WITH_ERROR(po, INITIALIZE,
675 	    sizeof(struct pmclog_initialize));
676 	PMCLOG_EMIT32(PMC_VERSION);
677 	PMCLOG_EMIT32(md->pmd_cputype);
678 	PMCLOG_DESPATCH(po);
679 
680 	return (0);
681 
682  error:
683 	KASSERT(po->po_kthread == NULL, ("[pmclog,%d] po=%p kthread not "
684 	    "stopped", __LINE__, po));
685 
686 	if (po->po_file)
687 		(void) fdrop(po->po_file, curthread);
688 	po->po_file  = NULL;	/* clear file and error state */
689 	po->po_error = 0;
690 	po->po_flags &= ~PMC_PO_OWNS_LOGFILE;
691 
692 	return (error);
693 }
694 
695 
696 /*
697  * De-configure a log file.  This will throw away any buffers queued
698  * for this owner process.
699  */
700 
701 int
pmclog_deconfigure_log(struct pmc_owner * po)702 pmclog_deconfigure_log(struct pmc_owner *po)
703 {
704 	int error;
705 	struct pmclog_buffer *lb;
706 
707 	PMCDBG1(LOG,CFG,1, "de-config po=%p", po);
708 
709 	if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
710 		return (EINVAL);
711 
712 	KASSERT(po->po_sscount == 0,
713 	    ("[pmclog,%d] po=%p still owning SS PMCs", __LINE__, po));
714 	KASSERT(po->po_file != NULL,
715 	    ("[pmclog,%d] po=%p no log file", __LINE__, po));
716 
717 	/* stop the kthread, this will reset the 'OWNS_LOGFILE' flag */
718 	pmclog_stop_kthread(po);
719 
720 	KASSERT(po->po_kthread == NULL,
721 	    ("[pmclog,%d] po=%p kthread not stopped", __LINE__, po));
722 
723 	/* return all queued log buffers to the global pool */
724 	while ((lb = TAILQ_FIRST(&po->po_logbuffers)) != NULL) {
725 		TAILQ_REMOVE(&po->po_logbuffers, lb, plb_next);
726 		PMCLOG_INIT_BUFFER_DESCRIPTOR(lb);
727 		mtx_lock_spin(&pmc_bufferlist_mtx);
728 		TAILQ_INSERT_HEAD(&pmc_bufferlist, lb, plb_next);
729 		mtx_unlock_spin(&pmc_bufferlist_mtx);
730 	}
731 
732 	/* return the 'current' buffer to the global pool */
733 	if ((lb = po->po_curbuf) != NULL) {
734 		PMCLOG_INIT_BUFFER_DESCRIPTOR(lb);
735 		mtx_lock_spin(&pmc_bufferlist_mtx);
736 		TAILQ_INSERT_HEAD(&pmc_bufferlist, lb, plb_next);
737 		mtx_unlock_spin(&pmc_bufferlist_mtx);
738 	}
739 
740 	/* drop a reference to the fd */
741 	if (po->po_file != NULL) {
742 		error = fdrop(po->po_file, curthread);
743 		po->po_file = NULL;
744 	} else
745 		error = 0;
746 	po->po_error = 0;
747 
748 	return (error);
749 }
750 
751 /*
752  * Flush a process' log buffer.
753  */
754 
755 int
pmclog_flush(struct pmc_owner * po)756 pmclog_flush(struct pmc_owner *po)
757 {
758 	int error;
759 	struct pmclog_buffer *lb;
760 
761 	PMCDBG1(LOG,FLS,1, "po=%p", po);
762 
763 	/*
764 	 * If there is a pending error recorded by the logger thread,
765 	 * return that.
766 	 */
767 	if (po->po_error)
768 		return (po->po_error);
769 
770 	error = 0;
771 
772 	/*
773 	 * Check that we do have an active log file.
774 	 */
775 	mtx_lock(&pmc_kthread_mtx);
776 	if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) {
777 		error = EINVAL;
778 		goto error;
779 	}
780 
781 	/*
782 	 * Schedule the current buffer if any and not empty.
783 	 */
784 	mtx_lock_spin(&po->po_mtx);
785 	lb = po->po_curbuf;
786 	if (lb && lb->plb_ptr != lb->plb_base) {
787 		pmclog_schedule_io(po);
788 	} else
789 		error = ENOBUFS;
790 	mtx_unlock_spin(&po->po_mtx);
791 
792  error:
793 	mtx_unlock(&pmc_kthread_mtx);
794 
795 	return (error);
796 }
797 
798 int
pmclog_close(struct pmc_owner * po)799 pmclog_close(struct pmc_owner *po)
800 {
801 
802 	PMCDBG1(LOG,CLO,1, "po=%p", po);
803 
804 	pmclog_process_closelog(po);
805 
806 	mtx_lock(&pmc_kthread_mtx);
807 
808 	/*
809 	 * Schedule the current buffer.
810 	 */
811 	mtx_lock_spin(&po->po_mtx);
812 	if (po->po_curbuf)
813 		pmclog_schedule_io(po);
814 	else
815 		wakeup_one(po);
816 	mtx_unlock_spin(&po->po_mtx);
817 
818 	/*
819 	 * Initiate shutdown: no new data queued,
820 	 * thread will close file on last block.
821 	 */
822 	po->po_flags |= PMC_PO_SHUTDOWN;
823 
824 	mtx_unlock(&pmc_kthread_mtx);
825 
826 	return (0);
827 }
828 
829 void
pmclog_process_callchain(struct pmc * pm,struct pmc_sample * ps)830 pmclog_process_callchain(struct pmc *pm, struct pmc_sample *ps)
831 {
832 	int n, recordlen;
833 	uint32_t flags;
834 	struct pmc_owner *po;
835 
836 	PMCDBG3(LOG,SAM,1,"pm=%p pid=%d n=%d", pm, ps->ps_pid,
837 	    ps->ps_nsamples);
838 
839 	recordlen = offsetof(struct pmclog_callchain, pl_pc) +
840 	    ps->ps_nsamples * sizeof(uintfptr_t);
841 	po = pm->pm_owner;
842 	flags = PMC_CALLCHAIN_TO_CPUFLAGS(ps->ps_cpu,ps->ps_flags);
843 	PMCLOG_RESERVE(po, CALLCHAIN, recordlen);
844 	PMCLOG_EMIT32(ps->ps_pid);
845 	PMCLOG_EMIT32(pm->pm_id);
846 	PMCLOG_EMIT32(flags);
847 	for (n = 0; n < ps->ps_nsamples; n++)
848 		PMCLOG_EMITADDR(ps->ps_pc[n]);
849 	PMCLOG_DESPATCH(po);
850 }
851 
852 void
pmclog_process_closelog(struct pmc_owner * po)853 pmclog_process_closelog(struct pmc_owner *po)
854 {
855 	PMCLOG_RESERVE(po,CLOSELOG,sizeof(struct pmclog_closelog));
856 	PMCLOG_DESPATCH(po);
857 }
858 
859 void
pmclog_process_dropnotify(struct pmc_owner * po)860 pmclog_process_dropnotify(struct pmc_owner *po)
861 {
862 	PMCLOG_RESERVE(po,DROPNOTIFY,sizeof(struct pmclog_dropnotify));
863 	PMCLOG_DESPATCH(po);
864 }
865 
866 void
pmclog_process_map_in(struct pmc_owner * po,pid_t pid,uintfptr_t start,const char * path)867 pmclog_process_map_in(struct pmc_owner *po, pid_t pid, uintfptr_t start,
868     const char *path)
869 {
870 	int pathlen, recordlen;
871 
872 	KASSERT(path != NULL, ("[pmclog,%d] map-in, null path", __LINE__));
873 
874 	pathlen = strlen(path) + 1;	/* #bytes for path name */
875 	recordlen = offsetof(struct pmclog_map_in, pl_pathname) +
876 	    pathlen;
877 
878 	PMCLOG_RESERVE(po, MAP_IN, recordlen);
879 	PMCLOG_EMIT32(pid);
880 	PMCLOG_EMITADDR(start);
881 	PMCLOG_EMITSTRING(path,pathlen);
882 	PMCLOG_DESPATCH(po);
883 }
884 
885 void
pmclog_process_map_out(struct pmc_owner * po,pid_t pid,uintfptr_t start,uintfptr_t end)886 pmclog_process_map_out(struct pmc_owner *po, pid_t pid, uintfptr_t start,
887     uintfptr_t end)
888 {
889 	KASSERT(start <= end, ("[pmclog,%d] start > end", __LINE__));
890 
891 	PMCLOG_RESERVE(po, MAP_OUT, sizeof(struct pmclog_map_out));
892 	PMCLOG_EMIT32(pid);
893 	PMCLOG_EMITADDR(start);
894 	PMCLOG_EMITADDR(end);
895 	PMCLOG_DESPATCH(po);
896 }
897 
898 void
pmclog_process_pmcallocate(struct pmc * pm)899 pmclog_process_pmcallocate(struct pmc *pm)
900 {
901 	struct pmc_owner *po;
902 	struct pmc_soft *ps;
903 
904 	po = pm->pm_owner;
905 
906 	PMCDBG1(LOG,ALL,1, "pm=%p", pm);
907 
908 	if (PMC_TO_CLASS(pm) == PMC_CLASS_SOFT) {
909 		PMCLOG_RESERVE(po, PMCALLOCATEDYN,
910 		    sizeof(struct pmclog_pmcallocatedyn));
911 		PMCLOG_EMIT32(pm->pm_id);
912 		PMCLOG_EMIT32(pm->pm_event);
913 		PMCLOG_EMIT32(pm->pm_flags);
914 		ps = pmc_soft_ev_acquire(pm->pm_event);
915 		if (ps != NULL)
916 			PMCLOG_EMITSTRING(ps->ps_ev.pm_ev_name,PMC_NAME_MAX);
917 		else
918 			PMCLOG_EMITNULLSTRING(PMC_NAME_MAX);
919 		pmc_soft_ev_release(ps);
920 		PMCLOG_DESPATCH(po);
921 	} else {
922 		PMCLOG_RESERVE(po, PMCALLOCATE,
923 		    sizeof(struct pmclog_pmcallocate));
924 		PMCLOG_EMIT32(pm->pm_id);
925 		PMCLOG_EMIT32(pm->pm_event);
926 		PMCLOG_EMIT32(pm->pm_flags);
927 		PMCLOG_DESPATCH(po);
928 	}
929 }
930 
931 void
pmclog_process_pmcattach(struct pmc * pm,pid_t pid,char * path)932 pmclog_process_pmcattach(struct pmc *pm, pid_t pid, char *path)
933 {
934 	int pathlen, recordlen;
935 	struct pmc_owner *po;
936 
937 	PMCDBG2(LOG,ATT,1,"pm=%p pid=%d", pm, pid);
938 
939 	po = pm->pm_owner;
940 
941 	pathlen = strlen(path) + 1;	/* #bytes for the string */
942 	recordlen = offsetof(struct pmclog_pmcattach, pl_pathname) + pathlen;
943 
944 	PMCLOG_RESERVE(po, PMCATTACH, recordlen);
945 	PMCLOG_EMIT32(pm->pm_id);
946 	PMCLOG_EMIT32(pid);
947 	PMCLOG_EMITSTRING(path, pathlen);
948 	PMCLOG_DESPATCH(po);
949 }
950 
951 void
pmclog_process_pmcdetach(struct pmc * pm,pid_t pid)952 pmclog_process_pmcdetach(struct pmc *pm, pid_t pid)
953 {
954 	struct pmc_owner *po;
955 
956 	PMCDBG2(LOG,ATT,1,"!pm=%p pid=%d", pm, pid);
957 
958 	po = pm->pm_owner;
959 
960 	PMCLOG_RESERVE(po, PMCDETACH, sizeof(struct pmclog_pmcdetach));
961 	PMCLOG_EMIT32(pm->pm_id);
962 	PMCLOG_EMIT32(pid);
963 	PMCLOG_DESPATCH(po);
964 }
965 
966 /*
967  * Log a context switch event to the log file.
968  */
969 
970 void
pmclog_process_proccsw(struct pmc * pm,struct pmc_process * pp,pmc_value_t v)971 pmclog_process_proccsw(struct pmc *pm, struct pmc_process *pp, pmc_value_t v)
972 {
973 	struct pmc_owner *po;
974 
975 	KASSERT(pm->pm_flags & PMC_F_LOG_PROCCSW,
976 	    ("[pmclog,%d] log-process-csw called gratuitously", __LINE__));
977 
978 	PMCDBG3(LOG,SWO,1,"pm=%p pid=%d v=%jx", pm, pp->pp_proc->p_pid,
979 	    v);
980 
981 	po = pm->pm_owner;
982 
983 	PMCLOG_RESERVE(po, PROCCSW, sizeof(struct pmclog_proccsw));
984 	PMCLOG_EMIT32(pm->pm_id);
985 	PMCLOG_EMIT64(v);
986 	PMCLOG_EMIT32(pp->pp_proc->p_pid);
987 	PMCLOG_DESPATCH(po);
988 }
989 
990 void
pmclog_process_procexec(struct pmc_owner * po,pmc_id_t pmid,pid_t pid,uintfptr_t startaddr,char * path)991 pmclog_process_procexec(struct pmc_owner *po, pmc_id_t pmid, pid_t pid,
992     uintfptr_t startaddr, char *path)
993 {
994 	int pathlen, recordlen;
995 
996 	PMCDBG3(LOG,EXC,1,"po=%p pid=%d path=\"%s\"", po, pid, path);
997 
998 	pathlen   = strlen(path) + 1;	/* #bytes for the path */
999 	recordlen = offsetof(struct pmclog_procexec, pl_pathname) + pathlen;
1000 
1001 	PMCLOG_RESERVE(po, PROCEXEC, recordlen);
1002 	PMCLOG_EMIT32(pid);
1003 	PMCLOG_EMITADDR(startaddr);
1004 	PMCLOG_EMIT32(pmid);
1005 	PMCLOG_EMITSTRING(path,pathlen);
1006 	PMCLOG_DESPATCH(po);
1007 }
1008 
1009 /*
1010  * Log a process exit event (and accumulated pmc value) to the log file.
1011  */
1012 
1013 void
pmclog_process_procexit(struct pmc * pm,struct pmc_process * pp)1014 pmclog_process_procexit(struct pmc *pm, struct pmc_process *pp)
1015 {
1016 	int ri;
1017 	struct pmc_owner *po;
1018 
1019 	ri = PMC_TO_ROWINDEX(pm);
1020 	PMCDBG3(LOG,EXT,1,"pm=%p pid=%d v=%jx", pm, pp->pp_proc->p_pid,
1021 	    pp->pp_pmcs[ri].pp_pmcval);
1022 
1023 	po = pm->pm_owner;
1024 
1025 	PMCLOG_RESERVE(po, PROCEXIT, sizeof(struct pmclog_procexit));
1026 	PMCLOG_EMIT32(pm->pm_id);
1027 	PMCLOG_EMIT64(pp->pp_pmcs[ri].pp_pmcval);
1028 	PMCLOG_EMIT32(pp->pp_proc->p_pid);
1029 	PMCLOG_DESPATCH(po);
1030 }
1031 
1032 /*
1033  * Log a fork event.
1034  */
1035 
1036 void
pmclog_process_procfork(struct pmc_owner * po,pid_t oldpid,pid_t newpid)1037 pmclog_process_procfork(struct pmc_owner *po, pid_t oldpid, pid_t newpid)
1038 {
1039 	PMCLOG_RESERVE(po, PROCFORK, sizeof(struct pmclog_procfork));
1040 	PMCLOG_EMIT32(oldpid);
1041 	PMCLOG_EMIT32(newpid);
1042 	PMCLOG_DESPATCH(po);
1043 }
1044 
1045 /*
1046  * Log a process exit event of the form suitable for system-wide PMCs.
1047  */
1048 
1049 void
pmclog_process_sysexit(struct pmc_owner * po,pid_t pid)1050 pmclog_process_sysexit(struct pmc_owner *po, pid_t pid)
1051 {
1052 	PMCLOG_RESERVE(po, SYSEXIT, sizeof(struct pmclog_sysexit));
1053 	PMCLOG_EMIT32(pid);
1054 	PMCLOG_DESPATCH(po);
1055 }
1056 
1057 /*
1058  * Write a user log entry.
1059  */
1060 
1061 int
pmclog_process_userlog(struct pmc_owner * po,struct pmc_op_writelog * wl)1062 pmclog_process_userlog(struct pmc_owner *po, struct pmc_op_writelog *wl)
1063 {
1064 	int error;
1065 
1066 	PMCDBG2(LOG,WRI,1, "writelog po=%p ud=0x%x", po, wl->pm_userdata);
1067 
1068 	error = 0;
1069 
1070 	PMCLOG_RESERVE_WITH_ERROR(po, USERDATA,
1071 	    sizeof(struct pmclog_userdata));
1072 	PMCLOG_EMIT32(wl->pm_userdata);
1073 	PMCLOG_DESPATCH(po);
1074 
1075  error:
1076 	return (error);
1077 }
1078 
1079 /*
1080  * Initialization.
1081  *
1082  * Create a pool of log buffers and initialize mutexes.
1083  */
1084 
1085 void
pmclog_initialize()1086 pmclog_initialize()
1087 {
1088 	int n;
1089 	struct pmclog_buffer *plb;
1090 
1091 	if (pmclog_buffer_size <= 0) {
1092 		(void) printf("hwpmc: tunable logbuffersize=%d must be "
1093 		    "greater than zero.\n", pmclog_buffer_size);
1094 		pmclog_buffer_size = PMC_LOG_BUFFER_SIZE;
1095 	}
1096 
1097 	if (pmc_nlogbuffers <= 0) {
1098 		(void) printf("hwpmc: tunable nlogbuffers=%d must be greater "
1099 		    "than zero.\n", pmc_nlogbuffers);
1100 		pmc_nlogbuffers = PMC_NLOGBUFFERS;
1101 	}
1102 
1103 	/* create global pool of log buffers */
1104 	for (n = 0; n < pmc_nlogbuffers; n++) {
1105 		plb = malloc(1024 * pmclog_buffer_size, M_PMC,
1106 		    M_WAITOK|M_ZERO);
1107 		PMCLOG_INIT_BUFFER_DESCRIPTOR(plb);
1108 		TAILQ_INSERT_HEAD(&pmc_bufferlist, plb, plb_next);
1109 	}
1110 	mtx_init(&pmc_bufferlist_mtx, "pmc-buffer-list", "pmc-leaf",
1111 	    MTX_SPIN);
1112 	mtx_init(&pmc_kthread_mtx, "pmc-kthread", "pmc-sleep", MTX_DEF);
1113 }
1114 
1115 /*
1116  * Shutdown logging.
1117  *
1118  * Destroy mutexes and release memory back the to free pool.
1119  */
1120 
1121 void
pmclog_shutdown()1122 pmclog_shutdown()
1123 {
1124 	struct pmclog_buffer *plb;
1125 
1126 	mtx_destroy(&pmc_kthread_mtx);
1127 	mtx_destroy(&pmc_bufferlist_mtx);
1128 
1129 	while ((plb = TAILQ_FIRST(&pmc_bufferlist)) != NULL) {
1130 		TAILQ_REMOVE(&pmc_bufferlist, plb, plb_next);
1131 		free(plb, M_PMC);
1132 	}
1133 }
1134