xref: /freebsd/sys/dev/hwpmc/hwpmc_logging.c (revision 3c29734502c330fc8f8e345213791722aeaf55c1)
1 /*-
2  * SPDX-License-Identifier: BSD-2-Clause
3  *
4  * Copyright (c) 2005-2007 Joseph Koshy
5  * Copyright (c) 2007 The FreeBSD Foundation
6  * Copyright (c) 2018 Matthew Macy
7  * All rights reserved.
8  *
9  * Portions of this software were developed by A. Joseph Koshy under
10  * sponsorship from the FreeBSD Foundation and Google, Inc.
11  *
12  * Redistribution and use in source and binary forms, with or without
13  * modification, are permitted provided that the following conditions
14  * are met:
15  * 1. Redistributions of source code must retain the above copyright
16  *    notice, this list of conditions and the following disclaimer.
17  * 2. Redistributions in binary form must reproduce the above copyright
18  *    notice, this list of conditions and the following disclaimer in the
19  *    documentation and/or other materials provided with the distribution.
20  *
21  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
24  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
31  * SUCH DAMAGE.
32  *
33  */
34 
35 /*
36  * Logging code for hwpmc(4)
37  */
38 
39 #include <sys/param.h>
40 #include <sys/capsicum.h>
41 #include <sys/domainset.h>
42 #include <sys/file.h>
43 #include <sys/kernel.h>
44 #include <sys/kthread.h>
45 #include <sys/lock.h>
46 #include <sys/module.h>
47 #include <sys/mutex.h>
48 #include <sys/pmc.h>
49 #include <sys/pmckern.h>
50 #include <sys/pmclog.h>
51 #include <sys/proc.h>
52 #include <sys/sched.h>
53 #include <sys/signalvar.h>
54 #include <sys/smp.h>
55 #include <sys/syscallsubr.h>
56 #include <sys/sysctl.h>
57 #include <sys/systm.h>
58 #include <sys/uio.h>
59 #include <sys/unistd.h>
60 #include <sys/vnode.h>
61 
62 #if defined(__i386__) || defined(__amd64__)
63 #include <machine/clock.h>
64 #endif
65 
66 #define curdomain PCPU_GET(domain)
67 
68 /*
69  * Sysctl tunables
70  */
71 
72 SYSCTL_DECL(_kern_hwpmc);
73 
74 /*
75  * kern.hwpmc.logbuffersize -- size of the per-cpu owner buffers.
76  */
77 
78 static int pmclog_buffer_size = PMC_LOG_BUFFER_SIZE;
79 SYSCTL_INT(_kern_hwpmc, OID_AUTO, logbuffersize, CTLFLAG_RDTUN,
80     &pmclog_buffer_size, 0, "size of log buffers in kilobytes");
81 
82 /*
83  * kern.hwpmc.nbuffers_pcpu -- number of global log buffers
84  */
85 
86 static int pmc_nlogbuffers_pcpu = PMC_NLOGBUFFERS_PCPU;
87 SYSCTL_INT(_kern_hwpmc, OID_AUTO, nbuffers_pcpu, CTLFLAG_RDTUN,
88     &pmc_nlogbuffers_pcpu, 0, "number of log buffers per cpu");
89 
90 /*
91  * Global log buffer list and associated spin lock.
92  */
93 
94 static struct mtx pmc_kthread_mtx;	/* sleep lock */
95 
96 #define	PMCLOG_INIT_BUFFER_DESCRIPTOR(D, buf, domain) do {		\
97 	(D)->plb_fence = ((char *)(buf)) + 1024 * pmclog_buffer_size;	\
98 	(D)->plb_base = (D)->plb_ptr = ((char *)(buf));			\
99 	(D)->plb_domain = domain;					\
100 } while (0)
101 
102 #define	PMCLOG_RESET_BUFFER_DESCRIPTOR(D) do {				\
103 	(D)->plb_ptr = (D)->plb_base;					\
104 } while (0)
105 
106 /*
107  * Log file record constructors.
108  */
109 #define	_PMCLOG_TO_HEADER(T, L)						\
110 	((PMCLOG_HEADER_MAGIC << 24) | (T << 16) | ((L) & 0xFFFF))
111 
112 /* reserve LEN bytes of space and initialize the entry header */
113 #define	_PMCLOG_RESERVE_SAFE(PO, TYPE, LEN, ACTION, TSC) do {		\
114 	uint32_t *_le;							\
115 	int _len = roundup((LEN), sizeof(uint32_t));			\
116 	struct pmclog_header *ph;					\
117 									\
118 	if ((_le = pmclog_reserve((PO), _len)) == NULL) {		\
119 		ACTION;							\
120 	}								\
121 	ph = (struct pmclog_header *)_le;				\
122 	ph->pl_header =_PMCLOG_TO_HEADER(TYPE,_len);			\
123 	ph->pl_tsc = (TSC);						\
124 	_le += sizeof(*ph) / 4	/* skip over timestamp */
125 
126 /* reserve LEN bytes of space and initialize the entry header */
127 #define	_PMCLOG_RESERVE(PO, TYPE, LEN, ACTION) do {			\
128 	uint32_t *_le;							\
129 	int _len = roundup((LEN), sizeof(uint32_t));			\
130 	uint64_t tsc;							\
131 	struct pmclog_header *ph;					\
132 									\
133 	tsc = pmc_rdtsc();						\
134 	spinlock_enter();						\
135 	if ((_le = pmclog_reserve((PO), _len)) == NULL) {		\
136 		spinlock_exit();					\
137 		ACTION;							\
138 	}								\
139 	ph = (struct pmclog_header *)_le;				\
140 	ph->pl_header =_PMCLOG_TO_HEADER(TYPE,_len);			\
141 	ph->pl_tsc = tsc;						\
142 
143 #define	PMCLOG_RESERVE_SAFE(P, T, L, TSC)				\
144 	_PMCLOG_RESERVE_SAFE(P, T, L, return, TSC)
145 #define	PMCLOG_RESERVE(P,T,L)						\
146 	_PMCLOG_RESERVE(P, T, L, return)
147 #define	PMCLOG_RESERVE_WITH_ERROR(P, T, L)				\
148 	_PMCLOG_RESERVE(P, T, L, error = ENOMEM; goto error)
149 
150 #define	PMCLOG_EMIT32(V)	do { *_le++ = (V); } while (0)
151 #define	PMCLOG_EMIT64(V)	do { 					\
152 	*_le++ = (uint32_t) ((V) & 0xFFFFFFFF);				\
153 	*_le++ = (uint32_t) (((V) >> 32) & 0xFFFFFFFF);			\
154 } while (0)
155 
156 
157 /* Emit a string.  Caution: does NOT update _le, so needs to be last */
158 #define	PMCLOG_EMITSTRING(S,L)	do {					\
159 	bcopy((S), _le, (L));						\
160 } while (0)
161 #define	PMCLOG_EMITNULLSTRING(L) do {					\
162 	bzero(_le, (L));						\
163 } while (0)
164 
165 #define	PMCLOG_DESPATCH_SAFE(PO)					\
166 	pmclog_release((PO));						\
167 } while (0)
168 
169 #define	PMCLOG_DESPATCH_SCHED_LOCK(PO)					\
170 	pmclog_release_flags((PO), 0);					\
171 } while (0)
172 
173 #define	PMCLOG_DESPATCH(PO)						\
174 	pmclog_release((PO));						\
175 	spinlock_exit();						\
176 } while (0)
177 
178 #define	PMCLOG_DESPATCH_SYNC(PO)					\
179 	pmclog_schedule_io((PO), 1);					\
180 	spinlock_exit();						\
181 } while (0)
182 
183 #define TSDELTA 4
184 /*
185  * Assertions about the log file format.
186  */
187 CTASSERT(sizeof(struct pmclog_callchain) == 7*4 + TSDELTA +
188     PMC_CALLCHAIN_DEPTH_MAX*sizeof(uintfptr_t));
189 CTASSERT(sizeof(struct pmclog_closelog) == 3*4 + TSDELTA);
190 CTASSERT(sizeof(struct pmclog_dropnotify) == 3*4 + TSDELTA);
191 CTASSERT(sizeof(struct pmclog_map_in) == PATH_MAX + TSDELTA +
192     5*4 + sizeof(uintfptr_t));
193 CTASSERT(offsetof(struct pmclog_map_in,pl_pathname) ==
194     5*4 + TSDELTA + sizeof(uintfptr_t));
195 CTASSERT(sizeof(struct pmclog_map_out) == 5*4 + 2*sizeof(uintfptr_t) + TSDELTA);
196 CTASSERT(sizeof(struct pmclog_pmcallocate) == 9*4 + TSDELTA);
197 CTASSERT(sizeof(struct pmclog_pmcattach) == 5*4 + PATH_MAX + TSDELTA);
198 CTASSERT(offsetof(struct pmclog_pmcattach,pl_pathname) == 5*4 + TSDELTA);
199 CTASSERT(sizeof(struct pmclog_pmcdetach) == 5*4 + TSDELTA);
200 CTASSERT(sizeof(struct pmclog_proccsw) == 7*4 + 8 + TSDELTA);
201 CTASSERT(sizeof(struct pmclog_procexec) == 5*4 + PATH_MAX +
202     2*sizeof(uintptr_t) + TSDELTA);
203 CTASSERT(offsetof(struct pmclog_procexec,pl_pathname) == 5*4 + TSDELTA +
204     2*sizeof(uintptr_t));
205 CTASSERT(sizeof(struct pmclog_procexit) == 5*4 + 8 + TSDELTA);
206 CTASSERT(sizeof(struct pmclog_procfork) == 5*4 + TSDELTA);
207 CTASSERT(sizeof(struct pmclog_sysexit) == 6*4);
208 CTASSERT(sizeof(struct pmclog_userdata) == 6*4);
209 
210 /*
211  * Log buffer structure
212  */
213 
214 struct pmclog_buffer {
215 	TAILQ_ENTRY(pmclog_buffer) plb_next;
216 	char 		*plb_base;
217 	char		*plb_ptr;
218 	char 		*plb_fence;
219 	uint16_t	 plb_domain;
220 } __aligned(CACHE_LINE_SIZE);
221 
222 /*
223  * Prototypes
224  */
225 
226 static int pmclog_get_buffer(struct pmc_owner *po);
227 static void pmclog_loop(void *arg);
228 static void pmclog_release(struct pmc_owner *po);
229 static uint32_t *pmclog_reserve(struct pmc_owner *po, int length);
230 static void pmclog_schedule_io(struct pmc_owner *po, int wakeup);
231 static void pmclog_schedule_all(struct pmc_owner *po);
232 static void pmclog_stop_kthread(struct pmc_owner *po);
233 
234 /*
235  * Helper functions
236  */
237 
238 static inline void
pmc_plb_rele_unlocked(struct pmclog_buffer * plb)239 pmc_plb_rele_unlocked(struct pmclog_buffer *plb)
240 {
241 	TAILQ_INSERT_HEAD(&pmc_dom_hdrs[plb->plb_domain]->pdbh_head, plb, plb_next);
242 }
243 
244 static inline void
pmc_plb_rele(struct pmclog_buffer * plb)245 pmc_plb_rele(struct pmclog_buffer *plb)
246 {
247 	mtx_lock_spin(&pmc_dom_hdrs[plb->plb_domain]->pdbh_mtx);
248 	pmc_plb_rele_unlocked(plb);
249 	mtx_unlock_spin(&pmc_dom_hdrs[plb->plb_domain]->pdbh_mtx);
250 }
251 
252 /*
253  * Get a log buffer
254  */
255 static int
pmclog_get_buffer(struct pmc_owner * po)256 pmclog_get_buffer(struct pmc_owner *po)
257 {
258 	struct pmclog_buffer *plb;
259 	int domain;
260 
261 	KASSERT(po->po_curbuf[curcpu] == NULL,
262 	    ("[pmclog,%d] po=%p current buffer still valid", __LINE__, po));
263 
264 	domain = curdomain;
265 	MPASS(pmc_dom_hdrs[domain]);
266 	mtx_lock_spin(&pmc_dom_hdrs[domain]->pdbh_mtx);
267 	if ((plb = TAILQ_FIRST(&pmc_dom_hdrs[domain]->pdbh_head)) != NULL)
268 		TAILQ_REMOVE(&pmc_dom_hdrs[domain]->pdbh_head, plb, plb_next);
269 	mtx_unlock_spin(&pmc_dom_hdrs[domain]->pdbh_mtx);
270 
271 	PMCDBG2(LOG,GTB,1, "po=%p plb=%p", po, plb);
272 
273 #ifdef	HWPMC_DEBUG
274 	if (plb)
275 		KASSERT(plb->plb_ptr == plb->plb_base &&
276 		    plb->plb_base < plb->plb_fence,
277 		    ("[pmclog,%d] po=%p buffer invariants: ptr=%p "
278 		    "base=%p fence=%p", __LINE__, po, plb->plb_ptr,
279 		    plb->plb_base, plb->plb_fence));
280 #endif
281 
282 	po->po_curbuf[curcpu] = plb;
283 
284 	/* update stats */
285 	counter_u64_add(pmc_stats.pm_buffer_requests, 1);
286 	if (plb == NULL)
287 		counter_u64_add(pmc_stats.pm_buffer_requests_failed, 1);
288 
289 	return (plb ? 0 : ENOMEM);
290 }
291 
292 struct pmclog_proc_init_args {
293 	struct proc *kthr;
294 	struct pmc_owner *po;
295 	bool exit;
296 	bool acted;
297 };
298 
299 int
pmclog_proc_create(struct thread * td,void ** handlep)300 pmclog_proc_create(struct thread *td, void **handlep)
301 {
302 	struct pmclog_proc_init_args *ia;
303 	int error;
304 
305 	ia = malloc(sizeof(*ia), M_TEMP, M_WAITOK | M_ZERO);
306 	error = kproc_create(pmclog_loop, ia, &ia->kthr,
307 	    RFHIGHPID, 0, "hwpmc: proc(%d)", td->td_proc->p_pid);
308 	if (error == 0)
309 		*handlep = ia;
310 	return (error);
311 }
312 
313 void
pmclog_proc_ignite(void * handle,struct pmc_owner * po)314 pmclog_proc_ignite(void *handle, struct pmc_owner *po)
315 {
316 	struct pmclog_proc_init_args *ia;
317 
318 	ia = handle;
319 	mtx_lock(&pmc_kthread_mtx);
320 	MPASS(!ia->acted);
321 	MPASS(ia->po == NULL);
322 	MPASS(!ia->exit);
323 	MPASS(ia->kthr != NULL);
324 	if (po == NULL) {
325 		ia->exit = true;
326 	} else {
327 		ia->po = po;
328 		KASSERT(po->po_kthread == NULL,
329 		    ("[pmclog,%d] po=%p kthread (%p) already present",
330 		    __LINE__, po, po->po_kthread));
331 		po->po_kthread = ia->kthr;
332 	}
333 	wakeup(ia);
334 	while (!ia->acted)
335 		msleep(ia, &pmc_kthread_mtx, PWAIT, "pmclogw", 0);
336 	mtx_unlock(&pmc_kthread_mtx);
337 	free(ia, M_TEMP);
338 }
339 
340 /*
341  * Log handler loop.
342  *
343  * This function is executed by each pmc owner's helper thread.
344  */
345 static void
pmclog_loop(void * arg)346 pmclog_loop(void *arg)
347 {
348 	struct pmclog_proc_init_args *ia;
349 	struct pmc_owner *po;
350 	struct pmclog_buffer *lb;
351 	struct proc *p;
352 	struct ucred *ownercred;
353 	struct ucred *mycred;
354 	struct thread *td;
355 	sigset_t unb;
356 	struct uio auio;
357 	struct iovec aiov;
358 	size_t nbytes;
359 	int error;
360 
361 	td = curthread;
362 
363 	SIGEMPTYSET(unb);
364 	SIGADDSET(unb, SIGHUP);
365 	(void)kern_sigprocmask(td, SIG_UNBLOCK, &unb, NULL, 0);
366 
367 	ia = arg;
368 	MPASS(ia->kthr == curproc);
369 	MPASS(!ia->acted);
370 	mtx_lock(&pmc_kthread_mtx);
371 	while (ia->po == NULL && !ia->exit)
372 		msleep(ia, &pmc_kthread_mtx, PWAIT, "pmclogi", 0);
373 	if (ia->exit) {
374 		ia->acted = true;
375 		wakeup(ia);
376 		mtx_unlock(&pmc_kthread_mtx);
377 		kproc_exit(0);
378 	}
379 	MPASS(ia->po != NULL);
380 	po = ia->po;
381 	ia->acted = true;
382 	wakeup(ia);
383 	mtx_unlock(&pmc_kthread_mtx);
384 	ia = NULL;
385 
386 	p = po->po_owner;
387 	mycred = td->td_ucred;
388 
389 	PROC_LOCK(p);
390 	ownercred = crhold(p->p_ucred);
391 	PROC_UNLOCK(p);
392 
393 	PMCDBG2(LOG,INI,1, "po=%p kt=%p", po, po->po_kthread);
394 	KASSERT(po->po_kthread == curthread->td_proc,
395 	    ("[pmclog,%d] proc mismatch po=%p po/kt=%p curproc=%p", __LINE__,
396 		po, po->po_kthread, curthread->td_proc));
397 
398 	lb = NULL;
399 
400 
401 	/*
402 	 * Loop waiting for I/O requests to be added to the owner
403 	 * struct's queue.  The loop is exited when the log file
404 	 * is deconfigured.
405 	 */
406 
407 	mtx_lock(&pmc_kthread_mtx);
408 
409 	for (;;) {
410 
411 		/* check if we've been asked to exit */
412 		if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
413 			break;
414 
415 		if (lb == NULL) { /* look for a fresh buffer to write */
416 			mtx_lock_spin(&po->po_mtx);
417 			if ((lb = TAILQ_FIRST(&po->po_logbuffers)) == NULL) {
418 				mtx_unlock_spin(&po->po_mtx);
419 
420 				/* No more buffers and shutdown required. */
421 				if (po->po_flags & PMC_PO_SHUTDOWN)
422 					break;
423 
424 				(void) msleep(po, &pmc_kthread_mtx, PWAIT,
425 				    "pmcloop", 250);
426 				continue;
427 			}
428 
429 			TAILQ_REMOVE(&po->po_logbuffers, lb, plb_next);
430 			mtx_unlock_spin(&po->po_mtx);
431 		}
432 
433 		mtx_unlock(&pmc_kthread_mtx);
434 
435 		/* process the request */
436 		PMCDBG3(LOG,WRI,2, "po=%p base=%p ptr=%p", po,
437 		    lb->plb_base, lb->plb_ptr);
438 		/* change our thread's credentials before issuing the I/O */
439 
440 		aiov.iov_base = lb->plb_base;
441 		aiov.iov_len  = nbytes = lb->plb_ptr - lb->plb_base;
442 
443 		auio.uio_iov    = &aiov;
444 		auio.uio_iovcnt = 1;
445 		auio.uio_offset = -1;
446 		auio.uio_resid  = nbytes;
447 		auio.uio_rw     = UIO_WRITE;
448 		auio.uio_segflg = UIO_SYSSPACE;
449 		auio.uio_td     = td;
450 
451 		/* switch thread credentials -- see kern_ktrace.c */
452 		td->td_ucred = ownercred;
453 		error = fo_write(po->po_file, &auio, ownercred, 0, td);
454 		td->td_ucred = mycred;
455 
456 		if (error) {
457 			/* XXX some errors are recoverable */
458 			/* send a SIGIO to the owner and exit */
459 			PROC_LOCK(p);
460 			kern_psignal(p, SIGIO);
461 			PROC_UNLOCK(p);
462 
463 			mtx_lock(&pmc_kthread_mtx);
464 
465 			po->po_error = error; /* save for flush log */
466 
467 			PMCDBG2(LOG,WRI,2, "po=%p error=%d", po, error);
468 
469 			break;
470 		}
471 
472 		mtx_lock(&pmc_kthread_mtx);
473 
474 		/* put the used buffer back into the global pool */
475 		PMCLOG_RESET_BUFFER_DESCRIPTOR(lb);
476 
477 		pmc_plb_rele(lb);
478 		lb = NULL;
479 	}
480 
481 	wakeup_one(po->po_kthread);
482 	po->po_kthread = NULL;
483 
484 	mtx_unlock(&pmc_kthread_mtx);
485 
486 	/* return the current I/O buffer to the global pool */
487 	if (lb) {
488 		PMCLOG_RESET_BUFFER_DESCRIPTOR(lb);
489 
490 		pmc_plb_rele(lb);
491 	}
492 
493 	/*
494 	 * Exit this thread, signalling the waiter
495 	 */
496 
497 	crfree(ownercred);
498 
499 	kproc_exit(0);
500 }
501 
502 /*
503  * Release and log entry and schedule an I/O if needed.
504  */
505 
506 static void
pmclog_release_flags(struct pmc_owner * po,int wakeup)507 pmclog_release_flags(struct pmc_owner *po, int wakeup)
508 {
509 	struct pmclog_buffer *plb;
510 
511 	plb = po->po_curbuf[curcpu];
512 	KASSERT(plb->plb_ptr >= plb->plb_base,
513 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p base=%p", __LINE__,
514 		po, plb->plb_ptr, plb->plb_base));
515 	KASSERT(plb->plb_ptr <= plb->plb_fence,
516 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p fenc=%p", __LINE__,
517 		po, plb->plb_ptr, plb->plb_fence));
518 
519 	/* schedule an I/O if we've filled a buffer */
520 	if (plb->plb_ptr >= plb->plb_fence)
521 		pmclog_schedule_io(po, wakeup);
522 
523 	PMCDBG1(LOG,REL,1, "po=%p", po);
524 }
525 
526 static void
pmclog_release(struct pmc_owner * po)527 pmclog_release(struct pmc_owner *po)
528 {
529 
530 	pmclog_release_flags(po, 1);
531 }
532 
533 
534 /*
535  * Attempt to reserve 'length' bytes of space in an owner's log
536  * buffer.  The function returns a pointer to 'length' bytes of space
537  * if there was enough space or returns NULL if no space was
538  * available.  Non-null returns do so with the po mutex locked.  The
539  * caller must invoke pmclog_release() on the pmc owner structure
540  * when done.
541  */
542 
543 static uint32_t *
pmclog_reserve(struct pmc_owner * po,int length)544 pmclog_reserve(struct pmc_owner *po, int length)
545 {
546 	uintptr_t newptr, oldptr __diagused;
547 	struct pmclog_buffer *plb, **pplb;
548 
549 	PMCDBG2(LOG,ALL,1, "po=%p len=%d", po, length);
550 
551 	KASSERT(length % sizeof(uint32_t) == 0,
552 	    ("[pmclog,%d] length not a multiple of word size", __LINE__));
553 
554 	/* No more data when shutdown in progress. */
555 	if (po->po_flags & PMC_PO_SHUTDOWN)
556 		return (NULL);
557 
558 	pplb = &po->po_curbuf[curcpu];
559 	if (*pplb == NULL && pmclog_get_buffer(po) != 0)
560 		goto fail;
561 
562 	KASSERT(*pplb != NULL,
563 	    ("[pmclog,%d] po=%p no current buffer", __LINE__, po));
564 
565 	plb = *pplb;
566 	KASSERT(plb->plb_ptr >= plb->plb_base &&
567 	    plb->plb_ptr <= plb->plb_fence,
568 	    ("[pmclog,%d] po=%p buffer invariants: ptr=%p base=%p fence=%p",
569 		__LINE__, po, plb->plb_ptr, plb->plb_base,
570 		plb->plb_fence));
571 
572 	oldptr = (uintptr_t) plb->plb_ptr;
573 	newptr = oldptr + length;
574 
575 	KASSERT(oldptr != (uintptr_t) NULL,
576 	    ("[pmclog,%d] po=%p Null log buffer pointer", __LINE__, po));
577 
578 	/*
579 	 * If we have space in the current buffer, return a pointer to
580 	 * available space with the PO structure locked.
581 	 */
582 	if (newptr <= (uintptr_t) plb->plb_fence) {
583 		plb->plb_ptr = (char *) newptr;
584 		goto done;
585 	}
586 
587 	/*
588 	 * Otherwise, schedule the current buffer for output and get a
589 	 * fresh buffer.
590 	 */
591 	pmclog_schedule_io(po, 0);
592 
593 	if (pmclog_get_buffer(po) != 0)
594 		goto fail;
595 
596 	plb = *pplb;
597 	KASSERT(plb != NULL,
598 	    ("[pmclog,%d] po=%p no current buffer", __LINE__, po));
599 
600 	KASSERT(plb->plb_ptr != NULL,
601 	    ("[pmclog,%d] null return from pmc_get_log_buffer", __LINE__));
602 
603 	KASSERT(plb->plb_ptr == plb->plb_base &&
604 	    plb->plb_ptr <= plb->plb_fence,
605 	    ("[pmclog,%d] po=%p buffer invariants: ptr=%p base=%p fence=%p",
606 		__LINE__, po, plb->plb_ptr, plb->plb_base,
607 		plb->plb_fence));
608 
609 	oldptr = (uintptr_t) plb->plb_ptr;
610 
611  done:
612 	return ((uint32_t *) oldptr);
613  fail:
614 	return (NULL);
615 }
616 
617 /*
618  * Schedule an I/O.
619  *
620  * Transfer the current buffer to the helper kthread.
621  */
622 
623 static void
pmclog_schedule_io(struct pmc_owner * po,int wakeup)624 pmclog_schedule_io(struct pmc_owner *po, int wakeup)
625 {
626 	struct pmclog_buffer *plb;
627 
628 	plb = po->po_curbuf[curcpu];
629 	po->po_curbuf[curcpu] = NULL;
630 	KASSERT(plb != NULL,
631 	    ("[pmclog,%d] schedule_io with null buffer po=%p", __LINE__, po));
632 	KASSERT(plb->plb_ptr >= plb->plb_base,
633 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p base=%p", __LINE__,
634 		po, plb->plb_ptr, plb->plb_base));
635 	KASSERT(plb->plb_ptr <= plb->plb_fence,
636 	    ("[pmclog,%d] buffer invariants po=%p ptr=%p fenc=%p", __LINE__,
637 		po, plb->plb_ptr, plb->plb_fence));
638 
639 	PMCDBG1(LOG,SIO, 1, "po=%p", po);
640 
641 	/*
642 	 * Add the current buffer to the tail of the buffer list and
643 	 * wakeup the helper.
644 	 */
645 	mtx_lock_spin(&po->po_mtx);
646 	TAILQ_INSERT_TAIL(&po->po_logbuffers, plb, plb_next);
647 	mtx_unlock_spin(&po->po_mtx);
648 	if (wakeup)
649 		wakeup_one(po);
650 }
651 
652 /*
653  * Stop the helper kthread.
654  */
655 
656 static void
pmclog_stop_kthread(struct pmc_owner * po)657 pmclog_stop_kthread(struct pmc_owner *po)
658 {
659 
660 	mtx_lock(&pmc_kthread_mtx);
661 	po->po_flags &= ~PMC_PO_OWNS_LOGFILE;
662 	if (po->po_kthread != NULL) {
663 		PROC_LOCK(po->po_kthread);
664 		kern_psignal(po->po_kthread, SIGHUP);
665 		PROC_UNLOCK(po->po_kthread);
666 	}
667 	wakeup_one(po);
668 	while (po->po_kthread)
669 		msleep(po->po_kthread, &pmc_kthread_mtx, PPAUSE, "pmckstp", 0);
670 	mtx_unlock(&pmc_kthread_mtx);
671 }
672 
673 /*
674  * Public functions
675  */
676 
677 /*
678  * Configure a log file for pmc owner 'po'.
679  *
680  * Parameter 'logfd' is a file handle referencing an open file in the
681  * owner process.  This file needs to have been opened for writing.
682  */
683 
684 int
pmclog_configure_log(struct pmc_mdep * md,struct pmc_owner * po,int logfd)685 pmclog_configure_log(struct pmc_mdep *md, struct pmc_owner *po, int logfd)
686 {
687 	struct proc *p;
688 	struct timespec ts;
689 	int error;
690 
691 	sx_assert(&pmc_sx, SA_XLOCKED);
692 	PMCDBG2(LOG,CFG,1, "config po=%p logfd=%d", po, logfd);
693 
694 	p = po->po_owner;
695 
696 	/* return EBUSY if a log file was already present */
697 	if (po->po_flags & PMC_PO_OWNS_LOGFILE)
698 		return (EBUSY);
699 
700 	KASSERT(po->po_file == NULL,
701 	    ("[pmclog,%d] po=%p file (%p) already present", __LINE__, po,
702 		po->po_file));
703 
704 	/* get a reference to the file state */
705 	error = fget_write(curthread, logfd, &cap_write_rights, &po->po_file);
706 	if (error)
707 		goto error;
708 
709 	/* mark process as owning a log file */
710 	po->po_flags |= PMC_PO_OWNS_LOGFILE;
711 
712 	/* mark process as using HWPMCs */
713 	PROC_LOCK(p);
714 	p->p_flag |= P_HWPMC;
715 	PROC_UNLOCK(p);
716 	nanotime(&ts);
717 	/* create a log initialization entry */
718 	PMCLOG_RESERVE_WITH_ERROR(po, PMCLOG_TYPE_INITIALIZE,
719 	    sizeof(struct pmclog_initialize));
720 	PMCLOG_EMIT32(PMC_VERSION);
721 	PMCLOG_EMIT32(md->pmd_cputype);
722 #if defined(__i386__) || defined(__amd64__)
723 	PMCLOG_EMIT64(tsc_freq);
724 #else
725 	/* other architectures will need to fill this in */
726 	PMCLOG_EMIT32(0);
727 	PMCLOG_EMIT32(0);
728 #endif
729 	memcpy(_le, &ts, sizeof(ts));
730 	_le += sizeof(ts)/4;
731 	PMCLOG_EMITSTRING(pmc_cpuid, PMC_CPUID_LEN);
732 	PMCLOG_DESPATCH_SYNC(po);
733 
734 	return (0);
735 
736  error:
737 	KASSERT(po->po_kthread == NULL, ("[pmclog,%d] po=%p kthread not "
738 	    "stopped", __LINE__, po));
739 
740 	if (po->po_file)
741 		(void) fdrop(po->po_file, curthread);
742 	po->po_file  = NULL;	/* clear file and error state */
743 	po->po_error = 0;
744 	po->po_flags &= ~PMC_PO_OWNS_LOGFILE;
745 
746 	return (error);
747 }
748 
749 
750 /*
751  * De-configure a log file.  This will throw away any buffers queued
752  * for this owner process.
753  */
754 
755 int
pmclog_deconfigure_log(struct pmc_owner * po)756 pmclog_deconfigure_log(struct pmc_owner *po)
757 {
758 	int error;
759 	struct pmclog_buffer *lb;
760 	struct pmc_binding pb;
761 
762 	PMCDBG1(LOG,CFG,1, "de-config po=%p", po);
763 
764 	if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
765 		return (EINVAL);
766 
767 	KASSERT(po->po_sscount == 0,
768 	    ("[pmclog,%d] po=%p still owning SS PMCs", __LINE__, po));
769 	KASSERT(po->po_file != NULL,
770 	    ("[pmclog,%d] po=%p no log file", __LINE__, po));
771 
772 	/* stop the kthread, this will reset the 'OWNS_LOGFILE' flag */
773 	pmclog_stop_kthread(po);
774 
775 	KASSERT(po->po_kthread == NULL,
776 	    ("[pmclog,%d] po=%p kthread not stopped", __LINE__, po));
777 
778 	/* return all queued log buffers to the global pool */
779 	while ((lb = TAILQ_FIRST(&po->po_logbuffers)) != NULL) {
780 		TAILQ_REMOVE(&po->po_logbuffers, lb, plb_next);
781 		PMCLOG_RESET_BUFFER_DESCRIPTOR(lb);
782 		pmc_plb_rele(lb);
783 	}
784 	pmc_save_cpu_binding(&pb);
785 	for (int i = 0; i < mp_ncpus; i++) {
786 		pmc_select_cpu(i);
787 		/* return the 'current' buffer to the global pool */
788 		if ((lb = po->po_curbuf[curcpu]) != NULL) {
789 			PMCLOG_RESET_BUFFER_DESCRIPTOR(lb);
790 			pmc_plb_rele(lb);
791 		}
792 	}
793 	pmc_restore_cpu_binding(&pb);
794 
795 	/* drop a reference to the fd */
796 	if (po->po_file != NULL) {
797 		error = fdrop(po->po_file, curthread);
798 		po->po_file = NULL;
799 	} else
800 		error = 0;
801 	po->po_error = 0;
802 
803 	return (error);
804 }
805 
806 /*
807  * Flush a process' log buffer.
808  */
809 
810 int
pmclog_flush(struct pmc_owner * po,int force)811 pmclog_flush(struct pmc_owner *po, int force)
812 {
813 	int error;
814 
815 	PMCDBG1(LOG,FLS,1, "po=%p", po);
816 
817 	/*
818 	 * If there is a pending error recorded by the logger thread,
819 	 * return that.
820 	 */
821 	if (po->po_error)
822 		return (po->po_error);
823 
824 	error = 0;
825 
826 	/*
827 	 * Check that we do have an active log file.
828 	 */
829 	mtx_lock(&pmc_kthread_mtx);
830 	if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) {
831 		error = EINVAL;
832 		goto error;
833 	}
834 
835 	pmclog_schedule_all(po);
836  error:
837 	mtx_unlock(&pmc_kthread_mtx);
838 
839 	return (error);
840 }
841 
842 static void
pmclog_schedule_one_cond(struct pmc_owner * po)843 pmclog_schedule_one_cond(struct pmc_owner *po)
844 {
845 	struct pmclog_buffer *plb;
846 	int cpu;
847 
848 	spinlock_enter();
849 	cpu = curcpu;
850 	/* tell hardclock not to run again */
851 	if (PMC_CPU_HAS_SAMPLES(cpu))
852 		PMC_CALL_HOOK_UNLOCKED(curthread, PMC_FN_DO_SAMPLES, NULL);
853 
854 	plb = po->po_curbuf[cpu];
855 	if (plb && plb->plb_ptr != plb->plb_base)
856 		pmclog_schedule_io(po, 1);
857 	spinlock_exit();
858 }
859 
860 static void
pmclog_schedule_all(struct pmc_owner * po)861 pmclog_schedule_all(struct pmc_owner *po)
862 {
863 	struct pmc_binding pb;
864 
865 	/*
866 	 * Schedule the current buffer if any and not empty.
867 	 */
868 	pmc_save_cpu_binding(&pb);
869 	for (int i = 0; i < mp_ncpus; i++) {
870 		pmc_select_cpu(i);
871 		pmclog_schedule_one_cond(po);
872 	}
873 	pmc_restore_cpu_binding(&pb);
874 }
875 
876 int
pmclog_close(struct pmc_owner * po)877 pmclog_close(struct pmc_owner *po)
878 {
879 
880 	PMCDBG1(LOG,CLO,1, "po=%p", po);
881 
882 	pmclog_process_closelog(po);
883 
884 	mtx_lock(&pmc_kthread_mtx);
885 	/*
886 	 * Initiate shutdown: no new data queued,
887 	 * thread will close file on last block.
888 	 */
889 	po->po_flags |= PMC_PO_SHUTDOWN;
890 	/* give time for all to see */
891 	DELAY(50);
892 
893 	/*
894 	 * Schedule the current buffer.
895 	 */
896 	pmclog_schedule_all(po);
897 	wakeup_one(po);
898 
899 	mtx_unlock(&pmc_kthread_mtx);
900 
901 	return (0);
902 }
903 
904 void
pmclog_process_callchain(struct pmc * pm,struct pmc_sample * ps)905 pmclog_process_callchain(struct pmc *pm, struct pmc_sample *ps)
906 {
907 	int n, recordlen;
908 	uint32_t flags;
909 	struct pmc_owner *po;
910 
911 	PMCDBG3(LOG,SAM,1,"pm=%p pid=%d n=%d", pm, ps->ps_pid,
912 	    ps->ps_nsamples);
913 
914 	recordlen = offsetof(struct pmclog_callchain, pl_pc) +
915 	    ps->ps_nsamples * sizeof(uintfptr_t);
916 	po = pm->pm_owner;
917 	flags = PMC_CALLCHAIN_TO_CPUFLAGS(ps->ps_cpu,ps->ps_flags);
918 	PMCLOG_RESERVE_SAFE(po, PMCLOG_TYPE_CALLCHAIN, recordlen, ps->ps_tsc);
919 	PMCLOG_EMIT32(ps->ps_pid);
920 	PMCLOG_EMIT32(ps->ps_tid);
921 	PMCLOG_EMIT32(pm->pm_id);
922 	PMCLOG_EMIT32(flags);
923 	for (n = 0; n < ps->ps_nsamples; n++)
924 		PMCLOG_EMITADDR(ps->ps_pc[n]);
925 	PMCLOG_DESPATCH_SAFE(po);
926 }
927 
928 void
pmclog_process_closelog(struct pmc_owner * po)929 pmclog_process_closelog(struct pmc_owner *po)
930 {
931 	PMCLOG_RESERVE(po, PMCLOG_TYPE_CLOSELOG,
932 	    sizeof(struct pmclog_closelog));
933 	PMCLOG_DESPATCH_SYNC(po);
934 }
935 
936 void
pmclog_process_dropnotify(struct pmc_owner * po)937 pmclog_process_dropnotify(struct pmc_owner *po)
938 {
939 	PMCLOG_RESERVE(po, PMCLOG_TYPE_DROPNOTIFY,
940 	    sizeof(struct pmclog_dropnotify));
941 	PMCLOG_DESPATCH(po);
942 }
943 
944 void
pmclog_process_map_in(struct pmc_owner * po,pid_t pid,uintfptr_t start,const char * path)945 pmclog_process_map_in(struct pmc_owner *po, pid_t pid, uintfptr_t start,
946     const char *path)
947 {
948 	int pathlen, recordlen;
949 
950 	KASSERT(path != NULL, ("[pmclog,%d] map-in, null path", __LINE__));
951 
952 	pathlen = strlen(path) + 1;	/* #bytes for path name */
953 	recordlen = offsetof(struct pmclog_map_in, pl_pathname) +
954 	    pathlen;
955 
956 	PMCLOG_RESERVE(po, PMCLOG_TYPE_MAP_IN, recordlen);
957 	PMCLOG_EMIT32(pid);
958 	PMCLOG_EMIT32(0);
959 	PMCLOG_EMITADDR(start);
960 	PMCLOG_EMITSTRING(path,pathlen);
961 	PMCLOG_DESPATCH_SYNC(po);
962 }
963 
964 void
pmclog_process_map_out(struct pmc_owner * po,pid_t pid,uintfptr_t start,uintfptr_t end)965 pmclog_process_map_out(struct pmc_owner *po, pid_t pid, uintfptr_t start,
966     uintfptr_t end)
967 {
968 	KASSERT(start <= end, ("[pmclog,%d] start > end", __LINE__));
969 
970 	PMCLOG_RESERVE(po, PMCLOG_TYPE_MAP_OUT, sizeof(struct pmclog_map_out));
971 	PMCLOG_EMIT32(pid);
972 	PMCLOG_EMIT32(0);
973 	PMCLOG_EMITADDR(start);
974 	PMCLOG_EMITADDR(end);
975 	PMCLOG_DESPATCH(po);
976 }
977 
978 void
pmclog_process_pmcallocate(struct pmc * pm)979 pmclog_process_pmcallocate(struct pmc *pm)
980 {
981 	struct pmc_owner *po;
982 	struct pmc_soft *ps;
983 
984 	po = pm->pm_owner;
985 
986 	PMCDBG1(LOG,ALL,1, "pm=%p", pm);
987 
988 	if (PMC_TO_CLASS(pm) == PMC_CLASS_SOFT) {
989 		PMCLOG_RESERVE(po, PMCLOG_TYPE_PMCALLOCATEDYN,
990 		    sizeof(struct pmclog_pmcallocatedyn));
991 		PMCLOG_EMIT32(pm->pm_id);
992 		PMCLOG_EMIT32(pm->pm_event);
993 		PMCLOG_EMIT32(pm->pm_flags);
994 		PMCLOG_EMIT32(0);
995 		PMCLOG_EMIT64(pm->pm_sc.pm_reloadcount);
996 		ps = pmc_soft_ev_acquire(pm->pm_event);
997 		if (ps != NULL)
998 			PMCLOG_EMITSTRING(ps->ps_ev.pm_ev_name,PMC_NAME_MAX);
999 		else
1000 			PMCLOG_EMITNULLSTRING(PMC_NAME_MAX);
1001 		pmc_soft_ev_release(ps);
1002 		PMCLOG_DESPATCH_SYNC(po);
1003 	} else {
1004 		PMCLOG_RESERVE(po, PMCLOG_TYPE_PMCALLOCATE,
1005 		    sizeof(struct pmclog_pmcallocate));
1006 		PMCLOG_EMIT32(pm->pm_id);
1007 		PMCLOG_EMIT32(pm->pm_event);
1008 		PMCLOG_EMIT32(pm->pm_flags);
1009 		PMCLOG_EMIT32(0);
1010 		PMCLOG_EMIT64(pm->pm_sc.pm_reloadcount);
1011 		PMCLOG_DESPATCH_SYNC(po);
1012 	}
1013 }
1014 
1015 void
pmclog_process_pmcattach(struct pmc * pm,pid_t pid,char * path)1016 pmclog_process_pmcattach(struct pmc *pm, pid_t pid, char *path)
1017 {
1018 	int pathlen, recordlen;
1019 	struct pmc_owner *po;
1020 
1021 	PMCDBG2(LOG,ATT,1,"pm=%p pid=%d", pm, pid);
1022 
1023 	po = pm->pm_owner;
1024 
1025 	pathlen = strlen(path) + 1;	/* #bytes for the string */
1026 	recordlen = offsetof(struct pmclog_pmcattach, pl_pathname) + pathlen;
1027 
1028 	PMCLOG_RESERVE(po, PMCLOG_TYPE_PMCATTACH, recordlen);
1029 	PMCLOG_EMIT32(pm->pm_id);
1030 	PMCLOG_EMIT32(pid);
1031 	PMCLOG_EMITSTRING(path, pathlen);
1032 	PMCLOG_DESPATCH_SYNC(po);
1033 }
1034 
1035 void
pmclog_process_pmcdetach(struct pmc * pm,pid_t pid)1036 pmclog_process_pmcdetach(struct pmc *pm, pid_t pid)
1037 {
1038 	struct pmc_owner *po;
1039 
1040 	PMCDBG2(LOG,ATT,1,"!pm=%p pid=%d", pm, pid);
1041 
1042 	po = pm->pm_owner;
1043 
1044 	PMCLOG_RESERVE(po, PMCLOG_TYPE_PMCDETACH,
1045 	    sizeof(struct pmclog_pmcdetach));
1046 	PMCLOG_EMIT32(pm->pm_id);
1047 	PMCLOG_EMIT32(pid);
1048 	PMCLOG_DESPATCH_SYNC(po);
1049 }
1050 
1051 void
pmclog_process_proccreate(struct pmc_owner * po,struct proc * p,int sync)1052 pmclog_process_proccreate(struct pmc_owner *po, struct proc *p, int sync)
1053 {
1054 	if (sync) {
1055 		PMCLOG_RESERVE(po, PMCLOG_TYPE_PROC_CREATE,
1056 		    sizeof(struct pmclog_proccreate));
1057 		PMCLOG_EMIT32(p->p_pid);
1058 		PMCLOG_EMIT32(p->p_flag);
1059 		PMCLOG_EMITSTRING(p->p_comm, MAXCOMLEN+1);
1060 		PMCLOG_DESPATCH_SYNC(po);
1061 	} else {
1062 		PMCLOG_RESERVE(po, PMCLOG_TYPE_PROC_CREATE,
1063 		    sizeof(struct pmclog_proccreate));
1064 		PMCLOG_EMIT32(p->p_pid);
1065 		PMCLOG_EMIT32(p->p_flag);
1066 		PMCLOG_EMITSTRING(p->p_comm, MAXCOMLEN+1);
1067 		PMCLOG_DESPATCH(po);
1068 	}
1069 }
1070 
1071 /*
1072  * Log a context switch event to the log file.
1073  */
1074 
1075 void
pmclog_process_proccsw(struct pmc * pm,struct pmc_process * pp,pmc_value_t v,struct thread * td)1076 pmclog_process_proccsw(struct pmc *pm, struct pmc_process *pp, pmc_value_t v, struct thread *td)
1077 {
1078 	struct pmc_owner *po;
1079 
1080 	KASSERT(pm->pm_flags & PMC_F_LOG_PROCCSW,
1081 	    ("[pmclog,%d] log-process-csw called gratuitously", __LINE__));
1082 
1083 	PMCDBG3(LOG,SWO,1,"pm=%p pid=%d v=%jx", pm, pp->pp_proc->p_pid,
1084 	    v);
1085 
1086 	po = pm->pm_owner;
1087 
1088 	PMCLOG_RESERVE_SAFE(po, PMCLOG_TYPE_PROCCSW,
1089 	    sizeof(struct pmclog_proccsw), pmc_rdtsc());
1090 	PMCLOG_EMIT64(v);
1091 	PMCLOG_EMIT32(pm->pm_id);
1092 	PMCLOG_EMIT32(pp->pp_proc->p_pid);
1093 	PMCLOG_EMIT32(td->td_tid);
1094 	PMCLOG_EMIT32(0);
1095 	PMCLOG_DESPATCH_SCHED_LOCK(po);
1096 }
1097 
1098 void
pmclog_process_procexec(struct pmc_owner * po,pmc_id_t pmid,pid_t pid,uintptr_t baseaddr,uintptr_t dynaddr,char * path)1099 pmclog_process_procexec(struct pmc_owner *po, pmc_id_t pmid, pid_t pid,
1100     uintptr_t baseaddr, uintptr_t dynaddr, char *path)
1101 {
1102 	int pathlen, recordlen;
1103 
1104 	PMCDBG3(LOG,EXC,1,"po=%p pid=%d path=\"%s\"", po, pid, path);
1105 
1106 	pathlen   = strlen(path) + 1;	/* #bytes for the path */
1107 	recordlen = offsetof(struct pmclog_procexec, pl_pathname) + pathlen;
1108 	PMCLOG_RESERVE(po, PMCLOG_TYPE_PROCEXEC, recordlen);
1109 	PMCLOG_EMIT32(pid);
1110 	PMCLOG_EMIT32(pmid);
1111 	PMCLOG_EMITADDR(baseaddr);
1112 	PMCLOG_EMITADDR(dynaddr);
1113 	PMCLOG_EMITSTRING(path,pathlen);
1114 	PMCLOG_DESPATCH_SYNC(po);
1115 }
1116 
1117 /*
1118  * Log a process exit event (and accumulated pmc value) to the log file.
1119  */
1120 
1121 void
pmclog_process_procexit(struct pmc * pm,struct pmc_process * pp)1122 pmclog_process_procexit(struct pmc *pm, struct pmc_process *pp)
1123 {
1124 	int ri;
1125 	struct pmc_owner *po;
1126 
1127 	ri = PMC_TO_ROWINDEX(pm);
1128 	PMCDBG3(LOG,EXT,1,"pm=%p pid=%d v=%jx", pm, pp->pp_proc->p_pid,
1129 	    pp->pp_pmcs[ri].pp_pmcval);
1130 
1131 	po = pm->pm_owner;
1132 
1133 	PMCLOG_RESERVE(po, PMCLOG_TYPE_PROCEXIT,
1134 	    sizeof(struct pmclog_procexit));
1135 	PMCLOG_EMIT32(pm->pm_id);
1136 	PMCLOG_EMIT32(pp->pp_proc->p_pid);
1137 	PMCLOG_EMIT64(pp->pp_pmcs[ri].pp_pmcval);
1138 	PMCLOG_DESPATCH(po);
1139 }
1140 
1141 /*
1142  * Log a fork event.
1143  */
1144 
1145 void
pmclog_process_procfork(struct pmc_owner * po,pid_t oldpid,pid_t newpid)1146 pmclog_process_procfork(struct pmc_owner *po, pid_t oldpid, pid_t newpid)
1147 {
1148 	PMCLOG_RESERVE(po, PMCLOG_TYPE_PROCFORK,
1149 	    sizeof(struct pmclog_procfork));
1150 	PMCLOG_EMIT32(oldpid);
1151 	PMCLOG_EMIT32(newpid);
1152 	PMCLOG_DESPATCH(po);
1153 }
1154 
1155 /*
1156  * Log a process exit event of the form suitable for system-wide PMCs.
1157  */
1158 
1159 void
pmclog_process_sysexit(struct pmc_owner * po,pid_t pid)1160 pmclog_process_sysexit(struct pmc_owner *po, pid_t pid)
1161 {
1162 	PMCLOG_RESERVE(po, PMCLOG_TYPE_SYSEXIT, sizeof(struct pmclog_sysexit));
1163 	PMCLOG_EMIT32(pid);
1164 	PMCLOG_DESPATCH(po);
1165 }
1166 
1167 void
pmclog_process_threadcreate(struct pmc_owner * po,struct thread * td,int sync)1168 pmclog_process_threadcreate(struct pmc_owner *po, struct thread *td, int sync)
1169 {
1170 	struct proc *p;
1171 
1172 	p = td->td_proc;
1173 	if (sync) {
1174 		PMCLOG_RESERVE(po, PMCLOG_TYPE_THR_CREATE,
1175 		    sizeof(struct pmclog_threadcreate));
1176 		PMCLOG_EMIT32(td->td_tid);
1177 		PMCLOG_EMIT32(p->p_pid);
1178 		PMCLOG_EMIT32(p->p_flag);
1179 		PMCLOG_EMIT32(0);
1180 		PMCLOG_EMITSTRING(td->td_name, MAXCOMLEN+1);
1181 		PMCLOG_DESPATCH_SYNC(po);
1182 	} else {
1183 		PMCLOG_RESERVE(po, PMCLOG_TYPE_THR_CREATE,
1184 		    sizeof(struct pmclog_threadcreate));
1185 		PMCLOG_EMIT32(td->td_tid);
1186 		PMCLOG_EMIT32(p->p_pid);
1187 		PMCLOG_EMIT32(p->p_flag);
1188 		PMCLOG_EMIT32(0);
1189 		PMCLOG_EMITSTRING(td->td_name, MAXCOMLEN+1);
1190 		PMCLOG_DESPATCH(po);
1191 	}
1192 }
1193 
1194 void
pmclog_process_threadexit(struct pmc_owner * po,struct thread * td)1195 pmclog_process_threadexit(struct pmc_owner *po, struct thread *td)
1196 {
1197 
1198 	PMCLOG_RESERVE(po, PMCLOG_TYPE_THR_EXIT,
1199 	    sizeof(struct pmclog_threadexit));
1200 	PMCLOG_EMIT32(td->td_tid);
1201 	PMCLOG_DESPATCH(po);
1202 }
1203 
1204 /*
1205  * Write a user log entry.
1206  */
1207 
1208 int
pmclog_process_userlog(struct pmc_owner * po,struct pmc_op_writelog * wl)1209 pmclog_process_userlog(struct pmc_owner *po, struct pmc_op_writelog *wl)
1210 {
1211 	int error;
1212 
1213 	PMCDBG2(LOG,WRI,1, "writelog po=%p ud=0x%x", po, wl->pm_userdata);
1214 
1215 	error = 0;
1216 
1217 	PMCLOG_RESERVE_WITH_ERROR(po, PMCLOG_TYPE_USERDATA,
1218 	    sizeof(struct pmclog_userdata));
1219 	PMCLOG_EMIT32(wl->pm_userdata);
1220 	PMCLOG_DESPATCH(po);
1221 
1222  error:
1223 	return (error);
1224 }
1225 
1226 /*
1227  * Initialization.
1228  *
1229  * Create a pool of log buffers and initialize mutexes.
1230  */
1231 
1232 void
pmclog_initialize(void)1233 pmclog_initialize(void)
1234 {
1235 	struct pmclog_buffer *plb;
1236 	int domain, ncpus, total;
1237 
1238 	if (pmclog_buffer_size <= 0 || pmclog_buffer_size > 16*1024) {
1239 		(void) printf("hwpmc: tunable logbuffersize=%d must be "
1240 					  "greater than zero and less than or equal to 16MB.\n",
1241 					  pmclog_buffer_size);
1242 		pmclog_buffer_size = PMC_LOG_BUFFER_SIZE;
1243 	}
1244 
1245 	if (pmc_nlogbuffers_pcpu <= 0) {
1246 		(void) printf("hwpmc: tunable nlogbuffers=%d must be greater "
1247 					  "than zero.\n", pmc_nlogbuffers_pcpu);
1248 		pmc_nlogbuffers_pcpu = PMC_NLOGBUFFERS_PCPU;
1249 	}
1250 	if (pmc_nlogbuffers_pcpu*pmclog_buffer_size > 32*1024) {
1251 		(void) printf("hwpmc: memory allocated pcpu must be less than 32MB (is %dK).\n",
1252 					  pmc_nlogbuffers_pcpu*pmclog_buffer_size);
1253 		pmc_nlogbuffers_pcpu = PMC_NLOGBUFFERS_PCPU;
1254 		pmclog_buffer_size = PMC_LOG_BUFFER_SIZE;
1255 	}
1256 	for (domain = 0; domain < vm_ndomains; domain++) {
1257 		ncpus = pmc_dom_hdrs[domain]->pdbh_ncpus;
1258 		total = ncpus * pmc_nlogbuffers_pcpu;
1259 
1260 		plb = malloc_domainset(sizeof(struct pmclog_buffer) * total,
1261 		    M_PMC, DOMAINSET_PREF(domain), M_WAITOK | M_ZERO);
1262 		pmc_dom_hdrs[domain]->pdbh_plbs = plb;
1263 		for (; total > 0; total--, plb++) {
1264 			void *buf;
1265 
1266 			buf = malloc_domainset(1024 * pmclog_buffer_size, M_PMC,
1267 			    DOMAINSET_PREF(domain), M_WAITOK | M_ZERO);
1268 			PMCLOG_INIT_BUFFER_DESCRIPTOR(plb, buf, domain);
1269 			pmc_plb_rele_unlocked(plb);
1270 		}
1271 	}
1272 	mtx_init(&pmc_kthread_mtx, "pmc-kthread", "pmc-sleep", MTX_DEF);
1273 }
1274 
1275 /*
1276  * Shutdown logging.
1277  *
1278  * Destroy mutexes and release memory back the to free pool.
1279  */
1280 
1281 void
pmclog_shutdown(void)1282 pmclog_shutdown(void)
1283 {
1284 	struct pmclog_buffer *plb;
1285 	int domain;
1286 
1287 	mtx_destroy(&pmc_kthread_mtx);
1288 
1289 	for (domain = 0; domain < vm_ndomains; domain++) {
1290 		while ((plb = TAILQ_FIRST(&pmc_dom_hdrs[domain]->pdbh_head)) != NULL) {
1291 			TAILQ_REMOVE(&pmc_dom_hdrs[domain]->pdbh_head, plb, plb_next);
1292 			free(plb->plb_base, M_PMC);
1293 		}
1294 		free(pmc_dom_hdrs[domain]->pdbh_plbs, M_PMC);
1295 	}
1296 }
1297