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