2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
4 * Copyright (c) 2005-2007 Joseph Koshy
5 * Copyright (c) 2007 The FreeBSD Foundation
6 * Copyright (c) 2018 Matthew Macy
9 * Portions of this software were developed by A. Joseph Koshy under
10 * sponsorship from the FreeBSD Foundation and Google, Inc.
12 * Redistribution and use in source and binary forms, with or without
13 * modification, are permitted provided that the following conditions
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.
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
36 * Logging code for hwpmc(4)
39 #include <sys/cdefs.h>
40 __FBSDID("$FreeBSD$");
42 #include <sys/param.h>
43 #include <sys/capsicum.h>
45 #include <sys/kernel.h>
46 #include <sys/kthread.h>
48 #include <sys/module.h>
49 #include <sys/mutex.h>
51 #include <sys/pmckern.h>
52 #include <sys/pmclog.h>
54 #include <sys/sched.h>
55 #include <sys/signalvar.h>
57 #include <sys/syscallsubr.h>
58 #include <sys/sysctl.h>
59 #include <sys/systm.h>
61 #include <sys/unistd.h>
62 #include <sys/vnode.h>
64 #if defined(__i386__) || defined(__amd64__)
65 #include <machine/clock.h>
68 #define curdomain PCPU_GET(domain)
74 SYSCTL_DECL(_kern_hwpmc);
77 * kern.hwpmc.logbuffersize -- size of the per-cpu owner buffers.
80 static int pmclog_buffer_size = PMC_LOG_BUFFER_SIZE;
81 #if (__FreeBSD_version < 1100000)
82 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "logbuffersize", &pmclog_buffer_size);
84 SYSCTL_INT(_kern_hwpmc, OID_AUTO, logbuffersize, CTLFLAG_RDTUN,
85 &pmclog_buffer_size, 0, "size of log buffers in kilobytes");
88 * kern.hwpmc.nbuffer -- number of global log buffers
91 static int pmc_nlogbuffers_pcpu = PMC_NLOGBUFFERS_PCPU;
92 #if (__FreeBSD_version < 1100000)
93 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "nbuffers", &pmc_nlogbuffers_pcpu);
95 SYSCTL_INT(_kern_hwpmc, OID_AUTO, nbuffers_pcpu, CTLFLAG_RDTUN,
96 &pmc_nlogbuffers_pcpu, 0, "number of log buffers per cpu");
99 * Global log buffer list and associated spin lock.
102 static struct mtx pmc_kthread_mtx; /* sleep lock */
104 #define PMCLOG_INIT_BUFFER_DESCRIPTOR(D, buf, domain) do { \
105 (D)->plb_fence = ((char *) (buf)) + 1024*pmclog_buffer_size; \
106 (D)->plb_base = (D)->plb_ptr = ((char *) (buf)); \
107 (D)->plb_domain = domain; \
110 #define PMCLOG_RESET_BUFFER_DESCRIPTOR(D) do { \
111 (D)->plb_ptr = (D)->plb_base; \
115 * Log file record constructors.
117 #define _PMCLOG_TO_HEADER(T,L) \
118 ((PMCLOG_HEADER_MAGIC << 24) | \
119 (PMCLOG_TYPE_ ## T << 16) | \
122 /* reserve LEN bytes of space and initialize the entry header */
123 #define _PMCLOG_RESERVE_SAFE(PO,TYPE,LEN,ACTION, TSC) do { \
125 int _len = roundup((LEN), sizeof(uint32_t)); \
126 struct pmclog_header *ph; \
127 if ((_le = pmclog_reserve((PO), _len)) == NULL) { \
130 ph = (struct pmclog_header *)_le; \
131 ph->pl_header =_PMCLOG_TO_HEADER(TYPE,_len); \
132 ph->pl_tsc = (TSC); \
133 _le += sizeof(*ph)/4 /* skip over timestamp */
135 /* reserve LEN bytes of space and initialize the entry header */
136 #define _PMCLOG_RESERVE(PO,TYPE,LEN,ACTION) do { \
138 int _len = roundup((LEN), sizeof(uint32_t)); \
140 struct pmclog_header *ph; \
143 if ((_le = pmclog_reserve((PO), _len)) == NULL) { \
147 ph = (struct pmclog_header *)_le; \
148 ph->pl_header =_PMCLOG_TO_HEADER(TYPE,_len); \
150 _le += sizeof(*ph)/4 /* skip over timestamp */
154 #define PMCLOG_RESERVE_SAFE(P,T,L,TSC) _PMCLOG_RESERVE_SAFE(P,T,L,return,TSC)
155 #define PMCLOG_RESERVE(P,T,L) _PMCLOG_RESERVE(P,T,L,return)
156 #define PMCLOG_RESERVE_WITH_ERROR(P,T,L) _PMCLOG_RESERVE(P,T,L, \
157 error=ENOMEM;goto error)
159 #define PMCLOG_EMIT32(V) do { *_le++ = (V); } while (0)
160 #define PMCLOG_EMIT64(V) do { \
161 *_le++ = (uint32_t) ((V) & 0xFFFFFFFF); \
162 *_le++ = (uint32_t) (((V) >> 32) & 0xFFFFFFFF); \
166 /* Emit a string. Caution: does NOT update _le, so needs to be last */
167 #define PMCLOG_EMITSTRING(S,L) do { bcopy((S), _le, (L)); } while (0)
168 #define PMCLOG_EMITNULLSTRING(L) do { bzero(_le, (L)); } while (0)
170 #define PMCLOG_DESPATCH_SAFE(PO) \
171 pmclog_release((PO)); \
174 #define PMCLOG_DESPATCH_SCHED_LOCK(PO) \
175 pmclog_release_flags((PO), 0); \
178 #define PMCLOG_DESPATCH(PO) \
179 pmclog_release((PO)); \
183 #define PMCLOG_DESPATCH_SYNC(PO) \
184 pmclog_schedule_io((PO), 1); \
191 * Assertions about the log file format.
193 CTASSERT(sizeof(struct pmclog_callchain) == 7*4 + TSDELTA +
194 PMC_CALLCHAIN_DEPTH_MAX*sizeof(uintfptr_t));
195 CTASSERT(sizeof(struct pmclog_closelog) == 3*4 + TSDELTA);
196 CTASSERT(sizeof(struct pmclog_dropnotify) == 3*4 + TSDELTA);
197 CTASSERT(sizeof(struct pmclog_map_in) == PATH_MAX + TSDELTA +
198 5*4 + sizeof(uintfptr_t));
199 CTASSERT(offsetof(struct pmclog_map_in,pl_pathname) ==
200 5*4 + TSDELTA + sizeof(uintfptr_t));
201 CTASSERT(sizeof(struct pmclog_map_out) == 5*4 + 2*sizeof(uintfptr_t) + TSDELTA);
202 CTASSERT(sizeof(struct pmclog_pmcallocate) == 9*4 + TSDELTA);
203 CTASSERT(sizeof(struct pmclog_pmcattach) == 5*4 + PATH_MAX + TSDELTA);
204 CTASSERT(offsetof(struct pmclog_pmcattach,pl_pathname) == 5*4 + TSDELTA);
205 CTASSERT(sizeof(struct pmclog_pmcdetach) == 5*4 + TSDELTA);
206 CTASSERT(sizeof(struct pmclog_proccsw) == 7*4 + 8 + TSDELTA);
207 CTASSERT(sizeof(struct pmclog_procexec) == 5*4 + PATH_MAX +
208 sizeof(uintfptr_t) + TSDELTA);
209 CTASSERT(offsetof(struct pmclog_procexec,pl_pathname) == 5*4 + TSDELTA +
211 CTASSERT(sizeof(struct pmclog_procexit) == 5*4 + 8 + TSDELTA);
212 CTASSERT(sizeof(struct pmclog_procfork) == 5*4 + TSDELTA);
213 CTASSERT(sizeof(struct pmclog_sysexit) == 6*4);
214 CTASSERT(sizeof(struct pmclog_userdata) == 6*4);
217 * Log buffer structure
220 struct pmclog_buffer {
221 TAILQ_ENTRY(pmclog_buffer) plb_next;
226 } __aligned(CACHE_LINE_SIZE);
232 static int pmclog_get_buffer(struct pmc_owner *po);
233 static void pmclog_loop(void *arg);
234 static void pmclog_release(struct pmc_owner *po);
235 static uint32_t *pmclog_reserve(struct pmc_owner *po, int length);
236 static void pmclog_schedule_io(struct pmc_owner *po, int wakeup);
237 static void pmclog_schedule_all(struct pmc_owner *po, int force);
238 static void pmclog_stop_kthread(struct pmc_owner *po);
245 pmc_plb_rele_unlocked(struct pmclog_buffer *plb)
247 TAILQ_INSERT_HEAD(&pmc_dom_hdrs[plb->plb_domain]->pdbh_head, plb, plb_next);
251 pmc_plb_rele(struct pmclog_buffer *plb)
253 mtx_lock_spin(&pmc_dom_hdrs[plb->plb_domain]->pdbh_mtx);
254 pmc_plb_rele_unlocked(plb);
255 mtx_unlock_spin(&pmc_dom_hdrs[plb->plb_domain]->pdbh_mtx);
262 pmclog_get_buffer(struct pmc_owner *po)
264 struct pmclog_buffer *plb;
267 KASSERT(po->po_curbuf[curcpu] == NULL,
268 ("[pmclog,%d] po=%p current buffer still valid", __LINE__, po));
271 MPASS(pmc_dom_hdrs[domain]);
272 mtx_lock_spin(&pmc_dom_hdrs[domain]->pdbh_mtx);
273 if ((plb = TAILQ_FIRST(&pmc_dom_hdrs[domain]->pdbh_head)) != NULL)
274 TAILQ_REMOVE(&pmc_dom_hdrs[domain]->pdbh_head, plb, plb_next);
275 mtx_unlock_spin(&pmc_dom_hdrs[domain]->pdbh_mtx);
277 PMCDBG2(LOG,GTB,1, "po=%p plb=%p", po, plb);
281 KASSERT(plb->plb_ptr == plb->plb_base &&
282 plb->plb_base < plb->plb_fence,
283 ("[pmclog,%d] po=%p buffer invariants: ptr=%p "
284 "base=%p fence=%p", __LINE__, po, plb->plb_ptr,
285 plb->plb_base, plb->plb_fence));
288 po->po_curbuf[curcpu] = plb;
291 counter_u64_add(pmc_stats.pm_buffer_requests, 1);
293 counter_u64_add(pmc_stats.pm_buffer_requests_failed, 1);
295 return (plb ? 0 : ENOMEM);
298 struct pmclog_proc_init_args {
300 struct pmc_owner *po;
306 pmclog_proc_create(struct thread *td, void **handlep)
308 struct pmclog_proc_init_args *ia;
311 ia = malloc(sizeof(*ia), M_TEMP, M_WAITOK | M_ZERO);
312 error = kproc_create(pmclog_loop, ia, &ia->kthr,
313 RFHIGHPID, 0, "hwpmc: proc(%d)", td->td_proc->p_pid);
320 pmclog_proc_ignite(void *handle, struct pmc_owner *po)
322 struct pmclog_proc_init_args *ia;
325 mtx_lock(&pmc_kthread_mtx);
327 MPASS(ia->po == NULL);
329 MPASS(ia->kthr != NULL);
334 KASSERT(po->po_kthread == NULL,
335 ("[pmclog,%d] po=%p kthread (%p) already present",
336 __LINE__, po, po->po_kthread));
337 po->po_kthread = ia->kthr;
341 msleep(ia, &pmc_kthread_mtx, PWAIT, "pmclogw", 0);
342 mtx_unlock(&pmc_kthread_mtx);
349 * This function is executed by each pmc owner's helper thread.
352 pmclog_loop(void *arg)
354 struct pmclog_proc_init_args *ia;
355 struct pmc_owner *po;
356 struct pmclog_buffer *lb;
358 struct ucred *ownercred;
359 struct ucred *mycred;
370 SIGADDSET(unb, SIGHUP);
371 (void)kern_sigprocmask(td, SIG_UNBLOCK, &unb, NULL, 0);
374 MPASS(ia->kthr == curproc);
376 mtx_lock(&pmc_kthread_mtx);
377 while (ia->po == NULL && !ia->exit)
378 msleep(ia, &pmc_kthread_mtx, PWAIT, "pmclogi", 0);
382 mtx_unlock(&pmc_kthread_mtx);
385 MPASS(ia->po != NULL);
389 mtx_unlock(&pmc_kthread_mtx);
393 mycred = td->td_ucred;
396 ownercred = crhold(p->p_ucred);
399 PMCDBG2(LOG,INI,1, "po=%p kt=%p", po, po->po_kthread);
400 KASSERT(po->po_kthread == curthread->td_proc,
401 ("[pmclog,%d] proc mismatch po=%p po/kt=%p curproc=%p", __LINE__,
402 po, po->po_kthread, curthread->td_proc));
408 * Loop waiting for I/O requests to be added to the owner
409 * struct's queue. The loop is exited when the log file
413 mtx_lock(&pmc_kthread_mtx);
417 /* check if we've been asked to exit */
418 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
421 if (lb == NULL) { /* look for a fresh buffer to write */
422 mtx_lock_spin(&po->po_mtx);
423 if ((lb = TAILQ_FIRST(&po->po_logbuffers)) == NULL) {
424 mtx_unlock_spin(&po->po_mtx);
426 /* No more buffers and shutdown required. */
427 if (po->po_flags & PMC_PO_SHUTDOWN)
430 (void) msleep(po, &pmc_kthread_mtx, PWAIT,
435 TAILQ_REMOVE(&po->po_logbuffers, lb, plb_next);
436 mtx_unlock_spin(&po->po_mtx);
439 mtx_unlock(&pmc_kthread_mtx);
441 /* process the request */
442 PMCDBG3(LOG,WRI,2, "po=%p base=%p ptr=%p", po,
443 lb->plb_base, lb->plb_ptr);
444 /* change our thread's credentials before issuing the I/O */
446 aiov.iov_base = lb->plb_base;
447 aiov.iov_len = nbytes = lb->plb_ptr - lb->plb_base;
449 auio.uio_iov = &aiov;
451 auio.uio_offset = -1;
452 auio.uio_resid = nbytes;
453 auio.uio_rw = UIO_WRITE;
454 auio.uio_segflg = UIO_SYSSPACE;
457 /* switch thread credentials -- see kern_ktrace.c */
458 td->td_ucred = ownercred;
459 error = fo_write(po->po_file, &auio, ownercred, 0, td);
460 td->td_ucred = mycred;
463 /* XXX some errors are recoverable */
464 /* send a SIGIO to the owner and exit */
466 kern_psignal(p, SIGIO);
469 mtx_lock(&pmc_kthread_mtx);
471 po->po_error = error; /* save for flush log */
473 PMCDBG2(LOG,WRI,2, "po=%p error=%d", po, error);
478 mtx_lock(&pmc_kthread_mtx);
480 /* put the used buffer back into the global pool */
481 PMCLOG_RESET_BUFFER_DESCRIPTOR(lb);
487 wakeup_one(po->po_kthread);
488 po->po_kthread = NULL;
490 mtx_unlock(&pmc_kthread_mtx);
492 /* return the current I/O buffer to the global pool */
494 PMCLOG_RESET_BUFFER_DESCRIPTOR(lb);
500 * Exit this thread, signalling the waiter
509 * Release and log entry and schedule an I/O if needed.
513 pmclog_release_flags(struct pmc_owner *po, int wakeup)
515 struct pmclog_buffer *plb;
517 plb = po->po_curbuf[curcpu];
518 KASSERT(plb->plb_ptr >= plb->plb_base,
519 ("[pmclog,%d] buffer invariants po=%p ptr=%p base=%p", __LINE__,
520 po, plb->plb_ptr, plb->plb_base));
521 KASSERT(plb->plb_ptr <= plb->plb_fence,
522 ("[pmclog,%d] buffer invariants po=%p ptr=%p fenc=%p", __LINE__,
523 po, plb->plb_ptr, plb->plb_fence));
525 /* schedule an I/O if we've filled a buffer */
526 if (plb->plb_ptr >= plb->plb_fence)
527 pmclog_schedule_io(po, wakeup);
529 PMCDBG1(LOG,REL,1, "po=%p", po);
533 pmclog_release(struct pmc_owner *po)
536 pmclog_release_flags(po, 1);
541 * Attempt to reserve 'length' bytes of space in an owner's log
542 * buffer. The function returns a pointer to 'length' bytes of space
543 * if there was enough space or returns NULL if no space was
544 * available. Non-null returns do so with the po mutex locked. The
545 * caller must invoke pmclog_release() on the pmc owner structure
550 pmclog_reserve(struct pmc_owner *po, int length)
552 uintptr_t newptr, oldptr;
553 struct pmclog_buffer *plb, **pplb;
555 PMCDBG2(LOG,ALL,1, "po=%p len=%d", po, length);
557 KASSERT(length % sizeof(uint32_t) == 0,
558 ("[pmclog,%d] length not a multiple of word size", __LINE__));
560 /* No more data when shutdown in progress. */
561 if (po->po_flags & PMC_PO_SHUTDOWN)
564 pplb = &po->po_curbuf[curcpu];
565 if (*pplb == NULL && pmclog_get_buffer(po) != 0)
568 KASSERT(*pplb != NULL,
569 ("[pmclog,%d] po=%p no current buffer", __LINE__, po));
572 KASSERT(plb->plb_ptr >= plb->plb_base &&
573 plb->plb_ptr <= plb->plb_fence,
574 ("[pmclog,%d] po=%p buffer invariants: ptr=%p base=%p fence=%p",
575 __LINE__, po, plb->plb_ptr, plb->plb_base,
578 oldptr = (uintptr_t) plb->plb_ptr;
579 newptr = oldptr + length;
581 KASSERT(oldptr != (uintptr_t) NULL,
582 ("[pmclog,%d] po=%p Null log buffer pointer", __LINE__, po));
585 * If we have space in the current buffer, return a pointer to
586 * available space with the PO structure locked.
588 if (newptr <= (uintptr_t) plb->plb_fence) {
589 plb->plb_ptr = (char *) newptr;
594 * Otherwise, schedule the current buffer for output and get a
597 pmclog_schedule_io(po, 0);
599 if (pmclog_get_buffer(po) != 0)
604 ("[pmclog,%d] po=%p no current buffer", __LINE__, po));
606 KASSERT(plb->plb_ptr != NULL,
607 ("[pmclog,%d] null return from pmc_get_log_buffer", __LINE__));
609 KASSERT(plb->plb_ptr == plb->plb_base &&
610 plb->plb_ptr <= plb->plb_fence,
611 ("[pmclog,%d] po=%p buffer invariants: ptr=%p base=%p fence=%p",
612 __LINE__, po, plb->plb_ptr, plb->plb_base,
615 oldptr = (uintptr_t) plb->plb_ptr;
618 return ((uint32_t *) oldptr);
626 * Transfer the current buffer to the helper kthread.
630 pmclog_schedule_io(struct pmc_owner *po, int wakeup)
632 struct pmclog_buffer *plb;
634 plb = po->po_curbuf[curcpu];
635 po->po_curbuf[curcpu] = NULL;
637 ("[pmclog,%d] schedule_io with null buffer po=%p", __LINE__, po));
638 KASSERT(plb->plb_ptr >= plb->plb_base,
639 ("[pmclog,%d] buffer invariants po=%p ptr=%p base=%p", __LINE__,
640 po, plb->plb_ptr, plb->plb_base));
641 KASSERT(plb->plb_ptr <= plb->plb_fence,
642 ("[pmclog,%d] buffer invariants po=%p ptr=%p fenc=%p", __LINE__,
643 po, plb->plb_ptr, plb->plb_fence));
645 PMCDBG1(LOG,SIO, 1, "po=%p", po);
648 * Add the current buffer to the tail of the buffer list and
651 mtx_lock_spin(&po->po_mtx);
652 TAILQ_INSERT_TAIL(&po->po_logbuffers, plb, plb_next);
653 mtx_unlock_spin(&po->po_mtx);
659 * Stop the helper kthread.
663 pmclog_stop_kthread(struct pmc_owner *po)
666 mtx_lock(&pmc_kthread_mtx);
667 po->po_flags &= ~PMC_PO_OWNS_LOGFILE;
668 if (po->po_kthread != NULL) {
669 PROC_LOCK(po->po_kthread);
670 kern_psignal(po->po_kthread, SIGHUP);
671 PROC_UNLOCK(po->po_kthread);
674 while (po->po_kthread)
675 msleep(po->po_kthread, &pmc_kthread_mtx, PPAUSE, "pmckstp", 0);
676 mtx_unlock(&pmc_kthread_mtx);
684 * Configure a log file for pmc owner 'po'.
686 * Parameter 'logfd' is a file handle referencing an open file in the
687 * owner process. This file needs to have been opened for writing.
691 pmclog_configure_log(struct pmc_mdep *md, struct pmc_owner *po, int logfd)
698 sx_assert(&pmc_sx, SA_XLOCKED);
699 PMCDBG2(LOG,CFG,1, "config po=%p logfd=%d", po, logfd);
703 /* return EBUSY if a log file was already present */
704 if (po->po_flags & PMC_PO_OWNS_LOGFILE)
707 KASSERT(po->po_file == NULL,
708 ("[pmclog,%d] po=%p file (%p) already present", __LINE__, po,
711 /* get a reference to the file state */
712 error = fget_write(curthread, logfd, &cap_write_rights, &po->po_file);
716 /* mark process as owning a log file */
717 po->po_flags |= PMC_PO_OWNS_LOGFILE;
719 /* mark process as using HWPMCs */
721 p->p_flag |= P_HWPMC;
725 /* create a log initialization entry */
726 PMCLOG_RESERVE_WITH_ERROR(po, INITIALIZE,
727 sizeof(struct pmclog_initialize));
728 PMCLOG_EMIT32(PMC_VERSION);
729 PMCLOG_EMIT32(md->pmd_cputype);
730 #if defined(__i386__) || defined(__amd64__)
731 PMCLOG_EMIT64(tsc_freq);
733 /* other architectures will need to fill this in */
737 memcpy(_le, &ts, sizeof(ts));
739 PMCLOG_EMITSTRING(pmc_cpuid, PMC_CPUID_LEN);
740 PMCLOG_DESPATCH_SYNC(po);
745 KASSERT(po->po_kthread == NULL, ("[pmclog,%d] po=%p kthread not "
746 "stopped", __LINE__, po));
749 (void) fdrop(po->po_file, curthread);
750 po->po_file = NULL; /* clear file and error state */
752 po->po_flags &= ~PMC_PO_OWNS_LOGFILE;
759 * De-configure a log file. This will throw away any buffers queued
760 * for this owner process.
764 pmclog_deconfigure_log(struct pmc_owner *po)
767 struct pmclog_buffer *lb;
769 PMCDBG1(LOG,CFG,1, "de-config po=%p", po);
771 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
774 KASSERT(po->po_sscount == 0,
775 ("[pmclog,%d] po=%p still owning SS PMCs", __LINE__, po));
776 KASSERT(po->po_file != NULL,
777 ("[pmclog,%d] po=%p no log file", __LINE__, po));
779 /* stop the kthread, this will reset the 'OWNS_LOGFILE' flag */
780 pmclog_stop_kthread(po);
782 KASSERT(po->po_kthread == NULL,
783 ("[pmclog,%d] po=%p kthread not stopped", __LINE__, po));
785 /* return all queued log buffers to the global pool */
786 while ((lb = TAILQ_FIRST(&po->po_logbuffers)) != NULL) {
787 TAILQ_REMOVE(&po->po_logbuffers, lb, plb_next);
788 PMCLOG_RESET_BUFFER_DESCRIPTOR(lb);
791 for (int i = 0; i < mp_ncpus; i++) {
792 thread_lock(curthread);
793 sched_bind(curthread, i);
794 thread_unlock(curthread);
795 /* return the 'current' buffer to the global pool */
796 if ((lb = po->po_curbuf[curcpu]) != NULL) {
797 PMCLOG_RESET_BUFFER_DESCRIPTOR(lb);
801 thread_lock(curthread);
802 sched_unbind(curthread);
803 thread_unlock(curthread);
805 /* drop a reference to the fd */
806 if (po->po_file != NULL) {
807 error = fdrop(po->po_file, curthread);
817 * Flush a process' log buffer.
821 pmclog_flush(struct pmc_owner *po, int force)
825 PMCDBG1(LOG,FLS,1, "po=%p", po);
828 * If there is a pending error recorded by the logger thread,
832 return (po->po_error);
837 * Check that we do have an active log file.
839 mtx_lock(&pmc_kthread_mtx);
840 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) {
845 pmclog_schedule_all(po, force);
847 mtx_unlock(&pmc_kthread_mtx);
853 pmclog_schedule_one_cond(struct pmc_owner *po, int force)
855 struct pmclog_buffer *plb;
860 /* tell hardclock not to run again */
861 if (PMC_CPU_HAS_SAMPLES(cpu))
862 PMC_CALL_HOOK_UNLOCKED(curthread, PMC_FN_DO_SAMPLES, NULL);
864 pmc_flush_samples(cpu);
865 plb = po->po_curbuf[cpu];
866 if (plb && plb->plb_ptr != plb->plb_base)
867 pmclog_schedule_io(po, 1);
872 pmclog_schedule_all(struct pmc_owner *po, int force)
875 * Schedule the current buffer if any and not empty.
877 for (int i = 0; i < mp_ncpus; i++) {
878 thread_lock(curthread);
879 sched_bind(curthread, i);
880 thread_unlock(curthread);
881 pmclog_schedule_one_cond(po, force);
883 thread_lock(curthread);
884 sched_unbind(curthread);
885 thread_unlock(curthread);
889 pmclog_close(struct pmc_owner *po)
892 PMCDBG1(LOG,CLO,1, "po=%p", po);
894 pmclog_process_closelog(po);
896 mtx_lock(&pmc_kthread_mtx);
898 * Initiate shutdown: no new data queued,
899 * thread will close file on last block.
901 po->po_flags |= PMC_PO_SHUTDOWN;
902 /* give time for all to see */
906 * Schedule the current buffer.
908 pmclog_schedule_all(po, 0);
911 mtx_unlock(&pmc_kthread_mtx);
917 pmclog_process_callchain(struct pmc *pm, struct pmc_sample *ps)
921 struct pmc_owner *po;
923 PMCDBG3(LOG,SAM,1,"pm=%p pid=%d n=%d", pm, ps->ps_pid,
926 recordlen = offsetof(struct pmclog_callchain, pl_pc) +
927 ps->ps_nsamples * sizeof(uintfptr_t);
929 flags = PMC_CALLCHAIN_TO_CPUFLAGS(ps->ps_cpu,ps->ps_flags);
930 PMCLOG_RESERVE_SAFE(po, CALLCHAIN, recordlen, ps->ps_tsc);
931 PMCLOG_EMIT32(ps->ps_pid);
932 PMCLOG_EMIT32(ps->ps_tid);
933 PMCLOG_EMIT32(pm->pm_id);
934 PMCLOG_EMIT32(flags);
935 for (n = 0; n < ps->ps_nsamples; n++)
936 PMCLOG_EMITADDR(ps->ps_pc[n]);
937 PMCLOG_DESPATCH_SAFE(po);
941 pmclog_process_closelog(struct pmc_owner *po)
943 PMCLOG_RESERVE(po,CLOSELOG,sizeof(struct pmclog_closelog));
944 PMCLOG_DESPATCH_SYNC(po);
948 pmclog_process_dropnotify(struct pmc_owner *po)
950 PMCLOG_RESERVE(po,DROPNOTIFY,sizeof(struct pmclog_dropnotify));
955 pmclog_process_map_in(struct pmc_owner *po, pid_t pid, uintfptr_t start,
958 int pathlen, recordlen;
960 KASSERT(path != NULL, ("[pmclog,%d] map-in, null path", __LINE__));
962 pathlen = strlen(path) + 1; /* #bytes for path name */
963 recordlen = offsetof(struct pmclog_map_in, pl_pathname) +
966 PMCLOG_RESERVE(po, MAP_IN, recordlen);
969 PMCLOG_EMITADDR(start);
970 PMCLOG_EMITSTRING(path,pathlen);
971 PMCLOG_DESPATCH_SYNC(po);
975 pmclog_process_map_out(struct pmc_owner *po, pid_t pid, uintfptr_t start,
978 KASSERT(start <= end, ("[pmclog,%d] start > end", __LINE__));
980 PMCLOG_RESERVE(po, MAP_OUT, sizeof(struct pmclog_map_out));
983 PMCLOG_EMITADDR(start);
984 PMCLOG_EMITADDR(end);
989 pmclog_process_pmcallocate(struct pmc *pm)
991 struct pmc_owner *po;
996 PMCDBG1(LOG,ALL,1, "pm=%p", pm);
998 if (PMC_TO_CLASS(pm) == PMC_CLASS_SOFT) {
999 PMCLOG_RESERVE(po, PMCALLOCATEDYN,
1000 sizeof(struct pmclog_pmcallocatedyn));
1001 PMCLOG_EMIT32(pm->pm_id);
1002 PMCLOG_EMIT32(pm->pm_event);
1003 PMCLOG_EMIT32(pm->pm_flags);
1005 PMCLOG_EMIT64(pm->pm_sc.pm_reloadcount);
1006 ps = pmc_soft_ev_acquire(pm->pm_event);
1008 PMCLOG_EMITSTRING(ps->ps_ev.pm_ev_name,PMC_NAME_MAX);
1010 PMCLOG_EMITNULLSTRING(PMC_NAME_MAX);
1011 pmc_soft_ev_release(ps);
1012 PMCLOG_DESPATCH_SYNC(po);
1014 PMCLOG_RESERVE(po, PMCALLOCATE,
1015 sizeof(struct pmclog_pmcallocate));
1016 PMCLOG_EMIT32(pm->pm_id);
1017 PMCLOG_EMIT32(pm->pm_event);
1018 PMCLOG_EMIT32(pm->pm_flags);
1020 PMCLOG_EMIT64(pm->pm_sc.pm_reloadcount);
1021 PMCLOG_DESPATCH_SYNC(po);
1026 pmclog_process_pmcattach(struct pmc *pm, pid_t pid, char *path)
1028 int pathlen, recordlen;
1029 struct pmc_owner *po;
1031 PMCDBG2(LOG,ATT,1,"pm=%p pid=%d", pm, pid);
1035 pathlen = strlen(path) + 1; /* #bytes for the string */
1036 recordlen = offsetof(struct pmclog_pmcattach, pl_pathname) + pathlen;
1038 PMCLOG_RESERVE(po, PMCATTACH, recordlen);
1039 PMCLOG_EMIT32(pm->pm_id);
1041 PMCLOG_EMITSTRING(path, pathlen);
1042 PMCLOG_DESPATCH_SYNC(po);
1046 pmclog_process_pmcdetach(struct pmc *pm, pid_t pid)
1048 struct pmc_owner *po;
1050 PMCDBG2(LOG,ATT,1,"!pm=%p pid=%d", pm, pid);
1054 PMCLOG_RESERVE(po, PMCDETACH, sizeof(struct pmclog_pmcdetach));
1055 PMCLOG_EMIT32(pm->pm_id);
1057 PMCLOG_DESPATCH_SYNC(po);
1061 pmclog_process_proccreate(struct pmc_owner *po, struct proc *p, int sync)
1064 PMCLOG_RESERVE(po, PROC_CREATE, sizeof(struct pmclog_proccreate));
1065 PMCLOG_EMIT32(p->p_pid);
1066 PMCLOG_EMIT32(p->p_flag);
1067 PMCLOG_EMITSTRING(p->p_comm, MAXCOMLEN+1);
1068 PMCLOG_DESPATCH_SYNC(po);
1070 PMCLOG_RESERVE(po, PROC_CREATE, sizeof(struct pmclog_proccreate));
1071 PMCLOG_EMIT32(p->p_pid);
1072 PMCLOG_EMIT32(p->p_flag);
1073 PMCLOG_EMITSTRING(p->p_comm, MAXCOMLEN+1);
1074 PMCLOG_DESPATCH(po);
1079 * Log a context switch event to the log file.
1083 pmclog_process_proccsw(struct pmc *pm, struct pmc_process *pp, pmc_value_t v, struct thread *td)
1085 struct pmc_owner *po;
1087 KASSERT(pm->pm_flags & PMC_F_LOG_PROCCSW,
1088 ("[pmclog,%d] log-process-csw called gratuitously", __LINE__));
1090 PMCDBG3(LOG,SWO,1,"pm=%p pid=%d v=%jx", pm, pp->pp_proc->p_pid,
1095 PMCLOG_RESERVE_SAFE(po, PROCCSW, sizeof(struct pmclog_proccsw), pmc_rdtsc());
1097 PMCLOG_EMIT32(pm->pm_id);
1098 PMCLOG_EMIT32(pp->pp_proc->p_pid);
1099 PMCLOG_EMIT32(td->td_tid);
1101 PMCLOG_DESPATCH_SCHED_LOCK(po);
1105 pmclog_process_procexec(struct pmc_owner *po, pmc_id_t pmid, pid_t pid,
1106 uintfptr_t startaddr, char *path)
1108 int pathlen, recordlen;
1110 PMCDBG3(LOG,EXC,1,"po=%p pid=%d path=\"%s\"", po, pid, path);
1112 pathlen = strlen(path) + 1; /* #bytes for the path */
1113 recordlen = offsetof(struct pmclog_procexec, pl_pathname) + pathlen;
1114 PMCLOG_RESERVE(po, PROCEXEC, recordlen);
1116 PMCLOG_EMIT32(pmid);
1117 PMCLOG_EMITADDR(startaddr);
1118 PMCLOG_EMITSTRING(path,pathlen);
1119 PMCLOG_DESPATCH_SYNC(po);
1123 * Log a process exit event (and accumulated pmc value) to the log file.
1127 pmclog_process_procexit(struct pmc *pm, struct pmc_process *pp)
1130 struct pmc_owner *po;
1132 ri = PMC_TO_ROWINDEX(pm);
1133 PMCDBG3(LOG,EXT,1,"pm=%p pid=%d v=%jx", pm, pp->pp_proc->p_pid,
1134 pp->pp_pmcs[ri].pp_pmcval);
1138 PMCLOG_RESERVE(po, PROCEXIT, sizeof(struct pmclog_procexit));
1139 PMCLOG_EMIT32(pm->pm_id);
1140 PMCLOG_EMIT32(pp->pp_proc->p_pid);
1141 PMCLOG_EMIT64(pp->pp_pmcs[ri].pp_pmcval);
1142 PMCLOG_DESPATCH(po);
1150 pmclog_process_procfork(struct pmc_owner *po, pid_t oldpid, pid_t newpid)
1152 PMCLOG_RESERVE(po, PROCFORK, sizeof(struct pmclog_procfork));
1153 PMCLOG_EMIT32(oldpid);
1154 PMCLOG_EMIT32(newpid);
1155 PMCLOG_DESPATCH(po);
1159 * Log a process exit event of the form suitable for system-wide PMCs.
1163 pmclog_process_sysexit(struct pmc_owner *po, pid_t pid)
1165 PMCLOG_RESERVE(po, SYSEXIT, sizeof(struct pmclog_sysexit));
1167 PMCLOG_DESPATCH(po);
1171 pmclog_process_threadcreate(struct pmc_owner *po, struct thread *td, int sync)
1177 PMCLOG_RESERVE(po, THR_CREATE, sizeof(struct pmclog_threadcreate));
1178 PMCLOG_EMIT32(td->td_tid);
1179 PMCLOG_EMIT32(p->p_pid);
1180 PMCLOG_EMIT32(p->p_flag);
1182 PMCLOG_EMITSTRING(td->td_name, MAXCOMLEN+1);
1183 PMCLOG_DESPATCH_SYNC(po);
1185 PMCLOG_RESERVE(po, THR_CREATE, sizeof(struct pmclog_threadcreate));
1186 PMCLOG_EMIT32(td->td_tid);
1187 PMCLOG_EMIT32(p->p_pid);
1188 PMCLOG_EMIT32(p->p_flag);
1190 PMCLOG_EMITSTRING(td->td_name, MAXCOMLEN+1);
1191 PMCLOG_DESPATCH(po);
1196 pmclog_process_threadexit(struct pmc_owner *po, struct thread *td)
1199 PMCLOG_RESERVE(po, THR_EXIT, sizeof(struct pmclog_threadexit));
1200 PMCLOG_EMIT32(td->td_tid);
1201 PMCLOG_DESPATCH(po);
1205 * Write a user log entry.
1209 pmclog_process_userlog(struct pmc_owner *po, struct pmc_op_writelog *wl)
1213 PMCDBG2(LOG,WRI,1, "writelog po=%p ud=0x%x", po, wl->pm_userdata);
1217 PMCLOG_RESERVE_WITH_ERROR(po, USERDATA,
1218 sizeof(struct pmclog_userdata));
1219 PMCLOG_EMIT32(wl->pm_userdata);
1220 PMCLOG_DESPATCH(po);
1229 * Create a pool of log buffers and initialize mutexes.
1236 struct pmclog_buffer *plb;
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;
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;
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;
1256 for (domain = 0; domain < vm_ndomains; domain++) {
1257 int ncpus = pmc_dom_hdrs[domain]->pdbh_ncpus;
1258 int total = ncpus*pmc_nlogbuffers_pcpu;
1260 plb = malloc_domain(sizeof(struct pmclog_buffer)*total, M_PMC, domain, M_WAITOK|M_ZERO);
1261 pmc_dom_hdrs[domain]->pdbh_plbs = plb;
1262 for (int i = 0; i < total; i++, plb++) {
1265 buf = malloc_domain(1024 * pmclog_buffer_size, M_PMC, domain,
1267 PMCLOG_INIT_BUFFER_DESCRIPTOR(plb, buf, domain);
1268 pmc_plb_rele_unlocked(plb);
1271 mtx_init(&pmc_kthread_mtx, "pmc-kthread", "pmc-sleep", MTX_DEF);
1277 * Destroy mutexes and release memory back the to free pool.
1283 struct pmclog_buffer *plb;
1286 mtx_destroy(&pmc_kthread_mtx);
1288 for (domain = 0; domain < vm_ndomains; domain++) {
1289 while ((plb = TAILQ_FIRST(&pmc_dom_hdrs[domain]->pdbh_head)) != NULL) {
1290 TAILQ_REMOVE(&pmc_dom_hdrs[domain]->pdbh_head, plb, plb_next);
1291 free(plb->plb_base, M_PMC);
1293 free(pmc_dom_hdrs[domain]->pdbh_plbs, M_PMC);