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>
44 #include <sys/domainset.h>
46 #include <sys/kernel.h>
47 #include <sys/kthread.h>
49 #include <sys/module.h>
50 #include <sys/mutex.h>
52 #include <sys/pmckern.h>
53 #include <sys/pmclog.h>
55 #include <sys/sched.h>
56 #include <sys/signalvar.h>
58 #include <sys/syscallsubr.h>
59 #include <sys/sysctl.h>
60 #include <sys/systm.h>
62 #include <sys/unistd.h>
63 #include <sys/vnode.h>
65 #if defined(__i386__) || defined(__amd64__)
66 #include <machine/clock.h>
69 #define curdomain PCPU_GET(domain)
75 SYSCTL_DECL(_kern_hwpmc);
78 * kern.hwpmc.logbuffersize -- size of the per-cpu owner buffers.
81 static int pmclog_buffer_size = PMC_LOG_BUFFER_SIZE;
82 #if (__FreeBSD_version < 1100000)
83 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "logbuffersize", &pmclog_buffer_size);
85 SYSCTL_INT(_kern_hwpmc, OID_AUTO, logbuffersize, CTLFLAG_RDTUN,
86 &pmclog_buffer_size, 0, "size of log buffers in kilobytes");
89 * kern.hwpmc.nbuffers_pcpu -- number of global log buffers
92 static int pmc_nlogbuffers_pcpu = PMC_NLOGBUFFERS_PCPU;
93 #if (__FreeBSD_version < 1100000)
94 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "nbuffers", &pmc_nlogbuffers_pcpu);
96 SYSCTL_INT(_kern_hwpmc, OID_AUTO, nbuffers_pcpu, CTLFLAG_RDTUN,
97 &pmc_nlogbuffers_pcpu, 0, "number of log buffers per cpu");
100 * Global log buffer list and associated spin lock.
103 static struct mtx pmc_kthread_mtx; /* sleep lock */
105 #define PMCLOG_INIT_BUFFER_DESCRIPTOR(D, buf, domain) do { \
106 (D)->plb_fence = ((char *) (buf)) + 1024*pmclog_buffer_size; \
107 (D)->plb_base = (D)->plb_ptr = ((char *) (buf)); \
108 (D)->plb_domain = domain; \
111 #define PMCLOG_RESET_BUFFER_DESCRIPTOR(D) do { \
112 (D)->plb_ptr = (D)->plb_base; \
116 * Log file record constructors.
118 #define _PMCLOG_TO_HEADER(T,L) \
119 ((PMCLOG_HEADER_MAGIC << 24) | \
120 (PMCLOG_TYPE_ ## T << 16) | \
123 /* reserve LEN bytes of space and initialize the entry header */
124 #define _PMCLOG_RESERVE_SAFE(PO,TYPE,LEN,ACTION, TSC) do { \
126 int _len = roundup((LEN), sizeof(uint32_t)); \
127 struct pmclog_header *ph; \
128 if ((_le = pmclog_reserve((PO), _len)) == NULL) { \
131 ph = (struct pmclog_header *)_le; \
132 ph->pl_header =_PMCLOG_TO_HEADER(TYPE,_len); \
133 ph->pl_tsc = (TSC); \
134 _le += sizeof(*ph)/4 /* skip over timestamp */
136 /* reserve LEN bytes of space and initialize the entry header */
137 #define _PMCLOG_RESERVE(PO,TYPE,LEN,ACTION) do { \
139 int _len = roundup((LEN), sizeof(uint32_t)); \
141 struct pmclog_header *ph; \
144 if ((_le = pmclog_reserve((PO), _len)) == NULL) { \
148 ph = (struct pmclog_header *)_le; \
149 ph->pl_header =_PMCLOG_TO_HEADER(TYPE,_len); \
151 _le += sizeof(*ph)/4 /* skip over timestamp */
155 #define PMCLOG_RESERVE_SAFE(P,T,L,TSC) _PMCLOG_RESERVE_SAFE(P,T,L,return,TSC)
156 #define PMCLOG_RESERVE(P,T,L) _PMCLOG_RESERVE(P,T,L,return)
157 #define PMCLOG_RESERVE_WITH_ERROR(P,T,L) _PMCLOG_RESERVE(P,T,L, \
158 error=ENOMEM;goto error)
160 #define PMCLOG_EMIT32(V) do { *_le++ = (V); } while (0)
161 #define PMCLOG_EMIT64(V) do { \
162 *_le++ = (uint32_t) ((V) & 0xFFFFFFFF); \
163 *_le++ = (uint32_t) (((V) >> 32) & 0xFFFFFFFF); \
167 /* Emit a string. Caution: does NOT update _le, so needs to be last */
168 #define PMCLOG_EMITSTRING(S,L) do { bcopy((S), _le, (L)); } while (0)
169 #define PMCLOG_EMITNULLSTRING(L) do { bzero(_le, (L)); } while (0)
171 #define PMCLOG_DESPATCH_SAFE(PO) \
172 pmclog_release((PO)); \
175 #define PMCLOG_DESPATCH_SCHED_LOCK(PO) \
176 pmclog_release_flags((PO), 0); \
179 #define PMCLOG_DESPATCH(PO) \
180 pmclog_release((PO)); \
184 #define PMCLOG_DESPATCH_SYNC(PO) \
185 pmclog_schedule_io((PO), 1); \
192 * Assertions about the log file format.
194 CTASSERT(sizeof(struct pmclog_callchain) == 7*4 + TSDELTA +
195 PMC_CALLCHAIN_DEPTH_MAX*sizeof(uintfptr_t));
196 CTASSERT(sizeof(struct pmclog_closelog) == 3*4 + TSDELTA);
197 CTASSERT(sizeof(struct pmclog_dropnotify) == 3*4 + TSDELTA);
198 CTASSERT(sizeof(struct pmclog_map_in) == PATH_MAX + TSDELTA +
199 5*4 + sizeof(uintfptr_t));
200 CTASSERT(offsetof(struct pmclog_map_in,pl_pathname) ==
201 5*4 + TSDELTA + sizeof(uintfptr_t));
202 CTASSERT(sizeof(struct pmclog_map_out) == 5*4 + 2*sizeof(uintfptr_t) + TSDELTA);
203 CTASSERT(sizeof(struct pmclog_pmcallocate) == 9*4 + TSDELTA);
204 CTASSERT(sizeof(struct pmclog_pmcattach) == 5*4 + PATH_MAX + TSDELTA);
205 CTASSERT(offsetof(struct pmclog_pmcattach,pl_pathname) == 5*4 + TSDELTA);
206 CTASSERT(sizeof(struct pmclog_pmcdetach) == 5*4 + TSDELTA);
207 CTASSERT(sizeof(struct pmclog_proccsw) == 7*4 + 8 + TSDELTA);
208 CTASSERT(sizeof(struct pmclog_procexec) == 5*4 + PATH_MAX +
209 sizeof(uintfptr_t) + TSDELTA);
210 CTASSERT(offsetof(struct pmclog_procexec,pl_pathname) == 5*4 + TSDELTA +
212 CTASSERT(sizeof(struct pmclog_procexit) == 5*4 + 8 + TSDELTA);
213 CTASSERT(sizeof(struct pmclog_procfork) == 5*4 + TSDELTA);
214 CTASSERT(sizeof(struct pmclog_sysexit) == 6*4);
215 CTASSERT(sizeof(struct pmclog_userdata) == 6*4);
218 * Log buffer structure
221 struct pmclog_buffer {
222 TAILQ_ENTRY(pmclog_buffer) plb_next;
227 } __aligned(CACHE_LINE_SIZE);
233 static int pmclog_get_buffer(struct pmc_owner *po);
234 static void pmclog_loop(void *arg);
235 static void pmclog_release(struct pmc_owner *po);
236 static uint32_t *pmclog_reserve(struct pmc_owner *po, int length);
237 static void pmclog_schedule_io(struct pmc_owner *po, int wakeup);
238 static void pmclog_schedule_all(struct pmc_owner *po);
239 static void pmclog_stop_kthread(struct pmc_owner *po);
246 pmc_plb_rele_unlocked(struct pmclog_buffer *plb)
248 TAILQ_INSERT_HEAD(&pmc_dom_hdrs[plb->plb_domain]->pdbh_head, plb, plb_next);
252 pmc_plb_rele(struct pmclog_buffer *plb)
254 mtx_lock_spin(&pmc_dom_hdrs[plb->plb_domain]->pdbh_mtx);
255 pmc_plb_rele_unlocked(plb);
256 mtx_unlock_spin(&pmc_dom_hdrs[plb->plb_domain]->pdbh_mtx);
263 pmclog_get_buffer(struct pmc_owner *po)
265 struct pmclog_buffer *plb;
268 KASSERT(po->po_curbuf[curcpu] == NULL,
269 ("[pmclog,%d] po=%p current buffer still valid", __LINE__, po));
272 MPASS(pmc_dom_hdrs[domain]);
273 mtx_lock_spin(&pmc_dom_hdrs[domain]->pdbh_mtx);
274 if ((plb = TAILQ_FIRST(&pmc_dom_hdrs[domain]->pdbh_head)) != NULL)
275 TAILQ_REMOVE(&pmc_dom_hdrs[domain]->pdbh_head, plb, plb_next);
276 mtx_unlock_spin(&pmc_dom_hdrs[domain]->pdbh_mtx);
278 PMCDBG2(LOG,GTB,1, "po=%p plb=%p", po, plb);
282 KASSERT(plb->plb_ptr == plb->plb_base &&
283 plb->plb_base < plb->plb_fence,
284 ("[pmclog,%d] po=%p buffer invariants: ptr=%p "
285 "base=%p fence=%p", __LINE__, po, plb->plb_ptr,
286 plb->plb_base, plb->plb_fence));
289 po->po_curbuf[curcpu] = plb;
292 counter_u64_add(pmc_stats.pm_buffer_requests, 1);
294 counter_u64_add(pmc_stats.pm_buffer_requests_failed, 1);
296 return (plb ? 0 : ENOMEM);
299 struct pmclog_proc_init_args {
301 struct pmc_owner *po;
307 pmclog_proc_create(struct thread *td, void **handlep)
309 struct pmclog_proc_init_args *ia;
312 ia = malloc(sizeof(*ia), M_TEMP, M_WAITOK | M_ZERO);
313 error = kproc_create(pmclog_loop, ia, &ia->kthr,
314 RFHIGHPID, 0, "hwpmc: proc(%d)", td->td_proc->p_pid);
321 pmclog_proc_ignite(void *handle, struct pmc_owner *po)
323 struct pmclog_proc_init_args *ia;
326 mtx_lock(&pmc_kthread_mtx);
328 MPASS(ia->po == NULL);
330 MPASS(ia->kthr != NULL);
335 KASSERT(po->po_kthread == NULL,
336 ("[pmclog,%d] po=%p kthread (%p) already present",
337 __LINE__, po, po->po_kthread));
338 po->po_kthread = ia->kthr;
342 msleep(ia, &pmc_kthread_mtx, PWAIT, "pmclogw", 0);
343 mtx_unlock(&pmc_kthread_mtx);
350 * This function is executed by each pmc owner's helper thread.
353 pmclog_loop(void *arg)
355 struct pmclog_proc_init_args *ia;
356 struct pmc_owner *po;
357 struct pmclog_buffer *lb;
359 struct ucred *ownercred;
360 struct ucred *mycred;
371 SIGADDSET(unb, SIGHUP);
372 (void)kern_sigprocmask(td, SIG_UNBLOCK, &unb, NULL, 0);
375 MPASS(ia->kthr == curproc);
377 mtx_lock(&pmc_kthread_mtx);
378 while (ia->po == NULL && !ia->exit)
379 msleep(ia, &pmc_kthread_mtx, PWAIT, "pmclogi", 0);
383 mtx_unlock(&pmc_kthread_mtx);
386 MPASS(ia->po != NULL);
390 mtx_unlock(&pmc_kthread_mtx);
394 mycred = td->td_ucred;
397 ownercred = crhold(p->p_ucred);
400 PMCDBG2(LOG,INI,1, "po=%p kt=%p", po, po->po_kthread);
401 KASSERT(po->po_kthread == curthread->td_proc,
402 ("[pmclog,%d] proc mismatch po=%p po/kt=%p curproc=%p", __LINE__,
403 po, po->po_kthread, curthread->td_proc));
409 * Loop waiting for I/O requests to be added to the owner
410 * struct's queue. The loop is exited when the log file
414 mtx_lock(&pmc_kthread_mtx);
418 /* check if we've been asked to exit */
419 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
422 if (lb == NULL) { /* look for a fresh buffer to write */
423 mtx_lock_spin(&po->po_mtx);
424 if ((lb = TAILQ_FIRST(&po->po_logbuffers)) == NULL) {
425 mtx_unlock_spin(&po->po_mtx);
427 /* No more buffers and shutdown required. */
428 if (po->po_flags & PMC_PO_SHUTDOWN)
431 (void) msleep(po, &pmc_kthread_mtx, PWAIT,
436 TAILQ_REMOVE(&po->po_logbuffers, lb, plb_next);
437 mtx_unlock_spin(&po->po_mtx);
440 mtx_unlock(&pmc_kthread_mtx);
442 /* process the request */
443 PMCDBG3(LOG,WRI,2, "po=%p base=%p ptr=%p", po,
444 lb->plb_base, lb->plb_ptr);
445 /* change our thread's credentials before issuing the I/O */
447 aiov.iov_base = lb->plb_base;
448 aiov.iov_len = nbytes = lb->plb_ptr - lb->plb_base;
450 auio.uio_iov = &aiov;
452 auio.uio_offset = -1;
453 auio.uio_resid = nbytes;
454 auio.uio_rw = UIO_WRITE;
455 auio.uio_segflg = UIO_SYSSPACE;
458 /* switch thread credentials -- see kern_ktrace.c */
459 td->td_ucred = ownercred;
460 error = fo_write(po->po_file, &auio, ownercred, 0, td);
461 td->td_ucred = mycred;
464 /* XXX some errors are recoverable */
465 /* send a SIGIO to the owner and exit */
467 kern_psignal(p, SIGIO);
470 mtx_lock(&pmc_kthread_mtx);
472 po->po_error = error; /* save for flush log */
474 PMCDBG2(LOG,WRI,2, "po=%p error=%d", po, error);
479 mtx_lock(&pmc_kthread_mtx);
481 /* put the used buffer back into the global pool */
482 PMCLOG_RESET_BUFFER_DESCRIPTOR(lb);
488 wakeup_one(po->po_kthread);
489 po->po_kthread = NULL;
491 mtx_unlock(&pmc_kthread_mtx);
493 /* return the current I/O buffer to the global pool */
495 PMCLOG_RESET_BUFFER_DESCRIPTOR(lb);
501 * Exit this thread, signalling the waiter
510 * Release and log entry and schedule an I/O if needed.
514 pmclog_release_flags(struct pmc_owner *po, int wakeup)
516 struct pmclog_buffer *plb;
518 plb = po->po_curbuf[curcpu];
519 KASSERT(plb->plb_ptr >= plb->plb_base,
520 ("[pmclog,%d] buffer invariants po=%p ptr=%p base=%p", __LINE__,
521 po, plb->plb_ptr, plb->plb_base));
522 KASSERT(plb->plb_ptr <= plb->plb_fence,
523 ("[pmclog,%d] buffer invariants po=%p ptr=%p fenc=%p", __LINE__,
524 po, plb->plb_ptr, plb->plb_fence));
526 /* schedule an I/O if we've filled a buffer */
527 if (plb->plb_ptr >= plb->plb_fence)
528 pmclog_schedule_io(po, wakeup);
530 PMCDBG1(LOG,REL,1, "po=%p", po);
534 pmclog_release(struct pmc_owner *po)
537 pmclog_release_flags(po, 1);
542 * Attempt to reserve 'length' bytes of space in an owner's log
543 * buffer. The function returns a pointer to 'length' bytes of space
544 * if there was enough space or returns NULL if no space was
545 * available. Non-null returns do so with the po mutex locked. The
546 * caller must invoke pmclog_release() on the pmc owner structure
551 pmclog_reserve(struct pmc_owner *po, int length)
553 uintptr_t newptr, oldptr __diagused;
554 struct pmclog_buffer *plb, **pplb;
556 PMCDBG2(LOG,ALL,1, "po=%p len=%d", po, length);
558 KASSERT(length % sizeof(uint32_t) == 0,
559 ("[pmclog,%d] length not a multiple of word size", __LINE__));
561 /* No more data when shutdown in progress. */
562 if (po->po_flags & PMC_PO_SHUTDOWN)
565 pplb = &po->po_curbuf[curcpu];
566 if (*pplb == NULL && pmclog_get_buffer(po) != 0)
569 KASSERT(*pplb != NULL,
570 ("[pmclog,%d] po=%p no current buffer", __LINE__, po));
573 KASSERT(plb->plb_ptr >= plb->plb_base &&
574 plb->plb_ptr <= plb->plb_fence,
575 ("[pmclog,%d] po=%p buffer invariants: ptr=%p base=%p fence=%p",
576 __LINE__, po, plb->plb_ptr, plb->plb_base,
579 oldptr = (uintptr_t) plb->plb_ptr;
580 newptr = oldptr + length;
582 KASSERT(oldptr != (uintptr_t) NULL,
583 ("[pmclog,%d] po=%p Null log buffer pointer", __LINE__, po));
586 * If we have space in the current buffer, return a pointer to
587 * available space with the PO structure locked.
589 if (newptr <= (uintptr_t) plb->plb_fence) {
590 plb->plb_ptr = (char *) newptr;
595 * Otherwise, schedule the current buffer for output and get a
598 pmclog_schedule_io(po, 0);
600 if (pmclog_get_buffer(po) != 0)
605 ("[pmclog,%d] po=%p no current buffer", __LINE__, po));
607 KASSERT(plb->plb_ptr != NULL,
608 ("[pmclog,%d] null return from pmc_get_log_buffer", __LINE__));
610 KASSERT(plb->plb_ptr == plb->plb_base &&
611 plb->plb_ptr <= plb->plb_fence,
612 ("[pmclog,%d] po=%p buffer invariants: ptr=%p base=%p fence=%p",
613 __LINE__, po, plb->plb_ptr, plb->plb_base,
616 oldptr = (uintptr_t) plb->plb_ptr;
619 return ((uint32_t *) oldptr);
627 * Transfer the current buffer to the helper kthread.
631 pmclog_schedule_io(struct pmc_owner *po, int wakeup)
633 struct pmclog_buffer *plb;
635 plb = po->po_curbuf[curcpu];
636 po->po_curbuf[curcpu] = NULL;
638 ("[pmclog,%d] schedule_io with null buffer po=%p", __LINE__, po));
639 KASSERT(plb->plb_ptr >= plb->plb_base,
640 ("[pmclog,%d] buffer invariants po=%p ptr=%p base=%p", __LINE__,
641 po, plb->plb_ptr, plb->plb_base));
642 KASSERT(plb->plb_ptr <= plb->plb_fence,
643 ("[pmclog,%d] buffer invariants po=%p ptr=%p fenc=%p", __LINE__,
644 po, plb->plb_ptr, plb->plb_fence));
646 PMCDBG1(LOG,SIO, 1, "po=%p", po);
649 * Add the current buffer to the tail of the buffer list and
652 mtx_lock_spin(&po->po_mtx);
653 TAILQ_INSERT_TAIL(&po->po_logbuffers, plb, plb_next);
654 mtx_unlock_spin(&po->po_mtx);
660 * Stop the helper kthread.
664 pmclog_stop_kthread(struct pmc_owner *po)
667 mtx_lock(&pmc_kthread_mtx);
668 po->po_flags &= ~PMC_PO_OWNS_LOGFILE;
669 if (po->po_kthread != NULL) {
670 PROC_LOCK(po->po_kthread);
671 kern_psignal(po->po_kthread, SIGHUP);
672 PROC_UNLOCK(po->po_kthread);
675 while (po->po_kthread)
676 msleep(po->po_kthread, &pmc_kthread_mtx, PPAUSE, "pmckstp", 0);
677 mtx_unlock(&pmc_kthread_mtx);
685 * Configure a log file for pmc owner 'po'.
687 * Parameter 'logfd' is a file handle referencing an open file in the
688 * owner process. This file needs to have been opened for writing.
692 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;
724 /* create a log initialization entry */
725 PMCLOG_RESERVE_WITH_ERROR(po, INITIALIZE,
726 sizeof(struct pmclog_initialize));
727 PMCLOG_EMIT32(PMC_VERSION);
728 PMCLOG_EMIT32(md->pmd_cputype);
729 #if defined(__i386__) || defined(__amd64__)
730 PMCLOG_EMIT64(tsc_freq);
732 /* other architectures will need to fill this in */
736 memcpy(_le, &ts, sizeof(ts));
738 PMCLOG_EMITSTRING(pmc_cpuid, PMC_CPUID_LEN);
739 PMCLOG_DESPATCH_SYNC(po);
744 KASSERT(po->po_kthread == NULL, ("[pmclog,%d] po=%p kthread not "
745 "stopped", __LINE__, po));
748 (void) fdrop(po->po_file, curthread);
749 po->po_file = NULL; /* clear file and error state */
751 po->po_flags &= ~PMC_PO_OWNS_LOGFILE;
758 * De-configure a log file. This will throw away any buffers queued
759 * for this owner process.
763 pmclog_deconfigure_log(struct pmc_owner *po)
766 struct pmclog_buffer *lb;
768 PMCDBG1(LOG,CFG,1, "de-config po=%p", po);
770 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
773 KASSERT(po->po_sscount == 0,
774 ("[pmclog,%d] po=%p still owning SS PMCs", __LINE__, po));
775 KASSERT(po->po_file != NULL,
776 ("[pmclog,%d] po=%p no log file", __LINE__, po));
778 /* stop the kthread, this will reset the 'OWNS_LOGFILE' flag */
779 pmclog_stop_kthread(po);
781 KASSERT(po->po_kthread == NULL,
782 ("[pmclog,%d] po=%p kthread not stopped", __LINE__, po));
784 /* return all queued log buffers to the global pool */
785 while ((lb = TAILQ_FIRST(&po->po_logbuffers)) != NULL) {
786 TAILQ_REMOVE(&po->po_logbuffers, lb, plb_next);
787 PMCLOG_RESET_BUFFER_DESCRIPTOR(lb);
790 for (int i = 0; i < mp_ncpus; i++) {
791 thread_lock(curthread);
792 sched_bind(curthread, i);
793 thread_unlock(curthread);
794 /* return the 'current' buffer to the global pool */
795 if ((lb = po->po_curbuf[curcpu]) != NULL) {
796 PMCLOG_RESET_BUFFER_DESCRIPTOR(lb);
800 thread_lock(curthread);
801 sched_unbind(curthread);
802 thread_unlock(curthread);
804 /* drop a reference to the fd */
805 if (po->po_file != NULL) {
806 error = fdrop(po->po_file, curthread);
816 * Flush a process' log buffer.
820 pmclog_flush(struct pmc_owner *po, int force)
824 PMCDBG1(LOG,FLS,1, "po=%p", po);
827 * If there is a pending error recorded by the logger thread,
831 return (po->po_error);
836 * Check that we do have an active log file.
838 mtx_lock(&pmc_kthread_mtx);
839 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) {
844 pmclog_schedule_all(po);
846 mtx_unlock(&pmc_kthread_mtx);
852 pmclog_schedule_one_cond(struct pmc_owner *po)
854 struct pmclog_buffer *plb;
859 /* tell hardclock not to run again */
860 if (PMC_CPU_HAS_SAMPLES(cpu))
861 PMC_CALL_HOOK_UNLOCKED(curthread, PMC_FN_DO_SAMPLES, NULL);
863 plb = po->po_curbuf[cpu];
864 if (plb && plb->plb_ptr != plb->plb_base)
865 pmclog_schedule_io(po, 1);
870 pmclog_schedule_all(struct pmc_owner *po)
873 * Schedule the current buffer if any and not empty.
875 for (int i = 0; i < mp_ncpus; i++) {
876 thread_lock(curthread);
877 sched_bind(curthread, i);
878 thread_unlock(curthread);
879 pmclog_schedule_one_cond(po);
881 thread_lock(curthread);
882 sched_unbind(curthread);
883 thread_unlock(curthread);
887 pmclog_close(struct pmc_owner *po)
890 PMCDBG1(LOG,CLO,1, "po=%p", po);
892 pmclog_process_closelog(po);
894 mtx_lock(&pmc_kthread_mtx);
896 * Initiate shutdown: no new data queued,
897 * thread will close file on last block.
899 po->po_flags |= PMC_PO_SHUTDOWN;
900 /* give time for all to see */
904 * Schedule the current buffer.
906 pmclog_schedule_all(po);
909 mtx_unlock(&pmc_kthread_mtx);
915 pmclog_process_callchain(struct pmc *pm, struct pmc_sample *ps)
919 struct pmc_owner *po;
921 PMCDBG3(LOG,SAM,1,"pm=%p pid=%d n=%d", pm, ps->ps_pid,
924 recordlen = offsetof(struct pmclog_callchain, pl_pc) +
925 ps->ps_nsamples * sizeof(uintfptr_t);
927 flags = PMC_CALLCHAIN_TO_CPUFLAGS(ps->ps_cpu,ps->ps_flags);
928 PMCLOG_RESERVE_SAFE(po, CALLCHAIN, recordlen, ps->ps_tsc);
929 PMCLOG_EMIT32(ps->ps_pid);
930 PMCLOG_EMIT32(ps->ps_tid);
931 PMCLOG_EMIT32(pm->pm_id);
932 PMCLOG_EMIT32(flags);
933 for (n = 0; n < ps->ps_nsamples; n++)
934 PMCLOG_EMITADDR(ps->ps_pc[n]);
935 PMCLOG_DESPATCH_SAFE(po);
939 pmclog_process_closelog(struct pmc_owner *po)
941 PMCLOG_RESERVE(po,CLOSELOG,sizeof(struct pmclog_closelog));
942 PMCLOG_DESPATCH_SYNC(po);
946 pmclog_process_dropnotify(struct pmc_owner *po)
948 PMCLOG_RESERVE(po,DROPNOTIFY,sizeof(struct pmclog_dropnotify));
953 pmclog_process_map_in(struct pmc_owner *po, pid_t pid, uintfptr_t start,
956 int pathlen, recordlen;
958 KASSERT(path != NULL, ("[pmclog,%d] map-in, null path", __LINE__));
960 pathlen = strlen(path) + 1; /* #bytes for path name */
961 recordlen = offsetof(struct pmclog_map_in, pl_pathname) +
964 PMCLOG_RESERVE(po, MAP_IN, recordlen);
967 PMCLOG_EMITADDR(start);
968 PMCLOG_EMITSTRING(path,pathlen);
969 PMCLOG_DESPATCH_SYNC(po);
973 pmclog_process_map_out(struct pmc_owner *po, pid_t pid, uintfptr_t start,
976 KASSERT(start <= end, ("[pmclog,%d] start > end", __LINE__));
978 PMCLOG_RESERVE(po, MAP_OUT, sizeof(struct pmclog_map_out));
981 PMCLOG_EMITADDR(start);
982 PMCLOG_EMITADDR(end);
987 pmclog_process_pmcallocate(struct pmc *pm)
989 struct pmc_owner *po;
994 PMCDBG1(LOG,ALL,1, "pm=%p", pm);
996 if (PMC_TO_CLASS(pm) == PMC_CLASS_SOFT) {
997 PMCLOG_RESERVE(po, PMCALLOCATEDYN,
998 sizeof(struct pmclog_pmcallocatedyn));
999 PMCLOG_EMIT32(pm->pm_id);
1000 PMCLOG_EMIT32(pm->pm_event);
1001 PMCLOG_EMIT32(pm->pm_flags);
1003 PMCLOG_EMIT64(pm->pm_sc.pm_reloadcount);
1004 ps = pmc_soft_ev_acquire(pm->pm_event);
1006 PMCLOG_EMITSTRING(ps->ps_ev.pm_ev_name,PMC_NAME_MAX);
1008 PMCLOG_EMITNULLSTRING(PMC_NAME_MAX);
1009 pmc_soft_ev_release(ps);
1010 PMCLOG_DESPATCH_SYNC(po);
1012 PMCLOG_RESERVE(po, PMCALLOCATE,
1013 sizeof(struct pmclog_pmcallocate));
1014 PMCLOG_EMIT32(pm->pm_id);
1015 PMCLOG_EMIT32(pm->pm_event);
1016 PMCLOG_EMIT32(pm->pm_flags);
1018 PMCLOG_EMIT64(pm->pm_sc.pm_reloadcount);
1019 PMCLOG_DESPATCH_SYNC(po);
1024 pmclog_process_pmcattach(struct pmc *pm, pid_t pid, char *path)
1026 int pathlen, recordlen;
1027 struct pmc_owner *po;
1029 PMCDBG2(LOG,ATT,1,"pm=%p pid=%d", pm, pid);
1033 pathlen = strlen(path) + 1; /* #bytes for the string */
1034 recordlen = offsetof(struct pmclog_pmcattach, pl_pathname) + pathlen;
1036 PMCLOG_RESERVE(po, PMCATTACH, recordlen);
1037 PMCLOG_EMIT32(pm->pm_id);
1039 PMCLOG_EMITSTRING(path, pathlen);
1040 PMCLOG_DESPATCH_SYNC(po);
1044 pmclog_process_pmcdetach(struct pmc *pm, pid_t pid)
1046 struct pmc_owner *po;
1048 PMCDBG2(LOG,ATT,1,"!pm=%p pid=%d", pm, pid);
1052 PMCLOG_RESERVE(po, PMCDETACH, sizeof(struct pmclog_pmcdetach));
1053 PMCLOG_EMIT32(pm->pm_id);
1055 PMCLOG_DESPATCH_SYNC(po);
1059 pmclog_process_proccreate(struct pmc_owner *po, struct proc *p, int sync)
1062 PMCLOG_RESERVE(po, PROC_CREATE, sizeof(struct pmclog_proccreate));
1063 PMCLOG_EMIT32(p->p_pid);
1064 PMCLOG_EMIT32(p->p_flag);
1065 PMCLOG_EMITSTRING(p->p_comm, MAXCOMLEN+1);
1066 PMCLOG_DESPATCH_SYNC(po);
1068 PMCLOG_RESERVE(po, PROC_CREATE, sizeof(struct pmclog_proccreate));
1069 PMCLOG_EMIT32(p->p_pid);
1070 PMCLOG_EMIT32(p->p_flag);
1071 PMCLOG_EMITSTRING(p->p_comm, MAXCOMLEN+1);
1072 PMCLOG_DESPATCH(po);
1077 * Log a context switch event to the log file.
1081 pmclog_process_proccsw(struct pmc *pm, struct pmc_process *pp, pmc_value_t v, struct thread *td)
1083 struct pmc_owner *po;
1085 KASSERT(pm->pm_flags & PMC_F_LOG_PROCCSW,
1086 ("[pmclog,%d] log-process-csw called gratuitously", __LINE__));
1088 PMCDBG3(LOG,SWO,1,"pm=%p pid=%d v=%jx", pm, pp->pp_proc->p_pid,
1093 PMCLOG_RESERVE_SAFE(po, PROCCSW, sizeof(struct pmclog_proccsw), pmc_rdtsc());
1095 PMCLOG_EMIT32(pm->pm_id);
1096 PMCLOG_EMIT32(pp->pp_proc->p_pid);
1097 PMCLOG_EMIT32(td->td_tid);
1099 PMCLOG_DESPATCH_SCHED_LOCK(po);
1103 pmclog_process_procexec(struct pmc_owner *po, pmc_id_t pmid, pid_t pid,
1104 uintfptr_t startaddr, char *path)
1106 int pathlen, recordlen;
1108 PMCDBG3(LOG,EXC,1,"po=%p pid=%d path=\"%s\"", po, pid, path);
1110 pathlen = strlen(path) + 1; /* #bytes for the path */
1111 recordlen = offsetof(struct pmclog_procexec, pl_pathname) + pathlen;
1112 PMCLOG_RESERVE(po, PROCEXEC, recordlen);
1114 PMCLOG_EMIT32(pmid);
1115 PMCLOG_EMITADDR(startaddr);
1116 PMCLOG_EMITSTRING(path,pathlen);
1117 PMCLOG_DESPATCH_SYNC(po);
1121 * Log a process exit event (and accumulated pmc value) to the log file.
1125 pmclog_process_procexit(struct pmc *pm, struct pmc_process *pp)
1128 struct pmc_owner *po;
1130 ri = PMC_TO_ROWINDEX(pm);
1131 PMCDBG3(LOG,EXT,1,"pm=%p pid=%d v=%jx", pm, pp->pp_proc->p_pid,
1132 pp->pp_pmcs[ri].pp_pmcval);
1136 PMCLOG_RESERVE(po, PROCEXIT, sizeof(struct pmclog_procexit));
1137 PMCLOG_EMIT32(pm->pm_id);
1138 PMCLOG_EMIT32(pp->pp_proc->p_pid);
1139 PMCLOG_EMIT64(pp->pp_pmcs[ri].pp_pmcval);
1140 PMCLOG_DESPATCH(po);
1148 pmclog_process_procfork(struct pmc_owner *po, pid_t oldpid, pid_t newpid)
1150 PMCLOG_RESERVE(po, PROCFORK, sizeof(struct pmclog_procfork));
1151 PMCLOG_EMIT32(oldpid);
1152 PMCLOG_EMIT32(newpid);
1153 PMCLOG_DESPATCH(po);
1157 * Log a process exit event of the form suitable for system-wide PMCs.
1161 pmclog_process_sysexit(struct pmc_owner *po, pid_t pid)
1163 PMCLOG_RESERVE(po, SYSEXIT, sizeof(struct pmclog_sysexit));
1165 PMCLOG_DESPATCH(po);
1169 pmclog_process_threadcreate(struct pmc_owner *po, struct thread *td, int sync)
1175 PMCLOG_RESERVE(po, THR_CREATE, sizeof(struct pmclog_threadcreate));
1176 PMCLOG_EMIT32(td->td_tid);
1177 PMCLOG_EMIT32(p->p_pid);
1178 PMCLOG_EMIT32(p->p_flag);
1180 PMCLOG_EMITSTRING(td->td_name, MAXCOMLEN+1);
1181 PMCLOG_DESPATCH_SYNC(po);
1183 PMCLOG_RESERVE(po, THR_CREATE, sizeof(struct pmclog_threadcreate));
1184 PMCLOG_EMIT32(td->td_tid);
1185 PMCLOG_EMIT32(p->p_pid);
1186 PMCLOG_EMIT32(p->p_flag);
1188 PMCLOG_EMITSTRING(td->td_name, MAXCOMLEN+1);
1189 PMCLOG_DESPATCH(po);
1194 pmclog_process_threadexit(struct pmc_owner *po, struct thread *td)
1197 PMCLOG_RESERVE(po, THR_EXIT, sizeof(struct pmclog_threadexit));
1198 PMCLOG_EMIT32(td->td_tid);
1199 PMCLOG_DESPATCH(po);
1203 * Write a user log entry.
1207 pmclog_process_userlog(struct pmc_owner *po, struct pmc_op_writelog *wl)
1211 PMCDBG2(LOG,WRI,1, "writelog po=%p ud=0x%x", po, wl->pm_userdata);
1215 PMCLOG_RESERVE_WITH_ERROR(po, USERDATA,
1216 sizeof(struct pmclog_userdata));
1217 PMCLOG_EMIT32(wl->pm_userdata);
1218 PMCLOG_DESPATCH(po);
1227 * Create a pool of log buffers and initialize mutexes.
1233 struct pmclog_buffer *plb;
1234 int domain, ncpus, total;
1236 if (pmclog_buffer_size <= 0 || pmclog_buffer_size > 16*1024) {
1237 (void) printf("hwpmc: tunable logbuffersize=%d must be "
1238 "greater than zero and less than or equal to 16MB.\n",
1239 pmclog_buffer_size);
1240 pmclog_buffer_size = PMC_LOG_BUFFER_SIZE;
1243 if (pmc_nlogbuffers_pcpu <= 0) {
1244 (void) printf("hwpmc: tunable nlogbuffers=%d must be greater "
1245 "than zero.\n", pmc_nlogbuffers_pcpu);
1246 pmc_nlogbuffers_pcpu = PMC_NLOGBUFFERS_PCPU;
1248 if (pmc_nlogbuffers_pcpu*pmclog_buffer_size > 32*1024) {
1249 (void) printf("hwpmc: memory allocated pcpu must be less than 32MB (is %dK).\n",
1250 pmc_nlogbuffers_pcpu*pmclog_buffer_size);
1251 pmc_nlogbuffers_pcpu = PMC_NLOGBUFFERS_PCPU;
1252 pmclog_buffer_size = PMC_LOG_BUFFER_SIZE;
1254 for (domain = 0; domain < vm_ndomains; domain++) {
1255 ncpus = pmc_dom_hdrs[domain]->pdbh_ncpus;
1256 total = ncpus * pmc_nlogbuffers_pcpu;
1258 plb = malloc_domainset(sizeof(struct pmclog_buffer) * total,
1259 M_PMC, DOMAINSET_PREF(domain), M_WAITOK | M_ZERO);
1260 pmc_dom_hdrs[domain]->pdbh_plbs = plb;
1261 for (; total > 0; total--, plb++) {
1264 buf = malloc_domainset(1024 * pmclog_buffer_size, M_PMC,
1265 DOMAINSET_PREF(domain), M_WAITOK | M_ZERO);
1266 PMCLOG_INIT_BUFFER_DESCRIPTOR(plb, buf, domain);
1267 pmc_plb_rele_unlocked(plb);
1270 mtx_init(&pmc_kthread_mtx, "pmc-kthread", "pmc-sleep", MTX_DEF);
1276 * Destroy mutexes and release memory back the to free pool.
1282 struct pmclog_buffer *plb;
1285 mtx_destroy(&pmc_kthread_mtx);
1287 for (domain = 0; domain < vm_ndomains; domain++) {
1288 while ((plb = TAILQ_FIRST(&pmc_dom_hdrs[domain]->pdbh_head)) != NULL) {
1289 TAILQ_REMOVE(&pmc_dom_hdrs[domain]->pdbh_head, plb, plb_next);
1290 free(plb->plb_base, M_PMC);
1292 free(pmc_dom_hdrs[domain]->pdbh_plbs, M_PMC);