sys/dev/hwpmc/hwpmc_logging.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
   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/cdefs.h>
  40 __FBSDID("$FreeBSD$");
  41
  42 #include <sys/param.h>
  43 #include <sys/capsicum.h>
  44 #include <sys/domainset.h>
  45 #include <sys/file.h>
  46 #include <sys/kernel.h>
  47 #include <sys/kthread.h>
  48 #include <sys/lock.h>
  49 #include <sys/module.h>
  50 #include <sys/mutex.h>
  51 #include <sys/pmc.h>
  52 #include <sys/pmckern.h>
  53 #include <sys/pmclog.h>
  54 #include <sys/proc.h>
  55 #include <sys/sched.h>
  56 #include <sys/signalvar.h>
  57 #include <sys/smp.h>
  58 #include <sys/syscallsubr.h>
  59 #include <sys/sysctl.h>
  60 #include <sys/systm.h>
  61 #include <sys/uio.h>
  62 #include <sys/unistd.h>
  63 #include <sys/vnode.h>
  64
  65 #if defined(__i386__) || defined(__amd64__)
  66 #include <machine/clock.h>
  67 #endif
  68
  69 #define curdomain PCPU_GET(domain)
  70
  71 /*
  72  * Sysctl tunables
  73  */
  74
  75 SYSCTL_DECL(_kern_hwpmc);
  76
  77 /*
  78  * kern.hwpmc.logbuffersize -- size of the per-cpu owner buffers.
  79  */
  80
  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);
  84 #endif
  85 SYSCTL_INT(_kern_hwpmc, OID_AUTO, logbuffersize, CTLFLAG_RDTUN,
  86     &pmclog_buffer_size, 0, "size of log buffers in kilobytes");
  87
  88 /*
  89  * kern.hwpmc.nbuffers_pcpu -- number of global log buffers
  90  */
  91
  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);
  95 #endif
  96 SYSCTL_INT(_kern_hwpmc, OID_AUTO, nbuffers_pcpu, CTLFLAG_RDTUN,
  97     &pmc_nlogbuffers_pcpu, 0, "number of log buffers per cpu");
  98
  99 /*
 100  * Global log buffer list and associated spin lock.
 101  */
 102
 103 static struct mtx pmc_kthread_mtx;      /* sleep lock */
 104
 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; \
 109         } while (0)
 110
 111 #define PMCLOG_RESET_BUFFER_DESCRIPTOR(D) do {                  \
 112                 (D)->plb_ptr  = (D)->plb_base; \
 113         } while (0)
 114
 115 /*
 116  * Log file record constructors.
 117  */
 118 #define _PMCLOG_TO_HEADER(T,L)                                          \
 119         ((PMCLOG_HEADER_MAGIC << 24) |                                  \
 120          (PMCLOG_TYPE_ ## T << 16)   |                                  \
 121          ((L) & 0xFFFF))
 122
 123 /* reserve LEN bytes of space and initialize the entry header */
 124 #define _PMCLOG_RESERVE_SAFE(PO,TYPE,LEN,ACTION, TSC) do {      \
 125                 uint32_t *_le;                                          \
 126                 int _len = roundup((LEN), sizeof(uint32_t));    \
 127                 struct pmclog_header *ph;                                                       \
 128                 if ((_le = pmclog_reserve((PO), _len)) == NULL) {       \
 129                         ACTION;                                                                                 \
 130                 }                                                                                                       \
 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 */
 135
 136 /* reserve LEN bytes of space and initialize the entry header */
 137 #define _PMCLOG_RESERVE(PO,TYPE,LEN,ACTION) do {                        \
 138                 uint32_t *_le;                                          \
 139                 int _len = roundup((LEN), sizeof(uint32_t));    \
 140                 uint64_t tsc;                                                                           \
 141                 struct pmclog_header *ph;                                                       \
 142                 tsc = pmc_rdtsc();                                                                      \
 143                 spinlock_enter();                                                                       \
 144                 if ((_le = pmclog_reserve((PO), _len)) == NULL) {       \
 145                         spinlock_exit();                                                                \
 146                         ACTION;                                                                                 \
 147                 }                                                                                               \
 148                 ph = (struct pmclog_header *)_le;                                       \
 149                 ph->pl_header =_PMCLOG_TO_HEADER(TYPE,_len);    \
 150                 ph->pl_tsc = tsc;                                                                       \
 151                 _le += sizeof(*ph)/4    /* skip over timestamp */
 152
 153
 154
 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)
 159
 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);         \
 164         } while (0)
 165
 166
 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)
 170
 171 #define PMCLOG_DESPATCH_SAFE(PO)                                                \
 172             pmclog_release((PO));                                               \
 173         } while (0)
 174
 175 #define PMCLOG_DESPATCH_SCHED_LOCK(PO)                                          \
 176              pmclog_release_flags((PO), 0);                                                     \
 177         } while (0)
 178
 179 #define PMCLOG_DESPATCH(PO)                                                     \
 180             pmclog_release((PO));                                               \
 181                 spinlock_exit();                                                        \
 182         } while (0)
 183
 184 #define PMCLOG_DESPATCH_SYNC(PO)                                                \
 185             pmclog_schedule_io((PO), 1);                                                \
 186                 spinlock_exit();                                                                \
 187                 } while (0)
 188
 189
 190 #define TSDELTA 4
 191 /*
 192  * Assertions about the log file format.
 193  */
 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 +
 211     sizeof(uintfptr_t));
 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);
 216
 217 /*
 218  * Log buffer structure
 219  */
 220
 221 struct pmclog_buffer {
 222         TAILQ_ENTRY(pmclog_buffer) plb_next;
 223         char            *plb_base;
 224         char            *plb_ptr;
 225         char            *plb_fence;
 226         uint16_t         plb_domain;
 227 } __aligned(CACHE_LINE_SIZE);
 228
 229 /*
 230  * Prototypes
 231  */
 232
 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);
 240
 241 /*
 242  * Helper functions
 243  */
 244
 245 static inline void
 246 pmc_plb_rele_unlocked(struct pmclog_buffer *plb)
 247 {
 248         TAILQ_INSERT_HEAD(&pmc_dom_hdrs[plb->plb_domain]->pdbh_head, plb, plb_next);
 249 }
 250
 251 static inline void
 252 pmc_plb_rele(struct pmclog_buffer *plb)
 253 {
 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);
 257 }
 258
 259 /*
 260  * Get a log buffer
 261  */
 262 static int
 263 pmclog_get_buffer(struct pmc_owner *po)
 264 {
 265         struct pmclog_buffer *plb;
 266         int domain;
 267
 268         KASSERT(po->po_curbuf[curcpu] == NULL,
 269             ("[pmclog,%d] po=%p current buffer still valid", __LINE__, po));
 270
 271         domain = curdomain;
 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);
 277
 278         PMCDBG2(LOG,GTB,1, "po=%p plb=%p", po, plb);
 279
 280 #ifdef  HWPMC_DEBUG
 281         if (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));
 287 #endif
 288
 289         po->po_curbuf[curcpu] = plb;
 290
 291         /* update stats */
 292         counter_u64_add(pmc_stats.pm_buffer_requests, 1);
 293         if (plb == NULL)
 294                 counter_u64_add(pmc_stats.pm_buffer_requests_failed, 1);
 295
 296         return (plb ? 0 : ENOMEM);
 297 }
 298
 299 struct pmclog_proc_init_args {
 300         struct proc *kthr;
 301         struct pmc_owner *po;
 302         bool exit;
 303         bool acted;
 304 };
 305
 306 int
 307 pmclog_proc_create(struct thread *td, void **handlep)
 308 {
 309         struct pmclog_proc_init_args *ia;
 310         int error;
 311
 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);
 315         if (error == 0)
 316                 *handlep = ia;
 317         return (error);
 318 }
 319
 320 void
 321 pmclog_proc_ignite(void *handle, struct pmc_owner *po)
 322 {
 323         struct pmclog_proc_init_args *ia;
 324
 325         ia = handle;
 326         mtx_lock(&pmc_kthread_mtx);
 327         MPASS(!ia->acted);
 328         MPASS(ia->po == NULL);
 329         MPASS(!ia->exit);
 330         MPASS(ia->kthr != NULL);
 331         if (po == NULL) {
 332                 ia->exit = true;
 333         } else {
 334                 ia->po = po;
 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;
 339         }
 340         wakeup(ia);
 341         while (!ia->acted)
 342                 msleep(ia, &pmc_kthread_mtx, PWAIT, "pmclogw", 0);
 343         mtx_unlock(&pmc_kthread_mtx);
 344         free(ia, M_TEMP);
 345 }
 346
 347 /*
 348  * Log handler loop.
 349  *
 350  * This function is executed by each pmc owner's helper thread.
 351  */
 352 static void
 353 pmclog_loop(void *arg)
 354 {
 355         struct pmclog_proc_init_args *ia;
 356         struct pmc_owner *po;
 357         struct pmclog_buffer *lb;
 358         struct proc *p;
 359         struct ucred *ownercred;
 360         struct ucred *mycred;
 361         struct thread *td;
 362         sigset_t unb;
 363         struct uio auio;
 364         struct iovec aiov;
 365         size_t nbytes;
 366         int error;
 367
 368         td = curthread;
 369
 370         SIGEMPTYSET(unb);
 371         SIGADDSET(unb, SIGHUP);
 372         (void)kern_sigprocmask(td, SIG_UNBLOCK, &unb, NULL, 0);
 373
 374         ia = arg;
 375         MPASS(ia->kthr == curproc);
 376         MPASS(!ia->acted);
 377         mtx_lock(&pmc_kthread_mtx);
 378         while (ia->po == NULL && !ia->exit)
 379                 msleep(ia, &pmc_kthread_mtx, PWAIT, "pmclogi", 0);
 380         if (ia->exit) {
 381                 ia->acted = true;
 382                 wakeup(ia);
 383                 mtx_unlock(&pmc_kthread_mtx);
 384                 kproc_exit(0);
 385         }
 386         MPASS(ia->po != NULL);
 387         po = ia->po;
 388         ia->acted = true;
 389         wakeup(ia);
 390         mtx_unlock(&pmc_kthread_mtx);
 391         ia = NULL;
 392
 393         p = po->po_owner;
 394         mycred = td->td_ucred;
 395
 396         PROC_LOCK(p);
 397         ownercred = crhold(p->p_ucred);
 398         PROC_UNLOCK(p);
 399
 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));
 404
 405         lb = NULL;
 406
 407
 408         /*
 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
 411          * is deconfigured.
 412          */
 413
 414         mtx_lock(&pmc_kthread_mtx);
 415
 416         for (;;) {
 417
 418                 /* check if we've been asked to exit */
 419                 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
 420                         break;
 421
 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);
 426
 427                                 /* No more buffers and shutdown required. */
 428                                 if (po->po_flags & PMC_PO_SHUTDOWN)
 429                                         break;
 430
 431                                 (void) msleep(po, &pmc_kthread_mtx, PWAIT,
 432                                     "pmcloop", 250);
 433                                 continue;
 434                         }
 435
 436                         TAILQ_REMOVE(&po->po_logbuffers, lb, plb_next);
 437                         mtx_unlock_spin(&po->po_mtx);
 438                 }
 439
 440                 mtx_unlock(&pmc_kthread_mtx);
 441
 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 */
 446
 447                 aiov.iov_base = lb->plb_base;
 448                 aiov.iov_len  = nbytes = lb->plb_ptr - lb->plb_base;
 449
 450                 auio.uio_iov    = &aiov;
 451                 auio.uio_iovcnt = 1;
 452                 auio.uio_offset = -1;
 453                 auio.uio_resid  = nbytes;
 454                 auio.uio_rw     = UIO_WRITE;
 455                 auio.uio_segflg = UIO_SYSSPACE;
 456                 auio.uio_td     = td;
 457
 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;
 462
 463                 if (error) {
 464                         /* XXX some errors are recoverable */
 465                         /* send a SIGIO to the owner and exit */
 466                         PROC_LOCK(p);
 467                         kern_psignal(p, SIGIO);
 468                         PROC_UNLOCK(p);
 469
 470                         mtx_lock(&pmc_kthread_mtx);
 471
 472                         po->po_error = error; /* save for flush log */
 473
 474                         PMCDBG2(LOG,WRI,2, "po=%p error=%d", po, error);
 475
 476                         break;
 477                 }
 478
 479                 mtx_lock(&pmc_kthread_mtx);
 480
 481                 /* put the used buffer back into the global pool */
 482                 PMCLOG_RESET_BUFFER_DESCRIPTOR(lb);
 483
 484                 pmc_plb_rele(lb);
 485                 lb = NULL;
 486         }
 487
 488         wakeup_one(po->po_kthread);
 489         po->po_kthread = NULL;
 490
 491         mtx_unlock(&pmc_kthread_mtx);
 492
 493         /* return the current I/O buffer to the global pool */
 494         if (lb) {
 495                 PMCLOG_RESET_BUFFER_DESCRIPTOR(lb);
 496
 497                 pmc_plb_rele(lb);
 498         }
 499
 500         /*
 501          * Exit this thread, signalling the waiter
 502          */
 503
 504         crfree(ownercred);
 505
 506         kproc_exit(0);
 507 }
 508
 509 /*
 510  * Release and log entry and schedule an I/O if needed.
 511  */
 512
 513 static void
 514 pmclog_release_flags(struct pmc_owner *po, int wakeup)
 515 {
 516         struct pmclog_buffer *plb;
 517
 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));
 525
 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);
 529
 530         PMCDBG1(LOG,REL,1, "po=%p", po);
 531 }
 532
 533 static void
 534 pmclog_release(struct pmc_owner *po)
 535 {
 536
 537         pmclog_release_flags(po, 1);
 538 }
 539
 540
 541 /*
 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
 547  * when done.
 548  */
 549
 550 static uint32_t *
 551 pmclog_reserve(struct pmc_owner *po, int length)
 552 {
 553         uintptr_t newptr, oldptr __diagused;
 554         struct pmclog_buffer *plb, **pplb;
 555
 556         PMCDBG2(LOG,ALL,1, "po=%p len=%d", po, length);
 557
 558         KASSERT(length % sizeof(uint32_t) == 0,
 559             ("[pmclog,%d] length not a multiple of word size", __LINE__));
 560
 561         /* No more data when shutdown in progress. */
 562         if (po->po_flags & PMC_PO_SHUTDOWN)
 563                 return (NULL);
 564
 565         pplb = &po->po_curbuf[curcpu];
 566         if (*pplb == NULL && pmclog_get_buffer(po) != 0)
 567                 goto fail;
 568
 569         KASSERT(*pplb != NULL,
 570             ("[pmclog,%d] po=%p no current buffer", __LINE__, po));
 571
 572         plb = *pplb;
 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,
 577                 plb->plb_fence));
 578
 579         oldptr = (uintptr_t) plb->plb_ptr;
 580         newptr = oldptr + length;
 581
 582         KASSERT(oldptr != (uintptr_t) NULL,
 583             ("[pmclog,%d] po=%p Null log buffer pointer", __LINE__, po));
 584
 585         /*
 586          * If we have space in the current buffer, return a pointer to
 587          * available space with the PO structure locked.
 588          */
 589         if (newptr <= (uintptr_t) plb->plb_fence) {
 590                 plb->plb_ptr = (char *) newptr;
 591                 goto done;
 592         }
 593
 594         /*
 595          * Otherwise, schedule the current buffer for output and get a
 596          * fresh buffer.
 597          */
 598         pmclog_schedule_io(po, 0);
 599
 600         if (pmclog_get_buffer(po) != 0)
 601                 goto fail;
 602
 603         plb = *pplb;
 604         KASSERT(plb != NULL,
 605             ("[pmclog,%d] po=%p no current buffer", __LINE__, po));
 606
 607         KASSERT(plb->plb_ptr != NULL,
 608             ("[pmclog,%d] null return from pmc_get_log_buffer", __LINE__));
 609
 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,
 614                 plb->plb_fence));
 615
 616         oldptr = (uintptr_t) plb->plb_ptr;
 617
 618  done:
 619         return ((uint32_t *) oldptr);
 620  fail:
 621         return (NULL);
 622 }
 623
 624 /*
 625  * Schedule an I/O.
 626  *
 627  * Transfer the current buffer to the helper kthread.
 628  */
 629
 630 static void
 631 pmclog_schedule_io(struct pmc_owner *po, int wakeup)
 632 {
 633         struct pmclog_buffer *plb;
 634
 635         plb = po->po_curbuf[curcpu];
 636         po->po_curbuf[curcpu] = NULL;
 637         KASSERT(plb != 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));
 645
 646         PMCDBG1(LOG,SIO, 1, "po=%p", po);
 647
 648         /*
 649          * Add the current buffer to the tail of the buffer list and
 650          * wakeup the helper.
 651          */
 652         mtx_lock_spin(&po->po_mtx);
 653         TAILQ_INSERT_TAIL(&po->po_logbuffers, plb, plb_next);
 654         mtx_unlock_spin(&po->po_mtx);
 655         if (wakeup)
 656                 wakeup_one(po);
 657 }
 658
 659 /*
 660  * Stop the helper kthread.
 661  */
 662
 663 static void
 664 pmclog_stop_kthread(struct pmc_owner *po)
 665 {
 666
 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);
 673         }
 674         wakeup_one(po);
 675         while (po->po_kthread)
 676                 msleep(po->po_kthread, &pmc_kthread_mtx, PPAUSE, "pmckstp", 0);
 677         mtx_unlock(&pmc_kthread_mtx);
 678 }
 679
 680 /*
 681  * Public functions
 682  */
 683
 684 /*
 685  * Configure a log file for pmc owner 'po'.
 686  *
 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.
 689  */
 690
 691 int
 692 pmclog_configure_log(struct pmc_mdep *md, struct pmc_owner *po, int logfd)
 693 {
 694         struct proc *p;
 695         struct timespec ts;
 696         int error;
 697
 698         sx_assert(&pmc_sx, SA_XLOCKED);
 699         PMCDBG2(LOG,CFG,1, "config po=%p logfd=%d", po, logfd);
 700
 701         p = po->po_owner;
 702
 703         /* return EBUSY if a log file was already present */
 704         if (po->po_flags & PMC_PO_OWNS_LOGFILE)
 705                 return (EBUSY);
 706
 707         KASSERT(po->po_file == NULL,
 708             ("[pmclog,%d] po=%p file (%p) already present", __LINE__, po,
 709                 po->po_file));
 710
 711         /* get a reference to the file state */
 712         error = fget_write(curthread, logfd, &cap_write_rights, &po->po_file);
 713         if (error)
 714                 goto error;
 715
 716         /* mark process as owning a log file */
 717         po->po_flags |= PMC_PO_OWNS_LOGFILE;
 718
 719         /* mark process as using HWPMCs */
 720         PROC_LOCK(p);
 721         p->p_flag |= P_HWPMC;
 722         PROC_UNLOCK(p);
 723         nanotime(&ts);
 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);
 731 #else
 732         /* other architectures will need to fill this in */
 733         PMCLOG_EMIT32(0);
 734         PMCLOG_EMIT32(0);
 735 #endif
 736         memcpy(_le, &ts, sizeof(ts));
 737         _le += sizeof(ts)/4;
 738         PMCLOG_EMITSTRING(pmc_cpuid, PMC_CPUID_LEN);
 739         PMCLOG_DESPATCH_SYNC(po);
 740
 741         return (0);
 742
 743  error:
 744         KASSERT(po->po_kthread == NULL, ("[pmclog,%d] po=%p kthread not "
 745             "stopped", __LINE__, po));
 746
 747         if (po->po_file)
 748                 (void) fdrop(po->po_file, curthread);
 749         po->po_file  = NULL;    /* clear file and error state */
 750         po->po_error = 0;
 751         po->po_flags &= ~PMC_PO_OWNS_LOGFILE;
 752
 753         return (error);
 754 }
 755
 756
 757 /*
 758  * De-configure a log file.  This will throw away any buffers queued
 759  * for this owner process.
 760  */
 761
 762 int
 763 pmclog_deconfigure_log(struct pmc_owner *po)
 764 {
 765         int error;
 766         struct pmclog_buffer *lb;
 767
 768         PMCDBG1(LOG,CFG,1, "de-config po=%p", po);
 769
 770         if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
 771                 return (EINVAL);
 772
 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));
 777
 778         /* stop the kthread, this will reset the 'OWNS_LOGFILE' flag */
 779         pmclog_stop_kthread(po);
 780
 781         KASSERT(po->po_kthread == NULL,
 782             ("[pmclog,%d] po=%p kthread not stopped", __LINE__, po));
 783
 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);
 788                 pmc_plb_rele(lb);
 789         }
 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);
 797                         pmc_plb_rele(lb);
 798                 }
 799         }
 800         thread_lock(curthread);
 801         sched_unbind(curthread);
 802         thread_unlock(curthread);
 803
 804         /* drop a reference to the fd */
 805         if (po->po_file != NULL) {
 806                 error = fdrop(po->po_file, curthread);
 807                 po->po_file = NULL;
 808         } else
 809                 error = 0;
 810         po->po_error = 0;
 811
 812         return (error);
 813 }
 814
 815 /*
 816  * Flush a process' log buffer.
 817  */
 818
 819 int
 820 pmclog_flush(struct pmc_owner *po, int force)
 821 {
 822         int error;
 823
 824         PMCDBG1(LOG,FLS,1, "po=%p", po);
 825
 826         /*
 827          * If there is a pending error recorded by the logger thread,
 828          * return that.
 829          */
 830         if (po->po_error)
 831                 return (po->po_error);
 832
 833         error = 0;
 834
 835         /*
 836          * Check that we do have an active log file.
 837          */
 838         mtx_lock(&pmc_kthread_mtx);
 839         if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) {
 840                 error = EINVAL;
 841                 goto error;
 842         }
 843
 844         pmclog_schedule_all(po);
 845  error:
 846         mtx_unlock(&pmc_kthread_mtx);
 847
 848         return (error);
 849 }
 850
 851 static void
 852 pmclog_schedule_one_cond(struct pmc_owner *po)
 853 {
 854         struct pmclog_buffer *plb;
 855         int cpu;
 856
 857         spinlock_enter();
 858         cpu = curcpu;
 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);
 862
 863         plb = po->po_curbuf[cpu];
 864         if (plb && plb->plb_ptr != plb->plb_base)
 865                 pmclog_schedule_io(po, 1);
 866         spinlock_exit();
 867 }
 868
 869 static void
 870 pmclog_schedule_all(struct pmc_owner *po)
 871 {
 872         /*
 873          * Schedule the current buffer if any and not empty.
 874          */
 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);
 880         }
 881         thread_lock(curthread);
 882         sched_unbind(curthread);
 883         thread_unlock(curthread);
 884 }
 885
 886 int
 887 pmclog_close(struct pmc_owner *po)
 888 {
 889
 890         PMCDBG1(LOG,CLO,1, "po=%p", po);
 891
 892         pmclog_process_closelog(po);
 893
 894         mtx_lock(&pmc_kthread_mtx);
 895         /*
 896          * Initiate shutdown: no new data queued,
 897          * thread will close file on last block.
 898          */
 899         po->po_flags |= PMC_PO_SHUTDOWN;
 900         /* give time for all to see */
 901         DELAY(50);
 902
 903         /*
 904          * Schedule the current buffer.
 905          */
 906         pmclog_schedule_all(po);
 907         wakeup_one(po);
 908
 909         mtx_unlock(&pmc_kthread_mtx);
 910
 911         return (0);
 912 }
 913
 914 void
 915 pmclog_process_callchain(struct pmc *pm, struct pmc_sample *ps)
 916 {
 917         int n, recordlen;
 918         uint32_t flags;
 919         struct pmc_owner *po;
 920
 921         PMCDBG3(LOG,SAM,1,"pm=%p pid=%d n=%d", pm, ps->ps_pid,
 922             ps->ps_nsamples);
 923
 924         recordlen = offsetof(struct pmclog_callchain, pl_pc) +
 925             ps->ps_nsamples * sizeof(uintfptr_t);
 926         po = pm->pm_owner;
 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);
 936 }
 937
 938 void
 939 pmclog_process_closelog(struct pmc_owner *po)
 940 {
 941         PMCLOG_RESERVE(po,CLOSELOG,sizeof(struct pmclog_closelog));
 942         PMCLOG_DESPATCH_SYNC(po);
 943 }
 944
 945 void
 946 pmclog_process_dropnotify(struct pmc_owner *po)
 947 {
 948         PMCLOG_RESERVE(po,DROPNOTIFY,sizeof(struct pmclog_dropnotify));
 949         PMCLOG_DESPATCH(po);
 950 }
 951
 952 void
 953 pmclog_process_map_in(struct pmc_owner *po, pid_t pid, uintfptr_t start,
 954     const char *path)
 955 {
 956         int pathlen, recordlen;
 957
 958         KASSERT(path != NULL, ("[pmclog,%d] map-in, null path", __LINE__));
 959
 960         pathlen = strlen(path) + 1;     /* #bytes for path name */
 961         recordlen = offsetof(struct pmclog_map_in, pl_pathname) +
 962             pathlen;
 963
 964         PMCLOG_RESERVE(po, MAP_IN, recordlen);
 965         PMCLOG_EMIT32(pid);
 966         PMCLOG_EMIT32(0);
 967         PMCLOG_EMITADDR(start);
 968         PMCLOG_EMITSTRING(path,pathlen);
 969         PMCLOG_DESPATCH_SYNC(po);
 970 }
 971
 972 void
 973 pmclog_process_map_out(struct pmc_owner *po, pid_t pid, uintfptr_t start,
 974     uintfptr_t end)
 975 {
 976         KASSERT(start <= end, ("[pmclog,%d] start > end", __LINE__));
 977
 978         PMCLOG_RESERVE(po, MAP_OUT, sizeof(struct pmclog_map_out));
 979         PMCLOG_EMIT32(pid);
 980         PMCLOG_EMIT32(0);
 981         PMCLOG_EMITADDR(start);
 982         PMCLOG_EMITADDR(end);
 983         PMCLOG_DESPATCH(po);
 984 }
 985
 986 void
 987 pmclog_process_pmcallocate(struct pmc *pm)
 988 {
 989         struct pmc_owner *po;
 990         struct pmc_soft *ps;
 991
 992         po = pm->pm_owner;
 993
 994         PMCDBG1(LOG,ALL,1, "pm=%p", pm);
 995
 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);
1002                 PMCLOG_EMIT32(0);
1003                 PMCLOG_EMIT64(pm->pm_sc.pm_reloadcount);
1004                 ps = pmc_soft_ev_acquire(pm->pm_event);
1005                 if (ps != NULL)
1006                         PMCLOG_EMITSTRING(ps->ps_ev.pm_ev_name,PMC_NAME_MAX);
1007                 else
1008                         PMCLOG_EMITNULLSTRING(PMC_NAME_MAX);
1009                 pmc_soft_ev_release(ps);
1010                 PMCLOG_DESPATCH_SYNC(po);
1011         } else {
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);
1017                 PMCLOG_EMIT32(0);
1018                 PMCLOG_EMIT64(pm->pm_sc.pm_reloadcount);
1019                 PMCLOG_DESPATCH_SYNC(po);
1020         }
1021 }
1022
1023 void
1024 pmclog_process_pmcattach(struct pmc *pm, pid_t pid, char *path)
1025 {
1026         int pathlen, recordlen;
1027         struct pmc_owner *po;
1028
1029         PMCDBG2(LOG,ATT,1,"pm=%p pid=%d", pm, pid);
1030
1031         po = pm->pm_owner;
1032
1033         pathlen = strlen(path) + 1;     /* #bytes for the string */
1034         recordlen = offsetof(struct pmclog_pmcattach, pl_pathname) + pathlen;
1035
1036         PMCLOG_RESERVE(po, PMCATTACH, recordlen);
1037         PMCLOG_EMIT32(pm->pm_id);
1038         PMCLOG_EMIT32(pid);
1039         PMCLOG_EMITSTRING(path, pathlen);
1040         PMCLOG_DESPATCH_SYNC(po);
1041 }
1042
1043 void
1044 pmclog_process_pmcdetach(struct pmc *pm, pid_t pid)
1045 {
1046         struct pmc_owner *po;
1047
1048         PMCDBG2(LOG,ATT,1,"!pm=%p pid=%d", pm, pid);
1049
1050         po = pm->pm_owner;
1051
1052         PMCLOG_RESERVE(po, PMCDETACH, sizeof(struct pmclog_pmcdetach));
1053         PMCLOG_EMIT32(pm->pm_id);
1054         PMCLOG_EMIT32(pid);
1055         PMCLOG_DESPATCH_SYNC(po);
1056 }
1057
1058 void
1059 pmclog_process_proccreate(struct pmc_owner *po, struct proc *p, int sync)
1060 {
1061         if (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);
1067         } else {
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);
1073         }
1074 }
1075
1076 /*
1077  * Log a context switch event to the log file.
1078  */
1079
1080 void
1081 pmclog_process_proccsw(struct pmc *pm, struct pmc_process *pp, pmc_value_t v, struct thread *td)
1082 {
1083         struct pmc_owner *po;
1084
1085         KASSERT(pm->pm_flags & PMC_F_LOG_PROCCSW,
1086             ("[pmclog,%d] log-process-csw called gratuitously", __LINE__));
1087
1088         PMCDBG3(LOG,SWO,1,"pm=%p pid=%d v=%jx", pm, pp->pp_proc->p_pid,
1089             v);
1090
1091         po = pm->pm_owner;
1092
1093         PMCLOG_RESERVE_SAFE(po, PROCCSW, sizeof(struct pmclog_proccsw), pmc_rdtsc());
1094         PMCLOG_EMIT64(v);
1095         PMCLOG_EMIT32(pm->pm_id);
1096         PMCLOG_EMIT32(pp->pp_proc->p_pid);
1097         PMCLOG_EMIT32(td->td_tid);
1098         PMCLOG_EMIT32(0);
1099         PMCLOG_DESPATCH_SCHED_LOCK(po);
1100 }
1101
1102 void
1103 pmclog_process_procexec(struct pmc_owner *po, pmc_id_t pmid, pid_t pid,
1104     uintfptr_t startaddr, char *path)
1105 {
1106         int pathlen, recordlen;
1107
1108         PMCDBG3(LOG,EXC,1,"po=%p pid=%d path=\"%s\"", po, pid, path);
1109
1110         pathlen   = strlen(path) + 1;   /* #bytes for the path */
1111         recordlen = offsetof(struct pmclog_procexec, pl_pathname) + pathlen;
1112         PMCLOG_RESERVE(po, PROCEXEC, recordlen);
1113         PMCLOG_EMIT32(pid);
1114         PMCLOG_EMIT32(pmid);
1115         PMCLOG_EMITADDR(startaddr);
1116         PMCLOG_EMITSTRING(path,pathlen);
1117         PMCLOG_DESPATCH_SYNC(po);
1118 }
1119
1120 /*
1121  * Log a process exit event (and accumulated pmc value) to the log file.
1122  */
1123
1124 void
1125 pmclog_process_procexit(struct pmc *pm, struct pmc_process *pp)
1126 {
1127         int ri;
1128         struct pmc_owner *po;
1129
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);
1133
1134         po = pm->pm_owner;
1135
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);
1141 }
1142
1143 /*
1144  * Log a fork event.
1145  */
1146
1147 void
1148 pmclog_process_procfork(struct pmc_owner *po, pid_t oldpid, pid_t newpid)
1149 {
1150         PMCLOG_RESERVE(po, PROCFORK, sizeof(struct pmclog_procfork));
1151         PMCLOG_EMIT32(oldpid);
1152         PMCLOG_EMIT32(newpid);
1153         PMCLOG_DESPATCH(po);
1154 }
1155
1156 /*
1157  * Log a process exit event of the form suitable for system-wide PMCs.
1158  */
1159
1160 void
1161 pmclog_process_sysexit(struct pmc_owner *po, pid_t pid)
1162 {
1163         PMCLOG_RESERVE(po, SYSEXIT, sizeof(struct pmclog_sysexit));
1164         PMCLOG_EMIT32(pid);
1165         PMCLOG_DESPATCH(po);
1166 }
1167
1168 void
1169 pmclog_process_threadcreate(struct pmc_owner *po, struct thread *td, int sync)
1170 {
1171         struct proc *p;
1172
1173         p = td->td_proc;
1174         if (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);
1179                 PMCLOG_EMIT32(0);
1180                 PMCLOG_EMITSTRING(td->td_name, MAXCOMLEN+1);
1181                 PMCLOG_DESPATCH_SYNC(po);
1182         } else {
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);
1187                 PMCLOG_EMIT32(0);
1188                 PMCLOG_EMITSTRING(td->td_name, MAXCOMLEN+1);
1189                 PMCLOG_DESPATCH(po);
1190         }
1191 }
1192
1193 void
1194 pmclog_process_threadexit(struct pmc_owner *po, struct thread *td)
1195 {
1196
1197         PMCLOG_RESERVE(po, THR_EXIT, sizeof(struct pmclog_threadexit));
1198         PMCLOG_EMIT32(td->td_tid);
1199         PMCLOG_DESPATCH(po);
1200 }
1201
1202 /*
1203  * Write a user log entry.
1204  */
1205
1206 int
1207 pmclog_process_userlog(struct pmc_owner *po, struct pmc_op_writelog *wl)
1208 {
1209         int error;
1210
1211         PMCDBG2(LOG,WRI,1, "writelog po=%p ud=0x%x", po, wl->pm_userdata);
1212
1213         error = 0;
1214
1215         PMCLOG_RESERVE_WITH_ERROR(po, USERDATA,
1216             sizeof(struct pmclog_userdata));
1217         PMCLOG_EMIT32(wl->pm_userdata);
1218         PMCLOG_DESPATCH(po);
1219
1220  error:
1221         return (error);
1222 }
1223
1224 /*
1225  * Initialization.
1226  *
1227  * Create a pool of log buffers and initialize mutexes.
1228  */
1229
1230 void
1231 pmclog_initialize()
1232 {
1233         struct pmclog_buffer *plb;
1234         int domain, ncpus, total;
1235
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;
1241         }
1242
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;
1247         }
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;
1253         }
1254         for (domain = 0; domain < vm_ndomains; domain++) {
1255                 ncpus = pmc_dom_hdrs[domain]->pdbh_ncpus;
1256                 total = ncpus * pmc_nlogbuffers_pcpu;
1257
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++) {
1262                         void *buf;
1263
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);
1268                 }
1269         }
1270         mtx_init(&pmc_kthread_mtx, "pmc-kthread", "pmc-sleep", MTX_DEF);
1271 }
1272
1273 /*
1274  * Shutdown logging.
1275  *
1276  * Destroy mutexes and release memory back the to free pool.
1277  */
1278
1279 void
1280 pmclog_shutdown()
1281 {
1282         struct pmclog_buffer *plb;
1283         int domain;
1284
1285         mtx_destroy(&pmc_kthread_mtx);
1286
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);
1291                 }
1292                 free(pmc_dom_hdrs[domain]->pdbh_plbs, M_PMC);
1293         }
1294 }