1 /*- 2 * Copyright (c) 2005-2007 Joseph Koshy 3 * Copyright (c) 2007 The FreeBSD Foundation 4 * All rights reserved. 5 * 6 * Portions of this software were developed by A. Joseph Koshy under 7 * sponsorship from the FreeBSD Foundation and Google, Inc. 8 * 9 * Redistribution and use in source and binary forms, with or without 10 * modification, are permitted provided that the following conditions 11 * are met: 12 * 1. Redistributions of source code must retain the above copyright 13 * notice, this list of conditions and the following disclaimer. 14 * 2. Redistributions in binary form must reproduce the above copyright 15 * notice, this list of conditions and the following disclaimer in the 16 * documentation and/or other materials provided with the distribution. 17 * 18 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 19 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 20 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 21 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 22 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 23 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 24 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 25 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 26 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 27 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 28 * SUCH DAMAGE. 29 * 30 */ 31 32 /* 33 * Logging code for hwpmc(4) 34 */ 35 36 #include <sys/cdefs.h> 37 __FBSDID("$FreeBSD$"); 38 39 #include <sys/param.h> 40 #include <sys/capsicum.h> 41 #include <sys/file.h> 42 #include <sys/kernel.h> 43 #include <sys/kthread.h> 44 #include <sys/lock.h> 45 #include <sys/module.h> 46 #include <sys/mutex.h> 47 #include <sys/pmc.h> 48 #include <sys/pmckern.h> 49 #include <sys/pmclog.h> 50 #include <sys/proc.h> 51 #include <sys/signalvar.h> 52 #include <sys/sysctl.h> 53 #include <sys/systm.h> 54 #include <sys/uio.h> 55 #include <sys/unistd.h> 56 #include <sys/vnode.h> 57 58 /* 59 * Sysctl tunables 60 */ 61 62 SYSCTL_DECL(_kern_hwpmc); 63 64 /* 65 * kern.hwpmc.logbuffersize -- size of the per-cpu owner buffers. 66 */ 67 68 static int pmclog_buffer_size = PMC_LOG_BUFFER_SIZE; 69 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "logbuffersize", &pmclog_buffer_size); 70 SYSCTL_INT(_kern_hwpmc, OID_AUTO, logbuffersize, CTLFLAG_TUN|CTLFLAG_RD, 71 &pmclog_buffer_size, 0, "size of log buffers in kilobytes"); 72 73 /* 74 * kern.hwpmc.nbuffer -- number of global log buffers 75 */ 76 77 static int pmc_nlogbuffers = PMC_NLOGBUFFERS; 78 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "nbuffers", &pmc_nlogbuffers); 79 SYSCTL_INT(_kern_hwpmc, OID_AUTO, nbuffers, CTLFLAG_TUN|CTLFLAG_RD, 80 &pmc_nlogbuffers, 0, "number of global log buffers"); 81 82 /* 83 * Global log buffer list and associated spin lock. 84 */ 85 86 TAILQ_HEAD(, pmclog_buffer) pmc_bufferlist = 87 TAILQ_HEAD_INITIALIZER(pmc_bufferlist); 88 static struct mtx pmc_bufferlist_mtx; /* spin lock */ 89 static struct mtx pmc_kthread_mtx; /* sleep lock */ 90 91 #define PMCLOG_INIT_BUFFER_DESCRIPTOR(D) do { \ 92 const int __roundup = roundup(sizeof(*D), \ 93 sizeof(uint32_t)); \ 94 (D)->plb_fence = ((char *) (D)) + \ 95 1024*pmclog_buffer_size; \ 96 (D)->plb_base = (D)->plb_ptr = ((char *) (D)) + \ 97 __roundup; \ 98 } while (0) 99 100 101 /* 102 * Log file record constructors. 103 */ 104 #define _PMCLOG_TO_HEADER(T,L) \ 105 ((PMCLOG_HEADER_MAGIC << 24) | \ 106 (PMCLOG_TYPE_ ## T << 16) | \ 107 ((L) & 0xFFFF)) 108 109 /* reserve LEN bytes of space and initialize the entry header */ 110 #define _PMCLOG_RESERVE(PO,TYPE,LEN,ACTION) do { \ 111 uint32_t *_le; \ 112 int _len = roundup((LEN), sizeof(uint32_t)); \ 113 if ((_le = pmclog_reserve((PO), _len)) == NULL) { \ 114 ACTION; \ 115 } \ 116 *_le = _PMCLOG_TO_HEADER(TYPE,_len); \ 117 _le += 3 /* skip over timestamp */ 118 119 #define PMCLOG_RESERVE(P,T,L) _PMCLOG_RESERVE(P,T,L,return) 120 #define PMCLOG_RESERVE_WITH_ERROR(P,T,L) _PMCLOG_RESERVE(P,T,L, \ 121 error=ENOMEM;goto error) 122 123 #define PMCLOG_EMIT32(V) do { *_le++ = (V); } while (0) 124 #define PMCLOG_EMIT64(V) do { \ 125 *_le++ = (uint32_t) ((V) & 0xFFFFFFFF); \ 126 *_le++ = (uint32_t) (((V) >> 32) & 0xFFFFFFFF); \ 127 } while (0) 128 129 130 /* Emit a string. Caution: does NOT update _le, so needs to be last */ 131 #define PMCLOG_EMITSTRING(S,L) do { bcopy((S), _le, (L)); } while (0) 132 #define PMCLOG_EMITNULLSTRING(L) do { bzero(_le, (L)); } while (0) 133 134 #define PMCLOG_DESPATCH(PO) \ 135 pmclog_release((PO)); \ 136 } while (0) 137 138 139 /* 140 * Assertions about the log file format. 141 */ 142 143 CTASSERT(sizeof(struct pmclog_callchain) == 6*4 + 144 PMC_CALLCHAIN_DEPTH_MAX*sizeof(uintfptr_t)); 145 CTASSERT(sizeof(struct pmclog_closelog) == 3*4); 146 CTASSERT(sizeof(struct pmclog_dropnotify) == 3*4); 147 CTASSERT(sizeof(struct pmclog_map_in) == PATH_MAX + 148 4*4 + sizeof(uintfptr_t)); 149 CTASSERT(offsetof(struct pmclog_map_in,pl_pathname) == 150 4*4 + sizeof(uintfptr_t)); 151 CTASSERT(sizeof(struct pmclog_map_out) == 4*4 + 2*sizeof(uintfptr_t)); 152 CTASSERT(sizeof(struct pmclog_pcsample) == 6*4 + sizeof(uintfptr_t)); 153 CTASSERT(sizeof(struct pmclog_pmcallocate) == 6*4); 154 CTASSERT(sizeof(struct pmclog_pmcattach) == 5*4 + PATH_MAX); 155 CTASSERT(offsetof(struct pmclog_pmcattach,pl_pathname) == 5*4); 156 CTASSERT(sizeof(struct pmclog_pmcdetach) == 5*4); 157 CTASSERT(sizeof(struct pmclog_proccsw) == 5*4 + 8); 158 CTASSERT(sizeof(struct pmclog_procexec) == 5*4 + PATH_MAX + 159 sizeof(uintfptr_t)); 160 CTASSERT(offsetof(struct pmclog_procexec,pl_pathname) == 5*4 + 161 sizeof(uintfptr_t)); 162 CTASSERT(sizeof(struct pmclog_procexit) == 5*4 + 8); 163 CTASSERT(sizeof(struct pmclog_procfork) == 5*4); 164 CTASSERT(sizeof(struct pmclog_sysexit) == 4*4); 165 CTASSERT(sizeof(struct pmclog_userdata) == 4*4); 166 167 /* 168 * Log buffer structure 169 */ 170 171 struct pmclog_buffer { 172 TAILQ_ENTRY(pmclog_buffer) plb_next; 173 char *plb_base; 174 char *plb_ptr; 175 char *plb_fence; 176 }; 177 178 /* 179 * Prototypes 180 */ 181 182 static int pmclog_get_buffer(struct pmc_owner *po); 183 static void pmclog_loop(void *arg); 184 static void pmclog_release(struct pmc_owner *po); 185 static uint32_t *pmclog_reserve(struct pmc_owner *po, int length); 186 static void pmclog_schedule_io(struct pmc_owner *po); 187 static void pmclog_stop_kthread(struct pmc_owner *po); 188 189 /* 190 * Helper functions 191 */ 192 193 /* 194 * Get a log buffer 195 */ 196 197 static int 198 pmclog_get_buffer(struct pmc_owner *po) 199 { 200 struct pmclog_buffer *plb; 201 202 mtx_assert(&po->po_mtx, MA_OWNED); 203 204 KASSERT(po->po_curbuf == NULL, 205 ("[pmclog,%d] po=%p current buffer still valid", __LINE__, po)); 206 207 mtx_lock_spin(&pmc_bufferlist_mtx); 208 if ((plb = TAILQ_FIRST(&pmc_bufferlist)) != NULL) 209 TAILQ_REMOVE(&pmc_bufferlist, plb, plb_next); 210 mtx_unlock_spin(&pmc_bufferlist_mtx); 211 212 PMCDBG(LOG,GTB,1, "po=%p plb=%p", po, plb); 213 214 #ifdef DEBUG 215 if (plb) 216 KASSERT(plb->plb_ptr == plb->plb_base && 217 plb->plb_base < plb->plb_fence, 218 ("[pmclog,%d] po=%p buffer invariants: ptr=%p " 219 "base=%p fence=%p", __LINE__, po, plb->plb_ptr, 220 plb->plb_base, plb->plb_fence)); 221 #endif 222 223 po->po_curbuf = plb; 224 225 /* update stats */ 226 atomic_add_int(&pmc_stats.pm_buffer_requests, 1); 227 if (plb == NULL) 228 atomic_add_int(&pmc_stats.pm_buffer_requests_failed, 1); 229 230 return (plb ? 0 : ENOMEM); 231 } 232 233 /* 234 * Log handler loop. 235 * 236 * This function is executed by each pmc owner's helper thread. 237 */ 238 239 static void 240 pmclog_loop(void *arg) 241 { 242 int error; 243 struct pmc_owner *po; 244 struct pmclog_buffer *lb; 245 struct proc *p; 246 struct ucred *ownercred; 247 struct ucred *mycred; 248 struct thread *td; 249 struct uio auio; 250 struct iovec aiov; 251 size_t nbytes; 252 253 po = (struct pmc_owner *) arg; 254 p = po->po_owner; 255 td = curthread; 256 mycred = td->td_ucred; 257 258 PROC_LOCK(p); 259 ownercred = crhold(p->p_ucred); 260 PROC_UNLOCK(p); 261 262 PMCDBG(LOG,INI,1, "po=%p kt=%p", po, po->po_kthread); 263 KASSERT(po->po_kthread == curthread->td_proc, 264 ("[pmclog,%d] proc mismatch po=%p po/kt=%p curproc=%p", __LINE__, 265 po, po->po_kthread, curthread->td_proc)); 266 267 lb = NULL; 268 269 270 /* 271 * Loop waiting for I/O requests to be added to the owner 272 * struct's queue. The loop is exited when the log file 273 * is deconfigured. 274 */ 275 276 mtx_lock(&pmc_kthread_mtx); 277 278 for (;;) { 279 280 /* check if we've been asked to exit */ 281 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) 282 break; 283 284 if (lb == NULL) { /* look for a fresh buffer to write */ 285 mtx_lock_spin(&po->po_mtx); 286 if ((lb = TAILQ_FIRST(&po->po_logbuffers)) == NULL) { 287 mtx_unlock_spin(&po->po_mtx); 288 289 /* No more buffers and shutdown required. */ 290 if (po->po_flags & PMC_PO_SHUTDOWN) { 291 mtx_unlock(&pmc_kthread_mtx); 292 /* 293 * Close the file to get PMCLOG_EOF 294 * error in pmclog(3). 295 */ 296 fo_close(po->po_file, curthread); 297 mtx_lock(&pmc_kthread_mtx); 298 break; 299 } 300 301 (void) msleep(po, &pmc_kthread_mtx, PWAIT, 302 "pmcloop", 0); 303 continue; 304 } 305 306 TAILQ_REMOVE(&po->po_logbuffers, lb, plb_next); 307 mtx_unlock_spin(&po->po_mtx); 308 } 309 310 mtx_unlock(&pmc_kthread_mtx); 311 312 /* process the request */ 313 PMCDBG(LOG,WRI,2, "po=%p base=%p ptr=%p", po, 314 lb->plb_base, lb->plb_ptr); 315 /* change our thread's credentials before issuing the I/O */ 316 317 aiov.iov_base = lb->plb_base; 318 aiov.iov_len = nbytes = lb->plb_ptr - lb->plb_base; 319 320 auio.uio_iov = &aiov; 321 auio.uio_iovcnt = 1; 322 auio.uio_offset = -1; 323 auio.uio_resid = nbytes; 324 auio.uio_rw = UIO_WRITE; 325 auio.uio_segflg = UIO_SYSSPACE; 326 auio.uio_td = td; 327 328 /* switch thread credentials -- see kern_ktrace.c */ 329 td->td_ucred = ownercred; 330 error = fo_write(po->po_file, &auio, ownercred, 0, td); 331 td->td_ucred = mycred; 332 333 if (error) { 334 /* XXX some errors are recoverable */ 335 /* send a SIGIO to the owner and exit */ 336 PROC_LOCK(p); 337 kern_psignal(p, SIGIO); 338 PROC_UNLOCK(p); 339 340 mtx_lock(&pmc_kthread_mtx); 341 342 po->po_error = error; /* save for flush log */ 343 344 PMCDBG(LOG,WRI,2, "po=%p error=%d", po, error); 345 346 break; 347 } 348 349 mtx_lock(&pmc_kthread_mtx); 350 351 /* put the used buffer back into the global pool */ 352 PMCLOG_INIT_BUFFER_DESCRIPTOR(lb); 353 354 mtx_lock_spin(&pmc_bufferlist_mtx); 355 TAILQ_INSERT_HEAD(&pmc_bufferlist, lb, plb_next); 356 mtx_unlock_spin(&pmc_bufferlist_mtx); 357 358 lb = NULL; 359 } 360 361 wakeup_one(po->po_kthread); 362 po->po_kthread = NULL; 363 364 mtx_unlock(&pmc_kthread_mtx); 365 366 /* return the current I/O buffer to the global pool */ 367 if (lb) { 368 PMCLOG_INIT_BUFFER_DESCRIPTOR(lb); 369 370 mtx_lock_spin(&pmc_bufferlist_mtx); 371 TAILQ_INSERT_HEAD(&pmc_bufferlist, lb, plb_next); 372 mtx_unlock_spin(&pmc_bufferlist_mtx); 373 } 374 375 /* 376 * Exit this thread, signalling the waiter 377 */ 378 379 crfree(ownercred); 380 381 kproc_exit(0); 382 } 383 384 /* 385 * Release and log entry and schedule an I/O if needed. 386 */ 387 388 static void 389 pmclog_release(struct pmc_owner *po) 390 { 391 KASSERT(po->po_curbuf->plb_ptr >= po->po_curbuf->plb_base, 392 ("[pmclog,%d] buffer invariants po=%p ptr=%p base=%p", __LINE__, 393 po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_base)); 394 KASSERT(po->po_curbuf->plb_ptr <= po->po_curbuf->plb_fence, 395 ("[pmclog,%d] buffer invariants po=%p ptr=%p fenc=%p", __LINE__, 396 po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_fence)); 397 398 /* schedule an I/O if we've filled a buffer */ 399 if (po->po_curbuf->plb_ptr >= po->po_curbuf->plb_fence) 400 pmclog_schedule_io(po); 401 402 mtx_unlock_spin(&po->po_mtx); 403 404 PMCDBG(LOG,REL,1, "po=%p", po); 405 } 406 407 408 /* 409 * Attempt to reserve 'length' bytes of space in an owner's log 410 * buffer. The function returns a pointer to 'length' bytes of space 411 * if there was enough space or returns NULL if no space was 412 * available. Non-null returns do so with the po mutex locked. The 413 * caller must invoke pmclog_release() on the pmc owner structure 414 * when done. 415 */ 416 417 static uint32_t * 418 pmclog_reserve(struct pmc_owner *po, int length) 419 { 420 uintptr_t newptr, oldptr; 421 uint32_t *lh; 422 struct timespec ts; 423 424 PMCDBG(LOG,ALL,1, "po=%p len=%d", po, length); 425 426 KASSERT(length % sizeof(uint32_t) == 0, 427 ("[pmclog,%d] length not a multiple of word size", __LINE__)); 428 429 mtx_lock_spin(&po->po_mtx); 430 431 /* No more data when shutdown in progress. */ 432 if (po->po_flags & PMC_PO_SHUTDOWN) { 433 mtx_unlock_spin(&po->po_mtx); 434 return (NULL); 435 } 436 437 if (po->po_curbuf == NULL) 438 if (pmclog_get_buffer(po) != 0) { 439 mtx_unlock_spin(&po->po_mtx); 440 return (NULL); 441 } 442 443 KASSERT(po->po_curbuf != NULL, 444 ("[pmclog,%d] po=%p no current buffer", __LINE__, po)); 445 446 KASSERT(po->po_curbuf->plb_ptr >= po->po_curbuf->plb_base && 447 po->po_curbuf->plb_ptr <= po->po_curbuf->plb_fence, 448 ("[pmclog,%d] po=%p buffer invariants: ptr=%p base=%p fence=%p", 449 __LINE__, po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_base, 450 po->po_curbuf->plb_fence)); 451 452 oldptr = (uintptr_t) po->po_curbuf->plb_ptr; 453 newptr = oldptr + length; 454 455 KASSERT(oldptr != (uintptr_t) NULL, 456 ("[pmclog,%d] po=%p Null log buffer pointer", __LINE__, po)); 457 458 /* 459 * If we have space in the current buffer, return a pointer to 460 * available space with the PO structure locked. 461 */ 462 if (newptr <= (uintptr_t) po->po_curbuf->plb_fence) { 463 po->po_curbuf->plb_ptr = (char *) newptr; 464 goto done; 465 } 466 467 /* 468 * Otherwise, schedule the current buffer for output and get a 469 * fresh buffer. 470 */ 471 pmclog_schedule_io(po); 472 473 if (pmclog_get_buffer(po) != 0) { 474 mtx_unlock_spin(&po->po_mtx); 475 return (NULL); 476 } 477 478 KASSERT(po->po_curbuf != NULL, 479 ("[pmclog,%d] po=%p no current buffer", __LINE__, po)); 480 481 KASSERT(po->po_curbuf->plb_ptr != NULL, 482 ("[pmclog,%d] null return from pmc_get_log_buffer", __LINE__)); 483 484 KASSERT(po->po_curbuf->plb_ptr == po->po_curbuf->plb_base && 485 po->po_curbuf->plb_ptr <= po->po_curbuf->plb_fence, 486 ("[pmclog,%d] po=%p buffer invariants: ptr=%p base=%p fence=%p", 487 __LINE__, po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_base, 488 po->po_curbuf->plb_fence)); 489 490 oldptr = (uintptr_t) po->po_curbuf->plb_ptr; 491 492 done: 493 lh = (uint32_t *) oldptr; 494 lh++; /* skip header */ 495 getnanotime(&ts); /* fill in the timestamp */ 496 *lh++ = ts.tv_sec & 0xFFFFFFFF; 497 *lh++ = ts.tv_nsec & 0xFFFFFFF; 498 return ((uint32_t *) oldptr); 499 } 500 501 /* 502 * Schedule an I/O. 503 * 504 * Transfer the current buffer to the helper kthread. 505 */ 506 507 static void 508 pmclog_schedule_io(struct pmc_owner *po) 509 { 510 KASSERT(po->po_curbuf != NULL, 511 ("[pmclog,%d] schedule_io with null buffer po=%p", __LINE__, po)); 512 513 KASSERT(po->po_curbuf->plb_ptr >= po->po_curbuf->plb_base, 514 ("[pmclog,%d] buffer invariants po=%p ptr=%p base=%p", __LINE__, 515 po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_base)); 516 KASSERT(po->po_curbuf->plb_ptr <= po->po_curbuf->plb_fence, 517 ("[pmclog,%d] buffer invariants po=%p ptr=%p fenc=%p", __LINE__, 518 po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_fence)); 519 520 PMCDBG(LOG,SIO, 1, "po=%p", po); 521 522 mtx_assert(&po->po_mtx, MA_OWNED); 523 524 /* 525 * Add the current buffer to the tail of the buffer list and 526 * wakeup the helper. 527 */ 528 TAILQ_INSERT_TAIL(&po->po_logbuffers, po->po_curbuf, plb_next); 529 po->po_curbuf = NULL; 530 wakeup_one(po); 531 } 532 533 /* 534 * Stop the helper kthread. 535 */ 536 537 static void 538 pmclog_stop_kthread(struct pmc_owner *po) 539 { 540 /* 541 * Close the file to force the thread out of fo_write, 542 * unset flag, wakeup the helper thread, 543 * wait for it to exit 544 */ 545 546 if (po->po_file != NULL) 547 fo_close(po->po_file, curthread); 548 549 mtx_lock(&pmc_kthread_mtx); 550 po->po_flags &= ~PMC_PO_OWNS_LOGFILE; 551 wakeup_one(po); 552 if (po->po_kthread) 553 msleep(po->po_kthread, &pmc_kthread_mtx, PPAUSE, "pmckstp", 0); 554 mtx_unlock(&pmc_kthread_mtx); 555 } 556 557 /* 558 * Public functions 559 */ 560 561 /* 562 * Configure a log file for pmc owner 'po'. 563 * 564 * Parameter 'logfd' is a file handle referencing an open file in the 565 * owner process. This file needs to have been opened for writing. 566 */ 567 568 int 569 pmclog_configure_log(struct pmc_mdep *md, struct pmc_owner *po, int logfd) 570 { 571 int error; 572 struct proc *p; 573 cap_rights_t rights; 574 575 /* 576 * As long as it is possible to get a LOR between pmc_sx lock and 577 * proctree/allproc sx locks used for adding a new process, assure 578 * the former is not held here. 579 */ 580 sx_assert(&pmc_sx, SA_UNLOCKED); 581 PMCDBG(LOG,CFG,1, "config po=%p logfd=%d", po, logfd); 582 583 p = po->po_owner; 584 585 /* return EBUSY if a log file was already present */ 586 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 587 return (EBUSY); 588 589 KASSERT(po->po_kthread == NULL, 590 ("[pmclog,%d] po=%p kthread (%p) already present", __LINE__, po, 591 po->po_kthread)); 592 KASSERT(po->po_file == NULL, 593 ("[pmclog,%d] po=%p file (%p) already present", __LINE__, po, 594 po->po_file)); 595 596 /* get a reference to the file state */ 597 error = fget_write(curthread, logfd, 598 cap_rights_init(&rights, CAP_WRITE), &po->po_file); 599 if (error) 600 goto error; 601 602 /* mark process as owning a log file */ 603 po->po_flags |= PMC_PO_OWNS_LOGFILE; 604 error = kproc_create(pmclog_loop, po, &po->po_kthread, 605 RFHIGHPID, 0, "hwpmc: proc(%d)", p->p_pid); 606 if (error) 607 goto error; 608 609 /* mark process as using HWPMCs */ 610 PROC_LOCK(p); 611 p->p_flag |= P_HWPMC; 612 PROC_UNLOCK(p); 613 614 /* create a log initialization entry */ 615 PMCLOG_RESERVE_WITH_ERROR(po, INITIALIZE, 616 sizeof(struct pmclog_initialize)); 617 PMCLOG_EMIT32(PMC_VERSION); 618 PMCLOG_EMIT32(md->pmd_cputype); 619 PMCLOG_DESPATCH(po); 620 621 return (0); 622 623 error: 624 /* shutdown the thread */ 625 if (po->po_kthread) 626 pmclog_stop_kthread(po); 627 628 KASSERT(po->po_kthread == NULL, ("[pmclog,%d] po=%p kthread not " 629 "stopped", __LINE__, po)); 630 631 if (po->po_file) 632 (void) fdrop(po->po_file, curthread); 633 po->po_file = NULL; /* clear file and error state */ 634 po->po_error = 0; 635 636 return (error); 637 } 638 639 640 /* 641 * De-configure a log file. This will throw away any buffers queued 642 * for this owner process. 643 */ 644 645 int 646 pmclog_deconfigure_log(struct pmc_owner *po) 647 { 648 int error; 649 struct pmclog_buffer *lb; 650 651 PMCDBG(LOG,CFG,1, "de-config po=%p", po); 652 653 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) 654 return (EINVAL); 655 656 KASSERT(po->po_sscount == 0, 657 ("[pmclog,%d] po=%p still owning SS PMCs", __LINE__, po)); 658 KASSERT(po->po_file != NULL, 659 ("[pmclog,%d] po=%p no log file", __LINE__, po)); 660 661 /* stop the kthread, this will reset the 'OWNS_LOGFILE' flag */ 662 pmclog_stop_kthread(po); 663 664 KASSERT(po->po_kthread == NULL, 665 ("[pmclog,%d] po=%p kthread not stopped", __LINE__, po)); 666 667 /* return all queued log buffers to the global pool */ 668 while ((lb = TAILQ_FIRST(&po->po_logbuffers)) != NULL) { 669 TAILQ_REMOVE(&po->po_logbuffers, lb, plb_next); 670 PMCLOG_INIT_BUFFER_DESCRIPTOR(lb); 671 mtx_lock_spin(&pmc_bufferlist_mtx); 672 TAILQ_INSERT_HEAD(&pmc_bufferlist, lb, plb_next); 673 mtx_unlock_spin(&pmc_bufferlist_mtx); 674 } 675 676 /* return the 'current' buffer to the global pool */ 677 if ((lb = po->po_curbuf) != NULL) { 678 PMCLOG_INIT_BUFFER_DESCRIPTOR(lb); 679 mtx_lock_spin(&pmc_bufferlist_mtx); 680 TAILQ_INSERT_HEAD(&pmc_bufferlist, lb, plb_next); 681 mtx_unlock_spin(&pmc_bufferlist_mtx); 682 } 683 684 /* drop a reference to the fd */ 685 error = fdrop(po->po_file, curthread); 686 po->po_file = NULL; 687 po->po_error = 0; 688 689 return (error); 690 } 691 692 /* 693 * Flush a process' log buffer. 694 */ 695 696 int 697 pmclog_flush(struct pmc_owner *po) 698 { 699 int error; 700 struct pmclog_buffer *lb; 701 702 PMCDBG(LOG,FLS,1, "po=%p", po); 703 704 /* 705 * If there is a pending error recorded by the logger thread, 706 * return that. 707 */ 708 if (po->po_error) 709 return (po->po_error); 710 711 error = 0; 712 713 /* 714 * Check that we do have an active log file. 715 */ 716 mtx_lock(&pmc_kthread_mtx); 717 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) { 718 error = EINVAL; 719 goto error; 720 } 721 722 /* 723 * Schedule the current buffer if any and not empty. 724 */ 725 mtx_lock_spin(&po->po_mtx); 726 lb = po->po_curbuf; 727 if (lb && lb->plb_ptr != lb->plb_base) { 728 pmclog_schedule_io(po); 729 } else 730 error = ENOBUFS; 731 mtx_unlock_spin(&po->po_mtx); 732 733 error: 734 mtx_unlock(&pmc_kthread_mtx); 735 736 return (error); 737 } 738 739 int 740 pmclog_close(struct pmc_owner *po) 741 { 742 743 PMCDBG(LOG,CLO,1, "po=%p", po); 744 745 mtx_lock(&pmc_kthread_mtx); 746 747 /* 748 * Schedule the current buffer. 749 */ 750 mtx_lock_spin(&po->po_mtx); 751 if (po->po_curbuf) 752 pmclog_schedule_io(po); 753 else 754 wakeup_one(po); 755 mtx_unlock_spin(&po->po_mtx); 756 757 /* 758 * Initiate shutdown: no new data queued, 759 * thread will close file on last block. 760 */ 761 po->po_flags |= PMC_PO_SHUTDOWN; 762 763 mtx_unlock(&pmc_kthread_mtx); 764 765 return (0); 766 } 767 768 void 769 pmclog_process_callchain(struct pmc *pm, struct pmc_sample *ps) 770 { 771 int n, recordlen; 772 uint32_t flags; 773 struct pmc_owner *po; 774 775 PMCDBG(LOG,SAM,1,"pm=%p pid=%d n=%d", pm, ps->ps_pid, 776 ps->ps_nsamples); 777 778 recordlen = offsetof(struct pmclog_callchain, pl_pc) + 779 ps->ps_nsamples * sizeof(uintfptr_t); 780 po = pm->pm_owner; 781 flags = PMC_CALLCHAIN_TO_CPUFLAGS(ps->ps_cpu,ps->ps_flags); 782 PMCLOG_RESERVE(po, CALLCHAIN, recordlen); 783 PMCLOG_EMIT32(ps->ps_pid); 784 PMCLOG_EMIT32(pm->pm_id); 785 PMCLOG_EMIT32(flags); 786 for (n = 0; n < ps->ps_nsamples; n++) 787 PMCLOG_EMITADDR(ps->ps_pc[n]); 788 PMCLOG_DESPATCH(po); 789 } 790 791 void 792 pmclog_process_closelog(struct pmc_owner *po) 793 { 794 PMCLOG_RESERVE(po,CLOSELOG,sizeof(struct pmclog_closelog)); 795 PMCLOG_DESPATCH(po); 796 } 797 798 void 799 pmclog_process_dropnotify(struct pmc_owner *po) 800 { 801 PMCLOG_RESERVE(po,DROPNOTIFY,sizeof(struct pmclog_dropnotify)); 802 PMCLOG_DESPATCH(po); 803 } 804 805 void 806 pmclog_process_map_in(struct pmc_owner *po, pid_t pid, uintfptr_t start, 807 const char *path) 808 { 809 int pathlen, recordlen; 810 811 KASSERT(path != NULL, ("[pmclog,%d] map-in, null path", __LINE__)); 812 813 pathlen = strlen(path) + 1; /* #bytes for path name */ 814 recordlen = offsetof(struct pmclog_map_in, pl_pathname) + 815 pathlen; 816 817 PMCLOG_RESERVE(po, MAP_IN, recordlen); 818 PMCLOG_EMIT32(pid); 819 PMCLOG_EMITADDR(start); 820 PMCLOG_EMITSTRING(path,pathlen); 821 PMCLOG_DESPATCH(po); 822 } 823 824 void 825 pmclog_process_map_out(struct pmc_owner *po, pid_t pid, uintfptr_t start, 826 uintfptr_t end) 827 { 828 KASSERT(start <= end, ("[pmclog,%d] start > end", __LINE__)); 829 830 PMCLOG_RESERVE(po, MAP_OUT, sizeof(struct pmclog_map_out)); 831 PMCLOG_EMIT32(pid); 832 PMCLOG_EMITADDR(start); 833 PMCLOG_EMITADDR(end); 834 PMCLOG_DESPATCH(po); 835 } 836 837 void 838 pmclog_process_pmcallocate(struct pmc *pm) 839 { 840 struct pmc_owner *po; 841 struct pmc_soft *ps; 842 843 po = pm->pm_owner; 844 845 PMCDBG(LOG,ALL,1, "pm=%p", pm); 846 847 if (PMC_TO_CLASS(pm) == PMC_CLASS_SOFT) { 848 PMCLOG_RESERVE(po, PMCALLOCATEDYN, 849 sizeof(struct pmclog_pmcallocatedyn)); 850 PMCLOG_EMIT32(pm->pm_id); 851 PMCLOG_EMIT32(pm->pm_event); 852 PMCLOG_EMIT32(pm->pm_flags); 853 ps = pmc_soft_ev_acquire(pm->pm_event); 854 if (ps != NULL) 855 PMCLOG_EMITSTRING(ps->ps_ev.pm_ev_name,PMC_NAME_MAX); 856 else 857 PMCLOG_EMITNULLSTRING(PMC_NAME_MAX); 858 pmc_soft_ev_release(ps); 859 PMCLOG_DESPATCH(po); 860 } else { 861 PMCLOG_RESERVE(po, PMCALLOCATE, 862 sizeof(struct pmclog_pmcallocate)); 863 PMCLOG_EMIT32(pm->pm_id); 864 PMCLOG_EMIT32(pm->pm_event); 865 PMCLOG_EMIT32(pm->pm_flags); 866 PMCLOG_DESPATCH(po); 867 } 868 } 869 870 void 871 pmclog_process_pmcattach(struct pmc *pm, pid_t pid, char *path) 872 { 873 int pathlen, recordlen; 874 struct pmc_owner *po; 875 876 PMCDBG(LOG,ATT,1,"pm=%p pid=%d", pm, pid); 877 878 po = pm->pm_owner; 879 880 pathlen = strlen(path) + 1; /* #bytes for the string */ 881 recordlen = offsetof(struct pmclog_pmcattach, pl_pathname) + pathlen; 882 883 PMCLOG_RESERVE(po, PMCATTACH, recordlen); 884 PMCLOG_EMIT32(pm->pm_id); 885 PMCLOG_EMIT32(pid); 886 PMCLOG_EMITSTRING(path, pathlen); 887 PMCLOG_DESPATCH(po); 888 } 889 890 void 891 pmclog_process_pmcdetach(struct pmc *pm, pid_t pid) 892 { 893 struct pmc_owner *po; 894 895 PMCDBG(LOG,ATT,1,"!pm=%p pid=%d", pm, pid); 896 897 po = pm->pm_owner; 898 899 PMCLOG_RESERVE(po, PMCDETACH, sizeof(struct pmclog_pmcdetach)); 900 PMCLOG_EMIT32(pm->pm_id); 901 PMCLOG_EMIT32(pid); 902 PMCLOG_DESPATCH(po); 903 } 904 905 /* 906 * Log a context switch event to the log file. 907 */ 908 909 void 910 pmclog_process_proccsw(struct pmc *pm, struct pmc_process *pp, pmc_value_t v) 911 { 912 struct pmc_owner *po; 913 914 KASSERT(pm->pm_flags & PMC_F_LOG_PROCCSW, 915 ("[pmclog,%d] log-process-csw called gratuitously", __LINE__)); 916 917 PMCDBG(LOG,SWO,1,"pm=%p pid=%d v=%jx", pm, pp->pp_proc->p_pid, 918 v); 919 920 po = pm->pm_owner; 921 922 PMCLOG_RESERVE(po, PROCCSW, sizeof(struct pmclog_proccsw)); 923 PMCLOG_EMIT32(pm->pm_id); 924 PMCLOG_EMIT64(v); 925 PMCLOG_EMIT32(pp->pp_proc->p_pid); 926 PMCLOG_DESPATCH(po); 927 } 928 929 void 930 pmclog_process_procexec(struct pmc_owner *po, pmc_id_t pmid, pid_t pid, 931 uintfptr_t startaddr, char *path) 932 { 933 int pathlen, recordlen; 934 935 PMCDBG(LOG,EXC,1,"po=%p pid=%d path=\"%s\"", po, pid, path); 936 937 pathlen = strlen(path) + 1; /* #bytes for the path */ 938 recordlen = offsetof(struct pmclog_procexec, pl_pathname) + pathlen; 939 940 PMCLOG_RESERVE(po, PROCEXEC, recordlen); 941 PMCLOG_EMIT32(pid); 942 PMCLOG_EMITADDR(startaddr); 943 PMCLOG_EMIT32(pmid); 944 PMCLOG_EMITSTRING(path,pathlen); 945 PMCLOG_DESPATCH(po); 946 } 947 948 /* 949 * Log a process exit event (and accumulated pmc value) to the log file. 950 */ 951 952 void 953 pmclog_process_procexit(struct pmc *pm, struct pmc_process *pp) 954 { 955 int ri; 956 struct pmc_owner *po; 957 958 ri = PMC_TO_ROWINDEX(pm); 959 PMCDBG(LOG,EXT,1,"pm=%p pid=%d v=%jx", pm, pp->pp_proc->p_pid, 960 pp->pp_pmcs[ri].pp_pmcval); 961 962 po = pm->pm_owner; 963 964 PMCLOG_RESERVE(po, PROCEXIT, sizeof(struct pmclog_procexit)); 965 PMCLOG_EMIT32(pm->pm_id); 966 PMCLOG_EMIT64(pp->pp_pmcs[ri].pp_pmcval); 967 PMCLOG_EMIT32(pp->pp_proc->p_pid); 968 PMCLOG_DESPATCH(po); 969 } 970 971 /* 972 * Log a fork event. 973 */ 974 975 void 976 pmclog_process_procfork(struct pmc_owner *po, pid_t oldpid, pid_t newpid) 977 { 978 PMCLOG_RESERVE(po, PROCFORK, sizeof(struct pmclog_procfork)); 979 PMCLOG_EMIT32(oldpid); 980 PMCLOG_EMIT32(newpid); 981 PMCLOG_DESPATCH(po); 982 } 983 984 /* 985 * Log a process exit event of the form suitable for system-wide PMCs. 986 */ 987 988 void 989 pmclog_process_sysexit(struct pmc_owner *po, pid_t pid) 990 { 991 PMCLOG_RESERVE(po, SYSEXIT, sizeof(struct pmclog_sysexit)); 992 PMCLOG_EMIT32(pid); 993 PMCLOG_DESPATCH(po); 994 } 995 996 /* 997 * Write a user log entry. 998 */ 999 1000 int 1001 pmclog_process_userlog(struct pmc_owner *po, struct pmc_op_writelog *wl) 1002 { 1003 int error; 1004 1005 PMCDBG(LOG,WRI,1, "writelog po=%p ud=0x%x", po, wl->pm_userdata); 1006 1007 error = 0; 1008 1009 PMCLOG_RESERVE_WITH_ERROR(po, USERDATA, 1010 sizeof(struct pmclog_userdata)); 1011 PMCLOG_EMIT32(wl->pm_userdata); 1012 PMCLOG_DESPATCH(po); 1013 1014 error: 1015 return (error); 1016 } 1017 1018 /* 1019 * Initialization. 1020 * 1021 * Create a pool of log buffers and initialize mutexes. 1022 */ 1023 1024 void 1025 pmclog_initialize() 1026 { 1027 int n; 1028 struct pmclog_buffer *plb; 1029 1030 if (pmclog_buffer_size <= 0) { 1031 (void) printf("hwpmc: tunable logbuffersize=%d must be " 1032 "greater than zero.\n", pmclog_buffer_size); 1033 pmclog_buffer_size = PMC_LOG_BUFFER_SIZE; 1034 } 1035 1036 if (pmc_nlogbuffers <= 0) { 1037 (void) printf("hwpmc: tunable nlogbuffers=%d must be greater " 1038 "than zero.\n", pmc_nlogbuffers); 1039 pmc_nlogbuffers = PMC_NLOGBUFFERS; 1040 } 1041 1042 /* create global pool of log buffers */ 1043 for (n = 0; n < pmc_nlogbuffers; n++) { 1044 plb = malloc(1024 * pmclog_buffer_size, M_PMC, 1045 M_WAITOK|M_ZERO); 1046 PMCLOG_INIT_BUFFER_DESCRIPTOR(plb); 1047 TAILQ_INSERT_HEAD(&pmc_bufferlist, plb, plb_next); 1048 } 1049 mtx_init(&pmc_bufferlist_mtx, "pmc-buffer-list", "pmc-leaf", 1050 MTX_SPIN); 1051 mtx_init(&pmc_kthread_mtx, "pmc-kthread", "pmc-sleep", MTX_DEF); 1052 } 1053 1054 /* 1055 * Shutdown logging. 1056 * 1057 * Destroy mutexes and release memory back the to free pool. 1058 */ 1059 1060 void 1061 pmclog_shutdown() 1062 { 1063 struct pmclog_buffer *plb; 1064 1065 mtx_destroy(&pmc_kthread_mtx); 1066 mtx_destroy(&pmc_bufferlist_mtx); 1067 1068 while ((plb = TAILQ_FIRST(&pmc_bufferlist)) != NULL) { 1069 TAILQ_REMOVE(&pmc_bufferlist, plb, plb_next); 1070 free(plb, M_PMC); 1071 } 1072 } 1073