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 #if (__FreeBSD_version < 1100000) 70 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "logbuffersize", &pmclog_buffer_size); 71 #endif 72 SYSCTL_INT(_kern_hwpmc, OID_AUTO, logbuffersize, CTLFLAG_RDTUN, 73 &pmclog_buffer_size, 0, "size of log buffers in kilobytes"); 74 75 /* 76 * kern.hwpmc.nbuffer -- number of global log buffers 77 */ 78 79 static int pmc_nlogbuffers = PMC_NLOGBUFFERS; 80 #if (__FreeBSD_version < 1100000) 81 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "nbuffers", &pmc_nlogbuffers); 82 #endif 83 SYSCTL_INT(_kern_hwpmc, OID_AUTO, nbuffers, CTLFLAG_RDTUN, 84 &pmc_nlogbuffers, 0, "number of global log buffers"); 85 86 /* 87 * Global log buffer list and associated spin lock. 88 */ 89 90 TAILQ_HEAD(, pmclog_buffer) pmc_bufferlist = 91 TAILQ_HEAD_INITIALIZER(pmc_bufferlist); 92 static struct mtx pmc_bufferlist_mtx; /* spin lock */ 93 static struct mtx pmc_kthread_mtx; /* sleep lock */ 94 95 #define PMCLOG_INIT_BUFFER_DESCRIPTOR(D) do { \ 96 const int __roundup = roundup(sizeof(*D), \ 97 sizeof(uint32_t)); \ 98 (D)->plb_fence = ((char *) (D)) + \ 99 1024*pmclog_buffer_size; \ 100 (D)->plb_base = (D)->plb_ptr = ((char *) (D)) + \ 101 __roundup; \ 102 } while (0) 103 104 105 /* 106 * Log file record constructors. 107 */ 108 #define _PMCLOG_TO_HEADER(T,L) \ 109 ((PMCLOG_HEADER_MAGIC << 24) | \ 110 (PMCLOG_TYPE_ ## T << 16) | \ 111 ((L) & 0xFFFF)) 112 113 /* reserve LEN bytes of space and initialize the entry header */ 114 #define _PMCLOG_RESERVE(PO,TYPE,LEN,ACTION) do { \ 115 uint32_t *_le; \ 116 int _len = roundup((LEN), sizeof(uint32_t)); \ 117 if ((_le = pmclog_reserve((PO), _len)) == NULL) { \ 118 ACTION; \ 119 } \ 120 *_le = _PMCLOG_TO_HEADER(TYPE,_len); \ 121 _le += 3 /* skip over timestamp */ 122 123 #define PMCLOG_RESERVE(P,T,L) _PMCLOG_RESERVE(P,T,L,return) 124 #define PMCLOG_RESERVE_WITH_ERROR(P,T,L) _PMCLOG_RESERVE(P,T,L, \ 125 error=ENOMEM;goto error) 126 127 #define PMCLOG_EMIT32(V) do { *_le++ = (V); } while (0) 128 #define PMCLOG_EMIT64(V) do { \ 129 *_le++ = (uint32_t) ((V) & 0xFFFFFFFF); \ 130 *_le++ = (uint32_t) (((V) >> 32) & 0xFFFFFFFF); \ 131 } while (0) 132 133 134 /* Emit a string. Caution: does NOT update _le, so needs to be last */ 135 #define PMCLOG_EMITSTRING(S,L) do { bcopy((S), _le, (L)); } while (0) 136 #define PMCLOG_EMITNULLSTRING(L) do { bzero(_le, (L)); } while (0) 137 138 #define PMCLOG_DESPATCH(PO) \ 139 pmclog_release((PO)); \ 140 } while (0) 141 142 143 /* 144 * Assertions about the log file format. 145 */ 146 147 CTASSERT(sizeof(struct pmclog_callchain) == 6*4 + 148 PMC_CALLCHAIN_DEPTH_MAX*sizeof(uintfptr_t)); 149 CTASSERT(sizeof(struct pmclog_closelog) == 3*4); 150 CTASSERT(sizeof(struct pmclog_dropnotify) == 3*4); 151 CTASSERT(sizeof(struct pmclog_map_in) == PATH_MAX + 152 4*4 + sizeof(uintfptr_t)); 153 CTASSERT(offsetof(struct pmclog_map_in,pl_pathname) == 154 4*4 + sizeof(uintfptr_t)); 155 CTASSERT(sizeof(struct pmclog_map_out) == 4*4 + 2*sizeof(uintfptr_t)); 156 CTASSERT(sizeof(struct pmclog_pcsample) == 6*4 + sizeof(uintfptr_t)); 157 CTASSERT(sizeof(struct pmclog_pmcallocate) == 6*4); 158 CTASSERT(sizeof(struct pmclog_pmcattach) == 5*4 + PATH_MAX); 159 CTASSERT(offsetof(struct pmclog_pmcattach,pl_pathname) == 5*4); 160 CTASSERT(sizeof(struct pmclog_pmcdetach) == 5*4); 161 CTASSERT(sizeof(struct pmclog_proccsw) == 5*4 + 8); 162 CTASSERT(sizeof(struct pmclog_procexec) == 5*4 + PATH_MAX + 163 sizeof(uintfptr_t)); 164 CTASSERT(offsetof(struct pmclog_procexec,pl_pathname) == 5*4 + 165 sizeof(uintfptr_t)); 166 CTASSERT(sizeof(struct pmclog_procexit) == 5*4 + 8); 167 CTASSERT(sizeof(struct pmclog_procfork) == 5*4); 168 CTASSERT(sizeof(struct pmclog_sysexit) == 4*4); 169 CTASSERT(sizeof(struct pmclog_userdata) == 4*4); 170 171 /* 172 * Log buffer structure 173 */ 174 175 struct pmclog_buffer { 176 TAILQ_ENTRY(pmclog_buffer) plb_next; 177 char *plb_base; 178 char *plb_ptr; 179 char *plb_fence; 180 }; 181 182 /* 183 * Prototypes 184 */ 185 186 static int pmclog_get_buffer(struct pmc_owner *po); 187 static void pmclog_loop(void *arg); 188 static void pmclog_release(struct pmc_owner *po); 189 static uint32_t *pmclog_reserve(struct pmc_owner *po, int length); 190 static void pmclog_schedule_io(struct pmc_owner *po); 191 static void pmclog_stop_kthread(struct pmc_owner *po); 192 193 /* 194 * Helper functions 195 */ 196 197 /* 198 * Get a log buffer 199 */ 200 201 static int 202 pmclog_get_buffer(struct pmc_owner *po) 203 { 204 struct pmclog_buffer *plb; 205 206 mtx_assert(&po->po_mtx, MA_OWNED); 207 208 KASSERT(po->po_curbuf == NULL, 209 ("[pmclog,%d] po=%p current buffer still valid", __LINE__, po)); 210 211 mtx_lock_spin(&pmc_bufferlist_mtx); 212 if ((plb = TAILQ_FIRST(&pmc_bufferlist)) != NULL) 213 TAILQ_REMOVE(&pmc_bufferlist, plb, plb_next); 214 mtx_unlock_spin(&pmc_bufferlist_mtx); 215 216 PMCDBG2(LOG,GTB,1, "po=%p plb=%p", po, plb); 217 218 #ifdef HWPMC_DEBUG 219 if (plb) 220 KASSERT(plb->plb_ptr == plb->plb_base && 221 plb->plb_base < plb->plb_fence, 222 ("[pmclog,%d] po=%p buffer invariants: ptr=%p " 223 "base=%p fence=%p", __LINE__, po, plb->plb_ptr, 224 plb->plb_base, plb->plb_fence)); 225 #endif 226 227 po->po_curbuf = plb; 228 229 /* update stats */ 230 atomic_add_int(&pmc_stats.pm_buffer_requests, 1); 231 if (plb == NULL) 232 atomic_add_int(&pmc_stats.pm_buffer_requests_failed, 1); 233 234 return (plb ? 0 : ENOMEM); 235 } 236 237 /* 238 * Log handler loop. 239 * 240 * This function is executed by each pmc owner's helper thread. 241 */ 242 243 static void 244 pmclog_loop(void *arg) 245 { 246 int error; 247 struct pmc_owner *po; 248 struct pmclog_buffer *lb; 249 struct proc *p; 250 struct ucred *ownercred; 251 struct ucred *mycred; 252 struct thread *td; 253 struct uio auio; 254 struct iovec aiov; 255 size_t nbytes; 256 257 po = (struct pmc_owner *) arg; 258 p = po->po_owner; 259 td = curthread; 260 mycred = td->td_ucred; 261 262 PROC_LOCK(p); 263 ownercred = crhold(p->p_ucred); 264 PROC_UNLOCK(p); 265 266 PMCDBG2(LOG,INI,1, "po=%p kt=%p", po, po->po_kthread); 267 KASSERT(po->po_kthread == curthread->td_proc, 268 ("[pmclog,%d] proc mismatch po=%p po/kt=%p curproc=%p", __LINE__, 269 po, po->po_kthread, curthread->td_proc)); 270 271 lb = NULL; 272 273 274 /* 275 * Loop waiting for I/O requests to be added to the owner 276 * struct's queue. The loop is exited when the log file 277 * is deconfigured. 278 */ 279 280 mtx_lock(&pmc_kthread_mtx); 281 282 for (;;) { 283 284 /* check if we've been asked to exit */ 285 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) 286 break; 287 288 if (lb == NULL) { /* look for a fresh buffer to write */ 289 mtx_lock_spin(&po->po_mtx); 290 if ((lb = TAILQ_FIRST(&po->po_logbuffers)) == NULL) { 291 mtx_unlock_spin(&po->po_mtx); 292 293 /* No more buffers and shutdown required. */ 294 if (po->po_flags & PMC_PO_SHUTDOWN) { 295 mtx_unlock(&pmc_kthread_mtx); 296 /* 297 * Close the file to get PMCLOG_EOF 298 * error in pmclog(3). 299 */ 300 fo_close(po->po_file, curthread); 301 mtx_lock(&pmc_kthread_mtx); 302 break; 303 } 304 305 (void) msleep(po, &pmc_kthread_mtx, PWAIT, 306 "pmcloop", 0); 307 continue; 308 } 309 310 TAILQ_REMOVE(&po->po_logbuffers, lb, plb_next); 311 mtx_unlock_spin(&po->po_mtx); 312 } 313 314 mtx_unlock(&pmc_kthread_mtx); 315 316 /* process the request */ 317 PMCDBG3(LOG,WRI,2, "po=%p base=%p ptr=%p", po, 318 lb->plb_base, lb->plb_ptr); 319 /* change our thread's credentials before issuing the I/O */ 320 321 aiov.iov_base = lb->plb_base; 322 aiov.iov_len = nbytes = lb->plb_ptr - lb->plb_base; 323 324 auio.uio_iov = &aiov; 325 auio.uio_iovcnt = 1; 326 auio.uio_offset = -1; 327 auio.uio_resid = nbytes; 328 auio.uio_rw = UIO_WRITE; 329 auio.uio_segflg = UIO_SYSSPACE; 330 auio.uio_td = td; 331 332 /* switch thread credentials -- see kern_ktrace.c */ 333 td->td_ucred = ownercred; 334 error = fo_write(po->po_file, &auio, ownercred, 0, td); 335 td->td_ucred = mycred; 336 337 if (error) { 338 /* XXX some errors are recoverable */ 339 /* send a SIGIO to the owner and exit */ 340 PROC_LOCK(p); 341 kern_psignal(p, SIGIO); 342 PROC_UNLOCK(p); 343 344 mtx_lock(&pmc_kthread_mtx); 345 346 po->po_error = error; /* save for flush log */ 347 348 PMCDBG2(LOG,WRI,2, "po=%p error=%d", po, error); 349 350 break; 351 } 352 353 mtx_lock(&pmc_kthread_mtx); 354 355 /* put the used buffer back into the global pool */ 356 PMCLOG_INIT_BUFFER_DESCRIPTOR(lb); 357 358 mtx_lock_spin(&pmc_bufferlist_mtx); 359 TAILQ_INSERT_HEAD(&pmc_bufferlist, lb, plb_next); 360 mtx_unlock_spin(&pmc_bufferlist_mtx); 361 362 lb = NULL; 363 } 364 365 wakeup_one(po->po_kthread); 366 po->po_kthread = NULL; 367 368 mtx_unlock(&pmc_kthread_mtx); 369 370 /* return the current I/O buffer to the global pool */ 371 if (lb) { 372 PMCLOG_INIT_BUFFER_DESCRIPTOR(lb); 373 374 mtx_lock_spin(&pmc_bufferlist_mtx); 375 TAILQ_INSERT_HEAD(&pmc_bufferlist, lb, plb_next); 376 mtx_unlock_spin(&pmc_bufferlist_mtx); 377 } 378 379 /* 380 * Exit this thread, signalling the waiter 381 */ 382 383 crfree(ownercred); 384 385 kproc_exit(0); 386 } 387 388 /* 389 * Release and log entry and schedule an I/O if needed. 390 */ 391 392 static void 393 pmclog_release(struct pmc_owner *po) 394 { 395 KASSERT(po->po_curbuf->plb_ptr >= po->po_curbuf->plb_base, 396 ("[pmclog,%d] buffer invariants po=%p ptr=%p base=%p", __LINE__, 397 po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_base)); 398 KASSERT(po->po_curbuf->plb_ptr <= po->po_curbuf->plb_fence, 399 ("[pmclog,%d] buffer invariants po=%p ptr=%p fenc=%p", __LINE__, 400 po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_fence)); 401 402 /* schedule an I/O if we've filled a buffer */ 403 if (po->po_curbuf->plb_ptr >= po->po_curbuf->plb_fence) 404 pmclog_schedule_io(po); 405 406 mtx_unlock_spin(&po->po_mtx); 407 408 PMCDBG1(LOG,REL,1, "po=%p", po); 409 } 410 411 412 /* 413 * Attempt to reserve 'length' bytes of space in an owner's log 414 * buffer. The function returns a pointer to 'length' bytes of space 415 * if there was enough space or returns NULL if no space was 416 * available. Non-null returns do so with the po mutex locked. The 417 * caller must invoke pmclog_release() on the pmc owner structure 418 * when done. 419 */ 420 421 static uint32_t * 422 pmclog_reserve(struct pmc_owner *po, int length) 423 { 424 uintptr_t newptr, oldptr; 425 uint32_t *lh; 426 struct timespec ts; 427 428 PMCDBG2(LOG,ALL,1, "po=%p len=%d", po, length); 429 430 KASSERT(length % sizeof(uint32_t) == 0, 431 ("[pmclog,%d] length not a multiple of word size", __LINE__)); 432 433 mtx_lock_spin(&po->po_mtx); 434 435 /* No more data when shutdown in progress. */ 436 if (po->po_flags & PMC_PO_SHUTDOWN) { 437 mtx_unlock_spin(&po->po_mtx); 438 return (NULL); 439 } 440 441 if (po->po_curbuf == NULL) 442 if (pmclog_get_buffer(po) != 0) { 443 mtx_unlock_spin(&po->po_mtx); 444 return (NULL); 445 } 446 447 KASSERT(po->po_curbuf != NULL, 448 ("[pmclog,%d] po=%p no current buffer", __LINE__, po)); 449 450 KASSERT(po->po_curbuf->plb_ptr >= po->po_curbuf->plb_base && 451 po->po_curbuf->plb_ptr <= po->po_curbuf->plb_fence, 452 ("[pmclog,%d] po=%p buffer invariants: ptr=%p base=%p fence=%p", 453 __LINE__, po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_base, 454 po->po_curbuf->plb_fence)); 455 456 oldptr = (uintptr_t) po->po_curbuf->plb_ptr; 457 newptr = oldptr + length; 458 459 KASSERT(oldptr != (uintptr_t) NULL, 460 ("[pmclog,%d] po=%p Null log buffer pointer", __LINE__, po)); 461 462 /* 463 * If we have space in the current buffer, return a pointer to 464 * available space with the PO structure locked. 465 */ 466 if (newptr <= (uintptr_t) po->po_curbuf->plb_fence) { 467 po->po_curbuf->plb_ptr = (char *) newptr; 468 goto done; 469 } 470 471 /* 472 * Otherwise, schedule the current buffer for output and get a 473 * fresh buffer. 474 */ 475 pmclog_schedule_io(po); 476 477 if (pmclog_get_buffer(po) != 0) { 478 mtx_unlock_spin(&po->po_mtx); 479 return (NULL); 480 } 481 482 KASSERT(po->po_curbuf != NULL, 483 ("[pmclog,%d] po=%p no current buffer", __LINE__, po)); 484 485 KASSERT(po->po_curbuf->plb_ptr != NULL, 486 ("[pmclog,%d] null return from pmc_get_log_buffer", __LINE__)); 487 488 KASSERT(po->po_curbuf->plb_ptr == po->po_curbuf->plb_base && 489 po->po_curbuf->plb_ptr <= po->po_curbuf->plb_fence, 490 ("[pmclog,%d] po=%p buffer invariants: ptr=%p base=%p fence=%p", 491 __LINE__, po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_base, 492 po->po_curbuf->plb_fence)); 493 494 oldptr = (uintptr_t) po->po_curbuf->plb_ptr; 495 496 done: 497 lh = (uint32_t *) oldptr; 498 lh++; /* skip header */ 499 getnanotime(&ts); /* fill in the timestamp */ 500 *lh++ = ts.tv_sec & 0xFFFFFFFF; 501 *lh++ = ts.tv_nsec & 0xFFFFFFF; 502 return ((uint32_t *) oldptr); 503 } 504 505 /* 506 * Schedule an I/O. 507 * 508 * Transfer the current buffer to the helper kthread. 509 */ 510 511 static void 512 pmclog_schedule_io(struct pmc_owner *po) 513 { 514 KASSERT(po->po_curbuf != NULL, 515 ("[pmclog,%d] schedule_io with null buffer po=%p", __LINE__, po)); 516 517 KASSERT(po->po_curbuf->plb_ptr >= po->po_curbuf->plb_base, 518 ("[pmclog,%d] buffer invariants po=%p ptr=%p base=%p", __LINE__, 519 po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_base)); 520 KASSERT(po->po_curbuf->plb_ptr <= po->po_curbuf->plb_fence, 521 ("[pmclog,%d] buffer invariants po=%p ptr=%p fenc=%p", __LINE__, 522 po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_fence)); 523 524 PMCDBG1(LOG,SIO, 1, "po=%p", po); 525 526 mtx_assert(&po->po_mtx, MA_OWNED); 527 528 /* 529 * Add the current buffer to the tail of the buffer list and 530 * wakeup the helper. 531 */ 532 TAILQ_INSERT_TAIL(&po->po_logbuffers, po->po_curbuf, plb_next); 533 po->po_curbuf = NULL; 534 wakeup_one(po); 535 } 536 537 /* 538 * Stop the helper kthread. 539 */ 540 541 static void 542 pmclog_stop_kthread(struct pmc_owner *po) 543 { 544 /* 545 * Close the file to force the thread out of fo_write, 546 * unset flag, wakeup the helper thread, 547 * wait for it to exit 548 */ 549 550 if (po->po_file != NULL) 551 fo_close(po->po_file, curthread); 552 553 mtx_lock(&pmc_kthread_mtx); 554 po->po_flags &= ~PMC_PO_OWNS_LOGFILE; 555 wakeup_one(po); 556 if (po->po_kthread) 557 msleep(po->po_kthread, &pmc_kthread_mtx, PPAUSE, "pmckstp", 0); 558 mtx_unlock(&pmc_kthread_mtx); 559 } 560 561 /* 562 * Public functions 563 */ 564 565 /* 566 * Configure a log file for pmc owner 'po'. 567 * 568 * Parameter 'logfd' is a file handle referencing an open file in the 569 * owner process. This file needs to have been opened for writing. 570 */ 571 572 int 573 pmclog_configure_log(struct pmc_mdep *md, struct pmc_owner *po, int logfd) 574 { 575 int error; 576 struct proc *p; 577 cap_rights_t rights; 578 /* 579 * As long as it is possible to get a LOR between pmc_sx lock and 580 * proctree/allproc sx locks used for adding a new process, assure 581 * the former is not held here. 582 */ 583 sx_assert(&pmc_sx, SA_UNLOCKED); 584 PMCDBG2(LOG,CFG,1, "config po=%p logfd=%d", po, logfd); 585 586 p = po->po_owner; 587 588 /* return EBUSY if a log file was already present */ 589 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 590 return (EBUSY); 591 592 KASSERT(po->po_kthread == NULL, 593 ("[pmclog,%d] po=%p kthread (%p) already present", __LINE__, po, 594 po->po_kthread)); 595 KASSERT(po->po_file == NULL, 596 ("[pmclog,%d] po=%p file (%p) already present", __LINE__, po, 597 po->po_file)); 598 599 /* get a reference to the file state */ 600 error = fget_write(curthread, logfd, 601 cap_rights_init(&rights, CAP_WRITE), &po->po_file); 602 if (error) 603 goto error; 604 605 /* mark process as owning a log file */ 606 po->po_flags |= PMC_PO_OWNS_LOGFILE; 607 error = kproc_create(pmclog_loop, po, &po->po_kthread, 608 RFHIGHPID, 0, "hwpmc: proc(%d)", p->p_pid); 609 if (error) 610 goto error; 611 612 /* mark process as using HWPMCs */ 613 PROC_LOCK(p); 614 p->p_flag |= P_HWPMC; 615 PROC_UNLOCK(p); 616 617 /* create a log initialization entry */ 618 PMCLOG_RESERVE_WITH_ERROR(po, INITIALIZE, 619 sizeof(struct pmclog_initialize)); 620 PMCLOG_EMIT32(PMC_VERSION); 621 PMCLOG_EMIT32(md->pmd_cputype); 622 PMCLOG_DESPATCH(po); 623 624 return (0); 625 626 error: 627 /* shutdown the thread */ 628 if (po->po_kthread) 629 pmclog_stop_kthread(po); 630 631 KASSERT(po->po_kthread == NULL, ("[pmclog,%d] po=%p kthread not " 632 "stopped", __LINE__, po)); 633 634 if (po->po_file) 635 (void) fdrop(po->po_file, curthread); 636 po->po_file = NULL; /* clear file and error state */ 637 po->po_error = 0; 638 639 return (error); 640 } 641 642 643 /* 644 * De-configure a log file. This will throw away any buffers queued 645 * for this owner process. 646 */ 647 648 int 649 pmclog_deconfigure_log(struct pmc_owner *po) 650 { 651 int error; 652 struct pmclog_buffer *lb; 653 654 PMCDBG1(LOG,CFG,1, "de-config po=%p", po); 655 656 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) 657 return (EINVAL); 658 659 KASSERT(po->po_sscount == 0, 660 ("[pmclog,%d] po=%p still owning SS PMCs", __LINE__, po)); 661 KASSERT(po->po_file != NULL, 662 ("[pmclog,%d] po=%p no log file", __LINE__, po)); 663 664 /* stop the kthread, this will reset the 'OWNS_LOGFILE' flag */ 665 pmclog_stop_kthread(po); 666 667 KASSERT(po->po_kthread == NULL, 668 ("[pmclog,%d] po=%p kthread not stopped", __LINE__, po)); 669 670 /* return all queued log buffers to the global pool */ 671 while ((lb = TAILQ_FIRST(&po->po_logbuffers)) != NULL) { 672 TAILQ_REMOVE(&po->po_logbuffers, lb, plb_next); 673 PMCLOG_INIT_BUFFER_DESCRIPTOR(lb); 674 mtx_lock_spin(&pmc_bufferlist_mtx); 675 TAILQ_INSERT_HEAD(&pmc_bufferlist, lb, plb_next); 676 mtx_unlock_spin(&pmc_bufferlist_mtx); 677 } 678 679 /* return the 'current' buffer to the global pool */ 680 if ((lb = po->po_curbuf) != NULL) { 681 PMCLOG_INIT_BUFFER_DESCRIPTOR(lb); 682 mtx_lock_spin(&pmc_bufferlist_mtx); 683 TAILQ_INSERT_HEAD(&pmc_bufferlist, lb, plb_next); 684 mtx_unlock_spin(&pmc_bufferlist_mtx); 685 } 686 687 /* drop a reference to the fd */ 688 error = fdrop(po->po_file, curthread); 689 po->po_file = NULL; 690 po->po_error = 0; 691 692 return (error); 693 } 694 695 /* 696 * Flush a process' log buffer. 697 */ 698 699 int 700 pmclog_flush(struct pmc_owner *po) 701 { 702 int error; 703 struct pmclog_buffer *lb; 704 705 PMCDBG1(LOG,FLS,1, "po=%p", po); 706 707 /* 708 * If there is a pending error recorded by the logger thread, 709 * return that. 710 */ 711 if (po->po_error) 712 return (po->po_error); 713 714 error = 0; 715 716 /* 717 * Check that we do have an active log file. 718 */ 719 mtx_lock(&pmc_kthread_mtx); 720 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) { 721 error = EINVAL; 722 goto error; 723 } 724 725 /* 726 * Schedule the current buffer if any and not empty. 727 */ 728 mtx_lock_spin(&po->po_mtx); 729 lb = po->po_curbuf; 730 if (lb && lb->plb_ptr != lb->plb_base) { 731 pmclog_schedule_io(po); 732 } else 733 error = ENOBUFS; 734 mtx_unlock_spin(&po->po_mtx); 735 736 error: 737 mtx_unlock(&pmc_kthread_mtx); 738 739 return (error); 740 } 741 742 int 743 pmclog_close(struct pmc_owner *po) 744 { 745 746 PMCDBG1(LOG,CLO,1, "po=%p", po); 747 748 mtx_lock(&pmc_kthread_mtx); 749 750 /* 751 * Schedule the current buffer. 752 */ 753 mtx_lock_spin(&po->po_mtx); 754 if (po->po_curbuf) 755 pmclog_schedule_io(po); 756 else 757 wakeup_one(po); 758 mtx_unlock_spin(&po->po_mtx); 759 760 /* 761 * Initiate shutdown: no new data queued, 762 * thread will close file on last block. 763 */ 764 po->po_flags |= PMC_PO_SHUTDOWN; 765 766 mtx_unlock(&pmc_kthread_mtx); 767 768 return (0); 769 } 770 771 void 772 pmclog_process_callchain(struct pmc *pm, struct pmc_sample *ps) 773 { 774 int n, recordlen; 775 uint32_t flags; 776 struct pmc_owner *po; 777 778 PMCDBG3(LOG,SAM,1,"pm=%p pid=%d n=%d", pm, ps->ps_pid, 779 ps->ps_nsamples); 780 781 recordlen = offsetof(struct pmclog_callchain, pl_pc) + 782 ps->ps_nsamples * sizeof(uintfptr_t); 783 po = pm->pm_owner; 784 flags = PMC_CALLCHAIN_TO_CPUFLAGS(ps->ps_cpu,ps->ps_flags); 785 PMCLOG_RESERVE(po, CALLCHAIN, recordlen); 786 PMCLOG_EMIT32(ps->ps_pid); 787 PMCLOG_EMIT32(pm->pm_id); 788 PMCLOG_EMIT32(flags); 789 for (n = 0; n < ps->ps_nsamples; n++) 790 PMCLOG_EMITADDR(ps->ps_pc[n]); 791 PMCLOG_DESPATCH(po); 792 } 793 794 void 795 pmclog_process_closelog(struct pmc_owner *po) 796 { 797 PMCLOG_RESERVE(po,CLOSELOG,sizeof(struct pmclog_closelog)); 798 PMCLOG_DESPATCH(po); 799 } 800 801 void 802 pmclog_process_dropnotify(struct pmc_owner *po) 803 { 804 PMCLOG_RESERVE(po,DROPNOTIFY,sizeof(struct pmclog_dropnotify)); 805 PMCLOG_DESPATCH(po); 806 } 807 808 void 809 pmclog_process_map_in(struct pmc_owner *po, pid_t pid, uintfptr_t start, 810 const char *path) 811 { 812 int pathlen, recordlen; 813 814 KASSERT(path != NULL, ("[pmclog,%d] map-in, null path", __LINE__)); 815 816 pathlen = strlen(path) + 1; /* #bytes for path name */ 817 recordlen = offsetof(struct pmclog_map_in, pl_pathname) + 818 pathlen; 819 820 PMCLOG_RESERVE(po, MAP_IN, recordlen); 821 PMCLOG_EMIT32(pid); 822 PMCLOG_EMITADDR(start); 823 PMCLOG_EMITSTRING(path,pathlen); 824 PMCLOG_DESPATCH(po); 825 } 826 827 void 828 pmclog_process_map_out(struct pmc_owner *po, pid_t pid, uintfptr_t start, 829 uintfptr_t end) 830 { 831 KASSERT(start <= end, ("[pmclog,%d] start > end", __LINE__)); 832 833 PMCLOG_RESERVE(po, MAP_OUT, sizeof(struct pmclog_map_out)); 834 PMCLOG_EMIT32(pid); 835 PMCLOG_EMITADDR(start); 836 PMCLOG_EMITADDR(end); 837 PMCLOG_DESPATCH(po); 838 } 839 840 void 841 pmclog_process_pmcallocate(struct pmc *pm) 842 { 843 struct pmc_owner *po; 844 struct pmc_soft *ps; 845 846 po = pm->pm_owner; 847 848 PMCDBG1(LOG,ALL,1, "pm=%p", pm); 849 850 if (PMC_TO_CLASS(pm) == PMC_CLASS_SOFT) { 851 PMCLOG_RESERVE(po, PMCALLOCATEDYN, 852 sizeof(struct pmclog_pmcallocatedyn)); 853 PMCLOG_EMIT32(pm->pm_id); 854 PMCLOG_EMIT32(pm->pm_event); 855 PMCLOG_EMIT32(pm->pm_flags); 856 ps = pmc_soft_ev_acquire(pm->pm_event); 857 if (ps != NULL) 858 PMCLOG_EMITSTRING(ps->ps_ev.pm_ev_name,PMC_NAME_MAX); 859 else 860 PMCLOG_EMITNULLSTRING(PMC_NAME_MAX); 861 pmc_soft_ev_release(ps); 862 PMCLOG_DESPATCH(po); 863 } else { 864 PMCLOG_RESERVE(po, PMCALLOCATE, 865 sizeof(struct pmclog_pmcallocate)); 866 PMCLOG_EMIT32(pm->pm_id); 867 PMCLOG_EMIT32(pm->pm_event); 868 PMCLOG_EMIT32(pm->pm_flags); 869 PMCLOG_DESPATCH(po); 870 } 871 } 872 873 void 874 pmclog_process_pmcattach(struct pmc *pm, pid_t pid, char *path) 875 { 876 int pathlen, recordlen; 877 struct pmc_owner *po; 878 879 PMCDBG2(LOG,ATT,1,"pm=%p pid=%d", pm, pid); 880 881 po = pm->pm_owner; 882 883 pathlen = strlen(path) + 1; /* #bytes for the string */ 884 recordlen = offsetof(struct pmclog_pmcattach, pl_pathname) + pathlen; 885 886 PMCLOG_RESERVE(po, PMCATTACH, recordlen); 887 PMCLOG_EMIT32(pm->pm_id); 888 PMCLOG_EMIT32(pid); 889 PMCLOG_EMITSTRING(path, pathlen); 890 PMCLOG_DESPATCH(po); 891 } 892 893 void 894 pmclog_process_pmcdetach(struct pmc *pm, pid_t pid) 895 { 896 struct pmc_owner *po; 897 898 PMCDBG2(LOG,ATT,1,"!pm=%p pid=%d", pm, pid); 899 900 po = pm->pm_owner; 901 902 PMCLOG_RESERVE(po, PMCDETACH, sizeof(struct pmclog_pmcdetach)); 903 PMCLOG_EMIT32(pm->pm_id); 904 PMCLOG_EMIT32(pid); 905 PMCLOG_DESPATCH(po); 906 } 907 908 /* 909 * Log a context switch event to the log file. 910 */ 911 912 void 913 pmclog_process_proccsw(struct pmc *pm, struct pmc_process *pp, pmc_value_t v) 914 { 915 struct pmc_owner *po; 916 917 KASSERT(pm->pm_flags & PMC_F_LOG_PROCCSW, 918 ("[pmclog,%d] log-process-csw called gratuitously", __LINE__)); 919 920 PMCDBG3(LOG,SWO,1,"pm=%p pid=%d v=%jx", pm, pp->pp_proc->p_pid, 921 v); 922 923 po = pm->pm_owner; 924 925 PMCLOG_RESERVE(po, PROCCSW, sizeof(struct pmclog_proccsw)); 926 PMCLOG_EMIT32(pm->pm_id); 927 PMCLOG_EMIT64(v); 928 PMCLOG_EMIT32(pp->pp_proc->p_pid); 929 PMCLOG_DESPATCH(po); 930 } 931 932 void 933 pmclog_process_procexec(struct pmc_owner *po, pmc_id_t pmid, pid_t pid, 934 uintfptr_t startaddr, char *path) 935 { 936 int pathlen, recordlen; 937 938 PMCDBG3(LOG,EXC,1,"po=%p pid=%d path=\"%s\"", po, pid, path); 939 940 pathlen = strlen(path) + 1; /* #bytes for the path */ 941 recordlen = offsetof(struct pmclog_procexec, pl_pathname) + pathlen; 942 943 PMCLOG_RESERVE(po, PROCEXEC, recordlen); 944 PMCLOG_EMIT32(pid); 945 PMCLOG_EMITADDR(startaddr); 946 PMCLOG_EMIT32(pmid); 947 PMCLOG_EMITSTRING(path,pathlen); 948 PMCLOG_DESPATCH(po); 949 } 950 951 /* 952 * Log a process exit event (and accumulated pmc value) to the log file. 953 */ 954 955 void 956 pmclog_process_procexit(struct pmc *pm, struct pmc_process *pp) 957 { 958 int ri; 959 struct pmc_owner *po; 960 961 ri = PMC_TO_ROWINDEX(pm); 962 PMCDBG3(LOG,EXT,1,"pm=%p pid=%d v=%jx", pm, pp->pp_proc->p_pid, 963 pp->pp_pmcs[ri].pp_pmcval); 964 965 po = pm->pm_owner; 966 967 PMCLOG_RESERVE(po, PROCEXIT, sizeof(struct pmclog_procexit)); 968 PMCLOG_EMIT32(pm->pm_id); 969 PMCLOG_EMIT64(pp->pp_pmcs[ri].pp_pmcval); 970 PMCLOG_EMIT32(pp->pp_proc->p_pid); 971 PMCLOG_DESPATCH(po); 972 } 973 974 /* 975 * Log a fork event. 976 */ 977 978 void 979 pmclog_process_procfork(struct pmc_owner *po, pid_t oldpid, pid_t newpid) 980 { 981 PMCLOG_RESERVE(po, PROCFORK, sizeof(struct pmclog_procfork)); 982 PMCLOG_EMIT32(oldpid); 983 PMCLOG_EMIT32(newpid); 984 PMCLOG_DESPATCH(po); 985 } 986 987 /* 988 * Log a process exit event of the form suitable for system-wide PMCs. 989 */ 990 991 void 992 pmclog_process_sysexit(struct pmc_owner *po, pid_t pid) 993 { 994 PMCLOG_RESERVE(po, SYSEXIT, sizeof(struct pmclog_sysexit)); 995 PMCLOG_EMIT32(pid); 996 PMCLOG_DESPATCH(po); 997 } 998 999 /* 1000 * Write a user log entry. 1001 */ 1002 1003 int 1004 pmclog_process_userlog(struct pmc_owner *po, struct pmc_op_writelog *wl) 1005 { 1006 int error; 1007 1008 PMCDBG2(LOG,WRI,1, "writelog po=%p ud=0x%x", po, wl->pm_userdata); 1009 1010 error = 0; 1011 1012 PMCLOG_RESERVE_WITH_ERROR(po, USERDATA, 1013 sizeof(struct pmclog_userdata)); 1014 PMCLOG_EMIT32(wl->pm_userdata); 1015 PMCLOG_DESPATCH(po); 1016 1017 error: 1018 return (error); 1019 } 1020 1021 /* 1022 * Initialization. 1023 * 1024 * Create a pool of log buffers and initialize mutexes. 1025 */ 1026 1027 void 1028 pmclog_initialize() 1029 { 1030 int n; 1031 struct pmclog_buffer *plb; 1032 1033 if (pmclog_buffer_size <= 0) { 1034 (void) printf("hwpmc: tunable logbuffersize=%d must be " 1035 "greater than zero.\n", pmclog_buffer_size); 1036 pmclog_buffer_size = PMC_LOG_BUFFER_SIZE; 1037 } 1038 1039 if (pmc_nlogbuffers <= 0) { 1040 (void) printf("hwpmc: tunable nlogbuffers=%d must be greater " 1041 "than zero.\n", pmc_nlogbuffers); 1042 pmc_nlogbuffers = PMC_NLOGBUFFERS; 1043 } 1044 1045 /* create global pool of log buffers */ 1046 for (n = 0; n < pmc_nlogbuffers; n++) { 1047 plb = malloc(1024 * pmclog_buffer_size, M_PMC, 1048 M_WAITOK|M_ZERO); 1049 PMCLOG_INIT_BUFFER_DESCRIPTOR(plb); 1050 TAILQ_INSERT_HEAD(&pmc_bufferlist, plb, plb_next); 1051 } 1052 mtx_init(&pmc_bufferlist_mtx, "pmc-buffer-list", "pmc-leaf", 1053 MTX_SPIN); 1054 mtx_init(&pmc_kthread_mtx, "pmc-kthread", "pmc-sleep", MTX_DEF); 1055 } 1056 1057 /* 1058 * Shutdown logging. 1059 * 1060 * Destroy mutexes and release memory back the to free pool. 1061 */ 1062 1063 void 1064 pmclog_shutdown() 1065 { 1066 struct pmclog_buffer *plb; 1067 1068 mtx_destroy(&pmc_kthread_mtx); 1069 mtx_destroy(&pmc_bufferlist_mtx); 1070 1071 while ((plb = TAILQ_FIRST(&pmc_bufferlist)) != NULL) { 1072 TAILQ_REMOVE(&pmc_bufferlist, plb, plb_next); 1073 free(plb, M_PMC); 1074 } 1075 } 1076