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#if (__FreeBSD_version >= 1100000) 41#include <sys/capsicum.h> 42#else 43#include <sys/capability.h> 44#endif 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/signalvar.h> 56#include <sys/sysctl.h> 57#include <sys/systm.h> 58#include <sys/uio.h> 59#include <sys/unistd.h> 60#include <sys/vnode.h> 61 62/* 63 * Sysctl tunables 64 */ 65 66SYSCTL_DECL(_kern_hwpmc); 67 68/* 69 * kern.hwpmc.logbuffersize -- size of the per-cpu owner buffers. 70 */ 71 72static int pmclog_buffer_size = PMC_LOG_BUFFER_SIZE; 73SYSCTL_INT(_kern_hwpmc, OID_AUTO, logbuffersize, CTLFLAG_RDTUN, 74 &pmclog_buffer_size, 0, "size of log buffers in kilobytes"); 75 76/* 77 * kern.hwpmc.nbuffer -- number of global log buffers 78 */ 79 80static int pmc_nlogbuffers = PMC_NLOGBUFFERS; 81SYSCTL_INT(_kern_hwpmc, OID_AUTO, nbuffers, CTLFLAG_RDTUN, 82 &pmc_nlogbuffers, 0, "number of global log buffers"); 83 84/* 85 * Global log buffer list and associated spin lock. 86 */ 87 88TAILQ_HEAD(, pmclog_buffer) pmc_bufferlist = 89 TAILQ_HEAD_INITIALIZER(pmc_bufferlist); 90static struct mtx pmc_bufferlist_mtx; /* spin lock */ 91static struct mtx pmc_kthread_mtx; /* sleep lock */ 92 93#define PMCLOG_INIT_BUFFER_DESCRIPTOR(D) do { \ 94 const int __roundup = roundup(sizeof(*D), \ 95 sizeof(uint32_t)); \ 96 (D)->plb_fence = ((char *) (D)) + \ 97 1024*pmclog_buffer_size; \ 98 (D)->plb_base = (D)->plb_ptr = ((char *) (D)) + \ 99 __roundup; \ 100 } while (0) 101 102 103/* 104 * Log file record constructors. 105 */ 106#define _PMCLOG_TO_HEADER(T,L) \ 107 ((PMCLOG_HEADER_MAGIC << 24) | \ 108 (PMCLOG_TYPE_ ## T << 16) | \ 109 ((L) & 0xFFFF)) 110 111/* reserve LEN bytes of space and initialize the entry header */ 112#define _PMCLOG_RESERVE(PO,TYPE,LEN,ACTION) do { \ 113 uint32_t *_le; \ 114 int _len = roundup((LEN), sizeof(uint32_t)); \ 115 if ((_le = pmclog_reserve((PO), _len)) == NULL) { \ 116 ACTION; \ 117 } \ 118 *_le = _PMCLOG_TO_HEADER(TYPE,_len); \ 119 _le += 3 /* skip over timestamp */ 120 121#define PMCLOG_RESERVE(P,T,L) _PMCLOG_RESERVE(P,T,L,return) 122#define PMCLOG_RESERVE_WITH_ERROR(P,T,L) _PMCLOG_RESERVE(P,T,L, \ 123 error=ENOMEM;goto error) 124 125#define PMCLOG_EMIT32(V) do { *_le++ = (V); } while (0) 126#define PMCLOG_EMIT64(V) do { \ 127 *_le++ = (uint32_t) ((V) & 0xFFFFFFFF); \ 128 *_le++ = (uint32_t) (((V) >> 32) & 0xFFFFFFFF); \ 129 } while (0) 130 131 132/* Emit a string. Caution: does NOT update _le, so needs to be last */ 133#define PMCLOG_EMITSTRING(S,L) do { bcopy((S), _le, (L)); } while (0) 134#define PMCLOG_EMITNULLSTRING(L) do { bzero(_le, (L)); } while (0) 135 136#define PMCLOG_DESPATCH(PO) \ 137 pmclog_release((PO)); \ 138 } while (0) 139 140 141/* 142 * Assertions about the log file format. 143 */ 144 145CTASSERT(sizeof(struct pmclog_callchain) == 6*4 + 146 PMC_CALLCHAIN_DEPTH_MAX*sizeof(uintfptr_t)); 147CTASSERT(sizeof(struct pmclog_closelog) == 3*4); 148CTASSERT(sizeof(struct pmclog_dropnotify) == 3*4); 149CTASSERT(sizeof(struct pmclog_map_in) == PATH_MAX + 150 4*4 + sizeof(uintfptr_t)); 151CTASSERT(offsetof(struct pmclog_map_in,pl_pathname) == 152 4*4 + sizeof(uintfptr_t)); 153CTASSERT(sizeof(struct pmclog_map_out) == 4*4 + 2*sizeof(uintfptr_t)); 154CTASSERT(sizeof(struct pmclog_pcsample) == 6*4 + sizeof(uintfptr_t)); 155CTASSERT(sizeof(struct pmclog_pmcallocate) == 6*4); 156CTASSERT(sizeof(struct pmclog_pmcattach) == 5*4 + PATH_MAX); 157CTASSERT(offsetof(struct pmclog_pmcattach,pl_pathname) == 5*4); 158CTASSERT(sizeof(struct pmclog_pmcdetach) == 5*4); 159CTASSERT(sizeof(struct pmclog_proccsw) == 5*4 + 8); 160CTASSERT(sizeof(struct pmclog_procexec) == 5*4 + PATH_MAX + 161 sizeof(uintfptr_t)); 162CTASSERT(offsetof(struct pmclog_procexec,pl_pathname) == 5*4 + 163 sizeof(uintfptr_t)); 164CTASSERT(sizeof(struct pmclog_procexit) == 5*4 + 8); 165CTASSERT(sizeof(struct pmclog_procfork) == 5*4); 166CTASSERT(sizeof(struct pmclog_sysexit) == 4*4); 167CTASSERT(sizeof(struct pmclog_userdata) == 4*4); 168 169/* 170 * Log buffer structure 171 */ 172 173struct pmclog_buffer { 174 TAILQ_ENTRY(pmclog_buffer) plb_next; 175 char *plb_base; 176 char *plb_ptr; 177 char *plb_fence; 178}; 179 180/* 181 * Prototypes 182 */ 183 184static int pmclog_get_buffer(struct pmc_owner *po); 185static void pmclog_loop(void *arg); 186static void pmclog_release(struct pmc_owner *po); 187static uint32_t *pmclog_reserve(struct pmc_owner *po, int length); 188static void pmclog_schedule_io(struct pmc_owner *po); 189static void pmclog_stop_kthread(struct pmc_owner *po); 190 191/* 192 * Helper functions 193 */ 194 195/* 196 * Get a log buffer 197 */ 198 199static int 200pmclog_get_buffer(struct pmc_owner *po) 201{ 202 struct pmclog_buffer *plb; 203 204 mtx_assert(&po->po_mtx, MA_OWNED); 205 206 KASSERT(po->po_curbuf == NULL, 207 ("[pmclog,%d] po=%p current buffer still valid", __LINE__, po)); 208 209 mtx_lock_spin(&pmc_bufferlist_mtx); 210 if ((plb = TAILQ_FIRST(&pmc_bufferlist)) != NULL) 211 TAILQ_REMOVE(&pmc_bufferlist, plb, plb_next); 212 mtx_unlock_spin(&pmc_bufferlist_mtx); 213 214 PMCDBG2(LOG,GTB,1, "po=%p plb=%p", po, plb); 215 216#ifdef HWPMC_DEBUG 217 if (plb) 218 KASSERT(plb->plb_ptr == plb->plb_base && 219 plb->plb_base < plb->plb_fence, 220 ("[pmclog,%d] po=%p buffer invariants: ptr=%p " 221 "base=%p fence=%p", __LINE__, po, plb->plb_ptr, 222 plb->plb_base, plb->plb_fence)); 223#endif 224 225 po->po_curbuf = plb; 226 227 /* update stats */ 228 atomic_add_int(&pmc_stats.pm_buffer_requests, 1); 229 if (plb == NULL) 230 atomic_add_int(&pmc_stats.pm_buffer_requests_failed, 1); 231 232 return (plb ? 0 : ENOMEM); 233} 234 235/* 236 * Log handler loop. 237 * 238 * This function is executed by each pmc owner's helper thread. 239 */ 240 241static void 242pmclog_loop(void *arg) 243{ 244 int error; 245 struct pmc_owner *po; 246 struct pmclog_buffer *lb; 247 struct proc *p; 248 struct ucred *ownercred; 249 struct ucred *mycred; 250 struct thread *td; 251 struct uio auio; 252 struct iovec aiov; 253 size_t nbytes; 254 255 po = (struct pmc_owner *) arg; 256 p = po->po_owner; 257 td = curthread; 258 mycred = td->td_ucred; 259 260 PROC_LOCK(p); 261 ownercred = crhold(p->p_ucred); 262 PROC_UNLOCK(p); 263 264 PMCDBG2(LOG,INI,1, "po=%p kt=%p", po, po->po_kthread); 265 KASSERT(po->po_kthread == curthread->td_proc, 266 ("[pmclog,%d] proc mismatch po=%p po/kt=%p curproc=%p", __LINE__, 267 po, po->po_kthread, curthread->td_proc)); 268 269 lb = NULL; 270 271 272 /* 273 * Loop waiting for I/O requests to be added to the owner 274 * struct's queue. The loop is exited when the log file 275 * is deconfigured. 276 */ 277 278 mtx_lock(&pmc_kthread_mtx); 279 280 for (;;) { 281 282 /* check if we've been asked to exit */ 283 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) 284 break; 285 286 if (lb == NULL) { /* look for a fresh buffer to write */ 287 mtx_lock_spin(&po->po_mtx); 288 if ((lb = TAILQ_FIRST(&po->po_logbuffers)) == NULL) { 289 mtx_unlock_spin(&po->po_mtx); 290 291 /* No more buffers and shutdown required. */ 292 if (po->po_flags & PMC_PO_SHUTDOWN) { 293 mtx_unlock(&pmc_kthread_mtx); 294 /* 295 * Close the file to get PMCLOG_EOF 296 * error in pmclog(3). 297 */ 298 fo_close(po->po_file, curthread); 299 mtx_lock(&pmc_kthread_mtx); 300 break; 301 } 302 303 (void) msleep(po, &pmc_kthread_mtx, PWAIT, 304 "pmcloop", 0); 305 continue; 306 } 307 308 TAILQ_REMOVE(&po->po_logbuffers, lb, plb_next); 309 mtx_unlock_spin(&po->po_mtx); 310 } 311 312 mtx_unlock(&pmc_kthread_mtx); 313 314 /* process the request */ 315 PMCDBG3(LOG,WRI,2, "po=%p base=%p ptr=%p", po, 316 lb->plb_base, lb->plb_ptr); 317 /* change our thread's credentials before issuing the I/O */ 318 319 aiov.iov_base = lb->plb_base; 320 aiov.iov_len = nbytes = lb->plb_ptr - lb->plb_base; 321 322 auio.uio_iov = &aiov; 323 auio.uio_iovcnt = 1; 324 auio.uio_offset = -1; 325 auio.uio_resid = nbytes; 326 auio.uio_rw = UIO_WRITE; 327 auio.uio_segflg = UIO_SYSSPACE; 328 auio.uio_td = td; 329 330 /* switch thread credentials -- see kern_ktrace.c */ 331 td->td_ucred = ownercred; 332 error = fo_write(po->po_file, &auio, ownercred, 0, td); 333 td->td_ucred = mycred; 334 335 if (error) { 336 /* XXX some errors are recoverable */ 337 /* send a SIGIO to the owner and exit */ 338 PROC_LOCK(p); 339 kern_psignal(p, SIGIO); 340 PROC_UNLOCK(p); 341 342 mtx_lock(&pmc_kthread_mtx); 343 344 po->po_error = error; /* save for flush log */ 345 346 PMCDBG2(LOG,WRI,2, "po=%p error=%d", po, error); 347 348 break; 349 } 350 351 mtx_lock(&pmc_kthread_mtx); 352 353 /* put the used buffer back into the global pool */ 354 PMCLOG_INIT_BUFFER_DESCRIPTOR(lb); 355 356 mtx_lock_spin(&pmc_bufferlist_mtx); 357 TAILQ_INSERT_HEAD(&pmc_bufferlist, lb, plb_next); 358 mtx_unlock_spin(&pmc_bufferlist_mtx); 359 360 lb = NULL; 361 } 362 363 wakeup_one(po->po_kthread); 364 po->po_kthread = NULL; 365 366 mtx_unlock(&pmc_kthread_mtx); 367 368 /* return the current I/O buffer to the global pool */ 369 if (lb) { 370 PMCLOG_INIT_BUFFER_DESCRIPTOR(lb); 371 372 mtx_lock_spin(&pmc_bufferlist_mtx); 373 TAILQ_INSERT_HEAD(&pmc_bufferlist, lb, plb_next); 374 mtx_unlock_spin(&pmc_bufferlist_mtx); 375 } 376 377 /* 378 * Exit this thread, signalling the waiter 379 */ 380 381 crfree(ownercred); 382 383 kproc_exit(0); 384} 385 386/* 387 * Release and log entry and schedule an I/O if needed. 388 */ 389 390static void 391pmclog_release(struct pmc_owner *po) 392{ 393 KASSERT(po->po_curbuf->plb_ptr >= po->po_curbuf->plb_base, 394 ("[pmclog,%d] buffer invariants po=%p ptr=%p base=%p", __LINE__, 395 po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_base)); 396 KASSERT(po->po_curbuf->plb_ptr <= po->po_curbuf->plb_fence, 397 ("[pmclog,%d] buffer invariants po=%p ptr=%p fenc=%p", __LINE__, 398 po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_fence)); 399 400 /* schedule an I/O if we've filled a buffer */ 401 if (po->po_curbuf->plb_ptr >= po->po_curbuf->plb_fence) 402 pmclog_schedule_io(po); 403 404 mtx_unlock_spin(&po->po_mtx); 405 406 PMCDBG1(LOG,REL,1, "po=%p", po); 407} 408 409 410/* 411 * Attempt to reserve 'length' bytes of space in an owner's log 412 * buffer. The function returns a pointer to 'length' bytes of space 413 * if there was enough space or returns NULL if no space was 414 * available. Non-null returns do so with the po mutex locked. The 415 * caller must invoke pmclog_release() on the pmc owner structure 416 * when done. 417 */ 418 419static uint32_t * 420pmclog_reserve(struct pmc_owner *po, int length) 421{ 422 uintptr_t newptr, oldptr; 423 uint32_t *lh; 424 struct timespec ts; 425 426 PMCDBG2(LOG,ALL,1, "po=%p len=%d", po, length); 427 428 KASSERT(length % sizeof(uint32_t) == 0, 429 ("[pmclog,%d] length not a multiple of word size", __LINE__)); 430 431 mtx_lock_spin(&po->po_mtx); 432 433 /* No more data when shutdown in progress. */ 434 if (po->po_flags & PMC_PO_SHUTDOWN) { 435 mtx_unlock_spin(&po->po_mtx); 436 return (NULL); 437 } 438 439 if (po->po_curbuf == NULL) 440 if (pmclog_get_buffer(po) != 0) { 441 mtx_unlock_spin(&po->po_mtx); 442 return (NULL); 443 } 444 445 KASSERT(po->po_curbuf != NULL, 446 ("[pmclog,%d] po=%p no current buffer", __LINE__, po)); 447 448 KASSERT(po->po_curbuf->plb_ptr >= po->po_curbuf->plb_base && 449 po->po_curbuf->plb_ptr <= po->po_curbuf->plb_fence, 450 ("[pmclog,%d] po=%p buffer invariants: ptr=%p base=%p fence=%p", 451 __LINE__, po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_base, 452 po->po_curbuf->plb_fence)); 453 454 oldptr = (uintptr_t) po->po_curbuf->plb_ptr; 455 newptr = oldptr + length; 456 457 KASSERT(oldptr != (uintptr_t) NULL, 458 ("[pmclog,%d] po=%p Null log buffer pointer", __LINE__, po)); 459 460 /* 461 * If we have space in the current buffer, return a pointer to 462 * available space with the PO structure locked. 463 */ 464 if (newptr <= (uintptr_t) po->po_curbuf->plb_fence) { 465 po->po_curbuf->plb_ptr = (char *) newptr; 466 goto done; 467 } 468 469 /* 470 * Otherwise, schedule the current buffer for output and get a 471 * fresh buffer. 472 */ 473 pmclog_schedule_io(po); 474 475 if (pmclog_get_buffer(po) != 0) { 476 mtx_unlock_spin(&po->po_mtx); 477 return (NULL); 478 } 479 480 KASSERT(po->po_curbuf != NULL, 481 ("[pmclog,%d] po=%p no current buffer", __LINE__, po)); 482 483 KASSERT(po->po_curbuf->plb_ptr != NULL, 484 ("[pmclog,%d] null return from pmc_get_log_buffer", __LINE__)); 485 486 KASSERT(po->po_curbuf->plb_ptr == po->po_curbuf->plb_base && 487 po->po_curbuf->plb_ptr <= po->po_curbuf->plb_fence, 488 ("[pmclog,%d] po=%p buffer invariants: ptr=%p base=%p fence=%p", 489 __LINE__, po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_base, 490 po->po_curbuf->plb_fence)); 491 492 oldptr = (uintptr_t) po->po_curbuf->plb_ptr; 493 494 done: 495 lh = (uint32_t *) oldptr; 496 lh++; /* skip header */ 497 getnanotime(&ts); /* fill in the timestamp */ 498 *lh++ = ts.tv_sec & 0xFFFFFFFF; 499 *lh++ = ts.tv_nsec & 0xFFFFFFF; 500 return ((uint32_t *) oldptr); 501} 502 503/* 504 * Schedule an I/O. 505 * 506 * Transfer the current buffer to the helper kthread. 507 */ 508 509static void 510pmclog_schedule_io(struct pmc_owner *po) 511{ 512 KASSERT(po->po_curbuf != NULL, 513 ("[pmclog,%d] schedule_io with null buffer po=%p", __LINE__, po)); 514 515 KASSERT(po->po_curbuf->plb_ptr >= po->po_curbuf->plb_base, 516 ("[pmclog,%d] buffer invariants po=%p ptr=%p base=%p", __LINE__, 517 po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_base)); 518 KASSERT(po->po_curbuf->plb_ptr <= po->po_curbuf->plb_fence, 519 ("[pmclog,%d] buffer invariants po=%p ptr=%p fenc=%p", __LINE__, 520 po, po->po_curbuf->plb_ptr, po->po_curbuf->plb_fence)); 521 522 PMCDBG1(LOG,SIO, 1, "po=%p", po); 523 524 mtx_assert(&po->po_mtx, MA_OWNED); 525 526 /* 527 * Add the current buffer to the tail of the buffer list and 528 * wakeup the helper. 529 */ 530 TAILQ_INSERT_TAIL(&po->po_logbuffers, po->po_curbuf, plb_next); 531 po->po_curbuf = NULL; 532 wakeup_one(po); 533} 534 535/* 536 * Stop the helper kthread. 537 */ 538 539static void 540pmclog_stop_kthread(struct pmc_owner *po) 541{ 542 /* 543 * Close the file to force the thread out of fo_write, 544 * unset flag, wakeup the helper thread, 545 * wait for it to exit 546 */ 547 548 if (po->po_file != NULL) 549 fo_close(po->po_file, curthread); 550 551 mtx_lock(&pmc_kthread_mtx); 552 po->po_flags &= ~PMC_PO_OWNS_LOGFILE; 553 wakeup_one(po); 554 if (po->po_kthread) 555 msleep(po->po_kthread, &pmc_kthread_mtx, PPAUSE, "pmckstp", 0); 556 mtx_unlock(&pmc_kthread_mtx); 557} 558 559/* 560 * Public functions 561 */ 562 563/* 564 * Configure a log file for pmc owner 'po'. 565 * 566 * Parameter 'logfd' is a file handle referencing an open file in the 567 * owner process. This file needs to have been opened for writing. 568 */ 569 570int 571pmclog_configure_log(struct pmc_mdep *md, struct pmc_owner *po, int logfd) 572{ 573 int error; 574 struct proc *p; 575 cap_rights_t rights; 576 /* 577 * As long as it is possible to get a LOR between pmc_sx lock and 578 * proctree/allproc sx locks used for adding a new process, assure 579 * the former is not held here. 580 */ 581 sx_assert(&pmc_sx, SA_UNLOCKED); 582 PMCDBG2(LOG,CFG,1, "config po=%p logfd=%d", po, logfd); 583 584 p = po->po_owner; 585 586 /* return EBUSY if a log file was already present */ 587 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 588 return (EBUSY); 589 590 KASSERT(po->po_kthread == NULL, 591 ("[pmclog,%d] po=%p kthread (%p) already present", __LINE__, po, 592 po->po_kthread)); 593 KASSERT(po->po_file == NULL, 594 ("[pmclog,%d] po=%p file (%p) already present", __LINE__, po, 595 po->po_file)); 596 597 /* get a reference to the file state */ 598 error = fget_write(curthread, logfd, 599 cap_rights_init(&rights, CAP_WRITE), &po->po_file); 600 if (error) 601 goto error; 602 603 /* mark process as owning a log file */ 604 po->po_flags |= PMC_PO_OWNS_LOGFILE; 605 error = kproc_create(pmclog_loop, po, &po->po_kthread, 606 RFHIGHPID, 0, "hwpmc: proc(%d)", p->p_pid); 607 if (error) 608 goto error; 609 610 /* mark process as using HWPMCs */ 611 PROC_LOCK(p); 612 p->p_flag |= P_HWPMC; 613 PROC_UNLOCK(p); 614 615 /* create a log initialization entry */ 616 PMCLOG_RESERVE_WITH_ERROR(po, INITIALIZE, 617 sizeof(struct pmclog_initialize)); 618 PMCLOG_EMIT32(PMC_VERSION); 619 PMCLOG_EMIT32(md->pmd_cputype); 620 PMCLOG_DESPATCH(po); 621 622 return (0); 623 624 error: 625 /* shutdown the thread */ 626 if (po->po_kthread) 627 pmclog_stop_kthread(po); 628 629 KASSERT(po->po_kthread == NULL, ("[pmclog,%d] po=%p kthread not " 630 "stopped", __LINE__, po)); 631 632 if (po->po_file) 633 (void) fdrop(po->po_file, curthread); 634 po->po_file = NULL; /* clear file and error state */ 635 po->po_error = 0; 636 637 return (error); 638} 639 640 641/* 642 * De-configure a log file. This will throw away any buffers queued 643 * for this owner process. 644 */ 645 646int 647pmclog_deconfigure_log(struct pmc_owner *po) 648{ 649 int error; 650 struct pmclog_buffer *lb; 651 652 PMCDBG1(LOG,CFG,1, "de-config po=%p", po); 653 654 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) 655 return (EINVAL); 656 657 KASSERT(po->po_sscount == 0, 658 ("[pmclog,%d] po=%p still owning SS PMCs", __LINE__, po)); 659 KASSERT(po->po_file != NULL, 660 ("[pmclog,%d] po=%p no log file", __LINE__, po)); 661 662 /* stop the kthread, this will reset the 'OWNS_LOGFILE' flag */ 663 pmclog_stop_kthread(po); 664 665 KASSERT(po->po_kthread == NULL, 666 ("[pmclog,%d] po=%p kthread not stopped", __LINE__, po)); 667 668 /* return all queued log buffers to the global pool */ 669 while ((lb = TAILQ_FIRST(&po->po_logbuffers)) != NULL) { 670 TAILQ_REMOVE(&po->po_logbuffers, lb, plb_next); 671 PMCLOG_INIT_BUFFER_DESCRIPTOR(lb); 672 mtx_lock_spin(&pmc_bufferlist_mtx); 673 TAILQ_INSERT_HEAD(&pmc_bufferlist, lb, plb_next); 674 mtx_unlock_spin(&pmc_bufferlist_mtx); 675 } 676 677 /* return the 'current' buffer to the global pool */ 678 if ((lb = po->po_curbuf) != NULL) { 679 PMCLOG_INIT_BUFFER_DESCRIPTOR(lb); 680 mtx_lock_spin(&pmc_bufferlist_mtx); 681 TAILQ_INSERT_HEAD(&pmc_bufferlist, lb, plb_next); 682 mtx_unlock_spin(&pmc_bufferlist_mtx); 683 } 684 685 /* drop a reference to the fd */ 686 error = fdrop(po->po_file, curthread); 687 po->po_file = NULL; 688 po->po_error = 0; 689 690 return (error); 691} 692 693/* 694 * Flush a process' log buffer. 695 */ 696 697int 698pmclog_flush(struct pmc_owner *po) 699{ 700 int error; 701 struct pmclog_buffer *lb; 702 703 PMCDBG1(LOG,FLS,1, "po=%p", po); 704 705 /* 706 * If there is a pending error recorded by the logger thread, 707 * return that. 708 */ 709 if (po->po_error) 710 return (po->po_error); 711 712 error = 0; 713 714 /* 715 * Check that we do have an active log file. 716 */ 717 mtx_lock(&pmc_kthread_mtx); 718 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) { 719 error = EINVAL; 720 goto error; 721 } 722 723 /* 724 * Schedule the current buffer if any and not empty. 725 */ 726 mtx_lock_spin(&po->po_mtx); 727 lb = po->po_curbuf; 728 if (lb && lb->plb_ptr != lb->plb_base) { 729 pmclog_schedule_io(po); 730 } else 731 error = ENOBUFS; 732 mtx_unlock_spin(&po->po_mtx); 733 734 error: 735 mtx_unlock(&pmc_kthread_mtx); 736 737 return (error); 738} 739 740int 741pmclog_close(struct pmc_owner *po) 742{ 743 744 PMCDBG1(LOG,CLO,1, "po=%p", po); 745 746 mtx_lock(&pmc_kthread_mtx); 747 748 /* 749 * Schedule the current buffer. 750 */ 751 mtx_lock_spin(&po->po_mtx); 752 if (po->po_curbuf) 753 pmclog_schedule_io(po); 754 else 755 wakeup_one(po); 756 mtx_unlock_spin(&po->po_mtx); 757 758 /* 759 * Initiate shutdown: no new data queued, 760 * thread will close file on last block. 761 */ 762 po->po_flags |= PMC_PO_SHUTDOWN; 763 764 mtx_unlock(&pmc_kthread_mtx); 765 766 return (0); 767} 768 769void 770pmclog_process_callchain(struct pmc *pm, struct pmc_sample *ps) 771{ 772 int n, recordlen; 773 uint32_t flags; 774 struct pmc_owner *po; 775 776 PMCDBG3(LOG,SAM,1,"pm=%p pid=%d n=%d", pm, ps->ps_pid, 777 ps->ps_nsamples); 778 779 recordlen = offsetof(struct pmclog_callchain, pl_pc) + 780 ps->ps_nsamples * sizeof(uintfptr_t); 781 po = pm->pm_owner; 782 flags = PMC_CALLCHAIN_TO_CPUFLAGS(ps->ps_cpu,ps->ps_flags); 783 PMCLOG_RESERVE(po, CALLCHAIN, recordlen); 784 PMCLOG_EMIT32(ps->ps_pid); 785 PMCLOG_EMIT32(pm->pm_id); 786 PMCLOG_EMIT32(flags); 787 for (n = 0; n < ps->ps_nsamples; n++) 788 PMCLOG_EMITADDR(ps->ps_pc[n]); 789 PMCLOG_DESPATCH(po); 790} 791 792void 793pmclog_process_closelog(struct pmc_owner *po) 794{ 795 PMCLOG_RESERVE(po,CLOSELOG,sizeof(struct pmclog_closelog)); 796 PMCLOG_DESPATCH(po); 797} 798 799void 800pmclog_process_dropnotify(struct pmc_owner *po) 801{ 802 PMCLOG_RESERVE(po,DROPNOTIFY,sizeof(struct pmclog_dropnotify)); 803 PMCLOG_DESPATCH(po); 804} 805 806void 807pmclog_process_map_in(struct pmc_owner *po, pid_t pid, uintfptr_t start, 808 const char *path) 809{ 810 int pathlen, recordlen; 811 812 KASSERT(path != NULL, ("[pmclog,%d] map-in, null path", __LINE__)); 813 814 pathlen = strlen(path) + 1; /* #bytes for path name */ 815 recordlen = offsetof(struct pmclog_map_in, pl_pathname) + 816 pathlen; 817 818 PMCLOG_RESERVE(po, MAP_IN, recordlen); 819 PMCLOG_EMIT32(pid); 820 PMCLOG_EMITADDR(start); 821 PMCLOG_EMITSTRING(path,pathlen); 822 PMCLOG_DESPATCH(po); 823} 824 825void 826pmclog_process_map_out(struct pmc_owner *po, pid_t pid, uintfptr_t start, 827 uintfptr_t end) 828{ 829 KASSERT(start <= end, ("[pmclog,%d] start > end", __LINE__)); 830 831 PMCLOG_RESERVE(po, MAP_OUT, sizeof(struct pmclog_map_out)); 832 PMCLOG_EMIT32(pid); 833 PMCLOG_EMITADDR(start); 834 PMCLOG_EMITADDR(end); 835 PMCLOG_DESPATCH(po); 836} 837 838void 839pmclog_process_pmcallocate(struct pmc *pm) 840{ 841 struct pmc_owner *po; 842 struct pmc_soft *ps; 843 844 po = pm->pm_owner; 845 846 PMCDBG1(LOG,ALL,1, "pm=%p", pm); 847 848 if (PMC_TO_CLASS(pm) == PMC_CLASS_SOFT) { 849 PMCLOG_RESERVE(po, PMCALLOCATEDYN, 850 sizeof(struct pmclog_pmcallocatedyn)); 851 PMCLOG_EMIT32(pm->pm_id); 852 PMCLOG_EMIT32(pm->pm_event); 853 PMCLOG_EMIT32(pm->pm_flags); 854 ps = pmc_soft_ev_acquire(pm->pm_event); 855 if (ps != NULL) 856 PMCLOG_EMITSTRING(ps->ps_ev.pm_ev_name,PMC_NAME_MAX); 857 else 858 PMCLOG_EMITNULLSTRING(PMC_NAME_MAX); 859 pmc_soft_ev_release(ps); 860 PMCLOG_DESPATCH(po); 861 } else { 862 PMCLOG_RESERVE(po, PMCALLOCATE, 863 sizeof(struct pmclog_pmcallocate)); 864 PMCLOG_EMIT32(pm->pm_id); 865 PMCLOG_EMIT32(pm->pm_event); 866 PMCLOG_EMIT32(pm->pm_flags); 867 PMCLOG_DESPATCH(po); 868 } 869} 870 871void 872pmclog_process_pmcattach(struct pmc *pm, pid_t pid, char *path) 873{ 874 int pathlen, recordlen; 875 struct pmc_owner *po; 876 877 PMCDBG2(LOG,ATT,1,"pm=%p pid=%d", pm, pid); 878 879 po = pm->pm_owner; 880 881 pathlen = strlen(path) + 1; /* #bytes for the string */ 882 recordlen = offsetof(struct pmclog_pmcattach, pl_pathname) + pathlen; 883 884 PMCLOG_RESERVE(po, PMCATTACH, recordlen); 885 PMCLOG_EMIT32(pm->pm_id); 886 PMCLOG_EMIT32(pid); 887 PMCLOG_EMITSTRING(path, pathlen); 888 PMCLOG_DESPATCH(po); 889} 890 891void 892pmclog_process_pmcdetach(struct pmc *pm, pid_t pid) 893{ 894 struct pmc_owner *po; 895 896 PMCDBG2(LOG,ATT,1,"!pm=%p pid=%d", pm, pid); 897 898 po = pm->pm_owner; 899 900 PMCLOG_RESERVE(po, PMCDETACH, sizeof(struct pmclog_pmcdetach)); 901 PMCLOG_EMIT32(pm->pm_id); 902 PMCLOG_EMIT32(pid); 903 PMCLOG_DESPATCH(po); 904} 905 906/* 907 * Log a context switch event to the log file. 908 */ 909 910void 911pmclog_process_proccsw(struct pmc *pm, struct pmc_process *pp, pmc_value_t v) 912{ 913 struct pmc_owner *po; 914 915 KASSERT(pm->pm_flags & PMC_F_LOG_PROCCSW, 916 ("[pmclog,%d] log-process-csw called gratuitously", __LINE__)); 917 918 PMCDBG3(LOG,SWO,1,"pm=%p pid=%d v=%jx", pm, pp->pp_proc->p_pid, 919 v); 920 921 po = pm->pm_owner; 922 923 PMCLOG_RESERVE(po, PROCCSW, sizeof(struct pmclog_proccsw)); 924 PMCLOG_EMIT32(pm->pm_id); 925 PMCLOG_EMIT64(v); 926 PMCLOG_EMIT32(pp->pp_proc->p_pid); 927 PMCLOG_DESPATCH(po); 928} 929 930void 931pmclog_process_procexec(struct pmc_owner *po, pmc_id_t pmid, pid_t pid, 932 uintfptr_t startaddr, char *path) 933{ 934 int pathlen, recordlen; 935 936 PMCDBG3(LOG,EXC,1,"po=%p pid=%d path=\"%s\"", po, pid, path); 937 938 pathlen = strlen(path) + 1; /* #bytes for the path */ 939 recordlen = offsetof(struct pmclog_procexec, pl_pathname) + pathlen; 940 941 PMCLOG_RESERVE(po, PROCEXEC, recordlen); 942 PMCLOG_EMIT32(pid); 943 PMCLOG_EMITADDR(startaddr); 944 PMCLOG_EMIT32(pmid); 945 PMCLOG_EMITSTRING(path,pathlen); 946 PMCLOG_DESPATCH(po); 947} 948 949/* 950 * Log a process exit event (and accumulated pmc value) to the log file. 951 */ 952 953void 954pmclog_process_procexit(struct pmc *pm, struct pmc_process *pp) 955{ 956 int ri; 957 struct pmc_owner *po; 958 959 ri = PMC_TO_ROWINDEX(pm); 960 PMCDBG3(LOG,EXT,1,"pm=%p pid=%d v=%jx", pm, pp->pp_proc->p_pid, 961 pp->pp_pmcs[ri].pp_pmcval); 962 963 po = pm->pm_owner; 964 965 PMCLOG_RESERVE(po, PROCEXIT, sizeof(struct pmclog_procexit)); 966 PMCLOG_EMIT32(pm->pm_id); 967 PMCLOG_EMIT64(pp->pp_pmcs[ri].pp_pmcval); 968 PMCLOG_EMIT32(pp->pp_proc->p_pid); 969 PMCLOG_DESPATCH(po); 970} 971 972/* 973 * Log a fork event. 974 */ 975 976void 977pmclog_process_procfork(struct pmc_owner *po, pid_t oldpid, pid_t newpid) 978{ 979 PMCLOG_RESERVE(po, PROCFORK, sizeof(struct pmclog_procfork)); 980 PMCLOG_EMIT32(oldpid); 981 PMCLOG_EMIT32(newpid); 982 PMCLOG_DESPATCH(po); 983} 984 985/* 986 * Log a process exit event of the form suitable for system-wide PMCs. 987 */ 988 989void 990pmclog_process_sysexit(struct pmc_owner *po, pid_t pid) 991{ 992 PMCLOG_RESERVE(po, SYSEXIT, sizeof(struct pmclog_sysexit)); 993 PMCLOG_EMIT32(pid); 994 PMCLOG_DESPATCH(po); 995} 996 997/* 998 * Write a user log entry. 999 */ 1000 1001int 1002pmclog_process_userlog(struct pmc_owner *po, struct pmc_op_writelog *wl) 1003{ 1004 int error; 1005 1006 PMCDBG2(LOG,WRI,1, "writelog po=%p ud=0x%x", po, wl->pm_userdata); 1007 1008 error = 0; 1009 1010 PMCLOG_RESERVE_WITH_ERROR(po, USERDATA, 1011 sizeof(struct pmclog_userdata)); 1012 PMCLOG_EMIT32(wl->pm_userdata); 1013 PMCLOG_DESPATCH(po); 1014 1015 error: 1016 return (error); 1017} 1018 1019/* 1020 * Initialization. 1021 * 1022 * Create a pool of log buffers and initialize mutexes. 1023 */ 1024 1025void 1026pmclog_initialize() 1027{ 1028 int n; 1029 struct pmclog_buffer *plb; 1030 1031 if (pmclog_buffer_size <= 0) { 1032 (void) printf("hwpmc: tunable logbuffersize=%d must be " 1033 "greater than zero.\n", pmclog_buffer_size); 1034 pmclog_buffer_size = PMC_LOG_BUFFER_SIZE; 1035 } 1036 1037 if (pmc_nlogbuffers <= 0) { 1038 (void) printf("hwpmc: tunable nlogbuffers=%d must be greater " 1039 "than zero.\n", pmc_nlogbuffers); 1040 pmc_nlogbuffers = PMC_NLOGBUFFERS; 1041 } 1042 1043 /* create global pool of log buffers */ 1044 for (n = 0; n < pmc_nlogbuffers; n++) { 1045 plb = malloc(1024 * pmclog_buffer_size, M_PMC, 1046 M_WAITOK|M_ZERO); 1047 PMCLOG_INIT_BUFFER_DESCRIPTOR(plb); 1048 TAILQ_INSERT_HEAD(&pmc_bufferlist, plb, plb_next); 1049 } 1050 mtx_init(&pmc_bufferlist_mtx, "pmc-buffer-list", "pmc-leaf", 1051 MTX_SPIN); 1052 mtx_init(&pmc_kthread_mtx, "pmc-kthread", "pmc-sleep", MTX_DEF); 1053} 1054 1055/* 1056 * Shutdown logging. 1057 * 1058 * Destroy mutexes and release memory back the to free pool. 1059 */ 1060 1061void 1062pmclog_shutdown() 1063{ 1064 struct pmclog_buffer *plb; 1065 1066 mtx_destroy(&pmc_kthread_mtx); 1067 mtx_destroy(&pmc_bufferlist_mtx); 1068 1069 while ((plb = TAILQ_FIRST(&pmc_bufferlist)) != NULL) { 1070 TAILQ_REMOVE(&pmc_bufferlist, plb, plb_next); 1071 free(plb, M_PMC); 1072 } 1073} 1074