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/file.h> 45 #include <sys/kernel.h> 46 #include <sys/kthread.h> 47 #include <sys/lock.h> 48 #include <sys/module.h> 49 #include <sys/mutex.h> 50 #include <sys/pmc.h> 51 #include <sys/pmckern.h> 52 #include <sys/pmclog.h> 53 #include <sys/proc.h> 54 #include <sys/sched.h> 55 #include <sys/signalvar.h> 56 #include <sys/smp.h> 57 #include <sys/syscallsubr.h> 58 #include <sys/sysctl.h> 59 #include <sys/systm.h> 60 #include <sys/uio.h> 61 #include <sys/unistd.h> 62 #include <sys/vnode.h> 63 64 #ifdef NUMA 65 #define NDOMAINS vm_ndomains 66 #define curdomain PCPU_GET(domain) 67 #else 68 #define NDOMAINS 1 69 #define curdomain 0 70 #define malloc_domain(size, type, domain, flags) malloc((size), (type), (flags)) 71 #define free_domain(addr, type) free(addr, type) 72 #endif 73 74 /* 75 * Sysctl tunables 76 */ 77 78 SYSCTL_DECL(_kern_hwpmc); 79 80 /* 81 * kern.hwpmc.logbuffersize -- size of the per-cpu owner buffers. 82 */ 83 84 static int pmclog_buffer_size = PMC_LOG_BUFFER_SIZE; 85 #if (__FreeBSD_version < 1100000) 86 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "logbuffersize", &pmclog_buffer_size); 87 #endif 88 SYSCTL_INT(_kern_hwpmc, OID_AUTO, logbuffersize, CTLFLAG_RDTUN, 89 &pmclog_buffer_size, 0, "size of log buffers in kilobytes"); 90 91 /* 92 * kern.hwpmc.nbuffer -- number of global log buffers 93 */ 94 95 static int pmc_nlogbuffers_pcpu = PMC_NLOGBUFFERS_PCPU; 96 #if (__FreeBSD_version < 1100000) 97 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "nbuffers", &pmc_nlogbuffers_pcpu); 98 #endif 99 SYSCTL_INT(_kern_hwpmc, OID_AUTO, nbuffers_pcpu, CTLFLAG_RDTUN, 100 &pmc_nlogbuffers_pcpu, 0, "number of log buffers per cpu"); 101 102 /* 103 * Global log buffer list and associated spin lock. 104 */ 105 106 static struct mtx pmc_kthread_mtx; /* sleep lock */ 107 108 #define PMCLOG_INIT_BUFFER_DESCRIPTOR(D, buf, domain) do { \ 109 (D)->plb_fence = ((char *) (buf)) + 1024*pmclog_buffer_size; \ 110 (D)->plb_base = (D)->plb_ptr = ((char *) (buf)); \ 111 (D)->plb_domain = domain; \ 112 } while (0) 113 114 #define PMCLOG_RESET_BUFFER_DESCRIPTOR(D) do { \ 115 (D)->plb_ptr = (D)->plb_base; \ 116 } while (0) 117 118 /* 119 * Log file record constructors. 120 */ 121 #define _PMCLOG_TO_HEADER(T,L) \ 122 ((PMCLOG_HEADER_MAGIC << 24) | \ 123 (PMCLOG_TYPE_ ## T << 16) | \ 124 ((L) & 0xFFFF)) 125 126 /* reserve LEN bytes of space and initialize the entry header */ 127 #define _PMCLOG_RESERVE_SAFE(PO,TYPE,LEN,ACTION) do { \ 128 uint32_t *_le; \ 129 int _len = roundup((LEN), sizeof(uint32_t)); \ 130 if ((_le = pmclog_reserve((PO), _len)) == NULL) { \ 131 ACTION; \ 132 } \ 133 *_le = _PMCLOG_TO_HEADER(TYPE,_len); \ 134 _le += 3 /* 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 spinlock_enter(); \ 141 if ((_le = pmclog_reserve((PO), _len)) == NULL) { \ 142 spinlock_exit(); \ 143 ACTION; \ 144 } \ 145 *_le = _PMCLOG_TO_HEADER(TYPE,_len); \ 146 _le += 3 /* skip over timestamp */ 147 148 149 #define PMCLOG_RESERVE_SAFE(P,T,L) _PMCLOG_RESERVE_SAFE(P,T,L,return) 150 #define PMCLOG_RESERVE(P,T,L) _PMCLOG_RESERVE(P,T,L,return) 151 #define PMCLOG_RESERVE_WITH_ERROR(P,T,L) _PMCLOG_RESERVE(P,T,L, \ 152 error=ENOMEM;goto error) 153 154 #define PMCLOG_EMIT32(V) do { *_le++ = (V); } while (0) 155 #define PMCLOG_EMIT64(V) do { \ 156 *_le++ = (uint32_t) ((V) & 0xFFFFFFFF); \ 157 *_le++ = (uint32_t) (((V) >> 32) & 0xFFFFFFFF); \ 158 } while (0) 159 160 161 /* Emit a string. Caution: does NOT update _le, so needs to be last */ 162 #define PMCLOG_EMITSTRING(S,L) do { bcopy((S), _le, (L)); } while (0) 163 #define PMCLOG_EMITNULLSTRING(L) do { bzero(_le, (L)); } while (0) 164 165 #define PMCLOG_DESPATCH_SAFE(PO) \ 166 pmclog_release((PO)); \ 167 } while (0) 168 169 #define PMCLOG_DESPATCH(PO) \ 170 pmclog_release((PO)); \ 171 spinlock_exit(); \ 172 } while (0) 173 174 #define PMCLOG_DESPATCH_SYNC(PO) \ 175 pmclog_schedule_io((PO)); \ 176 spinlock_exit(); \ 177 } while (0) 178 179 180 /* 181 * Assertions about the log file format. 182 */ 183 184 CTASSERT(sizeof(struct pmclog_callchain) == 6*4 + 185 PMC_CALLCHAIN_DEPTH_MAX*sizeof(uintfptr_t)); 186 CTASSERT(sizeof(struct pmclog_closelog) == 3*4); 187 CTASSERT(sizeof(struct pmclog_dropnotify) == 3*4); 188 CTASSERT(sizeof(struct pmclog_map_in) == PATH_MAX + 189 4*4 + sizeof(uintfptr_t)); 190 CTASSERT(offsetof(struct pmclog_map_in,pl_pathname) == 191 4*4 + sizeof(uintfptr_t)); 192 CTASSERT(sizeof(struct pmclog_map_out) == 4*4 + 2*sizeof(uintfptr_t)); 193 CTASSERT(sizeof(struct pmclog_pcsample) == 6*4 + sizeof(uintfptr_t)); 194 CTASSERT(sizeof(struct pmclog_pmcallocate) == 6*4); 195 CTASSERT(sizeof(struct pmclog_pmcattach) == 5*4 + PATH_MAX); 196 CTASSERT(offsetof(struct pmclog_pmcattach,pl_pathname) == 5*4); 197 CTASSERT(sizeof(struct pmclog_pmcdetach) == 5*4); 198 CTASSERT(sizeof(struct pmclog_proccsw) == 5*4 + 8); 199 CTASSERT(sizeof(struct pmclog_procexec) == 5*4 + PATH_MAX + 200 sizeof(uintfptr_t)); 201 CTASSERT(offsetof(struct pmclog_procexec,pl_pathname) == 5*4 + 202 sizeof(uintfptr_t)); 203 CTASSERT(sizeof(struct pmclog_procexit) == 5*4 + 8); 204 CTASSERT(sizeof(struct pmclog_procfork) == 5*4); 205 CTASSERT(sizeof(struct pmclog_sysexit) == 4*4); 206 CTASSERT(sizeof(struct pmclog_userdata) == 4*4); 207 208 /* 209 * Log buffer structure 210 */ 211 212 struct pmclog_buffer { 213 TAILQ_ENTRY(pmclog_buffer) plb_next; 214 char *plb_base; 215 char *plb_ptr; 216 char *plb_fence; 217 uint16_t plb_domain; 218 } __aligned(CACHE_LINE_SIZE); 219 220 /* 221 * Prototypes 222 */ 223 224 static int pmclog_get_buffer(struct pmc_owner *po); 225 static void pmclog_loop(void *arg); 226 static void pmclog_release(struct pmc_owner *po); 227 static uint32_t *pmclog_reserve(struct pmc_owner *po, int length); 228 static void pmclog_schedule_io(struct pmc_owner *po); 229 static void pmclog_schedule_all(struct pmc_owner *po); 230 static void pmclog_stop_kthread(struct pmc_owner *po); 231 232 /* 233 * Helper functions 234 */ 235 236 static inline void 237 pmc_plb_rele_unlocked(struct pmclog_buffer *plb) 238 { 239 TAILQ_INSERT_HEAD(&pmc_dom_hdrs[plb->plb_domain]->pdbh_head, plb, plb_next); 240 } 241 242 static inline void 243 pmc_plb_rele(struct pmclog_buffer *plb) 244 { 245 mtx_lock_spin(&pmc_dom_hdrs[plb->plb_domain]->pdbh_mtx); 246 pmc_plb_rele_unlocked(plb); 247 mtx_unlock_spin(&pmc_dom_hdrs[plb->plb_domain]->pdbh_mtx); 248 } 249 250 251 /* 252 * Get a log buffer 253 */ 254 255 static int 256 pmclog_get_buffer(struct pmc_owner *po) 257 { 258 struct pmclog_buffer *plb; 259 int domain; 260 261 KASSERT(po->po_curbuf[curcpu] == NULL, 262 ("[pmclog,%d] po=%p current buffer still valid", __LINE__, po)); 263 264 domain = curdomain; 265 MPASS(pmc_dom_hdrs[domain]); 266 mtx_lock_spin(&pmc_dom_hdrs[domain]->pdbh_mtx); 267 if ((plb = TAILQ_FIRST(&pmc_dom_hdrs[domain]->pdbh_head)) != NULL) 268 TAILQ_REMOVE(&pmc_dom_hdrs[domain]->pdbh_head, plb, plb_next); 269 mtx_unlock_spin(&pmc_dom_hdrs[domain]->pdbh_mtx); 270 271 PMCDBG2(LOG,GTB,1, "po=%p plb=%p", po, plb); 272 273 #ifdef HWPMC_DEBUG 274 if (plb) 275 KASSERT(plb->plb_ptr == plb->plb_base && 276 plb->plb_base < plb->plb_fence, 277 ("[pmclog,%d] po=%p buffer invariants: ptr=%p " 278 "base=%p fence=%p", __LINE__, po, plb->plb_ptr, 279 plb->plb_base, plb->plb_fence)); 280 #endif 281 282 po->po_curbuf[curcpu] = plb; 283 284 /* update stats */ 285 counter_u64_add(pmc_stats.pm_buffer_requests, 1); 286 if (plb == NULL) 287 counter_u64_add(pmc_stats.pm_buffer_requests_failed, 1); 288 289 return (plb ? 0 : ENOMEM); 290 } 291 292 struct pmclog_proc_init_args { 293 struct proc *kthr; 294 struct pmc_owner *po; 295 bool exit; 296 bool acted; 297 }; 298 299 int 300 pmclog_proc_create(struct thread *td, void **handlep) 301 { 302 struct pmclog_proc_init_args *ia; 303 int error; 304 305 ia = malloc(sizeof(*ia), M_TEMP, M_WAITOK | M_ZERO); 306 error = kproc_create(pmclog_loop, ia, &ia->kthr, 307 RFHIGHPID, 0, "hwpmc: proc(%d)", td->td_proc->p_pid); 308 if (error == 0) 309 *handlep = ia; 310 return (error); 311 } 312 313 void 314 pmclog_proc_ignite(void *handle, struct pmc_owner *po) 315 { 316 struct pmclog_proc_init_args *ia; 317 318 ia = handle; 319 mtx_lock(&pmc_kthread_mtx); 320 MPASS(!ia->acted); 321 MPASS(ia->po == NULL); 322 MPASS(!ia->exit); 323 MPASS(ia->kthr != NULL); 324 if (po == NULL) { 325 ia->exit = true; 326 } else { 327 ia->po = po; 328 KASSERT(po->po_kthread == NULL, 329 ("[pmclog,%d] po=%p kthread (%p) already present", 330 __LINE__, po, po->po_kthread)); 331 po->po_kthread = ia->kthr; 332 } 333 wakeup(ia); 334 while (!ia->acted) 335 msleep(ia, &pmc_kthread_mtx, PWAIT, "pmclogw", 0); 336 mtx_unlock(&pmc_kthread_mtx); 337 free(ia, M_TEMP); 338 } 339 340 /* 341 * Log handler loop. 342 * 343 * This function is executed by each pmc owner's helper thread. 344 */ 345 346 static void 347 pmclog_loop(void *arg) 348 { 349 struct pmclog_proc_init_args *ia; 350 struct pmc_owner *po; 351 struct pmclog_buffer *lb; 352 struct proc *p; 353 struct ucred *ownercred; 354 struct ucred *mycred; 355 struct thread *td; 356 sigset_t unb; 357 struct uio auio; 358 struct iovec aiov; 359 size_t nbytes; 360 int error; 361 362 td = curthread; 363 364 SIGEMPTYSET(unb); 365 SIGADDSET(unb, SIGHUP); 366 (void)kern_sigprocmask(td, SIG_UNBLOCK, &unb, NULL, 0); 367 368 ia = arg; 369 MPASS(ia->kthr == curproc); 370 MPASS(!ia->acted); 371 mtx_lock(&pmc_kthread_mtx); 372 while (ia->po == NULL && !ia->exit) 373 msleep(ia, &pmc_kthread_mtx, PWAIT, "pmclogi", 0); 374 if (ia->exit) { 375 ia->acted = true; 376 wakeup(ia); 377 mtx_unlock(&pmc_kthread_mtx); 378 kproc_exit(0); 379 } 380 MPASS(ia->po != NULL); 381 po = ia->po; 382 ia->acted = true; 383 wakeup(ia); 384 mtx_unlock(&pmc_kthread_mtx); 385 ia = NULL; 386 387 p = po->po_owner; 388 mycred = td->td_ucred; 389 390 PROC_LOCK(p); 391 ownercred = crhold(p->p_ucred); 392 PROC_UNLOCK(p); 393 394 PMCDBG2(LOG,INI,1, "po=%p kt=%p", po, po->po_kthread); 395 KASSERT(po->po_kthread == curthread->td_proc, 396 ("[pmclog,%d] proc mismatch po=%p po/kt=%p curproc=%p", __LINE__, 397 po, po->po_kthread, curthread->td_proc)); 398 399 lb = NULL; 400 401 402 /* 403 * Loop waiting for I/O requests to be added to the owner 404 * struct's queue. The loop is exited when the log file 405 * is deconfigured. 406 */ 407 408 mtx_lock(&pmc_kthread_mtx); 409 410 for (;;) { 411 412 /* check if we've been asked to exit */ 413 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) 414 break; 415 416 if (lb == NULL) { /* look for a fresh buffer to write */ 417 mtx_lock_spin(&po->po_mtx); 418 if ((lb = TAILQ_FIRST(&po->po_logbuffers)) == NULL) { 419 mtx_unlock_spin(&po->po_mtx); 420 421 /* No more buffers and shutdown required. */ 422 if (po->po_flags & PMC_PO_SHUTDOWN) 423 break; 424 425 (void) msleep(po, &pmc_kthread_mtx, PWAIT, 426 "pmcloop", 0); 427 continue; 428 } 429 430 TAILQ_REMOVE(&po->po_logbuffers, lb, plb_next); 431 mtx_unlock_spin(&po->po_mtx); 432 } 433 434 mtx_unlock(&pmc_kthread_mtx); 435 436 /* process the request */ 437 PMCDBG3(LOG,WRI,2, "po=%p base=%p ptr=%p", po, 438 lb->plb_base, lb->plb_ptr); 439 /* change our thread's credentials before issuing the I/O */ 440 441 aiov.iov_base = lb->plb_base; 442 aiov.iov_len = nbytes = lb->plb_ptr - lb->plb_base; 443 444 auio.uio_iov = &aiov; 445 auio.uio_iovcnt = 1; 446 auio.uio_offset = -1; 447 auio.uio_resid = nbytes; 448 auio.uio_rw = UIO_WRITE; 449 auio.uio_segflg = UIO_SYSSPACE; 450 auio.uio_td = td; 451 452 /* switch thread credentials -- see kern_ktrace.c */ 453 td->td_ucred = ownercred; 454 error = fo_write(po->po_file, &auio, ownercred, 0, td); 455 td->td_ucred = mycred; 456 457 if (error) { 458 /* XXX some errors are recoverable */ 459 /* send a SIGIO to the owner and exit */ 460 PROC_LOCK(p); 461 kern_psignal(p, SIGIO); 462 PROC_UNLOCK(p); 463 464 mtx_lock(&pmc_kthread_mtx); 465 466 po->po_error = error; /* save for flush log */ 467 468 PMCDBG2(LOG,WRI,2, "po=%p error=%d", po, error); 469 470 break; 471 } 472 473 mtx_lock(&pmc_kthread_mtx); 474 475 /* put the used buffer back into the global pool */ 476 PMCLOG_RESET_BUFFER_DESCRIPTOR(lb); 477 478 pmc_plb_rele(lb); 479 lb = NULL; 480 } 481 482 wakeup_one(po->po_kthread); 483 po->po_kthread = NULL; 484 485 mtx_unlock(&pmc_kthread_mtx); 486 487 /* return the current I/O buffer to the global pool */ 488 if (lb) { 489 PMCLOG_RESET_BUFFER_DESCRIPTOR(lb); 490 491 pmc_plb_rele(lb); 492 } 493 494 /* 495 * Exit this thread, signalling the waiter 496 */ 497 498 crfree(ownercred); 499 500 kproc_exit(0); 501 } 502 503 /* 504 * Release and log entry and schedule an I/O if needed. 505 */ 506 507 static void 508 pmclog_release(struct pmc_owner *po) 509 { 510 struct pmclog_buffer *plb; 511 512 plb = po->po_curbuf[curcpu]; 513 KASSERT(plb->plb_ptr >= plb->plb_base, 514 ("[pmclog,%d] buffer invariants po=%p ptr=%p base=%p", __LINE__, 515 po, plb->plb_ptr, plb->plb_base)); 516 KASSERT(plb->plb_ptr <= plb->plb_fence, 517 ("[pmclog,%d] buffer invariants po=%p ptr=%p fenc=%p", __LINE__, 518 po, plb->plb_ptr, plb->plb_fence)); 519 520 /* schedule an I/O if we've filled a buffer */ 521 if (plb->plb_ptr >= plb->plb_fence) 522 pmclog_schedule_io(po); 523 524 PMCDBG1(LOG,REL,1, "po=%p", po); 525 } 526 527 528 /* 529 * Attempt to reserve 'length' bytes of space in an owner's log 530 * buffer. The function returns a pointer to 'length' bytes of space 531 * if there was enough space or returns NULL if no space was 532 * available. Non-null returns do so with the po mutex locked. The 533 * caller must invoke pmclog_release() on the pmc owner structure 534 * when done. 535 */ 536 537 static uint32_t * 538 pmclog_reserve(struct pmc_owner *po, int length) 539 { 540 uintptr_t newptr, oldptr; 541 uint32_t *lh; 542 struct timespec ts; 543 struct pmclog_buffer *plb, **pplb; 544 545 PMCDBG2(LOG,ALL,1, "po=%p len=%d", po, length); 546 547 KASSERT(length % sizeof(uint32_t) == 0, 548 ("[pmclog,%d] length not a multiple of word size", __LINE__)); 549 550 /* No more data when shutdown in progress. */ 551 if (po->po_flags & PMC_PO_SHUTDOWN) 552 return (NULL); 553 554 pplb = &po->po_curbuf[curcpu]; 555 if (*pplb == NULL && pmclog_get_buffer(po) != 0) 556 goto fail; 557 558 KASSERT(*pplb != NULL, 559 ("[pmclog,%d] po=%p no current buffer", __LINE__, po)); 560 561 plb = *pplb; 562 KASSERT(plb->plb_ptr >= plb->plb_base && 563 plb->plb_ptr <= plb->plb_fence, 564 ("[pmclog,%d] po=%p buffer invariants: ptr=%p base=%p fence=%p", 565 __LINE__, po, plb->plb_ptr, plb->plb_base, 566 plb->plb_fence)); 567 568 oldptr = (uintptr_t) plb->plb_ptr; 569 newptr = oldptr + length; 570 571 KASSERT(oldptr != (uintptr_t) NULL, 572 ("[pmclog,%d] po=%p Null log buffer pointer", __LINE__, po)); 573 574 /* 575 * If we have space in the current buffer, return a pointer to 576 * available space with the PO structure locked. 577 */ 578 if (newptr <= (uintptr_t) plb->plb_fence) { 579 plb->plb_ptr = (char *) newptr; 580 goto done; 581 } 582 583 /* 584 * Otherwise, schedule the current buffer for output and get a 585 * fresh buffer. 586 */ 587 pmclog_schedule_io(po); 588 589 if (pmclog_get_buffer(po) != 0) 590 goto fail; 591 592 plb = *pplb; 593 KASSERT(plb != NULL, 594 ("[pmclog,%d] po=%p no current buffer", __LINE__, po)); 595 596 KASSERT(plb->plb_ptr != NULL, 597 ("[pmclog,%d] null return from pmc_get_log_buffer", __LINE__)); 598 599 KASSERT(plb->plb_ptr == plb->plb_base && 600 plb->plb_ptr <= plb->plb_fence, 601 ("[pmclog,%d] po=%p buffer invariants: ptr=%p base=%p fence=%p", 602 __LINE__, po, plb->plb_ptr, plb->plb_base, 603 plb->plb_fence)); 604 605 oldptr = (uintptr_t) plb->plb_ptr; 606 607 done: 608 lh = (uint32_t *) oldptr; 609 lh++; /* skip header */ 610 getnanotime(&ts); /* fill in the timestamp */ 611 *lh++ = ts.tv_sec & 0xFFFFFFFF; 612 *lh++ = ts.tv_nsec & 0xFFFFFFF; 613 return ((uint32_t *) oldptr); 614 fail: 615 return (NULL); 616 } 617 618 /* 619 * Schedule an I/O. 620 * 621 * Transfer the current buffer to the helper kthread. 622 */ 623 624 static void 625 pmclog_schedule_io(struct pmc_owner *po) 626 { 627 struct pmclog_buffer *plb; 628 629 plb = po->po_curbuf[curcpu]; 630 po->po_curbuf[curcpu] = NULL; 631 KASSERT(plb != NULL, 632 ("[pmclog,%d] schedule_io with null buffer po=%p", __LINE__, po)); 633 KASSERT(plb->plb_ptr >= plb->plb_base, 634 ("[pmclog,%d] buffer invariants po=%p ptr=%p base=%p", __LINE__, 635 po, plb->plb_ptr, plb->plb_base)); 636 KASSERT(plb->plb_ptr <= plb->plb_fence, 637 ("[pmclog,%d] buffer invariants po=%p ptr=%p fenc=%p", __LINE__, 638 po, plb->plb_ptr, plb->plb_fence)); 639 640 PMCDBG1(LOG,SIO, 1, "po=%p", po); 641 642 /* 643 * Add the current buffer to the tail of the buffer list and 644 * wakeup the helper. 645 */ 646 mtx_lock_spin(&po->po_mtx); 647 TAILQ_INSERT_TAIL(&po->po_logbuffers, plb, plb_next); 648 mtx_unlock_spin(&po->po_mtx); 649 wakeup_one(po); 650 } 651 652 /* 653 * Stop the helper kthread. 654 */ 655 656 static void 657 pmclog_stop_kthread(struct pmc_owner *po) 658 { 659 660 mtx_lock(&pmc_kthread_mtx); 661 po->po_flags &= ~PMC_PO_OWNS_LOGFILE; 662 if (po->po_kthread != NULL) { 663 PROC_LOCK(po->po_kthread); 664 kern_psignal(po->po_kthread, SIGHUP); 665 PROC_UNLOCK(po->po_kthread); 666 } 667 wakeup_one(po); 668 while (po->po_kthread) 669 msleep(po->po_kthread, &pmc_kthread_mtx, PPAUSE, "pmckstp", 0); 670 mtx_unlock(&pmc_kthread_mtx); 671 } 672 673 /* 674 * Public functions 675 */ 676 677 /* 678 * Configure a log file for pmc owner 'po'. 679 * 680 * Parameter 'logfd' is a file handle referencing an open file in the 681 * owner process. This file needs to have been opened for writing. 682 */ 683 684 int 685 pmclog_configure_log(struct pmc_mdep *md, struct pmc_owner *po, int logfd) 686 { 687 struct proc *p; 688 int error; 689 690 sx_assert(&pmc_sx, SA_XLOCKED); 691 PMCDBG2(LOG,CFG,1, "config po=%p logfd=%d", po, logfd); 692 693 p = po->po_owner; 694 695 /* return EBUSY if a log file was already present */ 696 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 697 return (EBUSY); 698 699 KASSERT(po->po_file == NULL, 700 ("[pmclog,%d] po=%p file (%p) already present", __LINE__, po, 701 po->po_file)); 702 703 /* get a reference to the file state */ 704 error = fget_write(curthread, logfd, &cap_write_rights, &po->po_file); 705 if (error) 706 goto error; 707 708 /* mark process as owning a log file */ 709 po->po_flags |= PMC_PO_OWNS_LOGFILE; 710 711 /* mark process as using HWPMCs */ 712 PROC_LOCK(p); 713 p->p_flag |= P_HWPMC; 714 PROC_UNLOCK(p); 715 716 /* create a log initialization entry */ 717 PMCLOG_RESERVE_WITH_ERROR(po, INITIALIZE, 718 sizeof(struct pmclog_initialize)); 719 PMCLOG_EMIT32(PMC_VERSION); 720 PMCLOG_EMIT32(md->pmd_cputype); 721 PMCLOG_DESPATCH_SYNC(po); 722 723 return (0); 724 725 error: 726 KASSERT(po->po_kthread == NULL, ("[pmclog,%d] po=%p kthread not " 727 "stopped", __LINE__, po)); 728 729 if (po->po_file) 730 (void) fdrop(po->po_file, curthread); 731 po->po_file = NULL; /* clear file and error state */ 732 po->po_error = 0; 733 po->po_flags &= ~PMC_PO_OWNS_LOGFILE; 734 735 return (error); 736 } 737 738 739 /* 740 * De-configure a log file. This will throw away any buffers queued 741 * for this owner process. 742 */ 743 744 int 745 pmclog_deconfigure_log(struct pmc_owner *po) 746 { 747 int error; 748 struct pmclog_buffer *lb; 749 750 PMCDBG1(LOG,CFG,1, "de-config po=%p", po); 751 752 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) 753 return (EINVAL); 754 755 KASSERT(po->po_sscount == 0, 756 ("[pmclog,%d] po=%p still owning SS PMCs", __LINE__, po)); 757 KASSERT(po->po_file != NULL, 758 ("[pmclog,%d] po=%p no log file", __LINE__, po)); 759 760 /* stop the kthread, this will reset the 'OWNS_LOGFILE' flag */ 761 pmclog_stop_kthread(po); 762 763 KASSERT(po->po_kthread == NULL, 764 ("[pmclog,%d] po=%p kthread not stopped", __LINE__, po)); 765 766 /* return all queued log buffers to the global pool */ 767 while ((lb = TAILQ_FIRST(&po->po_logbuffers)) != NULL) { 768 TAILQ_REMOVE(&po->po_logbuffers, lb, plb_next); 769 PMCLOG_RESET_BUFFER_DESCRIPTOR(lb); 770 pmc_plb_rele(lb); 771 } 772 for (int i = 0; i < mp_ncpus; i++) { 773 thread_lock(curthread); 774 sched_bind(curthread, i); 775 thread_unlock(curthread); 776 /* return the 'current' buffer to the global pool */ 777 if ((lb = po->po_curbuf[curcpu]) != NULL) { 778 PMCLOG_RESET_BUFFER_DESCRIPTOR(lb); 779 pmc_plb_rele(lb); 780 } 781 } 782 thread_lock(curthread); 783 sched_unbind(curthread); 784 thread_unlock(curthread); 785 786 /* drop a reference to the fd */ 787 if (po->po_file != NULL) { 788 error = fdrop(po->po_file, curthread); 789 po->po_file = NULL; 790 } else 791 error = 0; 792 po->po_error = 0; 793 794 return (error); 795 } 796 797 /* 798 * Flush a process' log buffer. 799 */ 800 801 int 802 pmclog_flush(struct pmc_owner *po) 803 { 804 int error; 805 806 PMCDBG1(LOG,FLS,1, "po=%p", po); 807 808 /* 809 * If there is a pending error recorded by the logger thread, 810 * return that. 811 */ 812 if (po->po_error) 813 return (po->po_error); 814 815 error = 0; 816 817 /* 818 * Check that we do have an active log file. 819 */ 820 mtx_lock(&pmc_kthread_mtx); 821 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) { 822 error = EINVAL; 823 goto error; 824 } 825 826 pmclog_schedule_all(po); 827 error: 828 mtx_unlock(&pmc_kthread_mtx); 829 830 return (error); 831 } 832 833 static void 834 pmclog_schedule_one_cond(void *arg) 835 { 836 struct pmc_owner *po = arg; 837 struct pmclog_buffer *plb; 838 839 spinlock_enter(); 840 /* tell hardclock not to run again */ 841 if (PMC_CPU_HAS_SAMPLES(PCPU_GET(cpuid))) 842 PMC_CALL_HOOK_UNLOCKED(curthread, PMC_FN_DO_SAMPLES, NULL); 843 plb = po->po_curbuf[curcpu]; 844 if (plb && plb->plb_ptr != plb->plb_base) 845 pmclog_schedule_io(po); 846 spinlock_exit(); 847 } 848 849 static void 850 pmclog_schedule_all(struct pmc_owner *po) 851 { 852 /* 853 * Schedule the current buffer if any and not empty. 854 */ 855 for (int i = 0; i < mp_ncpus; i++) { 856 thread_lock(curthread); 857 sched_bind(curthread, i); 858 thread_unlock(curthread); 859 pmclog_schedule_one_cond(po); 860 } 861 thread_lock(curthread); 862 sched_unbind(curthread); 863 thread_unlock(curthread); 864 } 865 866 int 867 pmclog_close(struct pmc_owner *po) 868 { 869 870 PMCDBG1(LOG,CLO,1, "po=%p", po); 871 872 pmclog_process_closelog(po); 873 874 mtx_lock(&pmc_kthread_mtx); 875 876 /* 877 * Schedule the current buffer. 878 */ 879 pmclog_schedule_all(po); 880 wakeup_one(po); 881 882 /* 883 * Initiate shutdown: no new data queued, 884 * thread will close file on last block. 885 */ 886 po->po_flags |= PMC_PO_SHUTDOWN; 887 mtx_unlock(&pmc_kthread_mtx); 888 889 return (0); 890 } 891 892 void 893 pmclog_process_callchain(struct pmc *pm, struct pmc_sample *ps) 894 { 895 int n, recordlen; 896 uint32_t flags; 897 struct pmc_owner *po; 898 899 PMCDBG3(LOG,SAM,1,"pm=%p pid=%d n=%d", pm, ps->ps_pid, 900 ps->ps_nsamples); 901 902 recordlen = offsetof(struct pmclog_callchain, pl_pc) + 903 ps->ps_nsamples * sizeof(uintfptr_t); 904 po = pm->pm_owner; 905 flags = PMC_CALLCHAIN_TO_CPUFLAGS(ps->ps_cpu,ps->ps_flags); 906 PMCLOG_RESERVE_SAFE(po, CALLCHAIN, recordlen); 907 PMCLOG_EMIT32(ps->ps_pid); 908 PMCLOG_EMIT32(pm->pm_id); 909 PMCLOG_EMIT32(flags); 910 for (n = 0; n < ps->ps_nsamples; n++) 911 PMCLOG_EMITADDR(ps->ps_pc[n]); 912 PMCLOG_DESPATCH_SAFE(po); 913 } 914 915 void 916 pmclog_process_closelog(struct pmc_owner *po) 917 { 918 PMCLOG_RESERVE(po,CLOSELOG,sizeof(struct pmclog_closelog)); 919 PMCLOG_DESPATCH_SYNC(po); 920 } 921 922 void 923 pmclog_process_dropnotify(struct pmc_owner *po) 924 { 925 PMCLOG_RESERVE(po,DROPNOTIFY,sizeof(struct pmclog_dropnotify)); 926 PMCLOG_DESPATCH(po); 927 } 928 929 void 930 pmclog_process_map_in(struct pmc_owner *po, pid_t pid, uintfptr_t start, 931 const char *path) 932 { 933 int pathlen, recordlen; 934 935 KASSERT(path != NULL, ("[pmclog,%d] map-in, null path", __LINE__)); 936 937 pathlen = strlen(path) + 1; /* #bytes for path name */ 938 recordlen = offsetof(struct pmclog_map_in, pl_pathname) + 939 pathlen; 940 941 PMCLOG_RESERVE(po, MAP_IN, recordlen); 942 PMCLOG_EMIT32(pid); 943 PMCLOG_EMITADDR(start); 944 PMCLOG_EMITSTRING(path,pathlen); 945 PMCLOG_DESPATCH(po); 946 } 947 948 void 949 pmclog_process_map_out(struct pmc_owner *po, pid_t pid, uintfptr_t start, 950 uintfptr_t end) 951 { 952 KASSERT(start <= end, ("[pmclog,%d] start > end", __LINE__)); 953 954 PMCLOG_RESERVE(po, MAP_OUT, sizeof(struct pmclog_map_out)); 955 PMCLOG_EMIT32(pid); 956 PMCLOG_EMITADDR(start); 957 PMCLOG_EMITADDR(end); 958 PMCLOG_DESPATCH(po); 959 } 960 961 void 962 pmclog_process_pmcallocate(struct pmc *pm) 963 { 964 struct pmc_owner *po; 965 struct pmc_soft *ps; 966 967 po = pm->pm_owner; 968 969 PMCDBG1(LOG,ALL,1, "pm=%p", pm); 970 971 if (PMC_TO_CLASS(pm) == PMC_CLASS_SOFT) { 972 PMCLOG_RESERVE(po, PMCALLOCATEDYN, 973 sizeof(struct pmclog_pmcallocatedyn)); 974 PMCLOG_EMIT32(pm->pm_id); 975 PMCLOG_EMIT32(pm->pm_event); 976 PMCLOG_EMIT32(pm->pm_flags); 977 ps = pmc_soft_ev_acquire(pm->pm_event); 978 if (ps != NULL) 979 PMCLOG_EMITSTRING(ps->ps_ev.pm_ev_name,PMC_NAME_MAX); 980 else 981 PMCLOG_EMITNULLSTRING(PMC_NAME_MAX); 982 pmc_soft_ev_release(ps); 983 PMCLOG_DESPATCH_SYNC(po); 984 } else { 985 PMCLOG_RESERVE(po, PMCALLOCATE, 986 sizeof(struct pmclog_pmcallocate)); 987 PMCLOG_EMIT32(pm->pm_id); 988 PMCLOG_EMIT32(pm->pm_event); 989 PMCLOG_EMIT32(pm->pm_flags); 990 PMCLOG_DESPATCH_SYNC(po); 991 } 992 } 993 994 void 995 pmclog_process_pmcattach(struct pmc *pm, pid_t pid, char *path) 996 { 997 int pathlen, recordlen; 998 struct pmc_owner *po; 999 1000 PMCDBG2(LOG,ATT,1,"pm=%p pid=%d", pm, pid); 1001 1002 po = pm->pm_owner; 1003 1004 pathlen = strlen(path) + 1; /* #bytes for the string */ 1005 recordlen = offsetof(struct pmclog_pmcattach, pl_pathname) + pathlen; 1006 1007 PMCLOG_RESERVE(po, PMCATTACH, recordlen); 1008 PMCLOG_EMIT32(pm->pm_id); 1009 PMCLOG_EMIT32(pid); 1010 PMCLOG_EMITSTRING(path, pathlen); 1011 PMCLOG_DESPATCH_SYNC(po); 1012 } 1013 1014 void 1015 pmclog_process_pmcdetach(struct pmc *pm, pid_t pid) 1016 { 1017 struct pmc_owner *po; 1018 1019 PMCDBG2(LOG,ATT,1,"!pm=%p pid=%d", pm, pid); 1020 1021 po = pm->pm_owner; 1022 1023 PMCLOG_RESERVE(po, PMCDETACH, sizeof(struct pmclog_pmcdetach)); 1024 PMCLOG_EMIT32(pm->pm_id); 1025 PMCLOG_EMIT32(pid); 1026 PMCLOG_DESPATCH_SYNC(po); 1027 } 1028 1029 /* 1030 * Log a context switch event to the log file. 1031 */ 1032 1033 void 1034 pmclog_process_proccsw(struct pmc *pm, struct pmc_process *pp, pmc_value_t v) 1035 { 1036 struct pmc_owner *po; 1037 1038 KASSERT(pm->pm_flags & PMC_F_LOG_PROCCSW, 1039 ("[pmclog,%d] log-process-csw called gratuitously", __LINE__)); 1040 1041 PMCDBG3(LOG,SWO,1,"pm=%p pid=%d v=%jx", pm, pp->pp_proc->p_pid, 1042 v); 1043 1044 po = pm->pm_owner; 1045 1046 PMCLOG_RESERVE(po, PROCCSW, sizeof(struct pmclog_proccsw)); 1047 PMCLOG_EMIT32(pm->pm_id); 1048 PMCLOG_EMIT64(v); 1049 PMCLOG_EMIT32(pp->pp_proc->p_pid); 1050 PMCLOG_DESPATCH(po); 1051 } 1052 1053 void 1054 pmclog_process_procexec(struct pmc_owner *po, pmc_id_t pmid, pid_t pid, 1055 uintfptr_t startaddr, char *path) 1056 { 1057 int pathlen, recordlen; 1058 1059 PMCDBG3(LOG,EXC,1,"po=%p pid=%d path=\"%s\"", po, pid, path); 1060 1061 pathlen = strlen(path) + 1; /* #bytes for the path */ 1062 recordlen = offsetof(struct pmclog_procexec, pl_pathname) + pathlen; 1063 1064 PMCLOG_RESERVE(po, PROCEXEC, recordlen); 1065 PMCLOG_EMIT32(pid); 1066 PMCLOG_EMITADDR(startaddr); 1067 PMCLOG_EMIT32(pmid); 1068 PMCLOG_EMITSTRING(path,pathlen); 1069 PMCLOG_DESPATCH(po); 1070 } 1071 1072 /* 1073 * Log a process exit event (and accumulated pmc value) to the log file. 1074 */ 1075 1076 void 1077 pmclog_process_procexit(struct pmc *pm, struct pmc_process *pp) 1078 { 1079 int ri; 1080 struct pmc_owner *po; 1081 1082 ri = PMC_TO_ROWINDEX(pm); 1083 PMCDBG3(LOG,EXT,1,"pm=%p pid=%d v=%jx", pm, pp->pp_proc->p_pid, 1084 pp->pp_pmcs[ri].pp_pmcval); 1085 1086 po = pm->pm_owner; 1087 1088 PMCLOG_RESERVE(po, PROCEXIT, sizeof(struct pmclog_procexit)); 1089 PMCLOG_EMIT32(pm->pm_id); 1090 PMCLOG_EMIT64(pp->pp_pmcs[ri].pp_pmcval); 1091 PMCLOG_EMIT32(pp->pp_proc->p_pid); 1092 PMCLOG_DESPATCH(po); 1093 } 1094 1095 /* 1096 * Log a fork event. 1097 */ 1098 1099 void 1100 pmclog_process_procfork(struct pmc_owner *po, pid_t oldpid, pid_t newpid) 1101 { 1102 PMCLOG_RESERVE(po, PROCFORK, sizeof(struct pmclog_procfork)); 1103 PMCLOG_EMIT32(oldpid); 1104 PMCLOG_EMIT32(newpid); 1105 PMCLOG_DESPATCH(po); 1106 } 1107 1108 /* 1109 * Log a process exit event of the form suitable for system-wide PMCs. 1110 */ 1111 1112 void 1113 pmclog_process_sysexit(struct pmc_owner *po, pid_t pid) 1114 { 1115 PMCLOG_RESERVE(po, SYSEXIT, sizeof(struct pmclog_sysexit)); 1116 PMCLOG_EMIT32(pid); 1117 PMCLOG_DESPATCH(po); 1118 } 1119 1120 /* 1121 * Write a user log entry. 1122 */ 1123 1124 int 1125 pmclog_process_userlog(struct pmc_owner *po, struct pmc_op_writelog *wl) 1126 { 1127 int error; 1128 1129 PMCDBG2(LOG,WRI,1, "writelog po=%p ud=0x%x", po, wl->pm_userdata); 1130 1131 error = 0; 1132 1133 PMCLOG_RESERVE_WITH_ERROR(po, USERDATA, 1134 sizeof(struct pmclog_userdata)); 1135 PMCLOG_EMIT32(wl->pm_userdata); 1136 PMCLOG_DESPATCH(po); 1137 1138 error: 1139 return (error); 1140 } 1141 1142 /* 1143 * Initialization. 1144 * 1145 * Create a pool of log buffers and initialize mutexes. 1146 */ 1147 1148 void 1149 pmclog_initialize() 1150 { 1151 int domain; 1152 struct pmclog_buffer *plb; 1153 1154 if (pmclog_buffer_size <= 0 || pmclog_buffer_size > 16*1024) { 1155 (void) printf("hwpmc: tunable logbuffersize=%d must be " 1156 "greater than zero and less than or equal to 16MB.\n", 1157 pmclog_buffer_size); 1158 pmclog_buffer_size = PMC_LOG_BUFFER_SIZE; 1159 } 1160 1161 if (pmc_nlogbuffers_pcpu <= 0) { 1162 (void) printf("hwpmc: tunable nlogbuffers=%d must be greater " 1163 "than zero.\n", pmc_nlogbuffers_pcpu); 1164 pmc_nlogbuffers_pcpu = PMC_NLOGBUFFERS_PCPU; 1165 } 1166 if (pmc_nlogbuffers_pcpu*pmclog_buffer_size > 32*1024) { 1167 (void) printf("hwpmc: memory allocated pcpu must be less than 32MB (is %dK).\n", 1168 pmc_nlogbuffers_pcpu*pmclog_buffer_size); 1169 pmc_nlogbuffers_pcpu = PMC_NLOGBUFFERS_PCPU; 1170 pmclog_buffer_size = PMC_LOG_BUFFER_SIZE; 1171 } 1172 for (domain = 0; domain < NDOMAINS; domain++) { 1173 int ncpus = pmc_dom_hdrs[domain]->pdbh_ncpus; 1174 int total = ncpus*pmc_nlogbuffers_pcpu; 1175 1176 plb = malloc_domain(sizeof(struct pmclog_buffer)*total, M_PMC, domain, M_WAITOK|M_ZERO); 1177 pmc_dom_hdrs[domain]->pdbh_plbs = plb; 1178 for (int i = 0; i < total; i++, plb++) { 1179 void *buf; 1180 1181 buf = malloc_domain(1024 * pmclog_buffer_size, M_PMC, domain, 1182 M_WAITOK|M_ZERO); 1183 PMCLOG_INIT_BUFFER_DESCRIPTOR(plb, buf, domain); 1184 pmc_plb_rele_unlocked(plb); 1185 } 1186 } 1187 mtx_init(&pmc_kthread_mtx, "pmc-kthread", "pmc-sleep", MTX_DEF); 1188 } 1189 1190 /* 1191 * Shutdown logging. 1192 * 1193 * Destroy mutexes and release memory back the to free pool. 1194 */ 1195 1196 void 1197 pmclog_shutdown() 1198 { 1199 struct pmclog_buffer *plb; 1200 int domain; 1201 1202 mtx_destroy(&pmc_kthread_mtx); 1203 1204 for (domain = 0; domain < NDOMAINS; domain++) { 1205 while ((plb = TAILQ_FIRST(&pmc_dom_hdrs[domain]->pdbh_head)) != NULL) { 1206 TAILQ_REMOVE(&pmc_dom_hdrs[domain]->pdbh_head, plb, plb_next); 1207 free(plb->plb_base, M_PMC); 1208 } 1209 free(pmc_dom_hdrs[domain]->pdbh_plbs, M_PMC); 1210 } 1211 } 1212