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) == 8*4 + 185 PMC_CALLCHAIN_DEPTH_MAX*sizeof(uintfptr_t)); 186 CTASSERT(sizeof(struct pmclog_closelog) == 4*4); 187 CTASSERT(sizeof(struct pmclog_dropnotify) == 4*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_pmcallocate) == 6*4); 194 CTASSERT(sizeof(struct pmclog_pmcattach) == 6*4 + PATH_MAX); 195 CTASSERT(offsetof(struct pmclog_pmcattach,pl_pathname) == 6*4); 196 CTASSERT(sizeof(struct pmclog_pmcdetach) == 6*4); 197 CTASSERT(sizeof(struct pmclog_proccsw) == 6*4 + 8); 198 CTASSERT(sizeof(struct pmclog_procexec) == 6*4 + PATH_MAX + 199 sizeof(uintfptr_t)); 200 CTASSERT(offsetof(struct pmclog_procexec,pl_pathname) == 6*4 + 201 sizeof(uintfptr_t)); 202 CTASSERT(sizeof(struct pmclog_procexit) == 6*4 + 8); 203 CTASSERT(sizeof(struct pmclog_procfork) == 6*4); 204 CTASSERT(sizeof(struct pmclog_sysexit) == 4*4); 205 CTASSERT(sizeof(struct pmclog_userdata) == 4*4); 206 207 /* 208 * Log buffer structure 209 */ 210 211 struct pmclog_buffer { 212 TAILQ_ENTRY(pmclog_buffer) plb_next; 213 char *plb_base; 214 char *plb_ptr; 215 char *plb_fence; 216 uint16_t plb_domain; 217 } __aligned(CACHE_LINE_SIZE); 218 219 /* 220 * Prototypes 221 */ 222 223 static int pmclog_get_buffer(struct pmc_owner *po); 224 static void pmclog_loop(void *arg); 225 static void pmclog_release(struct pmc_owner *po); 226 static uint32_t *pmclog_reserve(struct pmc_owner *po, int length); 227 static void pmclog_schedule_io(struct pmc_owner *po); 228 static void pmclog_schedule_all(struct pmc_owner *po); 229 static void pmclog_stop_kthread(struct pmc_owner *po); 230 231 /* 232 * Helper functions 233 */ 234 235 static inline void 236 pmc_plb_rele_unlocked(struct pmclog_buffer *plb) 237 { 238 TAILQ_INSERT_HEAD(&pmc_dom_hdrs[plb->plb_domain]->pdbh_head, plb, plb_next); 239 } 240 241 static inline void 242 pmc_plb_rele(struct pmclog_buffer *plb) 243 { 244 mtx_lock_spin(&pmc_dom_hdrs[plb->plb_domain]->pdbh_mtx); 245 pmc_plb_rele_unlocked(plb); 246 mtx_unlock_spin(&pmc_dom_hdrs[plb->plb_domain]->pdbh_mtx); 247 } 248 249 250 /* 251 * Get a log buffer 252 */ 253 254 static int 255 pmclog_get_buffer(struct pmc_owner *po) 256 { 257 struct pmclog_buffer *plb; 258 int domain; 259 260 KASSERT(po->po_curbuf[curcpu] == NULL, 261 ("[pmclog,%d] po=%p current buffer still valid", __LINE__, po)); 262 263 domain = curdomain; 264 MPASS(pmc_dom_hdrs[domain]); 265 mtx_lock_spin(&pmc_dom_hdrs[domain]->pdbh_mtx); 266 if ((plb = TAILQ_FIRST(&pmc_dom_hdrs[domain]->pdbh_head)) != NULL) 267 TAILQ_REMOVE(&pmc_dom_hdrs[domain]->pdbh_head, plb, plb_next); 268 mtx_unlock_spin(&pmc_dom_hdrs[domain]->pdbh_mtx); 269 270 PMCDBG2(LOG,GTB,1, "po=%p plb=%p", po, plb); 271 272 #ifdef HWPMC_DEBUG 273 if (plb) 274 KASSERT(plb->plb_ptr == plb->plb_base && 275 plb->plb_base < plb->plb_fence, 276 ("[pmclog,%d] po=%p buffer invariants: ptr=%p " 277 "base=%p fence=%p", __LINE__, po, plb->plb_ptr, 278 plb->plb_base, plb->plb_fence)); 279 #endif 280 281 po->po_curbuf[curcpu] = plb; 282 283 /* update stats */ 284 counter_u64_add(pmc_stats.pm_buffer_requests, 1); 285 if (plb == NULL) 286 counter_u64_add(pmc_stats.pm_buffer_requests_failed, 1); 287 288 return (plb ? 0 : ENOMEM); 289 } 290 291 struct pmclog_proc_init_args { 292 struct proc *kthr; 293 struct pmc_owner *po; 294 bool exit; 295 bool acted; 296 }; 297 298 int 299 pmclog_proc_create(struct thread *td, void **handlep) 300 { 301 struct pmclog_proc_init_args *ia; 302 int error; 303 304 ia = malloc(sizeof(*ia), M_TEMP, M_WAITOK | M_ZERO); 305 error = kproc_create(pmclog_loop, ia, &ia->kthr, 306 RFHIGHPID, 0, "hwpmc: proc(%d)", td->td_proc->p_pid); 307 if (error == 0) 308 *handlep = ia; 309 return (error); 310 } 311 312 void 313 pmclog_proc_ignite(void *handle, struct pmc_owner *po) 314 { 315 struct pmclog_proc_init_args *ia; 316 317 ia = handle; 318 mtx_lock(&pmc_kthread_mtx); 319 MPASS(!ia->acted); 320 MPASS(ia->po == NULL); 321 MPASS(!ia->exit); 322 MPASS(ia->kthr != NULL); 323 if (po == NULL) { 324 ia->exit = true; 325 } else { 326 ia->po = po; 327 KASSERT(po->po_kthread == NULL, 328 ("[pmclog,%d] po=%p kthread (%p) already present", 329 __LINE__, po, po->po_kthread)); 330 po->po_kthread = ia->kthr; 331 } 332 wakeup(ia); 333 while (!ia->acted) 334 msleep(ia, &pmc_kthread_mtx, PWAIT, "pmclogw", 0); 335 mtx_unlock(&pmc_kthread_mtx); 336 free(ia, M_TEMP); 337 } 338 339 /* 340 * Log handler loop. 341 * 342 * This function is executed by each pmc owner's helper thread. 343 */ 344 345 static void 346 pmclog_loop(void *arg) 347 { 348 struct pmclog_proc_init_args *ia; 349 struct pmc_owner *po; 350 struct pmclog_buffer *lb; 351 struct proc *p; 352 struct ucred *ownercred; 353 struct ucred *mycred; 354 struct thread *td; 355 sigset_t unb; 356 struct uio auio; 357 struct iovec aiov; 358 size_t nbytes; 359 int error; 360 361 td = curthread; 362 363 SIGEMPTYSET(unb); 364 SIGADDSET(unb, SIGHUP); 365 (void)kern_sigprocmask(td, SIG_UNBLOCK, &unb, NULL, 0); 366 367 ia = arg; 368 MPASS(ia->kthr == curproc); 369 MPASS(!ia->acted); 370 mtx_lock(&pmc_kthread_mtx); 371 while (ia->po == NULL && !ia->exit) 372 msleep(ia, &pmc_kthread_mtx, PWAIT, "pmclogi", 0); 373 if (ia->exit) { 374 ia->acted = true; 375 wakeup(ia); 376 mtx_unlock(&pmc_kthread_mtx); 377 kproc_exit(0); 378 } 379 MPASS(ia->po != NULL); 380 po = ia->po; 381 ia->acted = true; 382 wakeup(ia); 383 mtx_unlock(&pmc_kthread_mtx); 384 ia = NULL; 385 386 p = po->po_owner; 387 mycred = td->td_ucred; 388 389 PROC_LOCK(p); 390 ownercred = crhold(p->p_ucred); 391 PROC_UNLOCK(p); 392 393 PMCDBG2(LOG,INI,1, "po=%p kt=%p", po, po->po_kthread); 394 KASSERT(po->po_kthread == curthread->td_proc, 395 ("[pmclog,%d] proc mismatch po=%p po/kt=%p curproc=%p", __LINE__, 396 po, po->po_kthread, curthread->td_proc)); 397 398 lb = NULL; 399 400 401 /* 402 * Loop waiting for I/O requests to be added to the owner 403 * struct's queue. The loop is exited when the log file 404 * is deconfigured. 405 */ 406 407 mtx_lock(&pmc_kthread_mtx); 408 409 for (;;) { 410 411 /* check if we've been asked to exit */ 412 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) 413 break; 414 415 if (lb == NULL) { /* look for a fresh buffer to write */ 416 mtx_lock_spin(&po->po_mtx); 417 if ((lb = TAILQ_FIRST(&po->po_logbuffers)) == NULL) { 418 mtx_unlock_spin(&po->po_mtx); 419 420 /* No more buffers and shutdown required. */ 421 if (po->po_flags & PMC_PO_SHUTDOWN) 422 break; 423 424 (void) msleep(po, &pmc_kthread_mtx, PWAIT, 425 "pmcloop", 0); 426 continue; 427 } 428 429 TAILQ_REMOVE(&po->po_logbuffers, lb, plb_next); 430 mtx_unlock_spin(&po->po_mtx); 431 } 432 433 mtx_unlock(&pmc_kthread_mtx); 434 435 /* process the request */ 436 PMCDBG3(LOG,WRI,2, "po=%p base=%p ptr=%p", po, 437 lb->plb_base, lb->plb_ptr); 438 /* change our thread's credentials before issuing the I/O */ 439 440 aiov.iov_base = lb->plb_base; 441 aiov.iov_len = nbytes = lb->plb_ptr - lb->plb_base; 442 443 auio.uio_iov = &aiov; 444 auio.uio_iovcnt = 1; 445 auio.uio_offset = -1; 446 auio.uio_resid = nbytes; 447 auio.uio_rw = UIO_WRITE; 448 auio.uio_segflg = UIO_SYSSPACE; 449 auio.uio_td = td; 450 451 /* switch thread credentials -- see kern_ktrace.c */ 452 td->td_ucred = ownercred; 453 error = fo_write(po->po_file, &auio, ownercred, 0, td); 454 td->td_ucred = mycred; 455 456 if (error) { 457 /* XXX some errors are recoverable */ 458 /* send a SIGIO to the owner and exit */ 459 PROC_LOCK(p); 460 kern_psignal(p, SIGIO); 461 PROC_UNLOCK(p); 462 463 mtx_lock(&pmc_kthread_mtx); 464 465 po->po_error = error; /* save for flush log */ 466 467 PMCDBG2(LOG,WRI,2, "po=%p error=%d", po, error); 468 469 break; 470 } 471 472 mtx_lock(&pmc_kthread_mtx); 473 474 /* put the used buffer back into the global pool */ 475 PMCLOG_RESET_BUFFER_DESCRIPTOR(lb); 476 477 pmc_plb_rele(lb); 478 lb = NULL; 479 } 480 481 wakeup_one(po->po_kthread); 482 po->po_kthread = NULL; 483 484 mtx_unlock(&pmc_kthread_mtx); 485 486 /* return the current I/O buffer to the global pool */ 487 if (lb) { 488 PMCLOG_RESET_BUFFER_DESCRIPTOR(lb); 489 490 pmc_plb_rele(lb); 491 } 492 493 /* 494 * Exit this thread, signalling the waiter 495 */ 496 497 crfree(ownercred); 498 499 kproc_exit(0); 500 } 501 502 /* 503 * Release and log entry and schedule an I/O if needed. 504 */ 505 506 static void 507 pmclog_release(struct pmc_owner *po) 508 { 509 struct pmclog_buffer *plb; 510 511 plb = po->po_curbuf[curcpu]; 512 KASSERT(plb->plb_ptr >= plb->plb_base, 513 ("[pmclog,%d] buffer invariants po=%p ptr=%p base=%p", __LINE__, 514 po, plb->plb_ptr, plb->plb_base)); 515 KASSERT(plb->plb_ptr <= plb->plb_fence, 516 ("[pmclog,%d] buffer invariants po=%p ptr=%p fenc=%p", __LINE__, 517 po, plb->plb_ptr, plb->plb_fence)); 518 519 /* schedule an I/O if we've filled a buffer */ 520 if (plb->plb_ptr >= plb->plb_fence) 521 pmclog_schedule_io(po); 522 523 PMCDBG1(LOG,REL,1, "po=%p", po); 524 } 525 526 527 /* 528 * Attempt to reserve 'length' bytes of space in an owner's log 529 * buffer. The function returns a pointer to 'length' bytes of space 530 * if there was enough space or returns NULL if no space was 531 * available. Non-null returns do so with the po mutex locked. The 532 * caller must invoke pmclog_release() on the pmc owner structure 533 * when done. 534 */ 535 536 static uint32_t * 537 pmclog_reserve(struct pmc_owner *po, int length) 538 { 539 uintptr_t newptr, oldptr; 540 uint32_t *lh; 541 struct timespec ts; 542 struct pmclog_buffer *plb, **pplb; 543 544 PMCDBG2(LOG,ALL,1, "po=%p len=%d", po, length); 545 546 KASSERT(length % sizeof(uint32_t) == 0, 547 ("[pmclog,%d] length not a multiple of word size", __LINE__)); 548 549 /* No more data when shutdown in progress. */ 550 if (po->po_flags & PMC_PO_SHUTDOWN) 551 return (NULL); 552 553 pplb = &po->po_curbuf[curcpu]; 554 if (*pplb == NULL && pmclog_get_buffer(po) != 0) 555 goto fail; 556 557 KASSERT(*pplb != NULL, 558 ("[pmclog,%d] po=%p no current buffer", __LINE__, po)); 559 560 plb = *pplb; 561 KASSERT(plb->plb_ptr >= plb->plb_base && 562 plb->plb_ptr <= plb->plb_fence, 563 ("[pmclog,%d] po=%p buffer invariants: ptr=%p base=%p fence=%p", 564 __LINE__, po, plb->plb_ptr, plb->plb_base, 565 plb->plb_fence)); 566 567 oldptr = (uintptr_t) plb->plb_ptr; 568 newptr = oldptr + length; 569 570 KASSERT(oldptr != (uintptr_t) NULL, 571 ("[pmclog,%d] po=%p Null log buffer pointer", __LINE__, po)); 572 573 /* 574 * If we have space in the current buffer, return a pointer to 575 * available space with the PO structure locked. 576 */ 577 if (newptr <= (uintptr_t) plb->plb_fence) { 578 plb->plb_ptr = (char *) newptr; 579 goto done; 580 } 581 582 /* 583 * Otherwise, schedule the current buffer for output and get a 584 * fresh buffer. 585 */ 586 pmclog_schedule_io(po); 587 588 if (pmclog_get_buffer(po) != 0) 589 goto fail; 590 591 plb = *pplb; 592 KASSERT(plb != NULL, 593 ("[pmclog,%d] po=%p no current buffer", __LINE__, po)); 594 595 KASSERT(plb->plb_ptr != NULL, 596 ("[pmclog,%d] null return from pmc_get_log_buffer", __LINE__)); 597 598 KASSERT(plb->plb_ptr == plb->plb_base && 599 plb->plb_ptr <= plb->plb_fence, 600 ("[pmclog,%d] po=%p buffer invariants: ptr=%p base=%p fence=%p", 601 __LINE__, po, plb->plb_ptr, plb->plb_base, 602 plb->plb_fence)); 603 604 oldptr = (uintptr_t) plb->plb_ptr; 605 606 done: 607 lh = (uint32_t *) oldptr; 608 lh++; /* skip header */ 609 getnanotime(&ts); /* fill in the timestamp */ 610 *lh++ = ts.tv_sec & 0xFFFFFFFF; 611 *lh++ = ts.tv_nsec & 0xFFFFFFF; 612 return ((uint32_t *) oldptr); 613 fail: 614 return (NULL); 615 } 616 617 /* 618 * Schedule an I/O. 619 * 620 * Transfer the current buffer to the helper kthread. 621 */ 622 623 static void 624 pmclog_schedule_io(struct pmc_owner *po) 625 { 626 struct pmclog_buffer *plb; 627 628 plb = po->po_curbuf[curcpu]; 629 po->po_curbuf[curcpu] = NULL; 630 KASSERT(plb != NULL, 631 ("[pmclog,%d] schedule_io with null buffer po=%p", __LINE__, po)); 632 KASSERT(plb->plb_ptr >= plb->plb_base, 633 ("[pmclog,%d] buffer invariants po=%p ptr=%p base=%p", __LINE__, 634 po, plb->plb_ptr, plb->plb_base)); 635 KASSERT(plb->plb_ptr <= plb->plb_fence, 636 ("[pmclog,%d] buffer invariants po=%p ptr=%p fenc=%p", __LINE__, 637 po, plb->plb_ptr, plb->plb_fence)); 638 639 PMCDBG1(LOG,SIO, 1, "po=%p", po); 640 641 /* 642 * Add the current buffer to the tail of the buffer list and 643 * wakeup the helper. 644 */ 645 mtx_lock_spin(&po->po_mtx); 646 TAILQ_INSERT_TAIL(&po->po_logbuffers, plb, plb_next); 647 mtx_unlock_spin(&po->po_mtx); 648 wakeup_one(po); 649 } 650 651 /* 652 * Stop the helper kthread. 653 */ 654 655 static void 656 pmclog_stop_kthread(struct pmc_owner *po) 657 { 658 659 mtx_lock(&pmc_kthread_mtx); 660 po->po_flags &= ~PMC_PO_OWNS_LOGFILE; 661 if (po->po_kthread != NULL) { 662 PROC_LOCK(po->po_kthread); 663 kern_psignal(po->po_kthread, SIGHUP); 664 PROC_UNLOCK(po->po_kthread); 665 } 666 wakeup_one(po); 667 while (po->po_kthread) 668 msleep(po->po_kthread, &pmc_kthread_mtx, PPAUSE, "pmckstp", 0); 669 mtx_unlock(&pmc_kthread_mtx); 670 } 671 672 /* 673 * Public functions 674 */ 675 676 /* 677 * Configure a log file for pmc owner 'po'. 678 * 679 * Parameter 'logfd' is a file handle referencing an open file in the 680 * owner process. This file needs to have been opened for writing. 681 */ 682 683 int 684 pmclog_configure_log(struct pmc_mdep *md, struct pmc_owner *po, int logfd) 685 { 686 struct proc *p; 687 int error; 688 689 sx_assert(&pmc_sx, SA_XLOCKED); 690 PMCDBG2(LOG,CFG,1, "config po=%p logfd=%d", po, logfd); 691 692 p = po->po_owner; 693 694 /* return EBUSY if a log file was already present */ 695 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 696 return (EBUSY); 697 698 KASSERT(po->po_file == NULL, 699 ("[pmclog,%d] po=%p file (%p) already present", __LINE__, po, 700 po->po_file)); 701 702 /* get a reference to the file state */ 703 error = fget_write(curthread, logfd, &cap_write_rights, &po->po_file); 704 if (error) 705 goto error; 706 707 /* mark process as owning a log file */ 708 po->po_flags |= PMC_PO_OWNS_LOGFILE; 709 710 /* mark process as using HWPMCs */ 711 PROC_LOCK(p); 712 p->p_flag |= P_HWPMC; 713 PROC_UNLOCK(p); 714 715 /* create a log initialization entry */ 716 PMCLOG_RESERVE_WITH_ERROR(po, INITIALIZE, 717 sizeof(struct pmclog_initialize)); 718 PMCLOG_EMIT32(PMC_VERSION); 719 PMCLOG_EMIT32(md->pmd_cputype); 720 PMCLOG_DESPATCH_SYNC(po); 721 722 return (0); 723 724 error: 725 KASSERT(po->po_kthread == NULL, ("[pmclog,%d] po=%p kthread not " 726 "stopped", __LINE__, po)); 727 728 if (po->po_file) 729 (void) fdrop(po->po_file, curthread); 730 po->po_file = NULL; /* clear file and error state */ 731 po->po_error = 0; 732 po->po_flags &= ~PMC_PO_OWNS_LOGFILE; 733 734 return (error); 735 } 736 737 738 /* 739 * De-configure a log file. This will throw away any buffers queued 740 * for this owner process. 741 */ 742 743 int 744 pmclog_deconfigure_log(struct pmc_owner *po) 745 { 746 int error; 747 struct pmclog_buffer *lb; 748 749 PMCDBG1(LOG,CFG,1, "de-config po=%p", po); 750 751 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) 752 return (EINVAL); 753 754 KASSERT(po->po_sscount == 0, 755 ("[pmclog,%d] po=%p still owning SS PMCs", __LINE__, po)); 756 KASSERT(po->po_file != NULL, 757 ("[pmclog,%d] po=%p no log file", __LINE__, po)); 758 759 /* stop the kthread, this will reset the 'OWNS_LOGFILE' flag */ 760 pmclog_stop_kthread(po); 761 762 KASSERT(po->po_kthread == NULL, 763 ("[pmclog,%d] po=%p kthread not stopped", __LINE__, po)); 764 765 /* return all queued log buffers to the global pool */ 766 while ((lb = TAILQ_FIRST(&po->po_logbuffers)) != NULL) { 767 TAILQ_REMOVE(&po->po_logbuffers, lb, plb_next); 768 PMCLOG_RESET_BUFFER_DESCRIPTOR(lb); 769 pmc_plb_rele(lb); 770 } 771 for (int i = 0; i < mp_ncpus; i++) { 772 thread_lock(curthread); 773 sched_bind(curthread, i); 774 thread_unlock(curthread); 775 /* return the 'current' buffer to the global pool */ 776 if ((lb = po->po_curbuf[curcpu]) != NULL) { 777 PMCLOG_RESET_BUFFER_DESCRIPTOR(lb); 778 pmc_plb_rele(lb); 779 } 780 } 781 thread_lock(curthread); 782 sched_unbind(curthread); 783 thread_unlock(curthread); 784 785 /* drop a reference to the fd */ 786 if (po->po_file != NULL) { 787 error = fdrop(po->po_file, curthread); 788 po->po_file = NULL; 789 } else 790 error = 0; 791 po->po_error = 0; 792 793 return (error); 794 } 795 796 /* 797 * Flush a process' log buffer. 798 */ 799 800 int 801 pmclog_flush(struct pmc_owner *po) 802 { 803 int error; 804 805 PMCDBG1(LOG,FLS,1, "po=%p", po); 806 807 /* 808 * If there is a pending error recorded by the logger thread, 809 * return that. 810 */ 811 if (po->po_error) 812 return (po->po_error); 813 814 error = 0; 815 816 /* 817 * Check that we do have an active log file. 818 */ 819 mtx_lock(&pmc_kthread_mtx); 820 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) { 821 error = EINVAL; 822 goto error; 823 } 824 825 pmclog_schedule_all(po); 826 error: 827 mtx_unlock(&pmc_kthread_mtx); 828 829 return (error); 830 } 831 832 static void 833 pmclog_schedule_one_cond(void *arg) 834 { 835 struct pmc_owner *po = arg; 836 struct pmclog_buffer *plb; 837 838 spinlock_enter(); 839 /* tell hardclock not to run again */ 840 if (PMC_CPU_HAS_SAMPLES(PCPU_GET(cpuid))) 841 PMC_CALL_HOOK_UNLOCKED(curthread, PMC_FN_DO_SAMPLES, NULL); 842 plb = po->po_curbuf[curcpu]; 843 if (plb && plb->plb_ptr != plb->plb_base) 844 pmclog_schedule_io(po); 845 spinlock_exit(); 846 } 847 848 static void 849 pmclog_schedule_all(struct pmc_owner *po) 850 { 851 /* 852 * Schedule the current buffer if any and not empty. 853 */ 854 for (int i = 0; i < mp_ncpus; i++) { 855 thread_lock(curthread); 856 sched_bind(curthread, i); 857 thread_unlock(curthread); 858 pmclog_schedule_one_cond(po); 859 } 860 thread_lock(curthread); 861 sched_unbind(curthread); 862 thread_unlock(curthread); 863 } 864 865 int 866 pmclog_close(struct pmc_owner *po) 867 { 868 869 PMCDBG1(LOG,CLO,1, "po=%p", po); 870 871 pmclog_process_closelog(po); 872 873 mtx_lock(&pmc_kthread_mtx); 874 /* 875 * Initiate shutdown: no new data queued, 876 * thread will close file on last block. 877 */ 878 po->po_flags |= PMC_PO_SHUTDOWN; 879 /* give time for all to see */ 880 DELAY(50); 881 882 /* 883 * Schedule the current buffer. 884 */ 885 pmclog_schedule_all(po); 886 wakeup_one(po); 887 888 mtx_unlock(&pmc_kthread_mtx); 889 890 return (0); 891 } 892 893 void 894 pmclog_process_callchain(struct pmc *pm, struct pmc_sample *ps) 895 { 896 int n, recordlen; 897 uint32_t flags; 898 struct pmc_owner *po; 899 900 PMCDBG3(LOG,SAM,1,"pm=%p pid=%d n=%d", pm, ps->ps_pid, 901 ps->ps_nsamples); 902 903 recordlen = offsetof(struct pmclog_callchain, pl_pc) + 904 ps->ps_nsamples * sizeof(uintfptr_t); 905 po = pm->pm_owner; 906 flags = PMC_CALLCHAIN_TO_CPUFLAGS(ps->ps_cpu,ps->ps_flags); 907 PMCLOG_RESERVE_SAFE(po, CALLCHAIN, recordlen); 908 PMCLOG_EMIT32(ps->ps_pid); 909 PMCLOG_EMIT32(ps->ps_tid); 910 PMCLOG_EMIT32(pm->pm_id); 911 PMCLOG_EMIT32(flags); 912 /* unused for now */ 913 PMCLOG_EMIT32(0); 914 for (n = 0; n < ps->ps_nsamples; n++) 915 PMCLOG_EMITADDR(ps->ps_pc[n]); 916 PMCLOG_DESPATCH_SAFE(po); 917 } 918 919 void 920 pmclog_process_closelog(struct pmc_owner *po) 921 { 922 PMCLOG_RESERVE(po,CLOSELOG,sizeof(struct pmclog_closelog)); 923 PMCLOG_DESPATCH_SYNC(po); 924 } 925 926 void 927 pmclog_process_dropnotify(struct pmc_owner *po) 928 { 929 PMCLOG_RESERVE(po,DROPNOTIFY,sizeof(struct pmclog_dropnotify)); 930 PMCLOG_DESPATCH(po); 931 } 932 933 void 934 pmclog_process_map_in(struct pmc_owner *po, pid_t pid, uintfptr_t start, 935 const char *path) 936 { 937 int pathlen, recordlen; 938 939 KASSERT(path != NULL, ("[pmclog,%d] map-in, null path", __LINE__)); 940 941 pathlen = strlen(path) + 1; /* #bytes for path name */ 942 recordlen = offsetof(struct pmclog_map_in, pl_pathname) + 943 pathlen; 944 945 PMCLOG_RESERVE(po, MAP_IN, recordlen); 946 PMCLOG_EMIT32(pid); 947 PMCLOG_EMITADDR(start); 948 PMCLOG_EMITSTRING(path,pathlen); 949 PMCLOG_DESPATCH(po); 950 } 951 952 void 953 pmclog_process_map_out(struct pmc_owner *po, pid_t pid, uintfptr_t start, 954 uintfptr_t end) 955 { 956 KASSERT(start <= end, ("[pmclog,%d] start > end", __LINE__)); 957 958 PMCLOG_RESERVE(po, MAP_OUT, sizeof(struct pmclog_map_out)); 959 PMCLOG_EMIT32(pid); 960 PMCLOG_EMITADDR(start); 961 PMCLOG_EMITADDR(end); 962 PMCLOG_DESPATCH(po); 963 } 964 965 void 966 pmclog_process_pmcallocate(struct pmc *pm) 967 { 968 struct pmc_owner *po; 969 struct pmc_soft *ps; 970 971 po = pm->pm_owner; 972 973 PMCDBG1(LOG,ALL,1, "pm=%p", pm); 974 975 if (PMC_TO_CLASS(pm) == PMC_CLASS_SOFT) { 976 PMCLOG_RESERVE(po, PMCALLOCATEDYN, 977 sizeof(struct pmclog_pmcallocatedyn)); 978 PMCLOG_EMIT32(pm->pm_id); 979 PMCLOG_EMIT32(pm->pm_event); 980 PMCLOG_EMIT32(pm->pm_flags); 981 ps = pmc_soft_ev_acquire(pm->pm_event); 982 if (ps != NULL) 983 PMCLOG_EMITSTRING(ps->ps_ev.pm_ev_name,PMC_NAME_MAX); 984 else 985 PMCLOG_EMITNULLSTRING(PMC_NAME_MAX); 986 pmc_soft_ev_release(ps); 987 PMCLOG_DESPATCH_SYNC(po); 988 } else { 989 PMCLOG_RESERVE(po, PMCALLOCATE, 990 sizeof(struct pmclog_pmcallocate)); 991 PMCLOG_EMIT32(pm->pm_id); 992 PMCLOG_EMIT32(pm->pm_event); 993 PMCLOG_EMIT32(pm->pm_flags); 994 PMCLOG_DESPATCH_SYNC(po); 995 } 996 } 997 998 void 999 pmclog_process_pmcattach(struct pmc *pm, pid_t pid, char *path) 1000 { 1001 int pathlen, recordlen; 1002 struct pmc_owner *po; 1003 1004 PMCDBG2(LOG,ATT,1,"pm=%p pid=%d", pm, pid); 1005 1006 po = pm->pm_owner; 1007 1008 pathlen = strlen(path) + 1; /* #bytes for the string */ 1009 recordlen = offsetof(struct pmclog_pmcattach, pl_pathname) + pathlen; 1010 1011 PMCLOG_RESERVE(po, PMCATTACH, recordlen); 1012 PMCLOG_EMIT32(pm->pm_id); 1013 PMCLOG_EMIT32(pid); 1014 PMCLOG_EMIT32(0); 1015 PMCLOG_EMITSTRING(path, pathlen); 1016 PMCLOG_DESPATCH_SYNC(po); 1017 } 1018 1019 void 1020 pmclog_process_pmcdetach(struct pmc *pm, pid_t pid) 1021 { 1022 struct pmc_owner *po; 1023 1024 PMCDBG2(LOG,ATT,1,"!pm=%p pid=%d", pm, pid); 1025 1026 po = pm->pm_owner; 1027 1028 PMCLOG_RESERVE(po, PMCDETACH, sizeof(struct pmclog_pmcdetach)); 1029 PMCLOG_EMIT32(pm->pm_id); 1030 PMCLOG_EMIT32(pid); 1031 PMCLOG_DESPATCH_SYNC(po); 1032 } 1033 1034 /* 1035 * Log a context switch event to the log file. 1036 */ 1037 1038 void 1039 pmclog_process_proccsw(struct pmc *pm, struct pmc_process *pp, pmc_value_t v, struct thread *td) 1040 { 1041 struct pmc_owner *po; 1042 1043 KASSERT(pm->pm_flags & PMC_F_LOG_PROCCSW, 1044 ("[pmclog,%d] log-process-csw called gratuitously", __LINE__)); 1045 1046 PMCDBG3(LOG,SWO,1,"pm=%p pid=%d v=%jx", pm, pp->pp_proc->p_pid, 1047 v); 1048 1049 po = pm->pm_owner; 1050 1051 PMCLOG_RESERVE(po, PROCCSW, sizeof(struct pmclog_proccsw)); 1052 PMCLOG_EMIT32(pm->pm_id); 1053 PMCLOG_EMIT64(v); 1054 PMCLOG_EMIT32(pp->pp_proc->p_pid); 1055 PMCLOG_EMIT32(td->td_tid); 1056 PMCLOG_DESPATCH(po); 1057 } 1058 1059 void 1060 pmclog_process_procexec(struct pmc_owner *po, pmc_id_t pmid, pid_t pid, 1061 uintfptr_t startaddr, char *path) 1062 { 1063 int pathlen, recordlen; 1064 1065 PMCDBG3(LOG,EXC,1,"po=%p pid=%d path=\"%s\"", po, pid, path); 1066 1067 pathlen = strlen(path) + 1; /* #bytes for the path */ 1068 recordlen = offsetof(struct pmclog_procexec, pl_pathname) + pathlen; 1069 1070 PMCLOG_RESERVE(po, PROCEXEC, recordlen); 1071 PMCLOG_EMIT32(pid); 1072 PMCLOG_EMIT32(pmid); 1073 PMCLOG_EMIT32(0); 1074 PMCLOG_EMITADDR(startaddr); 1075 PMCLOG_EMITSTRING(path,pathlen); 1076 PMCLOG_DESPATCH(po); 1077 } 1078 1079 /* 1080 * Log a process exit event (and accumulated pmc value) to the log file. 1081 */ 1082 1083 void 1084 pmclog_process_procexit(struct pmc *pm, struct pmc_process *pp) 1085 { 1086 int ri; 1087 struct pmc_owner *po; 1088 1089 ri = PMC_TO_ROWINDEX(pm); 1090 PMCDBG3(LOG,EXT,1,"pm=%p pid=%d v=%jx", pm, pp->pp_proc->p_pid, 1091 pp->pp_pmcs[ri].pp_pmcval); 1092 1093 po = pm->pm_owner; 1094 1095 PMCLOG_RESERVE(po, PROCEXIT, sizeof(struct pmclog_procexit)); 1096 PMCLOG_EMIT32(pm->pm_id); 1097 PMCLOG_EMIT32(pp->pp_proc->p_pid); 1098 PMCLOG_EMIT32(0); 1099 PMCLOG_EMIT64(pp->pp_pmcs[ri].pp_pmcval); 1100 PMCLOG_DESPATCH(po); 1101 } 1102 1103 /* 1104 * Log a fork event. 1105 */ 1106 1107 void 1108 pmclog_process_procfork(struct pmc_owner *po, pid_t oldpid, pid_t newpid) 1109 { 1110 PMCLOG_RESERVE(po, PROCFORK, sizeof(struct pmclog_procfork)); 1111 PMCLOG_EMIT32(oldpid); 1112 PMCLOG_EMIT32(newpid); 1113 PMCLOG_DESPATCH(po); 1114 } 1115 1116 /* 1117 * Log a process exit event of the form suitable for system-wide PMCs. 1118 */ 1119 1120 void 1121 pmclog_process_sysexit(struct pmc_owner *po, pid_t pid) 1122 { 1123 PMCLOG_RESERVE(po, SYSEXIT, sizeof(struct pmclog_sysexit)); 1124 PMCLOG_EMIT32(pid); 1125 PMCLOG_DESPATCH(po); 1126 } 1127 1128 /* 1129 * Write a user log entry. 1130 */ 1131 1132 int 1133 pmclog_process_userlog(struct pmc_owner *po, struct pmc_op_writelog *wl) 1134 { 1135 int error; 1136 1137 PMCDBG2(LOG,WRI,1, "writelog po=%p ud=0x%x", po, wl->pm_userdata); 1138 1139 error = 0; 1140 1141 PMCLOG_RESERVE_WITH_ERROR(po, USERDATA, 1142 sizeof(struct pmclog_userdata)); 1143 PMCLOG_EMIT32(wl->pm_userdata); 1144 PMCLOG_DESPATCH(po); 1145 1146 error: 1147 return (error); 1148 } 1149 1150 /* 1151 * Initialization. 1152 * 1153 * Create a pool of log buffers and initialize mutexes. 1154 */ 1155 1156 void 1157 pmclog_initialize() 1158 { 1159 int domain; 1160 struct pmclog_buffer *plb; 1161 1162 if (pmclog_buffer_size <= 0 || pmclog_buffer_size > 16*1024) { 1163 (void) printf("hwpmc: tunable logbuffersize=%d must be " 1164 "greater than zero and less than or equal to 16MB.\n", 1165 pmclog_buffer_size); 1166 pmclog_buffer_size = PMC_LOG_BUFFER_SIZE; 1167 } 1168 1169 if (pmc_nlogbuffers_pcpu <= 0) { 1170 (void) printf("hwpmc: tunable nlogbuffers=%d must be greater " 1171 "than zero.\n", pmc_nlogbuffers_pcpu); 1172 pmc_nlogbuffers_pcpu = PMC_NLOGBUFFERS_PCPU; 1173 } 1174 if (pmc_nlogbuffers_pcpu*pmclog_buffer_size > 32*1024) { 1175 (void) printf("hwpmc: memory allocated pcpu must be less than 32MB (is %dK).\n", 1176 pmc_nlogbuffers_pcpu*pmclog_buffer_size); 1177 pmc_nlogbuffers_pcpu = PMC_NLOGBUFFERS_PCPU; 1178 pmclog_buffer_size = PMC_LOG_BUFFER_SIZE; 1179 } 1180 for (domain = 0; domain < NDOMAINS; domain++) { 1181 int ncpus = pmc_dom_hdrs[domain]->pdbh_ncpus; 1182 int total = ncpus*pmc_nlogbuffers_pcpu; 1183 1184 plb = malloc_domain(sizeof(struct pmclog_buffer)*total, M_PMC, domain, M_WAITOK|M_ZERO); 1185 pmc_dom_hdrs[domain]->pdbh_plbs = plb; 1186 for (int i = 0; i < total; i++, plb++) { 1187 void *buf; 1188 1189 buf = malloc_domain(1024 * pmclog_buffer_size, M_PMC, domain, 1190 M_WAITOK|M_ZERO); 1191 PMCLOG_INIT_BUFFER_DESCRIPTOR(plb, buf, domain); 1192 pmc_plb_rele_unlocked(plb); 1193 } 1194 } 1195 mtx_init(&pmc_kthread_mtx, "pmc-kthread", "pmc-sleep", MTX_DEF); 1196 } 1197 1198 /* 1199 * Shutdown logging. 1200 * 1201 * Destroy mutexes and release memory back the to free pool. 1202 */ 1203 1204 void 1205 pmclog_shutdown() 1206 { 1207 struct pmclog_buffer *plb; 1208 int domain; 1209 1210 mtx_destroy(&pmc_kthread_mtx); 1211 1212 for (domain = 0; domain < NDOMAINS; domain++) { 1213 while ((plb = TAILQ_FIRST(&pmc_dom_hdrs[domain]->pdbh_head)) != NULL) { 1214 TAILQ_REMOVE(&pmc_dom_hdrs[domain]->pdbh_head, plb, plb_next); 1215 free(plb->plb_base, M_PMC); 1216 } 1217 free(pmc_dom_hdrs[domain]->pdbh_plbs, M_PMC); 1218 } 1219 } 1220