1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD 3 * 4 * Copyright (c) 2003-2008 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 #include <sys/cdefs.h> 36 __FBSDID("$FreeBSD$"); 37 38 #include <sys/param.h> 39 #include <sys/eventhandler.h> 40 #include <sys/gtaskqueue.h> 41 #include <sys/jail.h> 42 #include <sys/kernel.h> 43 #include <sys/kthread.h> 44 #include <sys/limits.h> 45 #include <sys/lock.h> 46 #include <sys/malloc.h> 47 #include <sys/module.h> 48 #include <sys/mount.h> 49 #include <sys/mutex.h> 50 #include <sys/pmc.h> 51 #include <sys/pmckern.h> 52 #include <sys/pmclog.h> 53 #include <sys/priv.h> 54 #include <sys/proc.h> 55 #include <sys/queue.h> 56 #include <sys/resourcevar.h> 57 #include <sys/rwlock.h> 58 #include <sys/sched.h> 59 #include <sys/signalvar.h> 60 #include <sys/smp.h> 61 #include <sys/sx.h> 62 #include <sys/sysctl.h> 63 #include <sys/sysent.h> 64 #include <sys/systm.h> 65 #include <sys/vnode.h> 66 67 #include <sys/linker.h> /* needs to be after <sys/malloc.h> */ 68 69 #include <machine/atomic.h> 70 #include <machine/md_var.h> 71 72 #include <vm/vm.h> 73 #include <vm/vm_extern.h> 74 #include <vm/pmap.h> 75 #include <vm/vm_map.h> 76 #include <vm/vm_object.h> 77 78 #include "hwpmc_soft.h" 79 80 #ifdef NUMA 81 #define NDOMAINS vm_ndomains 82 #else 83 #define NDOMAINS 1 84 #define malloc_domain(size, type, domain, flags) malloc((size), (type), (flags)) 85 #define free_domain(addr, type) free(addr, type) 86 #endif 87 88 /* 89 * Types 90 */ 91 92 enum pmc_flags { 93 PMC_FLAG_NONE = 0x00, /* do nothing */ 94 PMC_FLAG_REMOVE = 0x01, /* atomically remove entry from hash */ 95 PMC_FLAG_ALLOCATE = 0x02, /* add entry to hash if not found */ 96 PMC_FLAG_NOWAIT = 0x04, /* do not wait for mallocs */ 97 }; 98 99 /* 100 * The offset in sysent where the syscall is allocated. 101 */ 102 103 static int pmc_syscall_num = NO_SYSCALL; 104 struct pmc_cpu **pmc_pcpu; /* per-cpu state */ 105 pmc_value_t *pmc_pcpu_saved; /* saved PMC values: CSW handling */ 106 107 #define PMC_PCPU_SAVED(C,R) pmc_pcpu_saved[(R) + md->pmd_npmc*(C)] 108 109 struct mtx_pool *pmc_mtxpool; 110 static int *pmc_pmcdisp; /* PMC row dispositions */ 111 112 #define PMC_ROW_DISP_IS_FREE(R) (pmc_pmcdisp[(R)] == 0) 113 #define PMC_ROW_DISP_IS_THREAD(R) (pmc_pmcdisp[(R)] > 0) 114 #define PMC_ROW_DISP_IS_STANDALONE(R) (pmc_pmcdisp[(R)] < 0) 115 116 #define PMC_MARK_ROW_FREE(R) do { \ 117 pmc_pmcdisp[(R)] = 0; \ 118 } while (0) 119 120 #define PMC_MARK_ROW_STANDALONE(R) do { \ 121 KASSERT(pmc_pmcdisp[(R)] <= 0, ("[pmc,%d] row disposition error", \ 122 __LINE__)); \ 123 atomic_add_int(&pmc_pmcdisp[(R)], -1); \ 124 KASSERT(pmc_pmcdisp[(R)] >= (-pmc_cpu_max_active()), \ 125 ("[pmc,%d] row disposition error", __LINE__)); \ 126 } while (0) 127 128 #define PMC_UNMARK_ROW_STANDALONE(R) do { \ 129 atomic_add_int(&pmc_pmcdisp[(R)], 1); \ 130 KASSERT(pmc_pmcdisp[(R)] <= 0, ("[pmc,%d] row disposition error", \ 131 __LINE__)); \ 132 } while (0) 133 134 #define PMC_MARK_ROW_THREAD(R) do { \ 135 KASSERT(pmc_pmcdisp[(R)] >= 0, ("[pmc,%d] row disposition error", \ 136 __LINE__)); \ 137 atomic_add_int(&pmc_pmcdisp[(R)], 1); \ 138 } while (0) 139 140 #define PMC_UNMARK_ROW_THREAD(R) do { \ 141 atomic_add_int(&pmc_pmcdisp[(R)], -1); \ 142 KASSERT(pmc_pmcdisp[(R)] >= 0, ("[pmc,%d] row disposition error", \ 143 __LINE__)); \ 144 } while (0) 145 146 147 /* various event handlers */ 148 static eventhandler_tag pmc_exit_tag, pmc_fork_tag, pmc_kld_load_tag, 149 pmc_kld_unload_tag; 150 151 /* Module statistics */ 152 struct pmc_driverstats pmc_stats; 153 154 155 /* Machine/processor dependent operations */ 156 static struct pmc_mdep *md; 157 158 /* 159 * Hash tables mapping owner processes and target threads to PMCs. 160 */ 161 162 struct mtx pmc_processhash_mtx; /* spin mutex */ 163 static u_long pmc_processhashmask; 164 static LIST_HEAD(pmc_processhash, pmc_process) *pmc_processhash; 165 166 /* 167 * Hash table of PMC owner descriptors. This table is protected by 168 * the shared PMC "sx" lock. 169 */ 170 171 static u_long pmc_ownerhashmask; 172 static LIST_HEAD(pmc_ownerhash, pmc_owner) *pmc_ownerhash; 173 174 /* 175 * List of PMC owners with system-wide sampling PMCs. 176 */ 177 178 static LIST_HEAD(, pmc_owner) pmc_ss_owners; 179 180 /* 181 * List of free thread entries. This is protected by the spin 182 * mutex. 183 */ 184 static struct mtx pmc_threadfreelist_mtx; /* spin mutex */ 185 static LIST_HEAD(, pmc_thread) pmc_threadfreelist; 186 static int pmc_threadfreelist_entries=0; 187 #define THREADENTRY_SIZE \ 188 (sizeof(struct pmc_thread) + (md->pmd_npmc * sizeof(struct pmc_threadpmcstate))) 189 190 /* 191 * Task to free thread descriptors 192 */ 193 static struct grouptask free_gtask; 194 195 /* 196 * A map of row indices to classdep structures. 197 */ 198 static struct pmc_classdep **pmc_rowindex_to_classdep; 199 200 /* 201 * Prototypes 202 */ 203 204 #ifdef HWPMC_DEBUG 205 static int pmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS); 206 static int pmc_debugflags_parse(char *newstr, char *fence); 207 #endif 208 209 static int load(struct module *module, int cmd, void *arg); 210 static void pmc_add_thread_descriptors_from_proc(struct proc *p, 211 struct pmc_process *pp); 212 static int pmc_attach_process(struct proc *p, struct pmc *pm); 213 static struct pmc *pmc_allocate_pmc_descriptor(void); 214 static struct pmc_owner *pmc_allocate_owner_descriptor(struct proc *p); 215 static int pmc_attach_one_process(struct proc *p, struct pmc *pm); 216 static int pmc_can_allocate_rowindex(struct proc *p, unsigned int ri, 217 int cpu); 218 static int pmc_can_attach(struct pmc *pm, struct proc *p); 219 static void pmc_capture_user_callchain(int cpu, int soft, struct trapframe *tf); 220 static void pmc_cleanup(void); 221 static int pmc_detach_process(struct proc *p, struct pmc *pm); 222 static int pmc_detach_one_process(struct proc *p, struct pmc *pm, 223 int flags); 224 static void pmc_destroy_owner_descriptor(struct pmc_owner *po); 225 static void pmc_destroy_pmc_descriptor(struct pmc *pm); 226 static void pmc_destroy_process_descriptor(struct pmc_process *pp); 227 static struct pmc_owner *pmc_find_owner_descriptor(struct proc *p); 228 static int pmc_find_pmc(pmc_id_t pmcid, struct pmc **pm); 229 static struct pmc *pmc_find_pmc_descriptor_in_process(struct pmc_owner *po, 230 pmc_id_t pmc); 231 static struct pmc_process *pmc_find_process_descriptor(struct proc *p, 232 uint32_t mode); 233 static struct pmc_thread *pmc_find_thread_descriptor(struct pmc_process *pp, 234 struct thread *td, uint32_t mode); 235 static void pmc_force_context_switch(void); 236 static void pmc_link_target_process(struct pmc *pm, 237 struct pmc_process *pp); 238 static void pmc_log_all_process_mappings(struct pmc_owner *po); 239 static void pmc_log_kernel_mappings(struct pmc *pm); 240 static void pmc_log_process_mappings(struct pmc_owner *po, struct proc *p); 241 static void pmc_maybe_remove_owner(struct pmc_owner *po); 242 static void pmc_process_csw_in(struct thread *td); 243 static void pmc_process_csw_out(struct thread *td); 244 static void pmc_process_exit(void *arg, struct proc *p); 245 static void pmc_process_fork(void *arg, struct proc *p1, 246 struct proc *p2, int n); 247 static void pmc_process_samples(int cpu, int soft); 248 static void pmc_release_pmc_descriptor(struct pmc *pmc); 249 static void pmc_process_thread_add(struct thread *td); 250 static void pmc_process_thread_delete(struct thread *td); 251 static void pmc_remove_owner(struct pmc_owner *po); 252 static void pmc_remove_process_descriptor(struct pmc_process *pp); 253 static void pmc_restore_cpu_binding(struct pmc_binding *pb); 254 static void pmc_save_cpu_binding(struct pmc_binding *pb); 255 static void pmc_select_cpu(int cpu); 256 static int pmc_start(struct pmc *pm); 257 static int pmc_stop(struct pmc *pm); 258 static int pmc_syscall_handler(struct thread *td, void *syscall_args); 259 static struct pmc_thread *pmc_thread_descriptor_pool_alloc(void); 260 static void pmc_thread_descriptor_pool_drain(void); 261 static void pmc_thread_descriptor_pool_free(struct pmc_thread *pt); 262 static void pmc_unlink_target_process(struct pmc *pmc, 263 struct pmc_process *pp); 264 static int generic_switch_in(struct pmc_cpu *pc, struct pmc_process *pp); 265 static int generic_switch_out(struct pmc_cpu *pc, struct pmc_process *pp); 266 static struct pmc_mdep *pmc_generic_cpu_initialize(void); 267 static void pmc_generic_cpu_finalize(struct pmc_mdep *md); 268 269 /* 270 * Kernel tunables and sysctl(8) interface. 271 */ 272 273 SYSCTL_DECL(_kern_hwpmc); 274 SYSCTL_NODE(_kern_hwpmc, OID_AUTO, stats, CTLFLAG_RW, 0, "HWPMC stats"); 275 276 277 /* Stats. */ 278 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, intr_ignored, CTLFLAG_RW, 279 &pmc_stats.pm_intr_ignored, "# of interrupts ignored"); 280 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, intr_processed, CTLFLAG_RW, 281 &pmc_stats.pm_intr_processed, "# of interrupts processed"); 282 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, intr_bufferfull, CTLFLAG_RW, 283 &pmc_stats.pm_intr_bufferfull, "# of interrupts where buffer was full"); 284 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, syscalls, CTLFLAG_RW, 285 &pmc_stats.pm_syscalls, "# of syscalls"); 286 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, syscall_errors, CTLFLAG_RW, 287 &pmc_stats.pm_syscall_errors, "# of syscall_errors"); 288 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, buffer_requests, CTLFLAG_RW, 289 &pmc_stats.pm_buffer_requests, "# of buffer requests"); 290 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, buffer_requests_failed, CTLFLAG_RW, 291 &pmc_stats.pm_buffer_requests_failed, "# of buffer requests which failed"); 292 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, log_sweeps, CTLFLAG_RW, 293 &pmc_stats.pm_log_sweeps, "# of ?"); 294 295 static int pmc_callchaindepth = PMC_CALLCHAIN_DEPTH; 296 SYSCTL_INT(_kern_hwpmc, OID_AUTO, callchaindepth, CTLFLAG_RDTUN, 297 &pmc_callchaindepth, 0, "depth of call chain records"); 298 299 char pmc_cpuid[64]; 300 SYSCTL_STRING(_kern_hwpmc, OID_AUTO, cpuid, CTLFLAG_RD, 301 pmc_cpuid, 0, "cpu version string"); 302 #ifdef HWPMC_DEBUG 303 struct pmc_debugflags pmc_debugflags = PMC_DEBUG_DEFAULT_FLAGS; 304 char pmc_debugstr[PMC_DEBUG_STRSIZE]; 305 TUNABLE_STR(PMC_SYSCTL_NAME_PREFIX "debugflags", pmc_debugstr, 306 sizeof(pmc_debugstr)); 307 SYSCTL_PROC(_kern_hwpmc, OID_AUTO, debugflags, 308 CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NOFETCH, 309 0, 0, pmc_debugflags_sysctl_handler, "A", "debug flags"); 310 #endif 311 312 313 /* 314 * kern.hwpmc.hashrows -- determines the number of rows in the 315 * of the hash table used to look up threads 316 */ 317 318 static int pmc_hashsize = PMC_HASH_SIZE; 319 SYSCTL_INT(_kern_hwpmc, OID_AUTO, hashsize, CTLFLAG_RDTUN, 320 &pmc_hashsize, 0, "rows in hash tables"); 321 322 /* 323 * kern.hwpmc.nsamples --- number of PC samples/callchain stacks per CPU 324 */ 325 326 static int pmc_nsamples = PMC_NSAMPLES; 327 SYSCTL_INT(_kern_hwpmc, OID_AUTO, nsamples, CTLFLAG_RDTUN, 328 &pmc_nsamples, 0, "number of PC samples per CPU"); 329 330 331 /* 332 * kern.hwpmc.mtxpoolsize -- number of mutexes in the mutex pool. 333 */ 334 335 static int pmc_mtxpool_size = PMC_MTXPOOL_SIZE; 336 SYSCTL_INT(_kern_hwpmc, OID_AUTO, mtxpoolsize, CTLFLAG_RDTUN, 337 &pmc_mtxpool_size, 0, "size of spin mutex pool"); 338 339 340 /* 341 * kern.hwpmc.threadfreelist_entries -- number of free entries 342 */ 343 344 SYSCTL_INT(_kern_hwpmc, OID_AUTO, threadfreelist_entries, CTLFLAG_RD, 345 &pmc_threadfreelist_entries, 0, "number of avalable thread entries"); 346 347 348 /* 349 * kern.hwpmc.threadfreelist_max -- maximum number of free entries 350 */ 351 352 static int pmc_threadfreelist_max = PMC_THREADLIST_MAX; 353 SYSCTL_INT(_kern_hwpmc, OID_AUTO, threadfreelist_max, CTLFLAG_RW, 354 &pmc_threadfreelist_max, 0, 355 "maximum number of available thread entries before freeing some"); 356 357 358 /* 359 * security.bsd.unprivileged_syspmcs -- allow non-root processes to 360 * allocate system-wide PMCs. 361 * 362 * Allowing unprivileged processes to allocate system PMCs is convenient 363 * if system-wide measurements need to be taken concurrently with other 364 * per-process measurements. This feature is turned off by default. 365 */ 366 367 static int pmc_unprivileged_syspmcs = 0; 368 SYSCTL_INT(_security_bsd, OID_AUTO, unprivileged_syspmcs, CTLFLAG_RWTUN, 369 &pmc_unprivileged_syspmcs, 0, 370 "allow unprivileged process to allocate system PMCs"); 371 372 /* 373 * Hash function. Discard the lower 2 bits of the pointer since 374 * these are always zero for our uses. The hash multiplier is 375 * round((2^LONG_BIT) * ((sqrt(5)-1)/2)). 376 */ 377 378 #if LONG_BIT == 64 379 #define _PMC_HM 11400714819323198486u 380 #elif LONG_BIT == 32 381 #define _PMC_HM 2654435769u 382 #else 383 #error Must know the size of 'long' to compile 384 #endif 385 386 #define PMC_HASH_PTR(P,M) ((((unsigned long) (P) >> 2) * _PMC_HM) & (M)) 387 388 /* 389 * Syscall structures 390 */ 391 392 /* The `sysent' for the new syscall */ 393 static struct sysent pmc_sysent = { 394 .sy_narg = 2, 395 .sy_call = pmc_syscall_handler, 396 }; 397 398 static struct syscall_module_data pmc_syscall_mod = { 399 .chainevh = load, 400 .chainarg = NULL, 401 .offset = &pmc_syscall_num, 402 .new_sysent = &pmc_sysent, 403 .old_sysent = { .sy_narg = 0, .sy_call = NULL }, 404 .flags = SY_THR_STATIC_KLD, 405 }; 406 407 static moduledata_t pmc_mod = { 408 .name = PMC_MODULE_NAME, 409 .evhand = syscall_module_handler, 410 .priv = &pmc_syscall_mod, 411 }; 412 413 #ifdef EARLY_AP_STARTUP 414 DECLARE_MODULE(pmc, pmc_mod, SI_SUB_SYSCALLS, SI_ORDER_ANY); 415 #else 416 DECLARE_MODULE(pmc, pmc_mod, SI_SUB_SMP, SI_ORDER_ANY); 417 #endif 418 MODULE_VERSION(pmc, PMC_VERSION); 419 420 #ifdef HWPMC_DEBUG 421 enum pmc_dbgparse_state { 422 PMCDS_WS, /* in whitespace */ 423 PMCDS_MAJOR, /* seen a major keyword */ 424 PMCDS_MINOR 425 }; 426 427 static int 428 pmc_debugflags_parse(char *newstr, char *fence) 429 { 430 char c, *p, *q; 431 struct pmc_debugflags *tmpflags; 432 int error, found, *newbits, tmp; 433 size_t kwlen; 434 435 tmpflags = malloc(sizeof(*tmpflags), M_PMC, M_WAITOK|M_ZERO); 436 437 p = newstr; 438 error = 0; 439 440 for (; p < fence && (c = *p); p++) { 441 442 /* skip white space */ 443 if (c == ' ' || c == '\t') 444 continue; 445 446 /* look for a keyword followed by "=" */ 447 for (q = p; p < fence && (c = *p) && c != '='; p++) 448 ; 449 if (c != '=') { 450 error = EINVAL; 451 goto done; 452 } 453 454 kwlen = p - q; 455 newbits = NULL; 456 457 /* lookup flag group name */ 458 #define DBG_SET_FLAG_MAJ(S,F) \ 459 if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0) \ 460 newbits = &tmpflags->pdb_ ## F; 461 462 DBG_SET_FLAG_MAJ("cpu", CPU); 463 DBG_SET_FLAG_MAJ("csw", CSW); 464 DBG_SET_FLAG_MAJ("logging", LOG); 465 DBG_SET_FLAG_MAJ("module", MOD); 466 DBG_SET_FLAG_MAJ("md", MDP); 467 DBG_SET_FLAG_MAJ("owner", OWN); 468 DBG_SET_FLAG_MAJ("pmc", PMC); 469 DBG_SET_FLAG_MAJ("process", PRC); 470 DBG_SET_FLAG_MAJ("sampling", SAM); 471 472 if (newbits == NULL) { 473 error = EINVAL; 474 goto done; 475 } 476 477 p++; /* skip the '=' */ 478 479 /* Now parse the individual flags */ 480 tmp = 0; 481 newflag: 482 for (q = p; p < fence && (c = *p); p++) 483 if (c == ' ' || c == '\t' || c == ',') 484 break; 485 486 /* p == fence or c == ws or c == "," or c == 0 */ 487 488 if ((kwlen = p - q) == 0) { 489 *newbits = tmp; 490 continue; 491 } 492 493 found = 0; 494 #define DBG_SET_FLAG_MIN(S,F) \ 495 if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0) \ 496 tmp |= found = (1 << PMC_DEBUG_MIN_ ## F) 497 498 /* a '*' denotes all possible flags in the group */ 499 if (kwlen == 1 && *q == '*') 500 tmp = found = ~0; 501 /* look for individual flag names */ 502 DBG_SET_FLAG_MIN("allocaterow", ALR); 503 DBG_SET_FLAG_MIN("allocate", ALL); 504 DBG_SET_FLAG_MIN("attach", ATT); 505 DBG_SET_FLAG_MIN("bind", BND); 506 DBG_SET_FLAG_MIN("config", CFG); 507 DBG_SET_FLAG_MIN("exec", EXC); 508 DBG_SET_FLAG_MIN("exit", EXT); 509 DBG_SET_FLAG_MIN("find", FND); 510 DBG_SET_FLAG_MIN("flush", FLS); 511 DBG_SET_FLAG_MIN("fork", FRK); 512 DBG_SET_FLAG_MIN("getbuf", GTB); 513 DBG_SET_FLAG_MIN("hook", PMH); 514 DBG_SET_FLAG_MIN("init", INI); 515 DBG_SET_FLAG_MIN("intr", INT); 516 DBG_SET_FLAG_MIN("linktarget", TLK); 517 DBG_SET_FLAG_MIN("mayberemove", OMR); 518 DBG_SET_FLAG_MIN("ops", OPS); 519 DBG_SET_FLAG_MIN("read", REA); 520 DBG_SET_FLAG_MIN("register", REG); 521 DBG_SET_FLAG_MIN("release", REL); 522 DBG_SET_FLAG_MIN("remove", ORM); 523 DBG_SET_FLAG_MIN("sample", SAM); 524 DBG_SET_FLAG_MIN("scheduleio", SIO); 525 DBG_SET_FLAG_MIN("select", SEL); 526 DBG_SET_FLAG_MIN("signal", SIG); 527 DBG_SET_FLAG_MIN("swi", SWI); 528 DBG_SET_FLAG_MIN("swo", SWO); 529 DBG_SET_FLAG_MIN("start", STA); 530 DBG_SET_FLAG_MIN("stop", STO); 531 DBG_SET_FLAG_MIN("syscall", PMS); 532 DBG_SET_FLAG_MIN("unlinktarget", TUL); 533 DBG_SET_FLAG_MIN("write", WRI); 534 if (found == 0) { 535 /* unrecognized flag name */ 536 error = EINVAL; 537 goto done; 538 } 539 540 if (c == 0 || c == ' ' || c == '\t') { /* end of flag group */ 541 *newbits = tmp; 542 continue; 543 } 544 545 p++; 546 goto newflag; 547 } 548 549 /* save the new flag set */ 550 bcopy(tmpflags, &pmc_debugflags, sizeof(pmc_debugflags)); 551 552 done: 553 free(tmpflags, M_PMC); 554 return error; 555 } 556 557 static int 558 pmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS) 559 { 560 char *fence, *newstr; 561 int error; 562 unsigned int n; 563 564 (void) arg1; (void) arg2; /* unused parameters */ 565 566 n = sizeof(pmc_debugstr); 567 newstr = malloc(n, M_PMC, M_WAITOK|M_ZERO); 568 (void) strlcpy(newstr, pmc_debugstr, n); 569 570 error = sysctl_handle_string(oidp, newstr, n, req); 571 572 /* if there is a new string, parse and copy it */ 573 if (error == 0 && req->newptr != NULL) { 574 fence = newstr + (n < req->newlen ? n : req->newlen + 1); 575 if ((error = pmc_debugflags_parse(newstr, fence)) == 0) 576 (void) strlcpy(pmc_debugstr, newstr, 577 sizeof(pmc_debugstr)); 578 } 579 580 free(newstr, M_PMC); 581 582 return error; 583 } 584 #endif 585 586 /* 587 * Map a row index to a classdep structure and return the adjusted row 588 * index for the PMC class index. 589 */ 590 static struct pmc_classdep * 591 pmc_ri_to_classdep(struct pmc_mdep *md, int ri, int *adjri) 592 { 593 struct pmc_classdep *pcd; 594 595 (void) md; 596 597 KASSERT(ri >= 0 && ri < md->pmd_npmc, 598 ("[pmc,%d] illegal row-index %d", __LINE__, ri)); 599 600 pcd = pmc_rowindex_to_classdep[ri]; 601 602 KASSERT(pcd != NULL, 603 ("[pmc,%d] ri %d null pcd", __LINE__, ri)); 604 605 *adjri = ri - pcd->pcd_ri; 606 607 KASSERT(*adjri >= 0 && *adjri < pcd->pcd_num, 608 ("[pmc,%d] adjusted row-index %d", __LINE__, *adjri)); 609 610 return (pcd); 611 } 612 613 /* 614 * Concurrency Control 615 * 616 * The driver manages the following data structures: 617 * 618 * - target process descriptors, one per target process 619 * - owner process descriptors (and attached lists), one per owner process 620 * - lookup hash tables for owner and target processes 621 * - PMC descriptors (and attached lists) 622 * - per-cpu hardware state 623 * - the 'hook' variable through which the kernel calls into 624 * this module 625 * - the machine hardware state (managed by the MD layer) 626 * 627 * These data structures are accessed from: 628 * 629 * - thread context-switch code 630 * - interrupt handlers (possibly on multiple cpus) 631 * - kernel threads on multiple cpus running on behalf of user 632 * processes doing system calls 633 * - this driver's private kernel threads 634 * 635 * = Locks and Locking strategy = 636 * 637 * The driver uses four locking strategies for its operation: 638 * 639 * - The global SX lock "pmc_sx" is used to protect internal 640 * data structures. 641 * 642 * Calls into the module by syscall() start with this lock being 643 * held in exclusive mode. Depending on the requested operation, 644 * the lock may be downgraded to 'shared' mode to allow more 645 * concurrent readers into the module. Calls into the module from 646 * other parts of the kernel acquire the lock in shared mode. 647 * 648 * This SX lock is held in exclusive mode for any operations that 649 * modify the linkages between the driver's internal data structures. 650 * 651 * The 'pmc_hook' function pointer is also protected by this lock. 652 * It is only examined with the sx lock held in exclusive mode. The 653 * kernel module is allowed to be unloaded only with the sx lock held 654 * in exclusive mode. In normal syscall handling, after acquiring the 655 * pmc_sx lock we first check that 'pmc_hook' is non-null before 656 * proceeding. This prevents races between the thread unloading the module 657 * and other threads seeking to use the module. 658 * 659 * - Lookups of target process structures and owner process structures 660 * cannot use the global "pmc_sx" SX lock because these lookups need 661 * to happen during context switches and in other critical sections 662 * where sleeping is not allowed. We protect these lookup tables 663 * with their own private spin-mutexes, "pmc_processhash_mtx" and 664 * "pmc_ownerhash_mtx". 665 * 666 * - Interrupt handlers work in a lock free manner. At interrupt 667 * time, handlers look at the PMC pointer (phw->phw_pmc) configured 668 * when the PMC was started. If this pointer is NULL, the interrupt 669 * is ignored after updating driver statistics. We ensure that this 670 * pointer is set (using an atomic operation if necessary) before the 671 * PMC hardware is started. Conversely, this pointer is unset atomically 672 * only after the PMC hardware is stopped. 673 * 674 * We ensure that everything needed for the operation of an 675 * interrupt handler is available without it needing to acquire any 676 * locks. We also ensure that a PMC's software state is destroyed only 677 * after the PMC is taken off hardware (on all CPUs). 678 * 679 * - Context-switch handling with process-private PMCs needs more 680 * care. 681 * 682 * A given process may be the target of multiple PMCs. For example, 683 * PMCATTACH and PMCDETACH may be requested by a process on one CPU 684 * while the target process is running on another. A PMC could also 685 * be getting released because its owner is exiting. We tackle 686 * these situations in the following manner: 687 * 688 * - each target process structure 'pmc_process' has an array 689 * of 'struct pmc *' pointers, one for each hardware PMC. 690 * 691 * - At context switch IN time, each "target" PMC in RUNNING state 692 * gets started on hardware and a pointer to each PMC is copied into 693 * the per-cpu phw array. The 'runcount' for the PMC is 694 * incremented. 695 * 696 * - At context switch OUT time, all process-virtual PMCs are stopped 697 * on hardware. The saved value is added to the PMCs value field 698 * only if the PMC is in a non-deleted state (the PMCs state could 699 * have changed during the current time slice). 700 * 701 * Note that since in-between a switch IN on a processor and a switch 702 * OUT, the PMC could have been released on another CPU. Therefore 703 * context switch OUT always looks at the hardware state to turn 704 * OFF PMCs and will update a PMC's saved value only if reachable 705 * from the target process record. 706 * 707 * - OP PMCRELEASE could be called on a PMC at any time (the PMC could 708 * be attached to many processes at the time of the call and could 709 * be active on multiple CPUs). 710 * 711 * We prevent further scheduling of the PMC by marking it as in 712 * state 'DELETED'. If the runcount of the PMC is non-zero then 713 * this PMC is currently running on a CPU somewhere. The thread 714 * doing the PMCRELEASE operation waits by repeatedly doing a 715 * pause() till the runcount comes to zero. 716 * 717 * The contents of a PMC descriptor (struct pmc) are protected using 718 * a spin-mutex. In order to save space, we use a mutex pool. 719 * 720 * In terms of lock types used by witness(4), we use: 721 * - Type "pmc-sx", used by the global SX lock. 722 * - Type "pmc-sleep", for sleep mutexes used by logger threads. 723 * - Type "pmc-per-proc", for protecting PMC owner descriptors. 724 * - Type "pmc-leaf", used for all other spin mutexes. 725 */ 726 727 /* 728 * save the cpu binding of the current kthread 729 */ 730 731 static void 732 pmc_save_cpu_binding(struct pmc_binding *pb) 733 { 734 PMCDBG0(CPU,BND,2, "save-cpu"); 735 thread_lock(curthread); 736 pb->pb_bound = sched_is_bound(curthread); 737 pb->pb_cpu = curthread->td_oncpu; 738 thread_unlock(curthread); 739 PMCDBG1(CPU,BND,2, "save-cpu cpu=%d", pb->pb_cpu); 740 } 741 742 /* 743 * restore the cpu binding of the current thread 744 */ 745 746 static void 747 pmc_restore_cpu_binding(struct pmc_binding *pb) 748 { 749 PMCDBG2(CPU,BND,2, "restore-cpu curcpu=%d restore=%d", 750 curthread->td_oncpu, pb->pb_cpu); 751 thread_lock(curthread); 752 if (pb->pb_bound) 753 sched_bind(curthread, pb->pb_cpu); 754 else 755 sched_unbind(curthread); 756 thread_unlock(curthread); 757 PMCDBG0(CPU,BND,2, "restore-cpu done"); 758 } 759 760 /* 761 * move execution over the specified cpu and bind it there. 762 */ 763 764 static void 765 pmc_select_cpu(int cpu) 766 { 767 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), 768 ("[pmc,%d] bad cpu number %d", __LINE__, cpu)); 769 770 /* Never move to an inactive CPU. */ 771 KASSERT(pmc_cpu_is_active(cpu), ("[pmc,%d] selecting inactive " 772 "CPU %d", __LINE__, cpu)); 773 774 PMCDBG1(CPU,SEL,2, "select-cpu cpu=%d", cpu); 775 thread_lock(curthread); 776 sched_bind(curthread, cpu); 777 thread_unlock(curthread); 778 779 KASSERT(curthread->td_oncpu == cpu, 780 ("[pmc,%d] CPU not bound [cpu=%d, curr=%d]", __LINE__, 781 cpu, curthread->td_oncpu)); 782 783 PMCDBG1(CPU,SEL,2, "select-cpu cpu=%d ok", cpu); 784 } 785 786 /* 787 * Force a context switch. 788 * 789 * We do this by pause'ing for 1 tick -- invoking mi_switch() is not 790 * guaranteed to force a context switch. 791 */ 792 793 static void 794 pmc_force_context_switch(void) 795 { 796 797 pause("pmcctx", 1); 798 } 799 800 /* 801 * Get the file name for an executable. This is a simple wrapper 802 * around vn_fullpath(9). 803 */ 804 805 static void 806 pmc_getfilename(struct vnode *v, char **fullpath, char **freepath) 807 { 808 809 *fullpath = "unknown"; 810 *freepath = NULL; 811 vn_fullpath(curthread, v, fullpath, freepath); 812 } 813 814 /* 815 * remove an process owning PMCs 816 */ 817 818 void 819 pmc_remove_owner(struct pmc_owner *po) 820 { 821 struct pmc *pm, *tmp; 822 823 sx_assert(&pmc_sx, SX_XLOCKED); 824 825 PMCDBG1(OWN,ORM,1, "remove-owner po=%p", po); 826 827 /* Remove descriptor from the owner hash table */ 828 LIST_REMOVE(po, po_next); 829 830 /* release all owned PMC descriptors */ 831 LIST_FOREACH_SAFE(pm, &po->po_pmcs, pm_next, tmp) { 832 PMCDBG1(OWN,ORM,2, "pmc=%p", pm); 833 KASSERT(pm->pm_owner == po, 834 ("[pmc,%d] owner %p != po %p", __LINE__, pm->pm_owner, po)); 835 836 pmc_release_pmc_descriptor(pm); /* will unlink from the list */ 837 pmc_destroy_pmc_descriptor(pm); 838 } 839 840 KASSERT(po->po_sscount == 0, 841 ("[pmc,%d] SS count not zero", __LINE__)); 842 KASSERT(LIST_EMPTY(&po->po_pmcs), 843 ("[pmc,%d] PMC list not empty", __LINE__)); 844 845 /* de-configure the log file if present */ 846 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 847 pmclog_deconfigure_log(po); 848 } 849 850 /* 851 * remove an owner process record if all conditions are met. 852 */ 853 854 static void 855 pmc_maybe_remove_owner(struct pmc_owner *po) 856 { 857 858 PMCDBG1(OWN,OMR,1, "maybe-remove-owner po=%p", po); 859 860 /* 861 * Remove owner record if 862 * - this process does not own any PMCs 863 * - this process has not allocated a system-wide sampling buffer 864 */ 865 866 if (LIST_EMPTY(&po->po_pmcs) && 867 ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)) { 868 pmc_remove_owner(po); 869 pmc_destroy_owner_descriptor(po); 870 } 871 } 872 873 /* 874 * Add an association between a target process and a PMC. 875 */ 876 877 static void 878 pmc_link_target_process(struct pmc *pm, struct pmc_process *pp) 879 { 880 int ri; 881 struct pmc_target *pt; 882 #ifdef INVARIANTS 883 struct pmc_thread *pt_td; 884 #endif 885 886 sx_assert(&pmc_sx, SX_XLOCKED); 887 888 KASSERT(pm != NULL && pp != NULL, 889 ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp)); 890 KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)), 891 ("[pmc,%d] Attaching a non-process-virtual pmc=%p to pid=%d", 892 __LINE__, pm, pp->pp_proc->p_pid)); 893 KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= ((int) md->pmd_npmc - 1), 894 ("[pmc,%d] Illegal reference count %d for process record %p", 895 __LINE__, pp->pp_refcnt, (void *) pp)); 896 897 ri = PMC_TO_ROWINDEX(pm); 898 899 PMCDBG3(PRC,TLK,1, "link-target pmc=%p ri=%d pmc-process=%p", 900 pm, ri, pp); 901 902 #ifdef HWPMC_DEBUG 903 LIST_FOREACH(pt, &pm->pm_targets, pt_next) 904 if (pt->pt_process == pp) 905 KASSERT(0, ("[pmc,%d] pp %p already in pmc %p targets", 906 __LINE__, pp, pm)); 907 #endif 908 909 pt = malloc(sizeof(struct pmc_target), M_PMC, M_WAITOK|M_ZERO); 910 pt->pt_process = pp; 911 912 LIST_INSERT_HEAD(&pm->pm_targets, pt, pt_next); 913 914 atomic_store_rel_ptr((uintptr_t *)&pp->pp_pmcs[ri].pp_pmc, 915 (uintptr_t)pm); 916 917 if (pm->pm_owner->po_owner == pp->pp_proc) 918 pm->pm_flags |= PMC_F_ATTACHED_TO_OWNER; 919 920 /* 921 * Initialize the per-process values at this row index. 922 */ 923 pp->pp_pmcs[ri].pp_pmcval = PMC_TO_MODE(pm) == PMC_MODE_TS ? 924 pm->pm_sc.pm_reloadcount : 0; 925 926 pp->pp_refcnt++; 927 928 #ifdef INVARIANTS 929 /* Confirm that the per-thread values at this row index are cleared. */ 930 if (PMC_TO_MODE(pm) == PMC_MODE_TS) { 931 mtx_lock_spin(pp->pp_tdslock); 932 LIST_FOREACH(pt_td, &pp->pp_tds, pt_next) { 933 KASSERT(pt_td->pt_pmcs[ri].pt_pmcval == (pmc_value_t) 0, 934 ("[pmc,%d] pt_pmcval not cleared for pid=%d at " 935 "ri=%d", __LINE__, pp->pp_proc->p_pid, ri)); 936 } 937 mtx_unlock_spin(pp->pp_tdslock); 938 } 939 #endif 940 } 941 942 /* 943 * Removes the association between a target process and a PMC. 944 */ 945 946 static void 947 pmc_unlink_target_process(struct pmc *pm, struct pmc_process *pp) 948 { 949 int ri; 950 struct proc *p; 951 struct pmc_target *ptgt; 952 struct pmc_thread *pt; 953 954 sx_assert(&pmc_sx, SX_XLOCKED); 955 956 KASSERT(pm != NULL && pp != NULL, 957 ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp)); 958 959 KASSERT(pp->pp_refcnt >= 1 && pp->pp_refcnt <= (int) md->pmd_npmc, 960 ("[pmc,%d] Illegal ref count %d on process record %p", 961 __LINE__, pp->pp_refcnt, (void *) pp)); 962 963 ri = PMC_TO_ROWINDEX(pm); 964 965 PMCDBG3(PRC,TUL,1, "unlink-target pmc=%p ri=%d pmc-process=%p", 966 pm, ri, pp); 967 968 KASSERT(pp->pp_pmcs[ri].pp_pmc == pm, 969 ("[pmc,%d] PMC ri %d mismatch pmc %p pp->[ri] %p", __LINE__, 970 ri, pm, pp->pp_pmcs[ri].pp_pmc)); 971 972 pp->pp_pmcs[ri].pp_pmc = NULL; 973 pp->pp_pmcs[ri].pp_pmcval = (pmc_value_t) 0; 974 975 /* Clear the per-thread values at this row index. */ 976 if (PMC_TO_MODE(pm) == PMC_MODE_TS) { 977 mtx_lock_spin(pp->pp_tdslock); 978 LIST_FOREACH(pt, &pp->pp_tds, pt_next) 979 pt->pt_pmcs[ri].pt_pmcval = (pmc_value_t) 0; 980 mtx_unlock_spin(pp->pp_tdslock); 981 } 982 983 /* Remove owner-specific flags */ 984 if (pm->pm_owner->po_owner == pp->pp_proc) { 985 pp->pp_flags &= ~PMC_PP_ENABLE_MSR_ACCESS; 986 pm->pm_flags &= ~PMC_F_ATTACHED_TO_OWNER; 987 } 988 989 pp->pp_refcnt--; 990 991 /* Remove the target process from the PMC structure */ 992 LIST_FOREACH(ptgt, &pm->pm_targets, pt_next) 993 if (ptgt->pt_process == pp) 994 break; 995 996 KASSERT(ptgt != NULL, ("[pmc,%d] process %p (pp: %p) not found " 997 "in pmc %p", __LINE__, pp->pp_proc, pp, pm)); 998 999 LIST_REMOVE(ptgt, pt_next); 1000 free(ptgt, M_PMC); 1001 1002 /* if the PMC now lacks targets, send the owner a SIGIO */ 1003 if (LIST_EMPTY(&pm->pm_targets)) { 1004 p = pm->pm_owner->po_owner; 1005 PROC_LOCK(p); 1006 kern_psignal(p, SIGIO); 1007 PROC_UNLOCK(p); 1008 1009 PMCDBG2(PRC,SIG,2, "signalling proc=%p signal=%d", p, 1010 SIGIO); 1011 } 1012 } 1013 1014 /* 1015 * Check if PMC 'pm' may be attached to target process 't'. 1016 */ 1017 1018 static int 1019 pmc_can_attach(struct pmc *pm, struct proc *t) 1020 { 1021 struct proc *o; /* pmc owner */ 1022 struct ucred *oc, *tc; /* owner, target credentials */ 1023 int decline_attach, i; 1024 1025 /* 1026 * A PMC's owner can always attach that PMC to itself. 1027 */ 1028 1029 if ((o = pm->pm_owner->po_owner) == t) 1030 return 0; 1031 1032 PROC_LOCK(o); 1033 oc = o->p_ucred; 1034 crhold(oc); 1035 PROC_UNLOCK(o); 1036 1037 PROC_LOCK(t); 1038 tc = t->p_ucred; 1039 crhold(tc); 1040 PROC_UNLOCK(t); 1041 1042 /* 1043 * The effective uid of the PMC owner should match at least one 1044 * of the {effective,real,saved} uids of the target process. 1045 */ 1046 1047 decline_attach = oc->cr_uid != tc->cr_uid && 1048 oc->cr_uid != tc->cr_svuid && 1049 oc->cr_uid != tc->cr_ruid; 1050 1051 /* 1052 * Every one of the target's group ids, must be in the owner's 1053 * group list. 1054 */ 1055 for (i = 0; !decline_attach && i < tc->cr_ngroups; i++) 1056 decline_attach = !groupmember(tc->cr_groups[i], oc); 1057 1058 /* check the read and saved gids too */ 1059 if (decline_attach == 0) 1060 decline_attach = !groupmember(tc->cr_rgid, oc) || 1061 !groupmember(tc->cr_svgid, oc); 1062 1063 crfree(tc); 1064 crfree(oc); 1065 1066 return !decline_attach; 1067 } 1068 1069 /* 1070 * Attach a process to a PMC. 1071 */ 1072 1073 static int 1074 pmc_attach_one_process(struct proc *p, struct pmc *pm) 1075 { 1076 int ri, error; 1077 char *fullpath, *freepath; 1078 struct pmc_process *pp; 1079 1080 sx_assert(&pmc_sx, SX_XLOCKED); 1081 1082 PMCDBG5(PRC,ATT,2, "attach-one pm=%p ri=%d proc=%p (%d, %s)", pm, 1083 PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm); 1084 1085 /* 1086 * Locate the process descriptor corresponding to process 'p', 1087 * allocating space as needed. 1088 * 1089 * Verify that rowindex 'pm_rowindex' is free in the process 1090 * descriptor. 1091 * 1092 * If not, allocate space for a descriptor and link the 1093 * process descriptor and PMC. 1094 */ 1095 ri = PMC_TO_ROWINDEX(pm); 1096 1097 /* mark process as using HWPMCs */ 1098 PROC_LOCK(p); 1099 p->p_flag |= P_HWPMC; 1100 PROC_UNLOCK(p); 1101 1102 if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_ALLOCATE)) == NULL) { 1103 error = ENOMEM; 1104 goto fail; 1105 } 1106 1107 if (pp->pp_pmcs[ri].pp_pmc == pm) {/* already present at slot [ri] */ 1108 error = EEXIST; 1109 goto fail; 1110 } 1111 1112 if (pp->pp_pmcs[ri].pp_pmc != NULL) { 1113 error = EBUSY; 1114 goto fail; 1115 } 1116 1117 pmc_link_target_process(pm, pp); 1118 1119 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) && 1120 (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) == 0) 1121 pm->pm_flags |= PMC_F_NEEDS_LOGFILE; 1122 1123 pm->pm_flags |= PMC_F_ATTACH_DONE; /* mark as attached */ 1124 1125 /* issue an attach event to a configured log file */ 1126 if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE) { 1127 if (p->p_flag & P_KPROC) { 1128 fullpath = kernelname; 1129 freepath = NULL; 1130 } else { 1131 pmc_getfilename(p->p_textvp, &fullpath, &freepath); 1132 pmclog_process_pmcattach(pm, p->p_pid, fullpath); 1133 } 1134 free(freepath, M_TEMP); 1135 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) 1136 pmc_log_process_mappings(pm->pm_owner, p); 1137 } 1138 1139 return (0); 1140 fail: 1141 PROC_LOCK(p); 1142 p->p_flag &= ~P_HWPMC; 1143 PROC_UNLOCK(p); 1144 return (error); 1145 } 1146 1147 /* 1148 * Attach a process and optionally its children 1149 */ 1150 1151 static int 1152 pmc_attach_process(struct proc *p, struct pmc *pm) 1153 { 1154 int error; 1155 struct proc *top; 1156 1157 sx_assert(&pmc_sx, SX_XLOCKED); 1158 1159 PMCDBG5(PRC,ATT,1, "attach pm=%p ri=%d proc=%p (%d, %s)", pm, 1160 PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm); 1161 1162 1163 /* 1164 * If this PMC successfully allowed a GETMSR operation 1165 * in the past, disallow further ATTACHes. 1166 */ 1167 1168 if ((pm->pm_flags & PMC_PP_ENABLE_MSR_ACCESS) != 0) 1169 return EPERM; 1170 1171 if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0) 1172 return pmc_attach_one_process(p, pm); 1173 1174 /* 1175 * Traverse all child processes, attaching them to 1176 * this PMC. 1177 */ 1178 1179 sx_slock(&proctree_lock); 1180 1181 top = p; 1182 1183 for (;;) { 1184 if ((error = pmc_attach_one_process(p, pm)) != 0) 1185 break; 1186 if (!LIST_EMPTY(&p->p_children)) 1187 p = LIST_FIRST(&p->p_children); 1188 else for (;;) { 1189 if (p == top) 1190 goto done; 1191 if (LIST_NEXT(p, p_sibling)) { 1192 p = LIST_NEXT(p, p_sibling); 1193 break; 1194 } 1195 p = p->p_pptr; 1196 } 1197 } 1198 1199 if (error) 1200 (void) pmc_detach_process(top, pm); 1201 1202 done: 1203 sx_sunlock(&proctree_lock); 1204 return error; 1205 } 1206 1207 /* 1208 * Detach a process from a PMC. If there are no other PMCs tracking 1209 * this process, remove the process structure from its hash table. If 1210 * 'flags' contains PMC_FLAG_REMOVE, then free the process structure. 1211 */ 1212 1213 static int 1214 pmc_detach_one_process(struct proc *p, struct pmc *pm, int flags) 1215 { 1216 int ri; 1217 struct pmc_process *pp; 1218 1219 sx_assert(&pmc_sx, SX_XLOCKED); 1220 1221 KASSERT(pm != NULL, 1222 ("[pmc,%d] null pm pointer", __LINE__)); 1223 1224 ri = PMC_TO_ROWINDEX(pm); 1225 1226 PMCDBG6(PRC,ATT,2, "detach-one pm=%p ri=%d proc=%p (%d, %s) flags=0x%x", 1227 pm, ri, p, p->p_pid, p->p_comm, flags); 1228 1229 if ((pp = pmc_find_process_descriptor(p, 0)) == NULL) 1230 return ESRCH; 1231 1232 if (pp->pp_pmcs[ri].pp_pmc != pm) 1233 return EINVAL; 1234 1235 pmc_unlink_target_process(pm, pp); 1236 1237 /* Issue a detach entry if a log file is configured */ 1238 if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE) 1239 pmclog_process_pmcdetach(pm, p->p_pid); 1240 1241 /* 1242 * If there are no PMCs targeting this process, we remove its 1243 * descriptor from the target hash table and unset the P_HWPMC 1244 * flag in the struct proc. 1245 */ 1246 KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= (int) md->pmd_npmc, 1247 ("[pmc,%d] Illegal refcnt %d for process struct %p", 1248 __LINE__, pp->pp_refcnt, pp)); 1249 1250 if (pp->pp_refcnt != 0) /* still a target of some PMC */ 1251 return 0; 1252 1253 pmc_remove_process_descriptor(pp); 1254 1255 if (flags & PMC_FLAG_REMOVE) 1256 pmc_destroy_process_descriptor(pp); 1257 1258 PROC_LOCK(p); 1259 p->p_flag &= ~P_HWPMC; 1260 PROC_UNLOCK(p); 1261 1262 return 0; 1263 } 1264 1265 /* 1266 * Detach a process and optionally its descendants from a PMC. 1267 */ 1268 1269 static int 1270 pmc_detach_process(struct proc *p, struct pmc *pm) 1271 { 1272 struct proc *top; 1273 1274 sx_assert(&pmc_sx, SX_XLOCKED); 1275 1276 PMCDBG5(PRC,ATT,1, "detach pm=%p ri=%d proc=%p (%d, %s)", pm, 1277 PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm); 1278 1279 if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0) 1280 return pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE); 1281 1282 /* 1283 * Traverse all children, detaching them from this PMC. We 1284 * ignore errors since we could be detaching a PMC from a 1285 * partially attached proc tree. 1286 */ 1287 1288 sx_slock(&proctree_lock); 1289 1290 top = p; 1291 1292 for (;;) { 1293 (void) pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE); 1294 1295 if (!LIST_EMPTY(&p->p_children)) 1296 p = LIST_FIRST(&p->p_children); 1297 else for (;;) { 1298 if (p == top) 1299 goto done; 1300 if (LIST_NEXT(p, p_sibling)) { 1301 p = LIST_NEXT(p, p_sibling); 1302 break; 1303 } 1304 p = p->p_pptr; 1305 } 1306 } 1307 1308 done: 1309 sx_sunlock(&proctree_lock); 1310 1311 if (LIST_EMPTY(&pm->pm_targets)) 1312 pm->pm_flags &= ~PMC_F_ATTACH_DONE; 1313 1314 return 0; 1315 } 1316 1317 1318 /* 1319 * Thread context switch IN 1320 */ 1321 1322 static void 1323 pmc_process_csw_in(struct thread *td) 1324 { 1325 int cpu; 1326 unsigned int adjri, ri; 1327 struct pmc *pm; 1328 struct proc *p; 1329 struct pmc_cpu *pc; 1330 struct pmc_hw *phw; 1331 pmc_value_t newvalue; 1332 struct pmc_process *pp; 1333 struct pmc_thread *pt; 1334 struct pmc_classdep *pcd; 1335 1336 p = td->td_proc; 1337 pt = NULL; 1338 if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE)) == NULL) 1339 return; 1340 1341 KASSERT(pp->pp_proc == td->td_proc, 1342 ("[pmc,%d] not my thread state", __LINE__)); 1343 1344 critical_enter(); /* no preemption from this point */ 1345 1346 cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */ 1347 1348 PMCDBG5(CSW,SWI,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p, 1349 p->p_pid, p->p_comm, pp); 1350 1351 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), 1352 ("[pmc,%d] weird CPU id %d", __LINE__, cpu)); 1353 1354 pc = pmc_pcpu[cpu]; 1355 1356 for (ri = 0; ri < md->pmd_npmc; ri++) { 1357 1358 if ((pm = pp->pp_pmcs[ri].pp_pmc) == NULL) 1359 continue; 1360 1361 KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)), 1362 ("[pmc,%d] Target PMC in non-virtual mode (%d)", 1363 __LINE__, PMC_TO_MODE(pm))); 1364 1365 KASSERT(PMC_TO_ROWINDEX(pm) == ri, 1366 ("[pmc,%d] Row index mismatch pmc %d != ri %d", 1367 __LINE__, PMC_TO_ROWINDEX(pm), ri)); 1368 1369 /* 1370 * Only PMCs that are marked as 'RUNNING' need 1371 * be placed on hardware. 1372 */ 1373 1374 if (pm->pm_state != PMC_STATE_RUNNING) 1375 continue; 1376 1377 /* increment PMC runcount */ 1378 counter_u64_add(pm->pm_runcount, 1); 1379 1380 /* configure the HWPMC we are going to use. */ 1381 pcd = pmc_ri_to_classdep(md, ri, &adjri); 1382 pcd->pcd_config_pmc(cpu, adjri, pm); 1383 1384 phw = pc->pc_hwpmcs[ri]; 1385 1386 KASSERT(phw != NULL, 1387 ("[pmc,%d] null hw pointer", __LINE__)); 1388 1389 KASSERT(phw->phw_pmc == pm, 1390 ("[pmc,%d] hw->pmc %p != pmc %p", __LINE__, 1391 phw->phw_pmc, pm)); 1392 1393 /* 1394 * Write out saved value and start the PMC. 1395 * 1396 * Sampling PMCs use a per-thread value, while 1397 * counting mode PMCs use a per-pmc value that is 1398 * inherited across descendants. 1399 */ 1400 if (PMC_TO_MODE(pm) == PMC_MODE_TS) { 1401 if (pt == NULL) 1402 pt = pmc_find_thread_descriptor(pp, td, 1403 PMC_FLAG_NONE); 1404 1405 KASSERT(pt != NULL, 1406 ("[pmc,%d] No thread found for td=%p", __LINE__, 1407 td)); 1408 1409 mtx_pool_lock_spin(pmc_mtxpool, pm); 1410 1411 /* 1412 * If we have a thread descriptor, use the per-thread 1413 * counter in the descriptor. If not, we will use 1414 * a per-process counter. 1415 * 1416 * TODO: Remove the per-process "safety net" once 1417 * we have thoroughly tested that we don't hit the 1418 * above assert. 1419 */ 1420 if (pt != NULL) { 1421 if (pt->pt_pmcs[ri].pt_pmcval > 0) 1422 newvalue = pt->pt_pmcs[ri].pt_pmcval; 1423 else 1424 newvalue = pm->pm_sc.pm_reloadcount; 1425 } else { 1426 /* 1427 * Use the saved value calculated after the most 1428 * recent time a thread using the shared counter 1429 * switched out. Reset the saved count in case 1430 * another thread from this process switches in 1431 * before any threads switch out. 1432 */ 1433 1434 newvalue = pp->pp_pmcs[ri].pp_pmcval; 1435 pp->pp_pmcs[ri].pp_pmcval = 1436 pm->pm_sc.pm_reloadcount; 1437 } 1438 mtx_pool_unlock_spin(pmc_mtxpool, pm); 1439 KASSERT(newvalue > 0 && newvalue <= 1440 pm->pm_sc.pm_reloadcount, 1441 ("[pmc,%d] pmcval outside of expected range cpu=%d " 1442 "ri=%d pmcval=%jx pm_reloadcount=%jx", __LINE__, 1443 cpu, ri, newvalue, pm->pm_sc.pm_reloadcount)); 1444 } else { 1445 KASSERT(PMC_TO_MODE(pm) == PMC_MODE_TC, 1446 ("[pmc,%d] illegal mode=%d", __LINE__, 1447 PMC_TO_MODE(pm))); 1448 mtx_pool_lock_spin(pmc_mtxpool, pm); 1449 newvalue = PMC_PCPU_SAVED(cpu, ri) = 1450 pm->pm_gv.pm_savedvalue; 1451 mtx_pool_unlock_spin(pmc_mtxpool, pm); 1452 } 1453 1454 PMCDBG3(CSW,SWI,1,"cpu=%d ri=%d new=%jd", cpu, ri, newvalue); 1455 1456 pcd->pcd_write_pmc(cpu, adjri, newvalue); 1457 1458 /* If a sampling mode PMC, reset stalled state. */ 1459 if (PMC_TO_MODE(pm) == PMC_MODE_TS) 1460 pm->pm_pcpu_state[cpu].pps_stalled = 0; 1461 1462 /* Indicate that we desire this to run. */ 1463 pm->pm_pcpu_state[cpu].pps_cpustate = 1; 1464 1465 /* Start the PMC. */ 1466 pcd->pcd_start_pmc(cpu, adjri); 1467 } 1468 1469 /* 1470 * perform any other architecture/cpu dependent thread 1471 * switch-in actions. 1472 */ 1473 1474 (void) (*md->pmd_switch_in)(pc, pp); 1475 1476 critical_exit(); 1477 1478 } 1479 1480 /* 1481 * Thread context switch OUT. 1482 */ 1483 1484 static void 1485 pmc_process_csw_out(struct thread *td) 1486 { 1487 int cpu; 1488 int64_t tmp; 1489 struct pmc *pm; 1490 struct proc *p; 1491 enum pmc_mode mode; 1492 struct pmc_cpu *pc; 1493 pmc_value_t newvalue; 1494 unsigned int adjri, ri; 1495 struct pmc_process *pp; 1496 struct pmc_thread *pt = NULL; 1497 struct pmc_classdep *pcd; 1498 1499 1500 /* 1501 * Locate our process descriptor; this may be NULL if 1502 * this process is exiting and we have already removed 1503 * the process from the target process table. 1504 * 1505 * Note that due to kernel preemption, multiple 1506 * context switches may happen while the process is 1507 * exiting. 1508 * 1509 * Note also that if the target process cannot be 1510 * found we still need to deconfigure any PMCs that 1511 * are currently running on hardware. 1512 */ 1513 1514 p = td->td_proc; 1515 pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE); 1516 1517 /* 1518 * save PMCs 1519 */ 1520 1521 critical_enter(); 1522 1523 cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */ 1524 1525 PMCDBG5(CSW,SWO,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p, 1526 p->p_pid, p->p_comm, pp); 1527 1528 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), 1529 ("[pmc,%d weird CPU id %d", __LINE__, cpu)); 1530 1531 pc = pmc_pcpu[cpu]; 1532 1533 /* 1534 * When a PMC gets unlinked from a target PMC, it will 1535 * be removed from the target's pp_pmc[] array. 1536 * 1537 * However, on a MP system, the target could have been 1538 * executing on another CPU at the time of the unlink. 1539 * So, at context switch OUT time, we need to look at 1540 * the hardware to determine if a PMC is scheduled on 1541 * it. 1542 */ 1543 1544 for (ri = 0; ri < md->pmd_npmc; ri++) { 1545 1546 pcd = pmc_ri_to_classdep(md, ri, &adjri); 1547 pm = NULL; 1548 (void) (*pcd->pcd_get_config)(cpu, adjri, &pm); 1549 1550 if (pm == NULL) /* nothing at this row index */ 1551 continue; 1552 1553 mode = PMC_TO_MODE(pm); 1554 if (!PMC_IS_VIRTUAL_MODE(mode)) 1555 continue; /* not a process virtual PMC */ 1556 1557 KASSERT(PMC_TO_ROWINDEX(pm) == ri, 1558 ("[pmc,%d] ri mismatch pmc(%d) ri(%d)", 1559 __LINE__, PMC_TO_ROWINDEX(pm), ri)); 1560 1561 /* 1562 * Change desired state, and then stop if not stalled. 1563 * This two-step dance should avoid race conditions where 1564 * an interrupt re-enables the PMC after this code has 1565 * already checked the pm_stalled flag. 1566 */ 1567 pm->pm_pcpu_state[cpu].pps_cpustate = 0; 1568 if (pm->pm_pcpu_state[cpu].pps_stalled == 0) 1569 pcd->pcd_stop_pmc(cpu, adjri); 1570 1571 /* reduce this PMC's runcount */ 1572 counter_u64_add(pm->pm_runcount, -1); 1573 1574 /* 1575 * If this PMC is associated with this process, 1576 * save the reading. 1577 */ 1578 1579 if (pm->pm_state != PMC_STATE_DELETED && pp != NULL && 1580 pp->pp_pmcs[ri].pp_pmc != NULL) { 1581 KASSERT(pm == pp->pp_pmcs[ri].pp_pmc, 1582 ("[pmc,%d] pm %p != pp_pmcs[%d] %p", __LINE__, 1583 pm, ri, pp->pp_pmcs[ri].pp_pmc)); 1584 1585 KASSERT(pp->pp_refcnt > 0, 1586 ("[pmc,%d] pp refcnt = %d", __LINE__, 1587 pp->pp_refcnt)); 1588 1589 pcd->pcd_read_pmc(cpu, adjri, &newvalue); 1590 1591 if (mode == PMC_MODE_TS) { 1592 PMCDBG3(CSW,SWO,1,"cpu=%d ri=%d val=%jd (samp)", 1593 cpu, ri, newvalue); 1594 1595 if (pt == NULL) 1596 pt = pmc_find_thread_descriptor(pp, td, 1597 PMC_FLAG_NONE); 1598 1599 KASSERT(pt != NULL, 1600 ("[pmc,%d] No thread found for td=%p", 1601 __LINE__, td)); 1602 1603 mtx_pool_lock_spin(pmc_mtxpool, pm); 1604 1605 /* 1606 * If we have a thread descriptor, save the 1607 * per-thread counter in the descriptor. If not, 1608 * we will update the per-process counter. 1609 * 1610 * TODO: Remove the per-process "safety net" 1611 * once we have thoroughly tested that we 1612 * don't hit the above assert. 1613 */ 1614 if (pt != NULL) 1615 pt->pt_pmcs[ri].pt_pmcval = newvalue; 1616 else { 1617 /* 1618 * For sampling process-virtual PMCs, 1619 * newvalue is the number of events to 1620 * be seen until the next sampling 1621 * interrupt. We can just add the events 1622 * left from this invocation to the 1623 * counter, then adjust in case we 1624 * overflow our range. 1625 * 1626 * (Recall that we reload the counter 1627 * every time we use it.) 1628 */ 1629 pp->pp_pmcs[ri].pp_pmcval += newvalue; 1630 if (pp->pp_pmcs[ri].pp_pmcval > 1631 pm->pm_sc.pm_reloadcount) 1632 pp->pp_pmcs[ri].pp_pmcval -= 1633 pm->pm_sc.pm_reloadcount; 1634 } 1635 mtx_pool_unlock_spin(pmc_mtxpool, pm); 1636 } else { 1637 tmp = newvalue - PMC_PCPU_SAVED(cpu,ri); 1638 1639 PMCDBG3(CSW,SWO,1,"cpu=%d ri=%d tmp=%jd (count)", 1640 cpu, ri, tmp); 1641 1642 /* 1643 * For counting process-virtual PMCs, 1644 * we expect the count to be 1645 * increasing monotonically, modulo a 64 1646 * bit wraparound. 1647 */ 1648 KASSERT(tmp >= 0, 1649 ("[pmc,%d] negative increment cpu=%d " 1650 "ri=%d newvalue=%jx saved=%jx " 1651 "incr=%jx", __LINE__, cpu, ri, 1652 newvalue, PMC_PCPU_SAVED(cpu,ri), tmp)); 1653 1654 mtx_pool_lock_spin(pmc_mtxpool, pm); 1655 pm->pm_gv.pm_savedvalue += tmp; 1656 pp->pp_pmcs[ri].pp_pmcval += tmp; 1657 mtx_pool_unlock_spin(pmc_mtxpool, pm); 1658 1659 if (pm->pm_flags & PMC_F_LOG_PROCCSW) 1660 pmclog_process_proccsw(pm, pp, tmp); 1661 } 1662 } 1663 1664 /* mark hardware as free */ 1665 pcd->pcd_config_pmc(cpu, adjri, NULL); 1666 } 1667 1668 /* 1669 * perform any other architecture/cpu dependent thread 1670 * switch out functions. 1671 */ 1672 1673 (void) (*md->pmd_switch_out)(pc, pp); 1674 1675 critical_exit(); 1676 } 1677 1678 /* 1679 * A new thread for a process. 1680 */ 1681 static void 1682 pmc_process_thread_add(struct thread *td) 1683 { 1684 struct pmc_process *pmc; 1685 1686 pmc = pmc_find_process_descriptor(td->td_proc, PMC_FLAG_NONE); 1687 if (pmc != NULL) 1688 pmc_find_thread_descriptor(pmc, td, PMC_FLAG_ALLOCATE); 1689 } 1690 1691 /* 1692 * A thread delete for a process. 1693 */ 1694 static void 1695 pmc_process_thread_delete(struct thread *td) 1696 { 1697 struct pmc_process *pmc; 1698 1699 pmc = pmc_find_process_descriptor(td->td_proc, PMC_FLAG_NONE); 1700 if (pmc != NULL) 1701 pmc_thread_descriptor_pool_free(pmc_find_thread_descriptor(pmc, 1702 td, PMC_FLAG_REMOVE)); 1703 } 1704 1705 /* 1706 * A mapping change for a process. 1707 */ 1708 1709 static void 1710 pmc_process_mmap(struct thread *td, struct pmckern_map_in *pkm) 1711 { 1712 int ri; 1713 pid_t pid; 1714 char *fullpath, *freepath; 1715 const struct pmc *pm; 1716 struct pmc_owner *po; 1717 const struct pmc_process *pp; 1718 1719 freepath = fullpath = NULL; 1720 epoch_exit_preempt(global_epoch_preempt); 1721 pmc_getfilename((struct vnode *) pkm->pm_file, &fullpath, &freepath); 1722 1723 pid = td->td_proc->p_pid; 1724 1725 epoch_enter_preempt(global_epoch_preempt); 1726 /* Inform owners of all system-wide sampling PMCs. */ 1727 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) 1728 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 1729 pmclog_process_map_in(po, pid, pkm->pm_address, fullpath); 1730 1731 if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL) 1732 goto done; 1733 1734 /* 1735 * Inform sampling PMC owners tracking this process. 1736 */ 1737 for (ri = 0; ri < md->pmd_npmc; ri++) 1738 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL && 1739 PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) 1740 pmclog_process_map_in(pm->pm_owner, 1741 pid, pkm->pm_address, fullpath); 1742 1743 done: 1744 if (freepath) 1745 free(freepath, M_TEMP); 1746 } 1747 1748 1749 /* 1750 * Log an munmap request. 1751 */ 1752 1753 static void 1754 pmc_process_munmap(struct thread *td, struct pmckern_map_out *pkm) 1755 { 1756 int ri; 1757 pid_t pid; 1758 struct pmc_owner *po; 1759 const struct pmc *pm; 1760 const struct pmc_process *pp; 1761 1762 pid = td->td_proc->p_pid; 1763 1764 epoch_enter_preempt(global_epoch_preempt); 1765 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) 1766 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 1767 pmclog_process_map_out(po, pid, pkm->pm_address, 1768 pkm->pm_address + pkm->pm_size); 1769 epoch_exit_preempt(global_epoch_preempt); 1770 1771 if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL) 1772 return; 1773 1774 for (ri = 0; ri < md->pmd_npmc; ri++) 1775 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL && 1776 PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) 1777 pmclog_process_map_out(pm->pm_owner, pid, 1778 pkm->pm_address, pkm->pm_address + pkm->pm_size); 1779 } 1780 1781 /* 1782 * Log mapping information about the kernel. 1783 */ 1784 1785 static void 1786 pmc_log_kernel_mappings(struct pmc *pm) 1787 { 1788 struct pmc_owner *po; 1789 struct pmckern_map_in *km, *kmbase; 1790 1791 MPASS(in_epoch() || sx_xlocked(&pmc_sx)); 1792 KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)), 1793 ("[pmc,%d] non-sampling PMC (%p) desires mapping information", 1794 __LINE__, (void *) pm)); 1795 1796 po = pm->pm_owner; 1797 1798 if (po->po_flags & PMC_PO_INITIAL_MAPPINGS_DONE) 1799 return; 1800 1801 /* 1802 * Log the current set of kernel modules. 1803 */ 1804 kmbase = linker_hwpmc_list_objects(); 1805 for (km = kmbase; km->pm_file != NULL; km++) { 1806 PMCDBG2(LOG,REG,1,"%s %p", (char *) km->pm_file, 1807 (void *) km->pm_address); 1808 pmclog_process_map_in(po, (pid_t) -1, km->pm_address, 1809 km->pm_file); 1810 } 1811 free(kmbase, M_LINKER); 1812 1813 po->po_flags |= PMC_PO_INITIAL_MAPPINGS_DONE; 1814 } 1815 1816 /* 1817 * Log the mappings for a single process. 1818 */ 1819 1820 static void 1821 pmc_log_process_mappings(struct pmc_owner *po, struct proc *p) 1822 { 1823 vm_map_t map; 1824 struct vnode *vp; 1825 struct vmspace *vm; 1826 vm_map_entry_t entry; 1827 vm_offset_t last_end; 1828 u_int last_timestamp; 1829 struct vnode *last_vp; 1830 vm_offset_t start_addr; 1831 vm_object_t obj, lobj, tobj; 1832 char *fullpath, *freepath; 1833 1834 last_vp = NULL; 1835 last_end = (vm_offset_t) 0; 1836 fullpath = freepath = NULL; 1837 1838 if ((vm = vmspace_acquire_ref(p)) == NULL) 1839 return; 1840 1841 map = &vm->vm_map; 1842 vm_map_lock_read(map); 1843 1844 for (entry = map->header.next; entry != &map->header; entry = entry->next) { 1845 1846 if (entry == NULL) { 1847 PMCDBG2(LOG,OPS,2, "hwpmc: vm_map entry unexpectedly " 1848 "NULL! pid=%d vm_map=%p\n", p->p_pid, map); 1849 break; 1850 } 1851 1852 /* 1853 * We only care about executable map entries. 1854 */ 1855 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) || 1856 !(entry->protection & VM_PROT_EXECUTE) || 1857 (entry->object.vm_object == NULL)) { 1858 continue; 1859 } 1860 1861 obj = entry->object.vm_object; 1862 VM_OBJECT_RLOCK(obj); 1863 1864 /* 1865 * Walk the backing_object list to find the base 1866 * (non-shadowed) vm_object. 1867 */ 1868 for (lobj = tobj = obj; tobj != NULL; tobj = tobj->backing_object) { 1869 if (tobj != obj) 1870 VM_OBJECT_RLOCK(tobj); 1871 if (lobj != obj) 1872 VM_OBJECT_RUNLOCK(lobj); 1873 lobj = tobj; 1874 } 1875 1876 /* 1877 * At this point lobj is the base vm_object and it is locked. 1878 */ 1879 if (lobj == NULL) { 1880 PMCDBG3(LOG,OPS,2, "hwpmc: lobj unexpectedly NULL! pid=%d " 1881 "vm_map=%p vm_obj=%p\n", p->p_pid, map, obj); 1882 VM_OBJECT_RUNLOCK(obj); 1883 continue; 1884 } 1885 1886 vp = vm_object_vnode(lobj); 1887 if (vp == NULL) { 1888 if (lobj != obj) 1889 VM_OBJECT_RUNLOCK(lobj); 1890 VM_OBJECT_RUNLOCK(obj); 1891 continue; 1892 } 1893 1894 /* 1895 * Skip contiguous regions that point to the same 1896 * vnode, so we don't emit redundant MAP-IN 1897 * directives. 1898 */ 1899 if (entry->start == last_end && vp == last_vp) { 1900 last_end = entry->end; 1901 if (lobj != obj) 1902 VM_OBJECT_RUNLOCK(lobj); 1903 VM_OBJECT_RUNLOCK(obj); 1904 continue; 1905 } 1906 1907 /* 1908 * We don't want to keep the proc's vm_map or this 1909 * vm_object locked while we walk the pathname, since 1910 * vn_fullpath() can sleep. However, if we drop the 1911 * lock, it's possible for concurrent activity to 1912 * modify the vm_map list. To protect against this, 1913 * we save the vm_map timestamp before we release the 1914 * lock, and check it after we reacquire the lock 1915 * below. 1916 */ 1917 start_addr = entry->start; 1918 last_end = entry->end; 1919 last_timestamp = map->timestamp; 1920 vm_map_unlock_read(map); 1921 1922 vref(vp); 1923 if (lobj != obj) 1924 VM_OBJECT_RUNLOCK(lobj); 1925 1926 VM_OBJECT_RUNLOCK(obj); 1927 1928 freepath = NULL; 1929 pmc_getfilename(vp, &fullpath, &freepath); 1930 last_vp = vp; 1931 1932 vrele(vp); 1933 1934 vp = NULL; 1935 pmclog_process_map_in(po, p->p_pid, start_addr, fullpath); 1936 if (freepath) 1937 free(freepath, M_TEMP); 1938 1939 vm_map_lock_read(map); 1940 1941 /* 1942 * If our saved timestamp doesn't match, this means 1943 * that the vm_map was modified out from under us and 1944 * we can't trust our current "entry" pointer. Do a 1945 * new lookup for this entry. If there is no entry 1946 * for this address range, vm_map_lookup_entry() will 1947 * return the previous one, so we always want to go to 1948 * entry->next on the next loop iteration. 1949 * 1950 * There is an edge condition here that can occur if 1951 * there is no entry at or before this address. In 1952 * this situation, vm_map_lookup_entry returns 1953 * &map->header, which would cause our loop to abort 1954 * without processing the rest of the map. However, 1955 * in practice this will never happen for process 1956 * vm_map. This is because the executable's text 1957 * segment is the first mapping in the proc's address 1958 * space, and this mapping is never removed until the 1959 * process exits, so there will always be a non-header 1960 * entry at or before the requested address for 1961 * vm_map_lookup_entry to return. 1962 */ 1963 if (map->timestamp != last_timestamp) 1964 vm_map_lookup_entry(map, last_end - 1, &entry); 1965 } 1966 1967 vm_map_unlock_read(map); 1968 vmspace_free(vm); 1969 return; 1970 } 1971 1972 /* 1973 * Log mappings for all processes in the system. 1974 */ 1975 1976 static void 1977 pmc_log_all_process_mappings(struct pmc_owner *po) 1978 { 1979 struct proc *p, *top; 1980 1981 sx_assert(&pmc_sx, SX_XLOCKED); 1982 1983 if ((p = pfind(1)) == NULL) 1984 panic("[pmc,%d] Cannot find init", __LINE__); 1985 1986 PROC_UNLOCK(p); 1987 1988 sx_slock(&proctree_lock); 1989 1990 top = p; 1991 1992 for (;;) { 1993 pmc_log_process_mappings(po, p); 1994 if (!LIST_EMPTY(&p->p_children)) 1995 p = LIST_FIRST(&p->p_children); 1996 else for (;;) { 1997 if (p == top) 1998 goto done; 1999 if (LIST_NEXT(p, p_sibling)) { 2000 p = LIST_NEXT(p, p_sibling); 2001 break; 2002 } 2003 p = p->p_pptr; 2004 } 2005 } 2006 done: 2007 sx_sunlock(&proctree_lock); 2008 } 2009 2010 /* 2011 * The 'hook' invoked from the kernel proper 2012 */ 2013 2014 2015 #ifdef HWPMC_DEBUG 2016 const char *pmc_hooknames[] = { 2017 /* these strings correspond to PMC_FN_* in <sys/pmckern.h> */ 2018 "", 2019 "EXEC", 2020 "CSW-IN", 2021 "CSW-OUT", 2022 "SAMPLE", 2023 "UNUSED1", 2024 "UNUSED2", 2025 "MMAP", 2026 "MUNMAP", 2027 "CALLCHAIN-NMI", 2028 "CALLCHAIN-SOFT", 2029 "SOFTSAMPLING", 2030 "THR-CREATE", 2031 "THR-EXIT", 2032 }; 2033 #endif 2034 2035 static int 2036 pmc_hook_handler(struct thread *td, int function, void *arg) 2037 { 2038 int cpu; 2039 2040 PMCDBG4(MOD,PMH,1, "hook td=%p func=%d \"%s\" arg=%p", td, function, 2041 pmc_hooknames[function], arg); 2042 2043 switch (function) 2044 { 2045 2046 /* 2047 * Process exec() 2048 */ 2049 2050 case PMC_FN_PROCESS_EXEC: 2051 { 2052 char *fullpath, *freepath; 2053 unsigned int ri; 2054 int is_using_hwpmcs; 2055 struct pmc *pm; 2056 struct proc *p; 2057 struct pmc_owner *po; 2058 struct pmc_process *pp; 2059 struct pmckern_procexec *pk; 2060 2061 sx_assert(&pmc_sx, SX_XLOCKED); 2062 2063 p = td->td_proc; 2064 pmc_getfilename(p->p_textvp, &fullpath, &freepath); 2065 2066 pk = (struct pmckern_procexec *) arg; 2067 2068 epoch_enter_preempt(global_epoch_preempt); 2069 /* Inform owners of SS mode PMCs of the exec event. */ 2070 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) 2071 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 2072 pmclog_process_procexec(po, PMC_ID_INVALID, 2073 p->p_pid, pk->pm_entryaddr, fullpath); 2074 epoch_exit_preempt(global_epoch_preempt); 2075 2076 PROC_LOCK(p); 2077 is_using_hwpmcs = p->p_flag & P_HWPMC; 2078 PROC_UNLOCK(p); 2079 2080 if (!is_using_hwpmcs) { 2081 if (freepath) 2082 free(freepath, M_TEMP); 2083 break; 2084 } 2085 2086 /* 2087 * PMCs are not inherited across an exec(): remove any 2088 * PMCs that this process is the owner of. 2089 */ 2090 2091 if ((po = pmc_find_owner_descriptor(p)) != NULL) { 2092 pmc_remove_owner(po); 2093 pmc_destroy_owner_descriptor(po); 2094 } 2095 2096 /* 2097 * If the process being exec'ed is not the target of any 2098 * PMC, we are done. 2099 */ 2100 if ((pp = pmc_find_process_descriptor(p, 0)) == NULL) { 2101 if (freepath) 2102 free(freepath, M_TEMP); 2103 break; 2104 } 2105 2106 /* 2107 * Log the exec event to all monitoring owners. Skip 2108 * owners who have already received the event because 2109 * they had system sampling PMCs active. 2110 */ 2111 for (ri = 0; ri < md->pmd_npmc; ri++) 2112 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) { 2113 po = pm->pm_owner; 2114 if (po->po_sscount == 0 && 2115 po->po_flags & PMC_PO_OWNS_LOGFILE) 2116 pmclog_process_procexec(po, pm->pm_id, 2117 p->p_pid, pk->pm_entryaddr, 2118 fullpath); 2119 } 2120 2121 if (freepath) 2122 free(freepath, M_TEMP); 2123 2124 2125 PMCDBG4(PRC,EXC,1, "exec proc=%p (%d, %s) cred-changed=%d", 2126 p, p->p_pid, p->p_comm, pk->pm_credentialschanged); 2127 2128 if (pk->pm_credentialschanged == 0) /* no change */ 2129 break; 2130 2131 /* 2132 * If the newly exec()'ed process has a different credential 2133 * than before, allow it to be the target of a PMC only if 2134 * the PMC's owner has sufficient privilege. 2135 */ 2136 2137 for (ri = 0; ri < md->pmd_npmc; ri++) 2138 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) 2139 if (pmc_can_attach(pm, td->td_proc) != 0) 2140 pmc_detach_one_process(td->td_proc, 2141 pm, PMC_FLAG_NONE); 2142 2143 KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= (int) md->pmd_npmc, 2144 ("[pmc,%d] Illegal ref count %d on pp %p", __LINE__, 2145 pp->pp_refcnt, pp)); 2146 2147 /* 2148 * If this process is no longer the target of any 2149 * PMCs, we can remove the process entry and free 2150 * up space. 2151 */ 2152 2153 if (pp->pp_refcnt == 0) { 2154 pmc_remove_process_descriptor(pp); 2155 pmc_destroy_process_descriptor(pp); 2156 break; 2157 } 2158 2159 } 2160 break; 2161 2162 case PMC_FN_CSW_IN: 2163 pmc_process_csw_in(td); 2164 break; 2165 2166 case PMC_FN_CSW_OUT: 2167 pmc_process_csw_out(td); 2168 break; 2169 2170 /* 2171 * Process accumulated PC samples. 2172 * 2173 * This function is expected to be called by hardclock() for 2174 * each CPU that has accumulated PC samples. 2175 * 2176 * This function is to be executed on the CPU whose samples 2177 * are being processed. 2178 */ 2179 case PMC_FN_DO_SAMPLES: 2180 2181 /* 2182 * Clear the cpu specific bit in the CPU mask before 2183 * do the rest of the processing. If the NMI handler 2184 * gets invoked after the "atomic_clear_int()" call 2185 * below but before "pmc_process_samples()" gets 2186 * around to processing the interrupt, then we will 2187 * come back here at the next hardclock() tick (and 2188 * may find nothing to do if "pmc_process_samples()" 2189 * had already processed the interrupt). We don't 2190 * lose the interrupt sample. 2191 */ 2192 DPCPU_SET(pmc_sampled, 0); 2193 cpu = PCPU_GET(cpuid); 2194 pmc_process_samples(cpu, PMC_HR); 2195 pmc_process_samples(cpu, PMC_SR); 2196 break; 2197 2198 case PMC_FN_MMAP: 2199 MPASS(in_epoch() || sx_xlocked(&pmc_sx)); 2200 pmc_process_mmap(td, (struct pmckern_map_in *) arg); 2201 break; 2202 2203 case PMC_FN_MUNMAP: 2204 MPASS(in_epoch() || sx_xlocked(&pmc_sx)); 2205 pmc_process_munmap(td, (struct pmckern_map_out *) arg); 2206 break; 2207 2208 case PMC_FN_USER_CALLCHAIN: 2209 /* 2210 * Record a call chain. 2211 */ 2212 KASSERT(td == curthread, ("[pmc,%d] td != curthread", 2213 __LINE__)); 2214 2215 pmc_capture_user_callchain(PCPU_GET(cpuid), PMC_HR, 2216 (struct trapframe *) arg); 2217 td->td_pflags &= ~TDP_CALLCHAIN; 2218 break; 2219 2220 case PMC_FN_USER_CALLCHAIN_SOFT: 2221 /* 2222 * Record a call chain. 2223 */ 2224 KASSERT(td == curthread, ("[pmc,%d] td != curthread", 2225 __LINE__)); 2226 pmc_capture_user_callchain(PCPU_GET(cpuid), PMC_SR, 2227 (struct trapframe *) arg); 2228 td->td_pflags &= ~TDP_CALLCHAIN; 2229 break; 2230 2231 case PMC_FN_SOFT_SAMPLING: 2232 /* 2233 * Call soft PMC sampling intr. 2234 */ 2235 pmc_soft_intr((struct pmckern_soft *) arg); 2236 break; 2237 2238 case PMC_FN_THR_CREATE: 2239 pmc_process_thread_add(td); 2240 break; 2241 2242 case PMC_FN_THR_EXIT: 2243 KASSERT(td == curthread, ("[pmc,%d] td != curthread", 2244 __LINE__)); 2245 pmc_process_thread_delete(td); 2246 break; 2247 2248 default: 2249 #ifdef HWPMC_DEBUG 2250 KASSERT(0, ("[pmc,%d] unknown hook %d\n", __LINE__, function)); 2251 #endif 2252 break; 2253 2254 } 2255 2256 return 0; 2257 } 2258 2259 /* 2260 * allocate a 'struct pmc_owner' descriptor in the owner hash table. 2261 */ 2262 2263 static struct pmc_owner * 2264 pmc_allocate_owner_descriptor(struct proc *p) 2265 { 2266 uint32_t hindex; 2267 struct pmc_owner *po; 2268 struct pmc_ownerhash *poh; 2269 2270 hindex = PMC_HASH_PTR(p, pmc_ownerhashmask); 2271 poh = &pmc_ownerhash[hindex]; 2272 2273 /* allocate space for N pointers and one descriptor struct */ 2274 po = malloc(sizeof(struct pmc_owner), M_PMC, M_WAITOK|M_ZERO); 2275 po->po_owner = p; 2276 LIST_INSERT_HEAD(poh, po, po_next); /* insert into hash table */ 2277 2278 TAILQ_INIT(&po->po_logbuffers); 2279 mtx_init(&po->po_mtx, "pmc-owner-mtx", "pmc-per-proc", MTX_SPIN); 2280 2281 PMCDBG4(OWN,ALL,1, "allocate-owner proc=%p (%d, %s) pmc-owner=%p", 2282 p, p->p_pid, p->p_comm, po); 2283 2284 return po; 2285 } 2286 2287 static void 2288 pmc_destroy_owner_descriptor(struct pmc_owner *po) 2289 { 2290 2291 PMCDBG4(OWN,REL,1, "destroy-owner po=%p proc=%p (%d, %s)", 2292 po, po->po_owner, po->po_owner->p_pid, po->po_owner->p_comm); 2293 2294 mtx_destroy(&po->po_mtx); 2295 free(po, M_PMC); 2296 } 2297 2298 /* 2299 * Allocate a thread descriptor from the free pool. 2300 * 2301 * NOTE: This *can* return NULL. 2302 */ 2303 static struct pmc_thread * 2304 pmc_thread_descriptor_pool_alloc(void) 2305 { 2306 struct pmc_thread *pt; 2307 2308 mtx_lock_spin(&pmc_threadfreelist_mtx); 2309 if ((pt = LIST_FIRST(&pmc_threadfreelist)) != NULL) { 2310 LIST_REMOVE(pt, pt_next); 2311 pmc_threadfreelist_entries--; 2312 } 2313 mtx_unlock_spin(&pmc_threadfreelist_mtx); 2314 2315 return (pt); 2316 } 2317 2318 /* 2319 * Add a thread descriptor to the free pool. We use this instead of free() 2320 * to maintain a cache of free entries. Additionally, we can safely call 2321 * this function when we cannot call free(), such as in a critical section. 2322 * 2323 */ 2324 static void 2325 pmc_thread_descriptor_pool_free(struct pmc_thread *pt) 2326 { 2327 2328 if (pt == NULL) 2329 return; 2330 2331 memset(pt, 0, THREADENTRY_SIZE); 2332 mtx_lock_spin(&pmc_threadfreelist_mtx); 2333 LIST_INSERT_HEAD(&pmc_threadfreelist, pt, pt_next); 2334 pmc_threadfreelist_entries++; 2335 if (pmc_threadfreelist_entries > pmc_threadfreelist_max) 2336 GROUPTASK_ENQUEUE(&free_gtask); 2337 mtx_unlock_spin(&pmc_threadfreelist_mtx); 2338 } 2339 2340 /* 2341 * A callout to manage the free list. 2342 */ 2343 static void 2344 pmc_thread_descriptor_pool_free_task(void *arg __unused) 2345 { 2346 struct pmc_thread *pt; 2347 LIST_HEAD(, pmc_thread) tmplist; 2348 int delta; 2349 2350 LIST_INIT(&tmplist); 2351 /* Determine what changes, if any, we need to make. */ 2352 mtx_lock_spin(&pmc_threadfreelist_mtx); 2353 delta = pmc_threadfreelist_entries - pmc_threadfreelist_max; 2354 while (delta > 0) { 2355 pt = LIST_FIRST(&pmc_threadfreelist); 2356 MPASS(pt); 2357 LIST_REMOVE(pt, pt_next); 2358 LIST_INSERT_HEAD(&tmplist, pt, pt_next); 2359 } 2360 mtx_unlock_spin(&pmc_threadfreelist_mtx); 2361 2362 /* If there are entries to free, free them. */ 2363 while (!LIST_EMPTY(&tmplist)) { 2364 pt = LIST_FIRST(&pmc_threadfreelist); 2365 LIST_REMOVE(pt, pt_next); 2366 free(pt, M_PMC); 2367 } 2368 } 2369 2370 /* 2371 * Drain the thread free pool, freeing all allocations. 2372 */ 2373 static void 2374 pmc_thread_descriptor_pool_drain() 2375 { 2376 struct pmc_thread *pt, *next; 2377 2378 LIST_FOREACH_SAFE(pt, &pmc_threadfreelist, pt_next, next) { 2379 LIST_REMOVE(pt, pt_next); 2380 free(pt, M_PMC); 2381 } 2382 } 2383 2384 /* 2385 * find the descriptor corresponding to thread 'td', adding or removing it 2386 * as specified by 'mode'. 2387 * 2388 * Note that this supports additional mode flags in addition to those 2389 * supported by pmc_find_process_descriptor(): 2390 * PMC_FLAG_NOWAIT: Causes the function to not wait for mallocs. 2391 * This makes it safe to call while holding certain other locks. 2392 */ 2393 2394 static struct pmc_thread * 2395 pmc_find_thread_descriptor(struct pmc_process *pp, struct thread *td, 2396 uint32_t mode) 2397 { 2398 struct pmc_thread *pt = NULL, *ptnew = NULL; 2399 int wait_flag; 2400 2401 KASSERT(td != NULL, ("[pmc,%d] called to add NULL td", __LINE__)); 2402 2403 /* 2404 * Pre-allocate memory in the PMC_FLAG_ALLOCATE case prior to 2405 * acquiring the lock. 2406 */ 2407 if (mode & PMC_FLAG_ALLOCATE) { 2408 if ((ptnew = pmc_thread_descriptor_pool_alloc()) == NULL) { 2409 wait_flag = (mode & PMC_FLAG_NOWAIT) ? M_NOWAIT : 2410 M_WAITOK; 2411 ptnew = malloc(THREADENTRY_SIZE, M_PMC, 2412 wait_flag|M_ZERO); 2413 } 2414 } 2415 2416 mtx_lock_spin(pp->pp_tdslock); 2417 2418 LIST_FOREACH(pt, &pp->pp_tds, pt_next) 2419 if (pt->pt_td == td) 2420 break; 2421 2422 if ((mode & PMC_FLAG_REMOVE) && pt != NULL) 2423 LIST_REMOVE(pt, pt_next); 2424 2425 if ((mode & PMC_FLAG_ALLOCATE) && pt == NULL && ptnew != NULL) { 2426 pt = ptnew; 2427 ptnew = NULL; 2428 pt->pt_td = td; 2429 LIST_INSERT_HEAD(&pp->pp_tds, pt, pt_next); 2430 } 2431 2432 mtx_unlock_spin(pp->pp_tdslock); 2433 2434 if (ptnew != NULL) { 2435 free(ptnew, M_PMC); 2436 } 2437 2438 return pt; 2439 } 2440 2441 /* 2442 * Try to add thread descriptors for each thread in a process. 2443 */ 2444 2445 static void 2446 pmc_add_thread_descriptors_from_proc(struct proc *p, struct pmc_process *pp) 2447 { 2448 struct thread *curtd; 2449 struct pmc_thread **tdlist; 2450 int i, tdcnt, tdlistsz; 2451 2452 KASSERT(!PROC_LOCKED(p), ("[pmc,%d] proc unexpectedly locked", 2453 __LINE__)); 2454 tdcnt = 32; 2455 restart: 2456 tdlistsz = roundup2(tdcnt, 32); 2457 2458 tdcnt = 0; 2459 tdlist = malloc(sizeof(struct pmc_thread*) * tdlistsz, M_TEMP, M_WAITOK); 2460 2461 PROC_LOCK(p); 2462 FOREACH_THREAD_IN_PROC(p, curtd) 2463 tdcnt++; 2464 if (tdcnt >= tdlistsz) { 2465 PROC_UNLOCK(p); 2466 free(tdlist, M_TEMP); 2467 goto restart; 2468 } 2469 /* 2470 * Try to add each thread to the list without sleeping. If unable, 2471 * add to a queue to retry after dropping the process lock. 2472 */ 2473 tdcnt = 0; 2474 FOREACH_THREAD_IN_PROC(p, curtd) { 2475 tdlist[tdcnt] = pmc_find_thread_descriptor(pp, curtd, 2476 PMC_FLAG_ALLOCATE|PMC_FLAG_NOWAIT); 2477 if (tdlist[tdcnt] == NULL) { 2478 PROC_UNLOCK(p); 2479 for (i = 0; i <= tdcnt; i++) 2480 pmc_thread_descriptor_pool_free(tdlist[i]); 2481 free(tdlist, M_TEMP); 2482 goto restart; 2483 } 2484 tdcnt++; 2485 } 2486 PROC_UNLOCK(p); 2487 free(tdlist, M_TEMP); 2488 } 2489 2490 /* 2491 * find the descriptor corresponding to process 'p', adding or removing it 2492 * as specified by 'mode'. 2493 */ 2494 2495 static struct pmc_process * 2496 pmc_find_process_descriptor(struct proc *p, uint32_t mode) 2497 { 2498 uint32_t hindex; 2499 struct pmc_process *pp, *ppnew; 2500 struct pmc_processhash *pph; 2501 2502 hindex = PMC_HASH_PTR(p, pmc_processhashmask); 2503 pph = &pmc_processhash[hindex]; 2504 2505 ppnew = NULL; 2506 2507 /* 2508 * Pre-allocate memory in the PMC_FLAG_ALLOCATE case since we 2509 * cannot call malloc(9) once we hold a spin lock. 2510 */ 2511 if (mode & PMC_FLAG_ALLOCATE) 2512 ppnew = malloc(sizeof(struct pmc_process) + md->pmd_npmc * 2513 sizeof(struct pmc_targetstate), M_PMC, M_WAITOK|M_ZERO); 2514 2515 mtx_lock_spin(&pmc_processhash_mtx); 2516 LIST_FOREACH(pp, pph, pp_next) 2517 if (pp->pp_proc == p) 2518 break; 2519 2520 if ((mode & PMC_FLAG_REMOVE) && pp != NULL) 2521 LIST_REMOVE(pp, pp_next); 2522 2523 if ((mode & PMC_FLAG_ALLOCATE) && pp == NULL && 2524 ppnew != NULL) { 2525 ppnew->pp_proc = p; 2526 LIST_INIT(&ppnew->pp_tds); 2527 ppnew->pp_tdslock = mtx_pool_find(pmc_mtxpool, ppnew); 2528 LIST_INSERT_HEAD(pph, ppnew, pp_next); 2529 mtx_unlock_spin(&pmc_processhash_mtx); 2530 pp = ppnew; 2531 ppnew = NULL; 2532 2533 /* Add thread descriptors for this process' current threads. */ 2534 pmc_add_thread_descriptors_from_proc(p, pp); 2535 } 2536 else 2537 mtx_unlock_spin(&pmc_processhash_mtx); 2538 2539 if (ppnew != NULL) 2540 free(ppnew, M_PMC); 2541 2542 return pp; 2543 } 2544 2545 /* 2546 * remove a process descriptor from the process hash table. 2547 */ 2548 2549 static void 2550 pmc_remove_process_descriptor(struct pmc_process *pp) 2551 { 2552 KASSERT(pp->pp_refcnt == 0, 2553 ("[pmc,%d] Removing process descriptor %p with count %d", 2554 __LINE__, pp, pp->pp_refcnt)); 2555 2556 mtx_lock_spin(&pmc_processhash_mtx); 2557 LIST_REMOVE(pp, pp_next); 2558 mtx_unlock_spin(&pmc_processhash_mtx); 2559 } 2560 2561 /* 2562 * destroy a process descriptor. 2563 */ 2564 2565 static void 2566 pmc_destroy_process_descriptor(struct pmc_process *pp) 2567 { 2568 struct pmc_thread *pmc_td; 2569 2570 while ((pmc_td = LIST_FIRST(&pp->pp_tds)) != NULL) { 2571 LIST_REMOVE(pmc_td, pt_next); 2572 pmc_thread_descriptor_pool_free(pmc_td); 2573 } 2574 free(pp, M_PMC); 2575 } 2576 2577 2578 /* 2579 * find an owner descriptor corresponding to proc 'p' 2580 */ 2581 2582 static struct pmc_owner * 2583 pmc_find_owner_descriptor(struct proc *p) 2584 { 2585 uint32_t hindex; 2586 struct pmc_owner *po; 2587 struct pmc_ownerhash *poh; 2588 2589 hindex = PMC_HASH_PTR(p, pmc_ownerhashmask); 2590 poh = &pmc_ownerhash[hindex]; 2591 2592 po = NULL; 2593 LIST_FOREACH(po, poh, po_next) 2594 if (po->po_owner == p) 2595 break; 2596 2597 PMCDBG5(OWN,FND,1, "find-owner proc=%p (%d, %s) hindex=0x%x -> " 2598 "pmc-owner=%p", p, p->p_pid, p->p_comm, hindex, po); 2599 2600 return po; 2601 } 2602 2603 /* 2604 * pmc_allocate_pmc_descriptor 2605 * 2606 * Allocate a pmc descriptor and initialize its 2607 * fields. 2608 */ 2609 2610 static struct pmc * 2611 pmc_allocate_pmc_descriptor(void) 2612 { 2613 struct pmc *pmc; 2614 2615 pmc = malloc(sizeof(struct pmc), M_PMC, M_WAITOK|M_ZERO); 2616 pmc->pm_runcount = counter_u64_alloc(M_WAITOK); 2617 pmc->pm_pcpu_state = malloc(sizeof(struct pmc_pcpu_state)*mp_ncpus, M_PMC, M_WAITOK|M_ZERO); 2618 PMCDBG1(PMC,ALL,1, "allocate-pmc -> pmc=%p", pmc); 2619 2620 return pmc; 2621 } 2622 2623 /* 2624 * Destroy a pmc descriptor. 2625 */ 2626 2627 static void 2628 pmc_destroy_pmc_descriptor(struct pmc *pm) 2629 { 2630 2631 KASSERT(pm->pm_state == PMC_STATE_DELETED || 2632 pm->pm_state == PMC_STATE_FREE, 2633 ("[pmc,%d] destroying non-deleted PMC", __LINE__)); 2634 KASSERT(LIST_EMPTY(&pm->pm_targets), 2635 ("[pmc,%d] destroying pmc with targets", __LINE__)); 2636 KASSERT(pm->pm_owner == NULL, 2637 ("[pmc,%d] destroying pmc attached to an owner", __LINE__)); 2638 KASSERT(counter_u64_fetch(pm->pm_runcount) == 0, 2639 ("[pmc,%d] pmc has non-zero run count %ld", __LINE__, 2640 (unsigned long)counter_u64_fetch(pm->pm_runcount))); 2641 2642 counter_u64_free(pm->pm_runcount); 2643 free(pm->pm_pcpu_state, M_PMC); 2644 free(pm, M_PMC); 2645 } 2646 2647 static void 2648 pmc_wait_for_pmc_idle(struct pmc *pm) 2649 { 2650 #ifdef HWPMC_DEBUG 2651 volatile int maxloop; 2652 2653 maxloop = 100 * pmc_cpu_max(); 2654 #endif 2655 /* 2656 * Loop (with a forced context switch) till the PMC's runcount 2657 * comes down to zero. 2658 */ 2659 while (counter_u64_fetch(pm->pm_runcount) > 0) { 2660 #ifdef HWPMC_DEBUG 2661 maxloop--; 2662 KASSERT(maxloop > 0, 2663 ("[pmc,%d] (ri%d, rc%ld) waiting too long for " 2664 "pmc to be free", __LINE__, 2665 PMC_TO_ROWINDEX(pm), (unsigned long)counter_u64_fetch(pm->pm_runcount))); 2666 #endif 2667 pmc_force_context_switch(); 2668 } 2669 } 2670 2671 /* 2672 * This function does the following things: 2673 * 2674 * - detaches the PMC from hardware 2675 * - unlinks all target threads that were attached to it 2676 * - removes the PMC from its owner's list 2677 * - destroys the PMC private mutex 2678 * 2679 * Once this function completes, the given pmc pointer can be freed by 2680 * calling pmc_destroy_pmc_descriptor(). 2681 */ 2682 2683 static void 2684 pmc_release_pmc_descriptor(struct pmc *pm) 2685 { 2686 enum pmc_mode mode; 2687 struct pmc_hw *phw; 2688 u_int adjri, ri, cpu; 2689 struct pmc_owner *po; 2690 struct pmc_binding pb; 2691 struct pmc_process *pp; 2692 struct pmc_classdep *pcd; 2693 struct pmc_target *ptgt, *tmp; 2694 2695 sx_assert(&pmc_sx, SX_XLOCKED); 2696 2697 KASSERT(pm, ("[pmc,%d] null pmc", __LINE__)); 2698 2699 ri = PMC_TO_ROWINDEX(pm); 2700 pcd = pmc_ri_to_classdep(md, ri, &adjri); 2701 mode = PMC_TO_MODE(pm); 2702 2703 PMCDBG3(PMC,REL,1, "release-pmc pmc=%p ri=%d mode=%d", pm, ri, 2704 mode); 2705 2706 /* 2707 * First, we take the PMC off hardware. 2708 */ 2709 cpu = 0; 2710 if (PMC_IS_SYSTEM_MODE(mode)) { 2711 2712 /* 2713 * A system mode PMC runs on a specific CPU. Switch 2714 * to this CPU and turn hardware off. 2715 */ 2716 pmc_save_cpu_binding(&pb); 2717 2718 cpu = PMC_TO_CPU(pm); 2719 2720 pmc_select_cpu(cpu); 2721 2722 /* switch off non-stalled CPUs */ 2723 pm->pm_pcpu_state[cpu].pps_cpustate = 0; 2724 if (pm->pm_state == PMC_STATE_RUNNING && 2725 pm->pm_pcpu_state[cpu].pps_stalled == 0) { 2726 2727 phw = pmc_pcpu[cpu]->pc_hwpmcs[ri]; 2728 2729 KASSERT(phw->phw_pmc == pm, 2730 ("[pmc, %d] pmc ptr ri(%d) hw(%p) pm(%p)", 2731 __LINE__, ri, phw->phw_pmc, pm)); 2732 PMCDBG2(PMC,REL,2, "stopping cpu=%d ri=%d", cpu, ri); 2733 2734 critical_enter(); 2735 pcd->pcd_stop_pmc(cpu, adjri); 2736 critical_exit(); 2737 } 2738 2739 PMCDBG2(PMC,REL,2, "decfg cpu=%d ri=%d", cpu, ri); 2740 2741 critical_enter(); 2742 pcd->pcd_config_pmc(cpu, adjri, NULL); 2743 critical_exit(); 2744 2745 /* adjust the global and process count of SS mode PMCs */ 2746 if (mode == PMC_MODE_SS && pm->pm_state == PMC_STATE_RUNNING) { 2747 po = pm->pm_owner; 2748 po->po_sscount--; 2749 if (po->po_sscount == 0) { 2750 atomic_subtract_rel_int(&pmc_ss_count, 1); 2751 CK_LIST_REMOVE(po, po_ssnext); 2752 epoch_wait_preempt(global_epoch_preempt); 2753 } 2754 } 2755 2756 pm->pm_state = PMC_STATE_DELETED; 2757 2758 pmc_restore_cpu_binding(&pb); 2759 2760 /* 2761 * We could have references to this PMC structure in 2762 * the per-cpu sample queues. Wait for the queue to 2763 * drain. 2764 */ 2765 pmc_wait_for_pmc_idle(pm); 2766 2767 } else if (PMC_IS_VIRTUAL_MODE(mode)) { 2768 2769 /* 2770 * A virtual PMC could be running on multiple CPUs at 2771 * a given instant. 2772 * 2773 * By marking its state as DELETED, we ensure that 2774 * this PMC is never further scheduled on hardware. 2775 * 2776 * Then we wait till all CPUs are done with this PMC. 2777 */ 2778 pm->pm_state = PMC_STATE_DELETED; 2779 2780 2781 /* Wait for the PMCs runcount to come to zero. */ 2782 pmc_wait_for_pmc_idle(pm); 2783 2784 /* 2785 * At this point the PMC is off all CPUs and cannot be 2786 * freshly scheduled onto a CPU. It is now safe to 2787 * unlink all targets from this PMC. If a 2788 * process-record's refcount falls to zero, we remove 2789 * it from the hash table. The module-wide SX lock 2790 * protects us from races. 2791 */ 2792 LIST_FOREACH_SAFE(ptgt, &pm->pm_targets, pt_next, tmp) { 2793 pp = ptgt->pt_process; 2794 pmc_unlink_target_process(pm, pp); /* frees 'ptgt' */ 2795 2796 PMCDBG1(PMC,REL,3, "pp->refcnt=%d", pp->pp_refcnt); 2797 2798 /* 2799 * If the target process record shows that no 2800 * PMCs are attached to it, reclaim its space. 2801 */ 2802 2803 if (pp->pp_refcnt == 0) { 2804 pmc_remove_process_descriptor(pp); 2805 pmc_destroy_process_descriptor(pp); 2806 } 2807 } 2808 2809 cpu = curthread->td_oncpu; /* setup cpu for pmd_release() */ 2810 2811 } 2812 2813 /* 2814 * Release any MD resources 2815 */ 2816 (void) pcd->pcd_release_pmc(cpu, adjri, pm); 2817 2818 /* 2819 * Update row disposition 2820 */ 2821 2822 if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm))) 2823 PMC_UNMARK_ROW_STANDALONE(ri); 2824 else 2825 PMC_UNMARK_ROW_THREAD(ri); 2826 2827 /* unlink from the owner's list */ 2828 if (pm->pm_owner) { 2829 LIST_REMOVE(pm, pm_next); 2830 pm->pm_owner = NULL; 2831 } 2832 } 2833 2834 /* 2835 * Register an owner and a pmc. 2836 */ 2837 2838 static int 2839 pmc_register_owner(struct proc *p, struct pmc *pmc) 2840 { 2841 struct pmc_owner *po; 2842 2843 sx_assert(&pmc_sx, SX_XLOCKED); 2844 2845 if ((po = pmc_find_owner_descriptor(p)) == NULL) 2846 if ((po = pmc_allocate_owner_descriptor(p)) == NULL) 2847 return ENOMEM; 2848 2849 KASSERT(pmc->pm_owner == NULL, 2850 ("[pmc,%d] attempting to own an initialized PMC", __LINE__)); 2851 pmc->pm_owner = po; 2852 2853 LIST_INSERT_HEAD(&po->po_pmcs, pmc, pm_next); 2854 2855 PROC_LOCK(p); 2856 p->p_flag |= P_HWPMC; 2857 PROC_UNLOCK(p); 2858 2859 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 2860 pmclog_process_pmcallocate(pmc); 2861 2862 PMCDBG2(PMC,REG,1, "register-owner pmc-owner=%p pmc=%p", 2863 po, pmc); 2864 2865 return 0; 2866 } 2867 2868 /* 2869 * Return the current row disposition: 2870 * == 0 => FREE 2871 * > 0 => PROCESS MODE 2872 * < 0 => SYSTEM MODE 2873 */ 2874 2875 int 2876 pmc_getrowdisp(int ri) 2877 { 2878 return pmc_pmcdisp[ri]; 2879 } 2880 2881 /* 2882 * Check if a PMC at row index 'ri' can be allocated to the current 2883 * process. 2884 * 2885 * Allocation can fail if: 2886 * - the current process is already being profiled by a PMC at index 'ri', 2887 * attached to it via OP_PMCATTACH. 2888 * - the current process has already allocated a PMC at index 'ri' 2889 * via OP_ALLOCATE. 2890 */ 2891 2892 static int 2893 pmc_can_allocate_rowindex(struct proc *p, unsigned int ri, int cpu) 2894 { 2895 enum pmc_mode mode; 2896 struct pmc *pm; 2897 struct pmc_owner *po; 2898 struct pmc_process *pp; 2899 2900 PMCDBG5(PMC,ALR,1, "can-allocate-rowindex proc=%p (%d, %s) ri=%d " 2901 "cpu=%d", p, p->p_pid, p->p_comm, ri, cpu); 2902 2903 /* 2904 * We shouldn't have already allocated a process-mode PMC at 2905 * row index 'ri'. 2906 * 2907 * We shouldn't have allocated a system-wide PMC on the same 2908 * CPU and same RI. 2909 */ 2910 if ((po = pmc_find_owner_descriptor(p)) != NULL) 2911 LIST_FOREACH(pm, &po->po_pmcs, pm_next) { 2912 if (PMC_TO_ROWINDEX(pm) == ri) { 2913 mode = PMC_TO_MODE(pm); 2914 if (PMC_IS_VIRTUAL_MODE(mode)) 2915 return EEXIST; 2916 if (PMC_IS_SYSTEM_MODE(mode) && 2917 (int) PMC_TO_CPU(pm) == cpu) 2918 return EEXIST; 2919 } 2920 } 2921 2922 /* 2923 * We also shouldn't be the target of any PMC at this index 2924 * since otherwise a PMC_ATTACH to ourselves will fail. 2925 */ 2926 if ((pp = pmc_find_process_descriptor(p, 0)) != NULL) 2927 if (pp->pp_pmcs[ri].pp_pmc) 2928 return EEXIST; 2929 2930 PMCDBG4(PMC,ALR,2, "can-allocate-rowindex proc=%p (%d, %s) ri=%d ok", 2931 p, p->p_pid, p->p_comm, ri); 2932 2933 return 0; 2934 } 2935 2936 /* 2937 * Check if a given PMC at row index 'ri' can be currently used in 2938 * mode 'mode'. 2939 */ 2940 2941 static int 2942 pmc_can_allocate_row(int ri, enum pmc_mode mode) 2943 { 2944 enum pmc_disp disp; 2945 2946 sx_assert(&pmc_sx, SX_XLOCKED); 2947 2948 PMCDBG2(PMC,ALR,1, "can-allocate-row ri=%d mode=%d", ri, mode); 2949 2950 if (PMC_IS_SYSTEM_MODE(mode)) 2951 disp = PMC_DISP_STANDALONE; 2952 else 2953 disp = PMC_DISP_THREAD; 2954 2955 /* 2956 * check disposition for PMC row 'ri': 2957 * 2958 * Expected disposition Row-disposition Result 2959 * 2960 * STANDALONE STANDALONE or FREE proceed 2961 * STANDALONE THREAD fail 2962 * THREAD THREAD or FREE proceed 2963 * THREAD STANDALONE fail 2964 */ 2965 2966 if (!PMC_ROW_DISP_IS_FREE(ri) && 2967 !(disp == PMC_DISP_THREAD && PMC_ROW_DISP_IS_THREAD(ri)) && 2968 !(disp == PMC_DISP_STANDALONE && PMC_ROW_DISP_IS_STANDALONE(ri))) 2969 return EBUSY; 2970 2971 /* 2972 * All OK 2973 */ 2974 2975 PMCDBG2(PMC,ALR,2, "can-allocate-row ri=%d mode=%d ok", ri, mode); 2976 2977 return 0; 2978 2979 } 2980 2981 /* 2982 * Find a PMC descriptor with user handle 'pmcid' for thread 'td'. 2983 */ 2984 2985 static struct pmc * 2986 pmc_find_pmc_descriptor_in_process(struct pmc_owner *po, pmc_id_t pmcid) 2987 { 2988 struct pmc *pm; 2989 2990 KASSERT(PMC_ID_TO_ROWINDEX(pmcid) < md->pmd_npmc, 2991 ("[pmc,%d] Illegal pmc index %d (max %d)", __LINE__, 2992 PMC_ID_TO_ROWINDEX(pmcid), md->pmd_npmc)); 2993 2994 LIST_FOREACH(pm, &po->po_pmcs, pm_next) 2995 if (pm->pm_id == pmcid) 2996 return pm; 2997 2998 return NULL; 2999 } 3000 3001 static int 3002 pmc_find_pmc(pmc_id_t pmcid, struct pmc **pmc) 3003 { 3004 3005 struct pmc *pm, *opm; 3006 struct pmc_owner *po; 3007 struct pmc_process *pp; 3008 3009 PMCDBG1(PMC,FND,1, "find-pmc id=%d", pmcid); 3010 if (PMC_ID_TO_ROWINDEX(pmcid) >= md->pmd_npmc) 3011 return (EINVAL); 3012 3013 if ((po = pmc_find_owner_descriptor(curthread->td_proc)) == NULL) { 3014 /* 3015 * In case of PMC_F_DESCENDANTS child processes we will not find 3016 * the current process in the owners hash list. Find the owner 3017 * process first and from there lookup the po. 3018 */ 3019 if ((pp = pmc_find_process_descriptor(curthread->td_proc, 3020 PMC_FLAG_NONE)) == NULL) { 3021 return ESRCH; 3022 } else { 3023 opm = pp->pp_pmcs[PMC_ID_TO_ROWINDEX(pmcid)].pp_pmc; 3024 if (opm == NULL) 3025 return ESRCH; 3026 if ((opm->pm_flags & (PMC_F_ATTACHED_TO_OWNER| 3027 PMC_F_DESCENDANTS)) != (PMC_F_ATTACHED_TO_OWNER| 3028 PMC_F_DESCENDANTS)) 3029 return ESRCH; 3030 po = opm->pm_owner; 3031 } 3032 } 3033 3034 if ((pm = pmc_find_pmc_descriptor_in_process(po, pmcid)) == NULL) 3035 return EINVAL; 3036 3037 PMCDBG2(PMC,FND,2, "find-pmc id=%d -> pmc=%p", pmcid, pm); 3038 3039 *pmc = pm; 3040 return 0; 3041 } 3042 3043 /* 3044 * Start a PMC. 3045 */ 3046 3047 static int 3048 pmc_start(struct pmc *pm) 3049 { 3050 enum pmc_mode mode; 3051 struct pmc_owner *po; 3052 struct pmc_binding pb; 3053 struct pmc_classdep *pcd; 3054 int adjri, error, cpu, ri; 3055 3056 KASSERT(pm != NULL, 3057 ("[pmc,%d] null pm", __LINE__)); 3058 3059 mode = PMC_TO_MODE(pm); 3060 ri = PMC_TO_ROWINDEX(pm); 3061 pcd = pmc_ri_to_classdep(md, ri, &adjri); 3062 3063 error = 0; 3064 3065 PMCDBG3(PMC,OPS,1, "start pmc=%p mode=%d ri=%d", pm, mode, ri); 3066 3067 po = pm->pm_owner; 3068 3069 /* 3070 * Disallow PMCSTART if a logfile is required but has not been 3071 * configured yet. 3072 */ 3073 if ((pm->pm_flags & PMC_F_NEEDS_LOGFILE) && 3074 (po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) 3075 return (EDOOFUS); /* programming error */ 3076 3077 /* 3078 * If this is a sampling mode PMC, log mapping information for 3079 * the kernel modules that are currently loaded. 3080 */ 3081 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) 3082 pmc_log_kernel_mappings(pm); 3083 3084 if (PMC_IS_VIRTUAL_MODE(mode)) { 3085 3086 /* 3087 * If a PMCATTACH has never been done on this PMC, 3088 * attach it to its owner process. 3089 */ 3090 3091 if (LIST_EMPTY(&pm->pm_targets)) 3092 error = (pm->pm_flags & PMC_F_ATTACH_DONE) ? ESRCH : 3093 pmc_attach_process(po->po_owner, pm); 3094 3095 /* 3096 * If the PMC is attached to its owner, then force a context 3097 * switch to ensure that the MD state gets set correctly. 3098 */ 3099 3100 if (error == 0) { 3101 pm->pm_state = PMC_STATE_RUNNING; 3102 if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) 3103 pmc_force_context_switch(); 3104 } 3105 3106 return (error); 3107 } 3108 3109 3110 /* 3111 * A system-wide PMC. 3112 * 3113 * Add the owner to the global list if this is a system-wide 3114 * sampling PMC. 3115 */ 3116 3117 if (mode == PMC_MODE_SS) { 3118 /* 3119 * Log mapping information for all existing processes in the 3120 * system. Subsequent mappings are logged as they happen; 3121 * see pmc_process_mmap(). 3122 */ 3123 if (po->po_logprocmaps == 0) { 3124 pmc_log_all_process_mappings(po); 3125 po->po_logprocmaps = 1; 3126 } 3127 po->po_sscount++; 3128 if (po->po_sscount == 1) { 3129 atomic_add_rel_int(&pmc_ss_count, 1); 3130 CK_LIST_INSERT_HEAD(&pmc_ss_owners, po, po_ssnext); 3131 PMCDBG1(PMC,OPS,1, "po=%p in global list", po); 3132 } 3133 } 3134 3135 /* 3136 * Move to the CPU associated with this 3137 * PMC, and start the hardware. 3138 */ 3139 3140 pmc_save_cpu_binding(&pb); 3141 3142 cpu = PMC_TO_CPU(pm); 3143 3144 if (!pmc_cpu_is_active(cpu)) 3145 return (ENXIO); 3146 3147 pmc_select_cpu(cpu); 3148 3149 /* 3150 * global PMCs are configured at allocation time 3151 * so write out the initial value and start the PMC. 3152 */ 3153 3154 pm->pm_state = PMC_STATE_RUNNING; 3155 3156 critical_enter(); 3157 if ((error = pcd->pcd_write_pmc(cpu, adjri, 3158 PMC_IS_SAMPLING_MODE(mode) ? 3159 pm->pm_sc.pm_reloadcount : 3160 pm->pm_sc.pm_initial)) == 0) { 3161 /* If a sampling mode PMC, reset stalled state. */ 3162 if (PMC_IS_SAMPLING_MODE(mode)) 3163 pm->pm_pcpu_state[cpu].pps_stalled = 0; 3164 3165 /* Indicate that we desire this to run. Start it. */ 3166 pm->pm_pcpu_state[cpu].pps_cpustate = 1; 3167 error = pcd->pcd_start_pmc(cpu, adjri); 3168 } 3169 critical_exit(); 3170 3171 pmc_restore_cpu_binding(&pb); 3172 3173 return (error); 3174 } 3175 3176 /* 3177 * Stop a PMC. 3178 */ 3179 3180 static int 3181 pmc_stop(struct pmc *pm) 3182 { 3183 struct pmc_owner *po; 3184 struct pmc_binding pb; 3185 struct pmc_classdep *pcd; 3186 int adjri, cpu, error, ri; 3187 3188 KASSERT(pm != NULL, ("[pmc,%d] null pmc", __LINE__)); 3189 3190 PMCDBG3(PMC,OPS,1, "stop pmc=%p mode=%d ri=%d", pm, 3191 PMC_TO_MODE(pm), PMC_TO_ROWINDEX(pm)); 3192 3193 pm->pm_state = PMC_STATE_STOPPED; 3194 3195 /* 3196 * If the PMC is a virtual mode one, changing the state to 3197 * non-RUNNING is enough to ensure that the PMC never gets 3198 * scheduled. 3199 * 3200 * If this PMC is current running on a CPU, then it will 3201 * handled correctly at the time its target process is context 3202 * switched out. 3203 */ 3204 3205 if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) 3206 return 0; 3207 3208 /* 3209 * A system-mode PMC. Move to the CPU associated with 3210 * this PMC, and stop the hardware. We update the 3211 * 'initial count' so that a subsequent PMCSTART will 3212 * resume counting from the current hardware count. 3213 */ 3214 3215 pmc_save_cpu_binding(&pb); 3216 3217 cpu = PMC_TO_CPU(pm); 3218 3219 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), 3220 ("[pmc,%d] illegal cpu=%d", __LINE__, cpu)); 3221 3222 if (!pmc_cpu_is_active(cpu)) 3223 return ENXIO; 3224 3225 pmc_select_cpu(cpu); 3226 3227 ri = PMC_TO_ROWINDEX(pm); 3228 pcd = pmc_ri_to_classdep(md, ri, &adjri); 3229 3230 pm->pm_pcpu_state[cpu].pps_cpustate = 0; 3231 critical_enter(); 3232 if ((error = pcd->pcd_stop_pmc(cpu, adjri)) == 0) 3233 error = pcd->pcd_read_pmc(cpu, adjri, &pm->pm_sc.pm_initial); 3234 critical_exit(); 3235 3236 pmc_restore_cpu_binding(&pb); 3237 3238 po = pm->pm_owner; 3239 3240 /* remove this owner from the global list of SS PMC owners */ 3241 if (PMC_TO_MODE(pm) == PMC_MODE_SS) { 3242 po->po_sscount--; 3243 if (po->po_sscount == 0) { 3244 atomic_subtract_rel_int(&pmc_ss_count, 1); 3245 CK_LIST_REMOVE(po, po_ssnext); 3246 epoch_wait_preempt(global_epoch_preempt); 3247 PMCDBG1(PMC,OPS,2,"po=%p removed from global list", po); 3248 } 3249 } 3250 3251 return (error); 3252 } 3253 3254 3255 #ifdef HWPMC_DEBUG 3256 static const char *pmc_op_to_name[] = { 3257 #undef __PMC_OP 3258 #define __PMC_OP(N, D) #N , 3259 __PMC_OPS() 3260 NULL 3261 }; 3262 #endif 3263 3264 /* 3265 * The syscall interface 3266 */ 3267 3268 #define PMC_GET_SX_XLOCK(...) do { \ 3269 sx_xlock(&pmc_sx); \ 3270 if (pmc_hook == NULL) { \ 3271 sx_xunlock(&pmc_sx); \ 3272 return __VA_ARGS__; \ 3273 } \ 3274 } while (0) 3275 3276 #define PMC_DOWNGRADE_SX() do { \ 3277 sx_downgrade(&pmc_sx); \ 3278 is_sx_downgraded = 1; \ 3279 } while (0) 3280 3281 static int 3282 pmc_syscall_handler(struct thread *td, void *syscall_args) 3283 { 3284 int error, is_sx_downgraded, op; 3285 struct pmc_syscall_args *c; 3286 void *pmclog_proc_handle; 3287 void *arg; 3288 3289 c = (struct pmc_syscall_args *)syscall_args; 3290 op = c->pmop_code; 3291 arg = c->pmop_data; 3292 /* PMC isn't set up yet */ 3293 if (pmc_hook == NULL) 3294 return (EINVAL); 3295 if (op == PMC_OP_CONFIGURELOG) { 3296 /* 3297 * We cannot create the logging process inside 3298 * pmclog_configure_log() because there is a LOR 3299 * between pmc_sx and process structure locks. 3300 * Instead, pre-create the process and ignite the loop 3301 * if everything is fine, otherwise direct the process 3302 * to exit. 3303 */ 3304 error = pmclog_proc_create(td, &pmclog_proc_handle); 3305 if (error != 0) 3306 goto done_syscall; 3307 } 3308 3309 PMC_GET_SX_XLOCK(ENOSYS); 3310 is_sx_downgraded = 0; 3311 PMCDBG3(MOD,PMS,1, "syscall op=%d \"%s\" arg=%p", op, 3312 pmc_op_to_name[op], arg); 3313 3314 error = 0; 3315 counter_u64_add(pmc_stats.pm_syscalls, 1); 3316 3317 switch (op) { 3318 3319 3320 /* 3321 * Configure a log file. 3322 * 3323 * XXX This OP will be reworked. 3324 */ 3325 3326 case PMC_OP_CONFIGURELOG: 3327 { 3328 struct proc *p; 3329 struct pmc *pm; 3330 struct pmc_owner *po; 3331 struct pmc_op_configurelog cl; 3332 3333 if ((error = copyin(arg, &cl, sizeof(cl))) != 0) { 3334 pmclog_proc_ignite(pmclog_proc_handle, NULL); 3335 break; 3336 } 3337 3338 /* mark this process as owning a log file */ 3339 p = td->td_proc; 3340 if ((po = pmc_find_owner_descriptor(p)) == NULL) 3341 if ((po = pmc_allocate_owner_descriptor(p)) == NULL) { 3342 pmclog_proc_ignite(pmclog_proc_handle, NULL); 3343 error = ENOMEM; 3344 break; 3345 } 3346 3347 /* 3348 * If a valid fd was passed in, try to configure that, 3349 * otherwise if 'fd' was less than zero and there was 3350 * a log file configured, flush its buffers and 3351 * de-configure it. 3352 */ 3353 if (cl.pm_logfd >= 0) { 3354 error = pmclog_configure_log(md, po, cl.pm_logfd); 3355 pmclog_proc_ignite(pmclog_proc_handle, error == 0 ? 3356 po : NULL); 3357 } else if (po->po_flags & PMC_PO_OWNS_LOGFILE) { 3358 pmclog_proc_ignite(pmclog_proc_handle, NULL); 3359 error = pmclog_close(po); 3360 if (error == 0) { 3361 LIST_FOREACH(pm, &po->po_pmcs, pm_next) 3362 if (pm->pm_flags & PMC_F_NEEDS_LOGFILE && 3363 pm->pm_state == PMC_STATE_RUNNING) 3364 pmc_stop(pm); 3365 error = pmclog_deconfigure_log(po); 3366 } 3367 } else { 3368 pmclog_proc_ignite(pmclog_proc_handle, NULL); 3369 error = EINVAL; 3370 } 3371 } 3372 break; 3373 3374 /* 3375 * Flush a log file. 3376 */ 3377 3378 case PMC_OP_FLUSHLOG: 3379 { 3380 struct pmc_owner *po; 3381 3382 sx_assert(&pmc_sx, SX_XLOCKED); 3383 3384 if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) { 3385 error = EINVAL; 3386 break; 3387 } 3388 3389 error = pmclog_flush(po); 3390 } 3391 break; 3392 3393 /* 3394 * Close a log file. 3395 */ 3396 3397 case PMC_OP_CLOSELOG: 3398 { 3399 struct pmc_owner *po; 3400 3401 sx_assert(&pmc_sx, SX_XLOCKED); 3402 3403 if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) { 3404 error = EINVAL; 3405 break; 3406 } 3407 3408 error = pmclog_close(po); 3409 } 3410 break; 3411 3412 /* 3413 * Retrieve hardware configuration. 3414 */ 3415 3416 case PMC_OP_GETCPUINFO: /* CPU information */ 3417 { 3418 struct pmc_op_getcpuinfo gci; 3419 struct pmc_classinfo *pci; 3420 struct pmc_classdep *pcd; 3421 int cl; 3422 3423 gci.pm_cputype = md->pmd_cputype; 3424 gci.pm_ncpu = pmc_cpu_max(); 3425 gci.pm_npmc = md->pmd_npmc; 3426 gci.pm_nclass = md->pmd_nclass; 3427 pci = gci.pm_classes; 3428 pcd = md->pmd_classdep; 3429 for (cl = 0; cl < md->pmd_nclass; cl++, pci++, pcd++) { 3430 pci->pm_caps = pcd->pcd_caps; 3431 pci->pm_class = pcd->pcd_class; 3432 pci->pm_width = pcd->pcd_width; 3433 pci->pm_num = pcd->pcd_num; 3434 } 3435 error = copyout(&gci, arg, sizeof(gci)); 3436 } 3437 break; 3438 3439 /* 3440 * Retrieve soft events list. 3441 */ 3442 case PMC_OP_GETDYNEVENTINFO: 3443 { 3444 enum pmc_class cl; 3445 enum pmc_event ev; 3446 struct pmc_op_getdyneventinfo *gei; 3447 struct pmc_dyn_event_descr dev; 3448 struct pmc_soft *ps; 3449 uint32_t nevent; 3450 3451 sx_assert(&pmc_sx, SX_LOCKED); 3452 3453 gei = (struct pmc_op_getdyneventinfo *) arg; 3454 3455 if ((error = copyin(&gei->pm_class, &cl, sizeof(cl))) != 0) 3456 break; 3457 3458 /* Only SOFT class is dynamic. */ 3459 if (cl != PMC_CLASS_SOFT) { 3460 error = EINVAL; 3461 break; 3462 } 3463 3464 nevent = 0; 3465 for (ev = PMC_EV_SOFT_FIRST; (int)ev <= PMC_EV_SOFT_LAST; ev++) { 3466 ps = pmc_soft_ev_acquire(ev); 3467 if (ps == NULL) 3468 continue; 3469 bcopy(&ps->ps_ev, &dev, sizeof(dev)); 3470 pmc_soft_ev_release(ps); 3471 3472 error = copyout(&dev, 3473 &gei->pm_events[nevent], 3474 sizeof(struct pmc_dyn_event_descr)); 3475 if (error != 0) 3476 break; 3477 nevent++; 3478 } 3479 if (error != 0) 3480 break; 3481 3482 error = copyout(&nevent, &gei->pm_nevent, 3483 sizeof(nevent)); 3484 } 3485 break; 3486 3487 /* 3488 * Get module statistics 3489 */ 3490 3491 case PMC_OP_GETDRIVERSTATS: 3492 { 3493 struct pmc_op_getdriverstats gms; 3494 #define CFETCH(a, b, field) a.field = counter_u64_fetch(b.field) 3495 CFETCH(gms, pmc_stats, pm_intr_ignored); 3496 CFETCH(gms, pmc_stats, pm_intr_processed); 3497 CFETCH(gms, pmc_stats, pm_intr_bufferfull); 3498 CFETCH(gms, pmc_stats, pm_syscalls); 3499 CFETCH(gms, pmc_stats, pm_syscall_errors); 3500 CFETCH(gms, pmc_stats, pm_buffer_requests); 3501 CFETCH(gms, pmc_stats, pm_buffer_requests_failed); 3502 CFETCH(gms, pmc_stats, pm_log_sweeps); 3503 #undef CFETCH 3504 error = copyout(&gms, arg, sizeof(gms)); 3505 } 3506 break; 3507 3508 3509 /* 3510 * Retrieve module version number 3511 */ 3512 3513 case PMC_OP_GETMODULEVERSION: 3514 { 3515 uint32_t cv, modv; 3516 3517 /* retrieve the client's idea of the ABI version */ 3518 if ((error = copyin(arg, &cv, sizeof(uint32_t))) != 0) 3519 break; 3520 /* don't service clients newer than our driver */ 3521 modv = PMC_VERSION; 3522 if ((cv & 0xFFFF0000) > (modv & 0xFFFF0000)) { 3523 error = EPROGMISMATCH; 3524 break; 3525 } 3526 error = copyout(&modv, arg, sizeof(int)); 3527 } 3528 break; 3529 3530 3531 /* 3532 * Retrieve the state of all the PMCs on a given 3533 * CPU. 3534 */ 3535 3536 case PMC_OP_GETPMCINFO: 3537 { 3538 int ari; 3539 struct pmc *pm; 3540 size_t pmcinfo_size; 3541 uint32_t cpu, n, npmc; 3542 struct pmc_owner *po; 3543 struct pmc_binding pb; 3544 struct pmc_classdep *pcd; 3545 struct pmc_info *p, *pmcinfo; 3546 struct pmc_op_getpmcinfo *gpi; 3547 3548 PMC_DOWNGRADE_SX(); 3549 3550 gpi = (struct pmc_op_getpmcinfo *) arg; 3551 3552 if ((error = copyin(&gpi->pm_cpu, &cpu, sizeof(cpu))) != 0) 3553 break; 3554 3555 if (cpu >= pmc_cpu_max()) { 3556 error = EINVAL; 3557 break; 3558 } 3559 3560 if (!pmc_cpu_is_active(cpu)) { 3561 error = ENXIO; 3562 break; 3563 } 3564 3565 /* switch to CPU 'cpu' */ 3566 pmc_save_cpu_binding(&pb); 3567 pmc_select_cpu(cpu); 3568 3569 npmc = md->pmd_npmc; 3570 3571 pmcinfo_size = npmc * sizeof(struct pmc_info); 3572 pmcinfo = malloc(pmcinfo_size, M_PMC, M_WAITOK); 3573 3574 p = pmcinfo; 3575 3576 for (n = 0; n < md->pmd_npmc; n++, p++) { 3577 3578 pcd = pmc_ri_to_classdep(md, n, &ari); 3579 3580 KASSERT(pcd != NULL, 3581 ("[pmc,%d] null pcd ri=%d", __LINE__, n)); 3582 3583 if ((error = pcd->pcd_describe(cpu, ari, p, &pm)) != 0) 3584 break; 3585 3586 if (PMC_ROW_DISP_IS_STANDALONE(n)) 3587 p->pm_rowdisp = PMC_DISP_STANDALONE; 3588 else if (PMC_ROW_DISP_IS_THREAD(n)) 3589 p->pm_rowdisp = PMC_DISP_THREAD; 3590 else 3591 p->pm_rowdisp = PMC_DISP_FREE; 3592 3593 p->pm_ownerpid = -1; 3594 3595 if (pm == NULL) /* no PMC associated */ 3596 continue; 3597 3598 po = pm->pm_owner; 3599 3600 KASSERT(po->po_owner != NULL, 3601 ("[pmc,%d] pmc_owner had a null proc pointer", 3602 __LINE__)); 3603 3604 p->pm_ownerpid = po->po_owner->p_pid; 3605 p->pm_mode = PMC_TO_MODE(pm); 3606 p->pm_event = pm->pm_event; 3607 p->pm_flags = pm->pm_flags; 3608 3609 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) 3610 p->pm_reloadcount = 3611 pm->pm_sc.pm_reloadcount; 3612 } 3613 3614 pmc_restore_cpu_binding(&pb); 3615 3616 /* now copy out the PMC info collected */ 3617 if (error == 0) 3618 error = copyout(pmcinfo, &gpi->pm_pmcs, pmcinfo_size); 3619 3620 free(pmcinfo, M_PMC); 3621 } 3622 break; 3623 3624 3625 /* 3626 * Set the administrative state of a PMC. I.e. whether 3627 * the PMC is to be used or not. 3628 */ 3629 3630 case PMC_OP_PMCADMIN: 3631 { 3632 int cpu, ri; 3633 enum pmc_state request; 3634 struct pmc_cpu *pc; 3635 struct pmc_hw *phw; 3636 struct pmc_op_pmcadmin pma; 3637 struct pmc_binding pb; 3638 3639 sx_assert(&pmc_sx, SX_XLOCKED); 3640 3641 KASSERT(td == curthread, 3642 ("[pmc,%d] td != curthread", __LINE__)); 3643 3644 error = priv_check(td, PRIV_PMC_MANAGE); 3645 if (error) 3646 break; 3647 3648 if ((error = copyin(arg, &pma, sizeof(pma))) != 0) 3649 break; 3650 3651 cpu = pma.pm_cpu; 3652 3653 if (cpu < 0 || cpu >= (int) pmc_cpu_max()) { 3654 error = EINVAL; 3655 break; 3656 } 3657 3658 if (!pmc_cpu_is_active(cpu)) { 3659 error = ENXIO; 3660 break; 3661 } 3662 3663 request = pma.pm_state; 3664 3665 if (request != PMC_STATE_DISABLED && 3666 request != PMC_STATE_FREE) { 3667 error = EINVAL; 3668 break; 3669 } 3670 3671 ri = pma.pm_pmc; /* pmc id == row index */ 3672 if (ri < 0 || ri >= (int) md->pmd_npmc) { 3673 error = EINVAL; 3674 break; 3675 } 3676 3677 /* 3678 * We can't disable a PMC with a row-index allocated 3679 * for process virtual PMCs. 3680 */ 3681 3682 if (PMC_ROW_DISP_IS_THREAD(ri) && 3683 request == PMC_STATE_DISABLED) { 3684 error = EBUSY; 3685 break; 3686 } 3687 3688 /* 3689 * otherwise, this PMC on this CPU is either free or 3690 * in system-wide mode. 3691 */ 3692 3693 pmc_save_cpu_binding(&pb); 3694 pmc_select_cpu(cpu); 3695 3696 pc = pmc_pcpu[cpu]; 3697 phw = pc->pc_hwpmcs[ri]; 3698 3699 /* 3700 * XXX do we need some kind of 'forced' disable? 3701 */ 3702 3703 if (phw->phw_pmc == NULL) { 3704 if (request == PMC_STATE_DISABLED && 3705 (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED)) { 3706 phw->phw_state &= ~PMC_PHW_FLAG_IS_ENABLED; 3707 PMC_MARK_ROW_STANDALONE(ri); 3708 } else if (request == PMC_STATE_FREE && 3709 (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0) { 3710 phw->phw_state |= PMC_PHW_FLAG_IS_ENABLED; 3711 PMC_UNMARK_ROW_STANDALONE(ri); 3712 } 3713 /* other cases are a no-op */ 3714 } else 3715 error = EBUSY; 3716 3717 pmc_restore_cpu_binding(&pb); 3718 } 3719 break; 3720 3721 3722 /* 3723 * Allocate a PMC. 3724 */ 3725 3726 case PMC_OP_PMCALLOCATE: 3727 { 3728 int adjri, n; 3729 u_int cpu; 3730 uint32_t caps; 3731 struct pmc *pmc; 3732 enum pmc_mode mode; 3733 struct pmc_hw *phw; 3734 struct pmc_binding pb; 3735 struct pmc_classdep *pcd; 3736 struct pmc_op_pmcallocate pa; 3737 3738 if ((error = copyin(arg, &pa, sizeof(pa))) != 0) 3739 break; 3740 3741 caps = pa.pm_caps; 3742 mode = pa.pm_mode; 3743 cpu = pa.pm_cpu; 3744 3745 if ((mode != PMC_MODE_SS && mode != PMC_MODE_SC && 3746 mode != PMC_MODE_TS && mode != PMC_MODE_TC) || 3747 (cpu != (u_int) PMC_CPU_ANY && cpu >= pmc_cpu_max())) { 3748 error = EINVAL; 3749 break; 3750 } 3751 3752 /* 3753 * Virtual PMCs should only ask for a default CPU. 3754 * System mode PMCs need to specify a non-default CPU. 3755 */ 3756 3757 if ((PMC_IS_VIRTUAL_MODE(mode) && cpu != (u_int) PMC_CPU_ANY) || 3758 (PMC_IS_SYSTEM_MODE(mode) && cpu == (u_int) PMC_CPU_ANY)) { 3759 error = EINVAL; 3760 break; 3761 } 3762 3763 /* 3764 * Check that an inactive CPU is not being asked for. 3765 */ 3766 3767 if (PMC_IS_SYSTEM_MODE(mode) && !pmc_cpu_is_active(cpu)) { 3768 error = ENXIO; 3769 break; 3770 } 3771 3772 /* 3773 * Refuse an allocation for a system-wide PMC if this 3774 * process has been jailed, or if this process lacks 3775 * super-user credentials and the sysctl tunable 3776 * 'security.bsd.unprivileged_syspmcs' is zero. 3777 */ 3778 3779 if (PMC_IS_SYSTEM_MODE(mode)) { 3780 if (jailed(curthread->td_ucred)) { 3781 error = EPERM; 3782 break; 3783 } 3784 if (!pmc_unprivileged_syspmcs) { 3785 error = priv_check(curthread, 3786 PRIV_PMC_SYSTEM); 3787 if (error) 3788 break; 3789 } 3790 } 3791 3792 /* 3793 * Look for valid values for 'pm_flags' 3794 */ 3795 3796 if ((pa.pm_flags & ~(PMC_F_DESCENDANTS | PMC_F_LOG_PROCCSW | 3797 PMC_F_LOG_PROCEXIT | PMC_F_CALLCHAIN)) != 0) { 3798 error = EINVAL; 3799 break; 3800 } 3801 3802 /* process logging options are not allowed for system PMCs */ 3803 if (PMC_IS_SYSTEM_MODE(mode) && (pa.pm_flags & 3804 (PMC_F_LOG_PROCCSW | PMC_F_LOG_PROCEXIT))) { 3805 error = EINVAL; 3806 break; 3807 } 3808 3809 /* 3810 * All sampling mode PMCs need to be able to interrupt the 3811 * CPU. 3812 */ 3813 if (PMC_IS_SAMPLING_MODE(mode)) 3814 caps |= PMC_CAP_INTERRUPT; 3815 3816 /* A valid class specifier should have been passed in. */ 3817 for (n = 0; n < md->pmd_nclass; n++) 3818 if (md->pmd_classdep[n].pcd_class == pa.pm_class) 3819 break; 3820 if (n == md->pmd_nclass) { 3821 error = EINVAL; 3822 break; 3823 } 3824 3825 /* The requested PMC capabilities should be feasible. */ 3826 if ((md->pmd_classdep[n].pcd_caps & caps) != caps) { 3827 error = EOPNOTSUPP; 3828 break; 3829 } 3830 3831 PMCDBG4(PMC,ALL,2, "event=%d caps=0x%x mode=%d cpu=%d", 3832 pa.pm_ev, caps, mode, cpu); 3833 3834 pmc = pmc_allocate_pmc_descriptor(); 3835 pmc->pm_id = PMC_ID_MAKE_ID(cpu,pa.pm_mode,pa.pm_class, 3836 PMC_ID_INVALID); 3837 pmc->pm_event = pa.pm_ev; 3838 pmc->pm_state = PMC_STATE_FREE; 3839 pmc->pm_caps = caps; 3840 pmc->pm_flags = pa.pm_flags; 3841 3842 /* switch thread to CPU 'cpu' */ 3843 pmc_save_cpu_binding(&pb); 3844 3845 #define PMC_IS_SHAREABLE_PMC(cpu, n) \ 3846 (pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_state & \ 3847 PMC_PHW_FLAG_IS_SHAREABLE) 3848 #define PMC_IS_UNALLOCATED(cpu, n) \ 3849 (pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_pmc == NULL) 3850 3851 if (PMC_IS_SYSTEM_MODE(mode)) { 3852 pmc_select_cpu(cpu); 3853 for (n = 0; n < (int) md->pmd_npmc; n++) { 3854 pcd = pmc_ri_to_classdep(md, n, &adjri); 3855 if (pmc_can_allocate_row(n, mode) == 0 && 3856 pmc_can_allocate_rowindex( 3857 curthread->td_proc, n, cpu) == 0 && 3858 (PMC_IS_UNALLOCATED(cpu, n) || 3859 PMC_IS_SHAREABLE_PMC(cpu, n)) && 3860 pcd->pcd_allocate_pmc(cpu, adjri, pmc, 3861 &pa) == 0) 3862 break; 3863 } 3864 } else { 3865 /* Process virtual mode */ 3866 for (n = 0; n < (int) md->pmd_npmc; n++) { 3867 pcd = pmc_ri_to_classdep(md, n, &adjri); 3868 if (pmc_can_allocate_row(n, mode) == 0 && 3869 pmc_can_allocate_rowindex( 3870 curthread->td_proc, n, 3871 PMC_CPU_ANY) == 0 && 3872 pcd->pcd_allocate_pmc(curthread->td_oncpu, 3873 adjri, pmc, &pa) == 0) 3874 break; 3875 } 3876 } 3877 3878 #undef PMC_IS_UNALLOCATED 3879 #undef PMC_IS_SHAREABLE_PMC 3880 3881 pmc_restore_cpu_binding(&pb); 3882 3883 if (n == (int) md->pmd_npmc) { 3884 pmc_destroy_pmc_descriptor(pmc); 3885 pmc = NULL; 3886 error = EINVAL; 3887 break; 3888 } 3889 3890 /* Fill in the correct value in the ID field */ 3891 pmc->pm_id = PMC_ID_MAKE_ID(cpu,mode,pa.pm_class,n); 3892 3893 PMCDBG5(PMC,ALL,2, "ev=%d class=%d mode=%d n=%d -> pmcid=%x", 3894 pmc->pm_event, pa.pm_class, mode, n, pmc->pm_id); 3895 3896 /* Process mode PMCs with logging enabled need log files */ 3897 if (pmc->pm_flags & (PMC_F_LOG_PROCEXIT | PMC_F_LOG_PROCCSW)) 3898 pmc->pm_flags |= PMC_F_NEEDS_LOGFILE; 3899 3900 /* All system mode sampling PMCs require a log file */ 3901 if (PMC_IS_SAMPLING_MODE(mode) && PMC_IS_SYSTEM_MODE(mode)) 3902 pmc->pm_flags |= PMC_F_NEEDS_LOGFILE; 3903 3904 /* 3905 * Configure global pmc's immediately 3906 */ 3907 3908 if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pmc))) { 3909 3910 pmc_save_cpu_binding(&pb); 3911 pmc_select_cpu(cpu); 3912 3913 phw = pmc_pcpu[cpu]->pc_hwpmcs[n]; 3914 pcd = pmc_ri_to_classdep(md, n, &adjri); 3915 3916 if ((phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0 || 3917 (error = pcd->pcd_config_pmc(cpu, adjri, pmc)) != 0) { 3918 (void) pcd->pcd_release_pmc(cpu, adjri, pmc); 3919 pmc_destroy_pmc_descriptor(pmc); 3920 pmc = NULL; 3921 pmc_restore_cpu_binding(&pb); 3922 error = EPERM; 3923 break; 3924 } 3925 3926 pmc_restore_cpu_binding(&pb); 3927 } 3928 3929 pmc->pm_state = PMC_STATE_ALLOCATED; 3930 3931 /* 3932 * mark row disposition 3933 */ 3934 3935 if (PMC_IS_SYSTEM_MODE(mode)) 3936 PMC_MARK_ROW_STANDALONE(n); 3937 else 3938 PMC_MARK_ROW_THREAD(n); 3939 3940 /* 3941 * Register this PMC with the current thread as its owner. 3942 */ 3943 3944 if ((error = 3945 pmc_register_owner(curthread->td_proc, pmc)) != 0) { 3946 pmc_release_pmc_descriptor(pmc); 3947 pmc_destroy_pmc_descriptor(pmc); 3948 pmc = NULL; 3949 break; 3950 } 3951 3952 /* 3953 * Return the allocated index. 3954 */ 3955 3956 pa.pm_pmcid = pmc->pm_id; 3957 3958 error = copyout(&pa, arg, sizeof(pa)); 3959 } 3960 break; 3961 3962 3963 /* 3964 * Attach a PMC to a process. 3965 */ 3966 3967 case PMC_OP_PMCATTACH: 3968 { 3969 struct pmc *pm; 3970 struct proc *p; 3971 struct pmc_op_pmcattach a; 3972 3973 sx_assert(&pmc_sx, SX_XLOCKED); 3974 3975 if ((error = copyin(arg, &a, sizeof(a))) != 0) 3976 break; 3977 3978 if (a.pm_pid < 0) { 3979 error = EINVAL; 3980 break; 3981 } else if (a.pm_pid == 0) 3982 a.pm_pid = td->td_proc->p_pid; 3983 3984 if ((error = pmc_find_pmc(a.pm_pmc, &pm)) != 0) 3985 break; 3986 3987 if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm))) { 3988 error = EINVAL; 3989 break; 3990 } 3991 3992 /* PMCs may be (re)attached only when allocated or stopped */ 3993 if (pm->pm_state == PMC_STATE_RUNNING) { 3994 error = EBUSY; 3995 break; 3996 } else if (pm->pm_state != PMC_STATE_ALLOCATED && 3997 pm->pm_state != PMC_STATE_STOPPED) { 3998 error = EINVAL; 3999 break; 4000 } 4001 4002 /* lookup pid */ 4003 if ((p = pfind(a.pm_pid)) == NULL) { 4004 error = ESRCH; 4005 break; 4006 } 4007 4008 /* 4009 * Ignore processes that are working on exiting. 4010 */ 4011 if (p->p_flag & P_WEXIT) { 4012 error = ESRCH; 4013 PROC_UNLOCK(p); /* pfind() returns a locked process */ 4014 break; 4015 } 4016 4017 /* 4018 * we are allowed to attach a PMC to a process if 4019 * we can debug it. 4020 */ 4021 error = p_candebug(curthread, p); 4022 4023 PROC_UNLOCK(p); 4024 4025 if (error == 0) 4026 error = pmc_attach_process(p, pm); 4027 } 4028 break; 4029 4030 4031 /* 4032 * Detach an attached PMC from a process. 4033 */ 4034 4035 case PMC_OP_PMCDETACH: 4036 { 4037 struct pmc *pm; 4038 struct proc *p; 4039 struct pmc_op_pmcattach a; 4040 4041 if ((error = copyin(arg, &a, sizeof(a))) != 0) 4042 break; 4043 4044 if (a.pm_pid < 0) { 4045 error = EINVAL; 4046 break; 4047 } else if (a.pm_pid == 0) 4048 a.pm_pid = td->td_proc->p_pid; 4049 4050 if ((error = pmc_find_pmc(a.pm_pmc, &pm)) != 0) 4051 break; 4052 4053 if ((p = pfind(a.pm_pid)) == NULL) { 4054 error = ESRCH; 4055 break; 4056 } 4057 4058 /* 4059 * Treat processes that are in the process of exiting 4060 * as if they were not present. 4061 */ 4062 4063 if (p->p_flag & P_WEXIT) 4064 error = ESRCH; 4065 4066 PROC_UNLOCK(p); /* pfind() returns a locked process */ 4067 4068 if (error == 0) 4069 error = pmc_detach_process(p, pm); 4070 } 4071 break; 4072 4073 4074 /* 4075 * Retrieve the MSR number associated with the counter 4076 * 'pmc_id'. This allows processes to directly use RDPMC 4077 * instructions to read their PMCs, without the overhead of a 4078 * system call. 4079 */ 4080 4081 case PMC_OP_PMCGETMSR: 4082 { 4083 int adjri, ri; 4084 struct pmc *pm; 4085 struct pmc_target *pt; 4086 struct pmc_op_getmsr gm; 4087 struct pmc_classdep *pcd; 4088 4089 PMC_DOWNGRADE_SX(); 4090 4091 if ((error = copyin(arg, &gm, sizeof(gm))) != 0) 4092 break; 4093 4094 if ((error = pmc_find_pmc(gm.pm_pmcid, &pm)) != 0) 4095 break; 4096 4097 /* 4098 * The allocated PMC has to be a process virtual PMC, 4099 * i.e., of type MODE_T[CS]. Global PMCs can only be 4100 * read using the PMCREAD operation since they may be 4101 * allocated on a different CPU than the one we could 4102 * be running on at the time of the RDPMC instruction. 4103 * 4104 * The GETMSR operation is not allowed for PMCs that 4105 * are inherited across processes. 4106 */ 4107 4108 if (!PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)) || 4109 (pm->pm_flags & PMC_F_DESCENDANTS)) { 4110 error = EINVAL; 4111 break; 4112 } 4113 4114 /* 4115 * It only makes sense to use a RDPMC (or its 4116 * equivalent instruction on non-x86 architectures) on 4117 * a process that has allocated and attached a PMC to 4118 * itself. Conversely the PMC is only allowed to have 4119 * one process attached to it -- its owner. 4120 */ 4121 4122 if ((pt = LIST_FIRST(&pm->pm_targets)) == NULL || 4123 LIST_NEXT(pt, pt_next) != NULL || 4124 pt->pt_process->pp_proc != pm->pm_owner->po_owner) { 4125 error = EINVAL; 4126 break; 4127 } 4128 4129 ri = PMC_TO_ROWINDEX(pm); 4130 pcd = pmc_ri_to_classdep(md, ri, &adjri); 4131 4132 /* PMC class has no 'GETMSR' support */ 4133 if (pcd->pcd_get_msr == NULL) { 4134 error = ENOSYS; 4135 break; 4136 } 4137 4138 if ((error = (*pcd->pcd_get_msr)(adjri, &gm.pm_msr)) < 0) 4139 break; 4140 4141 if ((error = copyout(&gm, arg, sizeof(gm))) < 0) 4142 break; 4143 4144 /* 4145 * Mark our process as using MSRs. Update machine 4146 * state using a forced context switch. 4147 */ 4148 4149 pt->pt_process->pp_flags |= PMC_PP_ENABLE_MSR_ACCESS; 4150 pmc_force_context_switch(); 4151 4152 } 4153 break; 4154 4155 /* 4156 * Release an allocated PMC 4157 */ 4158 4159 case PMC_OP_PMCRELEASE: 4160 { 4161 pmc_id_t pmcid; 4162 struct pmc *pm; 4163 struct pmc_owner *po; 4164 struct pmc_op_simple sp; 4165 4166 /* 4167 * Find PMC pointer for the named PMC. 4168 * 4169 * Use pmc_release_pmc_descriptor() to switch off the 4170 * PMC, remove all its target threads, and remove the 4171 * PMC from its owner's list. 4172 * 4173 * Remove the owner record if this is the last PMC 4174 * owned. 4175 * 4176 * Free up space. 4177 */ 4178 4179 if ((error = copyin(arg, &sp, sizeof(sp))) != 0) 4180 break; 4181 4182 pmcid = sp.pm_pmcid; 4183 4184 if ((error = pmc_find_pmc(pmcid, &pm)) != 0) 4185 break; 4186 4187 po = pm->pm_owner; 4188 pmc_release_pmc_descriptor(pm); 4189 pmc_maybe_remove_owner(po); 4190 pmc_destroy_pmc_descriptor(pm); 4191 } 4192 break; 4193 4194 4195 /* 4196 * Read and/or write a PMC. 4197 */ 4198 4199 case PMC_OP_PMCRW: 4200 { 4201 int adjri; 4202 struct pmc *pm; 4203 uint32_t cpu, ri; 4204 pmc_value_t oldvalue; 4205 struct pmc_binding pb; 4206 struct pmc_op_pmcrw prw; 4207 struct pmc_classdep *pcd; 4208 struct pmc_op_pmcrw *pprw; 4209 4210 PMC_DOWNGRADE_SX(); 4211 4212 if ((error = copyin(arg, &prw, sizeof(prw))) != 0) 4213 break; 4214 4215 ri = 0; 4216 PMCDBG2(PMC,OPS,1, "rw id=%d flags=0x%x", prw.pm_pmcid, 4217 prw.pm_flags); 4218 4219 /* must have at least one flag set */ 4220 if ((prw.pm_flags & (PMC_F_OLDVALUE|PMC_F_NEWVALUE)) == 0) { 4221 error = EINVAL; 4222 break; 4223 } 4224 4225 /* locate pmc descriptor */ 4226 if ((error = pmc_find_pmc(prw.pm_pmcid, &pm)) != 0) 4227 break; 4228 4229 /* Can't read a PMC that hasn't been started. */ 4230 if (pm->pm_state != PMC_STATE_ALLOCATED && 4231 pm->pm_state != PMC_STATE_STOPPED && 4232 pm->pm_state != PMC_STATE_RUNNING) { 4233 error = EINVAL; 4234 break; 4235 } 4236 4237 /* writing a new value is allowed only for 'STOPPED' pmcs */ 4238 if (pm->pm_state == PMC_STATE_RUNNING && 4239 (prw.pm_flags & PMC_F_NEWVALUE)) { 4240 error = EBUSY; 4241 break; 4242 } 4243 4244 if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) { 4245 4246 /* 4247 * If this PMC is attached to its owner (i.e., 4248 * the process requesting this operation) and 4249 * is running, then attempt to get an 4250 * upto-date reading from hardware for a READ. 4251 * Writes are only allowed when the PMC is 4252 * stopped, so only update the saved value 4253 * field. 4254 * 4255 * If the PMC is not running, or is not 4256 * attached to its owner, read/write to the 4257 * savedvalue field. 4258 */ 4259 4260 ri = PMC_TO_ROWINDEX(pm); 4261 pcd = pmc_ri_to_classdep(md, ri, &adjri); 4262 4263 mtx_pool_lock_spin(pmc_mtxpool, pm); 4264 cpu = curthread->td_oncpu; 4265 4266 if (prw.pm_flags & PMC_F_OLDVALUE) { 4267 if ((pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) && 4268 (pm->pm_state == PMC_STATE_RUNNING)) 4269 error = (*pcd->pcd_read_pmc)(cpu, adjri, 4270 &oldvalue); 4271 else 4272 oldvalue = pm->pm_gv.pm_savedvalue; 4273 } 4274 if (prw.pm_flags & PMC_F_NEWVALUE) 4275 pm->pm_gv.pm_savedvalue = prw.pm_value; 4276 4277 mtx_pool_unlock_spin(pmc_mtxpool, pm); 4278 4279 } else { /* System mode PMCs */ 4280 cpu = PMC_TO_CPU(pm); 4281 ri = PMC_TO_ROWINDEX(pm); 4282 pcd = pmc_ri_to_classdep(md, ri, &adjri); 4283 4284 if (!pmc_cpu_is_active(cpu)) { 4285 error = ENXIO; 4286 break; 4287 } 4288 4289 /* move this thread to CPU 'cpu' */ 4290 pmc_save_cpu_binding(&pb); 4291 pmc_select_cpu(cpu); 4292 4293 critical_enter(); 4294 /* save old value */ 4295 if (prw.pm_flags & PMC_F_OLDVALUE) 4296 if ((error = (*pcd->pcd_read_pmc)(cpu, adjri, 4297 &oldvalue))) 4298 goto error; 4299 /* write out new value */ 4300 if (prw.pm_flags & PMC_F_NEWVALUE) 4301 error = (*pcd->pcd_write_pmc)(cpu, adjri, 4302 prw.pm_value); 4303 error: 4304 critical_exit(); 4305 pmc_restore_cpu_binding(&pb); 4306 if (error) 4307 break; 4308 } 4309 4310 pprw = (struct pmc_op_pmcrw *) arg; 4311 4312 #ifdef HWPMC_DEBUG 4313 if (prw.pm_flags & PMC_F_NEWVALUE) 4314 PMCDBG3(PMC,OPS,2, "rw id=%d new %jx -> old %jx", 4315 ri, prw.pm_value, oldvalue); 4316 else if (prw.pm_flags & PMC_F_OLDVALUE) 4317 PMCDBG2(PMC,OPS,2, "rw id=%d -> old %jx", ri, oldvalue); 4318 #endif 4319 4320 /* return old value if requested */ 4321 if (prw.pm_flags & PMC_F_OLDVALUE) 4322 if ((error = copyout(&oldvalue, &pprw->pm_value, 4323 sizeof(prw.pm_value)))) 4324 break; 4325 4326 } 4327 break; 4328 4329 4330 /* 4331 * Set the sampling rate for a sampling mode PMC and the 4332 * initial count for a counting mode PMC. 4333 */ 4334 4335 case PMC_OP_PMCSETCOUNT: 4336 { 4337 struct pmc *pm; 4338 struct pmc_op_pmcsetcount sc; 4339 4340 PMC_DOWNGRADE_SX(); 4341 4342 if ((error = copyin(arg, &sc, sizeof(sc))) != 0) 4343 break; 4344 4345 if ((error = pmc_find_pmc(sc.pm_pmcid, &pm)) != 0) 4346 break; 4347 4348 if (pm->pm_state == PMC_STATE_RUNNING) { 4349 error = EBUSY; 4350 break; 4351 } 4352 4353 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) 4354 pm->pm_sc.pm_reloadcount = sc.pm_count; 4355 else 4356 pm->pm_sc.pm_initial = sc.pm_count; 4357 } 4358 break; 4359 4360 4361 /* 4362 * Start a PMC. 4363 */ 4364 4365 case PMC_OP_PMCSTART: 4366 { 4367 pmc_id_t pmcid; 4368 struct pmc *pm; 4369 struct pmc_op_simple sp; 4370 4371 sx_assert(&pmc_sx, SX_XLOCKED); 4372 4373 if ((error = copyin(arg, &sp, sizeof(sp))) != 0) 4374 break; 4375 4376 pmcid = sp.pm_pmcid; 4377 4378 if ((error = pmc_find_pmc(pmcid, &pm)) != 0) 4379 break; 4380 4381 KASSERT(pmcid == pm->pm_id, 4382 ("[pmc,%d] pmcid %x != id %x", __LINE__, 4383 pm->pm_id, pmcid)); 4384 4385 if (pm->pm_state == PMC_STATE_RUNNING) /* already running */ 4386 break; 4387 else if (pm->pm_state != PMC_STATE_STOPPED && 4388 pm->pm_state != PMC_STATE_ALLOCATED) { 4389 error = EINVAL; 4390 break; 4391 } 4392 4393 error = pmc_start(pm); 4394 } 4395 break; 4396 4397 4398 /* 4399 * Stop a PMC. 4400 */ 4401 4402 case PMC_OP_PMCSTOP: 4403 { 4404 pmc_id_t pmcid; 4405 struct pmc *pm; 4406 struct pmc_op_simple sp; 4407 4408 PMC_DOWNGRADE_SX(); 4409 4410 if ((error = copyin(arg, &sp, sizeof(sp))) != 0) 4411 break; 4412 4413 pmcid = sp.pm_pmcid; 4414 4415 /* 4416 * Mark the PMC as inactive and invoke the MD stop 4417 * routines if needed. 4418 */ 4419 4420 if ((error = pmc_find_pmc(pmcid, &pm)) != 0) 4421 break; 4422 4423 KASSERT(pmcid == pm->pm_id, 4424 ("[pmc,%d] pmc id %x != pmcid %x", __LINE__, 4425 pm->pm_id, pmcid)); 4426 4427 if (pm->pm_state == PMC_STATE_STOPPED) /* already stopped */ 4428 break; 4429 else if (pm->pm_state != PMC_STATE_RUNNING) { 4430 error = EINVAL; 4431 break; 4432 } 4433 4434 error = pmc_stop(pm); 4435 } 4436 break; 4437 4438 4439 /* 4440 * Write a user supplied value to the log file. 4441 */ 4442 4443 case PMC_OP_WRITELOG: 4444 { 4445 struct pmc_op_writelog wl; 4446 struct pmc_owner *po; 4447 4448 PMC_DOWNGRADE_SX(); 4449 4450 if ((error = copyin(arg, &wl, sizeof(wl))) != 0) 4451 break; 4452 4453 if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) { 4454 error = EINVAL; 4455 break; 4456 } 4457 4458 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) { 4459 error = EINVAL; 4460 break; 4461 } 4462 4463 error = pmclog_process_userlog(po, &wl); 4464 } 4465 break; 4466 4467 4468 default: 4469 error = EINVAL; 4470 break; 4471 } 4472 4473 if (is_sx_downgraded) 4474 sx_sunlock(&pmc_sx); 4475 else 4476 sx_xunlock(&pmc_sx); 4477 done_syscall: 4478 if (error) 4479 counter_u64_add(pmc_stats.pm_syscall_errors, 1); 4480 4481 return (error); 4482 } 4483 4484 /* 4485 * Helper functions 4486 */ 4487 4488 4489 /* 4490 * Mark the thread as needing callchain capture and post an AST. The 4491 * actual callchain capture will be done in a context where it is safe 4492 * to take page faults. 4493 */ 4494 4495 static void 4496 pmc_post_callchain_callback(void) 4497 { 4498 struct thread *td; 4499 4500 td = curthread; 4501 4502 /* 4503 * If there is multiple PMCs for the same interrupt ignore new post 4504 */ 4505 if (td->td_pflags & TDP_CALLCHAIN) 4506 return; 4507 4508 /* 4509 * Mark this thread as needing callchain capture. 4510 * `td->td_pflags' will be safe to touch because this thread 4511 * was in user space when it was interrupted. 4512 */ 4513 td->td_pflags |= TDP_CALLCHAIN; 4514 4515 /* 4516 * Don't let this thread migrate between CPUs until callchain 4517 * capture completes. 4518 */ 4519 sched_pin(); 4520 4521 return; 4522 } 4523 4524 /* 4525 * Interrupt processing. 4526 * 4527 * Find a free slot in the per-cpu array of samples and capture the 4528 * current callchain there. If a sample was successfully added, a bit 4529 * is set in mask 'pmc_cpumask' denoting that the DO_SAMPLES hook 4530 * needs to be invoked from the clock handler. 4531 * 4532 * This function is meant to be called from an NMI handler. It cannot 4533 * use any of the locking primitives supplied by the OS. 4534 */ 4535 4536 int 4537 pmc_process_interrupt(int cpu, int ring, struct pmc *pm, struct trapframe *tf, 4538 int inuserspace) 4539 { 4540 int error, callchaindepth; 4541 struct thread *td; 4542 struct pmc_sample *ps; 4543 struct pmc_samplebuffer *psb; 4544 4545 error = 0; 4546 4547 /* 4548 * Allocate space for a sample buffer. 4549 */ 4550 psb = pmc_pcpu[cpu]->pc_sb[ring]; 4551 4552 ps = psb->ps_write; 4553 if (ps->ps_nsamples) { /* in use, reader hasn't caught up */ 4554 pm->pm_pcpu_state[cpu].pps_stalled = 1; 4555 counter_u64_add(pmc_stats.pm_intr_bufferfull, 1); 4556 PMCDBG6(SAM,INT,1,"(spc) cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d", 4557 cpu, pm, (void *) tf, inuserspace, 4558 (int) (psb->ps_write - psb->ps_samples), 4559 (int) (psb->ps_read - psb->ps_samples)); 4560 callchaindepth = 1; 4561 error = ENOMEM; 4562 goto done; 4563 } 4564 4565 4566 /* Fill in entry. */ 4567 PMCDBG6(SAM,INT,1,"cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d", cpu, pm, 4568 (void *) tf, inuserspace, 4569 (int) (psb->ps_write - psb->ps_samples), 4570 (int) (psb->ps_read - psb->ps_samples)); 4571 4572 KASSERT(counter_u64_fetch(pm->pm_runcount) >= 0, 4573 ("[pmc,%d] pm=%p runcount %ld", __LINE__, (void *) pm, 4574 (unsigned long)counter_u64_fetch(pm->pm_runcount))); 4575 4576 counter_u64_add(pm->pm_runcount, 1); /* hold onto PMC */ 4577 4578 ps->ps_pmc = pm; 4579 if ((td = curthread) && td->td_proc) 4580 ps->ps_pid = td->td_proc->p_pid; 4581 else 4582 ps->ps_pid = -1; 4583 ps->ps_cpu = cpu; 4584 ps->ps_td = td; 4585 ps->ps_flags = inuserspace ? PMC_CC_F_USERSPACE : 0; 4586 4587 callchaindepth = (pm->pm_flags & PMC_F_CALLCHAIN) ? 4588 pmc_callchaindepth : 1; 4589 4590 if (callchaindepth == 1) 4591 ps->ps_pc[0] = PMC_TRAPFRAME_TO_PC(tf); 4592 else { 4593 /* 4594 * Kernel stack traversals can be done immediately, 4595 * while we defer to an AST for user space traversals. 4596 */ 4597 if (!inuserspace) { 4598 callchaindepth = 4599 pmc_save_kernel_callchain(ps->ps_pc, 4600 callchaindepth, tf); 4601 } else { 4602 pmc_post_callchain_callback(); 4603 callchaindepth = PMC_SAMPLE_INUSE; 4604 } 4605 } 4606 4607 ps->ps_nsamples = callchaindepth; /* mark entry as in use */ 4608 4609 /* increment write pointer, modulo ring buffer size */ 4610 ps++; 4611 if (ps == psb->ps_fence) 4612 psb->ps_write = psb->ps_samples; 4613 else 4614 psb->ps_write = ps; 4615 4616 done: 4617 /* mark CPU as needing processing */ 4618 if (callchaindepth != PMC_SAMPLE_INUSE) 4619 DPCPU_SET(pmc_sampled, 1); 4620 4621 return (error); 4622 } 4623 4624 /* 4625 * Capture a user call chain. This function will be called from ast() 4626 * before control returns to userland and before the process gets 4627 * rescheduled. 4628 */ 4629 4630 static void 4631 pmc_capture_user_callchain(int cpu, int ring, struct trapframe *tf) 4632 { 4633 struct pmc *pm; 4634 struct thread *td; 4635 struct pmc_sample *ps, *ps_end; 4636 struct pmc_samplebuffer *psb; 4637 #ifdef INVARIANTS 4638 int ncallchains; 4639 int nfree; 4640 #endif 4641 4642 psb = pmc_pcpu[cpu]->pc_sb[ring]; 4643 td = curthread; 4644 4645 KASSERT(td->td_pflags & TDP_CALLCHAIN, 4646 ("[pmc,%d] Retrieving callchain for thread that doesn't want it", 4647 __LINE__)); 4648 4649 #ifdef INVARIANTS 4650 ncallchains = 0; 4651 nfree = 0; 4652 #endif 4653 4654 /* 4655 * Iterate through all deferred callchain requests. 4656 * Walk from the current read pointer to the current 4657 * write pointer. 4658 */ 4659 4660 ps = psb->ps_read; 4661 ps_end = psb->ps_write; 4662 do { 4663 #ifdef INVARIANTS 4664 if ((ps->ps_pmc == NULL) || 4665 (ps->ps_pmc->pm_state != PMC_STATE_RUNNING)) 4666 nfree++; 4667 #endif 4668 if (ps->ps_nsamples != PMC_SAMPLE_INUSE) 4669 goto next; 4670 if (ps->ps_td != td) 4671 goto next; 4672 4673 KASSERT(ps->ps_cpu == cpu, 4674 ("[pmc,%d] cpu mismatch ps_cpu=%d pcpu=%d", __LINE__, 4675 ps->ps_cpu, PCPU_GET(cpuid))); 4676 4677 pm = ps->ps_pmc; 4678 4679 KASSERT(pm->pm_flags & PMC_F_CALLCHAIN, 4680 ("[pmc,%d] Retrieving callchain for PMC that doesn't " 4681 "want it", __LINE__)); 4682 4683 KASSERT(counter_u64_fetch(pm->pm_runcount) > 0, 4684 ("[pmc,%d] runcount %ld", __LINE__, (unsigned long)counter_u64_fetch(pm->pm_runcount))); 4685 4686 /* 4687 * Retrieve the callchain and mark the sample buffer 4688 * as 'processable' by the timer tick sweep code. 4689 */ 4690 ps->ps_nsamples = pmc_save_user_callchain(ps->ps_pc, 4691 pmc_callchaindepth, tf); 4692 4693 #ifdef INVARIANTS 4694 ncallchains++; 4695 #endif 4696 4697 next: 4698 /* increment the pointer, modulo sample ring size */ 4699 if (++ps == psb->ps_fence) 4700 ps = psb->ps_samples; 4701 } while (ps != ps_end); 4702 4703 #ifdef INVARIANTS 4704 KASSERT(ncallchains > 0 || nfree > 0, 4705 ("[pmc,%d] cpu %d didn't find a sample to collect", __LINE__, 4706 cpu)); 4707 #endif 4708 4709 KASSERT(td->td_pinned == 1, 4710 ("[pmc,%d] invalid td_pinned value", __LINE__)); 4711 sched_unpin(); /* Can migrate safely now. */ 4712 4713 /* mark CPU as needing processing */ 4714 DPCPU_SET(pmc_sampled, 1); 4715 } 4716 4717 /* 4718 * Process saved PC samples. 4719 */ 4720 4721 static void 4722 pmc_process_samples(int cpu, int ring) 4723 { 4724 struct pmc *pm; 4725 int adjri, n; 4726 struct thread *td; 4727 struct pmc_owner *po; 4728 struct pmc_sample *ps; 4729 struct pmc_classdep *pcd; 4730 struct pmc_samplebuffer *psb; 4731 4732 KASSERT(PCPU_GET(cpuid) == cpu, 4733 ("[pmc,%d] not on the correct CPU pcpu=%d cpu=%d", __LINE__, 4734 PCPU_GET(cpuid), cpu)); 4735 4736 psb = pmc_pcpu[cpu]->pc_sb[ring]; 4737 4738 for (n = 0; n < pmc_nsamples; n++) { /* bound on #iterations */ 4739 4740 ps = psb->ps_read; 4741 if (ps->ps_nsamples == PMC_SAMPLE_FREE) 4742 break; 4743 4744 pm = ps->ps_pmc; 4745 4746 KASSERT(counter_u64_fetch(pm->pm_runcount) > 0, 4747 ("[pmc,%d] pm=%p runcount %ld", __LINE__, (void *) pm, 4748 (unsigned long)counter_u64_fetch(pm->pm_runcount))); 4749 4750 po = pm->pm_owner; 4751 4752 KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)), 4753 ("[pmc,%d] pmc=%p non-sampling mode=%d", __LINE__, 4754 pm, PMC_TO_MODE(pm))); 4755 4756 /* Ignore PMCs that have been switched off */ 4757 if (pm->pm_state != PMC_STATE_RUNNING) 4758 goto entrydone; 4759 4760 /* If there is a pending AST wait for completion */ 4761 if (ps->ps_nsamples == PMC_SAMPLE_INUSE) { 4762 /* Need a rescan at a later time. */ 4763 DPCPU_SET(pmc_sampled, 1); 4764 break; 4765 } 4766 4767 PMCDBG6(SAM,OPS,1,"cpu=%d pm=%p n=%d fl=%x wr=%d rd=%d", cpu, 4768 pm, ps->ps_nsamples, ps->ps_flags, 4769 (int) (psb->ps_write - psb->ps_samples), 4770 (int) (psb->ps_read - psb->ps_samples)); 4771 4772 /* 4773 * If this is a process-mode PMC that is attached to 4774 * its owner, and if the PC is in user mode, update 4775 * profiling statistics like timer-based profiling 4776 * would have done. 4777 */ 4778 if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) { 4779 if (ps->ps_flags & PMC_CC_F_USERSPACE) { 4780 td = FIRST_THREAD_IN_PROC(po->po_owner); 4781 addupc_intr(td, ps->ps_pc[0], 1); 4782 } 4783 goto entrydone; 4784 } 4785 4786 /* 4787 * Otherwise, this is either a sampling mode PMC that 4788 * is attached to a different process than its owner, 4789 * or a system-wide sampling PMC. Dispatch a log 4790 * entry to the PMC's owner process. 4791 */ 4792 pmclog_process_callchain(pm, ps); 4793 4794 entrydone: 4795 ps->ps_nsamples = 0; /* mark entry as free */ 4796 counter_u64_add(pm->pm_runcount, -1); 4797 4798 /* increment read pointer, modulo sample size */ 4799 if (++ps == psb->ps_fence) 4800 psb->ps_read = psb->ps_samples; 4801 else 4802 psb->ps_read = ps; 4803 } 4804 4805 counter_u64_add(pmc_stats.pm_log_sweeps, 1); 4806 4807 /* Do not re-enable stalled PMCs if we failed to process any samples */ 4808 if (n == 0) 4809 return; 4810 4811 /* 4812 * Restart any stalled sampling PMCs on this CPU. 4813 * 4814 * If the NMI handler sets the pm_stalled field of a PMC after 4815 * the check below, we'll end up processing the stalled PMC at 4816 * the next hardclock tick. 4817 */ 4818 for (n = 0; n < md->pmd_npmc; n++) { 4819 pcd = pmc_ri_to_classdep(md, n, &adjri); 4820 KASSERT(pcd != NULL, 4821 ("[pmc,%d] null pcd ri=%d", __LINE__, n)); 4822 (void) (*pcd->pcd_get_config)(cpu,adjri,&pm); 4823 4824 if (pm == NULL || /* !cfg'ed */ 4825 pm->pm_state != PMC_STATE_RUNNING || /* !active */ 4826 !PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) || /* !sampling */ 4827 !pm->pm_pcpu_state[cpu].pps_cpustate || /* !desired */ 4828 !pm->pm_pcpu_state[cpu].pps_stalled) /* !stalled */ 4829 continue; 4830 4831 pm->pm_pcpu_state[cpu].pps_stalled = 0; 4832 (*pcd->pcd_start_pmc)(cpu, adjri); 4833 } 4834 } 4835 4836 /* 4837 * Event handlers. 4838 */ 4839 4840 /* 4841 * Handle a process exit. 4842 * 4843 * Remove this process from all hash tables. If this process 4844 * owned any PMCs, turn off those PMCs and deallocate them, 4845 * removing any associations with target processes. 4846 * 4847 * This function will be called by the last 'thread' of a 4848 * process. 4849 * 4850 * XXX This eventhandler gets called early in the exit process. 4851 * Consider using a 'hook' invocation from thread_exit() or equivalent 4852 * spot. Another negative is that kse_exit doesn't seem to call 4853 * exit1() [??]. 4854 * 4855 */ 4856 4857 static void 4858 pmc_process_exit(void *arg __unused, struct proc *p) 4859 { 4860 struct pmc *pm; 4861 int adjri, cpu; 4862 unsigned int ri; 4863 int is_using_hwpmcs; 4864 struct pmc_owner *po; 4865 struct pmc_process *pp; 4866 struct pmc_classdep *pcd; 4867 pmc_value_t newvalue, tmp; 4868 4869 PROC_LOCK(p); 4870 is_using_hwpmcs = p->p_flag & P_HWPMC; 4871 PROC_UNLOCK(p); 4872 4873 /* 4874 * Log a sysexit event to all SS PMC owners. 4875 */ 4876 epoch_enter_preempt(global_epoch_preempt); 4877 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) 4878 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 4879 pmclog_process_sysexit(po, p->p_pid); 4880 epoch_exit_preempt(global_epoch_preempt); 4881 4882 if (!is_using_hwpmcs) 4883 return; 4884 4885 PMC_GET_SX_XLOCK(); 4886 PMCDBG3(PRC,EXT,1,"process-exit proc=%p (%d, %s)", p, p->p_pid, 4887 p->p_comm); 4888 4889 /* 4890 * Since this code is invoked by the last thread in an exiting 4891 * process, we would have context switched IN at some prior 4892 * point. However, with PREEMPTION, kernel mode context 4893 * switches may happen any time, so we want to disable a 4894 * context switch OUT till we get any PMCs targeting this 4895 * process off the hardware. 4896 * 4897 * We also need to atomically remove this process' 4898 * entry from our target process hash table, using 4899 * PMC_FLAG_REMOVE. 4900 */ 4901 PMCDBG3(PRC,EXT,1, "process-exit proc=%p (%d, %s)", p, p->p_pid, 4902 p->p_comm); 4903 4904 critical_enter(); /* no preemption */ 4905 4906 cpu = curthread->td_oncpu; 4907 4908 if ((pp = pmc_find_process_descriptor(p, 4909 PMC_FLAG_REMOVE)) != NULL) { 4910 4911 PMCDBG2(PRC,EXT,2, 4912 "process-exit proc=%p pmc-process=%p", p, pp); 4913 4914 /* 4915 * The exiting process could the target of 4916 * some PMCs which will be running on 4917 * currently executing CPU. 4918 * 4919 * We need to turn these PMCs off like we 4920 * would do at context switch OUT time. 4921 */ 4922 for (ri = 0; ri < md->pmd_npmc; ri++) { 4923 4924 /* 4925 * Pick up the pmc pointer from hardware 4926 * state similar to the CSW_OUT code. 4927 */ 4928 pm = NULL; 4929 4930 pcd = pmc_ri_to_classdep(md, ri, &adjri); 4931 4932 (void) (*pcd->pcd_get_config)(cpu, adjri, &pm); 4933 4934 PMCDBG2(PRC,EXT,2, "ri=%d pm=%p", ri, pm); 4935 4936 if (pm == NULL || 4937 !PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) 4938 continue; 4939 4940 PMCDBG4(PRC,EXT,2, "ppmcs[%d]=%p pm=%p " 4941 "state=%d", ri, pp->pp_pmcs[ri].pp_pmc, 4942 pm, pm->pm_state); 4943 4944 KASSERT(PMC_TO_ROWINDEX(pm) == ri, 4945 ("[pmc,%d] ri mismatch pmc(%d) ri(%d)", 4946 __LINE__, PMC_TO_ROWINDEX(pm), ri)); 4947 4948 KASSERT(pm == pp->pp_pmcs[ri].pp_pmc, 4949 ("[pmc,%d] pm %p != pp_pmcs[%d] %p", 4950 __LINE__, pm, ri, pp->pp_pmcs[ri].pp_pmc)); 4951 4952 KASSERT(counter_u64_fetch(pm->pm_runcount) > 0, 4953 ("[pmc,%d] bad runcount ri %d rc %ld", 4954 __LINE__, ri, (unsigned long)counter_u64_fetch(pm->pm_runcount))); 4955 4956 /* 4957 * Change desired state, and then stop if not 4958 * stalled. This two-step dance should avoid 4959 * race conditions where an interrupt re-enables 4960 * the PMC after this code has already checked 4961 * the pm_stalled flag. 4962 */ 4963 if (pm->pm_pcpu_state[cpu].pps_cpustate) { 4964 pm->pm_pcpu_state[cpu].pps_cpustate = 0; 4965 if (!pm->pm_pcpu_state[cpu].pps_stalled) { 4966 (void) pcd->pcd_stop_pmc(cpu, adjri); 4967 4968 if (PMC_TO_MODE(pm) == PMC_MODE_TC) { 4969 pcd->pcd_read_pmc(cpu, adjri, 4970 &newvalue); 4971 tmp = newvalue - 4972 PMC_PCPU_SAVED(cpu,ri); 4973 4974 mtx_pool_lock_spin(pmc_mtxpool, 4975 pm); 4976 pm->pm_gv.pm_savedvalue += tmp; 4977 pp->pp_pmcs[ri].pp_pmcval += 4978 tmp; 4979 mtx_pool_unlock_spin( 4980 pmc_mtxpool, pm); 4981 } 4982 } 4983 } 4984 4985 counter_u64_add(pm->pm_runcount, -1); 4986 4987 KASSERT((int) counter_u64_fetch(pm->pm_runcount) >= 0, 4988 ("[pmc,%d] runcount is %d", __LINE__, ri)); 4989 4990 (void) pcd->pcd_config_pmc(cpu, adjri, NULL); 4991 } 4992 4993 /* 4994 * Inform the MD layer of this pseudo "context switch 4995 * out" 4996 */ 4997 (void) md->pmd_switch_out(pmc_pcpu[cpu], pp); 4998 4999 critical_exit(); /* ok to be pre-empted now */ 5000 5001 /* 5002 * Unlink this process from the PMCs that are 5003 * targeting it. This will send a signal to 5004 * all PMC owner's whose PMCs are orphaned. 5005 * 5006 * Log PMC value at exit time if requested. 5007 */ 5008 for (ri = 0; ri < md->pmd_npmc; ri++) 5009 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) { 5010 if (pm->pm_flags & PMC_F_NEEDS_LOGFILE && 5011 PMC_IS_COUNTING_MODE(PMC_TO_MODE(pm))) 5012 pmclog_process_procexit(pm, pp); 5013 pmc_unlink_target_process(pm, pp); 5014 } 5015 free(pp, M_PMC); 5016 5017 } else 5018 critical_exit(); /* pp == NULL */ 5019 5020 5021 /* 5022 * If the process owned PMCs, free them up and free up 5023 * memory. 5024 */ 5025 if ((po = pmc_find_owner_descriptor(p)) != NULL) { 5026 pmc_remove_owner(po); 5027 pmc_destroy_owner_descriptor(po); 5028 } 5029 5030 sx_xunlock(&pmc_sx); 5031 } 5032 5033 /* 5034 * Handle a process fork. 5035 * 5036 * If the parent process 'p1' is under HWPMC monitoring, then copy 5037 * over any attached PMCs that have 'do_descendants' semantics. 5038 */ 5039 5040 static void 5041 pmc_process_fork(void *arg __unused, struct proc *p1, struct proc *newproc, 5042 int flags) 5043 { 5044 int is_using_hwpmcs; 5045 unsigned int ri; 5046 uint32_t do_descendants; 5047 struct pmc *pm; 5048 struct pmc_owner *po; 5049 struct pmc_process *ppnew, *ppold; 5050 5051 (void) flags; /* unused parameter */ 5052 5053 PROC_LOCK(p1); 5054 is_using_hwpmcs = p1->p_flag & P_HWPMC; 5055 PROC_UNLOCK(p1); 5056 5057 /* 5058 * If there are system-wide sampling PMCs active, we need to 5059 * log all fork events to their owner's logs. 5060 */ 5061 epoch_enter_preempt(global_epoch_preempt); 5062 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) 5063 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 5064 pmclog_process_procfork(po, p1->p_pid, newproc->p_pid); 5065 epoch_exit_preempt(global_epoch_preempt); 5066 5067 if (!is_using_hwpmcs) 5068 return; 5069 5070 PMC_GET_SX_XLOCK(); 5071 PMCDBG4(PMC,FRK,1, "process-fork proc=%p (%d, %s) -> %p", p1, 5072 p1->p_pid, p1->p_comm, newproc); 5073 5074 /* 5075 * If the parent process (curthread->td_proc) is a 5076 * target of any PMCs, look for PMCs that are to be 5077 * inherited, and link these into the new process 5078 * descriptor. 5079 */ 5080 if ((ppold = pmc_find_process_descriptor(curthread->td_proc, 5081 PMC_FLAG_NONE)) == NULL) 5082 goto done; /* nothing to do */ 5083 5084 do_descendants = 0; 5085 for (ri = 0; ri < md->pmd_npmc; ri++) 5086 if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL) 5087 do_descendants |= pm->pm_flags & PMC_F_DESCENDANTS; 5088 if (do_descendants == 0) /* nothing to do */ 5089 goto done; 5090 5091 /* 5092 * Now mark the new process as being tracked by this driver. 5093 */ 5094 PROC_LOCK(newproc); 5095 newproc->p_flag |= P_HWPMC; 5096 PROC_UNLOCK(newproc); 5097 5098 /* allocate a descriptor for the new process */ 5099 if ((ppnew = pmc_find_process_descriptor(newproc, 5100 PMC_FLAG_ALLOCATE)) == NULL) 5101 goto done; 5102 5103 /* 5104 * Run through all PMCs that were targeting the old process 5105 * and which specified F_DESCENDANTS and attach them to the 5106 * new process. 5107 * 5108 * Log the fork event to all owners of PMCs attached to this 5109 * process, if not already logged. 5110 */ 5111 for (ri = 0; ri < md->pmd_npmc; ri++) 5112 if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL && 5113 (pm->pm_flags & PMC_F_DESCENDANTS)) { 5114 pmc_link_target_process(pm, ppnew); 5115 po = pm->pm_owner; 5116 if (po->po_sscount == 0 && 5117 po->po_flags & PMC_PO_OWNS_LOGFILE) 5118 pmclog_process_procfork(po, p1->p_pid, 5119 newproc->p_pid); 5120 } 5121 5122 done: 5123 sx_xunlock(&pmc_sx); 5124 } 5125 5126 static void 5127 pmc_kld_load(void *arg __unused, linker_file_t lf) 5128 { 5129 struct pmc_owner *po; 5130 5131 /* 5132 * Notify owners of system sampling PMCs about KLD operations. 5133 */ 5134 epoch_enter_preempt(global_epoch_preempt); 5135 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) 5136 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 5137 pmclog_process_map_in(po, (pid_t) -1, 5138 (uintfptr_t) lf->address, lf->filename); 5139 epoch_exit_preempt(global_epoch_preempt); 5140 5141 /* 5142 * TODO: Notify owners of (all) process-sampling PMCs too. 5143 */ 5144 } 5145 5146 static void 5147 pmc_kld_unload(void *arg __unused, const char *filename __unused, 5148 caddr_t address, size_t size) 5149 { 5150 struct pmc_owner *po; 5151 5152 epoch_enter_preempt(global_epoch_preempt); 5153 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) 5154 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 5155 pmclog_process_map_out(po, (pid_t) -1, 5156 (uintfptr_t) address, (uintfptr_t) address + size); 5157 epoch_exit_preempt(global_epoch_preempt); 5158 5159 /* 5160 * TODO: Notify owners of process-sampling PMCs. 5161 */ 5162 } 5163 5164 /* 5165 * initialization 5166 */ 5167 static const char * 5168 pmc_name_of_pmcclass(enum pmc_class class) 5169 { 5170 5171 switch (class) { 5172 #undef __PMC_CLASS 5173 #define __PMC_CLASS(S,V,D) \ 5174 case PMC_CLASS_##S: \ 5175 return #S; 5176 __PMC_CLASSES(); 5177 default: 5178 return ("<unknown>"); 5179 } 5180 } 5181 5182 /* 5183 * Base class initializer: allocate structure and set default classes. 5184 */ 5185 struct pmc_mdep * 5186 pmc_mdep_alloc(int nclasses) 5187 { 5188 struct pmc_mdep *md; 5189 int n; 5190 5191 /* SOFT + md classes */ 5192 n = 1 + nclasses; 5193 md = malloc(sizeof(struct pmc_mdep) + n * 5194 sizeof(struct pmc_classdep), M_PMC, M_WAITOK|M_ZERO); 5195 md->pmd_nclass = n; 5196 5197 /* Add base class. */ 5198 pmc_soft_initialize(md); 5199 return md; 5200 } 5201 5202 void 5203 pmc_mdep_free(struct pmc_mdep *md) 5204 { 5205 pmc_soft_finalize(md); 5206 free(md, M_PMC); 5207 } 5208 5209 static int 5210 generic_switch_in(struct pmc_cpu *pc, struct pmc_process *pp) 5211 { 5212 (void) pc; (void) pp; 5213 5214 return (0); 5215 } 5216 5217 static int 5218 generic_switch_out(struct pmc_cpu *pc, struct pmc_process *pp) 5219 { 5220 (void) pc; (void) pp; 5221 5222 return (0); 5223 } 5224 5225 static struct pmc_mdep * 5226 pmc_generic_cpu_initialize(void) 5227 { 5228 struct pmc_mdep *md; 5229 5230 md = pmc_mdep_alloc(0); 5231 5232 md->pmd_cputype = PMC_CPU_GENERIC; 5233 5234 md->pmd_pcpu_init = NULL; 5235 md->pmd_pcpu_fini = NULL; 5236 md->pmd_switch_in = generic_switch_in; 5237 md->pmd_switch_out = generic_switch_out; 5238 5239 return (md); 5240 } 5241 5242 static void 5243 pmc_generic_cpu_finalize(struct pmc_mdep *md) 5244 { 5245 (void) md; 5246 } 5247 5248 5249 static int 5250 pmc_initialize(void) 5251 { 5252 int c, cpu, error, n, ri; 5253 unsigned int maxcpu, domain; 5254 struct pcpu *pc; 5255 struct pmc_binding pb; 5256 struct pmc_sample *ps; 5257 struct pmc_classdep *pcd; 5258 struct pmc_samplebuffer *sb; 5259 5260 md = NULL; 5261 error = 0; 5262 5263 pmc_stats.pm_intr_ignored = counter_u64_alloc(M_WAITOK); 5264 pmc_stats.pm_intr_processed = counter_u64_alloc(M_WAITOK); 5265 pmc_stats.pm_intr_bufferfull = counter_u64_alloc(M_WAITOK); 5266 pmc_stats.pm_syscalls = counter_u64_alloc(M_WAITOK); 5267 pmc_stats.pm_syscall_errors = counter_u64_alloc(M_WAITOK); 5268 pmc_stats.pm_buffer_requests = counter_u64_alloc(M_WAITOK); 5269 pmc_stats.pm_buffer_requests_failed = counter_u64_alloc(M_WAITOK); 5270 pmc_stats.pm_log_sweeps = counter_u64_alloc(M_WAITOK); 5271 5272 #ifdef HWPMC_DEBUG 5273 /* parse debug flags first */ 5274 if (TUNABLE_STR_FETCH(PMC_SYSCTL_NAME_PREFIX "debugflags", 5275 pmc_debugstr, sizeof(pmc_debugstr))) 5276 pmc_debugflags_parse(pmc_debugstr, 5277 pmc_debugstr+strlen(pmc_debugstr)); 5278 #endif 5279 5280 PMCDBG1(MOD,INI,0, "PMC Initialize (version %x)", PMC_VERSION); 5281 5282 /* check kernel version */ 5283 if (pmc_kernel_version != PMC_VERSION) { 5284 if (pmc_kernel_version == 0) 5285 printf("hwpmc: this kernel has not been compiled with " 5286 "'options HWPMC_HOOKS'.\n"); 5287 else 5288 printf("hwpmc: kernel version (0x%x) does not match " 5289 "module version (0x%x).\n", pmc_kernel_version, 5290 PMC_VERSION); 5291 return EPROGMISMATCH; 5292 } 5293 5294 /* 5295 * check sysctl parameters 5296 */ 5297 5298 if (pmc_hashsize <= 0) { 5299 (void) printf("hwpmc: tunable \"hashsize\"=%d must be " 5300 "greater than zero.\n", pmc_hashsize); 5301 pmc_hashsize = PMC_HASH_SIZE; 5302 } 5303 5304 if (pmc_nsamples <= 0 || pmc_nsamples > 65535) { 5305 (void) printf("hwpmc: tunable \"nsamples\"=%d out of " 5306 "range.\n", pmc_nsamples); 5307 pmc_nsamples = PMC_NSAMPLES; 5308 } 5309 5310 if (pmc_callchaindepth <= 0 || 5311 pmc_callchaindepth > PMC_CALLCHAIN_DEPTH_MAX) { 5312 (void) printf("hwpmc: tunable \"callchaindepth\"=%d out of " 5313 "range - using %d.\n", pmc_callchaindepth, 5314 PMC_CALLCHAIN_DEPTH_MAX); 5315 pmc_callchaindepth = PMC_CALLCHAIN_DEPTH_MAX; 5316 } 5317 5318 md = pmc_md_initialize(); 5319 if (md == NULL) { 5320 /* Default to generic CPU. */ 5321 md = pmc_generic_cpu_initialize(); 5322 if (md == NULL) 5323 return (ENOSYS); 5324 } 5325 5326 KASSERT(md->pmd_nclass >= 1 && md->pmd_npmc >= 1, 5327 ("[pmc,%d] no classes or pmcs", __LINE__)); 5328 5329 /* Compute the map from row-indices to classdep pointers. */ 5330 pmc_rowindex_to_classdep = malloc(sizeof(struct pmc_classdep *) * 5331 md->pmd_npmc, M_PMC, M_WAITOK|M_ZERO); 5332 5333 for (n = 0; n < md->pmd_npmc; n++) 5334 pmc_rowindex_to_classdep[n] = NULL; 5335 for (ri = c = 0; c < md->pmd_nclass; c++) { 5336 pcd = &md->pmd_classdep[c]; 5337 for (n = 0; n < pcd->pcd_num; n++, ri++) 5338 pmc_rowindex_to_classdep[ri] = pcd; 5339 } 5340 5341 KASSERT(ri == md->pmd_npmc, 5342 ("[pmc,%d] npmc miscomputed: ri=%d, md->npmc=%d", __LINE__, 5343 ri, md->pmd_npmc)); 5344 5345 maxcpu = pmc_cpu_max(); 5346 5347 /* allocate space for the per-cpu array */ 5348 pmc_pcpu = malloc(maxcpu * sizeof(struct pmc_cpu *), M_PMC, 5349 M_WAITOK|M_ZERO); 5350 5351 /* per-cpu 'saved values' for managing process-mode PMCs */ 5352 pmc_pcpu_saved = malloc(sizeof(pmc_value_t) * maxcpu * md->pmd_npmc, 5353 M_PMC, M_WAITOK); 5354 5355 /* Perform CPU-dependent initialization. */ 5356 pmc_save_cpu_binding(&pb); 5357 error = 0; 5358 for (cpu = 0; error == 0 && cpu < maxcpu; cpu++) { 5359 if (!pmc_cpu_is_active(cpu)) 5360 continue; 5361 pmc_select_cpu(cpu); 5362 pmc_pcpu[cpu] = malloc(sizeof(struct pmc_cpu) + 5363 md->pmd_npmc * sizeof(struct pmc_hw *), M_PMC, 5364 M_WAITOK|M_ZERO); 5365 if (md->pmd_pcpu_init) 5366 error = md->pmd_pcpu_init(md, cpu); 5367 for (n = 0; error == 0 && n < md->pmd_nclass; n++) 5368 error = md->pmd_classdep[n].pcd_pcpu_init(md, cpu); 5369 } 5370 pmc_restore_cpu_binding(&pb); 5371 5372 if (error) 5373 return (error); 5374 5375 /* allocate space for the sample array */ 5376 for (cpu = 0; cpu < maxcpu; cpu++) { 5377 if (!pmc_cpu_is_active(cpu)) 5378 continue; 5379 pc = pcpu_find(cpu); 5380 domain = pc->pc_domain; 5381 sb = malloc_domain(sizeof(struct pmc_samplebuffer) + 5382 pmc_nsamples * sizeof(struct pmc_sample), M_PMC, domain, 5383 M_WAITOK|M_ZERO); 5384 sb->ps_read = sb->ps_write = sb->ps_samples; 5385 sb->ps_fence = sb->ps_samples + pmc_nsamples; 5386 5387 KASSERT(pmc_pcpu[cpu] != NULL, 5388 ("[pmc,%d] cpu=%d Null per-cpu data", __LINE__, cpu)); 5389 5390 sb->ps_callchains = malloc_domain(pmc_callchaindepth * pmc_nsamples * 5391 sizeof(uintptr_t), M_PMC, domain, M_WAITOK|M_ZERO); 5392 5393 for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++) 5394 ps->ps_pc = sb->ps_callchains + 5395 (n * pmc_callchaindepth); 5396 5397 pmc_pcpu[cpu]->pc_sb[PMC_HR] = sb; 5398 5399 sb = malloc_domain(sizeof(struct pmc_samplebuffer) + 5400 pmc_nsamples * sizeof(struct pmc_sample), M_PMC, domain, 5401 M_WAITOK|M_ZERO); 5402 sb->ps_read = sb->ps_write = sb->ps_samples; 5403 sb->ps_fence = sb->ps_samples + pmc_nsamples; 5404 5405 KASSERT(pmc_pcpu[cpu] != NULL, 5406 ("[pmc,%d] cpu=%d Null per-cpu data", __LINE__, cpu)); 5407 5408 sb->ps_callchains = malloc_domain(pmc_callchaindepth * pmc_nsamples * 5409 sizeof(uintptr_t), M_PMC, domain, M_WAITOK|M_ZERO); 5410 5411 for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++) 5412 ps->ps_pc = sb->ps_callchains + 5413 (n * pmc_callchaindepth); 5414 5415 pmc_pcpu[cpu]->pc_sb[PMC_SR] = sb; 5416 } 5417 5418 /* allocate space for the row disposition array */ 5419 pmc_pmcdisp = malloc(sizeof(enum pmc_mode) * md->pmd_npmc, 5420 M_PMC, M_WAITOK|M_ZERO); 5421 5422 /* mark all PMCs as available */ 5423 for (n = 0; n < (int) md->pmd_npmc; n++) 5424 PMC_MARK_ROW_FREE(n); 5425 5426 /* allocate thread hash tables */ 5427 pmc_ownerhash = hashinit(pmc_hashsize, M_PMC, 5428 &pmc_ownerhashmask); 5429 5430 pmc_processhash = hashinit(pmc_hashsize, M_PMC, 5431 &pmc_processhashmask); 5432 mtx_init(&pmc_processhash_mtx, "pmc-process-hash", "pmc-leaf", 5433 MTX_SPIN); 5434 5435 LIST_INIT(&pmc_ss_owners); 5436 pmc_ss_count = 0; 5437 5438 /* allocate a pool of spin mutexes */ 5439 pmc_mtxpool = mtx_pool_create("pmc-leaf", pmc_mtxpool_size, 5440 MTX_SPIN); 5441 5442 PMCDBG4(MOD,INI,1, "pmc_ownerhash=%p, mask=0x%lx " 5443 "targethash=%p mask=0x%lx", pmc_ownerhash, pmc_ownerhashmask, 5444 pmc_processhash, pmc_processhashmask); 5445 5446 /* Initialize a spin mutex for the thread free list. */ 5447 mtx_init(&pmc_threadfreelist_mtx, "pmc-threadfreelist", "pmc-leaf", 5448 MTX_SPIN); 5449 5450 /* 5451 * Initialize the callout to monitor the thread free list. 5452 * This callout will also handle the initial population of the list. 5453 */ 5454 taskqgroup_config_gtask_init(NULL, &free_gtask, pmc_thread_descriptor_pool_free_task, "thread descriptor pool free task"); 5455 5456 /* register process {exit,fork,exec} handlers */ 5457 pmc_exit_tag = EVENTHANDLER_REGISTER(process_exit, 5458 pmc_process_exit, NULL, EVENTHANDLER_PRI_ANY); 5459 pmc_fork_tag = EVENTHANDLER_REGISTER(process_fork, 5460 pmc_process_fork, NULL, EVENTHANDLER_PRI_ANY); 5461 5462 /* register kld event handlers */ 5463 pmc_kld_load_tag = EVENTHANDLER_REGISTER(kld_load, pmc_kld_load, 5464 NULL, EVENTHANDLER_PRI_ANY); 5465 pmc_kld_unload_tag = EVENTHANDLER_REGISTER(kld_unload, pmc_kld_unload, 5466 NULL, EVENTHANDLER_PRI_ANY); 5467 5468 /* initialize logging */ 5469 pmclog_initialize(); 5470 5471 /* set hook functions */ 5472 pmc_intr = md->pmd_intr; 5473 wmb(); 5474 pmc_hook = pmc_hook_handler; 5475 5476 if (error == 0) { 5477 printf(PMC_MODULE_NAME ":"); 5478 for (n = 0; n < (int) md->pmd_nclass; n++) { 5479 pcd = &md->pmd_classdep[n]; 5480 printf(" %s/%d/%d/0x%b", 5481 pmc_name_of_pmcclass(pcd->pcd_class), 5482 pcd->pcd_num, 5483 pcd->pcd_width, 5484 pcd->pcd_caps, 5485 "\20" 5486 "\1INT\2USR\3SYS\4EDG\5THR" 5487 "\6REA\7WRI\10INV\11QUA\12PRC" 5488 "\13TAG\14CSC"); 5489 } 5490 printf("\n"); 5491 } 5492 5493 return (error); 5494 } 5495 5496 /* prepare to be unloaded */ 5497 static void 5498 pmc_cleanup(void) 5499 { 5500 int c, cpu; 5501 unsigned int maxcpu; 5502 struct pmc_ownerhash *ph; 5503 struct pmc_owner *po, *tmp; 5504 struct pmc_binding pb; 5505 #ifdef HWPMC_DEBUG 5506 struct pmc_processhash *prh; 5507 #endif 5508 5509 PMCDBG0(MOD,INI,0, "cleanup"); 5510 5511 /* switch off sampling */ 5512 CPU_FOREACH(cpu) 5513 DPCPU_ID_SET(cpu, pmc_sampled, 0); 5514 pmc_intr = NULL; 5515 5516 sx_xlock(&pmc_sx); 5517 if (pmc_hook == NULL) { /* being unloaded already */ 5518 sx_xunlock(&pmc_sx); 5519 return; 5520 } 5521 5522 pmc_hook = NULL; /* prevent new threads from entering module */ 5523 5524 /* deregister event handlers */ 5525 EVENTHANDLER_DEREGISTER(process_fork, pmc_fork_tag); 5526 EVENTHANDLER_DEREGISTER(process_exit, pmc_exit_tag); 5527 EVENTHANDLER_DEREGISTER(kld_load, pmc_kld_load_tag); 5528 EVENTHANDLER_DEREGISTER(kld_unload, pmc_kld_unload_tag); 5529 5530 /* send SIGBUS to all owner threads, free up allocations */ 5531 if (pmc_ownerhash) 5532 for (ph = pmc_ownerhash; 5533 ph <= &pmc_ownerhash[pmc_ownerhashmask]; 5534 ph++) { 5535 LIST_FOREACH_SAFE(po, ph, po_next, tmp) { 5536 pmc_remove_owner(po); 5537 5538 /* send SIGBUS to owner processes */ 5539 PMCDBG3(MOD,INI,2, "cleanup signal proc=%p " 5540 "(%d, %s)", po->po_owner, 5541 po->po_owner->p_pid, 5542 po->po_owner->p_comm); 5543 5544 PROC_LOCK(po->po_owner); 5545 kern_psignal(po->po_owner, SIGBUS); 5546 PROC_UNLOCK(po->po_owner); 5547 5548 pmc_destroy_owner_descriptor(po); 5549 } 5550 } 5551 5552 /* reclaim allocated data structures */ 5553 mtx_destroy(&pmc_threadfreelist_mtx); 5554 pmc_thread_descriptor_pool_drain(); 5555 5556 if (pmc_mtxpool) 5557 mtx_pool_destroy(&pmc_mtxpool); 5558 5559 mtx_destroy(&pmc_processhash_mtx); 5560 if (pmc_processhash) { 5561 #ifdef HWPMC_DEBUG 5562 struct pmc_process *pp; 5563 5564 PMCDBG0(MOD,INI,3, "destroy process hash"); 5565 for (prh = pmc_processhash; 5566 prh <= &pmc_processhash[pmc_processhashmask]; 5567 prh++) 5568 LIST_FOREACH(pp, prh, pp_next) 5569 PMCDBG1(MOD,INI,3, "pid=%d", pp->pp_proc->p_pid); 5570 #endif 5571 5572 hashdestroy(pmc_processhash, M_PMC, pmc_processhashmask); 5573 pmc_processhash = NULL; 5574 } 5575 5576 if (pmc_ownerhash) { 5577 PMCDBG0(MOD,INI,3, "destroy owner hash"); 5578 hashdestroy(pmc_ownerhash, M_PMC, pmc_ownerhashmask); 5579 pmc_ownerhash = NULL; 5580 } 5581 5582 KASSERT(LIST_EMPTY(&pmc_ss_owners), 5583 ("[pmc,%d] Global SS owner list not empty", __LINE__)); 5584 KASSERT(pmc_ss_count == 0, 5585 ("[pmc,%d] Global SS count not empty", __LINE__)); 5586 5587 /* do processor and pmc-class dependent cleanup */ 5588 maxcpu = pmc_cpu_max(); 5589 5590 PMCDBG0(MOD,INI,3, "md cleanup"); 5591 if (md) { 5592 pmc_save_cpu_binding(&pb); 5593 for (cpu = 0; cpu < maxcpu; cpu++) { 5594 PMCDBG2(MOD,INI,1,"pmc-cleanup cpu=%d pcs=%p", 5595 cpu, pmc_pcpu[cpu]); 5596 if (!pmc_cpu_is_active(cpu) || pmc_pcpu[cpu] == NULL) 5597 continue; 5598 pmc_select_cpu(cpu); 5599 for (c = 0; c < md->pmd_nclass; c++) 5600 md->pmd_classdep[c].pcd_pcpu_fini(md, cpu); 5601 if (md->pmd_pcpu_fini) 5602 md->pmd_pcpu_fini(md, cpu); 5603 } 5604 5605 if (md->pmd_cputype == PMC_CPU_GENERIC) 5606 pmc_generic_cpu_finalize(md); 5607 else 5608 pmc_md_finalize(md); 5609 5610 pmc_mdep_free(md); 5611 md = NULL; 5612 pmc_restore_cpu_binding(&pb); 5613 } 5614 5615 /* Free per-cpu descriptors. */ 5616 for (cpu = 0; cpu < maxcpu; cpu++) { 5617 if (!pmc_cpu_is_active(cpu)) 5618 continue; 5619 KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_HR] != NULL, 5620 ("[pmc,%d] Null hw cpu sample buffer cpu=%d", __LINE__, 5621 cpu)); 5622 KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_SR] != NULL, 5623 ("[pmc,%d] Null sw cpu sample buffer cpu=%d", __LINE__, 5624 cpu)); 5625 free_domain(pmc_pcpu[cpu]->pc_sb[PMC_HR]->ps_callchains, M_PMC); 5626 free_domain(pmc_pcpu[cpu]->pc_sb[PMC_HR], M_PMC); 5627 free_domain(pmc_pcpu[cpu]->pc_sb[PMC_SR]->ps_callchains, M_PMC); 5628 free_domain(pmc_pcpu[cpu]->pc_sb[PMC_SR], M_PMC); 5629 free_domain(pmc_pcpu[cpu], M_PMC); 5630 } 5631 5632 free(pmc_pcpu, M_PMC); 5633 pmc_pcpu = NULL; 5634 5635 free(pmc_pcpu_saved, M_PMC); 5636 pmc_pcpu_saved = NULL; 5637 5638 if (pmc_pmcdisp) { 5639 free(pmc_pmcdisp, M_PMC); 5640 pmc_pmcdisp = NULL; 5641 } 5642 5643 if (pmc_rowindex_to_classdep) { 5644 free(pmc_rowindex_to_classdep, M_PMC); 5645 pmc_rowindex_to_classdep = NULL; 5646 } 5647 5648 pmclog_shutdown(); 5649 counter_u64_free(pmc_stats.pm_intr_ignored); 5650 counter_u64_free(pmc_stats.pm_intr_processed); 5651 counter_u64_free(pmc_stats.pm_intr_bufferfull); 5652 counter_u64_free(pmc_stats.pm_syscalls); 5653 counter_u64_free(pmc_stats.pm_syscall_errors); 5654 counter_u64_free(pmc_stats.pm_buffer_requests); 5655 counter_u64_free(pmc_stats.pm_buffer_requests_failed); 5656 counter_u64_free(pmc_stats.pm_log_sweeps); 5657 sx_xunlock(&pmc_sx); /* we are done */ 5658 } 5659 5660 /* 5661 * The function called at load/unload. 5662 */ 5663 5664 static int 5665 load (struct module *module __unused, int cmd, void *arg __unused) 5666 { 5667 int error; 5668 5669 error = 0; 5670 5671 switch (cmd) { 5672 case MOD_LOAD : 5673 /* initialize the subsystem */ 5674 error = pmc_initialize(); 5675 if (error != 0) 5676 break; 5677 PMCDBG2(MOD,INI,1, "syscall=%d maxcpu=%d", 5678 pmc_syscall_num, pmc_cpu_max()); 5679 break; 5680 5681 5682 case MOD_UNLOAD : 5683 case MOD_SHUTDOWN: 5684 pmc_cleanup(); 5685 PMCDBG0(MOD,INI,1, "unloaded"); 5686 break; 5687 5688 default : 5689 error = EINVAL; /* XXX should panic(9) */ 5690 break; 5691 } 5692 5693 return error; 5694 } 5695