1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause 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 #include <sys/param.h> 35 #include <sys/systm.h> 36 #include <sys/domainset.h> 37 #include <sys/eventhandler.h> 38 #include <sys/jail.h> 39 #include <sys/kernel.h> 40 #include <sys/kthread.h> 41 #include <sys/limits.h> 42 #include <sys/lock.h> 43 #include <sys/malloc.h> 44 #include <sys/module.h> 45 #include <sys/mount.h> 46 #include <sys/mutex.h> 47 #include <sys/pmc.h> 48 #include <sys/pmckern.h> 49 #include <sys/pmclog.h> 50 #include <sys/priv.h> 51 #include <sys/proc.h> 52 #include <sys/queue.h> 53 #include <sys/resourcevar.h> 54 #include <sys/rwlock.h> 55 #include <sys/sched.h> 56 #include <sys/signalvar.h> 57 #include <sys/smp.h> 58 #include <sys/sx.h> 59 #include <sys/sysctl.h> 60 #include <sys/sysent.h> 61 #include <sys/syslog.h> 62 #include <sys/taskqueue.h> 63 #include <sys/vnode.h> 64 65 #include <sys/linker.h> /* needs to be after <sys/malloc.h> */ 66 67 #include <machine/atomic.h> 68 #include <machine/md_var.h> 69 70 #include <vm/vm.h> 71 #include <vm/vm_extern.h> 72 #include <vm/pmap.h> 73 #include <vm/vm_map.h> 74 #include <vm/vm_object.h> 75 76 #include "hwpmc_soft.h" 77 78 #define PMC_EPOCH_ENTER() \ 79 struct epoch_tracker pmc_et; \ 80 epoch_enter_preempt(global_epoch_preempt, &pmc_et) 81 82 #define PMC_EPOCH_EXIT() \ 83 epoch_exit_preempt(global_epoch_preempt, &pmc_et) 84 85 /* 86 * Types 87 */ 88 89 enum pmc_flags { 90 PMC_FLAG_NONE = 0x00, /* do nothing */ 91 PMC_FLAG_REMOVE = 0x01, /* atomically remove entry from hash */ 92 PMC_FLAG_ALLOCATE = 0x02, /* add entry to hash if not found */ 93 PMC_FLAG_NOWAIT = 0x04, /* do not wait for mallocs */ 94 }; 95 96 /* 97 * The offset in sysent where the syscall is allocated. 98 */ 99 static int pmc_syscall_num = NO_SYSCALL; 100 101 struct pmc_cpu **pmc_pcpu; /* per-cpu state */ 102 pmc_value_t *pmc_pcpu_saved; /* saved PMC values: CSW handling */ 103 104 #define PMC_PCPU_SAVED(C, R) pmc_pcpu_saved[(R) + md->pmd_npmc * (C)] 105 106 struct mtx_pool *pmc_mtxpool; 107 static int *pmc_pmcdisp; /* PMC row dispositions */ 108 109 #define PMC_ROW_DISP_IS_FREE(R) (pmc_pmcdisp[(R)] == 0) 110 #define PMC_ROW_DISP_IS_THREAD(R) (pmc_pmcdisp[(R)] > 0) 111 #define PMC_ROW_DISP_IS_STANDALONE(R) (pmc_pmcdisp[(R)] < 0) 112 113 #define PMC_MARK_ROW_FREE(R) do { \ 114 pmc_pmcdisp[(R)] = 0; \ 115 } while (0) 116 117 #define PMC_MARK_ROW_STANDALONE(R) do { \ 118 KASSERT(pmc_pmcdisp[(R)] <= 0, ("[pmc,%d] row disposition error", \ 119 __LINE__)); \ 120 atomic_add_int(&pmc_pmcdisp[(R)], -1); \ 121 KASSERT(pmc_pmcdisp[(R)] >= (-pmc_cpu_max_active()), \ 122 ("[pmc,%d] row disposition error", __LINE__)); \ 123 } while (0) 124 125 #define PMC_UNMARK_ROW_STANDALONE(R) do { \ 126 atomic_add_int(&pmc_pmcdisp[(R)], 1); \ 127 KASSERT(pmc_pmcdisp[(R)] <= 0, ("[pmc,%d] row disposition error", \ 128 __LINE__)); \ 129 } while (0) 130 131 #define PMC_MARK_ROW_THREAD(R) do { \ 132 KASSERT(pmc_pmcdisp[(R)] >= 0, ("[pmc,%d] row disposition error", \ 133 __LINE__)); \ 134 atomic_add_int(&pmc_pmcdisp[(R)], 1); \ 135 } while (0) 136 137 #define PMC_UNMARK_ROW_THREAD(R) do { \ 138 atomic_add_int(&pmc_pmcdisp[(R)], -1); \ 139 KASSERT(pmc_pmcdisp[(R)] >= 0, ("[pmc,%d] row disposition error", \ 140 __LINE__)); \ 141 } while (0) 142 143 /* various event handlers */ 144 static eventhandler_tag pmc_exit_tag, pmc_fork_tag, pmc_kld_load_tag, 145 pmc_kld_unload_tag; 146 147 /* Module statistics */ 148 struct pmc_driverstats pmc_stats; 149 150 /* Machine/processor dependent operations */ 151 static struct pmc_mdep *md; 152 153 /* 154 * Hash tables mapping owner processes and target threads to PMCs. 155 */ 156 struct mtx pmc_processhash_mtx; /* spin mutex */ 157 static u_long pmc_processhashmask; 158 static LIST_HEAD(pmc_processhash, pmc_process) *pmc_processhash; 159 160 /* 161 * Hash table of PMC owner descriptors. This table is protected by 162 * the shared PMC "sx" lock. 163 */ 164 static u_long pmc_ownerhashmask; 165 static LIST_HEAD(pmc_ownerhash, pmc_owner) *pmc_ownerhash; 166 167 /* 168 * List of PMC owners with system-wide sampling PMCs. 169 */ 170 static CK_LIST_HEAD(, pmc_owner) pmc_ss_owners; 171 172 /* 173 * List of free thread entries. This is protected by the spin 174 * mutex. 175 */ 176 static struct mtx pmc_threadfreelist_mtx; /* spin mutex */ 177 static LIST_HEAD(, pmc_thread) pmc_threadfreelist; 178 static int pmc_threadfreelist_entries = 0; 179 #define THREADENTRY_SIZE (sizeof(struct pmc_thread) + \ 180 (md->pmd_npmc * sizeof(struct pmc_threadpmcstate))) 181 182 /* 183 * Task to free thread descriptors 184 */ 185 static struct task free_task; 186 187 /* 188 * A map of row indices to classdep structures. 189 */ 190 static struct pmc_classdep **pmc_rowindex_to_classdep; 191 192 /* 193 * Prototypes 194 */ 195 196 #ifdef HWPMC_DEBUG 197 static int pmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS); 198 static int pmc_debugflags_parse(char *newstr, char *fence); 199 #endif 200 201 static int load(struct module *module, int cmd, void *arg); 202 static int pmc_add_sample(ring_type_t ring, struct pmc *pm, 203 struct trapframe *tf); 204 static void pmc_add_thread_descriptors_from_proc(struct proc *p, 205 struct pmc_process *pp); 206 static int pmc_attach_process(struct proc *p, struct pmc *pm); 207 static struct pmc *pmc_allocate_pmc_descriptor(void); 208 static struct pmc_owner *pmc_allocate_owner_descriptor(struct proc *p); 209 static int pmc_attach_one_process(struct proc *p, struct pmc *pm); 210 static bool pmc_can_allocate_row(int ri, enum pmc_mode mode); 211 static bool pmc_can_allocate_rowindex(struct proc *p, unsigned int ri, 212 int cpu); 213 static bool pmc_can_attach(struct pmc *pm, struct proc *p); 214 static void pmc_capture_user_callchain(int cpu, int soft, 215 struct trapframe *tf); 216 static void pmc_cleanup(void); 217 static int pmc_detach_process(struct proc *p, struct pmc *pm); 218 static int pmc_detach_one_process(struct proc *p, struct pmc *pm, 219 int flags); 220 static void pmc_destroy_owner_descriptor(struct pmc_owner *po); 221 static void pmc_destroy_pmc_descriptor(struct pmc *pm); 222 static void pmc_destroy_process_descriptor(struct pmc_process *pp); 223 static struct pmc_owner *pmc_find_owner_descriptor(struct proc *p); 224 static int pmc_find_pmc(pmc_id_t pmcid, struct pmc **pm); 225 static struct pmc *pmc_find_pmc_descriptor_in_process(struct pmc_owner *po, 226 pmc_id_t pmc); 227 static struct pmc_process *pmc_find_process_descriptor(struct proc *p, 228 uint32_t mode); 229 static struct pmc_thread *pmc_find_thread_descriptor(struct pmc_process *pp, 230 struct thread *td, uint32_t mode); 231 static void pmc_force_context_switch(void); 232 static void pmc_link_target_process(struct pmc *pm, 233 struct pmc_process *pp); 234 static void pmc_log_all_process_mappings(struct pmc_owner *po); 235 static void pmc_log_kernel_mappings(struct pmc *pm); 236 static void pmc_log_process_mappings(struct pmc_owner *po, struct proc *p); 237 static void pmc_maybe_remove_owner(struct pmc_owner *po); 238 static void pmc_post_callchain_callback(void); 239 static void pmc_process_allproc(struct pmc *pm); 240 static void pmc_process_csw_in(struct thread *td); 241 static void pmc_process_csw_out(struct thread *td); 242 static void pmc_process_exec(struct thread *td, 243 struct pmckern_procexec *pk); 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_proccreate(struct proc *p); 248 static void pmc_process_samples(int cpu, ring_type_t soft); 249 static void pmc_process_threadcreate(struct thread *td); 250 static void pmc_process_threadexit(struct thread *td); 251 static void pmc_process_thread_add(struct thread *td); 252 static void pmc_process_thread_delete(struct thread *td); 253 static void pmc_process_thread_userret(struct thread *td); 254 static void pmc_release_pmc_descriptor(struct pmc *pmc); 255 static void pmc_remove_owner(struct pmc_owner *po); 256 static void pmc_remove_process_descriptor(struct pmc_process *pp); 257 static int pmc_start(struct pmc *pm); 258 static int pmc_stop(struct pmc *pm); 259 static int pmc_syscall_handler(struct thread *td, void *syscall_args); 260 static struct pmc_thread *pmc_thread_descriptor_pool_alloc(void); 261 static void pmc_thread_descriptor_pool_drain(void); 262 static void pmc_thread_descriptor_pool_free(struct pmc_thread *pt); 263 static void pmc_unlink_target_process(struct pmc *pmc, 264 struct pmc_process *pp); 265 266 static int generic_switch_in(struct pmc_cpu *pc, struct pmc_process *pp); 267 static int generic_switch_out(struct pmc_cpu *pc, struct pmc_process *pp); 268 static struct pmc_mdep *pmc_generic_cpu_initialize(void); 269 static void pmc_generic_cpu_finalize(struct pmc_mdep *md); 270 271 /* 272 * Kernel tunables and sysctl(8) interface. 273 */ 274 275 SYSCTL_DECL(_kern_hwpmc); 276 SYSCTL_NODE(_kern_hwpmc, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 277 "HWPMC stats"); 278 279 /* Stats. */ 280 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, intr_ignored, CTLFLAG_RW, 281 &pmc_stats.pm_intr_ignored, 282 "# of interrupts ignored"); 283 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, intr_processed, CTLFLAG_RW, 284 &pmc_stats.pm_intr_processed, 285 "# of interrupts processed"); 286 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, intr_bufferfull, CTLFLAG_RW, 287 &pmc_stats.pm_intr_bufferfull, 288 "# of interrupts where buffer was full"); 289 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, syscalls, CTLFLAG_RW, 290 &pmc_stats.pm_syscalls, 291 "# of syscalls"); 292 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, syscall_errors, CTLFLAG_RW, 293 &pmc_stats.pm_syscall_errors, 294 "# of syscall_errors"); 295 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, buffer_requests, CTLFLAG_RW, 296 &pmc_stats.pm_buffer_requests, 297 "# of buffer requests"); 298 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, buffer_requests_failed, 299 CTLFLAG_RW, &pmc_stats.pm_buffer_requests_failed, 300 "# of buffer requests which failed"); 301 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, log_sweeps, CTLFLAG_RW, 302 &pmc_stats.pm_log_sweeps, 303 "# of times samples were processed"); 304 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, merges, CTLFLAG_RW, 305 &pmc_stats.pm_merges, 306 "# of times kernel stack was found for user trace"); 307 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, overwrites, CTLFLAG_RW, 308 &pmc_stats.pm_overwrites, 309 "# of times a sample was overwritten before being logged"); 310 311 static int pmc_callchaindepth = PMC_CALLCHAIN_DEPTH; 312 SYSCTL_INT(_kern_hwpmc, OID_AUTO, callchaindepth, CTLFLAG_RDTUN, 313 &pmc_callchaindepth, 0, 314 "depth of call chain records"); 315 316 char pmc_cpuid[PMC_CPUID_LEN]; 317 SYSCTL_STRING(_kern_hwpmc, OID_AUTO, cpuid, CTLFLAG_RD, 318 pmc_cpuid, 0, 319 "cpu version string"); 320 321 #ifdef HWPMC_DEBUG 322 struct pmc_debugflags pmc_debugflags = PMC_DEBUG_DEFAULT_FLAGS; 323 char pmc_debugstr[PMC_DEBUG_STRSIZE]; 324 TUNABLE_STR(PMC_SYSCTL_NAME_PREFIX "debugflags", pmc_debugstr, 325 sizeof(pmc_debugstr)); 326 SYSCTL_PROC(_kern_hwpmc, OID_AUTO, debugflags, 327 CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, 328 0, 0, pmc_debugflags_sysctl_handler, "A", 329 "debug flags"); 330 #endif 331 332 /* 333 * kern.hwpmc.hashsize -- determines the number of rows in the 334 * of the hash table used to look up threads 335 */ 336 static int pmc_hashsize = PMC_HASH_SIZE; 337 SYSCTL_INT(_kern_hwpmc, OID_AUTO, hashsize, CTLFLAG_RDTUN, 338 &pmc_hashsize, 0, 339 "rows in hash tables"); 340 341 /* 342 * kern.hwpmc.nsamples --- number of PC samples/callchain stacks per CPU 343 */ 344 static int pmc_nsamples = PMC_NSAMPLES; 345 SYSCTL_INT(_kern_hwpmc, OID_AUTO, nsamples, CTLFLAG_RDTUN, 346 &pmc_nsamples, 0, 347 "number of PC samples per CPU"); 348 349 static uint64_t pmc_sample_mask = PMC_NSAMPLES - 1; 350 351 /* 352 * kern.hwpmc.mtxpoolsize -- number of mutexes in the mutex pool. 353 */ 354 static int pmc_mtxpool_size = PMC_MTXPOOL_SIZE; 355 SYSCTL_INT(_kern_hwpmc, OID_AUTO, mtxpoolsize, CTLFLAG_RDTUN, 356 &pmc_mtxpool_size, 0, 357 "size of spin mutex pool"); 358 359 /* 360 * kern.hwpmc.threadfreelist_entries -- number of free entries 361 */ 362 SYSCTL_INT(_kern_hwpmc, OID_AUTO, threadfreelist_entries, CTLFLAG_RD, 363 &pmc_threadfreelist_entries, 0, 364 "number of available thread entries"); 365 366 /* 367 * kern.hwpmc.threadfreelist_max -- maximum number of free entries 368 */ 369 static int pmc_threadfreelist_max = PMC_THREADLIST_MAX; 370 SYSCTL_INT(_kern_hwpmc, OID_AUTO, threadfreelist_max, CTLFLAG_RW, 371 &pmc_threadfreelist_max, 0, 372 "maximum number of available thread entries before freeing some"); 373 374 /* 375 * kern.hwpmc.mincount -- minimum sample count 376 */ 377 static u_int pmc_mincount = 1000; 378 SYSCTL_INT(_kern_hwpmc, OID_AUTO, mincount, CTLFLAG_RWTUN, 379 &pmc_mincount, 0, 380 "minimum count for sampling counters"); 381 382 /* 383 * security.bsd.unprivileged_syspmcs -- allow non-root processes to 384 * allocate system-wide PMCs. 385 * 386 * Allowing unprivileged processes to allocate system PMCs is convenient 387 * if system-wide measurements need to be taken concurrently with other 388 * per-process measurements. This feature is turned off by default. 389 */ 390 static int pmc_unprivileged_syspmcs = 0; 391 SYSCTL_INT(_security_bsd, OID_AUTO, unprivileged_syspmcs, CTLFLAG_RWTUN, 392 &pmc_unprivileged_syspmcs, 0, 393 "allow unprivileged process to allocate system PMCs"); 394 395 /* 396 * Hash function. Discard the lower 2 bits of the pointer since 397 * these are always zero for our uses. The hash multiplier is 398 * round((2^LONG_BIT) * ((sqrt(5)-1)/2)). 399 */ 400 #if LONG_BIT == 64 401 #define _PMC_HM 11400714819323198486u 402 #elif LONG_BIT == 32 403 #define _PMC_HM 2654435769u 404 #else 405 #error Must know the size of 'long' to compile 406 #endif 407 408 #define PMC_HASH_PTR(P,M) ((((unsigned long) (P) >> 2) * _PMC_HM) & (M)) 409 410 /* 411 * Syscall structures 412 */ 413 414 /* The `sysent' for the new syscall */ 415 static struct sysent pmc_sysent = { 416 .sy_narg = 2, 417 .sy_call = pmc_syscall_handler, 418 }; 419 420 static struct syscall_module_data pmc_syscall_mod = { 421 .chainevh = load, 422 .chainarg = NULL, 423 .offset = &pmc_syscall_num, 424 .new_sysent = &pmc_sysent, 425 .old_sysent = { .sy_narg = 0, .sy_call = NULL }, 426 .flags = SY_THR_STATIC_KLD, 427 }; 428 429 static moduledata_t pmc_mod = { 430 .name = PMC_MODULE_NAME, 431 .evhand = syscall_module_handler, 432 .priv = &pmc_syscall_mod, 433 }; 434 435 #ifdef EARLY_AP_STARTUP 436 DECLARE_MODULE(pmc, pmc_mod, SI_SUB_SYSCALLS, SI_ORDER_ANY); 437 #else 438 DECLARE_MODULE(pmc, pmc_mod, SI_SUB_SMP, SI_ORDER_ANY); 439 #endif 440 MODULE_VERSION(pmc, PMC_VERSION); 441 442 #ifdef HWPMC_DEBUG 443 enum pmc_dbgparse_state { 444 PMCDS_WS, /* in whitespace */ 445 PMCDS_MAJOR, /* seen a major keyword */ 446 PMCDS_MINOR 447 }; 448 449 static int 450 pmc_debugflags_parse(char *newstr, char *fence) 451 { 452 struct pmc_debugflags *tmpflags; 453 size_t kwlen; 454 char c, *p, *q; 455 int error, *newbits, tmp; 456 int found; 457 458 tmpflags = malloc(sizeof(*tmpflags), M_PMC, M_WAITOK | M_ZERO); 459 460 error = 0; 461 for (p = newstr; p < fence && (c = *p); p++) { 462 /* skip white space */ 463 if (c == ' ' || c == '\t') 464 continue; 465 466 /* look for a keyword followed by "=" */ 467 for (q = p; p < fence && (c = *p) && c != '='; p++) 468 ; 469 if (c != '=') { 470 error = EINVAL; 471 goto done; 472 } 473 474 kwlen = p - q; 475 newbits = NULL; 476 477 /* lookup flag group name */ 478 #define DBG_SET_FLAG_MAJ(S,F) \ 479 if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0) \ 480 newbits = &tmpflags->pdb_ ## F; 481 482 DBG_SET_FLAG_MAJ("cpu", CPU); 483 DBG_SET_FLAG_MAJ("csw", CSW); 484 DBG_SET_FLAG_MAJ("logging", LOG); 485 DBG_SET_FLAG_MAJ("module", MOD); 486 DBG_SET_FLAG_MAJ("md", MDP); 487 DBG_SET_FLAG_MAJ("owner", OWN); 488 DBG_SET_FLAG_MAJ("pmc", PMC); 489 DBG_SET_FLAG_MAJ("process", PRC); 490 DBG_SET_FLAG_MAJ("sampling", SAM); 491 #undef DBG_SET_FLAG_MAJ 492 493 if (newbits == NULL) { 494 error = EINVAL; 495 goto done; 496 } 497 498 p++; /* skip the '=' */ 499 500 /* Now parse the individual flags */ 501 tmp = 0; 502 newflag: 503 for (q = p; p < fence && (c = *p); p++) 504 if (c == ' ' || c == '\t' || c == ',') 505 break; 506 507 /* p == fence or c == ws or c == "," or c == 0 */ 508 509 if ((kwlen = p - q) == 0) { 510 *newbits = tmp; 511 continue; 512 } 513 514 found = 0; 515 #define DBG_SET_FLAG_MIN(S,F) \ 516 if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0) \ 517 tmp |= found = (1 << PMC_DEBUG_MIN_ ## F) 518 519 /* a '*' denotes all possible flags in the group */ 520 if (kwlen == 1 && *q == '*') 521 tmp = found = ~0; 522 /* look for individual flag names */ 523 DBG_SET_FLAG_MIN("allocaterow", ALR); 524 DBG_SET_FLAG_MIN("allocate", ALL); 525 DBG_SET_FLAG_MIN("attach", ATT); 526 DBG_SET_FLAG_MIN("bind", BND); 527 DBG_SET_FLAG_MIN("config", CFG); 528 DBG_SET_FLAG_MIN("exec", EXC); 529 DBG_SET_FLAG_MIN("exit", EXT); 530 DBG_SET_FLAG_MIN("find", FND); 531 DBG_SET_FLAG_MIN("flush", FLS); 532 DBG_SET_FLAG_MIN("fork", FRK); 533 DBG_SET_FLAG_MIN("getbuf", GTB); 534 DBG_SET_FLAG_MIN("hook", PMH); 535 DBG_SET_FLAG_MIN("init", INI); 536 DBG_SET_FLAG_MIN("intr", INT); 537 DBG_SET_FLAG_MIN("linktarget", TLK); 538 DBG_SET_FLAG_MIN("mayberemove", OMR); 539 DBG_SET_FLAG_MIN("ops", OPS); 540 DBG_SET_FLAG_MIN("read", REA); 541 DBG_SET_FLAG_MIN("register", REG); 542 DBG_SET_FLAG_MIN("release", REL); 543 DBG_SET_FLAG_MIN("remove", ORM); 544 DBG_SET_FLAG_MIN("sample", SAM); 545 DBG_SET_FLAG_MIN("scheduleio", SIO); 546 DBG_SET_FLAG_MIN("select", SEL); 547 DBG_SET_FLAG_MIN("signal", SIG); 548 DBG_SET_FLAG_MIN("swi", SWI); 549 DBG_SET_FLAG_MIN("swo", SWO); 550 DBG_SET_FLAG_MIN("start", STA); 551 DBG_SET_FLAG_MIN("stop", STO); 552 DBG_SET_FLAG_MIN("syscall", PMS); 553 DBG_SET_FLAG_MIN("unlinktarget", TUL); 554 DBG_SET_FLAG_MIN("write", WRI); 555 #undef DBG_SET_FLAG_MIN 556 if (found == 0) { 557 /* unrecognized flag name */ 558 error = EINVAL; 559 goto done; 560 } 561 562 if (c == 0 || c == ' ' || c == '\t') { /* end of flag group */ 563 *newbits = tmp; 564 continue; 565 } 566 567 p++; 568 goto newflag; 569 } 570 571 /* save the new flag set */ 572 bcopy(tmpflags, &pmc_debugflags, sizeof(pmc_debugflags)); 573 done: 574 free(tmpflags, M_PMC); 575 return (error); 576 } 577 578 static int 579 pmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS) 580 { 581 char *fence, *newstr; 582 int error; 583 u_int n; 584 585 n = sizeof(pmc_debugstr); 586 newstr = malloc(n, M_PMC, M_WAITOK | M_ZERO); 587 strlcpy(newstr, pmc_debugstr, n); 588 589 error = sysctl_handle_string(oidp, newstr, n, req); 590 591 /* if there is a new string, parse and copy it */ 592 if (error == 0 && req->newptr != NULL) { 593 fence = newstr + (n < req->newlen ? n : req->newlen + 1); 594 error = pmc_debugflags_parse(newstr, fence); 595 if (error == 0) 596 strlcpy(pmc_debugstr, newstr, sizeof(pmc_debugstr)); 597 } 598 free(newstr, M_PMC); 599 600 return (error); 601 } 602 #endif 603 604 /* 605 * Map a row index to a classdep structure and return the adjusted row 606 * index for the PMC class index. 607 */ 608 static struct pmc_classdep * 609 pmc_ri_to_classdep(struct pmc_mdep *md __unused, int ri, int *adjri) 610 { 611 struct pmc_classdep *pcd; 612 613 KASSERT(ri >= 0 && ri < md->pmd_npmc, 614 ("[pmc,%d] illegal row-index %d", __LINE__, ri)); 615 616 pcd = pmc_rowindex_to_classdep[ri]; 617 KASSERT(pcd != NULL, 618 ("[pmc,%d] ri %d null pcd", __LINE__, ri)); 619 620 *adjri = ri - pcd->pcd_ri; 621 KASSERT(*adjri >= 0 && *adjri < pcd->pcd_num, 622 ("[pmc,%d] adjusted row-index %d", __LINE__, *adjri)); 623 624 return (pcd); 625 } 626 627 /* 628 * Concurrency Control 629 * 630 * The driver manages the following data structures: 631 * 632 * - target process descriptors, one per target process 633 * - owner process descriptors (and attached lists), one per owner process 634 * - lookup hash tables for owner and target processes 635 * - PMC descriptors (and attached lists) 636 * - per-cpu hardware state 637 * - the 'hook' variable through which the kernel calls into 638 * this module 639 * - the machine hardware state (managed by the MD layer) 640 * 641 * These data structures are accessed from: 642 * 643 * - thread context-switch code 644 * - interrupt handlers (possibly on multiple cpus) 645 * - kernel threads on multiple cpus running on behalf of user 646 * processes doing system calls 647 * - this driver's private kernel threads 648 * 649 * = Locks and Locking strategy = 650 * 651 * The driver uses four locking strategies for its operation: 652 * 653 * - The global SX lock "pmc_sx" is used to protect internal 654 * data structures. 655 * 656 * Calls into the module by syscall() start with this lock being 657 * held in exclusive mode. Depending on the requested operation, 658 * the lock may be downgraded to 'shared' mode to allow more 659 * concurrent readers into the module. Calls into the module from 660 * other parts of the kernel acquire the lock in shared mode. 661 * 662 * This SX lock is held in exclusive mode for any operations that 663 * modify the linkages between the driver's internal data structures. 664 * 665 * The 'pmc_hook' function pointer is also protected by this lock. 666 * It is only examined with the sx lock held in exclusive mode. The 667 * kernel module is allowed to be unloaded only with the sx lock held 668 * in exclusive mode. In normal syscall handling, after acquiring the 669 * pmc_sx lock we first check that 'pmc_hook' is non-null before 670 * proceeding. This prevents races between the thread unloading the module 671 * and other threads seeking to use the module. 672 * 673 * - Lookups of target process structures and owner process structures 674 * cannot use the global "pmc_sx" SX lock because these lookups need 675 * to happen during context switches and in other critical sections 676 * where sleeping is not allowed. We protect these lookup tables 677 * with their own private spin-mutexes, "pmc_processhash_mtx" and 678 * "pmc_ownerhash_mtx". 679 * 680 * - Interrupt handlers work in a lock free manner. At interrupt 681 * time, handlers look at the PMC pointer (phw->phw_pmc) configured 682 * when the PMC was started. If this pointer is NULL, the interrupt 683 * is ignored after updating driver statistics. We ensure that this 684 * pointer is set (using an atomic operation if necessary) before the 685 * PMC hardware is started. Conversely, this pointer is unset atomically 686 * only after the PMC hardware is stopped. 687 * 688 * We ensure that everything needed for the operation of an 689 * interrupt handler is available without it needing to acquire any 690 * locks. We also ensure that a PMC's software state is destroyed only 691 * after the PMC is taken off hardware (on all CPUs). 692 * 693 * - Context-switch handling with process-private PMCs needs more 694 * care. 695 * 696 * A given process may be the target of multiple PMCs. For example, 697 * PMCATTACH and PMCDETACH may be requested by a process on one CPU 698 * while the target process is running on another. A PMC could also 699 * be getting released because its owner is exiting. We tackle 700 * these situations in the following manner: 701 * 702 * - each target process structure 'pmc_process' has an array 703 * of 'struct pmc *' pointers, one for each hardware PMC. 704 * 705 * - At context switch IN time, each "target" PMC in RUNNING state 706 * gets started on hardware and a pointer to each PMC is copied into 707 * the per-cpu phw array. The 'runcount' for the PMC is 708 * incremented. 709 * 710 * - At context switch OUT time, all process-virtual PMCs are stopped 711 * on hardware. The saved value is added to the PMCs value field 712 * only if the PMC is in a non-deleted state (the PMCs state could 713 * have changed during the current time slice). 714 * 715 * Note that since in-between a switch IN on a processor and a switch 716 * OUT, the PMC could have been released on another CPU. Therefore 717 * context switch OUT always looks at the hardware state to turn 718 * OFF PMCs and will update a PMC's saved value only if reachable 719 * from the target process record. 720 * 721 * - OP PMCRELEASE could be called on a PMC at any time (the PMC could 722 * be attached to many processes at the time of the call and could 723 * be active on multiple CPUs). 724 * 725 * We prevent further scheduling of the PMC by marking it as in 726 * state 'DELETED'. If the runcount of the PMC is non-zero then 727 * this PMC is currently running on a CPU somewhere. The thread 728 * doing the PMCRELEASE operation waits by repeatedly doing a 729 * pause() till the runcount comes to zero. 730 * 731 * The contents of a PMC descriptor (struct pmc) are protected using 732 * a spin-mutex. In order to save space, we use a mutex pool. 733 * 734 * In terms of lock types used by witness(4), we use: 735 * - Type "pmc-sx", used by the global SX lock. 736 * - Type "pmc-sleep", for sleep mutexes used by logger threads. 737 * - Type "pmc-per-proc", for protecting PMC owner descriptors. 738 * - Type "pmc-leaf", used for all other spin mutexes. 739 */ 740 741 /* 742 * Save the CPU binding of the current kthread. 743 */ 744 void 745 pmc_save_cpu_binding(struct pmc_binding *pb) 746 { 747 PMCDBG0(CPU,BND,2, "save-cpu"); 748 thread_lock(curthread); 749 pb->pb_bound = sched_is_bound(curthread); 750 pb->pb_cpu = curthread->td_oncpu; 751 pb->pb_priority = curthread->td_priority; 752 thread_unlock(curthread); 753 PMCDBG1(CPU,BND,2, "save-cpu cpu=%d", pb->pb_cpu); 754 } 755 756 /* 757 * Restore the CPU binding of the current thread. 758 */ 759 void 760 pmc_restore_cpu_binding(struct pmc_binding *pb) 761 { 762 PMCDBG2(CPU,BND,2, "restore-cpu curcpu=%d restore=%d", 763 curthread->td_oncpu, pb->pb_cpu); 764 thread_lock(curthread); 765 sched_bind(curthread, pb->pb_cpu); 766 if (!pb->pb_bound) 767 sched_unbind(curthread); 768 sched_prio(curthread, pb->pb_priority); 769 thread_unlock(curthread); 770 PMCDBG0(CPU,BND,2, "restore-cpu done"); 771 } 772 773 /* 774 * Move execution over to the specified CPU and bind it there. 775 */ 776 void 777 pmc_select_cpu(int cpu) 778 { 779 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), 780 ("[pmc,%d] bad cpu number %d", __LINE__, cpu)); 781 782 /* Never move to an inactive CPU. */ 783 KASSERT(pmc_cpu_is_active(cpu), ("[pmc,%d] selecting inactive " 784 "CPU %d", __LINE__, cpu)); 785 786 PMCDBG1(CPU,SEL,2, "select-cpu cpu=%d", cpu); 787 thread_lock(curthread); 788 sched_prio(curthread, PRI_MIN); 789 sched_bind(curthread, cpu); 790 thread_unlock(curthread); 791 792 KASSERT(curthread->td_oncpu == cpu, 793 ("[pmc,%d] CPU not bound [cpu=%d, curr=%d]", __LINE__, 794 cpu, curthread->td_oncpu)); 795 796 PMCDBG1(CPU,SEL,2, "select-cpu cpu=%d ok", cpu); 797 } 798 799 /* 800 * Force a context switch. 801 * 802 * We do this by pause'ing for 1 tick -- invoking mi_switch() is not 803 * guaranteed to force a context switch. 804 */ 805 static void 806 pmc_force_context_switch(void) 807 { 808 809 pause("pmcctx", 1); 810 } 811 812 uint64_t 813 pmc_rdtsc(void) 814 { 815 #if defined(__i386__) || defined(__amd64__) 816 if (__predict_true(amd_feature & AMDID_RDTSCP)) 817 return (rdtscp()); 818 else 819 return (rdtsc()); 820 #else 821 return (get_cyclecount()); 822 #endif 823 } 824 825 /* 826 * Get the file name for an executable. This is a simple wrapper 827 * around vn_fullpath(9). 828 */ 829 static void 830 pmc_getfilename(struct vnode *v, char **fullpath, char **freepath) 831 { 832 833 *fullpath = "unknown"; 834 *freepath = NULL; 835 vn_fullpath(v, fullpath, freepath); 836 } 837 838 /* 839 * Remove a process owning PMCs. 840 */ 841 void 842 pmc_remove_owner(struct pmc_owner *po) 843 { 844 struct pmc *pm, *tmp; 845 846 sx_assert(&pmc_sx, SX_XLOCKED); 847 848 PMCDBG1(OWN,ORM,1, "remove-owner po=%p", po); 849 850 /* Remove descriptor from the owner hash table */ 851 LIST_REMOVE(po, po_next); 852 853 /* release all owned PMC descriptors */ 854 LIST_FOREACH_SAFE(pm, &po->po_pmcs, pm_next, tmp) { 855 PMCDBG1(OWN,ORM,2, "pmc=%p", pm); 856 KASSERT(pm->pm_owner == po, 857 ("[pmc,%d] owner %p != po %p", __LINE__, pm->pm_owner, po)); 858 859 pmc_release_pmc_descriptor(pm); /* will unlink from the list */ 860 pmc_destroy_pmc_descriptor(pm); 861 } 862 863 KASSERT(po->po_sscount == 0, 864 ("[pmc,%d] SS count not zero", __LINE__)); 865 KASSERT(LIST_EMPTY(&po->po_pmcs), 866 ("[pmc,%d] PMC list not empty", __LINE__)); 867 868 /* de-configure the log file if present */ 869 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 870 pmclog_deconfigure_log(po); 871 } 872 873 /* 874 * Remove an owner process record if all conditions are met. 875 */ 876 static void 877 pmc_maybe_remove_owner(struct pmc_owner *po) 878 { 879 880 PMCDBG1(OWN,OMR,1, "maybe-remove-owner po=%p", po); 881 882 /* 883 * Remove owner record if 884 * - this process does not own any PMCs 885 * - this process has not allocated a system-wide sampling buffer 886 */ 887 if (LIST_EMPTY(&po->po_pmcs) && 888 ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)) { 889 pmc_remove_owner(po); 890 pmc_destroy_owner_descriptor(po); 891 } 892 } 893 894 /* 895 * Add an association between a target process and a PMC. 896 */ 897 static void 898 pmc_link_target_process(struct pmc *pm, struct pmc_process *pp) 899 { 900 struct pmc_target *pt; 901 struct pmc_thread *pt_td __diagused; 902 int ri; 903 904 sx_assert(&pmc_sx, SX_XLOCKED); 905 KASSERT(pm != NULL && pp != NULL, 906 ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp)); 907 KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)), 908 ("[pmc,%d] Attaching a non-process-virtual pmc=%p to pid=%d", 909 __LINE__, pm, pp->pp_proc->p_pid)); 910 KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= ((int) md->pmd_npmc - 1), 911 ("[pmc,%d] Illegal reference count %d for process record %p", 912 __LINE__, pp->pp_refcnt, (void *) pp)); 913 914 ri = PMC_TO_ROWINDEX(pm); 915 916 PMCDBG3(PRC,TLK,1, "link-target pmc=%p ri=%d pmc-process=%p", 917 pm, ri, pp); 918 919 #ifdef HWPMC_DEBUG 920 LIST_FOREACH(pt, &pm->pm_targets, pt_next) { 921 if (pt->pt_process == pp) 922 KASSERT(0, ("[pmc,%d] pp %p already in pmc %p targets", 923 __LINE__, pp, pm)); 924 } 925 #endif 926 pt = malloc(sizeof(struct pmc_target), M_PMC, M_WAITOK | M_ZERO); 927 pt->pt_process = pp; 928 929 LIST_INSERT_HEAD(&pm->pm_targets, pt, pt_next); 930 931 atomic_store_rel_ptr((uintptr_t *)&pp->pp_pmcs[ri].pp_pmc, 932 (uintptr_t)pm); 933 934 if (pm->pm_owner->po_owner == pp->pp_proc) 935 pm->pm_flags |= PMC_F_ATTACHED_TO_OWNER; 936 937 /* 938 * Initialize the per-process values at this row index. 939 */ 940 pp->pp_pmcs[ri].pp_pmcval = PMC_TO_MODE(pm) == PMC_MODE_TS ? 941 pm->pm_sc.pm_reloadcount : 0; 942 pp->pp_refcnt++; 943 944 #ifdef INVARIANTS 945 /* Confirm that the per-thread values at this row index are cleared. */ 946 if (PMC_TO_MODE(pm) == PMC_MODE_TS) { 947 mtx_lock_spin(pp->pp_tdslock); 948 LIST_FOREACH(pt_td, &pp->pp_tds, pt_next) { 949 KASSERT(pt_td->pt_pmcs[ri].pt_pmcval == (pmc_value_t) 0, 950 ("[pmc,%d] pt_pmcval not cleared for pid=%d at " 951 "ri=%d", __LINE__, pp->pp_proc->p_pid, ri)); 952 } 953 mtx_unlock_spin(pp->pp_tdslock); 954 } 955 #endif 956 } 957 958 /* 959 * Removes the association between a target process and a PMC. 960 */ 961 static void 962 pmc_unlink_target_process(struct pmc *pm, struct pmc_process *pp) 963 { 964 int ri; 965 struct proc *p; 966 struct pmc_target *ptgt; 967 struct pmc_thread *pt; 968 969 sx_assert(&pmc_sx, SX_XLOCKED); 970 971 KASSERT(pm != NULL && pp != NULL, 972 ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp)); 973 974 KASSERT(pp->pp_refcnt >= 1 && pp->pp_refcnt <= (int) md->pmd_npmc, 975 ("[pmc,%d] Illegal ref count %d on process record %p", 976 __LINE__, pp->pp_refcnt, (void *) pp)); 977 978 ri = PMC_TO_ROWINDEX(pm); 979 980 PMCDBG3(PRC,TUL,1, "unlink-target pmc=%p ri=%d pmc-process=%p", 981 pm, ri, pp); 982 983 KASSERT(pp->pp_pmcs[ri].pp_pmc == pm, 984 ("[pmc,%d] PMC ri %d mismatch pmc %p pp->[ri] %p", __LINE__, 985 ri, pm, pp->pp_pmcs[ri].pp_pmc)); 986 987 pp->pp_pmcs[ri].pp_pmc = NULL; 988 pp->pp_pmcs[ri].pp_pmcval = (pmc_value_t)0; 989 990 /* Clear the per-thread values at this row index. */ 991 if (PMC_TO_MODE(pm) == PMC_MODE_TS) { 992 mtx_lock_spin(pp->pp_tdslock); 993 LIST_FOREACH(pt, &pp->pp_tds, pt_next) 994 pt->pt_pmcs[ri].pt_pmcval = (pmc_value_t)0; 995 mtx_unlock_spin(pp->pp_tdslock); 996 } 997 998 /* Remove owner-specific flags */ 999 if (pm->pm_owner->po_owner == pp->pp_proc) { 1000 pp->pp_flags &= ~PMC_PP_ENABLE_MSR_ACCESS; 1001 pm->pm_flags &= ~PMC_F_ATTACHED_TO_OWNER; 1002 } 1003 1004 pp->pp_refcnt--; 1005 1006 /* Remove the target process from the PMC structure */ 1007 LIST_FOREACH(ptgt, &pm->pm_targets, pt_next) 1008 if (ptgt->pt_process == pp) 1009 break; 1010 1011 KASSERT(ptgt != NULL, ("[pmc,%d] process %p (pp: %p) not found " 1012 "in pmc %p", __LINE__, pp->pp_proc, pp, pm)); 1013 1014 LIST_REMOVE(ptgt, pt_next); 1015 free(ptgt, M_PMC); 1016 1017 /* if the PMC now lacks targets, send the owner a SIGIO */ 1018 if (LIST_EMPTY(&pm->pm_targets)) { 1019 p = pm->pm_owner->po_owner; 1020 PROC_LOCK(p); 1021 kern_psignal(p, SIGIO); 1022 PROC_UNLOCK(p); 1023 1024 PMCDBG2(PRC,SIG,2, "signalling proc=%p signal=%d", p, SIGIO); 1025 } 1026 } 1027 1028 /* 1029 * Check if PMC 'pm' may be attached to target process 't'. 1030 */ 1031 1032 static bool 1033 pmc_can_attach(struct pmc *pm, struct proc *t) 1034 { 1035 struct proc *o; /* pmc owner */ 1036 struct ucred *oc, *tc; /* owner, target credentials */ 1037 bool decline_attach; 1038 1039 /* 1040 * A PMC's owner can always attach that PMC to itself. 1041 */ 1042 1043 if ((o = pm->pm_owner->po_owner) == t) 1044 return (true); 1045 1046 PROC_LOCK(o); 1047 oc = o->p_ucred; 1048 crhold(oc); 1049 PROC_UNLOCK(o); 1050 1051 PROC_LOCK(t); 1052 tc = t->p_ucred; 1053 crhold(tc); 1054 PROC_UNLOCK(t); 1055 1056 /* 1057 * The effective uid of the PMC owner should match at least one 1058 * of the {effective,real,saved} uids of the target process. 1059 */ 1060 1061 decline_attach = oc->cr_uid != tc->cr_uid && 1062 oc->cr_uid != tc->cr_svuid && 1063 oc->cr_uid != tc->cr_ruid; 1064 1065 /* 1066 * Every one of the target's group ids, must be in the owner's 1067 * group list. 1068 */ 1069 for (int i = 0; !decline_attach && i < tc->cr_ngroups; i++) 1070 decline_attach = !groupmember(tc->cr_groups[i], oc); 1071 if (!decline_attach) 1072 decline_attach = !groupmember(tc->cr_gid, oc) || 1073 !groupmember(tc->cr_rgid, oc) || 1074 !groupmember(tc->cr_svgid, oc); 1075 1076 crfree(tc); 1077 crfree(oc); 1078 1079 return (!decline_attach); 1080 } 1081 1082 /* 1083 * Attach a process to a PMC. 1084 */ 1085 static int 1086 pmc_attach_one_process(struct proc *p, struct pmc *pm) 1087 { 1088 int ri, error; 1089 char *fullpath, *freepath; 1090 struct pmc_process *pp; 1091 1092 sx_assert(&pmc_sx, SX_XLOCKED); 1093 1094 PMCDBG5(PRC,ATT,2, "attach-one pm=%p ri=%d proc=%p (%d, %s)", pm, 1095 PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm); 1096 1097 /* 1098 * Locate the process descriptor corresponding to process 'p', 1099 * allocating space as needed. 1100 * 1101 * Verify that rowindex 'pm_rowindex' is free in the process 1102 * descriptor. 1103 * 1104 * If not, allocate space for a descriptor and link the 1105 * process descriptor and PMC. 1106 */ 1107 ri = PMC_TO_ROWINDEX(pm); 1108 1109 /* mark process as using HWPMCs */ 1110 PROC_LOCK(p); 1111 p->p_flag |= P_HWPMC; 1112 PROC_UNLOCK(p); 1113 1114 if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_ALLOCATE)) == NULL) { 1115 error = ENOMEM; 1116 goto fail; 1117 } 1118 1119 if (pp->pp_pmcs[ri].pp_pmc == pm) {/* already present at slot [ri] */ 1120 error = EEXIST; 1121 goto fail; 1122 } 1123 1124 if (pp->pp_pmcs[ri].pp_pmc != NULL) { 1125 error = EBUSY; 1126 goto fail; 1127 } 1128 1129 pmc_link_target_process(pm, pp); 1130 1131 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) && 1132 (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) == 0) 1133 pm->pm_flags |= PMC_F_NEEDS_LOGFILE; 1134 1135 pm->pm_flags |= PMC_F_ATTACH_DONE; /* mark as attached */ 1136 1137 /* issue an attach event to a configured log file */ 1138 if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE) { 1139 if (p->p_flag & P_KPROC) { 1140 fullpath = kernelname; 1141 freepath = NULL; 1142 } else { 1143 pmc_getfilename(p->p_textvp, &fullpath, &freepath); 1144 pmclog_process_pmcattach(pm, p->p_pid, fullpath); 1145 } 1146 free(freepath, M_TEMP); 1147 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) 1148 pmc_log_process_mappings(pm->pm_owner, p); 1149 } 1150 1151 return (0); 1152 fail: 1153 PROC_LOCK(p); 1154 p->p_flag &= ~P_HWPMC; 1155 PROC_UNLOCK(p); 1156 return (error); 1157 } 1158 1159 /* 1160 * Attach a process and optionally its children 1161 */ 1162 static int 1163 pmc_attach_process(struct proc *p, struct pmc *pm) 1164 { 1165 int error; 1166 struct proc *top; 1167 1168 sx_assert(&pmc_sx, SX_XLOCKED); 1169 1170 PMCDBG5(PRC,ATT,1, "attach pm=%p ri=%d proc=%p (%d, %s)", pm, 1171 PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm); 1172 1173 /* 1174 * If this PMC successfully allowed a GETMSR operation 1175 * in the past, disallow further ATTACHes. 1176 */ 1177 if ((pm->pm_flags & PMC_PP_ENABLE_MSR_ACCESS) != 0) 1178 return (EPERM); 1179 1180 if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0) 1181 return (pmc_attach_one_process(p, pm)); 1182 1183 /* 1184 * Traverse all child processes, attaching them to 1185 * this PMC. 1186 */ 1187 sx_slock(&proctree_lock); 1188 1189 top = p; 1190 for (;;) { 1191 if ((error = pmc_attach_one_process(p, pm)) != 0) 1192 break; 1193 if (!LIST_EMPTY(&p->p_children)) 1194 p = LIST_FIRST(&p->p_children); 1195 else for (;;) { 1196 if (p == top) 1197 goto done; 1198 if (LIST_NEXT(p, p_sibling)) { 1199 p = LIST_NEXT(p, p_sibling); 1200 break; 1201 } 1202 p = p->p_pptr; 1203 } 1204 } 1205 1206 if (error != 0) 1207 (void)pmc_detach_process(top, pm); 1208 1209 done: 1210 sx_sunlock(&proctree_lock); 1211 return (error); 1212 } 1213 1214 /* 1215 * Detach a process from a PMC. If there are no other PMCs tracking 1216 * this process, remove the process structure from its hash table. If 1217 * 'flags' contains PMC_FLAG_REMOVE, then free the process structure. 1218 */ 1219 static int 1220 pmc_detach_one_process(struct proc *p, struct pmc *pm, int flags) 1221 { 1222 int ri; 1223 struct pmc_process *pp; 1224 1225 sx_assert(&pmc_sx, SX_XLOCKED); 1226 1227 KASSERT(pm != NULL, 1228 ("[pmc,%d] null pm pointer", __LINE__)); 1229 1230 ri = PMC_TO_ROWINDEX(pm); 1231 1232 PMCDBG6(PRC,ATT,2, "detach-one pm=%p ri=%d proc=%p (%d, %s) flags=0x%x", 1233 pm, ri, p, p->p_pid, p->p_comm, flags); 1234 1235 if ((pp = pmc_find_process_descriptor(p, 0)) == NULL) 1236 return (ESRCH); 1237 1238 if (pp->pp_pmcs[ri].pp_pmc != pm) 1239 return (EINVAL); 1240 1241 pmc_unlink_target_process(pm, pp); 1242 1243 /* Issue a detach entry if a log file is configured */ 1244 if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE) 1245 pmclog_process_pmcdetach(pm, p->p_pid); 1246 1247 /* 1248 * If there are no PMCs targeting this process, we remove its 1249 * descriptor from the target hash table and unset the P_HWPMC 1250 * flag in the struct proc. 1251 */ 1252 KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= (int) md->pmd_npmc, 1253 ("[pmc,%d] Illegal refcnt %d for process struct %p", 1254 __LINE__, pp->pp_refcnt, pp)); 1255 1256 if (pp->pp_refcnt != 0) /* still a target of some PMC */ 1257 return (0); 1258 1259 pmc_remove_process_descriptor(pp); 1260 1261 if (flags & PMC_FLAG_REMOVE) 1262 pmc_destroy_process_descriptor(pp); 1263 1264 PROC_LOCK(p); 1265 p->p_flag &= ~P_HWPMC; 1266 PROC_UNLOCK(p); 1267 1268 return (0); 1269 } 1270 1271 /* 1272 * Detach a process and optionally its descendants from a PMC. 1273 */ 1274 static int 1275 pmc_detach_process(struct proc *p, struct pmc *pm) 1276 { 1277 struct proc *top; 1278 1279 sx_assert(&pmc_sx, SX_XLOCKED); 1280 1281 PMCDBG5(PRC,ATT,1, "detach pm=%p ri=%d proc=%p (%d, %s)", pm, 1282 PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm); 1283 1284 if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0) 1285 return (pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE)); 1286 1287 /* 1288 * Traverse all children, detaching them from this PMC. We 1289 * ignore errors since we could be detaching a PMC from a 1290 * partially attached proc tree. 1291 */ 1292 sx_slock(&proctree_lock); 1293 1294 top = p; 1295 for (;;) { 1296 (void)pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE); 1297 1298 if (!LIST_EMPTY(&p->p_children)) { 1299 p = LIST_FIRST(&p->p_children); 1300 } else { 1301 for (;;) { 1302 if (p == top) 1303 goto done; 1304 if (LIST_NEXT(p, p_sibling)) { 1305 p = LIST_NEXT(p, p_sibling); 1306 break; 1307 } 1308 p = p->p_pptr; 1309 } 1310 } 1311 } 1312 done: 1313 sx_sunlock(&proctree_lock); 1314 if (LIST_EMPTY(&pm->pm_targets)) 1315 pm->pm_flags &= ~PMC_F_ATTACH_DONE; 1316 1317 return (0); 1318 } 1319 1320 /* 1321 * Handle events after an exec() for a process: 1322 * - Inform log owners of the new exec() event 1323 * - Release any PMCs owned by the process before the exec() 1324 * - Detach PMCs from the target if required 1325 */ 1326 static void 1327 pmc_process_exec(struct thread *td, struct pmckern_procexec *pk) 1328 { 1329 struct pmc *pm; 1330 struct pmc_owner *po; 1331 struct pmc_process *pp; 1332 struct proc *p; 1333 char *fullpath, *freepath; 1334 u_int ri; 1335 bool is_using_hwpmcs; 1336 1337 sx_assert(&pmc_sx, SX_XLOCKED); 1338 1339 p = td->td_proc; 1340 pmc_getfilename(p->p_textvp, &fullpath, &freepath); 1341 1342 PMC_EPOCH_ENTER(); 1343 /* Inform owners of SS mode PMCs of the exec event. */ 1344 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) { 1345 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) != 0) { 1346 pmclog_process_procexec(po, PMC_ID_INVALID, p->p_pid, 1347 pk->pm_baseaddr, pk->pm_dynaddr, fullpath); 1348 } 1349 } 1350 PMC_EPOCH_EXIT(); 1351 1352 PROC_LOCK(p); 1353 is_using_hwpmcs = (p->p_flag & P_HWPMC) != 0; 1354 PROC_UNLOCK(p); 1355 1356 if (!is_using_hwpmcs) { 1357 if (freepath != NULL) 1358 free(freepath, M_TEMP); 1359 return; 1360 } 1361 1362 /* 1363 * PMCs are not inherited across an exec(): remove any PMCs that this 1364 * process is the owner of. 1365 */ 1366 if ((po = pmc_find_owner_descriptor(p)) != NULL) { 1367 pmc_remove_owner(po); 1368 pmc_destroy_owner_descriptor(po); 1369 } 1370 1371 /* 1372 * If the process being exec'ed is not the target of any PMC, we are 1373 * done. 1374 */ 1375 if ((pp = pmc_find_process_descriptor(p, 0)) == NULL) { 1376 if (freepath != NULL) 1377 free(freepath, M_TEMP); 1378 return; 1379 } 1380 1381 /* 1382 * Log the exec event to all monitoring owners. Skip owners who have 1383 * already received the event because they had system sampling PMCs 1384 * active. 1385 */ 1386 for (ri = 0; ri < md->pmd_npmc; ri++) { 1387 if ((pm = pp->pp_pmcs[ri].pp_pmc) == NULL) 1388 continue; 1389 1390 po = pm->pm_owner; 1391 if (po->po_sscount == 0 && 1392 (po->po_flags & PMC_PO_OWNS_LOGFILE) != 0) { 1393 pmclog_process_procexec(po, pm->pm_id, p->p_pid, 1394 pk->pm_baseaddr, pk->pm_dynaddr, fullpath); 1395 } 1396 } 1397 1398 if (freepath != NULL) 1399 free(freepath, M_TEMP); 1400 1401 PMCDBG4(PRC,EXC,1, "exec proc=%p (%d, %s) cred-changed=%d", 1402 p, p->p_pid, p->p_comm, pk->pm_credentialschanged); 1403 1404 if (pk->pm_credentialschanged == 0) /* no change */ 1405 return; 1406 1407 /* 1408 * If the newly exec()'ed process has a different credential 1409 * than before, allow it to be the target of a PMC only if 1410 * the PMC's owner has sufficient privilege. 1411 */ 1412 for (ri = 0; ri < md->pmd_npmc; ri++) { 1413 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) { 1414 if (pmc_can_attach(pm, td->td_proc)) { 1415 pmc_detach_one_process(td->td_proc, pm, 1416 PMC_FLAG_NONE); 1417 } 1418 } 1419 } 1420 1421 KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= md->pmd_npmc, 1422 ("[pmc,%d] Illegal ref count %u on pp %p", __LINE__, 1423 pp->pp_refcnt, pp)); 1424 1425 /* 1426 * If this process is no longer the target of any 1427 * PMCs, we can remove the process entry and free 1428 * up space. 1429 */ 1430 if (pp->pp_refcnt == 0) { 1431 pmc_remove_process_descriptor(pp); 1432 pmc_destroy_process_descriptor(pp); 1433 } 1434 } 1435 1436 /* 1437 * Thread context switch IN. 1438 */ 1439 static void 1440 pmc_process_csw_in(struct thread *td) 1441 { 1442 struct pmc *pm; 1443 struct pmc_classdep *pcd; 1444 struct pmc_cpu *pc; 1445 struct pmc_hw *phw __diagused; 1446 struct pmc_process *pp; 1447 struct pmc_thread *pt; 1448 struct proc *p; 1449 pmc_value_t newvalue; 1450 int cpu; 1451 u_int adjri, ri; 1452 1453 p = td->td_proc; 1454 pt = NULL; 1455 if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE)) == NULL) 1456 return; 1457 1458 KASSERT(pp->pp_proc == td->td_proc, 1459 ("[pmc,%d] not my thread state", __LINE__)); 1460 1461 critical_enter(); /* no preemption from this point */ 1462 1463 cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */ 1464 1465 PMCDBG5(CSW,SWI,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p, 1466 p->p_pid, p->p_comm, pp); 1467 1468 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), 1469 ("[pmc,%d] weird CPU id %d", __LINE__, cpu)); 1470 1471 pc = pmc_pcpu[cpu]; 1472 for (ri = 0; ri < md->pmd_npmc; ri++) { 1473 if ((pm = pp->pp_pmcs[ri].pp_pmc) == NULL) 1474 continue; 1475 1476 KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)), 1477 ("[pmc,%d] Target PMC in non-virtual mode (%d)", 1478 __LINE__, PMC_TO_MODE(pm))); 1479 KASSERT(PMC_TO_ROWINDEX(pm) == ri, 1480 ("[pmc,%d] Row index mismatch pmc %d != ri %d", 1481 __LINE__, PMC_TO_ROWINDEX(pm), ri)); 1482 1483 /* 1484 * Only PMCs that are marked as 'RUNNING' need 1485 * be placed on hardware. 1486 */ 1487 if (pm->pm_state != PMC_STATE_RUNNING) 1488 continue; 1489 1490 KASSERT(counter_u64_fetch(pm->pm_runcount) >= 0, 1491 ("[pmc,%d] pm=%p runcount %ju", __LINE__, pm, 1492 (uintmax_t)counter_u64_fetch(pm->pm_runcount))); 1493 1494 /* increment PMC runcount */ 1495 counter_u64_add(pm->pm_runcount, 1); 1496 1497 /* configure the HWPMC we are going to use. */ 1498 pcd = pmc_ri_to_classdep(md, ri, &adjri); 1499 (void)pcd->pcd_config_pmc(cpu, adjri, pm); 1500 1501 phw = pc->pc_hwpmcs[ri]; 1502 1503 KASSERT(phw != NULL, 1504 ("[pmc,%d] null hw pointer", __LINE__)); 1505 1506 KASSERT(phw->phw_pmc == pm, 1507 ("[pmc,%d] hw->pmc %p != pmc %p", __LINE__, 1508 phw->phw_pmc, pm)); 1509 1510 /* 1511 * Write out saved value and start the PMC. 1512 * 1513 * Sampling PMCs use a per-thread value, while 1514 * counting mode PMCs use a per-pmc value that is 1515 * inherited across descendants. 1516 */ 1517 if (PMC_TO_MODE(pm) == PMC_MODE_TS) { 1518 if (pt == NULL) 1519 pt = pmc_find_thread_descriptor(pp, td, 1520 PMC_FLAG_NONE); 1521 1522 KASSERT(pt != NULL, 1523 ("[pmc,%d] No thread found for td=%p", __LINE__, 1524 td)); 1525 1526 mtx_pool_lock_spin(pmc_mtxpool, pm); 1527 1528 /* 1529 * If we have a thread descriptor, use the per-thread 1530 * counter in the descriptor. If not, we will use 1531 * a per-process counter. 1532 * 1533 * TODO: Remove the per-process "safety net" once 1534 * we have thoroughly tested that we don't hit the 1535 * above assert. 1536 */ 1537 if (pt != NULL) { 1538 if (pt->pt_pmcs[ri].pt_pmcval > 0) 1539 newvalue = pt->pt_pmcs[ri].pt_pmcval; 1540 else 1541 newvalue = pm->pm_sc.pm_reloadcount; 1542 } else { 1543 /* 1544 * Use the saved value calculated after the most 1545 * recent time a thread using the shared counter 1546 * switched out. Reset the saved count in case 1547 * another thread from this process switches in 1548 * before any threads switch out. 1549 */ 1550 newvalue = pp->pp_pmcs[ri].pp_pmcval; 1551 pp->pp_pmcs[ri].pp_pmcval = 1552 pm->pm_sc.pm_reloadcount; 1553 } 1554 mtx_pool_unlock_spin(pmc_mtxpool, pm); 1555 KASSERT(newvalue > 0 && newvalue <= 1556 pm->pm_sc.pm_reloadcount, 1557 ("[pmc,%d] pmcval outside of expected range cpu=%d " 1558 "ri=%d pmcval=%jx pm_reloadcount=%jx", __LINE__, 1559 cpu, ri, newvalue, pm->pm_sc.pm_reloadcount)); 1560 } else { 1561 KASSERT(PMC_TO_MODE(pm) == PMC_MODE_TC, 1562 ("[pmc,%d] illegal mode=%d", __LINE__, 1563 PMC_TO_MODE(pm))); 1564 mtx_pool_lock_spin(pmc_mtxpool, pm); 1565 newvalue = PMC_PCPU_SAVED(cpu, ri) = 1566 pm->pm_gv.pm_savedvalue; 1567 mtx_pool_unlock_spin(pmc_mtxpool, pm); 1568 } 1569 1570 PMCDBG3(CSW,SWI,1,"cpu=%d ri=%d new=%jd", cpu, ri, newvalue); 1571 1572 (void)pcd->pcd_write_pmc(cpu, adjri, pm, newvalue); 1573 1574 /* If a sampling mode PMC, reset stalled state. */ 1575 if (PMC_TO_MODE(pm) == PMC_MODE_TS) 1576 pm->pm_pcpu_state[cpu].pps_stalled = 0; 1577 1578 /* Indicate that we desire this to run. */ 1579 pm->pm_pcpu_state[cpu].pps_cpustate = 1; 1580 1581 /* Start the PMC. */ 1582 (void)pcd->pcd_start_pmc(cpu, adjri, pm); 1583 } 1584 1585 /* 1586 * Perform any other architecture/cpu dependent thread 1587 * switch-in actions. 1588 */ 1589 (void)(*md->pmd_switch_in)(pc, pp); 1590 1591 critical_exit(); 1592 } 1593 1594 /* 1595 * Thread context switch OUT. 1596 */ 1597 static void 1598 pmc_process_csw_out(struct thread *td) 1599 { 1600 struct pmc *pm; 1601 struct pmc_classdep *pcd; 1602 struct pmc_cpu *pc; 1603 struct pmc_process *pp; 1604 struct pmc_thread *pt = NULL; 1605 struct proc *p; 1606 pmc_value_t newvalue; 1607 int64_t tmp; 1608 enum pmc_mode mode; 1609 int cpu; 1610 u_int adjri, ri; 1611 1612 /* 1613 * Locate our process descriptor; this may be NULL if 1614 * this process is exiting and we have already removed 1615 * the process from the target process table. 1616 * 1617 * Note that due to kernel preemption, multiple 1618 * context switches may happen while the process is 1619 * exiting. 1620 * 1621 * Note also that if the target process cannot be 1622 * found we still need to deconfigure any PMCs that 1623 * are currently running on hardware. 1624 */ 1625 p = td->td_proc; 1626 pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE); 1627 1628 critical_enter(); 1629 1630 cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */ 1631 1632 PMCDBG5(CSW,SWO,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p, 1633 p->p_pid, p->p_comm, pp); 1634 1635 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), 1636 ("[pmc,%d weird CPU id %d", __LINE__, cpu)); 1637 1638 pc = pmc_pcpu[cpu]; 1639 1640 /* 1641 * When a PMC gets unlinked from a target PMC, it will 1642 * be removed from the target's pp_pmc[] array. 1643 * 1644 * However, on a MP system, the target could have been 1645 * executing on another CPU at the time of the unlink. 1646 * So, at context switch OUT time, we need to look at 1647 * the hardware to determine if a PMC is scheduled on 1648 * it. 1649 */ 1650 for (ri = 0; ri < md->pmd_npmc; ri++) { 1651 pcd = pmc_ri_to_classdep(md, ri, &adjri); 1652 pm = NULL; 1653 (void)(*pcd->pcd_get_config)(cpu, adjri, &pm); 1654 1655 if (pm == NULL) /* nothing at this row index */ 1656 continue; 1657 1658 mode = PMC_TO_MODE(pm); 1659 if (!PMC_IS_VIRTUAL_MODE(mode)) 1660 continue; /* not a process virtual PMC */ 1661 1662 KASSERT(PMC_TO_ROWINDEX(pm) == ri, 1663 ("[pmc,%d] ri mismatch pmc(%d) ri(%d)", 1664 __LINE__, PMC_TO_ROWINDEX(pm), ri)); 1665 1666 /* 1667 * Change desired state, and then stop if not stalled. 1668 * This two-step dance should avoid race conditions where 1669 * an interrupt re-enables the PMC after this code has 1670 * already checked the pm_stalled flag. 1671 */ 1672 pm->pm_pcpu_state[cpu].pps_cpustate = 0; 1673 if (pm->pm_pcpu_state[cpu].pps_stalled == 0) 1674 (void)pcd->pcd_stop_pmc(cpu, adjri, pm); 1675 1676 KASSERT(counter_u64_fetch(pm->pm_runcount) > 0, 1677 ("[pmc,%d] pm=%p runcount %ju", __LINE__, pm, 1678 (uintmax_t)counter_u64_fetch(pm->pm_runcount))); 1679 1680 /* reduce this PMC's runcount */ 1681 counter_u64_add(pm->pm_runcount, -1); 1682 1683 /* 1684 * If this PMC is associated with this process, 1685 * save the reading. 1686 */ 1687 if (pm->pm_state != PMC_STATE_DELETED && pp != NULL && 1688 pp->pp_pmcs[ri].pp_pmc != NULL) { 1689 KASSERT(pm == pp->pp_pmcs[ri].pp_pmc, 1690 ("[pmc,%d] pm %p != pp_pmcs[%d] %p", __LINE__, 1691 pm, ri, pp->pp_pmcs[ri].pp_pmc)); 1692 KASSERT(pp->pp_refcnt > 0, 1693 ("[pmc,%d] pp refcnt = %d", __LINE__, 1694 pp->pp_refcnt)); 1695 1696 (void)pcd->pcd_read_pmc(cpu, adjri, pm, &newvalue); 1697 1698 if (mode == PMC_MODE_TS) { 1699 PMCDBG3(CSW,SWO,1,"cpu=%d ri=%d val=%jd (samp)", 1700 cpu, ri, newvalue); 1701 1702 if (pt == NULL) 1703 pt = pmc_find_thread_descriptor(pp, td, 1704 PMC_FLAG_NONE); 1705 1706 KASSERT(pt != NULL, 1707 ("[pmc,%d] No thread found for td=%p", 1708 __LINE__, td)); 1709 1710 mtx_pool_lock_spin(pmc_mtxpool, pm); 1711 1712 /* 1713 * If we have a thread descriptor, save the 1714 * per-thread counter in the descriptor. If not, 1715 * we will update the per-process counter. 1716 * 1717 * TODO: Remove the per-process "safety net" 1718 * once we have thoroughly tested that we 1719 * don't hit the above assert. 1720 */ 1721 if (pt != NULL) { 1722 pt->pt_pmcs[ri].pt_pmcval = newvalue; 1723 } else { 1724 /* 1725 * For sampling process-virtual PMCs, 1726 * newvalue is the number of events to 1727 * be seen until the next sampling 1728 * interrupt. We can just add the events 1729 * left from this invocation to the 1730 * counter, then adjust in case we 1731 * overflow our range. 1732 * 1733 * (Recall that we reload the counter 1734 * every time we use it.) 1735 */ 1736 pp->pp_pmcs[ri].pp_pmcval += newvalue; 1737 if (pp->pp_pmcs[ri].pp_pmcval > 1738 pm->pm_sc.pm_reloadcount) { 1739 pp->pp_pmcs[ri].pp_pmcval -= 1740 pm->pm_sc.pm_reloadcount; 1741 } 1742 } 1743 mtx_pool_unlock_spin(pmc_mtxpool, pm); 1744 } else { 1745 tmp = newvalue - PMC_PCPU_SAVED(cpu, ri); 1746 1747 PMCDBG3(CSW,SWO,1,"cpu=%d ri=%d tmp=%jd (count)", 1748 cpu, ri, tmp); 1749 1750 /* 1751 * For counting process-virtual PMCs, 1752 * we expect the count to be 1753 * increasing monotonically, modulo a 64 1754 * bit wraparound. 1755 */ 1756 KASSERT(tmp >= 0, 1757 ("[pmc,%d] negative increment cpu=%d " 1758 "ri=%d newvalue=%jx saved=%jx " 1759 "incr=%jx", __LINE__, cpu, ri, 1760 newvalue, PMC_PCPU_SAVED(cpu, ri), tmp)); 1761 1762 mtx_pool_lock_spin(pmc_mtxpool, pm); 1763 pm->pm_gv.pm_savedvalue += tmp; 1764 pp->pp_pmcs[ri].pp_pmcval += tmp; 1765 mtx_pool_unlock_spin(pmc_mtxpool, pm); 1766 1767 if (pm->pm_flags & PMC_F_LOG_PROCCSW) 1768 pmclog_process_proccsw(pm, pp, tmp, td); 1769 } 1770 } 1771 1772 /* Mark hardware as free. */ 1773 (void)pcd->pcd_config_pmc(cpu, adjri, NULL); 1774 } 1775 1776 /* 1777 * Perform any other architecture/cpu dependent thread 1778 * switch out functions. 1779 */ 1780 (void)(*md->pmd_switch_out)(pc, pp); 1781 1782 critical_exit(); 1783 } 1784 1785 /* 1786 * A new thread for a process. 1787 */ 1788 static void 1789 pmc_process_thread_add(struct thread *td) 1790 { 1791 struct pmc_process *pmc; 1792 1793 pmc = pmc_find_process_descriptor(td->td_proc, PMC_FLAG_NONE); 1794 if (pmc != NULL) 1795 pmc_find_thread_descriptor(pmc, td, PMC_FLAG_ALLOCATE); 1796 } 1797 1798 /* 1799 * A thread delete for a process. 1800 */ 1801 static void 1802 pmc_process_thread_delete(struct thread *td) 1803 { 1804 struct pmc_process *pmc; 1805 1806 pmc = pmc_find_process_descriptor(td->td_proc, PMC_FLAG_NONE); 1807 if (pmc != NULL) 1808 pmc_thread_descriptor_pool_free(pmc_find_thread_descriptor(pmc, 1809 td, PMC_FLAG_REMOVE)); 1810 } 1811 1812 /* 1813 * A userret() call for a thread. 1814 */ 1815 static void 1816 pmc_process_thread_userret(struct thread *td) 1817 { 1818 sched_pin(); 1819 pmc_capture_user_callchain(curcpu, PMC_UR, td->td_frame); 1820 sched_unpin(); 1821 } 1822 1823 /* 1824 * A mapping change for a process. 1825 */ 1826 static void 1827 pmc_process_mmap(struct thread *td, struct pmckern_map_in *pkm) 1828 { 1829 const struct pmc *pm; 1830 const struct pmc_process *pp; 1831 struct pmc_owner *po; 1832 char *fullpath, *freepath; 1833 pid_t pid; 1834 int ri; 1835 1836 MPASS(!in_epoch(global_epoch_preempt)); 1837 1838 freepath = fullpath = NULL; 1839 pmc_getfilename((struct vnode *)pkm->pm_file, &fullpath, &freepath); 1840 1841 pid = td->td_proc->p_pid; 1842 1843 PMC_EPOCH_ENTER(); 1844 /* Inform owners of all system-wide sampling PMCs. */ 1845 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) { 1846 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 1847 pmclog_process_map_in(po, pid, pkm->pm_address, 1848 fullpath); 1849 } 1850 1851 if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL) 1852 goto done; 1853 1854 /* 1855 * Inform sampling PMC owners tracking this process. 1856 */ 1857 for (ri = 0; ri < md->pmd_npmc; ri++) { 1858 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL && 1859 PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) { 1860 pmclog_process_map_in(pm->pm_owner, 1861 pid, pkm->pm_address, fullpath); 1862 } 1863 } 1864 1865 done: 1866 if (freepath != NULL) 1867 free(freepath, M_TEMP); 1868 PMC_EPOCH_EXIT(); 1869 } 1870 1871 /* 1872 * Log an munmap request. 1873 */ 1874 static void 1875 pmc_process_munmap(struct thread *td, struct pmckern_map_out *pkm) 1876 { 1877 const struct pmc *pm; 1878 const struct pmc_process *pp; 1879 struct pmc_owner *po; 1880 pid_t pid; 1881 int ri; 1882 1883 pid = td->td_proc->p_pid; 1884 1885 PMC_EPOCH_ENTER(); 1886 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) { 1887 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 1888 pmclog_process_map_out(po, pid, pkm->pm_address, 1889 pkm->pm_address + pkm->pm_size); 1890 } 1891 PMC_EPOCH_EXIT(); 1892 1893 if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL) 1894 return; 1895 1896 for (ri = 0; ri < md->pmd_npmc; ri++) { 1897 pm = pp->pp_pmcs[ri].pp_pmc; 1898 if (pm != NULL && PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) { 1899 pmclog_process_map_out(pm->pm_owner, pid, 1900 pkm->pm_address, pkm->pm_address + pkm->pm_size); 1901 } 1902 } 1903 } 1904 1905 /* 1906 * Log mapping information about the kernel. 1907 */ 1908 static void 1909 pmc_log_kernel_mappings(struct pmc *pm) 1910 { 1911 struct pmc_owner *po; 1912 struct pmckern_map_in *km, *kmbase; 1913 1914 MPASS(in_epoch(global_epoch_preempt) || sx_xlocked(&pmc_sx)); 1915 KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)), 1916 ("[pmc,%d] non-sampling PMC (%p) desires mapping information", 1917 __LINE__, (void *) pm)); 1918 1919 po = pm->pm_owner; 1920 if ((po->po_flags & PMC_PO_INITIAL_MAPPINGS_DONE) != 0) 1921 return; 1922 1923 if (PMC_TO_MODE(pm) == PMC_MODE_SS) 1924 pmc_process_allproc(pm); 1925 1926 /* 1927 * Log the current set of kernel modules. 1928 */ 1929 kmbase = linker_hwpmc_list_objects(); 1930 for (km = kmbase; km->pm_file != NULL; km++) { 1931 PMCDBG2(LOG,REG,1,"%s %p", (char *)km->pm_file, 1932 (void *)km->pm_address); 1933 pmclog_process_map_in(po, (pid_t)-1, km->pm_address, 1934 km->pm_file); 1935 } 1936 free(kmbase, M_LINKER); 1937 1938 po->po_flags |= PMC_PO_INITIAL_MAPPINGS_DONE; 1939 } 1940 1941 /* 1942 * Log the mappings for a single process. 1943 */ 1944 static void 1945 pmc_log_process_mappings(struct pmc_owner *po, struct proc *p) 1946 { 1947 vm_map_t map; 1948 vm_map_entry_t entry; 1949 vm_object_t obj, lobj, tobj; 1950 vm_offset_t last_end; 1951 vm_offset_t start_addr; 1952 struct vnode *vp, *last_vp; 1953 struct vmspace *vm; 1954 char *fullpath, *freepath; 1955 u_int last_timestamp; 1956 1957 last_vp = NULL; 1958 last_end = (vm_offset_t)0; 1959 fullpath = freepath = NULL; 1960 1961 if ((vm = vmspace_acquire_ref(p)) == NULL) 1962 return; 1963 1964 map = &vm->vm_map; 1965 vm_map_lock_read(map); 1966 VM_MAP_ENTRY_FOREACH(entry, map) { 1967 if (entry == NULL) { 1968 PMCDBG2(LOG,OPS,2, "hwpmc: vm_map entry unexpectedly " 1969 "NULL! pid=%d vm_map=%p\n", p->p_pid, map); 1970 break; 1971 } 1972 1973 /* 1974 * We only care about executable map entries. 1975 */ 1976 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0 || 1977 (entry->protection & VM_PROT_EXECUTE) == 0 || 1978 entry->object.vm_object == NULL) { 1979 continue; 1980 } 1981 1982 obj = entry->object.vm_object; 1983 VM_OBJECT_RLOCK(obj); 1984 1985 /* 1986 * Walk the backing_object list to find the base (non-shadowed) 1987 * vm_object. 1988 */ 1989 for (lobj = tobj = obj; tobj != NULL; 1990 tobj = tobj->backing_object) { 1991 if (tobj != obj) 1992 VM_OBJECT_RLOCK(tobj); 1993 if (lobj != obj) 1994 VM_OBJECT_RUNLOCK(lobj); 1995 lobj = tobj; 1996 } 1997 1998 /* 1999 * At this point lobj is the base vm_object and it is locked. 2000 */ 2001 if (lobj == NULL) { 2002 PMCDBG3(LOG,OPS,2, 2003 "hwpmc: lobj unexpectedly NULL! pid=%d " 2004 "vm_map=%p vm_obj=%p\n", p->p_pid, map, obj); 2005 VM_OBJECT_RUNLOCK(obj); 2006 continue; 2007 } 2008 2009 vp = vm_object_vnode(lobj); 2010 if (vp == NULL) { 2011 if (lobj != obj) 2012 VM_OBJECT_RUNLOCK(lobj); 2013 VM_OBJECT_RUNLOCK(obj); 2014 continue; 2015 } 2016 2017 /* 2018 * Skip contiguous regions that point to the same vnode, so we 2019 * don't emit redundant MAP-IN directives. 2020 */ 2021 if (entry->start == last_end && vp == last_vp) { 2022 last_end = entry->end; 2023 if (lobj != obj) 2024 VM_OBJECT_RUNLOCK(lobj); 2025 VM_OBJECT_RUNLOCK(obj); 2026 continue; 2027 } 2028 2029 /* 2030 * We don't want to keep the proc's vm_map or this vm_object 2031 * locked while we walk the pathname, since vn_fullpath() can 2032 * sleep. However, if we drop the lock, it's possible for 2033 * concurrent activity to modify the vm_map list. To protect 2034 * against this, we save the vm_map timestamp before we release 2035 * the lock, and check it after we reacquire the lock below. 2036 */ 2037 start_addr = entry->start; 2038 last_end = entry->end; 2039 last_timestamp = map->timestamp; 2040 vm_map_unlock_read(map); 2041 2042 vref(vp); 2043 if (lobj != obj) 2044 VM_OBJECT_RUNLOCK(lobj); 2045 VM_OBJECT_RUNLOCK(obj); 2046 2047 freepath = NULL; 2048 pmc_getfilename(vp, &fullpath, &freepath); 2049 last_vp = vp; 2050 2051 vrele(vp); 2052 2053 vp = NULL; 2054 pmclog_process_map_in(po, p->p_pid, start_addr, fullpath); 2055 if (freepath != NULL) 2056 free(freepath, M_TEMP); 2057 2058 vm_map_lock_read(map); 2059 2060 /* 2061 * If our saved timestamp doesn't match, this means 2062 * that the vm_map was modified out from under us and 2063 * we can't trust our current "entry" pointer. Do a 2064 * new lookup for this entry. If there is no entry 2065 * for this address range, vm_map_lookup_entry() will 2066 * return the previous one, so we always want to go to 2067 * the next entry on the next loop iteration. 2068 * 2069 * There is an edge condition here that can occur if 2070 * there is no entry at or before this address. In 2071 * this situation, vm_map_lookup_entry returns 2072 * &map->header, which would cause our loop to abort 2073 * without processing the rest of the map. However, 2074 * in practice this will never happen for process 2075 * vm_map. This is because the executable's text 2076 * segment is the first mapping in the proc's address 2077 * space, and this mapping is never removed until the 2078 * process exits, so there will always be a non-header 2079 * entry at or before the requested address for 2080 * vm_map_lookup_entry to return. 2081 */ 2082 if (map->timestamp != last_timestamp) 2083 vm_map_lookup_entry(map, last_end - 1, &entry); 2084 } 2085 2086 vm_map_unlock_read(map); 2087 vmspace_free(vm); 2088 return; 2089 } 2090 2091 /* 2092 * Log mappings for all processes in the system. 2093 */ 2094 static void 2095 pmc_log_all_process_mappings(struct pmc_owner *po) 2096 { 2097 struct proc *p, *top; 2098 2099 sx_assert(&pmc_sx, SX_XLOCKED); 2100 2101 if ((p = pfind(1)) == NULL) 2102 panic("[pmc,%d] Cannot find init", __LINE__); 2103 2104 PROC_UNLOCK(p); 2105 2106 sx_slock(&proctree_lock); 2107 2108 top = p; 2109 for (;;) { 2110 pmc_log_process_mappings(po, p); 2111 if (!LIST_EMPTY(&p->p_children)) 2112 p = LIST_FIRST(&p->p_children); 2113 else for (;;) { 2114 if (p == top) 2115 goto done; 2116 if (LIST_NEXT(p, p_sibling)) { 2117 p = LIST_NEXT(p, p_sibling); 2118 break; 2119 } 2120 p = p->p_pptr; 2121 } 2122 } 2123 done: 2124 sx_sunlock(&proctree_lock); 2125 } 2126 2127 #ifdef HWPMC_DEBUG 2128 const char *pmc_hooknames[] = { 2129 /* these strings correspond to PMC_FN_* in <sys/pmckern.h> */ 2130 "", 2131 "EXEC", 2132 "CSW-IN", 2133 "CSW-OUT", 2134 "SAMPLE", 2135 "UNUSED1", 2136 "UNUSED2", 2137 "MMAP", 2138 "MUNMAP", 2139 "CALLCHAIN-NMI", 2140 "CALLCHAIN-SOFT", 2141 "SOFTSAMPLING", 2142 "THR-CREATE", 2143 "THR-EXIT", 2144 "THR-USERRET", 2145 "THR-CREATE-LOG", 2146 "THR-EXIT-LOG", 2147 "PROC-CREATE-LOG" 2148 }; 2149 #endif 2150 2151 /* 2152 * The 'hook' invoked from the kernel proper 2153 */ 2154 static int 2155 pmc_hook_handler(struct thread *td, int function, void *arg) 2156 { 2157 int cpu; 2158 2159 PMCDBG4(MOD,PMH,1, "hook td=%p func=%d \"%s\" arg=%p", td, function, 2160 pmc_hooknames[function], arg); 2161 2162 switch (function) { 2163 case PMC_FN_PROCESS_EXEC: 2164 pmc_process_exec(td, (struct pmckern_procexec *)arg); 2165 break; 2166 2167 case PMC_FN_CSW_IN: 2168 pmc_process_csw_in(td); 2169 break; 2170 2171 case PMC_FN_CSW_OUT: 2172 pmc_process_csw_out(td); 2173 break; 2174 2175 /* 2176 * Process accumulated PC samples. 2177 * 2178 * This function is expected to be called by hardclock() for 2179 * each CPU that has accumulated PC samples. 2180 * 2181 * This function is to be executed on the CPU whose samples 2182 * are being processed. 2183 */ 2184 case PMC_FN_DO_SAMPLES: 2185 /* 2186 * Clear the cpu specific bit in the CPU mask before 2187 * do the rest of the processing. If the NMI handler 2188 * gets invoked after the "atomic_clear_int()" call 2189 * below but before "pmc_process_samples()" gets 2190 * around to processing the interrupt, then we will 2191 * come back here at the next hardclock() tick (and 2192 * may find nothing to do if "pmc_process_samples()" 2193 * had already processed the interrupt). We don't 2194 * lose the interrupt sample. 2195 */ 2196 DPCPU_SET(pmc_sampled, 0); 2197 cpu = PCPU_GET(cpuid); 2198 pmc_process_samples(cpu, PMC_HR); 2199 pmc_process_samples(cpu, PMC_SR); 2200 pmc_process_samples(cpu, PMC_UR); 2201 break; 2202 2203 case PMC_FN_MMAP: 2204 pmc_process_mmap(td, (struct pmckern_map_in *)arg); 2205 break; 2206 2207 case PMC_FN_MUNMAP: 2208 MPASS(in_epoch(global_epoch_preempt) || sx_xlocked(&pmc_sx)); 2209 pmc_process_munmap(td, (struct pmckern_map_out *)arg); 2210 break; 2211 2212 case PMC_FN_PROC_CREATE_LOG: 2213 pmc_process_proccreate((struct proc *)arg); 2214 break; 2215 2216 case PMC_FN_USER_CALLCHAIN: 2217 /* 2218 * Record a call chain. 2219 */ 2220 KASSERT(td == curthread, ("[pmc,%d] td != curthread", 2221 __LINE__)); 2222 2223 pmc_capture_user_callchain(PCPU_GET(cpuid), PMC_HR, 2224 (struct trapframe *)arg); 2225 2226 KASSERT(td->td_pinned == 1, 2227 ("[pmc,%d] invalid td_pinned value", __LINE__)); 2228 sched_unpin(); /* Can migrate safely now. */ 2229 2230 td->td_pflags &= ~TDP_CALLCHAIN; 2231 break; 2232 2233 case PMC_FN_USER_CALLCHAIN_SOFT: 2234 /* 2235 * Record a call chain. 2236 */ 2237 KASSERT(td == curthread, ("[pmc,%d] td != curthread", 2238 __LINE__)); 2239 2240 cpu = PCPU_GET(cpuid); 2241 pmc_capture_user_callchain(cpu, PMC_SR, 2242 (struct trapframe *) arg); 2243 2244 KASSERT(td->td_pinned == 1, 2245 ("[pmc,%d] invalid td_pinned value", __LINE__)); 2246 2247 sched_unpin(); /* Can migrate safely now. */ 2248 2249 td->td_pflags &= ~TDP_CALLCHAIN; 2250 break; 2251 2252 case PMC_FN_SOFT_SAMPLING: 2253 /* 2254 * Call soft PMC sampling intr. 2255 */ 2256 pmc_soft_intr((struct pmckern_soft *)arg); 2257 break; 2258 2259 case PMC_FN_THR_CREATE: 2260 pmc_process_thread_add(td); 2261 pmc_process_threadcreate(td); 2262 break; 2263 2264 case PMC_FN_THR_CREATE_LOG: 2265 pmc_process_threadcreate(td); 2266 break; 2267 2268 case PMC_FN_THR_EXIT: 2269 KASSERT(td == curthread, ("[pmc,%d] td != curthread", 2270 __LINE__)); 2271 pmc_process_thread_delete(td); 2272 pmc_process_threadexit(td); 2273 break; 2274 case PMC_FN_THR_EXIT_LOG: 2275 pmc_process_threadexit(td); 2276 break; 2277 case PMC_FN_THR_USERRET: 2278 KASSERT(td == curthread, ("[pmc,%d] td != curthread", 2279 __LINE__)); 2280 pmc_process_thread_userret(td); 2281 break; 2282 default: 2283 #ifdef HWPMC_DEBUG 2284 KASSERT(0, ("[pmc,%d] unknown hook %d\n", __LINE__, function)); 2285 #endif 2286 break; 2287 } 2288 2289 return (0); 2290 } 2291 2292 /* 2293 * Allocate a 'struct pmc_owner' descriptor in the owner hash table. 2294 */ 2295 static struct pmc_owner * 2296 pmc_allocate_owner_descriptor(struct proc *p) 2297 { 2298 struct pmc_owner *po; 2299 struct pmc_ownerhash *poh; 2300 uint32_t hindex; 2301 2302 hindex = PMC_HASH_PTR(p, pmc_ownerhashmask); 2303 poh = &pmc_ownerhash[hindex]; 2304 2305 /* Allocate space for N pointers and one descriptor struct. */ 2306 po = malloc(sizeof(struct pmc_owner), M_PMC, M_WAITOK | M_ZERO); 2307 po->po_owner = p; 2308 LIST_INSERT_HEAD(poh, po, po_next); /* insert into hash table */ 2309 2310 TAILQ_INIT(&po->po_logbuffers); 2311 mtx_init(&po->po_mtx, "pmc-owner-mtx", "pmc-per-proc", MTX_SPIN); 2312 2313 PMCDBG4(OWN,ALL,1, "allocate-owner proc=%p (%d, %s) pmc-owner=%p", 2314 p, p->p_pid, p->p_comm, po); 2315 2316 return (po); 2317 } 2318 2319 static void 2320 pmc_destroy_owner_descriptor(struct pmc_owner *po) 2321 { 2322 2323 PMCDBG4(OWN,REL,1, "destroy-owner po=%p proc=%p (%d, %s)", 2324 po, po->po_owner, po->po_owner->p_pid, po->po_owner->p_comm); 2325 2326 mtx_destroy(&po->po_mtx); 2327 free(po, M_PMC); 2328 } 2329 2330 /* 2331 * Allocate a thread descriptor from the free pool. 2332 * 2333 * NOTE: This *can* return NULL. 2334 */ 2335 static struct pmc_thread * 2336 pmc_thread_descriptor_pool_alloc(void) 2337 { 2338 struct pmc_thread *pt; 2339 2340 mtx_lock_spin(&pmc_threadfreelist_mtx); 2341 if ((pt = LIST_FIRST(&pmc_threadfreelist)) != NULL) { 2342 LIST_REMOVE(pt, pt_next); 2343 pmc_threadfreelist_entries--; 2344 } 2345 mtx_unlock_spin(&pmc_threadfreelist_mtx); 2346 2347 return (pt); 2348 } 2349 2350 /* 2351 * Add a thread descriptor to the free pool. We use this instead of free() 2352 * to maintain a cache of free entries. Additionally, we can safely call 2353 * this function when we cannot call free(), such as in a critical section. 2354 */ 2355 static void 2356 pmc_thread_descriptor_pool_free(struct pmc_thread *pt) 2357 { 2358 2359 if (pt == NULL) 2360 return; 2361 2362 memset(pt, 0, THREADENTRY_SIZE); 2363 mtx_lock_spin(&pmc_threadfreelist_mtx); 2364 LIST_INSERT_HEAD(&pmc_threadfreelist, pt, pt_next); 2365 pmc_threadfreelist_entries++; 2366 if (pmc_threadfreelist_entries > pmc_threadfreelist_max) 2367 taskqueue_enqueue(taskqueue_fast, &free_task); 2368 mtx_unlock_spin(&pmc_threadfreelist_mtx); 2369 } 2370 2371 /* 2372 * An asynchronous task to manage the free list. 2373 */ 2374 static void 2375 pmc_thread_descriptor_pool_free_task(void *arg __unused, int pending __unused) 2376 { 2377 struct pmc_thread *pt; 2378 LIST_HEAD(, pmc_thread) tmplist; 2379 int delta; 2380 2381 LIST_INIT(&tmplist); 2382 2383 /* Determine what changes, if any, we need to make. */ 2384 mtx_lock_spin(&pmc_threadfreelist_mtx); 2385 delta = pmc_threadfreelist_entries - pmc_threadfreelist_max; 2386 while (delta > 0 && (pt = LIST_FIRST(&pmc_threadfreelist)) != NULL) { 2387 delta--; 2388 pmc_threadfreelist_entries--; 2389 LIST_REMOVE(pt, pt_next); 2390 LIST_INSERT_HEAD(&tmplist, pt, pt_next); 2391 } 2392 mtx_unlock_spin(&pmc_threadfreelist_mtx); 2393 2394 /* If there are entries to free, free them. */ 2395 while (!LIST_EMPTY(&tmplist)) { 2396 pt = LIST_FIRST(&tmplist); 2397 LIST_REMOVE(pt, pt_next); 2398 free(pt, M_PMC); 2399 } 2400 } 2401 2402 /* 2403 * Drain the thread free pool, freeing all allocations. 2404 */ 2405 static void 2406 pmc_thread_descriptor_pool_drain(void) 2407 { 2408 struct pmc_thread *pt, *next; 2409 2410 LIST_FOREACH_SAFE(pt, &pmc_threadfreelist, pt_next, next) { 2411 LIST_REMOVE(pt, pt_next); 2412 free(pt, M_PMC); 2413 } 2414 } 2415 2416 /* 2417 * find the descriptor corresponding to thread 'td', adding or removing it 2418 * as specified by 'mode'. 2419 * 2420 * Note that this supports additional mode flags in addition to those 2421 * supported by pmc_find_process_descriptor(): 2422 * PMC_FLAG_NOWAIT: Causes the function to not wait for mallocs. 2423 * This makes it safe to call while holding certain other locks. 2424 */ 2425 static struct pmc_thread * 2426 pmc_find_thread_descriptor(struct pmc_process *pp, struct thread *td, 2427 uint32_t mode) 2428 { 2429 struct pmc_thread *pt = NULL, *ptnew = NULL; 2430 int wait_flag; 2431 2432 KASSERT(td != NULL, ("[pmc,%d] called to add NULL td", __LINE__)); 2433 2434 /* 2435 * Pre-allocate memory in the PMC_FLAG_ALLOCATE case prior to 2436 * acquiring the lock. 2437 */ 2438 if ((mode & PMC_FLAG_ALLOCATE) != 0) { 2439 if ((ptnew = pmc_thread_descriptor_pool_alloc()) == NULL) { 2440 wait_flag = M_WAITOK; 2441 if ((mode & PMC_FLAG_NOWAIT) != 0 || 2442 in_epoch(global_epoch_preempt)) 2443 wait_flag = M_NOWAIT; 2444 2445 ptnew = malloc(THREADENTRY_SIZE, M_PMC, 2446 wait_flag | M_ZERO); 2447 } 2448 } 2449 2450 mtx_lock_spin(pp->pp_tdslock); 2451 LIST_FOREACH(pt, &pp->pp_tds, pt_next) { 2452 if (pt->pt_td == td) 2453 break; 2454 } 2455 2456 if ((mode & PMC_FLAG_REMOVE) != 0 && pt != NULL) 2457 LIST_REMOVE(pt, pt_next); 2458 2459 if ((mode & PMC_FLAG_ALLOCATE) != 0 && pt == NULL && ptnew != NULL) { 2460 pt = ptnew; 2461 ptnew = NULL; 2462 pt->pt_td = td; 2463 LIST_INSERT_HEAD(&pp->pp_tds, pt, pt_next); 2464 } 2465 2466 mtx_unlock_spin(pp->pp_tdslock); 2467 2468 if (ptnew != NULL) { 2469 free(ptnew, M_PMC); 2470 } 2471 2472 return (pt); 2473 } 2474 2475 /* 2476 * Try to add thread descriptors for each thread in a process. 2477 */ 2478 static void 2479 pmc_add_thread_descriptors_from_proc(struct proc *p, struct pmc_process *pp) 2480 { 2481 struct pmc_thread **tdlist; 2482 struct thread *curtd; 2483 int i, tdcnt, tdlistsz; 2484 2485 KASSERT(!PROC_LOCKED(p), ("[pmc,%d] proc unexpectedly locked", 2486 __LINE__)); 2487 tdcnt = 32; 2488 restart: 2489 tdlistsz = roundup2(tdcnt, 32); 2490 2491 tdcnt = 0; 2492 tdlist = malloc(sizeof(struct pmc_thread *) * tdlistsz, M_TEMP, 2493 M_WAITOK); 2494 2495 PROC_LOCK(p); 2496 FOREACH_THREAD_IN_PROC(p, curtd) 2497 tdcnt++; 2498 if (tdcnt >= tdlistsz) { 2499 PROC_UNLOCK(p); 2500 free(tdlist, M_TEMP); 2501 goto restart; 2502 } 2503 2504 /* 2505 * Try to add each thread to the list without sleeping. If unable, 2506 * add to a queue to retry after dropping the process lock. 2507 */ 2508 tdcnt = 0; 2509 FOREACH_THREAD_IN_PROC(p, curtd) { 2510 tdlist[tdcnt] = pmc_find_thread_descriptor(pp, curtd, 2511 PMC_FLAG_ALLOCATE | PMC_FLAG_NOWAIT); 2512 if (tdlist[tdcnt] == NULL) { 2513 PROC_UNLOCK(p); 2514 for (i = 0; i <= tdcnt; i++) 2515 pmc_thread_descriptor_pool_free(tdlist[i]); 2516 free(tdlist, M_TEMP); 2517 goto restart; 2518 } 2519 tdcnt++; 2520 } 2521 PROC_UNLOCK(p); 2522 free(tdlist, M_TEMP); 2523 } 2524 2525 /* 2526 * Find the descriptor corresponding to process 'p', adding or removing it 2527 * as specified by 'mode'. 2528 */ 2529 static struct pmc_process * 2530 pmc_find_process_descriptor(struct proc *p, uint32_t mode) 2531 { 2532 struct pmc_process *pp, *ppnew; 2533 struct pmc_processhash *pph; 2534 uint32_t hindex; 2535 2536 hindex = PMC_HASH_PTR(p, pmc_processhashmask); 2537 pph = &pmc_processhash[hindex]; 2538 2539 ppnew = NULL; 2540 2541 /* 2542 * Pre-allocate memory in the PMC_FLAG_ALLOCATE case since we 2543 * cannot call malloc(9) once we hold a spin lock. 2544 */ 2545 if ((mode & PMC_FLAG_ALLOCATE) != 0) 2546 ppnew = malloc(sizeof(struct pmc_process) + md->pmd_npmc * 2547 sizeof(struct pmc_targetstate), M_PMC, M_WAITOK | M_ZERO); 2548 2549 mtx_lock_spin(&pmc_processhash_mtx); 2550 LIST_FOREACH(pp, pph, pp_next) { 2551 if (pp->pp_proc == p) 2552 break; 2553 } 2554 2555 if ((mode & PMC_FLAG_REMOVE) != 0 && pp != NULL) 2556 LIST_REMOVE(pp, pp_next); 2557 2558 if ((mode & PMC_FLAG_ALLOCATE) != 0 && pp == NULL && ppnew != NULL) { 2559 ppnew->pp_proc = p; 2560 LIST_INIT(&ppnew->pp_tds); 2561 ppnew->pp_tdslock = mtx_pool_find(pmc_mtxpool, ppnew); 2562 LIST_INSERT_HEAD(pph, ppnew, pp_next); 2563 mtx_unlock_spin(&pmc_processhash_mtx); 2564 pp = ppnew; 2565 ppnew = NULL; 2566 2567 /* Add thread descriptors for this process' current threads. */ 2568 pmc_add_thread_descriptors_from_proc(p, pp); 2569 } else 2570 mtx_unlock_spin(&pmc_processhash_mtx); 2571 2572 if (ppnew != NULL) 2573 free(ppnew, M_PMC); 2574 return (pp); 2575 } 2576 2577 /* 2578 * Remove a process descriptor from the process hash table. 2579 */ 2580 static void 2581 pmc_remove_process_descriptor(struct pmc_process *pp) 2582 { 2583 KASSERT(pp->pp_refcnt == 0, 2584 ("[pmc,%d] Removing process descriptor %p with count %d", 2585 __LINE__, pp, pp->pp_refcnt)); 2586 2587 mtx_lock_spin(&pmc_processhash_mtx); 2588 LIST_REMOVE(pp, pp_next); 2589 mtx_unlock_spin(&pmc_processhash_mtx); 2590 } 2591 2592 /* 2593 * Destroy a process descriptor. 2594 */ 2595 static void 2596 pmc_destroy_process_descriptor(struct pmc_process *pp) 2597 { 2598 struct pmc_thread *pmc_td; 2599 2600 while ((pmc_td = LIST_FIRST(&pp->pp_tds)) != NULL) { 2601 LIST_REMOVE(pmc_td, pt_next); 2602 pmc_thread_descriptor_pool_free(pmc_td); 2603 } 2604 free(pp, M_PMC); 2605 } 2606 2607 /* 2608 * Find an owner descriptor corresponding to proc 'p'. 2609 */ 2610 static struct pmc_owner * 2611 pmc_find_owner_descriptor(struct proc *p) 2612 { 2613 struct pmc_owner *po; 2614 struct pmc_ownerhash *poh; 2615 uint32_t hindex; 2616 2617 hindex = PMC_HASH_PTR(p, pmc_ownerhashmask); 2618 poh = &pmc_ownerhash[hindex]; 2619 2620 po = NULL; 2621 LIST_FOREACH(po, poh, po_next) { 2622 if (po->po_owner == p) 2623 break; 2624 } 2625 2626 PMCDBG5(OWN,FND,1, "find-owner proc=%p (%d, %s) hindex=0x%x -> " 2627 "pmc-owner=%p", p, p->p_pid, p->p_comm, hindex, po); 2628 2629 return (po); 2630 } 2631 2632 /* 2633 * Allocate a pmc descriptor and initialize its fields. 2634 */ 2635 static struct pmc * 2636 pmc_allocate_pmc_descriptor(void) 2637 { 2638 struct pmc *pmc; 2639 2640 pmc = malloc(sizeof(struct pmc), M_PMC, M_WAITOK | M_ZERO); 2641 pmc->pm_runcount = counter_u64_alloc(M_WAITOK); 2642 pmc->pm_pcpu_state = malloc(sizeof(struct pmc_pcpu_state) * mp_ncpus, 2643 M_PMC, M_WAITOK | M_ZERO); 2644 PMCDBG1(PMC,ALL,1, "allocate-pmc -> pmc=%p", pmc); 2645 2646 return (pmc); 2647 } 2648 2649 /* 2650 * Destroy a pmc descriptor. 2651 */ 2652 static void 2653 pmc_destroy_pmc_descriptor(struct pmc *pm) 2654 { 2655 2656 KASSERT(pm->pm_state == PMC_STATE_DELETED || 2657 pm->pm_state == PMC_STATE_FREE, 2658 ("[pmc,%d] destroying non-deleted PMC", __LINE__)); 2659 KASSERT(LIST_EMPTY(&pm->pm_targets), 2660 ("[pmc,%d] destroying pmc with targets", __LINE__)); 2661 KASSERT(pm->pm_owner == NULL, 2662 ("[pmc,%d] destroying pmc attached to an owner", __LINE__)); 2663 KASSERT(counter_u64_fetch(pm->pm_runcount) == 0, 2664 ("[pmc,%d] pmc has non-zero run count %ju", __LINE__, 2665 (uintmax_t)counter_u64_fetch(pm->pm_runcount))); 2666 2667 counter_u64_free(pm->pm_runcount); 2668 free(pm->pm_pcpu_state, M_PMC); 2669 free(pm, M_PMC); 2670 } 2671 2672 static void 2673 pmc_wait_for_pmc_idle(struct pmc *pm) 2674 { 2675 #ifdef INVARIANTS 2676 volatile int maxloop; 2677 2678 maxloop = 100 * pmc_cpu_max(); 2679 #endif 2680 /* 2681 * Loop (with a forced context switch) till the PMC's runcount 2682 * comes down to zero. 2683 */ 2684 pmclog_flush(pm->pm_owner, 1); 2685 while (counter_u64_fetch(pm->pm_runcount) > 0) { 2686 pmclog_flush(pm->pm_owner, 1); 2687 #ifdef INVARIANTS 2688 maxloop--; 2689 KASSERT(maxloop > 0, 2690 ("[pmc,%d] (ri%d, rc%ju) waiting too long for " 2691 "pmc to be free", __LINE__, PMC_TO_ROWINDEX(pm), 2692 (uintmax_t)counter_u64_fetch(pm->pm_runcount))); 2693 #endif 2694 pmc_force_context_switch(); 2695 } 2696 } 2697 2698 /* 2699 * This function does the following things: 2700 * 2701 * - detaches the PMC from hardware 2702 * - unlinks all target threads that were attached to it 2703 * - removes the PMC from its owner's list 2704 * - destroys the PMC private mutex 2705 * 2706 * Once this function completes, the given pmc pointer can be freed by 2707 * calling pmc_destroy_pmc_descriptor(). 2708 */ 2709 static void 2710 pmc_release_pmc_descriptor(struct pmc *pm) 2711 { 2712 struct pmc_binding pb; 2713 struct pmc_classdep *pcd; 2714 struct pmc_hw *phw __diagused; 2715 struct pmc_owner *po; 2716 struct pmc_process *pp; 2717 struct pmc_target *ptgt, *tmp; 2718 enum pmc_mode mode; 2719 u_int adjri, ri, cpu; 2720 2721 sx_assert(&pmc_sx, SX_XLOCKED); 2722 KASSERT(pm, ("[pmc,%d] null pmc", __LINE__)); 2723 2724 ri = PMC_TO_ROWINDEX(pm); 2725 pcd = pmc_ri_to_classdep(md, ri, &adjri); 2726 mode = PMC_TO_MODE(pm); 2727 2728 PMCDBG3(PMC,REL,1, "release-pmc pmc=%p ri=%d mode=%d", pm, ri, 2729 mode); 2730 2731 /* 2732 * First, we take the PMC off hardware. 2733 */ 2734 cpu = 0; 2735 if (PMC_IS_SYSTEM_MODE(mode)) { 2736 /* 2737 * A system mode PMC runs on a specific CPU. Switch 2738 * to this CPU and turn hardware off. 2739 */ 2740 pmc_save_cpu_binding(&pb); 2741 cpu = PMC_TO_CPU(pm); 2742 pmc_select_cpu(cpu); 2743 2744 /* switch off non-stalled CPUs */ 2745 pm->pm_pcpu_state[cpu].pps_cpustate = 0; 2746 if (pm->pm_state == PMC_STATE_RUNNING && 2747 pm->pm_pcpu_state[cpu].pps_stalled == 0) { 2748 2749 phw = pmc_pcpu[cpu]->pc_hwpmcs[ri]; 2750 2751 KASSERT(phw->phw_pmc == pm, 2752 ("[pmc, %d] pmc ptr ri(%d) hw(%p) pm(%p)", 2753 __LINE__, ri, phw->phw_pmc, pm)); 2754 PMCDBG2(PMC,REL,2, "stopping cpu=%d ri=%d", cpu, ri); 2755 2756 critical_enter(); 2757 (void)pcd->pcd_stop_pmc(cpu, adjri, pm); 2758 critical_exit(); 2759 } 2760 2761 PMCDBG2(PMC,REL,2, "decfg cpu=%d ri=%d", cpu, ri); 2762 2763 critical_enter(); 2764 (void)pcd->pcd_config_pmc(cpu, adjri, NULL); 2765 critical_exit(); 2766 2767 /* adjust the global and process count of SS mode PMCs */ 2768 if (mode == PMC_MODE_SS && pm->pm_state == PMC_STATE_RUNNING) { 2769 po = pm->pm_owner; 2770 po->po_sscount--; 2771 if (po->po_sscount == 0) { 2772 atomic_subtract_rel_int(&pmc_ss_count, 1); 2773 CK_LIST_REMOVE(po, po_ssnext); 2774 epoch_wait_preempt(global_epoch_preempt); 2775 } 2776 } 2777 pm->pm_state = PMC_STATE_DELETED; 2778 2779 pmc_restore_cpu_binding(&pb); 2780 2781 /* 2782 * We could have references to this PMC structure in the 2783 * per-cpu sample queues. Wait for the queue to drain. 2784 */ 2785 pmc_wait_for_pmc_idle(pm); 2786 2787 } else if (PMC_IS_VIRTUAL_MODE(mode)) { 2788 /* 2789 * A virtual PMC could be running on multiple CPUs at a given 2790 * instant. 2791 * 2792 * By marking its state as DELETED, we ensure that this PMC is 2793 * never further scheduled on hardware. 2794 * 2795 * Then we wait till all CPUs are done with this PMC. 2796 */ 2797 pm->pm_state = PMC_STATE_DELETED; 2798 2799 /* Wait for the PMCs runcount to come to zero. */ 2800 pmc_wait_for_pmc_idle(pm); 2801 2802 /* 2803 * At this point the PMC is off all CPUs and cannot be freshly 2804 * scheduled onto a CPU. It is now safe to unlink all targets 2805 * from this PMC. If a process-record's refcount falls to zero, 2806 * we remove it from the hash table. The module-wide SX lock 2807 * protects us from races. 2808 */ 2809 LIST_FOREACH_SAFE(ptgt, &pm->pm_targets, pt_next, tmp) { 2810 pp = ptgt->pt_process; 2811 pmc_unlink_target_process(pm, pp); /* frees 'ptgt' */ 2812 2813 PMCDBG1(PMC,REL,3, "pp->refcnt=%d", pp->pp_refcnt); 2814 2815 /* 2816 * If the target process record shows that no PMCs are 2817 * attached to it, reclaim its space. 2818 */ 2819 if (pp->pp_refcnt == 0) { 2820 pmc_remove_process_descriptor(pp); 2821 pmc_destroy_process_descriptor(pp); 2822 } 2823 } 2824 2825 cpu = curthread->td_oncpu; /* setup cpu for pmd_release() */ 2826 } 2827 2828 /* 2829 * Release any MD resources. 2830 */ 2831 (void)pcd->pcd_release_pmc(cpu, adjri, pm); 2832 2833 /* 2834 * Update row disposition. 2835 */ 2836 if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm))) 2837 PMC_UNMARK_ROW_STANDALONE(ri); 2838 else 2839 PMC_UNMARK_ROW_THREAD(ri); 2840 2841 /* Unlink from the owner's list. */ 2842 if (pm->pm_owner != NULL) { 2843 LIST_REMOVE(pm, pm_next); 2844 pm->pm_owner = NULL; 2845 } 2846 } 2847 2848 /* 2849 * Register an owner and a pmc. 2850 */ 2851 static int 2852 pmc_register_owner(struct proc *p, struct pmc *pmc) 2853 { 2854 struct pmc_owner *po; 2855 2856 sx_assert(&pmc_sx, SX_XLOCKED); 2857 2858 if ((po = pmc_find_owner_descriptor(p)) == NULL) { 2859 if ((po = pmc_allocate_owner_descriptor(p)) == NULL) 2860 return (ENOMEM); 2861 } 2862 2863 KASSERT(pmc->pm_owner == NULL, 2864 ("[pmc,%d] attempting to own an initialized PMC", __LINE__)); 2865 pmc->pm_owner = po; 2866 2867 LIST_INSERT_HEAD(&po->po_pmcs, pmc, pm_next); 2868 2869 PROC_LOCK(p); 2870 p->p_flag |= P_HWPMC; 2871 PROC_UNLOCK(p); 2872 2873 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) != 0) 2874 pmclog_process_pmcallocate(pmc); 2875 2876 PMCDBG2(PMC,REG,1, "register-owner pmc-owner=%p pmc=%p", 2877 po, pmc); 2878 2879 return (0); 2880 } 2881 2882 /* 2883 * Return the current row disposition: 2884 * == 0 => FREE 2885 * > 0 => PROCESS MODE 2886 * < 0 => SYSTEM MODE 2887 */ 2888 int 2889 pmc_getrowdisp(int ri) 2890 { 2891 return (pmc_pmcdisp[ri]); 2892 } 2893 2894 /* 2895 * Check if a PMC at row index 'ri' can be allocated to the current 2896 * process. 2897 * 2898 * Allocation can fail if: 2899 * - the current process is already being profiled by a PMC at index 'ri', 2900 * attached to it via OP_PMCATTACH. 2901 * - the current process has already allocated a PMC at index 'ri' 2902 * via OP_ALLOCATE. 2903 */ 2904 static bool 2905 pmc_can_allocate_rowindex(struct proc *p, unsigned int ri, int cpu) 2906 { 2907 struct pmc *pm; 2908 struct pmc_owner *po; 2909 struct pmc_process *pp; 2910 enum pmc_mode mode; 2911 2912 PMCDBG5(PMC,ALR,1, "can-allocate-rowindex proc=%p (%d, %s) ri=%d " 2913 "cpu=%d", p, p->p_pid, p->p_comm, ri, cpu); 2914 2915 /* 2916 * We shouldn't have already allocated a process-mode PMC at 2917 * row index 'ri'. 2918 * 2919 * We shouldn't have allocated a system-wide PMC on the same 2920 * CPU and same RI. 2921 */ 2922 if ((po = pmc_find_owner_descriptor(p)) != NULL) { 2923 LIST_FOREACH(pm, &po->po_pmcs, pm_next) { 2924 if (PMC_TO_ROWINDEX(pm) == ri) { 2925 mode = PMC_TO_MODE(pm); 2926 if (PMC_IS_VIRTUAL_MODE(mode)) 2927 return (false); 2928 if (PMC_IS_SYSTEM_MODE(mode) && 2929 PMC_TO_CPU(pm) == cpu) 2930 return (false); 2931 } 2932 } 2933 } 2934 2935 /* 2936 * We also shouldn't be the target of any PMC at this index 2937 * since otherwise a PMC_ATTACH to ourselves will fail. 2938 */ 2939 if ((pp = pmc_find_process_descriptor(p, 0)) != NULL) 2940 if (pp->pp_pmcs[ri].pp_pmc != NULL) 2941 return (false); 2942 2943 PMCDBG4(PMC,ALR,2, "can-allocate-rowindex proc=%p (%d, %s) ri=%d ok", 2944 p, p->p_pid, p->p_comm, ri); 2945 return (true); 2946 } 2947 2948 /* 2949 * Check if a given PMC at row index 'ri' can be currently used in 2950 * mode 'mode'. 2951 */ 2952 static bool 2953 pmc_can_allocate_row(int ri, enum pmc_mode mode) 2954 { 2955 enum pmc_disp disp; 2956 2957 sx_assert(&pmc_sx, SX_XLOCKED); 2958 2959 PMCDBG2(PMC,ALR,1, "can-allocate-row ri=%d mode=%d", ri, mode); 2960 2961 if (PMC_IS_SYSTEM_MODE(mode)) 2962 disp = PMC_DISP_STANDALONE; 2963 else 2964 disp = PMC_DISP_THREAD; 2965 2966 /* 2967 * check disposition for PMC row 'ri': 2968 * 2969 * Expected disposition Row-disposition Result 2970 * 2971 * STANDALONE STANDALONE or FREE proceed 2972 * STANDALONE THREAD fail 2973 * THREAD THREAD or FREE proceed 2974 * THREAD STANDALONE fail 2975 */ 2976 if (!PMC_ROW_DISP_IS_FREE(ri) && 2977 !(disp == PMC_DISP_THREAD && PMC_ROW_DISP_IS_THREAD(ri)) && 2978 !(disp == PMC_DISP_STANDALONE && PMC_ROW_DISP_IS_STANDALONE(ri))) 2979 return (false); 2980 2981 /* 2982 * All OK 2983 */ 2984 PMCDBG2(PMC,ALR,2, "can-allocate-row ri=%d mode=%d ok", ri, mode); 2985 return (true); 2986 } 2987 2988 /* 2989 * Find a PMC descriptor with user handle 'pmcid' for thread 'td'. 2990 */ 2991 static struct pmc * 2992 pmc_find_pmc_descriptor_in_process(struct pmc_owner *po, pmc_id_t pmcid) 2993 { 2994 struct pmc *pm; 2995 2996 KASSERT(PMC_ID_TO_ROWINDEX(pmcid) < md->pmd_npmc, 2997 ("[pmc,%d] Illegal pmc index %d (max %d)", __LINE__, 2998 PMC_ID_TO_ROWINDEX(pmcid), md->pmd_npmc)); 2999 3000 LIST_FOREACH(pm, &po->po_pmcs, pm_next) { 3001 if (pm->pm_id == pmcid) 3002 return (pm); 3003 } 3004 3005 return (NULL); 3006 } 3007 3008 static int 3009 pmc_find_pmc(pmc_id_t pmcid, struct pmc **pmc) 3010 { 3011 struct pmc *pm, *opm; 3012 struct pmc_owner *po; 3013 struct pmc_process *pp; 3014 3015 PMCDBG1(PMC,FND,1, "find-pmc id=%d", pmcid); 3016 if (PMC_ID_TO_ROWINDEX(pmcid) >= md->pmd_npmc) 3017 return (EINVAL); 3018 3019 if ((po = pmc_find_owner_descriptor(curthread->td_proc)) == NULL) { 3020 /* 3021 * In case of PMC_F_DESCENDANTS child processes we will not find 3022 * the current process in the owners hash list. Find the owner 3023 * process first and from there lookup the po. 3024 */ 3025 pp = pmc_find_process_descriptor(curthread->td_proc, 3026 PMC_FLAG_NONE); 3027 if (pp == NULL) 3028 return (ESRCH); 3029 opm = pp->pp_pmcs[PMC_ID_TO_ROWINDEX(pmcid)].pp_pmc; 3030 if (opm == NULL) 3031 return (ESRCH); 3032 if ((opm->pm_flags & 3033 (PMC_F_ATTACHED_TO_OWNER | PMC_F_DESCENDANTS)) != 3034 (PMC_F_ATTACHED_TO_OWNER | PMC_F_DESCENDANTS)) 3035 return (ESRCH); 3036 3037 po = opm->pm_owner; 3038 } 3039 3040 if ((pm = pmc_find_pmc_descriptor_in_process(po, pmcid)) == NULL) 3041 return (EINVAL); 3042 3043 PMCDBG2(PMC,FND,2, "find-pmc id=%d -> pmc=%p", pmcid, pm); 3044 3045 *pmc = pm; 3046 return (0); 3047 } 3048 3049 /* 3050 * Start a PMC. 3051 */ 3052 static int 3053 pmc_start(struct pmc *pm) 3054 { 3055 struct pmc_binding pb; 3056 struct pmc_classdep *pcd; 3057 struct pmc_owner *po; 3058 pmc_value_t v; 3059 enum pmc_mode mode; 3060 int adjri, error, cpu, ri; 3061 3062 KASSERT(pm != NULL, 3063 ("[pmc,%d] null pm", __LINE__)); 3064 3065 mode = PMC_TO_MODE(pm); 3066 ri = PMC_TO_ROWINDEX(pm); 3067 pcd = pmc_ri_to_classdep(md, ri, &adjri); 3068 3069 error = 0; 3070 po = pm->pm_owner; 3071 3072 PMCDBG3(PMC,OPS,1, "start pmc=%p mode=%d ri=%d", pm, mode, ri); 3073 3074 po = pm->pm_owner; 3075 3076 /* 3077 * Disallow PMCSTART if a logfile is required but has not been 3078 * configured yet. 3079 */ 3080 if ((pm->pm_flags & PMC_F_NEEDS_LOGFILE) != 0 && 3081 (po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) 3082 return (EDOOFUS); /* programming error */ 3083 3084 /* 3085 * If this is a sampling mode PMC, log mapping information for 3086 * the kernel modules that are currently loaded. 3087 */ 3088 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) 3089 pmc_log_kernel_mappings(pm); 3090 3091 if (PMC_IS_VIRTUAL_MODE(mode)) { 3092 /* 3093 * If a PMCATTACH has never been done on this PMC, 3094 * attach it to its owner process. 3095 */ 3096 if (LIST_EMPTY(&pm->pm_targets)) { 3097 error = (pm->pm_flags & PMC_F_ATTACH_DONE) != 0 ? 3098 ESRCH : pmc_attach_process(po->po_owner, pm); 3099 } 3100 3101 /* 3102 * If the PMC is attached to its owner, then force a context 3103 * switch to ensure that the MD state gets set correctly. 3104 */ 3105 if (error == 0) { 3106 pm->pm_state = PMC_STATE_RUNNING; 3107 if ((pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) != 0) 3108 pmc_force_context_switch(); 3109 } 3110 3111 return (error); 3112 } 3113 3114 /* 3115 * A system-wide PMC. 3116 * 3117 * Add the owner to the global list if this is a system-wide 3118 * sampling PMC. 3119 */ 3120 if (mode == PMC_MODE_SS) { 3121 /* 3122 * Log mapping information for all existing processes in the 3123 * system. Subsequent mappings are logged as they happen; 3124 * see pmc_process_mmap(). 3125 */ 3126 if (po->po_logprocmaps == 0) { 3127 pmc_log_all_process_mappings(po); 3128 po->po_logprocmaps = 1; 3129 } 3130 po->po_sscount++; 3131 if (po->po_sscount == 1) { 3132 atomic_add_rel_int(&pmc_ss_count, 1); 3133 CK_LIST_INSERT_HEAD(&pmc_ss_owners, po, po_ssnext); 3134 PMCDBG1(PMC,OPS,1, "po=%p in global list", po); 3135 } 3136 } 3137 3138 /* 3139 * Move to the CPU associated with this 3140 * PMC, and start the hardware. 3141 */ 3142 pmc_save_cpu_binding(&pb); 3143 cpu = PMC_TO_CPU(pm); 3144 if (!pmc_cpu_is_active(cpu)) 3145 return (ENXIO); 3146 pmc_select_cpu(cpu); 3147 3148 /* 3149 * global PMCs are configured at allocation time 3150 * so write out the initial value and start the PMC. 3151 */ 3152 pm->pm_state = PMC_STATE_RUNNING; 3153 3154 critical_enter(); 3155 v = PMC_IS_SAMPLING_MODE(mode) ? pm->pm_sc.pm_reloadcount : 3156 pm->pm_sc.pm_initial; 3157 if ((error = pcd->pcd_write_pmc(cpu, adjri, pm, v)) == 0) { 3158 /* If a sampling mode PMC, reset stalled state. */ 3159 if (PMC_IS_SAMPLING_MODE(mode)) 3160 pm->pm_pcpu_state[cpu].pps_stalled = 0; 3161 3162 /* Indicate that we desire this to run. Start it. */ 3163 pm->pm_pcpu_state[cpu].pps_cpustate = 1; 3164 error = pcd->pcd_start_pmc(cpu, adjri, pm); 3165 } 3166 critical_exit(); 3167 3168 pmc_restore_cpu_binding(&pb); 3169 return (error); 3170 } 3171 3172 /* 3173 * Stop a PMC. 3174 */ 3175 static int 3176 pmc_stop(struct pmc *pm) 3177 { 3178 struct pmc_binding pb; 3179 struct pmc_classdep *pcd; 3180 struct pmc_owner *po; 3181 int adjri, cpu, error, ri; 3182 3183 KASSERT(pm != NULL, ("[pmc,%d] null pmc", __LINE__)); 3184 3185 PMCDBG3(PMC,OPS,1, "stop pmc=%p mode=%d ri=%d", pm, PMC_TO_MODE(pm), 3186 PMC_TO_ROWINDEX(pm)); 3187 3188 pm->pm_state = PMC_STATE_STOPPED; 3189 3190 /* 3191 * If the PMC is a virtual mode one, changing the state to non-RUNNING 3192 * is enough to ensure that the PMC never gets scheduled. 3193 * 3194 * If this PMC is current running on a CPU, then it will handled 3195 * correctly at the time its target process is context switched out. 3196 */ 3197 if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) 3198 return (0); 3199 3200 /* 3201 * A system-mode PMC. Move to the CPU associated with this PMC, and 3202 * stop the hardware. We update the 'initial count' so that a 3203 * subsequent PMCSTART will resume counting from the current hardware 3204 * count. 3205 */ 3206 pmc_save_cpu_binding(&pb); 3207 3208 cpu = PMC_TO_CPU(pm); 3209 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), 3210 ("[pmc,%d] illegal cpu=%d", __LINE__, cpu)); 3211 if (!pmc_cpu_is_active(cpu)) 3212 return (ENXIO); 3213 3214 pmc_select_cpu(cpu); 3215 3216 ri = PMC_TO_ROWINDEX(pm); 3217 pcd = pmc_ri_to_classdep(md, ri, &adjri); 3218 3219 pm->pm_pcpu_state[cpu].pps_cpustate = 0; 3220 critical_enter(); 3221 if ((error = pcd->pcd_stop_pmc(cpu, adjri, pm)) == 0) { 3222 error = pcd->pcd_read_pmc(cpu, adjri, pm, 3223 &pm->pm_sc.pm_initial); 3224 } 3225 critical_exit(); 3226 3227 pmc_restore_cpu_binding(&pb); 3228 3229 /* Remove this owner from the global list of SS PMC owners. */ 3230 po = pm->pm_owner; 3231 if (PMC_TO_MODE(pm) == PMC_MODE_SS) { 3232 po->po_sscount--; 3233 if (po->po_sscount == 0) { 3234 atomic_subtract_rel_int(&pmc_ss_count, 1); 3235 CK_LIST_REMOVE(po, po_ssnext); 3236 epoch_wait_preempt(global_epoch_preempt); 3237 PMCDBG1(PMC,OPS,2,"po=%p removed from global list", po); 3238 } 3239 } 3240 3241 return (error); 3242 } 3243 3244 static struct pmc_classdep * 3245 pmc_class_to_classdep(enum pmc_class class) 3246 { 3247 int n; 3248 3249 for (n = 0; n < md->pmd_nclass; n++) { 3250 if (md->pmd_classdep[n].pcd_class == class) 3251 return (&md->pmd_classdep[n]); 3252 } 3253 return (NULL); 3254 } 3255 3256 #if defined(HWPMC_DEBUG) && defined(KTR) 3257 static const char *pmc_op_to_name[] = { 3258 #undef __PMC_OP 3259 #define __PMC_OP(N, D) #N , 3260 __PMC_OPS() 3261 NULL 3262 }; 3263 #endif 3264 3265 /* 3266 * The syscall interface 3267 */ 3268 3269 #define PMC_GET_SX_XLOCK(...) do { \ 3270 sx_xlock(&pmc_sx); \ 3271 if (pmc_hook == NULL) { \ 3272 sx_xunlock(&pmc_sx); \ 3273 return __VA_ARGS__; \ 3274 } \ 3275 } while (0) 3276 3277 #define PMC_DOWNGRADE_SX() do { \ 3278 sx_downgrade(&pmc_sx); \ 3279 is_sx_downgraded = true; \ 3280 } while (0) 3281 3282 /* 3283 * Main body of PMC_OP_PMCALLOCATE. 3284 */ 3285 static int 3286 pmc_do_op_pmcallocate(struct thread *td, struct pmc_op_pmcallocate *pa) 3287 { 3288 struct proc *p; 3289 struct pmc *pmc; 3290 struct pmc_binding pb; 3291 struct pmc_classdep *pcd; 3292 struct pmc_hw *phw; 3293 enum pmc_mode mode; 3294 enum pmc_class class; 3295 uint32_t caps, flags; 3296 u_int cpu; 3297 int adjri, n; 3298 int error; 3299 3300 class = pa->pm_class; 3301 caps = pa->pm_caps; 3302 flags = pa->pm_flags; 3303 mode = pa->pm_mode; 3304 cpu = pa->pm_cpu; 3305 3306 p = td->td_proc; 3307 3308 /* Requested mode must exist. */ 3309 if ((mode != PMC_MODE_SS && mode != PMC_MODE_SC && 3310 mode != PMC_MODE_TS && mode != PMC_MODE_TC)) 3311 return (EINVAL); 3312 3313 /* Requested CPU must be valid. */ 3314 if (cpu != PMC_CPU_ANY && cpu >= pmc_cpu_max()) 3315 return (EINVAL); 3316 3317 /* 3318 * Virtual PMCs should only ask for a default CPU. 3319 * System mode PMCs need to specify a non-default CPU. 3320 */ 3321 if ((PMC_IS_VIRTUAL_MODE(mode) && cpu != PMC_CPU_ANY) || 3322 (PMC_IS_SYSTEM_MODE(mode) && cpu == PMC_CPU_ANY)) 3323 return (EINVAL); 3324 3325 /* 3326 * Check that an inactive CPU is not being asked for. 3327 */ 3328 if (PMC_IS_SYSTEM_MODE(mode) && !pmc_cpu_is_active(cpu)) 3329 return (ENXIO); 3330 3331 /* 3332 * Refuse an allocation for a system-wide PMC if this process has been 3333 * jailed, or if this process lacks super-user credentials and the 3334 * sysctl tunable 'security.bsd.unprivileged_syspmcs' is zero. 3335 */ 3336 if (PMC_IS_SYSTEM_MODE(mode)) { 3337 if (jailed(td->td_ucred)) 3338 return (EPERM); 3339 if (!pmc_unprivileged_syspmcs) { 3340 error = priv_check(td, PRIV_PMC_SYSTEM); 3341 if (error != 0) 3342 return (error); 3343 } 3344 } 3345 3346 /* 3347 * Look for valid values for 'pm_flags'. 3348 */ 3349 if ((flags & ~(PMC_F_DESCENDANTS | PMC_F_LOG_PROCCSW | 3350 PMC_F_LOG_PROCEXIT | PMC_F_CALLCHAIN | PMC_F_USERCALLCHAIN | 3351 PMC_F_EV_PMU)) != 0) 3352 return (EINVAL); 3353 3354 /* PMC_F_USERCALLCHAIN is only valid with PMC_F_CALLCHAIN. */ 3355 if ((flags & (PMC_F_CALLCHAIN | PMC_F_USERCALLCHAIN)) == 3356 PMC_F_USERCALLCHAIN) 3357 return (EINVAL); 3358 3359 /* PMC_F_USERCALLCHAIN is only valid for sampling mode. */ 3360 if ((flags & PMC_F_USERCALLCHAIN) != 0 && mode != PMC_MODE_TS && 3361 mode != PMC_MODE_SS) 3362 return (EINVAL); 3363 3364 /* Process logging options are not allowed for system PMCs. */ 3365 if (PMC_IS_SYSTEM_MODE(mode) && 3366 (flags & (PMC_F_LOG_PROCCSW | PMC_F_LOG_PROCEXIT)) != 0) 3367 return (EINVAL); 3368 3369 /* 3370 * All sampling mode PMCs need to be able to interrupt the CPU. 3371 */ 3372 if (PMC_IS_SAMPLING_MODE(mode)) 3373 caps |= PMC_CAP_INTERRUPT; 3374 3375 /* A valid class specifier should have been passed in. */ 3376 pcd = pmc_class_to_classdep(class); 3377 if (pcd == NULL) 3378 return (EINVAL); 3379 3380 /* The requested PMC capabilities should be feasible. */ 3381 if ((pcd->pcd_caps & caps) != caps) 3382 return (EOPNOTSUPP); 3383 3384 PMCDBG4(PMC,ALL,2, "event=%d caps=0x%x mode=%d cpu=%d", pa->pm_ev, 3385 caps, mode, cpu); 3386 3387 pmc = pmc_allocate_pmc_descriptor(); 3388 pmc->pm_id = PMC_ID_MAKE_ID(cpu, pa->pm_mode, class, PMC_ID_INVALID); 3389 pmc->pm_event = pa->pm_ev; 3390 pmc->pm_state = PMC_STATE_FREE; 3391 pmc->pm_caps = caps; 3392 pmc->pm_flags = flags; 3393 3394 /* XXX set lower bound on sampling for process counters */ 3395 if (PMC_IS_SAMPLING_MODE(mode)) { 3396 /* 3397 * Don't permit requested sample rate to be less than 3398 * pmc_mincount. 3399 */ 3400 if (pa->pm_count < MAX(1, pmc_mincount)) 3401 log(LOG_WARNING, "pmcallocate: passed sample " 3402 "rate %ju - setting to %u\n", 3403 (uintmax_t)pa->pm_count, 3404 MAX(1, pmc_mincount)); 3405 pmc->pm_sc.pm_reloadcount = MAX(MAX(1, pmc_mincount), 3406 pa->pm_count); 3407 } else 3408 pmc->pm_sc.pm_initial = pa->pm_count; 3409 3410 /* switch thread to CPU 'cpu' */ 3411 pmc_save_cpu_binding(&pb); 3412 3413 #define PMC_IS_SHAREABLE_PMC(cpu, n) \ 3414 (pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_state & \ 3415 PMC_PHW_FLAG_IS_SHAREABLE) 3416 #define PMC_IS_UNALLOCATED(cpu, n) \ 3417 (pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_pmc == NULL) 3418 3419 if (PMC_IS_SYSTEM_MODE(mode)) { 3420 pmc_select_cpu(cpu); 3421 for (n = pcd->pcd_ri; n < md->pmd_npmc; n++) { 3422 pcd = pmc_ri_to_classdep(md, n, &adjri); 3423 3424 if (!pmc_can_allocate_row(n, mode) || 3425 !pmc_can_allocate_rowindex(p, n, cpu)) 3426 continue; 3427 if (!PMC_IS_UNALLOCATED(cpu, n) && 3428 !PMC_IS_SHAREABLE_PMC(cpu, n)) 3429 continue; 3430 3431 if (pcd->pcd_allocate_pmc(cpu, adjri, pmc, pa) == 0) { 3432 /* Success. */ 3433 break; 3434 } 3435 } 3436 } else { 3437 /* Process virtual mode */ 3438 for (n = pcd->pcd_ri; n < md->pmd_npmc; n++) { 3439 pcd = pmc_ri_to_classdep(md, n, &adjri); 3440 3441 if (!pmc_can_allocate_row(n, mode) || 3442 !pmc_can_allocate_rowindex(p, n, PMC_CPU_ANY)) 3443 continue; 3444 3445 if (pcd->pcd_allocate_pmc(td->td_oncpu, adjri, pmc, 3446 pa) == 0) { 3447 /* Success. */ 3448 break; 3449 } 3450 } 3451 } 3452 3453 #undef PMC_IS_UNALLOCATED 3454 #undef PMC_IS_SHAREABLE_PMC 3455 3456 pmc_restore_cpu_binding(&pb); 3457 3458 if (n == md->pmd_npmc) { 3459 pmc_destroy_pmc_descriptor(pmc); 3460 return (EINVAL); 3461 } 3462 3463 /* Fill in the correct value in the ID field. */ 3464 pmc->pm_id = PMC_ID_MAKE_ID(cpu, mode, class, n); 3465 3466 PMCDBG5(PMC,ALL,2, "ev=%d class=%d mode=%d n=%d -> pmcid=%x", 3467 pmc->pm_event, class, mode, n, pmc->pm_id); 3468 3469 /* Process mode PMCs with logging enabled need log files. */ 3470 if ((pmc->pm_flags & (PMC_F_LOG_PROCEXIT | PMC_F_LOG_PROCCSW)) != 0) 3471 pmc->pm_flags |= PMC_F_NEEDS_LOGFILE; 3472 3473 /* All system mode sampling PMCs require a log file. */ 3474 if (PMC_IS_SAMPLING_MODE(mode) && PMC_IS_SYSTEM_MODE(mode)) 3475 pmc->pm_flags |= PMC_F_NEEDS_LOGFILE; 3476 3477 /* 3478 * Configure global pmc's immediately. 3479 */ 3480 if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pmc))) { 3481 pmc_save_cpu_binding(&pb); 3482 pmc_select_cpu(cpu); 3483 3484 phw = pmc_pcpu[cpu]->pc_hwpmcs[n]; 3485 pcd = pmc_ri_to_classdep(md, n, &adjri); 3486 3487 if ((phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0 || 3488 (error = pcd->pcd_config_pmc(cpu, adjri, pmc)) != 0) { 3489 (void)pcd->pcd_release_pmc(cpu, adjri, pmc); 3490 pmc_destroy_pmc_descriptor(pmc); 3491 pmc_restore_cpu_binding(&pb); 3492 return (EPERM); 3493 } 3494 3495 pmc_restore_cpu_binding(&pb); 3496 } 3497 3498 pmc->pm_state = PMC_STATE_ALLOCATED; 3499 pmc->pm_class = class; 3500 3501 /* 3502 * Mark row disposition. 3503 */ 3504 if (PMC_IS_SYSTEM_MODE(mode)) 3505 PMC_MARK_ROW_STANDALONE(n); 3506 else 3507 PMC_MARK_ROW_THREAD(n); 3508 3509 /* 3510 * Register this PMC with the current thread as its owner. 3511 */ 3512 error = pmc_register_owner(p, pmc); 3513 if (error != 0) { 3514 pmc_release_pmc_descriptor(pmc); 3515 pmc_destroy_pmc_descriptor(pmc); 3516 return (error); 3517 } 3518 3519 /* 3520 * Return the allocated index. 3521 */ 3522 pa->pm_pmcid = pmc->pm_id; 3523 return (0); 3524 } 3525 3526 /* 3527 * Main body of PMC_OP_PMCATTACH. 3528 */ 3529 static int 3530 pmc_do_op_pmcattach(struct thread *td, struct pmc_op_pmcattach a) 3531 { 3532 struct pmc *pm; 3533 struct proc *p; 3534 int error; 3535 3536 sx_assert(&pmc_sx, SX_XLOCKED); 3537 3538 if (a.pm_pid < 0) { 3539 return (EINVAL); 3540 } else if (a.pm_pid == 0) { 3541 a.pm_pid = td->td_proc->p_pid; 3542 } 3543 3544 error = pmc_find_pmc(a.pm_pmc, &pm); 3545 if (error != 0) 3546 return (error); 3547 3548 if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm))) 3549 return (EINVAL); 3550 3551 /* PMCs may be (re)attached only when allocated or stopped */ 3552 if (pm->pm_state == PMC_STATE_RUNNING) { 3553 return (EBUSY); 3554 } else if (pm->pm_state != PMC_STATE_ALLOCATED && 3555 pm->pm_state != PMC_STATE_STOPPED) { 3556 return (EINVAL); 3557 } 3558 3559 /* lookup pid */ 3560 if ((p = pfind(a.pm_pid)) == NULL) 3561 return (ESRCH); 3562 3563 /* 3564 * Ignore processes that are working on exiting. 3565 */ 3566 if ((p->p_flag & P_WEXIT) != 0) { 3567 PROC_UNLOCK(p); /* pfind() returns a locked process */ 3568 return (ESRCH); 3569 } 3570 3571 /* 3572 * We are allowed to attach a PMC to a process if we can debug it. 3573 */ 3574 error = p_candebug(curthread, p); 3575 3576 PROC_UNLOCK(p); 3577 3578 if (error == 0) 3579 error = pmc_attach_process(p, pm); 3580 3581 return (error); 3582 } 3583 3584 /* 3585 * Main body of PMC_OP_PMCDETACH. 3586 */ 3587 static int 3588 pmc_do_op_pmcdetach(struct thread *td, struct pmc_op_pmcattach a) 3589 { 3590 struct pmc *pm; 3591 struct proc *p; 3592 int error; 3593 3594 if (a.pm_pid < 0) { 3595 return (EINVAL); 3596 } else if (a.pm_pid == 0) 3597 a.pm_pid = td->td_proc->p_pid; 3598 3599 error = pmc_find_pmc(a.pm_pmc, &pm); 3600 if (error != 0) 3601 return (error); 3602 3603 if ((p = pfind(a.pm_pid)) == NULL) 3604 return (ESRCH); 3605 3606 /* 3607 * Treat processes that are in the process of exiting as if they were 3608 * not present. 3609 */ 3610 if ((p->p_flag & P_WEXIT) != 0) { 3611 PROC_UNLOCK(p); 3612 return (ESRCH); 3613 } 3614 3615 PROC_UNLOCK(p); /* pfind() returns a locked process */ 3616 3617 if (error == 0) 3618 error = pmc_detach_process(p, pm); 3619 3620 return (error); 3621 } 3622 3623 /* 3624 * Main body of PMC_OP_PMCRELEASE. 3625 */ 3626 static int 3627 pmc_do_op_pmcrelease(pmc_id_t pmcid) 3628 { 3629 struct pmc_owner *po; 3630 struct pmc *pm; 3631 int error; 3632 3633 /* 3634 * Find PMC pointer for the named PMC. 3635 * 3636 * Use pmc_release_pmc_descriptor() to switch off the 3637 * PMC, remove all its target threads, and remove the 3638 * PMC from its owner's list. 3639 * 3640 * Remove the owner record if this is the last PMC 3641 * owned. 3642 * 3643 * Free up space. 3644 */ 3645 error = pmc_find_pmc(pmcid, &pm); 3646 if (error != 0) 3647 return (error); 3648 3649 po = pm->pm_owner; 3650 pmc_release_pmc_descriptor(pm); 3651 pmc_maybe_remove_owner(po); 3652 pmc_destroy_pmc_descriptor(pm); 3653 3654 return (error); 3655 } 3656 3657 /* 3658 * Main body of PMC_OP_PMCRW. 3659 */ 3660 static int 3661 pmc_do_op_pmcrw(const struct pmc_op_pmcrw *prw, pmc_value_t *valp) 3662 { 3663 struct pmc_binding pb; 3664 struct pmc_classdep *pcd; 3665 struct pmc *pm; 3666 u_int cpu, ri, adjri; 3667 int error; 3668 3669 PMCDBG2(PMC,OPS,1, "rw id=%d flags=0x%x", prw->pm_pmcid, prw->pm_flags); 3670 3671 /* Must have at least one flag set. */ 3672 if ((prw->pm_flags & (PMC_F_OLDVALUE | PMC_F_NEWVALUE)) == 0) 3673 return (EINVAL); 3674 3675 /* Locate PMC descriptor. */ 3676 error = pmc_find_pmc(prw->pm_pmcid, &pm); 3677 if (error != 0) 3678 return (error); 3679 3680 /* Can't read a PMC that hasn't been started. */ 3681 if (pm->pm_state != PMC_STATE_ALLOCATED && 3682 pm->pm_state != PMC_STATE_STOPPED && 3683 pm->pm_state != PMC_STATE_RUNNING) 3684 return (EINVAL); 3685 3686 /* Writing a new value is allowed only for 'STOPPED' PMCs. */ 3687 if (pm->pm_state == PMC_STATE_RUNNING && 3688 (prw->pm_flags & PMC_F_NEWVALUE) != 0) 3689 return (EBUSY); 3690 3691 if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) { 3692 /* 3693 * If this PMC is attached to its owner (i.e., the process 3694 * requesting this operation) and is running, then attempt to 3695 * get an upto-date reading from hardware for a READ. Writes 3696 * are only allowed when the PMC is stopped, so only update the 3697 * saved value field. 3698 * 3699 * If the PMC is not running, or is not attached to its owner, 3700 * read/write to the savedvalue field. 3701 */ 3702 3703 ri = PMC_TO_ROWINDEX(pm); 3704 pcd = pmc_ri_to_classdep(md, ri, &adjri); 3705 3706 mtx_pool_lock_spin(pmc_mtxpool, pm); 3707 cpu = curthread->td_oncpu; 3708 3709 if ((prw->pm_flags & PMC_F_OLDVALUE) != 0) { 3710 if ((pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) && 3711 (pm->pm_state == PMC_STATE_RUNNING)) { 3712 error = (*pcd->pcd_read_pmc)(cpu, adjri, pm, 3713 valp); 3714 } else { 3715 *valp = pm->pm_gv.pm_savedvalue; 3716 } 3717 } 3718 3719 if ((prw->pm_flags & PMC_F_NEWVALUE) != 0) 3720 pm->pm_gv.pm_savedvalue = prw->pm_value; 3721 3722 mtx_pool_unlock_spin(pmc_mtxpool, pm); 3723 } else { /* System mode PMCs */ 3724 cpu = PMC_TO_CPU(pm); 3725 ri = PMC_TO_ROWINDEX(pm); 3726 pcd = pmc_ri_to_classdep(md, ri, &adjri); 3727 3728 if (!pmc_cpu_is_active(cpu)) 3729 return (ENXIO); 3730 3731 /* Move this thread to CPU 'cpu'. */ 3732 pmc_save_cpu_binding(&pb); 3733 pmc_select_cpu(cpu); 3734 critical_enter(); 3735 3736 /* Save old value. */ 3737 if ((prw->pm_flags & PMC_F_OLDVALUE) != 0) 3738 error = (*pcd->pcd_read_pmc)(cpu, adjri, pm, valp); 3739 3740 /* Write out new value. */ 3741 if (error == 0 && (prw->pm_flags & PMC_F_NEWVALUE) != 0) 3742 error = (*pcd->pcd_write_pmc)(cpu, adjri, pm, 3743 prw->pm_value); 3744 3745 critical_exit(); 3746 pmc_restore_cpu_binding(&pb); 3747 if (error != 0) 3748 return (error); 3749 } 3750 3751 #ifdef HWPMC_DEBUG 3752 if ((prw->pm_flags & PMC_F_NEWVALUE) != 0) 3753 PMCDBG3(PMC,OPS,2, "rw id=%d new %jx -> old %jx", 3754 ri, prw->pm_value, *valp); 3755 else 3756 PMCDBG2(PMC,OPS,2, "rw id=%d -> old %jx", ri, *valp); 3757 #endif 3758 return (error); 3759 } 3760 3761 static int 3762 pmc_syscall_handler(struct thread *td, void *syscall_args) 3763 { 3764 struct pmc_syscall_args *c; 3765 void *pmclog_proc_handle; 3766 void *arg; 3767 int error, op; 3768 bool is_sx_downgraded; 3769 3770 c = (struct pmc_syscall_args *)syscall_args; 3771 op = c->pmop_code; 3772 arg = c->pmop_data; 3773 3774 /* PMC isn't set up yet */ 3775 if (pmc_hook == NULL) 3776 return (EINVAL); 3777 3778 if (op == PMC_OP_CONFIGURELOG) { 3779 /* 3780 * We cannot create the logging process inside 3781 * pmclog_configure_log() because there is a LOR 3782 * between pmc_sx and process structure locks. 3783 * Instead, pre-create the process and ignite the loop 3784 * if everything is fine, otherwise direct the process 3785 * to exit. 3786 */ 3787 error = pmclog_proc_create(td, &pmclog_proc_handle); 3788 if (error != 0) 3789 goto done_syscall; 3790 } 3791 3792 PMC_GET_SX_XLOCK(ENOSYS); 3793 is_sx_downgraded = false; 3794 PMCDBG3(MOD,PMS,1, "syscall op=%d \"%s\" arg=%p", op, 3795 pmc_op_to_name[op], arg); 3796 3797 error = 0; 3798 counter_u64_add(pmc_stats.pm_syscalls, 1); 3799 3800 switch (op) { 3801 3802 3803 /* 3804 * Configure a log file. 3805 * 3806 * XXX This OP will be reworked. 3807 */ 3808 3809 case PMC_OP_CONFIGURELOG: 3810 { 3811 struct proc *p; 3812 struct pmc *pm; 3813 struct pmc_owner *po; 3814 struct pmc_op_configurelog cl; 3815 3816 if ((error = copyin(arg, &cl, sizeof(cl))) != 0) { 3817 pmclog_proc_ignite(pmclog_proc_handle, NULL); 3818 break; 3819 } 3820 3821 /* No flags currently implemented */ 3822 if (cl.pm_flags != 0) { 3823 pmclog_proc_ignite(pmclog_proc_handle, NULL); 3824 error = EINVAL; 3825 break; 3826 } 3827 3828 /* mark this process as owning a log file */ 3829 p = td->td_proc; 3830 if ((po = pmc_find_owner_descriptor(p)) == NULL) 3831 if ((po = pmc_allocate_owner_descriptor(p)) == NULL) { 3832 pmclog_proc_ignite(pmclog_proc_handle, NULL); 3833 error = ENOMEM; 3834 break; 3835 } 3836 3837 /* 3838 * If a valid fd was passed in, try to configure that, 3839 * otherwise if 'fd' was less than zero and there was 3840 * a log file configured, flush its buffers and 3841 * de-configure it. 3842 */ 3843 if (cl.pm_logfd >= 0) { 3844 error = pmclog_configure_log(md, po, cl.pm_logfd); 3845 pmclog_proc_ignite(pmclog_proc_handle, error == 0 ? 3846 po : NULL); 3847 } else if (po->po_flags & PMC_PO_OWNS_LOGFILE) { 3848 pmclog_proc_ignite(pmclog_proc_handle, NULL); 3849 error = pmclog_close(po); 3850 if (error == 0) { 3851 LIST_FOREACH(pm, &po->po_pmcs, pm_next) 3852 if (pm->pm_flags & PMC_F_NEEDS_LOGFILE && 3853 pm->pm_state == PMC_STATE_RUNNING) 3854 pmc_stop(pm); 3855 error = pmclog_deconfigure_log(po); 3856 } 3857 } else { 3858 pmclog_proc_ignite(pmclog_proc_handle, NULL); 3859 error = EINVAL; 3860 } 3861 } 3862 break; 3863 3864 /* 3865 * Flush a log file. 3866 */ 3867 3868 case PMC_OP_FLUSHLOG: 3869 { 3870 struct pmc_owner *po; 3871 3872 sx_assert(&pmc_sx, SX_XLOCKED); 3873 3874 if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) { 3875 error = EINVAL; 3876 break; 3877 } 3878 3879 error = pmclog_flush(po, 0); 3880 } 3881 break; 3882 3883 /* 3884 * Close a log file. 3885 */ 3886 3887 case PMC_OP_CLOSELOG: 3888 { 3889 struct pmc_owner *po; 3890 3891 sx_assert(&pmc_sx, SX_XLOCKED); 3892 3893 if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) { 3894 error = EINVAL; 3895 break; 3896 } 3897 3898 error = pmclog_close(po); 3899 } 3900 break; 3901 3902 /* 3903 * Retrieve hardware configuration. 3904 */ 3905 3906 case PMC_OP_GETCPUINFO: /* CPU information */ 3907 { 3908 struct pmc_op_getcpuinfo gci; 3909 struct pmc_classinfo *pci; 3910 struct pmc_classdep *pcd; 3911 int cl; 3912 3913 memset(&gci, 0, sizeof(gci)); 3914 gci.pm_cputype = md->pmd_cputype; 3915 gci.pm_ncpu = pmc_cpu_max(); 3916 gci.pm_npmc = md->pmd_npmc; 3917 gci.pm_nclass = md->pmd_nclass; 3918 pci = gci.pm_classes; 3919 pcd = md->pmd_classdep; 3920 for (cl = 0; cl < md->pmd_nclass; cl++, pci++, pcd++) { 3921 pci->pm_caps = pcd->pcd_caps; 3922 pci->pm_class = pcd->pcd_class; 3923 pci->pm_width = pcd->pcd_width; 3924 pci->pm_num = pcd->pcd_num; 3925 } 3926 error = copyout(&gci, arg, sizeof(gci)); 3927 } 3928 break; 3929 3930 /* 3931 * Retrieve soft events list. 3932 */ 3933 case PMC_OP_GETDYNEVENTINFO: 3934 { 3935 enum pmc_class cl; 3936 enum pmc_event ev; 3937 struct pmc_op_getdyneventinfo *gei; 3938 struct pmc_dyn_event_descr dev; 3939 struct pmc_soft *ps; 3940 uint32_t nevent; 3941 3942 sx_assert(&pmc_sx, SX_LOCKED); 3943 3944 gei = (struct pmc_op_getdyneventinfo *) arg; 3945 3946 if ((error = copyin(&gei->pm_class, &cl, sizeof(cl))) != 0) 3947 break; 3948 3949 /* Only SOFT class is dynamic. */ 3950 if (cl != PMC_CLASS_SOFT) { 3951 error = EINVAL; 3952 break; 3953 } 3954 3955 nevent = 0; 3956 for (ev = PMC_EV_SOFT_FIRST; (int)ev <= PMC_EV_SOFT_LAST; ev++) { 3957 ps = pmc_soft_ev_acquire(ev); 3958 if (ps == NULL) 3959 continue; 3960 bcopy(&ps->ps_ev, &dev, sizeof(dev)); 3961 pmc_soft_ev_release(ps); 3962 3963 error = copyout(&dev, 3964 &gei->pm_events[nevent], 3965 sizeof(struct pmc_dyn_event_descr)); 3966 if (error != 0) 3967 break; 3968 nevent++; 3969 } 3970 if (error != 0) 3971 break; 3972 3973 error = copyout(&nevent, &gei->pm_nevent, 3974 sizeof(nevent)); 3975 } 3976 break; 3977 3978 /* 3979 * Get module statistics 3980 */ 3981 3982 case PMC_OP_GETDRIVERSTATS: 3983 { 3984 struct pmc_op_getdriverstats gms; 3985 #define CFETCH(a, b, field) a.field = counter_u64_fetch(b.field) 3986 CFETCH(gms, pmc_stats, pm_intr_ignored); 3987 CFETCH(gms, pmc_stats, pm_intr_processed); 3988 CFETCH(gms, pmc_stats, pm_intr_bufferfull); 3989 CFETCH(gms, pmc_stats, pm_syscalls); 3990 CFETCH(gms, pmc_stats, pm_syscall_errors); 3991 CFETCH(gms, pmc_stats, pm_buffer_requests); 3992 CFETCH(gms, pmc_stats, pm_buffer_requests_failed); 3993 CFETCH(gms, pmc_stats, pm_log_sweeps); 3994 #undef CFETCH 3995 error = copyout(&gms, arg, sizeof(gms)); 3996 } 3997 break; 3998 3999 4000 /* 4001 * Retrieve module version number 4002 */ 4003 4004 case PMC_OP_GETMODULEVERSION: 4005 { 4006 uint32_t cv, modv; 4007 4008 /* retrieve the client's idea of the ABI version */ 4009 if ((error = copyin(arg, &cv, sizeof(uint32_t))) != 0) 4010 break; 4011 /* don't service clients newer than our driver */ 4012 modv = PMC_VERSION; 4013 if ((cv & 0xFFFF0000) > (modv & 0xFFFF0000)) { 4014 error = EPROGMISMATCH; 4015 break; 4016 } 4017 error = copyout(&modv, arg, sizeof(int)); 4018 } 4019 break; 4020 4021 4022 /* 4023 * Retrieve the state of all the PMCs on a given 4024 * CPU. 4025 */ 4026 4027 case PMC_OP_GETPMCINFO: 4028 { 4029 int ari; 4030 struct pmc *pm; 4031 size_t pmcinfo_size; 4032 uint32_t cpu, n, npmc; 4033 struct pmc_owner *po; 4034 struct pmc_binding pb; 4035 struct pmc_classdep *pcd; 4036 struct pmc_info *p, *pmcinfo; 4037 struct pmc_op_getpmcinfo *gpi; 4038 4039 PMC_DOWNGRADE_SX(); 4040 4041 gpi = (struct pmc_op_getpmcinfo *) arg; 4042 4043 if ((error = copyin(&gpi->pm_cpu, &cpu, sizeof(cpu))) != 0) 4044 break; 4045 4046 if (cpu >= pmc_cpu_max()) { 4047 error = EINVAL; 4048 break; 4049 } 4050 4051 if (!pmc_cpu_is_active(cpu)) { 4052 error = ENXIO; 4053 break; 4054 } 4055 4056 /* switch to CPU 'cpu' */ 4057 pmc_save_cpu_binding(&pb); 4058 pmc_select_cpu(cpu); 4059 4060 npmc = md->pmd_npmc; 4061 4062 pmcinfo_size = npmc * sizeof(struct pmc_info); 4063 pmcinfo = malloc(pmcinfo_size, M_PMC, M_WAITOK | M_ZERO); 4064 4065 p = pmcinfo; 4066 4067 for (n = 0; n < md->pmd_npmc; n++, p++) { 4068 4069 pcd = pmc_ri_to_classdep(md, n, &ari); 4070 4071 KASSERT(pcd != NULL, 4072 ("[pmc,%d] null pcd ri=%d", __LINE__, n)); 4073 4074 if ((error = pcd->pcd_describe(cpu, ari, p, &pm)) != 0) 4075 break; 4076 4077 if (PMC_ROW_DISP_IS_STANDALONE(n)) 4078 p->pm_rowdisp = PMC_DISP_STANDALONE; 4079 else if (PMC_ROW_DISP_IS_THREAD(n)) 4080 p->pm_rowdisp = PMC_DISP_THREAD; 4081 else 4082 p->pm_rowdisp = PMC_DISP_FREE; 4083 4084 p->pm_ownerpid = -1; 4085 4086 if (pm == NULL) /* no PMC associated */ 4087 continue; 4088 4089 po = pm->pm_owner; 4090 4091 KASSERT(po->po_owner != NULL, 4092 ("[pmc,%d] pmc_owner had a null proc pointer", 4093 __LINE__)); 4094 4095 p->pm_ownerpid = po->po_owner->p_pid; 4096 p->pm_mode = PMC_TO_MODE(pm); 4097 p->pm_event = pm->pm_event; 4098 p->pm_flags = pm->pm_flags; 4099 4100 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) 4101 p->pm_reloadcount = 4102 pm->pm_sc.pm_reloadcount; 4103 } 4104 4105 pmc_restore_cpu_binding(&pb); 4106 4107 /* now copy out the PMC info collected */ 4108 if (error == 0) 4109 error = copyout(pmcinfo, &gpi->pm_pmcs, pmcinfo_size); 4110 4111 free(pmcinfo, M_PMC); 4112 } 4113 break; 4114 4115 4116 /* 4117 * Set the administrative state of a PMC. I.e. whether 4118 * the PMC is to be used or not. 4119 */ 4120 4121 case PMC_OP_PMCADMIN: 4122 { 4123 int cpu, ri; 4124 enum pmc_state request; 4125 struct pmc_cpu *pc; 4126 struct pmc_hw *phw; 4127 struct pmc_op_pmcadmin pma; 4128 struct pmc_binding pb; 4129 4130 sx_assert(&pmc_sx, SX_XLOCKED); 4131 4132 KASSERT(td == curthread, 4133 ("[pmc,%d] td != curthread", __LINE__)); 4134 4135 error = priv_check(td, PRIV_PMC_MANAGE); 4136 if (error) 4137 break; 4138 4139 if ((error = copyin(arg, &pma, sizeof(pma))) != 0) 4140 break; 4141 4142 cpu = pma.pm_cpu; 4143 4144 if (cpu < 0 || cpu >= (int) pmc_cpu_max()) { 4145 error = EINVAL; 4146 break; 4147 } 4148 4149 if (!pmc_cpu_is_active(cpu)) { 4150 error = ENXIO; 4151 break; 4152 } 4153 4154 request = pma.pm_state; 4155 4156 if (request != PMC_STATE_DISABLED && 4157 request != PMC_STATE_FREE) { 4158 error = EINVAL; 4159 break; 4160 } 4161 4162 ri = pma.pm_pmc; /* pmc id == row index */ 4163 if (ri < 0 || ri >= (int) md->pmd_npmc) { 4164 error = EINVAL; 4165 break; 4166 } 4167 4168 /* 4169 * We can't disable a PMC with a row-index allocated 4170 * for process virtual PMCs. 4171 */ 4172 4173 if (PMC_ROW_DISP_IS_THREAD(ri) && 4174 request == PMC_STATE_DISABLED) { 4175 error = EBUSY; 4176 break; 4177 } 4178 4179 /* 4180 * otherwise, this PMC on this CPU is either free or 4181 * in system-wide mode. 4182 */ 4183 4184 pmc_save_cpu_binding(&pb); 4185 pmc_select_cpu(cpu); 4186 4187 pc = pmc_pcpu[cpu]; 4188 phw = pc->pc_hwpmcs[ri]; 4189 4190 /* 4191 * XXX do we need some kind of 'forced' disable? 4192 */ 4193 4194 if (phw->phw_pmc == NULL) { 4195 if (request == PMC_STATE_DISABLED && 4196 (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED)) { 4197 phw->phw_state &= ~PMC_PHW_FLAG_IS_ENABLED; 4198 PMC_MARK_ROW_STANDALONE(ri); 4199 } else if (request == PMC_STATE_FREE && 4200 (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0) { 4201 phw->phw_state |= PMC_PHW_FLAG_IS_ENABLED; 4202 PMC_UNMARK_ROW_STANDALONE(ri); 4203 } 4204 /* other cases are a no-op */ 4205 } else 4206 error = EBUSY; 4207 4208 pmc_restore_cpu_binding(&pb); 4209 } 4210 break; 4211 4212 4213 /* 4214 * Allocate a PMC. 4215 */ 4216 case PMC_OP_PMCALLOCATE: 4217 { 4218 struct pmc_op_pmcallocate pa; 4219 4220 error = copyin(arg, &pa, sizeof(pa)); 4221 if (error != 0) 4222 break; 4223 4224 error = pmc_do_op_pmcallocate(td, &pa); 4225 if (error != 0) 4226 break; 4227 4228 error = copyout(&pa, arg, sizeof(pa)); 4229 } 4230 break; 4231 4232 /* 4233 * Attach a PMC to a process. 4234 */ 4235 case PMC_OP_PMCATTACH: 4236 { 4237 struct pmc_op_pmcattach a; 4238 4239 error = copyin(arg, &a, sizeof(a)); 4240 if (error != 0) 4241 break; 4242 4243 error = pmc_do_op_pmcattach(td, a); 4244 } 4245 break; 4246 4247 /* 4248 * Detach an attached PMC from a process. 4249 */ 4250 case PMC_OP_PMCDETACH: 4251 { 4252 struct pmc_op_pmcattach a; 4253 4254 error = copyin(arg, &a, sizeof(a)); 4255 if (error != 0) 4256 break; 4257 4258 error = pmc_do_op_pmcdetach(td, a); 4259 } 4260 break; 4261 4262 4263 /* 4264 * Retrieve the MSR number associated with the counter 4265 * 'pmc_id'. This allows processes to directly use RDPMC 4266 * instructions to read their PMCs, without the overhead of a 4267 * system call. 4268 */ 4269 4270 case PMC_OP_PMCGETMSR: 4271 { 4272 int adjri, ri; 4273 struct pmc *pm; 4274 struct pmc_target *pt; 4275 struct pmc_op_getmsr gm; 4276 struct pmc_classdep *pcd; 4277 4278 PMC_DOWNGRADE_SX(); 4279 4280 if ((error = copyin(arg, &gm, sizeof(gm))) != 0) 4281 break; 4282 4283 if ((error = pmc_find_pmc(gm.pm_pmcid, &pm)) != 0) 4284 break; 4285 4286 /* 4287 * The allocated PMC has to be a process virtual PMC, 4288 * i.e., of type MODE_T[CS]. Global PMCs can only be 4289 * read using the PMCREAD operation since they may be 4290 * allocated on a different CPU than the one we could 4291 * be running on at the time of the RDPMC instruction. 4292 * 4293 * The GETMSR operation is not allowed for PMCs that 4294 * are inherited across processes. 4295 */ 4296 4297 if (!PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)) || 4298 (pm->pm_flags & PMC_F_DESCENDANTS)) { 4299 error = EINVAL; 4300 break; 4301 } 4302 4303 /* 4304 * It only makes sense to use a RDPMC (or its 4305 * equivalent instruction on non-x86 architectures) on 4306 * a process that has allocated and attached a PMC to 4307 * itself. Conversely the PMC is only allowed to have 4308 * one process attached to it -- its owner. 4309 */ 4310 4311 if ((pt = LIST_FIRST(&pm->pm_targets)) == NULL || 4312 LIST_NEXT(pt, pt_next) != NULL || 4313 pt->pt_process->pp_proc != pm->pm_owner->po_owner) { 4314 error = EINVAL; 4315 break; 4316 } 4317 4318 ri = PMC_TO_ROWINDEX(pm); 4319 pcd = pmc_ri_to_classdep(md, ri, &adjri); 4320 4321 /* PMC class has no 'GETMSR' support */ 4322 if (pcd->pcd_get_msr == NULL) { 4323 error = ENOSYS; 4324 break; 4325 } 4326 4327 if ((error = (*pcd->pcd_get_msr)(adjri, &gm.pm_msr)) < 0) 4328 break; 4329 4330 if ((error = copyout(&gm, arg, sizeof(gm))) < 0) 4331 break; 4332 4333 /* 4334 * Mark our process as using MSRs. Update machine 4335 * state using a forced context switch. 4336 */ 4337 4338 pt->pt_process->pp_flags |= PMC_PP_ENABLE_MSR_ACCESS; 4339 pmc_force_context_switch(); 4340 4341 } 4342 break; 4343 4344 /* 4345 * Release an allocated PMC. 4346 */ 4347 case PMC_OP_PMCRELEASE: 4348 { 4349 struct pmc_op_simple sp; 4350 4351 error = copyin(arg, &sp, sizeof(sp)); 4352 if (error != 0) 4353 break; 4354 4355 error = pmc_do_op_pmcrelease(sp.pm_pmcid); 4356 } 4357 break; 4358 4359 /* 4360 * Read and/or write a PMC. 4361 */ 4362 case PMC_OP_PMCRW: 4363 { 4364 struct pmc_op_pmcrw prw; 4365 struct pmc_op_pmcrw *pprw; 4366 pmc_value_t oldvalue; 4367 4368 PMC_DOWNGRADE_SX(); 4369 4370 error = copyin(arg, &prw, sizeof(prw)); 4371 if (error != 0) 4372 break; 4373 4374 error = pmc_do_op_pmcrw(&prw, &oldvalue); 4375 if (error != 0) 4376 break; 4377 4378 /* Return old value if requested. */ 4379 if ((prw.pm_flags & PMC_F_OLDVALUE) != 0) { 4380 pprw = arg; 4381 error = copyout(&oldvalue, &pprw->pm_value, 4382 sizeof(prw.pm_value)); 4383 } 4384 } 4385 break; 4386 4387 4388 /* 4389 * Set the sampling rate for a sampling mode PMC and the 4390 * initial count for a counting mode PMC. 4391 */ 4392 4393 case PMC_OP_PMCSETCOUNT: 4394 { 4395 struct pmc *pm; 4396 struct pmc_op_pmcsetcount sc; 4397 4398 PMC_DOWNGRADE_SX(); 4399 4400 if ((error = copyin(arg, &sc, sizeof(sc))) != 0) 4401 break; 4402 4403 if ((error = pmc_find_pmc(sc.pm_pmcid, &pm)) != 0) 4404 break; 4405 4406 if (pm->pm_state == PMC_STATE_RUNNING) { 4407 error = EBUSY; 4408 break; 4409 } 4410 4411 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) { 4412 /* 4413 * Don't permit requested sample rate to be 4414 * less than pmc_mincount. 4415 */ 4416 if (sc.pm_count < MAX(1, pmc_mincount)) 4417 log(LOG_WARNING, "pmcsetcount: passed sample " 4418 "rate %ju - setting to %u\n", 4419 (uintmax_t)sc.pm_count, 4420 MAX(1, pmc_mincount)); 4421 pm->pm_sc.pm_reloadcount = MAX(MAX(1, pmc_mincount), 4422 sc.pm_count); 4423 } else 4424 pm->pm_sc.pm_initial = sc.pm_count; 4425 } 4426 break; 4427 4428 4429 /* 4430 * Start a PMC. 4431 */ 4432 4433 case PMC_OP_PMCSTART: 4434 { 4435 pmc_id_t pmcid; 4436 struct pmc *pm; 4437 struct pmc_op_simple sp; 4438 4439 sx_assert(&pmc_sx, SX_XLOCKED); 4440 4441 if ((error = copyin(arg, &sp, sizeof(sp))) != 0) 4442 break; 4443 4444 pmcid = sp.pm_pmcid; 4445 4446 if ((error = pmc_find_pmc(pmcid, &pm)) != 0) 4447 break; 4448 4449 KASSERT(pmcid == pm->pm_id, 4450 ("[pmc,%d] pmcid %x != id %x", __LINE__, 4451 pm->pm_id, pmcid)); 4452 4453 if (pm->pm_state == PMC_STATE_RUNNING) /* already running */ 4454 break; 4455 else if (pm->pm_state != PMC_STATE_STOPPED && 4456 pm->pm_state != PMC_STATE_ALLOCATED) { 4457 error = EINVAL; 4458 break; 4459 } 4460 4461 error = pmc_start(pm); 4462 } 4463 break; 4464 4465 4466 /* 4467 * Stop a PMC. 4468 */ 4469 4470 case PMC_OP_PMCSTOP: 4471 { 4472 pmc_id_t pmcid; 4473 struct pmc *pm; 4474 struct pmc_op_simple sp; 4475 4476 PMC_DOWNGRADE_SX(); 4477 4478 if ((error = copyin(arg, &sp, sizeof(sp))) != 0) 4479 break; 4480 4481 pmcid = sp.pm_pmcid; 4482 4483 /* 4484 * Mark the PMC as inactive and invoke the MD stop 4485 * routines if needed. 4486 */ 4487 4488 if ((error = pmc_find_pmc(pmcid, &pm)) != 0) 4489 break; 4490 4491 KASSERT(pmcid == pm->pm_id, 4492 ("[pmc,%d] pmc id %x != pmcid %x", __LINE__, 4493 pm->pm_id, pmcid)); 4494 4495 if (pm->pm_state == PMC_STATE_STOPPED) /* already stopped */ 4496 break; 4497 else if (pm->pm_state != PMC_STATE_RUNNING) { 4498 error = EINVAL; 4499 break; 4500 } 4501 4502 error = pmc_stop(pm); 4503 } 4504 break; 4505 4506 4507 /* 4508 * Write a user supplied value to the log file. 4509 */ 4510 4511 case PMC_OP_WRITELOG: 4512 { 4513 struct pmc_op_writelog wl; 4514 struct pmc_owner *po; 4515 4516 PMC_DOWNGRADE_SX(); 4517 4518 if ((error = copyin(arg, &wl, sizeof(wl))) != 0) 4519 break; 4520 4521 if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) { 4522 error = EINVAL; 4523 break; 4524 } 4525 4526 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) { 4527 error = EINVAL; 4528 break; 4529 } 4530 4531 error = pmclog_process_userlog(po, &wl); 4532 } 4533 break; 4534 4535 4536 default: 4537 error = EINVAL; 4538 break; 4539 } 4540 4541 if (is_sx_downgraded) 4542 sx_sunlock(&pmc_sx); 4543 else 4544 sx_xunlock(&pmc_sx); 4545 done_syscall: 4546 if (error) 4547 counter_u64_add(pmc_stats.pm_syscall_errors, 1); 4548 4549 return (error); 4550 } 4551 4552 /* 4553 * Helper functions 4554 */ 4555 4556 /* 4557 * Mark the thread as needing callchain capture and post an AST. The 4558 * actual callchain capture will be done in a context where it is safe 4559 * to take page faults. 4560 */ 4561 static void 4562 pmc_post_callchain_callback(void) 4563 { 4564 struct thread *td; 4565 4566 td = curthread; 4567 4568 /* 4569 * If there is multiple PMCs for the same interrupt ignore new post 4570 */ 4571 if ((td->td_pflags & TDP_CALLCHAIN) != 0) 4572 return; 4573 4574 /* 4575 * Mark this thread as needing callchain capture. 4576 * `td->td_pflags' will be safe to touch because this thread 4577 * was in user space when it was interrupted. 4578 */ 4579 td->td_pflags |= TDP_CALLCHAIN; 4580 4581 /* 4582 * Don't let this thread migrate between CPUs until callchain 4583 * capture completes. 4584 */ 4585 sched_pin(); 4586 4587 return; 4588 } 4589 4590 /* 4591 * Find a free slot in the per-cpu array of samples and capture the 4592 * current callchain there. If a sample was successfully added, a bit 4593 * is set in mask 'pmc_cpumask' denoting that the DO_SAMPLES hook 4594 * needs to be invoked from the clock handler. 4595 * 4596 * This function is meant to be called from an NMI handler. It cannot 4597 * use any of the locking primitives supplied by the OS. 4598 */ 4599 static int 4600 pmc_add_sample(ring_type_t ring, struct pmc *pm, struct trapframe *tf) 4601 { 4602 struct pmc_sample *ps; 4603 struct pmc_samplebuffer *psb; 4604 struct thread *td; 4605 int error, cpu, callchaindepth; 4606 bool inuserspace; 4607 4608 error = 0; 4609 4610 /* 4611 * Allocate space for a sample buffer. 4612 */ 4613 cpu = curcpu; 4614 psb = pmc_pcpu[cpu]->pc_sb[ring]; 4615 inuserspace = TRAPF_USERMODE(tf); 4616 ps = PMC_PROD_SAMPLE(psb); 4617 if (psb->ps_considx != psb->ps_prodidx && 4618 ps->ps_nsamples) { /* in use, reader hasn't caught up */ 4619 pm->pm_pcpu_state[cpu].pps_stalled = 1; 4620 counter_u64_add(pmc_stats.pm_intr_bufferfull, 1); 4621 PMCDBG6(SAM,INT,1,"(spc) cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d", 4622 cpu, pm, tf, inuserspace, 4623 (int)(psb->ps_prodidx & pmc_sample_mask), 4624 (int)(psb->ps_considx & pmc_sample_mask)); 4625 callchaindepth = 1; 4626 error = ENOMEM; 4627 goto done; 4628 } 4629 4630 /* Fill in entry. */ 4631 PMCDBG6(SAM,INT,1,"cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d", cpu, pm, tf, 4632 inuserspace, (int)(psb->ps_prodidx & pmc_sample_mask), 4633 (int)(psb->ps_considx & pmc_sample_mask)); 4634 4635 td = curthread; 4636 ps->ps_pmc = pm; 4637 ps->ps_td = td; 4638 ps->ps_pid = td->td_proc->p_pid; 4639 ps->ps_tid = td->td_tid; 4640 ps->ps_tsc = pmc_rdtsc(); 4641 ps->ps_ticks = ticks; 4642 ps->ps_cpu = cpu; 4643 ps->ps_flags = inuserspace ? PMC_CC_F_USERSPACE : 0; 4644 4645 callchaindepth = (pm->pm_flags & PMC_F_CALLCHAIN) ? 4646 pmc_callchaindepth : 1; 4647 4648 MPASS(ps->ps_pc != NULL); 4649 if (callchaindepth == 1) { 4650 ps->ps_pc[0] = PMC_TRAPFRAME_TO_PC(tf); 4651 } else { 4652 /* 4653 * Kernel stack traversals can be done immediately, while we 4654 * defer to an AST for user space traversals. 4655 */ 4656 if (!inuserspace) { 4657 callchaindepth = pmc_save_kernel_callchain(ps->ps_pc, 4658 callchaindepth, tf); 4659 } else { 4660 pmc_post_callchain_callback(); 4661 callchaindepth = PMC_USER_CALLCHAIN_PENDING; 4662 } 4663 } 4664 4665 ps->ps_nsamples = callchaindepth; /* mark entry as in-use */ 4666 if (ring == PMC_UR) { 4667 ps->ps_nsamples_actual = callchaindepth; 4668 ps->ps_nsamples = PMC_USER_CALLCHAIN_PENDING; 4669 } 4670 4671 KASSERT(counter_u64_fetch(pm->pm_runcount) >= 0, 4672 ("[pmc,%d] pm=%p runcount %ju", __LINE__, pm, 4673 (uintmax_t)counter_u64_fetch(pm->pm_runcount))); 4674 4675 counter_u64_add(pm->pm_runcount, 1); /* hold onto PMC */ 4676 /* increment write pointer */ 4677 psb->ps_prodidx++; 4678 done: 4679 /* mark CPU as needing processing */ 4680 if (callchaindepth != PMC_USER_CALLCHAIN_PENDING) 4681 DPCPU_SET(pmc_sampled, 1); 4682 4683 return (error); 4684 } 4685 4686 /* 4687 * Interrupt processing. 4688 * 4689 * This function may be called from an NMI handler. It cannot use any of the 4690 * locking primitives supplied by the OS. 4691 */ 4692 int 4693 pmc_process_interrupt(int ring, struct pmc *pm, struct trapframe *tf) 4694 { 4695 struct thread *td; 4696 4697 td = curthread; 4698 if ((pm->pm_flags & PMC_F_USERCALLCHAIN) && 4699 (td->td_proc->p_flag & P_KPROC) == 0 && !TRAPF_USERMODE(tf)) { 4700 atomic_add_int(&td->td_pmcpend, 1); 4701 return (pmc_add_sample(PMC_UR, pm, tf)); 4702 } 4703 return (pmc_add_sample(ring, pm, tf)); 4704 } 4705 4706 /* 4707 * Capture a user call chain. This function will be called from ast() 4708 * before control returns to userland and before the process gets 4709 * rescheduled. 4710 */ 4711 static void 4712 pmc_capture_user_callchain(int cpu, int ring, struct trapframe *tf) 4713 { 4714 struct pmc *pm; 4715 struct pmc_sample *ps; 4716 struct pmc_samplebuffer *psb; 4717 struct thread *td; 4718 uint64_t considx, prodidx; 4719 int nsamples, nrecords, pass, iter; 4720 int start_ticks __diagused; 4721 4722 psb = pmc_pcpu[cpu]->pc_sb[ring]; 4723 td = curthread; 4724 nrecords = INT_MAX; 4725 pass = 0; 4726 start_ticks = ticks; 4727 4728 KASSERT(td->td_pflags & TDP_CALLCHAIN, 4729 ("[pmc,%d] Retrieving callchain for thread that doesn't want it", 4730 __LINE__)); 4731 restart: 4732 if (ring == PMC_UR) 4733 nrecords = atomic_readandclear_32(&td->td_pmcpend); 4734 4735 for (iter = 0, considx = psb->ps_considx, prodidx = psb->ps_prodidx; 4736 considx < prodidx && iter < pmc_nsamples; considx++, iter++) { 4737 ps = PMC_CONS_SAMPLE_OFF(psb, considx); 4738 4739 /* 4740 * Iterate through all deferred callchain requests. Walk from 4741 * the current read pointer to the current write pointer. 4742 */ 4743 #ifdef INVARIANTS 4744 if (ps->ps_nsamples == PMC_SAMPLE_FREE) { 4745 continue; 4746 } 4747 #endif 4748 if (ps->ps_td != td || 4749 ps->ps_nsamples != PMC_USER_CALLCHAIN_PENDING || 4750 ps->ps_pmc->pm_state != PMC_STATE_RUNNING) 4751 continue; 4752 4753 KASSERT(ps->ps_cpu == cpu, 4754 ("[pmc,%d] cpu mismatch ps_cpu=%d pcpu=%d", __LINE__, 4755 ps->ps_cpu, PCPU_GET(cpuid))); 4756 4757 pm = ps->ps_pmc; 4758 KASSERT(pm->pm_flags & PMC_F_CALLCHAIN, 4759 ("[pmc,%d] Retrieving callchain for PMC that doesn't " 4760 "want it", __LINE__)); 4761 KASSERT(counter_u64_fetch(pm->pm_runcount) > 0, 4762 ("[pmc,%d] runcount %ju", __LINE__, 4763 (uintmax_t)counter_u64_fetch(pm->pm_runcount))); 4764 4765 if (ring == PMC_UR) { 4766 nsamples = ps->ps_nsamples_actual; 4767 counter_u64_add(pmc_stats.pm_merges, 1); 4768 } else 4769 nsamples = 0; 4770 4771 /* 4772 * Retrieve the callchain and mark the sample buffer 4773 * as 'processable' by the timer tick sweep code. 4774 */ 4775 if (__predict_true(nsamples < pmc_callchaindepth - 1)) 4776 nsamples += pmc_save_user_callchain(ps->ps_pc + nsamples, 4777 pmc_callchaindepth - nsamples - 1, tf); 4778 4779 /* 4780 * We have to prevent hardclock from potentially overwriting 4781 * this sample between when we read the value and when we set 4782 * it. 4783 */ 4784 spinlock_enter(); 4785 4786 /* 4787 * Verify that the sample hasn't been dropped in the meantime. 4788 */ 4789 if (ps->ps_nsamples == PMC_USER_CALLCHAIN_PENDING) { 4790 ps->ps_nsamples = nsamples; 4791 /* 4792 * If we couldn't get a sample, simply drop the 4793 * reference. 4794 */ 4795 if (nsamples == 0) 4796 counter_u64_add(pm->pm_runcount, -1); 4797 } 4798 spinlock_exit(); 4799 if (nrecords-- == 1) 4800 break; 4801 } 4802 if (__predict_false(ring == PMC_UR && td->td_pmcpend)) { 4803 if (pass == 0) { 4804 pass = 1; 4805 goto restart; 4806 } 4807 /* only collect samples for this part once */ 4808 td->td_pmcpend = 0; 4809 } 4810 4811 #ifdef INVARIANTS 4812 if ((ticks - start_ticks) > hz) 4813 log(LOG_ERR, "%s took %d ticks\n", __func__, (ticks - start_ticks)); 4814 #endif 4815 /* mark CPU as needing processing */ 4816 DPCPU_SET(pmc_sampled, 1); 4817 } 4818 4819 /* 4820 * Process saved PC samples. 4821 */ 4822 static void 4823 pmc_process_samples(int cpu, ring_type_t ring) 4824 { 4825 struct pmc *pm; 4826 struct thread *td; 4827 struct pmc_owner *po; 4828 struct pmc_sample *ps; 4829 struct pmc_classdep *pcd; 4830 struct pmc_samplebuffer *psb; 4831 uint64_t delta __diagused; 4832 int adjri, n; 4833 4834 KASSERT(PCPU_GET(cpuid) == cpu, 4835 ("[pmc,%d] not on the correct CPU pcpu=%d cpu=%d", __LINE__, 4836 PCPU_GET(cpuid), cpu)); 4837 4838 psb = pmc_pcpu[cpu]->pc_sb[ring]; 4839 delta = psb->ps_prodidx - psb->ps_considx; 4840 MPASS(delta <= pmc_nsamples); 4841 MPASS(psb->ps_considx <= psb->ps_prodidx); 4842 for (n = 0; psb->ps_considx < psb->ps_prodidx; psb->ps_considx++, n++) { 4843 ps = PMC_CONS_SAMPLE(psb); 4844 4845 if (__predict_false(ps->ps_nsamples == PMC_SAMPLE_FREE)) 4846 continue; 4847 4848 /* skip non-running samples */ 4849 pm = ps->ps_pmc; 4850 if (pm->pm_state != PMC_STATE_RUNNING) 4851 goto entrydone; 4852 4853 KASSERT(counter_u64_fetch(pm->pm_runcount) > 0, 4854 ("[pmc,%d] pm=%p runcount %ju", __LINE__, pm, 4855 (uintmax_t)counter_u64_fetch(pm->pm_runcount))); 4856 KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)), 4857 ("[pmc,%d] pmc=%p non-sampling mode=%d", __LINE__, 4858 pm, PMC_TO_MODE(pm))); 4859 4860 po = pm->pm_owner; 4861 4862 /* If there is a pending AST wait for completion */ 4863 if (ps->ps_nsamples == PMC_USER_CALLCHAIN_PENDING) { 4864 /* 4865 * If we've been waiting more than 1 tick to 4866 * collect a callchain for this record then 4867 * drop it and move on. 4868 */ 4869 if (ticks - ps->ps_ticks > 1) { 4870 /* 4871 * Track how often we hit this as it will 4872 * preferentially lose user samples 4873 * for long running system calls. 4874 */ 4875 counter_u64_add(pmc_stats.pm_overwrites, 1); 4876 goto entrydone; 4877 } 4878 /* Need a rescan at a later time. */ 4879 DPCPU_SET(pmc_sampled, 1); 4880 break; 4881 } 4882 4883 PMCDBG6(SAM,OPS,1,"cpu=%d pm=%p n=%d fl=%x wr=%d rd=%d", cpu, 4884 pm, ps->ps_nsamples, ps->ps_flags, 4885 (int)(psb->ps_prodidx & pmc_sample_mask), 4886 (int)(psb->ps_considx & pmc_sample_mask)); 4887 4888 /* 4889 * If this is a process-mode PMC that is attached to 4890 * its owner, and if the PC is in user mode, update 4891 * profiling statistics like timer-based profiling 4892 * would have done. 4893 * 4894 * Otherwise, this is either a sampling-mode PMC that 4895 * is attached to a different process than its owner, 4896 * or a system-wide sampling PMC. Dispatch a log 4897 * entry to the PMC's owner process. 4898 */ 4899 if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) { 4900 if (ps->ps_flags & PMC_CC_F_USERSPACE) { 4901 td = FIRST_THREAD_IN_PROC(po->po_owner); 4902 addupc_intr(td, ps->ps_pc[0], 1); 4903 } 4904 } else 4905 pmclog_process_callchain(pm, ps); 4906 4907 entrydone: 4908 ps->ps_nsamples = 0; /* mark entry as free */ 4909 KASSERT(counter_u64_fetch(pm->pm_runcount) > 0, 4910 ("[pmc,%d] pm=%p runcount %ju", __LINE__, pm, 4911 (uintmax_t)counter_u64_fetch(pm->pm_runcount))); 4912 4913 counter_u64_add(pm->pm_runcount, -1); 4914 } 4915 4916 counter_u64_add(pmc_stats.pm_log_sweeps, 1); 4917 4918 /* Do not re-enable stalled PMCs if we failed to process any samples */ 4919 if (n == 0) 4920 return; 4921 4922 /* 4923 * Restart any stalled sampling PMCs on this CPU. 4924 * 4925 * If the NMI handler sets the pm_stalled field of a PMC after 4926 * the check below, we'll end up processing the stalled PMC at 4927 * the next hardclock tick. 4928 */ 4929 for (n = 0; n < md->pmd_npmc; n++) { 4930 pcd = pmc_ri_to_classdep(md, n, &adjri); 4931 KASSERT(pcd != NULL, 4932 ("[pmc,%d] null pcd ri=%d", __LINE__, n)); 4933 (void)(*pcd->pcd_get_config)(cpu, adjri, &pm); 4934 4935 if (pm == NULL || /* !cfg'ed */ 4936 pm->pm_state != PMC_STATE_RUNNING || /* !active */ 4937 !PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) || /* !sampling */ 4938 !pm->pm_pcpu_state[cpu].pps_cpustate || /* !desired */ 4939 !pm->pm_pcpu_state[cpu].pps_stalled) /* !stalled */ 4940 continue; 4941 4942 pm->pm_pcpu_state[cpu].pps_stalled = 0; 4943 (void)(*pcd->pcd_start_pmc)(cpu, adjri, pm); 4944 } 4945 } 4946 4947 /* 4948 * Event handlers. 4949 */ 4950 4951 /* 4952 * Handle a process exit. 4953 * 4954 * Remove this process from all hash tables. If this process 4955 * owned any PMCs, turn off those PMCs and deallocate them, 4956 * removing any associations with target processes. 4957 * 4958 * This function will be called by the last 'thread' of a 4959 * process. 4960 * 4961 * XXX This eventhandler gets called early in the exit process. 4962 * Consider using a 'hook' invocation from thread_exit() or equivalent 4963 * spot. Another negative is that kse_exit doesn't seem to call 4964 * exit1() [??]. 4965 */ 4966 static void 4967 pmc_process_exit(void *arg __unused, struct proc *p) 4968 { 4969 struct pmc *pm; 4970 struct pmc_owner *po; 4971 struct pmc_process *pp; 4972 struct pmc_classdep *pcd; 4973 pmc_value_t newvalue, tmp; 4974 int ri, adjri, cpu; 4975 bool is_using_hwpmcs; 4976 4977 PROC_LOCK(p); 4978 is_using_hwpmcs = (p->p_flag & P_HWPMC) != 0; 4979 PROC_UNLOCK(p); 4980 4981 /* 4982 * Log a sysexit event to all SS PMC owners. 4983 */ 4984 PMC_EPOCH_ENTER(); 4985 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) { 4986 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) != 0) 4987 pmclog_process_sysexit(po, p->p_pid); 4988 } 4989 PMC_EPOCH_EXIT(); 4990 4991 PMC_GET_SX_XLOCK(); 4992 PMCDBG3(PRC,EXT,1,"process-exit proc=%p (%d, %s)", p, p->p_pid, 4993 p->p_comm); 4994 4995 if (!is_using_hwpmcs) 4996 goto out; 4997 4998 /* 4999 * Since this code is invoked by the last thread in an exiting process, 5000 * we would have context switched IN at some prior point. However, with 5001 * PREEMPTION, kernel mode context switches may happen any time, so we 5002 * want to disable a context switch OUT till we get any PMCs targeting 5003 * this process off the hardware. 5004 * 5005 * We also need to atomically remove this process' entry from our 5006 * target process hash table, using PMC_FLAG_REMOVE. 5007 */ 5008 PMCDBG3(PRC,EXT,1, "process-exit proc=%p (%d, %s)", p, p->p_pid, 5009 p->p_comm); 5010 5011 critical_enter(); /* no preemption */ 5012 5013 cpu = curthread->td_oncpu; 5014 5015 pp = pmc_find_process_descriptor(p, PMC_FLAG_REMOVE); 5016 if (pp == NULL) { 5017 critical_exit(); 5018 goto out; 5019 } 5020 5021 PMCDBG2(PRC,EXT,2, "process-exit proc=%p pmc-process=%p", p, pp); 5022 5023 /* 5024 * The exiting process could be the target of some PMCs which will be 5025 * running on currently executing CPU. 5026 * 5027 * We need to turn these PMCs off like we would do at context switch 5028 * OUT time. 5029 */ 5030 for (ri = 0; ri < md->pmd_npmc; ri++) { 5031 /* 5032 * Pick up the pmc pointer from hardware state similar to the 5033 * CSW_OUT code. 5034 */ 5035 pm = NULL; 5036 pcd = pmc_ri_to_classdep(md, ri, &adjri); 5037 5038 (void)(*pcd->pcd_get_config)(cpu, adjri, &pm); 5039 5040 PMCDBG2(PRC,EXT,2, "ri=%d pm=%p", ri, pm); 5041 5042 if (pm == NULL || !PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) 5043 continue; 5044 5045 PMCDBG4(PRC,EXT,2, "ppmcs[%d]=%p pm=%p state=%d", ri, 5046 pp->pp_pmcs[ri].pp_pmc, pm, pm->pm_state); 5047 5048 KASSERT(PMC_TO_ROWINDEX(pm) == ri, 5049 ("[pmc,%d] ri mismatch pmc(%d) ri(%d)", __LINE__, 5050 PMC_TO_ROWINDEX(pm), ri)); 5051 KASSERT(pm == pp->pp_pmcs[ri].pp_pmc, 5052 ("[pmc,%d] pm %p != pp_pmcs[%d] %p", __LINE__, pm, ri, 5053 pp->pp_pmcs[ri].pp_pmc)); 5054 KASSERT(counter_u64_fetch(pm->pm_runcount) > 0, 5055 ("[pmc,%d] bad runcount ri %d rc %ju", __LINE__, ri, 5056 (uintmax_t)counter_u64_fetch(pm->pm_runcount))); 5057 5058 /* 5059 * Change desired state, and then stop if not stalled. This 5060 * two-step dance should avoid race conditions where an 5061 * interrupt re-enables the PMC after this code has already 5062 * checked the pm_stalled flag. 5063 */ 5064 if (pm->pm_pcpu_state[cpu].pps_cpustate) { 5065 pm->pm_pcpu_state[cpu].pps_cpustate = 0; 5066 if (!pm->pm_pcpu_state[cpu].pps_stalled) { 5067 (void)pcd->pcd_stop_pmc(cpu, adjri, pm); 5068 5069 if (PMC_TO_MODE(pm) == PMC_MODE_TC) { 5070 pcd->pcd_read_pmc(cpu, adjri, pm, 5071 &newvalue); 5072 tmp = newvalue - PMC_PCPU_SAVED(cpu, ri); 5073 5074 mtx_pool_lock_spin(pmc_mtxpool, pm); 5075 pm->pm_gv.pm_savedvalue += tmp; 5076 pp->pp_pmcs[ri].pp_pmcval += tmp; 5077 mtx_pool_unlock_spin(pmc_mtxpool, pm); 5078 } 5079 } 5080 } 5081 5082 KASSERT(counter_u64_fetch(pm->pm_runcount) > 0, 5083 ("[pmc,%d] runcount is %d", __LINE__, ri)); 5084 5085 counter_u64_add(pm->pm_runcount, -1); 5086 (void)pcd->pcd_config_pmc(cpu, adjri, NULL); 5087 } 5088 5089 /* 5090 * Inform the MD layer of this pseudo "context switch out". 5091 */ 5092 (void)md->pmd_switch_out(pmc_pcpu[cpu], pp); 5093 5094 critical_exit(); /* ok to be pre-empted now */ 5095 5096 /* 5097 * Unlink this process from the PMCs that are targeting it. This will 5098 * send a signal to all PMC owner's whose PMCs are orphaned. 5099 * 5100 * Log PMC value at exit time if requested. 5101 */ 5102 for (ri = 0; ri < md->pmd_npmc; ri++) { 5103 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) { 5104 if ((pm->pm_flags & PMC_F_NEEDS_LOGFILE) != 0 && 5105 PMC_IS_COUNTING_MODE(PMC_TO_MODE(pm))) { 5106 pmclog_process_procexit(pm, pp); 5107 } 5108 pmc_unlink_target_process(pm, pp); 5109 } 5110 } 5111 free(pp, M_PMC); 5112 5113 out: 5114 /* 5115 * If the process owned PMCs, free them up and free up memory. 5116 */ 5117 if ((po = pmc_find_owner_descriptor(p)) != NULL) { 5118 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) != 0) 5119 pmclog_close(po); 5120 pmc_remove_owner(po); 5121 pmc_destroy_owner_descriptor(po); 5122 } 5123 5124 sx_xunlock(&pmc_sx); 5125 } 5126 5127 /* 5128 * Handle a process fork. 5129 * 5130 * If the parent process 'p1' is under HWPMC monitoring, then copy 5131 * over any attached PMCs that have 'do_descendants' semantics. 5132 */ 5133 static void 5134 pmc_process_fork(void *arg __unused, struct proc *p1, struct proc *newproc, 5135 int flags __unused) 5136 { 5137 struct pmc *pm; 5138 struct pmc_owner *po; 5139 struct pmc_process *ppnew, *ppold; 5140 unsigned int ri; 5141 bool is_using_hwpmcs, do_descendants; 5142 5143 PROC_LOCK(p1); 5144 is_using_hwpmcs = (p1->p_flag & P_HWPMC) != 0; 5145 PROC_UNLOCK(p1); 5146 5147 /* 5148 * If there are system-wide sampling PMCs active, we need to 5149 * log all fork events to their owner's logs. 5150 */ 5151 PMC_EPOCH_ENTER(); 5152 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) { 5153 if (po->po_flags & PMC_PO_OWNS_LOGFILE) { 5154 pmclog_process_procfork(po, p1->p_pid, newproc->p_pid); 5155 pmclog_process_proccreate(po, newproc, 1); 5156 } 5157 } 5158 PMC_EPOCH_EXIT(); 5159 5160 if (!is_using_hwpmcs) 5161 return; 5162 5163 PMC_GET_SX_XLOCK(); 5164 PMCDBG4(PMC,FRK,1, "process-fork proc=%p (%d, %s) -> %p", p1, 5165 p1->p_pid, p1->p_comm, newproc); 5166 5167 /* 5168 * If the parent process (curthread->td_proc) is a 5169 * target of any PMCs, look for PMCs that are to be 5170 * inherited, and link these into the new process 5171 * descriptor. 5172 */ 5173 ppold = pmc_find_process_descriptor(curthread->td_proc, PMC_FLAG_NONE); 5174 if (ppold == NULL) 5175 goto done; /* nothing to do */ 5176 5177 do_descendants = false; 5178 for (ri = 0; ri < md->pmd_npmc; ri++) { 5179 if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL && 5180 (pm->pm_flags & PMC_F_DESCENDANTS) != 0) { 5181 do_descendants = true; 5182 break; 5183 } 5184 } 5185 if (!do_descendants) /* nothing to do */ 5186 goto done; 5187 5188 /* 5189 * Now mark the new process as being tracked by this driver. 5190 */ 5191 PROC_LOCK(newproc); 5192 newproc->p_flag |= P_HWPMC; 5193 PROC_UNLOCK(newproc); 5194 5195 /* Allocate a descriptor for the new process. */ 5196 ppnew = pmc_find_process_descriptor(newproc, PMC_FLAG_ALLOCATE); 5197 if (ppnew == NULL) 5198 goto done; 5199 5200 /* 5201 * Run through all PMCs that were targeting the old process 5202 * and which specified F_DESCENDANTS and attach them to the 5203 * new process. 5204 * 5205 * Log the fork event to all owners of PMCs attached to this 5206 * process, if not already logged. 5207 */ 5208 for (ri = 0; ri < md->pmd_npmc; ri++) { 5209 if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL && 5210 (pm->pm_flags & PMC_F_DESCENDANTS) != 0) { 5211 pmc_link_target_process(pm, ppnew); 5212 po = pm->pm_owner; 5213 if (po->po_sscount == 0 && 5214 (po->po_flags & PMC_PO_OWNS_LOGFILE) != 0) { 5215 pmclog_process_procfork(po, p1->p_pid, 5216 newproc->p_pid); 5217 } 5218 } 5219 } 5220 5221 done: 5222 sx_xunlock(&pmc_sx); 5223 } 5224 5225 static void 5226 pmc_process_threadcreate(struct thread *td) 5227 { 5228 struct pmc_owner *po; 5229 5230 PMC_EPOCH_ENTER(); 5231 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) { 5232 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) != 0) 5233 pmclog_process_threadcreate(po, td, 1); 5234 } 5235 PMC_EPOCH_EXIT(); 5236 } 5237 5238 static void 5239 pmc_process_threadexit(struct thread *td) 5240 { 5241 struct pmc_owner *po; 5242 5243 PMC_EPOCH_ENTER(); 5244 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) { 5245 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) != 0) 5246 pmclog_process_threadexit(po, td); 5247 } 5248 PMC_EPOCH_EXIT(); 5249 } 5250 5251 static void 5252 pmc_process_proccreate(struct proc *p) 5253 { 5254 struct pmc_owner *po; 5255 5256 PMC_EPOCH_ENTER(); 5257 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) { 5258 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) != 0) 5259 pmclog_process_proccreate(po, p, 1 /* sync */); 5260 } 5261 PMC_EPOCH_EXIT(); 5262 } 5263 5264 static void 5265 pmc_process_allproc(struct pmc *pm) 5266 { 5267 struct pmc_owner *po; 5268 struct thread *td; 5269 struct proc *p; 5270 5271 po = pm->pm_owner; 5272 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) 5273 return; 5274 5275 sx_slock(&allproc_lock); 5276 FOREACH_PROC_IN_SYSTEM(p) { 5277 pmclog_process_proccreate(po, p, 0 /* sync */); 5278 PROC_LOCK(p); 5279 FOREACH_THREAD_IN_PROC(p, td) 5280 pmclog_process_threadcreate(po, td, 0 /* sync */); 5281 PROC_UNLOCK(p); 5282 } 5283 sx_sunlock(&allproc_lock); 5284 pmclog_flush(po, 0); 5285 } 5286 5287 static void 5288 pmc_kld_load(void *arg __unused, linker_file_t lf) 5289 { 5290 struct pmc_owner *po; 5291 5292 /* 5293 * Notify owners of system sampling PMCs about KLD operations. 5294 */ 5295 PMC_EPOCH_ENTER(); 5296 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) { 5297 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 5298 pmclog_process_map_in(po, (pid_t) -1, 5299 (uintfptr_t) lf->address, lf->pathname); 5300 } 5301 PMC_EPOCH_EXIT(); 5302 5303 /* 5304 * TODO: Notify owners of (all) process-sampling PMCs too. 5305 */ 5306 } 5307 5308 static void 5309 pmc_kld_unload(void *arg __unused, const char *filename __unused, 5310 caddr_t address, size_t size) 5311 { 5312 struct pmc_owner *po; 5313 5314 PMC_EPOCH_ENTER(); 5315 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) { 5316 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) != 0) { 5317 pmclog_process_map_out(po, (pid_t)-1, 5318 (uintfptr_t)address, (uintfptr_t)address + size); 5319 } 5320 } 5321 PMC_EPOCH_EXIT(); 5322 5323 /* 5324 * TODO: Notify owners of process-sampling PMCs. 5325 */ 5326 } 5327 5328 /* 5329 * initialization 5330 */ 5331 static const char * 5332 pmc_name_of_pmcclass(enum pmc_class class) 5333 { 5334 5335 switch (class) { 5336 #undef __PMC_CLASS 5337 #define __PMC_CLASS(S,V,D) \ 5338 case PMC_CLASS_##S: \ 5339 return #S; 5340 __PMC_CLASSES(); 5341 default: 5342 return ("<unknown>"); 5343 } 5344 } 5345 5346 /* 5347 * Base class initializer: allocate structure and set default classes. 5348 */ 5349 struct pmc_mdep * 5350 pmc_mdep_alloc(int nclasses) 5351 { 5352 struct pmc_mdep *md; 5353 int n; 5354 5355 /* SOFT + md classes */ 5356 n = 1 + nclasses; 5357 md = malloc(sizeof(struct pmc_mdep) + n * sizeof(struct pmc_classdep), 5358 M_PMC, M_WAITOK | M_ZERO); 5359 md->pmd_nclass = n; 5360 5361 /* Default methods */ 5362 md->pmd_switch_in = generic_switch_in; 5363 md->pmd_switch_out = generic_switch_out; 5364 5365 /* Add base class. */ 5366 pmc_soft_initialize(md); 5367 return (md); 5368 } 5369 5370 void 5371 pmc_mdep_free(struct pmc_mdep *md) 5372 { 5373 pmc_soft_finalize(md); 5374 free(md, M_PMC); 5375 } 5376 5377 static int 5378 generic_switch_in(struct pmc_cpu *pc __unused, struct pmc_process *pp __unused) 5379 { 5380 5381 return (0); 5382 } 5383 5384 static int 5385 generic_switch_out(struct pmc_cpu *pc __unused, struct pmc_process *pp __unused) 5386 { 5387 5388 return (0); 5389 } 5390 5391 static struct pmc_mdep * 5392 pmc_generic_cpu_initialize(void) 5393 { 5394 struct pmc_mdep *md; 5395 5396 md = pmc_mdep_alloc(0); 5397 5398 md->pmd_cputype = PMC_CPU_GENERIC; 5399 5400 return (md); 5401 } 5402 5403 static void 5404 pmc_generic_cpu_finalize(struct pmc_mdep *md __unused) 5405 { 5406 5407 } 5408 5409 static int 5410 pmc_initialize(void) 5411 { 5412 struct pcpu *pc; 5413 struct pmc_binding pb; 5414 struct pmc_classdep *pcd; 5415 struct pmc_sample *ps; 5416 struct pmc_samplebuffer *sb; 5417 int c, cpu, error, n, ri; 5418 u_int maxcpu, domain; 5419 5420 md = NULL; 5421 error = 0; 5422 5423 pmc_stats.pm_intr_ignored = counter_u64_alloc(M_WAITOK); 5424 pmc_stats.pm_intr_processed = counter_u64_alloc(M_WAITOK); 5425 pmc_stats.pm_intr_bufferfull = counter_u64_alloc(M_WAITOK); 5426 pmc_stats.pm_syscalls = counter_u64_alloc(M_WAITOK); 5427 pmc_stats.pm_syscall_errors = counter_u64_alloc(M_WAITOK); 5428 pmc_stats.pm_buffer_requests = counter_u64_alloc(M_WAITOK); 5429 pmc_stats.pm_buffer_requests_failed = counter_u64_alloc(M_WAITOK); 5430 pmc_stats.pm_log_sweeps = counter_u64_alloc(M_WAITOK); 5431 pmc_stats.pm_merges = counter_u64_alloc(M_WAITOK); 5432 pmc_stats.pm_overwrites = counter_u64_alloc(M_WAITOK); 5433 5434 #ifdef HWPMC_DEBUG 5435 /* parse debug flags first */ 5436 if (TUNABLE_STR_FETCH(PMC_SYSCTL_NAME_PREFIX "debugflags", 5437 pmc_debugstr, sizeof(pmc_debugstr))) { 5438 pmc_debugflags_parse(pmc_debugstr, pmc_debugstr + 5439 strlen(pmc_debugstr)); 5440 } 5441 #endif 5442 5443 PMCDBG1(MOD,INI,0, "PMC Initialize (version %x)", PMC_VERSION); 5444 5445 /* check kernel version */ 5446 if (pmc_kernel_version != PMC_VERSION) { 5447 if (pmc_kernel_version == 0) 5448 printf("hwpmc: this kernel has not been compiled with " 5449 "'options HWPMC_HOOKS'.\n"); 5450 else 5451 printf("hwpmc: kernel version (0x%x) does not match " 5452 "module version (0x%x).\n", pmc_kernel_version, 5453 PMC_VERSION); 5454 return (EPROGMISMATCH); 5455 } 5456 5457 /* 5458 * check sysctl parameters 5459 */ 5460 if (pmc_hashsize <= 0) { 5461 printf("hwpmc: tunable \"hashsize\"=%d must be " 5462 "greater than zero.\n", pmc_hashsize); 5463 pmc_hashsize = PMC_HASH_SIZE; 5464 } 5465 5466 if (pmc_nsamples <= 0 || pmc_nsamples > 65535) { 5467 printf("hwpmc: tunable \"nsamples\"=%d out of " 5468 "range.\n", pmc_nsamples); 5469 pmc_nsamples = PMC_NSAMPLES; 5470 } 5471 pmc_sample_mask = pmc_nsamples - 1; 5472 5473 if (pmc_callchaindepth <= 0 || 5474 pmc_callchaindepth > PMC_CALLCHAIN_DEPTH_MAX) { 5475 printf("hwpmc: tunable \"callchaindepth\"=%d out of " 5476 "range - using %d.\n", pmc_callchaindepth, 5477 PMC_CALLCHAIN_DEPTH_MAX); 5478 pmc_callchaindepth = PMC_CALLCHAIN_DEPTH_MAX; 5479 } 5480 5481 md = pmc_md_initialize(); 5482 if (md == NULL) { 5483 /* Default to generic CPU. */ 5484 md = pmc_generic_cpu_initialize(); 5485 if (md == NULL) 5486 return (ENOSYS); 5487 } 5488 5489 /* 5490 * Refresh classes base ri. Optional classes may come in different 5491 * order. 5492 */ 5493 for (ri = c = 0; c < md->pmd_nclass; c++) { 5494 pcd = &md->pmd_classdep[c]; 5495 pcd->pcd_ri = ri; 5496 ri += pcd->pcd_num; 5497 } 5498 5499 KASSERT(md->pmd_nclass >= 1 && md->pmd_npmc >= 1, 5500 ("[pmc,%d] no classes or pmcs", __LINE__)); 5501 5502 /* Compute the map from row-indices to classdep pointers. */ 5503 pmc_rowindex_to_classdep = malloc(sizeof(struct pmc_classdep *) * 5504 md->pmd_npmc, M_PMC, M_WAITOK | M_ZERO); 5505 5506 for (n = 0; n < md->pmd_npmc; n++) 5507 pmc_rowindex_to_classdep[n] = NULL; 5508 5509 for (ri = c = 0; c < md->pmd_nclass; c++) { 5510 pcd = &md->pmd_classdep[c]; 5511 for (n = 0; n < pcd->pcd_num; n++, ri++) 5512 pmc_rowindex_to_classdep[ri] = pcd; 5513 } 5514 5515 KASSERT(ri == md->pmd_npmc, 5516 ("[pmc,%d] npmc miscomputed: ri=%d, md->npmc=%d", __LINE__, 5517 ri, md->pmd_npmc)); 5518 5519 maxcpu = pmc_cpu_max(); 5520 5521 /* allocate space for the per-cpu array */ 5522 pmc_pcpu = malloc(maxcpu * sizeof(struct pmc_cpu *), M_PMC, 5523 M_WAITOK | M_ZERO); 5524 5525 /* per-cpu 'saved values' for managing process-mode PMCs */ 5526 pmc_pcpu_saved = malloc(sizeof(pmc_value_t) * maxcpu * md->pmd_npmc, 5527 M_PMC, M_WAITOK); 5528 5529 /* Perform CPU-dependent initialization. */ 5530 pmc_save_cpu_binding(&pb); 5531 error = 0; 5532 for (cpu = 0; error == 0 && cpu < maxcpu; cpu++) { 5533 if (!pmc_cpu_is_active(cpu)) 5534 continue; 5535 pmc_select_cpu(cpu); 5536 pmc_pcpu[cpu] = malloc(sizeof(struct pmc_cpu) + 5537 md->pmd_npmc * sizeof(struct pmc_hw *), M_PMC, 5538 M_WAITOK | M_ZERO); 5539 for (n = 0; error == 0 && n < md->pmd_nclass; n++) 5540 if (md->pmd_classdep[n].pcd_num > 0) 5541 error = md->pmd_classdep[n].pcd_pcpu_init(md, 5542 cpu); 5543 } 5544 pmc_restore_cpu_binding(&pb); 5545 5546 if (error != 0) 5547 return (error); 5548 5549 /* allocate space for the sample array */ 5550 for (cpu = 0; cpu < maxcpu; cpu++) { 5551 if (!pmc_cpu_is_active(cpu)) 5552 continue; 5553 pc = pcpu_find(cpu); 5554 domain = pc->pc_domain; 5555 sb = malloc_domainset(sizeof(struct pmc_samplebuffer) + 5556 pmc_nsamples * sizeof(struct pmc_sample), M_PMC, 5557 DOMAINSET_PREF(domain), M_WAITOK | M_ZERO); 5558 5559 KASSERT(pmc_pcpu[cpu] != NULL, 5560 ("[pmc,%d] cpu=%d Null per-cpu data", __LINE__, cpu)); 5561 5562 sb->ps_callchains = malloc_domainset(pmc_callchaindepth * 5563 pmc_nsamples * sizeof(uintptr_t), M_PMC, 5564 DOMAINSET_PREF(domain), M_WAITOK | M_ZERO); 5565 5566 for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++) 5567 ps->ps_pc = sb->ps_callchains + 5568 (n * pmc_callchaindepth); 5569 5570 pmc_pcpu[cpu]->pc_sb[PMC_HR] = sb; 5571 5572 sb = malloc_domainset(sizeof(struct pmc_samplebuffer) + 5573 pmc_nsamples * sizeof(struct pmc_sample), M_PMC, 5574 DOMAINSET_PREF(domain), M_WAITOK | M_ZERO); 5575 5576 sb->ps_callchains = malloc_domainset(pmc_callchaindepth * 5577 pmc_nsamples * sizeof(uintptr_t), M_PMC, 5578 DOMAINSET_PREF(domain), M_WAITOK | M_ZERO); 5579 for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++) 5580 ps->ps_pc = sb->ps_callchains + 5581 (n * pmc_callchaindepth); 5582 5583 pmc_pcpu[cpu]->pc_sb[PMC_SR] = sb; 5584 5585 sb = malloc_domainset(sizeof(struct pmc_samplebuffer) + 5586 pmc_nsamples * sizeof(struct pmc_sample), M_PMC, 5587 DOMAINSET_PREF(domain), M_WAITOK | M_ZERO); 5588 sb->ps_callchains = malloc_domainset(pmc_callchaindepth * 5589 pmc_nsamples * sizeof(uintptr_t), M_PMC, 5590 DOMAINSET_PREF(domain), M_WAITOK | M_ZERO); 5591 for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++) 5592 ps->ps_pc = sb->ps_callchains + n * pmc_callchaindepth; 5593 5594 pmc_pcpu[cpu]->pc_sb[PMC_UR] = sb; 5595 } 5596 5597 /* allocate space for the row disposition array */ 5598 pmc_pmcdisp = malloc(sizeof(enum pmc_mode) * md->pmd_npmc, 5599 M_PMC, M_WAITOK | M_ZERO); 5600 5601 /* mark all PMCs as available */ 5602 for (n = 0; n < md->pmd_npmc; n++) 5603 PMC_MARK_ROW_FREE(n); 5604 5605 /* allocate thread hash tables */ 5606 pmc_ownerhash = hashinit(pmc_hashsize, M_PMC, 5607 &pmc_ownerhashmask); 5608 5609 pmc_processhash = hashinit(pmc_hashsize, M_PMC, 5610 &pmc_processhashmask); 5611 mtx_init(&pmc_processhash_mtx, "pmc-process-hash", "pmc-leaf", 5612 MTX_SPIN); 5613 5614 CK_LIST_INIT(&pmc_ss_owners); 5615 pmc_ss_count = 0; 5616 5617 /* allocate a pool of spin mutexes */ 5618 pmc_mtxpool = mtx_pool_create("pmc-leaf", pmc_mtxpool_size, 5619 MTX_SPIN); 5620 5621 PMCDBG4(MOD,INI,1, "pmc_ownerhash=%p, mask=0x%lx " 5622 "targethash=%p mask=0x%lx", pmc_ownerhash, pmc_ownerhashmask, 5623 pmc_processhash, pmc_processhashmask); 5624 5625 /* Initialize a spin mutex for the thread free list. */ 5626 mtx_init(&pmc_threadfreelist_mtx, "pmc-threadfreelist", "pmc-leaf", 5627 MTX_SPIN); 5628 5629 /* Initialize the task to prune the thread free list. */ 5630 TASK_INIT(&free_task, 0, pmc_thread_descriptor_pool_free_task, NULL); 5631 5632 /* register process {exit,fork,exec} handlers */ 5633 pmc_exit_tag = EVENTHANDLER_REGISTER(process_exit, 5634 pmc_process_exit, NULL, EVENTHANDLER_PRI_ANY); 5635 pmc_fork_tag = EVENTHANDLER_REGISTER(process_fork, 5636 pmc_process_fork, NULL, EVENTHANDLER_PRI_ANY); 5637 5638 /* register kld event handlers */ 5639 pmc_kld_load_tag = EVENTHANDLER_REGISTER(kld_load, pmc_kld_load, 5640 NULL, EVENTHANDLER_PRI_ANY); 5641 pmc_kld_unload_tag = EVENTHANDLER_REGISTER(kld_unload, pmc_kld_unload, 5642 NULL, EVENTHANDLER_PRI_ANY); 5643 5644 /* initialize logging */ 5645 pmclog_initialize(); 5646 5647 /* set hook functions */ 5648 pmc_intr = md->pmd_intr; 5649 wmb(); 5650 pmc_hook = pmc_hook_handler; 5651 5652 if (error == 0) { 5653 printf(PMC_MODULE_NAME ":"); 5654 for (n = 0; n < md->pmd_nclass; n++) { 5655 if (md->pmd_classdep[n].pcd_num == 0) 5656 continue; 5657 pcd = &md->pmd_classdep[n]; 5658 printf(" %s/%d/%d/0x%b", 5659 pmc_name_of_pmcclass(pcd->pcd_class), 5660 pcd->pcd_num, 5661 pcd->pcd_width, 5662 pcd->pcd_caps, 5663 "\20" 5664 "\1INT\2USR\3SYS\4EDG\5THR" 5665 "\6REA\7WRI\10INV\11QUA\12PRC" 5666 "\13TAG\14CSC"); 5667 } 5668 printf("\n"); 5669 } 5670 5671 return (error); 5672 } 5673 5674 /* prepare to be unloaded */ 5675 static void 5676 pmc_cleanup(void) 5677 { 5678 struct pmc_binding pb; 5679 struct pmc_owner *po, *tmp; 5680 struct pmc_ownerhash *ph; 5681 struct pmc_processhash *prh __pmcdbg_used; 5682 u_int maxcpu; 5683 int cpu, c; 5684 5685 PMCDBG0(MOD,INI,0, "cleanup"); 5686 5687 /* switch off sampling */ 5688 CPU_FOREACH(cpu) 5689 DPCPU_ID_SET(cpu, pmc_sampled, 0); 5690 pmc_intr = NULL; 5691 5692 sx_xlock(&pmc_sx); 5693 if (pmc_hook == NULL) { /* being unloaded already */ 5694 sx_xunlock(&pmc_sx); 5695 return; 5696 } 5697 5698 pmc_hook = NULL; /* prevent new threads from entering module */ 5699 5700 /* deregister event handlers */ 5701 EVENTHANDLER_DEREGISTER(process_fork, pmc_fork_tag); 5702 EVENTHANDLER_DEREGISTER(process_exit, pmc_exit_tag); 5703 EVENTHANDLER_DEREGISTER(kld_load, pmc_kld_load_tag); 5704 EVENTHANDLER_DEREGISTER(kld_unload, pmc_kld_unload_tag); 5705 5706 /* send SIGBUS to all owner threads, free up allocations */ 5707 if (pmc_ownerhash != NULL) { 5708 for (ph = pmc_ownerhash; 5709 ph <= &pmc_ownerhash[pmc_ownerhashmask]; 5710 ph++) { 5711 LIST_FOREACH_SAFE(po, ph, po_next, tmp) { 5712 pmc_remove_owner(po); 5713 5714 PMCDBG3(MOD,INI,2, 5715 "cleanup signal proc=%p (%d, %s)", 5716 po->po_owner, po->po_owner->p_pid, 5717 po->po_owner->p_comm); 5718 5719 PROC_LOCK(po->po_owner); 5720 kern_psignal(po->po_owner, SIGBUS); 5721 PROC_UNLOCK(po->po_owner); 5722 5723 pmc_destroy_owner_descriptor(po); 5724 } 5725 } 5726 } 5727 5728 /* reclaim allocated data structures */ 5729 taskqueue_drain(taskqueue_fast, &free_task); 5730 mtx_destroy(&pmc_threadfreelist_mtx); 5731 pmc_thread_descriptor_pool_drain(); 5732 5733 if (pmc_mtxpool != NULL) 5734 mtx_pool_destroy(&pmc_mtxpool); 5735 5736 mtx_destroy(&pmc_processhash_mtx); 5737 if (pmc_processhash != NULL) { 5738 #ifdef HWPMC_DEBUG 5739 struct pmc_process *pp; 5740 5741 PMCDBG0(MOD,INI,3, "destroy process hash"); 5742 for (prh = pmc_processhash; 5743 prh <= &pmc_processhash[pmc_processhashmask]; 5744 prh++) 5745 LIST_FOREACH(pp, prh, pp_next) 5746 PMCDBG1(MOD,INI,3, "pid=%d", pp->pp_proc->p_pid); 5747 #endif 5748 5749 hashdestroy(pmc_processhash, M_PMC, pmc_processhashmask); 5750 pmc_processhash = NULL; 5751 } 5752 5753 if (pmc_ownerhash != NULL) { 5754 PMCDBG0(MOD,INI,3, "destroy owner hash"); 5755 hashdestroy(pmc_ownerhash, M_PMC, pmc_ownerhashmask); 5756 pmc_ownerhash = NULL; 5757 } 5758 5759 KASSERT(CK_LIST_EMPTY(&pmc_ss_owners), 5760 ("[pmc,%d] Global SS owner list not empty", __LINE__)); 5761 KASSERT(pmc_ss_count == 0, 5762 ("[pmc,%d] Global SS count not empty", __LINE__)); 5763 5764 /* do processor and pmc-class dependent cleanup */ 5765 maxcpu = pmc_cpu_max(); 5766 5767 PMCDBG0(MOD,INI,3, "md cleanup"); 5768 if (md) { 5769 pmc_save_cpu_binding(&pb); 5770 for (cpu = 0; cpu < maxcpu; cpu++) { 5771 PMCDBG2(MOD,INI,1,"pmc-cleanup cpu=%d pcs=%p", 5772 cpu, pmc_pcpu[cpu]); 5773 if (!pmc_cpu_is_active(cpu) || pmc_pcpu[cpu] == NULL) 5774 continue; 5775 5776 pmc_select_cpu(cpu); 5777 for (c = 0; c < md->pmd_nclass; c++) { 5778 if (md->pmd_classdep[c].pcd_num > 0) { 5779 md->pmd_classdep[c].pcd_pcpu_fini(md, 5780 cpu); 5781 } 5782 } 5783 } 5784 5785 if (md->pmd_cputype == PMC_CPU_GENERIC) 5786 pmc_generic_cpu_finalize(md); 5787 else 5788 pmc_md_finalize(md); 5789 5790 pmc_mdep_free(md); 5791 md = NULL; 5792 pmc_restore_cpu_binding(&pb); 5793 } 5794 5795 /* Free per-cpu descriptors. */ 5796 for (cpu = 0; cpu < maxcpu; cpu++) { 5797 if (!pmc_cpu_is_active(cpu)) 5798 continue; 5799 KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_HR] != NULL, 5800 ("[pmc,%d] Null hw cpu sample buffer cpu=%d", __LINE__, 5801 cpu)); 5802 KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_SR] != NULL, 5803 ("[pmc,%d] Null sw cpu sample buffer cpu=%d", __LINE__, 5804 cpu)); 5805 KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_UR] != NULL, 5806 ("[pmc,%d] Null userret cpu sample buffer cpu=%d", __LINE__, 5807 cpu)); 5808 free(pmc_pcpu[cpu]->pc_sb[PMC_HR]->ps_callchains, M_PMC); 5809 free(pmc_pcpu[cpu]->pc_sb[PMC_HR], M_PMC); 5810 free(pmc_pcpu[cpu]->pc_sb[PMC_SR]->ps_callchains, M_PMC); 5811 free(pmc_pcpu[cpu]->pc_sb[PMC_SR], M_PMC); 5812 free(pmc_pcpu[cpu]->pc_sb[PMC_UR]->ps_callchains, M_PMC); 5813 free(pmc_pcpu[cpu]->pc_sb[PMC_UR], M_PMC); 5814 free(pmc_pcpu[cpu], M_PMC); 5815 } 5816 5817 free(pmc_pcpu, M_PMC); 5818 pmc_pcpu = NULL; 5819 5820 free(pmc_pcpu_saved, M_PMC); 5821 pmc_pcpu_saved = NULL; 5822 5823 if (pmc_pmcdisp != NULL) { 5824 free(pmc_pmcdisp, M_PMC); 5825 pmc_pmcdisp = NULL; 5826 } 5827 5828 if (pmc_rowindex_to_classdep != NULL) { 5829 free(pmc_rowindex_to_classdep, M_PMC); 5830 pmc_rowindex_to_classdep = NULL; 5831 } 5832 5833 pmclog_shutdown(); 5834 counter_u64_free(pmc_stats.pm_intr_ignored); 5835 counter_u64_free(pmc_stats.pm_intr_processed); 5836 counter_u64_free(pmc_stats.pm_intr_bufferfull); 5837 counter_u64_free(pmc_stats.pm_syscalls); 5838 counter_u64_free(pmc_stats.pm_syscall_errors); 5839 counter_u64_free(pmc_stats.pm_buffer_requests); 5840 counter_u64_free(pmc_stats.pm_buffer_requests_failed); 5841 counter_u64_free(pmc_stats.pm_log_sweeps); 5842 counter_u64_free(pmc_stats.pm_merges); 5843 counter_u64_free(pmc_stats.pm_overwrites); 5844 sx_xunlock(&pmc_sx); /* we are done */ 5845 } 5846 5847 /* 5848 * The function called at load/unload. 5849 */ 5850 static int 5851 load(struct module *module __unused, int cmd, void *arg __unused) 5852 { 5853 int error; 5854 5855 error = 0; 5856 5857 switch (cmd) { 5858 case MOD_LOAD: 5859 /* initialize the subsystem */ 5860 error = pmc_initialize(); 5861 if (error != 0) 5862 break; 5863 PMCDBG2(MOD,INI,1, "syscall=%d maxcpu=%d", pmc_syscall_num, 5864 pmc_cpu_max()); 5865 break; 5866 case MOD_UNLOAD: 5867 case MOD_SHUTDOWN: 5868 pmc_cleanup(); 5869 PMCDBG0(MOD,INI,1, "unloaded"); 5870 break; 5871 default: 5872 error = EINVAL; 5873 break; 5874 } 5875 5876 return (error); 5877 } 5878