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 #define EXTERR_CATEGORY EXTERR_CAT_HWPMC_MOD 66 #include <sys/exterrvar.h> 67 68 #include <sys/linker.h> /* needs to be after <sys/malloc.h> */ 69 70 #include <machine/atomic.h> 71 #include <machine/md_var.h> 72 73 #include <vm/vm.h> 74 #include <vm/vm_extern.h> 75 #include <vm/pmap.h> 76 #include <vm/vm_map.h> 77 #include <vm/vm_object.h> 78 79 #include "hwpmc_soft.h" 80 81 #define PMC_EPOCH_ENTER() \ 82 struct epoch_tracker pmc_et; \ 83 epoch_enter_preempt(global_epoch_preempt, &pmc_et) 84 85 #define PMC_EPOCH_EXIT() \ 86 epoch_exit_preempt(global_epoch_preempt, &pmc_et) 87 88 /* 89 * Types 90 */ 91 92 enum pmc_flags { 93 PMC_FLAG_NONE = 0x00, /* do nothing */ 94 PMC_FLAG_REMOVE = 0x01, /* atomically remove entry from hash */ 95 PMC_FLAG_ALLOCATE = 0x02, /* add entry to hash if not found */ 96 PMC_FLAG_NOWAIT = 0x04, /* do not wait for mallocs */ 97 }; 98 99 /* 100 * The offset in sysent where the syscall is allocated. 101 */ 102 static int pmc_syscall_num = NO_SYSCALL; 103 104 struct pmc_cpu **pmc_pcpu; /* per-cpu state */ 105 pmc_value_t *pmc_pcpu_saved; /* saved PMC values: CSW handling */ 106 107 #define PMC_PCPU_SAVED(C, R) pmc_pcpu_saved[(R) + md->pmd_npmc * (C)] 108 109 struct mtx_pool *pmc_mtxpool; 110 static int *pmc_pmcdisp; /* PMC row dispositions */ 111 112 #define PMC_ROW_DISP_IS_FREE(R) (pmc_pmcdisp[(R)] == 0) 113 #define PMC_ROW_DISP_IS_THREAD(R) (pmc_pmcdisp[(R)] > 0) 114 #define PMC_ROW_DISP_IS_STANDALONE(R) (pmc_pmcdisp[(R)] < 0) 115 116 #define PMC_MARK_ROW_FREE(R) do { \ 117 pmc_pmcdisp[(R)] = 0; \ 118 } while (0) 119 120 #define PMC_MARK_ROW_STANDALONE(R) do { \ 121 KASSERT(pmc_pmcdisp[(R)] <= 0, ("[pmc,%d] row disposition error", \ 122 __LINE__)); \ 123 atomic_add_int(&pmc_pmcdisp[(R)], -1); \ 124 KASSERT(pmc_pmcdisp[(R)] >= (-pmc_cpu_max_active()), \ 125 ("[pmc,%d] row disposition error", __LINE__)); \ 126 } while (0) 127 128 #define PMC_UNMARK_ROW_STANDALONE(R) do { \ 129 atomic_add_int(&pmc_pmcdisp[(R)], 1); \ 130 KASSERT(pmc_pmcdisp[(R)] <= 0, ("[pmc,%d] row disposition error", \ 131 __LINE__)); \ 132 } while (0) 133 134 #define PMC_MARK_ROW_THREAD(R) do { \ 135 KASSERT(pmc_pmcdisp[(R)] >= 0, ("[pmc,%d] row disposition error", \ 136 __LINE__)); \ 137 atomic_add_int(&pmc_pmcdisp[(R)], 1); \ 138 } while (0) 139 140 #define PMC_UNMARK_ROW_THREAD(R) do { \ 141 atomic_add_int(&pmc_pmcdisp[(R)], -1); \ 142 KASSERT(pmc_pmcdisp[(R)] >= 0, ("[pmc,%d] row disposition error", \ 143 __LINE__)); \ 144 } while (0) 145 146 /* various event handlers */ 147 static eventhandler_tag pmc_exit_tag, pmc_fork_tag, pmc_kld_load_tag, 148 pmc_kld_unload_tag; 149 150 /* Module statistics */ 151 struct pmc_driverstats pmc_stats; 152 153 /* Machine/processor dependent operations */ 154 static struct pmc_mdep *md; 155 156 /* 157 * Hash tables mapping owner processes and target threads to PMCs. 158 */ 159 struct mtx pmc_processhash_mtx; /* spin mutex */ 160 static u_long pmc_processhashmask; 161 static LIST_HEAD(pmc_processhash, pmc_process) *pmc_processhash; 162 163 /* 164 * Hash table of PMC owner descriptors. This table is protected by 165 * the shared PMC "sx" lock. 166 */ 167 static u_long pmc_ownerhashmask; 168 static LIST_HEAD(pmc_ownerhash, pmc_owner) *pmc_ownerhash; 169 170 /* 171 * List of PMC owners with system-wide sampling PMCs. 172 */ 173 static CK_LIST_HEAD(, pmc_owner) pmc_ss_owners; 174 175 /* 176 * List of free thread entries. This is protected by the spin 177 * mutex. 178 */ 179 static struct mtx pmc_threadfreelist_mtx; /* spin mutex */ 180 static LIST_HEAD(, pmc_thread) pmc_threadfreelist; 181 static int pmc_threadfreelist_entries = 0; 182 #define THREADENTRY_SIZE (sizeof(struct pmc_thread) + \ 183 (md->pmd_npmc * sizeof(struct pmc_threadpmcstate))) 184 185 /* 186 * Task to free thread descriptors 187 */ 188 static struct task free_task; 189 190 /* 191 * A map of row indices to classdep structures. 192 */ 193 static struct pmc_classdep **pmc_rowindex_to_classdep; 194 195 /* 196 * Prototypes 197 */ 198 199 #ifdef HWPMC_DEBUG 200 static int pmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS); 201 static int pmc_debugflags_parse(char *newstr, char *fence); 202 #endif 203 204 static void pmc_multipart_add(struct pmc_sample *ps, int type, 205 int length); 206 static void pmc_multipart_copydata(struct pmc_sample *ps, 207 struct pmc_multipart *mp); 208 209 static int load(struct module *module, int cmd, void *arg); 210 static int pmc_add_sample(ring_type_t ring, struct pmc *pm, 211 struct trapframe *tf, struct pmc_multipart *mp); 212 static void pmc_add_thread_descriptors_from_proc(struct proc *p, 213 struct pmc_process *pp); 214 static int pmc_attach_process(struct proc *p, struct pmc *pm); 215 static struct pmc *pmc_allocate_pmc_descriptor(void); 216 static struct pmc_owner *pmc_allocate_owner_descriptor(struct proc *p); 217 static int pmc_attach_one_process(struct proc *p, struct pmc *pm); 218 static bool pmc_can_allocate_row(int ri, enum pmc_mode mode); 219 static bool pmc_can_allocate_rowindex(struct proc *p, unsigned int ri, 220 int cpu); 221 static void pmc_capture_user_callchain(int cpu, int soft, 222 struct trapframe *tf); 223 static void pmc_cleanup(void); 224 static int pmc_detach_process(struct proc *p, struct pmc *pm); 225 static int pmc_detach_one_process(struct proc *p, struct pmc *pm, 226 int flags); 227 static void pmc_destroy_owner_descriptor(struct pmc_owner *po); 228 static void pmc_destroy_pmc_descriptor(struct pmc *pm); 229 static void pmc_destroy_process_descriptor(struct pmc_process *pp); 230 static void pmc_reclaim_pmc_from_cpu(struct pmc *pm, 231 struct pmc_process *pp, int cpu); 232 static struct pmc_owner *pmc_find_owner_descriptor(struct proc *p); 233 static int pmc_find_pmc(pmc_id_t pmcid, struct pmc **pm); 234 static struct pmc *pmc_find_pmc_descriptor_in_process(struct pmc_owner *po, 235 pmc_id_t pmc); 236 static struct pmc_process *pmc_find_process_descriptor(struct proc *p, 237 uint32_t mode); 238 static struct pmc_thread *pmc_find_thread_descriptor(struct pmc_process *pp, 239 struct thread *td, uint32_t mode); 240 static void pmc_force_context_switch(void); 241 static void pmc_link_target_process(struct pmc *pm, 242 struct pmc_process *pp); 243 static void pmc_log_all_process_mappings(struct pmc_owner *po); 244 static void pmc_log_kernel_mappings(struct pmc *pm); 245 static void pmc_log_process_mappings(struct pmc_owner *po, struct proc *p); 246 static void pmc_maybe_remove_owner(struct pmc_owner *po); 247 static void pmc_post_callchain_callback(void); 248 static void pmc_process_allproc(struct pmc *pm); 249 static void pmc_process_csw_in(struct thread *td); 250 static void pmc_process_csw_out(struct thread *td); 251 static void pmc_process_csw_start_all(int cpu); 252 static void pmc_process_csw_stop_all(int cpu); 253 static void pmc_process_exec(struct thread *td, 254 struct pmckern_procexec *pk); 255 static void pmc_process_exit(void *arg, struct proc *p); 256 static void pmc_process_fork(void *arg, struct proc *p1, 257 struct proc *p2, int n); 258 static void pmc_process_proccreate(struct proc *p); 259 static void pmc_process_samples(int cpu, ring_type_t soft); 260 static void pmc_process_threadcreate(struct thread *td); 261 static void pmc_process_threadexit(struct thread *td); 262 static void pmc_process_thread_add(struct thread *td); 263 static void pmc_process_thread_delete(struct thread *td); 264 static void pmc_process_thread_userret(struct thread *td); 265 static void pmc_release_pmc_descriptor(struct pmc *pmc); 266 static void pmc_remove_owner(struct pmc_owner *po); 267 static void pmc_remove_process_descriptor(struct pmc_process *pp); 268 static int pmc_start(struct pmc *pm); 269 static int pmc_stop(struct pmc *pm); 270 static int pmc_syscall_handler(struct thread *td, void *syscall_args); 271 static struct pmc_thread *pmc_thread_descriptor_pool_alloc(void); 272 static void pmc_thread_descriptor_pool_drain(void); 273 static void pmc_thread_descriptor_pool_free(struct pmc_thread *pt); 274 static void pmc_unlink_target_process(struct pmc *pmc, 275 struct pmc_process *pp); 276 277 static int generic_switch_in(struct pmc_cpu *pc, struct pmc_process *pp); 278 static int generic_switch_out(struct pmc_cpu *pc, struct pmc_process *pp); 279 static struct pmc_mdep *pmc_generic_cpu_initialize(void); 280 static void pmc_generic_cpu_finalize(struct pmc_mdep *md); 281 282 /* 283 * Kernel tunables and sysctl(8) interface. 284 */ 285 286 SYSCTL_DECL(_kern_hwpmc); 287 SYSCTL_NODE(_kern_hwpmc, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 288 "HWPMC stats"); 289 290 /* Stats. */ 291 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, intr_ignored, CTLFLAG_RW, 292 &pmc_stats.pm_intr_ignored, 293 "# of interrupts ignored"); 294 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, intr_processed, CTLFLAG_RW, 295 &pmc_stats.pm_intr_processed, 296 "# of interrupts processed"); 297 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, intr_bufferfull, CTLFLAG_RW, 298 &pmc_stats.pm_intr_bufferfull, 299 "# of interrupts where buffer was full"); 300 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, syscalls, CTLFLAG_RW, 301 &pmc_stats.pm_syscalls, 302 "# of syscalls"); 303 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, syscall_errors, CTLFLAG_RW, 304 &pmc_stats.pm_syscall_errors, 305 "# of syscall_errors"); 306 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, buffer_requests, CTLFLAG_RW, 307 &pmc_stats.pm_buffer_requests, 308 "# of buffer requests"); 309 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, buffer_requests_failed, 310 CTLFLAG_RW, &pmc_stats.pm_buffer_requests_failed, 311 "# of buffer requests which failed"); 312 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, log_sweeps, CTLFLAG_RW, 313 &pmc_stats.pm_log_sweeps, 314 "# of times samples were processed"); 315 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, merges, CTLFLAG_RW, 316 &pmc_stats.pm_merges, 317 "# of times kernel stack was found for user trace"); 318 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, overwrites, CTLFLAG_RW, 319 &pmc_stats.pm_overwrites, 320 "# of times a sample was overwritten before being logged"); 321 322 static int pmc_callchaindepth = PMC_CALLCHAIN_DEPTH; 323 SYSCTL_INT(_kern_hwpmc, OID_AUTO, callchaindepth, CTLFLAG_RDTUN, 324 &pmc_callchaindepth, 0, 325 "depth of call chain records"); 326 327 char pmc_cpuid[PMC_CPUID_LEN]; 328 SYSCTL_STRING(_kern_hwpmc, OID_AUTO, cpuid, CTLFLAG_RD, 329 pmc_cpuid, 0, 330 "cpu version string"); 331 332 #ifdef HWPMC_DEBUG 333 struct pmc_debugflags pmc_debugflags = PMC_DEBUG_DEFAULT_FLAGS; 334 char pmc_debugstr[PMC_DEBUG_STRSIZE]; 335 TUNABLE_STR(PMC_SYSCTL_NAME_PREFIX "debugflags", pmc_debugstr, 336 sizeof(pmc_debugstr)); 337 SYSCTL_PROC(_kern_hwpmc, OID_AUTO, debugflags, 338 CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, 339 0, 0, pmc_debugflags_sysctl_handler, "A", 340 "debug flags"); 341 #endif 342 343 /* 344 * kern.hwpmc.hashsize -- determines the number of rows in the 345 * of the hash table used to look up threads 346 */ 347 static int pmc_hashsize = PMC_HASH_SIZE; 348 SYSCTL_INT(_kern_hwpmc, OID_AUTO, hashsize, CTLFLAG_RDTUN, 349 &pmc_hashsize, 0, 350 "rows in hash tables"); 351 352 /* 353 * kern.hwpmc.nsamples --- number of PC samples/callchain stacks per CPU 354 */ 355 static int pmc_nsamples = PMC_NSAMPLES; 356 SYSCTL_INT(_kern_hwpmc, OID_AUTO, nsamples, CTLFLAG_RDTUN, 357 &pmc_nsamples, 0, 358 "number of PC samples per CPU"); 359 360 static uint64_t pmc_sample_mask = PMC_NSAMPLES - 1; 361 362 /* 363 * kern.hwpmc.mtxpoolsize -- number of mutexes in the mutex pool. 364 */ 365 static int pmc_mtxpool_size = PMC_MTXPOOL_SIZE; 366 SYSCTL_INT(_kern_hwpmc, OID_AUTO, mtxpoolsize, CTLFLAG_RDTUN, 367 &pmc_mtxpool_size, 0, 368 "size of spin mutex pool"); 369 370 /* 371 * kern.hwpmc.threadfreelist_entries -- number of free entries 372 */ 373 SYSCTL_INT(_kern_hwpmc, OID_AUTO, threadfreelist_entries, CTLFLAG_RD, 374 &pmc_threadfreelist_entries, 0, 375 "number of available thread entries"); 376 377 /* 378 * kern.hwpmc.threadfreelist_max -- maximum number of free entries 379 */ 380 static int pmc_threadfreelist_max = PMC_THREADLIST_MAX; 381 SYSCTL_INT(_kern_hwpmc, OID_AUTO, threadfreelist_max, CTLFLAG_RW, 382 &pmc_threadfreelist_max, 0, 383 "maximum number of available thread entries before freeing some"); 384 385 /* 386 * kern.hwpmc.mincount -- minimum sample count 387 */ 388 static u_int pmc_mincount = 1000; 389 SYSCTL_INT(_kern_hwpmc, OID_AUTO, mincount, CTLFLAG_RWTUN, 390 &pmc_mincount, 0, 391 "minimum count for sampling counters"); 392 393 /* 394 * security.bsd.unprivileged_syspmcs -- allow non-root processes to 395 * allocate system-wide PMCs. 396 * 397 * Allowing unprivileged processes to allocate system PMCs is convenient 398 * if system-wide measurements need to be taken concurrently with other 399 * per-process measurements. This feature is turned off by default. 400 */ 401 static int pmc_unprivileged_syspmcs = 0; 402 SYSCTL_INT(_security_bsd, OID_AUTO, unprivileged_syspmcs, CTLFLAG_RWTUN, 403 &pmc_unprivileged_syspmcs, 0, 404 "allow unprivileged process to allocate system PMCs"); 405 406 /* 407 * Hash function. Discard the lower 2 bits of the pointer since 408 * these are always zero for our uses. The hash multiplier is 409 * round((2^LONG_BIT) * ((sqrt(5)-1)/2)). 410 */ 411 #if LONG_BIT == 64 412 #define _PMC_HM 11400714819323198486u 413 #elif LONG_BIT == 32 414 #define _PMC_HM 2654435769u 415 #else 416 #error Must know the size of 'long' to compile 417 #endif 418 419 #define PMC_HASH_PTR(P,M) ((((unsigned long) (P) >> 2) * _PMC_HM) & (M)) 420 421 /* 422 * Syscall structures 423 */ 424 425 /* The `sysent' for the new syscall */ 426 static struct sysent pmc_sysent = { 427 .sy_narg = 2, 428 .sy_call = pmc_syscall_handler, 429 }; 430 431 static struct syscall_module_data pmc_syscall_mod = { 432 .chainevh = load, 433 .chainarg = NULL, 434 .offset = &pmc_syscall_num, 435 .new_sysent = &pmc_sysent, 436 .old_sysent = { .sy_narg = 0, .sy_call = NULL }, 437 .flags = SY_THR_STATIC_KLD, 438 }; 439 440 static moduledata_t pmc_mod = { 441 .name = PMC_MODULE_NAME, 442 .evhand = syscall_module_handler, 443 .priv = &pmc_syscall_mod, 444 }; 445 446 #ifdef EARLY_AP_STARTUP 447 DECLARE_MODULE(pmc, pmc_mod, SI_SUB_SYSCALLS, SI_ORDER_ANY); 448 #else 449 DECLARE_MODULE(pmc, pmc_mod, SI_SUB_SMP, SI_ORDER_ANY); 450 #endif 451 MODULE_VERSION(pmc, PMC_VERSION); 452 453 #ifdef HWPMC_DEBUG 454 enum pmc_dbgparse_state { 455 PMCDS_WS, /* in whitespace */ 456 PMCDS_MAJOR, /* seen a major keyword */ 457 PMCDS_MINOR 458 }; 459 460 static int 461 pmc_debugflags_parse(char *newstr, char *fence) 462 { 463 struct pmc_debugflags *tmpflags; 464 size_t kwlen; 465 char c, *p, *q; 466 int error, *newbits, tmp; 467 int found; 468 469 tmpflags = malloc(sizeof(*tmpflags), M_PMC, M_WAITOK | M_ZERO); 470 471 error = 0; 472 for (p = newstr; p < fence && (c = *p); p++) { 473 /* skip white space */ 474 if (c == ' ' || c == '\t') 475 continue; 476 477 /* look for a keyword followed by "=" */ 478 for (q = p; p < fence && (c = *p) && c != '='; p++) 479 ; 480 if (c != '=') { 481 error = EINVAL; 482 goto done; 483 } 484 485 kwlen = p - q; 486 newbits = NULL; 487 488 /* lookup flag group name */ 489 #define DBG_SET_FLAG_MAJ(S,F) \ 490 if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0) \ 491 newbits = &tmpflags->pdb_ ## F; 492 493 DBG_SET_FLAG_MAJ("cpu", CPU); 494 DBG_SET_FLAG_MAJ("csw", CSW); 495 DBG_SET_FLAG_MAJ("logging", LOG); 496 DBG_SET_FLAG_MAJ("module", MOD); 497 DBG_SET_FLAG_MAJ("md", MDP); 498 DBG_SET_FLAG_MAJ("owner", OWN); 499 DBG_SET_FLAG_MAJ("pmc", PMC); 500 DBG_SET_FLAG_MAJ("process", PRC); 501 DBG_SET_FLAG_MAJ("sampling", SAM); 502 #undef DBG_SET_FLAG_MAJ 503 504 if (newbits == NULL) { 505 error = EINVAL; 506 goto done; 507 } 508 509 p++; /* skip the '=' */ 510 511 /* Now parse the individual flags */ 512 tmp = 0; 513 newflag: 514 for (q = p; p < fence && (c = *p); p++) 515 if (c == ' ' || c == '\t' || c == ',') 516 break; 517 518 /* p == fence or c == ws or c == "," or c == 0 */ 519 520 if ((kwlen = p - q) == 0) { 521 *newbits = tmp; 522 continue; 523 } 524 525 found = 0; 526 #define DBG_SET_FLAG_MIN(S,F) \ 527 if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0) \ 528 tmp |= found = (1 << PMC_DEBUG_MIN_ ## F) 529 530 /* a '*' denotes all possible flags in the group */ 531 if (kwlen == 1 && *q == '*') 532 tmp = found = ~0; 533 /* look for individual flag names */ 534 DBG_SET_FLAG_MIN("allocaterow", ALR); 535 DBG_SET_FLAG_MIN("allocate", ALL); 536 DBG_SET_FLAG_MIN("attach", ATT); 537 DBG_SET_FLAG_MIN("bind", BND); 538 DBG_SET_FLAG_MIN("config", CFG); 539 DBG_SET_FLAG_MIN("exec", EXC); 540 DBG_SET_FLAG_MIN("exit", EXT); 541 DBG_SET_FLAG_MIN("find", FND); 542 DBG_SET_FLAG_MIN("flush", FLS); 543 DBG_SET_FLAG_MIN("fork", FRK); 544 DBG_SET_FLAG_MIN("getbuf", GTB); 545 DBG_SET_FLAG_MIN("hook", PMH); 546 DBG_SET_FLAG_MIN("init", INI); 547 DBG_SET_FLAG_MIN("intr", INT); 548 DBG_SET_FLAG_MIN("linktarget", TLK); 549 DBG_SET_FLAG_MIN("mayberemove", OMR); 550 DBG_SET_FLAG_MIN("ops", OPS); 551 DBG_SET_FLAG_MIN("read", REA); 552 DBG_SET_FLAG_MIN("register", REG); 553 DBG_SET_FLAG_MIN("release", REL); 554 DBG_SET_FLAG_MIN("remove", ORM); 555 DBG_SET_FLAG_MIN("sample", SAM); 556 DBG_SET_FLAG_MIN("scheduleio", SIO); 557 DBG_SET_FLAG_MIN("select", SEL); 558 DBG_SET_FLAG_MIN("signal", SIG); 559 DBG_SET_FLAG_MIN("swi", SWI); 560 DBG_SET_FLAG_MIN("swo", SWO); 561 DBG_SET_FLAG_MIN("start", STA); 562 DBG_SET_FLAG_MIN("stop", STO); 563 DBG_SET_FLAG_MIN("syscall", PMS); 564 DBG_SET_FLAG_MIN("unlinktarget", TUL); 565 DBG_SET_FLAG_MIN("write", WRI); 566 #undef DBG_SET_FLAG_MIN 567 if (found == 0) { 568 /* unrecognized flag name */ 569 error = EINVAL; 570 goto done; 571 } 572 573 if (c == 0 || c == ' ' || c == '\t') { /* end of flag group */ 574 *newbits = tmp; 575 continue; 576 } 577 578 p++; 579 goto newflag; 580 } 581 582 /* save the new flag set */ 583 bcopy(tmpflags, &pmc_debugflags, sizeof(pmc_debugflags)); 584 done: 585 free(tmpflags, M_PMC); 586 return (error); 587 } 588 589 static int 590 pmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS) 591 { 592 char *fence, *newstr; 593 int error; 594 u_int n; 595 596 n = sizeof(pmc_debugstr); 597 newstr = malloc(n, M_PMC, M_WAITOK | M_ZERO); 598 strlcpy(newstr, pmc_debugstr, n); 599 600 error = sysctl_handle_string(oidp, newstr, n, req); 601 602 /* if there is a new string, parse and copy it */ 603 if (error == 0 && req->newptr != NULL) { 604 fence = newstr + (n < req->newlen ? n : req->newlen + 1); 605 error = pmc_debugflags_parse(newstr, fence); 606 if (error == 0) 607 strlcpy(pmc_debugstr, newstr, sizeof(pmc_debugstr)); 608 } 609 free(newstr, M_PMC); 610 611 return (error); 612 } 613 #endif 614 615 /* 616 * Map a row index to a classdep structure and return the adjusted row 617 * index for the PMC class index. 618 */ 619 static struct pmc_classdep * 620 pmc_ri_to_classdep(struct pmc_mdep *md __unused, int ri, int *adjri) 621 { 622 struct pmc_classdep *pcd; 623 624 KASSERT(ri >= 0 && ri < md->pmd_npmc, 625 ("[pmc,%d] illegal row-index %d", __LINE__, ri)); 626 627 pcd = pmc_rowindex_to_classdep[ri]; 628 KASSERT(pcd != NULL, 629 ("[pmc,%d] ri %d null pcd", __LINE__, ri)); 630 631 *adjri = ri - pcd->pcd_ri; 632 KASSERT(*adjri >= 0 && *adjri < pcd->pcd_num, 633 ("[pmc,%d] adjusted row-index %d", __LINE__, *adjri)); 634 635 return (pcd); 636 } 637 638 /* 639 * Concurrency Control 640 * 641 * The driver manages the following data structures: 642 * 643 * - target process descriptors, one per target process 644 * - owner process descriptors (and attached lists), one per owner process 645 * - lookup hash tables for owner and target processes 646 * - PMC descriptors (and attached lists) 647 * - per-cpu hardware state 648 * - the 'hook' variable through which the kernel calls into 649 * this module 650 * - the machine hardware state (managed by the MD layer) 651 * 652 * These data structures are accessed from: 653 * 654 * - thread context-switch code 655 * - interrupt handlers (possibly on multiple cpus) 656 * - kernel threads on multiple cpus running on behalf of user 657 * processes doing system calls 658 * - this driver's private kernel threads 659 * 660 * = Locks and Locking strategy = 661 * 662 * The driver uses four locking strategies for its operation: 663 * 664 * - The global SX lock "pmc_sx" is used to protect internal 665 * data structures. 666 * 667 * Calls into the module by syscall() start with this lock being 668 * held in exclusive mode. Depending on the requested operation, 669 * the lock may be downgraded to 'shared' mode to allow more 670 * concurrent readers into the module. Calls into the module from 671 * other parts of the kernel acquire the lock in shared mode. 672 * 673 * This SX lock is held in exclusive mode for any operations that 674 * modify the linkages between the driver's internal data structures. 675 * 676 * The 'pmc_hook' function pointer is also protected by this lock. 677 * It is only examined with the sx lock held in exclusive mode. The 678 * kernel module is allowed to be unloaded only with the sx lock held 679 * in exclusive mode. In normal syscall handling, after acquiring the 680 * pmc_sx lock we first check that 'pmc_hook' is non-null before 681 * proceeding. This prevents races between the thread unloading the module 682 * and other threads seeking to use the module. 683 * 684 * - Lookups of target process structures and owner process structures 685 * cannot use the global "pmc_sx" SX lock because these lookups need 686 * to happen during context switches and in other critical sections 687 * where sleeping is not allowed. We protect these lookup tables 688 * with their own private spin-mutexes, "pmc_processhash_mtx" and 689 * "pmc_ownerhash_mtx". 690 * 691 * - Interrupt handlers work in a lock free manner. At interrupt 692 * time, handlers look at the PMC pointer (phw->phw_pmc) configured 693 * when the PMC was started. If this pointer is NULL, the interrupt 694 * is ignored after updating driver statistics. We ensure that this 695 * pointer is set (using an atomic operation if necessary) before the 696 * PMC hardware is started. Conversely, this pointer is unset atomically 697 * only after the PMC hardware is stopped. 698 * 699 * We ensure that everything needed for the operation of an 700 * interrupt handler is available without it needing to acquire any 701 * locks. We also ensure that a PMC's software state is destroyed only 702 * after the PMC is taken off hardware (on all CPUs). 703 * 704 * - Context-switch handling with process-private PMCs needs more 705 * care. 706 * 707 * A given process may be the target of multiple PMCs. For example, 708 * PMCATTACH and PMCDETACH may be requested by a process on one CPU 709 * while the target process is running on another. A PMC could also 710 * be getting released because its owner is exiting. We tackle 711 * these situations in the following manner: 712 * 713 * - each target process structure 'pmc_process' has an array 714 * of 'struct pmc *' pointers, one for each hardware PMC. 715 * 716 * - At context switch IN time, each "target" PMC in RUNNING state 717 * gets started on hardware and a pointer to each PMC is copied into 718 * the per-cpu phw array. The 'runcount' for the PMC is 719 * incremented. 720 * 721 * - At context switch OUT time, all process-virtual PMCs are stopped 722 * on hardware. The saved value is added to the PMCs value field 723 * only if the PMC is in a non-deleted state (the PMCs state could 724 * have changed during the current time slice). 725 * 726 * Note that since in-between a switch IN on a processor and a switch 727 * OUT, the PMC could have been released on another CPU. Therefore 728 * context switch OUT always looks at the hardware state to turn 729 * OFF PMCs and will update a PMC's saved value only if reachable 730 * from the target process record. 731 * 732 * - OP PMCRELEASE could be called on a PMC at any time (the PMC could 733 * be attached to many processes at the time of the call and could 734 * be active on multiple CPUs). 735 * 736 * We prevent further scheduling of the PMC by marking it as in 737 * state 'DELETED'. If the runcount of the PMC is non-zero then 738 * this PMC is currently running on a CPU somewhere. The thread 739 * doing the PMCRELEASE operation waits by repeatedly doing a 740 * pause() till the runcount comes to zero. 741 * 742 * The contents of a PMC descriptor (struct pmc) are protected using 743 * a spin-mutex. In order to save space, we use a mutex pool. 744 * 745 * In terms of lock types used by witness(4), we use: 746 * - Type "pmc-sx", used by the global SX lock. 747 * - Type "pmc-sleep", for sleep mutexes used by logger threads. 748 * - Type "pmc-per-proc", for protecting PMC owner descriptors. 749 * - Type "pmc-leaf", used for all other spin mutexes. 750 */ 751 752 /* 753 * Save the CPU binding of the current kthread. 754 */ 755 void 756 pmc_save_cpu_binding(struct pmc_binding *pb) 757 { 758 PMCDBG0(CPU,BND,2, "save-cpu"); 759 thread_lock(curthread); 760 pb->pb_bound = sched_is_bound(curthread); 761 pb->pb_cpu = curthread->td_oncpu; 762 pb->pb_priority = curthread->td_priority; 763 thread_unlock(curthread); 764 PMCDBG1(CPU,BND,2, "save-cpu cpu=%d", pb->pb_cpu); 765 } 766 767 /* 768 * Restore the CPU binding of the current thread. 769 */ 770 void 771 pmc_restore_cpu_binding(struct pmc_binding *pb) 772 { 773 PMCDBG2(CPU,BND,2, "restore-cpu curcpu=%d restore=%d", 774 curthread->td_oncpu, pb->pb_cpu); 775 thread_lock(curthread); 776 sched_bind(curthread, pb->pb_cpu); 777 if (!pb->pb_bound) 778 sched_unbind(curthread); 779 sched_prio(curthread, pb->pb_priority); 780 thread_unlock(curthread); 781 PMCDBG0(CPU,BND,2, "restore-cpu done"); 782 } 783 784 /* 785 * Move execution over to the specified CPU and bind it there. 786 */ 787 void 788 pmc_select_cpu(int cpu) 789 { 790 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), 791 ("[pmc,%d] bad cpu number %d", __LINE__, cpu)); 792 793 /* Never move to an inactive CPU. */ 794 KASSERT(pmc_cpu_is_active(cpu), ("[pmc,%d] selecting inactive " 795 "CPU %d", __LINE__, cpu)); 796 797 PMCDBG1(CPU,SEL,2, "select-cpu cpu=%d", cpu); 798 thread_lock(curthread); 799 sched_prio(curthread, PRI_MIN); 800 sched_bind(curthread, cpu); 801 thread_unlock(curthread); 802 803 KASSERT(curthread->td_oncpu == cpu, 804 ("[pmc,%d] CPU not bound [cpu=%d, curr=%d]", __LINE__, 805 cpu, curthread->td_oncpu)); 806 807 PMCDBG1(CPU,SEL,2, "select-cpu cpu=%d ok", cpu); 808 } 809 810 /* 811 * Force a context switch. 812 * 813 * We do this by pause'ing for 1 tick -- invoking mi_switch() is not 814 * guaranteed to force a context switch. 815 */ 816 static void 817 pmc_force_context_switch(void) 818 { 819 820 pause("pmcctx", 1); 821 } 822 823 uint64_t 824 pmc_rdtsc(void) 825 { 826 #if defined(__i386__) 827 /* Unfortunately get_cyclecount on i386 uses cpu_ticks. */ 828 return (rdtsc()); 829 #else 830 return (get_cyclecount()); 831 #endif 832 } 833 834 /* 835 * Get the file name for an executable. This is a simple wrapper 836 * around vn_fullpath(9). 837 */ 838 static void 839 pmc_getfilename(struct vnode *v, char **fullpath, char **freepath) 840 { 841 842 *fullpath = "unknown"; 843 *freepath = NULL; 844 vn_fullpath(v, fullpath, freepath); 845 } 846 847 /* 848 * Remove a process owning PMCs. 849 */ 850 void 851 pmc_remove_owner(struct pmc_owner *po) 852 { 853 struct pmc *pm, *tmp; 854 855 sx_assert(&pmc_sx, SX_XLOCKED); 856 857 PMCDBG1(OWN,ORM,1, "remove-owner po=%p", po); 858 859 /* Remove descriptor from the owner hash table */ 860 LIST_REMOVE(po, po_next); 861 862 /* release all owned PMC descriptors */ 863 LIST_FOREACH_SAFE(pm, &po->po_pmcs, pm_next, tmp) { 864 PMCDBG1(OWN,ORM,2, "pmc=%p", pm); 865 KASSERT(pm->pm_owner == po, 866 ("[pmc,%d] owner %p != po %p", __LINE__, pm->pm_owner, po)); 867 868 pmc_release_pmc_descriptor(pm); /* will unlink from the list */ 869 pmc_destroy_pmc_descriptor(pm); 870 } 871 872 KASSERT(po->po_sscount == 0, 873 ("[pmc,%d] SS count not zero", __LINE__)); 874 KASSERT(LIST_EMPTY(&po->po_pmcs), 875 ("[pmc,%d] PMC list not empty", __LINE__)); 876 877 /* de-configure the log file if present */ 878 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 879 pmclog_deconfigure_log(po); 880 } 881 882 /* 883 * Remove an owner process record if all conditions are met. 884 */ 885 static void 886 pmc_maybe_remove_owner(struct pmc_owner *po) 887 { 888 889 PMCDBG1(OWN,OMR,1, "maybe-remove-owner po=%p", po); 890 891 /* 892 * Remove owner record if 893 * - this process does not own any PMCs 894 * - this process has not allocated a system-wide sampling buffer 895 */ 896 if (LIST_EMPTY(&po->po_pmcs) && 897 ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)) { 898 pmc_remove_owner(po); 899 pmc_destroy_owner_descriptor(po); 900 } 901 } 902 903 /* 904 * Add an association between a target process and a PMC. 905 */ 906 static void 907 pmc_link_target_process(struct pmc *pm, struct pmc_process *pp) 908 { 909 struct pmc_target *pt; 910 struct pmc_thread *pt_td __diagused; 911 int ri; 912 913 sx_assert(&pmc_sx, SX_XLOCKED); 914 KASSERT(pm != NULL && pp != NULL, 915 ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp)); 916 KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)), 917 ("[pmc,%d] Attaching a non-process-virtual pmc=%p to pid=%d", 918 __LINE__, pm, pp->pp_proc->p_pid)); 919 KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= ((int) md->pmd_npmc - 1), 920 ("[pmc,%d] Illegal reference count %d for process record %p", 921 __LINE__, pp->pp_refcnt, (void *) pp)); 922 923 ri = PMC_TO_ROWINDEX(pm); 924 925 PMCDBG3(PRC,TLK,1, "link-target pmc=%p ri=%d pmc-process=%p", 926 pm, ri, pp); 927 928 #ifdef HWPMC_DEBUG 929 LIST_FOREACH(pt, &pm->pm_targets, pt_next) { 930 if (pt->pt_process == pp) 931 KASSERT(0, ("[pmc,%d] pp %p already in pmc %p targets", 932 __LINE__, pp, pm)); 933 } 934 #endif 935 pt = malloc(sizeof(struct pmc_target), M_PMC, M_WAITOK | M_ZERO); 936 pt->pt_process = pp; 937 938 LIST_INSERT_HEAD(&pm->pm_targets, pt, pt_next); 939 940 atomic_store_rel_ptr((uintptr_t *)&pp->pp_pmcs[ri].pp_pmc, 941 (uintptr_t)pm); 942 943 if (pm->pm_owner->po_owner == pp->pp_proc) 944 pm->pm_flags |= PMC_F_ATTACHED_TO_OWNER; 945 946 /* 947 * Initialize the per-process values at this row index. 948 */ 949 pp->pp_pmcs[ri].pp_pmcval = PMC_TO_MODE(pm) == PMC_MODE_TS ? 950 pm->pm_sc.pm_reloadcount : 0; 951 pp->pp_refcnt++; 952 953 #ifdef INVARIANTS 954 /* Confirm that the per-thread values at this row index are cleared. */ 955 if (PMC_TO_MODE(pm) == PMC_MODE_TS) { 956 mtx_lock_spin(pp->pp_tdslock); 957 LIST_FOREACH(pt_td, &pp->pp_tds, pt_next) { 958 KASSERT(pt_td->pt_pmcs[ri].pt_pmcval == (pmc_value_t) 0, 959 ("[pmc,%d] pt_pmcval not cleared for pid=%d at " 960 "ri=%d", __LINE__, pp->pp_proc->p_pid, ri)); 961 } 962 mtx_unlock_spin(pp->pp_tdslock); 963 } 964 #endif 965 } 966 967 /* 968 * Removes the association between a target process and a PMC. 969 */ 970 static void 971 pmc_unlink_target_process(struct pmc *pm, struct pmc_process *pp) 972 { 973 int ri; 974 struct proc *p; 975 struct pmc_target *ptgt; 976 struct pmc_thread *pt; 977 978 sx_assert(&pmc_sx, SX_XLOCKED); 979 980 KASSERT(pm != NULL && pp != NULL, 981 ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp)); 982 983 KASSERT(pp->pp_refcnt >= 1 && pp->pp_refcnt <= (int) md->pmd_npmc, 984 ("[pmc,%d] Illegal ref count %d on process record %p", 985 __LINE__, pp->pp_refcnt, (void *) pp)); 986 987 ri = PMC_TO_ROWINDEX(pm); 988 989 PMCDBG3(PRC,TUL,1, "unlink-target pmc=%p ri=%d pmc-process=%p", 990 pm, ri, pp); 991 992 KASSERT(pp->pp_pmcs[ri].pp_pmc == pm, 993 ("[pmc,%d] PMC ri %d mismatch pmc %p pp->[ri] %p", __LINE__, 994 ri, pm, pp->pp_pmcs[ri].pp_pmc)); 995 996 pp->pp_pmcs[ri].pp_pmc = NULL; 997 pp->pp_pmcs[ri].pp_pmcval = (pmc_value_t)0; 998 999 /* Clear the per-thread values at this row index. */ 1000 if (PMC_TO_MODE(pm) == PMC_MODE_TS) { 1001 mtx_lock_spin(pp->pp_tdslock); 1002 LIST_FOREACH(pt, &pp->pp_tds, pt_next) 1003 pt->pt_pmcs[ri].pt_pmcval = (pmc_value_t)0; 1004 mtx_unlock_spin(pp->pp_tdslock); 1005 } 1006 1007 /* Remove owner-specific flags */ 1008 if (pm->pm_owner->po_owner == pp->pp_proc) { 1009 pp->pp_flags &= ~PMC_PP_ENABLE_MSR_ACCESS; 1010 pm->pm_flags &= ~PMC_F_ATTACHED_TO_OWNER; 1011 } 1012 1013 pp->pp_refcnt--; 1014 1015 /* Remove the target process from the PMC structure */ 1016 LIST_FOREACH(ptgt, &pm->pm_targets, pt_next) 1017 if (ptgt->pt_process == pp) 1018 break; 1019 1020 KASSERT(ptgt != NULL, ("[pmc,%d] process %p (pp: %p) not found " 1021 "in pmc %p", __LINE__, pp->pp_proc, pp, pm)); 1022 1023 LIST_REMOVE(ptgt, pt_next); 1024 free(ptgt, M_PMC); 1025 1026 /* if the PMC now lacks targets, send the owner a SIGIO */ 1027 if (LIST_EMPTY(&pm->pm_targets)) { 1028 p = pm->pm_owner->po_owner; 1029 PROC_LOCK(p); 1030 kern_psignal(p, SIGIO); 1031 PROC_UNLOCK(p); 1032 1033 PMCDBG2(PRC,SIG,2, "signalling proc=%p signal=%d", p, SIGIO); 1034 } 1035 } 1036 1037 /* 1038 * Attach a process to a PMC. 1039 */ 1040 static int 1041 pmc_attach_one_process(struct proc *p, struct pmc *pm) 1042 { 1043 int ri, error; 1044 char *fullpath, *freepath; 1045 struct pmc_process *pp; 1046 1047 sx_assert(&pmc_sx, SX_XLOCKED); 1048 1049 PMCDBG5(PRC,ATT,2, "attach-one pm=%p ri=%d proc=%p (%d, %s)", pm, 1050 PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm); 1051 1052 /* 1053 * Locate the process descriptor corresponding to process 'p', 1054 * allocating space as needed. 1055 * 1056 * Verify that rowindex 'pm_rowindex' is free in the process 1057 * descriptor. 1058 * 1059 * If not, allocate space for a descriptor and link the 1060 * process descriptor and PMC. 1061 */ 1062 ri = PMC_TO_ROWINDEX(pm); 1063 1064 /* mark process as using HWPMCs */ 1065 PROC_LOCK(p); 1066 p->p_flag |= P_HWPMC; 1067 PROC_UNLOCK(p); 1068 1069 if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_ALLOCATE)) == NULL) { 1070 error = ENOMEM; 1071 goto fail; 1072 } 1073 1074 if (pp->pp_pmcs[ri].pp_pmc == pm) {/* already present at slot [ri] */ 1075 error = EEXIST; 1076 goto fail; 1077 } 1078 1079 if (pp->pp_pmcs[ri].pp_pmc != NULL) { 1080 error = EBUSY; 1081 goto fail; 1082 } 1083 1084 pmc_link_target_process(pm, pp); 1085 1086 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) && 1087 (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) == 0) 1088 pm->pm_flags |= PMC_F_NEEDS_LOGFILE; 1089 1090 pm->pm_flags |= PMC_F_ATTACH_DONE; /* mark as attached */ 1091 1092 /* issue an attach event to a configured log file */ 1093 if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE) { 1094 if (p->p_flag & P_KPROC) { 1095 fullpath = kernelname; 1096 freepath = NULL; 1097 } else { 1098 pmc_getfilename(p->p_textvp, &fullpath, &freepath); 1099 pmclog_process_pmcattach(pm, p->p_pid, fullpath); 1100 } 1101 free(freepath, M_TEMP); 1102 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) 1103 pmc_log_process_mappings(pm->pm_owner, p); 1104 } 1105 1106 return (0); 1107 fail: 1108 PROC_LOCK(p); 1109 p->p_flag &= ~P_HWPMC; 1110 PROC_UNLOCK(p); 1111 return (error); 1112 } 1113 1114 /* 1115 * Attach a process and optionally its children 1116 */ 1117 static int 1118 pmc_attach_process(struct proc *p, struct pmc *pm) 1119 { 1120 int error; 1121 struct proc *top; 1122 1123 sx_assert(&pmc_sx, SX_XLOCKED); 1124 1125 PMCDBG5(PRC,ATT,1, "attach pm=%p ri=%d proc=%p (%d, %s)", pm, 1126 PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm); 1127 1128 /* 1129 * If this PMC successfully allowed a GETMSR operation 1130 * in the past, disallow further ATTACHes. 1131 */ 1132 if ((pm->pm_flags & PMC_PP_ENABLE_MSR_ACCESS) != 0) 1133 return (EPERM); 1134 1135 if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0) 1136 return (pmc_attach_one_process(p, pm)); 1137 1138 /* 1139 * Traverse all child processes, attaching them to 1140 * this PMC. 1141 */ 1142 sx_slock(&proctree_lock); 1143 1144 top = p; 1145 for (;;) { 1146 if ((error = pmc_attach_one_process(p, pm)) != 0) 1147 break; 1148 if (!LIST_EMPTY(&p->p_children)) 1149 p = LIST_FIRST(&p->p_children); 1150 else for (;;) { 1151 if (p == top) 1152 goto done; 1153 if (LIST_NEXT(p, p_sibling)) { 1154 p = LIST_NEXT(p, p_sibling); 1155 break; 1156 } 1157 p = p->p_pptr; 1158 } 1159 } 1160 1161 if (error != 0) 1162 (void)pmc_detach_process(top, pm); 1163 1164 done: 1165 sx_sunlock(&proctree_lock); 1166 return (error); 1167 } 1168 1169 /* 1170 * Detach a process from a PMC. If there are no other PMCs tracking 1171 * this process, remove the process structure from its hash table. If 1172 * 'flags' contains PMC_FLAG_REMOVE, then free the process structure. 1173 */ 1174 static int 1175 pmc_detach_one_process(struct proc *p, struct pmc *pm, int flags) 1176 { 1177 int ri; 1178 struct pmc_process *pp; 1179 1180 sx_assert(&pmc_sx, SX_XLOCKED); 1181 1182 KASSERT(pm != NULL, 1183 ("[pmc,%d] null pm pointer", __LINE__)); 1184 1185 ri = PMC_TO_ROWINDEX(pm); 1186 1187 PMCDBG6(PRC,ATT,2, "detach-one pm=%p ri=%d proc=%p (%d, %s) flags=0x%x", 1188 pm, ri, p, p->p_pid, p->p_comm, flags); 1189 1190 if ((pp = pmc_find_process_descriptor(p, 0)) == NULL) 1191 return (ESRCH); 1192 1193 if (pp->pp_pmcs[ri].pp_pmc != pm) 1194 return (EINVAL); 1195 1196 /* 1197 * If this is a process-virtual PMC that is still loaded on the 1198 * hardware of the CPU we are running on (the common case when a 1199 * process detaches a PMC from itself), take it off and drop its 1200 * runcount reference now. The reference is otherwise only 1201 * dropped by the switch-out reclaim, which the scheduler stops 1202 * calling once P_HWPMC is cleared below - leaking it and later 1203 * wedging pmc_wait_for_pmc_idle() at release time. 1204 */ 1205 if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) { 1206 critical_enter(); 1207 pmc_reclaim_pmc_from_cpu(pm, pp, curthread->td_oncpu); 1208 critical_exit(); 1209 } 1210 1211 pmc_unlink_target_process(pm, pp); 1212 1213 /* Issue a detach entry if a log file is configured */ 1214 if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE) 1215 pmclog_process_pmcdetach(pm, p->p_pid); 1216 1217 /* 1218 * If there are no PMCs targeting this process, we remove its 1219 * descriptor from the target hash table and unset the P_HWPMC 1220 * flag in the struct proc. 1221 */ 1222 KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= (int) md->pmd_npmc, 1223 ("[pmc,%d] Illegal refcnt %d for process struct %p", 1224 __LINE__, pp->pp_refcnt, pp)); 1225 1226 if (pp->pp_refcnt != 0) /* still a target of some PMC */ 1227 return (0); 1228 1229 /* 1230 * This detach removed the process' last PMC and we are about to 1231 * clear P_HWPMC. If the detached PMC was its last target and is 1232 * still loaded on other CPUs (e.g. sibling threads of a 1233 * multi-threaded target, or a target running on another CPU), 1234 * drain those references first: the target is already unlinked so 1235 * it cannot reload the PMC, and P_HWPMC is still set so those 1236 * CPUs' switch-out reclaim still runs. Bounded by the target 1237 * threads being scheduled out. 1238 */ 1239 if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)) && 1240 LIST_EMPTY(&pm->pm_targets)) { 1241 while (counter_u64_fetch(pm->pm_runcount) > 0) 1242 pmc_force_context_switch(); 1243 } 1244 1245 pmc_remove_process_descriptor(pp); 1246 1247 if (flags & PMC_FLAG_REMOVE) 1248 pmc_destroy_process_descriptor(pp); 1249 1250 PROC_LOCK(p); 1251 p->p_flag &= ~P_HWPMC; 1252 PROC_UNLOCK(p); 1253 1254 return (0); 1255 } 1256 1257 /* 1258 * Detach a process and optionally its descendants from a PMC. 1259 */ 1260 static int 1261 pmc_detach_process(struct proc *p, struct pmc *pm) 1262 { 1263 struct proc *top; 1264 1265 sx_assert(&pmc_sx, SX_XLOCKED); 1266 1267 PMCDBG5(PRC,ATT,1, "detach pm=%p ri=%d proc=%p (%d, %s)", pm, 1268 PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm); 1269 1270 if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0) 1271 return (pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE)); 1272 1273 /* 1274 * Traverse all children, detaching them from this PMC. We 1275 * ignore errors since we could be detaching a PMC from a 1276 * partially attached proc tree. 1277 */ 1278 sx_slock(&proctree_lock); 1279 1280 top = p; 1281 for (;;) { 1282 (void)pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE); 1283 1284 if (!LIST_EMPTY(&p->p_children)) { 1285 p = LIST_FIRST(&p->p_children); 1286 } else { 1287 for (;;) { 1288 if (p == top) 1289 goto done; 1290 if (LIST_NEXT(p, p_sibling)) { 1291 p = LIST_NEXT(p, p_sibling); 1292 break; 1293 } 1294 p = p->p_pptr; 1295 } 1296 } 1297 } 1298 done: 1299 sx_sunlock(&proctree_lock); 1300 if (LIST_EMPTY(&pm->pm_targets)) 1301 pm->pm_flags &= ~PMC_F_ATTACH_DONE; 1302 1303 return (0); 1304 } 1305 1306 /* 1307 * Handle events after an exec() for a process: 1308 * - Inform log owners of the new exec() event 1309 * - Release any PMCs owned by the process before the exec() 1310 * - Detach PMCs from the target if required 1311 */ 1312 static void 1313 pmc_process_exec(struct thread *td, struct pmckern_procexec *pk) 1314 { 1315 struct pmc *pm; 1316 struct pmc_owner *po; 1317 struct pmc_process *pp; 1318 struct proc *p; 1319 char *fullpath, *freepath; 1320 u_int ri; 1321 bool is_using_hwpmcs; 1322 1323 sx_assert(&pmc_sx, SX_XLOCKED); 1324 1325 p = td->td_proc; 1326 pmc_getfilename(p->p_textvp, &fullpath, &freepath); 1327 1328 PMC_EPOCH_ENTER(); 1329 /* Inform owners of SS mode PMCs of the exec event. */ 1330 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) { 1331 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) != 0) { 1332 pmclog_process_procexec(po, PMC_ID_INVALID, p->p_pid, 1333 pk->pm_baseaddr, pk->pm_dynaddr, fullpath); 1334 } 1335 } 1336 PMC_EPOCH_EXIT(); 1337 1338 PROC_LOCK(p); 1339 is_using_hwpmcs = (p->p_flag & P_HWPMC) != 0; 1340 PROC_UNLOCK(p); 1341 1342 if (!is_using_hwpmcs) { 1343 if (freepath != NULL) 1344 free(freepath, M_TEMP); 1345 return; 1346 } 1347 1348 /* 1349 * PMCs are not inherited across an exec(): remove any PMCs that this 1350 * process is the owner of. 1351 */ 1352 if ((po = pmc_find_owner_descriptor(p)) != NULL) { 1353 pmc_remove_owner(po); 1354 pmc_destroy_owner_descriptor(po); 1355 } 1356 1357 /* 1358 * If the process being exec'ed is not the target of any PMC, we are 1359 * done. 1360 */ 1361 if ((pp = pmc_find_process_descriptor(p, 0)) == NULL) { 1362 if (freepath != NULL) 1363 free(freepath, M_TEMP); 1364 return; 1365 } 1366 1367 /* 1368 * Log the exec event to all monitoring owners. Skip owners who have 1369 * already received the event because they had system sampling PMCs 1370 * active. 1371 */ 1372 for (ri = 0; ri < md->pmd_npmc; ri++) { 1373 if ((pm = pp->pp_pmcs[ri].pp_pmc) == NULL) 1374 continue; 1375 1376 po = pm->pm_owner; 1377 if (po->po_sscount == 0 && 1378 (po->po_flags & PMC_PO_OWNS_LOGFILE) != 0) { 1379 pmclog_process_procexec(po, pm->pm_id, p->p_pid, 1380 pk->pm_baseaddr, pk->pm_dynaddr, fullpath); 1381 } 1382 } 1383 1384 if (freepath != NULL) 1385 free(freepath, M_TEMP); 1386 1387 PMCDBG4(PRC,EXC,1, "exec proc=%p (%d, %s) cred-changed=%d", 1388 p, p->p_pid, p->p_comm, pk->pm_credentialschanged); 1389 1390 if (pk->pm_credentialschanged == 0) /* no change */ 1391 return; 1392 1393 /* 1394 * If the newly exec()'ed process has a different credential 1395 * than before, allow it to be the target of a PMC only if 1396 * the PMC's owner has sufficient privilege. 1397 */ 1398 for (ri = 0; ri < md->pmd_npmc; ri++) { 1399 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) { 1400 struct proc *owner; 1401 struct ucred *cred; 1402 1403 owner = pm->pm_owner->po_owner; 1404 PROC_LOCK(owner); 1405 cred = crhold(owner->p_ucred); 1406 PROC_UNLOCK(owner); 1407 1408 if (priv_check_cred(cred, PRIV_DEBUG_DIFFCRED) != 0) 1409 pmc_detach_one_process(td->td_proc, pm, 1410 PMC_FLAG_NONE); 1411 1412 crfree(cred); 1413 } 1414 } 1415 1416 KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= md->pmd_npmc, 1417 ("[pmc,%d] Illegal ref count %u on pp %p", __LINE__, 1418 pp->pp_refcnt, pp)); 1419 1420 /* 1421 * If this process is no longer the target of any 1422 * PMCs, we can remove the process entry and free 1423 * up space. 1424 */ 1425 if (pp->pp_refcnt == 0) 1426 pmc_destroy_process_descriptor(pp); 1427 } 1428 1429 /* 1430 * Execute optional context-switch batch operations for all classes. 1431 */ 1432 static void 1433 pmc_process_csw_start_all(int cpu) 1434 { 1435 struct pmc_classdep *pcd; 1436 u_int class; 1437 1438 for (class = 0; class < md->pmd_nclass; class++) { 1439 pcd = &md->pmd_classdep[class]; 1440 if (pcd->pcd_start_all != NULL) 1441 (void)pcd->pcd_start_all(cpu); 1442 } 1443 } 1444 1445 static void 1446 pmc_process_csw_stop_all(int cpu) 1447 { 1448 struct pmc_classdep *pcd; 1449 u_int class; 1450 1451 for (class = 0; class < md->pmd_nclass; class++) { 1452 pcd = &md->pmd_classdep[class]; 1453 if (pcd->pcd_stop_all != NULL) 1454 (void)pcd->pcd_stop_all(cpu); 1455 } 1456 } 1457 1458 /* 1459 * Thread context switch IN. 1460 */ 1461 static void 1462 pmc_process_csw_in(struct thread *td) 1463 { 1464 struct pmc *pm; 1465 struct pmc_classdep *pcd; 1466 struct pmc_cpu *pc; 1467 struct pmc_hw *phw __diagused; 1468 struct pmc_process *pp; 1469 struct pmc_thread *pt; 1470 struct proc *p; 1471 pmc_value_t newvalue; 1472 int cpu; 1473 u_int adjri, ri; 1474 1475 p = td->td_proc; 1476 pt = NULL; 1477 if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE)) == NULL) 1478 return; 1479 1480 KASSERT(pp->pp_proc == td->td_proc, 1481 ("[pmc,%d] not my thread state", __LINE__)); 1482 1483 critical_enter(); /* no preemption from this point */ 1484 1485 cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */ 1486 1487 PMCDBG5(CSW,SWI,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p, 1488 p->p_pid, p->p_comm, pp); 1489 1490 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), 1491 ("[pmc,%d] weird CPU id %d", __LINE__, cpu)); 1492 1493 pc = pmc_pcpu[cpu]; 1494 for (ri = 0; ri < md->pmd_npmc; ri++) { 1495 if ((pm = pp->pp_pmcs[ri].pp_pmc) == NULL) 1496 continue; 1497 1498 KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)), 1499 ("[pmc,%d] Target PMC in non-virtual mode (%d)", 1500 __LINE__, PMC_TO_MODE(pm))); 1501 KASSERT(PMC_TO_ROWINDEX(pm) == ri, 1502 ("[pmc,%d] Row index mismatch pmc %d != ri %d", 1503 __LINE__, PMC_TO_ROWINDEX(pm), ri)); 1504 1505 /* 1506 * Only PMCs that are marked as 'RUNNING' need 1507 * be placed on hardware. 1508 */ 1509 if (pm->pm_state != PMC_STATE_RUNNING) 1510 continue; 1511 1512 KASSERT(counter_u64_fetch(pm->pm_runcount) >= 0, 1513 ("[pmc,%d] pm=%p runcount %ju", __LINE__, pm, 1514 (uintmax_t)counter_u64_fetch(pm->pm_runcount))); 1515 1516 /* increment PMC runcount */ 1517 counter_u64_add(pm->pm_runcount, 1); 1518 1519 /* configure the HWPMC we are going to use. */ 1520 pcd = pmc_ri_to_classdep(md, ri, &adjri); 1521 (void)pcd->pcd_config_pmc(cpu, adjri, pm); 1522 1523 phw = pc->pc_hwpmcs[ri]; 1524 1525 KASSERT(phw != NULL, 1526 ("[pmc,%d] null hw pointer", __LINE__)); 1527 1528 KASSERT(phw->phw_pmc == pm, 1529 ("[pmc,%d] hw->pmc %p != pmc %p", __LINE__, 1530 phw->phw_pmc, pm)); 1531 1532 /* 1533 * Write out saved value and start the PMC. 1534 * 1535 * Sampling PMCs use a per-thread value, while 1536 * counting mode PMCs use a per-pmc value that is 1537 * inherited across descendants. 1538 */ 1539 if (PMC_TO_MODE(pm) == PMC_MODE_TS) { 1540 if (pt == NULL) 1541 pt = pmc_find_thread_descriptor(pp, td, 1542 PMC_FLAG_NONE); 1543 1544 KASSERT(pt != NULL, 1545 ("[pmc,%d] No thread found for td=%p", __LINE__, 1546 td)); 1547 1548 mtx_pool_lock_spin(pmc_mtxpool, pm); 1549 1550 /* 1551 * If we have a thread descriptor, use the per-thread 1552 * counter in the descriptor. If not, we will use 1553 * a per-process counter. 1554 * 1555 * TODO: Remove the per-process "safety net" once 1556 * we have thoroughly tested that we don't hit the 1557 * above assert. 1558 */ 1559 if (pt != NULL) { 1560 if (pt->pt_pmcs[ri].pt_pmcval > 0) 1561 newvalue = pt->pt_pmcs[ri].pt_pmcval; 1562 else 1563 newvalue = pm->pm_sc.pm_reloadcount; 1564 } else { 1565 /* 1566 * Use the saved value calculated after the most 1567 * recent time a thread using the shared counter 1568 * switched out. Reset the saved count in case 1569 * another thread from this process switches in 1570 * before any threads switch out. 1571 */ 1572 newvalue = pp->pp_pmcs[ri].pp_pmcval; 1573 pp->pp_pmcs[ri].pp_pmcval = 1574 pm->pm_sc.pm_reloadcount; 1575 } 1576 mtx_pool_unlock_spin(pmc_mtxpool, pm); 1577 KASSERT(newvalue > 0 && newvalue <= 1578 pm->pm_sc.pm_reloadcount, 1579 ("[pmc,%d] pmcval outside of expected range cpu=%d " 1580 "ri=%d pmcval=%jx pm_reloadcount=%jx", __LINE__, 1581 cpu, ri, newvalue, pm->pm_sc.pm_reloadcount)); 1582 } else { 1583 KASSERT(PMC_TO_MODE(pm) == PMC_MODE_TC, 1584 ("[pmc,%d] illegal mode=%d", __LINE__, 1585 PMC_TO_MODE(pm))); 1586 mtx_pool_lock_spin(pmc_mtxpool, pm); 1587 newvalue = PMC_PCPU_SAVED(cpu, ri) = 1588 pm->pm_gv.pm_savedvalue; 1589 mtx_pool_unlock_spin(pmc_mtxpool, pm); 1590 } 1591 1592 PMCDBG3(CSW,SWI,1,"cpu=%d ri=%d new=%jd", cpu, ri, newvalue); 1593 1594 (void)pcd->pcd_write_pmc(cpu, adjri, pm, newvalue); 1595 1596 /* If a sampling mode PMC, reset stalled state. */ 1597 if (PMC_TO_MODE(pm) == PMC_MODE_TS) 1598 pm->pm_pcpu_state[cpu].pps_stalled = 0; 1599 1600 /* Indicate that we desire this to run. */ 1601 pm->pm_pcpu_state[cpu].pps_cpustate = 1; 1602 1603 /* Start the PMC. */ 1604 (void)pcd->pcd_start_pmc(cpu, adjri, pm); 1605 } 1606 1607 /* 1608 * Perform any other architecture/cpu dependent thread 1609 * switch-in actions. 1610 */ 1611 (void)(*md->pmd_switch_in)(pc, pp); 1612 1613 /* Commit all class PMC start updates at one boundary. */ 1614 pmc_process_csw_start_all(cpu); 1615 1616 critical_exit(); 1617 } 1618 1619 /* 1620 * Compute the change in a counter's value since it was last written. 1621 * The hardware counter is only pcd_width bits wide and wraps around, 1622 * while the value seeded into it may occupy the full 64-bit range, so 1623 * take the difference modulo the counter width. 1624 */ 1625 static pmc_value_t 1626 pmc_delta(const struct pmc_classdep *pcd, pmc_value_t newvalue, 1627 pmc_value_t oldvalue) 1628 { 1629 pmc_value_t delta; 1630 1631 delta = newvalue - oldvalue; 1632 if (pcd->pcd_width < 64) 1633 delta &= ((pmc_value_t)1 << pcd->pcd_width) - 1; 1634 return (delta); 1635 } 1636 1637 /* 1638 * Take a process-virtual PMC off the hardware of 'cpu' if it is 1639 * currently loaded there for process 'pp', accumulating its final 1640 * count and dropping its runcount reference. This is the same reclaim 1641 * that context switch out and process exit perform, factored out so it 1642 * can also run when a target is detached while the PMC may still be 1643 * live: the runcount reference is decremented by the switch-out reclaim 1644 * only, which the scheduler gates on P_HWPMC, so a detach that clears 1645 * P_HWPMC without draining would leak the reference and later wedge 1646 * pmc_wait_for_pmc_idle(). Must be called in a critical section. 1647 */ 1648 static void 1649 pmc_reclaim_pmc_from_cpu(struct pmc *pm, struct pmc_process *pp, int cpu) 1650 { 1651 struct pmc_classdep *pcd; 1652 struct pmc *phw_pm; 1653 pmc_value_t newvalue, tmp; 1654 u_int adjri, ri; 1655 1656 ri = PMC_TO_ROWINDEX(pm); 1657 pcd = pmc_ri_to_classdep(md, ri, &adjri); 1658 1659 /* Only reclaim if this PMC is actually loaded on this CPU. */ 1660 phw_pm = NULL; 1661 (void)(*pcd->pcd_get_config)(cpu, adjri, &phw_pm); 1662 if (phw_pm != pm) 1663 return; 1664 1665 KASSERT(counter_u64_fetch(pm->pm_runcount) > 0, 1666 ("[pmc,%d] pm=%p runcount %ju", __LINE__, pm, 1667 (uintmax_t)counter_u64_fetch(pm->pm_runcount))); 1668 1669 if (pm->pm_pcpu_state[cpu].pps_cpustate) { 1670 pm->pm_pcpu_state[cpu].pps_cpustate = 0; 1671 if (pm->pm_pcpu_state[cpu].pps_stalled == 0) { 1672 (void)pcd->pcd_stop_pmc(cpu, adjri, pm); 1673 1674 if (PMC_TO_MODE(pm) == PMC_MODE_TC) { 1675 (void)pcd->pcd_read_pmc(cpu, adjri, pm, 1676 &newvalue); 1677 tmp = pmc_delta(pcd, newvalue, 1678 PMC_PCPU_SAVED(cpu, ri)); 1679 1680 mtx_pool_lock_spin(pmc_mtxpool, pm); 1681 pm->pm_gv.pm_savedvalue += tmp; 1682 pp->pp_pmcs[ri].pp_pmcval += tmp; 1683 mtx_pool_unlock_spin(pmc_mtxpool, pm); 1684 } 1685 } 1686 } 1687 1688 counter_u64_add(pm->pm_runcount, -1); 1689 (void)pcd->pcd_config_pmc(cpu, adjri, NULL); 1690 } 1691 1692 /* 1693 * Thread context switch OUT. 1694 */ 1695 static void 1696 pmc_process_csw_out(struct thread *td) 1697 { 1698 struct pmc *pm; 1699 struct pmc_classdep *pcd; 1700 struct pmc_cpu *pc; 1701 struct pmc_process *pp; 1702 struct pmc_thread *pt = NULL; 1703 struct proc *p; 1704 pmc_value_t newvalue, tmp; 1705 enum pmc_mode mode; 1706 int cpu; 1707 u_int adjri, ri; 1708 1709 /* 1710 * Locate our process descriptor; this may be NULL if 1711 * this process is exiting and we have already removed 1712 * the process from the target process table. 1713 * 1714 * Note that due to kernel preemption, multiple 1715 * context switches may happen while the process is 1716 * exiting. 1717 * 1718 * Note also that if the target process cannot be 1719 * found we still need to deconfigure any PMCs that 1720 * are currently running on hardware. 1721 */ 1722 p = td->td_proc; 1723 pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE); 1724 1725 critical_enter(); 1726 1727 cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */ 1728 1729 PMCDBG5(CSW,SWO,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p, 1730 p->p_pid, p->p_comm, pp); 1731 1732 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), 1733 ("[pmc,%d weird CPU id %d", __LINE__, cpu)); 1734 1735 pc = pmc_pcpu[cpu]; 1736 1737 /* Close shared class gates before any PMC stop or read. */ 1738 pmc_process_csw_stop_all(cpu); 1739 1740 /* 1741 * When a PMC gets unlinked from a target PMC, it will 1742 * be removed from the target's pp_pmc[] array. 1743 * 1744 * However, on a MP system, the target could have been 1745 * executing on another CPU at the time of the unlink. 1746 * So, at context switch OUT time, we need to look at 1747 * the hardware to determine if a PMC is scheduled on 1748 * it. 1749 */ 1750 for (ri = 0; ri < md->pmd_npmc; ri++) { 1751 pcd = pmc_ri_to_classdep(md, ri, &adjri); 1752 pm = NULL; 1753 (void)(*pcd->pcd_get_config)(cpu, adjri, &pm); 1754 1755 if (pm == NULL) /* nothing at this row index */ 1756 continue; 1757 1758 mode = PMC_TO_MODE(pm); 1759 if (!PMC_IS_VIRTUAL_MODE(mode)) 1760 continue; /* not a process virtual PMC */ 1761 1762 KASSERT(PMC_TO_ROWINDEX(pm) == ri, 1763 ("[pmc,%d] ri mismatch pmc(%d) ri(%d)", 1764 __LINE__, PMC_TO_ROWINDEX(pm), ri)); 1765 1766 /* 1767 * Change desired state, and then stop if not stalled. 1768 * This two-step dance should avoid race conditions where 1769 * an interrupt re-enables the PMC after this code has 1770 * already checked the pm_stalled flag. 1771 */ 1772 pm->pm_pcpu_state[cpu].pps_cpustate = 0; 1773 if (pm->pm_pcpu_state[cpu].pps_stalled == 0) 1774 (void)pcd->pcd_stop_pmc(cpu, adjri, pm); 1775 1776 KASSERT(counter_u64_fetch(pm->pm_runcount) > 0, 1777 ("[pmc,%d] pm=%p runcount %ju", __LINE__, pm, 1778 (uintmax_t)counter_u64_fetch(pm->pm_runcount))); 1779 1780 /* reduce this PMC's runcount */ 1781 counter_u64_add(pm->pm_runcount, -1); 1782 1783 /* 1784 * If this PMC is associated with this process, 1785 * save the reading. 1786 */ 1787 if (pm->pm_state != PMC_STATE_DELETED && pp != NULL && 1788 pp->pp_pmcs[ri].pp_pmc != NULL) { 1789 KASSERT(pm == pp->pp_pmcs[ri].pp_pmc, 1790 ("[pmc,%d] pm %p != pp_pmcs[%d] %p", __LINE__, 1791 pm, ri, pp->pp_pmcs[ri].pp_pmc)); 1792 KASSERT(pp->pp_refcnt > 0, 1793 ("[pmc,%d] pp refcnt = %d", __LINE__, 1794 pp->pp_refcnt)); 1795 1796 (void)pcd->pcd_read_pmc(cpu, adjri, pm, &newvalue); 1797 1798 if (mode == PMC_MODE_TS) { 1799 PMCDBG3(CSW,SWO,1,"cpu=%d ri=%d val=%jd (samp)", 1800 cpu, ri, newvalue); 1801 1802 if (pt == NULL) 1803 pt = pmc_find_thread_descriptor(pp, td, 1804 PMC_FLAG_NONE); 1805 1806 KASSERT(pt != NULL, 1807 ("[pmc,%d] No thread found for td=%p", 1808 __LINE__, td)); 1809 1810 mtx_pool_lock_spin(pmc_mtxpool, pm); 1811 1812 /* 1813 * If we have a thread descriptor, save the 1814 * per-thread counter in the descriptor. If not, 1815 * we will update the per-process counter. 1816 * 1817 * TODO: Remove the per-process "safety net" 1818 * once we have thoroughly tested that we 1819 * don't hit the above assert. 1820 */ 1821 if (pt != NULL) { 1822 pt->pt_pmcs[ri].pt_pmcval = newvalue; 1823 } else { 1824 /* 1825 * For sampling process-virtual PMCs, 1826 * newvalue is the number of events to 1827 * be seen until the next sampling 1828 * interrupt. We can just add the events 1829 * left from this invocation to the 1830 * counter, then adjust in case we 1831 * overflow our range. 1832 * 1833 * (Recall that we reload the counter 1834 * every time we use it.) 1835 */ 1836 pp->pp_pmcs[ri].pp_pmcval += newvalue; 1837 if (pp->pp_pmcs[ri].pp_pmcval > 1838 pm->pm_sc.pm_reloadcount) { 1839 pp->pp_pmcs[ri].pp_pmcval -= 1840 pm->pm_sc.pm_reloadcount; 1841 } 1842 } 1843 mtx_pool_unlock_spin(pmc_mtxpool, pm); 1844 } else { 1845 /* 1846 * For counting process-virtual PMCs, the 1847 * hardware counter's value increases 1848 * monotonically modulo the counter width; 1849 * pmc_delta() recovers the increment even 1850 * when the counter wrapped during the run. 1851 */ 1852 tmp = pmc_delta(pcd, newvalue, 1853 PMC_PCPU_SAVED(cpu, ri)); 1854 1855 PMCDBG3(CSW,SWO,1,"cpu=%d ri=%d tmp=%jd (count)", 1856 cpu, ri, tmp); 1857 1858 mtx_pool_lock_spin(pmc_mtxpool, pm); 1859 pm->pm_gv.pm_savedvalue += tmp; 1860 pp->pp_pmcs[ri].pp_pmcval += tmp; 1861 mtx_pool_unlock_spin(pmc_mtxpool, pm); 1862 1863 if (pm->pm_flags & PMC_F_LOG_PROCCSW) 1864 pmclog_process_proccsw(pm, pp, tmp, td); 1865 } 1866 } 1867 1868 /* Mark hardware as free. */ 1869 (void)pcd->pcd_config_pmc(cpu, adjri, NULL); 1870 } 1871 1872 /* 1873 * Perform any other architecture/cpu dependent thread 1874 * switch out functions. 1875 */ 1876 (void)(*md->pmd_switch_out)(pc, pp); 1877 1878 critical_exit(); 1879 } 1880 1881 /* 1882 * A new thread for a process. 1883 */ 1884 static void 1885 pmc_process_thread_add(struct thread *td) 1886 { 1887 struct pmc_process *pmc; 1888 1889 pmc = pmc_find_process_descriptor(td->td_proc, PMC_FLAG_NONE); 1890 if (pmc != NULL) 1891 pmc_find_thread_descriptor(pmc, td, PMC_FLAG_ALLOCATE); 1892 } 1893 1894 /* 1895 * A thread delete for a process. 1896 */ 1897 static void 1898 pmc_process_thread_delete(struct thread *td) 1899 { 1900 struct pmc_process *pmc; 1901 1902 pmc = pmc_find_process_descriptor(td->td_proc, PMC_FLAG_NONE); 1903 if (pmc != NULL) 1904 pmc_thread_descriptor_pool_free(pmc_find_thread_descriptor(pmc, 1905 td, PMC_FLAG_REMOVE)); 1906 } 1907 1908 /* 1909 * A userret() call for a thread. 1910 */ 1911 static void 1912 pmc_process_thread_userret(struct thread *td) 1913 { 1914 sched_pin(); 1915 pmc_capture_user_callchain(curcpu, PMC_UR, td->td_frame); 1916 sched_unpin(); 1917 } 1918 1919 /* 1920 * A mapping change for a process. 1921 */ 1922 static void 1923 pmc_process_mmap(struct thread *td, struct pmckern_map_in *pkm) 1924 { 1925 const struct pmc *pm; 1926 const struct pmc_process *pp; 1927 struct pmc_owner *po; 1928 char *fullpath, *freepath; 1929 pid_t pid; 1930 int ri; 1931 1932 MPASS(!in_epoch(global_epoch_preempt)); 1933 1934 freepath = fullpath = NULL; 1935 pmc_getfilename((struct vnode *)pkm->pm_file, &fullpath, &freepath); 1936 1937 pid = td->td_proc->p_pid; 1938 1939 PMC_EPOCH_ENTER(); 1940 /* Inform owners of all system-wide sampling PMCs. */ 1941 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) { 1942 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 1943 pmclog_process_map_in(po, pid, pkm->pm_address, 1944 fullpath); 1945 } 1946 1947 if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL) 1948 goto done; 1949 1950 /* 1951 * Inform sampling PMC owners tracking this process. 1952 */ 1953 for (ri = 0; ri < md->pmd_npmc; ri++) { 1954 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL && 1955 PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) { 1956 pmclog_process_map_in(pm->pm_owner, 1957 pid, pkm->pm_address, fullpath); 1958 } 1959 } 1960 1961 done: 1962 if (freepath != NULL) 1963 free(freepath, M_TEMP); 1964 PMC_EPOCH_EXIT(); 1965 } 1966 1967 /* 1968 * Log an munmap request. 1969 */ 1970 static void 1971 pmc_process_munmap(struct thread *td, struct pmckern_map_out *pkm) 1972 { 1973 const struct pmc *pm; 1974 const struct pmc_process *pp; 1975 struct pmc_owner *po; 1976 pid_t pid; 1977 int ri; 1978 1979 pid = td->td_proc->p_pid; 1980 1981 PMC_EPOCH_ENTER(); 1982 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) { 1983 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 1984 pmclog_process_map_out(po, pid, pkm->pm_address, 1985 pkm->pm_address + pkm->pm_size); 1986 } 1987 PMC_EPOCH_EXIT(); 1988 1989 if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL) 1990 return; 1991 1992 for (ri = 0; ri < md->pmd_npmc; ri++) { 1993 pm = pp->pp_pmcs[ri].pp_pmc; 1994 if (pm != NULL && PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) { 1995 pmclog_process_map_out(pm->pm_owner, pid, 1996 pkm->pm_address, pkm->pm_address + pkm->pm_size); 1997 } 1998 } 1999 } 2000 2001 /* 2002 * Log mapping information about the kernel. 2003 */ 2004 static void 2005 pmc_log_kernel_mappings(struct pmc *pm) 2006 { 2007 struct pmc_owner *po; 2008 struct pmckern_map_in *km, *kmbase; 2009 2010 MPASS(in_epoch(global_epoch_preempt) || sx_xlocked(&pmc_sx)); 2011 KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)), 2012 ("[pmc,%d] non-sampling PMC (%p) desires mapping information", 2013 __LINE__, (void *) pm)); 2014 2015 po = pm->pm_owner; 2016 if ((po->po_flags & PMC_PO_INITIAL_MAPPINGS_DONE) != 0) 2017 return; 2018 2019 if (PMC_TO_MODE(pm) == PMC_MODE_SS) 2020 pmc_process_allproc(pm); 2021 2022 /* 2023 * Log the current set of kernel modules. 2024 */ 2025 kmbase = linker_hwpmc_list_objects(); 2026 for (km = kmbase; km->pm_file != NULL; km++) { 2027 PMCDBG2(LOG,REG,1,"%s %p", (char *)km->pm_file, 2028 (void *)km->pm_address); 2029 pmclog_process_map_in(po, (pid_t)-1, km->pm_address, 2030 km->pm_file); 2031 } 2032 free(kmbase, M_LINKER); 2033 2034 po->po_flags |= PMC_PO_INITIAL_MAPPINGS_DONE; 2035 } 2036 2037 /* 2038 * Log the mappings for a single process. 2039 */ 2040 static void 2041 pmc_log_process_mappings(struct pmc_owner *po, struct proc *p) 2042 { 2043 vm_map_t map; 2044 vm_map_entry_t entry; 2045 vm_object_t obj, lobj, tobj; 2046 vm_offset_t last_end; 2047 vm_offset_t start_addr; 2048 struct vnode *vp, *last_vp; 2049 struct vmspace *vm; 2050 char *fullpath, *freepath; 2051 u_int last_timestamp; 2052 2053 last_vp = NULL; 2054 last_end = (vm_offset_t)0; 2055 fullpath = freepath = NULL; 2056 2057 if ((vm = vmspace_acquire_ref(p)) == NULL) 2058 return; 2059 2060 map = &vm->vm_map; 2061 vm_map_lock_read(map); 2062 VM_MAP_ENTRY_FOREACH(entry, map) { 2063 if (entry == NULL) { 2064 PMCDBG2(LOG,OPS,2, "hwpmc: vm_map entry unexpectedly " 2065 "NULL! pid=%d vm_map=%p\n", p->p_pid, map); 2066 break; 2067 } 2068 2069 /* 2070 * We only care about executable map entries. 2071 */ 2072 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0 || 2073 (entry->protection & VM_PROT_EXECUTE) == 0 || 2074 entry->object.vm_object == NULL) { 2075 continue; 2076 } 2077 2078 obj = entry->object.vm_object; 2079 VM_OBJECT_RLOCK(obj); 2080 2081 /* 2082 * Walk the backing_object list to find the base (non-shadowed) 2083 * vm_object. 2084 */ 2085 for (lobj = tobj = obj; tobj != NULL; 2086 tobj = tobj->backing_object) { 2087 if (tobj != obj) 2088 VM_OBJECT_RLOCK(tobj); 2089 if (lobj != obj) 2090 VM_OBJECT_RUNLOCK(lobj); 2091 lobj = tobj; 2092 } 2093 2094 /* 2095 * At this point lobj is the base vm_object and it is locked. 2096 */ 2097 if (lobj == NULL) { 2098 PMCDBG3(LOG,OPS,2, 2099 "hwpmc: lobj unexpectedly NULL! pid=%d " 2100 "vm_map=%p vm_obj=%p\n", p->p_pid, map, obj); 2101 VM_OBJECT_RUNLOCK(obj); 2102 continue; 2103 } 2104 2105 vp = vm_object_vnode(lobj); 2106 if (vp == NULL) { 2107 if (lobj != obj) 2108 VM_OBJECT_RUNLOCK(lobj); 2109 VM_OBJECT_RUNLOCK(obj); 2110 continue; 2111 } 2112 2113 /* 2114 * Skip contiguous regions that point to the same vnode, so we 2115 * don't emit redundant MAP-IN directives. 2116 */ 2117 if (entry->start == last_end && vp == last_vp) { 2118 last_end = entry->end; 2119 if (lobj != obj) 2120 VM_OBJECT_RUNLOCK(lobj); 2121 VM_OBJECT_RUNLOCK(obj); 2122 continue; 2123 } 2124 2125 /* 2126 * We don't want to keep the proc's vm_map or this vm_object 2127 * locked while we walk the pathname, since vn_fullpath() can 2128 * sleep. However, if we drop the lock, it's possible for 2129 * concurrent activity to modify the vm_map list. To protect 2130 * against this, we save the vm_map timestamp before we release 2131 * the lock, and check it after we reacquire the lock below. 2132 */ 2133 start_addr = entry->start; 2134 last_end = entry->end; 2135 last_timestamp = map->timestamp; 2136 vm_map_unlock_read(map); 2137 2138 vref(vp); 2139 if (lobj != obj) 2140 VM_OBJECT_RUNLOCK(lobj); 2141 VM_OBJECT_RUNLOCK(obj); 2142 2143 freepath = NULL; 2144 pmc_getfilename(vp, &fullpath, &freepath); 2145 last_vp = vp; 2146 2147 vrele(vp); 2148 2149 vp = NULL; 2150 pmclog_process_map_in(po, p->p_pid, start_addr, fullpath); 2151 if (freepath != NULL) 2152 free(freepath, M_TEMP); 2153 2154 vm_map_lock_read(map); 2155 2156 /* 2157 * If our saved timestamp doesn't match, this means 2158 * that the vm_map was modified out from under us and 2159 * we can't trust our current "entry" pointer. Do a 2160 * new lookup for this entry. If there is no entry 2161 * for this address range, vm_map_lookup_entry() will 2162 * return the previous one, so we always want to go to 2163 * the next entry on the next loop iteration. 2164 * 2165 * There is an edge condition here that can occur if 2166 * there is no entry at or before this address. In 2167 * this situation, vm_map_lookup_entry returns 2168 * &map->header, which would cause our loop to abort 2169 * without processing the rest of the map. However, 2170 * in practice this will never happen for process 2171 * vm_map. This is because the executable's text 2172 * segment is the first mapping in the proc's address 2173 * space, and this mapping is never removed until the 2174 * process exits, so there will always be a non-header 2175 * entry at or before the requested address for 2176 * vm_map_lookup_entry to return. 2177 */ 2178 if (map->timestamp != last_timestamp) 2179 vm_map_lookup_entry(map, last_end - 1, &entry); 2180 } 2181 2182 vm_map_unlock_read(map); 2183 vmspace_free(vm); 2184 return; 2185 } 2186 2187 /* 2188 * Log mappings for all processes in the system. 2189 */ 2190 static void 2191 pmc_log_all_process_mappings(struct pmc_owner *po) 2192 { 2193 struct proc *p, *top; 2194 2195 sx_assert(&pmc_sx, SX_XLOCKED); 2196 2197 if ((p = pfind(1)) == NULL) 2198 panic("[pmc,%d] Cannot find init", __LINE__); 2199 2200 PROC_UNLOCK(p); 2201 2202 sx_slock(&proctree_lock); 2203 2204 top = p; 2205 for (;;) { 2206 pmc_log_process_mappings(po, p); 2207 if (!LIST_EMPTY(&p->p_children)) 2208 p = LIST_FIRST(&p->p_children); 2209 else for (;;) { 2210 if (p == top) 2211 goto done; 2212 if (LIST_NEXT(p, p_sibling)) { 2213 p = LIST_NEXT(p, p_sibling); 2214 break; 2215 } 2216 p = p->p_pptr; 2217 } 2218 } 2219 done: 2220 sx_sunlock(&proctree_lock); 2221 } 2222 2223 #ifdef HWPMC_DEBUG 2224 const char *pmc_hooknames[] = { 2225 /* these strings correspond to PMC_FN_* in <sys/pmckern.h> */ 2226 "", 2227 "EXEC", 2228 "CSW-IN", 2229 "CSW-OUT", 2230 "SAMPLE", 2231 "UNUSED1", 2232 "UNUSED2", 2233 "MMAP", 2234 "MUNMAP", 2235 "CALLCHAIN-NMI", 2236 "CALLCHAIN-SOFT", 2237 "SOFTSAMPLING", 2238 "THR-CREATE", 2239 "THR-EXIT", 2240 "THR-USERRET", 2241 "THR-CREATE-LOG", 2242 "THR-EXIT-LOG", 2243 "PROC-CREATE-LOG" 2244 }; 2245 #endif 2246 2247 /* 2248 * The 'hook' invoked from the kernel proper 2249 */ 2250 static int 2251 pmc_hook_handler(struct thread *td, int function, void *arg) 2252 { 2253 int cpu; 2254 2255 PMCDBG4(MOD,PMH,1, "hook td=%p func=%d \"%s\" arg=%p", td, function, 2256 pmc_hooknames[function], arg); 2257 2258 switch (function) { 2259 case PMC_FN_PROCESS_EXEC: 2260 pmc_process_exec(td, (struct pmckern_procexec *)arg); 2261 break; 2262 2263 case PMC_FN_CSW_IN: 2264 pmc_process_csw_in(td); 2265 break; 2266 2267 case PMC_FN_CSW_OUT: 2268 pmc_process_csw_out(td); 2269 break; 2270 2271 /* 2272 * Process accumulated PC samples. 2273 * 2274 * This function is expected to be called by hardclock() for 2275 * each CPU that has accumulated PC samples. 2276 * 2277 * This function is to be executed on the CPU whose samples 2278 * are being processed. 2279 */ 2280 case PMC_FN_DO_SAMPLES: 2281 /* 2282 * Clear the cpu specific bit in the CPU mask before 2283 * do the rest of the processing. If the NMI handler 2284 * gets invoked after the "atomic_clear_int()" call 2285 * below but before "pmc_process_samples()" gets 2286 * around to processing the interrupt, then we will 2287 * come back here at the next hardclock() tick (and 2288 * may find nothing to do if "pmc_process_samples()" 2289 * had already processed the interrupt). We don't 2290 * lose the interrupt sample. 2291 */ 2292 DPCPU_SET(pmc_sampled, 0); 2293 cpu = PCPU_GET(cpuid); 2294 pmc_process_samples(cpu, PMC_HR); 2295 pmc_process_samples(cpu, PMC_SR); 2296 pmc_process_samples(cpu, PMC_UR); 2297 break; 2298 2299 case PMC_FN_MMAP: 2300 pmc_process_mmap(td, (struct pmckern_map_in *)arg); 2301 break; 2302 2303 case PMC_FN_MUNMAP: 2304 MPASS(in_epoch(global_epoch_preempt) || sx_xlocked(&pmc_sx)); 2305 pmc_process_munmap(td, (struct pmckern_map_out *)arg); 2306 break; 2307 2308 case PMC_FN_PROC_CREATE_LOG: 2309 pmc_process_proccreate((struct proc *)arg); 2310 break; 2311 2312 case PMC_FN_USER_CALLCHAIN: 2313 /* 2314 * Record a call chain. 2315 */ 2316 KASSERT(td == curthread, ("[pmc,%d] td != curthread", 2317 __LINE__)); 2318 2319 pmc_capture_user_callchain(PCPU_GET(cpuid), PMC_HR, 2320 (struct trapframe *)arg); 2321 2322 KASSERT(td->td_pinned == 1, 2323 ("[pmc,%d] invalid td_pinned value", __LINE__)); 2324 sched_unpin(); /* Can migrate safely now. */ 2325 2326 td->td_pflags &= ~TDP_CALLCHAIN; 2327 break; 2328 2329 case PMC_FN_USER_CALLCHAIN_SOFT: 2330 /* 2331 * Record a call chain. 2332 */ 2333 KASSERT(td == curthread, ("[pmc,%d] td != curthread", 2334 __LINE__)); 2335 2336 cpu = PCPU_GET(cpuid); 2337 pmc_capture_user_callchain(cpu, PMC_SR, 2338 (struct trapframe *) arg); 2339 2340 KASSERT(td->td_pinned == 1, 2341 ("[pmc,%d] invalid td_pinned value", __LINE__)); 2342 2343 sched_unpin(); /* Can migrate safely now. */ 2344 2345 td->td_pflags &= ~TDP_CALLCHAIN; 2346 break; 2347 2348 case PMC_FN_SOFT_SAMPLING: 2349 /* 2350 * Call soft PMC sampling intr. 2351 */ 2352 pmc_soft_intr((struct pmckern_soft *)arg); 2353 break; 2354 2355 case PMC_FN_THR_CREATE: 2356 pmc_process_thread_add(td); 2357 pmc_process_threadcreate(td); 2358 break; 2359 2360 case PMC_FN_THR_CREATE_LOG: 2361 pmc_process_threadcreate(td); 2362 break; 2363 2364 case PMC_FN_THR_EXIT: 2365 KASSERT(td == curthread, ("[pmc,%d] td != curthread", 2366 __LINE__)); 2367 pmc_process_thread_delete(td); 2368 pmc_process_threadexit(td); 2369 break; 2370 case PMC_FN_THR_EXIT_LOG: 2371 pmc_process_threadexit(td); 2372 break; 2373 case PMC_FN_THR_USERRET: 2374 KASSERT(td == curthread, ("[pmc,%d] td != curthread", 2375 __LINE__)); 2376 pmc_process_thread_userret(td); 2377 break; 2378 default: 2379 #ifdef HWPMC_DEBUG 2380 KASSERT(0, ("[pmc,%d] unknown hook %d\n", __LINE__, function)); 2381 #endif 2382 break; 2383 } 2384 2385 return (0); 2386 } 2387 2388 /* 2389 * Allocate a 'struct pmc_owner' descriptor in the owner hash table. 2390 */ 2391 static struct pmc_owner * 2392 pmc_allocate_owner_descriptor(struct proc *p) 2393 { 2394 struct pmc_owner *po; 2395 struct pmc_ownerhash *poh; 2396 uint32_t hindex; 2397 2398 hindex = PMC_HASH_PTR(p, pmc_ownerhashmask); 2399 poh = &pmc_ownerhash[hindex]; 2400 2401 /* Allocate space for N pointers and one descriptor struct. */ 2402 po = malloc(sizeof(struct pmc_owner), M_PMC, M_WAITOK | M_ZERO); 2403 po->po_owner = p; 2404 LIST_INSERT_HEAD(poh, po, po_next); /* insert into hash table */ 2405 2406 TAILQ_INIT(&po->po_logbuffers); 2407 mtx_init(&po->po_mtx, "pmc-owner-mtx", "pmc-per-proc", MTX_SPIN); 2408 2409 PMCDBG4(OWN,ALL,1, "allocate-owner proc=%p (%d, %s) pmc-owner=%p", 2410 p, p->p_pid, p->p_comm, po); 2411 2412 return (po); 2413 } 2414 2415 static void 2416 pmc_destroy_owner_descriptor(struct pmc_owner *po) 2417 { 2418 2419 PMCDBG4(OWN,REL,1, "destroy-owner po=%p proc=%p (%d, %s)", 2420 po, po->po_owner, po->po_owner->p_pid, po->po_owner->p_comm); 2421 2422 mtx_destroy(&po->po_mtx); 2423 free(po, M_PMC); 2424 } 2425 2426 /* 2427 * Allocate a thread descriptor from the free pool. 2428 * 2429 * NOTE: This *can* return NULL. 2430 */ 2431 static struct pmc_thread * 2432 pmc_thread_descriptor_pool_alloc(void) 2433 { 2434 struct pmc_thread *pt; 2435 2436 mtx_lock_spin(&pmc_threadfreelist_mtx); 2437 if ((pt = LIST_FIRST(&pmc_threadfreelist)) != NULL) { 2438 LIST_REMOVE(pt, pt_next); 2439 pmc_threadfreelist_entries--; 2440 } 2441 mtx_unlock_spin(&pmc_threadfreelist_mtx); 2442 2443 return (pt); 2444 } 2445 2446 /* 2447 * Add a thread descriptor to the free pool. We use this instead of free() 2448 * to maintain a cache of free entries. Additionally, we can safely call 2449 * this function when we cannot call free(), such as in a critical section. 2450 */ 2451 static void 2452 pmc_thread_descriptor_pool_free(struct pmc_thread *pt) 2453 { 2454 2455 if (pt == NULL) 2456 return; 2457 2458 memset(pt, 0, THREADENTRY_SIZE); 2459 mtx_lock_spin(&pmc_threadfreelist_mtx); 2460 LIST_INSERT_HEAD(&pmc_threadfreelist, pt, pt_next); 2461 pmc_threadfreelist_entries++; 2462 if (pmc_threadfreelist_entries > pmc_threadfreelist_max) 2463 taskqueue_enqueue(taskqueue_fast, &free_task); 2464 mtx_unlock_spin(&pmc_threadfreelist_mtx); 2465 } 2466 2467 /* 2468 * An asynchronous task to manage the free list. 2469 */ 2470 static void 2471 pmc_thread_descriptor_pool_free_task(void *arg __unused, int pending __unused) 2472 { 2473 struct pmc_thread *pt; 2474 LIST_HEAD(, pmc_thread) tmplist; 2475 int delta; 2476 2477 LIST_INIT(&tmplist); 2478 2479 /* Determine what changes, if any, we need to make. */ 2480 mtx_lock_spin(&pmc_threadfreelist_mtx); 2481 delta = pmc_threadfreelist_entries - pmc_threadfreelist_max; 2482 while (delta > 0 && (pt = LIST_FIRST(&pmc_threadfreelist)) != NULL) { 2483 delta--; 2484 pmc_threadfreelist_entries--; 2485 LIST_REMOVE(pt, pt_next); 2486 LIST_INSERT_HEAD(&tmplist, pt, pt_next); 2487 } 2488 mtx_unlock_spin(&pmc_threadfreelist_mtx); 2489 2490 /* If there are entries to free, free them. */ 2491 while (!LIST_EMPTY(&tmplist)) { 2492 pt = LIST_FIRST(&tmplist); 2493 LIST_REMOVE(pt, pt_next); 2494 free(pt, M_PMC); 2495 } 2496 } 2497 2498 /* 2499 * Drain the thread free pool, freeing all allocations. 2500 */ 2501 static void 2502 pmc_thread_descriptor_pool_drain(void) 2503 { 2504 struct pmc_thread *pt, *next; 2505 2506 LIST_FOREACH_SAFE(pt, &pmc_threadfreelist, pt_next, next) { 2507 LIST_REMOVE(pt, pt_next); 2508 free(pt, M_PMC); 2509 } 2510 } 2511 2512 /* 2513 * find the descriptor corresponding to thread 'td', adding or removing it 2514 * as specified by 'mode'. 2515 * 2516 * Note that this supports additional mode flags in addition to those 2517 * supported by pmc_find_process_descriptor(): 2518 * PMC_FLAG_NOWAIT: Causes the function to not wait for mallocs. 2519 * This makes it safe to call while holding certain other locks. 2520 */ 2521 static struct pmc_thread * 2522 pmc_find_thread_descriptor(struct pmc_process *pp, struct thread *td, 2523 uint32_t mode) 2524 { 2525 struct pmc_thread *pt = NULL, *ptnew = NULL; 2526 int wait_flag; 2527 2528 KASSERT(td != NULL, ("[pmc,%d] called to add NULL td", __LINE__)); 2529 2530 /* 2531 * Pre-allocate memory in the PMC_FLAG_ALLOCATE case prior to 2532 * acquiring the lock. 2533 */ 2534 if ((mode & PMC_FLAG_ALLOCATE) != 0) { 2535 if ((ptnew = pmc_thread_descriptor_pool_alloc()) == NULL) { 2536 wait_flag = M_WAITOK; 2537 if ((mode & PMC_FLAG_NOWAIT) != 0 || 2538 in_epoch(global_epoch_preempt)) 2539 wait_flag = M_NOWAIT; 2540 2541 ptnew = malloc(THREADENTRY_SIZE, M_PMC, 2542 wait_flag | M_ZERO); 2543 } 2544 } 2545 2546 mtx_lock_spin(pp->pp_tdslock); 2547 LIST_FOREACH(pt, &pp->pp_tds, pt_next) { 2548 if (pt->pt_td == td) 2549 break; 2550 } 2551 2552 if ((mode & PMC_FLAG_REMOVE) != 0 && pt != NULL) 2553 LIST_REMOVE(pt, pt_next); 2554 2555 if ((mode & PMC_FLAG_ALLOCATE) != 0 && pt == NULL && ptnew != NULL) { 2556 pt = ptnew; 2557 ptnew = NULL; 2558 pt->pt_td = td; 2559 LIST_INSERT_HEAD(&pp->pp_tds, pt, pt_next); 2560 } 2561 2562 mtx_unlock_spin(pp->pp_tdslock); 2563 2564 if (ptnew != NULL) { 2565 free(ptnew, M_PMC); 2566 } 2567 2568 return (pt); 2569 } 2570 2571 /* 2572 * Try to add thread descriptors for each thread in a process. 2573 */ 2574 static void 2575 pmc_add_thread_descriptors_from_proc(struct proc *p, struct pmc_process *pp) 2576 { 2577 struct pmc_thread **tdlist; 2578 struct thread *curtd; 2579 int i, tdcnt, tdlistsz; 2580 2581 KASSERT(!PROC_LOCKED(p), ("[pmc,%d] proc unexpectedly locked", 2582 __LINE__)); 2583 tdcnt = 32; 2584 restart: 2585 tdlistsz = roundup2(tdcnt, 32); 2586 2587 tdcnt = 0; 2588 tdlist = malloc(sizeof(struct pmc_thread *) * tdlistsz, M_TEMP, 2589 M_WAITOK); 2590 2591 PROC_LOCK(p); 2592 FOREACH_THREAD_IN_PROC(p, curtd) 2593 tdcnt++; 2594 if (tdcnt >= tdlistsz) { 2595 PROC_UNLOCK(p); 2596 free(tdlist, M_TEMP); 2597 goto restart; 2598 } 2599 2600 /* 2601 * Try to add each thread to the list without sleeping. If unable, 2602 * add to a queue to retry after dropping the process lock. 2603 */ 2604 tdcnt = 0; 2605 FOREACH_THREAD_IN_PROC(p, curtd) { 2606 tdlist[tdcnt] = pmc_find_thread_descriptor(pp, curtd, 2607 PMC_FLAG_ALLOCATE | PMC_FLAG_NOWAIT); 2608 if (tdlist[tdcnt] == NULL) { 2609 PROC_UNLOCK(p); 2610 for (i = 0; i <= tdcnt; i++) 2611 pmc_thread_descriptor_pool_free(tdlist[i]); 2612 free(tdlist, M_TEMP); 2613 goto restart; 2614 } 2615 tdcnt++; 2616 } 2617 PROC_UNLOCK(p); 2618 free(tdlist, M_TEMP); 2619 } 2620 2621 /* 2622 * Find the descriptor corresponding to process 'p', adding or removing it 2623 * as specified by 'mode'. 2624 */ 2625 static struct pmc_process * 2626 pmc_find_process_descriptor(struct proc *p, uint32_t mode) 2627 { 2628 struct pmc_process *pp, *ppnew; 2629 struct pmc_processhash *pph; 2630 uint32_t hindex; 2631 2632 hindex = PMC_HASH_PTR(p, pmc_processhashmask); 2633 pph = &pmc_processhash[hindex]; 2634 2635 ppnew = NULL; 2636 2637 /* 2638 * Pre-allocate memory in the PMC_FLAG_ALLOCATE case since we 2639 * cannot call malloc(9) once we hold a spin lock. 2640 */ 2641 if ((mode & PMC_FLAG_ALLOCATE) != 0) 2642 ppnew = malloc(sizeof(struct pmc_process) + md->pmd_npmc * 2643 sizeof(struct pmc_targetstate), M_PMC, M_WAITOK | M_ZERO); 2644 2645 mtx_lock_spin(&pmc_processhash_mtx); 2646 LIST_FOREACH(pp, pph, pp_next) { 2647 if (pp->pp_proc == p) 2648 break; 2649 } 2650 2651 if ((mode & PMC_FLAG_REMOVE) != 0 && pp != NULL) 2652 LIST_REMOVE(pp, pp_next); 2653 2654 if ((mode & PMC_FLAG_ALLOCATE) != 0 && pp == NULL && ppnew != NULL) { 2655 ppnew->pp_proc = p; 2656 LIST_INIT(&ppnew->pp_tds); 2657 ppnew->pp_tdslock = mtx_pool_find(pmc_mtxpool, ppnew); 2658 LIST_INSERT_HEAD(pph, ppnew, pp_next); 2659 mtx_unlock_spin(&pmc_processhash_mtx); 2660 pp = ppnew; 2661 ppnew = NULL; 2662 2663 /* Add thread descriptors for this process' current threads. */ 2664 pmc_add_thread_descriptors_from_proc(p, pp); 2665 } else 2666 mtx_unlock_spin(&pmc_processhash_mtx); 2667 2668 if (ppnew != NULL) 2669 free(ppnew, M_PMC); 2670 return (pp); 2671 } 2672 2673 /* 2674 * Remove a process descriptor from the process hash table. 2675 */ 2676 static void 2677 pmc_remove_process_descriptor(struct pmc_process *pp) 2678 { 2679 KASSERT(pp->pp_refcnt == 0, 2680 ("[pmc,%d] Removing process descriptor %p with count %d", 2681 __LINE__, pp, pp->pp_refcnt)); 2682 2683 mtx_lock_spin(&pmc_processhash_mtx); 2684 LIST_REMOVE(pp, pp_next); 2685 mtx_unlock_spin(&pmc_processhash_mtx); 2686 } 2687 2688 /* 2689 * Destroy a process descriptor. 2690 */ 2691 static void 2692 pmc_destroy_process_descriptor(struct pmc_process *pp) 2693 { 2694 struct pmc_thread *pmc_td; 2695 2696 while ((pmc_td = LIST_FIRST(&pp->pp_tds)) != NULL) { 2697 LIST_REMOVE(pmc_td, pt_next); 2698 pmc_thread_descriptor_pool_free(pmc_td); 2699 } 2700 free(pp, M_PMC); 2701 } 2702 2703 /* 2704 * Find an owner descriptor corresponding to proc 'p'. 2705 */ 2706 static struct pmc_owner * 2707 pmc_find_owner_descriptor(struct proc *p) 2708 { 2709 struct pmc_owner *po; 2710 struct pmc_ownerhash *poh; 2711 uint32_t hindex; 2712 2713 hindex = PMC_HASH_PTR(p, pmc_ownerhashmask); 2714 poh = &pmc_ownerhash[hindex]; 2715 2716 po = NULL; 2717 LIST_FOREACH(po, poh, po_next) { 2718 if (po->po_owner == p) 2719 break; 2720 } 2721 2722 PMCDBG5(OWN,FND,1, "find-owner proc=%p (%d, %s) hindex=0x%x -> " 2723 "pmc-owner=%p", p, p->p_pid, p->p_comm, hindex, po); 2724 2725 return (po); 2726 } 2727 2728 /* 2729 * Allocate a pmc descriptor and initialize its fields. 2730 */ 2731 static struct pmc * 2732 pmc_allocate_pmc_descriptor(void) 2733 { 2734 struct pmc *pmc; 2735 2736 pmc = malloc(sizeof(struct pmc), M_PMC, M_WAITOK | M_ZERO); 2737 pmc->pm_runcount = counter_u64_alloc(M_WAITOK); 2738 pmc->pm_pcpu_state = malloc(sizeof(struct pmc_pcpu_state) * mp_ncpus, 2739 M_PMC, M_WAITOK | M_ZERO); 2740 PMCDBG1(PMC,ALL,1, "allocate-pmc -> pmc=%p", pmc); 2741 2742 return (pmc); 2743 } 2744 2745 /* 2746 * Destroy a pmc descriptor. 2747 */ 2748 static void 2749 pmc_destroy_pmc_descriptor(struct pmc *pm) 2750 { 2751 2752 KASSERT(pm->pm_state == PMC_STATE_DELETED || 2753 pm->pm_state == PMC_STATE_FREE, 2754 ("[pmc,%d] destroying non-deleted PMC", __LINE__)); 2755 KASSERT(LIST_EMPTY(&pm->pm_targets), 2756 ("[pmc,%d] destroying pmc with targets", __LINE__)); 2757 KASSERT(pm->pm_owner == NULL, 2758 ("[pmc,%d] destroying pmc attached to an owner", __LINE__)); 2759 KASSERT(counter_u64_fetch(pm->pm_runcount) == 0, 2760 ("[pmc,%d] pmc has non-zero run count %ju", __LINE__, 2761 (uintmax_t)counter_u64_fetch(pm->pm_runcount))); 2762 2763 counter_u64_free(pm->pm_runcount); 2764 free(pm->pm_pcpu_state, M_PMC); 2765 free(pm, M_PMC); 2766 } 2767 2768 static void 2769 pmc_wait_for_pmc_idle(struct pmc *pm) 2770 { 2771 #ifdef INVARIANTS 2772 volatile int maxloop; 2773 2774 maxloop = 100 * pmc_cpu_max(); 2775 #endif 2776 /* 2777 * Loop (with a forced context switch) till the PMC's runcount 2778 * comes down to zero. 2779 */ 2780 pmclog_flush(pm->pm_owner, 1); 2781 while (counter_u64_fetch(pm->pm_runcount) > 0) { 2782 pmclog_flush(pm->pm_owner, 1); 2783 #ifdef INVARIANTS 2784 maxloop--; 2785 KASSERT(maxloop > 0, 2786 ("[pmc,%d] (ri%d, rc%ju) waiting too long for " 2787 "pmc to be free", __LINE__, PMC_TO_ROWINDEX(pm), 2788 (uintmax_t)counter_u64_fetch(pm->pm_runcount))); 2789 #endif 2790 pmc_force_context_switch(); 2791 } 2792 } 2793 2794 /* 2795 * This function does the following things: 2796 * 2797 * - detaches the PMC from hardware 2798 * - unlinks all target threads that were attached to it 2799 * - removes the PMC from its owner's list 2800 * - destroys the PMC private mutex 2801 * 2802 * Once this function completes, the given pmc pointer can be freed by 2803 * calling pmc_destroy_pmc_descriptor(). 2804 */ 2805 static void 2806 pmc_release_pmc_descriptor(struct pmc *pm) 2807 { 2808 struct pmc_binding pb; 2809 struct pmc_classdep *pcd; 2810 struct pmc_hw *phw __diagused; 2811 struct pmc_owner *po; 2812 struct pmc_process *pp; 2813 struct pmc_target *ptgt, *tmp; 2814 enum pmc_mode mode; 2815 u_int adjri, ri, cpu; 2816 2817 sx_assert(&pmc_sx, SX_XLOCKED); 2818 KASSERT(pm, ("[pmc,%d] null pmc", __LINE__)); 2819 2820 ri = PMC_TO_ROWINDEX(pm); 2821 pcd = pmc_ri_to_classdep(md, ri, &adjri); 2822 mode = PMC_TO_MODE(pm); 2823 2824 PMCDBG3(PMC,REL,1, "release-pmc pmc=%p ri=%d mode=%d", pm, ri, 2825 mode); 2826 2827 /* 2828 * First, we take the PMC off hardware. 2829 */ 2830 cpu = 0; 2831 if (PMC_IS_SYSTEM_MODE(mode)) { 2832 /* 2833 * A system mode PMC runs on a specific CPU. Switch 2834 * to this CPU and turn hardware off. 2835 */ 2836 pmc_save_cpu_binding(&pb); 2837 cpu = PMC_TO_CPU(pm); 2838 pmc_select_cpu(cpu); 2839 2840 /* switch off non-stalled CPUs */ 2841 pm->pm_pcpu_state[cpu].pps_cpustate = 0; 2842 if (pm->pm_state == PMC_STATE_RUNNING && 2843 pm->pm_pcpu_state[cpu].pps_stalled == 0) { 2844 2845 phw = pmc_pcpu[cpu]->pc_hwpmcs[ri]; 2846 2847 KASSERT(phw->phw_pmc == pm, 2848 ("[pmc, %d] pmc ptr ri(%d) hw(%p) pm(%p)", 2849 __LINE__, ri, phw->phw_pmc, pm)); 2850 PMCDBG2(PMC,REL,2, "stopping cpu=%d ri=%d", cpu, ri); 2851 2852 critical_enter(); 2853 (void)pcd->pcd_stop_pmc(cpu, adjri, pm); 2854 critical_exit(); 2855 } 2856 2857 PMCDBG2(PMC,REL,2, "decfg cpu=%d ri=%d", cpu, ri); 2858 2859 critical_enter(); 2860 (void)pcd->pcd_config_pmc(cpu, adjri, NULL); 2861 critical_exit(); 2862 2863 /* adjust the global and process count of SS mode PMCs */ 2864 if (mode == PMC_MODE_SS && pm->pm_state == PMC_STATE_RUNNING) { 2865 po = pm->pm_owner; 2866 po->po_sscount--; 2867 if (po->po_sscount == 0) { 2868 atomic_subtract_rel_int(&pmc_ss_count, 1); 2869 CK_LIST_REMOVE(po, po_ssnext); 2870 epoch_wait_preempt(global_epoch_preempt); 2871 } 2872 } 2873 pm->pm_state = PMC_STATE_DELETED; 2874 2875 pmc_restore_cpu_binding(&pb); 2876 2877 /* 2878 * We could have references to this PMC structure in the 2879 * per-cpu sample queues. Wait for the queue to drain. 2880 */ 2881 pmc_wait_for_pmc_idle(pm); 2882 2883 } else if (PMC_IS_VIRTUAL_MODE(mode)) { 2884 /* 2885 * A virtual PMC could be running on multiple CPUs at a given 2886 * instant. 2887 * 2888 * By marking its state as DELETED, we ensure that this PMC is 2889 * never further scheduled on hardware. 2890 * 2891 * Then we wait till all CPUs are done with this PMC. 2892 */ 2893 pm->pm_state = PMC_STATE_DELETED; 2894 2895 /* Wait for the PMCs runcount to come to zero. */ 2896 pmc_wait_for_pmc_idle(pm); 2897 2898 /* 2899 * At this point the PMC is off all CPUs and cannot be freshly 2900 * scheduled onto a CPU. It is now safe to unlink all targets 2901 * from this PMC. If a process-record's refcount falls to zero, 2902 * we remove it from the hash table. The module-wide SX lock 2903 * protects us from races. 2904 */ 2905 LIST_FOREACH_SAFE(ptgt, &pm->pm_targets, pt_next, tmp) { 2906 pp = ptgt->pt_process; 2907 pmc_unlink_target_process(pm, pp); /* frees 'ptgt' */ 2908 2909 PMCDBG1(PMC,REL,3, "pp->refcnt=%d", pp->pp_refcnt); 2910 2911 /* 2912 * If the target process record shows that no PMCs are 2913 * attached to it, reclaim its space. 2914 */ 2915 if (pp->pp_refcnt == 0) { 2916 pmc_remove_process_descriptor(pp); 2917 pmc_destroy_process_descriptor(pp); 2918 } 2919 } 2920 2921 cpu = curthread->td_oncpu; /* setup cpu for pmd_release() */ 2922 } 2923 2924 /* 2925 * Release any MD resources. 2926 */ 2927 (void)pcd->pcd_release_pmc(cpu, adjri, pm); 2928 2929 /* 2930 * Update row disposition. 2931 */ 2932 if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm))) 2933 PMC_UNMARK_ROW_STANDALONE(ri); 2934 else 2935 PMC_UNMARK_ROW_THREAD(ri); 2936 2937 /* Unlink from the owner's list. */ 2938 if (pm->pm_owner != NULL) { 2939 LIST_REMOVE(pm, pm_next); 2940 pm->pm_owner = NULL; 2941 } 2942 } 2943 2944 /* 2945 * Register an owner and a pmc. 2946 */ 2947 static int 2948 pmc_register_owner(struct proc *p, struct pmc *pmc) 2949 { 2950 struct pmc_owner *po; 2951 2952 sx_assert(&pmc_sx, SX_XLOCKED); 2953 2954 if ((po = pmc_find_owner_descriptor(p)) == NULL) { 2955 if ((po = pmc_allocate_owner_descriptor(p)) == NULL) 2956 return (ENOMEM); 2957 } 2958 2959 KASSERT(pmc->pm_owner == NULL, 2960 ("[pmc,%d] attempting to own an initialized PMC", __LINE__)); 2961 pmc->pm_owner = po; 2962 2963 LIST_INSERT_HEAD(&po->po_pmcs, pmc, pm_next); 2964 2965 PROC_LOCK(p); 2966 p->p_flag |= P_HWPMC; 2967 PROC_UNLOCK(p); 2968 2969 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) != 0) 2970 pmclog_process_pmcallocate(pmc); 2971 2972 PMCDBG2(PMC,REG,1, "register-owner pmc-owner=%p pmc=%p", 2973 po, pmc); 2974 2975 return (0); 2976 } 2977 2978 /* 2979 * Return the current row disposition: 2980 * == 0 => FREE 2981 * > 0 => PROCESS MODE 2982 * < 0 => SYSTEM MODE 2983 */ 2984 int 2985 pmc_getrowdisp(int ri) 2986 { 2987 return (pmc_pmcdisp[ri]); 2988 } 2989 2990 /* 2991 * Check if a PMC at row index 'ri' can be allocated to the current 2992 * process. 2993 * 2994 * Allocation can fail if: 2995 * - the current process is already being profiled by a PMC at index 'ri', 2996 * attached to it via OP_PMCATTACH. 2997 * - the current process has already allocated a PMC at index 'ri' 2998 * via OP_ALLOCATE. 2999 */ 3000 static bool 3001 pmc_can_allocate_rowindex(struct proc *p, unsigned int ri, int cpu) 3002 { 3003 struct pmc *pm; 3004 struct pmc_owner *po; 3005 struct pmc_process *pp; 3006 enum pmc_mode mode; 3007 3008 PMCDBG5(PMC,ALR,1, "can-allocate-rowindex proc=%p (%d, %s) ri=%d " 3009 "cpu=%d", p, p->p_pid, p->p_comm, ri, cpu); 3010 3011 /* 3012 * We shouldn't have already allocated a process-mode PMC at 3013 * row index 'ri'. 3014 * 3015 * We shouldn't have allocated a system-wide PMC on the same 3016 * CPU and same RI. 3017 */ 3018 if ((po = pmc_find_owner_descriptor(p)) != NULL) { 3019 LIST_FOREACH(pm, &po->po_pmcs, pm_next) { 3020 if (PMC_TO_ROWINDEX(pm) == ri) { 3021 mode = PMC_TO_MODE(pm); 3022 if (PMC_IS_VIRTUAL_MODE(mode)) 3023 return (false); 3024 if (PMC_IS_SYSTEM_MODE(mode) && 3025 PMC_TO_CPU(pm) == cpu) 3026 return (false); 3027 } 3028 } 3029 } 3030 3031 /* 3032 * We also shouldn't be the target of any PMC at this index 3033 * since otherwise a PMC_ATTACH to ourselves will fail. 3034 */ 3035 if ((pp = pmc_find_process_descriptor(p, 0)) != NULL) 3036 if (pp->pp_pmcs[ri].pp_pmc != NULL) 3037 return (false); 3038 3039 PMCDBG4(PMC,ALR,2, "can-allocate-rowindex proc=%p (%d, %s) ri=%d ok", 3040 p, p->p_pid, p->p_comm, ri); 3041 return (true); 3042 } 3043 3044 /* 3045 * Check if a given PMC at row index 'ri' can be currently used in 3046 * mode 'mode'. 3047 */ 3048 static bool 3049 pmc_can_allocate_row(int ri, enum pmc_mode mode) 3050 { 3051 enum pmc_disp disp; 3052 3053 sx_assert(&pmc_sx, SX_XLOCKED); 3054 3055 PMCDBG2(PMC,ALR,1, "can-allocate-row ri=%d mode=%d", ri, mode); 3056 3057 if (PMC_IS_SYSTEM_MODE(mode)) 3058 disp = PMC_DISP_STANDALONE; 3059 else 3060 disp = PMC_DISP_THREAD; 3061 3062 /* 3063 * check disposition for PMC row 'ri': 3064 * 3065 * Expected disposition Row-disposition Result 3066 * 3067 * STANDALONE STANDALONE or FREE proceed 3068 * STANDALONE THREAD fail 3069 * THREAD THREAD or FREE proceed 3070 * THREAD STANDALONE fail 3071 */ 3072 if (!PMC_ROW_DISP_IS_FREE(ri) && 3073 !(disp == PMC_DISP_THREAD && PMC_ROW_DISP_IS_THREAD(ri)) && 3074 !(disp == PMC_DISP_STANDALONE && PMC_ROW_DISP_IS_STANDALONE(ri))) 3075 return (false); 3076 3077 /* 3078 * All OK 3079 */ 3080 PMCDBG2(PMC,ALR,2, "can-allocate-row ri=%d mode=%d ok", ri, mode); 3081 return (true); 3082 } 3083 3084 /* 3085 * Find a PMC descriptor with user handle 'pmcid' for thread 'td'. 3086 */ 3087 static struct pmc * 3088 pmc_find_pmc_descriptor_in_process(struct pmc_owner *po, pmc_id_t pmcid) 3089 { 3090 struct pmc *pm; 3091 3092 KASSERT(PMC_ID_TO_ROWINDEX(pmcid) < md->pmd_npmc, 3093 ("[pmc,%d] Illegal pmc index %d (max %d)", __LINE__, 3094 PMC_ID_TO_ROWINDEX(pmcid), md->pmd_npmc)); 3095 3096 LIST_FOREACH(pm, &po->po_pmcs, pm_next) { 3097 if (pm->pm_id == pmcid) 3098 return (pm); 3099 } 3100 3101 return (NULL); 3102 } 3103 3104 static int 3105 pmc_find_pmc(pmc_id_t pmcid, struct pmc **pmc) 3106 { 3107 struct pmc *pm, *opm; 3108 struct pmc_owner *po; 3109 struct pmc_process *pp; 3110 3111 PMCDBG1(PMC,FND,1, "find-pmc id=%d", pmcid); 3112 if (PMC_ID_TO_ROWINDEX(pmcid) >= md->pmd_npmc) 3113 return (EINVAL); 3114 3115 if ((po = pmc_find_owner_descriptor(curthread->td_proc)) == NULL) { 3116 /* 3117 * In case of PMC_F_DESCENDANTS child processes we will not find 3118 * the current process in the owners hash list. Find the owner 3119 * process first and from there lookup the po. 3120 */ 3121 pp = pmc_find_process_descriptor(curthread->td_proc, 3122 PMC_FLAG_NONE); 3123 if (pp == NULL) 3124 return (ESRCH); 3125 opm = pp->pp_pmcs[PMC_ID_TO_ROWINDEX(pmcid)].pp_pmc; 3126 if (opm == NULL) 3127 return (ESRCH); 3128 if ((opm->pm_flags & 3129 (PMC_F_ATTACHED_TO_OWNER | PMC_F_DESCENDANTS)) != 3130 (PMC_F_ATTACHED_TO_OWNER | PMC_F_DESCENDANTS)) 3131 return (ESRCH); 3132 3133 po = opm->pm_owner; 3134 } 3135 3136 if ((pm = pmc_find_pmc_descriptor_in_process(po, pmcid)) == NULL) 3137 return (EINVAL); 3138 3139 PMCDBG2(PMC,FND,2, "find-pmc id=%d -> pmc=%p", pmcid, pm); 3140 3141 *pmc = pm; 3142 return (0); 3143 } 3144 3145 /* 3146 * Start a PMC. 3147 */ 3148 static int 3149 pmc_start(struct pmc *pm) 3150 { 3151 struct pmc_binding pb; 3152 struct pmc_classdep *pcd; 3153 struct pmc_owner *po; 3154 pmc_value_t v; 3155 enum pmc_mode mode; 3156 int adjri, error, cpu, ri; 3157 3158 KASSERT(pm != NULL, 3159 ("[pmc,%d] null pm", __LINE__)); 3160 3161 mode = PMC_TO_MODE(pm); 3162 ri = PMC_TO_ROWINDEX(pm); 3163 pcd = pmc_ri_to_classdep(md, ri, &adjri); 3164 3165 error = 0; 3166 po = pm->pm_owner; 3167 3168 PMCDBG3(PMC,OPS,1, "start pmc=%p mode=%d ri=%d", pm, mode, ri); 3169 3170 po = pm->pm_owner; 3171 3172 /* 3173 * Disallow PMCSTART if a logfile is required but has not been 3174 * configured yet. 3175 */ 3176 if ((pm->pm_flags & PMC_F_NEEDS_LOGFILE) != 0 && 3177 (po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) 3178 return (EDOOFUS); /* programming error */ 3179 3180 /* 3181 * If this is a sampling mode PMC, log mapping information for 3182 * the kernel modules that are currently loaded. 3183 */ 3184 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) 3185 pmc_log_kernel_mappings(pm); 3186 3187 if (PMC_IS_VIRTUAL_MODE(mode)) { 3188 /* 3189 * If a PMCATTACH has never been done on this PMC, 3190 * attach it to its owner process. 3191 */ 3192 if (LIST_EMPTY(&pm->pm_targets)) { 3193 error = (pm->pm_flags & PMC_F_ATTACH_DONE) != 0 ? 3194 ESRCH : pmc_attach_process(po->po_owner, pm); 3195 } 3196 3197 /* 3198 * If the PMC is attached to its owner, then force a context 3199 * switch to ensure that the MD state gets set correctly. 3200 */ 3201 if (error == 0) { 3202 pm->pm_state = PMC_STATE_RUNNING; 3203 if ((pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) != 0) 3204 pmc_force_context_switch(); 3205 } 3206 3207 return (error); 3208 } 3209 3210 /* 3211 * A system-wide PMC. 3212 * 3213 * Add the owner to the global list if this is a system-wide 3214 * sampling PMC. 3215 */ 3216 if (mode == PMC_MODE_SS) { 3217 /* 3218 * Log mapping information for all existing processes in the 3219 * system. Subsequent mappings are logged as they happen; 3220 * see pmc_process_mmap(). 3221 */ 3222 if (po->po_logprocmaps == 0) { 3223 pmc_log_all_process_mappings(po); 3224 po->po_logprocmaps = 1; 3225 } 3226 po->po_sscount++; 3227 if (po->po_sscount == 1) { 3228 atomic_add_rel_int(&pmc_ss_count, 1); 3229 CK_LIST_INSERT_HEAD(&pmc_ss_owners, po, po_ssnext); 3230 PMCDBG1(PMC,OPS,1, "po=%p in global list", po); 3231 } 3232 } 3233 3234 /* 3235 * Move to the CPU associated with this 3236 * PMC, and start the hardware. 3237 */ 3238 pmc_save_cpu_binding(&pb); 3239 cpu = PMC_TO_CPU(pm); 3240 if (!pmc_cpu_is_active(cpu)) { 3241 return (EXTERROR(ENXIO, "PMC CPU %ju is not active for start", 3242 (uintmax_t)cpu)); 3243 } 3244 pmc_select_cpu(cpu); 3245 3246 /* 3247 * global PMCs are configured at allocation time 3248 * so write out the initial value and start the PMC. 3249 */ 3250 pm->pm_state = PMC_STATE_RUNNING; 3251 3252 critical_enter(); 3253 v = PMC_IS_SAMPLING_MODE(mode) ? pm->pm_sc.pm_reloadcount : 3254 pm->pm_sc.pm_initial; 3255 if ((error = pcd->pcd_write_pmc(cpu, adjri, pm, v)) == 0) { 3256 /* If a sampling mode PMC, reset stalled state. */ 3257 if (PMC_IS_SAMPLING_MODE(mode)) 3258 pm->pm_pcpu_state[cpu].pps_stalled = 0; 3259 3260 /* Indicate that we desire this to run. Start it. */ 3261 pm->pm_pcpu_state[cpu].pps_cpustate = 1; 3262 error = pcd->pcd_start_pmc(cpu, adjri, pm); 3263 } 3264 critical_exit(); 3265 3266 pmc_restore_cpu_binding(&pb); 3267 return (error); 3268 } 3269 3270 /* 3271 * Stop a PMC. 3272 */ 3273 static int 3274 pmc_stop(struct pmc *pm) 3275 { 3276 struct pmc_binding pb; 3277 struct pmc_classdep *pcd; 3278 struct pmc_owner *po; 3279 int adjri, cpu, error, ri; 3280 3281 KASSERT(pm != NULL, ("[pmc,%d] null pmc", __LINE__)); 3282 3283 PMCDBG3(PMC,OPS,1, "stop pmc=%p mode=%d ri=%d", pm, PMC_TO_MODE(pm), 3284 PMC_TO_ROWINDEX(pm)); 3285 3286 pm->pm_state = PMC_STATE_STOPPED; 3287 3288 /* 3289 * If the PMC is a virtual mode one, changing the state to non-RUNNING 3290 * is enough to ensure that the PMC never gets scheduled. 3291 * 3292 * If this PMC is current running on a CPU, then it will handled 3293 * correctly at the time its target process is context switched out. 3294 */ 3295 if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) 3296 return (0); 3297 3298 /* 3299 * A system-mode PMC. Move to the CPU associated with this PMC, and 3300 * stop the hardware. We update the 'initial count' so that a 3301 * subsequent PMCSTART will resume counting from the current hardware 3302 * count. 3303 */ 3304 pmc_save_cpu_binding(&pb); 3305 3306 cpu = PMC_TO_CPU(pm); 3307 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), 3308 ("[pmc,%d] illegal cpu=%d", __LINE__, cpu)); 3309 if (!pmc_cpu_is_active(cpu)) { 3310 return (EXTERROR(ENXIO, "PMC CPU %ju is not active for stop", 3311 (uintmax_t)cpu)); 3312 } 3313 pmc_select_cpu(cpu); 3314 3315 ri = PMC_TO_ROWINDEX(pm); 3316 pcd = pmc_ri_to_classdep(md, ri, &adjri); 3317 3318 pm->pm_pcpu_state[cpu].pps_cpustate = 0; 3319 critical_enter(); 3320 if ((error = pcd->pcd_stop_pmc(cpu, adjri, pm)) == 0) { 3321 error = pcd->pcd_read_pmc(cpu, adjri, pm, 3322 &pm->pm_sc.pm_initial); 3323 } 3324 critical_exit(); 3325 3326 pmc_restore_cpu_binding(&pb); 3327 3328 /* Remove this owner from the global list of SS PMC owners. */ 3329 po = pm->pm_owner; 3330 if (PMC_TO_MODE(pm) == PMC_MODE_SS) { 3331 po->po_sscount--; 3332 if (po->po_sscount == 0) { 3333 atomic_subtract_rel_int(&pmc_ss_count, 1); 3334 CK_LIST_REMOVE(po, po_ssnext); 3335 epoch_wait_preempt(global_epoch_preempt); 3336 PMCDBG1(PMC,OPS,2,"po=%p removed from global list", po); 3337 } 3338 } 3339 3340 return (error); 3341 } 3342 3343 static struct pmc_classdep * 3344 pmc_class_to_classdep(enum pmc_class class) 3345 { 3346 int n; 3347 3348 for (n = 0; n < md->pmd_nclass; n++) { 3349 if (md->pmd_classdep[n].pcd_class == class) 3350 return (&md->pmd_classdep[n]); 3351 } 3352 return (NULL); 3353 } 3354 3355 #if defined(HWPMC_DEBUG) && defined(KTR) 3356 static const char *pmc_op_to_name[] = { 3357 #undef __PMC_OP 3358 #define __PMC_OP(N, D) #N , 3359 __PMC_OPS() 3360 NULL 3361 }; 3362 #endif 3363 3364 /* 3365 * The syscall interface 3366 */ 3367 3368 #define PMC_GET_SX_XLOCK(...) do { \ 3369 sx_xlock(&pmc_sx); \ 3370 if (pmc_hook == NULL) { \ 3371 sx_xunlock(&pmc_sx); \ 3372 return __VA_ARGS__; \ 3373 } \ 3374 } while (0) 3375 3376 #define PMC_DOWNGRADE_SX() do { \ 3377 sx_downgrade(&pmc_sx); \ 3378 is_sx_downgraded = true; \ 3379 } while (0) 3380 3381 /* 3382 * Main body of PMC_OP_PMCALLOCATE. 3383 */ 3384 static int 3385 pmc_do_op_pmcallocate(struct thread *td, struct pmc_op_pmcallocate *pa) 3386 { 3387 struct proc *p; 3388 struct pmc *pmc; 3389 struct pmc_binding pb; 3390 struct pmc_classdep *pcd; 3391 struct pmc_hw *phw; 3392 enum pmc_mode mode; 3393 enum pmc_class class; 3394 uint32_t caps, flags; 3395 u_int cpu; 3396 int adjri, n; 3397 int error; 3398 3399 class = pa->pm_class; 3400 caps = pa->pm_caps; 3401 flags = pa->pm_flags; 3402 mode = pa->pm_mode; 3403 cpu = pa->pm_cpu; 3404 3405 p = td->td_proc; 3406 /* Requested mode must exist. */ 3407 if ((mode != PMC_MODE_SS && mode != PMC_MODE_SC && 3408 mode != PMC_MODE_TS && mode != PMC_MODE_TC)) 3409 return (EXTERROR(EINVAL, "Invalid PMC mode %ju", 3410 (uintmax_t)mode)); 3411 3412 /* Requested CPU must be valid. */ 3413 if (cpu != PMC_CPU_ANY && cpu >= pmc_cpu_max()) 3414 return (EXTERROR(EINVAL, "Invalid PMC CPU %ju", 3415 (uintmax_t)cpu)); 3416 3417 /* 3418 * Virtual PMCs should only ask for a default CPU. 3419 * System mode PMCs need to specify a non-default CPU. 3420 */ 3421 if ((PMC_IS_VIRTUAL_MODE(mode) && cpu != PMC_CPU_ANY) || 3422 (PMC_IS_SYSTEM_MODE(mode) && cpu == PMC_CPU_ANY)) { 3423 if (PMC_IS_VIRTUAL_MODE(mode)) { 3424 return (EXTERROR(EINVAL, 3425 "PMC mode %ju requires the default CPU", 3426 (uintmax_t)mode)); 3427 } 3428 return (EXTERROR(EINVAL, 3429 "PMC mode %ju requires an explicit CPU", 3430 (uintmax_t)mode)); 3431 } 3432 3433 /* 3434 * Check that an inactive CPU is not being asked for. 3435 */ 3436 if (PMC_IS_SYSTEM_MODE(mode) && !pmc_cpu_is_active(cpu)) 3437 return (EXTERROR(ENXIO, "PMC CPU %ju is not active", 3438 (uintmax_t)cpu)); 3439 3440 /* 3441 * Refuse an allocation for a system-wide PMC if this process has been 3442 * jailed, or if this process lacks super-user credentials and the 3443 * sysctl tunable 'security.bsd.unprivileged_syspmcs' is zero. 3444 */ 3445 if (PMC_IS_SYSTEM_MODE(mode)) { 3446 if (jailed(td->td_ucred)) 3447 return (EPERM); 3448 if (!pmc_unprivileged_syspmcs) { 3449 error = priv_check(td, PRIV_PMC_SYSTEM); 3450 if (error != 0) 3451 return (error); 3452 } 3453 } 3454 3455 /* 3456 * Look for valid values for 'pm_flags'. 3457 */ 3458 if ((flags & ~(PMC_F_DESCENDANTS | PMC_F_LOG_PROCCSW | 3459 PMC_F_LOG_PROCEXIT | PMC_F_CALLCHAIN | PMC_F_USERCALLCHAIN | 3460 PMC_F_EV_PMU)) != 0) 3461 return (EXTERROR(EINVAL, "Invalid PMC flags %#jx", 3462 (uintmax_t)flags)); 3463 3464 /* PMC_F_USERCALLCHAIN is only valid with PMC_F_CALLCHAIN. */ 3465 if ((flags & (PMC_F_CALLCHAIN | PMC_F_USERCALLCHAIN)) == 3466 PMC_F_USERCALLCHAIN) 3467 return (EXTERROR(EINVAL, 3468 "PMC_F_USERCALLCHAIN requires PMC_F_CALLCHAIN")); 3469 3470 /* PMC_F_USERCALLCHAIN is only valid for sampling mode. */ 3471 if ((flags & PMC_F_USERCALLCHAIN) != 0 && mode != PMC_MODE_TS && 3472 mode != PMC_MODE_SS) 3473 return (EXTERROR(EINVAL, 3474 "PMC_F_USERCALLCHAIN requires sampling mode")); 3475 3476 /* Process logging options are not allowed for system PMCs. */ 3477 if (PMC_IS_SYSTEM_MODE(mode) && 3478 (flags & (PMC_F_LOG_PROCCSW | PMC_F_LOG_PROCEXIT)) != 0) 3479 return (EXTERROR(EINVAL, 3480 "Process logging flags are not valid for system PMCs")); 3481 3482 /* 3483 * All sampling mode PMCs need to be able to interrupt the CPU. 3484 */ 3485 if (PMC_IS_SAMPLING_MODE(mode)) 3486 caps |= PMC_CAP_INTERRUPT; 3487 3488 /* A valid class specifier should have been passed in. */ 3489 pcd = pmc_class_to_classdep(class); 3490 if (pcd == NULL) 3491 return (EXTERROR(EINVAL, "Invalid PMC class %ju", 3492 (uintmax_t)class)); 3493 3494 /* The requested PMC capabilities should be feasible. */ 3495 if ((pcd->pcd_caps & caps) != caps) 3496 return (EXTERROR(EOPNOTSUPP, 3497 "Requested PMC capabilities %#jx are not supported", 3498 (uintmax_t)caps)); 3499 3500 PMCDBG4(PMC,ALL,2, "event=%d caps=0x%x mode=%d cpu=%d", pa->pm_ev, 3501 caps, mode, cpu); 3502 3503 pmc = pmc_allocate_pmc_descriptor(); 3504 pmc->pm_id = PMC_ID_MAKE_ID(cpu, pa->pm_mode, class, PMC_ID_INVALID); 3505 pmc->pm_event = pa->pm_ev; 3506 pmc->pm_state = PMC_STATE_FREE; 3507 pmc->pm_caps = caps; 3508 pmc->pm_flags = flags; 3509 3510 /* XXX set lower bound on sampling for process counters */ 3511 if (PMC_IS_SAMPLING_MODE(mode)) { 3512 /* 3513 * Don't permit requested sample rate to be less than 3514 * pmc_mincount. 3515 */ 3516 if (pa->pm_count < MAX(1, pmc_mincount)) 3517 log(LOG_WARNING, "pmcallocate: passed sample " 3518 "rate %ju - setting to %u\n", 3519 (uintmax_t)pa->pm_count, 3520 MAX(1, pmc_mincount)); 3521 pmc->pm_sc.pm_reloadcount = MAX(MAX(1, pmc_mincount), 3522 pa->pm_count); 3523 } else 3524 pmc->pm_sc.pm_initial = pa->pm_count; 3525 3526 /* switch thread to CPU 'cpu' */ 3527 pmc_save_cpu_binding(&pb); 3528 3529 #define PMC_IS_SHAREABLE_PMC(cpu, n) \ 3530 (pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_state & \ 3531 PMC_PHW_FLAG_IS_SHAREABLE) 3532 #define PMC_IS_UNALLOCATED(cpu, n) \ 3533 (pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_pmc == NULL) 3534 3535 if (PMC_IS_SYSTEM_MODE(mode)) { 3536 pmc_select_cpu(cpu); 3537 for (n = pcd->pcd_ri; n < md->pmd_npmc; n++) { 3538 pcd = pmc_ri_to_classdep(md, n, &adjri); 3539 3540 if (!pmc_can_allocate_row(n, mode) || 3541 !pmc_can_allocate_rowindex(p, n, cpu)) 3542 continue; 3543 if (!PMC_IS_UNALLOCATED(cpu, n) && 3544 !PMC_IS_SHAREABLE_PMC(cpu, n)) 3545 continue; 3546 3547 if (pcd->pcd_allocate_pmc(cpu, adjri, pmc, pa) == 0) { 3548 /* Success. */ 3549 break; 3550 } 3551 } 3552 } else { 3553 /* Process virtual mode */ 3554 for (n = pcd->pcd_ri; n < md->pmd_npmc; n++) { 3555 pcd = pmc_ri_to_classdep(md, n, &adjri); 3556 3557 if (!pmc_can_allocate_row(n, mode) || 3558 !pmc_can_allocate_rowindex(p, n, PMC_CPU_ANY)) 3559 continue; 3560 3561 if (pcd->pcd_allocate_pmc(td->td_oncpu, adjri, pmc, 3562 pa) == 0) { 3563 /* Success. */ 3564 break; 3565 } 3566 } 3567 } 3568 3569 #undef PMC_IS_UNALLOCATED 3570 #undef PMC_IS_SHAREABLE_PMC 3571 3572 pmc_restore_cpu_binding(&pb); 3573 3574 if (n == md->pmd_npmc) { 3575 pmc_destroy_pmc_descriptor(pmc); 3576 /* Preserve a more specific error from the class allocator. */ 3577 if ((td->td_pflags2 & TDP2_EXTERR) != 0) 3578 return (EINVAL); 3579 return (EXTERROR(EINVAL, 3580 "No PMC row accepted the allocation request")); 3581 } 3582 3583 /* Fill in the correct value in the ID field. */ 3584 pmc->pm_id = PMC_ID_MAKE_ID(cpu, mode, class, n); 3585 3586 PMCDBG5(PMC,ALL,2, "ev=%d class=%d mode=%d n=%d -> pmcid=%x", 3587 pmc->pm_event, class, mode, n, pmc->pm_id); 3588 3589 /* Process mode PMCs with logging enabled need log files. */ 3590 if ((pmc->pm_flags & (PMC_F_LOG_PROCEXIT | PMC_F_LOG_PROCCSW)) != 0) 3591 pmc->pm_flags |= PMC_F_NEEDS_LOGFILE; 3592 3593 /* All system mode sampling PMCs require a log file. */ 3594 if (PMC_IS_SAMPLING_MODE(mode) && PMC_IS_SYSTEM_MODE(mode)) 3595 pmc->pm_flags |= PMC_F_NEEDS_LOGFILE; 3596 3597 /* 3598 * Configure global pmc's immediately. 3599 */ 3600 if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pmc))) { 3601 pmc_save_cpu_binding(&pb); 3602 pmc_select_cpu(cpu); 3603 3604 phw = pmc_pcpu[cpu]->pc_hwpmcs[n]; 3605 pcd = pmc_ri_to_classdep(md, n, &adjri); 3606 3607 if ((phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0) { 3608 (void)pcd->pcd_release_pmc(cpu, adjri, pmc); 3609 pmc_destroy_pmc_descriptor(pmc); 3610 pmc_restore_cpu_binding(&pb); 3611 return (EXTERROR(EPERM, 3612 "PMC row %ju on CPU %ju is not enabled", 3613 (uintmax_t)n, (uintmax_t)cpu)); 3614 } 3615 if ((error = pcd->pcd_config_pmc(cpu, adjri, pmc)) != 0) { 3616 (void)pcd->pcd_release_pmc(cpu, adjri, pmc); 3617 pmc_destroy_pmc_descriptor(pmc); 3618 pmc_restore_cpu_binding(&pb); 3619 return (EXTERROR(EPERM, 3620 "PMC configuration failed for row %ju on CPU %ju", 3621 (uintmax_t)n, (uintmax_t)cpu)); 3622 } 3623 3624 pmc_restore_cpu_binding(&pb); 3625 } 3626 3627 pmc->pm_state = PMC_STATE_ALLOCATED; 3628 pmc->pm_class = class; 3629 3630 /* 3631 * Mark row disposition. 3632 */ 3633 if (PMC_IS_SYSTEM_MODE(mode)) 3634 PMC_MARK_ROW_STANDALONE(n); 3635 else 3636 PMC_MARK_ROW_THREAD(n); 3637 3638 /* 3639 * Register this PMC with the current thread as its owner. 3640 */ 3641 error = pmc_register_owner(p, pmc); 3642 if (error != 0) { 3643 pmc_release_pmc_descriptor(pmc); 3644 pmc_destroy_pmc_descriptor(pmc); 3645 return (EXTERROR(error, "Failed to register PMC owner")); 3646 } 3647 3648 /* 3649 * Return the allocated index. 3650 */ 3651 pa->pm_pmcid = pmc->pm_id; 3652 return (0); 3653 } 3654 3655 /* 3656 * Main body of PMC_OP_PMCATTACH. 3657 */ 3658 static int 3659 pmc_do_op_pmcattach(struct thread *td, struct pmc_op_pmcattach a) 3660 { 3661 struct pmc *pm; 3662 struct proc *p; 3663 int error; 3664 3665 sx_assert(&pmc_sx, SX_XLOCKED); 3666 3667 if (a.pm_pid < 0) { 3668 return (EXTERROR(EINVAL, "Invalid PMC attach pid %jd", 3669 (intmax_t)a.pm_pid)); 3670 } else if (a.pm_pid == 0) { 3671 a.pm_pid = td->td_proc->p_pid; 3672 } 3673 3674 error = pmc_find_pmc(a.pm_pmc, &pm); 3675 if (error != 0) 3676 return (error); 3677 3678 if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm))) 3679 return (EXTERROR(EINVAL, 3680 "Cannot attach a system-mode PMC to a process")); 3681 3682 /* PMCs may be (re)attached only when allocated or stopped */ 3683 if (pm->pm_state == PMC_STATE_RUNNING) { 3684 return (EXTERROR(EBUSY, 3685 "PMC must be stopped before attach")); 3686 } else if (pm->pm_state != PMC_STATE_ALLOCATED && 3687 pm->pm_state != PMC_STATE_STOPPED) { 3688 return (EXTERROR(EINVAL, 3689 "PMC state %ju does not allow attach", 3690 (uintmax_t)pm->pm_state)); 3691 } 3692 3693 /* lookup pid */ 3694 if ((p = pfind(a.pm_pid)) == NULL) 3695 return (ESRCH); 3696 3697 /* 3698 * Ignore processes that are working on exiting. 3699 */ 3700 if ((p->p_flag & P_WEXIT) != 0) { 3701 PROC_UNLOCK(p); /* pfind() returns a locked process */ 3702 return (ESRCH); 3703 } 3704 3705 /* 3706 * We are allowed to attach a PMC to a process if we can debug it. 3707 */ 3708 error = p_candebug(curthread, p); 3709 3710 PROC_UNLOCK(p); 3711 3712 if (error == 0) 3713 error = pmc_attach_process(p, pm); 3714 3715 return (error); 3716 } 3717 3718 /* 3719 * Main body of PMC_OP_PMCDETACH. 3720 */ 3721 static int 3722 pmc_do_op_pmcdetach(struct thread *td, struct pmc_op_pmcattach a) 3723 { 3724 struct pmc *pm; 3725 struct proc *p; 3726 int error; 3727 3728 if (a.pm_pid < 0) { 3729 return (EXTERROR(EINVAL, "Invalid PMC detach pid %jd", 3730 (intmax_t)a.pm_pid)); 3731 } else if (a.pm_pid == 0) 3732 a.pm_pid = td->td_proc->p_pid; 3733 3734 error = pmc_find_pmc(a.pm_pmc, &pm); 3735 if (error != 0) 3736 return (error); 3737 3738 if ((p = pfind(a.pm_pid)) == NULL) 3739 return (ESRCH); 3740 3741 /* 3742 * Treat processes that are in the process of exiting as if they were 3743 * not present. 3744 */ 3745 if ((p->p_flag & P_WEXIT) != 0) { 3746 PROC_UNLOCK(p); 3747 return (ESRCH); 3748 } 3749 3750 PROC_UNLOCK(p); /* pfind() returns a locked process */ 3751 3752 if (error == 0) 3753 error = pmc_detach_process(p, pm); 3754 3755 return (error); 3756 } 3757 3758 /* 3759 * Main body of PMC_OP_PMCRELEASE. 3760 */ 3761 static int 3762 pmc_do_op_pmcrelease(pmc_id_t pmcid) 3763 { 3764 struct pmc_owner *po; 3765 struct pmc *pm; 3766 int error; 3767 3768 /* 3769 * Find PMC pointer for the named PMC. 3770 * 3771 * Use pmc_release_pmc_descriptor() to switch off the 3772 * PMC, remove all its target threads, and remove the 3773 * PMC from its owner's list. 3774 * 3775 * Remove the owner record if this is the last PMC 3776 * owned. 3777 * 3778 * Free up space. 3779 */ 3780 error = pmc_find_pmc(pmcid, &pm); 3781 if (error != 0) 3782 return (error); 3783 3784 po = pm->pm_owner; 3785 pmc_release_pmc_descriptor(pm); 3786 pmc_maybe_remove_owner(po); 3787 pmc_destroy_pmc_descriptor(pm); 3788 3789 return (error); 3790 } 3791 3792 /* 3793 * Main body of PMC_OP_PMCRW. 3794 */ 3795 static int 3796 pmc_do_op_pmcrw(const struct pmc_op_pmcrw *prw, pmc_value_t *valp) 3797 { 3798 struct pmc_binding pb; 3799 struct pmc_classdep *pcd; 3800 struct pmc *pm; 3801 u_int cpu, ri, adjri; 3802 int error; 3803 3804 PMCDBG2(PMC,OPS,1, "rw id=%d flags=0x%x", prw->pm_pmcid, prw->pm_flags); 3805 3806 /* Must have at least one flag set. */ 3807 if ((prw->pm_flags & (PMC_F_OLDVALUE | PMC_F_NEWVALUE)) == 0) 3808 return (EXTERROR(EINVAL, 3809 "PMCRW requires PMC_F_OLDVALUE and/or PMC_F_NEWVALUE")); 3810 3811 /* Locate PMC descriptor. */ 3812 error = pmc_find_pmc(prw->pm_pmcid, &pm); 3813 if (error != 0) 3814 return (error); 3815 3816 /* Can't read a PMC that hasn't been started. */ 3817 if (pm->pm_state != PMC_STATE_ALLOCATED && 3818 pm->pm_state != PMC_STATE_STOPPED && 3819 pm->pm_state != PMC_STATE_RUNNING) 3820 return (EXTERROR(EINVAL, 3821 "PMC state %ju does not allow read/write", 3822 (uintmax_t)pm->pm_state)); 3823 3824 /* Writing a new value is allowed only for 'STOPPED' PMCs. */ 3825 if (pm->pm_state == PMC_STATE_RUNNING && 3826 (prw->pm_flags & PMC_F_NEWVALUE) != 0) 3827 return (EXTERROR(EBUSY, 3828 "Cannot write a PMC while it is running")); 3829 3830 if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) { 3831 /* 3832 * If this PMC is attached to its owner (i.e., the process 3833 * requesting this operation) and is running, then attempt to 3834 * get an upto-date reading from hardware for a READ. Writes 3835 * are only allowed when the PMC is stopped, so only update the 3836 * saved value field. 3837 * 3838 * If the PMC is not running, or is not attached to its owner, 3839 * read/write to the savedvalue field. 3840 */ 3841 3842 ri = PMC_TO_ROWINDEX(pm); 3843 pcd = pmc_ri_to_classdep(md, ri, &adjri); 3844 3845 mtx_pool_lock_spin(pmc_mtxpool, pm); 3846 cpu = curthread->td_oncpu; 3847 3848 if ((prw->pm_flags & PMC_F_OLDVALUE) != 0) { 3849 if ((pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) && 3850 (pm->pm_state == PMC_STATE_RUNNING)) { 3851 error = (*pcd->pcd_read_pmc)(cpu, adjri, pm, 3852 valp); 3853 } else { 3854 *valp = pm->pm_gv.pm_savedvalue; 3855 } 3856 } 3857 3858 if ((prw->pm_flags & PMC_F_NEWVALUE) != 0) 3859 pm->pm_gv.pm_savedvalue = prw->pm_value; 3860 3861 mtx_pool_unlock_spin(pmc_mtxpool, pm); 3862 } else { /* System mode PMCs */ 3863 cpu = PMC_TO_CPU(pm); 3864 ri = PMC_TO_ROWINDEX(pm); 3865 pcd = pmc_ri_to_classdep(md, ri, &adjri); 3866 3867 if (!pmc_cpu_is_active(cpu)) 3868 return (EXTERROR(ENXIO, 3869 "PMC CPU %ju is not active for read/write", 3870 (uintmax_t)cpu)); 3871 3872 /* Move this thread to CPU 'cpu'. */ 3873 pmc_save_cpu_binding(&pb); 3874 pmc_select_cpu(cpu); 3875 critical_enter(); 3876 3877 /* Save old value. */ 3878 if ((prw->pm_flags & PMC_F_OLDVALUE) != 0) 3879 error = (*pcd->pcd_read_pmc)(cpu, adjri, pm, valp); 3880 3881 /* Write out new value. */ 3882 if (error == 0 && (prw->pm_flags & PMC_F_NEWVALUE) != 0) 3883 error = (*pcd->pcd_write_pmc)(cpu, adjri, pm, 3884 prw->pm_value); 3885 3886 critical_exit(); 3887 pmc_restore_cpu_binding(&pb); 3888 if (error != 0) 3889 return (error); 3890 } 3891 3892 #ifdef HWPMC_DEBUG 3893 if ((prw->pm_flags & PMC_F_NEWVALUE) != 0) 3894 PMCDBG3(PMC,OPS,2, "rw id=%d new %jx -> old %jx", 3895 ri, prw->pm_value, *valp); 3896 else 3897 PMCDBG2(PMC,OPS,2, "rw id=%d -> old %jx", ri, *valp); 3898 #endif 3899 return (error); 3900 } 3901 3902 static int 3903 pmc_syscall_handler(struct thread *td, void *syscall_args) 3904 { 3905 struct pmc_syscall_args *c; 3906 void *pmclog_proc_handle; 3907 void *arg; 3908 int error, op; 3909 bool is_sx_downgraded; 3910 3911 c = (struct pmc_syscall_args *)syscall_args; 3912 op = c->pmop_code; 3913 arg = c->pmop_data; 3914 3915 /* PMC isn't set up yet */ 3916 if (pmc_hook == NULL) 3917 return (EINVAL); 3918 3919 if (op == PMC_OP_CONFIGURELOG) { 3920 /* 3921 * We cannot create the logging process inside 3922 * pmclog_configure_log() because there is a LOR 3923 * between pmc_sx and process structure locks. 3924 * Instead, pre-create the process and ignite the loop 3925 * if everything is fine, otherwise direct the process 3926 * to exit. 3927 */ 3928 error = pmclog_proc_create(td, &pmclog_proc_handle); 3929 if (error != 0) 3930 goto done_syscall; 3931 } 3932 3933 PMC_GET_SX_XLOCK(ENOSYS); 3934 is_sx_downgraded = false; 3935 PMCDBG3(MOD,PMS,1, "syscall op=%d \"%s\" arg=%p", op, 3936 pmc_op_to_name[op], arg); 3937 3938 error = 0; 3939 counter_u64_add(pmc_stats.pm_syscalls, 1); 3940 3941 switch (op) { 3942 3943 3944 /* 3945 * Configure a log file. 3946 * 3947 * XXX This OP will be reworked. 3948 */ 3949 3950 case PMC_OP_CONFIGURELOG: 3951 { 3952 struct proc *p; 3953 struct pmc *pm; 3954 struct pmc_owner *po; 3955 struct pmc_op_configurelog cl; 3956 3957 if ((error = copyin(arg, &cl, sizeof(cl))) != 0) { 3958 pmclog_proc_ignite(pmclog_proc_handle, NULL); 3959 break; 3960 } 3961 3962 /* No flags currently implemented */ 3963 if (cl.pm_flags != 0) { 3964 pmclog_proc_ignite(pmclog_proc_handle, NULL); 3965 error = EINVAL; 3966 break; 3967 } 3968 3969 /* mark this process as owning a log file */ 3970 p = td->td_proc; 3971 if ((po = pmc_find_owner_descriptor(p)) == NULL) 3972 if ((po = pmc_allocate_owner_descriptor(p)) == NULL) { 3973 pmclog_proc_ignite(pmclog_proc_handle, NULL); 3974 error = ENOMEM; 3975 break; 3976 } 3977 3978 /* 3979 * If a valid fd was passed in, try to configure that, 3980 * otherwise if 'fd' was less than zero and there was 3981 * a log file configured, flush its buffers and 3982 * de-configure it. 3983 */ 3984 if (cl.pm_logfd >= 0) { 3985 error = pmclog_configure_log(md, po, cl.pm_logfd); 3986 pmclog_proc_ignite(pmclog_proc_handle, error == 0 ? 3987 po : NULL); 3988 } else if (po->po_flags & PMC_PO_OWNS_LOGFILE) { 3989 pmclog_proc_ignite(pmclog_proc_handle, NULL); 3990 error = pmclog_close(po); 3991 if (error == 0) { 3992 LIST_FOREACH(pm, &po->po_pmcs, pm_next) 3993 if (pm->pm_flags & PMC_F_NEEDS_LOGFILE && 3994 pm->pm_state == PMC_STATE_RUNNING) 3995 pmc_stop(pm); 3996 error = pmclog_deconfigure_log(po); 3997 } 3998 } else { 3999 pmclog_proc_ignite(pmclog_proc_handle, NULL); 4000 error = EINVAL; 4001 } 4002 } 4003 break; 4004 4005 /* 4006 * Flush a log file. 4007 */ 4008 4009 case PMC_OP_FLUSHLOG: 4010 { 4011 struct pmc_owner *po; 4012 4013 sx_assert(&pmc_sx, SX_XLOCKED); 4014 4015 if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) { 4016 error = EINVAL; 4017 break; 4018 } 4019 4020 error = pmclog_flush(po, 0); 4021 } 4022 break; 4023 4024 /* 4025 * Close a log file. 4026 */ 4027 4028 case PMC_OP_CLOSELOG: 4029 { 4030 struct pmc_owner *po; 4031 4032 sx_assert(&pmc_sx, SX_XLOCKED); 4033 4034 if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) { 4035 error = EINVAL; 4036 break; 4037 } 4038 4039 error = pmclog_close(po); 4040 } 4041 break; 4042 4043 /* 4044 * Retrieve hardware configuration. 4045 */ 4046 4047 case PMC_OP_GETCPUINFO: /* CPU information */ 4048 { 4049 struct pmc_op_getcpuinfo gci; 4050 struct pmc_classinfo *pci; 4051 struct pmc_classdep *pcd; 4052 int cl; 4053 4054 memset(&gci, 0, sizeof(gci)); 4055 gci.pm_cputype = md->pmd_cputype; 4056 gci.pm_ncpu = pmc_cpu_max(); 4057 gci.pm_npmc = md->pmd_npmc; 4058 gci.pm_nclass = md->pmd_nclass; 4059 pci = gci.pm_classes; 4060 pcd = md->pmd_classdep; 4061 for (cl = 0; cl < md->pmd_nclass; cl++, pci++, pcd++) { 4062 pci->pm_caps = pcd->pcd_caps; 4063 pci->pm_class = pcd->pcd_class; 4064 pci->pm_width = pcd->pcd_width; 4065 pci->pm_num = pcd->pcd_num; 4066 } 4067 error = copyout(&gci, arg, sizeof(gci)); 4068 } 4069 break; 4070 4071 /* 4072 * Retrieve soft events list. 4073 */ 4074 case PMC_OP_GETDYNEVENTINFO: 4075 { 4076 enum pmc_class cl; 4077 enum pmc_event ev; 4078 struct pmc_op_getdyneventinfo *gei; 4079 struct pmc_dyn_event_descr dev; 4080 struct pmc_soft *ps; 4081 uint32_t nevent; 4082 4083 sx_assert(&pmc_sx, SX_LOCKED); 4084 4085 gei = (struct pmc_op_getdyneventinfo *) arg; 4086 4087 if ((error = copyin(&gei->pm_class, &cl, sizeof(cl))) != 0) 4088 break; 4089 4090 /* Only SOFT class is dynamic. */ 4091 if (cl != PMC_CLASS_SOFT) { 4092 error = EINVAL; 4093 break; 4094 } 4095 4096 nevent = 0; 4097 for (ev = PMC_EV_SOFT_FIRST; (int)ev <= PMC_EV_SOFT_LAST; ev++) { 4098 ps = pmc_soft_ev_acquire(ev); 4099 if (ps == NULL) 4100 continue; 4101 bcopy(&ps->ps_ev, &dev, sizeof(dev)); 4102 pmc_soft_ev_release(ps); 4103 4104 error = copyout(&dev, 4105 &gei->pm_events[nevent], 4106 sizeof(struct pmc_dyn_event_descr)); 4107 if (error != 0) 4108 break; 4109 nevent++; 4110 } 4111 if (error != 0) 4112 break; 4113 4114 error = copyout(&nevent, &gei->pm_nevent, 4115 sizeof(nevent)); 4116 } 4117 break; 4118 4119 /* 4120 * Get module statistics 4121 */ 4122 4123 case PMC_OP_GETDRIVERSTATS: 4124 { 4125 struct pmc_op_getdriverstats gms; 4126 #define CFETCH(a, b, field) a.field = counter_u64_fetch(b.field) 4127 CFETCH(gms, pmc_stats, pm_intr_ignored); 4128 CFETCH(gms, pmc_stats, pm_intr_processed); 4129 CFETCH(gms, pmc_stats, pm_intr_bufferfull); 4130 CFETCH(gms, pmc_stats, pm_syscalls); 4131 CFETCH(gms, pmc_stats, pm_syscall_errors); 4132 CFETCH(gms, pmc_stats, pm_buffer_requests); 4133 CFETCH(gms, pmc_stats, pm_buffer_requests_failed); 4134 CFETCH(gms, pmc_stats, pm_log_sweeps); 4135 #undef CFETCH 4136 error = copyout(&gms, arg, sizeof(gms)); 4137 } 4138 break; 4139 4140 4141 /* 4142 * Retrieve module version number 4143 */ 4144 4145 case PMC_OP_GETMODULEVERSION: 4146 { 4147 uint32_t cv, modv; 4148 4149 /* retrieve the client's idea of the ABI version */ 4150 if ((error = copyin(arg, &cv, sizeof(uint32_t))) != 0) 4151 break; 4152 /* don't service clients newer than our driver */ 4153 modv = PMC_VERSION; 4154 if ((cv & 0xFFFF0000) > (modv & 0xFFFF0000)) { 4155 error = EPROGMISMATCH; 4156 break; 4157 } 4158 error = copyout(&modv, arg, sizeof(int)); 4159 } 4160 break; 4161 4162 4163 /* 4164 * Retrieve the state of all the PMCs on a given 4165 * CPU. 4166 */ 4167 4168 case PMC_OP_GETPMCINFO: 4169 { 4170 int ari; 4171 struct pmc *pm; 4172 size_t pmcinfo_size; 4173 uint32_t cpu, n, npmc; 4174 struct pmc_owner *po; 4175 struct pmc_binding pb; 4176 struct pmc_classdep *pcd; 4177 struct pmc_info *p, *pmcinfo; 4178 struct pmc_op_getpmcinfo *gpi; 4179 4180 PMC_DOWNGRADE_SX(); 4181 4182 gpi = (struct pmc_op_getpmcinfo *) arg; 4183 4184 if ((error = copyin(&gpi->pm_cpu, &cpu, sizeof(cpu))) != 0) 4185 break; 4186 4187 if (cpu >= pmc_cpu_max()) { 4188 error = EINVAL; 4189 break; 4190 } 4191 4192 if (!pmc_cpu_is_active(cpu)) { 4193 error = ENXIO; 4194 break; 4195 } 4196 4197 /* switch to CPU 'cpu' */ 4198 pmc_save_cpu_binding(&pb); 4199 pmc_select_cpu(cpu); 4200 4201 npmc = md->pmd_npmc; 4202 4203 pmcinfo_size = npmc * sizeof(struct pmc_info); 4204 pmcinfo = malloc(pmcinfo_size, M_PMC, M_WAITOK | M_ZERO); 4205 4206 p = pmcinfo; 4207 4208 for (n = 0; n < md->pmd_npmc; n++, p++) { 4209 4210 pcd = pmc_ri_to_classdep(md, n, &ari); 4211 4212 KASSERT(pcd != NULL, 4213 ("[pmc,%d] null pcd ri=%d", __LINE__, n)); 4214 4215 if ((error = pcd->pcd_describe(cpu, ari, p, &pm)) != 0) 4216 break; 4217 4218 if (PMC_ROW_DISP_IS_STANDALONE(n)) 4219 p->pm_rowdisp = PMC_DISP_STANDALONE; 4220 else if (PMC_ROW_DISP_IS_THREAD(n)) 4221 p->pm_rowdisp = PMC_DISP_THREAD; 4222 else 4223 p->pm_rowdisp = PMC_DISP_FREE; 4224 4225 p->pm_ownerpid = -1; 4226 4227 if (pm == NULL) /* no PMC associated */ 4228 continue; 4229 4230 po = pm->pm_owner; 4231 4232 KASSERT(po->po_owner != NULL, 4233 ("[pmc,%d] pmc_owner had a null proc pointer", 4234 __LINE__)); 4235 4236 p->pm_ownerpid = po->po_owner->p_pid; 4237 p->pm_mode = PMC_TO_MODE(pm); 4238 p->pm_event = pm->pm_event; 4239 p->pm_flags = pm->pm_flags; 4240 4241 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) 4242 p->pm_reloadcount = 4243 pm->pm_sc.pm_reloadcount; 4244 } 4245 4246 pmc_restore_cpu_binding(&pb); 4247 4248 /* now copy out the PMC info collected */ 4249 if (error == 0) 4250 error = copyout(pmcinfo, &gpi->pm_pmcs, pmcinfo_size); 4251 4252 free(pmcinfo, M_PMC); 4253 } 4254 break; 4255 4256 4257 /* 4258 * Set the administrative state of a PMC. I.e. whether 4259 * the PMC is to be used or not. 4260 */ 4261 4262 case PMC_OP_PMCADMIN: 4263 { 4264 int cpu, ri; 4265 enum pmc_state request; 4266 struct pmc_cpu *pc; 4267 struct pmc_hw *phw; 4268 struct pmc_op_pmcadmin pma; 4269 struct pmc_binding pb; 4270 4271 sx_assert(&pmc_sx, SX_XLOCKED); 4272 4273 KASSERT(td == curthread, 4274 ("[pmc,%d] td != curthread", __LINE__)); 4275 4276 error = priv_check(td, PRIV_PMC_MANAGE); 4277 if (error) 4278 break; 4279 4280 if ((error = copyin(arg, &pma, sizeof(pma))) != 0) 4281 break; 4282 4283 cpu = pma.pm_cpu; 4284 4285 if (cpu < 0 || cpu >= (int) pmc_cpu_max()) { 4286 error = EINVAL; 4287 break; 4288 } 4289 4290 if (!pmc_cpu_is_active(cpu)) { 4291 error = ENXIO; 4292 break; 4293 } 4294 4295 request = pma.pm_state; 4296 4297 if (request != PMC_STATE_DISABLED && 4298 request != PMC_STATE_FREE) { 4299 error = EINVAL; 4300 break; 4301 } 4302 4303 ri = pma.pm_pmc; /* pmc id == row index */ 4304 if (ri < 0 || ri >= (int) md->pmd_npmc) { 4305 error = EINVAL; 4306 break; 4307 } 4308 4309 /* 4310 * We can't disable a PMC with a row-index allocated 4311 * for process virtual PMCs. 4312 */ 4313 4314 if (PMC_ROW_DISP_IS_THREAD(ri) && 4315 request == PMC_STATE_DISABLED) { 4316 error = EBUSY; 4317 break; 4318 } 4319 4320 /* 4321 * otherwise, this PMC on this CPU is either free or 4322 * in system-wide mode. 4323 */ 4324 4325 pmc_save_cpu_binding(&pb); 4326 pmc_select_cpu(cpu); 4327 4328 pc = pmc_pcpu[cpu]; 4329 phw = pc->pc_hwpmcs[ri]; 4330 4331 /* 4332 * XXX do we need some kind of 'forced' disable? 4333 */ 4334 4335 if (phw->phw_pmc == NULL) { 4336 if (request == PMC_STATE_DISABLED && 4337 (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED)) { 4338 phw->phw_state &= ~PMC_PHW_FLAG_IS_ENABLED; 4339 PMC_MARK_ROW_STANDALONE(ri); 4340 } else if (request == PMC_STATE_FREE && 4341 (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0) { 4342 phw->phw_state |= PMC_PHW_FLAG_IS_ENABLED; 4343 PMC_UNMARK_ROW_STANDALONE(ri); 4344 } 4345 /* other cases are a no-op */ 4346 } else 4347 error = EBUSY; 4348 4349 pmc_restore_cpu_binding(&pb); 4350 } 4351 break; 4352 4353 4354 /* 4355 * Allocate a PMC. 4356 */ 4357 case PMC_OP_PMCALLOCATE: 4358 { 4359 struct pmc_op_pmcallocate pa; 4360 4361 error = copyin(arg, &pa, sizeof(pa)); 4362 if (error != 0) 4363 break; 4364 4365 error = pmc_do_op_pmcallocate(td, &pa); 4366 if (error != 0) 4367 break; 4368 4369 error = copyout(&pa, arg, sizeof(pa)); 4370 } 4371 break; 4372 4373 /* 4374 * Attach a PMC to a process. 4375 */ 4376 case PMC_OP_PMCATTACH: 4377 { 4378 struct pmc_op_pmcattach a; 4379 4380 error = copyin(arg, &a, sizeof(a)); 4381 if (error != 0) 4382 break; 4383 4384 error = pmc_do_op_pmcattach(td, a); 4385 } 4386 break; 4387 4388 /* 4389 * Detach an attached PMC from a process. 4390 */ 4391 case PMC_OP_PMCDETACH: 4392 { 4393 struct pmc_op_pmcattach a; 4394 4395 error = copyin(arg, &a, sizeof(a)); 4396 if (error != 0) 4397 break; 4398 4399 error = pmc_do_op_pmcdetach(td, a); 4400 } 4401 break; 4402 4403 4404 /* 4405 * Retrieve the MSR number associated with the counter 4406 * 'pmc_id'. This allows processes to directly use RDPMC 4407 * instructions to read their PMCs, without the overhead of a 4408 * system call. 4409 */ 4410 4411 case PMC_OP_PMCGETMSR: 4412 { 4413 int adjri, ri; 4414 struct pmc *pm; 4415 struct pmc_target *pt; 4416 struct pmc_op_getmsr gm; 4417 struct pmc_classdep *pcd; 4418 4419 PMC_DOWNGRADE_SX(); 4420 4421 if ((error = copyin(arg, &gm, sizeof(gm))) != 0) 4422 break; 4423 4424 if ((error = pmc_find_pmc(gm.pm_pmcid, &pm)) != 0) 4425 break; 4426 4427 /* 4428 * The allocated PMC has to be a process virtual PMC, 4429 * i.e., of type MODE_T[CS]. Global PMCs can only be 4430 * read using the PMCREAD operation since they may be 4431 * allocated on a different CPU than the one we could 4432 * be running on at the time of the RDPMC instruction. 4433 * 4434 * The GETMSR operation is not allowed for PMCs that 4435 * are inherited across processes. 4436 */ 4437 4438 if (!PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)) || 4439 (pm->pm_flags & PMC_F_DESCENDANTS)) { 4440 error = EINVAL; 4441 break; 4442 } 4443 4444 /* 4445 * It only makes sense to use a RDPMC (or its 4446 * equivalent instruction on non-x86 architectures) on 4447 * a process that has allocated and attached a PMC to 4448 * itself. Conversely the PMC is only allowed to have 4449 * one process attached to it -- its owner. 4450 */ 4451 4452 if ((pt = LIST_FIRST(&pm->pm_targets)) == NULL || 4453 LIST_NEXT(pt, pt_next) != NULL || 4454 pt->pt_process->pp_proc != pm->pm_owner->po_owner) { 4455 error = EINVAL; 4456 break; 4457 } 4458 4459 ri = PMC_TO_ROWINDEX(pm); 4460 pcd = pmc_ri_to_classdep(md, ri, &adjri); 4461 4462 /* PMC class has no 'GETMSR' support */ 4463 if (pcd->pcd_get_msr == NULL) { 4464 error = ENOSYS; 4465 break; 4466 } 4467 4468 if ((error = (*pcd->pcd_get_msr)(adjri, &gm.pm_msr)) != 0) 4469 break; 4470 4471 if ((error = copyout(&gm, arg, sizeof(gm))) < 0) 4472 break; 4473 4474 /* 4475 * Mark our process as using MSRs. Update machine 4476 * state using a forced context switch. 4477 */ 4478 4479 pt->pt_process->pp_flags |= PMC_PP_ENABLE_MSR_ACCESS; 4480 pmc_force_context_switch(); 4481 4482 } 4483 break; 4484 4485 /* 4486 * Release an allocated PMC. 4487 */ 4488 case PMC_OP_PMCRELEASE: 4489 { 4490 struct pmc_op_simple sp; 4491 4492 error = copyin(arg, &sp, sizeof(sp)); 4493 if (error != 0) 4494 break; 4495 4496 error = pmc_do_op_pmcrelease(sp.pm_pmcid); 4497 } 4498 break; 4499 4500 /* 4501 * Read and/or write a PMC. 4502 */ 4503 case PMC_OP_PMCRW: 4504 { 4505 struct pmc_op_pmcrw prw; 4506 struct pmc_op_pmcrw *pprw; 4507 pmc_value_t oldvalue; 4508 4509 PMC_DOWNGRADE_SX(); 4510 4511 error = copyin(arg, &prw, sizeof(prw)); 4512 if (error != 0) 4513 break; 4514 4515 error = pmc_do_op_pmcrw(&prw, &oldvalue); 4516 if (error != 0) 4517 break; 4518 4519 /* Return old value if requested. */ 4520 if ((prw.pm_flags & PMC_F_OLDVALUE) != 0) { 4521 pprw = arg; 4522 error = copyout(&oldvalue, &pprw->pm_value, 4523 sizeof(prw.pm_value)); 4524 } 4525 } 4526 break; 4527 4528 4529 /* 4530 * Set the sampling rate for a sampling mode PMC and the 4531 * initial count for a counting mode PMC. 4532 */ 4533 4534 case PMC_OP_PMCSETCOUNT: 4535 { 4536 struct pmc *pm; 4537 struct pmc_op_pmcsetcount sc; 4538 4539 PMC_DOWNGRADE_SX(); 4540 4541 if ((error = copyin(arg, &sc, sizeof(sc))) != 0) 4542 break; 4543 4544 if ((error = pmc_find_pmc(sc.pm_pmcid, &pm)) != 0) 4545 break; 4546 4547 if (pm->pm_state == PMC_STATE_RUNNING) { 4548 error = EBUSY; 4549 break; 4550 } 4551 4552 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) { 4553 /* 4554 * Don't permit requested sample rate to be 4555 * less than pmc_mincount. 4556 */ 4557 if (sc.pm_count < MAX(1, pmc_mincount)) 4558 log(LOG_WARNING, "pmcsetcount: passed sample " 4559 "rate %ju - setting to %u\n", 4560 (uintmax_t)sc.pm_count, 4561 MAX(1, pmc_mincount)); 4562 pm->pm_sc.pm_reloadcount = MAX(MAX(1, pmc_mincount), 4563 sc.pm_count); 4564 } else 4565 pm->pm_sc.pm_initial = sc.pm_count; 4566 } 4567 break; 4568 4569 4570 /* 4571 * Start a PMC. 4572 */ 4573 4574 case PMC_OP_PMCSTART: 4575 { 4576 pmc_id_t pmcid; 4577 struct pmc *pm; 4578 struct pmc_op_simple sp; 4579 4580 sx_assert(&pmc_sx, SX_XLOCKED); 4581 4582 if ((error = copyin(arg, &sp, sizeof(sp))) != 0) 4583 break; 4584 4585 pmcid = sp.pm_pmcid; 4586 4587 if ((error = pmc_find_pmc(pmcid, &pm)) != 0) 4588 break; 4589 4590 KASSERT(pmcid == pm->pm_id, 4591 ("[pmc,%d] pmcid %x != id %x", __LINE__, 4592 pm->pm_id, pmcid)); 4593 4594 if (pm->pm_state == PMC_STATE_RUNNING) /* already running */ 4595 break; 4596 else if (pm->pm_state != PMC_STATE_STOPPED && 4597 pm->pm_state != PMC_STATE_ALLOCATED) { 4598 error = EINVAL; 4599 break; 4600 } 4601 4602 error = pmc_start(pm); 4603 } 4604 break; 4605 4606 4607 /* 4608 * Stop a PMC. 4609 */ 4610 4611 case PMC_OP_PMCSTOP: 4612 { 4613 pmc_id_t pmcid; 4614 struct pmc *pm; 4615 struct pmc_op_simple sp; 4616 4617 PMC_DOWNGRADE_SX(); 4618 4619 if ((error = copyin(arg, &sp, sizeof(sp))) != 0) 4620 break; 4621 4622 pmcid = sp.pm_pmcid; 4623 4624 /* 4625 * Mark the PMC as inactive and invoke the MD stop 4626 * routines if needed. 4627 */ 4628 4629 if ((error = pmc_find_pmc(pmcid, &pm)) != 0) 4630 break; 4631 4632 KASSERT(pmcid == pm->pm_id, 4633 ("[pmc,%d] pmc id %x != pmcid %x", __LINE__, 4634 pm->pm_id, pmcid)); 4635 4636 if (pm->pm_state == PMC_STATE_STOPPED) /* already stopped */ 4637 break; 4638 else if (pm->pm_state != PMC_STATE_RUNNING) { 4639 error = EINVAL; 4640 break; 4641 } 4642 4643 error = pmc_stop(pm); 4644 } 4645 break; 4646 4647 4648 /* 4649 * Write a user supplied value to the log file. 4650 */ 4651 4652 case PMC_OP_WRITELOG: 4653 { 4654 struct pmc_op_writelog wl; 4655 struct pmc_owner *po; 4656 4657 PMC_DOWNGRADE_SX(); 4658 4659 if ((error = copyin(arg, &wl, sizeof(wl))) != 0) 4660 break; 4661 4662 if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) { 4663 error = EINVAL; 4664 break; 4665 } 4666 4667 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) { 4668 error = EINVAL; 4669 break; 4670 } 4671 4672 error = pmclog_process_userlog(po, &wl); 4673 } 4674 break; 4675 4676 /* 4677 * Get the PMC capabilities 4678 */ 4679 4680 case PMC_OP_GETCAPS: 4681 { 4682 struct pmc_op_caps c; 4683 struct pmc *pm; 4684 struct pmc_classdep *pcd; 4685 pmc_id_t pmcid; 4686 int adjri, ri; 4687 4688 PMC_DOWNGRADE_SX(); 4689 4690 if ((error = copyin(arg, &c, sizeof(c))) != 0) 4691 break; 4692 4693 pmcid = c.pm_pmcid; 4694 4695 if ((error = pmc_find_pmc(pmcid, &pm)) != 0) 4696 break; 4697 4698 KASSERT(pmcid == pm->pm_id, 4699 ("[pmc,%d] pmc id %x != pmcid %x", __LINE__, 4700 pm->pm_id, pmcid)); 4701 4702 ri = PMC_TO_ROWINDEX(pm); 4703 pcd = pmc_ri_to_classdep(md, ri, &adjri); 4704 4705 /* 4706 * If PMC class has no GETCAPS return the class capabilities 4707 * otherwise get the per counter capabilities. 4708 */ 4709 if (pcd->pcd_get_caps == NULL) { 4710 c.pm_caps = pcd->pcd_caps; 4711 } else { 4712 error = (*pcd->pcd_get_caps)(adjri, &c.pm_caps); 4713 if (error < 0) 4714 break; 4715 } 4716 4717 if ((error = copyout(&c, arg, sizeof(c))) < 0) 4718 break; 4719 } 4720 break; 4721 4722 default: 4723 error = EINVAL; 4724 break; 4725 } 4726 4727 if (is_sx_downgraded) 4728 sx_sunlock(&pmc_sx); 4729 else 4730 sx_xunlock(&pmc_sx); 4731 done_syscall: 4732 if (error) 4733 counter_u64_add(pmc_stats.pm_syscall_errors, 1); 4734 4735 return (error); 4736 } 4737 4738 /* 4739 * Helper functions 4740 */ 4741 4742 /* 4743 * Mark the thread as needing callchain capture and post an AST. The 4744 * actual callchain capture will be done in a context where it is safe 4745 * to take page faults. 4746 */ 4747 static void 4748 pmc_post_callchain_callback(void) 4749 { 4750 struct thread *td; 4751 4752 td = curthread; 4753 4754 /* 4755 * If there is multiple PMCs for the same interrupt ignore new post 4756 */ 4757 if ((td->td_pflags & TDP_CALLCHAIN) != 0) 4758 return; 4759 4760 /* 4761 * Mark this thread as needing callchain capture. 4762 * `td->td_pflags' will be safe to touch because this thread 4763 * was in user space when it was interrupted. 4764 */ 4765 td->td_pflags |= TDP_CALLCHAIN; 4766 4767 /* 4768 * Don't let this thread migrate between CPUs until callchain 4769 * capture completes. 4770 */ 4771 sched_pin(); 4772 4773 return; 4774 } 4775 4776 static void 4777 pmc_multipart_add(struct pmc_sample *ps, int type, int length) 4778 { 4779 int i; 4780 uint8_t *hdr; 4781 4782 MPASS(ps->ps_pc != NULL); 4783 MPASS(ps->ps_nsamples_actual != 0); 4784 4785 hdr = (uint8_t *)ps->ps_pc; 4786 4787 for (i = 0; i < PMC_MULTIPART_HEADER_ENTRIES; i++) { 4788 if (hdr[2 * i] == PMC_CC_MULTIPART_NONE) { 4789 hdr[2 * i] = type; 4790 hdr[2 * i + 1] = length; 4791 ps->ps_nsamples_actual += length; 4792 return; 4793 } 4794 } 4795 4796 KASSERT(false, ("Too many parts in the multipart header!")); 4797 } 4798 4799 static void 4800 pmc_multipart_copydata(struct pmc_sample *ps, struct pmc_multipart *mp) 4801 { 4802 int i, scale; 4803 uint64_t *ps_pc; 4804 4805 MPASS(ps->ps_pc != NULL); 4806 MPASS(ps->ps_nsamples_actual != 0); 4807 4808 ps_pc = (uint64_t *)ps->ps_pc; 4809 4810 for (i = 0; i < mp->pl_length; i++) 4811 ps_pc[i + 1] = mp->pl_mpdata[i]; 4812 4813 scale = sizeof(uint64_t) / sizeof(uintptr_t); 4814 pmc_multipart_add(ps, mp->pl_type, scale * mp->pl_length); 4815 } 4816 4817 /* 4818 * Find a free slot in the per-cpu array of samples and capture the 4819 * current callchain there. If a sample was successfully added, a bit 4820 * is set in mask 'pmc_cpumask' denoting that the DO_SAMPLES hook 4821 * needs to be invoked from the clock handler. 4822 * 4823 * This function is meant to be called from an NMI handler. It cannot 4824 * use any of the locking primitives supplied by the OS. 4825 */ 4826 static int 4827 pmc_add_sample(ring_type_t ring, struct pmc *pm, struct trapframe *tf, 4828 struct pmc_multipart *mp) 4829 { 4830 struct pmc_sample *ps; 4831 struct pmc_samplebuffer *psb; 4832 struct thread *td; 4833 int error, cpu, callchaindepth; 4834 bool inuserspace; 4835 4836 error = 0; 4837 4838 /* 4839 * Allocate space for a sample buffer. 4840 */ 4841 cpu = curcpu; 4842 psb = pmc_pcpu[cpu]->pc_sb[ring]; 4843 inuserspace = TRAPF_USERMODE(tf); 4844 ps = PMC_PROD_SAMPLE(psb); 4845 if (psb->ps_considx != psb->ps_prodidx && 4846 ps->ps_nsamples) { /* in use, reader hasn't caught up */ 4847 pm->pm_pcpu_state[cpu].pps_stalled = 1; 4848 counter_u64_add(pmc_stats.pm_intr_bufferfull, 1); 4849 PMCDBG6(SAM,INT,1,"(spc) cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d", 4850 cpu, pm, tf, inuserspace, 4851 (int)(psb->ps_prodidx & pmc_sample_mask), 4852 (int)(psb->ps_considx & pmc_sample_mask)); 4853 callchaindepth = 1; 4854 error = ENOMEM; 4855 goto done; 4856 } 4857 4858 /* Fill in entry. */ 4859 PMCDBG6(SAM,INT,1,"cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d", cpu, pm, tf, 4860 inuserspace, (int)(psb->ps_prodidx & pmc_sample_mask), 4861 (int)(psb->ps_considx & pmc_sample_mask)); 4862 4863 td = curthread; 4864 ps->ps_pmc = pm; 4865 ps->ps_td = td; 4866 ps->ps_pid = td->td_proc->p_pid; 4867 ps->ps_tid = td->td_tid; 4868 ps->ps_tsc = pmc_rdtsc(); 4869 ps->ps_ticks = ticks; 4870 ps->ps_cpu = cpu; 4871 ps->ps_flags = inuserspace ? PMC_CC_F_USERSPACE : 0; 4872 ps->ps_nsamples_actual = 0; 4873 4874 callchaindepth = (pm->pm_flags & PMC_F_CALLCHAIN) ? 4875 pmc_callchaindepth : 1; 4876 4877 MPASS(ps->ps_pc != NULL); 4878 4879 if (mp != NULL) { 4880 /* Set multipart flag, clear header and copy data */ 4881 ps->ps_flags |= PMC_CC_F_MULTIPART; 4882 ps->ps_pc[0] = 0; 4883 ps->ps_nsamples_actual = 1; 4884 pmc_multipart_copydata(ps, mp); 4885 } 4886 4887 if (callchaindepth == 1) { 4888 ps->ps_pc[ps->ps_nsamples_actual] = PMC_TRAPFRAME_TO_PC(tf); 4889 } else { 4890 /* 4891 * Kernel stack traversals can be done immediately, while we 4892 * defer to an AST for user space traversals. 4893 */ 4894 if (!inuserspace) { 4895 callchaindepth = pmc_save_kernel_callchain( 4896 ps->ps_pc + ps->ps_nsamples_actual, 4897 callchaindepth - ps->ps_nsamples_actual, tf); 4898 callchaindepth += ps->ps_nsamples_actual; 4899 } else { 4900 pmc_post_callchain_callback(); 4901 callchaindepth = PMC_USER_CALLCHAIN_PENDING; 4902 } 4903 } 4904 4905 ps->ps_nsamples = callchaindepth; /* mark entry as in-use */ 4906 if (ring == PMC_UR) { 4907 ps->ps_nsamples_actual = ps->ps_nsamples; 4908 ps->ps_nsamples = PMC_USER_CALLCHAIN_PENDING; 4909 } 4910 4911 KASSERT(counter_u64_fetch(pm->pm_runcount) >= 0, 4912 ("[pmc,%d] pm=%p runcount %ju", __LINE__, pm, 4913 (uintmax_t)counter_u64_fetch(pm->pm_runcount))); 4914 4915 counter_u64_add(pm->pm_runcount, 1); /* hold onto PMC */ 4916 /* increment write pointer */ 4917 psb->ps_prodidx++; 4918 done: 4919 /* mark CPU as needing processing */ 4920 if (callchaindepth != PMC_USER_CALLCHAIN_PENDING) 4921 DPCPU_SET(pmc_sampled, 1); 4922 4923 return (error); 4924 } 4925 4926 /* 4927 * Interrupt processing. 4928 * 4929 * This function may be called from an NMI handler. It cannot use any of the 4930 * locking primitives supplied by the OS. 4931 */ 4932 int 4933 pmc_process_interrupt_mp(int ring, struct pmc *pm, struct trapframe *tf, 4934 struct pmc_multipart *mp) 4935 { 4936 struct thread *td; 4937 4938 td = curthread; 4939 if ((pm->pm_flags & PMC_F_USERCALLCHAIN) && 4940 (td->td_proc->p_flag & P_KPROC) == 0 && !TRAPF_USERMODE(tf)) { 4941 atomic_add_int(&td->td_pmcpend, 1); 4942 return (pmc_add_sample(PMC_UR, pm, tf, mp)); 4943 } 4944 return (pmc_add_sample(ring, pm, tf, mp)); 4945 } 4946 4947 int 4948 pmc_process_interrupt(int ring, struct pmc *pm, struct trapframe *tf) 4949 { 4950 return (pmc_process_interrupt_mp(ring, pm, tf, NULL)); 4951 } 4952 4953 /* 4954 * Capture a user call chain. This function will be called from ast() 4955 * before control returns to userland and before the process gets 4956 * rescheduled. 4957 */ 4958 static void 4959 pmc_capture_user_callchain(int cpu, int ring, struct trapframe *tf) 4960 { 4961 struct pmc *pm; 4962 struct pmc_sample *ps; 4963 struct pmc_samplebuffer *psb; 4964 struct thread *td; 4965 uint64_t considx, prodidx; 4966 int nsamples, nrecords, pass, iter; 4967 int start_ticks __diagused; 4968 4969 psb = pmc_pcpu[cpu]->pc_sb[ring]; 4970 td = curthread; 4971 nrecords = INT_MAX; 4972 pass = 0; 4973 start_ticks = ticks; 4974 4975 KASSERT(ring == PMC_UR || (td->td_pflags & TDP_CALLCHAIN) != 0, 4976 ("[pmc,%d] Retrieving callchain for thread that doesn't want it", 4977 __LINE__)); 4978 restart: 4979 if (ring == PMC_UR) 4980 nrecords = atomic_readandclear_32(&td->td_pmcpend); 4981 4982 for (iter = 0, considx = psb->ps_considx, prodidx = psb->ps_prodidx; 4983 considx < prodidx && iter < pmc_nsamples; considx++, iter++) { 4984 ps = PMC_CONS_SAMPLE_OFF(psb, considx); 4985 4986 /* 4987 * Iterate through all deferred callchain requests. Walk from 4988 * the current read pointer to the current write pointer. 4989 */ 4990 #ifdef INVARIANTS 4991 if (ps->ps_nsamples == PMC_SAMPLE_FREE) { 4992 continue; 4993 } 4994 #endif 4995 if (ps->ps_td != td || 4996 ps->ps_nsamples != PMC_USER_CALLCHAIN_PENDING || 4997 ps->ps_pmc->pm_state != PMC_STATE_RUNNING) 4998 continue; 4999 5000 KASSERT(ps->ps_cpu == cpu, 5001 ("[pmc,%d] cpu mismatch ps_cpu=%d pcpu=%d", __LINE__, 5002 ps->ps_cpu, PCPU_GET(cpuid))); 5003 5004 pm = ps->ps_pmc; 5005 KASSERT(pm->pm_flags & PMC_F_CALLCHAIN, 5006 ("[pmc,%d] Retrieving callchain for PMC that doesn't " 5007 "want it", __LINE__)); 5008 5009 if (ring == PMC_UR) { 5010 counter_u64_add(pmc_stats.pm_merges, 1); 5011 } 5012 nsamples = ps->ps_nsamples_actual; 5013 5014 /* 5015 * Retrieve the callchain and mark the sample buffer 5016 * as 'processable' by the timer tick sweep code. 5017 */ 5018 if (__predict_true(nsamples < pmc_callchaindepth - 1)) 5019 nsamples += pmc_save_user_callchain(ps->ps_pc + nsamples, 5020 pmc_callchaindepth - nsamples - 1, tf); 5021 5022 /* 5023 * We have to prevent hardclock from potentially overwriting 5024 * this sample between when we read the value and when we set 5025 * it. 5026 */ 5027 spinlock_enter(); 5028 5029 /* 5030 * Verify that the sample hasn't been dropped in the meantime. 5031 */ 5032 if (ps->ps_nsamples == PMC_USER_CALLCHAIN_PENDING) { 5033 KASSERT(counter_u64_fetch(pm->pm_runcount) > 0, 5034 ("[pmc,%d] runcount %ju", __LINE__, 5035 (uintmax_t)counter_u64_fetch(pm->pm_runcount))); 5036 5037 ps->ps_nsamples = nsamples; 5038 /* 5039 * If we couldn't get a sample, simply drop the 5040 * reference. 5041 */ 5042 if (nsamples == 0) 5043 counter_u64_add(pm->pm_runcount, -1); 5044 } 5045 spinlock_exit(); 5046 if (nrecords-- == 1) 5047 break; 5048 } 5049 if (__predict_false(ring == PMC_UR && td->td_pmcpend)) { 5050 if (pass == 0) { 5051 pass = 1; 5052 goto restart; 5053 } 5054 /* only collect samples for this part once */ 5055 td->td_pmcpend = 0; 5056 } 5057 5058 #ifdef INVARIANTS 5059 if ((ticks - start_ticks) > hz) 5060 log(LOG_ERR, "%s took %d ticks\n", __func__, (ticks - start_ticks)); 5061 #endif 5062 /* mark CPU as needing processing */ 5063 DPCPU_SET(pmc_sampled, 1); 5064 } 5065 5066 /* 5067 * Process saved PC samples. 5068 */ 5069 static void 5070 pmc_process_samples(int cpu, ring_type_t ring) 5071 { 5072 struct pmc *pm; 5073 struct thread *td; 5074 struct pmc_owner *po; 5075 struct pmc_sample *ps; 5076 struct pmc_classdep *pcd; 5077 struct pmc_samplebuffer *psb; 5078 uint64_t delta __diagused; 5079 int adjri, n; 5080 5081 KASSERT(PCPU_GET(cpuid) == cpu, 5082 ("[pmc,%d] not on the correct CPU pcpu=%d cpu=%d", __LINE__, 5083 PCPU_GET(cpuid), cpu)); 5084 5085 psb = pmc_pcpu[cpu]->pc_sb[ring]; 5086 delta = psb->ps_prodidx - psb->ps_considx; 5087 MPASS(delta <= pmc_nsamples); 5088 MPASS(psb->ps_considx <= psb->ps_prodidx); 5089 for (n = 0; psb->ps_considx < psb->ps_prodidx; psb->ps_considx++, n++) { 5090 ps = PMC_CONS_SAMPLE(psb); 5091 5092 if (__predict_false(ps->ps_nsamples == PMC_SAMPLE_FREE)) 5093 continue; 5094 5095 /* skip non-running samples */ 5096 pm = ps->ps_pmc; 5097 if (pm->pm_state != PMC_STATE_RUNNING) 5098 goto entrydone; 5099 5100 KASSERT(counter_u64_fetch(pm->pm_runcount) > 0, 5101 ("[pmc,%d] pm=%p runcount %ju", __LINE__, pm, 5102 (uintmax_t)counter_u64_fetch(pm->pm_runcount))); 5103 KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)), 5104 ("[pmc,%d] pmc=%p non-sampling mode=%d", __LINE__, 5105 pm, PMC_TO_MODE(pm))); 5106 5107 po = pm->pm_owner; 5108 5109 /* If there is a pending AST wait for completion */ 5110 if (ps->ps_nsamples == PMC_USER_CALLCHAIN_PENDING) { 5111 /* 5112 * If we've been waiting more than 1 tick to 5113 * collect a callchain for this record then 5114 * drop it and move on. 5115 */ 5116 if (ticks - ps->ps_ticks > 1) { 5117 /* 5118 * Track how often we hit this as it will 5119 * preferentially lose user samples 5120 * for long running system calls. 5121 */ 5122 counter_u64_add(pmc_stats.pm_overwrites, 1); 5123 goto entrydone; 5124 } 5125 /* Need a rescan at a later time. */ 5126 DPCPU_SET(pmc_sampled, 1); 5127 break; 5128 } 5129 5130 PMCDBG6(SAM,OPS,1,"cpu=%d pm=%p n=%d fl=%x wr=%d rd=%d", cpu, 5131 pm, ps->ps_nsamples, ps->ps_flags, 5132 (int)(psb->ps_prodidx & pmc_sample_mask), 5133 (int)(psb->ps_considx & pmc_sample_mask)); 5134 5135 /* 5136 * If this is a process-mode PMC that is attached to 5137 * its owner, and if the PC is in user mode, update 5138 * profiling statistics like timer-based profiling 5139 * would have done. 5140 * 5141 * Otherwise, this is either a sampling-mode PMC that 5142 * is attached to a different process than its owner, 5143 * or a system-wide sampling PMC. Dispatch a log 5144 * entry to the PMC's owner process. 5145 */ 5146 if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) { 5147 if (ps->ps_flags & PMC_CC_F_USERSPACE) { 5148 td = FIRST_THREAD_IN_PROC(po->po_owner); 5149 addupc_intr(td, ps->ps_pc[0], 1); 5150 } 5151 } else 5152 pmclog_process_callchain(pm, ps); 5153 5154 entrydone: 5155 ps->ps_nsamples = 0; /* mark entry as free */ 5156 KASSERT(counter_u64_fetch(pm->pm_runcount) > 0, 5157 ("[pmc,%d] pm=%p runcount %ju", __LINE__, pm, 5158 (uintmax_t)counter_u64_fetch(pm->pm_runcount))); 5159 5160 counter_u64_add(pm->pm_runcount, -1); 5161 } 5162 5163 counter_u64_add(pmc_stats.pm_log_sweeps, 1); 5164 5165 /* Do not re-enable stalled PMCs if we failed to process any samples */ 5166 if (n == 0) 5167 return; 5168 5169 /* 5170 * Restart any stalled sampling PMCs on this CPU. 5171 * 5172 * If the NMI handler sets the pm_stalled field of a PMC after 5173 * the check below, we'll end up processing the stalled PMC at 5174 * the next hardclock tick. 5175 */ 5176 for (n = 0; n < md->pmd_npmc; n++) { 5177 pcd = pmc_ri_to_classdep(md, n, &adjri); 5178 KASSERT(pcd != NULL, 5179 ("[pmc,%d] null pcd ri=%d", __LINE__, n)); 5180 (void)(*pcd->pcd_get_config)(cpu, adjri, &pm); 5181 5182 if (pm == NULL || /* !cfg'ed */ 5183 pm->pm_state != PMC_STATE_RUNNING || /* !active */ 5184 !PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) || /* !sampling */ 5185 !pm->pm_pcpu_state[cpu].pps_cpustate || /* !desired */ 5186 !pm->pm_pcpu_state[cpu].pps_stalled) /* !stalled */ 5187 continue; 5188 5189 pm->pm_pcpu_state[cpu].pps_stalled = 0; 5190 (void)(*pcd->pcd_start_pmc)(cpu, adjri, pm); 5191 } 5192 } 5193 5194 /* 5195 * Event handlers. 5196 */ 5197 5198 /* 5199 * Handle a process exit. 5200 * 5201 * Remove this process from all hash tables. If this process 5202 * owned any PMCs, turn off those PMCs and deallocate them, 5203 * removing any associations with target processes. 5204 * 5205 * This function will be called by the last 'thread' of a 5206 * process. 5207 * 5208 * XXX This eventhandler gets called early in the exit process. 5209 * Consider using a 'hook' invocation from thread_exit() or equivalent 5210 * spot. Another negative is that kse_exit doesn't seem to call 5211 * exit1() [??]. 5212 */ 5213 static void 5214 pmc_process_exit(void *arg __unused, struct proc *p) 5215 { 5216 struct pmc *pm; 5217 struct pmc_owner *po; 5218 struct pmc_process *pp; 5219 struct pmc_classdep *pcd; 5220 pmc_value_t newvalue, tmp; 5221 int ri, adjri, cpu; 5222 bool is_using_hwpmcs; 5223 5224 PROC_LOCK(p); 5225 is_using_hwpmcs = (p->p_flag & P_HWPMC) != 0; 5226 PROC_UNLOCK(p); 5227 5228 /* 5229 * Log a sysexit event to all SS PMC owners. 5230 */ 5231 PMC_EPOCH_ENTER(); 5232 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) { 5233 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) != 0) 5234 pmclog_process_sysexit(po, p->p_pid); 5235 } 5236 PMC_EPOCH_EXIT(); 5237 5238 PMC_GET_SX_XLOCK(); 5239 PMCDBG3(PRC,EXT,1,"process-exit proc=%p (%d, %s)", p, p->p_pid, 5240 p->p_comm); 5241 5242 if (!is_using_hwpmcs) 5243 goto out; 5244 5245 /* 5246 * Since this code is invoked by the last thread in an exiting process, 5247 * we would have context switched IN at some prior point. However, with 5248 * PREEMPTION, kernel mode context switches may happen any time, so we 5249 * want to disable a context switch OUT till we get any PMCs targeting 5250 * this process off the hardware. 5251 * 5252 * We also need to atomically remove this process' entry from our 5253 * target process hash table, using PMC_FLAG_REMOVE. 5254 */ 5255 PMCDBG3(PRC,EXT,1, "process-exit proc=%p (%d, %s)", p, p->p_pid, 5256 p->p_comm); 5257 5258 critical_enter(); /* no preemption */ 5259 5260 cpu = curthread->td_oncpu; 5261 5262 pp = pmc_find_process_descriptor(p, PMC_FLAG_REMOVE); 5263 if (pp == NULL) { 5264 critical_exit(); 5265 goto out; 5266 } 5267 5268 PMCDBG2(PRC,EXT,2, "process-exit proc=%p pmc-process=%p", p, pp); 5269 5270 /* Run the context-switch-out steps for process exit. */ 5271 pmc_process_csw_stop_all(cpu); 5272 5273 /* 5274 * The exiting process could be the target of some PMCs which will be 5275 * running on currently executing CPU. 5276 * 5277 * We need to turn these PMCs off like we would do at context switch 5278 * OUT time. 5279 */ 5280 for (ri = 0; ri < md->pmd_npmc; ri++) { 5281 /* 5282 * Pick up the pmc pointer from hardware state similar to the 5283 * CSW_OUT code. 5284 */ 5285 pm = NULL; 5286 pcd = pmc_ri_to_classdep(md, ri, &adjri); 5287 5288 (void)(*pcd->pcd_get_config)(cpu, adjri, &pm); 5289 5290 PMCDBG2(PRC,EXT,2, "ri=%d pm=%p", ri, pm); 5291 5292 if (pm == NULL || !PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) 5293 continue; 5294 5295 PMCDBG4(PRC,EXT,2, "ppmcs[%d]=%p pm=%p state=%d", ri, 5296 pp->pp_pmcs[ri].pp_pmc, pm, pm->pm_state); 5297 5298 KASSERT(PMC_TO_ROWINDEX(pm) == ri, 5299 ("[pmc,%d] ri mismatch pmc(%d) ri(%d)", __LINE__, 5300 PMC_TO_ROWINDEX(pm), ri)); 5301 KASSERT(pm == pp->pp_pmcs[ri].pp_pmc, 5302 ("[pmc,%d] pm %p != pp_pmcs[%d] %p", __LINE__, pm, ri, 5303 pp->pp_pmcs[ri].pp_pmc)); 5304 KASSERT(counter_u64_fetch(pm->pm_runcount) > 0, 5305 ("[pmc,%d] bad runcount ri %d rc %ju", __LINE__, ri, 5306 (uintmax_t)counter_u64_fetch(pm->pm_runcount))); 5307 5308 /* 5309 * Change desired state, and then stop if not stalled. This 5310 * two-step dance should avoid race conditions where an 5311 * interrupt re-enables the PMC after this code has already 5312 * checked the pm_stalled flag. 5313 */ 5314 if (pm->pm_pcpu_state[cpu].pps_cpustate) { 5315 pm->pm_pcpu_state[cpu].pps_cpustate = 0; 5316 if (!pm->pm_pcpu_state[cpu].pps_stalled) { 5317 (void)pcd->pcd_stop_pmc(cpu, adjri, pm); 5318 5319 if (PMC_TO_MODE(pm) == PMC_MODE_TC) { 5320 pcd->pcd_read_pmc(cpu, adjri, pm, 5321 &newvalue); 5322 tmp = pmc_delta(pcd, newvalue, 5323 PMC_PCPU_SAVED(cpu, ri)); 5324 5325 mtx_pool_lock_spin(pmc_mtxpool, pm); 5326 pm->pm_gv.pm_savedvalue += tmp; 5327 pp->pp_pmcs[ri].pp_pmcval += tmp; 5328 mtx_pool_unlock_spin(pmc_mtxpool, pm); 5329 } 5330 } 5331 } 5332 5333 KASSERT(counter_u64_fetch(pm->pm_runcount) > 0, 5334 ("[pmc,%d] runcount is %d", __LINE__, ri)); 5335 5336 counter_u64_add(pm->pm_runcount, -1); 5337 (void)pcd->pcd_config_pmc(cpu, adjri, NULL); 5338 } 5339 5340 /* 5341 * Inform the MD layer of this pseudo "context switch out". 5342 */ 5343 (void)md->pmd_switch_out(pmc_pcpu[cpu], pp); 5344 5345 critical_exit(); /* ok to be pre-empted now */ 5346 5347 /* 5348 * Unlink this process from the PMCs that are targeting it. This will 5349 * send a signal to all PMC owner's whose PMCs are orphaned. 5350 * 5351 * Log PMC value at exit time if requested. 5352 */ 5353 for (ri = 0; ri < md->pmd_npmc; ri++) { 5354 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) { 5355 if ((pm->pm_flags & PMC_F_NEEDS_LOGFILE) != 0 && 5356 PMC_IS_COUNTING_MODE(PMC_TO_MODE(pm))) { 5357 pmclog_process_procexit(pm, pp); 5358 } 5359 pmc_unlink_target_process(pm, pp); 5360 } 5361 } 5362 free(pp, M_PMC); 5363 5364 out: 5365 /* 5366 * If the process owned PMCs, free them up and free up memory. 5367 */ 5368 if ((po = pmc_find_owner_descriptor(p)) != NULL) { 5369 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) != 0) 5370 pmclog_close(po); 5371 pmc_remove_owner(po); 5372 pmc_destroy_owner_descriptor(po); 5373 } 5374 5375 sx_xunlock(&pmc_sx); 5376 } 5377 5378 /* 5379 * Handle a process fork. 5380 * 5381 * If the parent process 'p1' is under HWPMC monitoring, then copy 5382 * over any attached PMCs that have 'do_descendants' semantics. 5383 */ 5384 static void 5385 pmc_process_fork(void *arg __unused, struct proc *p1, struct proc *newproc, 5386 int flags __unused) 5387 { 5388 struct pmc *pm; 5389 struct pmc_owner *po; 5390 struct pmc_process *ppnew, *ppold; 5391 unsigned int ri; 5392 bool is_using_hwpmcs, do_descendants; 5393 5394 PROC_LOCK(p1); 5395 is_using_hwpmcs = (p1->p_flag & P_HWPMC) != 0; 5396 PROC_UNLOCK(p1); 5397 5398 /* 5399 * If there are system-wide sampling PMCs active, we need to 5400 * log all fork events to their owner's logs. 5401 */ 5402 PMC_EPOCH_ENTER(); 5403 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) { 5404 if (po->po_flags & PMC_PO_OWNS_LOGFILE) { 5405 pmclog_process_procfork(po, p1->p_pid, newproc->p_pid); 5406 pmclog_process_proccreate(po, newproc, 1); 5407 } 5408 } 5409 PMC_EPOCH_EXIT(); 5410 5411 if (!is_using_hwpmcs) 5412 return; 5413 5414 PMC_GET_SX_XLOCK(); 5415 PMCDBG4(PMC,FRK,1, "process-fork proc=%p (%d, %s) -> %p", p1, 5416 p1->p_pid, p1->p_comm, newproc); 5417 5418 /* 5419 * If the parent process (curthread->td_proc) is a 5420 * target of any PMCs, look for PMCs that are to be 5421 * inherited, and link these into the new process 5422 * descriptor. 5423 */ 5424 ppold = pmc_find_process_descriptor(curthread->td_proc, PMC_FLAG_NONE); 5425 if (ppold == NULL) 5426 goto done; /* nothing to do */ 5427 5428 do_descendants = false; 5429 for (ri = 0; ri < md->pmd_npmc; ri++) { 5430 if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL && 5431 (pm->pm_flags & PMC_F_DESCENDANTS) != 0) { 5432 do_descendants = true; 5433 break; 5434 } 5435 } 5436 if (!do_descendants) /* nothing to do */ 5437 goto done; 5438 5439 /* 5440 * Now mark the new process as being tracked by this driver. 5441 */ 5442 PROC_LOCK(newproc); 5443 newproc->p_flag |= P_HWPMC; 5444 PROC_UNLOCK(newproc); 5445 5446 /* Allocate a descriptor for the new process. */ 5447 ppnew = pmc_find_process_descriptor(newproc, PMC_FLAG_ALLOCATE); 5448 if (ppnew == NULL) 5449 goto done; 5450 5451 /* 5452 * Run through all PMCs that were targeting the old process 5453 * and which specified F_DESCENDANTS and attach them to the 5454 * new process. 5455 * 5456 * Log the fork event to all owners of PMCs attached to this 5457 * process, if not already logged. 5458 */ 5459 for (ri = 0; ri < md->pmd_npmc; ri++) { 5460 if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL && 5461 (pm->pm_flags & PMC_F_DESCENDANTS) != 0) { 5462 pmc_link_target_process(pm, ppnew); 5463 po = pm->pm_owner; 5464 if (po->po_sscount == 0 && 5465 (po->po_flags & PMC_PO_OWNS_LOGFILE) != 0) { 5466 pmclog_process_procfork(po, p1->p_pid, 5467 newproc->p_pid); 5468 } 5469 } 5470 } 5471 5472 done: 5473 sx_xunlock(&pmc_sx); 5474 } 5475 5476 static void 5477 pmc_process_threadcreate(struct thread *td) 5478 { 5479 struct pmc_owner *po; 5480 5481 PMC_EPOCH_ENTER(); 5482 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) { 5483 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) != 0) 5484 pmclog_process_threadcreate(po, td, 1); 5485 } 5486 PMC_EPOCH_EXIT(); 5487 } 5488 5489 static void 5490 pmc_process_threadexit(struct thread *td) 5491 { 5492 struct pmc_owner *po; 5493 5494 PMC_EPOCH_ENTER(); 5495 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) { 5496 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) != 0) 5497 pmclog_process_threadexit(po, td); 5498 } 5499 PMC_EPOCH_EXIT(); 5500 } 5501 5502 static void 5503 pmc_process_proccreate(struct proc *p) 5504 { 5505 struct pmc_owner *po; 5506 5507 PMC_EPOCH_ENTER(); 5508 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) { 5509 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) != 0) 5510 pmclog_process_proccreate(po, p, 1 /* sync */); 5511 } 5512 PMC_EPOCH_EXIT(); 5513 } 5514 5515 static void 5516 pmc_process_allproc(struct pmc *pm) 5517 { 5518 struct pmc_owner *po; 5519 struct thread *td; 5520 struct proc *p; 5521 5522 po = pm->pm_owner; 5523 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) 5524 return; 5525 5526 sx_slock(&allproc_lock); 5527 FOREACH_PROC_IN_SYSTEM(p) { 5528 pmclog_process_proccreate(po, p, 0 /* sync */); 5529 PROC_LOCK(p); 5530 FOREACH_THREAD_IN_PROC(p, td) 5531 pmclog_process_threadcreate(po, td, 0 /* sync */); 5532 PROC_UNLOCK(p); 5533 } 5534 sx_sunlock(&allproc_lock); 5535 pmclog_flush(po, 0); 5536 } 5537 5538 static void 5539 pmc_kld_load(void *arg __unused, linker_file_t lf) 5540 { 5541 struct pmc_owner *po; 5542 5543 /* 5544 * Notify owners of system sampling PMCs about KLD operations. 5545 */ 5546 PMC_EPOCH_ENTER(); 5547 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) { 5548 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 5549 pmclog_process_map_in(po, (pid_t) -1, 5550 (uintfptr_t) lf->address, lf->pathname); 5551 } 5552 PMC_EPOCH_EXIT(); 5553 5554 /* 5555 * TODO: Notify owners of (all) process-sampling PMCs too. 5556 */ 5557 } 5558 5559 static void 5560 pmc_kld_unload(void *arg __unused, const char *filename __unused, 5561 caddr_t address, size_t size) 5562 { 5563 struct pmc_owner *po; 5564 5565 PMC_EPOCH_ENTER(); 5566 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) { 5567 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) != 0) { 5568 pmclog_process_map_out(po, (pid_t)-1, 5569 (uintfptr_t)address, (uintfptr_t)address + size); 5570 } 5571 } 5572 PMC_EPOCH_EXIT(); 5573 5574 /* 5575 * TODO: Notify owners of process-sampling PMCs. 5576 */ 5577 } 5578 5579 /* 5580 * initialization 5581 */ 5582 static const char * 5583 pmc_name_of_pmcclass(enum pmc_class class) 5584 { 5585 5586 switch (class) { 5587 #undef __PMC_CLASS 5588 #define __PMC_CLASS(S,V,D) \ 5589 case PMC_CLASS_##S: \ 5590 return #S; 5591 __PMC_CLASSES(); 5592 default: 5593 return ("<unknown>"); 5594 } 5595 } 5596 5597 /* 5598 * Base class initializer: allocate structure and set default classes. 5599 */ 5600 struct pmc_mdep * 5601 pmc_mdep_alloc(int nclasses) 5602 { 5603 struct pmc_mdep *md; 5604 int n; 5605 5606 /* SOFT + md classes */ 5607 n = 1 + nclasses; 5608 md = malloc(sizeof(struct pmc_mdep) + n * sizeof(struct pmc_classdep), 5609 M_PMC, M_WAITOK | M_ZERO); 5610 md->pmd_nclass = n; 5611 5612 /* Default methods */ 5613 md->pmd_switch_in = generic_switch_in; 5614 md->pmd_switch_out = generic_switch_out; 5615 5616 /* Add base class. */ 5617 pmc_soft_initialize(md); 5618 return (md); 5619 } 5620 5621 void 5622 pmc_mdep_free(struct pmc_mdep *md) 5623 { 5624 pmc_soft_finalize(md); 5625 free(md, M_PMC); 5626 } 5627 5628 static int 5629 generic_switch_in(struct pmc_cpu *pc __unused, struct pmc_process *pp __unused) 5630 { 5631 5632 return (0); 5633 } 5634 5635 static int 5636 generic_switch_out(struct pmc_cpu *pc __unused, struct pmc_process *pp __unused) 5637 { 5638 5639 return (0); 5640 } 5641 5642 static struct pmc_mdep * 5643 pmc_generic_cpu_initialize(void) 5644 { 5645 struct pmc_mdep *md; 5646 5647 md = pmc_mdep_alloc(0); 5648 5649 md->pmd_cputype = PMC_CPU_GENERIC; 5650 5651 return (md); 5652 } 5653 5654 static void 5655 pmc_generic_cpu_finalize(struct pmc_mdep *md __unused) 5656 { 5657 5658 } 5659 5660 static int 5661 pmc_initialize(void) 5662 { 5663 struct pcpu *pc; 5664 struct pmc_binding pb; 5665 struct pmc_classdep *pcd; 5666 struct pmc_sample *ps; 5667 struct pmc_samplebuffer *sb; 5668 int c, cpu, error, n, ri; 5669 u_int maxcpu, domain; 5670 5671 md = NULL; 5672 error = 0; 5673 5674 pmc_stats.pm_intr_ignored = counter_u64_alloc(M_WAITOK); 5675 pmc_stats.pm_intr_processed = counter_u64_alloc(M_WAITOK); 5676 pmc_stats.pm_intr_bufferfull = counter_u64_alloc(M_WAITOK); 5677 pmc_stats.pm_syscalls = counter_u64_alloc(M_WAITOK); 5678 pmc_stats.pm_syscall_errors = counter_u64_alloc(M_WAITOK); 5679 pmc_stats.pm_buffer_requests = counter_u64_alloc(M_WAITOK); 5680 pmc_stats.pm_buffer_requests_failed = counter_u64_alloc(M_WAITOK); 5681 pmc_stats.pm_log_sweeps = counter_u64_alloc(M_WAITOK); 5682 pmc_stats.pm_merges = counter_u64_alloc(M_WAITOK); 5683 pmc_stats.pm_overwrites = counter_u64_alloc(M_WAITOK); 5684 5685 #ifdef HWPMC_DEBUG 5686 /* parse debug flags first */ 5687 if (TUNABLE_STR_FETCH(PMC_SYSCTL_NAME_PREFIX "debugflags", 5688 pmc_debugstr, sizeof(pmc_debugstr))) { 5689 pmc_debugflags_parse(pmc_debugstr, pmc_debugstr + 5690 strlen(pmc_debugstr)); 5691 } 5692 #endif 5693 5694 PMCDBG1(MOD,INI,0, "PMC Initialize (version %x)", PMC_VERSION); 5695 5696 /* check kernel version */ 5697 if (pmc_kernel_version != PMC_VERSION) { 5698 if (pmc_kernel_version == 0) 5699 printf("hwpmc: this kernel has not been compiled with " 5700 "'options HWPMC_HOOKS'.\n"); 5701 else 5702 printf("hwpmc: kernel version (0x%x) does not match " 5703 "module version (0x%x).\n", pmc_kernel_version, 5704 PMC_VERSION); 5705 return (EPROGMISMATCH); 5706 } 5707 5708 /* 5709 * check sysctl parameters 5710 */ 5711 if (pmc_hashsize <= 0) { 5712 printf("hwpmc: tunable \"hashsize\"=%d must be " 5713 "greater than zero.\n", pmc_hashsize); 5714 pmc_hashsize = PMC_HASH_SIZE; 5715 } 5716 5717 if (pmc_nsamples <= 0 || pmc_nsamples > 65535) { 5718 printf("hwpmc: tunable \"nsamples\"=%d out of " 5719 "range.\n", pmc_nsamples); 5720 pmc_nsamples = PMC_NSAMPLES; 5721 } 5722 pmc_sample_mask = pmc_nsamples - 1; 5723 5724 if (pmc_callchaindepth <= 0 || 5725 pmc_callchaindepth > PMC_CALLCHAIN_DEPTH_MAX) { 5726 printf("hwpmc: tunable \"callchaindepth\"=%d out of " 5727 "range - using %d.\n", pmc_callchaindepth, 5728 PMC_CALLCHAIN_DEPTH_MAX); 5729 pmc_callchaindepth = PMC_CALLCHAIN_DEPTH_MAX; 5730 } 5731 5732 md = pmc_md_initialize(); 5733 if (md == NULL) { 5734 /* Default to generic CPU. */ 5735 md = pmc_generic_cpu_initialize(); 5736 if (md == NULL) 5737 return (ENOSYS); 5738 } 5739 5740 /* 5741 * Refresh classes base ri. Optional classes may come in different 5742 * order. 5743 */ 5744 for (ri = c = 0; c < md->pmd_nclass; c++) { 5745 pcd = &md->pmd_classdep[c]; 5746 pcd->pcd_ri = ri; 5747 ri += pcd->pcd_num; 5748 } 5749 5750 KASSERT(md->pmd_nclass >= 1 && md->pmd_npmc >= 1, 5751 ("[pmc,%d] no classes or pmcs", __LINE__)); 5752 5753 /* Compute the map from row-indices to classdep pointers. */ 5754 pmc_rowindex_to_classdep = malloc(sizeof(struct pmc_classdep *) * 5755 md->pmd_npmc, M_PMC, M_WAITOK | M_ZERO); 5756 5757 for (n = 0; n < md->pmd_npmc; n++) 5758 pmc_rowindex_to_classdep[n] = NULL; 5759 5760 for (ri = c = 0; c < md->pmd_nclass; c++) { 5761 pcd = &md->pmd_classdep[c]; 5762 for (n = 0; n < pcd->pcd_num; n++, ri++) 5763 pmc_rowindex_to_classdep[ri] = pcd; 5764 } 5765 5766 KASSERT(ri == md->pmd_npmc, 5767 ("[pmc,%d] npmc miscomputed: ri=%d, md->npmc=%d", __LINE__, 5768 ri, md->pmd_npmc)); 5769 5770 maxcpu = pmc_cpu_max(); 5771 5772 /* allocate space for the per-cpu array */ 5773 pmc_pcpu = malloc(maxcpu * sizeof(struct pmc_cpu *), M_PMC, 5774 M_WAITOK | M_ZERO); 5775 5776 /* per-cpu 'saved values' for managing process-mode PMCs */ 5777 pmc_pcpu_saved = malloc(sizeof(pmc_value_t) * maxcpu * md->pmd_npmc, 5778 M_PMC, M_WAITOK); 5779 5780 /* Perform CPU-dependent initialization. */ 5781 pmc_save_cpu_binding(&pb); 5782 error = 0; 5783 for (cpu = 0; error == 0 && cpu < maxcpu; cpu++) { 5784 if (!pmc_cpu_is_active(cpu)) 5785 continue; 5786 pmc_select_cpu(cpu); 5787 pmc_pcpu[cpu] = malloc(sizeof(struct pmc_cpu) + 5788 md->pmd_npmc * sizeof(struct pmc_hw *), M_PMC, 5789 M_WAITOK | M_ZERO); 5790 for (n = 0; error == 0 && n < md->pmd_nclass; n++) 5791 if (md->pmd_classdep[n].pcd_num > 0) 5792 error = md->pmd_classdep[n].pcd_pcpu_init(md, 5793 cpu); 5794 } 5795 pmc_restore_cpu_binding(&pb); 5796 5797 if (error != 0) 5798 return (error); 5799 5800 /* allocate space for the sample array */ 5801 for (cpu = 0; cpu < maxcpu; cpu++) { 5802 if (!pmc_cpu_is_active(cpu)) 5803 continue; 5804 pc = pcpu_find(cpu); 5805 domain = pc->pc_domain; 5806 sb = malloc_domainset(sizeof(struct pmc_samplebuffer) + 5807 pmc_nsamples * sizeof(struct pmc_sample), M_PMC, 5808 DOMAINSET_PREF(domain), M_WAITOK | M_ZERO); 5809 5810 KASSERT(pmc_pcpu[cpu] != NULL, 5811 ("[pmc,%d] cpu=%d Null per-cpu data", __LINE__, cpu)); 5812 5813 sb->ps_callchains = malloc_domainset(pmc_callchaindepth * 5814 pmc_nsamples * sizeof(uintptr_t), M_PMC, 5815 DOMAINSET_PREF(domain), M_WAITOK | M_ZERO); 5816 5817 for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++) 5818 ps->ps_pc = sb->ps_callchains + 5819 (n * pmc_callchaindepth); 5820 5821 pmc_pcpu[cpu]->pc_sb[PMC_HR] = sb; 5822 5823 sb = malloc_domainset(sizeof(struct pmc_samplebuffer) + 5824 pmc_nsamples * sizeof(struct pmc_sample), M_PMC, 5825 DOMAINSET_PREF(domain), M_WAITOK | M_ZERO); 5826 5827 sb->ps_callchains = malloc_domainset(pmc_callchaindepth * 5828 pmc_nsamples * sizeof(uintptr_t), M_PMC, 5829 DOMAINSET_PREF(domain), M_WAITOK | M_ZERO); 5830 for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++) 5831 ps->ps_pc = sb->ps_callchains + 5832 (n * pmc_callchaindepth); 5833 5834 pmc_pcpu[cpu]->pc_sb[PMC_SR] = sb; 5835 5836 sb = malloc_domainset(sizeof(struct pmc_samplebuffer) + 5837 pmc_nsamples * sizeof(struct pmc_sample), M_PMC, 5838 DOMAINSET_PREF(domain), M_WAITOK | M_ZERO); 5839 sb->ps_callchains = malloc_domainset(pmc_callchaindepth * 5840 pmc_nsamples * sizeof(uintptr_t), M_PMC, 5841 DOMAINSET_PREF(domain), M_WAITOK | M_ZERO); 5842 for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++) 5843 ps->ps_pc = sb->ps_callchains + n * pmc_callchaindepth; 5844 5845 pmc_pcpu[cpu]->pc_sb[PMC_UR] = sb; 5846 } 5847 5848 /* allocate space for the row disposition array */ 5849 pmc_pmcdisp = malloc(sizeof(enum pmc_mode) * md->pmd_npmc, 5850 M_PMC, M_WAITOK | M_ZERO); 5851 5852 /* mark all PMCs as available */ 5853 for (n = 0; n < md->pmd_npmc; n++) 5854 PMC_MARK_ROW_FREE(n); 5855 5856 /* allocate thread hash tables */ 5857 pmc_ownerhash = hashinit(pmc_hashsize, M_PMC, 5858 &pmc_ownerhashmask); 5859 5860 pmc_processhash = hashinit(pmc_hashsize, M_PMC, 5861 &pmc_processhashmask); 5862 mtx_init(&pmc_processhash_mtx, "pmc-process-hash", "pmc-leaf", 5863 MTX_SPIN); 5864 5865 CK_LIST_INIT(&pmc_ss_owners); 5866 pmc_ss_count = 0; 5867 5868 /* allocate a pool of spin mutexes */ 5869 pmc_mtxpool = mtx_pool_create("pmc-leaf", pmc_mtxpool_size, 5870 MTX_SPIN); 5871 5872 PMCDBG4(MOD,INI,1, "pmc_ownerhash=%p, mask=0x%lx " 5873 "targethash=%p mask=0x%lx", pmc_ownerhash, pmc_ownerhashmask, 5874 pmc_processhash, pmc_processhashmask); 5875 5876 /* Initialize a spin mutex for the thread free list. */ 5877 mtx_init(&pmc_threadfreelist_mtx, "pmc-threadfreelist", "pmc-leaf", 5878 MTX_SPIN); 5879 5880 /* Initialize the task to prune the thread free list. */ 5881 TASK_INIT(&free_task, 0, pmc_thread_descriptor_pool_free_task, NULL); 5882 5883 /* register process {exit,fork,exec} handlers */ 5884 pmc_exit_tag = EVENTHANDLER_REGISTER(process_exit, 5885 pmc_process_exit, NULL, EVENTHANDLER_PRI_ANY); 5886 pmc_fork_tag = EVENTHANDLER_REGISTER(process_fork, 5887 pmc_process_fork, NULL, EVENTHANDLER_PRI_ANY); 5888 5889 /* register kld event handlers */ 5890 pmc_kld_load_tag = EVENTHANDLER_REGISTER(kld_load, pmc_kld_load, 5891 NULL, EVENTHANDLER_PRI_ANY); 5892 pmc_kld_unload_tag = EVENTHANDLER_REGISTER(kld_unload, pmc_kld_unload, 5893 NULL, EVENTHANDLER_PRI_ANY); 5894 5895 /* initialize logging */ 5896 pmclog_initialize(); 5897 5898 /* set hook functions */ 5899 pmc_intr = md->pmd_intr; 5900 wmb(); 5901 pmc_hook = pmc_hook_handler; 5902 5903 if (error == 0) { 5904 printf(PMC_MODULE_NAME ":"); 5905 for (n = 0; n < md->pmd_nclass; n++) { 5906 if (md->pmd_classdep[n].pcd_num == 0) 5907 continue; 5908 pcd = &md->pmd_classdep[n]; 5909 printf(" %s/%d/%d/0x%b", 5910 pmc_name_of_pmcclass(pcd->pcd_class), 5911 pcd->pcd_num, 5912 pcd->pcd_width, 5913 pcd->pcd_caps, 5914 "\20" 5915 "\1INT\2USR\3SYS\4EDG\5THR" 5916 "\6REA\7WRI\10INV\11QUA\12PRC" 5917 "\13TAG\14CSC"); 5918 } 5919 printf("\n"); 5920 } 5921 5922 return (error); 5923 } 5924 5925 /* prepare to be unloaded */ 5926 static void 5927 pmc_cleanup(void) 5928 { 5929 struct pmc_binding pb; 5930 struct pmc_owner *po, *tmp; 5931 struct pmc_ownerhash *ph; 5932 struct pmc_processhash *prh __pmcdbg_used; 5933 u_int maxcpu; 5934 int cpu, c; 5935 5936 PMCDBG0(MOD,INI,0, "cleanup"); 5937 5938 /* switch off sampling */ 5939 CPU_FOREACH(cpu) 5940 DPCPU_ID_SET(cpu, pmc_sampled, 0); 5941 pmc_intr = NULL; 5942 5943 sx_xlock(&pmc_sx); 5944 if (pmc_hook == NULL) { /* being unloaded already */ 5945 sx_xunlock(&pmc_sx); 5946 return; 5947 } 5948 5949 pmc_hook = NULL; /* prevent new threads from entering module */ 5950 5951 /* deregister event handlers */ 5952 EVENTHANDLER_DEREGISTER(process_fork, pmc_fork_tag); 5953 EVENTHANDLER_DEREGISTER(process_exit, pmc_exit_tag); 5954 EVENTHANDLER_DEREGISTER(kld_load, pmc_kld_load_tag); 5955 EVENTHANDLER_DEREGISTER(kld_unload, pmc_kld_unload_tag); 5956 5957 /* send SIGBUS to all owner threads, free up allocations */ 5958 if (pmc_ownerhash != NULL) { 5959 for (ph = pmc_ownerhash; 5960 ph <= &pmc_ownerhash[pmc_ownerhashmask]; 5961 ph++) { 5962 LIST_FOREACH_SAFE(po, ph, po_next, tmp) { 5963 pmc_remove_owner(po); 5964 5965 PMCDBG3(MOD,INI,2, 5966 "cleanup signal proc=%p (%d, %s)", 5967 po->po_owner, po->po_owner->p_pid, 5968 po->po_owner->p_comm); 5969 5970 PROC_LOCK(po->po_owner); 5971 kern_psignal(po->po_owner, SIGBUS); 5972 PROC_UNLOCK(po->po_owner); 5973 5974 pmc_destroy_owner_descriptor(po); 5975 } 5976 } 5977 } 5978 5979 /* reclaim allocated data structures */ 5980 taskqueue_drain(taskqueue_fast, &free_task); 5981 mtx_destroy(&pmc_threadfreelist_mtx); 5982 pmc_thread_descriptor_pool_drain(); 5983 5984 if (pmc_mtxpool != NULL) 5985 mtx_pool_destroy(&pmc_mtxpool); 5986 5987 mtx_destroy(&pmc_processhash_mtx); 5988 if (pmc_processhash != NULL) { 5989 #ifdef HWPMC_DEBUG 5990 struct pmc_process *pp; 5991 5992 PMCDBG0(MOD,INI,3, "destroy process hash"); 5993 for (prh = pmc_processhash; 5994 prh <= &pmc_processhash[pmc_processhashmask]; 5995 prh++) 5996 LIST_FOREACH(pp, prh, pp_next) 5997 PMCDBG1(MOD,INI,3, "pid=%d", pp->pp_proc->p_pid); 5998 #endif 5999 6000 hashdestroy(pmc_processhash, M_PMC, pmc_processhashmask); 6001 pmc_processhash = NULL; 6002 } 6003 6004 if (pmc_ownerhash != NULL) { 6005 PMCDBG0(MOD,INI,3, "destroy owner hash"); 6006 hashdestroy(pmc_ownerhash, M_PMC, pmc_ownerhashmask); 6007 pmc_ownerhash = NULL; 6008 } 6009 6010 KASSERT(CK_LIST_EMPTY(&pmc_ss_owners), 6011 ("[pmc,%d] Global SS owner list not empty", __LINE__)); 6012 KASSERT(pmc_ss_count == 0, 6013 ("[pmc,%d] Global SS count not empty", __LINE__)); 6014 6015 /* do processor and pmc-class dependent cleanup */ 6016 maxcpu = pmc_cpu_max(); 6017 6018 PMCDBG0(MOD,INI,3, "md cleanup"); 6019 if (md) { 6020 pmc_save_cpu_binding(&pb); 6021 for (cpu = 0; cpu < maxcpu; cpu++) { 6022 PMCDBG2(MOD,INI,1,"pmc-cleanup cpu=%d pcs=%p", 6023 cpu, pmc_pcpu[cpu]); 6024 if (!pmc_cpu_is_active(cpu) || pmc_pcpu[cpu] == NULL) 6025 continue; 6026 6027 pmc_select_cpu(cpu); 6028 for (c = 0; c < md->pmd_nclass; c++) { 6029 if (md->pmd_classdep[c].pcd_num > 0) { 6030 md->pmd_classdep[c].pcd_pcpu_fini(md, 6031 cpu); 6032 } 6033 } 6034 } 6035 6036 if (md->pmd_cputype == PMC_CPU_GENERIC) 6037 pmc_generic_cpu_finalize(md); 6038 else 6039 pmc_md_finalize(md); 6040 6041 pmc_mdep_free(md); 6042 md = NULL; 6043 pmc_restore_cpu_binding(&pb); 6044 } 6045 6046 /* Free per-cpu descriptors. */ 6047 for (cpu = 0; cpu < maxcpu; cpu++) { 6048 if (!pmc_cpu_is_active(cpu)) 6049 continue; 6050 KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_HR] != NULL, 6051 ("[pmc,%d] Null hw cpu sample buffer cpu=%d", __LINE__, 6052 cpu)); 6053 KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_SR] != NULL, 6054 ("[pmc,%d] Null sw cpu sample buffer cpu=%d", __LINE__, 6055 cpu)); 6056 KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_UR] != NULL, 6057 ("[pmc,%d] Null userret cpu sample buffer cpu=%d", __LINE__, 6058 cpu)); 6059 free(pmc_pcpu[cpu]->pc_sb[PMC_HR]->ps_callchains, M_PMC); 6060 free(pmc_pcpu[cpu]->pc_sb[PMC_HR], M_PMC); 6061 free(pmc_pcpu[cpu]->pc_sb[PMC_SR]->ps_callchains, M_PMC); 6062 free(pmc_pcpu[cpu]->pc_sb[PMC_SR], M_PMC); 6063 free(pmc_pcpu[cpu]->pc_sb[PMC_UR]->ps_callchains, M_PMC); 6064 free(pmc_pcpu[cpu]->pc_sb[PMC_UR], M_PMC); 6065 free(pmc_pcpu[cpu], M_PMC); 6066 } 6067 6068 free(pmc_pcpu, M_PMC); 6069 pmc_pcpu = NULL; 6070 6071 free(pmc_pcpu_saved, M_PMC); 6072 pmc_pcpu_saved = NULL; 6073 6074 if (pmc_pmcdisp != NULL) { 6075 free(pmc_pmcdisp, M_PMC); 6076 pmc_pmcdisp = NULL; 6077 } 6078 6079 if (pmc_rowindex_to_classdep != NULL) { 6080 free(pmc_rowindex_to_classdep, M_PMC); 6081 pmc_rowindex_to_classdep = NULL; 6082 } 6083 6084 pmclog_shutdown(); 6085 counter_u64_free(pmc_stats.pm_intr_ignored); 6086 counter_u64_free(pmc_stats.pm_intr_processed); 6087 counter_u64_free(pmc_stats.pm_intr_bufferfull); 6088 counter_u64_free(pmc_stats.pm_syscalls); 6089 counter_u64_free(pmc_stats.pm_syscall_errors); 6090 counter_u64_free(pmc_stats.pm_buffer_requests); 6091 counter_u64_free(pmc_stats.pm_buffer_requests_failed); 6092 counter_u64_free(pmc_stats.pm_log_sweeps); 6093 counter_u64_free(pmc_stats.pm_merges); 6094 counter_u64_free(pmc_stats.pm_overwrites); 6095 sx_xunlock(&pmc_sx); /* we are done */ 6096 } 6097 6098 /* 6099 * The function called at load/unload. 6100 */ 6101 static int 6102 load(struct module *module __unused, int cmd, void *arg __unused) 6103 { 6104 int error; 6105 6106 error = 0; 6107 6108 switch (cmd) { 6109 case MOD_LOAD: 6110 /* initialize the subsystem */ 6111 error = pmc_initialize(); 6112 if (error != 0) 6113 break; 6114 PMCDBG2(MOD,INI,1, "syscall=%d maxcpu=%d", pmc_syscall_num, 6115 pmc_cpu_max()); 6116 break; 6117 case MOD_UNLOAD: 6118 case MOD_SHUTDOWN: 6119 pmc_cleanup(); 6120 PMCDBG0(MOD,INI,1, "unloaded"); 6121 break; 6122 default: 6123 error = EINVAL; 6124 break; 6125 } 6126 6127 return (error); 6128 } 6129