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_out_prepare(int cpu); 252 static void pmc_process_csw_start_all(int cpu); 253 static void pmc_process_csw_stop_all(int cpu); 254 static void pmc_process_exec(struct thread *td, 255 struct pmckern_procexec *pk); 256 static void pmc_process_exit(void *arg, struct proc *p); 257 static void pmc_process_fork(void *arg, struct proc *p1, 258 struct proc *p2, int n); 259 static void pmc_process_proccreate(struct proc *p); 260 static void pmc_process_samples(int cpu, ring_type_t soft); 261 static void pmc_process_threadcreate(struct thread *td); 262 static void pmc_process_threadexit(struct thread *td); 263 static void pmc_process_thread_add(struct thread *td); 264 static void pmc_process_thread_delete(struct thread *td); 265 static void pmc_process_thread_userret(struct thread *td); 266 static void pmc_release_pmc_descriptor(struct pmc *pmc); 267 static void pmc_remove_owner(struct pmc_owner *po); 268 static void pmc_remove_process_descriptor(struct pmc_process *pp); 269 static int pmc_start(struct pmc *pm); 270 static int pmc_stop(struct pmc *pm); 271 static int pmc_syscall_handler(struct thread *td, void *syscall_args); 272 static struct pmc_thread *pmc_thread_descriptor_pool_alloc(void); 273 static void pmc_thread_descriptor_pool_drain(void); 274 static void pmc_thread_descriptor_pool_free(struct pmc_thread *pt); 275 static void pmc_unlink_target_process(struct pmc *pmc, 276 struct pmc_process *pp); 277 278 static int generic_switch_in(struct pmc_cpu *pc, struct pmc_process *pp); 279 static int generic_switch_out(struct pmc_cpu *pc, struct pmc_process *pp); 280 static struct pmc_mdep *pmc_generic_cpu_initialize(void); 281 static void pmc_generic_cpu_finalize(struct pmc_mdep *md); 282 283 /* 284 * Kernel tunables and sysctl(8) interface. 285 */ 286 287 SYSCTL_DECL(_kern_hwpmc); 288 SYSCTL_NODE(_kern_hwpmc, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 289 "HWPMC stats"); 290 291 /* Stats. */ 292 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, intr_ignored, CTLFLAG_RW, 293 &pmc_stats.pm_intr_ignored, 294 "# of interrupts ignored"); 295 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, intr_processed, CTLFLAG_RW, 296 &pmc_stats.pm_intr_processed, 297 "# of interrupts processed"); 298 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, intr_bufferfull, CTLFLAG_RW, 299 &pmc_stats.pm_intr_bufferfull, 300 "# of interrupts where buffer was full"); 301 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, syscalls, CTLFLAG_RW, 302 &pmc_stats.pm_syscalls, 303 "# of syscalls"); 304 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, syscall_errors, CTLFLAG_RW, 305 &pmc_stats.pm_syscall_errors, 306 "# of syscall_errors"); 307 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, buffer_requests, CTLFLAG_RW, 308 &pmc_stats.pm_buffer_requests, 309 "# of buffer requests"); 310 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, buffer_requests_failed, 311 CTLFLAG_RW, &pmc_stats.pm_buffer_requests_failed, 312 "# of buffer requests which failed"); 313 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, log_sweeps, CTLFLAG_RW, 314 &pmc_stats.pm_log_sweeps, 315 "# of times samples were processed"); 316 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, merges, CTLFLAG_RW, 317 &pmc_stats.pm_merges, 318 "# of times kernel stack was found for user trace"); 319 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, overwrites, CTLFLAG_RW, 320 &pmc_stats.pm_overwrites, 321 "# of times a sample was overwritten before being logged"); 322 323 static int pmc_callchaindepth = PMC_CALLCHAIN_DEPTH; 324 SYSCTL_INT(_kern_hwpmc, OID_AUTO, callchaindepth, CTLFLAG_RDTUN, 325 &pmc_callchaindepth, 0, 326 "depth of call chain records"); 327 328 char pmc_cpuid[PMC_CPUID_LEN]; 329 SYSCTL_STRING(_kern_hwpmc, OID_AUTO, cpuid, CTLFLAG_RD, 330 pmc_cpuid, 0, 331 "cpu version string"); 332 333 #ifdef HWPMC_DEBUG 334 struct pmc_debugflags pmc_debugflags = PMC_DEBUG_DEFAULT_FLAGS; 335 char pmc_debugstr[PMC_DEBUG_STRSIZE]; 336 TUNABLE_STR(PMC_SYSCTL_NAME_PREFIX "debugflags", pmc_debugstr, 337 sizeof(pmc_debugstr)); 338 SYSCTL_PROC(_kern_hwpmc, OID_AUTO, debugflags, 339 CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, 340 0, 0, pmc_debugflags_sysctl_handler, "A", 341 "debug flags"); 342 #endif 343 344 /* 345 * kern.hwpmc.hashsize -- determines the number of rows in the 346 * of the hash table used to look up threads 347 */ 348 static int pmc_hashsize = PMC_HASH_SIZE; 349 SYSCTL_INT(_kern_hwpmc, OID_AUTO, hashsize, CTLFLAG_RDTUN, 350 &pmc_hashsize, 0, 351 "rows in hash tables"); 352 353 /* 354 * kern.hwpmc.nsamples --- number of PC samples/callchain stacks per CPU 355 */ 356 static int pmc_nsamples = PMC_NSAMPLES; 357 SYSCTL_INT(_kern_hwpmc, OID_AUTO, nsamples, CTLFLAG_RDTUN, 358 &pmc_nsamples, 0, 359 "number of PC samples per CPU"); 360 361 static uint64_t pmc_sample_mask = PMC_NSAMPLES - 1; 362 363 /* 364 * kern.hwpmc.mtxpoolsize -- number of mutexes in the mutex pool. 365 */ 366 static int pmc_mtxpool_size = PMC_MTXPOOL_SIZE; 367 SYSCTL_INT(_kern_hwpmc, OID_AUTO, mtxpoolsize, CTLFLAG_RDTUN, 368 &pmc_mtxpool_size, 0, 369 "size of spin mutex pool"); 370 371 /* 372 * kern.hwpmc.threadfreelist_entries -- number of free entries 373 */ 374 SYSCTL_INT(_kern_hwpmc, OID_AUTO, threadfreelist_entries, CTLFLAG_RD, 375 &pmc_threadfreelist_entries, 0, 376 "number of available thread entries"); 377 378 /* 379 * kern.hwpmc.threadfreelist_max -- maximum number of free entries 380 */ 381 static int pmc_threadfreelist_max = PMC_THREADLIST_MAX; 382 SYSCTL_INT(_kern_hwpmc, OID_AUTO, threadfreelist_max, CTLFLAG_RW, 383 &pmc_threadfreelist_max, 0, 384 "maximum number of available thread entries before freeing some"); 385 386 /* 387 * kern.hwpmc.mincount -- minimum sample count 388 */ 389 static u_int pmc_mincount = 1000; 390 SYSCTL_INT(_kern_hwpmc, OID_AUTO, mincount, CTLFLAG_RWTUN, 391 &pmc_mincount, 0, 392 "minimum count for sampling counters"); 393 394 /* 395 * security.bsd.unprivileged_syspmcs -- allow non-root processes to 396 * allocate system-wide PMCs. 397 * 398 * Allowing unprivileged processes to allocate system PMCs is convenient 399 * if system-wide measurements need to be taken concurrently with other 400 * per-process measurements. This feature is turned off by default. 401 */ 402 static int pmc_unprivileged_syspmcs = 0; 403 SYSCTL_INT(_security_bsd, OID_AUTO, unprivileged_syspmcs, CTLFLAG_RWTUN, 404 &pmc_unprivileged_syspmcs, 0, 405 "allow unprivileged process to allocate system PMCs"); 406 407 /* 408 * Hash function. Discard the lower 2 bits of the pointer since 409 * these are always zero for our uses. The hash multiplier is 410 * round((2^LONG_BIT) * ((sqrt(5)-1)/2)). 411 */ 412 #if LONG_BIT == 64 413 #define _PMC_HM 11400714819323198486u 414 #elif LONG_BIT == 32 415 #define _PMC_HM 2654435769u 416 #else 417 #error Must know the size of 'long' to compile 418 #endif 419 420 #define PMC_HASH_PTR(P,M) ((((unsigned long) (P) >> 2) * _PMC_HM) & (M)) 421 422 /* 423 * Syscall structures 424 */ 425 426 /* The `sysent' for the new syscall */ 427 static struct sysent pmc_sysent = { 428 .sy_narg = 2, 429 .sy_call = pmc_syscall_handler, 430 }; 431 432 static struct syscall_module_data pmc_syscall_mod = { 433 .chainevh = load, 434 .chainarg = NULL, 435 .offset = &pmc_syscall_num, 436 .new_sysent = &pmc_sysent, 437 .old_sysent = { .sy_narg = 0, .sy_call = NULL }, 438 .flags = SY_THR_STATIC_KLD, 439 }; 440 441 static moduledata_t pmc_mod = { 442 .name = PMC_MODULE_NAME, 443 .evhand = syscall_module_handler, 444 .priv = &pmc_syscall_mod, 445 }; 446 447 #ifdef EARLY_AP_STARTUP 448 DECLARE_MODULE(pmc, pmc_mod, SI_SUB_SYSCALLS, SI_ORDER_ANY); 449 #else 450 DECLARE_MODULE(pmc, pmc_mod, SI_SUB_SMP, SI_ORDER_ANY); 451 #endif 452 MODULE_VERSION(pmc, PMC_VERSION); 453 454 #ifdef HWPMC_DEBUG 455 enum pmc_dbgparse_state { 456 PMCDS_WS, /* in whitespace */ 457 PMCDS_MAJOR, /* seen a major keyword */ 458 PMCDS_MINOR 459 }; 460 461 static int 462 pmc_debugflags_parse(char *newstr, char *fence) 463 { 464 struct pmc_debugflags *tmpflags; 465 size_t kwlen; 466 char c, *p, *q; 467 int error, *newbits, tmp; 468 int found; 469 470 tmpflags = malloc(sizeof(*tmpflags), M_PMC, M_WAITOK | M_ZERO); 471 472 error = 0; 473 for (p = newstr; p < fence && (c = *p); p++) { 474 /* skip white space */ 475 if (c == ' ' || c == '\t') 476 continue; 477 478 /* look for a keyword followed by "=" */ 479 for (q = p; p < fence && (c = *p) && c != '='; p++) 480 ; 481 if (c != '=') { 482 error = EINVAL; 483 goto done; 484 } 485 486 kwlen = p - q; 487 newbits = NULL; 488 489 /* lookup flag group name */ 490 #define DBG_SET_FLAG_MAJ(S,F) \ 491 if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0) \ 492 newbits = &tmpflags->pdb_ ## F; 493 494 DBG_SET_FLAG_MAJ("cpu", CPU); 495 DBG_SET_FLAG_MAJ("csw", CSW); 496 DBG_SET_FLAG_MAJ("logging", LOG); 497 DBG_SET_FLAG_MAJ("module", MOD); 498 DBG_SET_FLAG_MAJ("md", MDP); 499 DBG_SET_FLAG_MAJ("owner", OWN); 500 DBG_SET_FLAG_MAJ("pmc", PMC); 501 DBG_SET_FLAG_MAJ("process", PRC); 502 DBG_SET_FLAG_MAJ("sampling", SAM); 503 #undef DBG_SET_FLAG_MAJ 504 505 if (newbits == NULL) { 506 error = EINVAL; 507 goto done; 508 } 509 510 p++; /* skip the '=' */ 511 512 /* Now parse the individual flags */ 513 tmp = 0; 514 newflag: 515 for (q = p; p < fence && (c = *p); p++) 516 if (c == ' ' || c == '\t' || c == ',') 517 break; 518 519 /* p == fence or c == ws or c == "," or c == 0 */ 520 521 if ((kwlen = p - q) == 0) { 522 *newbits = tmp; 523 continue; 524 } 525 526 found = 0; 527 #define DBG_SET_FLAG_MIN(S,F) \ 528 if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0) \ 529 tmp |= found = (1 << PMC_DEBUG_MIN_ ## F) 530 531 /* a '*' denotes all possible flags in the group */ 532 if (kwlen == 1 && *q == '*') 533 tmp = found = ~0; 534 /* look for individual flag names */ 535 DBG_SET_FLAG_MIN("allocaterow", ALR); 536 DBG_SET_FLAG_MIN("allocate", ALL); 537 DBG_SET_FLAG_MIN("attach", ATT); 538 DBG_SET_FLAG_MIN("bind", BND); 539 DBG_SET_FLAG_MIN("config", CFG); 540 DBG_SET_FLAG_MIN("exec", EXC); 541 DBG_SET_FLAG_MIN("exit", EXT); 542 DBG_SET_FLAG_MIN("find", FND); 543 DBG_SET_FLAG_MIN("flush", FLS); 544 DBG_SET_FLAG_MIN("fork", FRK); 545 DBG_SET_FLAG_MIN("getbuf", GTB); 546 DBG_SET_FLAG_MIN("hook", PMH); 547 DBG_SET_FLAG_MIN("init", INI); 548 DBG_SET_FLAG_MIN("intr", INT); 549 DBG_SET_FLAG_MIN("linktarget", TLK); 550 DBG_SET_FLAG_MIN("mayberemove", OMR); 551 DBG_SET_FLAG_MIN("ops", OPS); 552 DBG_SET_FLAG_MIN("read", REA); 553 DBG_SET_FLAG_MIN("register", REG); 554 DBG_SET_FLAG_MIN("release", REL); 555 DBG_SET_FLAG_MIN("remove", ORM); 556 DBG_SET_FLAG_MIN("sample", SAM); 557 DBG_SET_FLAG_MIN("scheduleio", SIO); 558 DBG_SET_FLAG_MIN("select", SEL); 559 DBG_SET_FLAG_MIN("signal", SIG); 560 DBG_SET_FLAG_MIN("swi", SWI); 561 DBG_SET_FLAG_MIN("swo", SWO); 562 DBG_SET_FLAG_MIN("start", STA); 563 DBG_SET_FLAG_MIN("stop", STO); 564 DBG_SET_FLAG_MIN("syscall", PMS); 565 DBG_SET_FLAG_MIN("unlinktarget", TUL); 566 DBG_SET_FLAG_MIN("write", WRI); 567 #undef DBG_SET_FLAG_MIN 568 if (found == 0) { 569 /* unrecognized flag name */ 570 error = EINVAL; 571 goto done; 572 } 573 574 if (c == 0 || c == ' ' || c == '\t') { /* end of flag group */ 575 *newbits = tmp; 576 continue; 577 } 578 579 p++; 580 goto newflag; 581 } 582 583 /* save the new flag set */ 584 bcopy(tmpflags, &pmc_debugflags, sizeof(pmc_debugflags)); 585 done: 586 free(tmpflags, M_PMC); 587 return (error); 588 } 589 590 static int 591 pmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS) 592 { 593 char *fence, *newstr; 594 int error; 595 u_int n; 596 597 n = sizeof(pmc_debugstr); 598 newstr = malloc(n, M_PMC, M_WAITOK | M_ZERO); 599 strlcpy(newstr, pmc_debugstr, n); 600 601 error = sysctl_handle_string(oidp, newstr, n, req); 602 603 /* if there is a new string, parse and copy it */ 604 if (error == 0 && req->newptr != NULL) { 605 fence = newstr + (n < req->newlen ? n : req->newlen + 1); 606 error = pmc_debugflags_parse(newstr, fence); 607 if (error == 0) 608 strlcpy(pmc_debugstr, newstr, sizeof(pmc_debugstr)); 609 } 610 free(newstr, M_PMC); 611 612 return (error); 613 } 614 #endif 615 616 /* 617 * Map a row index to a classdep structure and return the adjusted row 618 * index for the PMC class index. 619 */ 620 static struct pmc_classdep * 621 pmc_ri_to_classdep(struct pmc_mdep *md __unused, int ri, int *adjri) 622 { 623 struct pmc_classdep *pcd; 624 625 KASSERT(ri >= 0 && ri < md->pmd_npmc, 626 ("[pmc,%d] illegal row-index %d", __LINE__, ri)); 627 628 pcd = pmc_rowindex_to_classdep[ri]; 629 KASSERT(pcd != NULL, 630 ("[pmc,%d] ri %d null pcd", __LINE__, ri)); 631 632 *adjri = ri - pcd->pcd_ri; 633 KASSERT(*adjri >= 0 && *adjri < pcd->pcd_num, 634 ("[pmc,%d] adjusted row-index %d", __LINE__, *adjri)); 635 636 return (pcd); 637 } 638 639 /* 640 * Concurrency Control 641 * 642 * The driver manages the following data structures: 643 * 644 * - target process descriptors, one per target process 645 * - owner process descriptors (and attached lists), one per owner process 646 * - lookup hash tables for owner and target processes 647 * - PMC descriptors (and attached lists) 648 * - per-cpu hardware state 649 * - the 'hook' variable through which the kernel calls into 650 * this module 651 * - the machine hardware state (managed by the MD layer) 652 * 653 * These data structures are accessed from: 654 * 655 * - thread context-switch code 656 * - interrupt handlers (possibly on multiple cpus) 657 * - kernel threads on multiple cpus running on behalf of user 658 * processes doing system calls 659 * - this driver's private kernel threads 660 * 661 * = Locks and Locking strategy = 662 * 663 * The driver uses four locking strategies for its operation: 664 * 665 * - The global SX lock "pmc_sx" is used to protect internal 666 * data structures. 667 * 668 * Calls into the module by syscall() start with this lock being 669 * held in exclusive mode. Depending on the requested operation, 670 * the lock may be downgraded to 'shared' mode to allow more 671 * concurrent readers into the module. Calls into the module from 672 * other parts of the kernel acquire the lock in shared mode. 673 * 674 * This SX lock is held in exclusive mode for any operations that 675 * modify the linkages between the driver's internal data structures. 676 * 677 * The 'pmc_hook' function pointer is also protected by this lock. 678 * It is only examined with the sx lock held in exclusive mode. The 679 * kernel module is allowed to be unloaded only with the sx lock held 680 * in exclusive mode. In normal syscall handling, after acquiring the 681 * pmc_sx lock we first check that 'pmc_hook' is non-null before 682 * proceeding. This prevents races between the thread unloading the module 683 * and other threads seeking to use the module. 684 * 685 * - Lookups of target process structures and owner process structures 686 * cannot use the global "pmc_sx" SX lock because these lookups need 687 * to happen during context switches and in other critical sections 688 * where sleeping is not allowed. We protect these lookup tables 689 * with their own private spin-mutexes, "pmc_processhash_mtx" and 690 * "pmc_ownerhash_mtx". 691 * 692 * - Interrupt handlers work in a lock free manner. At interrupt 693 * time, handlers look at the PMC pointer (phw->phw_pmc) configured 694 * when the PMC was started. If this pointer is NULL, the interrupt 695 * is ignored after updating driver statistics. We ensure that this 696 * pointer is set (using an atomic operation if necessary) before the 697 * PMC hardware is started. Conversely, this pointer is unset atomically 698 * only after the PMC hardware is stopped. 699 * 700 * We ensure that everything needed for the operation of an 701 * interrupt handler is available without it needing to acquire any 702 * locks. We also ensure that a PMC's software state is destroyed only 703 * after the PMC is taken off hardware (on all CPUs). 704 * 705 * - Context-switch handling with process-private PMCs needs more 706 * care. 707 * 708 * A given process may be the target of multiple PMCs. For example, 709 * PMCATTACH and PMCDETACH may be requested by a process on one CPU 710 * while the target process is running on another. A PMC could also 711 * be getting released because its owner is exiting. We tackle 712 * these situations in the following manner: 713 * 714 * - each target process structure 'pmc_process' has an array 715 * of 'struct pmc *' pointers, one for each hardware PMC. 716 * 717 * - At context switch IN time, each "target" PMC in RUNNING state 718 * gets started on hardware and a pointer to each PMC is copied into 719 * the per-cpu phw array. The 'runcount' for the PMC is 720 * incremented. 721 * 722 * - At context switch OUT time, all process-virtual PMCs are stopped 723 * on hardware. The saved value is added to the PMCs value field 724 * only if the PMC is in a non-deleted state (the PMCs state could 725 * have changed during the current time slice). 726 * 727 * Note that since in-between a switch IN on a processor and a switch 728 * OUT, the PMC could have been released on another CPU. Therefore 729 * context switch OUT always looks at the hardware state to turn 730 * OFF PMCs and will update a PMC's saved value only if reachable 731 * from the target process record. 732 * 733 * - OP PMCRELEASE could be called on a PMC at any time (the PMC could 734 * be attached to many processes at the time of the call and could 735 * be active on multiple CPUs). 736 * 737 * We prevent further scheduling of the PMC by marking it as in 738 * state 'DELETED'. If the runcount of the PMC is non-zero then 739 * this PMC is currently running on a CPU somewhere. The thread 740 * doing the PMCRELEASE operation waits by repeatedly doing a 741 * pause() till the runcount comes to zero. 742 * 743 * The contents of a PMC descriptor (struct pmc) are protected using 744 * a spin-mutex. In order to save space, we use a mutex pool. 745 * 746 * In terms of lock types used by witness(4), we use: 747 * - Type "pmc-sx", used by the global SX lock. 748 * - Type "pmc-sleep", for sleep mutexes used by logger threads. 749 * - Type "pmc-per-proc", for protecting PMC owner descriptors. 750 * - Type "pmc-leaf", used for all other spin mutexes. 751 */ 752 753 /* 754 * Save the CPU binding of the current kthread. 755 */ 756 void 757 pmc_save_cpu_binding(struct pmc_binding *pb) 758 { 759 PMCDBG0(CPU,BND,2, "save-cpu"); 760 thread_lock(curthread); 761 pb->pb_bound = sched_is_bound(curthread); 762 pb->pb_cpu = curthread->td_oncpu; 763 pb->pb_priority = curthread->td_priority; 764 thread_unlock(curthread); 765 PMCDBG1(CPU,BND,2, "save-cpu cpu=%d", pb->pb_cpu); 766 } 767 768 /* 769 * Restore the CPU binding of the current thread. 770 */ 771 void 772 pmc_restore_cpu_binding(struct pmc_binding *pb) 773 { 774 PMCDBG2(CPU,BND,2, "restore-cpu curcpu=%d restore=%d", 775 curthread->td_oncpu, pb->pb_cpu); 776 thread_lock(curthread); 777 sched_bind(curthread, pb->pb_cpu); 778 if (!pb->pb_bound) 779 sched_unbind(curthread); 780 sched_prio(curthread, pb->pb_priority); 781 thread_unlock(curthread); 782 PMCDBG0(CPU,BND,2, "restore-cpu done"); 783 } 784 785 /* 786 * Move execution over to the specified CPU and bind it there. 787 */ 788 void 789 pmc_select_cpu(int cpu) 790 { 791 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), 792 ("[pmc,%d] bad cpu number %d", __LINE__, cpu)); 793 794 /* Never move to an inactive CPU. */ 795 KASSERT(pmc_cpu_is_active(cpu), ("[pmc,%d] selecting inactive " 796 "CPU %d", __LINE__, cpu)); 797 798 PMCDBG1(CPU,SEL,2, "select-cpu cpu=%d", cpu); 799 thread_lock(curthread); 800 sched_prio(curthread, PRI_MIN); 801 sched_bind(curthread, cpu); 802 thread_unlock(curthread); 803 804 KASSERT(curthread->td_oncpu == cpu, 805 ("[pmc,%d] CPU not bound [cpu=%d, curr=%d]", __LINE__, 806 cpu, curthread->td_oncpu)); 807 808 PMCDBG1(CPU,SEL,2, "select-cpu cpu=%d ok", cpu); 809 } 810 811 /* 812 * Force a context switch. 813 * 814 * We do this by pause'ing for 1 tick -- invoking mi_switch() is not 815 * guaranteed to force a context switch. 816 */ 817 static void 818 pmc_force_context_switch(void) 819 { 820 821 pause("pmcctx", 1); 822 } 823 824 uint64_t 825 pmc_rdtsc(void) 826 { 827 #if defined(__i386__) 828 /* Unfortunately get_cyclecount on i386 uses cpu_ticks. */ 829 return (rdtsc()); 830 #else 831 return (get_cyclecount()); 832 #endif 833 } 834 835 /* 836 * Get the file name for an executable. This is a simple wrapper 837 * around vn_fullpath(9). 838 */ 839 static void 840 pmc_getfilename(struct vnode *v, char **fullpath, char **freepath) 841 { 842 843 *fullpath = "unknown"; 844 *freepath = NULL; 845 vn_fullpath(v, fullpath, freepath); 846 } 847 848 /* 849 * Remove a process owning PMCs. 850 */ 851 void 852 pmc_remove_owner(struct pmc_owner *po) 853 { 854 struct pmc *pm, *tmp; 855 856 sx_assert(&pmc_sx, SX_XLOCKED); 857 858 PMCDBG1(OWN,ORM,1, "remove-owner po=%p", po); 859 860 /* Remove descriptor from the owner hash table */ 861 LIST_REMOVE(po, po_next); 862 863 /* release all owned PMC descriptors */ 864 LIST_FOREACH_SAFE(pm, &po->po_pmcs, pm_next, tmp) { 865 PMCDBG1(OWN,ORM,2, "pmc=%p", pm); 866 KASSERT(pm->pm_owner == po, 867 ("[pmc,%d] owner %p != po %p", __LINE__, pm->pm_owner, po)); 868 869 pmc_release_pmc_descriptor(pm); /* will unlink from the list */ 870 pmc_destroy_pmc_descriptor(pm); 871 } 872 873 KASSERT(po->po_sscount == 0, 874 ("[pmc,%d] SS count not zero", __LINE__)); 875 KASSERT(LIST_EMPTY(&po->po_pmcs), 876 ("[pmc,%d] PMC list not empty", __LINE__)); 877 878 /* de-configure the log file if present */ 879 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 880 pmclog_deconfigure_log(po); 881 } 882 883 /* 884 * Remove an owner process record if all conditions are met. 885 */ 886 static void 887 pmc_maybe_remove_owner(struct pmc_owner *po) 888 { 889 890 PMCDBG1(OWN,OMR,1, "maybe-remove-owner po=%p", po); 891 892 /* 893 * Remove owner record if 894 * - this process does not own any PMCs 895 * - this process has not allocated a system-wide sampling buffer 896 */ 897 if (LIST_EMPTY(&po->po_pmcs) && 898 ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)) { 899 pmc_remove_owner(po); 900 pmc_destroy_owner_descriptor(po); 901 } 902 } 903 904 /* 905 * Add an association between a target process and a PMC. 906 */ 907 static void 908 pmc_link_target_process(struct pmc *pm, struct pmc_process *pp) 909 { 910 struct pmc_target *pt; 911 struct pmc_thread *pt_td __diagused; 912 int ri; 913 914 sx_assert(&pmc_sx, SX_XLOCKED); 915 KASSERT(pm != NULL && pp != NULL, 916 ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp)); 917 KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)), 918 ("[pmc,%d] Attaching a non-process-virtual pmc=%p to pid=%d", 919 __LINE__, pm, pp->pp_proc->p_pid)); 920 KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= ((int) md->pmd_npmc - 1), 921 ("[pmc,%d] Illegal reference count %d for process record %p", 922 __LINE__, pp->pp_refcnt, (void *) pp)); 923 924 ri = PMC_TO_ROWINDEX(pm); 925 926 PMCDBG3(PRC,TLK,1, "link-target pmc=%p ri=%d pmc-process=%p", 927 pm, ri, pp); 928 929 #ifdef HWPMC_DEBUG 930 LIST_FOREACH(pt, &pm->pm_targets, pt_next) { 931 if (pt->pt_process == pp) 932 KASSERT(0, ("[pmc,%d] pp %p already in pmc %p targets", 933 __LINE__, pp, pm)); 934 } 935 #endif 936 pt = malloc(sizeof(struct pmc_target), M_PMC, M_WAITOK | M_ZERO); 937 pt->pt_process = pp; 938 939 LIST_INSERT_HEAD(&pm->pm_targets, pt, pt_next); 940 941 atomic_store_rel_ptr((uintptr_t *)&pp->pp_pmcs[ri].pp_pmc, 942 (uintptr_t)pm); 943 944 if (pm->pm_owner->po_owner == pp->pp_proc) 945 pm->pm_flags |= PMC_F_ATTACHED_TO_OWNER; 946 947 /* 948 * Initialize the per-process values at this row index. 949 */ 950 pp->pp_pmcs[ri].pp_pmcval = PMC_TO_MODE(pm) == PMC_MODE_TS ? 951 pm->pm_sc.pm_reloadcount : 0; 952 pp->pp_refcnt++; 953 954 #ifdef INVARIANTS 955 /* Confirm that the per-thread values at this row index are cleared. */ 956 if (PMC_TO_MODE(pm) == PMC_MODE_TS) { 957 mtx_lock_spin(pp->pp_tdslock); 958 LIST_FOREACH(pt_td, &pp->pp_tds, pt_next) { 959 KASSERT(pt_td->pt_pmcs[ri].pt_pmcval == (pmc_value_t) 0, 960 ("[pmc,%d] pt_pmcval not cleared for pid=%d at " 961 "ri=%d", __LINE__, pp->pp_proc->p_pid, ri)); 962 } 963 mtx_unlock_spin(pp->pp_tdslock); 964 } 965 #endif 966 } 967 968 /* 969 * Removes the association between a target process and a PMC. 970 */ 971 static void 972 pmc_unlink_target_process(struct pmc *pm, struct pmc_process *pp) 973 { 974 int ri; 975 struct proc *p; 976 struct pmc_target *ptgt; 977 struct pmc_thread *pt; 978 979 sx_assert(&pmc_sx, SX_XLOCKED); 980 981 KASSERT(pm != NULL && pp != NULL, 982 ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp)); 983 984 KASSERT(pp->pp_refcnt >= 1 && pp->pp_refcnt <= (int) md->pmd_npmc, 985 ("[pmc,%d] Illegal ref count %d on process record %p", 986 __LINE__, pp->pp_refcnt, (void *) pp)); 987 988 ri = PMC_TO_ROWINDEX(pm); 989 990 PMCDBG3(PRC,TUL,1, "unlink-target pmc=%p ri=%d pmc-process=%p", 991 pm, ri, pp); 992 993 KASSERT(pp->pp_pmcs[ri].pp_pmc == pm, 994 ("[pmc,%d] PMC ri %d mismatch pmc %p pp->[ri] %p", __LINE__, 995 ri, pm, pp->pp_pmcs[ri].pp_pmc)); 996 997 pp->pp_pmcs[ri].pp_pmc = NULL; 998 pp->pp_pmcs[ri].pp_pmcval = (pmc_value_t)0; 999 1000 /* Clear the per-thread values at this row index. */ 1001 if (PMC_TO_MODE(pm) == PMC_MODE_TS) { 1002 mtx_lock_spin(pp->pp_tdslock); 1003 LIST_FOREACH(pt, &pp->pp_tds, pt_next) 1004 pt->pt_pmcs[ri].pt_pmcval = (pmc_value_t)0; 1005 mtx_unlock_spin(pp->pp_tdslock); 1006 } 1007 1008 /* Remove owner-specific flags */ 1009 if (pm->pm_owner->po_owner == pp->pp_proc) { 1010 pp->pp_flags &= ~PMC_PP_ENABLE_MSR_ACCESS; 1011 pm->pm_flags &= ~PMC_F_ATTACHED_TO_OWNER; 1012 } 1013 1014 pp->pp_refcnt--; 1015 1016 /* Remove the target process from the PMC structure */ 1017 LIST_FOREACH(ptgt, &pm->pm_targets, pt_next) 1018 if (ptgt->pt_process == pp) 1019 break; 1020 1021 KASSERT(ptgt != NULL, ("[pmc,%d] process %p (pp: %p) not found " 1022 "in pmc %p", __LINE__, pp->pp_proc, pp, pm)); 1023 1024 LIST_REMOVE(ptgt, pt_next); 1025 free(ptgt, M_PMC); 1026 1027 /* if the PMC now lacks targets, send the owner a SIGIO */ 1028 if (LIST_EMPTY(&pm->pm_targets)) { 1029 p = pm->pm_owner->po_owner; 1030 PROC_LOCK(p); 1031 kern_psignal(p, SIGIO); 1032 PROC_UNLOCK(p); 1033 1034 PMCDBG2(PRC,SIG,2, "signalling proc=%p signal=%d", p, SIGIO); 1035 } 1036 } 1037 1038 /* 1039 * Attach a process to a PMC. 1040 */ 1041 static int 1042 pmc_attach_one_process(struct proc *p, struct pmc *pm) 1043 { 1044 int ri, error; 1045 char *fullpath, *freepath; 1046 struct pmc_process *pp; 1047 1048 sx_assert(&pmc_sx, SX_XLOCKED); 1049 1050 PMCDBG5(PRC,ATT,2, "attach-one pm=%p ri=%d proc=%p (%d, %s)", pm, 1051 PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm); 1052 1053 /* 1054 * Locate the process descriptor corresponding to process 'p', 1055 * allocating space as needed. 1056 * 1057 * Verify that rowindex 'pm_rowindex' is free in the process 1058 * descriptor. 1059 * 1060 * If not, allocate space for a descriptor and link the 1061 * process descriptor and PMC. 1062 */ 1063 ri = PMC_TO_ROWINDEX(pm); 1064 1065 /* mark process as using HWPMCs */ 1066 PROC_LOCK(p); 1067 p->p_flag |= P_HWPMC; 1068 PROC_UNLOCK(p); 1069 1070 if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_ALLOCATE)) == NULL) { 1071 error = ENOMEM; 1072 goto fail; 1073 } 1074 1075 if (pp->pp_pmcs[ri].pp_pmc == pm) {/* already present at slot [ri] */ 1076 error = EEXIST; 1077 goto fail; 1078 } 1079 1080 if (pp->pp_pmcs[ri].pp_pmc != NULL) { 1081 error = EBUSY; 1082 goto fail; 1083 } 1084 1085 pmc_link_target_process(pm, pp); 1086 1087 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) && 1088 (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) == 0) 1089 pm->pm_flags |= PMC_F_NEEDS_LOGFILE; 1090 1091 pm->pm_flags |= PMC_F_ATTACH_DONE; /* mark as attached */ 1092 1093 /* issue an attach event to a configured log file */ 1094 if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE) { 1095 if (p->p_flag & P_KPROC) { 1096 fullpath = kernelname; 1097 freepath = NULL; 1098 } else { 1099 pmc_getfilename(p->p_textvp, &fullpath, &freepath); 1100 pmclog_process_pmcattach(pm, p->p_pid, fullpath); 1101 } 1102 free(freepath, M_TEMP); 1103 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) 1104 pmc_log_process_mappings(pm->pm_owner, p); 1105 } 1106 1107 return (0); 1108 fail: 1109 PROC_LOCK(p); 1110 p->p_flag &= ~P_HWPMC; 1111 PROC_UNLOCK(p); 1112 return (error); 1113 } 1114 1115 /* 1116 * Attach a process and optionally its children 1117 */ 1118 static int 1119 pmc_attach_process(struct proc *p, struct pmc *pm) 1120 { 1121 int error; 1122 struct proc *top; 1123 1124 sx_assert(&pmc_sx, SX_XLOCKED); 1125 1126 PMCDBG5(PRC,ATT,1, "attach pm=%p ri=%d proc=%p (%d, %s)", pm, 1127 PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm); 1128 1129 /* 1130 * If this PMC successfully allowed a GETMSR operation 1131 * in the past, disallow further ATTACHes. 1132 */ 1133 if ((pm->pm_flags & PMC_PP_ENABLE_MSR_ACCESS) != 0) 1134 return (EPERM); 1135 1136 if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0) 1137 return (pmc_attach_one_process(p, pm)); 1138 1139 /* 1140 * Traverse all child processes, attaching them to 1141 * this PMC. 1142 */ 1143 sx_slock(&proctree_lock); 1144 1145 top = p; 1146 for (;;) { 1147 if ((error = pmc_attach_one_process(p, pm)) != 0) 1148 break; 1149 if (!LIST_EMPTY(&p->p_children)) 1150 p = LIST_FIRST(&p->p_children); 1151 else for (;;) { 1152 if (p == top) 1153 goto done; 1154 if (LIST_NEXT(p, p_sibling)) { 1155 p = LIST_NEXT(p, p_sibling); 1156 break; 1157 } 1158 p = p->p_pptr; 1159 } 1160 } 1161 1162 if (error != 0) 1163 (void)pmc_detach_process(top, pm); 1164 1165 done: 1166 sx_sunlock(&proctree_lock); 1167 return (error); 1168 } 1169 1170 /* 1171 * Detach a process from a PMC. If there are no other PMCs tracking 1172 * this process, remove the process structure from its hash table. If 1173 * 'flags' contains PMC_FLAG_REMOVE, then free the process structure. 1174 */ 1175 static int 1176 pmc_detach_one_process(struct proc *p, struct pmc *pm, int flags) 1177 { 1178 int ri; 1179 struct pmc_process *pp; 1180 1181 sx_assert(&pmc_sx, SX_XLOCKED); 1182 1183 KASSERT(pm != NULL, 1184 ("[pmc,%d] null pm pointer", __LINE__)); 1185 1186 ri = PMC_TO_ROWINDEX(pm); 1187 1188 PMCDBG6(PRC,ATT,2, "detach-one pm=%p ri=%d proc=%p (%d, %s) flags=0x%x", 1189 pm, ri, p, p->p_pid, p->p_comm, flags); 1190 1191 if ((pp = pmc_find_process_descriptor(p, 0)) == NULL) 1192 return (ESRCH); 1193 1194 if (pp->pp_pmcs[ri].pp_pmc != pm) 1195 return (EINVAL); 1196 1197 /* 1198 * If this is a process-virtual PMC that is still loaded on the 1199 * hardware of the CPU we are running on (the common case when a 1200 * process detaches a PMC from itself), take it off and drop its 1201 * runcount reference now. The reference is otherwise only 1202 * dropped by the switch-out reclaim, which the scheduler stops 1203 * calling once P_HWPMC is cleared below - leaking it and later 1204 * wedging pmc_wait_for_pmc_idle() at release time. 1205 */ 1206 if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) { 1207 critical_enter(); 1208 pmc_reclaim_pmc_from_cpu(pm, pp, curthread->td_oncpu); 1209 critical_exit(); 1210 } 1211 1212 pmc_unlink_target_process(pm, pp); 1213 1214 /* Issue a detach entry if a log file is configured */ 1215 if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE) 1216 pmclog_process_pmcdetach(pm, p->p_pid); 1217 1218 /* 1219 * If there are no PMCs targeting this process, we remove its 1220 * descriptor from the target hash table and unset the P_HWPMC 1221 * flag in the struct proc. 1222 */ 1223 KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= (int) md->pmd_npmc, 1224 ("[pmc,%d] Illegal refcnt %d for process struct %p", 1225 __LINE__, pp->pp_refcnt, pp)); 1226 1227 if (pp->pp_refcnt != 0) /* still a target of some PMC */ 1228 return (0); 1229 1230 /* 1231 * This detach removed the process' last PMC and we are about to 1232 * clear P_HWPMC. If the detached PMC was its last target and is 1233 * still loaded on other CPUs (e.g. sibling threads of a 1234 * multi-threaded target, or a target running on another CPU), 1235 * drain those references first: the target is already unlinked so 1236 * it cannot reload the PMC, and P_HWPMC is still set so those 1237 * CPUs' switch-out reclaim still runs. Bounded by the target 1238 * threads being scheduled out. 1239 */ 1240 if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)) && 1241 LIST_EMPTY(&pm->pm_targets)) { 1242 while (counter_u64_fetch(pm->pm_runcount) > 0) 1243 pmc_force_context_switch(); 1244 } 1245 1246 pmc_remove_process_descriptor(pp); 1247 1248 if (flags & PMC_FLAG_REMOVE) 1249 pmc_destroy_process_descriptor(pp); 1250 1251 PROC_LOCK(p); 1252 p->p_flag &= ~P_HWPMC; 1253 PROC_UNLOCK(p); 1254 1255 return (0); 1256 } 1257 1258 /* 1259 * Detach a process and optionally its descendants from a PMC. 1260 */ 1261 static int 1262 pmc_detach_process(struct proc *p, struct pmc *pm) 1263 { 1264 struct proc *top; 1265 1266 sx_assert(&pmc_sx, SX_XLOCKED); 1267 1268 PMCDBG5(PRC,ATT,1, "detach pm=%p ri=%d proc=%p (%d, %s)", pm, 1269 PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm); 1270 1271 if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0) 1272 return (pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE)); 1273 1274 /* 1275 * Traverse all children, detaching them from this PMC. We 1276 * ignore errors since we could be detaching a PMC from a 1277 * partially attached proc tree. 1278 */ 1279 sx_slock(&proctree_lock); 1280 1281 top = p; 1282 for (;;) { 1283 (void)pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE); 1284 1285 if (!LIST_EMPTY(&p->p_children)) { 1286 p = LIST_FIRST(&p->p_children); 1287 } else { 1288 for (;;) { 1289 if (p == top) 1290 goto done; 1291 if (LIST_NEXT(p, p_sibling)) { 1292 p = LIST_NEXT(p, p_sibling); 1293 break; 1294 } 1295 p = p->p_pptr; 1296 } 1297 } 1298 } 1299 done: 1300 sx_sunlock(&proctree_lock); 1301 if (LIST_EMPTY(&pm->pm_targets)) 1302 pm->pm_flags &= ~PMC_F_ATTACH_DONE; 1303 1304 return (0); 1305 } 1306 1307 /* 1308 * Handle events after an exec() for a process: 1309 * - Inform log owners of the new exec() event 1310 * - Release any PMCs owned by the process before the exec() 1311 * - Detach PMCs from the target if required 1312 */ 1313 static void 1314 pmc_process_exec(struct thread *td, struct pmckern_procexec *pk) 1315 { 1316 struct pmc *pm; 1317 struct pmc_owner *po; 1318 struct pmc_process *pp; 1319 struct proc *p; 1320 char *fullpath, *freepath; 1321 u_int ri; 1322 bool is_using_hwpmcs; 1323 1324 sx_assert(&pmc_sx, SX_XLOCKED); 1325 1326 p = td->td_proc; 1327 pmc_getfilename(p->p_textvp, &fullpath, &freepath); 1328 1329 PMC_EPOCH_ENTER(); 1330 /* Inform owners of SS mode PMCs of the exec event. */ 1331 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) { 1332 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) != 0) { 1333 pmclog_process_procexec(po, PMC_ID_INVALID, p->p_pid, 1334 pk->pm_baseaddr, pk->pm_dynaddr, fullpath); 1335 } 1336 } 1337 PMC_EPOCH_EXIT(); 1338 1339 PROC_LOCK(p); 1340 is_using_hwpmcs = (p->p_flag & P_HWPMC) != 0; 1341 PROC_UNLOCK(p); 1342 1343 if (!is_using_hwpmcs) { 1344 if (freepath != NULL) 1345 free(freepath, M_TEMP); 1346 return; 1347 } 1348 1349 /* 1350 * PMCs are not inherited across an exec(): remove any PMCs that this 1351 * process is the owner of. 1352 */ 1353 if ((po = pmc_find_owner_descriptor(p)) != NULL) { 1354 pmc_remove_owner(po); 1355 pmc_destroy_owner_descriptor(po); 1356 } 1357 1358 /* 1359 * If the process being exec'ed is not the target of any PMC, we are 1360 * done. 1361 */ 1362 if ((pp = pmc_find_process_descriptor(p, 0)) == NULL) { 1363 if (freepath != NULL) 1364 free(freepath, M_TEMP); 1365 return; 1366 } 1367 1368 /* 1369 * Log the exec event to all monitoring owners. Skip owners who have 1370 * already received the event because they had system sampling PMCs 1371 * active. 1372 */ 1373 for (ri = 0; ri < md->pmd_npmc; ri++) { 1374 if ((pm = pp->pp_pmcs[ri].pp_pmc) == NULL) 1375 continue; 1376 1377 po = pm->pm_owner; 1378 if (po->po_sscount == 0 && 1379 (po->po_flags & PMC_PO_OWNS_LOGFILE) != 0) { 1380 pmclog_process_procexec(po, pm->pm_id, p->p_pid, 1381 pk->pm_baseaddr, pk->pm_dynaddr, fullpath); 1382 } 1383 } 1384 1385 if (freepath != NULL) 1386 free(freepath, M_TEMP); 1387 1388 PMCDBG4(PRC,EXC,1, "exec proc=%p (%d, %s) cred-changed=%d", 1389 p, p->p_pid, p->p_comm, pk->pm_credentialschanged); 1390 1391 if (pk->pm_credentialschanged == 0) /* no change */ 1392 return; 1393 1394 /* 1395 * If the newly exec()'ed process has a different credential 1396 * than before, allow it to be the target of a PMC only if 1397 * the PMC's owner has sufficient privilege. 1398 */ 1399 for (ri = 0; ri < md->pmd_npmc; ri++) { 1400 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) { 1401 struct proc *owner; 1402 struct ucred *cred; 1403 1404 owner = pm->pm_owner->po_owner; 1405 PROC_LOCK(owner); 1406 cred = crhold(owner->p_ucred); 1407 PROC_UNLOCK(owner); 1408 1409 if (priv_check_cred(cred, PRIV_DEBUG_DIFFCRED) != 0) 1410 pmc_detach_one_process(td->td_proc, pm, 1411 PMC_FLAG_NONE); 1412 1413 crfree(cred); 1414 } 1415 } 1416 1417 KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= md->pmd_npmc, 1418 ("[pmc,%d] Illegal ref count %u on pp %p", __LINE__, 1419 pp->pp_refcnt, pp)); 1420 1421 /* 1422 * If this process is no longer the target of any 1423 * PMCs, we can remove the process entry and free 1424 * up space. 1425 */ 1426 if (pp->pp_refcnt == 0) 1427 pmc_destroy_process_descriptor(pp); 1428 } 1429 1430 /* 1431 * Execute optional context-switch batch operations for all classes. 1432 */ 1433 static void 1434 pmc_process_csw_start_all(int cpu) 1435 { 1436 struct pmc_classdep *pcd; 1437 u_int class; 1438 1439 for (class = 0; class < md->pmd_nclass; class++) { 1440 pcd = &md->pmd_classdep[class]; 1441 if (pcd->pcd_start_all != NULL) 1442 (void)pcd->pcd_start_all(cpu); 1443 } 1444 } 1445 1446 static void 1447 pmc_process_csw_stop_all(int cpu) 1448 { 1449 struct pmc_classdep *pcd; 1450 u_int class; 1451 1452 pmc_process_csw_out_prepare(cpu); 1453 1454 for (class = 0; class < md->pmd_nclass; class++) { 1455 pcd = &md->pmd_classdep[class]; 1456 if (pcd->pcd_stop_all != NULL) 1457 (void)pcd->pcd_stop_all(cpu); 1458 } 1459 } 1460 1461 /* 1462 * Mark virtual PMCs stopped, before you close hardware gates. 1463 */ 1464 static void 1465 pmc_process_csw_out_prepare(int cpu) 1466 { 1467 struct pmc *pm; 1468 struct pmc_classdep *pcd; 1469 u_int class; 1470 int adjri; 1471 1472 for (class = 0; class < md->pmd_nclass; class++) { 1473 pcd = &md->pmd_classdep[class]; 1474 if (pcd->pcd_stop_all == NULL) 1475 continue; 1476 for (adjri = 0; adjri < pcd->pcd_num; adjri++) { 1477 pm = NULL; 1478 (void)pcd->pcd_get_config(cpu, adjri, &pm); 1479 if (pm == NULL || 1480 !PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) 1481 continue; 1482 pm->pm_pcpu_state[cpu].pps_cpustate = 0; 1483 } 1484 } 1485 } 1486 1487 /* 1488 * Thread context switch IN. 1489 */ 1490 static void 1491 pmc_process_csw_in(struct thread *td) 1492 { 1493 struct pmc *pm; 1494 struct pmc_classdep *pcd; 1495 struct pmc_cpu *pc; 1496 struct pmc_hw *phw __diagused; 1497 struct pmc_process *pp; 1498 struct pmc_thread *pt; 1499 struct proc *p; 1500 pmc_value_t newvalue; 1501 int cpu; 1502 u_int adjri, ri; 1503 1504 p = td->td_proc; 1505 pt = NULL; 1506 if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE)) == NULL) 1507 return; 1508 1509 KASSERT(pp->pp_proc == td->td_proc, 1510 ("[pmc,%d] not my thread state", __LINE__)); 1511 1512 critical_enter(); /* no preemption from this point */ 1513 1514 cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */ 1515 1516 PMCDBG5(CSW,SWI,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p, 1517 p->p_pid, p->p_comm, pp); 1518 1519 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), 1520 ("[pmc,%d] weird CPU id %d", __LINE__, cpu)); 1521 1522 pc = pmc_pcpu[cpu]; 1523 for (ri = 0; ri < md->pmd_npmc; ri++) { 1524 if ((pm = pp->pp_pmcs[ri].pp_pmc) == NULL) 1525 continue; 1526 1527 KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)), 1528 ("[pmc,%d] Target PMC in non-virtual mode (%d)", 1529 __LINE__, PMC_TO_MODE(pm))); 1530 KASSERT(PMC_TO_ROWINDEX(pm) == ri, 1531 ("[pmc,%d] Row index mismatch pmc %d != ri %d", 1532 __LINE__, PMC_TO_ROWINDEX(pm), ri)); 1533 1534 /* 1535 * Only PMCs that are marked as 'RUNNING' need 1536 * be placed on hardware. 1537 */ 1538 if (pm->pm_state != PMC_STATE_RUNNING) 1539 continue; 1540 1541 KASSERT(counter_u64_fetch(pm->pm_runcount) >= 0, 1542 ("[pmc,%d] pm=%p runcount %ju", __LINE__, pm, 1543 (uintmax_t)counter_u64_fetch(pm->pm_runcount))); 1544 1545 /* increment PMC runcount */ 1546 counter_u64_add(pm->pm_runcount, 1); 1547 1548 /* configure the HWPMC we are going to use. */ 1549 pcd = pmc_ri_to_classdep(md, ri, &adjri); 1550 (void)pcd->pcd_config_pmc(cpu, adjri, pm); 1551 1552 phw = pc->pc_hwpmcs[ri]; 1553 1554 KASSERT(phw != NULL, 1555 ("[pmc,%d] null hw pointer", __LINE__)); 1556 1557 KASSERT(phw->phw_pmc == pm, 1558 ("[pmc,%d] hw->pmc %p != pmc %p", __LINE__, 1559 phw->phw_pmc, pm)); 1560 1561 /* 1562 * Write out saved value and start the PMC. 1563 * 1564 * Sampling PMCs use a per-thread value, while 1565 * counting mode PMCs use a per-pmc value that is 1566 * inherited across descendants. 1567 */ 1568 if (PMC_TO_MODE(pm) == PMC_MODE_TS) { 1569 if (pt == NULL) 1570 pt = pmc_find_thread_descriptor(pp, td, 1571 PMC_FLAG_NONE); 1572 1573 KASSERT(pt != NULL, 1574 ("[pmc,%d] No thread found for td=%p", __LINE__, 1575 td)); 1576 1577 mtx_pool_lock_spin(pmc_mtxpool, pm); 1578 1579 /* 1580 * If we have a thread descriptor, use the per-thread 1581 * counter in the descriptor. If not, we will use 1582 * a per-process counter. 1583 * 1584 * TODO: Remove the per-process "safety net" once 1585 * we have thoroughly tested that we don't hit the 1586 * above assert. 1587 */ 1588 if (pt != NULL) { 1589 if (pt->pt_pmcs[ri].pt_pmcval > 0) 1590 newvalue = pt->pt_pmcs[ri].pt_pmcval; 1591 else 1592 newvalue = pm->pm_sc.pm_reloadcount; 1593 } else { 1594 /* 1595 * Use the saved value calculated after the most 1596 * recent time a thread using the shared counter 1597 * switched out. Reset the saved count in case 1598 * another thread from this process switches in 1599 * before any threads switch out. 1600 */ 1601 newvalue = pp->pp_pmcs[ri].pp_pmcval; 1602 pp->pp_pmcs[ri].pp_pmcval = 1603 pm->pm_sc.pm_reloadcount; 1604 } 1605 mtx_pool_unlock_spin(pmc_mtxpool, pm); 1606 KASSERT(newvalue > 0 && newvalue <= 1607 pm->pm_sc.pm_reloadcount, 1608 ("[pmc,%d] pmcval outside of expected range cpu=%d " 1609 "ri=%d pmcval=%jx pm_reloadcount=%jx", __LINE__, 1610 cpu, ri, newvalue, pm->pm_sc.pm_reloadcount)); 1611 } else { 1612 KASSERT(PMC_TO_MODE(pm) == PMC_MODE_TC, 1613 ("[pmc,%d] illegal mode=%d", __LINE__, 1614 PMC_TO_MODE(pm))); 1615 mtx_pool_lock_spin(pmc_mtxpool, pm); 1616 newvalue = PMC_PCPU_SAVED(cpu, ri) = 1617 pm->pm_gv.pm_savedvalue; 1618 mtx_pool_unlock_spin(pmc_mtxpool, pm); 1619 } 1620 1621 PMCDBG3(CSW,SWI,1,"cpu=%d ri=%d new=%jd", cpu, ri, newvalue); 1622 1623 (void)pcd->pcd_write_pmc(cpu, adjri, pm, newvalue); 1624 1625 /* If a sampling mode PMC, reset stalled state. */ 1626 if (PMC_TO_MODE(pm) == PMC_MODE_TS) 1627 pm->pm_pcpu_state[cpu].pps_stalled = 0; 1628 1629 /* Indicate that we desire this to run. */ 1630 pm->pm_pcpu_state[cpu].pps_cpustate = 1; 1631 1632 /* Start the PMC. */ 1633 (void)pcd->pcd_start_pmc(cpu, adjri, pm); 1634 } 1635 1636 /* 1637 * Perform any other architecture/cpu dependent thread 1638 * switch-in actions. 1639 */ 1640 (void)(*md->pmd_switch_in)(pc, pp); 1641 1642 /* Commit all class PMC start updates at one boundary. */ 1643 pmc_process_csw_start_all(cpu); 1644 1645 critical_exit(); 1646 } 1647 1648 /* 1649 * Compute the change in a counter's value since it was last written. 1650 * The hardware counter is only pcd_width bits wide and wraps around, 1651 * while the value seeded into it may occupy the full 64-bit range, so 1652 * take the difference modulo the counter width. 1653 */ 1654 static pmc_value_t 1655 pmc_delta(const struct pmc_classdep *pcd, pmc_value_t newvalue, 1656 pmc_value_t oldvalue) 1657 { 1658 pmc_value_t delta; 1659 1660 delta = newvalue - oldvalue; 1661 if (pcd->pcd_width < 64) 1662 delta &= ((pmc_value_t)1 << pcd->pcd_width) - 1; 1663 return (delta); 1664 } 1665 1666 /* 1667 * Take a process-virtual PMC off the hardware of 'cpu' if it is 1668 * currently loaded there for process 'pp', accumulating its final 1669 * count and dropping its runcount reference. This is the same reclaim 1670 * that context switch out and process exit perform, factored out so it 1671 * can also run when a target is detached while the PMC may still be 1672 * live: the runcount reference is decremented by the switch-out reclaim 1673 * only, which the scheduler gates on P_HWPMC, so a detach that clears 1674 * P_HWPMC without draining would leak the reference and later wedge 1675 * pmc_wait_for_pmc_idle(). Must be called in a critical section. 1676 */ 1677 static void 1678 pmc_reclaim_pmc_from_cpu(struct pmc *pm, struct pmc_process *pp, int cpu) 1679 { 1680 struct pmc_classdep *pcd; 1681 struct pmc *phw_pm; 1682 pmc_value_t newvalue, tmp; 1683 u_int adjri, ri; 1684 1685 ri = PMC_TO_ROWINDEX(pm); 1686 pcd = pmc_ri_to_classdep(md, ri, &adjri); 1687 1688 /* Only reclaim if this PMC is actually loaded on this CPU. */ 1689 phw_pm = NULL; 1690 (void)(*pcd->pcd_get_config)(cpu, adjri, &phw_pm); 1691 if (phw_pm != pm) 1692 return; 1693 1694 KASSERT(counter_u64_fetch(pm->pm_runcount) > 0, 1695 ("[pmc,%d] pm=%p runcount %ju", __LINE__, pm, 1696 (uintmax_t)counter_u64_fetch(pm->pm_runcount))); 1697 1698 if (pm->pm_pcpu_state[cpu].pps_cpustate) { 1699 pm->pm_pcpu_state[cpu].pps_cpustate = 0; 1700 if (pm->pm_pcpu_state[cpu].pps_stalled == 0) { 1701 (void)pcd->pcd_stop_pmc(cpu, adjri, pm); 1702 1703 if (PMC_TO_MODE(pm) == PMC_MODE_TC) { 1704 (void)pcd->pcd_read_pmc(cpu, adjri, pm, 1705 &newvalue); 1706 tmp = pmc_delta(pcd, newvalue, 1707 PMC_PCPU_SAVED(cpu, ri)); 1708 1709 mtx_pool_lock_spin(pmc_mtxpool, pm); 1710 pm->pm_gv.pm_savedvalue += tmp; 1711 pp->pp_pmcs[ri].pp_pmcval += tmp; 1712 mtx_pool_unlock_spin(pmc_mtxpool, pm); 1713 } 1714 } 1715 } 1716 1717 counter_u64_add(pm->pm_runcount, -1); 1718 (void)pcd->pcd_config_pmc(cpu, adjri, NULL); 1719 } 1720 1721 /* 1722 * Thread context switch OUT. 1723 */ 1724 static void 1725 pmc_process_csw_out(struct thread *td) 1726 { 1727 struct pmc *pm; 1728 struct pmc_classdep *pcd; 1729 struct pmc_cpu *pc; 1730 struct pmc_process *pp; 1731 struct pmc_thread *pt = NULL; 1732 struct proc *p; 1733 pmc_value_t newvalue, tmp; 1734 enum pmc_mode mode; 1735 int cpu; 1736 u_int adjri, ri; 1737 1738 /* 1739 * Locate our process descriptor; this may be NULL if 1740 * this process is exiting and we have already removed 1741 * the process from the target process table. 1742 * 1743 * Note that due to kernel preemption, multiple 1744 * context switches may happen while the process is 1745 * exiting. 1746 * 1747 * Note also that if the target process cannot be 1748 * found we still need to deconfigure any PMCs that 1749 * are currently running on hardware. 1750 */ 1751 p = td->td_proc; 1752 pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE); 1753 1754 critical_enter(); 1755 1756 cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */ 1757 1758 PMCDBG5(CSW,SWO,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p, 1759 p->p_pid, p->p_comm, pp); 1760 1761 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), 1762 ("[pmc,%d weird CPU id %d", __LINE__, cpu)); 1763 1764 pc = pmc_pcpu[cpu]; 1765 1766 /* Close shared class gates before any PMC stop or read. */ 1767 pmc_process_csw_stop_all(cpu); 1768 1769 /* 1770 * When a PMC gets unlinked from a target PMC, it will 1771 * be removed from the target's pp_pmc[] array. 1772 * 1773 * However, on a MP system, the target could have been 1774 * executing on another CPU at the time of the unlink. 1775 * So, at context switch OUT time, we need to look at 1776 * the hardware to determine if a PMC is scheduled on 1777 * it. 1778 */ 1779 for (ri = 0; ri < md->pmd_npmc; ri++) { 1780 pcd = pmc_ri_to_classdep(md, ri, &adjri); 1781 pm = NULL; 1782 (void)(*pcd->pcd_get_config)(cpu, adjri, &pm); 1783 1784 if (pm == NULL) /* nothing at this row index */ 1785 continue; 1786 1787 mode = PMC_TO_MODE(pm); 1788 if (!PMC_IS_VIRTUAL_MODE(mode)) 1789 continue; /* not a process virtual PMC */ 1790 1791 KASSERT(PMC_TO_ROWINDEX(pm) == ri, 1792 ("[pmc,%d] ri mismatch pmc(%d) ri(%d)", 1793 __LINE__, PMC_TO_ROWINDEX(pm), ri)); 1794 1795 /* 1796 * Change desired state, and then stop if not stalled. 1797 * This two-step dance should avoid race conditions where 1798 * an interrupt re-enables the PMC after this code has 1799 * already checked the pm_stalled flag. 1800 */ 1801 pm->pm_pcpu_state[cpu].pps_cpustate = 0; 1802 if (pm->pm_pcpu_state[cpu].pps_stalled == 0) 1803 (void)pcd->pcd_stop_pmc(cpu, adjri, pm); 1804 1805 KASSERT(counter_u64_fetch(pm->pm_runcount) > 0, 1806 ("[pmc,%d] pm=%p runcount %ju", __LINE__, pm, 1807 (uintmax_t)counter_u64_fetch(pm->pm_runcount))); 1808 1809 /* reduce this PMC's runcount */ 1810 counter_u64_add(pm->pm_runcount, -1); 1811 1812 /* 1813 * If this PMC is associated with this process, 1814 * save the reading. 1815 */ 1816 if (pm->pm_state != PMC_STATE_DELETED && pp != NULL && 1817 pp->pp_pmcs[ri].pp_pmc != NULL) { 1818 KASSERT(pm == pp->pp_pmcs[ri].pp_pmc, 1819 ("[pmc,%d] pm %p != pp_pmcs[%d] %p", __LINE__, 1820 pm, ri, pp->pp_pmcs[ri].pp_pmc)); 1821 KASSERT(pp->pp_refcnt > 0, 1822 ("[pmc,%d] pp refcnt = %d", __LINE__, 1823 pp->pp_refcnt)); 1824 1825 (void)pcd->pcd_read_pmc(cpu, adjri, pm, &newvalue); 1826 1827 if (mode == PMC_MODE_TS) { 1828 PMCDBG3(CSW,SWO,1,"cpu=%d ri=%d val=%jd (samp)", 1829 cpu, ri, newvalue); 1830 1831 if (pt == NULL) 1832 pt = pmc_find_thread_descriptor(pp, td, 1833 PMC_FLAG_NONE); 1834 1835 KASSERT(pt != NULL, 1836 ("[pmc,%d] No thread found for td=%p", 1837 __LINE__, td)); 1838 1839 mtx_pool_lock_spin(pmc_mtxpool, pm); 1840 1841 /* 1842 * If we have a thread descriptor, save the 1843 * per-thread counter in the descriptor. If not, 1844 * we will update the per-process counter. 1845 * 1846 * TODO: Remove the per-process "safety net" 1847 * once we have thoroughly tested that we 1848 * don't hit the above assert. 1849 */ 1850 if (pt != NULL) { 1851 pt->pt_pmcs[ri].pt_pmcval = newvalue; 1852 } else { 1853 /* 1854 * For sampling process-virtual PMCs, 1855 * newvalue is the number of events to 1856 * be seen until the next sampling 1857 * interrupt. We can just add the events 1858 * left from this invocation to the 1859 * counter, then adjust in case we 1860 * overflow our range. 1861 * 1862 * (Recall that we reload the counter 1863 * every time we use it.) 1864 */ 1865 pp->pp_pmcs[ri].pp_pmcval += newvalue; 1866 if (pp->pp_pmcs[ri].pp_pmcval > 1867 pm->pm_sc.pm_reloadcount) { 1868 pp->pp_pmcs[ri].pp_pmcval -= 1869 pm->pm_sc.pm_reloadcount; 1870 } 1871 } 1872 mtx_pool_unlock_spin(pmc_mtxpool, pm); 1873 } else { 1874 /* 1875 * For counting process-virtual PMCs, the 1876 * hardware counter's value increases 1877 * monotonically modulo the counter width; 1878 * pmc_delta() recovers the increment even 1879 * when the counter wrapped during the run. 1880 */ 1881 tmp = pmc_delta(pcd, newvalue, 1882 PMC_PCPU_SAVED(cpu, ri)); 1883 1884 PMCDBG3(CSW,SWO,1,"cpu=%d ri=%d tmp=%jd (count)", 1885 cpu, ri, tmp); 1886 1887 mtx_pool_lock_spin(pmc_mtxpool, pm); 1888 pm->pm_gv.pm_savedvalue += tmp; 1889 pp->pp_pmcs[ri].pp_pmcval += tmp; 1890 mtx_pool_unlock_spin(pmc_mtxpool, pm); 1891 1892 if (pm->pm_flags & PMC_F_LOG_PROCCSW) 1893 pmclog_process_proccsw(pm, pp, tmp, td); 1894 } 1895 } 1896 1897 /* Mark hardware as free. */ 1898 (void)pcd->pcd_config_pmc(cpu, adjri, NULL); 1899 } 1900 1901 /* 1902 * Perform any other architecture/cpu dependent thread 1903 * switch out functions. 1904 */ 1905 (void)(*md->pmd_switch_out)(pc, pp); 1906 1907 critical_exit(); 1908 } 1909 1910 /* 1911 * A new thread for a process. 1912 */ 1913 static void 1914 pmc_process_thread_add(struct thread *td) 1915 { 1916 struct pmc_process *pmc; 1917 1918 pmc = pmc_find_process_descriptor(td->td_proc, PMC_FLAG_NONE); 1919 if (pmc != NULL) 1920 pmc_find_thread_descriptor(pmc, td, PMC_FLAG_ALLOCATE); 1921 } 1922 1923 /* 1924 * A thread delete for a process. 1925 */ 1926 static void 1927 pmc_process_thread_delete(struct thread *td) 1928 { 1929 struct pmc_process *pmc; 1930 1931 pmc = pmc_find_process_descriptor(td->td_proc, PMC_FLAG_NONE); 1932 if (pmc != NULL) 1933 pmc_thread_descriptor_pool_free(pmc_find_thread_descriptor(pmc, 1934 td, PMC_FLAG_REMOVE)); 1935 } 1936 1937 /* 1938 * A userret() call for a thread. 1939 */ 1940 static void 1941 pmc_process_thread_userret(struct thread *td) 1942 { 1943 sched_pin(); 1944 pmc_capture_user_callchain(curcpu, PMC_UR, td->td_frame); 1945 sched_unpin(); 1946 } 1947 1948 /* 1949 * A mapping change for a process. 1950 */ 1951 static void 1952 pmc_process_mmap(struct thread *td, struct pmckern_map_in *pkm) 1953 { 1954 const struct pmc *pm; 1955 const struct pmc_process *pp; 1956 struct pmc_owner *po; 1957 char *fullpath, *freepath; 1958 pid_t pid; 1959 int ri; 1960 1961 MPASS(!in_epoch(global_epoch_preempt)); 1962 1963 freepath = fullpath = NULL; 1964 pmc_getfilename((struct vnode *)pkm->pm_file, &fullpath, &freepath); 1965 1966 pid = td->td_proc->p_pid; 1967 1968 PMC_EPOCH_ENTER(); 1969 /* Inform owners of all system-wide sampling PMCs. */ 1970 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) { 1971 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 1972 pmclog_process_map_in(po, pid, pkm->pm_address, 1973 fullpath); 1974 } 1975 1976 if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL) 1977 goto done; 1978 1979 /* 1980 * Inform sampling PMC owners tracking this process. 1981 */ 1982 for (ri = 0; ri < md->pmd_npmc; ri++) { 1983 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL && 1984 PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) { 1985 pmclog_process_map_in(pm->pm_owner, 1986 pid, pkm->pm_address, fullpath); 1987 } 1988 } 1989 1990 done: 1991 if (freepath != NULL) 1992 free(freepath, M_TEMP); 1993 PMC_EPOCH_EXIT(); 1994 } 1995 1996 /* 1997 * Log an munmap request. 1998 */ 1999 static void 2000 pmc_process_munmap(struct thread *td, struct pmckern_map_out *pkm) 2001 { 2002 const struct pmc *pm; 2003 const struct pmc_process *pp; 2004 struct pmc_owner *po; 2005 pid_t pid; 2006 int ri; 2007 2008 pid = td->td_proc->p_pid; 2009 2010 PMC_EPOCH_ENTER(); 2011 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) { 2012 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 2013 pmclog_process_map_out(po, pid, pkm->pm_address, 2014 pkm->pm_address + pkm->pm_size); 2015 } 2016 PMC_EPOCH_EXIT(); 2017 2018 if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL) 2019 return; 2020 2021 for (ri = 0; ri < md->pmd_npmc; ri++) { 2022 pm = pp->pp_pmcs[ri].pp_pmc; 2023 if (pm != NULL && PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) { 2024 pmclog_process_map_out(pm->pm_owner, pid, 2025 pkm->pm_address, pkm->pm_address + pkm->pm_size); 2026 } 2027 } 2028 } 2029 2030 /* 2031 * Log mapping information about the kernel. 2032 */ 2033 static void 2034 pmc_log_kernel_mappings(struct pmc *pm) 2035 { 2036 struct pmc_owner *po; 2037 struct pmckern_map_in *km, *kmbase; 2038 2039 MPASS(in_epoch(global_epoch_preempt) || sx_xlocked(&pmc_sx)); 2040 KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)), 2041 ("[pmc,%d] non-sampling PMC (%p) desires mapping information", 2042 __LINE__, (void *) pm)); 2043 2044 po = pm->pm_owner; 2045 if ((po->po_flags & PMC_PO_INITIAL_MAPPINGS_DONE) != 0) 2046 return; 2047 2048 if (PMC_TO_MODE(pm) == PMC_MODE_SS) 2049 pmc_process_allproc(pm); 2050 2051 /* 2052 * Log the current set of kernel modules. 2053 */ 2054 kmbase = linker_hwpmc_list_objects(); 2055 for (km = kmbase; km->pm_file != NULL; km++) { 2056 PMCDBG2(LOG,REG,1,"%s %p", (char *)km->pm_file, 2057 (void *)km->pm_address); 2058 pmclog_process_map_in(po, (pid_t)-1, km->pm_address, 2059 km->pm_file); 2060 } 2061 free(kmbase, M_LINKER); 2062 2063 po->po_flags |= PMC_PO_INITIAL_MAPPINGS_DONE; 2064 } 2065 2066 /* 2067 * Log the mappings for a single process. 2068 */ 2069 static void 2070 pmc_log_process_mappings(struct pmc_owner *po, struct proc *p) 2071 { 2072 vm_map_t map; 2073 vm_map_entry_t entry; 2074 vm_object_t obj, lobj, tobj; 2075 vm_offset_t last_end; 2076 vm_offset_t start_addr; 2077 struct vnode *vp, *last_vp; 2078 struct vmspace *vm; 2079 char *fullpath, *freepath; 2080 u_int last_timestamp; 2081 2082 last_vp = NULL; 2083 last_end = (vm_offset_t)0; 2084 fullpath = freepath = NULL; 2085 2086 if ((vm = vmspace_acquire_ref(p)) == NULL) 2087 return; 2088 2089 map = &vm->vm_map; 2090 vm_map_lock_read(map); 2091 VM_MAP_ENTRY_FOREACH(entry, map) { 2092 if (entry == NULL) { 2093 PMCDBG2(LOG,OPS,2, "hwpmc: vm_map entry unexpectedly " 2094 "NULL! pid=%d vm_map=%p\n", p->p_pid, map); 2095 break; 2096 } 2097 2098 /* 2099 * We only care about executable map entries. 2100 */ 2101 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0 || 2102 (entry->protection & VM_PROT_EXECUTE) == 0 || 2103 entry->object.vm_object == NULL) { 2104 continue; 2105 } 2106 2107 obj = entry->object.vm_object; 2108 VM_OBJECT_RLOCK(obj); 2109 2110 /* 2111 * Walk the backing_object list to find the base (non-shadowed) 2112 * vm_object. 2113 */ 2114 for (lobj = tobj = obj; tobj != NULL; 2115 tobj = tobj->backing_object) { 2116 if (tobj != obj) 2117 VM_OBJECT_RLOCK(tobj); 2118 if (lobj != obj) 2119 VM_OBJECT_RUNLOCK(lobj); 2120 lobj = tobj; 2121 } 2122 2123 /* 2124 * At this point lobj is the base vm_object and it is locked. 2125 */ 2126 if (lobj == NULL) { 2127 PMCDBG3(LOG,OPS,2, 2128 "hwpmc: lobj unexpectedly NULL! pid=%d " 2129 "vm_map=%p vm_obj=%p\n", p->p_pid, map, obj); 2130 VM_OBJECT_RUNLOCK(obj); 2131 continue; 2132 } 2133 2134 vp = vm_object_vnode(lobj); 2135 if (vp == NULL) { 2136 if (lobj != obj) 2137 VM_OBJECT_RUNLOCK(lobj); 2138 VM_OBJECT_RUNLOCK(obj); 2139 continue; 2140 } 2141 2142 /* 2143 * Skip contiguous regions that point to the same vnode, so we 2144 * don't emit redundant MAP-IN directives. 2145 */ 2146 if (entry->start == last_end && vp == last_vp) { 2147 last_end = entry->end; 2148 if (lobj != obj) 2149 VM_OBJECT_RUNLOCK(lobj); 2150 VM_OBJECT_RUNLOCK(obj); 2151 continue; 2152 } 2153 2154 /* 2155 * We don't want to keep the proc's vm_map or this vm_object 2156 * locked while we walk the pathname, since vn_fullpath() can 2157 * sleep. However, if we drop the lock, it's possible for 2158 * concurrent activity to modify the vm_map list. To protect 2159 * against this, we save the vm_map timestamp before we release 2160 * the lock, and check it after we reacquire the lock below. 2161 */ 2162 start_addr = entry->start; 2163 last_end = entry->end; 2164 last_timestamp = map->timestamp; 2165 vm_map_unlock_read(map); 2166 2167 vref(vp); 2168 if (lobj != obj) 2169 VM_OBJECT_RUNLOCK(lobj); 2170 VM_OBJECT_RUNLOCK(obj); 2171 2172 freepath = NULL; 2173 pmc_getfilename(vp, &fullpath, &freepath); 2174 last_vp = vp; 2175 2176 vrele(vp); 2177 2178 vp = NULL; 2179 pmclog_process_map_in(po, p->p_pid, start_addr, fullpath); 2180 if (freepath != NULL) 2181 free(freepath, M_TEMP); 2182 2183 vm_map_lock_read(map); 2184 2185 /* 2186 * If our saved timestamp doesn't match, this means 2187 * that the vm_map was modified out from under us and 2188 * we can't trust our current "entry" pointer. Do a 2189 * new lookup for this entry. If there is no entry 2190 * for this address range, vm_map_lookup_entry() will 2191 * return the previous one, so we always want to go to 2192 * the next entry on the next loop iteration. 2193 * 2194 * There is an edge condition here that can occur if 2195 * there is no entry at or before this address. In 2196 * this situation, vm_map_lookup_entry returns 2197 * &map->header, which would cause our loop to abort 2198 * without processing the rest of the map. However, 2199 * in practice this will never happen for process 2200 * vm_map. This is because the executable's text 2201 * segment is the first mapping in the proc's address 2202 * space, and this mapping is never removed until the 2203 * process exits, so there will always be a non-header 2204 * entry at or before the requested address for 2205 * vm_map_lookup_entry to return. 2206 */ 2207 if (map->timestamp != last_timestamp) 2208 vm_map_lookup_entry(map, last_end - 1, &entry); 2209 } 2210 2211 vm_map_unlock_read(map); 2212 vmspace_free(vm); 2213 return; 2214 } 2215 2216 /* 2217 * Log mappings for all processes in the system. 2218 */ 2219 static void 2220 pmc_log_all_process_mappings(struct pmc_owner *po) 2221 { 2222 struct proc *p, *top; 2223 2224 sx_assert(&pmc_sx, SX_XLOCKED); 2225 2226 if ((p = pfind(1)) == NULL) 2227 panic("[pmc,%d] Cannot find init", __LINE__); 2228 2229 PROC_UNLOCK(p); 2230 2231 sx_slock(&proctree_lock); 2232 2233 top = p; 2234 for (;;) { 2235 pmc_log_process_mappings(po, p); 2236 if (!LIST_EMPTY(&p->p_children)) 2237 p = LIST_FIRST(&p->p_children); 2238 else for (;;) { 2239 if (p == top) 2240 goto done; 2241 if (LIST_NEXT(p, p_sibling)) { 2242 p = LIST_NEXT(p, p_sibling); 2243 break; 2244 } 2245 p = p->p_pptr; 2246 } 2247 } 2248 done: 2249 sx_sunlock(&proctree_lock); 2250 } 2251 2252 #ifdef HWPMC_DEBUG 2253 const char *pmc_hooknames[] = { 2254 /* these strings correspond to PMC_FN_* in <sys/pmckern.h> */ 2255 "", 2256 "EXEC", 2257 "CSW-IN", 2258 "CSW-OUT", 2259 "SAMPLE", 2260 "UNUSED1", 2261 "UNUSED2", 2262 "MMAP", 2263 "MUNMAP", 2264 "CALLCHAIN-NMI", 2265 "CALLCHAIN-SOFT", 2266 "SOFTSAMPLING", 2267 "THR-CREATE", 2268 "THR-EXIT", 2269 "THR-USERRET", 2270 "THR-CREATE-LOG", 2271 "THR-EXIT-LOG", 2272 "PROC-CREATE-LOG" 2273 }; 2274 #endif 2275 2276 /* 2277 * The 'hook' invoked from the kernel proper 2278 */ 2279 static int 2280 pmc_hook_handler(struct thread *td, int function, void *arg) 2281 { 2282 int cpu; 2283 2284 PMCDBG4(MOD,PMH,1, "hook td=%p func=%d \"%s\" arg=%p", td, function, 2285 pmc_hooknames[function], arg); 2286 2287 switch (function) { 2288 case PMC_FN_PROCESS_EXEC: 2289 pmc_process_exec(td, (struct pmckern_procexec *)arg); 2290 break; 2291 2292 case PMC_FN_CSW_IN: 2293 pmc_process_csw_in(td); 2294 break; 2295 2296 case PMC_FN_CSW_OUT: 2297 pmc_process_csw_out(td); 2298 break; 2299 2300 /* 2301 * Process accumulated PC samples. 2302 * 2303 * This function is expected to be called by hardclock() for 2304 * each CPU that has accumulated PC samples. 2305 * 2306 * This function is to be executed on the CPU whose samples 2307 * are being processed. 2308 */ 2309 case PMC_FN_DO_SAMPLES: 2310 /* 2311 * Clear the cpu specific bit in the CPU mask before 2312 * do the rest of the processing. If the NMI handler 2313 * gets invoked after the "atomic_clear_int()" call 2314 * below but before "pmc_process_samples()" gets 2315 * around to processing the interrupt, then we will 2316 * come back here at the next hardclock() tick (and 2317 * may find nothing to do if "pmc_process_samples()" 2318 * had already processed the interrupt). We don't 2319 * lose the interrupt sample. 2320 */ 2321 DPCPU_SET(pmc_sampled, 0); 2322 cpu = PCPU_GET(cpuid); 2323 pmc_process_samples(cpu, PMC_HR); 2324 pmc_process_samples(cpu, PMC_SR); 2325 pmc_process_samples(cpu, PMC_UR); 2326 break; 2327 2328 case PMC_FN_MMAP: 2329 pmc_process_mmap(td, (struct pmckern_map_in *)arg); 2330 break; 2331 2332 case PMC_FN_MUNMAP: 2333 MPASS(in_epoch(global_epoch_preempt) || sx_xlocked(&pmc_sx)); 2334 pmc_process_munmap(td, (struct pmckern_map_out *)arg); 2335 break; 2336 2337 case PMC_FN_PROC_CREATE_LOG: 2338 pmc_process_proccreate((struct proc *)arg); 2339 break; 2340 2341 case PMC_FN_USER_CALLCHAIN: 2342 /* 2343 * Record a call chain. 2344 */ 2345 KASSERT(td == curthread, ("[pmc,%d] td != curthread", 2346 __LINE__)); 2347 2348 pmc_capture_user_callchain(PCPU_GET(cpuid), PMC_HR, 2349 (struct trapframe *)arg); 2350 2351 KASSERT(td->td_pinned == 1, 2352 ("[pmc,%d] invalid td_pinned value", __LINE__)); 2353 sched_unpin(); /* Can migrate safely now. */ 2354 2355 td->td_pflags &= ~TDP_CALLCHAIN; 2356 break; 2357 2358 case PMC_FN_USER_CALLCHAIN_SOFT: 2359 /* 2360 * Record a call chain. 2361 */ 2362 KASSERT(td == curthread, ("[pmc,%d] td != curthread", 2363 __LINE__)); 2364 2365 cpu = PCPU_GET(cpuid); 2366 pmc_capture_user_callchain(cpu, PMC_SR, 2367 (struct trapframe *) arg); 2368 2369 KASSERT(td->td_pinned == 1, 2370 ("[pmc,%d] invalid td_pinned value", __LINE__)); 2371 2372 sched_unpin(); /* Can migrate safely now. */ 2373 2374 td->td_pflags &= ~TDP_CALLCHAIN; 2375 break; 2376 2377 case PMC_FN_SOFT_SAMPLING: 2378 /* 2379 * Call soft PMC sampling intr. 2380 */ 2381 pmc_soft_intr((struct pmckern_soft *)arg); 2382 break; 2383 2384 case PMC_FN_THR_CREATE: 2385 pmc_process_thread_add(td); 2386 pmc_process_threadcreate(td); 2387 break; 2388 2389 case PMC_FN_THR_CREATE_LOG: 2390 pmc_process_threadcreate(td); 2391 break; 2392 2393 case PMC_FN_THR_EXIT: 2394 KASSERT(td == curthread, ("[pmc,%d] td != curthread", 2395 __LINE__)); 2396 pmc_process_thread_delete(td); 2397 pmc_process_threadexit(td); 2398 break; 2399 case PMC_FN_THR_EXIT_LOG: 2400 pmc_process_threadexit(td); 2401 break; 2402 case PMC_FN_THR_USERRET: 2403 KASSERT(td == curthread, ("[pmc,%d] td != curthread", 2404 __LINE__)); 2405 pmc_process_thread_userret(td); 2406 break; 2407 default: 2408 #ifdef HWPMC_DEBUG 2409 KASSERT(0, ("[pmc,%d] unknown hook %d\n", __LINE__, function)); 2410 #endif 2411 break; 2412 } 2413 2414 return (0); 2415 } 2416 2417 /* 2418 * Allocate a 'struct pmc_owner' descriptor in the owner hash table. 2419 */ 2420 static struct pmc_owner * 2421 pmc_allocate_owner_descriptor(struct proc *p) 2422 { 2423 struct pmc_owner *po; 2424 struct pmc_ownerhash *poh; 2425 uint32_t hindex; 2426 2427 hindex = PMC_HASH_PTR(p, pmc_ownerhashmask); 2428 poh = &pmc_ownerhash[hindex]; 2429 2430 /* Allocate space for N pointers and one descriptor struct. */ 2431 po = malloc(sizeof(struct pmc_owner), M_PMC, M_WAITOK | M_ZERO); 2432 po->po_owner = p; 2433 LIST_INSERT_HEAD(poh, po, po_next); /* insert into hash table */ 2434 2435 TAILQ_INIT(&po->po_logbuffers); 2436 mtx_init(&po->po_mtx, "pmc-owner-mtx", "pmc-per-proc", MTX_SPIN); 2437 2438 PMCDBG4(OWN,ALL,1, "allocate-owner proc=%p (%d, %s) pmc-owner=%p", 2439 p, p->p_pid, p->p_comm, po); 2440 2441 return (po); 2442 } 2443 2444 static void 2445 pmc_destroy_owner_descriptor(struct pmc_owner *po) 2446 { 2447 2448 PMCDBG4(OWN,REL,1, "destroy-owner po=%p proc=%p (%d, %s)", 2449 po, po->po_owner, po->po_owner->p_pid, po->po_owner->p_comm); 2450 2451 mtx_destroy(&po->po_mtx); 2452 free(po, M_PMC); 2453 } 2454 2455 /* 2456 * Allocate a thread descriptor from the free pool. 2457 * 2458 * NOTE: This *can* return NULL. 2459 */ 2460 static struct pmc_thread * 2461 pmc_thread_descriptor_pool_alloc(void) 2462 { 2463 struct pmc_thread *pt; 2464 2465 mtx_lock_spin(&pmc_threadfreelist_mtx); 2466 if ((pt = LIST_FIRST(&pmc_threadfreelist)) != NULL) { 2467 LIST_REMOVE(pt, pt_next); 2468 pmc_threadfreelist_entries--; 2469 } 2470 mtx_unlock_spin(&pmc_threadfreelist_mtx); 2471 2472 return (pt); 2473 } 2474 2475 /* 2476 * Add a thread descriptor to the free pool. We use this instead of free() 2477 * to maintain a cache of free entries. Additionally, we can safely call 2478 * this function when we cannot call free(), such as in a critical section. 2479 */ 2480 static void 2481 pmc_thread_descriptor_pool_free(struct pmc_thread *pt) 2482 { 2483 2484 if (pt == NULL) 2485 return; 2486 2487 memset(pt, 0, THREADENTRY_SIZE); 2488 mtx_lock_spin(&pmc_threadfreelist_mtx); 2489 LIST_INSERT_HEAD(&pmc_threadfreelist, pt, pt_next); 2490 pmc_threadfreelist_entries++; 2491 if (pmc_threadfreelist_entries > pmc_threadfreelist_max) 2492 taskqueue_enqueue(taskqueue_fast, &free_task); 2493 mtx_unlock_spin(&pmc_threadfreelist_mtx); 2494 } 2495 2496 /* 2497 * An asynchronous task to manage the free list. 2498 */ 2499 static void 2500 pmc_thread_descriptor_pool_free_task(void *arg __unused, int pending __unused) 2501 { 2502 struct pmc_thread *pt; 2503 LIST_HEAD(, pmc_thread) tmplist; 2504 int delta; 2505 2506 LIST_INIT(&tmplist); 2507 2508 /* Determine what changes, if any, we need to make. */ 2509 mtx_lock_spin(&pmc_threadfreelist_mtx); 2510 delta = pmc_threadfreelist_entries - pmc_threadfreelist_max; 2511 while (delta > 0 && (pt = LIST_FIRST(&pmc_threadfreelist)) != NULL) { 2512 delta--; 2513 pmc_threadfreelist_entries--; 2514 LIST_REMOVE(pt, pt_next); 2515 LIST_INSERT_HEAD(&tmplist, pt, pt_next); 2516 } 2517 mtx_unlock_spin(&pmc_threadfreelist_mtx); 2518 2519 /* If there are entries to free, free them. */ 2520 while (!LIST_EMPTY(&tmplist)) { 2521 pt = LIST_FIRST(&tmplist); 2522 LIST_REMOVE(pt, pt_next); 2523 free(pt, M_PMC); 2524 } 2525 } 2526 2527 /* 2528 * Drain the thread free pool, freeing all allocations. 2529 */ 2530 static void 2531 pmc_thread_descriptor_pool_drain(void) 2532 { 2533 struct pmc_thread *pt, *next; 2534 2535 LIST_FOREACH_SAFE(pt, &pmc_threadfreelist, pt_next, next) { 2536 LIST_REMOVE(pt, pt_next); 2537 free(pt, M_PMC); 2538 } 2539 } 2540 2541 /* 2542 * find the descriptor corresponding to thread 'td', adding or removing it 2543 * as specified by 'mode'. 2544 * 2545 * Note that this supports additional mode flags in addition to those 2546 * supported by pmc_find_process_descriptor(): 2547 * PMC_FLAG_NOWAIT: Causes the function to not wait for mallocs. 2548 * This makes it safe to call while holding certain other locks. 2549 */ 2550 static struct pmc_thread * 2551 pmc_find_thread_descriptor(struct pmc_process *pp, struct thread *td, 2552 uint32_t mode) 2553 { 2554 struct pmc_thread *pt = NULL, *ptnew = NULL; 2555 int wait_flag; 2556 2557 KASSERT(td != NULL, ("[pmc,%d] called to add NULL td", __LINE__)); 2558 2559 /* 2560 * Pre-allocate memory in the PMC_FLAG_ALLOCATE case prior to 2561 * acquiring the lock. 2562 */ 2563 if ((mode & PMC_FLAG_ALLOCATE) != 0) { 2564 if ((ptnew = pmc_thread_descriptor_pool_alloc()) == NULL) { 2565 wait_flag = M_WAITOK; 2566 if ((mode & PMC_FLAG_NOWAIT) != 0 || 2567 in_epoch(global_epoch_preempt)) 2568 wait_flag = M_NOWAIT; 2569 2570 ptnew = malloc(THREADENTRY_SIZE, M_PMC, 2571 wait_flag | M_ZERO); 2572 } 2573 } 2574 2575 mtx_lock_spin(pp->pp_tdslock); 2576 LIST_FOREACH(pt, &pp->pp_tds, pt_next) { 2577 if (pt->pt_td == td) 2578 break; 2579 } 2580 2581 if ((mode & PMC_FLAG_REMOVE) != 0 && pt != NULL) 2582 LIST_REMOVE(pt, pt_next); 2583 2584 if ((mode & PMC_FLAG_ALLOCATE) != 0 && pt == NULL && ptnew != NULL) { 2585 pt = ptnew; 2586 ptnew = NULL; 2587 pt->pt_td = td; 2588 LIST_INSERT_HEAD(&pp->pp_tds, pt, pt_next); 2589 } 2590 2591 mtx_unlock_spin(pp->pp_tdslock); 2592 2593 if (ptnew != NULL) { 2594 free(ptnew, M_PMC); 2595 } 2596 2597 return (pt); 2598 } 2599 2600 /* 2601 * Try to add thread descriptors for each thread in a process. 2602 */ 2603 static void 2604 pmc_add_thread_descriptors_from_proc(struct proc *p, struct pmc_process *pp) 2605 { 2606 struct pmc_thread **tdlist; 2607 struct thread *curtd; 2608 int i, tdcnt, tdlistsz; 2609 2610 KASSERT(!PROC_LOCKED(p), ("[pmc,%d] proc unexpectedly locked", 2611 __LINE__)); 2612 tdcnt = 32; 2613 restart: 2614 tdlistsz = roundup2(tdcnt, 32); 2615 2616 tdcnt = 0; 2617 tdlist = malloc(sizeof(struct pmc_thread *) * tdlistsz, M_TEMP, 2618 M_WAITOK); 2619 2620 PROC_LOCK(p); 2621 FOREACH_THREAD_IN_PROC(p, curtd) 2622 tdcnt++; 2623 if (tdcnt >= tdlistsz) { 2624 PROC_UNLOCK(p); 2625 free(tdlist, M_TEMP); 2626 goto restart; 2627 } 2628 2629 /* 2630 * Try to add each thread to the list without sleeping. If unable, 2631 * add to a queue to retry after dropping the process lock. 2632 */ 2633 tdcnt = 0; 2634 FOREACH_THREAD_IN_PROC(p, curtd) { 2635 tdlist[tdcnt] = pmc_find_thread_descriptor(pp, curtd, 2636 PMC_FLAG_ALLOCATE | PMC_FLAG_NOWAIT); 2637 if (tdlist[tdcnt] == NULL) { 2638 PROC_UNLOCK(p); 2639 for (i = 0; i <= tdcnt; i++) 2640 pmc_thread_descriptor_pool_free(tdlist[i]); 2641 free(tdlist, M_TEMP); 2642 goto restart; 2643 } 2644 tdcnt++; 2645 } 2646 PROC_UNLOCK(p); 2647 free(tdlist, M_TEMP); 2648 } 2649 2650 /* 2651 * Find the descriptor corresponding to process 'p', adding or removing it 2652 * as specified by 'mode'. 2653 */ 2654 static struct pmc_process * 2655 pmc_find_process_descriptor(struct proc *p, uint32_t mode) 2656 { 2657 struct pmc_process *pp, *ppnew; 2658 struct pmc_processhash *pph; 2659 uint32_t hindex; 2660 2661 hindex = PMC_HASH_PTR(p, pmc_processhashmask); 2662 pph = &pmc_processhash[hindex]; 2663 2664 ppnew = NULL; 2665 2666 /* 2667 * Pre-allocate memory in the PMC_FLAG_ALLOCATE case since we 2668 * cannot call malloc(9) once we hold a spin lock. 2669 */ 2670 if ((mode & PMC_FLAG_ALLOCATE) != 0) 2671 ppnew = malloc(sizeof(struct pmc_process) + md->pmd_npmc * 2672 sizeof(struct pmc_targetstate), M_PMC, M_WAITOK | M_ZERO); 2673 2674 mtx_lock_spin(&pmc_processhash_mtx); 2675 LIST_FOREACH(pp, pph, pp_next) { 2676 if (pp->pp_proc == p) 2677 break; 2678 } 2679 2680 if ((mode & PMC_FLAG_REMOVE) != 0 && pp != NULL) 2681 LIST_REMOVE(pp, pp_next); 2682 2683 if ((mode & PMC_FLAG_ALLOCATE) != 0 && pp == NULL && ppnew != NULL) { 2684 ppnew->pp_proc = p; 2685 LIST_INIT(&ppnew->pp_tds); 2686 ppnew->pp_tdslock = mtx_pool_find(pmc_mtxpool, ppnew); 2687 LIST_INSERT_HEAD(pph, ppnew, pp_next); 2688 mtx_unlock_spin(&pmc_processhash_mtx); 2689 pp = ppnew; 2690 ppnew = NULL; 2691 2692 /* Add thread descriptors for this process' current threads. */ 2693 pmc_add_thread_descriptors_from_proc(p, pp); 2694 } else 2695 mtx_unlock_spin(&pmc_processhash_mtx); 2696 2697 if (ppnew != NULL) 2698 free(ppnew, M_PMC); 2699 return (pp); 2700 } 2701 2702 /* 2703 * Remove a process descriptor from the process hash table. 2704 */ 2705 static void 2706 pmc_remove_process_descriptor(struct pmc_process *pp) 2707 { 2708 KASSERT(pp->pp_refcnt == 0, 2709 ("[pmc,%d] Removing process descriptor %p with count %d", 2710 __LINE__, pp, pp->pp_refcnt)); 2711 2712 mtx_lock_spin(&pmc_processhash_mtx); 2713 LIST_REMOVE(pp, pp_next); 2714 mtx_unlock_spin(&pmc_processhash_mtx); 2715 } 2716 2717 /* 2718 * Destroy a process descriptor. 2719 */ 2720 static void 2721 pmc_destroy_process_descriptor(struct pmc_process *pp) 2722 { 2723 struct pmc_thread *pmc_td; 2724 2725 while ((pmc_td = LIST_FIRST(&pp->pp_tds)) != NULL) { 2726 LIST_REMOVE(pmc_td, pt_next); 2727 pmc_thread_descriptor_pool_free(pmc_td); 2728 } 2729 free(pp, M_PMC); 2730 } 2731 2732 /* 2733 * Find an owner descriptor corresponding to proc 'p'. 2734 */ 2735 static struct pmc_owner * 2736 pmc_find_owner_descriptor(struct proc *p) 2737 { 2738 struct pmc_owner *po; 2739 struct pmc_ownerhash *poh; 2740 uint32_t hindex; 2741 2742 hindex = PMC_HASH_PTR(p, pmc_ownerhashmask); 2743 poh = &pmc_ownerhash[hindex]; 2744 2745 po = NULL; 2746 LIST_FOREACH(po, poh, po_next) { 2747 if (po->po_owner == p) 2748 break; 2749 } 2750 2751 PMCDBG5(OWN,FND,1, "find-owner proc=%p (%d, %s) hindex=0x%x -> " 2752 "pmc-owner=%p", p, p->p_pid, p->p_comm, hindex, po); 2753 2754 return (po); 2755 } 2756 2757 /* 2758 * Allocate a pmc descriptor and initialize its fields. 2759 */ 2760 static struct pmc * 2761 pmc_allocate_pmc_descriptor(void) 2762 { 2763 struct pmc *pmc; 2764 2765 pmc = malloc(sizeof(struct pmc), M_PMC, M_WAITOK | M_ZERO); 2766 pmc->pm_runcount = counter_u64_alloc(M_WAITOK); 2767 pmc->pm_pcpu_state = malloc(sizeof(struct pmc_pcpu_state) * mp_ncpus, 2768 M_PMC, M_WAITOK | M_ZERO); 2769 PMCDBG1(PMC,ALL,1, "allocate-pmc -> pmc=%p", pmc); 2770 2771 return (pmc); 2772 } 2773 2774 /* 2775 * Destroy a pmc descriptor. 2776 */ 2777 static void 2778 pmc_destroy_pmc_descriptor(struct pmc *pm) 2779 { 2780 2781 KASSERT(pm->pm_state == PMC_STATE_DELETED || 2782 pm->pm_state == PMC_STATE_FREE, 2783 ("[pmc,%d] destroying non-deleted PMC", __LINE__)); 2784 KASSERT(LIST_EMPTY(&pm->pm_targets), 2785 ("[pmc,%d] destroying pmc with targets", __LINE__)); 2786 KASSERT(pm->pm_owner == NULL, 2787 ("[pmc,%d] destroying pmc attached to an owner", __LINE__)); 2788 KASSERT(counter_u64_fetch(pm->pm_runcount) == 0, 2789 ("[pmc,%d] pmc has non-zero run count %ju", __LINE__, 2790 (uintmax_t)counter_u64_fetch(pm->pm_runcount))); 2791 2792 counter_u64_free(pm->pm_runcount); 2793 free(pm->pm_pcpu_state, M_PMC); 2794 free(pm, M_PMC); 2795 } 2796 2797 static void 2798 pmc_wait_for_pmc_idle(struct pmc *pm) 2799 { 2800 #ifdef INVARIANTS 2801 volatile int maxloop; 2802 2803 maxloop = 100 * pmc_cpu_max(); 2804 #endif 2805 /* 2806 * Loop (with a forced context switch) till the PMC's runcount 2807 * comes down to zero. 2808 */ 2809 pmclog_flush(pm->pm_owner, 1); 2810 while (counter_u64_fetch(pm->pm_runcount) > 0) { 2811 pmclog_flush(pm->pm_owner, 1); 2812 #ifdef INVARIANTS 2813 maxloop--; 2814 KASSERT(maxloop > 0, 2815 ("[pmc,%d] (ri%d, rc%ju) waiting too long for " 2816 "pmc to be free", __LINE__, PMC_TO_ROWINDEX(pm), 2817 (uintmax_t)counter_u64_fetch(pm->pm_runcount))); 2818 #endif 2819 pmc_force_context_switch(); 2820 } 2821 } 2822 2823 /* 2824 * This function does the following things: 2825 * 2826 * - detaches the PMC from hardware 2827 * - unlinks all target threads that were attached to it 2828 * - removes the PMC from its owner's list 2829 * - destroys the PMC private mutex 2830 * 2831 * Once this function completes, the given pmc pointer can be freed by 2832 * calling pmc_destroy_pmc_descriptor(). 2833 */ 2834 static void 2835 pmc_release_pmc_descriptor(struct pmc *pm) 2836 { 2837 struct pmc_binding pb; 2838 struct pmc_classdep *pcd; 2839 struct pmc_hw *phw __diagused; 2840 struct pmc_owner *po; 2841 struct pmc_process *pp; 2842 struct pmc_target *ptgt, *tmp; 2843 enum pmc_mode mode; 2844 u_int adjri, ri, cpu; 2845 2846 sx_assert(&pmc_sx, SX_XLOCKED); 2847 KASSERT(pm, ("[pmc,%d] null pmc", __LINE__)); 2848 2849 ri = PMC_TO_ROWINDEX(pm); 2850 pcd = pmc_ri_to_classdep(md, ri, &adjri); 2851 mode = PMC_TO_MODE(pm); 2852 2853 PMCDBG3(PMC,REL,1, "release-pmc pmc=%p ri=%d mode=%d", pm, ri, 2854 mode); 2855 2856 /* 2857 * First, we take the PMC off hardware. 2858 */ 2859 cpu = 0; 2860 if (PMC_IS_SYSTEM_MODE(mode)) { 2861 /* 2862 * A system mode PMC runs on a specific CPU. Switch 2863 * to this CPU and turn hardware off. 2864 */ 2865 pmc_save_cpu_binding(&pb); 2866 cpu = PMC_TO_CPU(pm); 2867 pmc_select_cpu(cpu); 2868 2869 /* switch off non-stalled CPUs */ 2870 pm->pm_pcpu_state[cpu].pps_cpustate = 0; 2871 if (pm->pm_state == PMC_STATE_RUNNING && 2872 pm->pm_pcpu_state[cpu].pps_stalled == 0) { 2873 2874 phw = pmc_pcpu[cpu]->pc_hwpmcs[ri]; 2875 2876 KASSERT(phw->phw_pmc == pm, 2877 ("[pmc, %d] pmc ptr ri(%d) hw(%p) pm(%p)", 2878 __LINE__, ri, phw->phw_pmc, pm)); 2879 PMCDBG2(PMC,REL,2, "stopping cpu=%d ri=%d", cpu, ri); 2880 2881 critical_enter(); 2882 (void)pcd->pcd_stop_pmc(cpu, adjri, pm); 2883 critical_exit(); 2884 } 2885 2886 PMCDBG2(PMC,REL,2, "decfg cpu=%d ri=%d", cpu, ri); 2887 2888 critical_enter(); 2889 (void)pcd->pcd_config_pmc(cpu, adjri, NULL); 2890 critical_exit(); 2891 2892 /* adjust the global and process count of SS mode PMCs */ 2893 if (mode == PMC_MODE_SS && pm->pm_state == PMC_STATE_RUNNING) { 2894 po = pm->pm_owner; 2895 po->po_sscount--; 2896 if (po->po_sscount == 0) { 2897 atomic_subtract_rel_int(&pmc_ss_count, 1); 2898 CK_LIST_REMOVE(po, po_ssnext); 2899 epoch_wait_preempt(global_epoch_preempt); 2900 } 2901 } 2902 pm->pm_state = PMC_STATE_DELETED; 2903 2904 pmc_restore_cpu_binding(&pb); 2905 2906 /* 2907 * We could have references to this PMC structure in the 2908 * per-cpu sample queues. Wait for the queue to drain. 2909 */ 2910 pmc_wait_for_pmc_idle(pm); 2911 2912 } else if (PMC_IS_VIRTUAL_MODE(mode)) { 2913 /* 2914 * A virtual PMC could be running on multiple CPUs at a given 2915 * instant. 2916 * 2917 * By marking its state as DELETED, we ensure that this PMC is 2918 * never further scheduled on hardware. 2919 * 2920 * Then we wait till all CPUs are done with this PMC. 2921 */ 2922 pm->pm_state = PMC_STATE_DELETED; 2923 2924 /* Wait for the PMCs runcount to come to zero. */ 2925 pmc_wait_for_pmc_idle(pm); 2926 2927 /* 2928 * At this point the PMC is off all CPUs and cannot be freshly 2929 * scheduled onto a CPU. It is now safe to unlink all targets 2930 * from this PMC. If a process-record's refcount falls to zero, 2931 * we remove it from the hash table. The module-wide SX lock 2932 * protects us from races. 2933 */ 2934 LIST_FOREACH_SAFE(ptgt, &pm->pm_targets, pt_next, tmp) { 2935 pp = ptgt->pt_process; 2936 pmc_unlink_target_process(pm, pp); /* frees 'ptgt' */ 2937 2938 PMCDBG1(PMC,REL,3, "pp->refcnt=%d", pp->pp_refcnt); 2939 2940 /* 2941 * If the target process record shows that no PMCs are 2942 * attached to it, reclaim its space. 2943 */ 2944 if (pp->pp_refcnt == 0) { 2945 pmc_remove_process_descriptor(pp); 2946 pmc_destroy_process_descriptor(pp); 2947 } 2948 } 2949 2950 cpu = curthread->td_oncpu; /* setup cpu for pmd_release() */ 2951 } 2952 2953 /* 2954 * Release any MD resources. 2955 */ 2956 (void)pcd->pcd_release_pmc(cpu, adjri, pm); 2957 2958 /* 2959 * Update row disposition. 2960 */ 2961 if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm))) 2962 PMC_UNMARK_ROW_STANDALONE(ri); 2963 else 2964 PMC_UNMARK_ROW_THREAD(ri); 2965 2966 /* Unlink from the owner's list. */ 2967 if (pm->pm_owner != NULL) { 2968 LIST_REMOVE(pm, pm_next); 2969 pm->pm_owner = NULL; 2970 } 2971 } 2972 2973 /* 2974 * Register an owner and a pmc. 2975 */ 2976 static int 2977 pmc_register_owner(struct proc *p, struct pmc *pmc) 2978 { 2979 struct pmc_owner *po; 2980 2981 sx_assert(&pmc_sx, SX_XLOCKED); 2982 2983 if ((po = pmc_find_owner_descriptor(p)) == NULL) { 2984 if ((po = pmc_allocate_owner_descriptor(p)) == NULL) 2985 return (ENOMEM); 2986 } 2987 2988 KASSERT(pmc->pm_owner == NULL, 2989 ("[pmc,%d] attempting to own an initialized PMC", __LINE__)); 2990 pmc->pm_owner = po; 2991 2992 LIST_INSERT_HEAD(&po->po_pmcs, pmc, pm_next); 2993 2994 PROC_LOCK(p); 2995 p->p_flag |= P_HWPMC; 2996 PROC_UNLOCK(p); 2997 2998 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) != 0) 2999 pmclog_process_pmcallocate(pmc); 3000 3001 PMCDBG2(PMC,REG,1, "register-owner pmc-owner=%p pmc=%p", 3002 po, pmc); 3003 3004 return (0); 3005 } 3006 3007 /* 3008 * Return the current row disposition: 3009 * == 0 => FREE 3010 * > 0 => PROCESS MODE 3011 * < 0 => SYSTEM MODE 3012 */ 3013 int 3014 pmc_getrowdisp(int ri) 3015 { 3016 return (pmc_pmcdisp[ri]); 3017 } 3018 3019 /* 3020 * Check if a PMC at row index 'ri' can be allocated to the current 3021 * process. 3022 * 3023 * Allocation can fail if: 3024 * - the current process is already being profiled by a PMC at index 'ri', 3025 * attached to it via OP_PMCATTACH. 3026 * - the current process has already allocated a PMC at index 'ri' 3027 * via OP_ALLOCATE. 3028 */ 3029 static bool 3030 pmc_can_allocate_rowindex(struct proc *p, unsigned int ri, int cpu) 3031 { 3032 struct pmc *pm; 3033 struct pmc_owner *po; 3034 struct pmc_process *pp; 3035 enum pmc_mode mode; 3036 3037 PMCDBG5(PMC,ALR,1, "can-allocate-rowindex proc=%p (%d, %s) ri=%d " 3038 "cpu=%d", p, p->p_pid, p->p_comm, ri, cpu); 3039 3040 /* 3041 * We shouldn't have already allocated a process-mode PMC at 3042 * row index 'ri'. 3043 * 3044 * We shouldn't have allocated a system-wide PMC on the same 3045 * CPU and same RI. 3046 */ 3047 if ((po = pmc_find_owner_descriptor(p)) != NULL) { 3048 LIST_FOREACH(pm, &po->po_pmcs, pm_next) { 3049 if (PMC_TO_ROWINDEX(pm) == ri) { 3050 mode = PMC_TO_MODE(pm); 3051 if (PMC_IS_VIRTUAL_MODE(mode)) 3052 return (false); 3053 if (PMC_IS_SYSTEM_MODE(mode) && 3054 PMC_TO_CPU(pm) == cpu) 3055 return (false); 3056 } 3057 } 3058 } 3059 3060 /* 3061 * We also shouldn't be the target of any PMC at this index 3062 * since otherwise a PMC_ATTACH to ourselves will fail. 3063 */ 3064 if ((pp = pmc_find_process_descriptor(p, 0)) != NULL) 3065 if (pp->pp_pmcs[ri].pp_pmc != NULL) 3066 return (false); 3067 3068 PMCDBG4(PMC,ALR,2, "can-allocate-rowindex proc=%p (%d, %s) ri=%d ok", 3069 p, p->p_pid, p->p_comm, ri); 3070 return (true); 3071 } 3072 3073 /* 3074 * Check if a given PMC at row index 'ri' can be currently used in 3075 * mode 'mode'. 3076 */ 3077 static bool 3078 pmc_can_allocate_row(int ri, enum pmc_mode mode) 3079 { 3080 enum pmc_disp disp; 3081 3082 sx_assert(&pmc_sx, SX_XLOCKED); 3083 3084 PMCDBG2(PMC,ALR,1, "can-allocate-row ri=%d mode=%d", ri, mode); 3085 3086 if (PMC_IS_SYSTEM_MODE(mode)) 3087 disp = PMC_DISP_STANDALONE; 3088 else 3089 disp = PMC_DISP_THREAD; 3090 3091 /* 3092 * check disposition for PMC row 'ri': 3093 * 3094 * Expected disposition Row-disposition Result 3095 * 3096 * STANDALONE STANDALONE or FREE proceed 3097 * STANDALONE THREAD fail 3098 * THREAD THREAD or FREE proceed 3099 * THREAD STANDALONE fail 3100 */ 3101 if (!PMC_ROW_DISP_IS_FREE(ri) && 3102 !(disp == PMC_DISP_THREAD && PMC_ROW_DISP_IS_THREAD(ri)) && 3103 !(disp == PMC_DISP_STANDALONE && PMC_ROW_DISP_IS_STANDALONE(ri))) 3104 return (false); 3105 3106 /* 3107 * All OK 3108 */ 3109 PMCDBG2(PMC,ALR,2, "can-allocate-row ri=%d mode=%d ok", ri, mode); 3110 return (true); 3111 } 3112 3113 /* 3114 * Find a PMC descriptor with user handle 'pmcid' for thread 'td'. 3115 */ 3116 static struct pmc * 3117 pmc_find_pmc_descriptor_in_process(struct pmc_owner *po, pmc_id_t pmcid) 3118 { 3119 struct pmc *pm; 3120 3121 KASSERT(PMC_ID_TO_ROWINDEX(pmcid) < md->pmd_npmc, 3122 ("[pmc,%d] Illegal pmc index %d (max %d)", __LINE__, 3123 PMC_ID_TO_ROWINDEX(pmcid), md->pmd_npmc)); 3124 3125 LIST_FOREACH(pm, &po->po_pmcs, pm_next) { 3126 if (pm->pm_id == pmcid) 3127 return (pm); 3128 } 3129 3130 return (NULL); 3131 } 3132 3133 static int 3134 pmc_find_pmc(pmc_id_t pmcid, struct pmc **pmc) 3135 { 3136 struct pmc *pm, *opm; 3137 struct pmc_owner *po; 3138 struct pmc_process *pp; 3139 3140 PMCDBG1(PMC,FND,1, "find-pmc id=%d", pmcid); 3141 if (PMC_ID_TO_ROWINDEX(pmcid) >= md->pmd_npmc) 3142 return (EINVAL); 3143 3144 if ((po = pmc_find_owner_descriptor(curthread->td_proc)) == NULL) { 3145 /* 3146 * In case of PMC_F_DESCENDANTS child processes we will not find 3147 * the current process in the owners hash list. Find the owner 3148 * process first and from there lookup the po. 3149 */ 3150 pp = pmc_find_process_descriptor(curthread->td_proc, 3151 PMC_FLAG_NONE); 3152 if (pp == NULL) 3153 return (ESRCH); 3154 opm = pp->pp_pmcs[PMC_ID_TO_ROWINDEX(pmcid)].pp_pmc; 3155 if (opm == NULL) 3156 return (ESRCH); 3157 if ((opm->pm_flags & 3158 (PMC_F_ATTACHED_TO_OWNER | PMC_F_DESCENDANTS)) != 3159 (PMC_F_ATTACHED_TO_OWNER | PMC_F_DESCENDANTS)) 3160 return (ESRCH); 3161 3162 po = opm->pm_owner; 3163 } 3164 3165 if ((pm = pmc_find_pmc_descriptor_in_process(po, pmcid)) == NULL) 3166 return (EINVAL); 3167 3168 PMCDBG2(PMC,FND,2, "find-pmc id=%d -> pmc=%p", pmcid, pm); 3169 3170 *pmc = pm; 3171 return (0); 3172 } 3173 3174 /* 3175 * Start a PMC. 3176 */ 3177 static int 3178 pmc_start(struct pmc *pm) 3179 { 3180 struct pmc_binding pb; 3181 struct pmc_classdep *pcd; 3182 struct pmc_owner *po; 3183 pmc_value_t v; 3184 enum pmc_mode mode; 3185 int adjri, error, cpu, ri; 3186 3187 KASSERT(pm != NULL, 3188 ("[pmc,%d] null pm", __LINE__)); 3189 3190 mode = PMC_TO_MODE(pm); 3191 ri = PMC_TO_ROWINDEX(pm); 3192 pcd = pmc_ri_to_classdep(md, ri, &adjri); 3193 3194 error = 0; 3195 po = pm->pm_owner; 3196 3197 PMCDBG3(PMC,OPS,1, "start pmc=%p mode=%d ri=%d", pm, mode, ri); 3198 3199 po = pm->pm_owner; 3200 3201 /* 3202 * Disallow PMCSTART if a logfile is required but has not been 3203 * configured yet. 3204 */ 3205 if ((pm->pm_flags & PMC_F_NEEDS_LOGFILE) != 0 && 3206 (po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) 3207 return (EDOOFUS); /* programming error */ 3208 3209 /* 3210 * If this is a sampling mode PMC, log mapping information for 3211 * the kernel modules that are currently loaded. 3212 */ 3213 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) 3214 pmc_log_kernel_mappings(pm); 3215 3216 if (PMC_IS_VIRTUAL_MODE(mode)) { 3217 /* 3218 * If a PMCATTACH has never been done on this PMC, 3219 * attach it to its owner process. 3220 */ 3221 if (LIST_EMPTY(&pm->pm_targets)) { 3222 error = (pm->pm_flags & PMC_F_ATTACH_DONE) != 0 ? 3223 ESRCH : pmc_attach_process(po->po_owner, pm); 3224 } 3225 3226 /* 3227 * If the PMC is attached to its owner, then force a context 3228 * switch to ensure that the MD state gets set correctly. 3229 */ 3230 if (error == 0) { 3231 pm->pm_state = PMC_STATE_RUNNING; 3232 if ((pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) != 0) 3233 pmc_force_context_switch(); 3234 } 3235 3236 return (error); 3237 } 3238 3239 /* 3240 * A system-wide PMC. 3241 * 3242 * Add the owner to the global list if this is a system-wide 3243 * sampling PMC. 3244 */ 3245 if (mode == PMC_MODE_SS) { 3246 /* 3247 * Log mapping information for all existing processes in the 3248 * system. Subsequent mappings are logged as they happen; 3249 * see pmc_process_mmap(). 3250 */ 3251 if (po->po_logprocmaps == 0) { 3252 pmc_log_all_process_mappings(po); 3253 po->po_logprocmaps = 1; 3254 } 3255 po->po_sscount++; 3256 if (po->po_sscount == 1) { 3257 atomic_add_rel_int(&pmc_ss_count, 1); 3258 CK_LIST_INSERT_HEAD(&pmc_ss_owners, po, po_ssnext); 3259 PMCDBG1(PMC,OPS,1, "po=%p in global list", po); 3260 } 3261 } 3262 3263 /* 3264 * Move to the CPU associated with this 3265 * PMC, and start the hardware. 3266 */ 3267 pmc_save_cpu_binding(&pb); 3268 cpu = PMC_TO_CPU(pm); 3269 if (!pmc_cpu_is_active(cpu)) { 3270 return (EXTERROR(ENXIO, "PMC CPU %ju is not active for start", 3271 (uintmax_t)cpu)); 3272 } 3273 pmc_select_cpu(cpu); 3274 3275 /* 3276 * global PMCs are configured at allocation time 3277 * so write out the initial value and start the PMC. 3278 */ 3279 pm->pm_state = PMC_STATE_RUNNING; 3280 3281 critical_enter(); 3282 v = PMC_IS_SAMPLING_MODE(mode) ? pm->pm_sc.pm_reloadcount : 3283 pm->pm_sc.pm_initial; 3284 if ((error = pcd->pcd_write_pmc(cpu, adjri, pm, v)) == 0) { 3285 /* If a sampling mode PMC, reset stalled state. */ 3286 if (PMC_IS_SAMPLING_MODE(mode)) 3287 pm->pm_pcpu_state[cpu].pps_stalled = 0; 3288 3289 /* Indicate that we desire this to run. Start it. */ 3290 pm->pm_pcpu_state[cpu].pps_cpustate = 1; 3291 error = pcd->pcd_start_pmc(cpu, adjri, pm); 3292 } 3293 critical_exit(); 3294 3295 pmc_restore_cpu_binding(&pb); 3296 return (error); 3297 } 3298 3299 /* 3300 * Stop a PMC. 3301 */ 3302 static int 3303 pmc_stop(struct pmc *pm) 3304 { 3305 struct pmc_binding pb; 3306 struct pmc_classdep *pcd; 3307 struct pmc_owner *po; 3308 int adjri, cpu, error, ri; 3309 3310 KASSERT(pm != NULL, ("[pmc,%d] null pmc", __LINE__)); 3311 3312 PMCDBG3(PMC,OPS,1, "stop pmc=%p mode=%d ri=%d", pm, PMC_TO_MODE(pm), 3313 PMC_TO_ROWINDEX(pm)); 3314 3315 pm->pm_state = PMC_STATE_STOPPED; 3316 3317 /* 3318 * If the PMC is a virtual mode one, changing the state to non-RUNNING 3319 * is enough to ensure that the PMC never gets scheduled. 3320 * 3321 * If this PMC is current running on a CPU, then it will handled 3322 * correctly at the time its target process is context switched out. 3323 */ 3324 if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) 3325 return (0); 3326 3327 /* 3328 * A system-mode PMC. Move to the CPU associated with this PMC, and 3329 * stop the hardware. We update the 'initial count' so that a 3330 * subsequent PMCSTART will resume counting from the current hardware 3331 * count. 3332 */ 3333 pmc_save_cpu_binding(&pb); 3334 3335 cpu = PMC_TO_CPU(pm); 3336 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), 3337 ("[pmc,%d] illegal cpu=%d", __LINE__, cpu)); 3338 if (!pmc_cpu_is_active(cpu)) { 3339 return (EXTERROR(ENXIO, "PMC CPU %ju is not active for stop", 3340 (uintmax_t)cpu)); 3341 } 3342 pmc_select_cpu(cpu); 3343 3344 ri = PMC_TO_ROWINDEX(pm); 3345 pcd = pmc_ri_to_classdep(md, ri, &adjri); 3346 3347 pm->pm_pcpu_state[cpu].pps_cpustate = 0; 3348 critical_enter(); 3349 if ((error = pcd->pcd_stop_pmc(cpu, adjri, pm)) == 0) { 3350 error = pcd->pcd_read_pmc(cpu, adjri, pm, 3351 &pm->pm_sc.pm_initial); 3352 } 3353 critical_exit(); 3354 3355 pmc_restore_cpu_binding(&pb); 3356 3357 /* Remove this owner from the global list of SS PMC owners. */ 3358 po = pm->pm_owner; 3359 if (PMC_TO_MODE(pm) == PMC_MODE_SS) { 3360 po->po_sscount--; 3361 if (po->po_sscount == 0) { 3362 atomic_subtract_rel_int(&pmc_ss_count, 1); 3363 CK_LIST_REMOVE(po, po_ssnext); 3364 epoch_wait_preempt(global_epoch_preempt); 3365 PMCDBG1(PMC,OPS,2,"po=%p removed from global list", po); 3366 } 3367 } 3368 3369 return (error); 3370 } 3371 3372 static struct pmc_classdep * 3373 pmc_class_to_classdep(enum pmc_class class) 3374 { 3375 int n; 3376 3377 for (n = 0; n < md->pmd_nclass; n++) { 3378 if (md->pmd_classdep[n].pcd_class == class) 3379 return (&md->pmd_classdep[n]); 3380 } 3381 return (NULL); 3382 } 3383 3384 #if defined(HWPMC_DEBUG) && defined(KTR) 3385 static const char *pmc_op_to_name[] = { 3386 #undef __PMC_OP 3387 #define __PMC_OP(N, D) #N , 3388 __PMC_OPS() 3389 NULL 3390 }; 3391 #endif 3392 3393 /* 3394 * The syscall interface 3395 */ 3396 3397 #define PMC_GET_SX_XLOCK(...) do { \ 3398 sx_xlock(&pmc_sx); \ 3399 if (pmc_hook == NULL) { \ 3400 sx_xunlock(&pmc_sx); \ 3401 return __VA_ARGS__; \ 3402 } \ 3403 } while (0) 3404 3405 #define PMC_DOWNGRADE_SX() do { \ 3406 sx_downgrade(&pmc_sx); \ 3407 is_sx_downgraded = true; \ 3408 } while (0) 3409 3410 /* 3411 * Main body of PMC_OP_PMCALLOCATE. 3412 */ 3413 static int 3414 pmc_do_op_pmcallocate(struct thread *td, struct pmc_op_pmcallocate *pa) 3415 { 3416 struct proc *p; 3417 struct pmc *pmc; 3418 struct pmc_binding pb; 3419 struct pmc_classdep *pcd; 3420 struct pmc_hw *phw; 3421 enum pmc_mode mode; 3422 enum pmc_class class; 3423 uint32_t caps, flags; 3424 u_int cpu; 3425 int adjri, n; 3426 int error; 3427 3428 class = pa->pm_class; 3429 caps = pa->pm_caps; 3430 flags = pa->pm_flags; 3431 mode = pa->pm_mode; 3432 cpu = pa->pm_cpu; 3433 3434 p = td->td_proc; 3435 /* Requested mode must exist. */ 3436 if ((mode != PMC_MODE_SS && mode != PMC_MODE_SC && 3437 mode != PMC_MODE_TS && mode != PMC_MODE_TC)) 3438 return (EXTERROR(EINVAL, "Invalid PMC mode %ju", 3439 (uintmax_t)mode)); 3440 3441 /* Requested CPU must be valid. */ 3442 if (cpu != PMC_CPU_ANY && cpu >= pmc_cpu_max()) 3443 return (EXTERROR(EINVAL, "Invalid PMC CPU %ju", 3444 (uintmax_t)cpu)); 3445 3446 /* 3447 * Virtual PMCs should only ask for a default CPU. 3448 * System mode PMCs need to specify a non-default CPU. 3449 */ 3450 if ((PMC_IS_VIRTUAL_MODE(mode) && cpu != PMC_CPU_ANY) || 3451 (PMC_IS_SYSTEM_MODE(mode) && cpu == PMC_CPU_ANY)) { 3452 if (PMC_IS_VIRTUAL_MODE(mode)) { 3453 return (EXTERROR(EINVAL, 3454 "PMC mode %ju requires the default CPU", 3455 (uintmax_t)mode)); 3456 } 3457 return (EXTERROR(EINVAL, 3458 "PMC mode %ju requires an explicit CPU", 3459 (uintmax_t)mode)); 3460 } 3461 3462 /* 3463 * Check that an inactive CPU is not being asked for. 3464 */ 3465 if (PMC_IS_SYSTEM_MODE(mode) && !pmc_cpu_is_active(cpu)) 3466 return (EXTERROR(ENXIO, "PMC CPU %ju is not active", 3467 (uintmax_t)cpu)); 3468 3469 /* 3470 * Refuse an allocation for a system-wide PMC if this process has been 3471 * jailed, or if this process lacks super-user credentials and the 3472 * sysctl tunable 'security.bsd.unprivileged_syspmcs' is zero. 3473 */ 3474 if (PMC_IS_SYSTEM_MODE(mode)) { 3475 if (jailed(td->td_ucred)) 3476 return (EPERM); 3477 if (!pmc_unprivileged_syspmcs) { 3478 error = priv_check(td, PRIV_PMC_SYSTEM); 3479 if (error != 0) 3480 return (error); 3481 } 3482 } 3483 3484 /* 3485 * Look for valid values for 'pm_flags'. 3486 */ 3487 if ((flags & ~(PMC_F_DESCENDANTS | PMC_F_LOG_PROCCSW | 3488 PMC_F_LOG_PROCEXIT | PMC_F_CALLCHAIN | PMC_F_USERCALLCHAIN | 3489 PMC_F_EV_PMU)) != 0) 3490 return (EXTERROR(EINVAL, "Invalid PMC flags %#jx", 3491 (uintmax_t)flags)); 3492 3493 /* PMC_F_USERCALLCHAIN is only valid with PMC_F_CALLCHAIN. */ 3494 if ((flags & (PMC_F_CALLCHAIN | PMC_F_USERCALLCHAIN)) == 3495 PMC_F_USERCALLCHAIN) 3496 return (EXTERROR(EINVAL, 3497 "PMC_F_USERCALLCHAIN requires PMC_F_CALLCHAIN")); 3498 3499 /* PMC_F_USERCALLCHAIN is only valid for sampling mode. */ 3500 if ((flags & PMC_F_USERCALLCHAIN) != 0 && mode != PMC_MODE_TS && 3501 mode != PMC_MODE_SS) 3502 return (EXTERROR(EINVAL, 3503 "PMC_F_USERCALLCHAIN requires sampling mode")); 3504 3505 /* Process logging options are not allowed for system PMCs. */ 3506 if (PMC_IS_SYSTEM_MODE(mode) && 3507 (flags & (PMC_F_LOG_PROCCSW | PMC_F_LOG_PROCEXIT)) != 0) 3508 return (EXTERROR(EINVAL, 3509 "Process logging flags are not valid for system PMCs")); 3510 3511 /* 3512 * All sampling mode PMCs need to be able to interrupt the CPU. 3513 */ 3514 if (PMC_IS_SAMPLING_MODE(mode)) 3515 caps |= PMC_CAP_INTERRUPT; 3516 3517 /* A valid class specifier should have been passed in. */ 3518 pcd = pmc_class_to_classdep(class); 3519 if (pcd == NULL) 3520 return (EXTERROR(EINVAL, "Invalid PMC class %ju", 3521 (uintmax_t)class)); 3522 3523 /* The requested PMC capabilities should be feasible. */ 3524 if ((pcd->pcd_caps & caps) != caps) 3525 return (EXTERROR(EOPNOTSUPP, 3526 "Requested PMC capabilities %#jx are not supported", 3527 (uintmax_t)caps)); 3528 3529 PMCDBG4(PMC,ALL,2, "event=%d caps=0x%x mode=%d cpu=%d", pa->pm_ev, 3530 caps, mode, cpu); 3531 3532 pmc = pmc_allocate_pmc_descriptor(); 3533 pmc->pm_id = PMC_ID_MAKE_ID(cpu, pa->pm_mode, class, PMC_ID_INVALID); 3534 pmc->pm_event = pa->pm_ev; 3535 pmc->pm_state = PMC_STATE_FREE; 3536 pmc->pm_caps = caps; 3537 pmc->pm_flags = flags; 3538 3539 /* XXX set lower bound on sampling for process counters */ 3540 if (PMC_IS_SAMPLING_MODE(mode)) { 3541 /* 3542 * Don't permit requested sample rate to be less than 3543 * pmc_mincount. 3544 */ 3545 if (pa->pm_count < MAX(1, pmc_mincount)) 3546 log(LOG_WARNING, "pmcallocate: passed sample " 3547 "rate %ju - setting to %u\n", 3548 (uintmax_t)pa->pm_count, 3549 MAX(1, pmc_mincount)); 3550 pmc->pm_sc.pm_reloadcount = MAX(MAX(1, pmc_mincount), 3551 pa->pm_count); 3552 } else 3553 pmc->pm_sc.pm_initial = pa->pm_count; 3554 3555 /* switch thread to CPU 'cpu' */ 3556 pmc_save_cpu_binding(&pb); 3557 3558 #define PMC_IS_SHAREABLE_PMC(cpu, n) \ 3559 (pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_state & \ 3560 PMC_PHW_FLAG_IS_SHAREABLE) 3561 #define PMC_IS_UNALLOCATED(cpu, n) \ 3562 (pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_pmc == NULL) 3563 3564 if (PMC_IS_SYSTEM_MODE(mode)) { 3565 pmc_select_cpu(cpu); 3566 for (n = pcd->pcd_ri; n < md->pmd_npmc; n++) { 3567 pcd = pmc_ri_to_classdep(md, n, &adjri); 3568 3569 if (!pmc_can_allocate_row(n, mode) || 3570 !pmc_can_allocate_rowindex(p, n, cpu)) 3571 continue; 3572 if (!PMC_IS_UNALLOCATED(cpu, n) && 3573 !PMC_IS_SHAREABLE_PMC(cpu, n)) 3574 continue; 3575 3576 if (pcd->pcd_allocate_pmc(cpu, adjri, pmc, pa) == 0) { 3577 /* Success. */ 3578 break; 3579 } 3580 } 3581 } else { 3582 /* Process virtual mode */ 3583 for (n = pcd->pcd_ri; n < md->pmd_npmc; n++) { 3584 pcd = pmc_ri_to_classdep(md, n, &adjri); 3585 3586 if (!pmc_can_allocate_row(n, mode) || 3587 !pmc_can_allocate_rowindex(p, n, PMC_CPU_ANY)) 3588 continue; 3589 3590 if (pcd->pcd_allocate_pmc(td->td_oncpu, adjri, pmc, 3591 pa) == 0) { 3592 /* Success. */ 3593 break; 3594 } 3595 } 3596 } 3597 3598 #undef PMC_IS_UNALLOCATED 3599 #undef PMC_IS_SHAREABLE_PMC 3600 3601 pmc_restore_cpu_binding(&pb); 3602 3603 if (n == md->pmd_npmc) { 3604 pmc_destroy_pmc_descriptor(pmc); 3605 /* Preserve a more specific error from the class allocator. */ 3606 if ((td->td_pflags2 & TDP2_EXTERR) != 0) 3607 return (EINVAL); 3608 return (EXTERROR(EINVAL, 3609 "No PMC row accepted the allocation request")); 3610 } 3611 3612 /* Fill in the correct value in the ID field. */ 3613 pmc->pm_id = PMC_ID_MAKE_ID(cpu, mode, class, n); 3614 3615 PMCDBG5(PMC,ALL,2, "ev=%d class=%d mode=%d n=%d -> pmcid=%x", 3616 pmc->pm_event, class, mode, n, pmc->pm_id); 3617 3618 /* Process mode PMCs with logging enabled need log files. */ 3619 if ((pmc->pm_flags & (PMC_F_LOG_PROCEXIT | PMC_F_LOG_PROCCSW)) != 0) 3620 pmc->pm_flags |= PMC_F_NEEDS_LOGFILE; 3621 3622 /* All system mode sampling PMCs require a log file. */ 3623 if (PMC_IS_SAMPLING_MODE(mode) && PMC_IS_SYSTEM_MODE(mode)) 3624 pmc->pm_flags |= PMC_F_NEEDS_LOGFILE; 3625 3626 /* 3627 * Configure global pmc's immediately. 3628 */ 3629 if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pmc))) { 3630 pmc_save_cpu_binding(&pb); 3631 pmc_select_cpu(cpu); 3632 3633 phw = pmc_pcpu[cpu]->pc_hwpmcs[n]; 3634 pcd = pmc_ri_to_classdep(md, n, &adjri); 3635 3636 if ((phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0) { 3637 (void)pcd->pcd_release_pmc(cpu, adjri, pmc); 3638 pmc_destroy_pmc_descriptor(pmc); 3639 pmc_restore_cpu_binding(&pb); 3640 return (EXTERROR(EPERM, 3641 "PMC row %ju on CPU %ju is not enabled", 3642 (uintmax_t)n, (uintmax_t)cpu)); 3643 } 3644 if ((error = pcd->pcd_config_pmc(cpu, adjri, pmc)) != 0) { 3645 (void)pcd->pcd_release_pmc(cpu, adjri, pmc); 3646 pmc_destroy_pmc_descriptor(pmc); 3647 pmc_restore_cpu_binding(&pb); 3648 return (EXTERROR(EPERM, 3649 "PMC configuration failed for row %ju on CPU %ju", 3650 (uintmax_t)n, (uintmax_t)cpu)); 3651 } 3652 3653 pmc_restore_cpu_binding(&pb); 3654 } 3655 3656 pmc->pm_state = PMC_STATE_ALLOCATED; 3657 pmc->pm_class = class; 3658 3659 /* 3660 * Mark row disposition. 3661 */ 3662 if (PMC_IS_SYSTEM_MODE(mode)) 3663 PMC_MARK_ROW_STANDALONE(n); 3664 else 3665 PMC_MARK_ROW_THREAD(n); 3666 3667 /* 3668 * Register this PMC with the current thread as its owner. 3669 */ 3670 error = pmc_register_owner(p, pmc); 3671 if (error != 0) { 3672 pmc_release_pmc_descriptor(pmc); 3673 pmc_destroy_pmc_descriptor(pmc); 3674 return (EXTERROR(error, "Failed to register PMC owner")); 3675 } 3676 3677 /* 3678 * Return the allocated index. 3679 */ 3680 pa->pm_pmcid = pmc->pm_id; 3681 return (0); 3682 } 3683 3684 /* 3685 * Main body of PMC_OP_PMCATTACH. 3686 */ 3687 static int 3688 pmc_do_op_pmcattach(struct thread *td, struct pmc_op_pmcattach a) 3689 { 3690 struct pmc *pm; 3691 struct proc *p; 3692 int error; 3693 3694 sx_assert(&pmc_sx, SX_XLOCKED); 3695 3696 if (a.pm_pid < 0) { 3697 return (EXTERROR(EINVAL, "Invalid PMC attach pid %jd", 3698 (intmax_t)a.pm_pid)); 3699 } else if (a.pm_pid == 0) { 3700 a.pm_pid = td->td_proc->p_pid; 3701 } 3702 3703 error = pmc_find_pmc(a.pm_pmc, &pm); 3704 if (error != 0) 3705 return (error); 3706 3707 if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm))) 3708 return (EXTERROR(EINVAL, 3709 "Cannot attach a system-mode PMC to a process")); 3710 3711 /* PMCs may be (re)attached only when allocated or stopped */ 3712 if (pm->pm_state == PMC_STATE_RUNNING) { 3713 return (EXTERROR(EBUSY, 3714 "PMC must be stopped before attach")); 3715 } else if (pm->pm_state != PMC_STATE_ALLOCATED && 3716 pm->pm_state != PMC_STATE_STOPPED) { 3717 return (EXTERROR(EINVAL, 3718 "PMC state %ju does not allow attach", 3719 (uintmax_t)pm->pm_state)); 3720 } 3721 3722 /* lookup pid */ 3723 if ((p = pfind(a.pm_pid)) == NULL) 3724 return (ESRCH); 3725 3726 /* 3727 * Ignore processes that are working on exiting. 3728 */ 3729 if ((p->p_flag & P_WEXIT) != 0) { 3730 PROC_UNLOCK(p); /* pfind() returns a locked process */ 3731 return (ESRCH); 3732 } 3733 3734 /* 3735 * We are allowed to attach a PMC to a process if we can debug it. 3736 */ 3737 error = p_candebug(curthread, p); 3738 3739 PROC_UNLOCK(p); 3740 3741 if (error == 0) 3742 error = pmc_attach_process(p, pm); 3743 3744 return (error); 3745 } 3746 3747 /* 3748 * Main body of PMC_OP_PMCDETACH. 3749 */ 3750 static int 3751 pmc_do_op_pmcdetach(struct thread *td, struct pmc_op_pmcattach a) 3752 { 3753 struct pmc *pm; 3754 struct proc *p; 3755 int error; 3756 3757 if (a.pm_pid < 0) { 3758 return (EXTERROR(EINVAL, "Invalid PMC detach pid %jd", 3759 (intmax_t)a.pm_pid)); 3760 } else if (a.pm_pid == 0) 3761 a.pm_pid = td->td_proc->p_pid; 3762 3763 error = pmc_find_pmc(a.pm_pmc, &pm); 3764 if (error != 0) 3765 return (error); 3766 3767 if ((p = pfind(a.pm_pid)) == NULL) 3768 return (ESRCH); 3769 3770 /* 3771 * Treat processes that are in the process of exiting as if they were 3772 * not present. 3773 */ 3774 if ((p->p_flag & P_WEXIT) != 0) { 3775 PROC_UNLOCK(p); 3776 return (ESRCH); 3777 } 3778 3779 PROC_UNLOCK(p); /* pfind() returns a locked process */ 3780 3781 if (error == 0) 3782 error = pmc_detach_process(p, pm); 3783 3784 return (error); 3785 } 3786 3787 /* 3788 * Main body of PMC_OP_PMCRELEASE. 3789 */ 3790 static int 3791 pmc_do_op_pmcrelease(pmc_id_t pmcid) 3792 { 3793 struct pmc_owner *po; 3794 struct pmc *pm; 3795 int error; 3796 3797 /* 3798 * Find PMC pointer for the named PMC. 3799 * 3800 * Use pmc_release_pmc_descriptor() to switch off the 3801 * PMC, remove all its target threads, and remove the 3802 * PMC from its owner's list. 3803 * 3804 * Remove the owner record if this is the last PMC 3805 * owned. 3806 * 3807 * Free up space. 3808 */ 3809 error = pmc_find_pmc(pmcid, &pm); 3810 if (error != 0) 3811 return (error); 3812 3813 po = pm->pm_owner; 3814 pmc_release_pmc_descriptor(pm); 3815 pmc_maybe_remove_owner(po); 3816 pmc_destroy_pmc_descriptor(pm); 3817 3818 return (error); 3819 } 3820 3821 /* 3822 * Main body of PMC_OP_PMCRW. 3823 */ 3824 static int 3825 pmc_do_op_pmcrw(const struct pmc_op_pmcrw *prw, pmc_value_t *valp) 3826 { 3827 struct pmc_binding pb; 3828 struct pmc_classdep *pcd; 3829 struct pmc *pm; 3830 u_int cpu, ri, adjri; 3831 int error; 3832 3833 PMCDBG2(PMC,OPS,1, "rw id=%d flags=0x%x", prw->pm_pmcid, prw->pm_flags); 3834 3835 /* Must have at least one flag set. */ 3836 if ((prw->pm_flags & (PMC_F_OLDVALUE | PMC_F_NEWVALUE)) == 0) 3837 return (EXTERROR(EINVAL, 3838 "PMCRW requires PMC_F_OLDVALUE and/or PMC_F_NEWVALUE")); 3839 3840 /* Locate PMC descriptor. */ 3841 error = pmc_find_pmc(prw->pm_pmcid, &pm); 3842 if (error != 0) 3843 return (error); 3844 3845 /* Can't read a PMC that hasn't been started. */ 3846 if (pm->pm_state != PMC_STATE_ALLOCATED && 3847 pm->pm_state != PMC_STATE_STOPPED && 3848 pm->pm_state != PMC_STATE_RUNNING) 3849 return (EXTERROR(EINVAL, 3850 "PMC state %ju does not allow read/write", 3851 (uintmax_t)pm->pm_state)); 3852 3853 /* Writing a new value is allowed only for 'STOPPED' PMCs. */ 3854 if (pm->pm_state == PMC_STATE_RUNNING && 3855 (prw->pm_flags & PMC_F_NEWVALUE) != 0) 3856 return (EXTERROR(EBUSY, 3857 "Cannot write a PMC while it is running")); 3858 3859 if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) { 3860 /* 3861 * If this PMC is attached to its owner (i.e., the process 3862 * requesting this operation) and is running, then attempt to 3863 * get an upto-date reading from hardware for a READ. Writes 3864 * are only allowed when the PMC is stopped, so only update the 3865 * saved value field. 3866 * 3867 * If the PMC is not running, or is not attached to its owner, 3868 * read/write to the savedvalue field. 3869 */ 3870 3871 ri = PMC_TO_ROWINDEX(pm); 3872 pcd = pmc_ri_to_classdep(md, ri, &adjri); 3873 3874 mtx_pool_lock_spin(pmc_mtxpool, pm); 3875 cpu = curthread->td_oncpu; 3876 3877 if ((prw->pm_flags & PMC_F_OLDVALUE) != 0) { 3878 if ((pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) && 3879 (pm->pm_state == PMC_STATE_RUNNING)) { 3880 error = (*pcd->pcd_read_pmc)(cpu, adjri, pm, 3881 valp); 3882 } else { 3883 *valp = pm->pm_gv.pm_savedvalue; 3884 } 3885 } 3886 3887 if ((prw->pm_flags & PMC_F_NEWVALUE) != 0) 3888 pm->pm_gv.pm_savedvalue = prw->pm_value; 3889 3890 mtx_pool_unlock_spin(pmc_mtxpool, pm); 3891 } else { /* System mode PMCs */ 3892 cpu = PMC_TO_CPU(pm); 3893 ri = PMC_TO_ROWINDEX(pm); 3894 pcd = pmc_ri_to_classdep(md, ri, &adjri); 3895 3896 if (!pmc_cpu_is_active(cpu)) 3897 return (EXTERROR(ENXIO, 3898 "PMC CPU %ju is not active for read/write", 3899 (uintmax_t)cpu)); 3900 3901 /* Move this thread to CPU 'cpu'. */ 3902 pmc_save_cpu_binding(&pb); 3903 pmc_select_cpu(cpu); 3904 critical_enter(); 3905 3906 /* Save old value. */ 3907 if ((prw->pm_flags & PMC_F_OLDVALUE) != 0) 3908 error = (*pcd->pcd_read_pmc)(cpu, adjri, pm, valp); 3909 3910 /* Write out new value. */ 3911 if (error == 0 && (prw->pm_flags & PMC_F_NEWVALUE) != 0) 3912 error = (*pcd->pcd_write_pmc)(cpu, adjri, pm, 3913 prw->pm_value); 3914 3915 critical_exit(); 3916 pmc_restore_cpu_binding(&pb); 3917 if (error != 0) 3918 return (error); 3919 } 3920 3921 #ifdef HWPMC_DEBUG 3922 if ((prw->pm_flags & PMC_F_NEWVALUE) != 0) 3923 PMCDBG3(PMC,OPS,2, "rw id=%d new %jx -> old %jx", 3924 ri, prw->pm_value, *valp); 3925 else 3926 PMCDBG2(PMC,OPS,2, "rw id=%d -> old %jx", ri, *valp); 3927 #endif 3928 return (error); 3929 } 3930 3931 static int 3932 pmc_syscall_handler(struct thread *td, void *syscall_args) 3933 { 3934 struct pmc_syscall_args *c; 3935 void *pmclog_proc_handle; 3936 void *arg; 3937 int error, op; 3938 bool is_sx_downgraded; 3939 3940 c = (struct pmc_syscall_args *)syscall_args; 3941 op = c->pmop_code; 3942 arg = c->pmop_data; 3943 3944 /* PMC isn't set up yet */ 3945 if (pmc_hook == NULL) 3946 return (EINVAL); 3947 3948 if (op == PMC_OP_CONFIGURELOG) { 3949 /* 3950 * We cannot create the logging process inside 3951 * pmclog_configure_log() because there is a LOR 3952 * between pmc_sx and process structure locks. 3953 * Instead, pre-create the process and ignite the loop 3954 * if everything is fine, otherwise direct the process 3955 * to exit. 3956 */ 3957 error = pmclog_proc_create(td, &pmclog_proc_handle); 3958 if (error != 0) 3959 goto done_syscall; 3960 } 3961 3962 PMC_GET_SX_XLOCK(ENOSYS); 3963 is_sx_downgraded = false; 3964 PMCDBG3(MOD,PMS,1, "syscall op=%d \"%s\" arg=%p", op, 3965 pmc_op_to_name[op], arg); 3966 3967 error = 0; 3968 counter_u64_add(pmc_stats.pm_syscalls, 1); 3969 3970 switch (op) { 3971 3972 3973 /* 3974 * Configure a log file. 3975 * 3976 * XXX This OP will be reworked. 3977 */ 3978 3979 case PMC_OP_CONFIGURELOG: 3980 { 3981 struct proc *p; 3982 struct pmc *pm; 3983 struct pmc_owner *po; 3984 struct pmc_op_configurelog cl; 3985 3986 if ((error = copyin(arg, &cl, sizeof(cl))) != 0) { 3987 pmclog_proc_ignite(pmclog_proc_handle, NULL); 3988 break; 3989 } 3990 3991 /* No flags currently implemented */ 3992 if (cl.pm_flags != 0) { 3993 pmclog_proc_ignite(pmclog_proc_handle, NULL); 3994 error = EINVAL; 3995 break; 3996 } 3997 3998 /* mark this process as owning a log file */ 3999 p = td->td_proc; 4000 if ((po = pmc_find_owner_descriptor(p)) == NULL) 4001 if ((po = pmc_allocate_owner_descriptor(p)) == NULL) { 4002 pmclog_proc_ignite(pmclog_proc_handle, NULL); 4003 error = ENOMEM; 4004 break; 4005 } 4006 4007 /* 4008 * If a valid fd was passed in, try to configure that, 4009 * otherwise if 'fd' was less than zero and there was 4010 * a log file configured, flush its buffers and 4011 * de-configure it. 4012 */ 4013 if (cl.pm_logfd >= 0) { 4014 error = pmclog_configure_log(md, po, cl.pm_logfd); 4015 pmclog_proc_ignite(pmclog_proc_handle, error == 0 ? 4016 po : NULL); 4017 } else if (po->po_flags & PMC_PO_OWNS_LOGFILE) { 4018 pmclog_proc_ignite(pmclog_proc_handle, NULL); 4019 error = pmclog_close(po); 4020 if (error == 0) { 4021 LIST_FOREACH(pm, &po->po_pmcs, pm_next) 4022 if (pm->pm_flags & PMC_F_NEEDS_LOGFILE && 4023 pm->pm_state == PMC_STATE_RUNNING) 4024 pmc_stop(pm); 4025 error = pmclog_deconfigure_log(po); 4026 } 4027 } else { 4028 pmclog_proc_ignite(pmclog_proc_handle, NULL); 4029 error = EINVAL; 4030 } 4031 } 4032 break; 4033 4034 /* 4035 * Flush a log file. 4036 */ 4037 4038 case PMC_OP_FLUSHLOG: 4039 { 4040 struct pmc_owner *po; 4041 4042 sx_assert(&pmc_sx, SX_XLOCKED); 4043 4044 if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) { 4045 error = EINVAL; 4046 break; 4047 } 4048 4049 error = pmclog_flush(po, 0); 4050 } 4051 break; 4052 4053 /* 4054 * Close a log file. 4055 */ 4056 4057 case PMC_OP_CLOSELOG: 4058 { 4059 struct pmc_owner *po; 4060 4061 sx_assert(&pmc_sx, SX_XLOCKED); 4062 4063 if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) { 4064 error = EINVAL; 4065 break; 4066 } 4067 4068 error = pmclog_close(po); 4069 } 4070 break; 4071 4072 /* 4073 * Retrieve hardware configuration. 4074 */ 4075 4076 case PMC_OP_GETCPUINFO: /* CPU information */ 4077 { 4078 struct pmc_op_getcpuinfo gci; 4079 struct pmc_classinfo *pci; 4080 struct pmc_classdep *pcd; 4081 int cl; 4082 4083 memset(&gci, 0, sizeof(gci)); 4084 gci.pm_cputype = md->pmd_cputype; 4085 gci.pm_ncpu = pmc_cpu_max(); 4086 gci.pm_npmc = md->pmd_npmc; 4087 gci.pm_nclass = md->pmd_nclass; 4088 pci = gci.pm_classes; 4089 pcd = md->pmd_classdep; 4090 for (cl = 0; cl < md->pmd_nclass; cl++, pci++, pcd++) { 4091 pci->pm_caps = pcd->pcd_caps; 4092 pci->pm_class = pcd->pcd_class; 4093 pci->pm_width = pcd->pcd_width; 4094 pci->pm_num = pcd->pcd_num; 4095 } 4096 error = copyout(&gci, arg, sizeof(gci)); 4097 } 4098 break; 4099 4100 /* 4101 * Retrieve soft events list. 4102 */ 4103 case PMC_OP_GETDYNEVENTINFO: 4104 { 4105 enum pmc_class cl; 4106 enum pmc_event ev; 4107 struct pmc_op_getdyneventinfo *gei; 4108 struct pmc_dyn_event_descr dev; 4109 struct pmc_soft *ps; 4110 uint32_t nevent; 4111 4112 sx_assert(&pmc_sx, SX_LOCKED); 4113 4114 gei = (struct pmc_op_getdyneventinfo *) arg; 4115 4116 if ((error = copyin(&gei->pm_class, &cl, sizeof(cl))) != 0) 4117 break; 4118 4119 /* Only SOFT class is dynamic. */ 4120 if (cl != PMC_CLASS_SOFT) { 4121 error = EINVAL; 4122 break; 4123 } 4124 4125 nevent = 0; 4126 for (ev = PMC_EV_SOFT_FIRST; (int)ev <= PMC_EV_SOFT_LAST; ev++) { 4127 ps = pmc_soft_ev_acquire(ev); 4128 if (ps == NULL) 4129 continue; 4130 bcopy(&ps->ps_ev, &dev, sizeof(dev)); 4131 pmc_soft_ev_release(ps); 4132 4133 error = copyout(&dev, 4134 &gei->pm_events[nevent], 4135 sizeof(struct pmc_dyn_event_descr)); 4136 if (error != 0) 4137 break; 4138 nevent++; 4139 } 4140 if (error != 0) 4141 break; 4142 4143 error = copyout(&nevent, &gei->pm_nevent, 4144 sizeof(nevent)); 4145 } 4146 break; 4147 4148 /* 4149 * Get module statistics 4150 */ 4151 4152 case PMC_OP_GETDRIVERSTATS: 4153 { 4154 struct pmc_op_getdriverstats gms; 4155 #define CFETCH(a, b, field) a.field = counter_u64_fetch(b.field) 4156 CFETCH(gms, pmc_stats, pm_intr_ignored); 4157 CFETCH(gms, pmc_stats, pm_intr_processed); 4158 CFETCH(gms, pmc_stats, pm_intr_bufferfull); 4159 CFETCH(gms, pmc_stats, pm_syscalls); 4160 CFETCH(gms, pmc_stats, pm_syscall_errors); 4161 CFETCH(gms, pmc_stats, pm_buffer_requests); 4162 CFETCH(gms, pmc_stats, pm_buffer_requests_failed); 4163 CFETCH(gms, pmc_stats, pm_log_sweeps); 4164 #undef CFETCH 4165 error = copyout(&gms, arg, sizeof(gms)); 4166 } 4167 break; 4168 4169 4170 /* 4171 * Retrieve module version number 4172 */ 4173 4174 case PMC_OP_GETMODULEVERSION: 4175 { 4176 uint32_t cv, modv; 4177 4178 /* retrieve the client's idea of the ABI version */ 4179 if ((error = copyin(arg, &cv, sizeof(uint32_t))) != 0) 4180 break; 4181 /* don't service clients newer than our driver */ 4182 modv = PMC_VERSION; 4183 if ((cv & 0xFFFF0000) > (modv & 0xFFFF0000)) { 4184 error = EPROGMISMATCH; 4185 break; 4186 } 4187 error = copyout(&modv, arg, sizeof(int)); 4188 } 4189 break; 4190 4191 4192 /* 4193 * Retrieve the state of all the PMCs on a given 4194 * CPU. 4195 */ 4196 4197 case PMC_OP_GETPMCINFO: 4198 { 4199 int ari; 4200 struct pmc *pm; 4201 size_t pmcinfo_size; 4202 uint32_t cpu, n, npmc; 4203 struct pmc_owner *po; 4204 struct pmc_binding pb; 4205 struct pmc_classdep *pcd; 4206 struct pmc_info *p, *pmcinfo; 4207 struct pmc_op_getpmcinfo *gpi; 4208 4209 PMC_DOWNGRADE_SX(); 4210 4211 gpi = (struct pmc_op_getpmcinfo *) arg; 4212 4213 if ((error = copyin(&gpi->pm_cpu, &cpu, sizeof(cpu))) != 0) 4214 break; 4215 4216 if (cpu >= pmc_cpu_max()) { 4217 error = EINVAL; 4218 break; 4219 } 4220 4221 if (!pmc_cpu_is_active(cpu)) { 4222 error = ENXIO; 4223 break; 4224 } 4225 4226 /* switch to CPU 'cpu' */ 4227 pmc_save_cpu_binding(&pb); 4228 pmc_select_cpu(cpu); 4229 4230 npmc = md->pmd_npmc; 4231 4232 pmcinfo_size = npmc * sizeof(struct pmc_info); 4233 pmcinfo = malloc(pmcinfo_size, M_PMC, M_WAITOK | M_ZERO); 4234 4235 p = pmcinfo; 4236 4237 for (n = 0; n < md->pmd_npmc; n++, p++) { 4238 4239 pcd = pmc_ri_to_classdep(md, n, &ari); 4240 4241 KASSERT(pcd != NULL, 4242 ("[pmc,%d] null pcd ri=%d", __LINE__, n)); 4243 4244 if ((error = pcd->pcd_describe(cpu, ari, p, &pm)) != 0) 4245 break; 4246 4247 if (PMC_ROW_DISP_IS_STANDALONE(n)) 4248 p->pm_rowdisp = PMC_DISP_STANDALONE; 4249 else if (PMC_ROW_DISP_IS_THREAD(n)) 4250 p->pm_rowdisp = PMC_DISP_THREAD; 4251 else 4252 p->pm_rowdisp = PMC_DISP_FREE; 4253 4254 p->pm_ownerpid = -1; 4255 4256 if (pm == NULL) /* no PMC associated */ 4257 continue; 4258 4259 po = pm->pm_owner; 4260 4261 KASSERT(po->po_owner != NULL, 4262 ("[pmc,%d] pmc_owner had a null proc pointer", 4263 __LINE__)); 4264 4265 p->pm_ownerpid = po->po_owner->p_pid; 4266 p->pm_mode = PMC_TO_MODE(pm); 4267 p->pm_event = pm->pm_event; 4268 p->pm_flags = pm->pm_flags; 4269 4270 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) 4271 p->pm_reloadcount = 4272 pm->pm_sc.pm_reloadcount; 4273 } 4274 4275 pmc_restore_cpu_binding(&pb); 4276 4277 /* now copy out the PMC info collected */ 4278 if (error == 0) 4279 error = copyout(pmcinfo, &gpi->pm_pmcs, pmcinfo_size); 4280 4281 free(pmcinfo, M_PMC); 4282 } 4283 break; 4284 4285 4286 /* 4287 * Set the administrative state of a PMC. I.e. whether 4288 * the PMC is to be used or not. 4289 */ 4290 4291 case PMC_OP_PMCADMIN: 4292 { 4293 int cpu, ri; 4294 enum pmc_state request; 4295 struct pmc_cpu *pc; 4296 struct pmc_hw *phw; 4297 struct pmc_op_pmcadmin pma; 4298 struct pmc_binding pb; 4299 4300 sx_assert(&pmc_sx, SX_XLOCKED); 4301 4302 KASSERT(td == curthread, 4303 ("[pmc,%d] td != curthread", __LINE__)); 4304 4305 error = priv_check(td, PRIV_PMC_MANAGE); 4306 if (error) 4307 break; 4308 4309 if ((error = copyin(arg, &pma, sizeof(pma))) != 0) 4310 break; 4311 4312 cpu = pma.pm_cpu; 4313 4314 if (cpu < 0 || cpu >= (int) pmc_cpu_max()) { 4315 error = EINVAL; 4316 break; 4317 } 4318 4319 if (!pmc_cpu_is_active(cpu)) { 4320 error = ENXIO; 4321 break; 4322 } 4323 4324 request = pma.pm_state; 4325 4326 if (request != PMC_STATE_DISABLED && 4327 request != PMC_STATE_FREE) { 4328 error = EINVAL; 4329 break; 4330 } 4331 4332 ri = pma.pm_pmc; /* pmc id == row index */ 4333 if (ri < 0 || ri >= (int) md->pmd_npmc) { 4334 error = EINVAL; 4335 break; 4336 } 4337 4338 /* 4339 * We can't disable a PMC with a row-index allocated 4340 * for process virtual PMCs. 4341 */ 4342 4343 if (PMC_ROW_DISP_IS_THREAD(ri) && 4344 request == PMC_STATE_DISABLED) { 4345 error = EBUSY; 4346 break; 4347 } 4348 4349 /* 4350 * otherwise, this PMC on this CPU is either free or 4351 * in system-wide mode. 4352 */ 4353 4354 pmc_save_cpu_binding(&pb); 4355 pmc_select_cpu(cpu); 4356 4357 pc = pmc_pcpu[cpu]; 4358 phw = pc->pc_hwpmcs[ri]; 4359 4360 /* 4361 * XXX do we need some kind of 'forced' disable? 4362 */ 4363 4364 if (phw->phw_pmc == NULL) { 4365 if (request == PMC_STATE_DISABLED && 4366 (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED)) { 4367 phw->phw_state &= ~PMC_PHW_FLAG_IS_ENABLED; 4368 PMC_MARK_ROW_STANDALONE(ri); 4369 } else if (request == PMC_STATE_FREE && 4370 (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0) { 4371 phw->phw_state |= PMC_PHW_FLAG_IS_ENABLED; 4372 PMC_UNMARK_ROW_STANDALONE(ri); 4373 } 4374 /* other cases are a no-op */ 4375 } else 4376 error = EBUSY; 4377 4378 pmc_restore_cpu_binding(&pb); 4379 } 4380 break; 4381 4382 4383 /* 4384 * Allocate a PMC. 4385 */ 4386 case PMC_OP_PMCALLOCATE: 4387 { 4388 struct pmc_op_pmcallocate pa; 4389 4390 error = copyin(arg, &pa, sizeof(pa)); 4391 if (error != 0) 4392 break; 4393 4394 error = pmc_do_op_pmcallocate(td, &pa); 4395 if (error != 0) 4396 break; 4397 4398 error = copyout(&pa, arg, sizeof(pa)); 4399 } 4400 break; 4401 4402 /* 4403 * Attach a PMC to a process. 4404 */ 4405 case PMC_OP_PMCATTACH: 4406 { 4407 struct pmc_op_pmcattach a; 4408 4409 error = copyin(arg, &a, sizeof(a)); 4410 if (error != 0) 4411 break; 4412 4413 error = pmc_do_op_pmcattach(td, a); 4414 } 4415 break; 4416 4417 /* 4418 * Detach an attached PMC from a process. 4419 */ 4420 case PMC_OP_PMCDETACH: 4421 { 4422 struct pmc_op_pmcattach a; 4423 4424 error = copyin(arg, &a, sizeof(a)); 4425 if (error != 0) 4426 break; 4427 4428 error = pmc_do_op_pmcdetach(td, a); 4429 } 4430 break; 4431 4432 4433 /* 4434 * Retrieve the MSR number associated with the counter 4435 * 'pmc_id'. This allows processes to directly use RDPMC 4436 * instructions to read their PMCs, without the overhead of a 4437 * system call. 4438 */ 4439 4440 case PMC_OP_PMCGETMSR: 4441 { 4442 int adjri, ri; 4443 struct pmc *pm; 4444 struct pmc_target *pt; 4445 struct pmc_op_getmsr gm; 4446 struct pmc_classdep *pcd; 4447 4448 PMC_DOWNGRADE_SX(); 4449 4450 if ((error = copyin(arg, &gm, sizeof(gm))) != 0) 4451 break; 4452 4453 if ((error = pmc_find_pmc(gm.pm_pmcid, &pm)) != 0) 4454 break; 4455 4456 /* 4457 * The allocated PMC has to be a process virtual PMC, 4458 * i.e., of type MODE_T[CS]. Global PMCs can only be 4459 * read using the PMCREAD operation since they may be 4460 * allocated on a different CPU than the one we could 4461 * be running on at the time of the RDPMC instruction. 4462 * 4463 * The GETMSR operation is not allowed for PMCs that 4464 * are inherited across processes. 4465 */ 4466 4467 if (!PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)) || 4468 (pm->pm_flags & PMC_F_DESCENDANTS)) { 4469 error = EINVAL; 4470 break; 4471 } 4472 4473 /* 4474 * It only makes sense to use a RDPMC (or its 4475 * equivalent instruction on non-x86 architectures) on 4476 * a process that has allocated and attached a PMC to 4477 * itself. Conversely the PMC is only allowed to have 4478 * one process attached to it -- its owner. 4479 */ 4480 4481 if ((pt = LIST_FIRST(&pm->pm_targets)) == NULL || 4482 LIST_NEXT(pt, pt_next) != NULL || 4483 pt->pt_process->pp_proc != pm->pm_owner->po_owner) { 4484 error = EINVAL; 4485 break; 4486 } 4487 4488 ri = PMC_TO_ROWINDEX(pm); 4489 pcd = pmc_ri_to_classdep(md, ri, &adjri); 4490 4491 /* PMC class has no 'GETMSR' support */ 4492 if (pcd->pcd_get_msr == NULL) { 4493 error = ENOSYS; 4494 break; 4495 } 4496 4497 if ((error = (*pcd->pcd_get_msr)(adjri, &gm.pm_msr)) < 0) 4498 break; 4499 4500 if ((error = copyout(&gm, arg, sizeof(gm))) < 0) 4501 break; 4502 4503 /* 4504 * Mark our process as using MSRs. Update machine 4505 * state using a forced context switch. 4506 */ 4507 4508 pt->pt_process->pp_flags |= PMC_PP_ENABLE_MSR_ACCESS; 4509 pmc_force_context_switch(); 4510 4511 } 4512 break; 4513 4514 /* 4515 * Release an allocated PMC. 4516 */ 4517 case PMC_OP_PMCRELEASE: 4518 { 4519 struct pmc_op_simple sp; 4520 4521 error = copyin(arg, &sp, sizeof(sp)); 4522 if (error != 0) 4523 break; 4524 4525 error = pmc_do_op_pmcrelease(sp.pm_pmcid); 4526 } 4527 break; 4528 4529 /* 4530 * Read and/or write a PMC. 4531 */ 4532 case PMC_OP_PMCRW: 4533 { 4534 struct pmc_op_pmcrw prw; 4535 struct pmc_op_pmcrw *pprw; 4536 pmc_value_t oldvalue; 4537 4538 PMC_DOWNGRADE_SX(); 4539 4540 error = copyin(arg, &prw, sizeof(prw)); 4541 if (error != 0) 4542 break; 4543 4544 error = pmc_do_op_pmcrw(&prw, &oldvalue); 4545 if (error != 0) 4546 break; 4547 4548 /* Return old value if requested. */ 4549 if ((prw.pm_flags & PMC_F_OLDVALUE) != 0) { 4550 pprw = arg; 4551 error = copyout(&oldvalue, &pprw->pm_value, 4552 sizeof(prw.pm_value)); 4553 } 4554 } 4555 break; 4556 4557 4558 /* 4559 * Set the sampling rate for a sampling mode PMC and the 4560 * initial count for a counting mode PMC. 4561 */ 4562 4563 case PMC_OP_PMCSETCOUNT: 4564 { 4565 struct pmc *pm; 4566 struct pmc_op_pmcsetcount sc; 4567 4568 PMC_DOWNGRADE_SX(); 4569 4570 if ((error = copyin(arg, &sc, sizeof(sc))) != 0) 4571 break; 4572 4573 if ((error = pmc_find_pmc(sc.pm_pmcid, &pm)) != 0) 4574 break; 4575 4576 if (pm->pm_state == PMC_STATE_RUNNING) { 4577 error = EBUSY; 4578 break; 4579 } 4580 4581 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) { 4582 /* 4583 * Don't permit requested sample rate to be 4584 * less than pmc_mincount. 4585 */ 4586 if (sc.pm_count < MAX(1, pmc_mincount)) 4587 log(LOG_WARNING, "pmcsetcount: passed sample " 4588 "rate %ju - setting to %u\n", 4589 (uintmax_t)sc.pm_count, 4590 MAX(1, pmc_mincount)); 4591 pm->pm_sc.pm_reloadcount = MAX(MAX(1, pmc_mincount), 4592 sc.pm_count); 4593 } else 4594 pm->pm_sc.pm_initial = sc.pm_count; 4595 } 4596 break; 4597 4598 4599 /* 4600 * Start a PMC. 4601 */ 4602 4603 case PMC_OP_PMCSTART: 4604 { 4605 pmc_id_t pmcid; 4606 struct pmc *pm; 4607 struct pmc_op_simple sp; 4608 4609 sx_assert(&pmc_sx, SX_XLOCKED); 4610 4611 if ((error = copyin(arg, &sp, sizeof(sp))) != 0) 4612 break; 4613 4614 pmcid = sp.pm_pmcid; 4615 4616 if ((error = pmc_find_pmc(pmcid, &pm)) != 0) 4617 break; 4618 4619 KASSERT(pmcid == pm->pm_id, 4620 ("[pmc,%d] pmcid %x != id %x", __LINE__, 4621 pm->pm_id, pmcid)); 4622 4623 if (pm->pm_state == PMC_STATE_RUNNING) /* already running */ 4624 break; 4625 else if (pm->pm_state != PMC_STATE_STOPPED && 4626 pm->pm_state != PMC_STATE_ALLOCATED) { 4627 error = EINVAL; 4628 break; 4629 } 4630 4631 error = pmc_start(pm); 4632 } 4633 break; 4634 4635 4636 /* 4637 * Stop a PMC. 4638 */ 4639 4640 case PMC_OP_PMCSTOP: 4641 { 4642 pmc_id_t pmcid; 4643 struct pmc *pm; 4644 struct pmc_op_simple sp; 4645 4646 PMC_DOWNGRADE_SX(); 4647 4648 if ((error = copyin(arg, &sp, sizeof(sp))) != 0) 4649 break; 4650 4651 pmcid = sp.pm_pmcid; 4652 4653 /* 4654 * Mark the PMC as inactive and invoke the MD stop 4655 * routines if needed. 4656 */ 4657 4658 if ((error = pmc_find_pmc(pmcid, &pm)) != 0) 4659 break; 4660 4661 KASSERT(pmcid == pm->pm_id, 4662 ("[pmc,%d] pmc id %x != pmcid %x", __LINE__, 4663 pm->pm_id, pmcid)); 4664 4665 if (pm->pm_state == PMC_STATE_STOPPED) /* already stopped */ 4666 break; 4667 else if (pm->pm_state != PMC_STATE_RUNNING) { 4668 error = EINVAL; 4669 break; 4670 } 4671 4672 error = pmc_stop(pm); 4673 } 4674 break; 4675 4676 4677 /* 4678 * Write a user supplied value to the log file. 4679 */ 4680 4681 case PMC_OP_WRITELOG: 4682 { 4683 struct pmc_op_writelog wl; 4684 struct pmc_owner *po; 4685 4686 PMC_DOWNGRADE_SX(); 4687 4688 if ((error = copyin(arg, &wl, sizeof(wl))) != 0) 4689 break; 4690 4691 if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) { 4692 error = EINVAL; 4693 break; 4694 } 4695 4696 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) { 4697 error = EINVAL; 4698 break; 4699 } 4700 4701 error = pmclog_process_userlog(po, &wl); 4702 } 4703 break; 4704 4705 /* 4706 * Get the PMC capabilities 4707 */ 4708 4709 case PMC_OP_GETCAPS: 4710 { 4711 struct pmc_op_caps c; 4712 struct pmc *pm; 4713 struct pmc_classdep *pcd; 4714 pmc_id_t pmcid; 4715 int adjri, ri; 4716 4717 PMC_DOWNGRADE_SX(); 4718 4719 if ((error = copyin(arg, &c, sizeof(c))) != 0) 4720 break; 4721 4722 pmcid = c.pm_pmcid; 4723 4724 if ((error = pmc_find_pmc(pmcid, &pm)) != 0) 4725 break; 4726 4727 KASSERT(pmcid == pm->pm_id, 4728 ("[pmc,%d] pmc id %x != pmcid %x", __LINE__, 4729 pm->pm_id, pmcid)); 4730 4731 ri = PMC_TO_ROWINDEX(pm); 4732 pcd = pmc_ri_to_classdep(md, ri, &adjri); 4733 4734 /* 4735 * If PMC class has no GETCAPS return the class capabilities 4736 * otherwise get the per counter capabilities. 4737 */ 4738 if (pcd->pcd_get_caps == NULL) { 4739 c.pm_caps = pcd->pcd_caps; 4740 } else { 4741 error = (*pcd->pcd_get_caps)(adjri, &c.pm_caps); 4742 if (error < 0) 4743 break; 4744 } 4745 4746 if ((error = copyout(&c, arg, sizeof(c))) < 0) 4747 break; 4748 } 4749 break; 4750 4751 default: 4752 error = EINVAL; 4753 break; 4754 } 4755 4756 if (is_sx_downgraded) 4757 sx_sunlock(&pmc_sx); 4758 else 4759 sx_xunlock(&pmc_sx); 4760 done_syscall: 4761 if (error) 4762 counter_u64_add(pmc_stats.pm_syscall_errors, 1); 4763 4764 return (error); 4765 } 4766 4767 /* 4768 * Helper functions 4769 */ 4770 4771 /* 4772 * Mark the thread as needing callchain capture and post an AST. The 4773 * actual callchain capture will be done in a context where it is safe 4774 * to take page faults. 4775 */ 4776 static void 4777 pmc_post_callchain_callback(void) 4778 { 4779 struct thread *td; 4780 4781 td = curthread; 4782 4783 /* 4784 * If there is multiple PMCs for the same interrupt ignore new post 4785 */ 4786 if ((td->td_pflags & TDP_CALLCHAIN) != 0) 4787 return; 4788 4789 /* 4790 * Mark this thread as needing callchain capture. 4791 * `td->td_pflags' will be safe to touch because this thread 4792 * was in user space when it was interrupted. 4793 */ 4794 td->td_pflags |= TDP_CALLCHAIN; 4795 4796 /* 4797 * Don't let this thread migrate between CPUs until callchain 4798 * capture completes. 4799 */ 4800 sched_pin(); 4801 4802 return; 4803 } 4804 4805 static void 4806 pmc_multipart_add(struct pmc_sample *ps, int type, int length) 4807 { 4808 int i; 4809 uint8_t *hdr; 4810 4811 MPASS(ps->ps_pc != NULL); 4812 MPASS(ps->ps_nsamples_actual != 0); 4813 4814 hdr = (uint8_t *)ps->ps_pc; 4815 4816 for (i = 0; i < PMC_MULTIPART_HEADER_ENTRIES; i++) { 4817 if (hdr[2 * i] == PMC_CC_MULTIPART_NONE) { 4818 hdr[2 * i] = type; 4819 hdr[2 * i + 1] = length; 4820 ps->ps_nsamples_actual += length; 4821 return; 4822 } 4823 } 4824 4825 KASSERT(false, ("Too many parts in the multipart header!")); 4826 } 4827 4828 static void 4829 pmc_multipart_copydata(struct pmc_sample *ps, struct pmc_multipart *mp) 4830 { 4831 int i, scale; 4832 uint64_t *ps_pc; 4833 4834 MPASS(ps->ps_pc != NULL); 4835 MPASS(ps->ps_nsamples_actual != 0); 4836 4837 ps_pc = (uint64_t *)ps->ps_pc; 4838 4839 for (i = 0; i < mp->pl_length; i++) 4840 ps_pc[i + 1] = mp->pl_mpdata[i]; 4841 4842 scale = sizeof(uint64_t) / sizeof(uintptr_t); 4843 pmc_multipart_add(ps, mp->pl_type, scale * mp->pl_length); 4844 } 4845 4846 /* 4847 * Find a free slot in the per-cpu array of samples and capture the 4848 * current callchain there. If a sample was successfully added, a bit 4849 * is set in mask 'pmc_cpumask' denoting that the DO_SAMPLES hook 4850 * needs to be invoked from the clock handler. 4851 * 4852 * This function is meant to be called from an NMI handler. It cannot 4853 * use any of the locking primitives supplied by the OS. 4854 */ 4855 static int 4856 pmc_add_sample(ring_type_t ring, struct pmc *pm, struct trapframe *tf, 4857 struct pmc_multipart *mp) 4858 { 4859 struct pmc_sample *ps; 4860 struct pmc_samplebuffer *psb; 4861 struct thread *td; 4862 int error, cpu, callchaindepth; 4863 bool inuserspace; 4864 4865 error = 0; 4866 4867 /* 4868 * Allocate space for a sample buffer. 4869 */ 4870 cpu = curcpu; 4871 psb = pmc_pcpu[cpu]->pc_sb[ring]; 4872 inuserspace = TRAPF_USERMODE(tf); 4873 ps = PMC_PROD_SAMPLE(psb); 4874 if (psb->ps_considx != psb->ps_prodidx && 4875 ps->ps_nsamples) { /* in use, reader hasn't caught up */ 4876 pm->pm_pcpu_state[cpu].pps_stalled = 1; 4877 counter_u64_add(pmc_stats.pm_intr_bufferfull, 1); 4878 PMCDBG6(SAM,INT,1,"(spc) cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d", 4879 cpu, pm, tf, inuserspace, 4880 (int)(psb->ps_prodidx & pmc_sample_mask), 4881 (int)(psb->ps_considx & pmc_sample_mask)); 4882 callchaindepth = 1; 4883 error = ENOMEM; 4884 goto done; 4885 } 4886 4887 /* Fill in entry. */ 4888 PMCDBG6(SAM,INT,1,"cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d", cpu, pm, tf, 4889 inuserspace, (int)(psb->ps_prodidx & pmc_sample_mask), 4890 (int)(psb->ps_considx & pmc_sample_mask)); 4891 4892 td = curthread; 4893 ps->ps_pmc = pm; 4894 ps->ps_td = td; 4895 ps->ps_pid = td->td_proc->p_pid; 4896 ps->ps_tid = td->td_tid; 4897 ps->ps_tsc = pmc_rdtsc(); 4898 ps->ps_ticks = ticks; 4899 ps->ps_cpu = cpu; 4900 ps->ps_flags = inuserspace ? PMC_CC_F_USERSPACE : 0; 4901 ps->ps_nsamples_actual = 0; 4902 4903 callchaindepth = (pm->pm_flags & PMC_F_CALLCHAIN) ? 4904 pmc_callchaindepth : 1; 4905 4906 MPASS(ps->ps_pc != NULL); 4907 4908 if (mp != NULL) { 4909 /* Set multipart flag, clear header and copy data */ 4910 ps->ps_flags |= PMC_CC_F_MULTIPART; 4911 ps->ps_pc[0] = 0; 4912 ps->ps_nsamples_actual = 1; 4913 pmc_multipart_copydata(ps, mp); 4914 } 4915 4916 if (callchaindepth == 1) { 4917 ps->ps_pc[ps->ps_nsamples_actual] = PMC_TRAPFRAME_TO_PC(tf); 4918 } else { 4919 /* 4920 * Kernel stack traversals can be done immediately, while we 4921 * defer to an AST for user space traversals. 4922 */ 4923 if (!inuserspace) { 4924 callchaindepth = pmc_save_kernel_callchain( 4925 ps->ps_pc + ps->ps_nsamples_actual, 4926 callchaindepth - ps->ps_nsamples_actual, tf); 4927 callchaindepth += ps->ps_nsamples_actual; 4928 } else { 4929 pmc_post_callchain_callback(); 4930 callchaindepth = PMC_USER_CALLCHAIN_PENDING; 4931 } 4932 } 4933 4934 ps->ps_nsamples = callchaindepth; /* mark entry as in-use */ 4935 if (ring == PMC_UR) { 4936 ps->ps_nsamples_actual = ps->ps_nsamples; 4937 ps->ps_nsamples = PMC_USER_CALLCHAIN_PENDING; 4938 } 4939 4940 KASSERT(counter_u64_fetch(pm->pm_runcount) >= 0, 4941 ("[pmc,%d] pm=%p runcount %ju", __LINE__, pm, 4942 (uintmax_t)counter_u64_fetch(pm->pm_runcount))); 4943 4944 counter_u64_add(pm->pm_runcount, 1); /* hold onto PMC */ 4945 /* increment write pointer */ 4946 psb->ps_prodidx++; 4947 done: 4948 /* mark CPU as needing processing */ 4949 if (callchaindepth != PMC_USER_CALLCHAIN_PENDING) 4950 DPCPU_SET(pmc_sampled, 1); 4951 4952 return (error); 4953 } 4954 4955 /* 4956 * Interrupt processing. 4957 * 4958 * This function may be called from an NMI handler. It cannot use any of the 4959 * locking primitives supplied by the OS. 4960 */ 4961 int 4962 pmc_process_interrupt_mp(int ring, struct pmc *pm, struct trapframe *tf, 4963 struct pmc_multipart *mp) 4964 { 4965 struct thread *td; 4966 4967 td = curthread; 4968 if ((pm->pm_flags & PMC_F_USERCALLCHAIN) && 4969 (td->td_proc->p_flag & P_KPROC) == 0 && !TRAPF_USERMODE(tf)) { 4970 atomic_add_int(&td->td_pmcpend, 1); 4971 return (pmc_add_sample(PMC_UR, pm, tf, mp)); 4972 } 4973 return (pmc_add_sample(ring, pm, tf, mp)); 4974 } 4975 4976 int 4977 pmc_process_interrupt(int ring, struct pmc *pm, struct trapframe *tf) 4978 { 4979 return (pmc_process_interrupt_mp(ring, pm, tf, NULL)); 4980 } 4981 4982 /* 4983 * Capture a user call chain. This function will be called from ast() 4984 * before control returns to userland and before the process gets 4985 * rescheduled. 4986 */ 4987 static void 4988 pmc_capture_user_callchain(int cpu, int ring, struct trapframe *tf) 4989 { 4990 struct pmc *pm; 4991 struct pmc_sample *ps; 4992 struct pmc_samplebuffer *psb; 4993 struct thread *td; 4994 uint64_t considx, prodidx; 4995 int nsamples, nrecords, pass, iter; 4996 int start_ticks __diagused; 4997 4998 psb = pmc_pcpu[cpu]->pc_sb[ring]; 4999 td = curthread; 5000 nrecords = INT_MAX; 5001 pass = 0; 5002 start_ticks = ticks; 5003 5004 KASSERT(ring == PMC_UR || (td->td_pflags & TDP_CALLCHAIN) != 0, 5005 ("[pmc,%d] Retrieving callchain for thread that doesn't want it", 5006 __LINE__)); 5007 restart: 5008 if (ring == PMC_UR) 5009 nrecords = atomic_readandclear_32(&td->td_pmcpend); 5010 5011 for (iter = 0, considx = psb->ps_considx, prodidx = psb->ps_prodidx; 5012 considx < prodidx && iter < pmc_nsamples; considx++, iter++) { 5013 ps = PMC_CONS_SAMPLE_OFF(psb, considx); 5014 5015 /* 5016 * Iterate through all deferred callchain requests. Walk from 5017 * the current read pointer to the current write pointer. 5018 */ 5019 #ifdef INVARIANTS 5020 if (ps->ps_nsamples == PMC_SAMPLE_FREE) { 5021 continue; 5022 } 5023 #endif 5024 if (ps->ps_td != td || 5025 ps->ps_nsamples != PMC_USER_CALLCHAIN_PENDING || 5026 ps->ps_pmc->pm_state != PMC_STATE_RUNNING) 5027 continue; 5028 5029 KASSERT(ps->ps_cpu == cpu, 5030 ("[pmc,%d] cpu mismatch ps_cpu=%d pcpu=%d", __LINE__, 5031 ps->ps_cpu, PCPU_GET(cpuid))); 5032 5033 pm = ps->ps_pmc; 5034 KASSERT(pm->pm_flags & PMC_F_CALLCHAIN, 5035 ("[pmc,%d] Retrieving callchain for PMC that doesn't " 5036 "want it", __LINE__)); 5037 5038 if (ring == PMC_UR) { 5039 counter_u64_add(pmc_stats.pm_merges, 1); 5040 } 5041 nsamples = ps->ps_nsamples_actual; 5042 5043 /* 5044 * Retrieve the callchain and mark the sample buffer 5045 * as 'processable' by the timer tick sweep code. 5046 */ 5047 if (__predict_true(nsamples < pmc_callchaindepth - 1)) 5048 nsamples += pmc_save_user_callchain(ps->ps_pc + nsamples, 5049 pmc_callchaindepth - nsamples - 1, tf); 5050 5051 /* 5052 * We have to prevent hardclock from potentially overwriting 5053 * this sample between when we read the value and when we set 5054 * it. 5055 */ 5056 spinlock_enter(); 5057 5058 /* 5059 * Verify that the sample hasn't been dropped in the meantime. 5060 */ 5061 if (ps->ps_nsamples == PMC_USER_CALLCHAIN_PENDING) { 5062 KASSERT(counter_u64_fetch(pm->pm_runcount) > 0, 5063 ("[pmc,%d] runcount %ju", __LINE__, 5064 (uintmax_t)counter_u64_fetch(pm->pm_runcount))); 5065 5066 ps->ps_nsamples = nsamples; 5067 /* 5068 * If we couldn't get a sample, simply drop the 5069 * reference. 5070 */ 5071 if (nsamples == 0) 5072 counter_u64_add(pm->pm_runcount, -1); 5073 } 5074 spinlock_exit(); 5075 if (nrecords-- == 1) 5076 break; 5077 } 5078 if (__predict_false(ring == PMC_UR && td->td_pmcpend)) { 5079 if (pass == 0) { 5080 pass = 1; 5081 goto restart; 5082 } 5083 /* only collect samples for this part once */ 5084 td->td_pmcpend = 0; 5085 } 5086 5087 #ifdef INVARIANTS 5088 if ((ticks - start_ticks) > hz) 5089 log(LOG_ERR, "%s took %d ticks\n", __func__, (ticks - start_ticks)); 5090 #endif 5091 /* mark CPU as needing processing */ 5092 DPCPU_SET(pmc_sampled, 1); 5093 } 5094 5095 /* 5096 * Process saved PC samples. 5097 */ 5098 static void 5099 pmc_process_samples(int cpu, ring_type_t ring) 5100 { 5101 struct pmc *pm; 5102 struct thread *td; 5103 struct pmc_owner *po; 5104 struct pmc_sample *ps; 5105 struct pmc_classdep *pcd; 5106 struct pmc_samplebuffer *psb; 5107 uint64_t delta __diagused; 5108 int adjri, n; 5109 5110 KASSERT(PCPU_GET(cpuid) == cpu, 5111 ("[pmc,%d] not on the correct CPU pcpu=%d cpu=%d", __LINE__, 5112 PCPU_GET(cpuid), cpu)); 5113 5114 psb = pmc_pcpu[cpu]->pc_sb[ring]; 5115 delta = psb->ps_prodidx - psb->ps_considx; 5116 MPASS(delta <= pmc_nsamples); 5117 MPASS(psb->ps_considx <= psb->ps_prodidx); 5118 for (n = 0; psb->ps_considx < psb->ps_prodidx; psb->ps_considx++, n++) { 5119 ps = PMC_CONS_SAMPLE(psb); 5120 5121 if (__predict_false(ps->ps_nsamples == PMC_SAMPLE_FREE)) 5122 continue; 5123 5124 /* skip non-running samples */ 5125 pm = ps->ps_pmc; 5126 if (pm->pm_state != PMC_STATE_RUNNING) 5127 goto entrydone; 5128 5129 KASSERT(counter_u64_fetch(pm->pm_runcount) > 0, 5130 ("[pmc,%d] pm=%p runcount %ju", __LINE__, pm, 5131 (uintmax_t)counter_u64_fetch(pm->pm_runcount))); 5132 KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)), 5133 ("[pmc,%d] pmc=%p non-sampling mode=%d", __LINE__, 5134 pm, PMC_TO_MODE(pm))); 5135 5136 po = pm->pm_owner; 5137 5138 /* If there is a pending AST wait for completion */ 5139 if (ps->ps_nsamples == PMC_USER_CALLCHAIN_PENDING) { 5140 /* 5141 * If we've been waiting more than 1 tick to 5142 * collect a callchain for this record then 5143 * drop it and move on. 5144 */ 5145 if (ticks - ps->ps_ticks > 1) { 5146 /* 5147 * Track how often we hit this as it will 5148 * preferentially lose user samples 5149 * for long running system calls. 5150 */ 5151 counter_u64_add(pmc_stats.pm_overwrites, 1); 5152 goto entrydone; 5153 } 5154 /* Need a rescan at a later time. */ 5155 DPCPU_SET(pmc_sampled, 1); 5156 break; 5157 } 5158 5159 PMCDBG6(SAM,OPS,1,"cpu=%d pm=%p n=%d fl=%x wr=%d rd=%d", cpu, 5160 pm, ps->ps_nsamples, ps->ps_flags, 5161 (int)(psb->ps_prodidx & pmc_sample_mask), 5162 (int)(psb->ps_considx & pmc_sample_mask)); 5163 5164 /* 5165 * If this is a process-mode PMC that is attached to 5166 * its owner, and if the PC is in user mode, update 5167 * profiling statistics like timer-based profiling 5168 * would have done. 5169 * 5170 * Otherwise, this is either a sampling-mode PMC that 5171 * is attached to a different process than its owner, 5172 * or a system-wide sampling PMC. Dispatch a log 5173 * entry to the PMC's owner process. 5174 */ 5175 if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) { 5176 if (ps->ps_flags & PMC_CC_F_USERSPACE) { 5177 td = FIRST_THREAD_IN_PROC(po->po_owner); 5178 addupc_intr(td, ps->ps_pc[0], 1); 5179 } 5180 } else 5181 pmclog_process_callchain(pm, ps); 5182 5183 entrydone: 5184 ps->ps_nsamples = 0; /* mark entry as free */ 5185 KASSERT(counter_u64_fetch(pm->pm_runcount) > 0, 5186 ("[pmc,%d] pm=%p runcount %ju", __LINE__, pm, 5187 (uintmax_t)counter_u64_fetch(pm->pm_runcount))); 5188 5189 counter_u64_add(pm->pm_runcount, -1); 5190 } 5191 5192 counter_u64_add(pmc_stats.pm_log_sweeps, 1); 5193 5194 /* Do not re-enable stalled PMCs if we failed to process any samples */ 5195 if (n == 0) 5196 return; 5197 5198 /* 5199 * Restart any stalled sampling PMCs on this CPU. 5200 * 5201 * If the NMI handler sets the pm_stalled field of a PMC after 5202 * the check below, we'll end up processing the stalled PMC at 5203 * the next hardclock tick. 5204 */ 5205 for (n = 0; n < md->pmd_npmc; n++) { 5206 pcd = pmc_ri_to_classdep(md, n, &adjri); 5207 KASSERT(pcd != NULL, 5208 ("[pmc,%d] null pcd ri=%d", __LINE__, n)); 5209 (void)(*pcd->pcd_get_config)(cpu, adjri, &pm); 5210 5211 if (pm == NULL || /* !cfg'ed */ 5212 pm->pm_state != PMC_STATE_RUNNING || /* !active */ 5213 !PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) || /* !sampling */ 5214 !pm->pm_pcpu_state[cpu].pps_cpustate || /* !desired */ 5215 !pm->pm_pcpu_state[cpu].pps_stalled) /* !stalled */ 5216 continue; 5217 5218 pm->pm_pcpu_state[cpu].pps_stalled = 0; 5219 (void)(*pcd->pcd_start_pmc)(cpu, adjri, pm); 5220 } 5221 } 5222 5223 /* 5224 * Event handlers. 5225 */ 5226 5227 /* 5228 * Handle a process exit. 5229 * 5230 * Remove this process from all hash tables. If this process 5231 * owned any PMCs, turn off those PMCs and deallocate them, 5232 * removing any associations with target processes. 5233 * 5234 * This function will be called by the last 'thread' of a 5235 * process. 5236 * 5237 * XXX This eventhandler gets called early in the exit process. 5238 * Consider using a 'hook' invocation from thread_exit() or equivalent 5239 * spot. Another negative is that kse_exit doesn't seem to call 5240 * exit1() [??]. 5241 */ 5242 static void 5243 pmc_process_exit(void *arg __unused, struct proc *p) 5244 { 5245 struct pmc *pm; 5246 struct pmc_owner *po; 5247 struct pmc_process *pp; 5248 struct pmc_classdep *pcd; 5249 pmc_value_t newvalue, tmp; 5250 int ri, adjri, cpu; 5251 bool is_using_hwpmcs; 5252 5253 PROC_LOCK(p); 5254 is_using_hwpmcs = (p->p_flag & P_HWPMC) != 0; 5255 PROC_UNLOCK(p); 5256 5257 /* 5258 * Log a sysexit event to all SS PMC owners. 5259 */ 5260 PMC_EPOCH_ENTER(); 5261 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) { 5262 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) != 0) 5263 pmclog_process_sysexit(po, p->p_pid); 5264 } 5265 PMC_EPOCH_EXIT(); 5266 5267 PMC_GET_SX_XLOCK(); 5268 PMCDBG3(PRC,EXT,1,"process-exit proc=%p (%d, %s)", p, p->p_pid, 5269 p->p_comm); 5270 5271 if (!is_using_hwpmcs) 5272 goto out; 5273 5274 /* 5275 * Since this code is invoked by the last thread in an exiting process, 5276 * we would have context switched IN at some prior point. However, with 5277 * PREEMPTION, kernel mode context switches may happen any time, so we 5278 * want to disable a context switch OUT till we get any PMCs targeting 5279 * this process off the hardware. 5280 * 5281 * We also need to atomically remove this process' entry from our 5282 * target process hash table, using PMC_FLAG_REMOVE. 5283 */ 5284 PMCDBG3(PRC,EXT,1, "process-exit proc=%p (%d, %s)", p, p->p_pid, 5285 p->p_comm); 5286 5287 critical_enter(); /* no preemption */ 5288 5289 cpu = curthread->td_oncpu; 5290 5291 pp = pmc_find_process_descriptor(p, PMC_FLAG_REMOVE); 5292 if (pp == NULL) { 5293 critical_exit(); 5294 goto out; 5295 } 5296 5297 PMCDBG2(PRC,EXT,2, "process-exit proc=%p pmc-process=%p", p, pp); 5298 5299 /* Run the context-switch-out steps for process exit. */ 5300 pmc_process_csw_stop_all(cpu); 5301 5302 /* 5303 * The exiting process could be the target of some PMCs which will be 5304 * running on currently executing CPU. 5305 * 5306 * We need to turn these PMCs off like we would do at context switch 5307 * OUT time. 5308 */ 5309 for (ri = 0; ri < md->pmd_npmc; ri++) { 5310 /* 5311 * Pick up the pmc pointer from hardware state similar to the 5312 * CSW_OUT code. 5313 */ 5314 pm = NULL; 5315 pcd = pmc_ri_to_classdep(md, ri, &adjri); 5316 5317 (void)(*pcd->pcd_get_config)(cpu, adjri, &pm); 5318 5319 PMCDBG2(PRC,EXT,2, "ri=%d pm=%p", ri, pm); 5320 5321 if (pm == NULL || !PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) 5322 continue; 5323 5324 PMCDBG4(PRC,EXT,2, "ppmcs[%d]=%p pm=%p state=%d", ri, 5325 pp->pp_pmcs[ri].pp_pmc, pm, pm->pm_state); 5326 5327 KASSERT(PMC_TO_ROWINDEX(pm) == ri, 5328 ("[pmc,%d] ri mismatch pmc(%d) ri(%d)", __LINE__, 5329 PMC_TO_ROWINDEX(pm), ri)); 5330 KASSERT(pm == pp->pp_pmcs[ri].pp_pmc, 5331 ("[pmc,%d] pm %p != pp_pmcs[%d] %p", __LINE__, pm, ri, 5332 pp->pp_pmcs[ri].pp_pmc)); 5333 KASSERT(counter_u64_fetch(pm->pm_runcount) > 0, 5334 ("[pmc,%d] bad runcount ri %d rc %ju", __LINE__, ri, 5335 (uintmax_t)counter_u64_fetch(pm->pm_runcount))); 5336 5337 /* 5338 * Change desired state, and then stop if not stalled. This 5339 * two-step dance should avoid race conditions where an 5340 * interrupt re-enables the PMC after this code has already 5341 * checked the pm_stalled flag. 5342 */ 5343 if (pm->pm_pcpu_state[cpu].pps_cpustate) { 5344 pm->pm_pcpu_state[cpu].pps_cpustate = 0; 5345 if (!pm->pm_pcpu_state[cpu].pps_stalled) { 5346 (void)pcd->pcd_stop_pmc(cpu, adjri, pm); 5347 5348 if (PMC_TO_MODE(pm) == PMC_MODE_TC) { 5349 pcd->pcd_read_pmc(cpu, adjri, pm, 5350 &newvalue); 5351 tmp = pmc_delta(pcd, newvalue, 5352 PMC_PCPU_SAVED(cpu, ri)); 5353 5354 mtx_pool_lock_spin(pmc_mtxpool, pm); 5355 pm->pm_gv.pm_savedvalue += tmp; 5356 pp->pp_pmcs[ri].pp_pmcval += tmp; 5357 mtx_pool_unlock_spin(pmc_mtxpool, pm); 5358 } 5359 } 5360 } 5361 5362 KASSERT(counter_u64_fetch(pm->pm_runcount) > 0, 5363 ("[pmc,%d] runcount is %d", __LINE__, ri)); 5364 5365 counter_u64_add(pm->pm_runcount, -1); 5366 (void)pcd->pcd_config_pmc(cpu, adjri, NULL); 5367 } 5368 5369 /* 5370 * Inform the MD layer of this pseudo "context switch out". 5371 */ 5372 (void)md->pmd_switch_out(pmc_pcpu[cpu], pp); 5373 5374 critical_exit(); /* ok to be pre-empted now */ 5375 5376 /* 5377 * Unlink this process from the PMCs that are targeting it. This will 5378 * send a signal to all PMC owner's whose PMCs are orphaned. 5379 * 5380 * Log PMC value at exit time if requested. 5381 */ 5382 for (ri = 0; ri < md->pmd_npmc; ri++) { 5383 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) { 5384 if ((pm->pm_flags & PMC_F_NEEDS_LOGFILE) != 0 && 5385 PMC_IS_COUNTING_MODE(PMC_TO_MODE(pm))) { 5386 pmclog_process_procexit(pm, pp); 5387 } 5388 pmc_unlink_target_process(pm, pp); 5389 } 5390 } 5391 free(pp, M_PMC); 5392 5393 out: 5394 /* 5395 * If the process owned PMCs, free them up and free up memory. 5396 */ 5397 if ((po = pmc_find_owner_descriptor(p)) != NULL) { 5398 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) != 0) 5399 pmclog_close(po); 5400 pmc_remove_owner(po); 5401 pmc_destroy_owner_descriptor(po); 5402 } 5403 5404 sx_xunlock(&pmc_sx); 5405 } 5406 5407 /* 5408 * Handle a process fork. 5409 * 5410 * If the parent process 'p1' is under HWPMC monitoring, then copy 5411 * over any attached PMCs that have 'do_descendants' semantics. 5412 */ 5413 static void 5414 pmc_process_fork(void *arg __unused, struct proc *p1, struct proc *newproc, 5415 int flags __unused) 5416 { 5417 struct pmc *pm; 5418 struct pmc_owner *po; 5419 struct pmc_process *ppnew, *ppold; 5420 unsigned int ri; 5421 bool is_using_hwpmcs, do_descendants; 5422 5423 PROC_LOCK(p1); 5424 is_using_hwpmcs = (p1->p_flag & P_HWPMC) != 0; 5425 PROC_UNLOCK(p1); 5426 5427 /* 5428 * If there are system-wide sampling PMCs active, we need to 5429 * log all fork events to their owner's logs. 5430 */ 5431 PMC_EPOCH_ENTER(); 5432 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) { 5433 if (po->po_flags & PMC_PO_OWNS_LOGFILE) { 5434 pmclog_process_procfork(po, p1->p_pid, newproc->p_pid); 5435 pmclog_process_proccreate(po, newproc, 1); 5436 } 5437 } 5438 PMC_EPOCH_EXIT(); 5439 5440 if (!is_using_hwpmcs) 5441 return; 5442 5443 PMC_GET_SX_XLOCK(); 5444 PMCDBG4(PMC,FRK,1, "process-fork proc=%p (%d, %s) -> %p", p1, 5445 p1->p_pid, p1->p_comm, newproc); 5446 5447 /* 5448 * If the parent process (curthread->td_proc) is a 5449 * target of any PMCs, look for PMCs that are to be 5450 * inherited, and link these into the new process 5451 * descriptor. 5452 */ 5453 ppold = pmc_find_process_descriptor(curthread->td_proc, PMC_FLAG_NONE); 5454 if (ppold == NULL) 5455 goto done; /* nothing to do */ 5456 5457 do_descendants = false; 5458 for (ri = 0; ri < md->pmd_npmc; ri++) { 5459 if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL && 5460 (pm->pm_flags & PMC_F_DESCENDANTS) != 0) { 5461 do_descendants = true; 5462 break; 5463 } 5464 } 5465 if (!do_descendants) /* nothing to do */ 5466 goto done; 5467 5468 /* 5469 * Now mark the new process as being tracked by this driver. 5470 */ 5471 PROC_LOCK(newproc); 5472 newproc->p_flag |= P_HWPMC; 5473 PROC_UNLOCK(newproc); 5474 5475 /* Allocate a descriptor for the new process. */ 5476 ppnew = pmc_find_process_descriptor(newproc, PMC_FLAG_ALLOCATE); 5477 if (ppnew == NULL) 5478 goto done; 5479 5480 /* 5481 * Run through all PMCs that were targeting the old process 5482 * and which specified F_DESCENDANTS and attach them to the 5483 * new process. 5484 * 5485 * Log the fork event to all owners of PMCs attached to this 5486 * process, if not already logged. 5487 */ 5488 for (ri = 0; ri < md->pmd_npmc; ri++) { 5489 if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL && 5490 (pm->pm_flags & PMC_F_DESCENDANTS) != 0) { 5491 pmc_link_target_process(pm, ppnew); 5492 po = pm->pm_owner; 5493 if (po->po_sscount == 0 && 5494 (po->po_flags & PMC_PO_OWNS_LOGFILE) != 0) { 5495 pmclog_process_procfork(po, p1->p_pid, 5496 newproc->p_pid); 5497 } 5498 } 5499 } 5500 5501 done: 5502 sx_xunlock(&pmc_sx); 5503 } 5504 5505 static void 5506 pmc_process_threadcreate(struct thread *td) 5507 { 5508 struct pmc_owner *po; 5509 5510 PMC_EPOCH_ENTER(); 5511 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) { 5512 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) != 0) 5513 pmclog_process_threadcreate(po, td, 1); 5514 } 5515 PMC_EPOCH_EXIT(); 5516 } 5517 5518 static void 5519 pmc_process_threadexit(struct thread *td) 5520 { 5521 struct pmc_owner *po; 5522 5523 PMC_EPOCH_ENTER(); 5524 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) { 5525 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) != 0) 5526 pmclog_process_threadexit(po, td); 5527 } 5528 PMC_EPOCH_EXIT(); 5529 } 5530 5531 static void 5532 pmc_process_proccreate(struct proc *p) 5533 { 5534 struct pmc_owner *po; 5535 5536 PMC_EPOCH_ENTER(); 5537 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) { 5538 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) != 0) 5539 pmclog_process_proccreate(po, p, 1 /* sync */); 5540 } 5541 PMC_EPOCH_EXIT(); 5542 } 5543 5544 static void 5545 pmc_process_allproc(struct pmc *pm) 5546 { 5547 struct pmc_owner *po; 5548 struct thread *td; 5549 struct proc *p; 5550 5551 po = pm->pm_owner; 5552 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) 5553 return; 5554 5555 sx_slock(&allproc_lock); 5556 FOREACH_PROC_IN_SYSTEM(p) { 5557 pmclog_process_proccreate(po, p, 0 /* sync */); 5558 PROC_LOCK(p); 5559 FOREACH_THREAD_IN_PROC(p, td) 5560 pmclog_process_threadcreate(po, td, 0 /* sync */); 5561 PROC_UNLOCK(p); 5562 } 5563 sx_sunlock(&allproc_lock); 5564 pmclog_flush(po, 0); 5565 } 5566 5567 static void 5568 pmc_kld_load(void *arg __unused, linker_file_t lf) 5569 { 5570 struct pmc_owner *po; 5571 5572 /* 5573 * Notify owners of system sampling PMCs about KLD operations. 5574 */ 5575 PMC_EPOCH_ENTER(); 5576 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) { 5577 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 5578 pmclog_process_map_in(po, (pid_t) -1, 5579 (uintfptr_t) lf->address, lf->pathname); 5580 } 5581 PMC_EPOCH_EXIT(); 5582 5583 /* 5584 * TODO: Notify owners of (all) process-sampling PMCs too. 5585 */ 5586 } 5587 5588 static void 5589 pmc_kld_unload(void *arg __unused, const char *filename __unused, 5590 caddr_t address, size_t size) 5591 { 5592 struct pmc_owner *po; 5593 5594 PMC_EPOCH_ENTER(); 5595 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) { 5596 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) != 0) { 5597 pmclog_process_map_out(po, (pid_t)-1, 5598 (uintfptr_t)address, (uintfptr_t)address + size); 5599 } 5600 } 5601 PMC_EPOCH_EXIT(); 5602 5603 /* 5604 * TODO: Notify owners of process-sampling PMCs. 5605 */ 5606 } 5607 5608 /* 5609 * initialization 5610 */ 5611 static const char * 5612 pmc_name_of_pmcclass(enum pmc_class class) 5613 { 5614 5615 switch (class) { 5616 #undef __PMC_CLASS 5617 #define __PMC_CLASS(S,V,D) \ 5618 case PMC_CLASS_##S: \ 5619 return #S; 5620 __PMC_CLASSES(); 5621 default: 5622 return ("<unknown>"); 5623 } 5624 } 5625 5626 /* 5627 * Base class initializer: allocate structure and set default classes. 5628 */ 5629 struct pmc_mdep * 5630 pmc_mdep_alloc(int nclasses) 5631 { 5632 struct pmc_mdep *md; 5633 int n; 5634 5635 /* SOFT + md classes */ 5636 n = 1 + nclasses; 5637 md = malloc(sizeof(struct pmc_mdep) + n * sizeof(struct pmc_classdep), 5638 M_PMC, M_WAITOK | M_ZERO); 5639 md->pmd_nclass = n; 5640 5641 /* Default methods */ 5642 md->pmd_switch_in = generic_switch_in; 5643 md->pmd_switch_out = generic_switch_out; 5644 5645 /* Add base class. */ 5646 pmc_soft_initialize(md); 5647 return (md); 5648 } 5649 5650 void 5651 pmc_mdep_free(struct pmc_mdep *md) 5652 { 5653 pmc_soft_finalize(md); 5654 free(md, M_PMC); 5655 } 5656 5657 static int 5658 generic_switch_in(struct pmc_cpu *pc __unused, struct pmc_process *pp __unused) 5659 { 5660 5661 return (0); 5662 } 5663 5664 static int 5665 generic_switch_out(struct pmc_cpu *pc __unused, struct pmc_process *pp __unused) 5666 { 5667 5668 return (0); 5669 } 5670 5671 static struct pmc_mdep * 5672 pmc_generic_cpu_initialize(void) 5673 { 5674 struct pmc_mdep *md; 5675 5676 md = pmc_mdep_alloc(0); 5677 5678 md->pmd_cputype = PMC_CPU_GENERIC; 5679 5680 return (md); 5681 } 5682 5683 static void 5684 pmc_generic_cpu_finalize(struct pmc_mdep *md __unused) 5685 { 5686 5687 } 5688 5689 static int 5690 pmc_initialize(void) 5691 { 5692 struct pcpu *pc; 5693 struct pmc_binding pb; 5694 struct pmc_classdep *pcd; 5695 struct pmc_sample *ps; 5696 struct pmc_samplebuffer *sb; 5697 int c, cpu, error, n, ri; 5698 u_int maxcpu, domain; 5699 5700 md = NULL; 5701 error = 0; 5702 5703 pmc_stats.pm_intr_ignored = counter_u64_alloc(M_WAITOK); 5704 pmc_stats.pm_intr_processed = counter_u64_alloc(M_WAITOK); 5705 pmc_stats.pm_intr_bufferfull = counter_u64_alloc(M_WAITOK); 5706 pmc_stats.pm_syscalls = counter_u64_alloc(M_WAITOK); 5707 pmc_stats.pm_syscall_errors = counter_u64_alloc(M_WAITOK); 5708 pmc_stats.pm_buffer_requests = counter_u64_alloc(M_WAITOK); 5709 pmc_stats.pm_buffer_requests_failed = counter_u64_alloc(M_WAITOK); 5710 pmc_stats.pm_log_sweeps = counter_u64_alloc(M_WAITOK); 5711 pmc_stats.pm_merges = counter_u64_alloc(M_WAITOK); 5712 pmc_stats.pm_overwrites = counter_u64_alloc(M_WAITOK); 5713 5714 #ifdef HWPMC_DEBUG 5715 /* parse debug flags first */ 5716 if (TUNABLE_STR_FETCH(PMC_SYSCTL_NAME_PREFIX "debugflags", 5717 pmc_debugstr, sizeof(pmc_debugstr))) { 5718 pmc_debugflags_parse(pmc_debugstr, pmc_debugstr + 5719 strlen(pmc_debugstr)); 5720 } 5721 #endif 5722 5723 PMCDBG1(MOD,INI,0, "PMC Initialize (version %x)", PMC_VERSION); 5724 5725 /* check kernel version */ 5726 if (pmc_kernel_version != PMC_VERSION) { 5727 if (pmc_kernel_version == 0) 5728 printf("hwpmc: this kernel has not been compiled with " 5729 "'options HWPMC_HOOKS'.\n"); 5730 else 5731 printf("hwpmc: kernel version (0x%x) does not match " 5732 "module version (0x%x).\n", pmc_kernel_version, 5733 PMC_VERSION); 5734 return (EPROGMISMATCH); 5735 } 5736 5737 /* 5738 * check sysctl parameters 5739 */ 5740 if (pmc_hashsize <= 0) { 5741 printf("hwpmc: tunable \"hashsize\"=%d must be " 5742 "greater than zero.\n", pmc_hashsize); 5743 pmc_hashsize = PMC_HASH_SIZE; 5744 } 5745 5746 if (pmc_nsamples <= 0 || pmc_nsamples > 65535) { 5747 printf("hwpmc: tunable \"nsamples\"=%d out of " 5748 "range.\n", pmc_nsamples); 5749 pmc_nsamples = PMC_NSAMPLES; 5750 } 5751 pmc_sample_mask = pmc_nsamples - 1; 5752 5753 if (pmc_callchaindepth <= 0 || 5754 pmc_callchaindepth > PMC_CALLCHAIN_DEPTH_MAX) { 5755 printf("hwpmc: tunable \"callchaindepth\"=%d out of " 5756 "range - using %d.\n", pmc_callchaindepth, 5757 PMC_CALLCHAIN_DEPTH_MAX); 5758 pmc_callchaindepth = PMC_CALLCHAIN_DEPTH_MAX; 5759 } 5760 5761 md = pmc_md_initialize(); 5762 if (md == NULL) { 5763 /* Default to generic CPU. */ 5764 md = pmc_generic_cpu_initialize(); 5765 if (md == NULL) 5766 return (ENOSYS); 5767 } 5768 5769 /* 5770 * Refresh classes base ri. Optional classes may come in different 5771 * order. 5772 */ 5773 for (ri = c = 0; c < md->pmd_nclass; c++) { 5774 pcd = &md->pmd_classdep[c]; 5775 pcd->pcd_ri = ri; 5776 ri += pcd->pcd_num; 5777 } 5778 5779 KASSERT(md->pmd_nclass >= 1 && md->pmd_npmc >= 1, 5780 ("[pmc,%d] no classes or pmcs", __LINE__)); 5781 5782 /* Compute the map from row-indices to classdep pointers. */ 5783 pmc_rowindex_to_classdep = malloc(sizeof(struct pmc_classdep *) * 5784 md->pmd_npmc, M_PMC, M_WAITOK | M_ZERO); 5785 5786 for (n = 0; n < md->pmd_npmc; n++) 5787 pmc_rowindex_to_classdep[n] = NULL; 5788 5789 for (ri = c = 0; c < md->pmd_nclass; c++) { 5790 pcd = &md->pmd_classdep[c]; 5791 for (n = 0; n < pcd->pcd_num; n++, ri++) 5792 pmc_rowindex_to_classdep[ri] = pcd; 5793 } 5794 5795 KASSERT(ri == md->pmd_npmc, 5796 ("[pmc,%d] npmc miscomputed: ri=%d, md->npmc=%d", __LINE__, 5797 ri, md->pmd_npmc)); 5798 5799 maxcpu = pmc_cpu_max(); 5800 5801 /* allocate space for the per-cpu array */ 5802 pmc_pcpu = malloc(maxcpu * sizeof(struct pmc_cpu *), M_PMC, 5803 M_WAITOK | M_ZERO); 5804 5805 /* per-cpu 'saved values' for managing process-mode PMCs */ 5806 pmc_pcpu_saved = malloc(sizeof(pmc_value_t) * maxcpu * md->pmd_npmc, 5807 M_PMC, M_WAITOK); 5808 5809 /* Perform CPU-dependent initialization. */ 5810 pmc_save_cpu_binding(&pb); 5811 error = 0; 5812 for (cpu = 0; error == 0 && cpu < maxcpu; cpu++) { 5813 if (!pmc_cpu_is_active(cpu)) 5814 continue; 5815 pmc_select_cpu(cpu); 5816 pmc_pcpu[cpu] = malloc(sizeof(struct pmc_cpu) + 5817 md->pmd_npmc * sizeof(struct pmc_hw *), M_PMC, 5818 M_WAITOK | M_ZERO); 5819 for (n = 0; error == 0 && n < md->pmd_nclass; n++) 5820 if (md->pmd_classdep[n].pcd_num > 0) 5821 error = md->pmd_classdep[n].pcd_pcpu_init(md, 5822 cpu); 5823 } 5824 pmc_restore_cpu_binding(&pb); 5825 5826 if (error != 0) 5827 return (error); 5828 5829 /* allocate space for the sample array */ 5830 for (cpu = 0; cpu < maxcpu; cpu++) { 5831 if (!pmc_cpu_is_active(cpu)) 5832 continue; 5833 pc = pcpu_find(cpu); 5834 domain = pc->pc_domain; 5835 sb = malloc_domainset(sizeof(struct pmc_samplebuffer) + 5836 pmc_nsamples * sizeof(struct pmc_sample), M_PMC, 5837 DOMAINSET_PREF(domain), M_WAITOK | M_ZERO); 5838 5839 KASSERT(pmc_pcpu[cpu] != NULL, 5840 ("[pmc,%d] cpu=%d Null per-cpu data", __LINE__, cpu)); 5841 5842 sb->ps_callchains = malloc_domainset(pmc_callchaindepth * 5843 pmc_nsamples * sizeof(uintptr_t), M_PMC, 5844 DOMAINSET_PREF(domain), M_WAITOK | M_ZERO); 5845 5846 for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++) 5847 ps->ps_pc = sb->ps_callchains + 5848 (n * pmc_callchaindepth); 5849 5850 pmc_pcpu[cpu]->pc_sb[PMC_HR] = sb; 5851 5852 sb = malloc_domainset(sizeof(struct pmc_samplebuffer) + 5853 pmc_nsamples * sizeof(struct pmc_sample), M_PMC, 5854 DOMAINSET_PREF(domain), M_WAITOK | M_ZERO); 5855 5856 sb->ps_callchains = malloc_domainset(pmc_callchaindepth * 5857 pmc_nsamples * sizeof(uintptr_t), M_PMC, 5858 DOMAINSET_PREF(domain), M_WAITOK | M_ZERO); 5859 for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++) 5860 ps->ps_pc = sb->ps_callchains + 5861 (n * pmc_callchaindepth); 5862 5863 pmc_pcpu[cpu]->pc_sb[PMC_SR] = sb; 5864 5865 sb = malloc_domainset(sizeof(struct pmc_samplebuffer) + 5866 pmc_nsamples * sizeof(struct pmc_sample), M_PMC, 5867 DOMAINSET_PREF(domain), M_WAITOK | M_ZERO); 5868 sb->ps_callchains = malloc_domainset(pmc_callchaindepth * 5869 pmc_nsamples * sizeof(uintptr_t), M_PMC, 5870 DOMAINSET_PREF(domain), M_WAITOK | M_ZERO); 5871 for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++) 5872 ps->ps_pc = sb->ps_callchains + n * pmc_callchaindepth; 5873 5874 pmc_pcpu[cpu]->pc_sb[PMC_UR] = sb; 5875 } 5876 5877 /* allocate space for the row disposition array */ 5878 pmc_pmcdisp = malloc(sizeof(enum pmc_mode) * md->pmd_npmc, 5879 M_PMC, M_WAITOK | M_ZERO); 5880 5881 /* mark all PMCs as available */ 5882 for (n = 0; n < md->pmd_npmc; n++) 5883 PMC_MARK_ROW_FREE(n); 5884 5885 /* allocate thread hash tables */ 5886 pmc_ownerhash = hashinit(pmc_hashsize, M_PMC, 5887 &pmc_ownerhashmask); 5888 5889 pmc_processhash = hashinit(pmc_hashsize, M_PMC, 5890 &pmc_processhashmask); 5891 mtx_init(&pmc_processhash_mtx, "pmc-process-hash", "pmc-leaf", 5892 MTX_SPIN); 5893 5894 CK_LIST_INIT(&pmc_ss_owners); 5895 pmc_ss_count = 0; 5896 5897 /* allocate a pool of spin mutexes */ 5898 pmc_mtxpool = mtx_pool_create("pmc-leaf", pmc_mtxpool_size, 5899 MTX_SPIN); 5900 5901 PMCDBG4(MOD,INI,1, "pmc_ownerhash=%p, mask=0x%lx " 5902 "targethash=%p mask=0x%lx", pmc_ownerhash, pmc_ownerhashmask, 5903 pmc_processhash, pmc_processhashmask); 5904 5905 /* Initialize a spin mutex for the thread free list. */ 5906 mtx_init(&pmc_threadfreelist_mtx, "pmc-threadfreelist", "pmc-leaf", 5907 MTX_SPIN); 5908 5909 /* Initialize the task to prune the thread free list. */ 5910 TASK_INIT(&free_task, 0, pmc_thread_descriptor_pool_free_task, NULL); 5911 5912 /* register process {exit,fork,exec} handlers */ 5913 pmc_exit_tag = EVENTHANDLER_REGISTER(process_exit, 5914 pmc_process_exit, NULL, EVENTHANDLER_PRI_ANY); 5915 pmc_fork_tag = EVENTHANDLER_REGISTER(process_fork, 5916 pmc_process_fork, NULL, EVENTHANDLER_PRI_ANY); 5917 5918 /* register kld event handlers */ 5919 pmc_kld_load_tag = EVENTHANDLER_REGISTER(kld_load, pmc_kld_load, 5920 NULL, EVENTHANDLER_PRI_ANY); 5921 pmc_kld_unload_tag = EVENTHANDLER_REGISTER(kld_unload, pmc_kld_unload, 5922 NULL, EVENTHANDLER_PRI_ANY); 5923 5924 /* initialize logging */ 5925 pmclog_initialize(); 5926 5927 /* set hook functions */ 5928 pmc_intr = md->pmd_intr; 5929 wmb(); 5930 pmc_hook = pmc_hook_handler; 5931 5932 if (error == 0) { 5933 printf(PMC_MODULE_NAME ":"); 5934 for (n = 0; n < md->pmd_nclass; n++) { 5935 if (md->pmd_classdep[n].pcd_num == 0) 5936 continue; 5937 pcd = &md->pmd_classdep[n]; 5938 printf(" %s/%d/%d/0x%b", 5939 pmc_name_of_pmcclass(pcd->pcd_class), 5940 pcd->pcd_num, 5941 pcd->pcd_width, 5942 pcd->pcd_caps, 5943 "\20" 5944 "\1INT\2USR\3SYS\4EDG\5THR" 5945 "\6REA\7WRI\10INV\11QUA\12PRC" 5946 "\13TAG\14CSC"); 5947 } 5948 printf("\n"); 5949 } 5950 5951 return (error); 5952 } 5953 5954 /* prepare to be unloaded */ 5955 static void 5956 pmc_cleanup(void) 5957 { 5958 struct pmc_binding pb; 5959 struct pmc_owner *po, *tmp; 5960 struct pmc_ownerhash *ph; 5961 struct pmc_processhash *prh __pmcdbg_used; 5962 u_int maxcpu; 5963 int cpu, c; 5964 5965 PMCDBG0(MOD,INI,0, "cleanup"); 5966 5967 /* switch off sampling */ 5968 CPU_FOREACH(cpu) 5969 DPCPU_ID_SET(cpu, pmc_sampled, 0); 5970 pmc_intr = NULL; 5971 5972 sx_xlock(&pmc_sx); 5973 if (pmc_hook == NULL) { /* being unloaded already */ 5974 sx_xunlock(&pmc_sx); 5975 return; 5976 } 5977 5978 pmc_hook = NULL; /* prevent new threads from entering module */ 5979 5980 /* deregister event handlers */ 5981 EVENTHANDLER_DEREGISTER(process_fork, pmc_fork_tag); 5982 EVENTHANDLER_DEREGISTER(process_exit, pmc_exit_tag); 5983 EVENTHANDLER_DEREGISTER(kld_load, pmc_kld_load_tag); 5984 EVENTHANDLER_DEREGISTER(kld_unload, pmc_kld_unload_tag); 5985 5986 /* send SIGBUS to all owner threads, free up allocations */ 5987 if (pmc_ownerhash != NULL) { 5988 for (ph = pmc_ownerhash; 5989 ph <= &pmc_ownerhash[pmc_ownerhashmask]; 5990 ph++) { 5991 LIST_FOREACH_SAFE(po, ph, po_next, tmp) { 5992 pmc_remove_owner(po); 5993 5994 PMCDBG3(MOD,INI,2, 5995 "cleanup signal proc=%p (%d, %s)", 5996 po->po_owner, po->po_owner->p_pid, 5997 po->po_owner->p_comm); 5998 5999 PROC_LOCK(po->po_owner); 6000 kern_psignal(po->po_owner, SIGBUS); 6001 PROC_UNLOCK(po->po_owner); 6002 6003 pmc_destroy_owner_descriptor(po); 6004 } 6005 } 6006 } 6007 6008 /* reclaim allocated data structures */ 6009 taskqueue_drain(taskqueue_fast, &free_task); 6010 mtx_destroy(&pmc_threadfreelist_mtx); 6011 pmc_thread_descriptor_pool_drain(); 6012 6013 if (pmc_mtxpool != NULL) 6014 mtx_pool_destroy(&pmc_mtxpool); 6015 6016 mtx_destroy(&pmc_processhash_mtx); 6017 if (pmc_processhash != NULL) { 6018 #ifdef HWPMC_DEBUG 6019 struct pmc_process *pp; 6020 6021 PMCDBG0(MOD,INI,3, "destroy process hash"); 6022 for (prh = pmc_processhash; 6023 prh <= &pmc_processhash[pmc_processhashmask]; 6024 prh++) 6025 LIST_FOREACH(pp, prh, pp_next) 6026 PMCDBG1(MOD,INI,3, "pid=%d", pp->pp_proc->p_pid); 6027 #endif 6028 6029 hashdestroy(pmc_processhash, M_PMC, pmc_processhashmask); 6030 pmc_processhash = NULL; 6031 } 6032 6033 if (pmc_ownerhash != NULL) { 6034 PMCDBG0(MOD,INI,3, "destroy owner hash"); 6035 hashdestroy(pmc_ownerhash, M_PMC, pmc_ownerhashmask); 6036 pmc_ownerhash = NULL; 6037 } 6038 6039 KASSERT(CK_LIST_EMPTY(&pmc_ss_owners), 6040 ("[pmc,%d] Global SS owner list not empty", __LINE__)); 6041 KASSERT(pmc_ss_count == 0, 6042 ("[pmc,%d] Global SS count not empty", __LINE__)); 6043 6044 /* do processor and pmc-class dependent cleanup */ 6045 maxcpu = pmc_cpu_max(); 6046 6047 PMCDBG0(MOD,INI,3, "md cleanup"); 6048 if (md) { 6049 pmc_save_cpu_binding(&pb); 6050 for (cpu = 0; cpu < maxcpu; cpu++) { 6051 PMCDBG2(MOD,INI,1,"pmc-cleanup cpu=%d pcs=%p", 6052 cpu, pmc_pcpu[cpu]); 6053 if (!pmc_cpu_is_active(cpu) || pmc_pcpu[cpu] == NULL) 6054 continue; 6055 6056 pmc_select_cpu(cpu); 6057 for (c = 0; c < md->pmd_nclass; c++) { 6058 if (md->pmd_classdep[c].pcd_num > 0) { 6059 md->pmd_classdep[c].pcd_pcpu_fini(md, 6060 cpu); 6061 } 6062 } 6063 } 6064 6065 if (md->pmd_cputype == PMC_CPU_GENERIC) 6066 pmc_generic_cpu_finalize(md); 6067 else 6068 pmc_md_finalize(md); 6069 6070 pmc_mdep_free(md); 6071 md = NULL; 6072 pmc_restore_cpu_binding(&pb); 6073 } 6074 6075 /* Free per-cpu descriptors. */ 6076 for (cpu = 0; cpu < maxcpu; cpu++) { 6077 if (!pmc_cpu_is_active(cpu)) 6078 continue; 6079 KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_HR] != NULL, 6080 ("[pmc,%d] Null hw cpu sample buffer cpu=%d", __LINE__, 6081 cpu)); 6082 KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_SR] != NULL, 6083 ("[pmc,%d] Null sw cpu sample buffer cpu=%d", __LINE__, 6084 cpu)); 6085 KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_UR] != NULL, 6086 ("[pmc,%d] Null userret cpu sample buffer cpu=%d", __LINE__, 6087 cpu)); 6088 free(pmc_pcpu[cpu]->pc_sb[PMC_HR]->ps_callchains, M_PMC); 6089 free(pmc_pcpu[cpu]->pc_sb[PMC_HR], M_PMC); 6090 free(pmc_pcpu[cpu]->pc_sb[PMC_SR]->ps_callchains, M_PMC); 6091 free(pmc_pcpu[cpu]->pc_sb[PMC_SR], M_PMC); 6092 free(pmc_pcpu[cpu]->pc_sb[PMC_UR]->ps_callchains, M_PMC); 6093 free(pmc_pcpu[cpu]->pc_sb[PMC_UR], M_PMC); 6094 free(pmc_pcpu[cpu], M_PMC); 6095 } 6096 6097 free(pmc_pcpu, M_PMC); 6098 pmc_pcpu = NULL; 6099 6100 free(pmc_pcpu_saved, M_PMC); 6101 pmc_pcpu_saved = NULL; 6102 6103 if (pmc_pmcdisp != NULL) { 6104 free(pmc_pmcdisp, M_PMC); 6105 pmc_pmcdisp = NULL; 6106 } 6107 6108 if (pmc_rowindex_to_classdep != NULL) { 6109 free(pmc_rowindex_to_classdep, M_PMC); 6110 pmc_rowindex_to_classdep = NULL; 6111 } 6112 6113 pmclog_shutdown(); 6114 counter_u64_free(pmc_stats.pm_intr_ignored); 6115 counter_u64_free(pmc_stats.pm_intr_processed); 6116 counter_u64_free(pmc_stats.pm_intr_bufferfull); 6117 counter_u64_free(pmc_stats.pm_syscalls); 6118 counter_u64_free(pmc_stats.pm_syscall_errors); 6119 counter_u64_free(pmc_stats.pm_buffer_requests); 6120 counter_u64_free(pmc_stats.pm_buffer_requests_failed); 6121 counter_u64_free(pmc_stats.pm_log_sweeps); 6122 counter_u64_free(pmc_stats.pm_merges); 6123 counter_u64_free(pmc_stats.pm_overwrites); 6124 sx_xunlock(&pmc_sx); /* we are done */ 6125 } 6126 6127 /* 6128 * The function called at load/unload. 6129 */ 6130 static int 6131 load(struct module *module __unused, int cmd, void *arg __unused) 6132 { 6133 int error; 6134 6135 error = 0; 6136 6137 switch (cmd) { 6138 case MOD_LOAD: 6139 /* initialize the subsystem */ 6140 error = pmc_initialize(); 6141 if (error != 0) 6142 break; 6143 PMCDBG2(MOD,INI,1, "syscall=%d maxcpu=%d", pmc_syscall_num, 6144 pmc_cpu_max()); 6145 break; 6146 case MOD_UNLOAD: 6147 case MOD_SHUTDOWN: 6148 pmc_cleanup(); 6149 PMCDBG0(MOD,INI,1, "unloaded"); 6150 break; 6151 default: 6152 error = EINVAL; 6153 break; 6154 } 6155 6156 return (error); 6157 } 6158