1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause 3 * 4 * Copyright (c) 2003-2008 Joseph Koshy 5 * Copyright (c) 2007 The FreeBSD Foundation 6 * Copyright (c) 2018 Matthew Macy 7 * All rights reserved. 8 * 9 * Portions of this software were developed by A. Joseph Koshy under 10 * sponsorship from the FreeBSD Foundation and Google, Inc. 11 * 12 * Redistribution and use in source and binary forms, with or without 13 * modification, are permitted provided that the following conditions 14 * are met: 15 * 1. Redistributions of source code must retain the above copyright 16 * notice, this list of conditions and the following disclaimer. 17 * 2. Redistributions in binary form must reproduce the above copyright 18 * notice, this list of conditions and the following disclaimer in the 19 * documentation and/or other materials provided with the distribution. 20 * 21 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 22 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 23 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 24 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 25 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 26 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 27 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 28 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 29 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 30 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 31 * SUCH DAMAGE. 32 */ 33 34 #include <sys/param.h> 35 #include <sys/systm.h> 36 #include <sys/domainset.h> 37 #include <sys/eventhandler.h> 38 #include <sys/jail.h> 39 #include <sys/kernel.h> 40 #include <sys/kthread.h> 41 #include <sys/limits.h> 42 #include <sys/lock.h> 43 #include <sys/malloc.h> 44 #include <sys/module.h> 45 #include <sys/mount.h> 46 #include <sys/mutex.h> 47 #include <sys/pmc.h> 48 #include <sys/pmckern.h> 49 #include <sys/pmclog.h> 50 #include <sys/priv.h> 51 #include <sys/proc.h> 52 #include <sys/queue.h> 53 #include <sys/resourcevar.h> 54 #include <sys/rwlock.h> 55 #include <sys/sched.h> 56 #include <sys/signalvar.h> 57 #include <sys/smp.h> 58 #include <sys/sx.h> 59 #include <sys/sysctl.h> 60 #include <sys/sysent.h> 61 #include <sys/syslog.h> 62 #include <sys/taskqueue.h> 63 #include <sys/vnode.h> 64 65 #include <sys/linker.h> /* needs to be after <sys/malloc.h> */ 66 67 #include <machine/atomic.h> 68 #include <machine/md_var.h> 69 70 #include <vm/vm.h> 71 #include <vm/vm_extern.h> 72 #include <vm/pmap.h> 73 #include <vm/vm_map.h> 74 #include <vm/vm_object.h> 75 76 #include "hwpmc_soft.h" 77 78 #define PMC_EPOCH_ENTER() \ 79 struct epoch_tracker pmc_et; \ 80 epoch_enter_preempt(global_epoch_preempt, &pmc_et) 81 82 #define PMC_EPOCH_EXIT() \ 83 epoch_exit_preempt(global_epoch_preempt, &pmc_et) 84 85 /* 86 * Types 87 */ 88 89 enum pmc_flags { 90 PMC_FLAG_NONE = 0x00, /* do nothing */ 91 PMC_FLAG_REMOVE = 0x01, /* atomically remove entry from hash */ 92 PMC_FLAG_ALLOCATE = 0x02, /* add entry to hash if not found */ 93 PMC_FLAG_NOWAIT = 0x04, /* do not wait for mallocs */ 94 }; 95 96 /* 97 * The offset in sysent where the syscall is allocated. 98 */ 99 static int pmc_syscall_num = NO_SYSCALL; 100 101 struct pmc_cpu **pmc_pcpu; /* per-cpu state */ 102 pmc_value_t *pmc_pcpu_saved; /* saved PMC values: CSW handling */ 103 104 #define PMC_PCPU_SAVED(C, R) pmc_pcpu_saved[(R) + md->pmd_npmc * (C)] 105 106 struct mtx_pool *pmc_mtxpool; 107 static int *pmc_pmcdisp; /* PMC row dispositions */ 108 109 #define PMC_ROW_DISP_IS_FREE(R) (pmc_pmcdisp[(R)] == 0) 110 #define PMC_ROW_DISP_IS_THREAD(R) (pmc_pmcdisp[(R)] > 0) 111 #define PMC_ROW_DISP_IS_STANDALONE(R) (pmc_pmcdisp[(R)] < 0) 112 113 #define PMC_MARK_ROW_FREE(R) do { \ 114 pmc_pmcdisp[(R)] = 0; \ 115 } while (0) 116 117 #define PMC_MARK_ROW_STANDALONE(R) do { \ 118 KASSERT(pmc_pmcdisp[(R)] <= 0, ("[pmc,%d] row disposition error", \ 119 __LINE__)); \ 120 atomic_add_int(&pmc_pmcdisp[(R)], -1); \ 121 KASSERT(pmc_pmcdisp[(R)] >= (-pmc_cpu_max_active()), \ 122 ("[pmc,%d] row disposition error", __LINE__)); \ 123 } while (0) 124 125 #define PMC_UNMARK_ROW_STANDALONE(R) do { \ 126 atomic_add_int(&pmc_pmcdisp[(R)], 1); \ 127 KASSERT(pmc_pmcdisp[(R)] <= 0, ("[pmc,%d] row disposition error", \ 128 __LINE__)); \ 129 } while (0) 130 131 #define PMC_MARK_ROW_THREAD(R) do { \ 132 KASSERT(pmc_pmcdisp[(R)] >= 0, ("[pmc,%d] row disposition error", \ 133 __LINE__)); \ 134 atomic_add_int(&pmc_pmcdisp[(R)], 1); \ 135 } while (0) 136 137 #define PMC_UNMARK_ROW_THREAD(R) do { \ 138 atomic_add_int(&pmc_pmcdisp[(R)], -1); \ 139 KASSERT(pmc_pmcdisp[(R)] >= 0, ("[pmc,%d] row disposition error", \ 140 __LINE__)); \ 141 } while (0) 142 143 /* various event handlers */ 144 static eventhandler_tag pmc_exit_tag, pmc_fork_tag, pmc_kld_load_tag, 145 pmc_kld_unload_tag; 146 147 /* Module statistics */ 148 struct pmc_driverstats pmc_stats; 149 150 /* Machine/processor dependent operations */ 151 static struct pmc_mdep *md; 152 153 /* 154 * Hash tables mapping owner processes and target threads to PMCs. 155 */ 156 struct mtx pmc_processhash_mtx; /* spin mutex */ 157 static u_long pmc_processhashmask; 158 static LIST_HEAD(pmc_processhash, pmc_process) *pmc_processhash; 159 160 /* 161 * Hash table of PMC owner descriptors. This table is protected by 162 * the shared PMC "sx" lock. 163 */ 164 static u_long pmc_ownerhashmask; 165 static LIST_HEAD(pmc_ownerhash, pmc_owner) *pmc_ownerhash; 166 167 /* 168 * List of PMC owners with system-wide sampling PMCs. 169 */ 170 static CK_LIST_HEAD(, pmc_owner) pmc_ss_owners; 171 172 /* 173 * List of free thread entries. This is protected by the spin 174 * mutex. 175 */ 176 static struct mtx pmc_threadfreelist_mtx; /* spin mutex */ 177 static LIST_HEAD(, pmc_thread) pmc_threadfreelist; 178 static int pmc_threadfreelist_entries = 0; 179 #define THREADENTRY_SIZE (sizeof(struct pmc_thread) + \ 180 (md->pmd_npmc * sizeof(struct pmc_threadpmcstate))) 181 182 /* 183 * Task to free thread descriptors 184 */ 185 static struct task free_task; 186 187 /* 188 * A map of row indices to classdep structures. 189 */ 190 static struct pmc_classdep **pmc_rowindex_to_classdep; 191 192 /* 193 * Prototypes 194 */ 195 196 #ifdef HWPMC_DEBUG 197 static int pmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS); 198 static int pmc_debugflags_parse(char *newstr, char *fence); 199 #endif 200 201 static void pmc_multipart_add(struct pmc_sample *ps, int type, 202 int length); 203 static void pmc_multipart_copydata(struct pmc_sample *ps, 204 struct pmc_multipart *mp); 205 206 static int load(struct module *module, int cmd, void *arg); 207 static int pmc_add_sample(ring_type_t ring, struct pmc *pm, 208 struct trapframe *tf, struct pmc_multipart *mp); 209 static void pmc_add_thread_descriptors_from_proc(struct proc *p, 210 struct pmc_process *pp); 211 static int pmc_attach_process(struct proc *p, struct pmc *pm); 212 static struct pmc *pmc_allocate_pmc_descriptor(void); 213 static struct pmc_owner *pmc_allocate_owner_descriptor(struct proc *p); 214 static int pmc_attach_one_process(struct proc *p, struct pmc *pm); 215 static bool pmc_can_allocate_row(int ri, enum pmc_mode mode); 216 static bool pmc_can_allocate_rowindex(struct proc *p, unsigned int ri, 217 int cpu); 218 static bool pmc_can_attach(struct pmc *pm, struct proc *p); 219 static void pmc_capture_user_callchain(int cpu, int soft, 220 struct trapframe *tf); 221 static void pmc_cleanup(void); 222 static int pmc_detach_process(struct proc *p, struct pmc *pm); 223 static int pmc_detach_one_process(struct proc *p, struct pmc *pm, 224 int flags); 225 static void pmc_destroy_owner_descriptor(struct pmc_owner *po); 226 static void pmc_destroy_pmc_descriptor(struct pmc *pm); 227 static void pmc_destroy_process_descriptor(struct pmc_process *pp); 228 static struct pmc_owner *pmc_find_owner_descriptor(struct proc *p); 229 static int pmc_find_pmc(pmc_id_t pmcid, struct pmc **pm); 230 static struct pmc *pmc_find_pmc_descriptor_in_process(struct pmc_owner *po, 231 pmc_id_t pmc); 232 static struct pmc_process *pmc_find_process_descriptor(struct proc *p, 233 uint32_t mode); 234 static struct pmc_thread *pmc_find_thread_descriptor(struct pmc_process *pp, 235 struct thread *td, uint32_t mode); 236 static void pmc_force_context_switch(void); 237 static void pmc_link_target_process(struct pmc *pm, 238 struct pmc_process *pp); 239 static void pmc_log_all_process_mappings(struct pmc_owner *po); 240 static void pmc_log_kernel_mappings(struct pmc *pm); 241 static void pmc_log_process_mappings(struct pmc_owner *po, struct proc *p); 242 static void pmc_maybe_remove_owner(struct pmc_owner *po); 243 static void pmc_post_callchain_callback(void); 244 static void pmc_process_allproc(struct pmc *pm); 245 static void pmc_process_csw_in(struct thread *td); 246 static void pmc_process_csw_out(struct thread *td); 247 static void pmc_process_exec(struct thread *td, 248 struct pmckern_procexec *pk); 249 static void pmc_process_exit(void *arg, struct proc *p); 250 static void pmc_process_fork(void *arg, struct proc *p1, 251 struct proc *p2, int n); 252 static void pmc_process_proccreate(struct proc *p); 253 static void pmc_process_samples(int cpu, ring_type_t soft); 254 static void pmc_process_threadcreate(struct thread *td); 255 static void pmc_process_threadexit(struct thread *td); 256 static void pmc_process_thread_add(struct thread *td); 257 static void pmc_process_thread_delete(struct thread *td); 258 static void pmc_process_thread_userret(struct thread *td); 259 static void pmc_release_pmc_descriptor(struct pmc *pmc); 260 static void pmc_remove_owner(struct pmc_owner *po); 261 static void pmc_remove_process_descriptor(struct pmc_process *pp); 262 static int pmc_start(struct pmc *pm); 263 static int pmc_stop(struct pmc *pm); 264 static int pmc_syscall_handler(struct thread *td, void *syscall_args); 265 static struct pmc_thread *pmc_thread_descriptor_pool_alloc(void); 266 static void pmc_thread_descriptor_pool_drain(void); 267 static void pmc_thread_descriptor_pool_free(struct pmc_thread *pt); 268 static void pmc_unlink_target_process(struct pmc *pmc, 269 struct pmc_process *pp); 270 271 static int generic_switch_in(struct pmc_cpu *pc, struct pmc_process *pp); 272 static int generic_switch_out(struct pmc_cpu *pc, struct pmc_process *pp); 273 static struct pmc_mdep *pmc_generic_cpu_initialize(void); 274 static void pmc_generic_cpu_finalize(struct pmc_mdep *md); 275 276 /* 277 * Kernel tunables and sysctl(8) interface. 278 */ 279 280 SYSCTL_DECL(_kern_hwpmc); 281 SYSCTL_NODE(_kern_hwpmc, OID_AUTO, stats, CTLFLAG_RW | CTLFLAG_MPSAFE, 0, 282 "HWPMC stats"); 283 284 /* Stats. */ 285 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, intr_ignored, CTLFLAG_RW, 286 &pmc_stats.pm_intr_ignored, 287 "# of interrupts ignored"); 288 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, intr_processed, CTLFLAG_RW, 289 &pmc_stats.pm_intr_processed, 290 "# of interrupts processed"); 291 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, intr_bufferfull, CTLFLAG_RW, 292 &pmc_stats.pm_intr_bufferfull, 293 "# of interrupts where buffer was full"); 294 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, syscalls, CTLFLAG_RW, 295 &pmc_stats.pm_syscalls, 296 "# of syscalls"); 297 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, syscall_errors, CTLFLAG_RW, 298 &pmc_stats.pm_syscall_errors, 299 "# of syscall_errors"); 300 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, buffer_requests, CTLFLAG_RW, 301 &pmc_stats.pm_buffer_requests, 302 "# of buffer requests"); 303 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, buffer_requests_failed, 304 CTLFLAG_RW, &pmc_stats.pm_buffer_requests_failed, 305 "# of buffer requests which failed"); 306 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, log_sweeps, CTLFLAG_RW, 307 &pmc_stats.pm_log_sweeps, 308 "# of times samples were processed"); 309 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, merges, CTLFLAG_RW, 310 &pmc_stats.pm_merges, 311 "# of times kernel stack was found for user trace"); 312 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, overwrites, CTLFLAG_RW, 313 &pmc_stats.pm_overwrites, 314 "# of times a sample was overwritten before being logged"); 315 316 static int pmc_callchaindepth = PMC_CALLCHAIN_DEPTH; 317 SYSCTL_INT(_kern_hwpmc, OID_AUTO, callchaindepth, CTLFLAG_RDTUN, 318 &pmc_callchaindepth, 0, 319 "depth of call chain records"); 320 321 char pmc_cpuid[PMC_CPUID_LEN]; 322 SYSCTL_STRING(_kern_hwpmc, OID_AUTO, cpuid, CTLFLAG_RD, 323 pmc_cpuid, 0, 324 "cpu version string"); 325 326 #ifdef HWPMC_DEBUG 327 struct pmc_debugflags pmc_debugflags = PMC_DEBUG_DEFAULT_FLAGS; 328 char pmc_debugstr[PMC_DEBUG_STRSIZE]; 329 TUNABLE_STR(PMC_SYSCTL_NAME_PREFIX "debugflags", pmc_debugstr, 330 sizeof(pmc_debugstr)); 331 SYSCTL_PROC(_kern_hwpmc, OID_AUTO, debugflags, 332 CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, 333 0, 0, pmc_debugflags_sysctl_handler, "A", 334 "debug flags"); 335 #endif 336 337 /* 338 * kern.hwpmc.hashsize -- determines the number of rows in the 339 * of the hash table used to look up threads 340 */ 341 static int pmc_hashsize = PMC_HASH_SIZE; 342 SYSCTL_INT(_kern_hwpmc, OID_AUTO, hashsize, CTLFLAG_RDTUN, 343 &pmc_hashsize, 0, 344 "rows in hash tables"); 345 346 /* 347 * kern.hwpmc.nsamples --- number of PC samples/callchain stacks per CPU 348 */ 349 static int pmc_nsamples = PMC_NSAMPLES; 350 SYSCTL_INT(_kern_hwpmc, OID_AUTO, nsamples, CTLFLAG_RDTUN, 351 &pmc_nsamples, 0, 352 "number of PC samples per CPU"); 353 354 static uint64_t pmc_sample_mask = PMC_NSAMPLES - 1; 355 356 /* 357 * kern.hwpmc.mtxpoolsize -- number of mutexes in the mutex pool. 358 */ 359 static int pmc_mtxpool_size = PMC_MTXPOOL_SIZE; 360 SYSCTL_INT(_kern_hwpmc, OID_AUTO, mtxpoolsize, CTLFLAG_RDTUN, 361 &pmc_mtxpool_size, 0, 362 "size of spin mutex pool"); 363 364 /* 365 * kern.hwpmc.threadfreelist_entries -- number of free entries 366 */ 367 SYSCTL_INT(_kern_hwpmc, OID_AUTO, threadfreelist_entries, CTLFLAG_RD, 368 &pmc_threadfreelist_entries, 0, 369 "number of available thread entries"); 370 371 /* 372 * kern.hwpmc.threadfreelist_max -- maximum number of free entries 373 */ 374 static int pmc_threadfreelist_max = PMC_THREADLIST_MAX; 375 SYSCTL_INT(_kern_hwpmc, OID_AUTO, threadfreelist_max, CTLFLAG_RW, 376 &pmc_threadfreelist_max, 0, 377 "maximum number of available thread entries before freeing some"); 378 379 /* 380 * kern.hwpmc.mincount -- minimum sample count 381 */ 382 static u_int pmc_mincount = 1000; 383 SYSCTL_INT(_kern_hwpmc, OID_AUTO, mincount, CTLFLAG_RWTUN, 384 &pmc_mincount, 0, 385 "minimum count for sampling counters"); 386 387 /* 388 * security.bsd.unprivileged_syspmcs -- allow non-root processes to 389 * allocate system-wide PMCs. 390 * 391 * Allowing unprivileged processes to allocate system PMCs is convenient 392 * if system-wide measurements need to be taken concurrently with other 393 * per-process measurements. This feature is turned off by default. 394 */ 395 static int pmc_unprivileged_syspmcs = 0; 396 SYSCTL_INT(_security_bsd, OID_AUTO, unprivileged_syspmcs, CTLFLAG_RWTUN, 397 &pmc_unprivileged_syspmcs, 0, 398 "allow unprivileged process to allocate system PMCs"); 399 400 /* 401 * Hash function. Discard the lower 2 bits of the pointer since 402 * these are always zero for our uses. The hash multiplier is 403 * round((2^LONG_BIT) * ((sqrt(5)-1)/2)). 404 */ 405 #if LONG_BIT == 64 406 #define _PMC_HM 11400714819323198486u 407 #elif LONG_BIT == 32 408 #define _PMC_HM 2654435769u 409 #else 410 #error Must know the size of 'long' to compile 411 #endif 412 413 #define PMC_HASH_PTR(P,M) ((((unsigned long) (P) >> 2) * _PMC_HM) & (M)) 414 415 /* 416 * Syscall structures 417 */ 418 419 /* The `sysent' for the new syscall */ 420 static struct sysent pmc_sysent = { 421 .sy_narg = 2, 422 .sy_call = pmc_syscall_handler, 423 }; 424 425 static struct syscall_module_data pmc_syscall_mod = { 426 .chainevh = load, 427 .chainarg = NULL, 428 .offset = &pmc_syscall_num, 429 .new_sysent = &pmc_sysent, 430 .old_sysent = { .sy_narg = 0, .sy_call = NULL }, 431 .flags = SY_THR_STATIC_KLD, 432 }; 433 434 static moduledata_t pmc_mod = { 435 .name = PMC_MODULE_NAME, 436 .evhand = syscall_module_handler, 437 .priv = &pmc_syscall_mod, 438 }; 439 440 #ifdef EARLY_AP_STARTUP 441 DECLARE_MODULE(pmc, pmc_mod, SI_SUB_SYSCALLS, SI_ORDER_ANY); 442 #else 443 DECLARE_MODULE(pmc, pmc_mod, SI_SUB_SMP, SI_ORDER_ANY); 444 #endif 445 MODULE_VERSION(pmc, PMC_VERSION); 446 447 #ifdef HWPMC_DEBUG 448 enum pmc_dbgparse_state { 449 PMCDS_WS, /* in whitespace */ 450 PMCDS_MAJOR, /* seen a major keyword */ 451 PMCDS_MINOR 452 }; 453 454 static int 455 pmc_debugflags_parse(char *newstr, char *fence) 456 { 457 struct pmc_debugflags *tmpflags; 458 size_t kwlen; 459 char c, *p, *q; 460 int error, *newbits, tmp; 461 int found; 462 463 tmpflags = malloc(sizeof(*tmpflags), M_PMC, M_WAITOK | M_ZERO); 464 465 error = 0; 466 for (p = newstr; p < fence && (c = *p); p++) { 467 /* skip white space */ 468 if (c == ' ' || c == '\t') 469 continue; 470 471 /* look for a keyword followed by "=" */ 472 for (q = p; p < fence && (c = *p) && c != '='; p++) 473 ; 474 if (c != '=') { 475 error = EINVAL; 476 goto done; 477 } 478 479 kwlen = p - q; 480 newbits = NULL; 481 482 /* lookup flag group name */ 483 #define DBG_SET_FLAG_MAJ(S,F) \ 484 if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0) \ 485 newbits = &tmpflags->pdb_ ## F; 486 487 DBG_SET_FLAG_MAJ("cpu", CPU); 488 DBG_SET_FLAG_MAJ("csw", CSW); 489 DBG_SET_FLAG_MAJ("logging", LOG); 490 DBG_SET_FLAG_MAJ("module", MOD); 491 DBG_SET_FLAG_MAJ("md", MDP); 492 DBG_SET_FLAG_MAJ("owner", OWN); 493 DBG_SET_FLAG_MAJ("pmc", PMC); 494 DBG_SET_FLAG_MAJ("process", PRC); 495 DBG_SET_FLAG_MAJ("sampling", SAM); 496 #undef DBG_SET_FLAG_MAJ 497 498 if (newbits == NULL) { 499 error = EINVAL; 500 goto done; 501 } 502 503 p++; /* skip the '=' */ 504 505 /* Now parse the individual flags */ 506 tmp = 0; 507 newflag: 508 for (q = p; p < fence && (c = *p); p++) 509 if (c == ' ' || c == '\t' || c == ',') 510 break; 511 512 /* p == fence or c == ws or c == "," or c == 0 */ 513 514 if ((kwlen = p - q) == 0) { 515 *newbits = tmp; 516 continue; 517 } 518 519 found = 0; 520 #define DBG_SET_FLAG_MIN(S,F) \ 521 if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0) \ 522 tmp |= found = (1 << PMC_DEBUG_MIN_ ## F) 523 524 /* a '*' denotes all possible flags in the group */ 525 if (kwlen == 1 && *q == '*') 526 tmp = found = ~0; 527 /* look for individual flag names */ 528 DBG_SET_FLAG_MIN("allocaterow", ALR); 529 DBG_SET_FLAG_MIN("allocate", ALL); 530 DBG_SET_FLAG_MIN("attach", ATT); 531 DBG_SET_FLAG_MIN("bind", BND); 532 DBG_SET_FLAG_MIN("config", CFG); 533 DBG_SET_FLAG_MIN("exec", EXC); 534 DBG_SET_FLAG_MIN("exit", EXT); 535 DBG_SET_FLAG_MIN("find", FND); 536 DBG_SET_FLAG_MIN("flush", FLS); 537 DBG_SET_FLAG_MIN("fork", FRK); 538 DBG_SET_FLAG_MIN("getbuf", GTB); 539 DBG_SET_FLAG_MIN("hook", PMH); 540 DBG_SET_FLAG_MIN("init", INI); 541 DBG_SET_FLAG_MIN("intr", INT); 542 DBG_SET_FLAG_MIN("linktarget", TLK); 543 DBG_SET_FLAG_MIN("mayberemove", OMR); 544 DBG_SET_FLAG_MIN("ops", OPS); 545 DBG_SET_FLAG_MIN("read", REA); 546 DBG_SET_FLAG_MIN("register", REG); 547 DBG_SET_FLAG_MIN("release", REL); 548 DBG_SET_FLAG_MIN("remove", ORM); 549 DBG_SET_FLAG_MIN("sample", SAM); 550 DBG_SET_FLAG_MIN("scheduleio", SIO); 551 DBG_SET_FLAG_MIN("select", SEL); 552 DBG_SET_FLAG_MIN("signal", SIG); 553 DBG_SET_FLAG_MIN("swi", SWI); 554 DBG_SET_FLAG_MIN("swo", SWO); 555 DBG_SET_FLAG_MIN("start", STA); 556 DBG_SET_FLAG_MIN("stop", STO); 557 DBG_SET_FLAG_MIN("syscall", PMS); 558 DBG_SET_FLAG_MIN("unlinktarget", TUL); 559 DBG_SET_FLAG_MIN("write", WRI); 560 #undef DBG_SET_FLAG_MIN 561 if (found == 0) { 562 /* unrecognized flag name */ 563 error = EINVAL; 564 goto done; 565 } 566 567 if (c == 0 || c == ' ' || c == '\t') { /* end of flag group */ 568 *newbits = tmp; 569 continue; 570 } 571 572 p++; 573 goto newflag; 574 } 575 576 /* save the new flag set */ 577 bcopy(tmpflags, &pmc_debugflags, sizeof(pmc_debugflags)); 578 done: 579 free(tmpflags, M_PMC); 580 return (error); 581 } 582 583 static int 584 pmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS) 585 { 586 char *fence, *newstr; 587 int error; 588 u_int n; 589 590 n = sizeof(pmc_debugstr); 591 newstr = malloc(n, M_PMC, M_WAITOK | M_ZERO); 592 strlcpy(newstr, pmc_debugstr, n); 593 594 error = sysctl_handle_string(oidp, newstr, n, req); 595 596 /* if there is a new string, parse and copy it */ 597 if (error == 0 && req->newptr != NULL) { 598 fence = newstr + (n < req->newlen ? n : req->newlen + 1); 599 error = pmc_debugflags_parse(newstr, fence); 600 if (error == 0) 601 strlcpy(pmc_debugstr, newstr, sizeof(pmc_debugstr)); 602 } 603 free(newstr, M_PMC); 604 605 return (error); 606 } 607 #endif 608 609 /* 610 * Map a row index to a classdep structure and return the adjusted row 611 * index for the PMC class index. 612 */ 613 static struct pmc_classdep * 614 pmc_ri_to_classdep(struct pmc_mdep *md __unused, int ri, int *adjri) 615 { 616 struct pmc_classdep *pcd; 617 618 KASSERT(ri >= 0 && ri < md->pmd_npmc, 619 ("[pmc,%d] illegal row-index %d", __LINE__, ri)); 620 621 pcd = pmc_rowindex_to_classdep[ri]; 622 KASSERT(pcd != NULL, 623 ("[pmc,%d] ri %d null pcd", __LINE__, ri)); 624 625 *adjri = ri - pcd->pcd_ri; 626 KASSERT(*adjri >= 0 && *adjri < pcd->pcd_num, 627 ("[pmc,%d] adjusted row-index %d", __LINE__, *adjri)); 628 629 return (pcd); 630 } 631 632 /* 633 * Concurrency Control 634 * 635 * The driver manages the following data structures: 636 * 637 * - target process descriptors, one per target process 638 * - owner process descriptors (and attached lists), one per owner process 639 * - lookup hash tables for owner and target processes 640 * - PMC descriptors (and attached lists) 641 * - per-cpu hardware state 642 * - the 'hook' variable through which the kernel calls into 643 * this module 644 * - the machine hardware state (managed by the MD layer) 645 * 646 * These data structures are accessed from: 647 * 648 * - thread context-switch code 649 * - interrupt handlers (possibly on multiple cpus) 650 * - kernel threads on multiple cpus running on behalf of user 651 * processes doing system calls 652 * - this driver's private kernel threads 653 * 654 * = Locks and Locking strategy = 655 * 656 * The driver uses four locking strategies for its operation: 657 * 658 * - The global SX lock "pmc_sx" is used to protect internal 659 * data structures. 660 * 661 * Calls into the module by syscall() start with this lock being 662 * held in exclusive mode. Depending on the requested operation, 663 * the lock may be downgraded to 'shared' mode to allow more 664 * concurrent readers into the module. Calls into the module from 665 * other parts of the kernel acquire the lock in shared mode. 666 * 667 * This SX lock is held in exclusive mode for any operations that 668 * modify the linkages between the driver's internal data structures. 669 * 670 * The 'pmc_hook' function pointer is also protected by this lock. 671 * It is only examined with the sx lock held in exclusive mode. The 672 * kernel module is allowed to be unloaded only with the sx lock held 673 * in exclusive mode. In normal syscall handling, after acquiring the 674 * pmc_sx lock we first check that 'pmc_hook' is non-null before 675 * proceeding. This prevents races between the thread unloading the module 676 * and other threads seeking to use the module. 677 * 678 * - Lookups of target process structures and owner process structures 679 * cannot use the global "pmc_sx" SX lock because these lookups need 680 * to happen during context switches and in other critical sections 681 * where sleeping is not allowed. We protect these lookup tables 682 * with their own private spin-mutexes, "pmc_processhash_mtx" and 683 * "pmc_ownerhash_mtx". 684 * 685 * - Interrupt handlers work in a lock free manner. At interrupt 686 * time, handlers look at the PMC pointer (phw->phw_pmc) configured 687 * when the PMC was started. If this pointer is NULL, the interrupt 688 * is ignored after updating driver statistics. We ensure that this 689 * pointer is set (using an atomic operation if necessary) before the 690 * PMC hardware is started. Conversely, this pointer is unset atomically 691 * only after the PMC hardware is stopped. 692 * 693 * We ensure that everything needed for the operation of an 694 * interrupt handler is available without it needing to acquire any 695 * locks. We also ensure that a PMC's software state is destroyed only 696 * after the PMC is taken off hardware (on all CPUs). 697 * 698 * - Context-switch handling with process-private PMCs needs more 699 * care. 700 * 701 * A given process may be the target of multiple PMCs. For example, 702 * PMCATTACH and PMCDETACH may be requested by a process on one CPU 703 * while the target process is running on another. A PMC could also 704 * be getting released because its owner is exiting. We tackle 705 * these situations in the following manner: 706 * 707 * - each target process structure 'pmc_process' has an array 708 * of 'struct pmc *' pointers, one for each hardware PMC. 709 * 710 * - At context switch IN time, each "target" PMC in RUNNING state 711 * gets started on hardware and a pointer to each PMC is copied into 712 * the per-cpu phw array. The 'runcount' for the PMC is 713 * incremented. 714 * 715 * - At context switch OUT time, all process-virtual PMCs are stopped 716 * on hardware. The saved value is added to the PMCs value field 717 * only if the PMC is in a non-deleted state (the PMCs state could 718 * have changed during the current time slice). 719 * 720 * Note that since in-between a switch IN on a processor and a switch 721 * OUT, the PMC could have been released on another CPU. Therefore 722 * context switch OUT always looks at the hardware state to turn 723 * OFF PMCs and will update a PMC's saved value only if reachable 724 * from the target process record. 725 * 726 * - OP PMCRELEASE could be called on a PMC at any time (the PMC could 727 * be attached to many processes at the time of the call and could 728 * be active on multiple CPUs). 729 * 730 * We prevent further scheduling of the PMC by marking it as in 731 * state 'DELETED'. If the runcount of the PMC is non-zero then 732 * this PMC is currently running on a CPU somewhere. The thread 733 * doing the PMCRELEASE operation waits by repeatedly doing a 734 * pause() till the runcount comes to zero. 735 * 736 * The contents of a PMC descriptor (struct pmc) are protected using 737 * a spin-mutex. In order to save space, we use a mutex pool. 738 * 739 * In terms of lock types used by witness(4), we use: 740 * - Type "pmc-sx", used by the global SX lock. 741 * - Type "pmc-sleep", for sleep mutexes used by logger threads. 742 * - Type "pmc-per-proc", for protecting PMC owner descriptors. 743 * - Type "pmc-leaf", used for all other spin mutexes. 744 */ 745 746 /* 747 * Save the CPU binding of the current kthread. 748 */ 749 void 750 pmc_save_cpu_binding(struct pmc_binding *pb) 751 { 752 PMCDBG0(CPU,BND,2, "save-cpu"); 753 thread_lock(curthread); 754 pb->pb_bound = sched_is_bound(curthread); 755 pb->pb_cpu = curthread->td_oncpu; 756 pb->pb_priority = curthread->td_priority; 757 thread_unlock(curthread); 758 PMCDBG1(CPU,BND,2, "save-cpu cpu=%d", pb->pb_cpu); 759 } 760 761 /* 762 * Restore the CPU binding of the current thread. 763 */ 764 void 765 pmc_restore_cpu_binding(struct pmc_binding *pb) 766 { 767 PMCDBG2(CPU,BND,2, "restore-cpu curcpu=%d restore=%d", 768 curthread->td_oncpu, pb->pb_cpu); 769 thread_lock(curthread); 770 sched_bind(curthread, pb->pb_cpu); 771 if (!pb->pb_bound) 772 sched_unbind(curthread); 773 sched_prio(curthread, pb->pb_priority); 774 thread_unlock(curthread); 775 PMCDBG0(CPU,BND,2, "restore-cpu done"); 776 } 777 778 /* 779 * Move execution over to the specified CPU and bind it there. 780 */ 781 void 782 pmc_select_cpu(int cpu) 783 { 784 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), 785 ("[pmc,%d] bad cpu number %d", __LINE__, cpu)); 786 787 /* Never move to an inactive CPU. */ 788 KASSERT(pmc_cpu_is_active(cpu), ("[pmc,%d] selecting inactive " 789 "CPU %d", __LINE__, cpu)); 790 791 PMCDBG1(CPU,SEL,2, "select-cpu cpu=%d", cpu); 792 thread_lock(curthread); 793 sched_prio(curthread, PRI_MIN); 794 sched_bind(curthread, cpu); 795 thread_unlock(curthread); 796 797 KASSERT(curthread->td_oncpu == cpu, 798 ("[pmc,%d] CPU not bound [cpu=%d, curr=%d]", __LINE__, 799 cpu, curthread->td_oncpu)); 800 801 PMCDBG1(CPU,SEL,2, "select-cpu cpu=%d ok", cpu); 802 } 803 804 /* 805 * Force a context switch. 806 * 807 * We do this by pause'ing for 1 tick -- invoking mi_switch() is not 808 * guaranteed to force a context switch. 809 */ 810 static void 811 pmc_force_context_switch(void) 812 { 813 814 pause("pmcctx", 1); 815 } 816 817 uint64_t 818 pmc_rdtsc(void) 819 { 820 #if defined(__i386__) 821 /* Unfortunately get_cyclecount on i386 uses cpu_ticks. */ 822 return (rdtsc()); 823 #else 824 return (get_cyclecount()); 825 #endif 826 } 827 828 /* 829 * Get the file name for an executable. This is a simple wrapper 830 * around vn_fullpath(9). 831 */ 832 static void 833 pmc_getfilename(struct vnode *v, char **fullpath, char **freepath) 834 { 835 836 *fullpath = "unknown"; 837 *freepath = NULL; 838 vn_fullpath(v, fullpath, freepath); 839 } 840 841 /* 842 * Remove a process owning PMCs. 843 */ 844 void 845 pmc_remove_owner(struct pmc_owner *po) 846 { 847 struct pmc *pm, *tmp; 848 849 sx_assert(&pmc_sx, SX_XLOCKED); 850 851 PMCDBG1(OWN,ORM,1, "remove-owner po=%p", po); 852 853 /* Remove descriptor from the owner hash table */ 854 LIST_REMOVE(po, po_next); 855 856 /* release all owned PMC descriptors */ 857 LIST_FOREACH_SAFE(pm, &po->po_pmcs, pm_next, tmp) { 858 PMCDBG1(OWN,ORM,2, "pmc=%p", pm); 859 KASSERT(pm->pm_owner == po, 860 ("[pmc,%d] owner %p != po %p", __LINE__, pm->pm_owner, po)); 861 862 pmc_release_pmc_descriptor(pm); /* will unlink from the list */ 863 pmc_destroy_pmc_descriptor(pm); 864 } 865 866 KASSERT(po->po_sscount == 0, 867 ("[pmc,%d] SS count not zero", __LINE__)); 868 KASSERT(LIST_EMPTY(&po->po_pmcs), 869 ("[pmc,%d] PMC list not empty", __LINE__)); 870 871 /* de-configure the log file if present */ 872 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 873 pmclog_deconfigure_log(po); 874 } 875 876 /* 877 * Remove an owner process record if all conditions are met. 878 */ 879 static void 880 pmc_maybe_remove_owner(struct pmc_owner *po) 881 { 882 883 PMCDBG1(OWN,OMR,1, "maybe-remove-owner po=%p", po); 884 885 /* 886 * Remove owner record if 887 * - this process does not own any PMCs 888 * - this process has not allocated a system-wide sampling buffer 889 */ 890 if (LIST_EMPTY(&po->po_pmcs) && 891 ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)) { 892 pmc_remove_owner(po); 893 pmc_destroy_owner_descriptor(po); 894 } 895 } 896 897 /* 898 * Add an association between a target process and a PMC. 899 */ 900 static void 901 pmc_link_target_process(struct pmc *pm, struct pmc_process *pp) 902 { 903 struct pmc_target *pt; 904 struct pmc_thread *pt_td __diagused; 905 int ri; 906 907 sx_assert(&pmc_sx, SX_XLOCKED); 908 KASSERT(pm != NULL && pp != NULL, 909 ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp)); 910 KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)), 911 ("[pmc,%d] Attaching a non-process-virtual pmc=%p to pid=%d", 912 __LINE__, pm, pp->pp_proc->p_pid)); 913 KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= ((int) md->pmd_npmc - 1), 914 ("[pmc,%d] Illegal reference count %d for process record %p", 915 __LINE__, pp->pp_refcnt, (void *) pp)); 916 917 ri = PMC_TO_ROWINDEX(pm); 918 919 PMCDBG3(PRC,TLK,1, "link-target pmc=%p ri=%d pmc-process=%p", 920 pm, ri, pp); 921 922 #ifdef HWPMC_DEBUG 923 LIST_FOREACH(pt, &pm->pm_targets, pt_next) { 924 if (pt->pt_process == pp) 925 KASSERT(0, ("[pmc,%d] pp %p already in pmc %p targets", 926 __LINE__, pp, pm)); 927 } 928 #endif 929 pt = malloc(sizeof(struct pmc_target), M_PMC, M_WAITOK | M_ZERO); 930 pt->pt_process = pp; 931 932 LIST_INSERT_HEAD(&pm->pm_targets, pt, pt_next); 933 934 atomic_store_rel_ptr((uintptr_t *)&pp->pp_pmcs[ri].pp_pmc, 935 (uintptr_t)pm); 936 937 if (pm->pm_owner->po_owner == pp->pp_proc) 938 pm->pm_flags |= PMC_F_ATTACHED_TO_OWNER; 939 940 /* 941 * Initialize the per-process values at this row index. 942 */ 943 pp->pp_pmcs[ri].pp_pmcval = PMC_TO_MODE(pm) == PMC_MODE_TS ? 944 pm->pm_sc.pm_reloadcount : 0; 945 pp->pp_refcnt++; 946 947 #ifdef INVARIANTS 948 /* Confirm that the per-thread values at this row index are cleared. */ 949 if (PMC_TO_MODE(pm) == PMC_MODE_TS) { 950 mtx_lock_spin(pp->pp_tdslock); 951 LIST_FOREACH(pt_td, &pp->pp_tds, pt_next) { 952 KASSERT(pt_td->pt_pmcs[ri].pt_pmcval == (pmc_value_t) 0, 953 ("[pmc,%d] pt_pmcval not cleared for pid=%d at " 954 "ri=%d", __LINE__, pp->pp_proc->p_pid, ri)); 955 } 956 mtx_unlock_spin(pp->pp_tdslock); 957 } 958 #endif 959 } 960 961 /* 962 * Removes the association between a target process and a PMC. 963 */ 964 static void 965 pmc_unlink_target_process(struct pmc *pm, struct pmc_process *pp) 966 { 967 int ri; 968 struct proc *p; 969 struct pmc_target *ptgt; 970 struct pmc_thread *pt; 971 972 sx_assert(&pmc_sx, SX_XLOCKED); 973 974 KASSERT(pm != NULL && pp != NULL, 975 ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp)); 976 977 KASSERT(pp->pp_refcnt >= 1 && pp->pp_refcnt <= (int) md->pmd_npmc, 978 ("[pmc,%d] Illegal ref count %d on process record %p", 979 __LINE__, pp->pp_refcnt, (void *) pp)); 980 981 ri = PMC_TO_ROWINDEX(pm); 982 983 PMCDBG3(PRC,TUL,1, "unlink-target pmc=%p ri=%d pmc-process=%p", 984 pm, ri, pp); 985 986 KASSERT(pp->pp_pmcs[ri].pp_pmc == pm, 987 ("[pmc,%d] PMC ri %d mismatch pmc %p pp->[ri] %p", __LINE__, 988 ri, pm, pp->pp_pmcs[ri].pp_pmc)); 989 990 pp->pp_pmcs[ri].pp_pmc = NULL; 991 pp->pp_pmcs[ri].pp_pmcval = (pmc_value_t)0; 992 993 /* Clear the per-thread values at this row index. */ 994 if (PMC_TO_MODE(pm) == PMC_MODE_TS) { 995 mtx_lock_spin(pp->pp_tdslock); 996 LIST_FOREACH(pt, &pp->pp_tds, pt_next) 997 pt->pt_pmcs[ri].pt_pmcval = (pmc_value_t)0; 998 mtx_unlock_spin(pp->pp_tdslock); 999 } 1000 1001 /* Remove owner-specific flags */ 1002 if (pm->pm_owner->po_owner == pp->pp_proc) { 1003 pp->pp_flags &= ~PMC_PP_ENABLE_MSR_ACCESS; 1004 pm->pm_flags &= ~PMC_F_ATTACHED_TO_OWNER; 1005 } 1006 1007 pp->pp_refcnt--; 1008 1009 /* Remove the target process from the PMC structure */ 1010 LIST_FOREACH(ptgt, &pm->pm_targets, pt_next) 1011 if (ptgt->pt_process == pp) 1012 break; 1013 1014 KASSERT(ptgt != NULL, ("[pmc,%d] process %p (pp: %p) not found " 1015 "in pmc %p", __LINE__, pp->pp_proc, pp, pm)); 1016 1017 LIST_REMOVE(ptgt, pt_next); 1018 free(ptgt, M_PMC); 1019 1020 /* if the PMC now lacks targets, send the owner a SIGIO */ 1021 if (LIST_EMPTY(&pm->pm_targets)) { 1022 p = pm->pm_owner->po_owner; 1023 PROC_LOCK(p); 1024 kern_psignal(p, SIGIO); 1025 PROC_UNLOCK(p); 1026 1027 PMCDBG2(PRC,SIG,2, "signalling proc=%p signal=%d", p, SIGIO); 1028 } 1029 } 1030 1031 /* 1032 * Check if PMC 'pm' may be attached to target process 't'. 1033 */ 1034 1035 static bool 1036 pmc_can_attach(struct pmc *pm, struct proc *t) 1037 { 1038 struct proc *o; /* pmc owner */ 1039 struct ucred *oc, *tc; /* owner, target credentials */ 1040 bool decline_attach; 1041 1042 /* 1043 * A PMC's owner can always attach that PMC to itself. 1044 */ 1045 1046 if ((o = pm->pm_owner->po_owner) == t) 1047 return (true); 1048 1049 PROC_LOCK(o); 1050 oc = o->p_ucred; 1051 crhold(oc); 1052 PROC_UNLOCK(o); 1053 1054 PROC_LOCK(t); 1055 tc = t->p_ucred; 1056 crhold(tc); 1057 PROC_UNLOCK(t); 1058 1059 /* 1060 * The effective uid of the PMC owner should match at least one 1061 * of the {effective,real,saved} uids of the target process. 1062 */ 1063 1064 decline_attach = oc->cr_uid != tc->cr_uid && 1065 oc->cr_uid != tc->cr_svuid && 1066 oc->cr_uid != tc->cr_ruid; 1067 1068 /* 1069 * Every one of the target's group ids, must be in the owner's 1070 * group list. 1071 */ 1072 for (int i = 0; !decline_attach && i < tc->cr_ngroups; i++) 1073 decline_attach = !groupmember(tc->cr_groups[i], oc); 1074 if (!decline_attach) 1075 decline_attach = !groupmember(tc->cr_gid, oc) || 1076 !groupmember(tc->cr_rgid, oc) || 1077 !groupmember(tc->cr_svgid, oc); 1078 1079 crfree(tc); 1080 crfree(oc); 1081 1082 return (!decline_attach); 1083 } 1084 1085 /* 1086 * Attach a process to a PMC. 1087 */ 1088 static int 1089 pmc_attach_one_process(struct proc *p, struct pmc *pm) 1090 { 1091 int ri, error; 1092 char *fullpath, *freepath; 1093 struct pmc_process *pp; 1094 1095 sx_assert(&pmc_sx, SX_XLOCKED); 1096 1097 PMCDBG5(PRC,ATT,2, "attach-one pm=%p ri=%d proc=%p (%d, %s)", pm, 1098 PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm); 1099 1100 /* 1101 * Locate the process descriptor corresponding to process 'p', 1102 * allocating space as needed. 1103 * 1104 * Verify that rowindex 'pm_rowindex' is free in the process 1105 * descriptor. 1106 * 1107 * If not, allocate space for a descriptor and link the 1108 * process descriptor and PMC. 1109 */ 1110 ri = PMC_TO_ROWINDEX(pm); 1111 1112 /* mark process as using HWPMCs */ 1113 PROC_LOCK(p); 1114 p->p_flag |= P_HWPMC; 1115 PROC_UNLOCK(p); 1116 1117 if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_ALLOCATE)) == NULL) { 1118 error = ENOMEM; 1119 goto fail; 1120 } 1121 1122 if (pp->pp_pmcs[ri].pp_pmc == pm) {/* already present at slot [ri] */ 1123 error = EEXIST; 1124 goto fail; 1125 } 1126 1127 if (pp->pp_pmcs[ri].pp_pmc != NULL) { 1128 error = EBUSY; 1129 goto fail; 1130 } 1131 1132 pmc_link_target_process(pm, pp); 1133 1134 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) && 1135 (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) == 0) 1136 pm->pm_flags |= PMC_F_NEEDS_LOGFILE; 1137 1138 pm->pm_flags |= PMC_F_ATTACH_DONE; /* mark as attached */ 1139 1140 /* issue an attach event to a configured log file */ 1141 if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE) { 1142 if (p->p_flag & P_KPROC) { 1143 fullpath = kernelname; 1144 freepath = NULL; 1145 } else { 1146 pmc_getfilename(p->p_textvp, &fullpath, &freepath); 1147 pmclog_process_pmcattach(pm, p->p_pid, fullpath); 1148 } 1149 free(freepath, M_TEMP); 1150 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) 1151 pmc_log_process_mappings(pm->pm_owner, p); 1152 } 1153 1154 return (0); 1155 fail: 1156 PROC_LOCK(p); 1157 p->p_flag &= ~P_HWPMC; 1158 PROC_UNLOCK(p); 1159 return (error); 1160 } 1161 1162 /* 1163 * Attach a process and optionally its children 1164 */ 1165 static int 1166 pmc_attach_process(struct proc *p, struct pmc *pm) 1167 { 1168 int error; 1169 struct proc *top; 1170 1171 sx_assert(&pmc_sx, SX_XLOCKED); 1172 1173 PMCDBG5(PRC,ATT,1, "attach pm=%p ri=%d proc=%p (%d, %s)", pm, 1174 PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm); 1175 1176 /* 1177 * If this PMC successfully allowed a GETMSR operation 1178 * in the past, disallow further ATTACHes. 1179 */ 1180 if ((pm->pm_flags & PMC_PP_ENABLE_MSR_ACCESS) != 0) 1181 return (EPERM); 1182 1183 if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0) 1184 return (pmc_attach_one_process(p, pm)); 1185 1186 /* 1187 * Traverse all child processes, attaching them to 1188 * this PMC. 1189 */ 1190 sx_slock(&proctree_lock); 1191 1192 top = p; 1193 for (;;) { 1194 if ((error = pmc_attach_one_process(p, pm)) != 0) 1195 break; 1196 if (!LIST_EMPTY(&p->p_children)) 1197 p = LIST_FIRST(&p->p_children); 1198 else for (;;) { 1199 if (p == top) 1200 goto done; 1201 if (LIST_NEXT(p, p_sibling)) { 1202 p = LIST_NEXT(p, p_sibling); 1203 break; 1204 } 1205 p = p->p_pptr; 1206 } 1207 } 1208 1209 if (error != 0) 1210 (void)pmc_detach_process(top, pm); 1211 1212 done: 1213 sx_sunlock(&proctree_lock); 1214 return (error); 1215 } 1216 1217 /* 1218 * Detach a process from a PMC. If there are no other PMCs tracking 1219 * this process, remove the process structure from its hash table. If 1220 * 'flags' contains PMC_FLAG_REMOVE, then free the process structure. 1221 */ 1222 static int 1223 pmc_detach_one_process(struct proc *p, struct pmc *pm, int flags) 1224 { 1225 int ri; 1226 struct pmc_process *pp; 1227 1228 sx_assert(&pmc_sx, SX_XLOCKED); 1229 1230 KASSERT(pm != NULL, 1231 ("[pmc,%d] null pm pointer", __LINE__)); 1232 1233 ri = PMC_TO_ROWINDEX(pm); 1234 1235 PMCDBG6(PRC,ATT,2, "detach-one pm=%p ri=%d proc=%p (%d, %s) flags=0x%x", 1236 pm, ri, p, p->p_pid, p->p_comm, flags); 1237 1238 if ((pp = pmc_find_process_descriptor(p, 0)) == NULL) 1239 return (ESRCH); 1240 1241 if (pp->pp_pmcs[ri].pp_pmc != pm) 1242 return (EINVAL); 1243 1244 pmc_unlink_target_process(pm, pp); 1245 1246 /* Issue a detach entry if a log file is configured */ 1247 if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE) 1248 pmclog_process_pmcdetach(pm, p->p_pid); 1249 1250 /* 1251 * If there are no PMCs targeting this process, we remove its 1252 * descriptor from the target hash table and unset the P_HWPMC 1253 * flag in the struct proc. 1254 */ 1255 KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= (int) md->pmd_npmc, 1256 ("[pmc,%d] Illegal refcnt %d for process struct %p", 1257 __LINE__, pp->pp_refcnt, pp)); 1258 1259 if (pp->pp_refcnt != 0) /* still a target of some PMC */ 1260 return (0); 1261 1262 pmc_remove_process_descriptor(pp); 1263 1264 if (flags & PMC_FLAG_REMOVE) 1265 pmc_destroy_process_descriptor(pp); 1266 1267 PROC_LOCK(p); 1268 p->p_flag &= ~P_HWPMC; 1269 PROC_UNLOCK(p); 1270 1271 return (0); 1272 } 1273 1274 /* 1275 * Detach a process and optionally its descendants from a PMC. 1276 */ 1277 static int 1278 pmc_detach_process(struct proc *p, struct pmc *pm) 1279 { 1280 struct proc *top; 1281 1282 sx_assert(&pmc_sx, SX_XLOCKED); 1283 1284 PMCDBG5(PRC,ATT,1, "detach pm=%p ri=%d proc=%p (%d, %s)", pm, 1285 PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm); 1286 1287 if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0) 1288 return (pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE)); 1289 1290 /* 1291 * Traverse all children, detaching them from this PMC. We 1292 * ignore errors since we could be detaching a PMC from a 1293 * partially attached proc tree. 1294 */ 1295 sx_slock(&proctree_lock); 1296 1297 top = p; 1298 for (;;) { 1299 (void)pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE); 1300 1301 if (!LIST_EMPTY(&p->p_children)) { 1302 p = LIST_FIRST(&p->p_children); 1303 } else { 1304 for (;;) { 1305 if (p == top) 1306 goto done; 1307 if (LIST_NEXT(p, p_sibling)) { 1308 p = LIST_NEXT(p, p_sibling); 1309 break; 1310 } 1311 p = p->p_pptr; 1312 } 1313 } 1314 } 1315 done: 1316 sx_sunlock(&proctree_lock); 1317 if (LIST_EMPTY(&pm->pm_targets)) 1318 pm->pm_flags &= ~PMC_F_ATTACH_DONE; 1319 1320 return (0); 1321 } 1322 1323 /* 1324 * Handle events after an exec() for a process: 1325 * - Inform log owners of the new exec() event 1326 * - Release any PMCs owned by the process before the exec() 1327 * - Detach PMCs from the target if required 1328 */ 1329 static void 1330 pmc_process_exec(struct thread *td, struct pmckern_procexec *pk) 1331 { 1332 struct pmc *pm; 1333 struct pmc_owner *po; 1334 struct pmc_process *pp; 1335 struct proc *p; 1336 char *fullpath, *freepath; 1337 u_int ri; 1338 bool is_using_hwpmcs; 1339 1340 sx_assert(&pmc_sx, SX_XLOCKED); 1341 1342 p = td->td_proc; 1343 pmc_getfilename(p->p_textvp, &fullpath, &freepath); 1344 1345 PMC_EPOCH_ENTER(); 1346 /* Inform owners of SS mode PMCs of the exec event. */ 1347 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) { 1348 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) != 0) { 1349 pmclog_process_procexec(po, PMC_ID_INVALID, p->p_pid, 1350 pk->pm_baseaddr, pk->pm_dynaddr, fullpath); 1351 } 1352 } 1353 PMC_EPOCH_EXIT(); 1354 1355 PROC_LOCK(p); 1356 is_using_hwpmcs = (p->p_flag & P_HWPMC) != 0; 1357 PROC_UNLOCK(p); 1358 1359 if (!is_using_hwpmcs) { 1360 if (freepath != NULL) 1361 free(freepath, M_TEMP); 1362 return; 1363 } 1364 1365 /* 1366 * PMCs are not inherited across an exec(): remove any PMCs that this 1367 * process is the owner of. 1368 */ 1369 if ((po = pmc_find_owner_descriptor(p)) != NULL) { 1370 pmc_remove_owner(po); 1371 pmc_destroy_owner_descriptor(po); 1372 } 1373 1374 /* 1375 * If the process being exec'ed is not the target of any PMC, we are 1376 * done. 1377 */ 1378 if ((pp = pmc_find_process_descriptor(p, 0)) == NULL) { 1379 if (freepath != NULL) 1380 free(freepath, M_TEMP); 1381 return; 1382 } 1383 1384 /* 1385 * Log the exec event to all monitoring owners. Skip owners who have 1386 * already received the event because they had system sampling PMCs 1387 * active. 1388 */ 1389 for (ri = 0; ri < md->pmd_npmc; ri++) { 1390 if ((pm = pp->pp_pmcs[ri].pp_pmc) == NULL) 1391 continue; 1392 1393 po = pm->pm_owner; 1394 if (po->po_sscount == 0 && 1395 (po->po_flags & PMC_PO_OWNS_LOGFILE) != 0) { 1396 pmclog_process_procexec(po, pm->pm_id, p->p_pid, 1397 pk->pm_baseaddr, pk->pm_dynaddr, fullpath); 1398 } 1399 } 1400 1401 if (freepath != NULL) 1402 free(freepath, M_TEMP); 1403 1404 PMCDBG4(PRC,EXC,1, "exec proc=%p (%d, %s) cred-changed=%d", 1405 p, p->p_pid, p->p_comm, pk->pm_credentialschanged); 1406 1407 if (pk->pm_credentialschanged == 0) /* no change */ 1408 return; 1409 1410 /* 1411 * If the newly exec()'ed process has a different credential 1412 * than before, allow it to be the target of a PMC only if 1413 * the PMC's owner has sufficient privilege. 1414 */ 1415 for (ri = 0; ri < md->pmd_npmc; ri++) { 1416 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) { 1417 if (pmc_can_attach(pm, td->td_proc)) { 1418 pmc_detach_one_process(td->td_proc, pm, 1419 PMC_FLAG_NONE); 1420 } 1421 } 1422 } 1423 1424 KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= md->pmd_npmc, 1425 ("[pmc,%d] Illegal ref count %u on pp %p", __LINE__, 1426 pp->pp_refcnt, pp)); 1427 1428 /* 1429 * If this process is no longer the target of any 1430 * PMCs, we can remove the process entry and free 1431 * up space. 1432 */ 1433 if (pp->pp_refcnt == 0) { 1434 pmc_remove_process_descriptor(pp); 1435 pmc_destroy_process_descriptor(pp); 1436 } 1437 } 1438 1439 /* 1440 * Thread context switch IN. 1441 */ 1442 static void 1443 pmc_process_csw_in(struct thread *td) 1444 { 1445 struct pmc *pm; 1446 struct pmc_classdep *pcd; 1447 struct pmc_cpu *pc; 1448 struct pmc_hw *phw __diagused; 1449 struct pmc_process *pp; 1450 struct pmc_thread *pt; 1451 struct proc *p; 1452 pmc_value_t newvalue; 1453 int cpu; 1454 u_int adjri, ri; 1455 1456 p = td->td_proc; 1457 pt = NULL; 1458 if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE)) == NULL) 1459 return; 1460 1461 KASSERT(pp->pp_proc == td->td_proc, 1462 ("[pmc,%d] not my thread state", __LINE__)); 1463 1464 critical_enter(); /* no preemption from this point */ 1465 1466 cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */ 1467 1468 PMCDBG5(CSW,SWI,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p, 1469 p->p_pid, p->p_comm, pp); 1470 1471 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), 1472 ("[pmc,%d] weird CPU id %d", __LINE__, cpu)); 1473 1474 pc = pmc_pcpu[cpu]; 1475 for (ri = 0; ri < md->pmd_npmc; ri++) { 1476 if ((pm = pp->pp_pmcs[ri].pp_pmc) == NULL) 1477 continue; 1478 1479 KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)), 1480 ("[pmc,%d] Target PMC in non-virtual mode (%d)", 1481 __LINE__, PMC_TO_MODE(pm))); 1482 KASSERT(PMC_TO_ROWINDEX(pm) == ri, 1483 ("[pmc,%d] Row index mismatch pmc %d != ri %d", 1484 __LINE__, PMC_TO_ROWINDEX(pm), ri)); 1485 1486 /* 1487 * Only PMCs that are marked as 'RUNNING' need 1488 * be placed on hardware. 1489 */ 1490 if (pm->pm_state != PMC_STATE_RUNNING) 1491 continue; 1492 1493 KASSERT(counter_u64_fetch(pm->pm_runcount) >= 0, 1494 ("[pmc,%d] pm=%p runcount %ju", __LINE__, pm, 1495 (uintmax_t)counter_u64_fetch(pm->pm_runcount))); 1496 1497 /* increment PMC runcount */ 1498 counter_u64_add(pm->pm_runcount, 1); 1499 1500 /* configure the HWPMC we are going to use. */ 1501 pcd = pmc_ri_to_classdep(md, ri, &adjri); 1502 (void)pcd->pcd_config_pmc(cpu, adjri, pm); 1503 1504 phw = pc->pc_hwpmcs[ri]; 1505 1506 KASSERT(phw != NULL, 1507 ("[pmc,%d] null hw pointer", __LINE__)); 1508 1509 KASSERT(phw->phw_pmc == pm, 1510 ("[pmc,%d] hw->pmc %p != pmc %p", __LINE__, 1511 phw->phw_pmc, pm)); 1512 1513 /* 1514 * Write out saved value and start the PMC. 1515 * 1516 * Sampling PMCs use a per-thread value, while 1517 * counting mode PMCs use a per-pmc value that is 1518 * inherited across descendants. 1519 */ 1520 if (PMC_TO_MODE(pm) == PMC_MODE_TS) { 1521 if (pt == NULL) 1522 pt = pmc_find_thread_descriptor(pp, td, 1523 PMC_FLAG_NONE); 1524 1525 KASSERT(pt != NULL, 1526 ("[pmc,%d] No thread found for td=%p", __LINE__, 1527 td)); 1528 1529 mtx_pool_lock_spin(pmc_mtxpool, pm); 1530 1531 /* 1532 * If we have a thread descriptor, use the per-thread 1533 * counter in the descriptor. If not, we will use 1534 * a per-process counter. 1535 * 1536 * TODO: Remove the per-process "safety net" once 1537 * we have thoroughly tested that we don't hit the 1538 * above assert. 1539 */ 1540 if (pt != NULL) { 1541 if (pt->pt_pmcs[ri].pt_pmcval > 0) 1542 newvalue = pt->pt_pmcs[ri].pt_pmcval; 1543 else 1544 newvalue = pm->pm_sc.pm_reloadcount; 1545 } else { 1546 /* 1547 * Use the saved value calculated after the most 1548 * recent time a thread using the shared counter 1549 * switched out. Reset the saved count in case 1550 * another thread from this process switches in 1551 * before any threads switch out. 1552 */ 1553 newvalue = pp->pp_pmcs[ri].pp_pmcval; 1554 pp->pp_pmcs[ri].pp_pmcval = 1555 pm->pm_sc.pm_reloadcount; 1556 } 1557 mtx_pool_unlock_spin(pmc_mtxpool, pm); 1558 KASSERT(newvalue > 0 && newvalue <= 1559 pm->pm_sc.pm_reloadcount, 1560 ("[pmc,%d] pmcval outside of expected range cpu=%d " 1561 "ri=%d pmcval=%jx pm_reloadcount=%jx", __LINE__, 1562 cpu, ri, newvalue, pm->pm_sc.pm_reloadcount)); 1563 } else { 1564 KASSERT(PMC_TO_MODE(pm) == PMC_MODE_TC, 1565 ("[pmc,%d] illegal mode=%d", __LINE__, 1566 PMC_TO_MODE(pm))); 1567 mtx_pool_lock_spin(pmc_mtxpool, pm); 1568 newvalue = PMC_PCPU_SAVED(cpu, ri) = 1569 pm->pm_gv.pm_savedvalue; 1570 mtx_pool_unlock_spin(pmc_mtxpool, pm); 1571 } 1572 1573 PMCDBG3(CSW,SWI,1,"cpu=%d ri=%d new=%jd", cpu, ri, newvalue); 1574 1575 (void)pcd->pcd_write_pmc(cpu, adjri, pm, newvalue); 1576 1577 /* If a sampling mode PMC, reset stalled state. */ 1578 if (PMC_TO_MODE(pm) == PMC_MODE_TS) 1579 pm->pm_pcpu_state[cpu].pps_stalled = 0; 1580 1581 /* Indicate that we desire this to run. */ 1582 pm->pm_pcpu_state[cpu].pps_cpustate = 1; 1583 1584 /* Start the PMC. */ 1585 (void)pcd->pcd_start_pmc(cpu, adjri, pm); 1586 } 1587 1588 /* 1589 * Perform any other architecture/cpu dependent thread 1590 * switch-in actions. 1591 */ 1592 (void)(*md->pmd_switch_in)(pc, pp); 1593 1594 critical_exit(); 1595 } 1596 1597 /* 1598 * Thread context switch OUT. 1599 */ 1600 static void 1601 pmc_process_csw_out(struct thread *td) 1602 { 1603 struct pmc *pm; 1604 struct pmc_classdep *pcd; 1605 struct pmc_cpu *pc; 1606 struct pmc_process *pp; 1607 struct pmc_thread *pt = NULL; 1608 struct proc *p; 1609 pmc_value_t newvalue; 1610 int64_t tmp; 1611 enum pmc_mode mode; 1612 int cpu; 1613 u_int adjri, ri; 1614 1615 /* 1616 * Locate our process descriptor; this may be NULL if 1617 * this process is exiting and we have already removed 1618 * the process from the target process table. 1619 * 1620 * Note that due to kernel preemption, multiple 1621 * context switches may happen while the process is 1622 * exiting. 1623 * 1624 * Note also that if the target process cannot be 1625 * found we still need to deconfigure any PMCs that 1626 * are currently running on hardware. 1627 */ 1628 p = td->td_proc; 1629 pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE); 1630 1631 critical_enter(); 1632 1633 cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */ 1634 1635 PMCDBG5(CSW,SWO,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p, 1636 p->p_pid, p->p_comm, pp); 1637 1638 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), 1639 ("[pmc,%d weird CPU id %d", __LINE__, cpu)); 1640 1641 pc = pmc_pcpu[cpu]; 1642 1643 /* 1644 * When a PMC gets unlinked from a target PMC, it will 1645 * be removed from the target's pp_pmc[] array. 1646 * 1647 * However, on a MP system, the target could have been 1648 * executing on another CPU at the time of the unlink. 1649 * So, at context switch OUT time, we need to look at 1650 * the hardware to determine if a PMC is scheduled on 1651 * it. 1652 */ 1653 for (ri = 0; ri < md->pmd_npmc; ri++) { 1654 pcd = pmc_ri_to_classdep(md, ri, &adjri); 1655 pm = NULL; 1656 (void)(*pcd->pcd_get_config)(cpu, adjri, &pm); 1657 1658 if (pm == NULL) /* nothing at this row index */ 1659 continue; 1660 1661 mode = PMC_TO_MODE(pm); 1662 if (!PMC_IS_VIRTUAL_MODE(mode)) 1663 continue; /* not a process virtual PMC */ 1664 1665 KASSERT(PMC_TO_ROWINDEX(pm) == ri, 1666 ("[pmc,%d] ri mismatch pmc(%d) ri(%d)", 1667 __LINE__, PMC_TO_ROWINDEX(pm), ri)); 1668 1669 /* 1670 * Change desired state, and then stop if not stalled. 1671 * This two-step dance should avoid race conditions where 1672 * an interrupt re-enables the PMC after this code has 1673 * already checked the pm_stalled flag. 1674 */ 1675 pm->pm_pcpu_state[cpu].pps_cpustate = 0; 1676 if (pm->pm_pcpu_state[cpu].pps_stalled == 0) 1677 (void)pcd->pcd_stop_pmc(cpu, adjri, pm); 1678 1679 KASSERT(counter_u64_fetch(pm->pm_runcount) > 0, 1680 ("[pmc,%d] pm=%p runcount %ju", __LINE__, pm, 1681 (uintmax_t)counter_u64_fetch(pm->pm_runcount))); 1682 1683 /* reduce this PMC's runcount */ 1684 counter_u64_add(pm->pm_runcount, -1); 1685 1686 /* 1687 * If this PMC is associated with this process, 1688 * save the reading. 1689 */ 1690 if (pm->pm_state != PMC_STATE_DELETED && pp != NULL && 1691 pp->pp_pmcs[ri].pp_pmc != NULL) { 1692 KASSERT(pm == pp->pp_pmcs[ri].pp_pmc, 1693 ("[pmc,%d] pm %p != pp_pmcs[%d] %p", __LINE__, 1694 pm, ri, pp->pp_pmcs[ri].pp_pmc)); 1695 KASSERT(pp->pp_refcnt > 0, 1696 ("[pmc,%d] pp refcnt = %d", __LINE__, 1697 pp->pp_refcnt)); 1698 1699 (void)pcd->pcd_read_pmc(cpu, adjri, pm, &newvalue); 1700 1701 if (mode == PMC_MODE_TS) { 1702 PMCDBG3(CSW,SWO,1,"cpu=%d ri=%d val=%jd (samp)", 1703 cpu, ri, newvalue); 1704 1705 if (pt == NULL) 1706 pt = pmc_find_thread_descriptor(pp, td, 1707 PMC_FLAG_NONE); 1708 1709 KASSERT(pt != NULL, 1710 ("[pmc,%d] No thread found for td=%p", 1711 __LINE__, td)); 1712 1713 mtx_pool_lock_spin(pmc_mtxpool, pm); 1714 1715 /* 1716 * If we have a thread descriptor, save the 1717 * per-thread counter in the descriptor. If not, 1718 * we will update the per-process counter. 1719 * 1720 * TODO: Remove the per-process "safety net" 1721 * once we have thoroughly tested that we 1722 * don't hit the above assert. 1723 */ 1724 if (pt != NULL) { 1725 pt->pt_pmcs[ri].pt_pmcval = newvalue; 1726 } else { 1727 /* 1728 * For sampling process-virtual PMCs, 1729 * newvalue is the number of events to 1730 * be seen until the next sampling 1731 * interrupt. We can just add the events 1732 * left from this invocation to the 1733 * counter, then adjust in case we 1734 * overflow our range. 1735 * 1736 * (Recall that we reload the counter 1737 * every time we use it.) 1738 */ 1739 pp->pp_pmcs[ri].pp_pmcval += newvalue; 1740 if (pp->pp_pmcs[ri].pp_pmcval > 1741 pm->pm_sc.pm_reloadcount) { 1742 pp->pp_pmcs[ri].pp_pmcval -= 1743 pm->pm_sc.pm_reloadcount; 1744 } 1745 } 1746 mtx_pool_unlock_spin(pmc_mtxpool, pm); 1747 } else { 1748 tmp = newvalue - PMC_PCPU_SAVED(cpu, ri); 1749 1750 PMCDBG3(CSW,SWO,1,"cpu=%d ri=%d tmp=%jd (count)", 1751 cpu, ri, tmp); 1752 1753 /* 1754 * For counting process-virtual PMCs, 1755 * we expect the count to be 1756 * increasing monotonically, modulo a 64 1757 * bit wraparound. 1758 */ 1759 KASSERT(tmp >= 0, 1760 ("[pmc,%d] negative increment cpu=%d " 1761 "ri=%d newvalue=%jx saved=%jx " 1762 "incr=%jx", __LINE__, cpu, ri, 1763 newvalue, PMC_PCPU_SAVED(cpu, ri), tmp)); 1764 1765 mtx_pool_lock_spin(pmc_mtxpool, pm); 1766 pm->pm_gv.pm_savedvalue += tmp; 1767 pp->pp_pmcs[ri].pp_pmcval += tmp; 1768 mtx_pool_unlock_spin(pmc_mtxpool, pm); 1769 1770 if (pm->pm_flags & PMC_F_LOG_PROCCSW) 1771 pmclog_process_proccsw(pm, pp, tmp, td); 1772 } 1773 } 1774 1775 /* Mark hardware as free. */ 1776 (void)pcd->pcd_config_pmc(cpu, adjri, NULL); 1777 } 1778 1779 /* 1780 * Perform any other architecture/cpu dependent thread 1781 * switch out functions. 1782 */ 1783 (void)(*md->pmd_switch_out)(pc, pp); 1784 1785 critical_exit(); 1786 } 1787 1788 /* 1789 * A new thread for a process. 1790 */ 1791 static void 1792 pmc_process_thread_add(struct thread *td) 1793 { 1794 struct pmc_process *pmc; 1795 1796 pmc = pmc_find_process_descriptor(td->td_proc, PMC_FLAG_NONE); 1797 if (pmc != NULL) 1798 pmc_find_thread_descriptor(pmc, td, PMC_FLAG_ALLOCATE); 1799 } 1800 1801 /* 1802 * A thread delete for a process. 1803 */ 1804 static void 1805 pmc_process_thread_delete(struct thread *td) 1806 { 1807 struct pmc_process *pmc; 1808 1809 pmc = pmc_find_process_descriptor(td->td_proc, PMC_FLAG_NONE); 1810 if (pmc != NULL) 1811 pmc_thread_descriptor_pool_free(pmc_find_thread_descriptor(pmc, 1812 td, PMC_FLAG_REMOVE)); 1813 } 1814 1815 /* 1816 * A userret() call for a thread. 1817 */ 1818 static void 1819 pmc_process_thread_userret(struct thread *td) 1820 { 1821 sched_pin(); 1822 pmc_capture_user_callchain(curcpu, PMC_UR, td->td_frame); 1823 sched_unpin(); 1824 } 1825 1826 /* 1827 * A mapping change for a process. 1828 */ 1829 static void 1830 pmc_process_mmap(struct thread *td, struct pmckern_map_in *pkm) 1831 { 1832 const struct pmc *pm; 1833 const struct pmc_process *pp; 1834 struct pmc_owner *po; 1835 char *fullpath, *freepath; 1836 pid_t pid; 1837 int ri; 1838 1839 MPASS(!in_epoch(global_epoch_preempt)); 1840 1841 freepath = fullpath = NULL; 1842 pmc_getfilename((struct vnode *)pkm->pm_file, &fullpath, &freepath); 1843 1844 pid = td->td_proc->p_pid; 1845 1846 PMC_EPOCH_ENTER(); 1847 /* Inform owners of all system-wide sampling PMCs. */ 1848 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) { 1849 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 1850 pmclog_process_map_in(po, pid, pkm->pm_address, 1851 fullpath); 1852 } 1853 1854 if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL) 1855 goto done; 1856 1857 /* 1858 * Inform sampling PMC owners tracking this process. 1859 */ 1860 for (ri = 0; ri < md->pmd_npmc; ri++) { 1861 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL && 1862 PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) { 1863 pmclog_process_map_in(pm->pm_owner, 1864 pid, pkm->pm_address, fullpath); 1865 } 1866 } 1867 1868 done: 1869 if (freepath != NULL) 1870 free(freepath, M_TEMP); 1871 PMC_EPOCH_EXIT(); 1872 } 1873 1874 /* 1875 * Log an munmap request. 1876 */ 1877 static void 1878 pmc_process_munmap(struct thread *td, struct pmckern_map_out *pkm) 1879 { 1880 const struct pmc *pm; 1881 const struct pmc_process *pp; 1882 struct pmc_owner *po; 1883 pid_t pid; 1884 int ri; 1885 1886 pid = td->td_proc->p_pid; 1887 1888 PMC_EPOCH_ENTER(); 1889 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) { 1890 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 1891 pmclog_process_map_out(po, pid, pkm->pm_address, 1892 pkm->pm_address + pkm->pm_size); 1893 } 1894 PMC_EPOCH_EXIT(); 1895 1896 if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL) 1897 return; 1898 1899 for (ri = 0; ri < md->pmd_npmc; ri++) { 1900 pm = pp->pp_pmcs[ri].pp_pmc; 1901 if (pm != NULL && PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) { 1902 pmclog_process_map_out(pm->pm_owner, pid, 1903 pkm->pm_address, pkm->pm_address + pkm->pm_size); 1904 } 1905 } 1906 } 1907 1908 /* 1909 * Log mapping information about the kernel. 1910 */ 1911 static void 1912 pmc_log_kernel_mappings(struct pmc *pm) 1913 { 1914 struct pmc_owner *po; 1915 struct pmckern_map_in *km, *kmbase; 1916 1917 MPASS(in_epoch(global_epoch_preempt) || sx_xlocked(&pmc_sx)); 1918 KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)), 1919 ("[pmc,%d] non-sampling PMC (%p) desires mapping information", 1920 __LINE__, (void *) pm)); 1921 1922 po = pm->pm_owner; 1923 if ((po->po_flags & PMC_PO_INITIAL_MAPPINGS_DONE) != 0) 1924 return; 1925 1926 if (PMC_TO_MODE(pm) == PMC_MODE_SS) 1927 pmc_process_allproc(pm); 1928 1929 /* 1930 * Log the current set of kernel modules. 1931 */ 1932 kmbase = linker_hwpmc_list_objects(); 1933 for (km = kmbase; km->pm_file != NULL; km++) { 1934 PMCDBG2(LOG,REG,1,"%s %p", (char *)km->pm_file, 1935 (void *)km->pm_address); 1936 pmclog_process_map_in(po, (pid_t)-1, km->pm_address, 1937 km->pm_file); 1938 } 1939 free(kmbase, M_LINKER); 1940 1941 po->po_flags |= PMC_PO_INITIAL_MAPPINGS_DONE; 1942 } 1943 1944 /* 1945 * Log the mappings for a single process. 1946 */ 1947 static void 1948 pmc_log_process_mappings(struct pmc_owner *po, struct proc *p) 1949 { 1950 vm_map_t map; 1951 vm_map_entry_t entry; 1952 vm_object_t obj, lobj, tobj; 1953 vm_offset_t last_end; 1954 vm_offset_t start_addr; 1955 struct vnode *vp, *last_vp; 1956 struct vmspace *vm; 1957 char *fullpath, *freepath; 1958 u_int last_timestamp; 1959 1960 last_vp = NULL; 1961 last_end = (vm_offset_t)0; 1962 fullpath = freepath = NULL; 1963 1964 if ((vm = vmspace_acquire_ref(p)) == NULL) 1965 return; 1966 1967 map = &vm->vm_map; 1968 vm_map_lock_read(map); 1969 VM_MAP_ENTRY_FOREACH(entry, map) { 1970 if (entry == NULL) { 1971 PMCDBG2(LOG,OPS,2, "hwpmc: vm_map entry unexpectedly " 1972 "NULL! pid=%d vm_map=%p\n", p->p_pid, map); 1973 break; 1974 } 1975 1976 /* 1977 * We only care about executable map entries. 1978 */ 1979 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) != 0 || 1980 (entry->protection & VM_PROT_EXECUTE) == 0 || 1981 entry->object.vm_object == NULL) { 1982 continue; 1983 } 1984 1985 obj = entry->object.vm_object; 1986 VM_OBJECT_RLOCK(obj); 1987 1988 /* 1989 * Walk the backing_object list to find the base (non-shadowed) 1990 * vm_object. 1991 */ 1992 for (lobj = tobj = obj; tobj != NULL; 1993 tobj = tobj->backing_object) { 1994 if (tobj != obj) 1995 VM_OBJECT_RLOCK(tobj); 1996 if (lobj != obj) 1997 VM_OBJECT_RUNLOCK(lobj); 1998 lobj = tobj; 1999 } 2000 2001 /* 2002 * At this point lobj is the base vm_object and it is locked. 2003 */ 2004 if (lobj == NULL) { 2005 PMCDBG3(LOG,OPS,2, 2006 "hwpmc: lobj unexpectedly NULL! pid=%d " 2007 "vm_map=%p vm_obj=%p\n", p->p_pid, map, obj); 2008 VM_OBJECT_RUNLOCK(obj); 2009 continue; 2010 } 2011 2012 vp = vm_object_vnode(lobj); 2013 if (vp == NULL) { 2014 if (lobj != obj) 2015 VM_OBJECT_RUNLOCK(lobj); 2016 VM_OBJECT_RUNLOCK(obj); 2017 continue; 2018 } 2019 2020 /* 2021 * Skip contiguous regions that point to the same vnode, so we 2022 * don't emit redundant MAP-IN directives. 2023 */ 2024 if (entry->start == last_end && vp == last_vp) { 2025 last_end = entry->end; 2026 if (lobj != obj) 2027 VM_OBJECT_RUNLOCK(lobj); 2028 VM_OBJECT_RUNLOCK(obj); 2029 continue; 2030 } 2031 2032 /* 2033 * We don't want to keep the proc's vm_map or this vm_object 2034 * locked while we walk the pathname, since vn_fullpath() can 2035 * sleep. However, if we drop the lock, it's possible for 2036 * concurrent activity to modify the vm_map list. To protect 2037 * against this, we save the vm_map timestamp before we release 2038 * the lock, and check it after we reacquire the lock below. 2039 */ 2040 start_addr = entry->start; 2041 last_end = entry->end; 2042 last_timestamp = map->timestamp; 2043 vm_map_unlock_read(map); 2044 2045 vref(vp); 2046 if (lobj != obj) 2047 VM_OBJECT_RUNLOCK(lobj); 2048 VM_OBJECT_RUNLOCK(obj); 2049 2050 freepath = NULL; 2051 pmc_getfilename(vp, &fullpath, &freepath); 2052 last_vp = vp; 2053 2054 vrele(vp); 2055 2056 vp = NULL; 2057 pmclog_process_map_in(po, p->p_pid, start_addr, fullpath); 2058 if (freepath != NULL) 2059 free(freepath, M_TEMP); 2060 2061 vm_map_lock_read(map); 2062 2063 /* 2064 * If our saved timestamp doesn't match, this means 2065 * that the vm_map was modified out from under us and 2066 * we can't trust our current "entry" pointer. Do a 2067 * new lookup for this entry. If there is no entry 2068 * for this address range, vm_map_lookup_entry() will 2069 * return the previous one, so we always want to go to 2070 * the next entry on the next loop iteration. 2071 * 2072 * There is an edge condition here that can occur if 2073 * there is no entry at or before this address. In 2074 * this situation, vm_map_lookup_entry returns 2075 * &map->header, which would cause our loop to abort 2076 * without processing the rest of the map. However, 2077 * in practice this will never happen for process 2078 * vm_map. This is because the executable's text 2079 * segment is the first mapping in the proc's address 2080 * space, and this mapping is never removed until the 2081 * process exits, so there will always be a non-header 2082 * entry at or before the requested address for 2083 * vm_map_lookup_entry to return. 2084 */ 2085 if (map->timestamp != last_timestamp) 2086 vm_map_lookup_entry(map, last_end - 1, &entry); 2087 } 2088 2089 vm_map_unlock_read(map); 2090 vmspace_free(vm); 2091 return; 2092 } 2093 2094 /* 2095 * Log mappings for all processes in the system. 2096 */ 2097 static void 2098 pmc_log_all_process_mappings(struct pmc_owner *po) 2099 { 2100 struct proc *p, *top; 2101 2102 sx_assert(&pmc_sx, SX_XLOCKED); 2103 2104 if ((p = pfind(1)) == NULL) 2105 panic("[pmc,%d] Cannot find init", __LINE__); 2106 2107 PROC_UNLOCK(p); 2108 2109 sx_slock(&proctree_lock); 2110 2111 top = p; 2112 for (;;) { 2113 pmc_log_process_mappings(po, p); 2114 if (!LIST_EMPTY(&p->p_children)) 2115 p = LIST_FIRST(&p->p_children); 2116 else for (;;) { 2117 if (p == top) 2118 goto done; 2119 if (LIST_NEXT(p, p_sibling)) { 2120 p = LIST_NEXT(p, p_sibling); 2121 break; 2122 } 2123 p = p->p_pptr; 2124 } 2125 } 2126 done: 2127 sx_sunlock(&proctree_lock); 2128 } 2129 2130 #ifdef HWPMC_DEBUG 2131 const char *pmc_hooknames[] = { 2132 /* these strings correspond to PMC_FN_* in <sys/pmckern.h> */ 2133 "", 2134 "EXEC", 2135 "CSW-IN", 2136 "CSW-OUT", 2137 "SAMPLE", 2138 "UNUSED1", 2139 "UNUSED2", 2140 "MMAP", 2141 "MUNMAP", 2142 "CALLCHAIN-NMI", 2143 "CALLCHAIN-SOFT", 2144 "SOFTSAMPLING", 2145 "THR-CREATE", 2146 "THR-EXIT", 2147 "THR-USERRET", 2148 "THR-CREATE-LOG", 2149 "THR-EXIT-LOG", 2150 "PROC-CREATE-LOG" 2151 }; 2152 #endif 2153 2154 /* 2155 * The 'hook' invoked from the kernel proper 2156 */ 2157 static int 2158 pmc_hook_handler(struct thread *td, int function, void *arg) 2159 { 2160 int cpu; 2161 2162 PMCDBG4(MOD,PMH,1, "hook td=%p func=%d \"%s\" arg=%p", td, function, 2163 pmc_hooknames[function], arg); 2164 2165 switch (function) { 2166 case PMC_FN_PROCESS_EXEC: 2167 pmc_process_exec(td, (struct pmckern_procexec *)arg); 2168 break; 2169 2170 case PMC_FN_CSW_IN: 2171 pmc_process_csw_in(td); 2172 break; 2173 2174 case PMC_FN_CSW_OUT: 2175 pmc_process_csw_out(td); 2176 break; 2177 2178 /* 2179 * Process accumulated PC samples. 2180 * 2181 * This function is expected to be called by hardclock() for 2182 * each CPU that has accumulated PC samples. 2183 * 2184 * This function is to be executed on the CPU whose samples 2185 * are being processed. 2186 */ 2187 case PMC_FN_DO_SAMPLES: 2188 /* 2189 * Clear the cpu specific bit in the CPU mask before 2190 * do the rest of the processing. If the NMI handler 2191 * gets invoked after the "atomic_clear_int()" call 2192 * below but before "pmc_process_samples()" gets 2193 * around to processing the interrupt, then we will 2194 * come back here at the next hardclock() tick (and 2195 * may find nothing to do if "pmc_process_samples()" 2196 * had already processed the interrupt). We don't 2197 * lose the interrupt sample. 2198 */ 2199 DPCPU_SET(pmc_sampled, 0); 2200 cpu = PCPU_GET(cpuid); 2201 pmc_process_samples(cpu, PMC_HR); 2202 pmc_process_samples(cpu, PMC_SR); 2203 pmc_process_samples(cpu, PMC_UR); 2204 break; 2205 2206 case PMC_FN_MMAP: 2207 pmc_process_mmap(td, (struct pmckern_map_in *)arg); 2208 break; 2209 2210 case PMC_FN_MUNMAP: 2211 MPASS(in_epoch(global_epoch_preempt) || sx_xlocked(&pmc_sx)); 2212 pmc_process_munmap(td, (struct pmckern_map_out *)arg); 2213 break; 2214 2215 case PMC_FN_PROC_CREATE_LOG: 2216 pmc_process_proccreate((struct proc *)arg); 2217 break; 2218 2219 case PMC_FN_USER_CALLCHAIN: 2220 /* 2221 * Record a call chain. 2222 */ 2223 KASSERT(td == curthread, ("[pmc,%d] td != curthread", 2224 __LINE__)); 2225 2226 pmc_capture_user_callchain(PCPU_GET(cpuid), PMC_HR, 2227 (struct trapframe *)arg); 2228 2229 KASSERT(td->td_pinned == 1, 2230 ("[pmc,%d] invalid td_pinned value", __LINE__)); 2231 sched_unpin(); /* Can migrate safely now. */ 2232 2233 td->td_pflags &= ~TDP_CALLCHAIN; 2234 break; 2235 2236 case PMC_FN_USER_CALLCHAIN_SOFT: 2237 /* 2238 * Record a call chain. 2239 */ 2240 KASSERT(td == curthread, ("[pmc,%d] td != curthread", 2241 __LINE__)); 2242 2243 cpu = PCPU_GET(cpuid); 2244 pmc_capture_user_callchain(cpu, PMC_SR, 2245 (struct trapframe *) arg); 2246 2247 KASSERT(td->td_pinned == 1, 2248 ("[pmc,%d] invalid td_pinned value", __LINE__)); 2249 2250 sched_unpin(); /* Can migrate safely now. */ 2251 2252 td->td_pflags &= ~TDP_CALLCHAIN; 2253 break; 2254 2255 case PMC_FN_SOFT_SAMPLING: 2256 /* 2257 * Call soft PMC sampling intr. 2258 */ 2259 pmc_soft_intr((struct pmckern_soft *)arg); 2260 break; 2261 2262 case PMC_FN_THR_CREATE: 2263 pmc_process_thread_add(td); 2264 pmc_process_threadcreate(td); 2265 break; 2266 2267 case PMC_FN_THR_CREATE_LOG: 2268 pmc_process_threadcreate(td); 2269 break; 2270 2271 case PMC_FN_THR_EXIT: 2272 KASSERT(td == curthread, ("[pmc,%d] td != curthread", 2273 __LINE__)); 2274 pmc_process_thread_delete(td); 2275 pmc_process_threadexit(td); 2276 break; 2277 case PMC_FN_THR_EXIT_LOG: 2278 pmc_process_threadexit(td); 2279 break; 2280 case PMC_FN_THR_USERRET: 2281 KASSERT(td == curthread, ("[pmc,%d] td != curthread", 2282 __LINE__)); 2283 pmc_process_thread_userret(td); 2284 break; 2285 default: 2286 #ifdef HWPMC_DEBUG 2287 KASSERT(0, ("[pmc,%d] unknown hook %d\n", __LINE__, function)); 2288 #endif 2289 break; 2290 } 2291 2292 return (0); 2293 } 2294 2295 /* 2296 * Allocate a 'struct pmc_owner' descriptor in the owner hash table. 2297 */ 2298 static struct pmc_owner * 2299 pmc_allocate_owner_descriptor(struct proc *p) 2300 { 2301 struct pmc_owner *po; 2302 struct pmc_ownerhash *poh; 2303 uint32_t hindex; 2304 2305 hindex = PMC_HASH_PTR(p, pmc_ownerhashmask); 2306 poh = &pmc_ownerhash[hindex]; 2307 2308 /* Allocate space for N pointers and one descriptor struct. */ 2309 po = malloc(sizeof(struct pmc_owner), M_PMC, M_WAITOK | M_ZERO); 2310 po->po_owner = p; 2311 LIST_INSERT_HEAD(poh, po, po_next); /* insert into hash table */ 2312 2313 TAILQ_INIT(&po->po_logbuffers); 2314 mtx_init(&po->po_mtx, "pmc-owner-mtx", "pmc-per-proc", MTX_SPIN); 2315 2316 PMCDBG4(OWN,ALL,1, "allocate-owner proc=%p (%d, %s) pmc-owner=%p", 2317 p, p->p_pid, p->p_comm, po); 2318 2319 return (po); 2320 } 2321 2322 static void 2323 pmc_destroy_owner_descriptor(struct pmc_owner *po) 2324 { 2325 2326 PMCDBG4(OWN,REL,1, "destroy-owner po=%p proc=%p (%d, %s)", 2327 po, po->po_owner, po->po_owner->p_pid, po->po_owner->p_comm); 2328 2329 mtx_destroy(&po->po_mtx); 2330 free(po, M_PMC); 2331 } 2332 2333 /* 2334 * Allocate a thread descriptor from the free pool. 2335 * 2336 * NOTE: This *can* return NULL. 2337 */ 2338 static struct pmc_thread * 2339 pmc_thread_descriptor_pool_alloc(void) 2340 { 2341 struct pmc_thread *pt; 2342 2343 mtx_lock_spin(&pmc_threadfreelist_mtx); 2344 if ((pt = LIST_FIRST(&pmc_threadfreelist)) != NULL) { 2345 LIST_REMOVE(pt, pt_next); 2346 pmc_threadfreelist_entries--; 2347 } 2348 mtx_unlock_spin(&pmc_threadfreelist_mtx); 2349 2350 return (pt); 2351 } 2352 2353 /* 2354 * Add a thread descriptor to the free pool. We use this instead of free() 2355 * to maintain a cache of free entries. Additionally, we can safely call 2356 * this function when we cannot call free(), such as in a critical section. 2357 */ 2358 static void 2359 pmc_thread_descriptor_pool_free(struct pmc_thread *pt) 2360 { 2361 2362 if (pt == NULL) 2363 return; 2364 2365 memset(pt, 0, THREADENTRY_SIZE); 2366 mtx_lock_spin(&pmc_threadfreelist_mtx); 2367 LIST_INSERT_HEAD(&pmc_threadfreelist, pt, pt_next); 2368 pmc_threadfreelist_entries++; 2369 if (pmc_threadfreelist_entries > pmc_threadfreelist_max) 2370 taskqueue_enqueue(taskqueue_fast, &free_task); 2371 mtx_unlock_spin(&pmc_threadfreelist_mtx); 2372 } 2373 2374 /* 2375 * An asynchronous task to manage the free list. 2376 */ 2377 static void 2378 pmc_thread_descriptor_pool_free_task(void *arg __unused, int pending __unused) 2379 { 2380 struct pmc_thread *pt; 2381 LIST_HEAD(, pmc_thread) tmplist; 2382 int delta; 2383 2384 LIST_INIT(&tmplist); 2385 2386 /* Determine what changes, if any, we need to make. */ 2387 mtx_lock_spin(&pmc_threadfreelist_mtx); 2388 delta = pmc_threadfreelist_entries - pmc_threadfreelist_max; 2389 while (delta > 0 && (pt = LIST_FIRST(&pmc_threadfreelist)) != NULL) { 2390 delta--; 2391 pmc_threadfreelist_entries--; 2392 LIST_REMOVE(pt, pt_next); 2393 LIST_INSERT_HEAD(&tmplist, pt, pt_next); 2394 } 2395 mtx_unlock_spin(&pmc_threadfreelist_mtx); 2396 2397 /* If there are entries to free, free them. */ 2398 while (!LIST_EMPTY(&tmplist)) { 2399 pt = LIST_FIRST(&tmplist); 2400 LIST_REMOVE(pt, pt_next); 2401 free(pt, M_PMC); 2402 } 2403 } 2404 2405 /* 2406 * Drain the thread free pool, freeing all allocations. 2407 */ 2408 static void 2409 pmc_thread_descriptor_pool_drain(void) 2410 { 2411 struct pmc_thread *pt, *next; 2412 2413 LIST_FOREACH_SAFE(pt, &pmc_threadfreelist, pt_next, next) { 2414 LIST_REMOVE(pt, pt_next); 2415 free(pt, M_PMC); 2416 } 2417 } 2418 2419 /* 2420 * find the descriptor corresponding to thread 'td', adding or removing it 2421 * as specified by 'mode'. 2422 * 2423 * Note that this supports additional mode flags in addition to those 2424 * supported by pmc_find_process_descriptor(): 2425 * PMC_FLAG_NOWAIT: Causes the function to not wait for mallocs. 2426 * This makes it safe to call while holding certain other locks. 2427 */ 2428 static struct pmc_thread * 2429 pmc_find_thread_descriptor(struct pmc_process *pp, struct thread *td, 2430 uint32_t mode) 2431 { 2432 struct pmc_thread *pt = NULL, *ptnew = NULL; 2433 int wait_flag; 2434 2435 KASSERT(td != NULL, ("[pmc,%d] called to add NULL td", __LINE__)); 2436 2437 /* 2438 * Pre-allocate memory in the PMC_FLAG_ALLOCATE case prior to 2439 * acquiring the lock. 2440 */ 2441 if ((mode & PMC_FLAG_ALLOCATE) != 0) { 2442 if ((ptnew = pmc_thread_descriptor_pool_alloc()) == NULL) { 2443 wait_flag = M_WAITOK; 2444 if ((mode & PMC_FLAG_NOWAIT) != 0 || 2445 in_epoch(global_epoch_preempt)) 2446 wait_flag = M_NOWAIT; 2447 2448 ptnew = malloc(THREADENTRY_SIZE, M_PMC, 2449 wait_flag | M_ZERO); 2450 } 2451 } 2452 2453 mtx_lock_spin(pp->pp_tdslock); 2454 LIST_FOREACH(pt, &pp->pp_tds, pt_next) { 2455 if (pt->pt_td == td) 2456 break; 2457 } 2458 2459 if ((mode & PMC_FLAG_REMOVE) != 0 && pt != NULL) 2460 LIST_REMOVE(pt, pt_next); 2461 2462 if ((mode & PMC_FLAG_ALLOCATE) != 0 && pt == NULL && ptnew != NULL) { 2463 pt = ptnew; 2464 ptnew = NULL; 2465 pt->pt_td = td; 2466 LIST_INSERT_HEAD(&pp->pp_tds, pt, pt_next); 2467 } 2468 2469 mtx_unlock_spin(pp->pp_tdslock); 2470 2471 if (ptnew != NULL) { 2472 free(ptnew, M_PMC); 2473 } 2474 2475 return (pt); 2476 } 2477 2478 /* 2479 * Try to add thread descriptors for each thread in a process. 2480 */ 2481 static void 2482 pmc_add_thread_descriptors_from_proc(struct proc *p, struct pmc_process *pp) 2483 { 2484 struct pmc_thread **tdlist; 2485 struct thread *curtd; 2486 int i, tdcnt, tdlistsz; 2487 2488 KASSERT(!PROC_LOCKED(p), ("[pmc,%d] proc unexpectedly locked", 2489 __LINE__)); 2490 tdcnt = 32; 2491 restart: 2492 tdlistsz = roundup2(tdcnt, 32); 2493 2494 tdcnt = 0; 2495 tdlist = malloc(sizeof(struct pmc_thread *) * tdlistsz, M_TEMP, 2496 M_WAITOK); 2497 2498 PROC_LOCK(p); 2499 FOREACH_THREAD_IN_PROC(p, curtd) 2500 tdcnt++; 2501 if (tdcnt >= tdlistsz) { 2502 PROC_UNLOCK(p); 2503 free(tdlist, M_TEMP); 2504 goto restart; 2505 } 2506 2507 /* 2508 * Try to add each thread to the list without sleeping. If unable, 2509 * add to a queue to retry after dropping the process lock. 2510 */ 2511 tdcnt = 0; 2512 FOREACH_THREAD_IN_PROC(p, curtd) { 2513 tdlist[tdcnt] = pmc_find_thread_descriptor(pp, curtd, 2514 PMC_FLAG_ALLOCATE | PMC_FLAG_NOWAIT); 2515 if (tdlist[tdcnt] == NULL) { 2516 PROC_UNLOCK(p); 2517 for (i = 0; i <= tdcnt; i++) 2518 pmc_thread_descriptor_pool_free(tdlist[i]); 2519 free(tdlist, M_TEMP); 2520 goto restart; 2521 } 2522 tdcnt++; 2523 } 2524 PROC_UNLOCK(p); 2525 free(tdlist, M_TEMP); 2526 } 2527 2528 /* 2529 * Find the descriptor corresponding to process 'p', adding or removing it 2530 * as specified by 'mode'. 2531 */ 2532 static struct pmc_process * 2533 pmc_find_process_descriptor(struct proc *p, uint32_t mode) 2534 { 2535 struct pmc_process *pp, *ppnew; 2536 struct pmc_processhash *pph; 2537 uint32_t hindex; 2538 2539 hindex = PMC_HASH_PTR(p, pmc_processhashmask); 2540 pph = &pmc_processhash[hindex]; 2541 2542 ppnew = NULL; 2543 2544 /* 2545 * Pre-allocate memory in the PMC_FLAG_ALLOCATE case since we 2546 * cannot call malloc(9) once we hold a spin lock. 2547 */ 2548 if ((mode & PMC_FLAG_ALLOCATE) != 0) 2549 ppnew = malloc(sizeof(struct pmc_process) + md->pmd_npmc * 2550 sizeof(struct pmc_targetstate), M_PMC, M_WAITOK | M_ZERO); 2551 2552 mtx_lock_spin(&pmc_processhash_mtx); 2553 LIST_FOREACH(pp, pph, pp_next) { 2554 if (pp->pp_proc == p) 2555 break; 2556 } 2557 2558 if ((mode & PMC_FLAG_REMOVE) != 0 && pp != NULL) 2559 LIST_REMOVE(pp, pp_next); 2560 2561 if ((mode & PMC_FLAG_ALLOCATE) != 0 && pp == NULL && ppnew != NULL) { 2562 ppnew->pp_proc = p; 2563 LIST_INIT(&ppnew->pp_tds); 2564 ppnew->pp_tdslock = mtx_pool_find(pmc_mtxpool, ppnew); 2565 LIST_INSERT_HEAD(pph, ppnew, pp_next); 2566 mtx_unlock_spin(&pmc_processhash_mtx); 2567 pp = ppnew; 2568 ppnew = NULL; 2569 2570 /* Add thread descriptors for this process' current threads. */ 2571 pmc_add_thread_descriptors_from_proc(p, pp); 2572 } else 2573 mtx_unlock_spin(&pmc_processhash_mtx); 2574 2575 if (ppnew != NULL) 2576 free(ppnew, M_PMC); 2577 return (pp); 2578 } 2579 2580 /* 2581 * Remove a process descriptor from the process hash table. 2582 */ 2583 static void 2584 pmc_remove_process_descriptor(struct pmc_process *pp) 2585 { 2586 KASSERT(pp->pp_refcnt == 0, 2587 ("[pmc,%d] Removing process descriptor %p with count %d", 2588 __LINE__, pp, pp->pp_refcnt)); 2589 2590 mtx_lock_spin(&pmc_processhash_mtx); 2591 LIST_REMOVE(pp, pp_next); 2592 mtx_unlock_spin(&pmc_processhash_mtx); 2593 } 2594 2595 /* 2596 * Destroy a process descriptor. 2597 */ 2598 static void 2599 pmc_destroy_process_descriptor(struct pmc_process *pp) 2600 { 2601 struct pmc_thread *pmc_td; 2602 2603 while ((pmc_td = LIST_FIRST(&pp->pp_tds)) != NULL) { 2604 LIST_REMOVE(pmc_td, pt_next); 2605 pmc_thread_descriptor_pool_free(pmc_td); 2606 } 2607 free(pp, M_PMC); 2608 } 2609 2610 /* 2611 * Find an owner descriptor corresponding to proc 'p'. 2612 */ 2613 static struct pmc_owner * 2614 pmc_find_owner_descriptor(struct proc *p) 2615 { 2616 struct pmc_owner *po; 2617 struct pmc_ownerhash *poh; 2618 uint32_t hindex; 2619 2620 hindex = PMC_HASH_PTR(p, pmc_ownerhashmask); 2621 poh = &pmc_ownerhash[hindex]; 2622 2623 po = NULL; 2624 LIST_FOREACH(po, poh, po_next) { 2625 if (po->po_owner == p) 2626 break; 2627 } 2628 2629 PMCDBG5(OWN,FND,1, "find-owner proc=%p (%d, %s) hindex=0x%x -> " 2630 "pmc-owner=%p", p, p->p_pid, p->p_comm, hindex, po); 2631 2632 return (po); 2633 } 2634 2635 /* 2636 * Allocate a pmc descriptor and initialize its fields. 2637 */ 2638 static struct pmc * 2639 pmc_allocate_pmc_descriptor(void) 2640 { 2641 struct pmc *pmc; 2642 2643 pmc = malloc(sizeof(struct pmc), M_PMC, M_WAITOK | M_ZERO); 2644 pmc->pm_runcount = counter_u64_alloc(M_WAITOK); 2645 pmc->pm_pcpu_state = malloc(sizeof(struct pmc_pcpu_state) * mp_ncpus, 2646 M_PMC, M_WAITOK | M_ZERO); 2647 PMCDBG1(PMC,ALL,1, "allocate-pmc -> pmc=%p", pmc); 2648 2649 return (pmc); 2650 } 2651 2652 /* 2653 * Destroy a pmc descriptor. 2654 */ 2655 static void 2656 pmc_destroy_pmc_descriptor(struct pmc *pm) 2657 { 2658 2659 KASSERT(pm->pm_state == PMC_STATE_DELETED || 2660 pm->pm_state == PMC_STATE_FREE, 2661 ("[pmc,%d] destroying non-deleted PMC", __LINE__)); 2662 KASSERT(LIST_EMPTY(&pm->pm_targets), 2663 ("[pmc,%d] destroying pmc with targets", __LINE__)); 2664 KASSERT(pm->pm_owner == NULL, 2665 ("[pmc,%d] destroying pmc attached to an owner", __LINE__)); 2666 KASSERT(counter_u64_fetch(pm->pm_runcount) == 0, 2667 ("[pmc,%d] pmc has non-zero run count %ju", __LINE__, 2668 (uintmax_t)counter_u64_fetch(pm->pm_runcount))); 2669 2670 counter_u64_free(pm->pm_runcount); 2671 free(pm->pm_pcpu_state, M_PMC); 2672 free(pm, M_PMC); 2673 } 2674 2675 static void 2676 pmc_wait_for_pmc_idle(struct pmc *pm) 2677 { 2678 #ifdef INVARIANTS 2679 volatile int maxloop; 2680 2681 maxloop = 100 * pmc_cpu_max(); 2682 #endif 2683 /* 2684 * Loop (with a forced context switch) till the PMC's runcount 2685 * comes down to zero. 2686 */ 2687 pmclog_flush(pm->pm_owner, 1); 2688 while (counter_u64_fetch(pm->pm_runcount) > 0) { 2689 pmclog_flush(pm->pm_owner, 1); 2690 #ifdef INVARIANTS 2691 maxloop--; 2692 KASSERT(maxloop > 0, 2693 ("[pmc,%d] (ri%d, rc%ju) waiting too long for " 2694 "pmc to be free", __LINE__, PMC_TO_ROWINDEX(pm), 2695 (uintmax_t)counter_u64_fetch(pm->pm_runcount))); 2696 #endif 2697 pmc_force_context_switch(); 2698 } 2699 } 2700 2701 /* 2702 * This function does the following things: 2703 * 2704 * - detaches the PMC from hardware 2705 * - unlinks all target threads that were attached to it 2706 * - removes the PMC from its owner's list 2707 * - destroys the PMC private mutex 2708 * 2709 * Once this function completes, the given pmc pointer can be freed by 2710 * calling pmc_destroy_pmc_descriptor(). 2711 */ 2712 static void 2713 pmc_release_pmc_descriptor(struct pmc *pm) 2714 { 2715 struct pmc_binding pb; 2716 struct pmc_classdep *pcd; 2717 struct pmc_hw *phw __diagused; 2718 struct pmc_owner *po; 2719 struct pmc_process *pp; 2720 struct pmc_target *ptgt, *tmp; 2721 enum pmc_mode mode; 2722 u_int adjri, ri, cpu; 2723 2724 sx_assert(&pmc_sx, SX_XLOCKED); 2725 KASSERT(pm, ("[pmc,%d] null pmc", __LINE__)); 2726 2727 ri = PMC_TO_ROWINDEX(pm); 2728 pcd = pmc_ri_to_classdep(md, ri, &adjri); 2729 mode = PMC_TO_MODE(pm); 2730 2731 PMCDBG3(PMC,REL,1, "release-pmc pmc=%p ri=%d mode=%d", pm, ri, 2732 mode); 2733 2734 /* 2735 * First, we take the PMC off hardware. 2736 */ 2737 cpu = 0; 2738 if (PMC_IS_SYSTEM_MODE(mode)) { 2739 /* 2740 * A system mode PMC runs on a specific CPU. Switch 2741 * to this CPU and turn hardware off. 2742 */ 2743 pmc_save_cpu_binding(&pb); 2744 cpu = PMC_TO_CPU(pm); 2745 pmc_select_cpu(cpu); 2746 2747 /* switch off non-stalled CPUs */ 2748 pm->pm_pcpu_state[cpu].pps_cpustate = 0; 2749 if (pm->pm_state == PMC_STATE_RUNNING && 2750 pm->pm_pcpu_state[cpu].pps_stalled == 0) { 2751 2752 phw = pmc_pcpu[cpu]->pc_hwpmcs[ri]; 2753 2754 KASSERT(phw->phw_pmc == pm, 2755 ("[pmc, %d] pmc ptr ri(%d) hw(%p) pm(%p)", 2756 __LINE__, ri, phw->phw_pmc, pm)); 2757 PMCDBG2(PMC,REL,2, "stopping cpu=%d ri=%d", cpu, ri); 2758 2759 critical_enter(); 2760 (void)pcd->pcd_stop_pmc(cpu, adjri, pm); 2761 critical_exit(); 2762 } 2763 2764 PMCDBG2(PMC,REL,2, "decfg cpu=%d ri=%d", cpu, ri); 2765 2766 critical_enter(); 2767 (void)pcd->pcd_config_pmc(cpu, adjri, NULL); 2768 critical_exit(); 2769 2770 /* adjust the global and process count of SS mode PMCs */ 2771 if (mode == PMC_MODE_SS && pm->pm_state == PMC_STATE_RUNNING) { 2772 po = pm->pm_owner; 2773 po->po_sscount--; 2774 if (po->po_sscount == 0) { 2775 atomic_subtract_rel_int(&pmc_ss_count, 1); 2776 CK_LIST_REMOVE(po, po_ssnext); 2777 epoch_wait_preempt(global_epoch_preempt); 2778 } 2779 } 2780 pm->pm_state = PMC_STATE_DELETED; 2781 2782 pmc_restore_cpu_binding(&pb); 2783 2784 /* 2785 * We could have references to this PMC structure in the 2786 * per-cpu sample queues. Wait for the queue to drain. 2787 */ 2788 pmc_wait_for_pmc_idle(pm); 2789 2790 } else if (PMC_IS_VIRTUAL_MODE(mode)) { 2791 /* 2792 * A virtual PMC could be running on multiple CPUs at a given 2793 * instant. 2794 * 2795 * By marking its state as DELETED, we ensure that this PMC is 2796 * never further scheduled on hardware. 2797 * 2798 * Then we wait till all CPUs are done with this PMC. 2799 */ 2800 pm->pm_state = PMC_STATE_DELETED; 2801 2802 /* Wait for the PMCs runcount to come to zero. */ 2803 pmc_wait_for_pmc_idle(pm); 2804 2805 /* 2806 * At this point the PMC is off all CPUs and cannot be freshly 2807 * scheduled onto a CPU. It is now safe to unlink all targets 2808 * from this PMC. If a process-record's refcount falls to zero, 2809 * we remove it from the hash table. The module-wide SX lock 2810 * protects us from races. 2811 */ 2812 LIST_FOREACH_SAFE(ptgt, &pm->pm_targets, pt_next, tmp) { 2813 pp = ptgt->pt_process; 2814 pmc_unlink_target_process(pm, pp); /* frees 'ptgt' */ 2815 2816 PMCDBG1(PMC,REL,3, "pp->refcnt=%d", pp->pp_refcnt); 2817 2818 /* 2819 * If the target process record shows that no PMCs are 2820 * attached to it, reclaim its space. 2821 */ 2822 if (pp->pp_refcnt == 0) { 2823 pmc_remove_process_descriptor(pp); 2824 pmc_destroy_process_descriptor(pp); 2825 } 2826 } 2827 2828 cpu = curthread->td_oncpu; /* setup cpu for pmd_release() */ 2829 } 2830 2831 /* 2832 * Release any MD resources. 2833 */ 2834 (void)pcd->pcd_release_pmc(cpu, adjri, pm); 2835 2836 /* 2837 * Update row disposition. 2838 */ 2839 if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm))) 2840 PMC_UNMARK_ROW_STANDALONE(ri); 2841 else 2842 PMC_UNMARK_ROW_THREAD(ri); 2843 2844 /* Unlink from the owner's list. */ 2845 if (pm->pm_owner != NULL) { 2846 LIST_REMOVE(pm, pm_next); 2847 pm->pm_owner = NULL; 2848 } 2849 } 2850 2851 /* 2852 * Register an owner and a pmc. 2853 */ 2854 static int 2855 pmc_register_owner(struct proc *p, struct pmc *pmc) 2856 { 2857 struct pmc_owner *po; 2858 2859 sx_assert(&pmc_sx, SX_XLOCKED); 2860 2861 if ((po = pmc_find_owner_descriptor(p)) == NULL) { 2862 if ((po = pmc_allocate_owner_descriptor(p)) == NULL) 2863 return (ENOMEM); 2864 } 2865 2866 KASSERT(pmc->pm_owner == NULL, 2867 ("[pmc,%d] attempting to own an initialized PMC", __LINE__)); 2868 pmc->pm_owner = po; 2869 2870 LIST_INSERT_HEAD(&po->po_pmcs, pmc, pm_next); 2871 2872 PROC_LOCK(p); 2873 p->p_flag |= P_HWPMC; 2874 PROC_UNLOCK(p); 2875 2876 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) != 0) 2877 pmclog_process_pmcallocate(pmc); 2878 2879 PMCDBG2(PMC,REG,1, "register-owner pmc-owner=%p pmc=%p", 2880 po, pmc); 2881 2882 return (0); 2883 } 2884 2885 /* 2886 * Return the current row disposition: 2887 * == 0 => FREE 2888 * > 0 => PROCESS MODE 2889 * < 0 => SYSTEM MODE 2890 */ 2891 int 2892 pmc_getrowdisp(int ri) 2893 { 2894 return (pmc_pmcdisp[ri]); 2895 } 2896 2897 /* 2898 * Check if a PMC at row index 'ri' can be allocated to the current 2899 * process. 2900 * 2901 * Allocation can fail if: 2902 * - the current process is already being profiled by a PMC at index 'ri', 2903 * attached to it via OP_PMCATTACH. 2904 * - the current process has already allocated a PMC at index 'ri' 2905 * via OP_ALLOCATE. 2906 */ 2907 static bool 2908 pmc_can_allocate_rowindex(struct proc *p, unsigned int ri, int cpu) 2909 { 2910 struct pmc *pm; 2911 struct pmc_owner *po; 2912 struct pmc_process *pp; 2913 enum pmc_mode mode; 2914 2915 PMCDBG5(PMC,ALR,1, "can-allocate-rowindex proc=%p (%d, %s) ri=%d " 2916 "cpu=%d", p, p->p_pid, p->p_comm, ri, cpu); 2917 2918 /* 2919 * We shouldn't have already allocated a process-mode PMC at 2920 * row index 'ri'. 2921 * 2922 * We shouldn't have allocated a system-wide PMC on the same 2923 * CPU and same RI. 2924 */ 2925 if ((po = pmc_find_owner_descriptor(p)) != NULL) { 2926 LIST_FOREACH(pm, &po->po_pmcs, pm_next) { 2927 if (PMC_TO_ROWINDEX(pm) == ri) { 2928 mode = PMC_TO_MODE(pm); 2929 if (PMC_IS_VIRTUAL_MODE(mode)) 2930 return (false); 2931 if (PMC_IS_SYSTEM_MODE(mode) && 2932 PMC_TO_CPU(pm) == cpu) 2933 return (false); 2934 } 2935 } 2936 } 2937 2938 /* 2939 * We also shouldn't be the target of any PMC at this index 2940 * since otherwise a PMC_ATTACH to ourselves will fail. 2941 */ 2942 if ((pp = pmc_find_process_descriptor(p, 0)) != NULL) 2943 if (pp->pp_pmcs[ri].pp_pmc != NULL) 2944 return (false); 2945 2946 PMCDBG4(PMC,ALR,2, "can-allocate-rowindex proc=%p (%d, %s) ri=%d ok", 2947 p, p->p_pid, p->p_comm, ri); 2948 return (true); 2949 } 2950 2951 /* 2952 * Check if a given PMC at row index 'ri' can be currently used in 2953 * mode 'mode'. 2954 */ 2955 static bool 2956 pmc_can_allocate_row(int ri, enum pmc_mode mode) 2957 { 2958 enum pmc_disp disp; 2959 2960 sx_assert(&pmc_sx, SX_XLOCKED); 2961 2962 PMCDBG2(PMC,ALR,1, "can-allocate-row ri=%d mode=%d", ri, mode); 2963 2964 if (PMC_IS_SYSTEM_MODE(mode)) 2965 disp = PMC_DISP_STANDALONE; 2966 else 2967 disp = PMC_DISP_THREAD; 2968 2969 /* 2970 * check disposition for PMC row 'ri': 2971 * 2972 * Expected disposition Row-disposition Result 2973 * 2974 * STANDALONE STANDALONE or FREE proceed 2975 * STANDALONE THREAD fail 2976 * THREAD THREAD or FREE proceed 2977 * THREAD STANDALONE fail 2978 */ 2979 if (!PMC_ROW_DISP_IS_FREE(ri) && 2980 !(disp == PMC_DISP_THREAD && PMC_ROW_DISP_IS_THREAD(ri)) && 2981 !(disp == PMC_DISP_STANDALONE && PMC_ROW_DISP_IS_STANDALONE(ri))) 2982 return (false); 2983 2984 /* 2985 * All OK 2986 */ 2987 PMCDBG2(PMC,ALR,2, "can-allocate-row ri=%d mode=%d ok", ri, mode); 2988 return (true); 2989 } 2990 2991 /* 2992 * Find a PMC descriptor with user handle 'pmcid' for thread 'td'. 2993 */ 2994 static struct pmc * 2995 pmc_find_pmc_descriptor_in_process(struct pmc_owner *po, pmc_id_t pmcid) 2996 { 2997 struct pmc *pm; 2998 2999 KASSERT(PMC_ID_TO_ROWINDEX(pmcid) < md->pmd_npmc, 3000 ("[pmc,%d] Illegal pmc index %d (max %d)", __LINE__, 3001 PMC_ID_TO_ROWINDEX(pmcid), md->pmd_npmc)); 3002 3003 LIST_FOREACH(pm, &po->po_pmcs, pm_next) { 3004 if (pm->pm_id == pmcid) 3005 return (pm); 3006 } 3007 3008 return (NULL); 3009 } 3010 3011 static int 3012 pmc_find_pmc(pmc_id_t pmcid, struct pmc **pmc) 3013 { 3014 struct pmc *pm, *opm; 3015 struct pmc_owner *po; 3016 struct pmc_process *pp; 3017 3018 PMCDBG1(PMC,FND,1, "find-pmc id=%d", pmcid); 3019 if (PMC_ID_TO_ROWINDEX(pmcid) >= md->pmd_npmc) 3020 return (EINVAL); 3021 3022 if ((po = pmc_find_owner_descriptor(curthread->td_proc)) == NULL) { 3023 /* 3024 * In case of PMC_F_DESCENDANTS child processes we will not find 3025 * the current process in the owners hash list. Find the owner 3026 * process first and from there lookup the po. 3027 */ 3028 pp = pmc_find_process_descriptor(curthread->td_proc, 3029 PMC_FLAG_NONE); 3030 if (pp == NULL) 3031 return (ESRCH); 3032 opm = pp->pp_pmcs[PMC_ID_TO_ROWINDEX(pmcid)].pp_pmc; 3033 if (opm == NULL) 3034 return (ESRCH); 3035 if ((opm->pm_flags & 3036 (PMC_F_ATTACHED_TO_OWNER | PMC_F_DESCENDANTS)) != 3037 (PMC_F_ATTACHED_TO_OWNER | PMC_F_DESCENDANTS)) 3038 return (ESRCH); 3039 3040 po = opm->pm_owner; 3041 } 3042 3043 if ((pm = pmc_find_pmc_descriptor_in_process(po, pmcid)) == NULL) 3044 return (EINVAL); 3045 3046 PMCDBG2(PMC,FND,2, "find-pmc id=%d -> pmc=%p", pmcid, pm); 3047 3048 *pmc = pm; 3049 return (0); 3050 } 3051 3052 /* 3053 * Start a PMC. 3054 */ 3055 static int 3056 pmc_start(struct pmc *pm) 3057 { 3058 struct pmc_binding pb; 3059 struct pmc_classdep *pcd; 3060 struct pmc_owner *po; 3061 pmc_value_t v; 3062 enum pmc_mode mode; 3063 int adjri, error, cpu, ri; 3064 3065 KASSERT(pm != NULL, 3066 ("[pmc,%d] null pm", __LINE__)); 3067 3068 mode = PMC_TO_MODE(pm); 3069 ri = PMC_TO_ROWINDEX(pm); 3070 pcd = pmc_ri_to_classdep(md, ri, &adjri); 3071 3072 error = 0; 3073 po = pm->pm_owner; 3074 3075 PMCDBG3(PMC,OPS,1, "start pmc=%p mode=%d ri=%d", pm, mode, ri); 3076 3077 po = pm->pm_owner; 3078 3079 /* 3080 * Disallow PMCSTART if a logfile is required but has not been 3081 * configured yet. 3082 */ 3083 if ((pm->pm_flags & PMC_F_NEEDS_LOGFILE) != 0 && 3084 (po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) 3085 return (EDOOFUS); /* programming error */ 3086 3087 /* 3088 * If this is a sampling mode PMC, log mapping information for 3089 * the kernel modules that are currently loaded. 3090 */ 3091 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) 3092 pmc_log_kernel_mappings(pm); 3093 3094 if (PMC_IS_VIRTUAL_MODE(mode)) { 3095 /* 3096 * If a PMCATTACH has never been done on this PMC, 3097 * attach it to its owner process. 3098 */ 3099 if (LIST_EMPTY(&pm->pm_targets)) { 3100 error = (pm->pm_flags & PMC_F_ATTACH_DONE) != 0 ? 3101 ESRCH : pmc_attach_process(po->po_owner, pm); 3102 } 3103 3104 /* 3105 * If the PMC is attached to its owner, then force a context 3106 * switch to ensure that the MD state gets set correctly. 3107 */ 3108 if (error == 0) { 3109 pm->pm_state = PMC_STATE_RUNNING; 3110 if ((pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) != 0) 3111 pmc_force_context_switch(); 3112 } 3113 3114 return (error); 3115 } 3116 3117 /* 3118 * A system-wide PMC. 3119 * 3120 * Add the owner to the global list if this is a system-wide 3121 * sampling PMC. 3122 */ 3123 if (mode == PMC_MODE_SS) { 3124 /* 3125 * Log mapping information for all existing processes in the 3126 * system. Subsequent mappings are logged as they happen; 3127 * see pmc_process_mmap(). 3128 */ 3129 if (po->po_logprocmaps == 0) { 3130 pmc_log_all_process_mappings(po); 3131 po->po_logprocmaps = 1; 3132 } 3133 po->po_sscount++; 3134 if (po->po_sscount == 1) { 3135 atomic_add_rel_int(&pmc_ss_count, 1); 3136 CK_LIST_INSERT_HEAD(&pmc_ss_owners, po, po_ssnext); 3137 PMCDBG1(PMC,OPS,1, "po=%p in global list", po); 3138 } 3139 } 3140 3141 /* 3142 * Move to the CPU associated with this 3143 * PMC, and start the hardware. 3144 */ 3145 pmc_save_cpu_binding(&pb); 3146 cpu = PMC_TO_CPU(pm); 3147 if (!pmc_cpu_is_active(cpu)) 3148 return (ENXIO); 3149 pmc_select_cpu(cpu); 3150 3151 /* 3152 * global PMCs are configured at allocation time 3153 * so write out the initial value and start the PMC. 3154 */ 3155 pm->pm_state = PMC_STATE_RUNNING; 3156 3157 critical_enter(); 3158 v = PMC_IS_SAMPLING_MODE(mode) ? pm->pm_sc.pm_reloadcount : 3159 pm->pm_sc.pm_initial; 3160 if ((error = pcd->pcd_write_pmc(cpu, adjri, pm, v)) == 0) { 3161 /* If a sampling mode PMC, reset stalled state. */ 3162 if (PMC_IS_SAMPLING_MODE(mode)) 3163 pm->pm_pcpu_state[cpu].pps_stalled = 0; 3164 3165 /* Indicate that we desire this to run. Start it. */ 3166 pm->pm_pcpu_state[cpu].pps_cpustate = 1; 3167 error = pcd->pcd_start_pmc(cpu, adjri, pm); 3168 } 3169 critical_exit(); 3170 3171 pmc_restore_cpu_binding(&pb); 3172 return (error); 3173 } 3174 3175 /* 3176 * Stop a PMC. 3177 */ 3178 static int 3179 pmc_stop(struct pmc *pm) 3180 { 3181 struct pmc_binding pb; 3182 struct pmc_classdep *pcd; 3183 struct pmc_owner *po; 3184 int adjri, cpu, error, ri; 3185 3186 KASSERT(pm != NULL, ("[pmc,%d] null pmc", __LINE__)); 3187 3188 PMCDBG3(PMC,OPS,1, "stop pmc=%p mode=%d ri=%d", pm, PMC_TO_MODE(pm), 3189 PMC_TO_ROWINDEX(pm)); 3190 3191 pm->pm_state = PMC_STATE_STOPPED; 3192 3193 /* 3194 * If the PMC is a virtual mode one, changing the state to non-RUNNING 3195 * is enough to ensure that the PMC never gets scheduled. 3196 * 3197 * If this PMC is current running on a CPU, then it will handled 3198 * correctly at the time its target process is context switched out. 3199 */ 3200 if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) 3201 return (0); 3202 3203 /* 3204 * A system-mode PMC. Move to the CPU associated with this PMC, and 3205 * stop the hardware. We update the 'initial count' so that a 3206 * subsequent PMCSTART will resume counting from the current hardware 3207 * count. 3208 */ 3209 pmc_save_cpu_binding(&pb); 3210 3211 cpu = PMC_TO_CPU(pm); 3212 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), 3213 ("[pmc,%d] illegal cpu=%d", __LINE__, cpu)); 3214 if (!pmc_cpu_is_active(cpu)) 3215 return (ENXIO); 3216 3217 pmc_select_cpu(cpu); 3218 3219 ri = PMC_TO_ROWINDEX(pm); 3220 pcd = pmc_ri_to_classdep(md, ri, &adjri); 3221 3222 pm->pm_pcpu_state[cpu].pps_cpustate = 0; 3223 critical_enter(); 3224 if ((error = pcd->pcd_stop_pmc(cpu, adjri, pm)) == 0) { 3225 error = pcd->pcd_read_pmc(cpu, adjri, pm, 3226 &pm->pm_sc.pm_initial); 3227 } 3228 critical_exit(); 3229 3230 pmc_restore_cpu_binding(&pb); 3231 3232 /* Remove this owner from the global list of SS PMC owners. */ 3233 po = pm->pm_owner; 3234 if (PMC_TO_MODE(pm) == PMC_MODE_SS) { 3235 po->po_sscount--; 3236 if (po->po_sscount == 0) { 3237 atomic_subtract_rel_int(&pmc_ss_count, 1); 3238 CK_LIST_REMOVE(po, po_ssnext); 3239 epoch_wait_preempt(global_epoch_preempt); 3240 PMCDBG1(PMC,OPS,2,"po=%p removed from global list", po); 3241 } 3242 } 3243 3244 return (error); 3245 } 3246 3247 static struct pmc_classdep * 3248 pmc_class_to_classdep(enum pmc_class class) 3249 { 3250 int n; 3251 3252 for (n = 0; n < md->pmd_nclass; n++) { 3253 if (md->pmd_classdep[n].pcd_class == class) 3254 return (&md->pmd_classdep[n]); 3255 } 3256 return (NULL); 3257 } 3258 3259 #if defined(HWPMC_DEBUG) && defined(KTR) 3260 static const char *pmc_op_to_name[] = { 3261 #undef __PMC_OP 3262 #define __PMC_OP(N, D) #N , 3263 __PMC_OPS() 3264 NULL 3265 }; 3266 #endif 3267 3268 /* 3269 * The syscall interface 3270 */ 3271 3272 #define PMC_GET_SX_XLOCK(...) do { \ 3273 sx_xlock(&pmc_sx); \ 3274 if (pmc_hook == NULL) { \ 3275 sx_xunlock(&pmc_sx); \ 3276 return __VA_ARGS__; \ 3277 } \ 3278 } while (0) 3279 3280 #define PMC_DOWNGRADE_SX() do { \ 3281 sx_downgrade(&pmc_sx); \ 3282 is_sx_downgraded = true; \ 3283 } while (0) 3284 3285 /* 3286 * Main body of PMC_OP_PMCALLOCATE. 3287 */ 3288 static int 3289 pmc_do_op_pmcallocate(struct thread *td, struct pmc_op_pmcallocate *pa) 3290 { 3291 struct proc *p; 3292 struct pmc *pmc; 3293 struct pmc_binding pb; 3294 struct pmc_classdep *pcd; 3295 struct pmc_hw *phw; 3296 enum pmc_mode mode; 3297 enum pmc_class class; 3298 uint32_t caps, flags; 3299 u_int cpu; 3300 int adjri, n; 3301 int error; 3302 3303 class = pa->pm_class; 3304 caps = pa->pm_caps; 3305 flags = pa->pm_flags; 3306 mode = pa->pm_mode; 3307 cpu = pa->pm_cpu; 3308 3309 p = td->td_proc; 3310 3311 /* Requested mode must exist. */ 3312 if ((mode != PMC_MODE_SS && mode != PMC_MODE_SC && 3313 mode != PMC_MODE_TS && mode != PMC_MODE_TC)) 3314 return (EINVAL); 3315 3316 /* Requested CPU must be valid. */ 3317 if (cpu != PMC_CPU_ANY && cpu >= pmc_cpu_max()) 3318 return (EINVAL); 3319 3320 /* 3321 * Virtual PMCs should only ask for a default CPU. 3322 * System mode PMCs need to specify a non-default CPU. 3323 */ 3324 if ((PMC_IS_VIRTUAL_MODE(mode) && cpu != PMC_CPU_ANY) || 3325 (PMC_IS_SYSTEM_MODE(mode) && cpu == PMC_CPU_ANY)) 3326 return (EINVAL); 3327 3328 /* 3329 * Check that an inactive CPU is not being asked for. 3330 */ 3331 if (PMC_IS_SYSTEM_MODE(mode) && !pmc_cpu_is_active(cpu)) 3332 return (ENXIO); 3333 3334 /* 3335 * Refuse an allocation for a system-wide PMC if this process has been 3336 * jailed, or if this process lacks super-user credentials and the 3337 * sysctl tunable 'security.bsd.unprivileged_syspmcs' is zero. 3338 */ 3339 if (PMC_IS_SYSTEM_MODE(mode)) { 3340 if (jailed(td->td_ucred)) 3341 return (EPERM); 3342 if (!pmc_unprivileged_syspmcs) { 3343 error = priv_check(td, PRIV_PMC_SYSTEM); 3344 if (error != 0) 3345 return (error); 3346 } 3347 } 3348 3349 /* 3350 * Look for valid values for 'pm_flags'. 3351 */ 3352 if ((flags & ~(PMC_F_DESCENDANTS | PMC_F_LOG_PROCCSW | 3353 PMC_F_LOG_PROCEXIT | PMC_F_CALLCHAIN | PMC_F_USERCALLCHAIN | 3354 PMC_F_EV_PMU)) != 0) 3355 return (EINVAL); 3356 3357 /* PMC_F_USERCALLCHAIN is only valid with PMC_F_CALLCHAIN. */ 3358 if ((flags & (PMC_F_CALLCHAIN | PMC_F_USERCALLCHAIN)) == 3359 PMC_F_USERCALLCHAIN) 3360 return (EINVAL); 3361 3362 /* PMC_F_USERCALLCHAIN is only valid for sampling mode. */ 3363 if ((flags & PMC_F_USERCALLCHAIN) != 0 && mode != PMC_MODE_TS && 3364 mode != PMC_MODE_SS) 3365 return (EINVAL); 3366 3367 /* Process logging options are not allowed for system PMCs. */ 3368 if (PMC_IS_SYSTEM_MODE(mode) && 3369 (flags & (PMC_F_LOG_PROCCSW | PMC_F_LOG_PROCEXIT)) != 0) 3370 return (EINVAL); 3371 3372 /* 3373 * All sampling mode PMCs need to be able to interrupt the CPU. 3374 */ 3375 if (PMC_IS_SAMPLING_MODE(mode)) 3376 caps |= PMC_CAP_INTERRUPT; 3377 3378 /* A valid class specifier should have been passed in. */ 3379 pcd = pmc_class_to_classdep(class); 3380 if (pcd == NULL) 3381 return (EINVAL); 3382 3383 /* The requested PMC capabilities should be feasible. */ 3384 if ((pcd->pcd_caps & caps) != caps) 3385 return (EOPNOTSUPP); 3386 3387 PMCDBG4(PMC,ALL,2, "event=%d caps=0x%x mode=%d cpu=%d", pa->pm_ev, 3388 caps, mode, cpu); 3389 3390 pmc = pmc_allocate_pmc_descriptor(); 3391 pmc->pm_id = PMC_ID_MAKE_ID(cpu, pa->pm_mode, class, PMC_ID_INVALID); 3392 pmc->pm_event = pa->pm_ev; 3393 pmc->pm_state = PMC_STATE_FREE; 3394 pmc->pm_caps = caps; 3395 pmc->pm_flags = flags; 3396 3397 /* XXX set lower bound on sampling for process counters */ 3398 if (PMC_IS_SAMPLING_MODE(mode)) { 3399 /* 3400 * Don't permit requested sample rate to be less than 3401 * pmc_mincount. 3402 */ 3403 if (pa->pm_count < MAX(1, pmc_mincount)) 3404 log(LOG_WARNING, "pmcallocate: passed sample " 3405 "rate %ju - setting to %u\n", 3406 (uintmax_t)pa->pm_count, 3407 MAX(1, pmc_mincount)); 3408 pmc->pm_sc.pm_reloadcount = MAX(MAX(1, pmc_mincount), 3409 pa->pm_count); 3410 } else 3411 pmc->pm_sc.pm_initial = pa->pm_count; 3412 3413 /* switch thread to CPU 'cpu' */ 3414 pmc_save_cpu_binding(&pb); 3415 3416 #define PMC_IS_SHAREABLE_PMC(cpu, n) \ 3417 (pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_state & \ 3418 PMC_PHW_FLAG_IS_SHAREABLE) 3419 #define PMC_IS_UNALLOCATED(cpu, n) \ 3420 (pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_pmc == NULL) 3421 3422 if (PMC_IS_SYSTEM_MODE(mode)) { 3423 pmc_select_cpu(cpu); 3424 for (n = pcd->pcd_ri; n < md->pmd_npmc; n++) { 3425 pcd = pmc_ri_to_classdep(md, n, &adjri); 3426 3427 if (!pmc_can_allocate_row(n, mode) || 3428 !pmc_can_allocate_rowindex(p, n, cpu)) 3429 continue; 3430 if (!PMC_IS_UNALLOCATED(cpu, n) && 3431 !PMC_IS_SHAREABLE_PMC(cpu, n)) 3432 continue; 3433 3434 if (pcd->pcd_allocate_pmc(cpu, adjri, pmc, pa) == 0) { 3435 /* Success. */ 3436 break; 3437 } 3438 } 3439 } else { 3440 /* Process virtual mode */ 3441 for (n = pcd->pcd_ri; n < md->pmd_npmc; n++) { 3442 pcd = pmc_ri_to_classdep(md, n, &adjri); 3443 3444 if (!pmc_can_allocate_row(n, mode) || 3445 !pmc_can_allocate_rowindex(p, n, PMC_CPU_ANY)) 3446 continue; 3447 3448 if (pcd->pcd_allocate_pmc(td->td_oncpu, adjri, pmc, 3449 pa) == 0) { 3450 /* Success. */ 3451 break; 3452 } 3453 } 3454 } 3455 3456 #undef PMC_IS_UNALLOCATED 3457 #undef PMC_IS_SHAREABLE_PMC 3458 3459 pmc_restore_cpu_binding(&pb); 3460 3461 if (n == md->pmd_npmc) { 3462 pmc_destroy_pmc_descriptor(pmc); 3463 return (EINVAL); 3464 } 3465 3466 /* Fill in the correct value in the ID field. */ 3467 pmc->pm_id = PMC_ID_MAKE_ID(cpu, mode, class, n); 3468 3469 PMCDBG5(PMC,ALL,2, "ev=%d class=%d mode=%d n=%d -> pmcid=%x", 3470 pmc->pm_event, class, mode, n, pmc->pm_id); 3471 3472 /* Process mode PMCs with logging enabled need log files. */ 3473 if ((pmc->pm_flags & (PMC_F_LOG_PROCEXIT | PMC_F_LOG_PROCCSW)) != 0) 3474 pmc->pm_flags |= PMC_F_NEEDS_LOGFILE; 3475 3476 /* All system mode sampling PMCs require a log file. */ 3477 if (PMC_IS_SAMPLING_MODE(mode) && PMC_IS_SYSTEM_MODE(mode)) 3478 pmc->pm_flags |= PMC_F_NEEDS_LOGFILE; 3479 3480 /* 3481 * Configure global pmc's immediately. 3482 */ 3483 if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pmc))) { 3484 pmc_save_cpu_binding(&pb); 3485 pmc_select_cpu(cpu); 3486 3487 phw = pmc_pcpu[cpu]->pc_hwpmcs[n]; 3488 pcd = pmc_ri_to_classdep(md, n, &adjri); 3489 3490 if ((phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0 || 3491 (error = pcd->pcd_config_pmc(cpu, adjri, pmc)) != 0) { 3492 (void)pcd->pcd_release_pmc(cpu, adjri, pmc); 3493 pmc_destroy_pmc_descriptor(pmc); 3494 pmc_restore_cpu_binding(&pb); 3495 return (EPERM); 3496 } 3497 3498 pmc_restore_cpu_binding(&pb); 3499 } 3500 3501 pmc->pm_state = PMC_STATE_ALLOCATED; 3502 pmc->pm_class = class; 3503 3504 /* 3505 * Mark row disposition. 3506 */ 3507 if (PMC_IS_SYSTEM_MODE(mode)) 3508 PMC_MARK_ROW_STANDALONE(n); 3509 else 3510 PMC_MARK_ROW_THREAD(n); 3511 3512 /* 3513 * Register this PMC with the current thread as its owner. 3514 */ 3515 error = pmc_register_owner(p, pmc); 3516 if (error != 0) { 3517 pmc_release_pmc_descriptor(pmc); 3518 pmc_destroy_pmc_descriptor(pmc); 3519 return (error); 3520 } 3521 3522 /* 3523 * Return the allocated index. 3524 */ 3525 pa->pm_pmcid = pmc->pm_id; 3526 return (0); 3527 } 3528 3529 /* 3530 * Main body of PMC_OP_PMCATTACH. 3531 */ 3532 static int 3533 pmc_do_op_pmcattach(struct thread *td, struct pmc_op_pmcattach a) 3534 { 3535 struct pmc *pm; 3536 struct proc *p; 3537 int error; 3538 3539 sx_assert(&pmc_sx, SX_XLOCKED); 3540 3541 if (a.pm_pid < 0) { 3542 return (EINVAL); 3543 } else if (a.pm_pid == 0) { 3544 a.pm_pid = td->td_proc->p_pid; 3545 } 3546 3547 error = pmc_find_pmc(a.pm_pmc, &pm); 3548 if (error != 0) 3549 return (error); 3550 3551 if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm))) 3552 return (EINVAL); 3553 3554 /* PMCs may be (re)attached only when allocated or stopped */ 3555 if (pm->pm_state == PMC_STATE_RUNNING) { 3556 return (EBUSY); 3557 } else if (pm->pm_state != PMC_STATE_ALLOCATED && 3558 pm->pm_state != PMC_STATE_STOPPED) { 3559 return (EINVAL); 3560 } 3561 3562 /* lookup pid */ 3563 if ((p = pfind(a.pm_pid)) == NULL) 3564 return (ESRCH); 3565 3566 /* 3567 * Ignore processes that are working on exiting. 3568 */ 3569 if ((p->p_flag & P_WEXIT) != 0) { 3570 PROC_UNLOCK(p); /* pfind() returns a locked process */ 3571 return (ESRCH); 3572 } 3573 3574 /* 3575 * We are allowed to attach a PMC to a process if we can debug it. 3576 */ 3577 error = p_candebug(curthread, p); 3578 3579 PROC_UNLOCK(p); 3580 3581 if (error == 0) 3582 error = pmc_attach_process(p, pm); 3583 3584 return (error); 3585 } 3586 3587 /* 3588 * Main body of PMC_OP_PMCDETACH. 3589 */ 3590 static int 3591 pmc_do_op_pmcdetach(struct thread *td, struct pmc_op_pmcattach a) 3592 { 3593 struct pmc *pm; 3594 struct proc *p; 3595 int error; 3596 3597 if (a.pm_pid < 0) { 3598 return (EINVAL); 3599 } else if (a.pm_pid == 0) 3600 a.pm_pid = td->td_proc->p_pid; 3601 3602 error = pmc_find_pmc(a.pm_pmc, &pm); 3603 if (error != 0) 3604 return (error); 3605 3606 if ((p = pfind(a.pm_pid)) == NULL) 3607 return (ESRCH); 3608 3609 /* 3610 * Treat processes that are in the process of exiting as if they were 3611 * not present. 3612 */ 3613 if ((p->p_flag & P_WEXIT) != 0) { 3614 PROC_UNLOCK(p); 3615 return (ESRCH); 3616 } 3617 3618 PROC_UNLOCK(p); /* pfind() returns a locked process */ 3619 3620 if (error == 0) 3621 error = pmc_detach_process(p, pm); 3622 3623 return (error); 3624 } 3625 3626 /* 3627 * Main body of PMC_OP_PMCRELEASE. 3628 */ 3629 static int 3630 pmc_do_op_pmcrelease(pmc_id_t pmcid) 3631 { 3632 struct pmc_owner *po; 3633 struct pmc *pm; 3634 int error; 3635 3636 /* 3637 * Find PMC pointer for the named PMC. 3638 * 3639 * Use pmc_release_pmc_descriptor() to switch off the 3640 * PMC, remove all its target threads, and remove the 3641 * PMC from its owner's list. 3642 * 3643 * Remove the owner record if this is the last PMC 3644 * owned. 3645 * 3646 * Free up space. 3647 */ 3648 error = pmc_find_pmc(pmcid, &pm); 3649 if (error != 0) 3650 return (error); 3651 3652 po = pm->pm_owner; 3653 pmc_release_pmc_descriptor(pm); 3654 pmc_maybe_remove_owner(po); 3655 pmc_destroy_pmc_descriptor(pm); 3656 3657 return (error); 3658 } 3659 3660 /* 3661 * Main body of PMC_OP_PMCRW. 3662 */ 3663 static int 3664 pmc_do_op_pmcrw(const struct pmc_op_pmcrw *prw, pmc_value_t *valp) 3665 { 3666 struct pmc_binding pb; 3667 struct pmc_classdep *pcd; 3668 struct pmc *pm; 3669 u_int cpu, ri, adjri; 3670 int error; 3671 3672 PMCDBG2(PMC,OPS,1, "rw id=%d flags=0x%x", prw->pm_pmcid, prw->pm_flags); 3673 3674 /* Must have at least one flag set. */ 3675 if ((prw->pm_flags & (PMC_F_OLDVALUE | PMC_F_NEWVALUE)) == 0) 3676 return (EINVAL); 3677 3678 /* Locate PMC descriptor. */ 3679 error = pmc_find_pmc(prw->pm_pmcid, &pm); 3680 if (error != 0) 3681 return (error); 3682 3683 /* Can't read a PMC that hasn't been started. */ 3684 if (pm->pm_state != PMC_STATE_ALLOCATED && 3685 pm->pm_state != PMC_STATE_STOPPED && 3686 pm->pm_state != PMC_STATE_RUNNING) 3687 return (EINVAL); 3688 3689 /* Writing a new value is allowed only for 'STOPPED' PMCs. */ 3690 if (pm->pm_state == PMC_STATE_RUNNING && 3691 (prw->pm_flags & PMC_F_NEWVALUE) != 0) 3692 return (EBUSY); 3693 3694 if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) { 3695 /* 3696 * If this PMC is attached to its owner (i.e., the process 3697 * requesting this operation) and is running, then attempt to 3698 * get an upto-date reading from hardware for a READ. Writes 3699 * are only allowed when the PMC is stopped, so only update the 3700 * saved value field. 3701 * 3702 * If the PMC is not running, or is not attached to its owner, 3703 * read/write to the savedvalue field. 3704 */ 3705 3706 ri = PMC_TO_ROWINDEX(pm); 3707 pcd = pmc_ri_to_classdep(md, ri, &adjri); 3708 3709 mtx_pool_lock_spin(pmc_mtxpool, pm); 3710 cpu = curthread->td_oncpu; 3711 3712 if ((prw->pm_flags & PMC_F_OLDVALUE) != 0) { 3713 if ((pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) && 3714 (pm->pm_state == PMC_STATE_RUNNING)) { 3715 error = (*pcd->pcd_read_pmc)(cpu, adjri, pm, 3716 valp); 3717 } else { 3718 *valp = pm->pm_gv.pm_savedvalue; 3719 } 3720 } 3721 3722 if ((prw->pm_flags & PMC_F_NEWVALUE) != 0) 3723 pm->pm_gv.pm_savedvalue = prw->pm_value; 3724 3725 mtx_pool_unlock_spin(pmc_mtxpool, pm); 3726 } else { /* System mode PMCs */ 3727 cpu = PMC_TO_CPU(pm); 3728 ri = PMC_TO_ROWINDEX(pm); 3729 pcd = pmc_ri_to_classdep(md, ri, &adjri); 3730 3731 if (!pmc_cpu_is_active(cpu)) 3732 return (ENXIO); 3733 3734 /* Move this thread to CPU 'cpu'. */ 3735 pmc_save_cpu_binding(&pb); 3736 pmc_select_cpu(cpu); 3737 critical_enter(); 3738 3739 /* Save old value. */ 3740 if ((prw->pm_flags & PMC_F_OLDVALUE) != 0) 3741 error = (*pcd->pcd_read_pmc)(cpu, adjri, pm, valp); 3742 3743 /* Write out new value. */ 3744 if (error == 0 && (prw->pm_flags & PMC_F_NEWVALUE) != 0) 3745 error = (*pcd->pcd_write_pmc)(cpu, adjri, pm, 3746 prw->pm_value); 3747 3748 critical_exit(); 3749 pmc_restore_cpu_binding(&pb); 3750 if (error != 0) 3751 return (error); 3752 } 3753 3754 #ifdef HWPMC_DEBUG 3755 if ((prw->pm_flags & PMC_F_NEWVALUE) != 0) 3756 PMCDBG3(PMC,OPS,2, "rw id=%d new %jx -> old %jx", 3757 ri, prw->pm_value, *valp); 3758 else 3759 PMCDBG2(PMC,OPS,2, "rw id=%d -> old %jx", ri, *valp); 3760 #endif 3761 return (error); 3762 } 3763 3764 static int 3765 pmc_syscall_handler(struct thread *td, void *syscall_args) 3766 { 3767 struct pmc_syscall_args *c; 3768 void *pmclog_proc_handle; 3769 void *arg; 3770 int error, op; 3771 bool is_sx_downgraded; 3772 3773 c = (struct pmc_syscall_args *)syscall_args; 3774 op = c->pmop_code; 3775 arg = c->pmop_data; 3776 3777 /* PMC isn't set up yet */ 3778 if (pmc_hook == NULL) 3779 return (EINVAL); 3780 3781 if (op == PMC_OP_CONFIGURELOG) { 3782 /* 3783 * We cannot create the logging process inside 3784 * pmclog_configure_log() because there is a LOR 3785 * between pmc_sx and process structure locks. 3786 * Instead, pre-create the process and ignite the loop 3787 * if everything is fine, otherwise direct the process 3788 * to exit. 3789 */ 3790 error = pmclog_proc_create(td, &pmclog_proc_handle); 3791 if (error != 0) 3792 goto done_syscall; 3793 } 3794 3795 PMC_GET_SX_XLOCK(ENOSYS); 3796 is_sx_downgraded = false; 3797 PMCDBG3(MOD,PMS,1, "syscall op=%d \"%s\" arg=%p", op, 3798 pmc_op_to_name[op], arg); 3799 3800 error = 0; 3801 counter_u64_add(pmc_stats.pm_syscalls, 1); 3802 3803 switch (op) { 3804 3805 3806 /* 3807 * Configure a log file. 3808 * 3809 * XXX This OP will be reworked. 3810 */ 3811 3812 case PMC_OP_CONFIGURELOG: 3813 { 3814 struct proc *p; 3815 struct pmc *pm; 3816 struct pmc_owner *po; 3817 struct pmc_op_configurelog cl; 3818 3819 if ((error = copyin(arg, &cl, sizeof(cl))) != 0) { 3820 pmclog_proc_ignite(pmclog_proc_handle, NULL); 3821 break; 3822 } 3823 3824 /* No flags currently implemented */ 3825 if (cl.pm_flags != 0) { 3826 pmclog_proc_ignite(pmclog_proc_handle, NULL); 3827 error = EINVAL; 3828 break; 3829 } 3830 3831 /* mark this process as owning a log file */ 3832 p = td->td_proc; 3833 if ((po = pmc_find_owner_descriptor(p)) == NULL) 3834 if ((po = pmc_allocate_owner_descriptor(p)) == NULL) { 3835 pmclog_proc_ignite(pmclog_proc_handle, NULL); 3836 error = ENOMEM; 3837 break; 3838 } 3839 3840 /* 3841 * If a valid fd was passed in, try to configure that, 3842 * otherwise if 'fd' was less than zero and there was 3843 * a log file configured, flush its buffers and 3844 * de-configure it. 3845 */ 3846 if (cl.pm_logfd >= 0) { 3847 error = pmclog_configure_log(md, po, cl.pm_logfd); 3848 pmclog_proc_ignite(pmclog_proc_handle, error == 0 ? 3849 po : NULL); 3850 } else if (po->po_flags & PMC_PO_OWNS_LOGFILE) { 3851 pmclog_proc_ignite(pmclog_proc_handle, NULL); 3852 error = pmclog_close(po); 3853 if (error == 0) { 3854 LIST_FOREACH(pm, &po->po_pmcs, pm_next) 3855 if (pm->pm_flags & PMC_F_NEEDS_LOGFILE && 3856 pm->pm_state == PMC_STATE_RUNNING) 3857 pmc_stop(pm); 3858 error = pmclog_deconfigure_log(po); 3859 } 3860 } else { 3861 pmclog_proc_ignite(pmclog_proc_handle, NULL); 3862 error = EINVAL; 3863 } 3864 } 3865 break; 3866 3867 /* 3868 * Flush a log file. 3869 */ 3870 3871 case PMC_OP_FLUSHLOG: 3872 { 3873 struct pmc_owner *po; 3874 3875 sx_assert(&pmc_sx, SX_XLOCKED); 3876 3877 if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) { 3878 error = EINVAL; 3879 break; 3880 } 3881 3882 error = pmclog_flush(po, 0); 3883 } 3884 break; 3885 3886 /* 3887 * Close a log file. 3888 */ 3889 3890 case PMC_OP_CLOSELOG: 3891 { 3892 struct pmc_owner *po; 3893 3894 sx_assert(&pmc_sx, SX_XLOCKED); 3895 3896 if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) { 3897 error = EINVAL; 3898 break; 3899 } 3900 3901 error = pmclog_close(po); 3902 } 3903 break; 3904 3905 /* 3906 * Retrieve hardware configuration. 3907 */ 3908 3909 case PMC_OP_GETCPUINFO: /* CPU information */ 3910 { 3911 struct pmc_op_getcpuinfo gci; 3912 struct pmc_classinfo *pci; 3913 struct pmc_classdep *pcd; 3914 int cl; 3915 3916 memset(&gci, 0, sizeof(gci)); 3917 gci.pm_cputype = md->pmd_cputype; 3918 gci.pm_ncpu = pmc_cpu_max(); 3919 gci.pm_npmc = md->pmd_npmc; 3920 gci.pm_nclass = md->pmd_nclass; 3921 pci = gci.pm_classes; 3922 pcd = md->pmd_classdep; 3923 for (cl = 0; cl < md->pmd_nclass; cl++, pci++, pcd++) { 3924 pci->pm_caps = pcd->pcd_caps; 3925 pci->pm_class = pcd->pcd_class; 3926 pci->pm_width = pcd->pcd_width; 3927 pci->pm_num = pcd->pcd_num; 3928 } 3929 error = copyout(&gci, arg, sizeof(gci)); 3930 } 3931 break; 3932 3933 /* 3934 * Retrieve soft events list. 3935 */ 3936 case PMC_OP_GETDYNEVENTINFO: 3937 { 3938 enum pmc_class cl; 3939 enum pmc_event ev; 3940 struct pmc_op_getdyneventinfo *gei; 3941 struct pmc_dyn_event_descr dev; 3942 struct pmc_soft *ps; 3943 uint32_t nevent; 3944 3945 sx_assert(&pmc_sx, SX_LOCKED); 3946 3947 gei = (struct pmc_op_getdyneventinfo *) arg; 3948 3949 if ((error = copyin(&gei->pm_class, &cl, sizeof(cl))) != 0) 3950 break; 3951 3952 /* Only SOFT class is dynamic. */ 3953 if (cl != PMC_CLASS_SOFT) { 3954 error = EINVAL; 3955 break; 3956 } 3957 3958 nevent = 0; 3959 for (ev = PMC_EV_SOFT_FIRST; (int)ev <= PMC_EV_SOFT_LAST; ev++) { 3960 ps = pmc_soft_ev_acquire(ev); 3961 if (ps == NULL) 3962 continue; 3963 bcopy(&ps->ps_ev, &dev, sizeof(dev)); 3964 pmc_soft_ev_release(ps); 3965 3966 error = copyout(&dev, 3967 &gei->pm_events[nevent], 3968 sizeof(struct pmc_dyn_event_descr)); 3969 if (error != 0) 3970 break; 3971 nevent++; 3972 } 3973 if (error != 0) 3974 break; 3975 3976 error = copyout(&nevent, &gei->pm_nevent, 3977 sizeof(nevent)); 3978 } 3979 break; 3980 3981 /* 3982 * Get module statistics 3983 */ 3984 3985 case PMC_OP_GETDRIVERSTATS: 3986 { 3987 struct pmc_op_getdriverstats gms; 3988 #define CFETCH(a, b, field) a.field = counter_u64_fetch(b.field) 3989 CFETCH(gms, pmc_stats, pm_intr_ignored); 3990 CFETCH(gms, pmc_stats, pm_intr_processed); 3991 CFETCH(gms, pmc_stats, pm_intr_bufferfull); 3992 CFETCH(gms, pmc_stats, pm_syscalls); 3993 CFETCH(gms, pmc_stats, pm_syscall_errors); 3994 CFETCH(gms, pmc_stats, pm_buffer_requests); 3995 CFETCH(gms, pmc_stats, pm_buffer_requests_failed); 3996 CFETCH(gms, pmc_stats, pm_log_sweeps); 3997 #undef CFETCH 3998 error = copyout(&gms, arg, sizeof(gms)); 3999 } 4000 break; 4001 4002 4003 /* 4004 * Retrieve module version number 4005 */ 4006 4007 case PMC_OP_GETMODULEVERSION: 4008 { 4009 uint32_t cv, modv; 4010 4011 /* retrieve the client's idea of the ABI version */ 4012 if ((error = copyin(arg, &cv, sizeof(uint32_t))) != 0) 4013 break; 4014 /* don't service clients newer than our driver */ 4015 modv = PMC_VERSION; 4016 if ((cv & 0xFFFF0000) > (modv & 0xFFFF0000)) { 4017 error = EPROGMISMATCH; 4018 break; 4019 } 4020 error = copyout(&modv, arg, sizeof(int)); 4021 } 4022 break; 4023 4024 4025 /* 4026 * Retrieve the state of all the PMCs on a given 4027 * CPU. 4028 */ 4029 4030 case PMC_OP_GETPMCINFO: 4031 { 4032 int ari; 4033 struct pmc *pm; 4034 size_t pmcinfo_size; 4035 uint32_t cpu, n, npmc; 4036 struct pmc_owner *po; 4037 struct pmc_binding pb; 4038 struct pmc_classdep *pcd; 4039 struct pmc_info *p, *pmcinfo; 4040 struct pmc_op_getpmcinfo *gpi; 4041 4042 PMC_DOWNGRADE_SX(); 4043 4044 gpi = (struct pmc_op_getpmcinfo *) arg; 4045 4046 if ((error = copyin(&gpi->pm_cpu, &cpu, sizeof(cpu))) != 0) 4047 break; 4048 4049 if (cpu >= pmc_cpu_max()) { 4050 error = EINVAL; 4051 break; 4052 } 4053 4054 if (!pmc_cpu_is_active(cpu)) { 4055 error = ENXIO; 4056 break; 4057 } 4058 4059 /* switch to CPU 'cpu' */ 4060 pmc_save_cpu_binding(&pb); 4061 pmc_select_cpu(cpu); 4062 4063 npmc = md->pmd_npmc; 4064 4065 pmcinfo_size = npmc * sizeof(struct pmc_info); 4066 pmcinfo = malloc(pmcinfo_size, M_PMC, M_WAITOK | M_ZERO); 4067 4068 p = pmcinfo; 4069 4070 for (n = 0; n < md->pmd_npmc; n++, p++) { 4071 4072 pcd = pmc_ri_to_classdep(md, n, &ari); 4073 4074 KASSERT(pcd != NULL, 4075 ("[pmc,%d] null pcd ri=%d", __LINE__, n)); 4076 4077 if ((error = pcd->pcd_describe(cpu, ari, p, &pm)) != 0) 4078 break; 4079 4080 if (PMC_ROW_DISP_IS_STANDALONE(n)) 4081 p->pm_rowdisp = PMC_DISP_STANDALONE; 4082 else if (PMC_ROW_DISP_IS_THREAD(n)) 4083 p->pm_rowdisp = PMC_DISP_THREAD; 4084 else 4085 p->pm_rowdisp = PMC_DISP_FREE; 4086 4087 p->pm_ownerpid = -1; 4088 4089 if (pm == NULL) /* no PMC associated */ 4090 continue; 4091 4092 po = pm->pm_owner; 4093 4094 KASSERT(po->po_owner != NULL, 4095 ("[pmc,%d] pmc_owner had a null proc pointer", 4096 __LINE__)); 4097 4098 p->pm_ownerpid = po->po_owner->p_pid; 4099 p->pm_mode = PMC_TO_MODE(pm); 4100 p->pm_event = pm->pm_event; 4101 p->pm_flags = pm->pm_flags; 4102 4103 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) 4104 p->pm_reloadcount = 4105 pm->pm_sc.pm_reloadcount; 4106 } 4107 4108 pmc_restore_cpu_binding(&pb); 4109 4110 /* now copy out the PMC info collected */ 4111 if (error == 0) 4112 error = copyout(pmcinfo, &gpi->pm_pmcs, pmcinfo_size); 4113 4114 free(pmcinfo, M_PMC); 4115 } 4116 break; 4117 4118 4119 /* 4120 * Set the administrative state of a PMC. I.e. whether 4121 * the PMC is to be used or not. 4122 */ 4123 4124 case PMC_OP_PMCADMIN: 4125 { 4126 int cpu, ri; 4127 enum pmc_state request; 4128 struct pmc_cpu *pc; 4129 struct pmc_hw *phw; 4130 struct pmc_op_pmcadmin pma; 4131 struct pmc_binding pb; 4132 4133 sx_assert(&pmc_sx, SX_XLOCKED); 4134 4135 KASSERT(td == curthread, 4136 ("[pmc,%d] td != curthread", __LINE__)); 4137 4138 error = priv_check(td, PRIV_PMC_MANAGE); 4139 if (error) 4140 break; 4141 4142 if ((error = copyin(arg, &pma, sizeof(pma))) != 0) 4143 break; 4144 4145 cpu = pma.pm_cpu; 4146 4147 if (cpu < 0 || cpu >= (int) pmc_cpu_max()) { 4148 error = EINVAL; 4149 break; 4150 } 4151 4152 if (!pmc_cpu_is_active(cpu)) { 4153 error = ENXIO; 4154 break; 4155 } 4156 4157 request = pma.pm_state; 4158 4159 if (request != PMC_STATE_DISABLED && 4160 request != PMC_STATE_FREE) { 4161 error = EINVAL; 4162 break; 4163 } 4164 4165 ri = pma.pm_pmc; /* pmc id == row index */ 4166 if (ri < 0 || ri >= (int) md->pmd_npmc) { 4167 error = EINVAL; 4168 break; 4169 } 4170 4171 /* 4172 * We can't disable a PMC with a row-index allocated 4173 * for process virtual PMCs. 4174 */ 4175 4176 if (PMC_ROW_DISP_IS_THREAD(ri) && 4177 request == PMC_STATE_DISABLED) { 4178 error = EBUSY; 4179 break; 4180 } 4181 4182 /* 4183 * otherwise, this PMC on this CPU is either free or 4184 * in system-wide mode. 4185 */ 4186 4187 pmc_save_cpu_binding(&pb); 4188 pmc_select_cpu(cpu); 4189 4190 pc = pmc_pcpu[cpu]; 4191 phw = pc->pc_hwpmcs[ri]; 4192 4193 /* 4194 * XXX do we need some kind of 'forced' disable? 4195 */ 4196 4197 if (phw->phw_pmc == NULL) { 4198 if (request == PMC_STATE_DISABLED && 4199 (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED)) { 4200 phw->phw_state &= ~PMC_PHW_FLAG_IS_ENABLED; 4201 PMC_MARK_ROW_STANDALONE(ri); 4202 } else if (request == PMC_STATE_FREE && 4203 (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0) { 4204 phw->phw_state |= PMC_PHW_FLAG_IS_ENABLED; 4205 PMC_UNMARK_ROW_STANDALONE(ri); 4206 } 4207 /* other cases are a no-op */ 4208 } else 4209 error = EBUSY; 4210 4211 pmc_restore_cpu_binding(&pb); 4212 } 4213 break; 4214 4215 4216 /* 4217 * Allocate a PMC. 4218 */ 4219 case PMC_OP_PMCALLOCATE: 4220 { 4221 struct pmc_op_pmcallocate pa; 4222 4223 error = copyin(arg, &pa, sizeof(pa)); 4224 if (error != 0) 4225 break; 4226 4227 error = pmc_do_op_pmcallocate(td, &pa); 4228 if (error != 0) 4229 break; 4230 4231 error = copyout(&pa, arg, sizeof(pa)); 4232 } 4233 break; 4234 4235 /* 4236 * Attach a PMC to a process. 4237 */ 4238 case PMC_OP_PMCATTACH: 4239 { 4240 struct pmc_op_pmcattach a; 4241 4242 error = copyin(arg, &a, sizeof(a)); 4243 if (error != 0) 4244 break; 4245 4246 error = pmc_do_op_pmcattach(td, a); 4247 } 4248 break; 4249 4250 /* 4251 * Detach an attached PMC from a process. 4252 */ 4253 case PMC_OP_PMCDETACH: 4254 { 4255 struct pmc_op_pmcattach a; 4256 4257 error = copyin(arg, &a, sizeof(a)); 4258 if (error != 0) 4259 break; 4260 4261 error = pmc_do_op_pmcdetach(td, a); 4262 } 4263 break; 4264 4265 4266 /* 4267 * Retrieve the MSR number associated with the counter 4268 * 'pmc_id'. This allows processes to directly use RDPMC 4269 * instructions to read their PMCs, without the overhead of a 4270 * system call. 4271 */ 4272 4273 case PMC_OP_PMCGETMSR: 4274 { 4275 int adjri, ri; 4276 struct pmc *pm; 4277 struct pmc_target *pt; 4278 struct pmc_op_getmsr gm; 4279 struct pmc_classdep *pcd; 4280 4281 PMC_DOWNGRADE_SX(); 4282 4283 if ((error = copyin(arg, &gm, sizeof(gm))) != 0) 4284 break; 4285 4286 if ((error = pmc_find_pmc(gm.pm_pmcid, &pm)) != 0) 4287 break; 4288 4289 /* 4290 * The allocated PMC has to be a process virtual PMC, 4291 * i.e., of type MODE_T[CS]. Global PMCs can only be 4292 * read using the PMCREAD operation since they may be 4293 * allocated on a different CPU than the one we could 4294 * be running on at the time of the RDPMC instruction. 4295 * 4296 * The GETMSR operation is not allowed for PMCs that 4297 * are inherited across processes. 4298 */ 4299 4300 if (!PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)) || 4301 (pm->pm_flags & PMC_F_DESCENDANTS)) { 4302 error = EINVAL; 4303 break; 4304 } 4305 4306 /* 4307 * It only makes sense to use a RDPMC (or its 4308 * equivalent instruction on non-x86 architectures) on 4309 * a process that has allocated and attached a PMC to 4310 * itself. Conversely the PMC is only allowed to have 4311 * one process attached to it -- its owner. 4312 */ 4313 4314 if ((pt = LIST_FIRST(&pm->pm_targets)) == NULL || 4315 LIST_NEXT(pt, pt_next) != NULL || 4316 pt->pt_process->pp_proc != pm->pm_owner->po_owner) { 4317 error = EINVAL; 4318 break; 4319 } 4320 4321 ri = PMC_TO_ROWINDEX(pm); 4322 pcd = pmc_ri_to_classdep(md, ri, &adjri); 4323 4324 /* PMC class has no 'GETMSR' support */ 4325 if (pcd->pcd_get_msr == NULL) { 4326 error = ENOSYS; 4327 break; 4328 } 4329 4330 if ((error = (*pcd->pcd_get_msr)(adjri, &gm.pm_msr)) < 0) 4331 break; 4332 4333 if ((error = copyout(&gm, arg, sizeof(gm))) < 0) 4334 break; 4335 4336 /* 4337 * Mark our process as using MSRs. Update machine 4338 * state using a forced context switch. 4339 */ 4340 4341 pt->pt_process->pp_flags |= PMC_PP_ENABLE_MSR_ACCESS; 4342 pmc_force_context_switch(); 4343 4344 } 4345 break; 4346 4347 /* 4348 * Release an allocated PMC. 4349 */ 4350 case PMC_OP_PMCRELEASE: 4351 { 4352 struct pmc_op_simple sp; 4353 4354 error = copyin(arg, &sp, sizeof(sp)); 4355 if (error != 0) 4356 break; 4357 4358 error = pmc_do_op_pmcrelease(sp.pm_pmcid); 4359 } 4360 break; 4361 4362 /* 4363 * Read and/or write a PMC. 4364 */ 4365 case PMC_OP_PMCRW: 4366 { 4367 struct pmc_op_pmcrw prw; 4368 struct pmc_op_pmcrw *pprw; 4369 pmc_value_t oldvalue; 4370 4371 PMC_DOWNGRADE_SX(); 4372 4373 error = copyin(arg, &prw, sizeof(prw)); 4374 if (error != 0) 4375 break; 4376 4377 error = pmc_do_op_pmcrw(&prw, &oldvalue); 4378 if (error != 0) 4379 break; 4380 4381 /* Return old value if requested. */ 4382 if ((prw.pm_flags & PMC_F_OLDVALUE) != 0) { 4383 pprw = arg; 4384 error = copyout(&oldvalue, &pprw->pm_value, 4385 sizeof(prw.pm_value)); 4386 } 4387 } 4388 break; 4389 4390 4391 /* 4392 * Set the sampling rate for a sampling mode PMC and the 4393 * initial count for a counting mode PMC. 4394 */ 4395 4396 case PMC_OP_PMCSETCOUNT: 4397 { 4398 struct pmc *pm; 4399 struct pmc_op_pmcsetcount sc; 4400 4401 PMC_DOWNGRADE_SX(); 4402 4403 if ((error = copyin(arg, &sc, sizeof(sc))) != 0) 4404 break; 4405 4406 if ((error = pmc_find_pmc(sc.pm_pmcid, &pm)) != 0) 4407 break; 4408 4409 if (pm->pm_state == PMC_STATE_RUNNING) { 4410 error = EBUSY; 4411 break; 4412 } 4413 4414 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) { 4415 /* 4416 * Don't permit requested sample rate to be 4417 * less than pmc_mincount. 4418 */ 4419 if (sc.pm_count < MAX(1, pmc_mincount)) 4420 log(LOG_WARNING, "pmcsetcount: passed sample " 4421 "rate %ju - setting to %u\n", 4422 (uintmax_t)sc.pm_count, 4423 MAX(1, pmc_mincount)); 4424 pm->pm_sc.pm_reloadcount = MAX(MAX(1, pmc_mincount), 4425 sc.pm_count); 4426 } else 4427 pm->pm_sc.pm_initial = sc.pm_count; 4428 } 4429 break; 4430 4431 4432 /* 4433 * Start a PMC. 4434 */ 4435 4436 case PMC_OP_PMCSTART: 4437 { 4438 pmc_id_t pmcid; 4439 struct pmc *pm; 4440 struct pmc_op_simple sp; 4441 4442 sx_assert(&pmc_sx, SX_XLOCKED); 4443 4444 if ((error = copyin(arg, &sp, sizeof(sp))) != 0) 4445 break; 4446 4447 pmcid = sp.pm_pmcid; 4448 4449 if ((error = pmc_find_pmc(pmcid, &pm)) != 0) 4450 break; 4451 4452 KASSERT(pmcid == pm->pm_id, 4453 ("[pmc,%d] pmcid %x != id %x", __LINE__, 4454 pm->pm_id, pmcid)); 4455 4456 if (pm->pm_state == PMC_STATE_RUNNING) /* already running */ 4457 break; 4458 else if (pm->pm_state != PMC_STATE_STOPPED && 4459 pm->pm_state != PMC_STATE_ALLOCATED) { 4460 error = EINVAL; 4461 break; 4462 } 4463 4464 error = pmc_start(pm); 4465 } 4466 break; 4467 4468 4469 /* 4470 * Stop a PMC. 4471 */ 4472 4473 case PMC_OP_PMCSTOP: 4474 { 4475 pmc_id_t pmcid; 4476 struct pmc *pm; 4477 struct pmc_op_simple sp; 4478 4479 PMC_DOWNGRADE_SX(); 4480 4481 if ((error = copyin(arg, &sp, sizeof(sp))) != 0) 4482 break; 4483 4484 pmcid = sp.pm_pmcid; 4485 4486 /* 4487 * Mark the PMC as inactive and invoke the MD stop 4488 * routines if needed. 4489 */ 4490 4491 if ((error = pmc_find_pmc(pmcid, &pm)) != 0) 4492 break; 4493 4494 KASSERT(pmcid == pm->pm_id, 4495 ("[pmc,%d] pmc id %x != pmcid %x", __LINE__, 4496 pm->pm_id, pmcid)); 4497 4498 if (pm->pm_state == PMC_STATE_STOPPED) /* already stopped */ 4499 break; 4500 else if (pm->pm_state != PMC_STATE_RUNNING) { 4501 error = EINVAL; 4502 break; 4503 } 4504 4505 error = pmc_stop(pm); 4506 } 4507 break; 4508 4509 4510 /* 4511 * Write a user supplied value to the log file. 4512 */ 4513 4514 case PMC_OP_WRITELOG: 4515 { 4516 struct pmc_op_writelog wl; 4517 struct pmc_owner *po; 4518 4519 PMC_DOWNGRADE_SX(); 4520 4521 if ((error = copyin(arg, &wl, sizeof(wl))) != 0) 4522 break; 4523 4524 if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) { 4525 error = EINVAL; 4526 break; 4527 } 4528 4529 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) { 4530 error = EINVAL; 4531 break; 4532 } 4533 4534 error = pmclog_process_userlog(po, &wl); 4535 } 4536 break; 4537 4538 /* 4539 * Get the PMC capabilities 4540 */ 4541 4542 case PMC_OP_GETCAPS: 4543 { 4544 struct pmc_op_caps c; 4545 struct pmc *pm; 4546 struct pmc_classdep *pcd; 4547 pmc_id_t pmcid; 4548 int adjri, ri; 4549 4550 PMC_DOWNGRADE_SX(); 4551 4552 if ((error = copyin(arg, &c, sizeof(c))) != 0) 4553 break; 4554 4555 pmcid = c.pm_pmcid; 4556 4557 if ((error = pmc_find_pmc(pmcid, &pm)) != 0) 4558 break; 4559 4560 KASSERT(pmcid == pm->pm_id, 4561 ("[pmc,%d] pmc id %x != pmcid %x", __LINE__, 4562 pm->pm_id, pmcid)); 4563 4564 ri = PMC_TO_ROWINDEX(pm); 4565 pcd = pmc_ri_to_classdep(md, ri, &adjri); 4566 4567 /* 4568 * If PMC class has no GETCAPS return the class capabilities 4569 * otherwise get the per counter capabilities. 4570 */ 4571 if (pcd->pcd_get_caps == NULL) { 4572 c.pm_caps = pcd->pcd_caps; 4573 } else { 4574 error = (*pcd->pcd_get_caps)(adjri, &c.pm_caps); 4575 if (error < 0) 4576 break; 4577 } 4578 4579 if ((error = copyout(&c, arg, sizeof(c))) < 0) 4580 break; 4581 } 4582 break; 4583 4584 default: 4585 error = EINVAL; 4586 break; 4587 } 4588 4589 if (is_sx_downgraded) 4590 sx_sunlock(&pmc_sx); 4591 else 4592 sx_xunlock(&pmc_sx); 4593 done_syscall: 4594 if (error) 4595 counter_u64_add(pmc_stats.pm_syscall_errors, 1); 4596 4597 return (error); 4598 } 4599 4600 /* 4601 * Helper functions 4602 */ 4603 4604 /* 4605 * Mark the thread as needing callchain capture and post an AST. The 4606 * actual callchain capture will be done in a context where it is safe 4607 * to take page faults. 4608 */ 4609 static void 4610 pmc_post_callchain_callback(void) 4611 { 4612 struct thread *td; 4613 4614 td = curthread; 4615 4616 /* 4617 * If there is multiple PMCs for the same interrupt ignore new post 4618 */ 4619 if ((td->td_pflags & TDP_CALLCHAIN) != 0) 4620 return; 4621 4622 /* 4623 * Mark this thread as needing callchain capture. 4624 * `td->td_pflags' will be safe to touch because this thread 4625 * was in user space when it was interrupted. 4626 */ 4627 td->td_pflags |= TDP_CALLCHAIN; 4628 4629 /* 4630 * Don't let this thread migrate between CPUs until callchain 4631 * capture completes. 4632 */ 4633 sched_pin(); 4634 4635 return; 4636 } 4637 4638 static void 4639 pmc_multipart_add(struct pmc_sample *ps, int type, int length) 4640 { 4641 int i; 4642 uint8_t *hdr; 4643 4644 MPASS(ps->ps_pc != NULL); 4645 MPASS(ps->ps_nsamples_actual != 0); 4646 4647 hdr = (uint8_t *)ps->ps_pc; 4648 4649 for (i = 0; i < PMC_MULTIPART_HEADER_ENTRIES; i++) { 4650 if (hdr[2 * i] == PMC_CC_MULTIPART_NONE) { 4651 hdr[2 * i] = type; 4652 hdr[2 * i + 1] = length; 4653 ps->ps_nsamples_actual += length; 4654 return; 4655 } 4656 } 4657 4658 KASSERT(false, ("Too many parts in the multipart header!")); 4659 } 4660 4661 static void 4662 pmc_multipart_copydata(struct pmc_sample *ps, struct pmc_multipart *mp) 4663 { 4664 int i, scale; 4665 uint64_t *ps_pc; 4666 4667 MPASS(ps->ps_pc != NULL); 4668 MPASS(ps->ps_nsamples_actual != 0); 4669 4670 ps_pc = (uint64_t *)ps->ps_pc; 4671 4672 for (i = 0; i < mp->pl_length; i++) 4673 ps_pc[i + 1] = mp->pl_mpdata[i]; 4674 4675 scale = sizeof(uint64_t) / sizeof(uintptr_t); 4676 pmc_multipart_add(ps, mp->pl_type, scale * mp->pl_length); 4677 } 4678 4679 /* 4680 * Find a free slot in the per-cpu array of samples and capture the 4681 * current callchain there. If a sample was successfully added, a bit 4682 * is set in mask 'pmc_cpumask' denoting that the DO_SAMPLES hook 4683 * needs to be invoked from the clock handler. 4684 * 4685 * This function is meant to be called from an NMI handler. It cannot 4686 * use any of the locking primitives supplied by the OS. 4687 */ 4688 static int 4689 pmc_add_sample(ring_type_t ring, struct pmc *pm, struct trapframe *tf, 4690 struct pmc_multipart *mp) 4691 { 4692 struct pmc_sample *ps; 4693 struct pmc_samplebuffer *psb; 4694 struct thread *td; 4695 int error, cpu, callchaindepth; 4696 bool inuserspace; 4697 4698 error = 0; 4699 4700 /* 4701 * Allocate space for a sample buffer. 4702 */ 4703 cpu = curcpu; 4704 psb = pmc_pcpu[cpu]->pc_sb[ring]; 4705 inuserspace = TRAPF_USERMODE(tf); 4706 ps = PMC_PROD_SAMPLE(psb); 4707 if (psb->ps_considx != psb->ps_prodidx && 4708 ps->ps_nsamples) { /* in use, reader hasn't caught up */ 4709 pm->pm_pcpu_state[cpu].pps_stalled = 1; 4710 counter_u64_add(pmc_stats.pm_intr_bufferfull, 1); 4711 PMCDBG6(SAM,INT,1,"(spc) cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d", 4712 cpu, pm, tf, inuserspace, 4713 (int)(psb->ps_prodidx & pmc_sample_mask), 4714 (int)(psb->ps_considx & pmc_sample_mask)); 4715 callchaindepth = 1; 4716 error = ENOMEM; 4717 goto done; 4718 } 4719 4720 /* Fill in entry. */ 4721 PMCDBG6(SAM,INT,1,"cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d", cpu, pm, tf, 4722 inuserspace, (int)(psb->ps_prodidx & pmc_sample_mask), 4723 (int)(psb->ps_considx & pmc_sample_mask)); 4724 4725 td = curthread; 4726 ps->ps_pmc = pm; 4727 ps->ps_td = td; 4728 ps->ps_pid = td->td_proc->p_pid; 4729 ps->ps_tid = td->td_tid; 4730 ps->ps_tsc = pmc_rdtsc(); 4731 ps->ps_ticks = ticks; 4732 ps->ps_cpu = cpu; 4733 ps->ps_flags = inuserspace ? PMC_CC_F_USERSPACE : 0; 4734 ps->ps_nsamples_actual = 0; 4735 4736 callchaindepth = (pm->pm_flags & PMC_F_CALLCHAIN) ? 4737 pmc_callchaindepth : 1; 4738 4739 MPASS(ps->ps_pc != NULL); 4740 4741 if (mp != NULL) { 4742 /* Set multipart flag, clear header and copy data */ 4743 ps->ps_flags |= PMC_CC_F_MULTIPART; 4744 ps->ps_pc[0] = 0; 4745 ps->ps_nsamples_actual = 1; 4746 pmc_multipart_copydata(ps, mp); 4747 } 4748 4749 if (callchaindepth == 1) { 4750 ps->ps_pc[ps->ps_nsamples_actual] = PMC_TRAPFRAME_TO_PC(tf); 4751 } else { 4752 /* 4753 * Kernel stack traversals can be done immediately, while we 4754 * defer to an AST for user space traversals. 4755 */ 4756 if (!inuserspace) { 4757 callchaindepth = pmc_save_kernel_callchain( 4758 ps->ps_pc + ps->ps_nsamples_actual, 4759 callchaindepth - ps->ps_nsamples_actual, tf); 4760 callchaindepth += ps->ps_nsamples_actual; 4761 } else { 4762 pmc_post_callchain_callback(); 4763 callchaindepth = PMC_USER_CALLCHAIN_PENDING; 4764 } 4765 } 4766 4767 ps->ps_nsamples = callchaindepth; /* mark entry as in-use */ 4768 if (ring == PMC_UR) { 4769 ps->ps_nsamples_actual = ps->ps_nsamples; 4770 ps->ps_nsamples = PMC_USER_CALLCHAIN_PENDING; 4771 } 4772 4773 KASSERT(counter_u64_fetch(pm->pm_runcount) >= 0, 4774 ("[pmc,%d] pm=%p runcount %ju", __LINE__, pm, 4775 (uintmax_t)counter_u64_fetch(pm->pm_runcount))); 4776 4777 counter_u64_add(pm->pm_runcount, 1); /* hold onto PMC */ 4778 /* increment write pointer */ 4779 psb->ps_prodidx++; 4780 done: 4781 /* mark CPU as needing processing */ 4782 if (callchaindepth != PMC_USER_CALLCHAIN_PENDING) 4783 DPCPU_SET(pmc_sampled, 1); 4784 4785 return (error); 4786 } 4787 4788 /* 4789 * Interrupt processing. 4790 * 4791 * This function may be called from an NMI handler. It cannot use any of the 4792 * locking primitives supplied by the OS. 4793 */ 4794 int 4795 pmc_process_interrupt_mp(int ring, struct pmc *pm, struct trapframe *tf, 4796 struct pmc_multipart *mp) 4797 { 4798 struct thread *td; 4799 4800 td = curthread; 4801 if ((pm->pm_flags & PMC_F_USERCALLCHAIN) && 4802 (td->td_proc->p_flag & P_KPROC) == 0 && !TRAPF_USERMODE(tf)) { 4803 atomic_add_int(&td->td_pmcpend, 1); 4804 return (pmc_add_sample(PMC_UR, pm, tf, mp)); 4805 } 4806 return (pmc_add_sample(ring, pm, tf, mp)); 4807 } 4808 4809 int 4810 pmc_process_interrupt(int ring, struct pmc *pm, struct trapframe *tf) 4811 { 4812 return (pmc_process_interrupt_mp(ring, pm, tf, NULL)); 4813 } 4814 4815 /* 4816 * Capture a user call chain. This function will be called from ast() 4817 * before control returns to userland and before the process gets 4818 * rescheduled. 4819 */ 4820 static void 4821 pmc_capture_user_callchain(int cpu, int ring, struct trapframe *tf) 4822 { 4823 struct pmc *pm; 4824 struct pmc_sample *ps; 4825 struct pmc_samplebuffer *psb; 4826 struct thread *td; 4827 uint64_t considx, prodidx; 4828 int nsamples, nrecords, pass, iter; 4829 int start_ticks __diagused; 4830 4831 psb = pmc_pcpu[cpu]->pc_sb[ring]; 4832 td = curthread; 4833 nrecords = INT_MAX; 4834 pass = 0; 4835 start_ticks = ticks; 4836 4837 KASSERT(td->td_pflags & TDP_CALLCHAIN, 4838 ("[pmc,%d] Retrieving callchain for thread that doesn't want it", 4839 __LINE__)); 4840 restart: 4841 if (ring == PMC_UR) 4842 nrecords = atomic_readandclear_32(&td->td_pmcpend); 4843 4844 for (iter = 0, considx = psb->ps_considx, prodidx = psb->ps_prodidx; 4845 considx < prodidx && iter < pmc_nsamples; considx++, iter++) { 4846 ps = PMC_CONS_SAMPLE_OFF(psb, considx); 4847 4848 /* 4849 * Iterate through all deferred callchain requests. Walk from 4850 * the current read pointer to the current write pointer. 4851 */ 4852 #ifdef INVARIANTS 4853 if (ps->ps_nsamples == PMC_SAMPLE_FREE) { 4854 continue; 4855 } 4856 #endif 4857 if (ps->ps_td != td || 4858 ps->ps_nsamples != PMC_USER_CALLCHAIN_PENDING || 4859 ps->ps_pmc->pm_state != PMC_STATE_RUNNING) 4860 continue; 4861 4862 KASSERT(ps->ps_cpu == cpu, 4863 ("[pmc,%d] cpu mismatch ps_cpu=%d pcpu=%d", __LINE__, 4864 ps->ps_cpu, PCPU_GET(cpuid))); 4865 4866 pm = ps->ps_pmc; 4867 KASSERT(pm->pm_flags & PMC_F_CALLCHAIN, 4868 ("[pmc,%d] Retrieving callchain for PMC that doesn't " 4869 "want it", __LINE__)); 4870 KASSERT(counter_u64_fetch(pm->pm_runcount) > 0, 4871 ("[pmc,%d] runcount %ju", __LINE__, 4872 (uintmax_t)counter_u64_fetch(pm->pm_runcount))); 4873 4874 if (ring == PMC_UR) { 4875 counter_u64_add(pmc_stats.pm_merges, 1); 4876 } 4877 nsamples = ps->ps_nsamples_actual; 4878 4879 /* 4880 * Retrieve the callchain and mark the sample buffer 4881 * as 'processable' by the timer tick sweep code. 4882 */ 4883 if (__predict_true(nsamples < pmc_callchaindepth - 1)) 4884 nsamples += pmc_save_user_callchain(ps->ps_pc + nsamples, 4885 pmc_callchaindepth - nsamples - 1, tf); 4886 4887 /* 4888 * We have to prevent hardclock from potentially overwriting 4889 * this sample between when we read the value and when we set 4890 * it. 4891 */ 4892 spinlock_enter(); 4893 4894 /* 4895 * Verify that the sample hasn't been dropped in the meantime. 4896 */ 4897 if (ps->ps_nsamples == PMC_USER_CALLCHAIN_PENDING) { 4898 ps->ps_nsamples = nsamples; 4899 /* 4900 * If we couldn't get a sample, simply drop the 4901 * reference. 4902 */ 4903 if (nsamples == 0) 4904 counter_u64_add(pm->pm_runcount, -1); 4905 } 4906 spinlock_exit(); 4907 if (nrecords-- == 1) 4908 break; 4909 } 4910 if (__predict_false(ring == PMC_UR && td->td_pmcpend)) { 4911 if (pass == 0) { 4912 pass = 1; 4913 goto restart; 4914 } 4915 /* only collect samples for this part once */ 4916 td->td_pmcpend = 0; 4917 } 4918 4919 #ifdef INVARIANTS 4920 if ((ticks - start_ticks) > hz) 4921 log(LOG_ERR, "%s took %d ticks\n", __func__, (ticks - start_ticks)); 4922 #endif 4923 /* mark CPU as needing processing */ 4924 DPCPU_SET(pmc_sampled, 1); 4925 } 4926 4927 /* 4928 * Process saved PC samples. 4929 */ 4930 static void 4931 pmc_process_samples(int cpu, ring_type_t ring) 4932 { 4933 struct pmc *pm; 4934 struct thread *td; 4935 struct pmc_owner *po; 4936 struct pmc_sample *ps; 4937 struct pmc_classdep *pcd; 4938 struct pmc_samplebuffer *psb; 4939 uint64_t delta __diagused; 4940 int adjri, n; 4941 4942 KASSERT(PCPU_GET(cpuid) == cpu, 4943 ("[pmc,%d] not on the correct CPU pcpu=%d cpu=%d", __LINE__, 4944 PCPU_GET(cpuid), cpu)); 4945 4946 psb = pmc_pcpu[cpu]->pc_sb[ring]; 4947 delta = psb->ps_prodidx - psb->ps_considx; 4948 MPASS(delta <= pmc_nsamples); 4949 MPASS(psb->ps_considx <= psb->ps_prodidx); 4950 for (n = 0; psb->ps_considx < psb->ps_prodidx; psb->ps_considx++, n++) { 4951 ps = PMC_CONS_SAMPLE(psb); 4952 4953 if (__predict_false(ps->ps_nsamples == PMC_SAMPLE_FREE)) 4954 continue; 4955 4956 /* skip non-running samples */ 4957 pm = ps->ps_pmc; 4958 if (pm->pm_state != PMC_STATE_RUNNING) 4959 goto entrydone; 4960 4961 KASSERT(counter_u64_fetch(pm->pm_runcount) > 0, 4962 ("[pmc,%d] pm=%p runcount %ju", __LINE__, pm, 4963 (uintmax_t)counter_u64_fetch(pm->pm_runcount))); 4964 KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)), 4965 ("[pmc,%d] pmc=%p non-sampling mode=%d", __LINE__, 4966 pm, PMC_TO_MODE(pm))); 4967 4968 po = pm->pm_owner; 4969 4970 /* If there is a pending AST wait for completion */ 4971 if (ps->ps_nsamples == PMC_USER_CALLCHAIN_PENDING) { 4972 /* 4973 * If we've been waiting more than 1 tick to 4974 * collect a callchain for this record then 4975 * drop it and move on. 4976 */ 4977 if (ticks - ps->ps_ticks > 1) { 4978 /* 4979 * Track how often we hit this as it will 4980 * preferentially lose user samples 4981 * for long running system calls. 4982 */ 4983 counter_u64_add(pmc_stats.pm_overwrites, 1); 4984 goto entrydone; 4985 } 4986 /* Need a rescan at a later time. */ 4987 DPCPU_SET(pmc_sampled, 1); 4988 break; 4989 } 4990 4991 PMCDBG6(SAM,OPS,1,"cpu=%d pm=%p n=%d fl=%x wr=%d rd=%d", cpu, 4992 pm, ps->ps_nsamples, ps->ps_flags, 4993 (int)(psb->ps_prodidx & pmc_sample_mask), 4994 (int)(psb->ps_considx & pmc_sample_mask)); 4995 4996 /* 4997 * If this is a process-mode PMC that is attached to 4998 * its owner, and if the PC is in user mode, update 4999 * profiling statistics like timer-based profiling 5000 * would have done. 5001 * 5002 * Otherwise, this is either a sampling-mode PMC that 5003 * is attached to a different process than its owner, 5004 * or a system-wide sampling PMC. Dispatch a log 5005 * entry to the PMC's owner process. 5006 */ 5007 if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) { 5008 if (ps->ps_flags & PMC_CC_F_USERSPACE) { 5009 td = FIRST_THREAD_IN_PROC(po->po_owner); 5010 addupc_intr(td, ps->ps_pc[0], 1); 5011 } 5012 } else 5013 pmclog_process_callchain(pm, ps); 5014 5015 entrydone: 5016 ps->ps_nsamples = 0; /* mark entry as free */ 5017 KASSERT(counter_u64_fetch(pm->pm_runcount) > 0, 5018 ("[pmc,%d] pm=%p runcount %ju", __LINE__, pm, 5019 (uintmax_t)counter_u64_fetch(pm->pm_runcount))); 5020 5021 counter_u64_add(pm->pm_runcount, -1); 5022 } 5023 5024 counter_u64_add(pmc_stats.pm_log_sweeps, 1); 5025 5026 /* Do not re-enable stalled PMCs if we failed to process any samples */ 5027 if (n == 0) 5028 return; 5029 5030 /* 5031 * Restart any stalled sampling PMCs on this CPU. 5032 * 5033 * If the NMI handler sets the pm_stalled field of a PMC after 5034 * the check below, we'll end up processing the stalled PMC at 5035 * the next hardclock tick. 5036 */ 5037 for (n = 0; n < md->pmd_npmc; n++) { 5038 pcd = pmc_ri_to_classdep(md, n, &adjri); 5039 KASSERT(pcd != NULL, 5040 ("[pmc,%d] null pcd ri=%d", __LINE__, n)); 5041 (void)(*pcd->pcd_get_config)(cpu, adjri, &pm); 5042 5043 if (pm == NULL || /* !cfg'ed */ 5044 pm->pm_state != PMC_STATE_RUNNING || /* !active */ 5045 !PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) || /* !sampling */ 5046 !pm->pm_pcpu_state[cpu].pps_cpustate || /* !desired */ 5047 !pm->pm_pcpu_state[cpu].pps_stalled) /* !stalled */ 5048 continue; 5049 5050 pm->pm_pcpu_state[cpu].pps_stalled = 0; 5051 (void)(*pcd->pcd_start_pmc)(cpu, adjri, pm); 5052 } 5053 } 5054 5055 /* 5056 * Event handlers. 5057 */ 5058 5059 /* 5060 * Handle a process exit. 5061 * 5062 * Remove this process from all hash tables. If this process 5063 * owned any PMCs, turn off those PMCs and deallocate them, 5064 * removing any associations with target processes. 5065 * 5066 * This function will be called by the last 'thread' of a 5067 * process. 5068 * 5069 * XXX This eventhandler gets called early in the exit process. 5070 * Consider using a 'hook' invocation from thread_exit() or equivalent 5071 * spot. Another negative is that kse_exit doesn't seem to call 5072 * exit1() [??]. 5073 */ 5074 static void 5075 pmc_process_exit(void *arg __unused, struct proc *p) 5076 { 5077 struct pmc *pm; 5078 struct pmc_owner *po; 5079 struct pmc_process *pp; 5080 struct pmc_classdep *pcd; 5081 pmc_value_t newvalue, tmp; 5082 int ri, adjri, cpu; 5083 bool is_using_hwpmcs; 5084 5085 PROC_LOCK(p); 5086 is_using_hwpmcs = (p->p_flag & P_HWPMC) != 0; 5087 PROC_UNLOCK(p); 5088 5089 /* 5090 * Log a sysexit event to all SS PMC owners. 5091 */ 5092 PMC_EPOCH_ENTER(); 5093 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) { 5094 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) != 0) 5095 pmclog_process_sysexit(po, p->p_pid); 5096 } 5097 PMC_EPOCH_EXIT(); 5098 5099 PMC_GET_SX_XLOCK(); 5100 PMCDBG3(PRC,EXT,1,"process-exit proc=%p (%d, %s)", p, p->p_pid, 5101 p->p_comm); 5102 5103 if (!is_using_hwpmcs) 5104 goto out; 5105 5106 /* 5107 * Since this code is invoked by the last thread in an exiting process, 5108 * we would have context switched IN at some prior point. However, with 5109 * PREEMPTION, kernel mode context switches may happen any time, so we 5110 * want to disable a context switch OUT till we get any PMCs targeting 5111 * this process off the hardware. 5112 * 5113 * We also need to atomically remove this process' entry from our 5114 * target process hash table, using PMC_FLAG_REMOVE. 5115 */ 5116 PMCDBG3(PRC,EXT,1, "process-exit proc=%p (%d, %s)", p, p->p_pid, 5117 p->p_comm); 5118 5119 critical_enter(); /* no preemption */ 5120 5121 cpu = curthread->td_oncpu; 5122 5123 pp = pmc_find_process_descriptor(p, PMC_FLAG_REMOVE); 5124 if (pp == NULL) { 5125 critical_exit(); 5126 goto out; 5127 } 5128 5129 PMCDBG2(PRC,EXT,2, "process-exit proc=%p pmc-process=%p", p, pp); 5130 5131 /* 5132 * The exiting process could be the target of some PMCs which will be 5133 * running on currently executing CPU. 5134 * 5135 * We need to turn these PMCs off like we would do at context switch 5136 * OUT time. 5137 */ 5138 for (ri = 0; ri < md->pmd_npmc; ri++) { 5139 /* 5140 * Pick up the pmc pointer from hardware state similar to the 5141 * CSW_OUT code. 5142 */ 5143 pm = NULL; 5144 pcd = pmc_ri_to_classdep(md, ri, &adjri); 5145 5146 (void)(*pcd->pcd_get_config)(cpu, adjri, &pm); 5147 5148 PMCDBG2(PRC,EXT,2, "ri=%d pm=%p", ri, pm); 5149 5150 if (pm == NULL || !PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) 5151 continue; 5152 5153 PMCDBG4(PRC,EXT,2, "ppmcs[%d]=%p pm=%p state=%d", ri, 5154 pp->pp_pmcs[ri].pp_pmc, pm, pm->pm_state); 5155 5156 KASSERT(PMC_TO_ROWINDEX(pm) == ri, 5157 ("[pmc,%d] ri mismatch pmc(%d) ri(%d)", __LINE__, 5158 PMC_TO_ROWINDEX(pm), ri)); 5159 KASSERT(pm == pp->pp_pmcs[ri].pp_pmc, 5160 ("[pmc,%d] pm %p != pp_pmcs[%d] %p", __LINE__, pm, ri, 5161 pp->pp_pmcs[ri].pp_pmc)); 5162 KASSERT(counter_u64_fetch(pm->pm_runcount) > 0, 5163 ("[pmc,%d] bad runcount ri %d rc %ju", __LINE__, ri, 5164 (uintmax_t)counter_u64_fetch(pm->pm_runcount))); 5165 5166 /* 5167 * Change desired state, and then stop if not stalled. This 5168 * two-step dance should avoid race conditions where an 5169 * interrupt re-enables the PMC after this code has already 5170 * checked the pm_stalled flag. 5171 */ 5172 if (pm->pm_pcpu_state[cpu].pps_cpustate) { 5173 pm->pm_pcpu_state[cpu].pps_cpustate = 0; 5174 if (!pm->pm_pcpu_state[cpu].pps_stalled) { 5175 (void)pcd->pcd_stop_pmc(cpu, adjri, pm); 5176 5177 if (PMC_TO_MODE(pm) == PMC_MODE_TC) { 5178 pcd->pcd_read_pmc(cpu, adjri, pm, 5179 &newvalue); 5180 tmp = newvalue - PMC_PCPU_SAVED(cpu, ri); 5181 5182 mtx_pool_lock_spin(pmc_mtxpool, pm); 5183 pm->pm_gv.pm_savedvalue += tmp; 5184 pp->pp_pmcs[ri].pp_pmcval += tmp; 5185 mtx_pool_unlock_spin(pmc_mtxpool, pm); 5186 } 5187 } 5188 } 5189 5190 KASSERT(counter_u64_fetch(pm->pm_runcount) > 0, 5191 ("[pmc,%d] runcount is %d", __LINE__, ri)); 5192 5193 counter_u64_add(pm->pm_runcount, -1); 5194 (void)pcd->pcd_config_pmc(cpu, adjri, NULL); 5195 } 5196 5197 /* 5198 * Inform the MD layer of this pseudo "context switch out". 5199 */ 5200 (void)md->pmd_switch_out(pmc_pcpu[cpu], pp); 5201 5202 critical_exit(); /* ok to be pre-empted now */ 5203 5204 /* 5205 * Unlink this process from the PMCs that are targeting it. This will 5206 * send a signal to all PMC owner's whose PMCs are orphaned. 5207 * 5208 * Log PMC value at exit time if requested. 5209 */ 5210 for (ri = 0; ri < md->pmd_npmc; ri++) { 5211 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) { 5212 if ((pm->pm_flags & PMC_F_NEEDS_LOGFILE) != 0 && 5213 PMC_IS_COUNTING_MODE(PMC_TO_MODE(pm))) { 5214 pmclog_process_procexit(pm, pp); 5215 } 5216 pmc_unlink_target_process(pm, pp); 5217 } 5218 } 5219 free(pp, M_PMC); 5220 5221 out: 5222 /* 5223 * If the process owned PMCs, free them up and free up memory. 5224 */ 5225 if ((po = pmc_find_owner_descriptor(p)) != NULL) { 5226 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) != 0) 5227 pmclog_close(po); 5228 pmc_remove_owner(po); 5229 pmc_destroy_owner_descriptor(po); 5230 } 5231 5232 sx_xunlock(&pmc_sx); 5233 } 5234 5235 /* 5236 * Handle a process fork. 5237 * 5238 * If the parent process 'p1' is under HWPMC monitoring, then copy 5239 * over any attached PMCs that have 'do_descendants' semantics. 5240 */ 5241 static void 5242 pmc_process_fork(void *arg __unused, struct proc *p1, struct proc *newproc, 5243 int flags __unused) 5244 { 5245 struct pmc *pm; 5246 struct pmc_owner *po; 5247 struct pmc_process *ppnew, *ppold; 5248 unsigned int ri; 5249 bool is_using_hwpmcs, do_descendants; 5250 5251 PROC_LOCK(p1); 5252 is_using_hwpmcs = (p1->p_flag & P_HWPMC) != 0; 5253 PROC_UNLOCK(p1); 5254 5255 /* 5256 * If there are system-wide sampling PMCs active, we need to 5257 * log all fork events to their owner's logs. 5258 */ 5259 PMC_EPOCH_ENTER(); 5260 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) { 5261 if (po->po_flags & PMC_PO_OWNS_LOGFILE) { 5262 pmclog_process_procfork(po, p1->p_pid, newproc->p_pid); 5263 pmclog_process_proccreate(po, newproc, 1); 5264 } 5265 } 5266 PMC_EPOCH_EXIT(); 5267 5268 if (!is_using_hwpmcs) 5269 return; 5270 5271 PMC_GET_SX_XLOCK(); 5272 PMCDBG4(PMC,FRK,1, "process-fork proc=%p (%d, %s) -> %p", p1, 5273 p1->p_pid, p1->p_comm, newproc); 5274 5275 /* 5276 * If the parent process (curthread->td_proc) is a 5277 * target of any PMCs, look for PMCs that are to be 5278 * inherited, and link these into the new process 5279 * descriptor. 5280 */ 5281 ppold = pmc_find_process_descriptor(curthread->td_proc, PMC_FLAG_NONE); 5282 if (ppold == NULL) 5283 goto done; /* nothing to do */ 5284 5285 do_descendants = false; 5286 for (ri = 0; ri < md->pmd_npmc; ri++) { 5287 if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL && 5288 (pm->pm_flags & PMC_F_DESCENDANTS) != 0) { 5289 do_descendants = true; 5290 break; 5291 } 5292 } 5293 if (!do_descendants) /* nothing to do */ 5294 goto done; 5295 5296 /* 5297 * Now mark the new process as being tracked by this driver. 5298 */ 5299 PROC_LOCK(newproc); 5300 newproc->p_flag |= P_HWPMC; 5301 PROC_UNLOCK(newproc); 5302 5303 /* Allocate a descriptor for the new process. */ 5304 ppnew = pmc_find_process_descriptor(newproc, PMC_FLAG_ALLOCATE); 5305 if (ppnew == NULL) 5306 goto done; 5307 5308 /* 5309 * Run through all PMCs that were targeting the old process 5310 * and which specified F_DESCENDANTS and attach them to the 5311 * new process. 5312 * 5313 * Log the fork event to all owners of PMCs attached to this 5314 * process, if not already logged. 5315 */ 5316 for (ri = 0; ri < md->pmd_npmc; ri++) { 5317 if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL && 5318 (pm->pm_flags & PMC_F_DESCENDANTS) != 0) { 5319 pmc_link_target_process(pm, ppnew); 5320 po = pm->pm_owner; 5321 if (po->po_sscount == 0 && 5322 (po->po_flags & PMC_PO_OWNS_LOGFILE) != 0) { 5323 pmclog_process_procfork(po, p1->p_pid, 5324 newproc->p_pid); 5325 } 5326 } 5327 } 5328 5329 done: 5330 sx_xunlock(&pmc_sx); 5331 } 5332 5333 static void 5334 pmc_process_threadcreate(struct thread *td) 5335 { 5336 struct pmc_owner *po; 5337 5338 PMC_EPOCH_ENTER(); 5339 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) { 5340 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) != 0) 5341 pmclog_process_threadcreate(po, td, 1); 5342 } 5343 PMC_EPOCH_EXIT(); 5344 } 5345 5346 static void 5347 pmc_process_threadexit(struct thread *td) 5348 { 5349 struct pmc_owner *po; 5350 5351 PMC_EPOCH_ENTER(); 5352 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) { 5353 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) != 0) 5354 pmclog_process_threadexit(po, td); 5355 } 5356 PMC_EPOCH_EXIT(); 5357 } 5358 5359 static void 5360 pmc_process_proccreate(struct proc *p) 5361 { 5362 struct pmc_owner *po; 5363 5364 PMC_EPOCH_ENTER(); 5365 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) { 5366 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) != 0) 5367 pmclog_process_proccreate(po, p, 1 /* sync */); 5368 } 5369 PMC_EPOCH_EXIT(); 5370 } 5371 5372 static void 5373 pmc_process_allproc(struct pmc *pm) 5374 { 5375 struct pmc_owner *po; 5376 struct thread *td; 5377 struct proc *p; 5378 5379 po = pm->pm_owner; 5380 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) 5381 return; 5382 5383 sx_slock(&allproc_lock); 5384 FOREACH_PROC_IN_SYSTEM(p) { 5385 pmclog_process_proccreate(po, p, 0 /* sync */); 5386 PROC_LOCK(p); 5387 FOREACH_THREAD_IN_PROC(p, td) 5388 pmclog_process_threadcreate(po, td, 0 /* sync */); 5389 PROC_UNLOCK(p); 5390 } 5391 sx_sunlock(&allproc_lock); 5392 pmclog_flush(po, 0); 5393 } 5394 5395 static void 5396 pmc_kld_load(void *arg __unused, linker_file_t lf) 5397 { 5398 struct pmc_owner *po; 5399 5400 /* 5401 * Notify owners of system sampling PMCs about KLD operations. 5402 */ 5403 PMC_EPOCH_ENTER(); 5404 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) { 5405 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 5406 pmclog_process_map_in(po, (pid_t) -1, 5407 (uintfptr_t) lf->address, lf->pathname); 5408 } 5409 PMC_EPOCH_EXIT(); 5410 5411 /* 5412 * TODO: Notify owners of (all) process-sampling PMCs too. 5413 */ 5414 } 5415 5416 static void 5417 pmc_kld_unload(void *arg __unused, const char *filename __unused, 5418 caddr_t address, size_t size) 5419 { 5420 struct pmc_owner *po; 5421 5422 PMC_EPOCH_ENTER(); 5423 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) { 5424 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) != 0) { 5425 pmclog_process_map_out(po, (pid_t)-1, 5426 (uintfptr_t)address, (uintfptr_t)address + size); 5427 } 5428 } 5429 PMC_EPOCH_EXIT(); 5430 5431 /* 5432 * TODO: Notify owners of process-sampling PMCs. 5433 */ 5434 } 5435 5436 /* 5437 * initialization 5438 */ 5439 static const char * 5440 pmc_name_of_pmcclass(enum pmc_class class) 5441 { 5442 5443 switch (class) { 5444 #undef __PMC_CLASS 5445 #define __PMC_CLASS(S,V,D) \ 5446 case PMC_CLASS_##S: \ 5447 return #S; 5448 __PMC_CLASSES(); 5449 default: 5450 return ("<unknown>"); 5451 } 5452 } 5453 5454 /* 5455 * Base class initializer: allocate structure and set default classes. 5456 */ 5457 struct pmc_mdep * 5458 pmc_mdep_alloc(int nclasses) 5459 { 5460 struct pmc_mdep *md; 5461 int n; 5462 5463 /* SOFT + md classes */ 5464 n = 1 + nclasses; 5465 md = malloc(sizeof(struct pmc_mdep) + n * sizeof(struct pmc_classdep), 5466 M_PMC, M_WAITOK | M_ZERO); 5467 md->pmd_nclass = n; 5468 5469 /* Default methods */ 5470 md->pmd_switch_in = generic_switch_in; 5471 md->pmd_switch_out = generic_switch_out; 5472 5473 /* Add base class. */ 5474 pmc_soft_initialize(md); 5475 return (md); 5476 } 5477 5478 void 5479 pmc_mdep_free(struct pmc_mdep *md) 5480 { 5481 pmc_soft_finalize(md); 5482 free(md, M_PMC); 5483 } 5484 5485 static int 5486 generic_switch_in(struct pmc_cpu *pc __unused, struct pmc_process *pp __unused) 5487 { 5488 5489 return (0); 5490 } 5491 5492 static int 5493 generic_switch_out(struct pmc_cpu *pc __unused, struct pmc_process *pp __unused) 5494 { 5495 5496 return (0); 5497 } 5498 5499 static struct pmc_mdep * 5500 pmc_generic_cpu_initialize(void) 5501 { 5502 struct pmc_mdep *md; 5503 5504 md = pmc_mdep_alloc(0); 5505 5506 md->pmd_cputype = PMC_CPU_GENERIC; 5507 5508 return (md); 5509 } 5510 5511 static void 5512 pmc_generic_cpu_finalize(struct pmc_mdep *md __unused) 5513 { 5514 5515 } 5516 5517 static int 5518 pmc_initialize(void) 5519 { 5520 struct pcpu *pc; 5521 struct pmc_binding pb; 5522 struct pmc_classdep *pcd; 5523 struct pmc_sample *ps; 5524 struct pmc_samplebuffer *sb; 5525 int c, cpu, error, n, ri; 5526 u_int maxcpu, domain; 5527 5528 md = NULL; 5529 error = 0; 5530 5531 pmc_stats.pm_intr_ignored = counter_u64_alloc(M_WAITOK); 5532 pmc_stats.pm_intr_processed = counter_u64_alloc(M_WAITOK); 5533 pmc_stats.pm_intr_bufferfull = counter_u64_alloc(M_WAITOK); 5534 pmc_stats.pm_syscalls = counter_u64_alloc(M_WAITOK); 5535 pmc_stats.pm_syscall_errors = counter_u64_alloc(M_WAITOK); 5536 pmc_stats.pm_buffer_requests = counter_u64_alloc(M_WAITOK); 5537 pmc_stats.pm_buffer_requests_failed = counter_u64_alloc(M_WAITOK); 5538 pmc_stats.pm_log_sweeps = counter_u64_alloc(M_WAITOK); 5539 pmc_stats.pm_merges = counter_u64_alloc(M_WAITOK); 5540 pmc_stats.pm_overwrites = counter_u64_alloc(M_WAITOK); 5541 5542 #ifdef HWPMC_DEBUG 5543 /* parse debug flags first */ 5544 if (TUNABLE_STR_FETCH(PMC_SYSCTL_NAME_PREFIX "debugflags", 5545 pmc_debugstr, sizeof(pmc_debugstr))) { 5546 pmc_debugflags_parse(pmc_debugstr, pmc_debugstr + 5547 strlen(pmc_debugstr)); 5548 } 5549 #endif 5550 5551 PMCDBG1(MOD,INI,0, "PMC Initialize (version %x)", PMC_VERSION); 5552 5553 /* check kernel version */ 5554 if (pmc_kernel_version != PMC_VERSION) { 5555 if (pmc_kernel_version == 0) 5556 printf("hwpmc: this kernel has not been compiled with " 5557 "'options HWPMC_HOOKS'.\n"); 5558 else 5559 printf("hwpmc: kernel version (0x%x) does not match " 5560 "module version (0x%x).\n", pmc_kernel_version, 5561 PMC_VERSION); 5562 return (EPROGMISMATCH); 5563 } 5564 5565 /* 5566 * check sysctl parameters 5567 */ 5568 if (pmc_hashsize <= 0) { 5569 printf("hwpmc: tunable \"hashsize\"=%d must be " 5570 "greater than zero.\n", pmc_hashsize); 5571 pmc_hashsize = PMC_HASH_SIZE; 5572 } 5573 5574 if (pmc_nsamples <= 0 || pmc_nsamples > 65535) { 5575 printf("hwpmc: tunable \"nsamples\"=%d out of " 5576 "range.\n", pmc_nsamples); 5577 pmc_nsamples = PMC_NSAMPLES; 5578 } 5579 pmc_sample_mask = pmc_nsamples - 1; 5580 5581 if (pmc_callchaindepth <= 0 || 5582 pmc_callchaindepth > PMC_CALLCHAIN_DEPTH_MAX) { 5583 printf("hwpmc: tunable \"callchaindepth\"=%d out of " 5584 "range - using %d.\n", pmc_callchaindepth, 5585 PMC_CALLCHAIN_DEPTH_MAX); 5586 pmc_callchaindepth = PMC_CALLCHAIN_DEPTH_MAX; 5587 } 5588 5589 md = pmc_md_initialize(); 5590 if (md == NULL) { 5591 /* Default to generic CPU. */ 5592 md = pmc_generic_cpu_initialize(); 5593 if (md == NULL) 5594 return (ENOSYS); 5595 } 5596 5597 /* 5598 * Refresh classes base ri. Optional classes may come in different 5599 * order. 5600 */ 5601 for (ri = c = 0; c < md->pmd_nclass; c++) { 5602 pcd = &md->pmd_classdep[c]; 5603 pcd->pcd_ri = ri; 5604 ri += pcd->pcd_num; 5605 } 5606 5607 KASSERT(md->pmd_nclass >= 1 && md->pmd_npmc >= 1, 5608 ("[pmc,%d] no classes or pmcs", __LINE__)); 5609 5610 /* Compute the map from row-indices to classdep pointers. */ 5611 pmc_rowindex_to_classdep = malloc(sizeof(struct pmc_classdep *) * 5612 md->pmd_npmc, M_PMC, M_WAITOK | M_ZERO); 5613 5614 for (n = 0; n < md->pmd_npmc; n++) 5615 pmc_rowindex_to_classdep[n] = NULL; 5616 5617 for (ri = c = 0; c < md->pmd_nclass; c++) { 5618 pcd = &md->pmd_classdep[c]; 5619 for (n = 0; n < pcd->pcd_num; n++, ri++) 5620 pmc_rowindex_to_classdep[ri] = pcd; 5621 } 5622 5623 KASSERT(ri == md->pmd_npmc, 5624 ("[pmc,%d] npmc miscomputed: ri=%d, md->npmc=%d", __LINE__, 5625 ri, md->pmd_npmc)); 5626 5627 maxcpu = pmc_cpu_max(); 5628 5629 /* allocate space for the per-cpu array */ 5630 pmc_pcpu = malloc(maxcpu * sizeof(struct pmc_cpu *), M_PMC, 5631 M_WAITOK | M_ZERO); 5632 5633 /* per-cpu 'saved values' for managing process-mode PMCs */ 5634 pmc_pcpu_saved = malloc(sizeof(pmc_value_t) * maxcpu * md->pmd_npmc, 5635 M_PMC, M_WAITOK); 5636 5637 /* Perform CPU-dependent initialization. */ 5638 pmc_save_cpu_binding(&pb); 5639 error = 0; 5640 for (cpu = 0; error == 0 && cpu < maxcpu; cpu++) { 5641 if (!pmc_cpu_is_active(cpu)) 5642 continue; 5643 pmc_select_cpu(cpu); 5644 pmc_pcpu[cpu] = malloc(sizeof(struct pmc_cpu) + 5645 md->pmd_npmc * sizeof(struct pmc_hw *), M_PMC, 5646 M_WAITOK | M_ZERO); 5647 for (n = 0; error == 0 && n < md->pmd_nclass; n++) 5648 if (md->pmd_classdep[n].pcd_num > 0) 5649 error = md->pmd_classdep[n].pcd_pcpu_init(md, 5650 cpu); 5651 } 5652 pmc_restore_cpu_binding(&pb); 5653 5654 if (error != 0) 5655 return (error); 5656 5657 /* allocate space for the sample array */ 5658 for (cpu = 0; cpu < maxcpu; cpu++) { 5659 if (!pmc_cpu_is_active(cpu)) 5660 continue; 5661 pc = pcpu_find(cpu); 5662 domain = pc->pc_domain; 5663 sb = malloc_domainset(sizeof(struct pmc_samplebuffer) + 5664 pmc_nsamples * sizeof(struct pmc_sample), M_PMC, 5665 DOMAINSET_PREF(domain), M_WAITOK | M_ZERO); 5666 5667 KASSERT(pmc_pcpu[cpu] != NULL, 5668 ("[pmc,%d] cpu=%d Null per-cpu data", __LINE__, cpu)); 5669 5670 sb->ps_callchains = malloc_domainset(pmc_callchaindepth * 5671 pmc_nsamples * sizeof(uintptr_t), M_PMC, 5672 DOMAINSET_PREF(domain), M_WAITOK | M_ZERO); 5673 5674 for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++) 5675 ps->ps_pc = sb->ps_callchains + 5676 (n * pmc_callchaindepth); 5677 5678 pmc_pcpu[cpu]->pc_sb[PMC_HR] = sb; 5679 5680 sb = malloc_domainset(sizeof(struct pmc_samplebuffer) + 5681 pmc_nsamples * sizeof(struct pmc_sample), M_PMC, 5682 DOMAINSET_PREF(domain), M_WAITOK | M_ZERO); 5683 5684 sb->ps_callchains = malloc_domainset(pmc_callchaindepth * 5685 pmc_nsamples * sizeof(uintptr_t), M_PMC, 5686 DOMAINSET_PREF(domain), M_WAITOK | M_ZERO); 5687 for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++) 5688 ps->ps_pc = sb->ps_callchains + 5689 (n * pmc_callchaindepth); 5690 5691 pmc_pcpu[cpu]->pc_sb[PMC_SR] = sb; 5692 5693 sb = malloc_domainset(sizeof(struct pmc_samplebuffer) + 5694 pmc_nsamples * sizeof(struct pmc_sample), M_PMC, 5695 DOMAINSET_PREF(domain), M_WAITOK | M_ZERO); 5696 sb->ps_callchains = malloc_domainset(pmc_callchaindepth * 5697 pmc_nsamples * sizeof(uintptr_t), M_PMC, 5698 DOMAINSET_PREF(domain), M_WAITOK | M_ZERO); 5699 for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++) 5700 ps->ps_pc = sb->ps_callchains + n * pmc_callchaindepth; 5701 5702 pmc_pcpu[cpu]->pc_sb[PMC_UR] = sb; 5703 } 5704 5705 /* allocate space for the row disposition array */ 5706 pmc_pmcdisp = malloc(sizeof(enum pmc_mode) * md->pmd_npmc, 5707 M_PMC, M_WAITOK | M_ZERO); 5708 5709 /* mark all PMCs as available */ 5710 for (n = 0; n < md->pmd_npmc; n++) 5711 PMC_MARK_ROW_FREE(n); 5712 5713 /* allocate thread hash tables */ 5714 pmc_ownerhash = hashinit(pmc_hashsize, M_PMC, 5715 &pmc_ownerhashmask); 5716 5717 pmc_processhash = hashinit(pmc_hashsize, M_PMC, 5718 &pmc_processhashmask); 5719 mtx_init(&pmc_processhash_mtx, "pmc-process-hash", "pmc-leaf", 5720 MTX_SPIN); 5721 5722 CK_LIST_INIT(&pmc_ss_owners); 5723 pmc_ss_count = 0; 5724 5725 /* allocate a pool of spin mutexes */ 5726 pmc_mtxpool = mtx_pool_create("pmc-leaf", pmc_mtxpool_size, 5727 MTX_SPIN); 5728 5729 PMCDBG4(MOD,INI,1, "pmc_ownerhash=%p, mask=0x%lx " 5730 "targethash=%p mask=0x%lx", pmc_ownerhash, pmc_ownerhashmask, 5731 pmc_processhash, pmc_processhashmask); 5732 5733 /* Initialize a spin mutex for the thread free list. */ 5734 mtx_init(&pmc_threadfreelist_mtx, "pmc-threadfreelist", "pmc-leaf", 5735 MTX_SPIN); 5736 5737 /* Initialize the task to prune the thread free list. */ 5738 TASK_INIT(&free_task, 0, pmc_thread_descriptor_pool_free_task, NULL); 5739 5740 /* register process {exit,fork,exec} handlers */ 5741 pmc_exit_tag = EVENTHANDLER_REGISTER(process_exit, 5742 pmc_process_exit, NULL, EVENTHANDLER_PRI_ANY); 5743 pmc_fork_tag = EVENTHANDLER_REGISTER(process_fork, 5744 pmc_process_fork, NULL, EVENTHANDLER_PRI_ANY); 5745 5746 /* register kld event handlers */ 5747 pmc_kld_load_tag = EVENTHANDLER_REGISTER(kld_load, pmc_kld_load, 5748 NULL, EVENTHANDLER_PRI_ANY); 5749 pmc_kld_unload_tag = EVENTHANDLER_REGISTER(kld_unload, pmc_kld_unload, 5750 NULL, EVENTHANDLER_PRI_ANY); 5751 5752 /* initialize logging */ 5753 pmclog_initialize(); 5754 5755 /* set hook functions */ 5756 pmc_intr = md->pmd_intr; 5757 wmb(); 5758 pmc_hook = pmc_hook_handler; 5759 5760 if (error == 0) { 5761 printf(PMC_MODULE_NAME ":"); 5762 for (n = 0; n < md->pmd_nclass; n++) { 5763 if (md->pmd_classdep[n].pcd_num == 0) 5764 continue; 5765 pcd = &md->pmd_classdep[n]; 5766 printf(" %s/%d/%d/0x%b", 5767 pmc_name_of_pmcclass(pcd->pcd_class), 5768 pcd->pcd_num, 5769 pcd->pcd_width, 5770 pcd->pcd_caps, 5771 "\20" 5772 "\1INT\2USR\3SYS\4EDG\5THR" 5773 "\6REA\7WRI\10INV\11QUA\12PRC" 5774 "\13TAG\14CSC"); 5775 } 5776 printf("\n"); 5777 } 5778 5779 return (error); 5780 } 5781 5782 /* prepare to be unloaded */ 5783 static void 5784 pmc_cleanup(void) 5785 { 5786 struct pmc_binding pb; 5787 struct pmc_owner *po, *tmp; 5788 struct pmc_ownerhash *ph; 5789 struct pmc_processhash *prh __pmcdbg_used; 5790 u_int maxcpu; 5791 int cpu, c; 5792 5793 PMCDBG0(MOD,INI,0, "cleanup"); 5794 5795 /* switch off sampling */ 5796 CPU_FOREACH(cpu) 5797 DPCPU_ID_SET(cpu, pmc_sampled, 0); 5798 pmc_intr = NULL; 5799 5800 sx_xlock(&pmc_sx); 5801 if (pmc_hook == NULL) { /* being unloaded already */ 5802 sx_xunlock(&pmc_sx); 5803 return; 5804 } 5805 5806 pmc_hook = NULL; /* prevent new threads from entering module */ 5807 5808 /* deregister event handlers */ 5809 EVENTHANDLER_DEREGISTER(process_fork, pmc_fork_tag); 5810 EVENTHANDLER_DEREGISTER(process_exit, pmc_exit_tag); 5811 EVENTHANDLER_DEREGISTER(kld_load, pmc_kld_load_tag); 5812 EVENTHANDLER_DEREGISTER(kld_unload, pmc_kld_unload_tag); 5813 5814 /* send SIGBUS to all owner threads, free up allocations */ 5815 if (pmc_ownerhash != NULL) { 5816 for (ph = pmc_ownerhash; 5817 ph <= &pmc_ownerhash[pmc_ownerhashmask]; 5818 ph++) { 5819 LIST_FOREACH_SAFE(po, ph, po_next, tmp) { 5820 pmc_remove_owner(po); 5821 5822 PMCDBG3(MOD,INI,2, 5823 "cleanup signal proc=%p (%d, %s)", 5824 po->po_owner, po->po_owner->p_pid, 5825 po->po_owner->p_comm); 5826 5827 PROC_LOCK(po->po_owner); 5828 kern_psignal(po->po_owner, SIGBUS); 5829 PROC_UNLOCK(po->po_owner); 5830 5831 pmc_destroy_owner_descriptor(po); 5832 } 5833 } 5834 } 5835 5836 /* reclaim allocated data structures */ 5837 taskqueue_drain(taskqueue_fast, &free_task); 5838 mtx_destroy(&pmc_threadfreelist_mtx); 5839 pmc_thread_descriptor_pool_drain(); 5840 5841 if (pmc_mtxpool != NULL) 5842 mtx_pool_destroy(&pmc_mtxpool); 5843 5844 mtx_destroy(&pmc_processhash_mtx); 5845 if (pmc_processhash != NULL) { 5846 #ifdef HWPMC_DEBUG 5847 struct pmc_process *pp; 5848 5849 PMCDBG0(MOD,INI,3, "destroy process hash"); 5850 for (prh = pmc_processhash; 5851 prh <= &pmc_processhash[pmc_processhashmask]; 5852 prh++) 5853 LIST_FOREACH(pp, prh, pp_next) 5854 PMCDBG1(MOD,INI,3, "pid=%d", pp->pp_proc->p_pid); 5855 #endif 5856 5857 hashdestroy(pmc_processhash, M_PMC, pmc_processhashmask); 5858 pmc_processhash = NULL; 5859 } 5860 5861 if (pmc_ownerhash != NULL) { 5862 PMCDBG0(MOD,INI,3, "destroy owner hash"); 5863 hashdestroy(pmc_ownerhash, M_PMC, pmc_ownerhashmask); 5864 pmc_ownerhash = NULL; 5865 } 5866 5867 KASSERT(CK_LIST_EMPTY(&pmc_ss_owners), 5868 ("[pmc,%d] Global SS owner list not empty", __LINE__)); 5869 KASSERT(pmc_ss_count == 0, 5870 ("[pmc,%d] Global SS count not empty", __LINE__)); 5871 5872 /* do processor and pmc-class dependent cleanup */ 5873 maxcpu = pmc_cpu_max(); 5874 5875 PMCDBG0(MOD,INI,3, "md cleanup"); 5876 if (md) { 5877 pmc_save_cpu_binding(&pb); 5878 for (cpu = 0; cpu < maxcpu; cpu++) { 5879 PMCDBG2(MOD,INI,1,"pmc-cleanup cpu=%d pcs=%p", 5880 cpu, pmc_pcpu[cpu]); 5881 if (!pmc_cpu_is_active(cpu) || pmc_pcpu[cpu] == NULL) 5882 continue; 5883 5884 pmc_select_cpu(cpu); 5885 for (c = 0; c < md->pmd_nclass; c++) { 5886 if (md->pmd_classdep[c].pcd_num > 0) { 5887 md->pmd_classdep[c].pcd_pcpu_fini(md, 5888 cpu); 5889 } 5890 } 5891 } 5892 5893 if (md->pmd_cputype == PMC_CPU_GENERIC) 5894 pmc_generic_cpu_finalize(md); 5895 else 5896 pmc_md_finalize(md); 5897 5898 pmc_mdep_free(md); 5899 md = NULL; 5900 pmc_restore_cpu_binding(&pb); 5901 } 5902 5903 /* Free per-cpu descriptors. */ 5904 for (cpu = 0; cpu < maxcpu; cpu++) { 5905 if (!pmc_cpu_is_active(cpu)) 5906 continue; 5907 KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_HR] != NULL, 5908 ("[pmc,%d] Null hw cpu sample buffer cpu=%d", __LINE__, 5909 cpu)); 5910 KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_SR] != NULL, 5911 ("[pmc,%d] Null sw cpu sample buffer cpu=%d", __LINE__, 5912 cpu)); 5913 KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_UR] != NULL, 5914 ("[pmc,%d] Null userret cpu sample buffer cpu=%d", __LINE__, 5915 cpu)); 5916 free(pmc_pcpu[cpu]->pc_sb[PMC_HR]->ps_callchains, M_PMC); 5917 free(pmc_pcpu[cpu]->pc_sb[PMC_HR], M_PMC); 5918 free(pmc_pcpu[cpu]->pc_sb[PMC_SR]->ps_callchains, M_PMC); 5919 free(pmc_pcpu[cpu]->pc_sb[PMC_SR], M_PMC); 5920 free(pmc_pcpu[cpu]->pc_sb[PMC_UR]->ps_callchains, M_PMC); 5921 free(pmc_pcpu[cpu]->pc_sb[PMC_UR], M_PMC); 5922 free(pmc_pcpu[cpu], M_PMC); 5923 } 5924 5925 free(pmc_pcpu, M_PMC); 5926 pmc_pcpu = NULL; 5927 5928 free(pmc_pcpu_saved, M_PMC); 5929 pmc_pcpu_saved = NULL; 5930 5931 if (pmc_pmcdisp != NULL) { 5932 free(pmc_pmcdisp, M_PMC); 5933 pmc_pmcdisp = NULL; 5934 } 5935 5936 if (pmc_rowindex_to_classdep != NULL) { 5937 free(pmc_rowindex_to_classdep, M_PMC); 5938 pmc_rowindex_to_classdep = NULL; 5939 } 5940 5941 pmclog_shutdown(); 5942 counter_u64_free(pmc_stats.pm_intr_ignored); 5943 counter_u64_free(pmc_stats.pm_intr_processed); 5944 counter_u64_free(pmc_stats.pm_intr_bufferfull); 5945 counter_u64_free(pmc_stats.pm_syscalls); 5946 counter_u64_free(pmc_stats.pm_syscall_errors); 5947 counter_u64_free(pmc_stats.pm_buffer_requests); 5948 counter_u64_free(pmc_stats.pm_buffer_requests_failed); 5949 counter_u64_free(pmc_stats.pm_log_sweeps); 5950 counter_u64_free(pmc_stats.pm_merges); 5951 counter_u64_free(pmc_stats.pm_overwrites); 5952 sx_xunlock(&pmc_sx); /* we are done */ 5953 } 5954 5955 /* 5956 * The function called at load/unload. 5957 */ 5958 static int 5959 load(struct module *module __unused, int cmd, void *arg __unused) 5960 { 5961 int error; 5962 5963 error = 0; 5964 5965 switch (cmd) { 5966 case MOD_LOAD: 5967 /* initialize the subsystem */ 5968 error = pmc_initialize(); 5969 if (error != 0) 5970 break; 5971 PMCDBG2(MOD,INI,1, "syscall=%d maxcpu=%d", pmc_syscall_num, 5972 pmc_cpu_max()); 5973 break; 5974 case MOD_UNLOAD: 5975 case MOD_SHUTDOWN: 5976 pmc_cleanup(); 5977 PMCDBG0(MOD,INI,1, "unloaded"); 5978 break; 5979 default: 5980 error = EINVAL; 5981 break; 5982 } 5983 5984 return (error); 5985 } 5986