hwpmc_mod.c revision 156778
1/*- 2 * Copyright (c) 2003-2005 Joseph Koshy 3 * All rights reserved. 4 * 5 * Redistribution and use in source and binary forms, with or without 6 * modification, are permitted provided that the following conditions 7 * are met: 8 * 1. Redistributions of source code must retain the above copyright 9 * notice, this list of conditions and the following disclaimer. 10 * 2. Redistributions in binary form must reproduce the above copyright 11 * notice, this list of conditions and the following disclaimer in the 12 * documentation and/or other materials provided with the distribution. 13 * 14 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 15 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 16 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 17 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 18 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 19 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 20 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 21 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 22 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 23 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 24 * SUCH DAMAGE. 25 * 26 */ 27 28#include <sys/cdefs.h> 29__FBSDID("$FreeBSD: head/sys/dev/hwpmc/hwpmc_mod.c 156778 2006-03-16 16:32:56Z jkoshy $"); 30 31#include <sys/param.h> 32#include <sys/eventhandler.h> 33#include <sys/jail.h> 34#include <sys/kernel.h> 35#include <sys/kthread.h> 36#include <sys/limits.h> 37#include <sys/lock.h> 38#include <sys/malloc.h> 39#include <sys/module.h> 40#include <sys/mutex.h> 41#include <sys/pmc.h> 42#include <sys/pmckern.h> 43#include <sys/pmclog.h> 44#include <sys/proc.h> 45#include <sys/queue.h> 46#include <sys/resourcevar.h> 47#include <sys/sched.h> 48#include <sys/signalvar.h> 49#include <sys/smp.h> 50#include <sys/sx.h> 51#include <sys/sysctl.h> 52#include <sys/sysent.h> 53#include <sys/systm.h> 54#include <sys/vnode.h> 55 56#include <machine/atomic.h> 57#include <machine/md_var.h> 58 59/* 60 * Types 61 */ 62 63enum pmc_flags { 64 PMC_FLAG_NONE = 0x00, /* do nothing */ 65 PMC_FLAG_REMOVE = 0x01, /* atomically remove entry from hash */ 66 PMC_FLAG_ALLOCATE = 0x02, /* add entry to hash if not found */ 67}; 68 69/* 70 * The offset in sysent where the syscall is allocated. 71 */ 72 73static int pmc_syscall_num = NO_SYSCALL; 74struct pmc_cpu **pmc_pcpu; /* per-cpu state */ 75pmc_value_t *pmc_pcpu_saved; /* saved PMC values: CSW handling */ 76 77#define PMC_PCPU_SAVED(C,R) pmc_pcpu_saved[(R) + md->pmd_npmc*(C)] 78 79struct mtx_pool *pmc_mtxpool; 80static int *pmc_pmcdisp; /* PMC row dispositions */ 81 82#define PMC_ROW_DISP_IS_FREE(R) (pmc_pmcdisp[(R)] == 0) 83#define PMC_ROW_DISP_IS_THREAD(R) (pmc_pmcdisp[(R)] > 0) 84#define PMC_ROW_DISP_IS_STANDALONE(R) (pmc_pmcdisp[(R)] < 0) 85 86#define PMC_MARK_ROW_FREE(R) do { \ 87 pmc_pmcdisp[(R)] = 0; \ 88} while (0) 89 90#define PMC_MARK_ROW_STANDALONE(R) do { \ 91 KASSERT(pmc_pmcdisp[(R)] <= 0, ("[pmc,%d] row disposition error", \ 92 __LINE__)); \ 93 atomic_add_int(&pmc_pmcdisp[(R)], -1); \ 94 KASSERT(pmc_pmcdisp[(R)] >= (-mp_ncpus), ("[pmc,%d] row " \ 95 "disposition error", __LINE__)); \ 96} while (0) 97 98#define PMC_UNMARK_ROW_STANDALONE(R) do { \ 99 atomic_add_int(&pmc_pmcdisp[(R)], 1); \ 100 KASSERT(pmc_pmcdisp[(R)] <= 0, ("[pmc,%d] row disposition error", \ 101 __LINE__)); \ 102} while (0) 103 104#define PMC_MARK_ROW_THREAD(R) do { \ 105 KASSERT(pmc_pmcdisp[(R)] >= 0, ("[pmc,%d] row disposition error", \ 106 __LINE__)); \ 107 atomic_add_int(&pmc_pmcdisp[(R)], 1); \ 108} while (0) 109 110#define PMC_UNMARK_ROW_THREAD(R) do { \ 111 atomic_add_int(&pmc_pmcdisp[(R)], -1); \ 112 KASSERT(pmc_pmcdisp[(R)] >= 0, ("[pmc,%d] row disposition error", \ 113 __LINE__)); \ 114} while (0) 115 116 117/* various event handlers */ 118static eventhandler_tag pmc_exit_tag, pmc_fork_tag; 119 120/* Module statistics */ 121struct pmc_op_getdriverstats pmc_stats; 122 123/* Machine/processor dependent operations */ 124struct pmc_mdep *md; 125 126/* 127 * Hash tables mapping owner processes and target threads to PMCs. 128 */ 129 130struct mtx pmc_processhash_mtx; /* spin mutex */ 131static u_long pmc_processhashmask; 132static LIST_HEAD(pmc_processhash, pmc_process) *pmc_processhash; 133 134/* 135 * Hash table of PMC owner descriptors. This table is protected by 136 * the shared PMC "sx" lock. 137 */ 138 139static u_long pmc_ownerhashmask; 140static LIST_HEAD(pmc_ownerhash, pmc_owner) *pmc_ownerhash; 141 142/* 143 * List of PMC owners with system-wide sampling PMCs. 144 */ 145 146static LIST_HEAD(, pmc_owner) pmc_ss_owners; 147 148 149/* 150 * Prototypes 151 */ 152 153#ifdef DEBUG 154static int pmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS); 155static int pmc_debugflags_parse(char *newstr, char *fence); 156#endif 157 158static int load(struct module *module, int cmd, void *arg); 159static int pmc_attach_process(struct proc *p, struct pmc *pm); 160static struct pmc *pmc_allocate_pmc_descriptor(void); 161static struct pmc_owner *pmc_allocate_owner_descriptor(struct proc *p); 162static int pmc_attach_one_process(struct proc *p, struct pmc *pm); 163static int pmc_can_allocate_rowindex(struct proc *p, unsigned int ri, 164 int cpu); 165static int pmc_can_attach(struct pmc *pm, struct proc *p); 166static void pmc_cleanup(void); 167static int pmc_detach_process(struct proc *p, struct pmc *pm); 168static int pmc_detach_one_process(struct proc *p, struct pmc *pm, 169 int flags); 170static void pmc_destroy_owner_descriptor(struct pmc_owner *po); 171static struct pmc_owner *pmc_find_owner_descriptor(struct proc *p); 172static int pmc_find_pmc(pmc_id_t pmcid, struct pmc **pm); 173static struct pmc *pmc_find_pmc_descriptor_in_process(struct pmc_owner *po, 174 pmc_id_t pmc); 175static struct pmc_process *pmc_find_process_descriptor(struct proc *p, 176 uint32_t mode); 177static void pmc_force_context_switch(void); 178static void pmc_link_target_process(struct pmc *pm, 179 struct pmc_process *pp); 180static void pmc_maybe_remove_owner(struct pmc_owner *po); 181static void pmc_process_csw_in(struct thread *td); 182static void pmc_process_csw_out(struct thread *td); 183static void pmc_process_exit(void *arg, struct proc *p); 184static void pmc_process_fork(void *arg, struct proc *p1, 185 struct proc *p2, int n); 186static void pmc_process_samples(int cpu); 187static void pmc_release_pmc_descriptor(struct pmc *pmc); 188static void pmc_remove_owner(struct pmc_owner *po); 189static void pmc_remove_process_descriptor(struct pmc_process *pp); 190static void pmc_restore_cpu_binding(struct pmc_binding *pb); 191static void pmc_save_cpu_binding(struct pmc_binding *pb); 192static void pmc_select_cpu(int cpu); 193static int pmc_start(struct pmc *pm); 194static int pmc_stop(struct pmc *pm); 195static int pmc_syscall_handler(struct thread *td, void *syscall_args); 196static void pmc_unlink_target_process(struct pmc *pmc, 197 struct pmc_process *pp); 198 199/* 200 * Kernel tunables and sysctl(8) interface. 201 */ 202 203SYSCTL_NODE(_kern, OID_AUTO, hwpmc, CTLFLAG_RW, 0, "HWPMC parameters"); 204 205#ifdef DEBUG 206struct pmc_debugflags pmc_debugflags = PMC_DEBUG_DEFAULT_FLAGS; 207char pmc_debugstr[PMC_DEBUG_STRSIZE]; 208TUNABLE_STR(PMC_SYSCTL_NAME_PREFIX "debugflags", pmc_debugstr, 209 sizeof(pmc_debugstr)); 210SYSCTL_PROC(_kern_hwpmc, OID_AUTO, debugflags, 211 CTLTYPE_STRING|CTLFLAG_RW|CTLFLAG_TUN, 212 0, 0, pmc_debugflags_sysctl_handler, "A", "debug flags"); 213#endif 214 215/* 216 * kern.hwpmc.hashrows -- determines the number of rows in the 217 * of the hash table used to look up threads 218 */ 219 220static int pmc_hashsize = PMC_HASH_SIZE; 221TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "hashsize", &pmc_hashsize); 222SYSCTL_INT(_kern_hwpmc, OID_AUTO, hashsize, CTLFLAG_TUN|CTLFLAG_RD, 223 &pmc_hashsize, 0, "rows in hash tables"); 224 225/* 226 * kern.hwpmc.nsamples --- number of PC samples per CPU 227 */ 228 229static int pmc_nsamples = PMC_NSAMPLES; 230TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "nsamples", &pmc_nsamples); 231SYSCTL_INT(_kern_hwpmc, OID_AUTO, nsamples, CTLFLAG_TUN|CTLFLAG_RD, 232 &pmc_nsamples, 0, "number of PC samples per CPU"); 233 234/* 235 * kern.hwpmc.mtxpoolsize -- number of mutexes in the mutex pool. 236 */ 237 238static int pmc_mtxpool_size = PMC_MTXPOOL_SIZE; 239TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "mtxpoolsize", &pmc_mtxpool_size); 240SYSCTL_INT(_kern_hwpmc, OID_AUTO, mtxpoolsize, CTLFLAG_TUN|CTLFLAG_RD, 241 &pmc_mtxpool_size, 0, "size of spin mutex pool"); 242 243 244/* 245 * security.bsd.unprivileged_syspmcs -- allow non-root processes to 246 * allocate system-wide PMCs. 247 * 248 * Allowing unprivileged processes to allocate system PMCs is convenient 249 * if system-wide measurements need to be taken concurrently with other 250 * per-process measurements. This feature is turned off by default. 251 */ 252 253SYSCTL_DECL(_security_bsd); 254 255static int pmc_unprivileged_syspmcs = 0; 256TUNABLE_INT("security.bsd.unprivileged_syspmcs", &pmc_unprivileged_syspmcs); 257SYSCTL_INT(_security_bsd, OID_AUTO, unprivileged_syspmcs, CTLFLAG_RW, 258 &pmc_unprivileged_syspmcs, 0, 259 "allow unprivileged process to allocate system PMCs"); 260 261/* 262 * Hash function. Discard the lower 2 bits of the pointer since 263 * these are always zero for our uses. The hash multiplier is 264 * round((2^LONG_BIT) * ((sqrt(5)-1)/2)). 265 */ 266 267#if LONG_BIT == 64 268#define _PMC_HM 11400714819323198486u 269#elif LONG_BIT == 32 270#define _PMC_HM 2654435769u 271#else 272#error Must know the size of 'long' to compile 273#endif 274 275#define PMC_HASH_PTR(P,M) ((((unsigned long) (P) >> 2) * _PMC_HM) & (M)) 276 277/* 278 * Syscall structures 279 */ 280 281/* The `sysent' for the new syscall */ 282static struct sysent pmc_sysent = { 283 2, /* sy_narg */ 284 pmc_syscall_handler /* sy_call */ 285}; 286 287static struct syscall_module_data pmc_syscall_mod = { 288 load, 289 NULL, 290 &pmc_syscall_num, 291 &pmc_sysent, 292 { 0, NULL } 293}; 294 295static moduledata_t pmc_mod = { 296 PMC_MODULE_NAME, 297 syscall_module_handler, 298 &pmc_syscall_mod 299}; 300 301DECLARE_MODULE(pmc, pmc_mod, SI_SUB_SMP, SI_ORDER_ANY); 302MODULE_VERSION(pmc, PMC_VERSION); 303 304#ifdef DEBUG 305enum pmc_dbgparse_state { 306 PMCDS_WS, /* in whitespace */ 307 PMCDS_MAJOR, /* seen a major keyword */ 308 PMCDS_MINOR 309}; 310 311static int 312pmc_debugflags_parse(char *newstr, char *fence) 313{ 314 char c, *p, *q; 315 struct pmc_debugflags *tmpflags; 316 int error, found, *newbits, tmp; 317 size_t kwlen; 318 319 MALLOC(tmpflags, struct pmc_debugflags *, sizeof(*tmpflags), 320 M_PMC, M_WAITOK|M_ZERO); 321 322 p = newstr; 323 error = 0; 324 325 for (; p < fence && (c = *p); p++) { 326 327 /* skip white space */ 328 if (c == ' ' || c == '\t') 329 continue; 330 331 /* look for a keyword followed by "=" */ 332 for (q = p; p < fence && (c = *p) && c != '='; p++) 333 ; 334 if (c != '=') { 335 error = EINVAL; 336 goto done; 337 } 338 339 kwlen = p - q; 340 newbits = NULL; 341 342 /* lookup flag group name */ 343#define DBG_SET_FLAG_MAJ(S,F) \ 344 if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0) \ 345 newbits = &tmpflags->pdb_ ## F; 346 347 DBG_SET_FLAG_MAJ("cpu", CPU); 348 DBG_SET_FLAG_MAJ("csw", CSW); 349 DBG_SET_FLAG_MAJ("logging", LOG); 350 DBG_SET_FLAG_MAJ("module", MOD); 351 DBG_SET_FLAG_MAJ("md", MDP); 352 DBG_SET_FLAG_MAJ("owner", OWN); 353 DBG_SET_FLAG_MAJ("pmc", PMC); 354 DBG_SET_FLAG_MAJ("process", PRC); 355 DBG_SET_FLAG_MAJ("sampling", SAM); 356 357 if (newbits == NULL) { 358 error = EINVAL; 359 goto done; 360 } 361 362 p++; /* skip the '=' */ 363 364 /* Now parse the individual flags */ 365 tmp = 0; 366 newflag: 367 for (q = p; p < fence && (c = *p); p++) 368 if (c == ' ' || c == '\t' || c == ',') 369 break; 370 371 /* p == fence or c == ws or c == "," or c == 0 */ 372 373 if ((kwlen = p - q) == 0) { 374 *newbits = tmp; 375 continue; 376 } 377 378 found = 0; 379#define DBG_SET_FLAG_MIN(S,F) \ 380 if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0) \ 381 tmp |= found = (1 << PMC_DEBUG_MIN_ ## F) 382 383 /* a '*' denotes all possible flags in the group */ 384 if (kwlen == 1 && *q == '*') 385 tmp = found = ~0; 386 /* look for individual flag names */ 387 DBG_SET_FLAG_MIN("allocaterow", ALR); 388 DBG_SET_FLAG_MIN("allocate", ALL); 389 DBG_SET_FLAG_MIN("attach", ATT); 390 DBG_SET_FLAG_MIN("bind", BND); 391 DBG_SET_FLAG_MIN("config", CFG); 392 DBG_SET_FLAG_MIN("exec", EXC); 393 DBG_SET_FLAG_MIN("exit", EXT); 394 DBG_SET_FLAG_MIN("find", FND); 395 DBG_SET_FLAG_MIN("flush", FLS); 396 DBG_SET_FLAG_MIN("fork", FRK); 397 DBG_SET_FLAG_MIN("getbuf", GTB); 398 DBG_SET_FLAG_MIN("hook", PMH); 399 DBG_SET_FLAG_MIN("init", INI); 400 DBG_SET_FLAG_MIN("intr", INT); 401 DBG_SET_FLAG_MIN("linktarget", TLK); 402 DBG_SET_FLAG_MIN("mayberemove", OMR); 403 DBG_SET_FLAG_MIN("ops", OPS); 404 DBG_SET_FLAG_MIN("read", REA); 405 DBG_SET_FLAG_MIN("register", REG); 406 DBG_SET_FLAG_MIN("release", REL); 407 DBG_SET_FLAG_MIN("remove", ORM); 408 DBG_SET_FLAG_MIN("sample", SAM); 409 DBG_SET_FLAG_MIN("scheduleio", SIO); 410 DBG_SET_FLAG_MIN("select", SEL); 411 DBG_SET_FLAG_MIN("signal", SIG); 412 DBG_SET_FLAG_MIN("swi", SWI); 413 DBG_SET_FLAG_MIN("swo", SWO); 414 DBG_SET_FLAG_MIN("start", STA); 415 DBG_SET_FLAG_MIN("stop", STO); 416 DBG_SET_FLAG_MIN("syscall", PMS); 417 DBG_SET_FLAG_MIN("unlinktarget", TUL); 418 DBG_SET_FLAG_MIN("write", WRI); 419 if (found == 0) { 420 /* unrecognized flag name */ 421 error = EINVAL; 422 goto done; 423 } 424 425 if (c == 0 || c == ' ' || c == '\t') { /* end of flag group */ 426 *newbits = tmp; 427 continue; 428 } 429 430 p++; 431 goto newflag; 432 } 433 434 /* save the new flag set */ 435 bcopy(tmpflags, &pmc_debugflags, sizeof(pmc_debugflags)); 436 437 done: 438 FREE(tmpflags, M_PMC); 439 return error; 440} 441 442static int 443pmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS) 444{ 445 char *fence, *newstr; 446 int error; 447 unsigned int n; 448 449 (void) arg1; (void) arg2; /* unused parameters */ 450 451 n = sizeof(pmc_debugstr); 452 MALLOC(newstr, char *, n, M_PMC, M_ZERO|M_WAITOK); 453 (void) strlcpy(newstr, pmc_debugstr, n); 454 455 error = sysctl_handle_string(oidp, newstr, n, req); 456 457 /* if there is a new string, parse and copy it */ 458 if (error == 0 && req->newptr != NULL) { 459 fence = newstr + (n < req->newlen ? n : req->newlen + 1); 460 if ((error = pmc_debugflags_parse(newstr, fence)) == 0) 461 (void) strlcpy(pmc_debugstr, newstr, 462 sizeof(pmc_debugstr)); 463 } 464 465 FREE(newstr, M_PMC); 466 467 return error; 468} 469#endif 470 471/* 472 * Concurrency Control 473 * 474 * The driver manages the following data structures: 475 * 476 * - target process descriptors, one per target process 477 * - owner process descriptors (and attached lists), one per owner process 478 * - lookup hash tables for owner and target processes 479 * - PMC descriptors (and attached lists) 480 * - per-cpu hardware state 481 * - the 'hook' variable through which the kernel calls into 482 * this module 483 * - the machine hardware state (managed by the MD layer) 484 * 485 * These data structures are accessed from: 486 * 487 * - thread context-switch code 488 * - interrupt handlers (possibly on multiple cpus) 489 * - kernel threads on multiple cpus running on behalf of user 490 * processes doing system calls 491 * - this driver's private kernel threads 492 * 493 * = Locks and Locking strategy = 494 * 495 * The driver uses four locking strategies for its operation: 496 * 497 * - There is a 'global' SX lock "pmc_sx" that is used to protect 498 * the its 'meta-data'. 499 * 500 * Calls into the module (via syscall() or by the kernel) start with 501 * this lock being held in exclusive mode. Depending on the requested 502 * operation, the lock may be downgraded to 'shared' mode to allow 503 * more concurrent readers into the module. 504 * 505 * This SX lock is held in exclusive mode for any operations that 506 * modify the linkages between the driver's internal data structures. 507 * 508 * The 'pmc_hook' function pointer is also protected by this lock. 509 * It is only examined with the sx lock held in exclusive mode. The 510 * kernel module is allowed to be unloaded only with the sx lock 511 * held in exclusive mode. In normal syscall handling, after 512 * acquiring the pmc_sx lock we first check that 'pmc_hook' is 513 * non-null before proceeding. This prevents races between the 514 * thread unloading the module and other threads seeking to use the 515 * module. 516 * 517 * - Lookups of target process structures and owner process structures 518 * cannot use the global "pmc_sx" SX lock because these lookups need 519 * to happen during context switches and in other critical sections 520 * where sleeping is not allowed. We protect these lookup tables 521 * with their own private spin-mutexes, "pmc_processhash_mtx" and 522 * "pmc_ownerhash_mtx". These are 'leaf' mutexes, in that no other 523 * lock is acquired with these locks held. 524 * 525 * - Interrupt handlers work in a lock free manner. At interrupt 526 * time, handlers look at the PMC pointer (phw->phw_pmc) configured 527 * when the PMC was started. If this pointer is NULL, the interrupt 528 * is ignored after updating driver statistics. We ensure that this 529 * pointer is set (using an atomic operation if necessary) before the 530 * PMC hardware is started. Conversely, this pointer is unset atomically 531 * only after the PMC hardware is stopped. 532 * 533 * We ensure that everything needed for the operation of an 534 * interrupt handler is available without it needing to acquire any 535 * locks. We also ensure that a PMC's software state is destroyed only 536 * after the PMC is taken off hardware (on all CPUs). 537 * 538 * - Context-switch handling with process-private PMCs needs more 539 * care. 540 * 541 * A given process may be the target of multiple PMCs. For example, 542 * PMCATTACH and PMCDETACH may be requested by a process on one CPU 543 * while the target process is running on another. A PMC could also 544 * be getting released because its owner is exiting. We tackle 545 * these situations in the following manner: 546 * 547 * - each target process structure 'pmc_process' has an array 548 * of 'struct pmc *' pointers, one for each hardware PMC. 549 * 550 * - At context switch IN time, each "target" PMC in RUNNING state 551 * gets started on hardware and a pointer to each PMC is copied into 552 * the per-cpu phw array. The 'runcount' for the PMC is 553 * incremented. 554 * 555 * - At context switch OUT time, all process-virtual PMCs are stopped 556 * on hardware. The saved value is added to the PMCs value field 557 * only if the PMC is in a non-deleted state (the PMCs state could 558 * have changed during the current time slice). 559 * 560 * Note that since in-between a switch IN on a processor and a switch 561 * OUT, the PMC could have been released on another CPU. Therefore 562 * context switch OUT always looks at the hardware state to turn 563 * OFF PMCs and will update a PMC's saved value only if reachable 564 * from the target process record. 565 * 566 * - OP PMCRELEASE could be called on a PMC at any time (the PMC could 567 * be attached to many processes at the time of the call and could 568 * be active on multiple CPUs). 569 * 570 * We prevent further scheduling of the PMC by marking it as in 571 * state 'DELETED'. If the runcount of the PMC is non-zero then 572 * this PMC is currently running on a CPU somewhere. The thread 573 * doing the PMCRELEASE operation waits by repeatedly doing an 574 * tsleep() till the runcount comes to zero. 575 * 576 */ 577 578/* 579 * save the cpu binding of the current kthread 580 */ 581 582static void 583pmc_save_cpu_binding(struct pmc_binding *pb) 584{ 585 PMCDBG(CPU,BND,2, "%s", "save-cpu"); 586 mtx_lock_spin(&sched_lock); 587 pb->pb_bound = sched_is_bound(curthread); 588 pb->pb_cpu = curthread->td_oncpu; 589 mtx_unlock_spin(&sched_lock); 590 PMCDBG(CPU,BND,2, "save-cpu cpu=%d", pb->pb_cpu); 591} 592 593/* 594 * restore the cpu binding of the current thread 595 */ 596 597static void 598pmc_restore_cpu_binding(struct pmc_binding *pb) 599{ 600 PMCDBG(CPU,BND,2, "restore-cpu curcpu=%d restore=%d", 601 curthread->td_oncpu, pb->pb_cpu); 602 mtx_lock_spin(&sched_lock); 603 if (pb->pb_bound) 604 sched_bind(curthread, pb->pb_cpu); 605 else 606 sched_unbind(curthread); 607 mtx_unlock_spin(&sched_lock); 608 PMCDBG(CPU,BND,2, "%s", "restore-cpu done"); 609} 610 611/* 612 * move execution over the specified cpu and bind it there. 613 */ 614 615static void 616pmc_select_cpu(int cpu) 617{ 618 KASSERT(cpu >= 0 && cpu < mp_ncpus, 619 ("[pmc,%d] bad cpu number %d", __LINE__, cpu)); 620 621 /* never move to a disabled CPU */ 622 KASSERT(pmc_cpu_is_disabled(cpu) == 0, ("[pmc,%d] selecting " 623 "disabled CPU %d", __LINE__, cpu)); 624 625 PMCDBG(CPU,SEL,2, "select-cpu cpu=%d", cpu); 626 mtx_lock_spin(&sched_lock); 627 sched_bind(curthread, cpu); 628 mtx_unlock_spin(&sched_lock); 629 630 KASSERT(curthread->td_oncpu == cpu, 631 ("[pmc,%d] CPU not bound [cpu=%d, curr=%d]", __LINE__, 632 cpu, curthread->td_oncpu)); 633 634 PMCDBG(CPU,SEL,2, "select-cpu cpu=%d ok", cpu); 635} 636 637/* 638 * Force a context switch. 639 * 640 * We do this by tsleep'ing for 1 tick -- invoking mi_switch() is not 641 * guaranteed to force a context switch. 642 */ 643 644static void 645pmc_force_context_switch(void) 646{ 647 u_char curpri; 648 649 mtx_lock_spin(&sched_lock); 650 curpri = curthread->td_priority; 651 mtx_unlock_spin(&sched_lock); 652 653 (void) tsleep((void *) pmc_force_context_switch, curpri, 654 "pmcctx", 1); 655 656} 657 658/* 659 * Get the file name for an executable. This is a simple wrapper 660 * around vn_fullpath(9). 661 */ 662 663static void 664pmc_getfilename(struct vnode *v, char **fullpath, char **freepath) 665{ 666 struct thread *td; 667 668 td = curthread; 669 *fullpath = "unknown"; 670 *freepath = NULL; 671 vn_lock(v, LK_CANRECURSE | LK_EXCLUSIVE | LK_RETRY, td); 672 vn_fullpath(td, v, fullpath, freepath); 673 VOP_UNLOCK(v, 0, td); 674} 675 676/* 677 * remove an process owning PMCs 678 */ 679 680void 681pmc_remove_owner(struct pmc_owner *po) 682{ 683 struct pmc *pm, *tmp; 684 685 sx_assert(&pmc_sx, SX_XLOCKED); 686 687 PMCDBG(OWN,ORM,1, "remove-owner po=%p", po); 688 689 /* Remove descriptor from the owner hash table */ 690 LIST_REMOVE(po, po_next); 691 692 /* release all owned PMC descriptors */ 693 LIST_FOREACH_SAFE(pm, &po->po_pmcs, pm_next, tmp) { 694 PMCDBG(OWN,ORM,2, "pmc=%p", pm); 695 KASSERT(pm->pm_owner == po, 696 ("[pmc,%d] owner %p != po %p", __LINE__, pm->pm_owner, po)); 697 698 pmc_release_pmc_descriptor(pm); /* will unlink from the list */ 699 } 700 701 KASSERT(po->po_sscount == 0, 702 ("[pmc,%d] SS count not zero", __LINE__)); 703 KASSERT(LIST_EMPTY(&po->po_pmcs), 704 ("[pmc,%d] PMC list not empty", __LINE__)); 705 706 /* de-configure the log file if present */ 707 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 708 pmclog_deconfigure_log(po); 709} 710 711/* 712 * remove an owner process record if all conditions are met. 713 */ 714 715static void 716pmc_maybe_remove_owner(struct pmc_owner *po) 717{ 718 719 PMCDBG(OWN,OMR,1, "maybe-remove-owner po=%p", po); 720 721 /* 722 * Remove owner record if 723 * - this process does not own any PMCs 724 * - this process has not allocated a system-wide sampling buffer 725 */ 726 727 if (LIST_EMPTY(&po->po_pmcs) && 728 ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)) { 729 pmc_remove_owner(po); 730 pmc_destroy_owner_descriptor(po); 731 } 732} 733 734/* 735 * Add an association between a target process and a PMC. 736 */ 737 738static void 739pmc_link_target_process(struct pmc *pm, struct pmc_process *pp) 740{ 741 int ri; 742 struct pmc_target *pt; 743 744 sx_assert(&pmc_sx, SX_XLOCKED); 745 746 KASSERT(pm != NULL && pp != NULL, 747 ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp)); 748 KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)), 749 ("[pmc,%d] Attaching a non-process-virtual pmc=%p to pid=%d", 750 __LINE__, pm, pp->pp_proc->p_pid)); 751 KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt < ((int) md->pmd_npmc - 1), 752 ("[pmc,%d] Illegal reference count %d for process record %p", 753 __LINE__, pp->pp_refcnt, (void *) pp)); 754 755 ri = PMC_TO_ROWINDEX(pm); 756 757 PMCDBG(PRC,TLK,1, "link-target pmc=%p ri=%d pmc-process=%p", 758 pm, ri, pp); 759 760#ifdef DEBUG 761 LIST_FOREACH(pt, &pm->pm_targets, pt_next) 762 if (pt->pt_process == pp) 763 KASSERT(0, ("[pmc,%d] pp %p already in pmc %p targets", 764 __LINE__, pp, pm)); 765#endif 766 767 MALLOC(pt, struct pmc_target *, sizeof(struct pmc_target), 768 M_PMC, M_ZERO|M_WAITOK); 769 770 pt->pt_process = pp; 771 772 LIST_INSERT_HEAD(&pm->pm_targets, pt, pt_next); 773 774 atomic_store_rel_ptr((uintptr_t *)&pp->pp_pmcs[ri].pp_pmc, 775 (uintptr_t)pm); 776 777 if (pm->pm_owner->po_owner == pp->pp_proc) 778 pm->pm_flags |= PMC_F_ATTACHED_TO_OWNER; 779 780 /* 781 * Initialize the per-process values at this row index. 782 */ 783 pp->pp_pmcs[ri].pp_pmcval = PMC_TO_MODE(pm) == PMC_MODE_TS ? 784 pm->pm_sc.pm_reloadcount : 0; 785 786 pp->pp_refcnt++; 787 788} 789 790/* 791 * Removes the association between a target process and a PMC. 792 */ 793 794static void 795pmc_unlink_target_process(struct pmc *pm, struct pmc_process *pp) 796{ 797 int ri; 798 struct proc *p; 799 struct pmc_target *ptgt; 800 801 sx_assert(&pmc_sx, SX_XLOCKED); 802 803 KASSERT(pm != NULL && pp != NULL, 804 ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp)); 805 806 KASSERT(pp->pp_refcnt >= 1 && pp->pp_refcnt < (int) md->pmd_npmc, 807 ("[pmc,%d] Illegal ref count %d on process record %p", 808 __LINE__, pp->pp_refcnt, (void *) pp)); 809 810 ri = PMC_TO_ROWINDEX(pm); 811 812 PMCDBG(PRC,TUL,1, "unlink-target pmc=%p ri=%d pmc-process=%p", 813 pm, ri, pp); 814 815 KASSERT(pp->pp_pmcs[ri].pp_pmc == pm, 816 ("[pmc,%d] PMC ri %d mismatch pmc %p pp->[ri] %p", __LINE__, 817 ri, pm, pp->pp_pmcs[ri].pp_pmc)); 818 819 pp->pp_pmcs[ri].pp_pmc = NULL; 820 pp->pp_pmcs[ri].pp_pmcval = (pmc_value_t) 0; 821 822 /* Remove owner-specific flags */ 823 if (pm->pm_owner->po_owner == pp->pp_proc) { 824 pp->pp_flags &= ~PMC_PP_ENABLE_MSR_ACCESS; 825 pm->pm_flags &= ~PMC_F_ATTACHED_TO_OWNER; 826 } 827 828 pp->pp_refcnt--; 829 830 /* Remove the target process from the PMC structure */ 831 LIST_FOREACH(ptgt, &pm->pm_targets, pt_next) 832 if (ptgt->pt_process == pp) 833 break; 834 835 KASSERT(ptgt != NULL, ("[pmc,%d] process %p (pp: %p) not found " 836 "in pmc %p", __LINE__, pp->pp_proc, pp, pm)); 837 838 LIST_REMOVE(ptgt, pt_next); 839 FREE(ptgt, M_PMC); 840 841 /* if the PMC now lacks targets, send the owner a SIGIO */ 842 if (LIST_EMPTY(&pm->pm_targets)) { 843 p = pm->pm_owner->po_owner; 844 PROC_LOCK(p); 845 psignal(p, SIGIO); 846 PROC_UNLOCK(p); 847 848 PMCDBG(PRC,SIG,2, "signalling proc=%p signal=%d", p, 849 SIGIO); 850 } 851} 852 853/* 854 * Check if PMC 'pm' may be attached to target process 't'. 855 */ 856 857static int 858pmc_can_attach(struct pmc *pm, struct proc *t) 859{ 860 struct proc *o; /* pmc owner */ 861 struct ucred *oc, *tc; /* owner, target credentials */ 862 int decline_attach, i; 863 864 /* 865 * A PMC's owner can always attach that PMC to itself. 866 */ 867 868 if ((o = pm->pm_owner->po_owner) == t) 869 return 0; 870 871 PROC_LOCK(o); 872 oc = o->p_ucred; 873 crhold(oc); 874 PROC_UNLOCK(o); 875 876 PROC_LOCK(t); 877 tc = t->p_ucred; 878 crhold(tc); 879 PROC_UNLOCK(t); 880 881 /* 882 * The effective uid of the PMC owner should match at least one 883 * of the {effective,real,saved} uids of the target process. 884 */ 885 886 decline_attach = oc->cr_uid != tc->cr_uid && 887 oc->cr_uid != tc->cr_svuid && 888 oc->cr_uid != tc->cr_ruid; 889 890 /* 891 * Every one of the target's group ids, must be in the owner's 892 * group list. 893 */ 894 for (i = 0; !decline_attach && i < tc->cr_ngroups; i++) 895 decline_attach = !groupmember(tc->cr_groups[i], oc); 896 897 /* check the read and saved gids too */ 898 if (decline_attach == 0) 899 decline_attach = !groupmember(tc->cr_rgid, oc) || 900 !groupmember(tc->cr_svgid, oc); 901 902 crfree(tc); 903 crfree(oc); 904 905 return !decline_attach; 906} 907 908/* 909 * Attach a process to a PMC. 910 */ 911 912static int 913pmc_attach_one_process(struct proc *p, struct pmc *pm) 914{ 915 int ri; 916 char *fullpath, *freepath; 917 struct pmc_process *pp; 918 919 sx_assert(&pmc_sx, SX_XLOCKED); 920 921 PMCDBG(PRC,ATT,2, "attach-one pm=%p ri=%d proc=%p (%d, %s)", pm, 922 PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm); 923 924 /* 925 * Locate the process descriptor corresponding to process 'p', 926 * allocating space as needed. 927 * 928 * Verify that rowindex 'pm_rowindex' is free in the process 929 * descriptor. 930 * 931 * If not, allocate space for a descriptor and link the 932 * process descriptor and PMC. 933 */ 934 ri = PMC_TO_ROWINDEX(pm); 935 936 if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_ALLOCATE)) == NULL) 937 return ENOMEM; 938 939 if (pp->pp_pmcs[ri].pp_pmc == pm) /* already present at slot [ri] */ 940 return EEXIST; 941 942 if (pp->pp_pmcs[ri].pp_pmc != NULL) 943 return EBUSY; 944 945 pmc_link_target_process(pm, pp); 946 947 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) && 948 (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) == 0) 949 pm->pm_flags |= PMC_F_NEEDS_LOGFILE; 950 951 pm->pm_flags |= PMC_F_ATTACH_DONE; /* mark as attached */ 952 953 /* issue an attach event to a configured log file */ 954 if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE) { 955 pmc_getfilename(p->p_textvp, &fullpath, &freepath); 956 pmclog_process_pmcattach(pm, p->p_pid, fullpath); 957 if (freepath) 958 FREE(freepath, M_TEMP); 959 } 960 /* mark process as using HWPMCs */ 961 PROC_LOCK(p); 962 p->p_flag |= P_HWPMC; 963 PROC_UNLOCK(p); 964 965 return 0; 966} 967 968/* 969 * Attach a process and optionally its children 970 */ 971 972static int 973pmc_attach_process(struct proc *p, struct pmc *pm) 974{ 975 int error; 976 struct proc *top; 977 978 sx_assert(&pmc_sx, SX_XLOCKED); 979 980 PMCDBG(PRC,ATT,1, "attach pm=%p ri=%d proc=%p (%d, %s)", pm, 981 PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm); 982 983 984 /* 985 * If this PMC successfully allowed a GETMSR operation 986 * in the past, disallow further ATTACHes. 987 */ 988 989 if ((pm->pm_flags & PMC_PP_ENABLE_MSR_ACCESS) != 0) 990 return EPERM; 991 992 if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0) 993 return pmc_attach_one_process(p, pm); 994 995 /* 996 * Traverse all child processes, attaching them to 997 * this PMC. 998 */ 999 1000 sx_slock(&proctree_lock); 1001 1002 top = p; 1003 1004 for (;;) { 1005 if ((error = pmc_attach_one_process(p, pm)) != 0) 1006 break; 1007 if (!LIST_EMPTY(&p->p_children)) 1008 p = LIST_FIRST(&p->p_children); 1009 else for (;;) { 1010 if (p == top) 1011 goto done; 1012 if (LIST_NEXT(p, p_sibling)) { 1013 p = LIST_NEXT(p, p_sibling); 1014 break; 1015 } 1016 p = p->p_pptr; 1017 } 1018 } 1019 1020 if (error) 1021 (void) pmc_detach_process(top, pm); 1022 1023 done: 1024 sx_sunlock(&proctree_lock); 1025 return error; 1026} 1027 1028/* 1029 * Detach a process from a PMC. If there are no other PMCs tracking 1030 * this process, remove the process structure from its hash table. If 1031 * 'flags' contains PMC_FLAG_REMOVE, then free the process structure. 1032 */ 1033 1034static int 1035pmc_detach_one_process(struct proc *p, struct pmc *pm, int flags) 1036{ 1037 int ri; 1038 struct pmc_process *pp; 1039 1040 sx_assert(&pmc_sx, SX_XLOCKED); 1041 1042 KASSERT(pm != NULL, 1043 ("[pmc,%d] null pm pointer", __LINE__)); 1044 1045 ri = PMC_TO_ROWINDEX(pm); 1046 1047 PMCDBG(PRC,ATT,2, "detach-one pm=%p ri=%d proc=%p (%d, %s) flags=0x%x", 1048 pm, ri, p, p->p_pid, p->p_comm, flags); 1049 1050 if ((pp = pmc_find_process_descriptor(p, 0)) == NULL) 1051 return ESRCH; 1052 1053 if (pp->pp_pmcs[ri].pp_pmc != pm) 1054 return EINVAL; 1055 1056 pmc_unlink_target_process(pm, pp); 1057 1058 /* Issue a detach entry if a log file is configured */ 1059 if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE) 1060 pmclog_process_pmcdetach(pm, p->p_pid); 1061 1062 /* 1063 * If there are no PMCs targetting this process, we remove its 1064 * descriptor from the target hash table and unset the P_HWPMC 1065 * flag in the struct proc. 1066 */ 1067 KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt < (int) md->pmd_npmc, 1068 ("[pmc,%d] Illegal refcnt %d for process struct %p", 1069 __LINE__, pp->pp_refcnt, pp)); 1070 1071 if (pp->pp_refcnt != 0) /* still a target of some PMC */ 1072 return 0; 1073 1074 pmc_remove_process_descriptor(pp); 1075 1076 if (flags & PMC_FLAG_REMOVE) 1077 FREE(pp, M_PMC); 1078 1079 PROC_LOCK(p); 1080 p->p_flag &= ~P_HWPMC; 1081 PROC_UNLOCK(p); 1082 1083 return 0; 1084} 1085 1086/* 1087 * Detach a process and optionally its descendants from a PMC. 1088 */ 1089 1090static int 1091pmc_detach_process(struct proc *p, struct pmc *pm) 1092{ 1093 struct proc *top; 1094 1095 sx_assert(&pmc_sx, SX_XLOCKED); 1096 1097 PMCDBG(PRC,ATT,1, "detach pm=%p ri=%d proc=%p (%d, %s)", pm, 1098 PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm); 1099 1100 if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0) 1101 return pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE); 1102 1103 /* 1104 * Traverse all children, detaching them from this PMC. We 1105 * ignore errors since we could be detaching a PMC from a 1106 * partially attached proc tree. 1107 */ 1108 1109 sx_slock(&proctree_lock); 1110 1111 top = p; 1112 1113 for (;;) { 1114 (void) pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE); 1115 1116 if (!LIST_EMPTY(&p->p_children)) 1117 p = LIST_FIRST(&p->p_children); 1118 else for (;;) { 1119 if (p == top) 1120 goto done; 1121 if (LIST_NEXT(p, p_sibling)) { 1122 p = LIST_NEXT(p, p_sibling); 1123 break; 1124 } 1125 p = p->p_pptr; 1126 } 1127 } 1128 1129 done: 1130 sx_sunlock(&proctree_lock); 1131 1132 if (LIST_EMPTY(&pm->pm_targets)) 1133 pm->pm_flags &= ~PMC_F_ATTACH_DONE; 1134 1135 return 0; 1136} 1137 1138 1139/* 1140 * Thread context switch IN 1141 */ 1142 1143static void 1144pmc_process_csw_in(struct thread *td) 1145{ 1146 int cpu; 1147 unsigned int ri; 1148 struct pmc *pm; 1149 struct proc *p; 1150 struct pmc_cpu *pc; 1151 struct pmc_hw *phw; 1152 struct pmc_process *pp; 1153 pmc_value_t newvalue; 1154 1155 p = td->td_proc; 1156 1157 if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE)) == NULL) 1158 return; 1159 1160 KASSERT(pp->pp_proc == td->td_proc, 1161 ("[pmc,%d] not my thread state", __LINE__)); 1162 1163 critical_enter(); /* no preemption from this point */ 1164 1165 cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */ 1166 1167 PMCDBG(CSW,SWI,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p, 1168 p->p_pid, p->p_comm, pp); 1169 1170 KASSERT(cpu >= 0 && cpu < mp_ncpus, 1171 ("[pmc,%d] wierd CPU id %d", __LINE__, cpu)); 1172 1173 pc = pmc_pcpu[cpu]; 1174 1175 for (ri = 0; ri < md->pmd_npmc; ri++) { 1176 1177 if ((pm = pp->pp_pmcs[ri].pp_pmc) == NULL) 1178 continue; 1179 1180 KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)), 1181 ("[pmc,%d] Target PMC in non-virtual mode (%d)", 1182 __LINE__, PMC_TO_MODE(pm))); 1183 1184 KASSERT(PMC_TO_ROWINDEX(pm) == ri, 1185 ("[pmc,%d] Row index mismatch pmc %d != ri %d", 1186 __LINE__, PMC_TO_ROWINDEX(pm), ri)); 1187 1188 /* 1189 * Only PMCs that are marked as 'RUNNING' need 1190 * be placed on hardware. 1191 */ 1192 1193 if (pm->pm_state != PMC_STATE_RUNNING) 1194 continue; 1195 1196 /* increment PMC runcount */ 1197 atomic_add_rel_32(&pm->pm_runcount, 1); 1198 1199 /* configure the HWPMC we are going to use. */ 1200 md->pmd_config_pmc(cpu, ri, pm); 1201 1202 phw = pc->pc_hwpmcs[ri]; 1203 1204 KASSERT(phw != NULL, 1205 ("[pmc,%d] null hw pointer", __LINE__)); 1206 1207 KASSERT(phw->phw_pmc == pm, 1208 ("[pmc,%d] hw->pmc %p != pmc %p", __LINE__, 1209 phw->phw_pmc, pm)); 1210 1211 /* 1212 * Write out saved value and start the PMC. 1213 * 1214 * Sampling PMCs use a per-process value, while 1215 * counting mode PMCs use a per-pmc value that is 1216 * inherited across descendants. 1217 */ 1218 if (PMC_TO_MODE(pm) == PMC_MODE_TS) { 1219 mtx_pool_lock_spin(pmc_mtxpool, pm); 1220 newvalue = PMC_PCPU_SAVED(cpu,ri) = 1221 pp->pp_pmcs[ri].pp_pmcval; 1222 mtx_pool_unlock_spin(pmc_mtxpool, pm); 1223 } else { 1224 KASSERT(PMC_TO_MODE(pm) == PMC_MODE_TC, 1225 ("[pmc,%d] illegal mode=%d", __LINE__, 1226 PMC_TO_MODE(pm))); 1227 mtx_pool_lock_spin(pmc_mtxpool, pm); 1228 newvalue = PMC_PCPU_SAVED(cpu, ri) = 1229 pm->pm_gv.pm_savedvalue; 1230 mtx_pool_unlock_spin(pmc_mtxpool, pm); 1231 } 1232 1233 PMCDBG(CSW,SWI,1,"cpu=%d ri=%d new=%jd", cpu, ri, newvalue); 1234 1235 md->pmd_write_pmc(cpu, ri, newvalue); 1236 md->pmd_start_pmc(cpu, ri); 1237 } 1238 1239 /* 1240 * perform any other architecture/cpu dependent thread 1241 * switch-in actions. 1242 */ 1243 1244 (void) (*md->pmd_switch_in)(pc, pp); 1245 1246 critical_exit(); 1247 1248} 1249 1250/* 1251 * Thread context switch OUT. 1252 */ 1253 1254static void 1255pmc_process_csw_out(struct thread *td) 1256{ 1257 int cpu; 1258 enum pmc_mode mode; 1259 unsigned int ri; 1260 struct pmc *pm; 1261 struct proc *p; 1262 struct pmc_cpu *pc; 1263 struct pmc_process *pp; 1264 int64_t tmp; 1265 pmc_value_t newvalue; 1266 1267 /* 1268 * Locate our process descriptor; this may be NULL if 1269 * this process is exiting and we have already removed 1270 * the process from the target process table. 1271 * 1272 * Note that due to kernel preemption, multiple 1273 * context switches may happen while the process is 1274 * exiting. 1275 * 1276 * Note also that if the target process cannot be 1277 * found we still need to deconfigure any PMCs that 1278 * are currently running on hardware. 1279 */ 1280 1281 p = td->td_proc; 1282 pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE); 1283 1284 /* 1285 * save PMCs 1286 */ 1287 1288 critical_enter(); 1289 1290 cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */ 1291 1292 PMCDBG(CSW,SWO,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p, 1293 p->p_pid, p->p_comm, pp); 1294 1295 KASSERT(cpu >= 0 && cpu < mp_ncpus, 1296 ("[pmc,%d wierd CPU id %d", __LINE__, cpu)); 1297 1298 pc = pmc_pcpu[cpu]; 1299 1300 /* 1301 * When a PMC gets unlinked from a target PMC, it will 1302 * be removed from the target's pp_pmc[] array. 1303 * 1304 * However, on a MP system, the target could have been 1305 * executing on another CPU at the time of the unlink. 1306 * So, at context switch OUT time, we need to look at 1307 * the hardware to determine if a PMC is scheduled on 1308 * it. 1309 */ 1310 1311 for (ri = 0; ri < md->pmd_npmc; ri++) { 1312 1313 pm = NULL; 1314 (void) (*md->pmd_get_config)(cpu, ri, &pm); 1315 1316 if (pm == NULL) /* nothing at this row index */ 1317 continue; 1318 1319 mode = PMC_TO_MODE(pm); 1320 if (!PMC_IS_VIRTUAL_MODE(mode)) 1321 continue; /* not a process virtual PMC */ 1322 1323 KASSERT(PMC_TO_ROWINDEX(pm) == ri, 1324 ("[pmc,%d] ri mismatch pmc(%d) ri(%d)", 1325 __LINE__, PMC_TO_ROWINDEX(pm), ri)); 1326 1327 /* Stop hardware if not already stopped */ 1328 if (pm->pm_stalled == 0) 1329 md->pmd_stop_pmc(cpu, ri); 1330 1331 /* reduce this PMC's runcount */ 1332 atomic_subtract_rel_32(&pm->pm_runcount, 1); 1333 1334 /* 1335 * If this PMC is associated with this process, 1336 * save the reading. 1337 */ 1338 1339 if (pp != NULL && pp->pp_pmcs[ri].pp_pmc != NULL) { 1340 1341 KASSERT(pm == pp->pp_pmcs[ri].pp_pmc, 1342 ("[pmc,%d] pm %p != pp_pmcs[%d] %p", __LINE__, 1343 pm, ri, pp->pp_pmcs[ri].pp_pmc)); 1344 1345 KASSERT(pp->pp_refcnt > 0, 1346 ("[pmc,%d] pp refcnt = %d", __LINE__, 1347 pp->pp_refcnt)); 1348 1349 md->pmd_read_pmc(cpu, ri, &newvalue); 1350 1351 tmp = newvalue - PMC_PCPU_SAVED(cpu,ri); 1352 1353 PMCDBG(CSW,SWI,1,"cpu=%d ri=%d tmp=%jd", cpu, ri, 1354 tmp); 1355 1356 if (mode == PMC_MODE_TS) { 1357 1358 /* 1359 * For sampling process-virtual PMCs, 1360 * we expect the count to be 1361 * decreasing as the 'value' 1362 * programmed into the PMC is the 1363 * number of events to be seen till 1364 * the next sampling interrupt. 1365 */ 1366 if (tmp < 0) 1367 tmp += pm->pm_sc.pm_reloadcount; 1368 mtx_pool_lock_spin(pmc_mtxpool, pm); 1369 pp->pp_pmcs[ri].pp_pmcval -= tmp; 1370 if ((int64_t) pp->pp_pmcs[ri].pp_pmcval < 0) 1371 pp->pp_pmcs[ri].pp_pmcval += 1372 pm->pm_sc.pm_reloadcount; 1373 mtx_pool_unlock_spin(pmc_mtxpool, pm); 1374 1375 } else { 1376 1377 /* 1378 * For counting process-virtual PMCs, 1379 * we expect the count to be 1380 * increasing monotonically, modulo a 64 1381 * bit wraparound. 1382 */ 1383 KASSERT((int64_t) tmp >= 0, 1384 ("[pmc,%d] negative increment cpu=%d " 1385 "ri=%d newvalue=%jx saved=%jx " 1386 "incr=%jx", __LINE__, cpu, ri, 1387 newvalue, PMC_PCPU_SAVED(cpu,ri), tmp)); 1388 1389 mtx_pool_lock_spin(pmc_mtxpool, pm); 1390 pm->pm_gv.pm_savedvalue += tmp; 1391 pp->pp_pmcs[ri].pp_pmcval += tmp; 1392 mtx_pool_unlock_spin(pmc_mtxpool, pm); 1393 1394 if (pm->pm_flags & PMC_F_LOG_PROCCSW) 1395 pmclog_process_proccsw(pm, pp, tmp); 1396 } 1397 } 1398 1399 /* mark hardware as free */ 1400 md->pmd_config_pmc(cpu, ri, NULL); 1401 } 1402 1403 /* 1404 * perform any other architecture/cpu dependent thread 1405 * switch out functions. 1406 */ 1407 1408 (void) (*md->pmd_switch_out)(pc, pp); 1409 1410 critical_exit(); 1411} 1412 1413/* 1414 * The 'hook' invoked from the kernel proper 1415 */ 1416 1417 1418#ifdef DEBUG 1419const char *pmc_hooknames[] = { 1420 "", 1421 "EXIT", 1422 "EXEC", 1423 "FORK", 1424 "CSW-IN", 1425 "CSW-OUT", 1426 "SAMPLE" 1427}; 1428#endif 1429 1430static int 1431pmc_hook_handler(struct thread *td, int function, void *arg) 1432{ 1433 1434 PMCDBG(MOD,PMH,1, "hook td=%p func=%d \"%s\" arg=%p", td, function, 1435 pmc_hooknames[function], arg); 1436 1437 switch (function) 1438 { 1439 1440 /* 1441 * Process exec() 1442 */ 1443 1444 case PMC_FN_PROCESS_EXEC: 1445 { 1446 char *fullpath, *freepath; 1447 unsigned int ri; 1448 int is_using_hwpmcs; 1449 struct pmc *pm; 1450 struct proc *p; 1451 struct pmc_owner *po; 1452 struct pmc_process *pp; 1453 struct pmckern_procexec *pk; 1454 1455 sx_assert(&pmc_sx, SX_XLOCKED); 1456 1457 p = td->td_proc; 1458 pmc_getfilename(p->p_textvp, &fullpath, &freepath); 1459 1460 pk = (struct pmckern_procexec *) arg; 1461 1462 /* Inform owners of SS mode PMCs of the exec event. */ 1463 LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) 1464 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 1465 pmclog_process_procexec(po, PMC_ID_INVALID, 1466 p->p_pid, pk->pm_entryaddr, fullpath); 1467 1468 PROC_LOCK(p); 1469 is_using_hwpmcs = p->p_flag & P_HWPMC; 1470 PROC_UNLOCK(p); 1471 1472 if (!is_using_hwpmcs) { 1473 if (freepath) 1474 FREE(freepath, M_TEMP); 1475 break; 1476 } 1477 1478 /* 1479 * PMCs are not inherited across an exec(): remove any 1480 * PMCs that this process is the owner of. 1481 */ 1482 1483 if ((po = pmc_find_owner_descriptor(p)) != NULL) { 1484 pmc_remove_owner(po); 1485 pmc_destroy_owner_descriptor(po); 1486 } 1487 1488 /* 1489 * If the process being exec'ed is not the target of any 1490 * PMC, we are done. 1491 */ 1492 if ((pp = pmc_find_process_descriptor(p, 0)) == NULL) { 1493 if (freepath) 1494 FREE(freepath, M_TEMP); 1495 break; 1496 } 1497 1498 /* 1499 * Log the exec event to all monitoring owners. Skip 1500 * owners who have already recieved the event because 1501 * they had system sampling PMCs active. 1502 */ 1503 for (ri = 0; ri < md->pmd_npmc; ri++) 1504 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) { 1505 po = pm->pm_owner; 1506 if (po->po_sscount == 0 && 1507 po->po_flags & PMC_PO_OWNS_LOGFILE) 1508 pmclog_process_procexec(po, pm->pm_id, 1509 p->p_pid, pk->pm_entryaddr, 1510 fullpath); 1511 } 1512 1513 if (freepath) 1514 FREE(freepath, M_TEMP); 1515 1516 1517 PMCDBG(PRC,EXC,1, "exec proc=%p (%d, %s) cred-changed=%d", 1518 p, p->p_pid, p->p_comm, pk->pm_credentialschanged); 1519 1520 if (pk->pm_credentialschanged == 0) /* no change */ 1521 break; 1522 1523 /* 1524 * If the newly exec()'ed process has a different credential 1525 * than before, allow it to be the target of a PMC only if 1526 * the PMC's owner has sufficient priviledge. 1527 */ 1528 1529 for (ri = 0; ri < md->pmd_npmc; ri++) 1530 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) 1531 if (pmc_can_attach(pm, td->td_proc) != 0) 1532 pmc_detach_one_process(td->td_proc, 1533 pm, PMC_FLAG_NONE); 1534 1535 KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt < (int) md->pmd_npmc, 1536 ("[pmc,%d] Illegal ref count %d on pp %p", __LINE__, 1537 pp->pp_refcnt, pp)); 1538 1539 /* 1540 * If this process is no longer the target of any 1541 * PMCs, we can remove the process entry and free 1542 * up space. 1543 */ 1544 1545 if (pp->pp_refcnt == 0) { 1546 pmc_remove_process_descriptor(pp); 1547 FREE(pp, M_PMC); 1548 break; 1549 } 1550 1551 } 1552 break; 1553 1554 case PMC_FN_CSW_IN: 1555 pmc_process_csw_in(td); 1556 break; 1557 1558 case PMC_FN_CSW_OUT: 1559 pmc_process_csw_out(td); 1560 break; 1561 1562 /* 1563 * Process accumulated PC samples. 1564 * 1565 * This function is expected to be called by hardclock() for 1566 * each CPU that has accumulated PC samples. 1567 * 1568 * This function is to be executed on the CPU whose samples 1569 * are being processed. 1570 */ 1571 case PMC_FN_DO_SAMPLES: 1572 1573 /* 1574 * Clear the cpu specific bit in the CPU mask before 1575 * do the rest of the processing. If the NMI handler 1576 * gets invoked after the "atomic_clear_int()" call 1577 * below but before "pmc_process_samples()" gets 1578 * around to processing the interrupt, then we will 1579 * come back here at the next hardclock() tick (and 1580 * may find nothing to do if "pmc_process_samples()" 1581 * had already processed the interrupt). We don't 1582 * lose the interrupt sample. 1583 */ 1584 atomic_clear_int(&pmc_cpumask, (1 << PCPU_GET(cpuid))); 1585 pmc_process_samples(PCPU_GET(cpuid)); 1586 break; 1587 1588 default: 1589#ifdef DEBUG 1590 KASSERT(0, ("[pmc,%d] unknown hook %d\n", __LINE__, function)); 1591#endif 1592 break; 1593 1594 } 1595 1596 return 0; 1597} 1598 1599/* 1600 * allocate a 'struct pmc_owner' descriptor in the owner hash table. 1601 */ 1602 1603static struct pmc_owner * 1604pmc_allocate_owner_descriptor(struct proc *p) 1605{ 1606 uint32_t hindex; 1607 struct pmc_owner *po; 1608 struct pmc_ownerhash *poh; 1609 1610 hindex = PMC_HASH_PTR(p, pmc_ownerhashmask); 1611 poh = &pmc_ownerhash[hindex]; 1612 1613 /* allocate space for N pointers and one descriptor struct */ 1614 MALLOC(po, struct pmc_owner *, sizeof(struct pmc_owner), 1615 M_PMC, M_ZERO|M_WAITOK); 1616 1617 po->po_sscount = po->po_error = po->po_flags = 0; 1618 po->po_file = NULL; 1619 po->po_owner = p; 1620 po->po_kthread = NULL; 1621 LIST_INIT(&po->po_pmcs); 1622 LIST_INSERT_HEAD(poh, po, po_next); /* insert into hash table */ 1623 1624 TAILQ_INIT(&po->po_logbuffers); 1625 mtx_init(&po->po_mtx, "pmc-owner-mtx", "pmc", MTX_SPIN); 1626 1627 PMCDBG(OWN,ALL,1, "allocate-owner proc=%p (%d, %s) pmc-owner=%p", 1628 p, p->p_pid, p->p_comm, po); 1629 1630 return po; 1631} 1632 1633static void 1634pmc_destroy_owner_descriptor(struct pmc_owner *po) 1635{ 1636 1637 PMCDBG(OWN,REL,1, "destroy-owner po=%p proc=%p (%d, %s)", 1638 po, po->po_owner, po->po_owner->p_pid, po->po_owner->p_comm); 1639 1640 mtx_destroy(&po->po_mtx); 1641 FREE(po, M_PMC); 1642} 1643 1644/* 1645 * find the descriptor corresponding to process 'p', adding or removing it 1646 * as specified by 'mode'. 1647 */ 1648 1649static struct pmc_process * 1650pmc_find_process_descriptor(struct proc *p, uint32_t mode) 1651{ 1652 uint32_t hindex; 1653 struct pmc_process *pp, *ppnew; 1654 struct pmc_processhash *pph; 1655 1656 hindex = PMC_HASH_PTR(p, pmc_processhashmask); 1657 pph = &pmc_processhash[hindex]; 1658 1659 ppnew = NULL; 1660 1661 /* 1662 * Pre-allocate memory in the FIND_ALLOCATE case since we 1663 * cannot call malloc(9) once we hold a spin lock. 1664 */ 1665 1666 if (mode & PMC_FLAG_ALLOCATE) { 1667 /* allocate additional space for 'n' pmc pointers */ 1668 MALLOC(ppnew, struct pmc_process *, 1669 sizeof(struct pmc_process) + md->pmd_npmc * 1670 sizeof(struct pmc_targetstate), M_PMC, M_ZERO|M_WAITOK); 1671 } 1672 1673 mtx_lock_spin(&pmc_processhash_mtx); 1674 LIST_FOREACH(pp, pph, pp_next) 1675 if (pp->pp_proc == p) 1676 break; 1677 1678 if ((mode & PMC_FLAG_REMOVE) && pp != NULL) 1679 LIST_REMOVE(pp, pp_next); 1680 1681 if ((mode & PMC_FLAG_ALLOCATE) && pp == NULL && 1682 ppnew != NULL) { 1683 ppnew->pp_proc = p; 1684 LIST_INSERT_HEAD(pph, ppnew, pp_next); 1685 pp = ppnew; 1686 ppnew = NULL; 1687 } 1688 mtx_unlock_spin(&pmc_processhash_mtx); 1689 1690 if (pp != NULL && ppnew != NULL) 1691 FREE(ppnew, M_PMC); 1692 1693 return pp; 1694} 1695 1696/* 1697 * remove a process descriptor from the process hash table. 1698 */ 1699 1700static void 1701pmc_remove_process_descriptor(struct pmc_process *pp) 1702{ 1703 KASSERT(pp->pp_refcnt == 0, 1704 ("[pmc,%d] Removing process descriptor %p with count %d", 1705 __LINE__, pp, pp->pp_refcnt)); 1706 1707 mtx_lock_spin(&pmc_processhash_mtx); 1708 LIST_REMOVE(pp, pp_next); 1709 mtx_unlock_spin(&pmc_processhash_mtx); 1710} 1711 1712 1713/* 1714 * find an owner descriptor corresponding to proc 'p' 1715 */ 1716 1717static struct pmc_owner * 1718pmc_find_owner_descriptor(struct proc *p) 1719{ 1720 uint32_t hindex; 1721 struct pmc_owner *po; 1722 struct pmc_ownerhash *poh; 1723 1724 hindex = PMC_HASH_PTR(p, pmc_ownerhashmask); 1725 poh = &pmc_ownerhash[hindex]; 1726 1727 po = NULL; 1728 LIST_FOREACH(po, poh, po_next) 1729 if (po->po_owner == p) 1730 break; 1731 1732 PMCDBG(OWN,FND,1, "find-owner proc=%p (%d, %s) hindex=0x%x -> " 1733 "pmc-owner=%p", p, p->p_pid, p->p_comm, hindex, po); 1734 1735 return po; 1736} 1737 1738/* 1739 * pmc_allocate_pmc_descriptor 1740 * 1741 * Allocate a pmc descriptor and initialize its 1742 * fields. 1743 */ 1744 1745static struct pmc * 1746pmc_allocate_pmc_descriptor(void) 1747{ 1748 struct pmc *pmc; 1749 1750 MALLOC(pmc, struct pmc *, sizeof(struct pmc), M_PMC, M_ZERO|M_WAITOK); 1751 1752 if (pmc != NULL) { 1753 pmc->pm_owner = NULL; 1754 LIST_INIT(&pmc->pm_targets); 1755 } 1756 1757 PMCDBG(PMC,ALL,1, "allocate-pmc -> pmc=%p", pmc); 1758 1759 return pmc; 1760} 1761 1762/* 1763 * Destroy a pmc descriptor. 1764 */ 1765 1766static void 1767pmc_destroy_pmc_descriptor(struct pmc *pm) 1768{ 1769 (void) pm; 1770 1771#ifdef DEBUG 1772 KASSERT(pm->pm_state == PMC_STATE_DELETED || 1773 pm->pm_state == PMC_STATE_FREE, 1774 ("[pmc,%d] destroying non-deleted PMC", __LINE__)); 1775 KASSERT(LIST_EMPTY(&pm->pm_targets), 1776 ("[pmc,%d] destroying pmc with targets", __LINE__)); 1777 KASSERT(pm->pm_owner == NULL, 1778 ("[pmc,%d] destroying pmc attached to an owner", __LINE__)); 1779 KASSERT(pm->pm_runcount == 0, 1780 ("[pmc,%d] pmc has non-zero run count %d", __LINE__, 1781 pm->pm_runcount)); 1782#endif 1783} 1784 1785static void 1786pmc_wait_for_pmc_idle(struct pmc *pm) 1787{ 1788#ifdef DEBUG 1789 volatile int maxloop; 1790 1791 maxloop = 100 * mp_ncpus; 1792#endif 1793 1794 /* 1795 * Loop (with a forced context switch) till the PMC's runcount 1796 * comes down to zero. 1797 */ 1798 while (atomic_load_acq_32(&pm->pm_runcount) > 0) { 1799#ifdef DEBUG 1800 maxloop--; 1801 KASSERT(maxloop > 0, 1802 ("[pmc,%d] (ri%d, rc%d) waiting too long for " 1803 "pmc to be free", __LINE__, 1804 PMC_TO_ROWINDEX(pm), pm->pm_runcount)); 1805#endif 1806 pmc_force_context_switch(); 1807 } 1808} 1809 1810/* 1811 * This function does the following things: 1812 * 1813 * - detaches the PMC from hardware 1814 * - unlinks all target threads that were attached to it 1815 * - removes the PMC from its owner's list 1816 * - destroy's the PMC private mutex 1817 * 1818 * Once this function completes, the given pmc pointer can be safely 1819 * FREE'd by the caller. 1820 */ 1821 1822static void 1823pmc_release_pmc_descriptor(struct pmc *pm) 1824{ 1825 u_int ri, cpu; 1826 enum pmc_mode mode; 1827 struct pmc_hw *phw; 1828 struct pmc_owner *po; 1829 struct pmc_process *pp; 1830 struct pmc_target *ptgt, *tmp; 1831 struct pmc_binding pb; 1832 1833 sx_assert(&pmc_sx, SX_XLOCKED); 1834 1835 KASSERT(pm, ("[pmc,%d] null pmc", __LINE__)); 1836 1837 ri = PMC_TO_ROWINDEX(pm); 1838 mode = PMC_TO_MODE(pm); 1839 1840 PMCDBG(PMC,REL,1, "release-pmc pmc=%p ri=%d mode=%d", pm, ri, 1841 mode); 1842 1843 /* 1844 * First, we take the PMC off hardware. 1845 */ 1846 cpu = 0; 1847 if (PMC_IS_SYSTEM_MODE(mode)) { 1848 1849 /* 1850 * A system mode PMC runs on a specific CPU. Switch 1851 * to this CPU and turn hardware off. 1852 */ 1853 pmc_save_cpu_binding(&pb); 1854 1855 cpu = PMC_TO_CPU(pm); 1856 1857 pmc_select_cpu(cpu); 1858 1859 /* switch off non-stalled CPUs */ 1860 if (pm->pm_state == PMC_STATE_RUNNING && 1861 pm->pm_stalled == 0) { 1862 1863 phw = pmc_pcpu[cpu]->pc_hwpmcs[ri]; 1864 1865 KASSERT(phw->phw_pmc == pm, 1866 ("[pmc, %d] pmc ptr ri(%d) hw(%p) pm(%p)", 1867 __LINE__, ri, phw->phw_pmc, pm)); 1868 PMCDBG(PMC,REL,2, "stopping cpu=%d ri=%d", cpu, ri); 1869 1870 critical_enter(); 1871 md->pmd_stop_pmc(cpu, ri); 1872 critical_exit(); 1873 } 1874 1875 PMCDBG(PMC,REL,2, "decfg cpu=%d ri=%d", cpu, ri); 1876 1877 critical_enter(); 1878 md->pmd_config_pmc(cpu, ri, NULL); 1879 critical_exit(); 1880 1881 /* adjust the global and process count of SS mode PMCs */ 1882 if (mode == PMC_MODE_SS && pm->pm_state == PMC_STATE_RUNNING) { 1883 po = pm->pm_owner; 1884 po->po_sscount--; 1885 if (po->po_sscount == 0) { 1886 atomic_subtract_rel_int(&pmc_ss_count, 1); 1887 LIST_REMOVE(po, po_ssnext); 1888 } 1889 } 1890 1891 pm->pm_state = PMC_STATE_DELETED; 1892 1893 pmc_restore_cpu_binding(&pb); 1894 1895 /* 1896 * We could have references to this PMC structure in 1897 * the per-cpu sample queues. Wait for the queue to 1898 * drain. 1899 */ 1900 pmc_wait_for_pmc_idle(pm); 1901 1902 } else if (PMC_IS_VIRTUAL_MODE(mode)) { 1903 1904 /* 1905 * A virtual PMC could be running on multiple CPUs at 1906 * a given instant. 1907 * 1908 * By marking its state as DELETED, we ensure that 1909 * this PMC is never further scheduled on hardware. 1910 * 1911 * Then we wait till all CPUs are done with this PMC. 1912 */ 1913 pm->pm_state = PMC_STATE_DELETED; 1914 1915 1916 /* Wait for the PMCs runcount to come to zero. */ 1917 pmc_wait_for_pmc_idle(pm); 1918 1919 /* 1920 * At this point the PMC is off all CPUs and cannot be 1921 * freshly scheduled onto a CPU. It is now safe to 1922 * unlink all targets from this PMC. If a 1923 * process-record's refcount falls to zero, we remove 1924 * it from the hash table. The module-wide SX lock 1925 * protects us from races. 1926 */ 1927 LIST_FOREACH_SAFE(ptgt, &pm->pm_targets, pt_next, tmp) { 1928 pp = ptgt->pt_process; 1929 pmc_unlink_target_process(pm, pp); /* frees 'ptgt' */ 1930 1931 PMCDBG(PMC,REL,3, "pp->refcnt=%d", pp->pp_refcnt); 1932 1933 /* 1934 * If the target process record shows that no 1935 * PMCs are attached to it, reclaim its space. 1936 */ 1937 1938 if (pp->pp_refcnt == 0) { 1939 pmc_remove_process_descriptor(pp); 1940 FREE(pp, M_PMC); 1941 } 1942 } 1943 1944 cpu = curthread->td_oncpu; /* setup cpu for pmd_release() */ 1945 1946 } 1947 1948 /* 1949 * Release any MD resources 1950 */ 1951 1952 (void) md->pmd_release_pmc(cpu, ri, pm); 1953 1954 /* 1955 * Update row disposition 1956 */ 1957 1958 if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm))) 1959 PMC_UNMARK_ROW_STANDALONE(ri); 1960 else 1961 PMC_UNMARK_ROW_THREAD(ri); 1962 1963 /* unlink from the owner's list */ 1964 if (pm->pm_owner) { 1965 LIST_REMOVE(pm, pm_next); 1966 pm->pm_owner = NULL; 1967 } 1968 1969 pmc_destroy_pmc_descriptor(pm); 1970} 1971 1972/* 1973 * Register an owner and a pmc. 1974 */ 1975 1976static int 1977pmc_register_owner(struct proc *p, struct pmc *pmc) 1978{ 1979 struct pmc_owner *po; 1980 1981 sx_assert(&pmc_sx, SX_XLOCKED); 1982 1983 if ((po = pmc_find_owner_descriptor(p)) == NULL) 1984 if ((po = pmc_allocate_owner_descriptor(p)) == NULL) 1985 return ENOMEM; 1986 1987 KASSERT(pmc->pm_owner == NULL, 1988 ("[pmc,%d] attempting to own an initialized PMC", __LINE__)); 1989 pmc->pm_owner = po; 1990 1991 LIST_INSERT_HEAD(&po->po_pmcs, pmc, pm_next); 1992 1993 PROC_LOCK(p); 1994 p->p_flag |= P_HWPMC; 1995 PROC_UNLOCK(p); 1996 1997 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 1998 pmclog_process_pmcallocate(pmc); 1999 2000 PMCDBG(PMC,REG,1, "register-owner pmc-owner=%p pmc=%p", 2001 po, pmc); 2002 2003 return 0; 2004} 2005 2006/* 2007 * Return the current row disposition: 2008 * == 0 => FREE 2009 * > 0 => PROCESS MODE 2010 * < 0 => SYSTEM MODE 2011 */ 2012 2013int 2014pmc_getrowdisp(int ri) 2015{ 2016 return pmc_pmcdisp[ri]; 2017} 2018 2019/* 2020 * Check if a PMC at row index 'ri' can be allocated to the current 2021 * process. 2022 * 2023 * Allocation can fail if: 2024 * - the current process is already being profiled by a PMC at index 'ri', 2025 * attached to it via OP_PMCATTACH. 2026 * - the current process has already allocated a PMC at index 'ri' 2027 * via OP_ALLOCATE. 2028 */ 2029 2030static int 2031pmc_can_allocate_rowindex(struct proc *p, unsigned int ri, int cpu) 2032{ 2033 enum pmc_mode mode; 2034 struct pmc *pm; 2035 struct pmc_owner *po; 2036 struct pmc_process *pp; 2037 2038 PMCDBG(PMC,ALR,1, "can-allocate-rowindex proc=%p (%d, %s) ri=%d " 2039 "cpu=%d", p, p->p_pid, p->p_comm, ri, cpu); 2040 2041 /* 2042 * We shouldn't have already allocated a process-mode PMC at 2043 * row index 'ri'. 2044 * 2045 * We shouldn't have allocated a system-wide PMC on the same 2046 * CPU and same RI. 2047 */ 2048 if ((po = pmc_find_owner_descriptor(p)) != NULL) 2049 LIST_FOREACH(pm, &po->po_pmcs, pm_next) { 2050 if (PMC_TO_ROWINDEX(pm) == ri) { 2051 mode = PMC_TO_MODE(pm); 2052 if (PMC_IS_VIRTUAL_MODE(mode)) 2053 return EEXIST; 2054 if (PMC_IS_SYSTEM_MODE(mode) && 2055 (int) PMC_TO_CPU(pm) == cpu) 2056 return EEXIST; 2057 } 2058 } 2059 2060 /* 2061 * We also shouldn't be the target of any PMC at this index 2062 * since otherwise a PMC_ATTACH to ourselves will fail. 2063 */ 2064 if ((pp = pmc_find_process_descriptor(p, 0)) != NULL) 2065 if (pp->pp_pmcs[ri].pp_pmc) 2066 return EEXIST; 2067 2068 PMCDBG(PMC,ALR,2, "can-allocate-rowindex proc=%p (%d, %s) ri=%d ok", 2069 p, p->p_pid, p->p_comm, ri); 2070 2071 return 0; 2072} 2073 2074/* 2075 * Check if a given PMC at row index 'ri' can be currently used in 2076 * mode 'mode'. 2077 */ 2078 2079static int 2080pmc_can_allocate_row(int ri, enum pmc_mode mode) 2081{ 2082 enum pmc_disp disp; 2083 2084 sx_assert(&pmc_sx, SX_XLOCKED); 2085 2086 PMCDBG(PMC,ALR,1, "can-allocate-row ri=%d mode=%d", ri, mode); 2087 2088 if (PMC_IS_SYSTEM_MODE(mode)) 2089 disp = PMC_DISP_STANDALONE; 2090 else 2091 disp = PMC_DISP_THREAD; 2092 2093 /* 2094 * check disposition for PMC row 'ri': 2095 * 2096 * Expected disposition Row-disposition Result 2097 * 2098 * STANDALONE STANDALONE or FREE proceed 2099 * STANDALONE THREAD fail 2100 * THREAD THREAD or FREE proceed 2101 * THREAD STANDALONE fail 2102 */ 2103 2104 if (!PMC_ROW_DISP_IS_FREE(ri) && 2105 !(disp == PMC_DISP_THREAD && PMC_ROW_DISP_IS_THREAD(ri)) && 2106 !(disp == PMC_DISP_STANDALONE && PMC_ROW_DISP_IS_STANDALONE(ri))) 2107 return EBUSY; 2108 2109 /* 2110 * All OK 2111 */ 2112 2113 PMCDBG(PMC,ALR,2, "can-allocate-row ri=%d mode=%d ok", ri, mode); 2114 2115 return 0; 2116 2117} 2118 2119/* 2120 * Find a PMC descriptor with user handle 'pmcid' for thread 'td'. 2121 */ 2122 2123static struct pmc * 2124pmc_find_pmc_descriptor_in_process(struct pmc_owner *po, pmc_id_t pmcid) 2125{ 2126 struct pmc *pm; 2127 2128 KASSERT(PMC_ID_TO_ROWINDEX(pmcid) < md->pmd_npmc, 2129 ("[pmc,%d] Illegal pmc index %d (max %d)", __LINE__, 2130 PMC_ID_TO_ROWINDEX(pmcid), md->pmd_npmc)); 2131 2132 LIST_FOREACH(pm, &po->po_pmcs, pm_next) 2133 if (pm->pm_id == pmcid) 2134 return pm; 2135 2136 return NULL; 2137} 2138 2139static int 2140pmc_find_pmc(pmc_id_t pmcid, struct pmc **pmc) 2141{ 2142 2143 struct pmc *pm; 2144 struct pmc_owner *po; 2145 2146 PMCDBG(PMC,FND,1, "find-pmc id=%d", pmcid); 2147 2148 if ((po = pmc_find_owner_descriptor(curthread->td_proc)) == NULL) 2149 return ESRCH; 2150 2151 if ((pm = pmc_find_pmc_descriptor_in_process(po, pmcid)) == NULL) 2152 return EINVAL; 2153 2154 PMCDBG(PMC,FND,2, "find-pmc id=%d -> pmc=%p", pmcid, pm); 2155 2156 *pmc = pm; 2157 return 0; 2158} 2159 2160/* 2161 * Start a PMC. 2162 */ 2163 2164static int 2165pmc_start(struct pmc *pm) 2166{ 2167 int error, cpu, ri; 2168 enum pmc_mode mode; 2169 struct pmc_owner *po; 2170 struct pmc_binding pb; 2171 2172 KASSERT(pm != NULL, 2173 ("[pmc,%d] null pm", __LINE__)); 2174 2175 mode = PMC_TO_MODE(pm); 2176 ri = PMC_TO_ROWINDEX(pm); 2177 error = 0; 2178 2179 PMCDBG(PMC,OPS,1, "start pmc=%p mode=%d ri=%d", pm, mode, ri); 2180 2181 po = pm->pm_owner; 2182 2183 if (PMC_IS_VIRTUAL_MODE(mode)) { 2184 2185 /* 2186 * If a PMCATTACH has never been done on this PMC, 2187 * attach it to its owner process. 2188 */ 2189 2190 if (LIST_EMPTY(&pm->pm_targets)) 2191 error = (pm->pm_flags & PMC_F_ATTACH_DONE) ? ESRCH : 2192 pmc_attach_process(po->po_owner, pm); 2193 2194 /* 2195 * Disallow PMCSTART if a logfile is required but has not 2196 * been configured yet. 2197 */ 2198 2199 if (error == 0 && (pm->pm_flags & PMC_F_NEEDS_LOGFILE) && 2200 (po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) 2201 error = EDOOFUS; 2202 2203 /* 2204 * If the PMC is attached to its owner, then force a context 2205 * switch to ensure that the MD state gets set correctly. 2206 */ 2207 2208 if (error == 0) { 2209 pm->pm_state = PMC_STATE_RUNNING; 2210 if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) 2211 pmc_force_context_switch(); 2212 } 2213 2214 return error; 2215 } 2216 2217 2218 /* 2219 * A system-wide PMC. 2220 */ 2221 2222 if ((pm->pm_flags & PMC_F_NEEDS_LOGFILE) && 2223 (po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) 2224 return EDOOFUS; /* programming error */ 2225 2226 /* 2227 * Add the owner to the global list if this is a system-wide 2228 * sampling PMC. 2229 */ 2230 2231 if (mode == PMC_MODE_SS) { 2232 if (po->po_sscount == 0) { 2233 LIST_INSERT_HEAD(&pmc_ss_owners, po, po_ssnext); 2234 atomic_add_rel_int(&pmc_ss_count, 1); 2235 PMCDBG(PMC,OPS,1, "po=%p in global list", po); 2236 } 2237 po->po_sscount++; 2238 } 2239 2240 /* 2241 * Move to the CPU associated with this 2242 * PMC, and start the hardware. 2243 */ 2244 2245 pmc_save_cpu_binding(&pb); 2246 2247 cpu = PMC_TO_CPU(pm); 2248 2249 if (pmc_cpu_is_disabled(cpu)) 2250 return ENXIO; 2251 2252 pmc_select_cpu(cpu); 2253 2254 /* 2255 * global PMCs are configured at allocation time 2256 * so write out the initial value and start the PMC. 2257 */ 2258 2259 pm->pm_state = PMC_STATE_RUNNING; 2260 2261 critical_enter(); 2262 if ((error = md->pmd_write_pmc(cpu, ri, 2263 PMC_IS_SAMPLING_MODE(mode) ? 2264 pm->pm_sc.pm_reloadcount : 2265 pm->pm_sc.pm_initial)) == 0) 2266 error = md->pmd_start_pmc(cpu, ri); 2267 critical_exit(); 2268 2269 pmc_restore_cpu_binding(&pb); 2270 2271 return error; 2272} 2273 2274/* 2275 * Stop a PMC. 2276 */ 2277 2278static int 2279pmc_stop(struct pmc *pm) 2280{ 2281 int cpu, error, ri; 2282 struct pmc_owner *po; 2283 struct pmc_binding pb; 2284 2285 KASSERT(pm != NULL, ("[pmc,%d] null pmc", __LINE__)); 2286 2287 PMCDBG(PMC,OPS,1, "stop pmc=%p mode=%d ri=%d", pm, 2288 PMC_TO_MODE(pm), PMC_TO_ROWINDEX(pm)); 2289 2290 pm->pm_state = PMC_STATE_STOPPED; 2291 2292 /* 2293 * If the PMC is a virtual mode one, changing the state to 2294 * non-RUNNING is enough to ensure that the PMC never gets 2295 * scheduled. 2296 * 2297 * If this PMC is current running on a CPU, then it will 2298 * handled correctly at the time its target process is context 2299 * switched out. 2300 */ 2301 2302 if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) 2303 return 0; 2304 2305 /* 2306 * A system-mode PMC. Move to the CPU associated with 2307 * this PMC, and stop the hardware. We update the 2308 * 'initial count' so that a subsequent PMCSTART will 2309 * resume counting from the current hardware count. 2310 */ 2311 2312 pmc_save_cpu_binding(&pb); 2313 2314 cpu = PMC_TO_CPU(pm); 2315 2316 KASSERT(cpu >= 0 && cpu < mp_ncpus, 2317 ("[pmc,%d] illegal cpu=%d", __LINE__, cpu)); 2318 2319 if (pmc_cpu_is_disabled(cpu)) 2320 return ENXIO; 2321 2322 pmc_select_cpu(cpu); 2323 2324 ri = PMC_TO_ROWINDEX(pm); 2325 2326 critical_enter(); 2327 if ((error = md->pmd_stop_pmc(cpu, ri)) == 0) 2328 error = md->pmd_read_pmc(cpu, ri, &pm->pm_sc.pm_initial); 2329 critical_exit(); 2330 2331 pmc_restore_cpu_binding(&pb); 2332 2333 po = pm->pm_owner; 2334 2335 /* remove this owner from the global list of SS PMC owners */ 2336 if (PMC_TO_MODE(pm) == PMC_MODE_SS) { 2337 po->po_sscount--; 2338 if (po->po_sscount == 0) { 2339 atomic_subtract_rel_int(&pmc_ss_count, 1); 2340 LIST_REMOVE(po, po_ssnext); 2341 PMCDBG(PMC,OPS,2,"po=%p removed from global list", po); 2342 } 2343 } 2344 2345 return error; 2346} 2347 2348 2349#ifdef DEBUG 2350static const char *pmc_op_to_name[] = { 2351#undef __PMC_OP 2352#define __PMC_OP(N, D) #N , 2353 __PMC_OPS() 2354 NULL 2355}; 2356#endif 2357 2358/* 2359 * The syscall interface 2360 */ 2361 2362#define PMC_GET_SX_XLOCK(...) do { \ 2363 sx_xlock(&pmc_sx); \ 2364 if (pmc_hook == NULL) { \ 2365 sx_xunlock(&pmc_sx); \ 2366 return __VA_ARGS__; \ 2367 } \ 2368} while (0) 2369 2370#define PMC_DOWNGRADE_SX() do { \ 2371 sx_downgrade(&pmc_sx); \ 2372 is_sx_downgraded = 1; \ 2373} while (0) 2374 2375static int 2376pmc_syscall_handler(struct thread *td, void *syscall_args) 2377{ 2378 int error, is_sx_downgraded, op; 2379 struct pmc_syscall_args *c; 2380 void *arg; 2381 2382 PMC_GET_SX_XLOCK(ENOSYS); 2383 2384 DROP_GIANT(); 2385 2386 is_sx_downgraded = 0; 2387 2388 c = (struct pmc_syscall_args *) syscall_args; 2389 2390 op = c->pmop_code; 2391 arg = c->pmop_data; 2392 2393 PMCDBG(MOD,PMS,1, "syscall op=%d \"%s\" arg=%p", op, 2394 pmc_op_to_name[op], arg); 2395 2396 error = 0; 2397 atomic_add_int(&pmc_stats.pm_syscalls, 1); 2398 2399 switch(op) 2400 { 2401 2402 2403 /* 2404 * Configure a log file. 2405 * 2406 * XXX This OP will be reworked. 2407 */ 2408 2409 case PMC_OP_CONFIGURELOG: 2410 { 2411 struct pmc *pm; 2412 struct pmc_owner *po; 2413 struct pmc_op_configurelog cl; 2414 struct proc *p; 2415 2416 sx_assert(&pmc_sx, SX_XLOCKED); 2417 2418 if ((error = copyin(arg, &cl, sizeof(cl))) != 0) 2419 break; 2420 2421 /* mark this process as owning a log file */ 2422 p = td->td_proc; 2423 if ((po = pmc_find_owner_descriptor(p)) == NULL) 2424 if ((po = pmc_allocate_owner_descriptor(p)) == NULL) { 2425 error = ENOMEM; 2426 break; 2427 } 2428 2429 /* 2430 * If a valid fd was passed in, try to configure that, 2431 * otherwise if 'fd' was less than zero and there was 2432 * a log file configured, flush its buffers and 2433 * de-configure it. 2434 */ 2435 if (cl.pm_logfd >= 0) 2436 error = pmclog_configure_log(po, cl.pm_logfd); 2437 else if (po->po_flags & PMC_PO_OWNS_LOGFILE) { 2438 pmclog_process_closelog(po); 2439 error = pmclog_flush(po); 2440 if (error == 0) { 2441 LIST_FOREACH(pm, &po->po_pmcs, pm_next) 2442 if (pm->pm_flags & PMC_F_NEEDS_LOGFILE) 2443 pmc_stop(pm); 2444 error = pmclog_deconfigure_log(po); 2445 } 2446 } else 2447 error = EINVAL; 2448 } 2449 break; 2450 2451 2452 /* 2453 * Flush a log file. 2454 */ 2455 2456 case PMC_OP_FLUSHLOG: 2457 { 2458 struct pmc_owner *po; 2459 2460 sx_assert(&pmc_sx, SX_XLOCKED); 2461 2462 if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) { 2463 error = EINVAL; 2464 break; 2465 } 2466 2467 error = pmclog_flush(po); 2468 } 2469 break; 2470 2471 /* 2472 * Retrieve hardware configuration. 2473 */ 2474 2475 case PMC_OP_GETCPUINFO: /* CPU information */ 2476 { 2477 struct pmc_op_getcpuinfo gci; 2478 2479 gci.pm_cputype = md->pmd_cputype; 2480 gci.pm_ncpu = mp_ncpus; 2481 gci.pm_npmc = md->pmd_npmc; 2482 gci.pm_nclass = md->pmd_nclass; 2483 bcopy(md->pmd_classes, &gci.pm_classes, 2484 sizeof(gci.pm_classes)); 2485 error = copyout(&gci, arg, sizeof(gci)); 2486 } 2487 break; 2488 2489 2490 /* 2491 * Get module statistics 2492 */ 2493 2494 case PMC_OP_GETDRIVERSTATS: 2495 { 2496 struct pmc_op_getdriverstats gms; 2497 2498 bcopy(&pmc_stats, &gms, sizeof(gms)); 2499 error = copyout(&gms, arg, sizeof(gms)); 2500 } 2501 break; 2502 2503 2504 /* 2505 * Retrieve module version number 2506 */ 2507 2508 case PMC_OP_GETMODULEVERSION: 2509 { 2510 uint32_t cv, modv; 2511 2512 /* retrieve the client's idea of the ABI version */ 2513 if ((error = copyin(arg, &cv, sizeof(uint32_t))) != 0) 2514 break; 2515 /* don't service clients newer than our driver */ 2516 modv = PMC_VERSION; 2517 if ((cv & 0xFFFF0000) > (modv & 0xFFFF0000)) { 2518 error = EPROGMISMATCH; 2519 break; 2520 } 2521 error = copyout(&modv, arg, sizeof(int)); 2522 } 2523 break; 2524 2525 2526 /* 2527 * Retrieve the state of all the PMCs on a given 2528 * CPU. 2529 */ 2530 2531 case PMC_OP_GETPMCINFO: 2532 { 2533 uint32_t cpu, n, npmc; 2534 size_t pmcinfo_size; 2535 struct pmc *pm; 2536 struct pmc_info *p, *pmcinfo; 2537 struct pmc_op_getpmcinfo *gpi; 2538 struct pmc_owner *po; 2539 struct pmc_binding pb; 2540 2541 PMC_DOWNGRADE_SX(); 2542 2543 gpi = (struct pmc_op_getpmcinfo *) arg; 2544 2545 if ((error = copyin(&gpi->pm_cpu, &cpu, sizeof(cpu))) != 0) 2546 break; 2547 2548 if (cpu >= (unsigned int) mp_ncpus) { 2549 error = EINVAL; 2550 break; 2551 } 2552 2553 if (pmc_cpu_is_disabled(cpu)) { 2554 error = ENXIO; 2555 break; 2556 } 2557 2558 /* switch to CPU 'cpu' */ 2559 pmc_save_cpu_binding(&pb); 2560 pmc_select_cpu(cpu); 2561 2562 npmc = md->pmd_npmc; 2563 2564 pmcinfo_size = npmc * sizeof(struct pmc_info); 2565 MALLOC(pmcinfo, struct pmc_info *, pmcinfo_size, M_PMC, 2566 M_WAITOK); 2567 2568 p = pmcinfo; 2569 2570 for (n = 0; n < md->pmd_npmc; n++, p++) { 2571 2572 if ((error = md->pmd_describe(cpu, n, p, &pm)) != 0) 2573 break; 2574 2575 if (PMC_ROW_DISP_IS_STANDALONE(n)) 2576 p->pm_rowdisp = PMC_DISP_STANDALONE; 2577 else if (PMC_ROW_DISP_IS_THREAD(n)) 2578 p->pm_rowdisp = PMC_DISP_THREAD; 2579 else 2580 p->pm_rowdisp = PMC_DISP_FREE; 2581 2582 p->pm_ownerpid = -1; 2583 2584 if (pm == NULL) /* no PMC associated */ 2585 continue; 2586 2587 po = pm->pm_owner; 2588 2589 KASSERT(po->po_owner != NULL, 2590 ("[pmc,%d] pmc_owner had a null proc pointer", 2591 __LINE__)); 2592 2593 p->pm_ownerpid = po->po_owner->p_pid; 2594 p->pm_mode = PMC_TO_MODE(pm); 2595 p->pm_event = pm->pm_event; 2596 p->pm_flags = pm->pm_flags; 2597 2598 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) 2599 p->pm_reloadcount = 2600 pm->pm_sc.pm_reloadcount; 2601 } 2602 2603 pmc_restore_cpu_binding(&pb); 2604 2605 /* now copy out the PMC info collected */ 2606 if (error == 0) 2607 error = copyout(pmcinfo, &gpi->pm_pmcs, pmcinfo_size); 2608 2609 FREE(pmcinfo, M_PMC); 2610 } 2611 break; 2612 2613 2614 /* 2615 * Set the administrative state of a PMC. I.e. whether 2616 * the PMC is to be used or not. 2617 */ 2618 2619 case PMC_OP_PMCADMIN: 2620 { 2621 int cpu, ri; 2622 enum pmc_state request; 2623 struct pmc_cpu *pc; 2624 struct pmc_hw *phw; 2625 struct pmc_op_pmcadmin pma; 2626 struct pmc_binding pb; 2627 2628 sx_assert(&pmc_sx, SX_XLOCKED); 2629 2630 KASSERT(td == curthread, 2631 ("[pmc,%d] td != curthread", __LINE__)); 2632 2633 if (suser(td) || jailed(td->td_ucred)) { 2634 error = EPERM; 2635 break; 2636 } 2637 2638 if ((error = copyin(arg, &pma, sizeof(pma))) != 0) 2639 break; 2640 2641 cpu = pma.pm_cpu; 2642 2643 if (cpu < 0 || cpu >= mp_ncpus) { 2644 error = EINVAL; 2645 break; 2646 } 2647 2648 if (pmc_cpu_is_disabled(cpu)) { 2649 error = ENXIO; 2650 break; 2651 } 2652 2653 request = pma.pm_state; 2654 2655 if (request != PMC_STATE_DISABLED && 2656 request != PMC_STATE_FREE) { 2657 error = EINVAL; 2658 break; 2659 } 2660 2661 ri = pma.pm_pmc; /* pmc id == row index */ 2662 if (ri < 0 || ri >= (int) md->pmd_npmc) { 2663 error = EINVAL; 2664 break; 2665 } 2666 2667 /* 2668 * We can't disable a PMC with a row-index allocated 2669 * for process virtual PMCs. 2670 */ 2671 2672 if (PMC_ROW_DISP_IS_THREAD(ri) && 2673 request == PMC_STATE_DISABLED) { 2674 error = EBUSY; 2675 break; 2676 } 2677 2678 /* 2679 * otherwise, this PMC on this CPU is either free or 2680 * in system-wide mode. 2681 */ 2682 2683 pmc_save_cpu_binding(&pb); 2684 pmc_select_cpu(cpu); 2685 2686 pc = pmc_pcpu[cpu]; 2687 phw = pc->pc_hwpmcs[ri]; 2688 2689 /* 2690 * XXX do we need some kind of 'forced' disable? 2691 */ 2692 2693 if (phw->phw_pmc == NULL) { 2694 if (request == PMC_STATE_DISABLED && 2695 (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED)) { 2696 phw->phw_state &= ~PMC_PHW_FLAG_IS_ENABLED; 2697 PMC_MARK_ROW_STANDALONE(ri); 2698 } else if (request == PMC_STATE_FREE && 2699 (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0) { 2700 phw->phw_state |= PMC_PHW_FLAG_IS_ENABLED; 2701 PMC_UNMARK_ROW_STANDALONE(ri); 2702 } 2703 /* other cases are a no-op */ 2704 } else 2705 error = EBUSY; 2706 2707 pmc_restore_cpu_binding(&pb); 2708 } 2709 break; 2710 2711 2712 /* 2713 * Allocate a PMC. 2714 */ 2715 2716 case PMC_OP_PMCALLOCATE: 2717 { 2718 uint32_t caps; 2719 u_int cpu; 2720 int n; 2721 enum pmc_mode mode; 2722 struct pmc *pmc; 2723 struct pmc_hw *phw; 2724 struct pmc_op_pmcallocate pa; 2725 struct pmc_binding pb; 2726 2727 if ((error = copyin(arg, &pa, sizeof(pa))) != 0) 2728 break; 2729 2730 caps = pa.pm_caps; 2731 mode = pa.pm_mode; 2732 cpu = pa.pm_cpu; 2733 2734 if ((mode != PMC_MODE_SS && mode != PMC_MODE_SC && 2735 mode != PMC_MODE_TS && mode != PMC_MODE_TC) || 2736 (cpu != (u_int) PMC_CPU_ANY && cpu >= (u_int) mp_ncpus)) { 2737 error = EINVAL; 2738 break; 2739 } 2740 2741 /* 2742 * Virtual PMCs should only ask for a default CPU. 2743 * System mode PMCs need to specify a non-default CPU. 2744 */ 2745 2746 if ((PMC_IS_VIRTUAL_MODE(mode) && cpu != (u_int) PMC_CPU_ANY) || 2747 (PMC_IS_SYSTEM_MODE(mode) && cpu == (u_int) PMC_CPU_ANY)) { 2748 error = EINVAL; 2749 break; 2750 } 2751 2752 /* 2753 * Check that a disabled CPU is not being asked for. 2754 */ 2755 2756 if (PMC_IS_SYSTEM_MODE(mode) && pmc_cpu_is_disabled(cpu)) { 2757 error = ENXIO; 2758 break; 2759 } 2760 2761 /* 2762 * Refuse an allocation for a system-wide PMC if this 2763 * process has been jailed, or if this process lacks 2764 * super-user credentials and the sysctl tunable 2765 * 'security.bsd.unprivileged_syspmcs' is zero. 2766 */ 2767 2768 if (PMC_IS_SYSTEM_MODE(mode)) { 2769 if (jailed(curthread->td_ucred)) 2770 error = EPERM; 2771 else if (suser(curthread) && 2772 (pmc_unprivileged_syspmcs == 0)) 2773 error = EPERM; 2774 } 2775 2776 if (error) 2777 break; 2778 2779 /* 2780 * Look for valid values for 'pm_flags' 2781 */ 2782 2783 if ((pa.pm_flags & ~(PMC_F_DESCENDANTS | PMC_F_LOG_PROCCSW | 2784 PMC_F_LOG_PROCEXIT)) != 0) { 2785 error = EINVAL; 2786 break; 2787 } 2788 2789 /* process logging options are not allowed for system PMCs */ 2790 if (PMC_IS_SYSTEM_MODE(mode) && (pa.pm_flags & 2791 (PMC_F_LOG_PROCCSW | PMC_F_LOG_PROCEXIT))) { 2792 error = EINVAL; 2793 break; 2794 } 2795 2796 /* 2797 * All sampling mode PMCs need to be able to interrupt the 2798 * CPU. 2799 */ 2800 if (PMC_IS_SAMPLING_MODE(mode)) 2801 caps |= PMC_CAP_INTERRUPT; 2802 2803 /* A valid class specifier should have been passed in. */ 2804 for (n = 0; n < md->pmd_nclass; n++) 2805 if (md->pmd_classes[n].pm_class == pa.pm_class) 2806 break; 2807 if (n == md->pmd_nclass) { 2808 error = EINVAL; 2809 break; 2810 } 2811 2812 /* The requested PMC capabilities should be feasible. */ 2813 if ((md->pmd_classes[n].pm_caps & caps) != caps) { 2814 error = EOPNOTSUPP; 2815 break; 2816 } 2817 2818 PMCDBG(PMC,ALL,2, "event=%d caps=0x%x mode=%d cpu=%d", 2819 pa.pm_ev, caps, mode, cpu); 2820 2821 pmc = pmc_allocate_pmc_descriptor(); 2822 pmc->pm_id = PMC_ID_MAKE_ID(cpu,pa.pm_mode,pa.pm_class, 2823 PMC_ID_INVALID); 2824 pmc->pm_event = pa.pm_ev; 2825 pmc->pm_state = PMC_STATE_FREE; 2826 pmc->pm_caps = caps; 2827 pmc->pm_flags = pa.pm_flags; 2828 2829 /* switch thread to CPU 'cpu' */ 2830 pmc_save_cpu_binding(&pb); 2831 2832#define PMC_IS_SHAREABLE_PMC(cpu, n) \ 2833 (pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_state & \ 2834 PMC_PHW_FLAG_IS_SHAREABLE) 2835#define PMC_IS_UNALLOCATED(cpu, n) \ 2836 (pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_pmc == NULL) 2837 2838 if (PMC_IS_SYSTEM_MODE(mode)) { 2839 pmc_select_cpu(cpu); 2840 for (n = 0; n < (int) md->pmd_npmc; n++) 2841 if (pmc_can_allocate_row(n, mode) == 0 && 2842 pmc_can_allocate_rowindex( 2843 curthread->td_proc, n, cpu) == 0 && 2844 (PMC_IS_UNALLOCATED(cpu, n) || 2845 PMC_IS_SHAREABLE_PMC(cpu, n)) && 2846 md->pmd_allocate_pmc(cpu, n, pmc, 2847 &pa) == 0) 2848 break; 2849 } else { 2850 /* Process virtual mode */ 2851 for (n = 0; n < (int) md->pmd_npmc; n++) { 2852 if (pmc_can_allocate_row(n, mode) == 0 && 2853 pmc_can_allocate_rowindex( 2854 curthread->td_proc, n, 2855 PMC_CPU_ANY) == 0 && 2856 md->pmd_allocate_pmc(curthread->td_oncpu, 2857 n, pmc, &pa) == 0) 2858 break; 2859 } 2860 } 2861 2862#undef PMC_IS_UNALLOCATED 2863#undef PMC_IS_SHAREABLE_PMC 2864 2865 pmc_restore_cpu_binding(&pb); 2866 2867 if (n == (int) md->pmd_npmc) { 2868 pmc_destroy_pmc_descriptor(pmc); 2869 FREE(pmc, M_PMC); 2870 pmc = NULL; 2871 error = EINVAL; 2872 break; 2873 } 2874 2875 /* Fill in the correct value in the ID field */ 2876 pmc->pm_id = PMC_ID_MAKE_ID(cpu,mode,pa.pm_class,n); 2877 2878 PMCDBG(PMC,ALL,2, "ev=%d class=%d mode=%d n=%d -> pmcid=%x", 2879 pmc->pm_event, pa.pm_class, mode, n, pmc->pm_id); 2880 2881 /* Process mode PMCs with logging enabled need log files */ 2882 if (pmc->pm_flags & (PMC_F_LOG_PROCEXIT | PMC_F_LOG_PROCCSW)) 2883 pmc->pm_flags |= PMC_F_NEEDS_LOGFILE; 2884 2885 /* All system mode sampling PMCs require a log file */ 2886 if (PMC_IS_SAMPLING_MODE(mode) && PMC_IS_SYSTEM_MODE(mode)) 2887 pmc->pm_flags |= PMC_F_NEEDS_LOGFILE; 2888 2889 /* 2890 * Configure global pmc's immediately 2891 */ 2892 2893 if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pmc))) { 2894 2895 pmc_save_cpu_binding(&pb); 2896 pmc_select_cpu(cpu); 2897 2898 phw = pmc_pcpu[cpu]->pc_hwpmcs[n]; 2899 2900 if ((phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0 || 2901 (error = md->pmd_config_pmc(cpu, n, pmc)) != 0) { 2902 (void) md->pmd_release_pmc(cpu, n, pmc); 2903 pmc_destroy_pmc_descriptor(pmc); 2904 FREE(pmc, M_PMC); 2905 pmc = NULL; 2906 pmc_restore_cpu_binding(&pb); 2907 error = EPERM; 2908 break; 2909 } 2910 2911 pmc_restore_cpu_binding(&pb); 2912 } 2913 2914 pmc->pm_state = PMC_STATE_ALLOCATED; 2915 2916 /* 2917 * mark row disposition 2918 */ 2919 2920 if (PMC_IS_SYSTEM_MODE(mode)) 2921 PMC_MARK_ROW_STANDALONE(n); 2922 else 2923 PMC_MARK_ROW_THREAD(n); 2924 2925 /* 2926 * Register this PMC with the current thread as its owner. 2927 */ 2928 2929 if ((error = 2930 pmc_register_owner(curthread->td_proc, pmc)) != 0) { 2931 pmc_release_pmc_descriptor(pmc); 2932 FREE(pmc, M_PMC); 2933 pmc = NULL; 2934 break; 2935 } 2936 2937 /* 2938 * Return the allocated index. 2939 */ 2940 2941 pa.pm_pmcid = pmc->pm_id; 2942 2943 error = copyout(&pa, arg, sizeof(pa)); 2944 } 2945 break; 2946 2947 2948 /* 2949 * Attach a PMC to a process. 2950 */ 2951 2952 case PMC_OP_PMCATTACH: 2953 { 2954 struct pmc *pm; 2955 struct proc *p; 2956 struct pmc_op_pmcattach a; 2957 2958 sx_assert(&pmc_sx, SX_XLOCKED); 2959 2960 if ((error = copyin(arg, &a, sizeof(a))) != 0) 2961 break; 2962 2963 if (a.pm_pid < 0) { 2964 error = EINVAL; 2965 break; 2966 } else if (a.pm_pid == 0) 2967 a.pm_pid = td->td_proc->p_pid; 2968 2969 if ((error = pmc_find_pmc(a.pm_pmc, &pm)) != 0) 2970 break; 2971 2972 if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm))) { 2973 error = EINVAL; 2974 break; 2975 } 2976 2977 /* PMCs may be (re)attached only when allocated or stopped */ 2978 if (pm->pm_state == PMC_STATE_RUNNING) { 2979 error = EBUSY; 2980 break; 2981 } else if (pm->pm_state != PMC_STATE_ALLOCATED && 2982 pm->pm_state != PMC_STATE_STOPPED) { 2983 error = EINVAL; 2984 break; 2985 } 2986 2987 /* lookup pid */ 2988 if ((p = pfind(a.pm_pid)) == NULL) { 2989 error = ESRCH; 2990 break; 2991 } 2992 2993 /* 2994 * Ignore processes that are working on exiting. 2995 */ 2996 if (p->p_flag & P_WEXIT) { 2997 error = ESRCH; 2998 PROC_UNLOCK(p); /* pfind() returns a locked process */ 2999 break; 3000 } 3001 3002 /* 3003 * we are allowed to attach a PMC to a process if 3004 * we can debug it. 3005 */ 3006 error = p_candebug(curthread, p); 3007 3008 PROC_UNLOCK(p); 3009 3010 if (error == 0) 3011 error = pmc_attach_process(p, pm); 3012 } 3013 break; 3014 3015 3016 /* 3017 * Detach an attached PMC from a process. 3018 */ 3019 3020 case PMC_OP_PMCDETACH: 3021 { 3022 struct pmc *pm; 3023 struct proc *p; 3024 struct pmc_op_pmcattach a; 3025 3026 if ((error = copyin(arg, &a, sizeof(a))) != 0) 3027 break; 3028 3029 if (a.pm_pid < 0) { 3030 error = EINVAL; 3031 break; 3032 } else if (a.pm_pid == 0) 3033 a.pm_pid = td->td_proc->p_pid; 3034 3035 if ((error = pmc_find_pmc(a.pm_pmc, &pm)) != 0) 3036 break; 3037 3038 if ((p = pfind(a.pm_pid)) == NULL) { 3039 error = ESRCH; 3040 break; 3041 } 3042 3043 /* 3044 * Treat processes that are in the process of exiting 3045 * as if they were not present. 3046 */ 3047 3048 if (p->p_flag & P_WEXIT) 3049 error = ESRCH; 3050 3051 PROC_UNLOCK(p); /* pfind() returns a locked process */ 3052 3053 if (error == 0) 3054 error = pmc_detach_process(p, pm); 3055 } 3056 break; 3057 3058 3059 /* 3060 * Retrieve the MSR number associated with the counter 3061 * 'pmc_id'. This allows processes to directly use RDPMC 3062 * instructions to read their PMCs, without the overhead of a 3063 * system call. 3064 */ 3065 3066 case PMC_OP_PMCGETMSR: 3067 { 3068 int ri; 3069 struct pmc *pm; 3070 struct pmc_target *pt; 3071 struct pmc_op_getmsr gm; 3072 3073 PMC_DOWNGRADE_SX(); 3074 3075 /* CPU has no 'GETMSR' support */ 3076 if (md->pmd_get_msr == NULL) { 3077 error = ENOSYS; 3078 break; 3079 } 3080 3081 if ((error = copyin(arg, &gm, sizeof(gm))) != 0) 3082 break; 3083 3084 if ((error = pmc_find_pmc(gm.pm_pmcid, &pm)) != 0) 3085 break; 3086 3087 /* 3088 * The allocated PMC has to be a process virtual PMC, 3089 * i.e., of type MODE_T[CS]. Global PMCs can only be 3090 * read using the PMCREAD operation since they may be 3091 * allocated on a different CPU than the one we could 3092 * be running on at the time of the RDPMC instruction. 3093 * 3094 * The GETMSR operation is not allowed for PMCs that 3095 * are inherited across processes. 3096 */ 3097 3098 if (!PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)) || 3099 (pm->pm_flags & PMC_F_DESCENDANTS)) { 3100 error = EINVAL; 3101 break; 3102 } 3103 3104 /* 3105 * It only makes sense to use a RDPMC (or its 3106 * equivalent instruction on non-x86 architectures) on 3107 * a process that has allocated and attached a PMC to 3108 * itself. Conversely the PMC is only allowed to have 3109 * one process attached to it -- its owner. 3110 */ 3111 3112 if ((pt = LIST_FIRST(&pm->pm_targets)) == NULL || 3113 LIST_NEXT(pt, pt_next) != NULL || 3114 pt->pt_process->pp_proc != pm->pm_owner->po_owner) { 3115 error = EINVAL; 3116 break; 3117 } 3118 3119 ri = PMC_TO_ROWINDEX(pm); 3120 3121 if ((error = (*md->pmd_get_msr)(ri, &gm.pm_msr)) < 0) 3122 break; 3123 3124 if ((error = copyout(&gm, arg, sizeof(gm))) < 0) 3125 break; 3126 3127 /* 3128 * Mark our process as using MSRs. Update machine 3129 * state using a forced context switch. 3130 */ 3131 3132 pt->pt_process->pp_flags |= PMC_PP_ENABLE_MSR_ACCESS; 3133 pmc_force_context_switch(); 3134 3135 } 3136 break; 3137 3138 /* 3139 * Release an allocated PMC 3140 */ 3141 3142 case PMC_OP_PMCRELEASE: 3143 { 3144 pmc_id_t pmcid; 3145 struct pmc *pm; 3146 struct pmc_owner *po; 3147 struct pmc_op_simple sp; 3148 3149 /* 3150 * Find PMC pointer for the named PMC. 3151 * 3152 * Use pmc_release_pmc_descriptor() to switch off the 3153 * PMC, remove all its target threads, and remove the 3154 * PMC from its owner's list. 3155 * 3156 * Remove the owner record if this is the last PMC 3157 * owned. 3158 * 3159 * Free up space. 3160 */ 3161 3162 if ((error = copyin(arg, &sp, sizeof(sp))) != 0) 3163 break; 3164 3165 pmcid = sp.pm_pmcid; 3166 3167 if ((error = pmc_find_pmc(pmcid, &pm)) != 0) 3168 break; 3169 3170 po = pm->pm_owner; 3171 pmc_release_pmc_descriptor(pm); 3172 pmc_maybe_remove_owner(po); 3173 3174 FREE(pm, M_PMC); 3175 } 3176 break; 3177 3178 3179 /* 3180 * Read and/or write a PMC. 3181 */ 3182 3183 case PMC_OP_PMCRW: 3184 { 3185 uint32_t cpu, ri; 3186 struct pmc *pm; 3187 struct pmc_op_pmcrw *pprw; 3188 struct pmc_op_pmcrw prw; 3189 struct pmc_binding pb; 3190 pmc_value_t oldvalue; 3191 3192 PMC_DOWNGRADE_SX(); 3193 3194 if ((error = copyin(arg, &prw, sizeof(prw))) != 0) 3195 break; 3196 3197 ri = 0; 3198 PMCDBG(PMC,OPS,1, "rw id=%d flags=0x%x", prw.pm_pmcid, 3199 prw.pm_flags); 3200 3201 /* must have at least one flag set */ 3202 if ((prw.pm_flags & (PMC_F_OLDVALUE|PMC_F_NEWVALUE)) == 0) { 3203 error = EINVAL; 3204 break; 3205 } 3206 3207 /* locate pmc descriptor */ 3208 if ((error = pmc_find_pmc(prw.pm_pmcid, &pm)) != 0) 3209 break; 3210 3211 /* Can't read a PMC that hasn't been started. */ 3212 if (pm->pm_state != PMC_STATE_ALLOCATED && 3213 pm->pm_state != PMC_STATE_STOPPED && 3214 pm->pm_state != PMC_STATE_RUNNING) { 3215 error = EINVAL; 3216 break; 3217 } 3218 3219 /* writing a new value is allowed only for 'STOPPED' pmcs */ 3220 if (pm->pm_state == PMC_STATE_RUNNING && 3221 (prw.pm_flags & PMC_F_NEWVALUE)) { 3222 error = EBUSY; 3223 break; 3224 } 3225 3226 if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) { 3227 3228 /* 3229 * If this PMC is attached to its owner (i.e., 3230 * the process requesting this operation) and 3231 * is running, then attempt to get an 3232 * upto-date reading from hardware for a READ. 3233 * Writes are only allowed when the PMC is 3234 * stopped, so only update the saved value 3235 * field. 3236 * 3237 * If the PMC is not running, or is not 3238 * attached to its owner, read/write to the 3239 * savedvalue field. 3240 */ 3241 3242 ri = PMC_TO_ROWINDEX(pm); 3243 3244 mtx_pool_lock_spin(pmc_mtxpool, pm); 3245 cpu = curthread->td_oncpu; 3246 3247 if (prw.pm_flags & PMC_F_OLDVALUE) { 3248 if ((pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) && 3249 (pm->pm_state == PMC_STATE_RUNNING)) 3250 error = (*md->pmd_read_pmc)(cpu, ri, 3251 &oldvalue); 3252 else 3253 oldvalue = pm->pm_gv.pm_savedvalue; 3254 } 3255 if (prw.pm_flags & PMC_F_NEWVALUE) 3256 pm->pm_gv.pm_savedvalue = prw.pm_value; 3257 3258 mtx_pool_unlock_spin(pmc_mtxpool, pm); 3259 3260 } else { /* System mode PMCs */ 3261 cpu = PMC_TO_CPU(pm); 3262 ri = PMC_TO_ROWINDEX(pm); 3263 3264 if (pmc_cpu_is_disabled(cpu)) { 3265 error = ENXIO; 3266 break; 3267 } 3268 3269 /* move this thread to CPU 'cpu' */ 3270 pmc_save_cpu_binding(&pb); 3271 pmc_select_cpu(cpu); 3272 3273 critical_enter(); 3274 /* save old value */ 3275 if (prw.pm_flags & PMC_F_OLDVALUE) 3276 if ((error = (*md->pmd_read_pmc)(cpu, ri, 3277 &oldvalue))) 3278 goto error; 3279 /* write out new value */ 3280 if (prw.pm_flags & PMC_F_NEWVALUE) 3281 error = (*md->pmd_write_pmc)(cpu, ri, 3282 prw.pm_value); 3283 error: 3284 critical_exit(); 3285 pmc_restore_cpu_binding(&pb); 3286 if (error) 3287 break; 3288 } 3289 3290 pprw = (struct pmc_op_pmcrw *) arg; 3291 3292#ifdef DEBUG 3293 if (prw.pm_flags & PMC_F_NEWVALUE) 3294 PMCDBG(PMC,OPS,2, "rw id=%d new %jx -> old %jx", 3295 ri, prw.pm_value, oldvalue); 3296 else if (prw.pm_flags & PMC_F_OLDVALUE) 3297 PMCDBG(PMC,OPS,2, "rw id=%d -> old %jx", ri, oldvalue); 3298#endif 3299 3300 /* return old value if requested */ 3301 if (prw.pm_flags & PMC_F_OLDVALUE) 3302 if ((error = copyout(&oldvalue, &pprw->pm_value, 3303 sizeof(prw.pm_value)))) 3304 break; 3305 3306 } 3307 break; 3308 3309 3310 /* 3311 * Set the sampling rate for a sampling mode PMC and the 3312 * initial count for a counting mode PMC. 3313 */ 3314 3315 case PMC_OP_PMCSETCOUNT: 3316 { 3317 struct pmc *pm; 3318 struct pmc_op_pmcsetcount sc; 3319 3320 PMC_DOWNGRADE_SX(); 3321 3322 if ((error = copyin(arg, &sc, sizeof(sc))) != 0) 3323 break; 3324 3325 if ((error = pmc_find_pmc(sc.pm_pmcid, &pm)) != 0) 3326 break; 3327 3328 if (pm->pm_state == PMC_STATE_RUNNING) { 3329 error = EBUSY; 3330 break; 3331 } 3332 3333 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) 3334 pm->pm_sc.pm_reloadcount = sc.pm_count; 3335 else 3336 pm->pm_sc.pm_initial = sc.pm_count; 3337 } 3338 break; 3339 3340 3341 /* 3342 * Start a PMC. 3343 */ 3344 3345 case PMC_OP_PMCSTART: 3346 { 3347 pmc_id_t pmcid; 3348 struct pmc *pm; 3349 struct pmc_op_simple sp; 3350 3351 sx_assert(&pmc_sx, SX_XLOCKED); 3352 3353 if ((error = copyin(arg, &sp, sizeof(sp))) != 0) 3354 break; 3355 3356 pmcid = sp.pm_pmcid; 3357 3358 if ((error = pmc_find_pmc(pmcid, &pm)) != 0) 3359 break; 3360 3361 KASSERT(pmcid == pm->pm_id, 3362 ("[pmc,%d] pmcid %x != id %x", __LINE__, 3363 pm->pm_id, pmcid)); 3364 3365 if (pm->pm_state == PMC_STATE_RUNNING) /* already running */ 3366 break; 3367 else if (pm->pm_state != PMC_STATE_STOPPED && 3368 pm->pm_state != PMC_STATE_ALLOCATED) { 3369 error = EINVAL; 3370 break; 3371 } 3372 3373 error = pmc_start(pm); 3374 } 3375 break; 3376 3377 3378 /* 3379 * Stop a PMC. 3380 */ 3381 3382 case PMC_OP_PMCSTOP: 3383 { 3384 pmc_id_t pmcid; 3385 struct pmc *pm; 3386 struct pmc_op_simple sp; 3387 3388 PMC_DOWNGRADE_SX(); 3389 3390 if ((error = copyin(arg, &sp, sizeof(sp))) != 0) 3391 break; 3392 3393 pmcid = sp.pm_pmcid; 3394 3395 /* 3396 * Mark the PMC as inactive and invoke the MD stop 3397 * routines if needed. 3398 */ 3399 3400 if ((error = pmc_find_pmc(pmcid, &pm)) != 0) 3401 break; 3402 3403 KASSERT(pmcid == pm->pm_id, 3404 ("[pmc,%d] pmc id %x != pmcid %x", __LINE__, 3405 pm->pm_id, pmcid)); 3406 3407 if (pm->pm_state == PMC_STATE_STOPPED) /* already stopped */ 3408 break; 3409 else if (pm->pm_state != PMC_STATE_RUNNING) { 3410 error = EINVAL; 3411 break; 3412 } 3413 3414 error = pmc_stop(pm); 3415 } 3416 break; 3417 3418 3419 /* 3420 * Write a user supplied value to the log file. 3421 */ 3422 3423 case PMC_OP_WRITELOG: 3424 { 3425 struct pmc_op_writelog wl; 3426 struct pmc_owner *po; 3427 3428 PMC_DOWNGRADE_SX(); 3429 3430 if ((error = copyin(arg, &wl, sizeof(wl))) != 0) 3431 break; 3432 3433 if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) { 3434 error = EINVAL; 3435 break; 3436 } 3437 3438 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) { 3439 error = EINVAL; 3440 break; 3441 } 3442 3443 error = pmclog_process_userlog(po, &wl); 3444 } 3445 break; 3446 3447 3448 default: 3449 error = EINVAL; 3450 break; 3451 } 3452 3453 if (is_sx_downgraded) 3454 sx_sunlock(&pmc_sx); 3455 else 3456 sx_xunlock(&pmc_sx); 3457 3458 if (error) 3459 atomic_add_int(&pmc_stats.pm_syscall_errors, 1); 3460 3461 PICKUP_GIANT(); 3462 3463 return error; 3464} 3465 3466/* 3467 * Helper functions 3468 */ 3469 3470 3471/* 3472 * Interrupt processing. 3473 * 3474 * Find a free slot in the per-cpu array of PC samples and write the 3475 * current (PMC,PID,PC) triple to it. If an event was successfully 3476 * added, a bit is set in mask 'pmc_cpumask' denoting that the 3477 * DO_SAMPLES hook needs to be invoked from the clock handler. 3478 * 3479 * This function is meant to be called from an NMI handler. It cannot 3480 * use any of the locking primitives supplied by the OS. 3481 */ 3482 3483int 3484pmc_process_interrupt(int cpu, struct pmc *pm, uintfptr_t pc, int usermode) 3485{ 3486 int error, ri; 3487 struct thread *td; 3488 struct pmc_sample *ps; 3489 struct pmc_samplebuffer *psb; 3490 3491 error = 0; 3492 ri = PMC_TO_ROWINDEX(pm); 3493 3494 psb = pmc_pcpu[cpu]->pc_sb; 3495 3496 ps = psb->ps_write; 3497 if (ps->ps_pc) { /* in use, reader hasn't caught up */ 3498 pm->pm_stalled = 1; 3499 atomic_add_int(&pmc_stats.pm_intr_bufferfull, 1); 3500 PMCDBG(SAM,INT,1,"(spc) cpu=%d pm=%p pc=%jx um=%d wr=%d rd=%d", 3501 cpu, pm, (uint64_t) pc, usermode, 3502 (int) (psb->ps_write - psb->ps_samples), 3503 (int) (psb->ps_read - psb->ps_samples)); 3504 error = ENOMEM; 3505 goto done; 3506 } 3507 3508 /* fill in entry */ 3509 PMCDBG(SAM,INT,1,"cpu=%d pm=%p pc=%jx um=%d wr=%d rd=%d", cpu, pm, 3510 (uint64_t) pc, usermode, 3511 (int) (psb->ps_write - psb->ps_samples), 3512 (int) (psb->ps_read - psb->ps_samples)); 3513 3514 atomic_add_rel_32(&pm->pm_runcount, 1); /* hold onto PMC */ 3515 ps->ps_pmc = pm; 3516 if ((td = curthread) && td->td_proc) 3517 ps->ps_pid = td->td_proc->p_pid; 3518 else 3519 ps->ps_pid = -1; 3520 ps->ps_usermode = usermode; 3521 ps->ps_pc = pc; /* mark entry as in use */ 3522 3523 /* increment write pointer, modulo ring buffer size */ 3524 ps++; 3525 if (ps == psb->ps_fence) 3526 psb->ps_write = psb->ps_samples; 3527 else 3528 psb->ps_write = ps; 3529 3530 done: 3531 /* mark CPU as needing processing */ 3532 atomic_set_rel_int(&pmc_cpumask, (1 << cpu)); 3533 3534 return error; 3535} 3536 3537 3538/* 3539 * Process saved PC samples. 3540 */ 3541 3542static void 3543pmc_process_samples(int cpu) 3544{ 3545 int n, ri; 3546 struct pmc *pm; 3547 struct thread *td; 3548 struct pmc_owner *po; 3549 struct pmc_sample *ps; 3550 struct pmc_samplebuffer *psb; 3551 3552 KASSERT(PCPU_GET(cpuid) == cpu, 3553 ("[pmc,%d] not on the correct CPU pcpu=%d cpu=%d", __LINE__, 3554 PCPU_GET(cpuid), cpu)); 3555 3556 psb = pmc_pcpu[cpu]->pc_sb; 3557 3558 for (n = 0; n < pmc_nsamples; n++) { /* bound on #iterations */ 3559 3560 ps = psb->ps_read; 3561 if (ps->ps_pc == (uintfptr_t) 0) /* no data */ 3562 break; 3563 3564 pm = ps->ps_pmc; 3565 po = pm->pm_owner; 3566 3567 KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)), 3568 ("[pmc,%d] pmc=%p non-sampling mode=%d", __LINE__, 3569 pm, PMC_TO_MODE(pm))); 3570 3571 /* Ignore PMCs that have been switched off */ 3572 if (pm->pm_state != PMC_STATE_RUNNING) 3573 goto entrydone; 3574 3575 PMCDBG(SAM,OPS,1,"cpu=%d pm=%p pc=%jx um=%d wr=%d rd=%d", cpu, 3576 pm, (uint64_t) ps->ps_pc, ps->ps_usermode, 3577 (int) (psb->ps_write - psb->ps_samples), 3578 (int) (psb->ps_read - psb->ps_samples)); 3579 3580 /* 3581 * If this is a process-mode PMC that is attached to 3582 * its owner, and if the PC is in user mode, update 3583 * profiling statistics like timer-based profiling 3584 * would have done. 3585 */ 3586 if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) { 3587 if (ps->ps_usermode) { 3588 td = FIRST_THREAD_IN_PROC(po->po_owner); 3589 addupc_intr(td, ps->ps_pc, 1); 3590 } 3591 goto entrydone; 3592 } 3593 3594 /* 3595 * Otherwise, this is either a sampling mode PMC that 3596 * is attached to a different process than its owner, 3597 * or a system-wide sampling PMC. Dispatch a log 3598 * entry to the PMC's owner process. 3599 */ 3600 3601 pmclog_process_pcsample(pm, ps); 3602 3603 entrydone: 3604 ps->ps_pc = (uintfptr_t) 0; /* mark entry as free */ 3605 atomic_subtract_rel_32(&pm->pm_runcount, 1); 3606 3607 /* increment read pointer, modulo sample size */ 3608 if (++ps == psb->ps_fence) 3609 psb->ps_read = psb->ps_samples; 3610 else 3611 psb->ps_read = ps; 3612 } 3613 3614 atomic_add_int(&pmc_stats.pm_log_sweeps, 1); 3615 3616 /* Do not re-enable stalled PMCs if we failed to process any samples */ 3617 if (n == 0) 3618 return; 3619 3620 /* 3621 * Restart any stalled sampling PMCs on this CPU. 3622 * 3623 * If the NMI handler sets the pm_stalled field of a PMC after 3624 * the check below, we'll end up processing the stalled PMC at 3625 * the next hardclock tick. 3626 */ 3627 for (n = 0; n < md->pmd_npmc; n++) { 3628 (void) (*md->pmd_get_config)(cpu,n,&pm); 3629 if (pm == NULL || /* !cfg'ed */ 3630 pm->pm_state != PMC_STATE_RUNNING || /* !active */ 3631 !PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) || /* !sampling */ 3632 pm->pm_stalled == 0) /* !stalled */ 3633 continue; 3634 3635 pm->pm_stalled = 0; 3636 ri = PMC_TO_ROWINDEX(pm); 3637 (*md->pmd_start_pmc)(cpu, ri); 3638 } 3639} 3640 3641/* 3642 * Event handlers. 3643 */ 3644 3645/* 3646 * Handle a process exit. 3647 * 3648 * Remove this process from all hash tables. If this process 3649 * owned any PMCs, turn off those PMCs and deallocate them, 3650 * removing any associations with target processes. 3651 * 3652 * This function will be called by the last 'thread' of a 3653 * process. 3654 * 3655 * XXX This eventhandler gets called early in the exit process. 3656 * Consider using a 'hook' invocation from thread_exit() or equivalent 3657 * spot. Another negative is that kse_exit doesn't seem to call 3658 * exit1() [??]. 3659 * 3660 */ 3661 3662static void 3663pmc_process_exit(void *arg __unused, struct proc *p) 3664{ 3665 int is_using_hwpmcs; 3666 int cpu; 3667 unsigned int ri; 3668 struct pmc *pm; 3669 struct pmc_process *pp; 3670 struct pmc_owner *po; 3671 pmc_value_t newvalue, tmp; 3672 3673 PROC_LOCK(p); 3674 is_using_hwpmcs = p->p_flag & P_HWPMC; 3675 PROC_UNLOCK(p); 3676 3677 /* 3678 * Log a sysexit event to all SS PMC owners. 3679 */ 3680 LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) 3681 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 3682 pmclog_process_sysexit(po, p->p_pid); 3683 3684 if (!is_using_hwpmcs) 3685 return; 3686 3687 PMC_GET_SX_XLOCK(); 3688 PMCDBG(PRC,EXT,1,"process-exit proc=%p (%d, %s)", p, p->p_pid, 3689 p->p_comm); 3690 3691 /* 3692 * Since this code is invoked by the last thread in an exiting 3693 * process, we would have context switched IN at some prior 3694 * point. However, with PREEMPTION, kernel mode context 3695 * switches may happen any time, so we want to disable a 3696 * context switch OUT till we get any PMCs targetting this 3697 * process off the hardware. 3698 * 3699 * We also need to atomically remove this process' 3700 * entry from our target process hash table, using 3701 * PMC_FLAG_REMOVE. 3702 */ 3703 PMCDBG(PRC,EXT,1, "process-exit proc=%p (%d, %s)", p, p->p_pid, 3704 p->p_comm); 3705 3706 critical_enter(); /* no preemption */ 3707 3708 cpu = curthread->td_oncpu; 3709 3710 if ((pp = pmc_find_process_descriptor(p, 3711 PMC_FLAG_REMOVE)) != NULL) { 3712 3713 PMCDBG(PRC,EXT,2, 3714 "process-exit proc=%p pmc-process=%p", p, pp); 3715 3716 /* 3717 * The exiting process could the target of 3718 * some PMCs which will be running on 3719 * currently executing CPU. 3720 * 3721 * We need to turn these PMCs off like we 3722 * would do at context switch OUT time. 3723 */ 3724 for (ri = 0; ri < md->pmd_npmc; ri++) { 3725 3726 /* 3727 * Pick up the pmc pointer from hardware 3728 * state similar to the CSW_OUT code. 3729 */ 3730 pm = NULL; 3731 (void) (*md->pmd_get_config)(cpu, ri, &pm); 3732 3733 PMCDBG(PRC,EXT,2, "ri=%d pm=%p", ri, pm); 3734 3735 if (pm == NULL || 3736 !PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) 3737 continue; 3738 3739 PMCDBG(PRC,EXT,2, "ppmcs[%d]=%p pm=%p " 3740 "state=%d", ri, pp->pp_pmcs[ri].pp_pmc, 3741 pm, pm->pm_state); 3742 3743 KASSERT(PMC_TO_ROWINDEX(pm) == ri, 3744 ("[pmc,%d] ri mismatch pmc(%d) ri(%d)", 3745 __LINE__, PMC_TO_ROWINDEX(pm), ri)); 3746 3747 KASSERT(pm == pp->pp_pmcs[ri].pp_pmc, 3748 ("[pmc,%d] pm %p != pp_pmcs[%d] %p", 3749 __LINE__, pm, ri, pp->pp_pmcs[ri].pp_pmc)); 3750 3751 (void) md->pmd_stop_pmc(cpu, ri); 3752 3753 KASSERT(pm->pm_runcount > 0, 3754 ("[pmc,%d] bad runcount ri %d rc %d", 3755 __LINE__, ri, pm->pm_runcount)); 3756 3757 /* Stop hardware only if it is actually running */ 3758 if (pm->pm_state == PMC_STATE_RUNNING && 3759 pm->pm_stalled == 0) { 3760 md->pmd_read_pmc(cpu, ri, &newvalue); 3761 tmp = newvalue - 3762 PMC_PCPU_SAVED(cpu,ri); 3763 3764 mtx_pool_lock_spin(pmc_mtxpool, pm); 3765 pm->pm_gv.pm_savedvalue += tmp; 3766 pp->pp_pmcs[ri].pp_pmcval += tmp; 3767 mtx_pool_unlock_spin(pmc_mtxpool, pm); 3768 } 3769 3770 atomic_subtract_rel_32(&pm->pm_runcount,1); 3771 3772 KASSERT((int) pm->pm_runcount >= 0, 3773 ("[pmc,%d] runcount is %d", __LINE__, ri)); 3774 3775 (void) md->pmd_config_pmc(cpu, ri, NULL); 3776 } 3777 3778 /* 3779 * Inform the MD layer of this pseudo "context switch 3780 * out" 3781 */ 3782 (void) md->pmd_switch_out(pmc_pcpu[cpu], pp); 3783 3784 critical_exit(); /* ok to be pre-empted now */ 3785 3786 /* 3787 * Unlink this process from the PMCs that are 3788 * targetting it. This will send a signal to 3789 * all PMC owner's whose PMCs are orphaned. 3790 * 3791 * Log PMC value at exit time if requested. 3792 */ 3793 for (ri = 0; ri < md->pmd_npmc; ri++) 3794 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) { 3795 if (pm->pm_flags & PMC_F_NEEDS_LOGFILE && 3796 PMC_IS_COUNTING_MODE(PMC_TO_MODE(pm))) 3797 pmclog_process_procexit(pm, pp); 3798 pmc_unlink_target_process(pm, pp); 3799 } 3800 FREE(pp, M_PMC); 3801 3802 } else 3803 critical_exit(); /* pp == NULL */ 3804 3805 3806 /* 3807 * If the process owned PMCs, free them up and free up 3808 * memory. 3809 */ 3810 if ((po = pmc_find_owner_descriptor(p)) != NULL) { 3811 pmc_remove_owner(po); 3812 pmc_destroy_owner_descriptor(po); 3813 } 3814 3815 sx_xunlock(&pmc_sx); 3816} 3817 3818/* 3819 * Handle a process fork. 3820 * 3821 * If the parent process 'p1' is under HWPMC monitoring, then copy 3822 * over any attached PMCs that have 'do_descendants' semantics. 3823 */ 3824 3825static void 3826pmc_process_fork(void *arg __unused, struct proc *p1, struct proc *newproc, 3827 int flags) 3828{ 3829 int is_using_hwpmcs; 3830 unsigned int ri; 3831 uint32_t do_descendants; 3832 struct pmc *pm; 3833 struct pmc_owner *po; 3834 struct pmc_process *ppnew, *ppold; 3835 3836 (void) flags; /* unused parameter */ 3837 3838 PROC_LOCK(p1); 3839 is_using_hwpmcs = p1->p_flag & P_HWPMC; 3840 PROC_UNLOCK(p1); 3841 3842 /* 3843 * If there are system-wide sampling PMCs active, we need to 3844 * log all fork events to their owner's logs. 3845 */ 3846 3847 LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) 3848 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 3849 pmclog_process_procfork(po, p1->p_pid, newproc->p_pid); 3850 3851 if (!is_using_hwpmcs) 3852 return; 3853 3854 PMC_GET_SX_XLOCK(); 3855 PMCDBG(PMC,FRK,1, "process-fork proc=%p (%d, %s) -> %p", p1, 3856 p1->p_pid, p1->p_comm, newproc); 3857 3858 /* 3859 * If the parent process (curthread->td_proc) is a 3860 * target of any PMCs, look for PMCs that are to be 3861 * inherited, and link these into the new process 3862 * descriptor. 3863 */ 3864 if ((ppold = pmc_find_process_descriptor(curthread->td_proc, 3865 PMC_FLAG_NONE)) == NULL) 3866 goto done; /* nothing to do */ 3867 3868 do_descendants = 0; 3869 for (ri = 0; ri < md->pmd_npmc; ri++) 3870 if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL) 3871 do_descendants |= pm->pm_flags & PMC_F_DESCENDANTS; 3872 if (do_descendants == 0) /* nothing to do */ 3873 goto done; 3874 3875 /* allocate a descriptor for the new process */ 3876 if ((ppnew = pmc_find_process_descriptor(newproc, 3877 PMC_FLAG_ALLOCATE)) == NULL) 3878 goto done; 3879 3880 /* 3881 * Run through all PMCs that were targeting the old process 3882 * and which specified F_DESCENDANTS and attach them to the 3883 * new process. 3884 * 3885 * Log the fork event to all owners of PMCs attached to this 3886 * process, if not already logged. 3887 */ 3888 for (ri = 0; ri < md->pmd_npmc; ri++) 3889 if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL && 3890 (pm->pm_flags & PMC_F_DESCENDANTS)) { 3891 pmc_link_target_process(pm, ppnew); 3892 po = pm->pm_owner; 3893 if (po->po_sscount == 0 && 3894 po->po_flags & PMC_PO_OWNS_LOGFILE) 3895 pmclog_process_procfork(po, p1->p_pid, 3896 newproc->p_pid); 3897 } 3898 3899 /* 3900 * Now mark the new process as being tracked by this driver. 3901 */ 3902 PROC_LOCK(newproc); 3903 newproc->p_flag |= P_HWPMC; 3904 PROC_UNLOCK(newproc); 3905 3906 done: 3907 sx_xunlock(&pmc_sx); 3908} 3909 3910 3911/* 3912 * initialization 3913 */ 3914 3915static const char *pmc_name_of_pmcclass[] = { 3916#undef __PMC_CLASS 3917#define __PMC_CLASS(N) #N , 3918 __PMC_CLASSES() 3919}; 3920 3921static int 3922pmc_initialize(void) 3923{ 3924 int cpu, error, n; 3925 struct pmc_binding pb; 3926 struct pmc_samplebuffer *sb; 3927 3928 md = NULL; 3929 error = 0; 3930 3931#ifdef DEBUG 3932 /* parse debug flags first */ 3933 if (TUNABLE_STR_FETCH(PMC_SYSCTL_NAME_PREFIX "debugflags", 3934 pmc_debugstr, sizeof(pmc_debugstr))) 3935 pmc_debugflags_parse(pmc_debugstr, 3936 pmc_debugstr+strlen(pmc_debugstr)); 3937#endif 3938 3939 PMCDBG(MOD,INI,0, "PMC Initialize (version %x)", PMC_VERSION); 3940 3941 /* check kernel version */ 3942 if (pmc_kernel_version != PMC_VERSION) { 3943 if (pmc_kernel_version == 0) 3944 printf("hwpmc: this kernel has not been compiled with " 3945 "'options HWPMC_HOOKS'.\n"); 3946 else 3947 printf("hwpmc: kernel version (0x%x) does not match " 3948 "module version (0x%x).\n", pmc_kernel_version, 3949 PMC_VERSION); 3950 return EPROGMISMATCH; 3951 } 3952 3953 /* 3954 * check sysctl parameters 3955 */ 3956 3957 if (pmc_hashsize <= 0) { 3958 (void) printf("hwpmc: tunable hashsize=%d must be greater " 3959 "than zero.\n", pmc_hashsize); 3960 pmc_hashsize = PMC_HASH_SIZE; 3961 } 3962 3963 if (pmc_nsamples <= 0 || pmc_nsamples > 65535) { 3964 (void) printf("hwpmc: tunable nsamples=%d out of range.\n", 3965 pmc_nsamples); 3966 pmc_nsamples = PMC_NSAMPLES; 3967 } 3968 3969 md = pmc_md_initialize(); 3970 3971 if (md == NULL || md->pmd_init == NULL) 3972 return ENOSYS; 3973 3974 /* allocate space for the per-cpu array */ 3975 MALLOC(pmc_pcpu, struct pmc_cpu **, mp_ncpus * sizeof(struct pmc_cpu *), 3976 M_PMC, M_WAITOK|M_ZERO); 3977 3978 /* per-cpu 'saved values' for managing process-mode PMCs */ 3979 MALLOC(pmc_pcpu_saved, pmc_value_t *, 3980 sizeof(pmc_value_t) * mp_ncpus * md->pmd_npmc, M_PMC, M_WAITOK); 3981 3982 /* perform cpu dependent initialization */ 3983 pmc_save_cpu_binding(&pb); 3984 for (cpu = 0; cpu < mp_ncpus; cpu++) { 3985 if (pmc_cpu_is_disabled(cpu)) 3986 continue; 3987 pmc_select_cpu(cpu); 3988 if ((error = md->pmd_init(cpu)) != 0) 3989 break; 3990 } 3991 pmc_restore_cpu_binding(&pb); 3992 3993 if (error != 0) 3994 return error; 3995 3996 /* allocate space for the sample array */ 3997 for (cpu = 0; cpu < mp_ncpus; cpu++) { 3998 if (pmc_cpu_is_disabled(cpu)) 3999 continue; 4000 MALLOC(sb, struct pmc_samplebuffer *, 4001 sizeof(struct pmc_samplebuffer) + 4002 pmc_nsamples * sizeof(struct pmc_sample), M_PMC, 4003 M_WAITOK|M_ZERO); 4004 4005 sb->ps_read = sb->ps_write = sb->ps_samples; 4006 sb->ps_fence = sb->ps_samples + pmc_nsamples; 4007 KASSERT(pmc_pcpu[cpu] != NULL, 4008 ("[pmc,%d] cpu=%d Null per-cpu data", __LINE__, cpu)); 4009 4010 pmc_pcpu[cpu]->pc_sb = sb; 4011 } 4012 4013 /* allocate space for the row disposition array */ 4014 pmc_pmcdisp = malloc(sizeof(enum pmc_mode) * md->pmd_npmc, 4015 M_PMC, M_WAITOK|M_ZERO); 4016 4017 KASSERT(pmc_pmcdisp != NULL, 4018 ("[pmc,%d] pmcdisp allocation returned NULL", __LINE__)); 4019 4020 /* mark all PMCs as available */ 4021 for (n = 0; n < (int) md->pmd_npmc; n++) 4022 PMC_MARK_ROW_FREE(n); 4023 4024 /* allocate thread hash tables */ 4025 pmc_ownerhash = hashinit(pmc_hashsize, M_PMC, 4026 &pmc_ownerhashmask); 4027 4028 pmc_processhash = hashinit(pmc_hashsize, M_PMC, 4029 &pmc_processhashmask); 4030 mtx_init(&pmc_processhash_mtx, "pmc-process-hash", "pmc", MTX_SPIN); 4031 4032 LIST_INIT(&pmc_ss_owners); 4033 pmc_ss_count = 0; 4034 4035 /* allocate a pool of spin mutexes */ 4036 pmc_mtxpool = mtx_pool_create("pmc", pmc_mtxpool_size, MTX_SPIN); 4037 4038 PMCDBG(MOD,INI,1, "pmc_ownerhash=%p, mask=0x%lx " 4039 "targethash=%p mask=0x%lx", pmc_ownerhash, pmc_ownerhashmask, 4040 pmc_processhash, pmc_processhashmask); 4041 4042 /* register process {exit,fork,exec} handlers */ 4043 pmc_exit_tag = EVENTHANDLER_REGISTER(process_exit, 4044 pmc_process_exit, NULL, EVENTHANDLER_PRI_ANY); 4045 pmc_fork_tag = EVENTHANDLER_REGISTER(process_fork, 4046 pmc_process_fork, NULL, EVENTHANDLER_PRI_ANY); 4047 4048 /* initialize logging */ 4049 pmclog_initialize(); 4050 4051 /* set hook functions */ 4052 pmc_intr = md->pmd_intr; 4053 pmc_hook = pmc_hook_handler; 4054 4055 if (error == 0) { 4056 printf(PMC_MODULE_NAME ":"); 4057 for (n = 0; n < (int) md->pmd_nclass; n++) { 4058 printf(" %s/%d/0x%b", 4059 pmc_name_of_pmcclass[md->pmd_classes[n].pm_class], 4060 md->pmd_nclasspmcs[n], 4061 md->pmd_classes[n].pm_caps, 4062 "\20" 4063 "\1INT\2USR\3SYS\4EDG\5THR" 4064 "\6REA\7WRI\10INV\11QUA\12PRC" 4065 "\13TAG\14CSC"); 4066 } 4067 printf("\n"); 4068 } 4069 4070 return error; 4071} 4072 4073/* prepare to be unloaded */ 4074static void 4075pmc_cleanup(void) 4076{ 4077 int cpu; 4078 struct pmc_ownerhash *ph; 4079 struct pmc_owner *po, *tmp; 4080 struct pmc_binding pb; 4081#ifdef DEBUG 4082 struct pmc_processhash *prh; 4083#endif 4084 4085 PMCDBG(MOD,INI,0, "%s", "cleanup"); 4086 4087 /* switch off sampling */ 4088 atomic_store_rel_int(&pmc_cpumask, 0); 4089 pmc_intr = NULL; 4090 4091 sx_xlock(&pmc_sx); 4092 if (pmc_hook == NULL) { /* being unloaded already */ 4093 sx_xunlock(&pmc_sx); 4094 return; 4095 } 4096 4097 pmc_hook = NULL; /* prevent new threads from entering module */ 4098 4099 /* deregister event handlers */ 4100 EVENTHANDLER_DEREGISTER(process_fork, pmc_fork_tag); 4101 EVENTHANDLER_DEREGISTER(process_exit, pmc_exit_tag); 4102 4103 /* send SIGBUS to all owner threads, free up allocations */ 4104 if (pmc_ownerhash) 4105 for (ph = pmc_ownerhash; 4106 ph <= &pmc_ownerhash[pmc_ownerhashmask]; 4107 ph++) { 4108 LIST_FOREACH_SAFE(po, ph, po_next, tmp) { 4109 pmc_remove_owner(po); 4110 4111 /* send SIGBUS to owner processes */ 4112 PMCDBG(MOD,INI,2, "cleanup signal proc=%p " 4113 "(%d, %s)", po->po_owner, 4114 po->po_owner->p_pid, 4115 po->po_owner->p_comm); 4116 4117 PROC_LOCK(po->po_owner); 4118 psignal(po->po_owner, SIGBUS); 4119 PROC_UNLOCK(po->po_owner); 4120 4121 pmc_destroy_owner_descriptor(po); 4122 } 4123 } 4124 4125 /* reclaim allocated data structures */ 4126 if (pmc_mtxpool) 4127 mtx_pool_destroy(&pmc_mtxpool); 4128 4129 mtx_destroy(&pmc_processhash_mtx); 4130 if (pmc_processhash) { 4131#ifdef DEBUG 4132 struct pmc_process *pp; 4133 4134 PMCDBG(MOD,INI,3, "%s", "destroy process hash"); 4135 for (prh = pmc_processhash; 4136 prh <= &pmc_processhash[pmc_processhashmask]; 4137 prh++) 4138 LIST_FOREACH(pp, prh, pp_next) 4139 PMCDBG(MOD,INI,3, "pid=%d", pp->pp_proc->p_pid); 4140#endif 4141 4142 hashdestroy(pmc_processhash, M_PMC, pmc_processhashmask); 4143 pmc_processhash = NULL; 4144 } 4145 4146 if (pmc_ownerhash) { 4147 PMCDBG(MOD,INI,3, "%s", "destroy owner hash"); 4148 hashdestroy(pmc_ownerhash, M_PMC, pmc_ownerhashmask); 4149 pmc_ownerhash = NULL; 4150 } 4151 4152 KASSERT(LIST_EMPTY(&pmc_ss_owners), 4153 ("[pmc,%d] Global SS owner list not empty", __LINE__)); 4154 KASSERT(pmc_ss_count == 0, 4155 ("[pmc,%d] Global SS count not empty", __LINE__)); 4156 4157 /* free the per-cpu sample buffers */ 4158 for (cpu = 0; cpu < mp_ncpus; cpu++) { 4159 if (pmc_cpu_is_disabled(cpu)) 4160 continue; 4161 KASSERT(pmc_pcpu[cpu]->pc_sb != NULL, 4162 ("[pmc,%d] Null cpu sample buffer cpu=%d", __LINE__, 4163 cpu)); 4164 FREE(pmc_pcpu[cpu]->pc_sb, M_PMC); 4165 pmc_pcpu[cpu]->pc_sb = NULL; 4166 } 4167 4168 /* do processor dependent cleanup */ 4169 PMCDBG(MOD,INI,3, "%s", "md cleanup"); 4170 if (md) { 4171 pmc_save_cpu_binding(&pb); 4172 for (cpu = 0; cpu < mp_ncpus; cpu++) { 4173 PMCDBG(MOD,INI,1,"pmc-cleanup cpu=%d pcs=%p", 4174 cpu, pmc_pcpu[cpu]); 4175 if (pmc_cpu_is_disabled(cpu)) 4176 continue; 4177 pmc_select_cpu(cpu); 4178 if (pmc_pcpu[cpu]) 4179 (void) md->pmd_cleanup(cpu); 4180 } 4181 FREE(md, M_PMC); 4182 md = NULL; 4183 pmc_restore_cpu_binding(&pb); 4184 } 4185 4186 /* deallocate per-cpu structures */ 4187 FREE(pmc_pcpu, M_PMC); 4188 pmc_pcpu = NULL; 4189 4190 FREE(pmc_pcpu_saved, M_PMC); 4191 pmc_pcpu_saved = NULL; 4192 4193 if (pmc_pmcdisp) { 4194 FREE(pmc_pmcdisp, M_PMC); 4195 pmc_pmcdisp = NULL; 4196 } 4197 4198 pmclog_shutdown(); 4199 4200 sx_xunlock(&pmc_sx); /* we are done */ 4201} 4202 4203/* 4204 * The function called at load/unload. 4205 */ 4206 4207static int 4208load (struct module *module __unused, int cmd, void *arg __unused) 4209{ 4210 int error; 4211 4212 error = 0; 4213 4214 switch (cmd) { 4215 case MOD_LOAD : 4216 /* initialize the subsystem */ 4217 error = pmc_initialize(); 4218 if (error != 0) 4219 break; 4220 PMCDBG(MOD,INI,1, "syscall=%d ncpus=%d", 4221 pmc_syscall_num, mp_ncpus); 4222 break; 4223 4224 4225 case MOD_UNLOAD : 4226 case MOD_SHUTDOWN: 4227 pmc_cleanup(); 4228 PMCDBG(MOD,INI,1, "%s", "unloaded"); 4229 break; 4230 4231 default : 4232 error = EINVAL; /* XXX should panic(9) */ 4233 break; 4234 } 4235 4236 return error; 4237} 4238 4239/* memory pool */ 4240MALLOC_DEFINE(M_PMC, "pmc", "Memory space for the PMC module"); 4241