1 /*- 2 * Copyright (c) 2003-2008 Joseph Koshy 3 * Copyright (c) 2007 The FreeBSD Foundation 4 * All rights reserved. 5 * 6 * Portions of this software were developed by A. Joseph Koshy under 7 * sponsorship from the FreeBSD Foundation and Google, Inc. 8 * 9 * Redistribution and use in source and binary forms, with or without 10 * modification, are permitted provided that the following conditions 11 * are met: 12 * 1. Redistributions of source code must retain the above copyright 13 * notice, this list of conditions and the following disclaimer. 14 * 2. Redistributions in binary form must reproduce the above copyright 15 * notice, this list of conditions and the following disclaimer in the 16 * documentation and/or other materials provided with the distribution. 17 * 18 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 19 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 20 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 21 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE 22 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 23 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 24 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 25 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 26 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 27 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 28 * SUCH DAMAGE. 29 * 30 */ 31 32 #include <sys/cdefs.h> 33 __FBSDID("$FreeBSD$"); 34 35 #include <sys/param.h> 36 #include <sys/eventhandler.h> 37 #include <sys/jail.h> 38 #include <sys/kernel.h> 39 #include <sys/kthread.h> 40 #include <sys/limits.h> 41 #include <sys/lock.h> 42 #include <sys/malloc.h> 43 #include <sys/module.h> 44 #include <sys/mount.h> 45 #include <sys/mutex.h> 46 #include <sys/pmc.h> 47 #include <sys/pmckern.h> 48 #include <sys/pmclog.h> 49 #include <sys/priv.h> 50 #include <sys/proc.h> 51 #include <sys/queue.h> 52 #include <sys/resourcevar.h> 53 #include <sys/rwlock.h> 54 #include <sys/sched.h> 55 #include <sys/signalvar.h> 56 #include <sys/smp.h> 57 #include <sys/sx.h> 58 #include <sys/sysctl.h> 59 #include <sys/sysent.h> 60 #include <sys/systm.h> 61 #include <sys/vnode.h> 62 63 #include <sys/linker.h> /* needs to be after <sys/malloc.h> */ 64 65 #include <machine/atomic.h> 66 #include <machine/md_var.h> 67 68 #include <vm/vm.h> 69 #include <vm/vm_extern.h> 70 #include <vm/pmap.h> 71 #include <vm/vm_map.h> 72 #include <vm/vm_object.h> 73 74 #include "hwpmc_soft.h" 75 76 /* 77 * Types 78 */ 79 80 enum pmc_flags { 81 PMC_FLAG_NONE = 0x00, /* do nothing */ 82 PMC_FLAG_REMOVE = 0x01, /* atomically remove entry from hash */ 83 PMC_FLAG_ALLOCATE = 0x02, /* add entry to hash if not found */ 84 }; 85 86 /* 87 * The offset in sysent where the syscall is allocated. 88 */ 89 90 static int pmc_syscall_num = NO_SYSCALL; 91 struct pmc_cpu **pmc_pcpu; /* per-cpu state */ 92 pmc_value_t *pmc_pcpu_saved; /* saved PMC values: CSW handling */ 93 94 #define PMC_PCPU_SAVED(C,R) pmc_pcpu_saved[(R) + md->pmd_npmc*(C)] 95 96 struct mtx_pool *pmc_mtxpool; 97 static int *pmc_pmcdisp; /* PMC row dispositions */ 98 99 #define PMC_ROW_DISP_IS_FREE(R) (pmc_pmcdisp[(R)] == 0) 100 #define PMC_ROW_DISP_IS_THREAD(R) (pmc_pmcdisp[(R)] > 0) 101 #define PMC_ROW_DISP_IS_STANDALONE(R) (pmc_pmcdisp[(R)] < 0) 102 103 #define PMC_MARK_ROW_FREE(R) do { \ 104 pmc_pmcdisp[(R)] = 0; \ 105 } while (0) 106 107 #define PMC_MARK_ROW_STANDALONE(R) do { \ 108 KASSERT(pmc_pmcdisp[(R)] <= 0, ("[pmc,%d] row disposition error", \ 109 __LINE__)); \ 110 atomic_add_int(&pmc_pmcdisp[(R)], -1); \ 111 KASSERT(pmc_pmcdisp[(R)] >= (-pmc_cpu_max_active()), \ 112 ("[pmc,%d] row disposition error", __LINE__)); \ 113 } while (0) 114 115 #define PMC_UNMARK_ROW_STANDALONE(R) do { \ 116 atomic_add_int(&pmc_pmcdisp[(R)], 1); \ 117 KASSERT(pmc_pmcdisp[(R)] <= 0, ("[pmc,%d] row disposition error", \ 118 __LINE__)); \ 119 } while (0) 120 121 #define PMC_MARK_ROW_THREAD(R) do { \ 122 KASSERT(pmc_pmcdisp[(R)] >= 0, ("[pmc,%d] row disposition error", \ 123 __LINE__)); \ 124 atomic_add_int(&pmc_pmcdisp[(R)], 1); \ 125 } while (0) 126 127 #define PMC_UNMARK_ROW_THREAD(R) do { \ 128 atomic_add_int(&pmc_pmcdisp[(R)], -1); \ 129 KASSERT(pmc_pmcdisp[(R)] >= 0, ("[pmc,%d] row disposition error", \ 130 __LINE__)); \ 131 } while (0) 132 133 134 /* various event handlers */ 135 static eventhandler_tag pmc_exit_tag, pmc_fork_tag, pmc_kld_load_tag, 136 pmc_kld_unload_tag; 137 138 /* Module statistics */ 139 struct pmc_op_getdriverstats pmc_stats; 140 141 /* Machine/processor dependent operations */ 142 static struct pmc_mdep *md; 143 144 /* 145 * Hash tables mapping owner processes and target threads to PMCs. 146 */ 147 148 struct mtx pmc_processhash_mtx; /* spin mutex */ 149 static u_long pmc_processhashmask; 150 static LIST_HEAD(pmc_processhash, pmc_process) *pmc_processhash; 151 152 /* 153 * Hash table of PMC owner descriptors. This table is protected by 154 * the shared PMC "sx" lock. 155 */ 156 157 static u_long pmc_ownerhashmask; 158 static LIST_HEAD(pmc_ownerhash, pmc_owner) *pmc_ownerhash; 159 160 /* 161 * List of PMC owners with system-wide sampling PMCs. 162 */ 163 164 static LIST_HEAD(, pmc_owner) pmc_ss_owners; 165 166 167 /* 168 * A map of row indices to classdep structures. 169 */ 170 static struct pmc_classdep **pmc_rowindex_to_classdep; 171 172 /* 173 * Prototypes 174 */ 175 176 #ifdef DEBUG 177 static int pmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS); 178 static int pmc_debugflags_parse(char *newstr, char *fence); 179 #endif 180 181 static int load(struct module *module, int cmd, void *arg); 182 static int pmc_attach_process(struct proc *p, struct pmc *pm); 183 static struct pmc *pmc_allocate_pmc_descriptor(void); 184 static struct pmc_owner *pmc_allocate_owner_descriptor(struct proc *p); 185 static int pmc_attach_one_process(struct proc *p, struct pmc *pm); 186 static int pmc_can_allocate_rowindex(struct proc *p, unsigned int ri, 187 int cpu); 188 static int pmc_can_attach(struct pmc *pm, struct proc *p); 189 static void pmc_capture_user_callchain(int cpu, int soft, struct trapframe *tf); 190 static void pmc_cleanup(void); 191 static int pmc_detach_process(struct proc *p, struct pmc *pm); 192 static int pmc_detach_one_process(struct proc *p, struct pmc *pm, 193 int flags); 194 static void pmc_destroy_owner_descriptor(struct pmc_owner *po); 195 static struct pmc_owner *pmc_find_owner_descriptor(struct proc *p); 196 static int pmc_find_pmc(pmc_id_t pmcid, struct pmc **pm); 197 static struct pmc *pmc_find_pmc_descriptor_in_process(struct pmc_owner *po, 198 pmc_id_t pmc); 199 static struct pmc_process *pmc_find_process_descriptor(struct proc *p, 200 uint32_t mode); 201 static void pmc_force_context_switch(void); 202 static void pmc_link_target_process(struct pmc *pm, 203 struct pmc_process *pp); 204 static void pmc_log_all_process_mappings(struct pmc_owner *po); 205 static void pmc_log_kernel_mappings(struct pmc *pm); 206 static void pmc_log_process_mappings(struct pmc_owner *po, struct proc *p); 207 static void pmc_maybe_remove_owner(struct pmc_owner *po); 208 static void pmc_process_csw_in(struct thread *td); 209 static void pmc_process_csw_out(struct thread *td); 210 static void pmc_process_exit(void *arg, struct proc *p); 211 static void pmc_process_fork(void *arg, struct proc *p1, 212 struct proc *p2, int n); 213 static void pmc_process_samples(int cpu, int soft); 214 static void pmc_release_pmc_descriptor(struct pmc *pmc); 215 static void pmc_remove_owner(struct pmc_owner *po); 216 static void pmc_remove_process_descriptor(struct pmc_process *pp); 217 static void pmc_restore_cpu_binding(struct pmc_binding *pb); 218 static void pmc_save_cpu_binding(struct pmc_binding *pb); 219 static void pmc_select_cpu(int cpu); 220 static int pmc_start(struct pmc *pm); 221 static int pmc_stop(struct pmc *pm); 222 static int pmc_syscall_handler(struct thread *td, void *syscall_args); 223 static void pmc_unlink_target_process(struct pmc *pmc, 224 struct pmc_process *pp); 225 static int generic_switch_in(struct pmc_cpu *pc, struct pmc_process *pp); 226 static int generic_switch_out(struct pmc_cpu *pc, struct pmc_process *pp); 227 static struct pmc_mdep *pmc_generic_cpu_initialize(void); 228 static void pmc_generic_cpu_finalize(struct pmc_mdep *md); 229 230 /* 231 * Kernel tunables and sysctl(8) interface. 232 */ 233 234 SYSCTL_DECL(_kern_hwpmc); 235 236 static int pmc_callchaindepth = PMC_CALLCHAIN_DEPTH; 237 SYSCTL_INT(_kern_hwpmc, OID_AUTO, callchaindepth, CTLFLAG_RDTUN, 238 &pmc_callchaindepth, 0, "depth of call chain records"); 239 240 #ifdef DEBUG 241 struct pmc_debugflags pmc_debugflags = PMC_DEBUG_DEFAULT_FLAGS; 242 char pmc_debugstr[PMC_DEBUG_STRSIZE]; 243 TUNABLE_STR(PMC_SYSCTL_NAME_PREFIX "debugflags", pmc_debugstr, 244 sizeof(pmc_debugstr)); 245 SYSCTL_PROC(_kern_hwpmc, OID_AUTO, debugflags, 246 CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NOFETCH, 247 0, 0, pmc_debugflags_sysctl_handler, "A", "debug flags"); 248 #endif 249 250 /* 251 * kern.hwpmc.hashrows -- determines the number of rows in the 252 * of the hash table used to look up threads 253 */ 254 255 static int pmc_hashsize = PMC_HASH_SIZE; 256 SYSCTL_INT(_kern_hwpmc, OID_AUTO, hashsize, CTLFLAG_RDTUN, 257 &pmc_hashsize, 0, "rows in hash tables"); 258 259 /* 260 * kern.hwpmc.nsamples --- number of PC samples/callchain stacks per CPU 261 */ 262 263 static int pmc_nsamples = PMC_NSAMPLES; 264 SYSCTL_INT(_kern_hwpmc, OID_AUTO, nsamples, CTLFLAG_RDTUN, 265 &pmc_nsamples, 0, "number of PC samples per CPU"); 266 267 268 /* 269 * kern.hwpmc.mtxpoolsize -- number of mutexes in the mutex pool. 270 */ 271 272 static int pmc_mtxpool_size = PMC_MTXPOOL_SIZE; 273 SYSCTL_INT(_kern_hwpmc, OID_AUTO, mtxpoolsize, CTLFLAG_RDTUN, 274 &pmc_mtxpool_size, 0, "size of spin mutex pool"); 275 276 277 /* 278 * security.bsd.unprivileged_syspmcs -- allow non-root processes to 279 * allocate system-wide PMCs. 280 * 281 * Allowing unprivileged processes to allocate system PMCs is convenient 282 * if system-wide measurements need to be taken concurrently with other 283 * per-process measurements. This feature is turned off by default. 284 */ 285 286 static int pmc_unprivileged_syspmcs = 0; 287 SYSCTL_INT(_security_bsd, OID_AUTO, unprivileged_syspmcs, CTLFLAG_RWTUN, 288 &pmc_unprivileged_syspmcs, 0, 289 "allow unprivileged process to allocate system PMCs"); 290 291 /* 292 * Hash function. Discard the lower 2 bits of the pointer since 293 * these are always zero for our uses. The hash multiplier is 294 * round((2^LONG_BIT) * ((sqrt(5)-1)/2)). 295 */ 296 297 #if LONG_BIT == 64 298 #define _PMC_HM 11400714819323198486u 299 #elif LONG_BIT == 32 300 #define _PMC_HM 2654435769u 301 #else 302 #error Must know the size of 'long' to compile 303 #endif 304 305 #define PMC_HASH_PTR(P,M) ((((unsigned long) (P) >> 2) * _PMC_HM) & (M)) 306 307 /* 308 * Syscall structures 309 */ 310 311 /* The `sysent' for the new syscall */ 312 static struct sysent pmc_sysent = { 313 2, /* sy_narg */ 314 pmc_syscall_handler /* sy_call */ 315 }; 316 317 static struct syscall_module_data pmc_syscall_mod = { 318 load, 319 NULL, 320 &pmc_syscall_num, 321 &pmc_sysent, 322 { 0, NULL } 323 }; 324 325 static moduledata_t pmc_mod = { 326 PMC_MODULE_NAME, 327 syscall_module_handler, 328 &pmc_syscall_mod 329 }; 330 331 DECLARE_MODULE(pmc, pmc_mod, SI_SUB_SMP, SI_ORDER_ANY); 332 MODULE_VERSION(pmc, PMC_VERSION); 333 334 #ifdef DEBUG 335 enum pmc_dbgparse_state { 336 PMCDS_WS, /* in whitespace */ 337 PMCDS_MAJOR, /* seen a major keyword */ 338 PMCDS_MINOR 339 }; 340 341 static int 342 pmc_debugflags_parse(char *newstr, char *fence) 343 { 344 char c, *p, *q; 345 struct pmc_debugflags *tmpflags; 346 int error, found, *newbits, tmp; 347 size_t kwlen; 348 349 tmpflags = malloc(sizeof(*tmpflags), M_PMC, M_WAITOK|M_ZERO); 350 351 p = newstr; 352 error = 0; 353 354 for (; p < fence && (c = *p); p++) { 355 356 /* skip white space */ 357 if (c == ' ' || c == '\t') 358 continue; 359 360 /* look for a keyword followed by "=" */ 361 for (q = p; p < fence && (c = *p) && c != '='; p++) 362 ; 363 if (c != '=') { 364 error = EINVAL; 365 goto done; 366 } 367 368 kwlen = p - q; 369 newbits = NULL; 370 371 /* lookup flag group name */ 372 #define DBG_SET_FLAG_MAJ(S,F) \ 373 if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0) \ 374 newbits = &tmpflags->pdb_ ## F; 375 376 DBG_SET_FLAG_MAJ("cpu", CPU); 377 DBG_SET_FLAG_MAJ("csw", CSW); 378 DBG_SET_FLAG_MAJ("logging", LOG); 379 DBG_SET_FLAG_MAJ("module", MOD); 380 DBG_SET_FLAG_MAJ("md", MDP); 381 DBG_SET_FLAG_MAJ("owner", OWN); 382 DBG_SET_FLAG_MAJ("pmc", PMC); 383 DBG_SET_FLAG_MAJ("process", PRC); 384 DBG_SET_FLAG_MAJ("sampling", SAM); 385 386 if (newbits == NULL) { 387 error = EINVAL; 388 goto done; 389 } 390 391 p++; /* skip the '=' */ 392 393 /* Now parse the individual flags */ 394 tmp = 0; 395 newflag: 396 for (q = p; p < fence && (c = *p); p++) 397 if (c == ' ' || c == '\t' || c == ',') 398 break; 399 400 /* p == fence or c == ws or c == "," or c == 0 */ 401 402 if ((kwlen = p - q) == 0) { 403 *newbits = tmp; 404 continue; 405 } 406 407 found = 0; 408 #define DBG_SET_FLAG_MIN(S,F) \ 409 if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0) \ 410 tmp |= found = (1 << PMC_DEBUG_MIN_ ## F) 411 412 /* a '*' denotes all possible flags in the group */ 413 if (kwlen == 1 && *q == '*') 414 tmp = found = ~0; 415 /* look for individual flag names */ 416 DBG_SET_FLAG_MIN("allocaterow", ALR); 417 DBG_SET_FLAG_MIN("allocate", ALL); 418 DBG_SET_FLAG_MIN("attach", ATT); 419 DBG_SET_FLAG_MIN("bind", BND); 420 DBG_SET_FLAG_MIN("config", CFG); 421 DBG_SET_FLAG_MIN("exec", EXC); 422 DBG_SET_FLAG_MIN("exit", EXT); 423 DBG_SET_FLAG_MIN("find", FND); 424 DBG_SET_FLAG_MIN("flush", FLS); 425 DBG_SET_FLAG_MIN("fork", FRK); 426 DBG_SET_FLAG_MIN("getbuf", GTB); 427 DBG_SET_FLAG_MIN("hook", PMH); 428 DBG_SET_FLAG_MIN("init", INI); 429 DBG_SET_FLAG_MIN("intr", INT); 430 DBG_SET_FLAG_MIN("linktarget", TLK); 431 DBG_SET_FLAG_MIN("mayberemove", OMR); 432 DBG_SET_FLAG_MIN("ops", OPS); 433 DBG_SET_FLAG_MIN("read", REA); 434 DBG_SET_FLAG_MIN("register", REG); 435 DBG_SET_FLAG_MIN("release", REL); 436 DBG_SET_FLAG_MIN("remove", ORM); 437 DBG_SET_FLAG_MIN("sample", SAM); 438 DBG_SET_FLAG_MIN("scheduleio", SIO); 439 DBG_SET_FLAG_MIN("select", SEL); 440 DBG_SET_FLAG_MIN("signal", SIG); 441 DBG_SET_FLAG_MIN("swi", SWI); 442 DBG_SET_FLAG_MIN("swo", SWO); 443 DBG_SET_FLAG_MIN("start", STA); 444 DBG_SET_FLAG_MIN("stop", STO); 445 DBG_SET_FLAG_MIN("syscall", PMS); 446 DBG_SET_FLAG_MIN("unlinktarget", TUL); 447 DBG_SET_FLAG_MIN("write", WRI); 448 if (found == 0) { 449 /* unrecognized flag name */ 450 error = EINVAL; 451 goto done; 452 } 453 454 if (c == 0 || c == ' ' || c == '\t') { /* end of flag group */ 455 *newbits = tmp; 456 continue; 457 } 458 459 p++; 460 goto newflag; 461 } 462 463 /* save the new flag set */ 464 bcopy(tmpflags, &pmc_debugflags, sizeof(pmc_debugflags)); 465 466 done: 467 free(tmpflags, M_PMC); 468 return error; 469 } 470 471 static int 472 pmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS) 473 { 474 char *fence, *newstr; 475 int error; 476 unsigned int n; 477 478 (void) arg1; (void) arg2; /* unused parameters */ 479 480 n = sizeof(pmc_debugstr); 481 newstr = malloc(n, M_PMC, M_WAITOK|M_ZERO); 482 (void) strlcpy(newstr, pmc_debugstr, n); 483 484 error = sysctl_handle_string(oidp, newstr, n, req); 485 486 /* if there is a new string, parse and copy it */ 487 if (error == 0 && req->newptr != NULL) { 488 fence = newstr + (n < req->newlen ? n : req->newlen + 1); 489 if ((error = pmc_debugflags_parse(newstr, fence)) == 0) 490 (void) strlcpy(pmc_debugstr, newstr, 491 sizeof(pmc_debugstr)); 492 } 493 494 free(newstr, M_PMC); 495 496 return error; 497 } 498 #endif 499 500 /* 501 * Map a row index to a classdep structure and return the adjusted row 502 * index for the PMC class index. 503 */ 504 static struct pmc_classdep * 505 pmc_ri_to_classdep(struct pmc_mdep *md, int ri, int *adjri) 506 { 507 struct pmc_classdep *pcd; 508 509 (void) md; 510 511 KASSERT(ri >= 0 && ri < md->pmd_npmc, 512 ("[pmc,%d] illegal row-index %d", __LINE__, ri)); 513 514 pcd = pmc_rowindex_to_classdep[ri]; 515 516 KASSERT(pcd != NULL, 517 ("[pmc,%d] ri %d null pcd", __LINE__, ri)); 518 519 *adjri = ri - pcd->pcd_ri; 520 521 KASSERT(*adjri >= 0 && *adjri < pcd->pcd_num, 522 ("[pmc,%d] adjusted row-index %d", __LINE__, *adjri)); 523 524 return (pcd); 525 } 526 527 /* 528 * Concurrency Control 529 * 530 * The driver manages the following data structures: 531 * 532 * - target process descriptors, one per target process 533 * - owner process descriptors (and attached lists), one per owner process 534 * - lookup hash tables for owner and target processes 535 * - PMC descriptors (and attached lists) 536 * - per-cpu hardware state 537 * - the 'hook' variable through which the kernel calls into 538 * this module 539 * - the machine hardware state (managed by the MD layer) 540 * 541 * These data structures are accessed from: 542 * 543 * - thread context-switch code 544 * - interrupt handlers (possibly on multiple cpus) 545 * - kernel threads on multiple cpus running on behalf of user 546 * processes doing system calls 547 * - this driver's private kernel threads 548 * 549 * = Locks and Locking strategy = 550 * 551 * The driver uses four locking strategies for its operation: 552 * 553 * - The global SX lock "pmc_sx" is used to protect internal 554 * data structures. 555 * 556 * Calls into the module by syscall() start with this lock being 557 * held in exclusive mode. Depending on the requested operation, 558 * the lock may be downgraded to 'shared' mode to allow more 559 * concurrent readers into the module. Calls into the module from 560 * other parts of the kernel acquire the lock in shared mode. 561 * 562 * This SX lock is held in exclusive mode for any operations that 563 * modify the linkages between the driver's internal data structures. 564 * 565 * The 'pmc_hook' function pointer is also protected by this lock. 566 * It is only examined with the sx lock held in exclusive mode. The 567 * kernel module is allowed to be unloaded only with the sx lock held 568 * in exclusive mode. In normal syscall handling, after acquiring the 569 * pmc_sx lock we first check that 'pmc_hook' is non-null before 570 * proceeding. This prevents races between the thread unloading the module 571 * and other threads seeking to use the module. 572 * 573 * - Lookups of target process structures and owner process structures 574 * cannot use the global "pmc_sx" SX lock because these lookups need 575 * to happen during context switches and in other critical sections 576 * where sleeping is not allowed. We protect these lookup tables 577 * with their own private spin-mutexes, "pmc_processhash_mtx" and 578 * "pmc_ownerhash_mtx". 579 * 580 * - Interrupt handlers work in a lock free manner. At interrupt 581 * time, handlers look at the PMC pointer (phw->phw_pmc) configured 582 * when the PMC was started. If this pointer is NULL, the interrupt 583 * is ignored after updating driver statistics. We ensure that this 584 * pointer is set (using an atomic operation if necessary) before the 585 * PMC hardware is started. Conversely, this pointer is unset atomically 586 * only after the PMC hardware is stopped. 587 * 588 * We ensure that everything needed for the operation of an 589 * interrupt handler is available without it needing to acquire any 590 * locks. We also ensure that a PMC's software state is destroyed only 591 * after the PMC is taken off hardware (on all CPUs). 592 * 593 * - Context-switch handling with process-private PMCs needs more 594 * care. 595 * 596 * A given process may be the target of multiple PMCs. For example, 597 * PMCATTACH and PMCDETACH may be requested by a process on one CPU 598 * while the target process is running on another. A PMC could also 599 * be getting released because its owner is exiting. We tackle 600 * these situations in the following manner: 601 * 602 * - each target process structure 'pmc_process' has an array 603 * of 'struct pmc *' pointers, one for each hardware PMC. 604 * 605 * - At context switch IN time, each "target" PMC in RUNNING state 606 * gets started on hardware and a pointer to each PMC is copied into 607 * the per-cpu phw array. The 'runcount' for the PMC is 608 * incremented. 609 * 610 * - At context switch OUT time, all process-virtual PMCs are stopped 611 * on hardware. The saved value is added to the PMCs value field 612 * only if the PMC is in a non-deleted state (the PMCs state could 613 * have changed during the current time slice). 614 * 615 * Note that since in-between a switch IN on a processor and a switch 616 * OUT, the PMC could have been released on another CPU. Therefore 617 * context switch OUT always looks at the hardware state to turn 618 * OFF PMCs and will update a PMC's saved value only if reachable 619 * from the target process record. 620 * 621 * - OP PMCRELEASE could be called on a PMC at any time (the PMC could 622 * be attached to many processes at the time of the call and could 623 * be active on multiple CPUs). 624 * 625 * We prevent further scheduling of the PMC by marking it as in 626 * state 'DELETED'. If the runcount of the PMC is non-zero then 627 * this PMC is currently running on a CPU somewhere. The thread 628 * doing the PMCRELEASE operation waits by repeatedly doing a 629 * pause() till the runcount comes to zero. 630 * 631 * The contents of a PMC descriptor (struct pmc) are protected using 632 * a spin-mutex. In order to save space, we use a mutex pool. 633 * 634 * In terms of lock types used by witness(4), we use: 635 * - Type "pmc-sx", used by the global SX lock. 636 * - Type "pmc-sleep", for sleep mutexes used by logger threads. 637 * - Type "pmc-per-proc", for protecting PMC owner descriptors. 638 * - Type "pmc-leaf", used for all other spin mutexes. 639 */ 640 641 /* 642 * save the cpu binding of the current kthread 643 */ 644 645 static void 646 pmc_save_cpu_binding(struct pmc_binding *pb) 647 { 648 PMCDBG(CPU,BND,2, "%s", "save-cpu"); 649 thread_lock(curthread); 650 pb->pb_bound = sched_is_bound(curthread); 651 pb->pb_cpu = curthread->td_oncpu; 652 thread_unlock(curthread); 653 PMCDBG(CPU,BND,2, "save-cpu cpu=%d", pb->pb_cpu); 654 } 655 656 /* 657 * restore the cpu binding of the current thread 658 */ 659 660 static void 661 pmc_restore_cpu_binding(struct pmc_binding *pb) 662 { 663 PMCDBG(CPU,BND,2, "restore-cpu curcpu=%d restore=%d", 664 curthread->td_oncpu, pb->pb_cpu); 665 thread_lock(curthread); 666 if (pb->pb_bound) 667 sched_bind(curthread, pb->pb_cpu); 668 else 669 sched_unbind(curthread); 670 thread_unlock(curthread); 671 PMCDBG(CPU,BND,2, "%s", "restore-cpu done"); 672 } 673 674 /* 675 * move execution over the specified cpu and bind it there. 676 */ 677 678 static void 679 pmc_select_cpu(int cpu) 680 { 681 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), 682 ("[pmc,%d] bad cpu number %d", __LINE__, cpu)); 683 684 /* Never move to an inactive CPU. */ 685 KASSERT(pmc_cpu_is_active(cpu), ("[pmc,%d] selecting inactive " 686 "CPU %d", __LINE__, cpu)); 687 688 PMCDBG(CPU,SEL,2, "select-cpu cpu=%d", cpu); 689 thread_lock(curthread); 690 sched_bind(curthread, cpu); 691 thread_unlock(curthread); 692 693 KASSERT(curthread->td_oncpu == cpu, 694 ("[pmc,%d] CPU not bound [cpu=%d, curr=%d]", __LINE__, 695 cpu, curthread->td_oncpu)); 696 697 PMCDBG(CPU,SEL,2, "select-cpu cpu=%d ok", cpu); 698 } 699 700 /* 701 * Force a context switch. 702 * 703 * We do this by pause'ing for 1 tick -- invoking mi_switch() is not 704 * guaranteed to force a context switch. 705 */ 706 707 static void 708 pmc_force_context_switch(void) 709 { 710 711 pause("pmcctx", 1); 712 } 713 714 /* 715 * Get the file name for an executable. This is a simple wrapper 716 * around vn_fullpath(9). 717 */ 718 719 static void 720 pmc_getfilename(struct vnode *v, char **fullpath, char **freepath) 721 { 722 723 *fullpath = "unknown"; 724 *freepath = NULL; 725 vn_fullpath(curthread, v, fullpath, freepath); 726 } 727 728 /* 729 * remove an process owning PMCs 730 */ 731 732 void 733 pmc_remove_owner(struct pmc_owner *po) 734 { 735 struct pmc *pm, *tmp; 736 737 sx_assert(&pmc_sx, SX_XLOCKED); 738 739 PMCDBG(OWN,ORM,1, "remove-owner po=%p", po); 740 741 /* Remove descriptor from the owner hash table */ 742 LIST_REMOVE(po, po_next); 743 744 /* release all owned PMC descriptors */ 745 LIST_FOREACH_SAFE(pm, &po->po_pmcs, pm_next, tmp) { 746 PMCDBG(OWN,ORM,2, "pmc=%p", pm); 747 KASSERT(pm->pm_owner == po, 748 ("[pmc,%d] owner %p != po %p", __LINE__, pm->pm_owner, po)); 749 750 pmc_release_pmc_descriptor(pm); /* will unlink from the list */ 751 } 752 753 KASSERT(po->po_sscount == 0, 754 ("[pmc,%d] SS count not zero", __LINE__)); 755 KASSERT(LIST_EMPTY(&po->po_pmcs), 756 ("[pmc,%d] PMC list not empty", __LINE__)); 757 758 /* de-configure the log file if present */ 759 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 760 pmclog_deconfigure_log(po); 761 } 762 763 /* 764 * remove an owner process record if all conditions are met. 765 */ 766 767 static void 768 pmc_maybe_remove_owner(struct pmc_owner *po) 769 { 770 771 PMCDBG(OWN,OMR,1, "maybe-remove-owner po=%p", po); 772 773 /* 774 * Remove owner record if 775 * - this process does not own any PMCs 776 * - this process has not allocated a system-wide sampling buffer 777 */ 778 779 if (LIST_EMPTY(&po->po_pmcs) && 780 ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)) { 781 pmc_remove_owner(po); 782 pmc_destroy_owner_descriptor(po); 783 } 784 } 785 786 /* 787 * Add an association between a target process and a PMC. 788 */ 789 790 static void 791 pmc_link_target_process(struct pmc *pm, struct pmc_process *pp) 792 { 793 int ri; 794 struct pmc_target *pt; 795 796 sx_assert(&pmc_sx, SX_XLOCKED); 797 798 KASSERT(pm != NULL && pp != NULL, 799 ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp)); 800 KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)), 801 ("[pmc,%d] Attaching a non-process-virtual pmc=%p to pid=%d", 802 __LINE__, pm, pp->pp_proc->p_pid)); 803 KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= ((int) md->pmd_npmc - 1), 804 ("[pmc,%d] Illegal reference count %d for process record %p", 805 __LINE__, pp->pp_refcnt, (void *) pp)); 806 807 ri = PMC_TO_ROWINDEX(pm); 808 809 PMCDBG(PRC,TLK,1, "link-target pmc=%p ri=%d pmc-process=%p", 810 pm, ri, pp); 811 812 #ifdef DEBUG 813 LIST_FOREACH(pt, &pm->pm_targets, pt_next) 814 if (pt->pt_process == pp) 815 KASSERT(0, ("[pmc,%d] pp %p already in pmc %p targets", 816 __LINE__, pp, pm)); 817 #endif 818 819 pt = malloc(sizeof(struct pmc_target), M_PMC, M_WAITOK|M_ZERO); 820 pt->pt_process = pp; 821 822 LIST_INSERT_HEAD(&pm->pm_targets, pt, pt_next); 823 824 atomic_store_rel_ptr((uintptr_t *)&pp->pp_pmcs[ri].pp_pmc, 825 (uintptr_t)pm); 826 827 if (pm->pm_owner->po_owner == pp->pp_proc) 828 pm->pm_flags |= PMC_F_ATTACHED_TO_OWNER; 829 830 /* 831 * Initialize the per-process values at this row index. 832 */ 833 pp->pp_pmcs[ri].pp_pmcval = PMC_TO_MODE(pm) == PMC_MODE_TS ? 834 pm->pm_sc.pm_reloadcount : 0; 835 836 pp->pp_refcnt++; 837 838 } 839 840 /* 841 * Removes the association between a target process and a PMC. 842 */ 843 844 static void 845 pmc_unlink_target_process(struct pmc *pm, struct pmc_process *pp) 846 { 847 int ri; 848 struct proc *p; 849 struct pmc_target *ptgt; 850 851 sx_assert(&pmc_sx, SX_XLOCKED); 852 853 KASSERT(pm != NULL && pp != NULL, 854 ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp)); 855 856 KASSERT(pp->pp_refcnt >= 1 && pp->pp_refcnt <= (int) md->pmd_npmc, 857 ("[pmc,%d] Illegal ref count %d on process record %p", 858 __LINE__, pp->pp_refcnt, (void *) pp)); 859 860 ri = PMC_TO_ROWINDEX(pm); 861 862 PMCDBG(PRC,TUL,1, "unlink-target pmc=%p ri=%d pmc-process=%p", 863 pm, ri, pp); 864 865 KASSERT(pp->pp_pmcs[ri].pp_pmc == pm, 866 ("[pmc,%d] PMC ri %d mismatch pmc %p pp->[ri] %p", __LINE__, 867 ri, pm, pp->pp_pmcs[ri].pp_pmc)); 868 869 pp->pp_pmcs[ri].pp_pmc = NULL; 870 pp->pp_pmcs[ri].pp_pmcval = (pmc_value_t) 0; 871 872 /* Remove owner-specific flags */ 873 if (pm->pm_owner->po_owner == pp->pp_proc) { 874 pp->pp_flags &= ~PMC_PP_ENABLE_MSR_ACCESS; 875 pm->pm_flags &= ~PMC_F_ATTACHED_TO_OWNER; 876 } 877 878 pp->pp_refcnt--; 879 880 /* Remove the target process from the PMC structure */ 881 LIST_FOREACH(ptgt, &pm->pm_targets, pt_next) 882 if (ptgt->pt_process == pp) 883 break; 884 885 KASSERT(ptgt != NULL, ("[pmc,%d] process %p (pp: %p) not found " 886 "in pmc %p", __LINE__, pp->pp_proc, pp, pm)); 887 888 LIST_REMOVE(ptgt, pt_next); 889 free(ptgt, M_PMC); 890 891 /* if the PMC now lacks targets, send the owner a SIGIO */ 892 if (LIST_EMPTY(&pm->pm_targets)) { 893 p = pm->pm_owner->po_owner; 894 PROC_LOCK(p); 895 kern_psignal(p, SIGIO); 896 PROC_UNLOCK(p); 897 898 PMCDBG(PRC,SIG,2, "signalling proc=%p signal=%d", p, 899 SIGIO); 900 } 901 } 902 903 /* 904 * Check if PMC 'pm' may be attached to target process 't'. 905 */ 906 907 static int 908 pmc_can_attach(struct pmc *pm, struct proc *t) 909 { 910 struct proc *o; /* pmc owner */ 911 struct ucred *oc, *tc; /* owner, target credentials */ 912 int decline_attach, i; 913 914 /* 915 * A PMC's owner can always attach that PMC to itself. 916 */ 917 918 if ((o = pm->pm_owner->po_owner) == t) 919 return 0; 920 921 PROC_LOCK(o); 922 oc = o->p_ucred; 923 crhold(oc); 924 PROC_UNLOCK(o); 925 926 PROC_LOCK(t); 927 tc = t->p_ucred; 928 crhold(tc); 929 PROC_UNLOCK(t); 930 931 /* 932 * The effective uid of the PMC owner should match at least one 933 * of the {effective,real,saved} uids of the target process. 934 */ 935 936 decline_attach = oc->cr_uid != tc->cr_uid && 937 oc->cr_uid != tc->cr_svuid && 938 oc->cr_uid != tc->cr_ruid; 939 940 /* 941 * Every one of the target's group ids, must be in the owner's 942 * group list. 943 */ 944 for (i = 0; !decline_attach && i < tc->cr_ngroups; i++) 945 decline_attach = !groupmember(tc->cr_groups[i], oc); 946 947 /* check the read and saved gids too */ 948 if (decline_attach == 0) 949 decline_attach = !groupmember(tc->cr_rgid, oc) || 950 !groupmember(tc->cr_svgid, oc); 951 952 crfree(tc); 953 crfree(oc); 954 955 return !decline_attach; 956 } 957 958 /* 959 * Attach a process to a PMC. 960 */ 961 962 static int 963 pmc_attach_one_process(struct proc *p, struct pmc *pm) 964 { 965 int ri; 966 char *fullpath, *freepath; 967 struct pmc_process *pp; 968 969 sx_assert(&pmc_sx, SX_XLOCKED); 970 971 PMCDBG(PRC,ATT,2, "attach-one pm=%p ri=%d proc=%p (%d, %s)", pm, 972 PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm); 973 974 /* 975 * Locate the process descriptor corresponding to process 'p', 976 * allocating space as needed. 977 * 978 * Verify that rowindex 'pm_rowindex' is free in the process 979 * descriptor. 980 * 981 * If not, allocate space for a descriptor and link the 982 * process descriptor and PMC. 983 */ 984 ri = PMC_TO_ROWINDEX(pm); 985 986 if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_ALLOCATE)) == NULL) 987 return ENOMEM; 988 989 if (pp->pp_pmcs[ri].pp_pmc == pm) /* already present at slot [ri] */ 990 return EEXIST; 991 992 if (pp->pp_pmcs[ri].pp_pmc != NULL) 993 return EBUSY; 994 995 pmc_link_target_process(pm, pp); 996 997 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) && 998 (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) == 0) 999 pm->pm_flags |= PMC_F_NEEDS_LOGFILE; 1000 1001 pm->pm_flags |= PMC_F_ATTACH_DONE; /* mark as attached */ 1002 1003 /* issue an attach event to a configured log file */ 1004 if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE) { 1005 pmc_getfilename(p->p_textvp, &fullpath, &freepath); 1006 if (p->p_flag & P_KTHREAD) { 1007 fullpath = kernelname; 1008 freepath = NULL; 1009 } else 1010 pmclog_process_pmcattach(pm, p->p_pid, fullpath); 1011 if (freepath) 1012 free(freepath, M_TEMP); 1013 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) 1014 pmc_log_process_mappings(pm->pm_owner, p); 1015 } 1016 /* mark process as using HWPMCs */ 1017 PROC_LOCK(p); 1018 p->p_flag |= P_HWPMC; 1019 PROC_UNLOCK(p); 1020 1021 return 0; 1022 } 1023 1024 /* 1025 * Attach a process and optionally its children 1026 */ 1027 1028 static int 1029 pmc_attach_process(struct proc *p, struct pmc *pm) 1030 { 1031 int error; 1032 struct proc *top; 1033 1034 sx_assert(&pmc_sx, SX_XLOCKED); 1035 1036 PMCDBG(PRC,ATT,1, "attach pm=%p ri=%d proc=%p (%d, %s)", pm, 1037 PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm); 1038 1039 1040 /* 1041 * If this PMC successfully allowed a GETMSR operation 1042 * in the past, disallow further ATTACHes. 1043 */ 1044 1045 if ((pm->pm_flags & PMC_PP_ENABLE_MSR_ACCESS) != 0) 1046 return EPERM; 1047 1048 if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0) 1049 return pmc_attach_one_process(p, pm); 1050 1051 /* 1052 * Traverse all child processes, attaching them to 1053 * this PMC. 1054 */ 1055 1056 sx_slock(&proctree_lock); 1057 1058 top = p; 1059 1060 for (;;) { 1061 if ((error = pmc_attach_one_process(p, pm)) != 0) 1062 break; 1063 if (!LIST_EMPTY(&p->p_children)) 1064 p = LIST_FIRST(&p->p_children); 1065 else for (;;) { 1066 if (p == top) 1067 goto done; 1068 if (LIST_NEXT(p, p_sibling)) { 1069 p = LIST_NEXT(p, p_sibling); 1070 break; 1071 } 1072 p = p->p_pptr; 1073 } 1074 } 1075 1076 if (error) 1077 (void) pmc_detach_process(top, pm); 1078 1079 done: 1080 sx_sunlock(&proctree_lock); 1081 return error; 1082 } 1083 1084 /* 1085 * Detach a process from a PMC. If there are no other PMCs tracking 1086 * this process, remove the process structure from its hash table. If 1087 * 'flags' contains PMC_FLAG_REMOVE, then free the process structure. 1088 */ 1089 1090 static int 1091 pmc_detach_one_process(struct proc *p, struct pmc *pm, int flags) 1092 { 1093 int ri; 1094 struct pmc_process *pp; 1095 1096 sx_assert(&pmc_sx, SX_XLOCKED); 1097 1098 KASSERT(pm != NULL, 1099 ("[pmc,%d] null pm pointer", __LINE__)); 1100 1101 ri = PMC_TO_ROWINDEX(pm); 1102 1103 PMCDBG(PRC,ATT,2, "detach-one pm=%p ri=%d proc=%p (%d, %s) flags=0x%x", 1104 pm, ri, p, p->p_pid, p->p_comm, flags); 1105 1106 if ((pp = pmc_find_process_descriptor(p, 0)) == NULL) 1107 return ESRCH; 1108 1109 if (pp->pp_pmcs[ri].pp_pmc != pm) 1110 return EINVAL; 1111 1112 pmc_unlink_target_process(pm, pp); 1113 1114 /* Issue a detach entry if a log file is configured */ 1115 if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE) 1116 pmclog_process_pmcdetach(pm, p->p_pid); 1117 1118 /* 1119 * If there are no PMCs targetting this process, we remove its 1120 * descriptor from the target hash table and unset the P_HWPMC 1121 * flag in the struct proc. 1122 */ 1123 KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= (int) md->pmd_npmc, 1124 ("[pmc,%d] Illegal refcnt %d for process struct %p", 1125 __LINE__, pp->pp_refcnt, pp)); 1126 1127 if (pp->pp_refcnt != 0) /* still a target of some PMC */ 1128 return 0; 1129 1130 pmc_remove_process_descriptor(pp); 1131 1132 if (flags & PMC_FLAG_REMOVE) 1133 free(pp, M_PMC); 1134 1135 PROC_LOCK(p); 1136 p->p_flag &= ~P_HWPMC; 1137 PROC_UNLOCK(p); 1138 1139 return 0; 1140 } 1141 1142 /* 1143 * Detach a process and optionally its descendants from a PMC. 1144 */ 1145 1146 static int 1147 pmc_detach_process(struct proc *p, struct pmc *pm) 1148 { 1149 struct proc *top; 1150 1151 sx_assert(&pmc_sx, SX_XLOCKED); 1152 1153 PMCDBG(PRC,ATT,1, "detach pm=%p ri=%d proc=%p (%d, %s)", pm, 1154 PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm); 1155 1156 if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0) 1157 return pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE); 1158 1159 /* 1160 * Traverse all children, detaching them from this PMC. We 1161 * ignore errors since we could be detaching a PMC from a 1162 * partially attached proc tree. 1163 */ 1164 1165 sx_slock(&proctree_lock); 1166 1167 top = p; 1168 1169 for (;;) { 1170 (void) pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE); 1171 1172 if (!LIST_EMPTY(&p->p_children)) 1173 p = LIST_FIRST(&p->p_children); 1174 else for (;;) { 1175 if (p == top) 1176 goto done; 1177 if (LIST_NEXT(p, p_sibling)) { 1178 p = LIST_NEXT(p, p_sibling); 1179 break; 1180 } 1181 p = p->p_pptr; 1182 } 1183 } 1184 1185 done: 1186 sx_sunlock(&proctree_lock); 1187 1188 if (LIST_EMPTY(&pm->pm_targets)) 1189 pm->pm_flags &= ~PMC_F_ATTACH_DONE; 1190 1191 return 0; 1192 } 1193 1194 1195 /* 1196 * Thread context switch IN 1197 */ 1198 1199 static void 1200 pmc_process_csw_in(struct thread *td) 1201 { 1202 int cpu; 1203 unsigned int adjri, ri; 1204 struct pmc *pm; 1205 struct proc *p; 1206 struct pmc_cpu *pc; 1207 struct pmc_hw *phw; 1208 pmc_value_t newvalue; 1209 struct pmc_process *pp; 1210 struct pmc_classdep *pcd; 1211 1212 p = td->td_proc; 1213 1214 if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE)) == NULL) 1215 return; 1216 1217 KASSERT(pp->pp_proc == td->td_proc, 1218 ("[pmc,%d] not my thread state", __LINE__)); 1219 1220 critical_enter(); /* no preemption from this point */ 1221 1222 cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */ 1223 1224 PMCDBG(CSW,SWI,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p, 1225 p->p_pid, p->p_comm, pp); 1226 1227 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), 1228 ("[pmc,%d] wierd CPU id %d", __LINE__, cpu)); 1229 1230 pc = pmc_pcpu[cpu]; 1231 1232 for (ri = 0; ri < md->pmd_npmc; ri++) { 1233 1234 if ((pm = pp->pp_pmcs[ri].pp_pmc) == NULL) 1235 continue; 1236 1237 KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)), 1238 ("[pmc,%d] Target PMC in non-virtual mode (%d)", 1239 __LINE__, PMC_TO_MODE(pm))); 1240 1241 KASSERT(PMC_TO_ROWINDEX(pm) == ri, 1242 ("[pmc,%d] Row index mismatch pmc %d != ri %d", 1243 __LINE__, PMC_TO_ROWINDEX(pm), ri)); 1244 1245 /* 1246 * Only PMCs that are marked as 'RUNNING' need 1247 * be placed on hardware. 1248 */ 1249 1250 if (pm->pm_state != PMC_STATE_RUNNING) 1251 continue; 1252 1253 /* increment PMC runcount */ 1254 atomic_add_rel_int(&pm->pm_runcount, 1); 1255 1256 /* configure the HWPMC we are going to use. */ 1257 pcd = pmc_ri_to_classdep(md, ri, &adjri); 1258 pcd->pcd_config_pmc(cpu, adjri, pm); 1259 1260 phw = pc->pc_hwpmcs[ri]; 1261 1262 KASSERT(phw != NULL, 1263 ("[pmc,%d] null hw pointer", __LINE__)); 1264 1265 KASSERT(phw->phw_pmc == pm, 1266 ("[pmc,%d] hw->pmc %p != pmc %p", __LINE__, 1267 phw->phw_pmc, pm)); 1268 1269 /* 1270 * Write out saved value and start the PMC. 1271 * 1272 * Sampling PMCs use a per-process value, while 1273 * counting mode PMCs use a per-pmc value that is 1274 * inherited across descendants. 1275 */ 1276 if (PMC_TO_MODE(pm) == PMC_MODE_TS) { 1277 mtx_pool_lock_spin(pmc_mtxpool, pm); 1278 newvalue = PMC_PCPU_SAVED(cpu,ri) = 1279 pp->pp_pmcs[ri].pp_pmcval; 1280 mtx_pool_unlock_spin(pmc_mtxpool, pm); 1281 } else { 1282 KASSERT(PMC_TO_MODE(pm) == PMC_MODE_TC, 1283 ("[pmc,%d] illegal mode=%d", __LINE__, 1284 PMC_TO_MODE(pm))); 1285 mtx_pool_lock_spin(pmc_mtxpool, pm); 1286 newvalue = PMC_PCPU_SAVED(cpu, ri) = 1287 pm->pm_gv.pm_savedvalue; 1288 mtx_pool_unlock_spin(pmc_mtxpool, pm); 1289 } 1290 1291 PMCDBG(CSW,SWI,1,"cpu=%d ri=%d new=%jd", cpu, ri, newvalue); 1292 1293 pcd->pcd_write_pmc(cpu, adjri, newvalue); 1294 pcd->pcd_start_pmc(cpu, adjri); 1295 } 1296 1297 /* 1298 * perform any other architecture/cpu dependent thread 1299 * switch-in actions. 1300 */ 1301 1302 (void) (*md->pmd_switch_in)(pc, pp); 1303 1304 critical_exit(); 1305 1306 } 1307 1308 /* 1309 * Thread context switch OUT. 1310 */ 1311 1312 static void 1313 pmc_process_csw_out(struct thread *td) 1314 { 1315 int cpu; 1316 int64_t tmp; 1317 struct pmc *pm; 1318 struct proc *p; 1319 enum pmc_mode mode; 1320 struct pmc_cpu *pc; 1321 pmc_value_t newvalue; 1322 unsigned int adjri, ri; 1323 struct pmc_process *pp; 1324 struct pmc_classdep *pcd; 1325 1326 1327 /* 1328 * Locate our process descriptor; this may be NULL if 1329 * this process is exiting and we have already removed 1330 * the process from the target process table. 1331 * 1332 * Note that due to kernel preemption, multiple 1333 * context switches may happen while the process is 1334 * exiting. 1335 * 1336 * Note also that if the target process cannot be 1337 * found we still need to deconfigure any PMCs that 1338 * are currently running on hardware. 1339 */ 1340 1341 p = td->td_proc; 1342 pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE); 1343 1344 /* 1345 * save PMCs 1346 */ 1347 1348 critical_enter(); 1349 1350 cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */ 1351 1352 PMCDBG(CSW,SWO,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p, 1353 p->p_pid, p->p_comm, pp); 1354 1355 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), 1356 ("[pmc,%d wierd CPU id %d", __LINE__, cpu)); 1357 1358 pc = pmc_pcpu[cpu]; 1359 1360 /* 1361 * When a PMC gets unlinked from a target PMC, it will 1362 * be removed from the target's pp_pmc[] array. 1363 * 1364 * However, on a MP system, the target could have been 1365 * executing on another CPU at the time of the unlink. 1366 * So, at context switch OUT time, we need to look at 1367 * the hardware to determine if a PMC is scheduled on 1368 * it. 1369 */ 1370 1371 for (ri = 0; ri < md->pmd_npmc; ri++) { 1372 1373 pcd = pmc_ri_to_classdep(md, ri, &adjri); 1374 pm = NULL; 1375 (void) (*pcd->pcd_get_config)(cpu, adjri, &pm); 1376 1377 if (pm == NULL) /* nothing at this row index */ 1378 continue; 1379 1380 mode = PMC_TO_MODE(pm); 1381 if (!PMC_IS_VIRTUAL_MODE(mode)) 1382 continue; /* not a process virtual PMC */ 1383 1384 KASSERT(PMC_TO_ROWINDEX(pm) == ri, 1385 ("[pmc,%d] ri mismatch pmc(%d) ri(%d)", 1386 __LINE__, PMC_TO_ROWINDEX(pm), ri)); 1387 1388 /* Stop hardware if not already stopped */ 1389 if (pm->pm_stalled == 0) 1390 pcd->pcd_stop_pmc(cpu, adjri); 1391 1392 /* reduce this PMC's runcount */ 1393 atomic_subtract_rel_int(&pm->pm_runcount, 1); 1394 1395 /* 1396 * If this PMC is associated with this process, 1397 * save the reading. 1398 */ 1399 1400 if (pp != NULL && pp->pp_pmcs[ri].pp_pmc != NULL) { 1401 1402 KASSERT(pm == pp->pp_pmcs[ri].pp_pmc, 1403 ("[pmc,%d] pm %p != pp_pmcs[%d] %p", __LINE__, 1404 pm, ri, pp->pp_pmcs[ri].pp_pmc)); 1405 1406 KASSERT(pp->pp_refcnt > 0, 1407 ("[pmc,%d] pp refcnt = %d", __LINE__, 1408 pp->pp_refcnt)); 1409 1410 pcd->pcd_read_pmc(cpu, adjri, &newvalue); 1411 1412 tmp = newvalue - PMC_PCPU_SAVED(cpu,ri); 1413 1414 PMCDBG(CSW,SWO,1,"cpu=%d ri=%d tmp=%jd", cpu, ri, 1415 tmp); 1416 1417 if (mode == PMC_MODE_TS) { 1418 1419 /* 1420 * For sampling process-virtual PMCs, 1421 * we expect the count to be 1422 * decreasing as the 'value' 1423 * programmed into the PMC is the 1424 * number of events to be seen till 1425 * the next sampling interrupt. 1426 */ 1427 if (tmp < 0) 1428 tmp += pm->pm_sc.pm_reloadcount; 1429 mtx_pool_lock_spin(pmc_mtxpool, pm); 1430 pp->pp_pmcs[ri].pp_pmcval -= tmp; 1431 if ((int64_t) pp->pp_pmcs[ri].pp_pmcval < 0) 1432 pp->pp_pmcs[ri].pp_pmcval += 1433 pm->pm_sc.pm_reloadcount; 1434 mtx_pool_unlock_spin(pmc_mtxpool, pm); 1435 1436 } else { 1437 1438 /* 1439 * For counting process-virtual PMCs, 1440 * we expect the count to be 1441 * increasing monotonically, modulo a 64 1442 * bit wraparound. 1443 */ 1444 KASSERT((int64_t) tmp >= 0, 1445 ("[pmc,%d] negative increment cpu=%d " 1446 "ri=%d newvalue=%jx saved=%jx " 1447 "incr=%jx", __LINE__, cpu, ri, 1448 newvalue, PMC_PCPU_SAVED(cpu,ri), tmp)); 1449 1450 mtx_pool_lock_spin(pmc_mtxpool, pm); 1451 pm->pm_gv.pm_savedvalue += tmp; 1452 pp->pp_pmcs[ri].pp_pmcval += tmp; 1453 mtx_pool_unlock_spin(pmc_mtxpool, pm); 1454 1455 if (pm->pm_flags & PMC_F_LOG_PROCCSW) 1456 pmclog_process_proccsw(pm, pp, tmp); 1457 } 1458 } 1459 1460 /* mark hardware as free */ 1461 pcd->pcd_config_pmc(cpu, adjri, NULL); 1462 } 1463 1464 /* 1465 * perform any other architecture/cpu dependent thread 1466 * switch out functions. 1467 */ 1468 1469 (void) (*md->pmd_switch_out)(pc, pp); 1470 1471 critical_exit(); 1472 } 1473 1474 /* 1475 * A mapping change for a process. 1476 */ 1477 1478 static void 1479 pmc_process_mmap(struct thread *td, struct pmckern_map_in *pkm) 1480 { 1481 int ri; 1482 pid_t pid; 1483 char *fullpath, *freepath; 1484 const struct pmc *pm; 1485 struct pmc_owner *po; 1486 const struct pmc_process *pp; 1487 1488 freepath = fullpath = NULL; 1489 pmc_getfilename((struct vnode *) pkm->pm_file, &fullpath, &freepath); 1490 1491 pid = td->td_proc->p_pid; 1492 1493 /* Inform owners of all system-wide sampling PMCs. */ 1494 LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) 1495 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 1496 pmclog_process_map_in(po, pid, pkm->pm_address, fullpath); 1497 1498 if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL) 1499 goto done; 1500 1501 /* 1502 * Inform sampling PMC owners tracking this process. 1503 */ 1504 for (ri = 0; ri < md->pmd_npmc; ri++) 1505 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL && 1506 PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) 1507 pmclog_process_map_in(pm->pm_owner, 1508 pid, pkm->pm_address, fullpath); 1509 1510 done: 1511 if (freepath) 1512 free(freepath, M_TEMP); 1513 } 1514 1515 1516 /* 1517 * Log an munmap request. 1518 */ 1519 1520 static void 1521 pmc_process_munmap(struct thread *td, struct pmckern_map_out *pkm) 1522 { 1523 int ri; 1524 pid_t pid; 1525 struct pmc_owner *po; 1526 const struct pmc *pm; 1527 const struct pmc_process *pp; 1528 1529 pid = td->td_proc->p_pid; 1530 1531 LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) 1532 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 1533 pmclog_process_map_out(po, pid, pkm->pm_address, 1534 pkm->pm_address + pkm->pm_size); 1535 1536 if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL) 1537 return; 1538 1539 for (ri = 0; ri < md->pmd_npmc; ri++) 1540 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL && 1541 PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) 1542 pmclog_process_map_out(pm->pm_owner, pid, 1543 pkm->pm_address, pkm->pm_address + pkm->pm_size); 1544 } 1545 1546 /* 1547 * Log mapping information about the kernel. 1548 */ 1549 1550 static void 1551 pmc_log_kernel_mappings(struct pmc *pm) 1552 { 1553 struct pmc_owner *po; 1554 struct pmckern_map_in *km, *kmbase; 1555 1556 sx_assert(&pmc_sx, SX_LOCKED); 1557 KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)), 1558 ("[pmc,%d] non-sampling PMC (%p) desires mapping information", 1559 __LINE__, (void *) pm)); 1560 1561 po = pm->pm_owner; 1562 1563 if (po->po_flags & PMC_PO_INITIAL_MAPPINGS_DONE) 1564 return; 1565 1566 /* 1567 * Log the current set of kernel modules. 1568 */ 1569 kmbase = linker_hwpmc_list_objects(); 1570 for (km = kmbase; km->pm_file != NULL; km++) { 1571 PMCDBG(LOG,REG,1,"%s %p", (char *) km->pm_file, 1572 (void *) km->pm_address); 1573 pmclog_process_map_in(po, (pid_t) -1, km->pm_address, 1574 km->pm_file); 1575 } 1576 free(kmbase, M_LINKER); 1577 1578 po->po_flags |= PMC_PO_INITIAL_MAPPINGS_DONE; 1579 } 1580 1581 /* 1582 * Log the mappings for a single process. 1583 */ 1584 1585 static void 1586 pmc_log_process_mappings(struct pmc_owner *po, struct proc *p) 1587 { 1588 vm_map_t map; 1589 struct vnode *vp; 1590 struct vmspace *vm; 1591 vm_map_entry_t entry; 1592 vm_offset_t last_end; 1593 u_int last_timestamp; 1594 struct vnode *last_vp; 1595 vm_offset_t start_addr; 1596 vm_object_t obj, lobj, tobj; 1597 char *fullpath, *freepath; 1598 1599 last_vp = NULL; 1600 last_end = (vm_offset_t) 0; 1601 fullpath = freepath = NULL; 1602 1603 if ((vm = vmspace_acquire_ref(p)) == NULL) 1604 return; 1605 1606 map = &vm->vm_map; 1607 vm_map_lock_read(map); 1608 1609 for (entry = map->header.next; entry != &map->header; entry = entry->next) { 1610 1611 if (entry == NULL) { 1612 PMCDBG(LOG,OPS,2, "hwpmc: vm_map entry unexpectedly " 1613 "NULL! pid=%d vm_map=%p\n", p->p_pid, map); 1614 break; 1615 } 1616 1617 /* 1618 * We only care about executable map entries. 1619 */ 1620 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) || 1621 !(entry->protection & VM_PROT_EXECUTE) || 1622 (entry->object.vm_object == NULL)) { 1623 continue; 1624 } 1625 1626 obj = entry->object.vm_object; 1627 VM_OBJECT_RLOCK(obj); 1628 1629 /* 1630 * Walk the backing_object list to find the base 1631 * (non-shadowed) vm_object. 1632 */ 1633 for (lobj = tobj = obj; tobj != NULL; tobj = tobj->backing_object) { 1634 if (tobj != obj) 1635 VM_OBJECT_RLOCK(tobj); 1636 if (lobj != obj) 1637 VM_OBJECT_RUNLOCK(lobj); 1638 lobj = tobj; 1639 } 1640 1641 /* 1642 * At this point lobj is the base vm_object and it is locked. 1643 */ 1644 if (lobj == NULL) { 1645 PMCDBG(LOG,OPS,2, "hwpmc: lobj unexpectedly NULL! pid=%d " 1646 "vm_map=%p vm_obj=%p\n", p->p_pid, map, obj); 1647 VM_OBJECT_RUNLOCK(obj); 1648 continue; 1649 } 1650 1651 if (lobj->type != OBJT_VNODE || lobj->handle == NULL) { 1652 if (lobj != obj) 1653 VM_OBJECT_RUNLOCK(lobj); 1654 VM_OBJECT_RUNLOCK(obj); 1655 continue; 1656 } 1657 1658 /* 1659 * Skip contiguous regions that point to the same 1660 * vnode, so we don't emit redundant MAP-IN 1661 * directives. 1662 */ 1663 if (entry->start == last_end && lobj->handle == last_vp) { 1664 last_end = entry->end; 1665 if (lobj != obj) 1666 VM_OBJECT_RUNLOCK(lobj); 1667 VM_OBJECT_RUNLOCK(obj); 1668 continue; 1669 } 1670 1671 /* 1672 * We don't want to keep the proc's vm_map or this 1673 * vm_object locked while we walk the pathname, since 1674 * vn_fullpath() can sleep. However, if we drop the 1675 * lock, it's possible for concurrent activity to 1676 * modify the vm_map list. To protect against this, 1677 * we save the vm_map timestamp before we release the 1678 * lock, and check it after we reacquire the lock 1679 * below. 1680 */ 1681 start_addr = entry->start; 1682 last_end = entry->end; 1683 last_timestamp = map->timestamp; 1684 vm_map_unlock_read(map); 1685 1686 vp = lobj->handle; 1687 vref(vp); 1688 if (lobj != obj) 1689 VM_OBJECT_RUNLOCK(lobj); 1690 1691 VM_OBJECT_RUNLOCK(obj); 1692 1693 freepath = NULL; 1694 pmc_getfilename(vp, &fullpath, &freepath); 1695 last_vp = vp; 1696 1697 vrele(vp); 1698 1699 vp = NULL; 1700 pmclog_process_map_in(po, p->p_pid, start_addr, fullpath); 1701 if (freepath) 1702 free(freepath, M_TEMP); 1703 1704 vm_map_lock_read(map); 1705 1706 /* 1707 * If our saved timestamp doesn't match, this means 1708 * that the vm_map was modified out from under us and 1709 * we can't trust our current "entry" pointer. Do a 1710 * new lookup for this entry. If there is no entry 1711 * for this address range, vm_map_lookup_entry() will 1712 * return the previous one, so we always want to go to 1713 * entry->next on the next loop iteration. 1714 * 1715 * There is an edge condition here that can occur if 1716 * there is no entry at or before this address. In 1717 * this situation, vm_map_lookup_entry returns 1718 * &map->header, which would cause our loop to abort 1719 * without processing the rest of the map. However, 1720 * in practice this will never happen for process 1721 * vm_map. This is because the executable's text 1722 * segment is the first mapping in the proc's address 1723 * space, and this mapping is never removed until the 1724 * process exits, so there will always be a non-header 1725 * entry at or before the requested address for 1726 * vm_map_lookup_entry to return. 1727 */ 1728 if (map->timestamp != last_timestamp) 1729 vm_map_lookup_entry(map, last_end - 1, &entry); 1730 } 1731 1732 vm_map_unlock_read(map); 1733 vmspace_free(vm); 1734 return; 1735 } 1736 1737 /* 1738 * Log mappings for all processes in the system. 1739 */ 1740 1741 static void 1742 pmc_log_all_process_mappings(struct pmc_owner *po) 1743 { 1744 struct proc *p, *top; 1745 1746 sx_assert(&pmc_sx, SX_XLOCKED); 1747 1748 if ((p = pfind(1)) == NULL) 1749 panic("[pmc,%d] Cannot find init", __LINE__); 1750 1751 PROC_UNLOCK(p); 1752 1753 sx_slock(&proctree_lock); 1754 1755 top = p; 1756 1757 for (;;) { 1758 pmc_log_process_mappings(po, p); 1759 if (!LIST_EMPTY(&p->p_children)) 1760 p = LIST_FIRST(&p->p_children); 1761 else for (;;) { 1762 if (p == top) 1763 goto done; 1764 if (LIST_NEXT(p, p_sibling)) { 1765 p = LIST_NEXT(p, p_sibling); 1766 break; 1767 } 1768 p = p->p_pptr; 1769 } 1770 } 1771 done: 1772 sx_sunlock(&proctree_lock); 1773 } 1774 1775 /* 1776 * The 'hook' invoked from the kernel proper 1777 */ 1778 1779 1780 #ifdef DEBUG 1781 const char *pmc_hooknames[] = { 1782 /* these strings correspond to PMC_FN_* in <sys/pmckern.h> */ 1783 "", 1784 "EXEC", 1785 "CSW-IN", 1786 "CSW-OUT", 1787 "SAMPLE", 1788 "UNUSED1", 1789 "UNUSED2", 1790 "MMAP", 1791 "MUNMAP", 1792 "CALLCHAIN-NMI", 1793 "CALLCHAIN-SOFT", 1794 "SOFTSAMPLING" 1795 }; 1796 #endif 1797 1798 static int 1799 pmc_hook_handler(struct thread *td, int function, void *arg) 1800 { 1801 1802 PMCDBG(MOD,PMH,1, "hook td=%p func=%d \"%s\" arg=%p", td, function, 1803 pmc_hooknames[function], arg); 1804 1805 switch (function) 1806 { 1807 1808 /* 1809 * Process exec() 1810 */ 1811 1812 case PMC_FN_PROCESS_EXEC: 1813 { 1814 char *fullpath, *freepath; 1815 unsigned int ri; 1816 int is_using_hwpmcs; 1817 struct pmc *pm; 1818 struct proc *p; 1819 struct pmc_owner *po; 1820 struct pmc_process *pp; 1821 struct pmckern_procexec *pk; 1822 1823 sx_assert(&pmc_sx, SX_XLOCKED); 1824 1825 p = td->td_proc; 1826 pmc_getfilename(p->p_textvp, &fullpath, &freepath); 1827 1828 pk = (struct pmckern_procexec *) arg; 1829 1830 /* Inform owners of SS mode PMCs of the exec event. */ 1831 LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) 1832 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 1833 pmclog_process_procexec(po, PMC_ID_INVALID, 1834 p->p_pid, pk->pm_entryaddr, fullpath); 1835 1836 PROC_LOCK(p); 1837 is_using_hwpmcs = p->p_flag & P_HWPMC; 1838 PROC_UNLOCK(p); 1839 1840 if (!is_using_hwpmcs) { 1841 if (freepath) 1842 free(freepath, M_TEMP); 1843 break; 1844 } 1845 1846 /* 1847 * PMCs are not inherited across an exec(): remove any 1848 * PMCs that this process is the owner of. 1849 */ 1850 1851 if ((po = pmc_find_owner_descriptor(p)) != NULL) { 1852 pmc_remove_owner(po); 1853 pmc_destroy_owner_descriptor(po); 1854 } 1855 1856 /* 1857 * If the process being exec'ed is not the target of any 1858 * PMC, we are done. 1859 */ 1860 if ((pp = pmc_find_process_descriptor(p, 0)) == NULL) { 1861 if (freepath) 1862 free(freepath, M_TEMP); 1863 break; 1864 } 1865 1866 /* 1867 * Log the exec event to all monitoring owners. Skip 1868 * owners who have already recieved the event because 1869 * they had system sampling PMCs active. 1870 */ 1871 for (ri = 0; ri < md->pmd_npmc; ri++) 1872 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) { 1873 po = pm->pm_owner; 1874 if (po->po_sscount == 0 && 1875 po->po_flags & PMC_PO_OWNS_LOGFILE) 1876 pmclog_process_procexec(po, pm->pm_id, 1877 p->p_pid, pk->pm_entryaddr, 1878 fullpath); 1879 } 1880 1881 if (freepath) 1882 free(freepath, M_TEMP); 1883 1884 1885 PMCDBG(PRC,EXC,1, "exec proc=%p (%d, %s) cred-changed=%d", 1886 p, p->p_pid, p->p_comm, pk->pm_credentialschanged); 1887 1888 if (pk->pm_credentialschanged == 0) /* no change */ 1889 break; 1890 1891 /* 1892 * If the newly exec()'ed process has a different credential 1893 * than before, allow it to be the target of a PMC only if 1894 * the PMC's owner has sufficient priviledge. 1895 */ 1896 1897 for (ri = 0; ri < md->pmd_npmc; ri++) 1898 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) 1899 if (pmc_can_attach(pm, td->td_proc) != 0) 1900 pmc_detach_one_process(td->td_proc, 1901 pm, PMC_FLAG_NONE); 1902 1903 KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= (int) md->pmd_npmc, 1904 ("[pmc,%d] Illegal ref count %d on pp %p", __LINE__, 1905 pp->pp_refcnt, pp)); 1906 1907 /* 1908 * If this process is no longer the target of any 1909 * PMCs, we can remove the process entry and free 1910 * up space. 1911 */ 1912 1913 if (pp->pp_refcnt == 0) { 1914 pmc_remove_process_descriptor(pp); 1915 free(pp, M_PMC); 1916 break; 1917 } 1918 1919 } 1920 break; 1921 1922 case PMC_FN_CSW_IN: 1923 pmc_process_csw_in(td); 1924 break; 1925 1926 case PMC_FN_CSW_OUT: 1927 pmc_process_csw_out(td); 1928 break; 1929 1930 /* 1931 * Process accumulated PC samples. 1932 * 1933 * This function is expected to be called by hardclock() for 1934 * each CPU that has accumulated PC samples. 1935 * 1936 * This function is to be executed on the CPU whose samples 1937 * are being processed. 1938 */ 1939 case PMC_FN_DO_SAMPLES: 1940 1941 /* 1942 * Clear the cpu specific bit in the CPU mask before 1943 * do the rest of the processing. If the NMI handler 1944 * gets invoked after the "atomic_clear_int()" call 1945 * below but before "pmc_process_samples()" gets 1946 * around to processing the interrupt, then we will 1947 * come back here at the next hardclock() tick (and 1948 * may find nothing to do if "pmc_process_samples()" 1949 * had already processed the interrupt). We don't 1950 * lose the interrupt sample. 1951 */ 1952 CPU_CLR_ATOMIC(PCPU_GET(cpuid), &pmc_cpumask); 1953 pmc_process_samples(PCPU_GET(cpuid), PMC_HR); 1954 pmc_process_samples(PCPU_GET(cpuid), PMC_SR); 1955 break; 1956 1957 case PMC_FN_MMAP: 1958 sx_assert(&pmc_sx, SX_LOCKED); 1959 pmc_process_mmap(td, (struct pmckern_map_in *) arg); 1960 break; 1961 1962 case PMC_FN_MUNMAP: 1963 sx_assert(&pmc_sx, SX_LOCKED); 1964 pmc_process_munmap(td, (struct pmckern_map_out *) arg); 1965 break; 1966 1967 case PMC_FN_USER_CALLCHAIN: 1968 /* 1969 * Record a call chain. 1970 */ 1971 KASSERT(td == curthread, ("[pmc,%d] td != curthread", 1972 __LINE__)); 1973 1974 pmc_capture_user_callchain(PCPU_GET(cpuid), PMC_HR, 1975 (struct trapframe *) arg); 1976 td->td_pflags &= ~TDP_CALLCHAIN; 1977 break; 1978 1979 case PMC_FN_USER_CALLCHAIN_SOFT: 1980 /* 1981 * Record a call chain. 1982 */ 1983 KASSERT(td == curthread, ("[pmc,%d] td != curthread", 1984 __LINE__)); 1985 pmc_capture_user_callchain(PCPU_GET(cpuid), PMC_SR, 1986 (struct trapframe *) arg); 1987 td->td_pflags &= ~TDP_CALLCHAIN; 1988 break; 1989 1990 case PMC_FN_SOFT_SAMPLING: 1991 /* 1992 * Call soft PMC sampling intr. 1993 */ 1994 pmc_soft_intr((struct pmckern_soft *) arg); 1995 break; 1996 1997 default: 1998 #ifdef DEBUG 1999 KASSERT(0, ("[pmc,%d] unknown hook %d\n", __LINE__, function)); 2000 #endif 2001 break; 2002 2003 } 2004 2005 return 0; 2006 } 2007 2008 /* 2009 * allocate a 'struct pmc_owner' descriptor in the owner hash table. 2010 */ 2011 2012 static struct pmc_owner * 2013 pmc_allocate_owner_descriptor(struct proc *p) 2014 { 2015 uint32_t hindex; 2016 struct pmc_owner *po; 2017 struct pmc_ownerhash *poh; 2018 2019 hindex = PMC_HASH_PTR(p, pmc_ownerhashmask); 2020 poh = &pmc_ownerhash[hindex]; 2021 2022 /* allocate space for N pointers and one descriptor struct */ 2023 po = malloc(sizeof(struct pmc_owner), M_PMC, M_WAITOK|M_ZERO); 2024 po->po_owner = p; 2025 LIST_INSERT_HEAD(poh, po, po_next); /* insert into hash table */ 2026 2027 TAILQ_INIT(&po->po_logbuffers); 2028 mtx_init(&po->po_mtx, "pmc-owner-mtx", "pmc-per-proc", MTX_SPIN); 2029 2030 PMCDBG(OWN,ALL,1, "allocate-owner proc=%p (%d, %s) pmc-owner=%p", 2031 p, p->p_pid, p->p_comm, po); 2032 2033 return po; 2034 } 2035 2036 static void 2037 pmc_destroy_owner_descriptor(struct pmc_owner *po) 2038 { 2039 2040 PMCDBG(OWN,REL,1, "destroy-owner po=%p proc=%p (%d, %s)", 2041 po, po->po_owner, po->po_owner->p_pid, po->po_owner->p_comm); 2042 2043 mtx_destroy(&po->po_mtx); 2044 free(po, M_PMC); 2045 } 2046 2047 /* 2048 * find the descriptor corresponding to process 'p', adding or removing it 2049 * as specified by 'mode'. 2050 */ 2051 2052 static struct pmc_process * 2053 pmc_find_process_descriptor(struct proc *p, uint32_t mode) 2054 { 2055 uint32_t hindex; 2056 struct pmc_process *pp, *ppnew; 2057 struct pmc_processhash *pph; 2058 2059 hindex = PMC_HASH_PTR(p, pmc_processhashmask); 2060 pph = &pmc_processhash[hindex]; 2061 2062 ppnew = NULL; 2063 2064 /* 2065 * Pre-allocate memory in the FIND_ALLOCATE case since we 2066 * cannot call malloc(9) once we hold a spin lock. 2067 */ 2068 if (mode & PMC_FLAG_ALLOCATE) 2069 ppnew = malloc(sizeof(struct pmc_process) + md->pmd_npmc * 2070 sizeof(struct pmc_targetstate), M_PMC, M_WAITOK|M_ZERO); 2071 2072 mtx_lock_spin(&pmc_processhash_mtx); 2073 LIST_FOREACH(pp, pph, pp_next) 2074 if (pp->pp_proc == p) 2075 break; 2076 2077 if ((mode & PMC_FLAG_REMOVE) && pp != NULL) 2078 LIST_REMOVE(pp, pp_next); 2079 2080 if ((mode & PMC_FLAG_ALLOCATE) && pp == NULL && 2081 ppnew != NULL) { 2082 ppnew->pp_proc = p; 2083 LIST_INSERT_HEAD(pph, ppnew, pp_next); 2084 pp = ppnew; 2085 ppnew = NULL; 2086 } 2087 mtx_unlock_spin(&pmc_processhash_mtx); 2088 2089 if (pp != NULL && ppnew != NULL) 2090 free(ppnew, M_PMC); 2091 2092 return pp; 2093 } 2094 2095 /* 2096 * remove a process descriptor from the process hash table. 2097 */ 2098 2099 static void 2100 pmc_remove_process_descriptor(struct pmc_process *pp) 2101 { 2102 KASSERT(pp->pp_refcnt == 0, 2103 ("[pmc,%d] Removing process descriptor %p with count %d", 2104 __LINE__, pp, pp->pp_refcnt)); 2105 2106 mtx_lock_spin(&pmc_processhash_mtx); 2107 LIST_REMOVE(pp, pp_next); 2108 mtx_unlock_spin(&pmc_processhash_mtx); 2109 } 2110 2111 2112 /* 2113 * find an owner descriptor corresponding to proc 'p' 2114 */ 2115 2116 static struct pmc_owner * 2117 pmc_find_owner_descriptor(struct proc *p) 2118 { 2119 uint32_t hindex; 2120 struct pmc_owner *po; 2121 struct pmc_ownerhash *poh; 2122 2123 hindex = PMC_HASH_PTR(p, pmc_ownerhashmask); 2124 poh = &pmc_ownerhash[hindex]; 2125 2126 po = NULL; 2127 LIST_FOREACH(po, poh, po_next) 2128 if (po->po_owner == p) 2129 break; 2130 2131 PMCDBG(OWN,FND,1, "find-owner proc=%p (%d, %s) hindex=0x%x -> " 2132 "pmc-owner=%p", p, p->p_pid, p->p_comm, hindex, po); 2133 2134 return po; 2135 } 2136 2137 /* 2138 * pmc_allocate_pmc_descriptor 2139 * 2140 * Allocate a pmc descriptor and initialize its 2141 * fields. 2142 */ 2143 2144 static struct pmc * 2145 pmc_allocate_pmc_descriptor(void) 2146 { 2147 struct pmc *pmc; 2148 2149 pmc = malloc(sizeof(struct pmc), M_PMC, M_WAITOK|M_ZERO); 2150 2151 PMCDBG(PMC,ALL,1, "allocate-pmc -> pmc=%p", pmc); 2152 2153 return pmc; 2154 } 2155 2156 /* 2157 * Destroy a pmc descriptor. 2158 */ 2159 2160 static void 2161 pmc_destroy_pmc_descriptor(struct pmc *pm) 2162 { 2163 (void) pm; 2164 2165 #ifdef DEBUG 2166 KASSERT(pm->pm_state == PMC_STATE_DELETED || 2167 pm->pm_state == PMC_STATE_FREE, 2168 ("[pmc,%d] destroying non-deleted PMC", __LINE__)); 2169 KASSERT(LIST_EMPTY(&pm->pm_targets), 2170 ("[pmc,%d] destroying pmc with targets", __LINE__)); 2171 KASSERT(pm->pm_owner == NULL, 2172 ("[pmc,%d] destroying pmc attached to an owner", __LINE__)); 2173 KASSERT(pm->pm_runcount == 0, 2174 ("[pmc,%d] pmc has non-zero run count %d", __LINE__, 2175 pm->pm_runcount)); 2176 #endif 2177 } 2178 2179 static void 2180 pmc_wait_for_pmc_idle(struct pmc *pm) 2181 { 2182 #ifdef DEBUG 2183 volatile int maxloop; 2184 2185 maxloop = 100 * pmc_cpu_max(); 2186 #endif 2187 /* 2188 * Loop (with a forced context switch) till the PMC's runcount 2189 * comes down to zero. 2190 */ 2191 while (atomic_load_acq_32(&pm->pm_runcount) > 0) { 2192 #ifdef DEBUG 2193 maxloop--; 2194 KASSERT(maxloop > 0, 2195 ("[pmc,%d] (ri%d, rc%d) waiting too long for " 2196 "pmc to be free", __LINE__, 2197 PMC_TO_ROWINDEX(pm), pm->pm_runcount)); 2198 #endif 2199 pmc_force_context_switch(); 2200 } 2201 } 2202 2203 /* 2204 * This function does the following things: 2205 * 2206 * - detaches the PMC from hardware 2207 * - unlinks all target threads that were attached to it 2208 * - removes the PMC from its owner's list 2209 * - destroy's the PMC private mutex 2210 * 2211 * Once this function completes, the given pmc pointer can be safely 2212 * FREE'd by the caller. 2213 */ 2214 2215 static void 2216 pmc_release_pmc_descriptor(struct pmc *pm) 2217 { 2218 enum pmc_mode mode; 2219 struct pmc_hw *phw; 2220 u_int adjri, ri, cpu; 2221 struct pmc_owner *po; 2222 struct pmc_binding pb; 2223 struct pmc_process *pp; 2224 struct pmc_classdep *pcd; 2225 struct pmc_target *ptgt, *tmp; 2226 2227 sx_assert(&pmc_sx, SX_XLOCKED); 2228 2229 KASSERT(pm, ("[pmc,%d] null pmc", __LINE__)); 2230 2231 ri = PMC_TO_ROWINDEX(pm); 2232 pcd = pmc_ri_to_classdep(md, ri, &adjri); 2233 mode = PMC_TO_MODE(pm); 2234 2235 PMCDBG(PMC,REL,1, "release-pmc pmc=%p ri=%d mode=%d", pm, ri, 2236 mode); 2237 2238 /* 2239 * First, we take the PMC off hardware. 2240 */ 2241 cpu = 0; 2242 if (PMC_IS_SYSTEM_MODE(mode)) { 2243 2244 /* 2245 * A system mode PMC runs on a specific CPU. Switch 2246 * to this CPU and turn hardware off. 2247 */ 2248 pmc_save_cpu_binding(&pb); 2249 2250 cpu = PMC_TO_CPU(pm); 2251 2252 pmc_select_cpu(cpu); 2253 2254 /* switch off non-stalled CPUs */ 2255 if (pm->pm_state == PMC_STATE_RUNNING && 2256 pm->pm_stalled == 0) { 2257 2258 phw = pmc_pcpu[cpu]->pc_hwpmcs[ri]; 2259 2260 KASSERT(phw->phw_pmc == pm, 2261 ("[pmc, %d] pmc ptr ri(%d) hw(%p) pm(%p)", 2262 __LINE__, ri, phw->phw_pmc, pm)); 2263 PMCDBG(PMC,REL,2, "stopping cpu=%d ri=%d", cpu, ri); 2264 2265 critical_enter(); 2266 pcd->pcd_stop_pmc(cpu, adjri); 2267 critical_exit(); 2268 } 2269 2270 PMCDBG(PMC,REL,2, "decfg cpu=%d ri=%d", cpu, ri); 2271 2272 critical_enter(); 2273 pcd->pcd_config_pmc(cpu, adjri, NULL); 2274 critical_exit(); 2275 2276 /* adjust the global and process count of SS mode PMCs */ 2277 if (mode == PMC_MODE_SS && pm->pm_state == PMC_STATE_RUNNING) { 2278 po = pm->pm_owner; 2279 po->po_sscount--; 2280 if (po->po_sscount == 0) { 2281 atomic_subtract_rel_int(&pmc_ss_count, 1); 2282 LIST_REMOVE(po, po_ssnext); 2283 } 2284 } 2285 2286 pm->pm_state = PMC_STATE_DELETED; 2287 2288 pmc_restore_cpu_binding(&pb); 2289 2290 /* 2291 * We could have references to this PMC structure in 2292 * the per-cpu sample queues. Wait for the queue to 2293 * drain. 2294 */ 2295 pmc_wait_for_pmc_idle(pm); 2296 2297 } else if (PMC_IS_VIRTUAL_MODE(mode)) { 2298 2299 /* 2300 * A virtual PMC could be running on multiple CPUs at 2301 * a given instant. 2302 * 2303 * By marking its state as DELETED, we ensure that 2304 * this PMC is never further scheduled on hardware. 2305 * 2306 * Then we wait till all CPUs are done with this PMC. 2307 */ 2308 pm->pm_state = PMC_STATE_DELETED; 2309 2310 2311 /* Wait for the PMCs runcount to come to zero. */ 2312 pmc_wait_for_pmc_idle(pm); 2313 2314 /* 2315 * At this point the PMC is off all CPUs and cannot be 2316 * freshly scheduled onto a CPU. It is now safe to 2317 * unlink all targets from this PMC. If a 2318 * process-record's refcount falls to zero, we remove 2319 * it from the hash table. The module-wide SX lock 2320 * protects us from races. 2321 */ 2322 LIST_FOREACH_SAFE(ptgt, &pm->pm_targets, pt_next, tmp) { 2323 pp = ptgt->pt_process; 2324 pmc_unlink_target_process(pm, pp); /* frees 'ptgt' */ 2325 2326 PMCDBG(PMC,REL,3, "pp->refcnt=%d", pp->pp_refcnt); 2327 2328 /* 2329 * If the target process record shows that no 2330 * PMCs are attached to it, reclaim its space. 2331 */ 2332 2333 if (pp->pp_refcnt == 0) { 2334 pmc_remove_process_descriptor(pp); 2335 free(pp, M_PMC); 2336 } 2337 } 2338 2339 cpu = curthread->td_oncpu; /* setup cpu for pmd_release() */ 2340 2341 } 2342 2343 /* 2344 * Release any MD resources 2345 */ 2346 (void) pcd->pcd_release_pmc(cpu, adjri, pm); 2347 2348 /* 2349 * Update row disposition 2350 */ 2351 2352 if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm))) 2353 PMC_UNMARK_ROW_STANDALONE(ri); 2354 else 2355 PMC_UNMARK_ROW_THREAD(ri); 2356 2357 /* unlink from the owner's list */ 2358 if (pm->pm_owner) { 2359 LIST_REMOVE(pm, pm_next); 2360 pm->pm_owner = NULL; 2361 } 2362 2363 pmc_destroy_pmc_descriptor(pm); 2364 } 2365 2366 /* 2367 * Register an owner and a pmc. 2368 */ 2369 2370 static int 2371 pmc_register_owner(struct proc *p, struct pmc *pmc) 2372 { 2373 struct pmc_owner *po; 2374 2375 sx_assert(&pmc_sx, SX_XLOCKED); 2376 2377 if ((po = pmc_find_owner_descriptor(p)) == NULL) 2378 if ((po = pmc_allocate_owner_descriptor(p)) == NULL) 2379 return ENOMEM; 2380 2381 KASSERT(pmc->pm_owner == NULL, 2382 ("[pmc,%d] attempting to own an initialized PMC", __LINE__)); 2383 pmc->pm_owner = po; 2384 2385 LIST_INSERT_HEAD(&po->po_pmcs, pmc, pm_next); 2386 2387 PROC_LOCK(p); 2388 p->p_flag |= P_HWPMC; 2389 PROC_UNLOCK(p); 2390 2391 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 2392 pmclog_process_pmcallocate(pmc); 2393 2394 PMCDBG(PMC,REG,1, "register-owner pmc-owner=%p pmc=%p", 2395 po, pmc); 2396 2397 return 0; 2398 } 2399 2400 /* 2401 * Return the current row disposition: 2402 * == 0 => FREE 2403 * > 0 => PROCESS MODE 2404 * < 0 => SYSTEM MODE 2405 */ 2406 2407 int 2408 pmc_getrowdisp(int ri) 2409 { 2410 return pmc_pmcdisp[ri]; 2411 } 2412 2413 /* 2414 * Check if a PMC at row index 'ri' can be allocated to the current 2415 * process. 2416 * 2417 * Allocation can fail if: 2418 * - the current process is already being profiled by a PMC at index 'ri', 2419 * attached to it via OP_PMCATTACH. 2420 * - the current process has already allocated a PMC at index 'ri' 2421 * via OP_ALLOCATE. 2422 */ 2423 2424 static int 2425 pmc_can_allocate_rowindex(struct proc *p, unsigned int ri, int cpu) 2426 { 2427 enum pmc_mode mode; 2428 struct pmc *pm; 2429 struct pmc_owner *po; 2430 struct pmc_process *pp; 2431 2432 PMCDBG(PMC,ALR,1, "can-allocate-rowindex proc=%p (%d, %s) ri=%d " 2433 "cpu=%d", p, p->p_pid, p->p_comm, ri, cpu); 2434 2435 /* 2436 * We shouldn't have already allocated a process-mode PMC at 2437 * row index 'ri'. 2438 * 2439 * We shouldn't have allocated a system-wide PMC on the same 2440 * CPU and same RI. 2441 */ 2442 if ((po = pmc_find_owner_descriptor(p)) != NULL) 2443 LIST_FOREACH(pm, &po->po_pmcs, pm_next) { 2444 if (PMC_TO_ROWINDEX(pm) == ri) { 2445 mode = PMC_TO_MODE(pm); 2446 if (PMC_IS_VIRTUAL_MODE(mode)) 2447 return EEXIST; 2448 if (PMC_IS_SYSTEM_MODE(mode) && 2449 (int) PMC_TO_CPU(pm) == cpu) 2450 return EEXIST; 2451 } 2452 } 2453 2454 /* 2455 * We also shouldn't be the target of any PMC at this index 2456 * since otherwise a PMC_ATTACH to ourselves will fail. 2457 */ 2458 if ((pp = pmc_find_process_descriptor(p, 0)) != NULL) 2459 if (pp->pp_pmcs[ri].pp_pmc) 2460 return EEXIST; 2461 2462 PMCDBG(PMC,ALR,2, "can-allocate-rowindex proc=%p (%d, %s) ri=%d ok", 2463 p, p->p_pid, p->p_comm, ri); 2464 2465 return 0; 2466 } 2467 2468 /* 2469 * Check if a given PMC at row index 'ri' can be currently used in 2470 * mode 'mode'. 2471 */ 2472 2473 static int 2474 pmc_can_allocate_row(int ri, enum pmc_mode mode) 2475 { 2476 enum pmc_disp disp; 2477 2478 sx_assert(&pmc_sx, SX_XLOCKED); 2479 2480 PMCDBG(PMC,ALR,1, "can-allocate-row ri=%d mode=%d", ri, mode); 2481 2482 if (PMC_IS_SYSTEM_MODE(mode)) 2483 disp = PMC_DISP_STANDALONE; 2484 else 2485 disp = PMC_DISP_THREAD; 2486 2487 /* 2488 * check disposition for PMC row 'ri': 2489 * 2490 * Expected disposition Row-disposition Result 2491 * 2492 * STANDALONE STANDALONE or FREE proceed 2493 * STANDALONE THREAD fail 2494 * THREAD THREAD or FREE proceed 2495 * THREAD STANDALONE fail 2496 */ 2497 2498 if (!PMC_ROW_DISP_IS_FREE(ri) && 2499 !(disp == PMC_DISP_THREAD && PMC_ROW_DISP_IS_THREAD(ri)) && 2500 !(disp == PMC_DISP_STANDALONE && PMC_ROW_DISP_IS_STANDALONE(ri))) 2501 return EBUSY; 2502 2503 /* 2504 * All OK 2505 */ 2506 2507 PMCDBG(PMC,ALR,2, "can-allocate-row ri=%d mode=%d ok", ri, mode); 2508 2509 return 0; 2510 2511 } 2512 2513 /* 2514 * Find a PMC descriptor with user handle 'pmcid' for thread 'td'. 2515 */ 2516 2517 static struct pmc * 2518 pmc_find_pmc_descriptor_in_process(struct pmc_owner *po, pmc_id_t pmcid) 2519 { 2520 struct pmc *pm; 2521 2522 KASSERT(PMC_ID_TO_ROWINDEX(pmcid) < md->pmd_npmc, 2523 ("[pmc,%d] Illegal pmc index %d (max %d)", __LINE__, 2524 PMC_ID_TO_ROWINDEX(pmcid), md->pmd_npmc)); 2525 2526 LIST_FOREACH(pm, &po->po_pmcs, pm_next) 2527 if (pm->pm_id == pmcid) 2528 return pm; 2529 2530 return NULL; 2531 } 2532 2533 static int 2534 pmc_find_pmc(pmc_id_t pmcid, struct pmc **pmc) 2535 { 2536 2537 struct pmc *pm; 2538 struct pmc_owner *po; 2539 2540 PMCDBG(PMC,FND,1, "find-pmc id=%d", pmcid); 2541 2542 if ((po = pmc_find_owner_descriptor(curthread->td_proc)) == NULL) 2543 return ESRCH; 2544 2545 if ((pm = pmc_find_pmc_descriptor_in_process(po, pmcid)) == NULL) 2546 return EINVAL; 2547 2548 PMCDBG(PMC,FND,2, "find-pmc id=%d -> pmc=%p", pmcid, pm); 2549 2550 *pmc = pm; 2551 return 0; 2552 } 2553 2554 /* 2555 * Start a PMC. 2556 */ 2557 2558 static int 2559 pmc_start(struct pmc *pm) 2560 { 2561 enum pmc_mode mode; 2562 struct pmc_owner *po; 2563 struct pmc_binding pb; 2564 struct pmc_classdep *pcd; 2565 int adjri, error, cpu, ri; 2566 2567 KASSERT(pm != NULL, 2568 ("[pmc,%d] null pm", __LINE__)); 2569 2570 mode = PMC_TO_MODE(pm); 2571 ri = PMC_TO_ROWINDEX(pm); 2572 pcd = pmc_ri_to_classdep(md, ri, &adjri); 2573 2574 error = 0; 2575 2576 PMCDBG(PMC,OPS,1, "start pmc=%p mode=%d ri=%d", pm, mode, ri); 2577 2578 po = pm->pm_owner; 2579 2580 /* 2581 * Disallow PMCSTART if a logfile is required but has not been 2582 * configured yet. 2583 */ 2584 if ((pm->pm_flags & PMC_F_NEEDS_LOGFILE) && 2585 (po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) 2586 return (EDOOFUS); /* programming error */ 2587 2588 /* 2589 * If this is a sampling mode PMC, log mapping information for 2590 * the kernel modules that are currently loaded. 2591 */ 2592 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) 2593 pmc_log_kernel_mappings(pm); 2594 2595 if (PMC_IS_VIRTUAL_MODE(mode)) { 2596 2597 /* 2598 * If a PMCATTACH has never been done on this PMC, 2599 * attach it to its owner process. 2600 */ 2601 2602 if (LIST_EMPTY(&pm->pm_targets)) 2603 error = (pm->pm_flags & PMC_F_ATTACH_DONE) ? ESRCH : 2604 pmc_attach_process(po->po_owner, pm); 2605 2606 /* 2607 * If the PMC is attached to its owner, then force a context 2608 * switch to ensure that the MD state gets set correctly. 2609 */ 2610 2611 if (error == 0) { 2612 pm->pm_state = PMC_STATE_RUNNING; 2613 if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) 2614 pmc_force_context_switch(); 2615 } 2616 2617 return (error); 2618 } 2619 2620 2621 /* 2622 * A system-wide PMC. 2623 * 2624 * Add the owner to the global list if this is a system-wide 2625 * sampling PMC. 2626 */ 2627 2628 if (mode == PMC_MODE_SS) { 2629 if (po->po_sscount == 0) { 2630 LIST_INSERT_HEAD(&pmc_ss_owners, po, po_ssnext); 2631 atomic_add_rel_int(&pmc_ss_count, 1); 2632 PMCDBG(PMC,OPS,1, "po=%p in global list", po); 2633 } 2634 po->po_sscount++; 2635 2636 /* 2637 * Log mapping information for all existing processes in the 2638 * system. Subsequent mappings are logged as they happen; 2639 * see pmc_process_mmap(). 2640 */ 2641 if (po->po_logprocmaps == 0) { 2642 pmc_log_all_process_mappings(po); 2643 po->po_logprocmaps = 1; 2644 } 2645 } 2646 2647 /* 2648 * Move to the CPU associated with this 2649 * PMC, and start the hardware. 2650 */ 2651 2652 pmc_save_cpu_binding(&pb); 2653 2654 cpu = PMC_TO_CPU(pm); 2655 2656 if (!pmc_cpu_is_active(cpu)) 2657 return (ENXIO); 2658 2659 pmc_select_cpu(cpu); 2660 2661 /* 2662 * global PMCs are configured at allocation time 2663 * so write out the initial value and start the PMC. 2664 */ 2665 2666 pm->pm_state = PMC_STATE_RUNNING; 2667 2668 critical_enter(); 2669 if ((error = pcd->pcd_write_pmc(cpu, adjri, 2670 PMC_IS_SAMPLING_MODE(mode) ? 2671 pm->pm_sc.pm_reloadcount : 2672 pm->pm_sc.pm_initial)) == 0) 2673 error = pcd->pcd_start_pmc(cpu, adjri); 2674 critical_exit(); 2675 2676 pmc_restore_cpu_binding(&pb); 2677 2678 return (error); 2679 } 2680 2681 /* 2682 * Stop a PMC. 2683 */ 2684 2685 static int 2686 pmc_stop(struct pmc *pm) 2687 { 2688 struct pmc_owner *po; 2689 struct pmc_binding pb; 2690 struct pmc_classdep *pcd; 2691 int adjri, cpu, error, ri; 2692 2693 KASSERT(pm != NULL, ("[pmc,%d] null pmc", __LINE__)); 2694 2695 PMCDBG(PMC,OPS,1, "stop pmc=%p mode=%d ri=%d", pm, 2696 PMC_TO_MODE(pm), PMC_TO_ROWINDEX(pm)); 2697 2698 pm->pm_state = PMC_STATE_STOPPED; 2699 2700 /* 2701 * If the PMC is a virtual mode one, changing the state to 2702 * non-RUNNING is enough to ensure that the PMC never gets 2703 * scheduled. 2704 * 2705 * If this PMC is current running on a CPU, then it will 2706 * handled correctly at the time its target process is context 2707 * switched out. 2708 */ 2709 2710 if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) 2711 return 0; 2712 2713 /* 2714 * A system-mode PMC. Move to the CPU associated with 2715 * this PMC, and stop the hardware. We update the 2716 * 'initial count' so that a subsequent PMCSTART will 2717 * resume counting from the current hardware count. 2718 */ 2719 2720 pmc_save_cpu_binding(&pb); 2721 2722 cpu = PMC_TO_CPU(pm); 2723 2724 KASSERT(cpu >= 0 && cpu < pmc_cpu_max(), 2725 ("[pmc,%d] illegal cpu=%d", __LINE__, cpu)); 2726 2727 if (!pmc_cpu_is_active(cpu)) 2728 return ENXIO; 2729 2730 pmc_select_cpu(cpu); 2731 2732 ri = PMC_TO_ROWINDEX(pm); 2733 pcd = pmc_ri_to_classdep(md, ri, &adjri); 2734 2735 critical_enter(); 2736 if ((error = pcd->pcd_stop_pmc(cpu, adjri)) == 0) 2737 error = pcd->pcd_read_pmc(cpu, adjri, &pm->pm_sc.pm_initial); 2738 critical_exit(); 2739 2740 pmc_restore_cpu_binding(&pb); 2741 2742 po = pm->pm_owner; 2743 2744 /* remove this owner from the global list of SS PMC owners */ 2745 if (PMC_TO_MODE(pm) == PMC_MODE_SS) { 2746 po->po_sscount--; 2747 if (po->po_sscount == 0) { 2748 atomic_subtract_rel_int(&pmc_ss_count, 1); 2749 LIST_REMOVE(po, po_ssnext); 2750 PMCDBG(PMC,OPS,2,"po=%p removed from global list", po); 2751 } 2752 } 2753 2754 return (error); 2755 } 2756 2757 2758 #ifdef DEBUG 2759 static const char *pmc_op_to_name[] = { 2760 #undef __PMC_OP 2761 #define __PMC_OP(N, D) #N , 2762 __PMC_OPS() 2763 NULL 2764 }; 2765 #endif 2766 2767 /* 2768 * The syscall interface 2769 */ 2770 2771 #define PMC_GET_SX_XLOCK(...) do { \ 2772 sx_xlock(&pmc_sx); \ 2773 if (pmc_hook == NULL) { \ 2774 sx_xunlock(&pmc_sx); \ 2775 return __VA_ARGS__; \ 2776 } \ 2777 } while (0) 2778 2779 #define PMC_DOWNGRADE_SX() do { \ 2780 sx_downgrade(&pmc_sx); \ 2781 is_sx_downgraded = 1; \ 2782 } while (0) 2783 2784 static int 2785 pmc_syscall_handler(struct thread *td, void *syscall_args) 2786 { 2787 int error, is_sx_downgraded, is_sx_locked, op; 2788 struct pmc_syscall_args *c; 2789 void *arg; 2790 2791 PMC_GET_SX_XLOCK(ENOSYS); 2792 2793 DROP_GIANT(); 2794 2795 is_sx_downgraded = 0; 2796 is_sx_locked = 1; 2797 2798 c = (struct pmc_syscall_args *) syscall_args; 2799 2800 op = c->pmop_code; 2801 arg = c->pmop_data; 2802 2803 PMCDBG(MOD,PMS,1, "syscall op=%d \"%s\" arg=%p", op, 2804 pmc_op_to_name[op], arg); 2805 2806 error = 0; 2807 atomic_add_int(&pmc_stats.pm_syscalls, 1); 2808 2809 switch(op) 2810 { 2811 2812 2813 /* 2814 * Configure a log file. 2815 * 2816 * XXX This OP will be reworked. 2817 */ 2818 2819 case PMC_OP_CONFIGURELOG: 2820 { 2821 struct proc *p; 2822 struct pmc *pm; 2823 struct pmc_owner *po; 2824 struct pmc_op_configurelog cl; 2825 2826 sx_assert(&pmc_sx, SX_XLOCKED); 2827 2828 if ((error = copyin(arg, &cl, sizeof(cl))) != 0) 2829 break; 2830 2831 /* mark this process as owning a log file */ 2832 p = td->td_proc; 2833 if ((po = pmc_find_owner_descriptor(p)) == NULL) 2834 if ((po = pmc_allocate_owner_descriptor(p)) == NULL) { 2835 error = ENOMEM; 2836 break; 2837 } 2838 2839 /* 2840 * If a valid fd was passed in, try to configure that, 2841 * otherwise if 'fd' was less than zero and there was 2842 * a log file configured, flush its buffers and 2843 * de-configure it. 2844 */ 2845 if (cl.pm_logfd >= 0) { 2846 sx_xunlock(&pmc_sx); 2847 is_sx_locked = 0; 2848 error = pmclog_configure_log(md, po, cl.pm_logfd); 2849 } else if (po->po_flags & PMC_PO_OWNS_LOGFILE) { 2850 pmclog_process_closelog(po); 2851 error = pmclog_close(po); 2852 if (error == 0) { 2853 LIST_FOREACH(pm, &po->po_pmcs, pm_next) 2854 if (pm->pm_flags & PMC_F_NEEDS_LOGFILE && 2855 pm->pm_state == PMC_STATE_RUNNING) 2856 pmc_stop(pm); 2857 error = pmclog_deconfigure_log(po); 2858 } 2859 } else 2860 error = EINVAL; 2861 2862 if (error) 2863 break; 2864 } 2865 break; 2866 2867 /* 2868 * Flush a log file. 2869 */ 2870 2871 case PMC_OP_FLUSHLOG: 2872 { 2873 struct pmc_owner *po; 2874 2875 sx_assert(&pmc_sx, SX_XLOCKED); 2876 2877 if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) { 2878 error = EINVAL; 2879 break; 2880 } 2881 2882 error = pmclog_flush(po); 2883 } 2884 break; 2885 2886 /* 2887 * Close a log file. 2888 */ 2889 2890 case PMC_OP_CLOSELOG: 2891 { 2892 struct pmc_owner *po; 2893 2894 sx_assert(&pmc_sx, SX_XLOCKED); 2895 2896 if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) { 2897 error = EINVAL; 2898 break; 2899 } 2900 2901 error = pmclog_close(po); 2902 } 2903 break; 2904 2905 /* 2906 * Retrieve hardware configuration. 2907 */ 2908 2909 case PMC_OP_GETCPUINFO: /* CPU information */ 2910 { 2911 struct pmc_op_getcpuinfo gci; 2912 struct pmc_classinfo *pci; 2913 struct pmc_classdep *pcd; 2914 int cl; 2915 2916 gci.pm_cputype = md->pmd_cputype; 2917 gci.pm_ncpu = pmc_cpu_max(); 2918 gci.pm_npmc = md->pmd_npmc; 2919 gci.pm_nclass = md->pmd_nclass; 2920 pci = gci.pm_classes; 2921 pcd = md->pmd_classdep; 2922 for (cl = 0; cl < md->pmd_nclass; cl++, pci++, pcd++) { 2923 pci->pm_caps = pcd->pcd_caps; 2924 pci->pm_class = pcd->pcd_class; 2925 pci->pm_width = pcd->pcd_width; 2926 pci->pm_num = pcd->pcd_num; 2927 } 2928 error = copyout(&gci, arg, sizeof(gci)); 2929 } 2930 break; 2931 2932 /* 2933 * Retrieve soft events list. 2934 */ 2935 case PMC_OP_GETDYNEVENTINFO: 2936 { 2937 enum pmc_class cl; 2938 enum pmc_event ev; 2939 struct pmc_op_getdyneventinfo *gei; 2940 struct pmc_dyn_event_descr dev; 2941 struct pmc_soft *ps; 2942 uint32_t nevent; 2943 2944 sx_assert(&pmc_sx, SX_LOCKED); 2945 2946 gei = (struct pmc_op_getdyneventinfo *) arg; 2947 2948 if ((error = copyin(&gei->pm_class, &cl, sizeof(cl))) != 0) 2949 break; 2950 2951 /* Only SOFT class is dynamic. */ 2952 if (cl != PMC_CLASS_SOFT) { 2953 error = EINVAL; 2954 break; 2955 } 2956 2957 nevent = 0; 2958 for (ev = PMC_EV_SOFT_FIRST; (int)ev <= PMC_EV_SOFT_LAST; ev++) { 2959 ps = pmc_soft_ev_acquire(ev); 2960 if (ps == NULL) 2961 continue; 2962 bcopy(&ps->ps_ev, &dev, sizeof(dev)); 2963 pmc_soft_ev_release(ps); 2964 2965 error = copyout(&dev, 2966 &gei->pm_events[nevent], 2967 sizeof(struct pmc_dyn_event_descr)); 2968 if (error != 0) 2969 break; 2970 nevent++; 2971 } 2972 if (error != 0) 2973 break; 2974 2975 error = copyout(&nevent, &gei->pm_nevent, 2976 sizeof(nevent)); 2977 } 2978 break; 2979 2980 /* 2981 * Get module statistics 2982 */ 2983 2984 case PMC_OP_GETDRIVERSTATS: 2985 { 2986 struct pmc_op_getdriverstats gms; 2987 2988 bcopy(&pmc_stats, &gms, sizeof(gms)); 2989 error = copyout(&gms, arg, sizeof(gms)); 2990 } 2991 break; 2992 2993 2994 /* 2995 * Retrieve module version number 2996 */ 2997 2998 case PMC_OP_GETMODULEVERSION: 2999 { 3000 uint32_t cv, modv; 3001 3002 /* retrieve the client's idea of the ABI version */ 3003 if ((error = copyin(arg, &cv, sizeof(uint32_t))) != 0) 3004 break; 3005 /* don't service clients newer than our driver */ 3006 modv = PMC_VERSION; 3007 if ((cv & 0xFFFF0000) > (modv & 0xFFFF0000)) { 3008 error = EPROGMISMATCH; 3009 break; 3010 } 3011 error = copyout(&modv, arg, sizeof(int)); 3012 } 3013 break; 3014 3015 3016 /* 3017 * Retrieve the state of all the PMCs on a given 3018 * CPU. 3019 */ 3020 3021 case PMC_OP_GETPMCINFO: 3022 { 3023 int ari; 3024 struct pmc *pm; 3025 size_t pmcinfo_size; 3026 uint32_t cpu, n, npmc; 3027 struct pmc_owner *po; 3028 struct pmc_binding pb; 3029 struct pmc_classdep *pcd; 3030 struct pmc_info *p, *pmcinfo; 3031 struct pmc_op_getpmcinfo *gpi; 3032 3033 PMC_DOWNGRADE_SX(); 3034 3035 gpi = (struct pmc_op_getpmcinfo *) arg; 3036 3037 if ((error = copyin(&gpi->pm_cpu, &cpu, sizeof(cpu))) != 0) 3038 break; 3039 3040 if (cpu >= pmc_cpu_max()) { 3041 error = EINVAL; 3042 break; 3043 } 3044 3045 if (!pmc_cpu_is_active(cpu)) { 3046 error = ENXIO; 3047 break; 3048 } 3049 3050 /* switch to CPU 'cpu' */ 3051 pmc_save_cpu_binding(&pb); 3052 pmc_select_cpu(cpu); 3053 3054 npmc = md->pmd_npmc; 3055 3056 pmcinfo_size = npmc * sizeof(struct pmc_info); 3057 pmcinfo = malloc(pmcinfo_size, M_PMC, M_WAITOK); 3058 3059 p = pmcinfo; 3060 3061 for (n = 0; n < md->pmd_npmc; n++, p++) { 3062 3063 pcd = pmc_ri_to_classdep(md, n, &ari); 3064 3065 KASSERT(pcd != NULL, 3066 ("[pmc,%d] null pcd ri=%d", __LINE__, n)); 3067 3068 if ((error = pcd->pcd_describe(cpu, ari, p, &pm)) != 0) 3069 break; 3070 3071 if (PMC_ROW_DISP_IS_STANDALONE(n)) 3072 p->pm_rowdisp = PMC_DISP_STANDALONE; 3073 else if (PMC_ROW_DISP_IS_THREAD(n)) 3074 p->pm_rowdisp = PMC_DISP_THREAD; 3075 else 3076 p->pm_rowdisp = PMC_DISP_FREE; 3077 3078 p->pm_ownerpid = -1; 3079 3080 if (pm == NULL) /* no PMC associated */ 3081 continue; 3082 3083 po = pm->pm_owner; 3084 3085 KASSERT(po->po_owner != NULL, 3086 ("[pmc,%d] pmc_owner had a null proc pointer", 3087 __LINE__)); 3088 3089 p->pm_ownerpid = po->po_owner->p_pid; 3090 p->pm_mode = PMC_TO_MODE(pm); 3091 p->pm_event = pm->pm_event; 3092 p->pm_flags = pm->pm_flags; 3093 3094 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) 3095 p->pm_reloadcount = 3096 pm->pm_sc.pm_reloadcount; 3097 } 3098 3099 pmc_restore_cpu_binding(&pb); 3100 3101 /* now copy out the PMC info collected */ 3102 if (error == 0) 3103 error = copyout(pmcinfo, &gpi->pm_pmcs, pmcinfo_size); 3104 3105 free(pmcinfo, M_PMC); 3106 } 3107 break; 3108 3109 3110 /* 3111 * Set the administrative state of a PMC. I.e. whether 3112 * the PMC is to be used or not. 3113 */ 3114 3115 case PMC_OP_PMCADMIN: 3116 { 3117 int cpu, ri; 3118 enum pmc_state request; 3119 struct pmc_cpu *pc; 3120 struct pmc_hw *phw; 3121 struct pmc_op_pmcadmin pma; 3122 struct pmc_binding pb; 3123 3124 sx_assert(&pmc_sx, SX_XLOCKED); 3125 3126 KASSERT(td == curthread, 3127 ("[pmc,%d] td != curthread", __LINE__)); 3128 3129 error = priv_check(td, PRIV_PMC_MANAGE); 3130 if (error) 3131 break; 3132 3133 if ((error = copyin(arg, &pma, sizeof(pma))) != 0) 3134 break; 3135 3136 cpu = pma.pm_cpu; 3137 3138 if (cpu < 0 || cpu >= (int) pmc_cpu_max()) { 3139 error = EINVAL; 3140 break; 3141 } 3142 3143 if (!pmc_cpu_is_active(cpu)) { 3144 error = ENXIO; 3145 break; 3146 } 3147 3148 request = pma.pm_state; 3149 3150 if (request != PMC_STATE_DISABLED && 3151 request != PMC_STATE_FREE) { 3152 error = EINVAL; 3153 break; 3154 } 3155 3156 ri = pma.pm_pmc; /* pmc id == row index */ 3157 if (ri < 0 || ri >= (int) md->pmd_npmc) { 3158 error = EINVAL; 3159 break; 3160 } 3161 3162 /* 3163 * We can't disable a PMC with a row-index allocated 3164 * for process virtual PMCs. 3165 */ 3166 3167 if (PMC_ROW_DISP_IS_THREAD(ri) && 3168 request == PMC_STATE_DISABLED) { 3169 error = EBUSY; 3170 break; 3171 } 3172 3173 /* 3174 * otherwise, this PMC on this CPU is either free or 3175 * in system-wide mode. 3176 */ 3177 3178 pmc_save_cpu_binding(&pb); 3179 pmc_select_cpu(cpu); 3180 3181 pc = pmc_pcpu[cpu]; 3182 phw = pc->pc_hwpmcs[ri]; 3183 3184 /* 3185 * XXX do we need some kind of 'forced' disable? 3186 */ 3187 3188 if (phw->phw_pmc == NULL) { 3189 if (request == PMC_STATE_DISABLED && 3190 (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED)) { 3191 phw->phw_state &= ~PMC_PHW_FLAG_IS_ENABLED; 3192 PMC_MARK_ROW_STANDALONE(ri); 3193 } else if (request == PMC_STATE_FREE && 3194 (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0) { 3195 phw->phw_state |= PMC_PHW_FLAG_IS_ENABLED; 3196 PMC_UNMARK_ROW_STANDALONE(ri); 3197 } 3198 /* other cases are a no-op */ 3199 } else 3200 error = EBUSY; 3201 3202 pmc_restore_cpu_binding(&pb); 3203 } 3204 break; 3205 3206 3207 /* 3208 * Allocate a PMC. 3209 */ 3210 3211 case PMC_OP_PMCALLOCATE: 3212 { 3213 int adjri, n; 3214 u_int cpu; 3215 uint32_t caps; 3216 struct pmc *pmc; 3217 enum pmc_mode mode; 3218 struct pmc_hw *phw; 3219 struct pmc_binding pb; 3220 struct pmc_classdep *pcd; 3221 struct pmc_op_pmcallocate pa; 3222 3223 if ((error = copyin(arg, &pa, sizeof(pa))) != 0) 3224 break; 3225 3226 caps = pa.pm_caps; 3227 mode = pa.pm_mode; 3228 cpu = pa.pm_cpu; 3229 3230 if ((mode != PMC_MODE_SS && mode != PMC_MODE_SC && 3231 mode != PMC_MODE_TS && mode != PMC_MODE_TC) || 3232 (cpu != (u_int) PMC_CPU_ANY && cpu >= pmc_cpu_max())) { 3233 error = EINVAL; 3234 break; 3235 } 3236 3237 /* 3238 * Virtual PMCs should only ask for a default CPU. 3239 * System mode PMCs need to specify a non-default CPU. 3240 */ 3241 3242 if ((PMC_IS_VIRTUAL_MODE(mode) && cpu != (u_int) PMC_CPU_ANY) || 3243 (PMC_IS_SYSTEM_MODE(mode) && cpu == (u_int) PMC_CPU_ANY)) { 3244 error = EINVAL; 3245 break; 3246 } 3247 3248 /* 3249 * Check that an inactive CPU is not being asked for. 3250 */ 3251 3252 if (PMC_IS_SYSTEM_MODE(mode) && !pmc_cpu_is_active(cpu)) { 3253 error = ENXIO; 3254 break; 3255 } 3256 3257 /* 3258 * Refuse an allocation for a system-wide PMC if this 3259 * process has been jailed, or if this process lacks 3260 * super-user credentials and the sysctl tunable 3261 * 'security.bsd.unprivileged_syspmcs' is zero. 3262 */ 3263 3264 if (PMC_IS_SYSTEM_MODE(mode)) { 3265 if (jailed(curthread->td_ucred)) { 3266 error = EPERM; 3267 break; 3268 } 3269 if (!pmc_unprivileged_syspmcs) { 3270 error = priv_check(curthread, 3271 PRIV_PMC_SYSTEM); 3272 if (error) 3273 break; 3274 } 3275 } 3276 3277 /* 3278 * Look for valid values for 'pm_flags' 3279 */ 3280 3281 if ((pa.pm_flags & ~(PMC_F_DESCENDANTS | PMC_F_LOG_PROCCSW | 3282 PMC_F_LOG_PROCEXIT | PMC_F_CALLCHAIN)) != 0) { 3283 error = EINVAL; 3284 break; 3285 } 3286 3287 /* process logging options are not allowed for system PMCs */ 3288 if (PMC_IS_SYSTEM_MODE(mode) && (pa.pm_flags & 3289 (PMC_F_LOG_PROCCSW | PMC_F_LOG_PROCEXIT))) { 3290 error = EINVAL; 3291 break; 3292 } 3293 3294 /* 3295 * All sampling mode PMCs need to be able to interrupt the 3296 * CPU. 3297 */ 3298 if (PMC_IS_SAMPLING_MODE(mode)) 3299 caps |= PMC_CAP_INTERRUPT; 3300 3301 /* A valid class specifier should have been passed in. */ 3302 for (n = 0; n < md->pmd_nclass; n++) 3303 if (md->pmd_classdep[n].pcd_class == pa.pm_class) 3304 break; 3305 if (n == md->pmd_nclass) { 3306 error = EINVAL; 3307 break; 3308 } 3309 3310 /* The requested PMC capabilities should be feasible. */ 3311 if ((md->pmd_classdep[n].pcd_caps & caps) != caps) { 3312 error = EOPNOTSUPP; 3313 break; 3314 } 3315 3316 PMCDBG(PMC,ALL,2, "event=%d caps=0x%x mode=%d cpu=%d", 3317 pa.pm_ev, caps, mode, cpu); 3318 3319 pmc = pmc_allocate_pmc_descriptor(); 3320 pmc->pm_id = PMC_ID_MAKE_ID(cpu,pa.pm_mode,pa.pm_class, 3321 PMC_ID_INVALID); 3322 pmc->pm_event = pa.pm_ev; 3323 pmc->pm_state = PMC_STATE_FREE; 3324 pmc->pm_caps = caps; 3325 pmc->pm_flags = pa.pm_flags; 3326 3327 /* switch thread to CPU 'cpu' */ 3328 pmc_save_cpu_binding(&pb); 3329 3330 #define PMC_IS_SHAREABLE_PMC(cpu, n) \ 3331 (pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_state & \ 3332 PMC_PHW_FLAG_IS_SHAREABLE) 3333 #define PMC_IS_UNALLOCATED(cpu, n) \ 3334 (pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_pmc == NULL) 3335 3336 if (PMC_IS_SYSTEM_MODE(mode)) { 3337 pmc_select_cpu(cpu); 3338 for (n = 0; n < (int) md->pmd_npmc; n++) { 3339 pcd = pmc_ri_to_classdep(md, n, &adjri); 3340 if (pmc_can_allocate_row(n, mode) == 0 && 3341 pmc_can_allocate_rowindex( 3342 curthread->td_proc, n, cpu) == 0 && 3343 (PMC_IS_UNALLOCATED(cpu, n) || 3344 PMC_IS_SHAREABLE_PMC(cpu, n)) && 3345 pcd->pcd_allocate_pmc(cpu, adjri, pmc, 3346 &pa) == 0) 3347 break; 3348 } 3349 } else { 3350 /* Process virtual mode */ 3351 for (n = 0; n < (int) md->pmd_npmc; n++) { 3352 pcd = pmc_ri_to_classdep(md, n, &adjri); 3353 if (pmc_can_allocate_row(n, mode) == 0 && 3354 pmc_can_allocate_rowindex( 3355 curthread->td_proc, n, 3356 PMC_CPU_ANY) == 0 && 3357 pcd->pcd_allocate_pmc(curthread->td_oncpu, 3358 adjri, pmc, &pa) == 0) 3359 break; 3360 } 3361 } 3362 3363 #undef PMC_IS_UNALLOCATED 3364 #undef PMC_IS_SHAREABLE_PMC 3365 3366 pmc_restore_cpu_binding(&pb); 3367 3368 if (n == (int) md->pmd_npmc) { 3369 pmc_destroy_pmc_descriptor(pmc); 3370 free(pmc, M_PMC); 3371 pmc = NULL; 3372 error = EINVAL; 3373 break; 3374 } 3375 3376 /* Fill in the correct value in the ID field */ 3377 pmc->pm_id = PMC_ID_MAKE_ID(cpu,mode,pa.pm_class,n); 3378 3379 PMCDBG(PMC,ALL,2, "ev=%d class=%d mode=%d n=%d -> pmcid=%x", 3380 pmc->pm_event, pa.pm_class, mode, n, pmc->pm_id); 3381 3382 /* Process mode PMCs with logging enabled need log files */ 3383 if (pmc->pm_flags & (PMC_F_LOG_PROCEXIT | PMC_F_LOG_PROCCSW)) 3384 pmc->pm_flags |= PMC_F_NEEDS_LOGFILE; 3385 3386 /* All system mode sampling PMCs require a log file */ 3387 if (PMC_IS_SAMPLING_MODE(mode) && PMC_IS_SYSTEM_MODE(mode)) 3388 pmc->pm_flags |= PMC_F_NEEDS_LOGFILE; 3389 3390 /* 3391 * Configure global pmc's immediately 3392 */ 3393 3394 if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pmc))) { 3395 3396 pmc_save_cpu_binding(&pb); 3397 pmc_select_cpu(cpu); 3398 3399 phw = pmc_pcpu[cpu]->pc_hwpmcs[n]; 3400 pcd = pmc_ri_to_classdep(md, n, &adjri); 3401 3402 if ((phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0 || 3403 (error = pcd->pcd_config_pmc(cpu, adjri, pmc)) != 0) { 3404 (void) pcd->pcd_release_pmc(cpu, adjri, pmc); 3405 pmc_destroy_pmc_descriptor(pmc); 3406 free(pmc, M_PMC); 3407 pmc = NULL; 3408 pmc_restore_cpu_binding(&pb); 3409 error = EPERM; 3410 break; 3411 } 3412 3413 pmc_restore_cpu_binding(&pb); 3414 } 3415 3416 pmc->pm_state = PMC_STATE_ALLOCATED; 3417 3418 /* 3419 * mark row disposition 3420 */ 3421 3422 if (PMC_IS_SYSTEM_MODE(mode)) 3423 PMC_MARK_ROW_STANDALONE(n); 3424 else 3425 PMC_MARK_ROW_THREAD(n); 3426 3427 /* 3428 * Register this PMC with the current thread as its owner. 3429 */ 3430 3431 if ((error = 3432 pmc_register_owner(curthread->td_proc, pmc)) != 0) { 3433 pmc_release_pmc_descriptor(pmc); 3434 free(pmc, M_PMC); 3435 pmc = NULL; 3436 break; 3437 } 3438 3439 /* 3440 * Return the allocated index. 3441 */ 3442 3443 pa.pm_pmcid = pmc->pm_id; 3444 3445 error = copyout(&pa, arg, sizeof(pa)); 3446 } 3447 break; 3448 3449 3450 /* 3451 * Attach a PMC to a process. 3452 */ 3453 3454 case PMC_OP_PMCATTACH: 3455 { 3456 struct pmc *pm; 3457 struct proc *p; 3458 struct pmc_op_pmcattach a; 3459 3460 sx_assert(&pmc_sx, SX_XLOCKED); 3461 3462 if ((error = copyin(arg, &a, sizeof(a))) != 0) 3463 break; 3464 3465 if (a.pm_pid < 0) { 3466 error = EINVAL; 3467 break; 3468 } else if (a.pm_pid == 0) 3469 a.pm_pid = td->td_proc->p_pid; 3470 3471 if ((error = pmc_find_pmc(a.pm_pmc, &pm)) != 0) 3472 break; 3473 3474 if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm))) { 3475 error = EINVAL; 3476 break; 3477 } 3478 3479 /* PMCs may be (re)attached only when allocated or stopped */ 3480 if (pm->pm_state == PMC_STATE_RUNNING) { 3481 error = EBUSY; 3482 break; 3483 } else if (pm->pm_state != PMC_STATE_ALLOCATED && 3484 pm->pm_state != PMC_STATE_STOPPED) { 3485 error = EINVAL; 3486 break; 3487 } 3488 3489 /* lookup pid */ 3490 if ((p = pfind(a.pm_pid)) == NULL) { 3491 error = ESRCH; 3492 break; 3493 } 3494 3495 /* 3496 * Ignore processes that are working on exiting. 3497 */ 3498 if (p->p_flag & P_WEXIT) { 3499 error = ESRCH; 3500 PROC_UNLOCK(p); /* pfind() returns a locked process */ 3501 break; 3502 } 3503 3504 /* 3505 * we are allowed to attach a PMC to a process if 3506 * we can debug it. 3507 */ 3508 error = p_candebug(curthread, p); 3509 3510 PROC_UNLOCK(p); 3511 3512 if (error == 0) 3513 error = pmc_attach_process(p, pm); 3514 } 3515 break; 3516 3517 3518 /* 3519 * Detach an attached PMC from a process. 3520 */ 3521 3522 case PMC_OP_PMCDETACH: 3523 { 3524 struct pmc *pm; 3525 struct proc *p; 3526 struct pmc_op_pmcattach a; 3527 3528 if ((error = copyin(arg, &a, sizeof(a))) != 0) 3529 break; 3530 3531 if (a.pm_pid < 0) { 3532 error = EINVAL; 3533 break; 3534 } else if (a.pm_pid == 0) 3535 a.pm_pid = td->td_proc->p_pid; 3536 3537 if ((error = pmc_find_pmc(a.pm_pmc, &pm)) != 0) 3538 break; 3539 3540 if ((p = pfind(a.pm_pid)) == NULL) { 3541 error = ESRCH; 3542 break; 3543 } 3544 3545 /* 3546 * Treat processes that are in the process of exiting 3547 * as if they were not present. 3548 */ 3549 3550 if (p->p_flag & P_WEXIT) 3551 error = ESRCH; 3552 3553 PROC_UNLOCK(p); /* pfind() returns a locked process */ 3554 3555 if (error == 0) 3556 error = pmc_detach_process(p, pm); 3557 } 3558 break; 3559 3560 3561 /* 3562 * Retrieve the MSR number associated with the counter 3563 * 'pmc_id'. This allows processes to directly use RDPMC 3564 * instructions to read their PMCs, without the overhead of a 3565 * system call. 3566 */ 3567 3568 case PMC_OP_PMCGETMSR: 3569 { 3570 int adjri, ri; 3571 struct pmc *pm; 3572 struct pmc_target *pt; 3573 struct pmc_op_getmsr gm; 3574 struct pmc_classdep *pcd; 3575 3576 PMC_DOWNGRADE_SX(); 3577 3578 if ((error = copyin(arg, &gm, sizeof(gm))) != 0) 3579 break; 3580 3581 if ((error = pmc_find_pmc(gm.pm_pmcid, &pm)) != 0) 3582 break; 3583 3584 /* 3585 * The allocated PMC has to be a process virtual PMC, 3586 * i.e., of type MODE_T[CS]. Global PMCs can only be 3587 * read using the PMCREAD operation since they may be 3588 * allocated on a different CPU than the one we could 3589 * be running on at the time of the RDPMC instruction. 3590 * 3591 * The GETMSR operation is not allowed for PMCs that 3592 * are inherited across processes. 3593 */ 3594 3595 if (!PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)) || 3596 (pm->pm_flags & PMC_F_DESCENDANTS)) { 3597 error = EINVAL; 3598 break; 3599 } 3600 3601 /* 3602 * It only makes sense to use a RDPMC (or its 3603 * equivalent instruction on non-x86 architectures) on 3604 * a process that has allocated and attached a PMC to 3605 * itself. Conversely the PMC is only allowed to have 3606 * one process attached to it -- its owner. 3607 */ 3608 3609 if ((pt = LIST_FIRST(&pm->pm_targets)) == NULL || 3610 LIST_NEXT(pt, pt_next) != NULL || 3611 pt->pt_process->pp_proc != pm->pm_owner->po_owner) { 3612 error = EINVAL; 3613 break; 3614 } 3615 3616 ri = PMC_TO_ROWINDEX(pm); 3617 pcd = pmc_ri_to_classdep(md, ri, &adjri); 3618 3619 /* PMC class has no 'GETMSR' support */ 3620 if (pcd->pcd_get_msr == NULL) { 3621 error = ENOSYS; 3622 break; 3623 } 3624 3625 if ((error = (*pcd->pcd_get_msr)(adjri, &gm.pm_msr)) < 0) 3626 break; 3627 3628 if ((error = copyout(&gm, arg, sizeof(gm))) < 0) 3629 break; 3630 3631 /* 3632 * Mark our process as using MSRs. Update machine 3633 * state using a forced context switch. 3634 */ 3635 3636 pt->pt_process->pp_flags |= PMC_PP_ENABLE_MSR_ACCESS; 3637 pmc_force_context_switch(); 3638 3639 } 3640 break; 3641 3642 /* 3643 * Release an allocated PMC 3644 */ 3645 3646 case PMC_OP_PMCRELEASE: 3647 { 3648 pmc_id_t pmcid; 3649 struct pmc *pm; 3650 struct pmc_owner *po; 3651 struct pmc_op_simple sp; 3652 3653 /* 3654 * Find PMC pointer for the named PMC. 3655 * 3656 * Use pmc_release_pmc_descriptor() to switch off the 3657 * PMC, remove all its target threads, and remove the 3658 * PMC from its owner's list. 3659 * 3660 * Remove the owner record if this is the last PMC 3661 * owned. 3662 * 3663 * Free up space. 3664 */ 3665 3666 if ((error = copyin(arg, &sp, sizeof(sp))) != 0) 3667 break; 3668 3669 pmcid = sp.pm_pmcid; 3670 3671 if ((error = pmc_find_pmc(pmcid, &pm)) != 0) 3672 break; 3673 3674 po = pm->pm_owner; 3675 pmc_release_pmc_descriptor(pm); 3676 pmc_maybe_remove_owner(po); 3677 3678 free(pm, M_PMC); 3679 } 3680 break; 3681 3682 3683 /* 3684 * Read and/or write a PMC. 3685 */ 3686 3687 case PMC_OP_PMCRW: 3688 { 3689 int adjri; 3690 struct pmc *pm; 3691 uint32_t cpu, ri; 3692 pmc_value_t oldvalue; 3693 struct pmc_binding pb; 3694 struct pmc_op_pmcrw prw; 3695 struct pmc_classdep *pcd; 3696 struct pmc_op_pmcrw *pprw; 3697 3698 PMC_DOWNGRADE_SX(); 3699 3700 if ((error = copyin(arg, &prw, sizeof(prw))) != 0) 3701 break; 3702 3703 ri = 0; 3704 PMCDBG(PMC,OPS,1, "rw id=%d flags=0x%x", prw.pm_pmcid, 3705 prw.pm_flags); 3706 3707 /* must have at least one flag set */ 3708 if ((prw.pm_flags & (PMC_F_OLDVALUE|PMC_F_NEWVALUE)) == 0) { 3709 error = EINVAL; 3710 break; 3711 } 3712 3713 /* locate pmc descriptor */ 3714 if ((error = pmc_find_pmc(prw.pm_pmcid, &pm)) != 0) 3715 break; 3716 3717 /* Can't read a PMC that hasn't been started. */ 3718 if (pm->pm_state != PMC_STATE_ALLOCATED && 3719 pm->pm_state != PMC_STATE_STOPPED && 3720 pm->pm_state != PMC_STATE_RUNNING) { 3721 error = EINVAL; 3722 break; 3723 } 3724 3725 /* writing a new value is allowed only for 'STOPPED' pmcs */ 3726 if (pm->pm_state == PMC_STATE_RUNNING && 3727 (prw.pm_flags & PMC_F_NEWVALUE)) { 3728 error = EBUSY; 3729 break; 3730 } 3731 3732 if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) { 3733 3734 /* 3735 * If this PMC is attached to its owner (i.e., 3736 * the process requesting this operation) and 3737 * is running, then attempt to get an 3738 * upto-date reading from hardware for a READ. 3739 * Writes are only allowed when the PMC is 3740 * stopped, so only update the saved value 3741 * field. 3742 * 3743 * If the PMC is not running, or is not 3744 * attached to its owner, read/write to the 3745 * savedvalue field. 3746 */ 3747 3748 ri = PMC_TO_ROWINDEX(pm); 3749 pcd = pmc_ri_to_classdep(md, ri, &adjri); 3750 3751 mtx_pool_lock_spin(pmc_mtxpool, pm); 3752 cpu = curthread->td_oncpu; 3753 3754 if (prw.pm_flags & PMC_F_OLDVALUE) { 3755 if ((pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) && 3756 (pm->pm_state == PMC_STATE_RUNNING)) 3757 error = (*pcd->pcd_read_pmc)(cpu, adjri, 3758 &oldvalue); 3759 else 3760 oldvalue = pm->pm_gv.pm_savedvalue; 3761 } 3762 if (prw.pm_flags & PMC_F_NEWVALUE) 3763 pm->pm_gv.pm_savedvalue = prw.pm_value; 3764 3765 mtx_pool_unlock_spin(pmc_mtxpool, pm); 3766 3767 } else { /* System mode PMCs */ 3768 cpu = PMC_TO_CPU(pm); 3769 ri = PMC_TO_ROWINDEX(pm); 3770 pcd = pmc_ri_to_classdep(md, ri, &adjri); 3771 3772 if (!pmc_cpu_is_active(cpu)) { 3773 error = ENXIO; 3774 break; 3775 } 3776 3777 /* move this thread to CPU 'cpu' */ 3778 pmc_save_cpu_binding(&pb); 3779 pmc_select_cpu(cpu); 3780 3781 critical_enter(); 3782 /* save old value */ 3783 if (prw.pm_flags & PMC_F_OLDVALUE) 3784 if ((error = (*pcd->pcd_read_pmc)(cpu, adjri, 3785 &oldvalue))) 3786 goto error; 3787 /* write out new value */ 3788 if (prw.pm_flags & PMC_F_NEWVALUE) 3789 error = (*pcd->pcd_write_pmc)(cpu, adjri, 3790 prw.pm_value); 3791 error: 3792 critical_exit(); 3793 pmc_restore_cpu_binding(&pb); 3794 if (error) 3795 break; 3796 } 3797 3798 pprw = (struct pmc_op_pmcrw *) arg; 3799 3800 #ifdef DEBUG 3801 if (prw.pm_flags & PMC_F_NEWVALUE) 3802 PMCDBG(PMC,OPS,2, "rw id=%d new %jx -> old %jx", 3803 ri, prw.pm_value, oldvalue); 3804 else if (prw.pm_flags & PMC_F_OLDVALUE) 3805 PMCDBG(PMC,OPS,2, "rw id=%d -> old %jx", ri, oldvalue); 3806 #endif 3807 3808 /* return old value if requested */ 3809 if (prw.pm_flags & PMC_F_OLDVALUE) 3810 if ((error = copyout(&oldvalue, &pprw->pm_value, 3811 sizeof(prw.pm_value)))) 3812 break; 3813 3814 } 3815 break; 3816 3817 3818 /* 3819 * Set the sampling rate for a sampling mode PMC and the 3820 * initial count for a counting mode PMC. 3821 */ 3822 3823 case PMC_OP_PMCSETCOUNT: 3824 { 3825 struct pmc *pm; 3826 struct pmc_op_pmcsetcount sc; 3827 3828 PMC_DOWNGRADE_SX(); 3829 3830 if ((error = copyin(arg, &sc, sizeof(sc))) != 0) 3831 break; 3832 3833 if ((error = pmc_find_pmc(sc.pm_pmcid, &pm)) != 0) 3834 break; 3835 3836 if (pm->pm_state == PMC_STATE_RUNNING) { 3837 error = EBUSY; 3838 break; 3839 } 3840 3841 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) 3842 pm->pm_sc.pm_reloadcount = sc.pm_count; 3843 else 3844 pm->pm_sc.pm_initial = sc.pm_count; 3845 } 3846 break; 3847 3848 3849 /* 3850 * Start a PMC. 3851 */ 3852 3853 case PMC_OP_PMCSTART: 3854 { 3855 pmc_id_t pmcid; 3856 struct pmc *pm; 3857 struct pmc_op_simple sp; 3858 3859 sx_assert(&pmc_sx, SX_XLOCKED); 3860 3861 if ((error = copyin(arg, &sp, sizeof(sp))) != 0) 3862 break; 3863 3864 pmcid = sp.pm_pmcid; 3865 3866 if ((error = pmc_find_pmc(pmcid, &pm)) != 0) 3867 break; 3868 3869 KASSERT(pmcid == pm->pm_id, 3870 ("[pmc,%d] pmcid %x != id %x", __LINE__, 3871 pm->pm_id, pmcid)); 3872 3873 if (pm->pm_state == PMC_STATE_RUNNING) /* already running */ 3874 break; 3875 else if (pm->pm_state != PMC_STATE_STOPPED && 3876 pm->pm_state != PMC_STATE_ALLOCATED) { 3877 error = EINVAL; 3878 break; 3879 } 3880 3881 error = pmc_start(pm); 3882 } 3883 break; 3884 3885 3886 /* 3887 * Stop a PMC. 3888 */ 3889 3890 case PMC_OP_PMCSTOP: 3891 { 3892 pmc_id_t pmcid; 3893 struct pmc *pm; 3894 struct pmc_op_simple sp; 3895 3896 PMC_DOWNGRADE_SX(); 3897 3898 if ((error = copyin(arg, &sp, sizeof(sp))) != 0) 3899 break; 3900 3901 pmcid = sp.pm_pmcid; 3902 3903 /* 3904 * Mark the PMC as inactive and invoke the MD stop 3905 * routines if needed. 3906 */ 3907 3908 if ((error = pmc_find_pmc(pmcid, &pm)) != 0) 3909 break; 3910 3911 KASSERT(pmcid == pm->pm_id, 3912 ("[pmc,%d] pmc id %x != pmcid %x", __LINE__, 3913 pm->pm_id, pmcid)); 3914 3915 if (pm->pm_state == PMC_STATE_STOPPED) /* already stopped */ 3916 break; 3917 else if (pm->pm_state != PMC_STATE_RUNNING) { 3918 error = EINVAL; 3919 break; 3920 } 3921 3922 error = pmc_stop(pm); 3923 } 3924 break; 3925 3926 3927 /* 3928 * Write a user supplied value to the log file. 3929 */ 3930 3931 case PMC_OP_WRITELOG: 3932 { 3933 struct pmc_op_writelog wl; 3934 struct pmc_owner *po; 3935 3936 PMC_DOWNGRADE_SX(); 3937 3938 if ((error = copyin(arg, &wl, sizeof(wl))) != 0) 3939 break; 3940 3941 if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) { 3942 error = EINVAL; 3943 break; 3944 } 3945 3946 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) { 3947 error = EINVAL; 3948 break; 3949 } 3950 3951 error = pmclog_process_userlog(po, &wl); 3952 } 3953 break; 3954 3955 3956 default: 3957 error = EINVAL; 3958 break; 3959 } 3960 3961 if (is_sx_locked != 0) { 3962 if (is_sx_downgraded) 3963 sx_sunlock(&pmc_sx); 3964 else 3965 sx_xunlock(&pmc_sx); 3966 } 3967 3968 if (error) 3969 atomic_add_int(&pmc_stats.pm_syscall_errors, 1); 3970 3971 PICKUP_GIANT(); 3972 3973 return error; 3974 } 3975 3976 /* 3977 * Helper functions 3978 */ 3979 3980 3981 /* 3982 * Mark the thread as needing callchain capture and post an AST. The 3983 * actual callchain capture will be done in a context where it is safe 3984 * to take page faults. 3985 */ 3986 3987 static void 3988 pmc_post_callchain_callback(void) 3989 { 3990 struct thread *td; 3991 3992 td = curthread; 3993 3994 /* 3995 * If there is multiple PMCs for the same interrupt ignore new post 3996 */ 3997 if (td->td_pflags & TDP_CALLCHAIN) 3998 return; 3999 4000 /* 4001 * Mark this thread as needing callchain capture. 4002 * `td->td_pflags' will be safe to touch because this thread 4003 * was in user space when it was interrupted. 4004 */ 4005 td->td_pflags |= TDP_CALLCHAIN; 4006 4007 /* 4008 * Don't let this thread migrate between CPUs until callchain 4009 * capture completes. 4010 */ 4011 sched_pin(); 4012 4013 return; 4014 } 4015 4016 /* 4017 * Interrupt processing. 4018 * 4019 * Find a free slot in the per-cpu array of samples and capture the 4020 * current callchain there. If a sample was successfully added, a bit 4021 * is set in mask 'pmc_cpumask' denoting that the DO_SAMPLES hook 4022 * needs to be invoked from the clock handler. 4023 * 4024 * This function is meant to be called from an NMI handler. It cannot 4025 * use any of the locking primitives supplied by the OS. 4026 */ 4027 4028 int 4029 pmc_process_interrupt(int cpu, int ring, struct pmc *pm, struct trapframe *tf, 4030 int inuserspace) 4031 { 4032 int error, callchaindepth; 4033 struct thread *td; 4034 struct pmc_sample *ps; 4035 struct pmc_samplebuffer *psb; 4036 4037 error = 0; 4038 4039 /* 4040 * Allocate space for a sample buffer. 4041 */ 4042 psb = pmc_pcpu[cpu]->pc_sb[ring]; 4043 4044 ps = psb->ps_write; 4045 if (ps->ps_nsamples) { /* in use, reader hasn't caught up */ 4046 pm->pm_stalled = 1; 4047 atomic_add_int(&pmc_stats.pm_intr_bufferfull, 1); 4048 PMCDBG(SAM,INT,1,"(spc) cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d", 4049 cpu, pm, (void *) tf, inuserspace, 4050 (int) (psb->ps_write - psb->ps_samples), 4051 (int) (psb->ps_read - psb->ps_samples)); 4052 error = ENOMEM; 4053 goto done; 4054 } 4055 4056 4057 /* Fill in entry. */ 4058 PMCDBG(SAM,INT,1,"cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d", cpu, pm, 4059 (void *) tf, inuserspace, 4060 (int) (psb->ps_write - psb->ps_samples), 4061 (int) (psb->ps_read - psb->ps_samples)); 4062 4063 KASSERT(pm->pm_runcount >= 0, 4064 ("[pmc,%d] pm=%p runcount %d", __LINE__, (void *) pm, 4065 pm->pm_runcount)); 4066 4067 atomic_add_rel_int(&pm->pm_runcount, 1); /* hold onto PMC */ 4068 4069 ps->ps_pmc = pm; 4070 if ((td = curthread) && td->td_proc) 4071 ps->ps_pid = td->td_proc->p_pid; 4072 else 4073 ps->ps_pid = -1; 4074 ps->ps_cpu = cpu; 4075 ps->ps_td = td; 4076 ps->ps_flags = inuserspace ? PMC_CC_F_USERSPACE : 0; 4077 4078 callchaindepth = (pm->pm_flags & PMC_F_CALLCHAIN) ? 4079 pmc_callchaindepth : 1; 4080 4081 if (callchaindepth == 1) 4082 ps->ps_pc[0] = PMC_TRAPFRAME_TO_PC(tf); 4083 else { 4084 /* 4085 * Kernel stack traversals can be done immediately, 4086 * while we defer to an AST for user space traversals. 4087 */ 4088 if (!inuserspace) { 4089 callchaindepth = 4090 pmc_save_kernel_callchain(ps->ps_pc, 4091 callchaindepth, tf); 4092 } else { 4093 pmc_post_callchain_callback(); 4094 callchaindepth = PMC_SAMPLE_INUSE; 4095 } 4096 } 4097 4098 ps->ps_nsamples = callchaindepth; /* mark entry as in use */ 4099 4100 /* increment write pointer, modulo ring buffer size */ 4101 ps++; 4102 if (ps == psb->ps_fence) 4103 psb->ps_write = psb->ps_samples; 4104 else 4105 psb->ps_write = ps; 4106 4107 done: 4108 /* mark CPU as needing processing */ 4109 CPU_SET_ATOMIC(cpu, &pmc_cpumask); 4110 4111 return (error); 4112 } 4113 4114 /* 4115 * Capture a user call chain. This function will be called from ast() 4116 * before control returns to userland and before the process gets 4117 * rescheduled. 4118 */ 4119 4120 static void 4121 pmc_capture_user_callchain(int cpu, int ring, struct trapframe *tf) 4122 { 4123 int i; 4124 struct pmc *pm; 4125 struct thread *td; 4126 struct pmc_sample *ps; 4127 struct pmc_samplebuffer *psb; 4128 #ifdef INVARIANTS 4129 int ncallchains; 4130 #endif 4131 4132 psb = pmc_pcpu[cpu]->pc_sb[ring]; 4133 td = curthread; 4134 4135 KASSERT(td->td_pflags & TDP_CALLCHAIN, 4136 ("[pmc,%d] Retrieving callchain for thread that doesn't want it", 4137 __LINE__)); 4138 4139 #ifdef INVARIANTS 4140 ncallchains = 0; 4141 #endif 4142 4143 /* 4144 * Iterate through all deferred callchain requests. 4145 */ 4146 4147 ps = psb->ps_samples; 4148 for (i = 0; i < pmc_nsamples; i++, ps++) { 4149 4150 if (ps->ps_nsamples != PMC_SAMPLE_INUSE) 4151 continue; 4152 if (ps->ps_td != td) 4153 continue; 4154 4155 KASSERT(ps->ps_cpu == cpu, 4156 ("[pmc,%d] cpu mismatch ps_cpu=%d pcpu=%d", __LINE__, 4157 ps->ps_cpu, PCPU_GET(cpuid))); 4158 4159 pm = ps->ps_pmc; 4160 4161 KASSERT(pm->pm_flags & PMC_F_CALLCHAIN, 4162 ("[pmc,%d] Retrieving callchain for PMC that doesn't " 4163 "want it", __LINE__)); 4164 4165 KASSERT(pm->pm_runcount > 0, 4166 ("[pmc,%d] runcount %d", __LINE__, pm->pm_runcount)); 4167 4168 /* 4169 * Retrieve the callchain and mark the sample buffer 4170 * as 'processable' by the timer tick sweep code. 4171 */ 4172 ps->ps_nsamples = pmc_save_user_callchain(ps->ps_pc, 4173 pmc_callchaindepth, tf); 4174 4175 #ifdef INVARIANTS 4176 ncallchains++; 4177 #endif 4178 } 4179 4180 KASSERT(ncallchains > 0, 4181 ("[pmc,%d] cpu %d didn't find a sample to collect", __LINE__, 4182 cpu)); 4183 4184 KASSERT(td->td_pinned == 1, 4185 ("[pmc,%d] invalid td_pinned value", __LINE__)); 4186 sched_unpin(); /* Can migrate safely now. */ 4187 4188 return; 4189 } 4190 4191 /* 4192 * Process saved PC samples. 4193 */ 4194 4195 static void 4196 pmc_process_samples(int cpu, int ring) 4197 { 4198 struct pmc *pm; 4199 int adjri, n; 4200 struct thread *td; 4201 struct pmc_owner *po; 4202 struct pmc_sample *ps; 4203 struct pmc_classdep *pcd; 4204 struct pmc_samplebuffer *psb; 4205 4206 KASSERT(PCPU_GET(cpuid) == cpu, 4207 ("[pmc,%d] not on the correct CPU pcpu=%d cpu=%d", __LINE__, 4208 PCPU_GET(cpuid), cpu)); 4209 4210 psb = pmc_pcpu[cpu]->pc_sb[ring]; 4211 4212 for (n = 0; n < pmc_nsamples; n++) { /* bound on #iterations */ 4213 4214 ps = psb->ps_read; 4215 if (ps->ps_nsamples == PMC_SAMPLE_FREE) 4216 break; 4217 4218 pm = ps->ps_pmc; 4219 4220 KASSERT(pm->pm_runcount > 0, 4221 ("[pmc,%d] pm=%p runcount %d", __LINE__, (void *) pm, 4222 pm->pm_runcount)); 4223 4224 po = pm->pm_owner; 4225 4226 KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)), 4227 ("[pmc,%d] pmc=%p non-sampling mode=%d", __LINE__, 4228 pm, PMC_TO_MODE(pm))); 4229 4230 /* Ignore PMCs that have been switched off */ 4231 if (pm->pm_state != PMC_STATE_RUNNING) 4232 goto entrydone; 4233 4234 /* If there is a pending AST wait for completion */ 4235 if (ps->ps_nsamples == PMC_SAMPLE_INUSE) { 4236 /* Need a rescan at a later time. */ 4237 CPU_SET_ATOMIC(cpu, &pmc_cpumask); 4238 break; 4239 } 4240 4241 PMCDBG(SAM,OPS,1,"cpu=%d pm=%p n=%d fl=%x wr=%d rd=%d", cpu, 4242 pm, ps->ps_nsamples, ps->ps_flags, 4243 (int) (psb->ps_write - psb->ps_samples), 4244 (int) (psb->ps_read - psb->ps_samples)); 4245 4246 /* 4247 * If this is a process-mode PMC that is attached to 4248 * its owner, and if the PC is in user mode, update 4249 * profiling statistics like timer-based profiling 4250 * would have done. 4251 */ 4252 if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) { 4253 if (ps->ps_flags & PMC_CC_F_USERSPACE) { 4254 td = FIRST_THREAD_IN_PROC(po->po_owner); 4255 addupc_intr(td, ps->ps_pc[0], 1); 4256 } 4257 goto entrydone; 4258 } 4259 4260 /* 4261 * Otherwise, this is either a sampling mode PMC that 4262 * is attached to a different process than its owner, 4263 * or a system-wide sampling PMC. Dispatch a log 4264 * entry to the PMC's owner process. 4265 */ 4266 pmclog_process_callchain(pm, ps); 4267 4268 entrydone: 4269 ps->ps_nsamples = 0; /* mark entry as free */ 4270 atomic_subtract_rel_int(&pm->pm_runcount, 1); 4271 4272 /* increment read pointer, modulo sample size */ 4273 if (++ps == psb->ps_fence) 4274 psb->ps_read = psb->ps_samples; 4275 else 4276 psb->ps_read = ps; 4277 } 4278 4279 atomic_add_int(&pmc_stats.pm_log_sweeps, 1); 4280 4281 /* Do not re-enable stalled PMCs if we failed to process any samples */ 4282 if (n == 0) 4283 return; 4284 4285 /* 4286 * Restart any stalled sampling PMCs on this CPU. 4287 * 4288 * If the NMI handler sets the pm_stalled field of a PMC after 4289 * the check below, we'll end up processing the stalled PMC at 4290 * the next hardclock tick. 4291 */ 4292 for (n = 0; n < md->pmd_npmc; n++) { 4293 pcd = pmc_ri_to_classdep(md, n, &adjri); 4294 KASSERT(pcd != NULL, 4295 ("[pmc,%d] null pcd ri=%d", __LINE__, n)); 4296 (void) (*pcd->pcd_get_config)(cpu,adjri,&pm); 4297 4298 if (pm == NULL || /* !cfg'ed */ 4299 pm->pm_state != PMC_STATE_RUNNING || /* !active */ 4300 !PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) || /* !sampling */ 4301 pm->pm_stalled == 0) /* !stalled */ 4302 continue; 4303 4304 pm->pm_stalled = 0; 4305 (*pcd->pcd_start_pmc)(cpu, adjri); 4306 } 4307 } 4308 4309 /* 4310 * Event handlers. 4311 */ 4312 4313 /* 4314 * Handle a process exit. 4315 * 4316 * Remove this process from all hash tables. If this process 4317 * owned any PMCs, turn off those PMCs and deallocate them, 4318 * removing any associations with target processes. 4319 * 4320 * This function will be called by the last 'thread' of a 4321 * process. 4322 * 4323 * XXX This eventhandler gets called early in the exit process. 4324 * Consider using a 'hook' invocation from thread_exit() or equivalent 4325 * spot. Another negative is that kse_exit doesn't seem to call 4326 * exit1() [??]. 4327 * 4328 */ 4329 4330 static void 4331 pmc_process_exit(void *arg __unused, struct proc *p) 4332 { 4333 struct pmc *pm; 4334 int adjri, cpu; 4335 unsigned int ri; 4336 int is_using_hwpmcs; 4337 struct pmc_owner *po; 4338 struct pmc_process *pp; 4339 struct pmc_classdep *pcd; 4340 pmc_value_t newvalue, tmp; 4341 4342 PROC_LOCK(p); 4343 is_using_hwpmcs = p->p_flag & P_HWPMC; 4344 PROC_UNLOCK(p); 4345 4346 /* 4347 * Log a sysexit event to all SS PMC owners. 4348 */ 4349 LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) 4350 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 4351 pmclog_process_sysexit(po, p->p_pid); 4352 4353 if (!is_using_hwpmcs) 4354 return; 4355 4356 PMC_GET_SX_XLOCK(); 4357 PMCDBG(PRC,EXT,1,"process-exit proc=%p (%d, %s)", p, p->p_pid, 4358 p->p_comm); 4359 4360 /* 4361 * Since this code is invoked by the last thread in an exiting 4362 * process, we would have context switched IN at some prior 4363 * point. However, with PREEMPTION, kernel mode context 4364 * switches may happen any time, so we want to disable a 4365 * context switch OUT till we get any PMCs targetting this 4366 * process off the hardware. 4367 * 4368 * We also need to atomically remove this process' 4369 * entry from our target process hash table, using 4370 * PMC_FLAG_REMOVE. 4371 */ 4372 PMCDBG(PRC,EXT,1, "process-exit proc=%p (%d, %s)", p, p->p_pid, 4373 p->p_comm); 4374 4375 critical_enter(); /* no preemption */ 4376 4377 cpu = curthread->td_oncpu; 4378 4379 if ((pp = pmc_find_process_descriptor(p, 4380 PMC_FLAG_REMOVE)) != NULL) { 4381 4382 PMCDBG(PRC,EXT,2, 4383 "process-exit proc=%p pmc-process=%p", p, pp); 4384 4385 /* 4386 * The exiting process could the target of 4387 * some PMCs which will be running on 4388 * currently executing CPU. 4389 * 4390 * We need to turn these PMCs off like we 4391 * would do at context switch OUT time. 4392 */ 4393 for (ri = 0; ri < md->pmd_npmc; ri++) { 4394 4395 /* 4396 * Pick up the pmc pointer from hardware 4397 * state similar to the CSW_OUT code. 4398 */ 4399 pm = NULL; 4400 4401 pcd = pmc_ri_to_classdep(md, ri, &adjri); 4402 4403 (void) (*pcd->pcd_get_config)(cpu, adjri, &pm); 4404 4405 PMCDBG(PRC,EXT,2, "ri=%d pm=%p", ri, pm); 4406 4407 if (pm == NULL || 4408 !PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) 4409 continue; 4410 4411 PMCDBG(PRC,EXT,2, "ppmcs[%d]=%p pm=%p " 4412 "state=%d", ri, pp->pp_pmcs[ri].pp_pmc, 4413 pm, pm->pm_state); 4414 4415 KASSERT(PMC_TO_ROWINDEX(pm) == ri, 4416 ("[pmc,%d] ri mismatch pmc(%d) ri(%d)", 4417 __LINE__, PMC_TO_ROWINDEX(pm), ri)); 4418 4419 KASSERT(pm == pp->pp_pmcs[ri].pp_pmc, 4420 ("[pmc,%d] pm %p != pp_pmcs[%d] %p", 4421 __LINE__, pm, ri, pp->pp_pmcs[ri].pp_pmc)); 4422 4423 (void) pcd->pcd_stop_pmc(cpu, adjri); 4424 4425 KASSERT(pm->pm_runcount > 0, 4426 ("[pmc,%d] bad runcount ri %d rc %d", 4427 __LINE__, ri, pm->pm_runcount)); 4428 4429 /* Stop hardware only if it is actually running */ 4430 if (pm->pm_state == PMC_STATE_RUNNING && 4431 pm->pm_stalled == 0) { 4432 pcd->pcd_read_pmc(cpu, adjri, &newvalue); 4433 tmp = newvalue - 4434 PMC_PCPU_SAVED(cpu,ri); 4435 4436 mtx_pool_lock_spin(pmc_mtxpool, pm); 4437 pm->pm_gv.pm_savedvalue += tmp; 4438 pp->pp_pmcs[ri].pp_pmcval += tmp; 4439 mtx_pool_unlock_spin(pmc_mtxpool, pm); 4440 } 4441 4442 atomic_subtract_rel_int(&pm->pm_runcount,1); 4443 4444 KASSERT((int) pm->pm_runcount >= 0, 4445 ("[pmc,%d] runcount is %d", __LINE__, ri)); 4446 4447 (void) pcd->pcd_config_pmc(cpu, adjri, NULL); 4448 } 4449 4450 /* 4451 * Inform the MD layer of this pseudo "context switch 4452 * out" 4453 */ 4454 (void) md->pmd_switch_out(pmc_pcpu[cpu], pp); 4455 4456 critical_exit(); /* ok to be pre-empted now */ 4457 4458 /* 4459 * Unlink this process from the PMCs that are 4460 * targetting it. This will send a signal to 4461 * all PMC owner's whose PMCs are orphaned. 4462 * 4463 * Log PMC value at exit time if requested. 4464 */ 4465 for (ri = 0; ri < md->pmd_npmc; ri++) 4466 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) { 4467 if (pm->pm_flags & PMC_F_NEEDS_LOGFILE && 4468 PMC_IS_COUNTING_MODE(PMC_TO_MODE(pm))) 4469 pmclog_process_procexit(pm, pp); 4470 pmc_unlink_target_process(pm, pp); 4471 } 4472 free(pp, M_PMC); 4473 4474 } else 4475 critical_exit(); /* pp == NULL */ 4476 4477 4478 /* 4479 * If the process owned PMCs, free them up and free up 4480 * memory. 4481 */ 4482 if ((po = pmc_find_owner_descriptor(p)) != NULL) { 4483 pmc_remove_owner(po); 4484 pmc_destroy_owner_descriptor(po); 4485 } 4486 4487 sx_xunlock(&pmc_sx); 4488 } 4489 4490 /* 4491 * Handle a process fork. 4492 * 4493 * If the parent process 'p1' is under HWPMC monitoring, then copy 4494 * over any attached PMCs that have 'do_descendants' semantics. 4495 */ 4496 4497 static void 4498 pmc_process_fork(void *arg __unused, struct proc *p1, struct proc *newproc, 4499 int flags) 4500 { 4501 int is_using_hwpmcs; 4502 unsigned int ri; 4503 uint32_t do_descendants; 4504 struct pmc *pm; 4505 struct pmc_owner *po; 4506 struct pmc_process *ppnew, *ppold; 4507 4508 (void) flags; /* unused parameter */ 4509 4510 PROC_LOCK(p1); 4511 is_using_hwpmcs = p1->p_flag & P_HWPMC; 4512 PROC_UNLOCK(p1); 4513 4514 /* 4515 * If there are system-wide sampling PMCs active, we need to 4516 * log all fork events to their owner's logs. 4517 */ 4518 4519 LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) 4520 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 4521 pmclog_process_procfork(po, p1->p_pid, newproc->p_pid); 4522 4523 if (!is_using_hwpmcs) 4524 return; 4525 4526 PMC_GET_SX_XLOCK(); 4527 PMCDBG(PMC,FRK,1, "process-fork proc=%p (%d, %s) -> %p", p1, 4528 p1->p_pid, p1->p_comm, newproc); 4529 4530 /* 4531 * If the parent process (curthread->td_proc) is a 4532 * target of any PMCs, look for PMCs that are to be 4533 * inherited, and link these into the new process 4534 * descriptor. 4535 */ 4536 if ((ppold = pmc_find_process_descriptor(curthread->td_proc, 4537 PMC_FLAG_NONE)) == NULL) 4538 goto done; /* nothing to do */ 4539 4540 do_descendants = 0; 4541 for (ri = 0; ri < md->pmd_npmc; ri++) 4542 if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL) 4543 do_descendants |= pm->pm_flags & PMC_F_DESCENDANTS; 4544 if (do_descendants == 0) /* nothing to do */ 4545 goto done; 4546 4547 /* allocate a descriptor for the new process */ 4548 if ((ppnew = pmc_find_process_descriptor(newproc, 4549 PMC_FLAG_ALLOCATE)) == NULL) 4550 goto done; 4551 4552 /* 4553 * Run through all PMCs that were targeting the old process 4554 * and which specified F_DESCENDANTS and attach them to the 4555 * new process. 4556 * 4557 * Log the fork event to all owners of PMCs attached to this 4558 * process, if not already logged. 4559 */ 4560 for (ri = 0; ri < md->pmd_npmc; ri++) 4561 if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL && 4562 (pm->pm_flags & PMC_F_DESCENDANTS)) { 4563 pmc_link_target_process(pm, ppnew); 4564 po = pm->pm_owner; 4565 if (po->po_sscount == 0 && 4566 po->po_flags & PMC_PO_OWNS_LOGFILE) 4567 pmclog_process_procfork(po, p1->p_pid, 4568 newproc->p_pid); 4569 } 4570 4571 /* 4572 * Now mark the new process as being tracked by this driver. 4573 */ 4574 PROC_LOCK(newproc); 4575 newproc->p_flag |= P_HWPMC; 4576 PROC_UNLOCK(newproc); 4577 4578 done: 4579 sx_xunlock(&pmc_sx); 4580 } 4581 4582 static void 4583 pmc_kld_load(void *arg __unused, linker_file_t lf) 4584 { 4585 struct pmc_owner *po; 4586 4587 sx_slock(&pmc_sx); 4588 4589 /* 4590 * Notify owners of system sampling PMCs about KLD operations. 4591 */ 4592 LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) 4593 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 4594 pmclog_process_map_in(po, (pid_t) -1, 4595 (uintfptr_t) lf->address, lf->filename); 4596 4597 /* 4598 * TODO: Notify owners of (all) process-sampling PMCs too. 4599 */ 4600 4601 sx_sunlock(&pmc_sx); 4602 } 4603 4604 static void 4605 pmc_kld_unload(void *arg __unused, const char *filename __unused, 4606 caddr_t address, size_t size) 4607 { 4608 struct pmc_owner *po; 4609 4610 sx_slock(&pmc_sx); 4611 4612 LIST_FOREACH(po, &pmc_ss_owners, po_ssnext) 4613 if (po->po_flags & PMC_PO_OWNS_LOGFILE) 4614 pmclog_process_map_out(po, (pid_t) -1, 4615 (uintfptr_t) address, (uintfptr_t) address + size); 4616 4617 /* 4618 * TODO: Notify owners of process-sampling PMCs. 4619 */ 4620 4621 sx_sunlock(&pmc_sx); 4622 } 4623 4624 /* 4625 * initialization 4626 */ 4627 4628 static const char *pmc_name_of_pmcclass[] = { 4629 #undef __PMC_CLASS 4630 #define __PMC_CLASS(N) #N , 4631 __PMC_CLASSES() 4632 }; 4633 4634 /* 4635 * Base class initializer: allocate structure and set default classes. 4636 */ 4637 struct pmc_mdep * 4638 pmc_mdep_alloc(int nclasses) 4639 { 4640 struct pmc_mdep *md; 4641 int n; 4642 4643 /* SOFT + md classes */ 4644 n = 1 + nclasses; 4645 md = malloc(sizeof(struct pmc_mdep) + n * 4646 sizeof(struct pmc_classdep), M_PMC, M_WAITOK|M_ZERO); 4647 md->pmd_nclass = n; 4648 4649 /* Add base class. */ 4650 pmc_soft_initialize(md); 4651 return md; 4652 } 4653 4654 void 4655 pmc_mdep_free(struct pmc_mdep *md) 4656 { 4657 pmc_soft_finalize(md); 4658 free(md, M_PMC); 4659 } 4660 4661 static int 4662 generic_switch_in(struct pmc_cpu *pc, struct pmc_process *pp) 4663 { 4664 (void) pc; (void) pp; 4665 4666 return (0); 4667 } 4668 4669 static int 4670 generic_switch_out(struct pmc_cpu *pc, struct pmc_process *pp) 4671 { 4672 (void) pc; (void) pp; 4673 4674 return (0); 4675 } 4676 4677 static struct pmc_mdep * 4678 pmc_generic_cpu_initialize(void) 4679 { 4680 struct pmc_mdep *md; 4681 4682 md = pmc_mdep_alloc(0); 4683 4684 md->pmd_cputype = PMC_CPU_GENERIC; 4685 4686 md->pmd_pcpu_init = NULL; 4687 md->pmd_pcpu_fini = NULL; 4688 md->pmd_switch_in = generic_switch_in; 4689 md->pmd_switch_out = generic_switch_out; 4690 4691 return (md); 4692 } 4693 4694 static void 4695 pmc_generic_cpu_finalize(struct pmc_mdep *md) 4696 { 4697 (void) md; 4698 } 4699 4700 4701 static int 4702 pmc_initialize(void) 4703 { 4704 int c, cpu, error, n, ri; 4705 unsigned int maxcpu; 4706 struct pmc_binding pb; 4707 struct pmc_sample *ps; 4708 struct pmc_classdep *pcd; 4709 struct pmc_samplebuffer *sb; 4710 4711 md = NULL; 4712 error = 0; 4713 4714 #ifdef DEBUG 4715 /* parse debug flags first */ 4716 if (TUNABLE_STR_FETCH(PMC_SYSCTL_NAME_PREFIX "debugflags", 4717 pmc_debugstr, sizeof(pmc_debugstr))) 4718 pmc_debugflags_parse(pmc_debugstr, 4719 pmc_debugstr+strlen(pmc_debugstr)); 4720 #endif 4721 4722 PMCDBG(MOD,INI,0, "PMC Initialize (version %x)", PMC_VERSION); 4723 4724 /* check kernel version */ 4725 if (pmc_kernel_version != PMC_VERSION) { 4726 if (pmc_kernel_version == 0) 4727 printf("hwpmc: this kernel has not been compiled with " 4728 "'options HWPMC_HOOKS'.\n"); 4729 else 4730 printf("hwpmc: kernel version (0x%x) does not match " 4731 "module version (0x%x).\n", pmc_kernel_version, 4732 PMC_VERSION); 4733 return EPROGMISMATCH; 4734 } 4735 4736 /* 4737 * check sysctl parameters 4738 */ 4739 4740 if (pmc_hashsize <= 0) { 4741 (void) printf("hwpmc: tunable \"hashsize\"=%d must be " 4742 "greater than zero.\n", pmc_hashsize); 4743 pmc_hashsize = PMC_HASH_SIZE; 4744 } 4745 4746 if (pmc_nsamples <= 0 || pmc_nsamples > 65535) { 4747 (void) printf("hwpmc: tunable \"nsamples\"=%d out of " 4748 "range.\n", pmc_nsamples); 4749 pmc_nsamples = PMC_NSAMPLES; 4750 } 4751 4752 if (pmc_callchaindepth <= 0 || 4753 pmc_callchaindepth > PMC_CALLCHAIN_DEPTH_MAX) { 4754 (void) printf("hwpmc: tunable \"callchaindepth\"=%d out of " 4755 "range.\n", pmc_callchaindepth); 4756 pmc_callchaindepth = PMC_CALLCHAIN_DEPTH; 4757 } 4758 4759 md = pmc_md_initialize(); 4760 if (md == NULL) { 4761 /* Default to generic CPU. */ 4762 md = pmc_generic_cpu_initialize(); 4763 if (md == NULL) 4764 return (ENOSYS); 4765 } 4766 4767 KASSERT(md->pmd_nclass >= 1 && md->pmd_npmc >= 1, 4768 ("[pmc,%d] no classes or pmcs", __LINE__)); 4769 4770 /* Compute the map from row-indices to classdep pointers. */ 4771 pmc_rowindex_to_classdep = malloc(sizeof(struct pmc_classdep *) * 4772 md->pmd_npmc, M_PMC, M_WAITOK|M_ZERO); 4773 4774 for (n = 0; n < md->pmd_npmc; n++) 4775 pmc_rowindex_to_classdep[n] = NULL; 4776 for (ri = c = 0; c < md->pmd_nclass; c++) { 4777 pcd = &md->pmd_classdep[c]; 4778 for (n = 0; n < pcd->pcd_num; n++, ri++) 4779 pmc_rowindex_to_classdep[ri] = pcd; 4780 } 4781 4782 KASSERT(ri == md->pmd_npmc, 4783 ("[pmc,%d] npmc miscomputed: ri=%d, md->npmc=%d", __LINE__, 4784 ri, md->pmd_npmc)); 4785 4786 maxcpu = pmc_cpu_max(); 4787 4788 /* allocate space for the per-cpu array */ 4789 pmc_pcpu = malloc(maxcpu * sizeof(struct pmc_cpu *), M_PMC, 4790 M_WAITOK|M_ZERO); 4791 4792 /* per-cpu 'saved values' for managing process-mode PMCs */ 4793 pmc_pcpu_saved = malloc(sizeof(pmc_value_t) * maxcpu * md->pmd_npmc, 4794 M_PMC, M_WAITOK); 4795 4796 /* Perform CPU-dependent initialization. */ 4797 pmc_save_cpu_binding(&pb); 4798 error = 0; 4799 for (cpu = 0; error == 0 && cpu < maxcpu; cpu++) { 4800 if (!pmc_cpu_is_active(cpu)) 4801 continue; 4802 pmc_select_cpu(cpu); 4803 pmc_pcpu[cpu] = malloc(sizeof(struct pmc_cpu) + 4804 md->pmd_npmc * sizeof(struct pmc_hw *), M_PMC, 4805 M_WAITOK|M_ZERO); 4806 if (md->pmd_pcpu_init) 4807 error = md->pmd_pcpu_init(md, cpu); 4808 for (n = 0; error == 0 && n < md->pmd_nclass; n++) 4809 error = md->pmd_classdep[n].pcd_pcpu_init(md, cpu); 4810 } 4811 pmc_restore_cpu_binding(&pb); 4812 4813 if (error) 4814 return (error); 4815 4816 /* allocate space for the sample array */ 4817 for (cpu = 0; cpu < maxcpu; cpu++) { 4818 if (!pmc_cpu_is_active(cpu)) 4819 continue; 4820 4821 sb = malloc(sizeof(struct pmc_samplebuffer) + 4822 pmc_nsamples * sizeof(struct pmc_sample), M_PMC, 4823 M_WAITOK|M_ZERO); 4824 sb->ps_read = sb->ps_write = sb->ps_samples; 4825 sb->ps_fence = sb->ps_samples + pmc_nsamples; 4826 4827 KASSERT(pmc_pcpu[cpu] != NULL, 4828 ("[pmc,%d] cpu=%d Null per-cpu data", __LINE__, cpu)); 4829 4830 sb->ps_callchains = malloc(pmc_callchaindepth * pmc_nsamples * 4831 sizeof(uintptr_t), M_PMC, M_WAITOK|M_ZERO); 4832 4833 for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++) 4834 ps->ps_pc = sb->ps_callchains + 4835 (n * pmc_callchaindepth); 4836 4837 pmc_pcpu[cpu]->pc_sb[PMC_HR] = sb; 4838 4839 sb = malloc(sizeof(struct pmc_samplebuffer) + 4840 pmc_nsamples * sizeof(struct pmc_sample), M_PMC, 4841 M_WAITOK|M_ZERO); 4842 sb->ps_read = sb->ps_write = sb->ps_samples; 4843 sb->ps_fence = sb->ps_samples + pmc_nsamples; 4844 4845 KASSERT(pmc_pcpu[cpu] != NULL, 4846 ("[pmc,%d] cpu=%d Null per-cpu data", __LINE__, cpu)); 4847 4848 sb->ps_callchains = malloc(pmc_callchaindepth * pmc_nsamples * 4849 sizeof(uintptr_t), M_PMC, M_WAITOK|M_ZERO); 4850 4851 for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++) 4852 ps->ps_pc = sb->ps_callchains + 4853 (n * pmc_callchaindepth); 4854 4855 pmc_pcpu[cpu]->pc_sb[PMC_SR] = sb; 4856 } 4857 4858 /* allocate space for the row disposition array */ 4859 pmc_pmcdisp = malloc(sizeof(enum pmc_mode) * md->pmd_npmc, 4860 M_PMC, M_WAITOK|M_ZERO); 4861 4862 /* mark all PMCs as available */ 4863 for (n = 0; n < (int) md->pmd_npmc; n++) 4864 PMC_MARK_ROW_FREE(n); 4865 4866 /* allocate thread hash tables */ 4867 pmc_ownerhash = hashinit(pmc_hashsize, M_PMC, 4868 &pmc_ownerhashmask); 4869 4870 pmc_processhash = hashinit(pmc_hashsize, M_PMC, 4871 &pmc_processhashmask); 4872 mtx_init(&pmc_processhash_mtx, "pmc-process-hash", "pmc-leaf", 4873 MTX_SPIN); 4874 4875 LIST_INIT(&pmc_ss_owners); 4876 pmc_ss_count = 0; 4877 4878 /* allocate a pool of spin mutexes */ 4879 pmc_mtxpool = mtx_pool_create("pmc-leaf", pmc_mtxpool_size, 4880 MTX_SPIN); 4881 4882 PMCDBG(MOD,INI,1, "pmc_ownerhash=%p, mask=0x%lx " 4883 "targethash=%p mask=0x%lx", pmc_ownerhash, pmc_ownerhashmask, 4884 pmc_processhash, pmc_processhashmask); 4885 4886 /* register process {exit,fork,exec} handlers */ 4887 pmc_exit_tag = EVENTHANDLER_REGISTER(process_exit, 4888 pmc_process_exit, NULL, EVENTHANDLER_PRI_ANY); 4889 pmc_fork_tag = EVENTHANDLER_REGISTER(process_fork, 4890 pmc_process_fork, NULL, EVENTHANDLER_PRI_ANY); 4891 4892 /* register kld event handlers */ 4893 pmc_kld_load_tag = EVENTHANDLER_REGISTER(kld_load, pmc_kld_load, 4894 NULL, EVENTHANDLER_PRI_ANY); 4895 pmc_kld_unload_tag = EVENTHANDLER_REGISTER(kld_unload, pmc_kld_unload, 4896 NULL, EVENTHANDLER_PRI_ANY); 4897 4898 /* initialize logging */ 4899 pmclog_initialize(); 4900 4901 /* set hook functions */ 4902 pmc_intr = md->pmd_intr; 4903 pmc_hook = pmc_hook_handler; 4904 4905 if (error == 0) { 4906 printf(PMC_MODULE_NAME ":"); 4907 for (n = 0; n < (int) md->pmd_nclass; n++) { 4908 pcd = &md->pmd_classdep[n]; 4909 printf(" %s/%d/%d/0x%b", 4910 pmc_name_of_pmcclass[pcd->pcd_class], 4911 pcd->pcd_num, 4912 pcd->pcd_width, 4913 pcd->pcd_caps, 4914 "\20" 4915 "\1INT\2USR\3SYS\4EDG\5THR" 4916 "\6REA\7WRI\10INV\11QUA\12PRC" 4917 "\13TAG\14CSC"); 4918 } 4919 printf("\n"); 4920 } 4921 4922 return (error); 4923 } 4924 4925 /* prepare to be unloaded */ 4926 static void 4927 pmc_cleanup(void) 4928 { 4929 int c, cpu; 4930 unsigned int maxcpu; 4931 struct pmc_ownerhash *ph; 4932 struct pmc_owner *po, *tmp; 4933 struct pmc_binding pb; 4934 #ifdef DEBUG 4935 struct pmc_processhash *prh; 4936 #endif 4937 4938 PMCDBG(MOD,INI,0, "%s", "cleanup"); 4939 4940 /* switch off sampling */ 4941 CPU_ZERO(&pmc_cpumask); 4942 pmc_intr = NULL; 4943 4944 sx_xlock(&pmc_sx); 4945 if (pmc_hook == NULL) { /* being unloaded already */ 4946 sx_xunlock(&pmc_sx); 4947 return; 4948 } 4949 4950 pmc_hook = NULL; /* prevent new threads from entering module */ 4951 4952 /* deregister event handlers */ 4953 EVENTHANDLER_DEREGISTER(process_fork, pmc_fork_tag); 4954 EVENTHANDLER_DEREGISTER(process_exit, pmc_exit_tag); 4955 EVENTHANDLER_DEREGISTER(kld_load, pmc_kld_load_tag); 4956 EVENTHANDLER_DEREGISTER(kld_unload, pmc_kld_unload_tag); 4957 4958 /* send SIGBUS to all owner threads, free up allocations */ 4959 if (pmc_ownerhash) 4960 for (ph = pmc_ownerhash; 4961 ph <= &pmc_ownerhash[pmc_ownerhashmask]; 4962 ph++) { 4963 LIST_FOREACH_SAFE(po, ph, po_next, tmp) { 4964 pmc_remove_owner(po); 4965 4966 /* send SIGBUS to owner processes */ 4967 PMCDBG(MOD,INI,2, "cleanup signal proc=%p " 4968 "(%d, %s)", po->po_owner, 4969 po->po_owner->p_pid, 4970 po->po_owner->p_comm); 4971 4972 PROC_LOCK(po->po_owner); 4973 kern_psignal(po->po_owner, SIGBUS); 4974 PROC_UNLOCK(po->po_owner); 4975 4976 pmc_destroy_owner_descriptor(po); 4977 } 4978 } 4979 4980 /* reclaim allocated data structures */ 4981 if (pmc_mtxpool) 4982 mtx_pool_destroy(&pmc_mtxpool); 4983 4984 mtx_destroy(&pmc_processhash_mtx); 4985 if (pmc_processhash) { 4986 #ifdef DEBUG 4987 struct pmc_process *pp; 4988 4989 PMCDBG(MOD,INI,3, "%s", "destroy process hash"); 4990 for (prh = pmc_processhash; 4991 prh <= &pmc_processhash[pmc_processhashmask]; 4992 prh++) 4993 LIST_FOREACH(pp, prh, pp_next) 4994 PMCDBG(MOD,INI,3, "pid=%d", pp->pp_proc->p_pid); 4995 #endif 4996 4997 hashdestroy(pmc_processhash, M_PMC, pmc_processhashmask); 4998 pmc_processhash = NULL; 4999 } 5000 5001 if (pmc_ownerhash) { 5002 PMCDBG(MOD,INI,3, "%s", "destroy owner hash"); 5003 hashdestroy(pmc_ownerhash, M_PMC, pmc_ownerhashmask); 5004 pmc_ownerhash = NULL; 5005 } 5006 5007 KASSERT(LIST_EMPTY(&pmc_ss_owners), 5008 ("[pmc,%d] Global SS owner list not empty", __LINE__)); 5009 KASSERT(pmc_ss_count == 0, 5010 ("[pmc,%d] Global SS count not empty", __LINE__)); 5011 5012 /* do processor and pmc-class dependent cleanup */ 5013 maxcpu = pmc_cpu_max(); 5014 5015 PMCDBG(MOD,INI,3, "%s", "md cleanup"); 5016 if (md) { 5017 pmc_save_cpu_binding(&pb); 5018 for (cpu = 0; cpu < maxcpu; cpu++) { 5019 PMCDBG(MOD,INI,1,"pmc-cleanup cpu=%d pcs=%p", 5020 cpu, pmc_pcpu[cpu]); 5021 if (!pmc_cpu_is_active(cpu) || pmc_pcpu[cpu] == NULL) 5022 continue; 5023 pmc_select_cpu(cpu); 5024 for (c = 0; c < md->pmd_nclass; c++) 5025 md->pmd_classdep[c].pcd_pcpu_fini(md, cpu); 5026 if (md->pmd_pcpu_fini) 5027 md->pmd_pcpu_fini(md, cpu); 5028 } 5029 5030 if (md->pmd_cputype == PMC_CPU_GENERIC) 5031 pmc_generic_cpu_finalize(md); 5032 else 5033 pmc_md_finalize(md); 5034 5035 pmc_mdep_free(md); 5036 md = NULL; 5037 pmc_restore_cpu_binding(&pb); 5038 } 5039 5040 /* Free per-cpu descriptors. */ 5041 for (cpu = 0; cpu < maxcpu; cpu++) { 5042 if (!pmc_cpu_is_active(cpu)) 5043 continue; 5044 KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_HR] != NULL, 5045 ("[pmc,%d] Null hw cpu sample buffer cpu=%d", __LINE__, 5046 cpu)); 5047 KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_SR] != NULL, 5048 ("[pmc,%d] Null sw cpu sample buffer cpu=%d", __LINE__, 5049 cpu)); 5050 free(pmc_pcpu[cpu]->pc_sb[PMC_HR]->ps_callchains, M_PMC); 5051 free(pmc_pcpu[cpu]->pc_sb[PMC_HR], M_PMC); 5052 free(pmc_pcpu[cpu]->pc_sb[PMC_SR]->ps_callchains, M_PMC); 5053 free(pmc_pcpu[cpu]->pc_sb[PMC_SR], M_PMC); 5054 free(pmc_pcpu[cpu], M_PMC); 5055 } 5056 5057 free(pmc_pcpu, M_PMC); 5058 pmc_pcpu = NULL; 5059 5060 free(pmc_pcpu_saved, M_PMC); 5061 pmc_pcpu_saved = NULL; 5062 5063 if (pmc_pmcdisp) { 5064 free(pmc_pmcdisp, M_PMC); 5065 pmc_pmcdisp = NULL; 5066 } 5067 5068 if (pmc_rowindex_to_classdep) { 5069 free(pmc_rowindex_to_classdep, M_PMC); 5070 pmc_rowindex_to_classdep = NULL; 5071 } 5072 5073 pmclog_shutdown(); 5074 5075 sx_xunlock(&pmc_sx); /* we are done */ 5076 } 5077 5078 /* 5079 * The function called at load/unload. 5080 */ 5081 5082 static int 5083 load (struct module *module __unused, int cmd, void *arg __unused) 5084 { 5085 int error; 5086 5087 error = 0; 5088 5089 switch (cmd) { 5090 case MOD_LOAD : 5091 /* initialize the subsystem */ 5092 error = pmc_initialize(); 5093 if (error != 0) 5094 break; 5095 PMCDBG(MOD,INI,1, "syscall=%d maxcpu=%d", 5096 pmc_syscall_num, pmc_cpu_max()); 5097 break; 5098 5099 5100 case MOD_UNLOAD : 5101 case MOD_SHUTDOWN: 5102 pmc_cleanup(); 5103 PMCDBG(MOD,INI,1, "%s", "unloaded"); 5104 break; 5105 5106 default : 5107 error = EINVAL; /* XXX should panic(9) */ 5108 break; 5109 } 5110 5111 return error; 5112 } 5113 5114 /* memory pool */ 5115 MALLOC_DEFINE(M_PMC, "pmc", "Memory space for the PMC module"); 5116