sys/dev/hwpmc/hwpmc_mod.c

   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  HWPMC_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 void     pmc_destroy_pmc_descriptor(struct pmc *pm);
 196 static struct pmc_owner *pmc_find_owner_descriptor(struct proc *p);
 197 static int      pmc_find_pmc(pmc_id_t pmcid, struct pmc **pm);
 198 static struct pmc *pmc_find_pmc_descriptor_in_process(struct pmc_owner *po,
 199     pmc_id_t pmc);
 200 static struct pmc_process *pmc_find_process_descriptor(struct proc *p,
 201     uint32_t mode);
 202 static void     pmc_force_context_switch(void);
 203 static void     pmc_link_target_process(struct pmc *pm,
 204     struct pmc_process *pp);
 205 static void     pmc_log_all_process_mappings(struct pmc_owner *po);
 206 static void     pmc_log_kernel_mappings(struct pmc *pm);
 207 static void     pmc_log_process_mappings(struct pmc_owner *po, struct proc *p);
 208 static void     pmc_maybe_remove_owner(struct pmc_owner *po);
 209 static void     pmc_process_csw_in(struct thread *td);
 210 static void     pmc_process_csw_out(struct thread *td);
 211 static void     pmc_process_exit(void *arg, struct proc *p);
 212 static void     pmc_process_fork(void *arg, struct proc *p1,
 213     struct proc *p2, int n);
 214 static void     pmc_process_samples(int cpu, int soft);
 215 static void     pmc_release_pmc_descriptor(struct pmc *pmc);
 216 static void     pmc_remove_owner(struct pmc_owner *po);
 217 static void     pmc_remove_process_descriptor(struct pmc_process *pp);
 218 static void     pmc_restore_cpu_binding(struct pmc_binding *pb);
 219 static void     pmc_save_cpu_binding(struct pmc_binding *pb);
 220 static void     pmc_select_cpu(int cpu);
 221 static int      pmc_start(struct pmc *pm);
 222 static int      pmc_stop(struct pmc *pm);
 223 static int      pmc_syscall_handler(struct thread *td, void *syscall_args);
 224 static void     pmc_unlink_target_process(struct pmc *pmc,
 225     struct pmc_process *pp);
 226 static int generic_switch_in(struct pmc_cpu *pc, struct pmc_process *pp);
 227 static int generic_switch_out(struct pmc_cpu *pc, struct pmc_process *pp);
 228 static struct pmc_mdep *pmc_generic_cpu_initialize(void);
 229 static void pmc_generic_cpu_finalize(struct pmc_mdep *md);
 230
 231 /*
 232  * Kernel tunables and sysctl(8) interface.
 233  */
 234
 235 SYSCTL_DECL(_kern_hwpmc);
 236
 237 static int pmc_callchaindepth = PMC_CALLCHAIN_DEPTH;
 238 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "callchaindepth", &pmc_callchaindepth);
 239 SYSCTL_INT(_kern_hwpmc, OID_AUTO, callchaindepth, CTLFLAG_TUN|CTLFLAG_RD,
 240     &pmc_callchaindepth, 0, "depth of call chain records");
 241
 242 #ifdef  HWPMC_DEBUG
 243 struct pmc_debugflags pmc_debugflags = PMC_DEBUG_DEFAULT_FLAGS;
 244 char    pmc_debugstr[PMC_DEBUG_STRSIZE];
 245 TUNABLE_STR(PMC_SYSCTL_NAME_PREFIX "debugflags", pmc_debugstr,
 246     sizeof(pmc_debugstr));
 247 SYSCTL_PROC(_kern_hwpmc, OID_AUTO, debugflags,
 248     CTLTYPE_STRING|CTLFLAG_RW|CTLFLAG_TUN,
 249     0, 0, pmc_debugflags_sysctl_handler, "A", "debug flags");
 250 #endif
 251
 252 /*
 253  * kern.hwpmc.hashrows -- determines the number of rows in the
 254  * of the hash table used to look up threads
 255  */
 256
 257 static int pmc_hashsize = PMC_HASH_SIZE;
 258 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "hashsize", &pmc_hashsize);
 259 SYSCTL_INT(_kern_hwpmc, OID_AUTO, hashsize, CTLFLAG_TUN|CTLFLAG_RD,
 260     &pmc_hashsize, 0, "rows in hash tables");
 261
 262 /*
 263  * kern.hwpmc.nsamples --- number of PC samples/callchain stacks per CPU
 264  */
 265
 266 static int pmc_nsamples = PMC_NSAMPLES;
 267 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "nsamples", &pmc_nsamples);
 268 SYSCTL_INT(_kern_hwpmc, OID_AUTO, nsamples, CTLFLAG_TUN|CTLFLAG_RD,
 269     &pmc_nsamples, 0, "number of PC samples per CPU");
 270
 271
 272 /*
 273  * kern.hwpmc.mtxpoolsize -- number of mutexes in the mutex pool.
 274  */
 275
 276 static int pmc_mtxpool_size = PMC_MTXPOOL_SIZE;
 277 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "mtxpoolsize", &pmc_mtxpool_size);
 278 SYSCTL_INT(_kern_hwpmc, OID_AUTO, mtxpoolsize, CTLFLAG_TUN|CTLFLAG_RD,
 279     &pmc_mtxpool_size, 0, "size of spin mutex pool");
 280
 281
 282 /*
 283  * security.bsd.unprivileged_syspmcs -- allow non-root processes to
 284  * allocate system-wide PMCs.
 285  *
 286  * Allowing unprivileged processes to allocate system PMCs is convenient
 287  * if system-wide measurements need to be taken concurrently with other
 288  * per-process measurements.  This feature is turned off by default.
 289  */
 290
 291 static int pmc_unprivileged_syspmcs = 0;
 292 TUNABLE_INT("security.bsd.unprivileged_syspmcs", &pmc_unprivileged_syspmcs);
 293 SYSCTL_INT(_security_bsd, OID_AUTO, unprivileged_syspmcs, CTLFLAG_RW,
 294     &pmc_unprivileged_syspmcs, 0,
 295     "allow unprivileged process to allocate system PMCs");
 296
 297 /*
 298  * Hash function.  Discard the lower 2 bits of the pointer since
 299  * these are always zero for our uses.  The hash multiplier is
 300  * round((2^LONG_BIT) * ((sqrt(5)-1)/2)).
 301  */
 302
 303 #if     LONG_BIT == 64
 304 #define _PMC_HM         11400714819323198486u
 305 #elif   LONG_BIT == 32
 306 #define _PMC_HM         2654435769u
 307 #else
 308 #error  Must know the size of 'long' to compile
 309 #endif
 310
 311 #define PMC_HASH_PTR(P,M)       ((((unsigned long) (P) >> 2) * _PMC_HM) & (M))
 312
 313 /*
 314  * Syscall structures
 315  */
 316
 317 /* The `sysent' for the new syscall */
 318 static struct sysent pmc_sysent = {
 319         2,                      /* sy_narg */
 320         pmc_syscall_handler     /* sy_call */
 321 };
 322
 323 static struct syscall_module_data pmc_syscall_mod = {
 324         load,
 325         NULL,
 326         &pmc_syscall_num,
 327         &pmc_sysent,
 328 #if (__FreeBSD_version >= 1100000)
 329         { 0, NULL },
 330         SY_THR_STATIC_KLD,
 331 #else
 332         { 0, NULL }
 333 #endif
 334 };
 335
 336 static moduledata_t pmc_mod = {
 337         PMC_MODULE_NAME,
 338         syscall_module_handler,
 339         &pmc_syscall_mod
 340 };
 341
 342 DECLARE_MODULE(pmc, pmc_mod, SI_SUB_SMP, SI_ORDER_ANY);
 343 MODULE_VERSION(pmc, PMC_VERSION);
 344
 345 #ifdef  HWPMC_DEBUG
 346 enum pmc_dbgparse_state {
 347         PMCDS_WS,               /* in whitespace */
 348         PMCDS_MAJOR,            /* seen a major keyword */
 349         PMCDS_MINOR
 350 };
 351
 352 static int
 353 pmc_debugflags_parse(char *newstr, char *fence)
 354 {
 355         char c, *p, *q;
 356         struct pmc_debugflags *tmpflags;
 357         int error, found, *newbits, tmp;
 358         size_t kwlen;
 359
 360         tmpflags = malloc(sizeof(*tmpflags), M_PMC, M_WAITOK|M_ZERO);
 361
 362         p = newstr;
 363         error = 0;
 364
 365         for (; p < fence && (c = *p); p++) {
 366
 367                 /* skip white space */
 368                 if (c == ' ' || c == '\t')
 369                         continue;
 370
 371                 /* look for a keyword followed by "=" */
 372                 for (q = p; p < fence && (c = *p) && c != '='; p++)
 373                         ;
 374                 if (c != '=') {
 375                         error = EINVAL;
 376                         goto done;
 377                 }
 378
 379                 kwlen = p - q;
 380                 newbits = NULL;
 381
 382                 /* lookup flag group name */
 383 #define DBG_SET_FLAG_MAJ(S,F)                                           \
 384                 if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0)  \
 385                         newbits = &tmpflags->pdb_ ## F;
 386
 387                 DBG_SET_FLAG_MAJ("cpu",         CPU);
 388                 DBG_SET_FLAG_MAJ("csw",         CSW);
 389                 DBG_SET_FLAG_MAJ("logging",     LOG);
 390                 DBG_SET_FLAG_MAJ("module",      MOD);
 391                 DBG_SET_FLAG_MAJ("md",          MDP);
 392                 DBG_SET_FLAG_MAJ("owner",       OWN);
 393                 DBG_SET_FLAG_MAJ("pmc",         PMC);
 394                 DBG_SET_FLAG_MAJ("process",     PRC);
 395                 DBG_SET_FLAG_MAJ("sampling",    SAM);
 396
 397                 if (newbits == NULL) {
 398                         error = EINVAL;
 399                         goto done;
 400                 }
 401
 402                 p++;            /* skip the '=' */
 403
 404                 /* Now parse the individual flags */
 405                 tmp = 0;
 406         newflag:
 407                 for (q = p; p < fence && (c = *p); p++)
 408                         if (c == ' ' || c == '\t' || c == ',')
 409                                 break;
 410
 411                 /* p == fence or c == ws or c == "," or c == 0 */
 412
 413                 if ((kwlen = p - q) == 0) {
 414                         *newbits = tmp;
 415                         continue;
 416                 }
 417
 418                 found = 0;
 419 #define DBG_SET_FLAG_MIN(S,F)                                           \
 420                 if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0)  \
 421                         tmp |= found = (1 << PMC_DEBUG_MIN_ ## F)
 422
 423                 /* a '*' denotes all possible flags in the group */
 424                 if (kwlen == 1 && *q == '*')
 425                         tmp = found = ~0;
 426                 /* look for individual flag names */
 427                 DBG_SET_FLAG_MIN("allocaterow", ALR);
 428                 DBG_SET_FLAG_MIN("allocate",    ALL);
 429                 DBG_SET_FLAG_MIN("attach",      ATT);
 430                 DBG_SET_FLAG_MIN("bind",        BND);
 431                 DBG_SET_FLAG_MIN("config",      CFG);
 432                 DBG_SET_FLAG_MIN("exec",        EXC);
 433                 DBG_SET_FLAG_MIN("exit",        EXT);
 434                 DBG_SET_FLAG_MIN("find",        FND);
 435                 DBG_SET_FLAG_MIN("flush",       FLS);
 436                 DBG_SET_FLAG_MIN("fork",        FRK);
 437                 DBG_SET_FLAG_MIN("getbuf",      GTB);
 438                 DBG_SET_FLAG_MIN("hook",        PMH);
 439                 DBG_SET_FLAG_MIN("init",        INI);
 440                 DBG_SET_FLAG_MIN("intr",        INT);
 441                 DBG_SET_FLAG_MIN("linktarget",  TLK);
 442                 DBG_SET_FLAG_MIN("mayberemove", OMR);
 443                 DBG_SET_FLAG_MIN("ops",         OPS);
 444                 DBG_SET_FLAG_MIN("read",        REA);
 445                 DBG_SET_FLAG_MIN("register",    REG);
 446                 DBG_SET_FLAG_MIN("release",     REL);
 447                 DBG_SET_FLAG_MIN("remove",      ORM);
 448                 DBG_SET_FLAG_MIN("sample",      SAM);
 449                 DBG_SET_FLAG_MIN("scheduleio",  SIO);
 450                 DBG_SET_FLAG_MIN("select",      SEL);
 451                 DBG_SET_FLAG_MIN("signal",      SIG);
 452                 DBG_SET_FLAG_MIN("swi",         SWI);
 453                 DBG_SET_FLAG_MIN("swo",         SWO);
 454                 DBG_SET_FLAG_MIN("start",       STA);
 455                 DBG_SET_FLAG_MIN("stop",        STO);
 456                 DBG_SET_FLAG_MIN("syscall",     PMS);
 457                 DBG_SET_FLAG_MIN("unlinktarget", TUL);
 458                 DBG_SET_FLAG_MIN("write",       WRI);
 459                 if (found == 0) {
 460                         /* unrecognized flag name */
 461                         error = EINVAL;
 462                         goto done;
 463                 }
 464
 465                 if (c == 0 || c == ' ' || c == '\t') {  /* end of flag group */
 466                         *newbits = tmp;
 467                         continue;
 468                 }
 469
 470                 p++;
 471                 goto newflag;
 472         }
 473
 474         /* save the new flag set */
 475         bcopy(tmpflags, &pmc_debugflags, sizeof(pmc_debugflags));
 476
 477  done:
 478         free(tmpflags, M_PMC);
 479         return error;
 480 }
 481
 482 static int
 483 pmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS)
 484 {
 485         char *fence, *newstr;
 486         int error;
 487         unsigned int n;
 488
 489         (void) arg1; (void) arg2; /* unused parameters */
 490
 491         n = sizeof(pmc_debugstr);
 492         newstr = malloc(n, M_PMC, M_WAITOK|M_ZERO);
 493         (void) strlcpy(newstr, pmc_debugstr, n);
 494
 495         error = sysctl_handle_string(oidp, newstr, n, req);
 496
 497         /* if there is a new string, parse and copy it */
 498         if (error == 0 && req->newptr != NULL) {
 499                 fence = newstr + (n < req->newlen ? n : req->newlen + 1);
 500                 if ((error = pmc_debugflags_parse(newstr, fence)) == 0)
 501                         (void) strlcpy(pmc_debugstr, newstr,
 502                             sizeof(pmc_debugstr));
 503         }
 504
 505         free(newstr, M_PMC);
 506
 507         return error;
 508 }
 509 #endif
 510
 511 /*
 512  * Map a row index to a classdep structure and return the adjusted row
 513  * index for the PMC class index.
 514  */
 515 static struct pmc_classdep *
 516 pmc_ri_to_classdep(struct pmc_mdep *md, int ri, int *adjri)
 517 {
 518         struct pmc_classdep *pcd;
 519
 520         (void) md;
 521
 522         KASSERT(ri >= 0 && ri < md->pmd_npmc,
 523             ("[pmc,%d] illegal row-index %d", __LINE__, ri));
 524
 525         pcd = pmc_rowindex_to_classdep[ri];
 526
 527         KASSERT(pcd != NULL,
 528             ("[pmc,%d] ri %d null pcd", __LINE__, ri));
 529
 530         *adjri = ri - pcd->pcd_ri;
 531
 532         KASSERT(*adjri >= 0 && *adjri < pcd->pcd_num,
 533             ("[pmc,%d] adjusted row-index %d", __LINE__, *adjri));
 534
 535         return (pcd);
 536 }
 537
 538 /*
 539  * Concurrency Control
 540  *
 541  * The driver manages the following data structures:
 542  *
 543  *   - target process descriptors, one per target process
 544  *   - owner process descriptors (and attached lists), one per owner process
 545  *   - lookup hash tables for owner and target processes
 546  *   - PMC descriptors (and attached lists)
 547  *   - per-cpu hardware state
 548  *   - the 'hook' variable through which the kernel calls into
 549  *     this module
 550  *   - the machine hardware state (managed by the MD layer)
 551  *
 552  * These data structures are accessed from:
 553  *
 554  * - thread context-switch code
 555  * - interrupt handlers (possibly on multiple cpus)
 556  * - kernel threads on multiple cpus running on behalf of user
 557  *   processes doing system calls
 558  * - this driver's private kernel threads
 559  *
 560  * = Locks and Locking strategy =
 561  *
 562  * The driver uses four locking strategies for its operation:
 563  *
 564  * - The global SX lock "pmc_sx" is used to protect internal
 565  *   data structures.
 566  *
 567  *   Calls into the module by syscall() start with this lock being
 568  *   held in exclusive mode.  Depending on the requested operation,
 569  *   the lock may be downgraded to 'shared' mode to allow more
 570  *   concurrent readers into the module.  Calls into the module from
 571  *   other parts of the kernel acquire the lock in shared mode.
 572  *
 573  *   This SX lock is held in exclusive mode for any operations that
 574  *   modify the linkages between the driver's internal data structures.
 575  *
 576  *   The 'pmc_hook' function pointer is also protected by this lock.
 577  *   It is only examined with the sx lock held in exclusive mode.  The
 578  *   kernel module is allowed to be unloaded only with the sx lock held
 579  *   in exclusive mode.  In normal syscall handling, after acquiring the
 580  *   pmc_sx lock we first check that 'pmc_hook' is non-null before
 581  *   proceeding.  This prevents races between the thread unloading the module
 582  *   and other threads seeking to use the module.
 583  *
 584  * - Lookups of target process structures and owner process structures
 585  *   cannot use the global "pmc_sx" SX lock because these lookups need
 586  *   to happen during context switches and in other critical sections
 587  *   where sleeping is not allowed.  We protect these lookup tables
 588  *   with their own private spin-mutexes, "pmc_processhash_mtx" and
 589  *   "pmc_ownerhash_mtx".
 590  *
 591  * - Interrupt handlers work in a lock free manner.  At interrupt
 592  *   time, handlers look at the PMC pointer (phw->phw_pmc) configured
 593  *   when the PMC was started.  If this pointer is NULL, the interrupt
 594  *   is ignored after updating driver statistics.  We ensure that this
 595  *   pointer is set (using an atomic operation if necessary) before the
 596  *   PMC hardware is started.  Conversely, this pointer is unset atomically
 597  *   only after the PMC hardware is stopped.
 598  *
 599  *   We ensure that everything needed for the operation of an
 600  *   interrupt handler is available without it needing to acquire any
 601  *   locks.  We also ensure that a PMC's software state is destroyed only
 602  *   after the PMC is taken off hardware (on all CPUs).
 603  *
 604  * - Context-switch handling with process-private PMCs needs more
 605  *   care.
 606  *
 607  *   A given process may be the target of multiple PMCs.  For example,
 608  *   PMCATTACH and PMCDETACH may be requested by a process on one CPU
 609  *   while the target process is running on another.  A PMC could also
 610  *   be getting released because its owner is exiting.  We tackle
 611  *   these situations in the following manner:
 612  *
 613  *   - each target process structure 'pmc_process' has an array
 614  *     of 'struct pmc *' pointers, one for each hardware PMC.
 615  *
 616  *   - At context switch IN time, each "target" PMC in RUNNING state
 617  *     gets started on hardware and a pointer to each PMC is copied into
 618  *     the per-cpu phw array.  The 'runcount' for the PMC is
 619  *     incremented.
 620  *
 621  *   - At context switch OUT time, all process-virtual PMCs are stopped
 622  *     on hardware.  The saved value is added to the PMCs value field
 623  *     only if the PMC is in a non-deleted state (the PMCs state could
 624  *     have changed during the current time slice).
 625  *
 626  *     Note that since in-between a switch IN on a processor and a switch
 627  *     OUT, the PMC could have been released on another CPU.  Therefore
 628  *     context switch OUT always looks at the hardware state to turn
 629  *     OFF PMCs and will update a PMC's saved value only if reachable
 630  *     from the target process record.
 631  *
 632  *   - OP PMCRELEASE could be called on a PMC at any time (the PMC could
 633  *     be attached to many processes at the time of the call and could
 634  *     be active on multiple CPUs).
 635  *
 636  *     We prevent further scheduling of the PMC by marking it as in
 637  *     state 'DELETED'.  If the runcount of the PMC is non-zero then
 638  *     this PMC is currently running on a CPU somewhere.  The thread
 639  *     doing the PMCRELEASE operation waits by repeatedly doing a
 640  *     pause() till the runcount comes to zero.
 641  *
 642  * The contents of a PMC descriptor (struct pmc) are protected using
 643  * a spin-mutex.  In order to save space, we use a mutex pool.
 644  *
 645  * In terms of lock types used by witness(4), we use:
 646  * - Type "pmc-sx", used by the global SX lock.
 647  * - Type "pmc-sleep", for sleep mutexes used by logger threads.
 648  * - Type "pmc-per-proc", for protecting PMC owner descriptors.
 649  * - Type "pmc-leaf", used for all other spin mutexes.
 650  */
 651
 652 /*
 653  * save the cpu binding of the current kthread
 654  */
 655
 656 static void
 657 pmc_save_cpu_binding(struct pmc_binding *pb)
 658 {
 659         PMCDBG0(CPU,BND,2, "save-cpu");
 660         thread_lock(curthread);
 661         pb->pb_bound = sched_is_bound(curthread);
 662         pb->pb_cpu   = curthread->td_oncpu;
 663         thread_unlock(curthread);
 664         PMCDBG1(CPU,BND,2, "save-cpu cpu=%d", pb->pb_cpu);
 665 }
 666
 667 /*
 668  * restore the cpu binding of the current thread
 669  */
 670
 671 static void
 672 pmc_restore_cpu_binding(struct pmc_binding *pb)
 673 {
 674         PMCDBG2(CPU,BND,2, "restore-cpu curcpu=%d restore=%d",
 675             curthread->td_oncpu, pb->pb_cpu);
 676         thread_lock(curthread);
 677         if (pb->pb_bound)
 678                 sched_bind(curthread, pb->pb_cpu);
 679         else
 680                 sched_unbind(curthread);
 681         thread_unlock(curthread);
 682         PMCDBG0(CPU,BND,2, "restore-cpu done");
 683 }
 684
 685 /*
 686  * move execution over the specified cpu and bind it there.
 687  */
 688
 689 static void
 690 pmc_select_cpu(int cpu)
 691 {
 692         KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
 693             ("[pmc,%d] bad cpu number %d", __LINE__, cpu));
 694
 695         /* Never move to an inactive CPU. */
 696         KASSERT(pmc_cpu_is_active(cpu), ("[pmc,%d] selecting inactive "
 697             "CPU %d", __LINE__, cpu));
 698
 699         PMCDBG1(CPU,SEL,2, "select-cpu cpu=%d", cpu);
 700         thread_lock(curthread);
 701         sched_bind(curthread, cpu);
 702         thread_unlock(curthread);
 703
 704         KASSERT(curthread->td_oncpu == cpu,
 705             ("[pmc,%d] CPU not bound [cpu=%d, curr=%d]", __LINE__,
 706                 cpu, curthread->td_oncpu));
 707
 708         PMCDBG1(CPU,SEL,2, "select-cpu cpu=%d ok", cpu);
 709 }
 710
 711 /*
 712  * Force a context switch.
 713  *
 714  * We do this by pause'ing for 1 tick -- invoking mi_switch() is not
 715  * guaranteed to force a context switch.
 716  */
 717
 718 static void
 719 pmc_force_context_switch(void)
 720 {
 721
 722         pause("pmcctx", 1);
 723 }
 724
 725 /*
 726  * Get the file name for an executable.  This is a simple wrapper
 727  * around vn_fullpath(9).
 728  */
 729
 730 static void
 731 pmc_getfilename(struct vnode *v, char **fullpath, char **freepath)
 732 {
 733
 734         *fullpath = "unknown";
 735         *freepath = NULL;
 736         vn_fullpath(curthread, v, fullpath, freepath);
 737 }
 738
 739 /*
 740  * remove an process owning PMCs
 741  */
 742
 743 void
 744 pmc_remove_owner(struct pmc_owner *po)
 745 {
 746         struct pmc *pm, *tmp;
 747
 748         sx_assert(&pmc_sx, SX_XLOCKED);
 749
 750         PMCDBG1(OWN,ORM,1, "remove-owner po=%p", po);
 751
 752         /* Remove descriptor from the owner hash table */
 753         LIST_REMOVE(po, po_next);
 754
 755         /* release all owned PMC descriptors */
 756         LIST_FOREACH_SAFE(pm, &po->po_pmcs, pm_next, tmp) {
 757                 PMCDBG1(OWN,ORM,2, "pmc=%p", pm);
 758                 KASSERT(pm->pm_owner == po,
 759                     ("[pmc,%d] owner %p != po %p", __LINE__, pm->pm_owner, po));
 760
 761                 pmc_release_pmc_descriptor(pm); /* will unlink from the list */
 762                 pmc_destroy_pmc_descriptor(pm);
 763         }
 764
 765         KASSERT(po->po_sscount == 0,
 766             ("[pmc,%d] SS count not zero", __LINE__));
 767         KASSERT(LIST_EMPTY(&po->po_pmcs),
 768             ("[pmc,%d] PMC list not empty", __LINE__));
 769
 770         /* de-configure the log file if present */
 771         if (po->po_flags & PMC_PO_OWNS_LOGFILE)
 772                 pmclog_deconfigure_log(po);
 773 }
 774
 775 /*
 776  * remove an owner process record if all conditions are met.
 777  */
 778
 779 static void
 780 pmc_maybe_remove_owner(struct pmc_owner *po)
 781 {
 782
 783         PMCDBG1(OWN,OMR,1, "maybe-remove-owner po=%p", po);
 784
 785         /*
 786          * Remove owner record if
 787          * - this process does not own any PMCs
 788          * - this process has not allocated a system-wide sampling buffer
 789          */
 790
 791         if (LIST_EMPTY(&po->po_pmcs) &&
 792             ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)) {
 793                 pmc_remove_owner(po);
 794                 pmc_destroy_owner_descriptor(po);
 795         }
 796 }
 797
 798 /*
 799  * Add an association between a target process and a PMC.
 800  */
 801
 802 static void
 803 pmc_link_target_process(struct pmc *pm, struct pmc_process *pp)
 804 {
 805         int ri;
 806         struct pmc_target *pt;
 807
 808         sx_assert(&pmc_sx, SX_XLOCKED);
 809
 810         KASSERT(pm != NULL && pp != NULL,
 811             ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp));
 812         KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)),
 813             ("[pmc,%d] Attaching a non-process-virtual pmc=%p to pid=%d",
 814                 __LINE__, pm, pp->pp_proc->p_pid));
 815         KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= ((int) md->pmd_npmc - 1),
 816             ("[pmc,%d] Illegal reference count %d for process record %p",
 817                 __LINE__, pp->pp_refcnt, (void *) pp));
 818
 819         ri = PMC_TO_ROWINDEX(pm);
 820
 821         PMCDBG3(PRC,TLK,1, "link-target pmc=%p ri=%d pmc-process=%p",
 822             pm, ri, pp);
 823
 824 #ifdef  HWPMC_DEBUG
 825         LIST_FOREACH(pt, &pm->pm_targets, pt_next)
 826             if (pt->pt_process == pp)
 827                     KASSERT(0, ("[pmc,%d] pp %p already in pmc %p targets",
 828                                 __LINE__, pp, pm));
 829 #endif
 830
 831         pt = malloc(sizeof(struct pmc_target), M_PMC, M_WAITOK|M_ZERO);
 832         pt->pt_process = pp;
 833
 834         LIST_INSERT_HEAD(&pm->pm_targets, pt, pt_next);
 835
 836         atomic_store_rel_ptr((uintptr_t *)&pp->pp_pmcs[ri].pp_pmc,
 837             (uintptr_t)pm);
 838
 839         if (pm->pm_owner->po_owner == pp->pp_proc)
 840                 pm->pm_flags |= PMC_F_ATTACHED_TO_OWNER;
 841
 842         /*
 843          * Initialize the per-process values at this row index.
 844          */
 845         pp->pp_pmcs[ri].pp_pmcval = PMC_TO_MODE(pm) == PMC_MODE_TS ?
 846             pm->pm_sc.pm_reloadcount : 0;
 847
 848         pp->pp_refcnt++;
 849
 850 }
 851
 852 /*
 853  * Removes the association between a target process and a PMC.
 854  */
 855
 856 static void
 857 pmc_unlink_target_process(struct pmc *pm, struct pmc_process *pp)
 858 {
 859         int ri;
 860         struct proc *p;
 861         struct pmc_target *ptgt;
 862
 863         sx_assert(&pmc_sx, SX_XLOCKED);
 864
 865         KASSERT(pm != NULL && pp != NULL,
 866             ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp));
 867
 868         KASSERT(pp->pp_refcnt >= 1 && pp->pp_refcnt <= (int) md->pmd_npmc,
 869             ("[pmc,%d] Illegal ref count %d on process record %p",
 870                 __LINE__, pp->pp_refcnt, (void *) pp));
 871
 872         ri = PMC_TO_ROWINDEX(pm);
 873
 874         PMCDBG3(PRC,TUL,1, "unlink-target pmc=%p ri=%d pmc-process=%p",
 875             pm, ri, pp);
 876
 877         KASSERT(pp->pp_pmcs[ri].pp_pmc == pm,
 878             ("[pmc,%d] PMC ri %d mismatch pmc %p pp->[ri] %p", __LINE__,
 879                 ri, pm, pp->pp_pmcs[ri].pp_pmc));
 880
 881         pp->pp_pmcs[ri].pp_pmc = NULL;
 882         pp->pp_pmcs[ri].pp_pmcval = (pmc_value_t) 0;
 883
 884         /* Remove owner-specific flags */
 885         if (pm->pm_owner->po_owner == pp->pp_proc) {
 886                 pp->pp_flags &= ~PMC_PP_ENABLE_MSR_ACCESS;
 887                 pm->pm_flags &= ~PMC_F_ATTACHED_TO_OWNER;
 888         }
 889
 890         pp->pp_refcnt--;
 891
 892         /* Remove the target process from the PMC structure */
 893         LIST_FOREACH(ptgt, &pm->pm_targets, pt_next)
 894                 if (ptgt->pt_process == pp)
 895                         break;
 896
 897         KASSERT(ptgt != NULL, ("[pmc,%d] process %p (pp: %p) not found "
 898                     "in pmc %p", __LINE__, pp->pp_proc, pp, pm));
 899
 900         LIST_REMOVE(ptgt, pt_next);
 901         free(ptgt, M_PMC);
 902
 903         /* if the PMC now lacks targets, send the owner a SIGIO */
 904         if (LIST_EMPTY(&pm->pm_targets)) {
 905                 p = pm->pm_owner->po_owner;
 906                 PROC_LOCK(p);
 907                 kern_psignal(p, SIGIO);
 908                 PROC_UNLOCK(p);
 909
 910                 PMCDBG2(PRC,SIG,2, "signalling proc=%p signal=%d", p,
 911                     SIGIO);
 912         }
 913 }
 914
 915 /*
 916  * Check if PMC 'pm' may be attached to target process 't'.
 917  */
 918
 919 static int
 920 pmc_can_attach(struct pmc *pm, struct proc *t)
 921 {
 922         struct proc *o;         /* pmc owner */
 923         struct ucred *oc, *tc;  /* owner, target credentials */
 924         int decline_attach, i;
 925
 926         /*
 927          * A PMC's owner can always attach that PMC to itself.
 928          */
 929
 930         if ((o = pm->pm_owner->po_owner) == t)
 931                 return 0;
 932
 933         PROC_LOCK(o);
 934         oc = o->p_ucred;
 935         crhold(oc);
 936         PROC_UNLOCK(o);
 937
 938         PROC_LOCK(t);
 939         tc = t->p_ucred;
 940         crhold(tc);
 941         PROC_UNLOCK(t);
 942
 943         /*
 944          * The effective uid of the PMC owner should match at least one
 945          * of the {effective,real,saved} uids of the target process.
 946          */
 947
 948         decline_attach = oc->cr_uid != tc->cr_uid &&
 949             oc->cr_uid != tc->cr_svuid &&
 950             oc->cr_uid != tc->cr_ruid;
 951
 952         /*
 953          * Every one of the target's group ids, must be in the owner's
 954          * group list.
 955          */
 956         for (i = 0; !decline_attach && i < tc->cr_ngroups; i++)
 957                 decline_attach = !groupmember(tc->cr_groups[i], oc);
 958
 959         /* check the read and saved gids too */
 960         if (decline_attach == 0)
 961                 decline_attach = !groupmember(tc->cr_rgid, oc) ||
 962                     !groupmember(tc->cr_svgid, oc);
 963
 964         crfree(tc);
 965         crfree(oc);
 966
 967         return !decline_attach;
 968 }
 969
 970 /*
 971  * Attach a process to a PMC.
 972  */
 973
 974 static int
 975 pmc_attach_one_process(struct proc *p, struct pmc *pm)
 976 {
 977         int ri;
 978         char *fullpath, *freepath;
 979         struct pmc_process      *pp;
 980
 981         sx_assert(&pmc_sx, SX_XLOCKED);
 982
 983         PMCDBG5(PRC,ATT,2, "attach-one pm=%p ri=%d proc=%p (%d, %s)", pm,
 984             PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm);
 985
 986         /*
 987          * Locate the process descriptor corresponding to process 'p',
 988          * allocating space as needed.
 989          *
 990          * Verify that rowindex 'pm_rowindex' is free in the process
 991          * descriptor.
 992          *
 993          * If not, allocate space for a descriptor and link the
 994          * process descriptor and PMC.
 995          */
 996         ri = PMC_TO_ROWINDEX(pm);
 997
 998         if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_ALLOCATE)) == NULL)
 999                 return ENOMEM;
1000
1001         if (pp->pp_pmcs[ri].pp_pmc == pm) /* already present at slot [ri] */
1002                 return EEXIST;
1003
1004         if (pp->pp_pmcs[ri].pp_pmc != NULL)
1005                 return EBUSY;
1006
1007         pmc_link_target_process(pm, pp);
1008
1009         if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) &&
1010             (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) == 0)
1011                 pm->pm_flags |= PMC_F_NEEDS_LOGFILE;
1012
1013         pm->pm_flags |= PMC_F_ATTACH_DONE; /* mark as attached */
1014
1015         /* issue an attach event to a configured log file */
1016         if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE) {
1017                 pmc_getfilename(p->p_textvp, &fullpath, &freepath);
1018                 if (p->p_flag & P_KTHREAD) {
1019                         fullpath = kernelname;
1020                         freepath = NULL;
1021                 } else
1022                         pmclog_process_pmcattach(pm, p->p_pid, fullpath);
1023                 if (freepath)
1024                         free(freepath, M_TEMP);
1025                 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1026                         pmc_log_process_mappings(pm->pm_owner, p);
1027         }
1028         /* mark process as using HWPMCs */
1029         PROC_LOCK(p);
1030         p->p_flag |= P_HWPMC;
1031         PROC_UNLOCK(p);
1032
1033         return 0;
1034 }
1035
1036 /*
1037  * Attach a process and optionally its children
1038  */
1039
1040 static int
1041 pmc_attach_process(struct proc *p, struct pmc *pm)
1042 {
1043         int error;
1044         struct proc *top;
1045
1046         sx_assert(&pmc_sx, SX_XLOCKED);
1047
1048         PMCDBG5(PRC,ATT,1, "attach pm=%p ri=%d proc=%p (%d, %s)", pm,
1049             PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm);
1050
1051
1052         /*
1053          * If this PMC successfully allowed a GETMSR operation
1054          * in the past, disallow further ATTACHes.
1055          */
1056
1057         if ((pm->pm_flags & PMC_PP_ENABLE_MSR_ACCESS) != 0)
1058                 return EPERM;
1059
1060         if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0)
1061                 return pmc_attach_one_process(p, pm);
1062
1063         /*
1064          * Traverse all child processes, attaching them to
1065          * this PMC.
1066          */
1067
1068         sx_slock(&proctree_lock);
1069
1070         top = p;
1071
1072         for (;;) {
1073                 if ((error = pmc_attach_one_process(p, pm)) != 0)
1074                         break;
1075                 if (!LIST_EMPTY(&p->p_children))
1076                         p = LIST_FIRST(&p->p_children);
1077                 else for (;;) {
1078                         if (p == top)
1079                                 goto done;
1080                         if (LIST_NEXT(p, p_sibling)) {
1081                                 p = LIST_NEXT(p, p_sibling);
1082                                 break;
1083                         }
1084                         p = p->p_pptr;
1085                 }
1086         }
1087
1088         if (error)
1089                 (void) pmc_detach_process(top, pm);
1090
1091  done:
1092         sx_sunlock(&proctree_lock);
1093         return error;
1094 }
1095
1096 /*
1097  * Detach a process from a PMC.  If there are no other PMCs tracking
1098  * this process, remove the process structure from its hash table.  If
1099  * 'flags' contains PMC_FLAG_REMOVE, then free the process structure.
1100  */
1101
1102 static int
1103 pmc_detach_one_process(struct proc *p, struct pmc *pm, int flags)
1104 {
1105         int ri;
1106         struct pmc_process *pp;
1107
1108         sx_assert(&pmc_sx, SX_XLOCKED);
1109
1110         KASSERT(pm != NULL,
1111             ("[pmc,%d] null pm pointer", __LINE__));
1112
1113         ri = PMC_TO_ROWINDEX(pm);
1114
1115         PMCDBG6(PRC,ATT,2, "detach-one pm=%p ri=%d proc=%p (%d, %s) flags=0x%x",
1116             pm, ri, p, p->p_pid, p->p_comm, flags);
1117
1118         if ((pp = pmc_find_process_descriptor(p, 0)) == NULL)
1119                 return ESRCH;
1120
1121         if (pp->pp_pmcs[ri].pp_pmc != pm)
1122                 return EINVAL;
1123
1124         pmc_unlink_target_process(pm, pp);
1125
1126         /* Issue a detach entry if a log file is configured */
1127         if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE)
1128                 pmclog_process_pmcdetach(pm, p->p_pid);
1129
1130         /*
1131          * If there are no PMCs targetting this process, we remove its
1132          * descriptor from the target hash table and unset the P_HWPMC
1133          * flag in the struct proc.
1134          */
1135         KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= (int) md->pmd_npmc,
1136             ("[pmc,%d] Illegal refcnt %d for process struct %p",
1137                 __LINE__, pp->pp_refcnt, pp));
1138
1139         if (pp->pp_refcnt != 0) /* still a target of some PMC */
1140                 return 0;
1141
1142         pmc_remove_process_descriptor(pp);
1143
1144         if (flags & PMC_FLAG_REMOVE)
1145                 free(pp, M_PMC);
1146
1147         PROC_LOCK(p);
1148         p->p_flag &= ~P_HWPMC;
1149         PROC_UNLOCK(p);
1150
1151         return 0;
1152 }
1153
1154 /*
1155  * Detach a process and optionally its descendants from a PMC.
1156  */
1157
1158 static int
1159 pmc_detach_process(struct proc *p, struct pmc *pm)
1160 {
1161         struct proc *top;
1162
1163         sx_assert(&pmc_sx, SX_XLOCKED);
1164
1165         PMCDBG5(PRC,ATT,1, "detach pm=%p ri=%d proc=%p (%d, %s)", pm,
1166             PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm);
1167
1168         if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0)
1169                 return pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE);
1170
1171         /*
1172          * Traverse all children, detaching them from this PMC.  We
1173          * ignore errors since we could be detaching a PMC from a
1174          * partially attached proc tree.
1175          */
1176
1177         sx_slock(&proctree_lock);
1178
1179         top = p;
1180
1181         for (;;) {
1182                 (void) pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE);
1183
1184                 if (!LIST_EMPTY(&p->p_children))
1185                         p = LIST_FIRST(&p->p_children);
1186                 else for (;;) {
1187                         if (p == top)
1188                                 goto done;
1189                         if (LIST_NEXT(p, p_sibling)) {
1190                                 p = LIST_NEXT(p, p_sibling);
1191                                 break;
1192                         }
1193                         p = p->p_pptr;
1194                 }
1195         }
1196
1197  done:
1198         sx_sunlock(&proctree_lock);
1199
1200         if (LIST_EMPTY(&pm->pm_targets))
1201                 pm->pm_flags &= ~PMC_F_ATTACH_DONE;
1202
1203         return 0;
1204 }
1205
1206
1207 /*
1208  * Thread context switch IN
1209  */
1210
1211 static void
1212 pmc_process_csw_in(struct thread *td)
1213 {
1214         int cpu;
1215         unsigned int adjri, ri;
1216         struct pmc *pm;
1217         struct proc *p;
1218         struct pmc_cpu *pc;
1219         struct pmc_hw *phw;
1220         pmc_value_t newvalue;
1221         struct pmc_process *pp;
1222         struct pmc_classdep *pcd;
1223
1224         p = td->td_proc;
1225
1226         if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE)) == NULL)
1227                 return;
1228
1229         KASSERT(pp->pp_proc == td->td_proc,
1230             ("[pmc,%d] not my thread state", __LINE__));
1231
1232         critical_enter(); /* no preemption from this point */
1233
1234         cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */
1235
1236         PMCDBG5(CSW,SWI,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p,
1237             p->p_pid, p->p_comm, pp);
1238
1239         KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
1240             ("[pmc,%d] wierd CPU id %d", __LINE__, cpu));
1241
1242         pc = pmc_pcpu[cpu];
1243
1244         for (ri = 0; ri < md->pmd_npmc; ri++) {
1245
1246                 if ((pm = pp->pp_pmcs[ri].pp_pmc) == NULL)
1247                         continue;
1248
1249                 KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)),
1250                     ("[pmc,%d] Target PMC in non-virtual mode (%d)",
1251                         __LINE__, PMC_TO_MODE(pm)));
1252
1253                 KASSERT(PMC_TO_ROWINDEX(pm) == ri,
1254                     ("[pmc,%d] Row index mismatch pmc %d != ri %d",
1255                         __LINE__, PMC_TO_ROWINDEX(pm), ri));
1256
1257                 /*
1258                  * Only PMCs that are marked as 'RUNNING' need
1259                  * be placed on hardware.
1260                  */
1261
1262                 if (pm->pm_state != PMC_STATE_RUNNING)
1263                         continue;
1264
1265                 /* increment PMC runcount */
1266                 atomic_add_rel_int(&pm->pm_runcount, 1);
1267
1268                 /* configure the HWPMC we are going to use. */
1269                 pcd = pmc_ri_to_classdep(md, ri, &adjri);
1270                 pcd->pcd_config_pmc(cpu, adjri, pm);
1271
1272                 phw = pc->pc_hwpmcs[ri];
1273
1274                 KASSERT(phw != NULL,
1275                     ("[pmc,%d] null hw pointer", __LINE__));
1276
1277                 KASSERT(phw->phw_pmc == pm,
1278                     ("[pmc,%d] hw->pmc %p != pmc %p", __LINE__,
1279                         phw->phw_pmc, pm));
1280
1281                 /*
1282                  * Write out saved value and start the PMC.
1283                  *
1284                  * Sampling PMCs use a per-process value, while
1285                  * counting mode PMCs use a per-pmc value that is
1286                  * inherited across descendants.
1287                  */
1288                 if (PMC_TO_MODE(pm) == PMC_MODE_TS) {
1289                         mtx_pool_lock_spin(pmc_mtxpool, pm);
1290
1291                         /*
1292                          * Use the saved value calculated after the most recent
1293                          * thread switch out to start this counter.  Reset
1294                          * the saved count in case another thread from this
1295                          * process switches in before any threads switch out.
1296                          */
1297                         newvalue = PMC_PCPU_SAVED(cpu,ri) =
1298                             pp->pp_pmcs[ri].pp_pmcval;
1299                         pp->pp_pmcs[ri].pp_pmcval = pm->pm_sc.pm_reloadcount;
1300                         mtx_pool_unlock_spin(pmc_mtxpool, pm);
1301                 } else {
1302                         KASSERT(PMC_TO_MODE(pm) == PMC_MODE_TC,
1303                             ("[pmc,%d] illegal mode=%d", __LINE__,
1304                             PMC_TO_MODE(pm)));
1305                         mtx_pool_lock_spin(pmc_mtxpool, pm);
1306                         newvalue = PMC_PCPU_SAVED(cpu, ri) =
1307                             pm->pm_gv.pm_savedvalue;
1308                         mtx_pool_unlock_spin(pmc_mtxpool, pm);
1309                 }
1310
1311                 PMCDBG3(CSW,SWI,1,"cpu=%d ri=%d new=%jd", cpu, ri, newvalue);
1312
1313                 pcd->pcd_write_pmc(cpu, adjri, newvalue);
1314
1315                 /* If a sampling mode PMC, reset stalled state. */
1316                 if (PMC_TO_MODE(pm) == PMC_MODE_TS)
1317                         CPU_CLR_ATOMIC(cpu, &pm->pm_stalled);
1318
1319                 /* Indicate that we desire this to run. */
1320                 CPU_SET_ATOMIC(cpu, &pm->pm_cpustate);
1321
1322                 /* Start the PMC. */
1323                 pcd->pcd_start_pmc(cpu, adjri);
1324         }
1325
1326         /*
1327          * perform any other architecture/cpu dependent thread
1328          * switch-in actions.
1329          */
1330
1331         (void) (*md->pmd_switch_in)(pc, pp);
1332
1333         critical_exit();
1334
1335 }
1336
1337 /*
1338  * Thread context switch OUT.
1339  */
1340
1341 static void
1342 pmc_process_csw_out(struct thread *td)
1343 {
1344         int cpu;
1345         int64_t tmp;
1346         struct pmc *pm;
1347         struct proc *p;
1348         enum pmc_mode mode;
1349         struct pmc_cpu *pc;
1350         pmc_value_t newvalue;
1351         unsigned int adjri, ri;
1352         struct pmc_process *pp;
1353         struct pmc_classdep *pcd;
1354
1355
1356         /*
1357          * Locate our process descriptor; this may be NULL if
1358          * this process is exiting and we have already removed
1359          * the process from the target process table.
1360          *
1361          * Note that due to kernel preemption, multiple
1362          * context switches may happen while the process is
1363          * exiting.
1364          *
1365          * Note also that if the target process cannot be
1366          * found we still need to deconfigure any PMCs that
1367          * are currently running on hardware.
1368          */
1369
1370         p = td->td_proc;
1371         pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE);
1372
1373         /*
1374          * save PMCs
1375          */
1376
1377         critical_enter();
1378
1379         cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */
1380
1381         PMCDBG5(CSW,SWO,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p,
1382             p->p_pid, p->p_comm, pp);
1383
1384         KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
1385             ("[pmc,%d wierd CPU id %d", __LINE__, cpu));
1386
1387         pc = pmc_pcpu[cpu];
1388
1389         /*
1390          * When a PMC gets unlinked from a target PMC, it will
1391          * be removed from the target's pp_pmc[] array.
1392          *
1393          * However, on a MP system, the target could have been
1394          * executing on another CPU at the time of the unlink.
1395          * So, at context switch OUT time, we need to look at
1396          * the hardware to determine if a PMC is scheduled on
1397          * it.
1398          */
1399
1400         for (ri = 0; ri < md->pmd_npmc; ri++) {
1401
1402                 pcd = pmc_ri_to_classdep(md, ri, &adjri);
1403                 pm  = NULL;
1404                 (void) (*pcd->pcd_get_config)(cpu, adjri, &pm);
1405
1406                 if (pm == NULL) /* nothing at this row index */
1407                         continue;
1408
1409                 mode = PMC_TO_MODE(pm);
1410                 if (!PMC_IS_VIRTUAL_MODE(mode))
1411                         continue; /* not a process virtual PMC */
1412
1413                 KASSERT(PMC_TO_ROWINDEX(pm) == ri,
1414                     ("[pmc,%d] ri mismatch pmc(%d) ri(%d)",
1415                         __LINE__, PMC_TO_ROWINDEX(pm), ri));
1416
1417                 /*
1418                  * Change desired state, and then stop if not stalled.
1419                  * This two-step dance should avoid race conditions where
1420                  * an interrupt re-enables the PMC after this code has
1421                  * already checked the pm_stalled flag.
1422                  */
1423                 CPU_CLR_ATOMIC(cpu, &pm->pm_cpustate);
1424                 if (!CPU_ISSET(cpu, &pm->pm_stalled))
1425                         pcd->pcd_stop_pmc(cpu, adjri);
1426
1427                 /* reduce this PMC's runcount */
1428                 atomic_subtract_rel_int(&pm->pm_runcount, 1);
1429
1430                 /*
1431                  * If this PMC is associated with this process,
1432                  * save the reading.
1433                  */
1434
1435                 if (pp != NULL && pp->pp_pmcs[ri].pp_pmc != NULL) {
1436
1437                         KASSERT(pm == pp->pp_pmcs[ri].pp_pmc,
1438                             ("[pmc,%d] pm %p != pp_pmcs[%d] %p", __LINE__,
1439                                 pm, ri, pp->pp_pmcs[ri].pp_pmc));
1440
1441                         KASSERT(pp->pp_refcnt > 0,
1442                             ("[pmc,%d] pp refcnt = %d", __LINE__,
1443                                 pp->pp_refcnt));
1444
1445                         pcd->pcd_read_pmc(cpu, adjri, &newvalue);
1446
1447                         if (mode == PMC_MODE_TS) {
1448                                 PMCDBG3(CSW,SWO,1,"cpu=%d ri=%d tmp=%jd (samp)",
1449                                     cpu, ri, PMC_PCPU_SAVED(cpu,ri) - newvalue);
1450
1451                                 /*
1452                                  * For sampling process-virtual PMCs,
1453                                  * newvalue is the number of events to be seen
1454                                  * until the next sampling interrupt.
1455                                  * We can just add the events left from this
1456                                  * invocation to the counter, then adjust
1457                                  * in case we overflow our range.
1458                                  *
1459                                  * (Recall that we reload the counter every
1460                                  * time we use it.)
1461                                  */
1462                                 mtx_pool_lock_spin(pmc_mtxpool, pm);
1463
1464                                 pp->pp_pmcs[ri].pp_pmcval += newvalue;
1465                                 if (pp->pp_pmcs[ri].pp_pmcval >
1466                                     pm->pm_sc.pm_reloadcount)
1467                                         pp->pp_pmcs[ri].pp_pmcval -=
1468                                             pm->pm_sc.pm_reloadcount;
1469                                 KASSERT(pp->pp_pmcs[ri].pp_pmcval > 0 &&
1470                                     pp->pp_pmcs[ri].pp_pmcval <=
1471                                     pm->pm_sc.pm_reloadcount,
1472                                     ("[pmc,%d] pp_pmcval outside of expected "
1473                                     "range cpu=%d ri=%d pp_pmcval=%jx "
1474                                     "pm_reloadcount=%jx", __LINE__, cpu, ri,
1475                                     pp->pp_pmcs[ri].pp_pmcval,
1476                                     pm->pm_sc.pm_reloadcount));
1477                                 mtx_pool_unlock_spin(pmc_mtxpool, pm);
1478
1479                         } else {
1480                                 tmp = newvalue - PMC_PCPU_SAVED(cpu,ri);
1481
1482                                 PMCDBG3(CSW,SWO,1,"cpu=%d ri=%d tmp=%jd (count)",
1483                                     cpu, ri, tmp);
1484
1485                                 /*
1486                                  * For counting process-virtual PMCs,
1487                                  * we expect the count to be
1488                                  * increasing monotonically, modulo a 64
1489                                  * bit wraparound.
1490                                  */
1491                                 KASSERT((int64_t) tmp >= 0,
1492                                     ("[pmc,%d] negative increment cpu=%d "
1493                                      "ri=%d newvalue=%jx saved=%jx "
1494                                      "incr=%jx", __LINE__, cpu, ri,
1495                                      newvalue, PMC_PCPU_SAVED(cpu,ri), tmp));
1496
1497                                 mtx_pool_lock_spin(pmc_mtxpool, pm);
1498                                 pm->pm_gv.pm_savedvalue += tmp;
1499                                 pp->pp_pmcs[ri].pp_pmcval += tmp;
1500                                 mtx_pool_unlock_spin(pmc_mtxpool, pm);
1501
1502                                 if (pm->pm_flags & PMC_F_LOG_PROCCSW)
1503                                         pmclog_process_proccsw(pm, pp, tmp);
1504                         }
1505                 }
1506
1507                 /* mark hardware as free */
1508                 pcd->pcd_config_pmc(cpu, adjri, NULL);
1509         }
1510
1511         /*
1512          * perform any other architecture/cpu dependent thread
1513          * switch out functions.
1514          */
1515
1516         (void) (*md->pmd_switch_out)(pc, pp);
1517
1518         critical_exit();
1519 }
1520
1521 /*
1522  * A mapping change for a process.
1523  */
1524
1525 static void
1526 pmc_process_mmap(struct thread *td, struct pmckern_map_in *pkm)
1527 {
1528         int ri;
1529         pid_t pid;
1530         char *fullpath, *freepath;
1531         const struct pmc *pm;
1532         struct pmc_owner *po;
1533         const struct pmc_process *pp;
1534
1535         freepath = fullpath = NULL;
1536         pmc_getfilename((struct vnode *) pkm->pm_file, &fullpath, &freepath);
1537
1538         pid = td->td_proc->p_pid;
1539
1540         /* Inform owners of all system-wide sampling PMCs. */
1541         LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1542             if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1543                 pmclog_process_map_in(po, pid, pkm->pm_address, fullpath);
1544
1545         if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL)
1546                 goto done;
1547
1548         /*
1549          * Inform sampling PMC owners tracking this process.
1550          */
1551         for (ri = 0; ri < md->pmd_npmc; ri++)
1552                 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL &&
1553                     PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1554                         pmclog_process_map_in(pm->pm_owner,
1555                             pid, pkm->pm_address, fullpath);
1556
1557   done:
1558         if (freepath)
1559                 free(freepath, M_TEMP);
1560 }
1561
1562
1563 /*
1564  * Log an munmap request.
1565  */
1566
1567 static void
1568 pmc_process_munmap(struct thread *td, struct pmckern_map_out *pkm)
1569 {
1570         int ri;
1571         pid_t pid;
1572         struct pmc_owner *po;
1573         const struct pmc *pm;
1574         const struct pmc_process *pp;
1575
1576         pid = td->td_proc->p_pid;
1577
1578         LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1579             if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1580                 pmclog_process_map_out(po, pid, pkm->pm_address,
1581                     pkm->pm_address + pkm->pm_size);
1582
1583         if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL)
1584                 return;
1585
1586         for (ri = 0; ri < md->pmd_npmc; ri++)
1587                 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL &&
1588                     PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1589                         pmclog_process_map_out(pm->pm_owner, pid,
1590                             pkm->pm_address, pkm->pm_address + pkm->pm_size);
1591 }
1592
1593 /*
1594  * Log mapping information about the kernel.
1595  */
1596
1597 static void
1598 pmc_log_kernel_mappings(struct pmc *pm)
1599 {
1600         struct pmc_owner *po;
1601         struct pmckern_map_in *km, *kmbase;
1602
1603         sx_assert(&pmc_sx, SX_LOCKED);
1604         KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)),
1605             ("[pmc,%d] non-sampling PMC (%p) desires mapping information",
1606                 __LINE__, (void *) pm));
1607
1608         po = pm->pm_owner;
1609
1610         if (po->po_flags & PMC_PO_INITIAL_MAPPINGS_DONE)
1611                 return;
1612
1613         /*
1614          * Log the current set of kernel modules.
1615          */
1616         kmbase = linker_hwpmc_list_objects();
1617         for (km = kmbase; km->pm_file != NULL; km++) {
1618                 PMCDBG2(LOG,REG,1,"%s %p", (char *) km->pm_file,
1619                     (void *) km->pm_address);
1620                 pmclog_process_map_in(po, (pid_t) -1, km->pm_address,
1621                     km->pm_file);
1622         }
1623         free(kmbase, M_LINKER);
1624
1625         po->po_flags |= PMC_PO_INITIAL_MAPPINGS_DONE;
1626 }
1627
1628 /*
1629  * Log the mappings for a single process.
1630  */
1631
1632 static void
1633 pmc_log_process_mappings(struct pmc_owner *po, struct proc *p)
1634 {
1635         vm_map_t map;
1636         struct vnode *vp;
1637         struct vmspace *vm;
1638         vm_map_entry_t entry;
1639         vm_offset_t last_end;
1640         u_int last_timestamp;
1641         struct vnode *last_vp;
1642         vm_offset_t start_addr;
1643         vm_object_t obj, lobj, tobj;
1644         char *fullpath, *freepath;
1645
1646         last_vp = NULL;
1647         last_end = (vm_offset_t) 0;
1648         fullpath = freepath = NULL;
1649
1650         if ((vm = vmspace_acquire_ref(p)) == NULL)
1651                 return;
1652
1653         map = &vm->vm_map;
1654         vm_map_lock_read(map);
1655
1656         for (entry = map->header.next; entry != &map->header; entry = entry->next) {
1657
1658                 if (entry == NULL) {
1659                         PMCDBG2(LOG,OPS,2, "hwpmc: vm_map entry unexpectedly "
1660                             "NULL! pid=%d vm_map=%p\n", p->p_pid, map);
1661                         break;
1662                 }
1663
1664                 /*
1665                  * We only care about executable map entries.
1666                  */
1667                 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) ||
1668                     !(entry->protection & VM_PROT_EXECUTE) ||
1669                     (entry->object.vm_object == NULL)) {
1670                         continue;
1671                 }
1672
1673                 obj = entry->object.vm_object;
1674                 VM_OBJECT_RLOCK(obj);
1675
1676                 /*
1677                  * Walk the backing_object list to find the base
1678                  * (non-shadowed) vm_object.
1679                  */
1680                 for (lobj = tobj = obj; tobj != NULL; tobj = tobj->backing_object) {
1681                         if (tobj != obj)
1682                                 VM_OBJECT_RLOCK(tobj);
1683                         if (lobj != obj)
1684                                 VM_OBJECT_RUNLOCK(lobj);
1685                         lobj = tobj;
1686                 }
1687
1688                 /*
1689                  * At this point lobj is the base vm_object and it is locked.
1690                  */
1691                 if (lobj == NULL) {
1692                         PMCDBG3(LOG,OPS,2, "hwpmc: lobj unexpectedly NULL! pid=%d "
1693                             "vm_map=%p vm_obj=%p\n", p->p_pid, map, obj);
1694                         VM_OBJECT_RUNLOCK(obj);
1695                         continue;
1696                 }
1697
1698                 vp = vm_object_vnode(lobj);
1699                 if (vp == NULL) {
1700                         if (lobj != obj)
1701                                 VM_OBJECT_RUNLOCK(lobj);
1702                         VM_OBJECT_RUNLOCK(obj);
1703                         continue;
1704                 }
1705
1706                 /*
1707                  * Skip contiguous regions that point to the same
1708                  * vnode, so we don't emit redundant MAP-IN
1709                  * directives.
1710                  */
1711                 if (entry->start == last_end && vp == last_vp) {
1712                         last_end = entry->end;
1713                         if (lobj != obj)
1714                                 VM_OBJECT_RUNLOCK(lobj);
1715                         VM_OBJECT_RUNLOCK(obj);
1716                         continue;
1717                 }
1718
1719                 /*
1720                  * We don't want to keep the proc's vm_map or this
1721                  * vm_object locked while we walk the pathname, since
1722                  * vn_fullpath() can sleep.  However, if we drop the
1723                  * lock, it's possible for concurrent activity to
1724                  * modify the vm_map list.  To protect against this,
1725                  * we save the vm_map timestamp before we release the
1726                  * lock, and check it after we reacquire the lock
1727                  * below.
1728                  */
1729                 start_addr = entry->start;
1730                 last_end = entry->end;
1731                 last_timestamp = map->timestamp;
1732                 vm_map_unlock_read(map);
1733
1734                 vref(vp);
1735                 if (lobj != obj)
1736                         VM_OBJECT_RUNLOCK(lobj);
1737
1738                 VM_OBJECT_RUNLOCK(obj);
1739
1740                 freepath = NULL;
1741                 pmc_getfilename(vp, &fullpath, &freepath);
1742                 last_vp = vp;
1743
1744                 vrele(vp);
1745
1746                 vp = NULL;
1747                 pmclog_process_map_in(po, p->p_pid, start_addr, fullpath);
1748                 if (freepath)
1749                         free(freepath, M_TEMP);
1750
1751                 vm_map_lock_read(map);
1752
1753                 /*
1754                  * If our saved timestamp doesn't match, this means
1755                  * that the vm_map was modified out from under us and
1756                  * we can't trust our current "entry" pointer.  Do a
1757                  * new lookup for this entry.  If there is no entry
1758                  * for this address range, vm_map_lookup_entry() will
1759                  * return the previous one, so we always want to go to
1760                  * entry->next on the next loop iteration.
1761                  *
1762                  * There is an edge condition here that can occur if
1763                  * there is no entry at or before this address.  In
1764                  * this situation, vm_map_lookup_entry returns
1765                  * &map->header, which would cause our loop to abort
1766                  * without processing the rest of the map.  However,
1767                  * in practice this will never happen for process
1768                  * vm_map.  This is because the executable's text
1769                  * segment is the first mapping in the proc's address
1770                  * space, and this mapping is never removed until the
1771                  * process exits, so there will always be a non-header
1772                  * entry at or before the requested address for
1773                  * vm_map_lookup_entry to return.
1774                  */
1775                 if (map->timestamp != last_timestamp)
1776                         vm_map_lookup_entry(map, last_end - 1, &entry);
1777         }
1778
1779         vm_map_unlock_read(map);
1780         vmspace_free(vm);
1781         return;
1782 }
1783
1784 /*
1785  * Log mappings for all processes in the system.
1786  */
1787
1788 static void
1789 pmc_log_all_process_mappings(struct pmc_owner *po)
1790 {
1791         struct proc *p, *top;
1792
1793         sx_assert(&pmc_sx, SX_XLOCKED);
1794
1795         if ((p = pfind(1)) == NULL)
1796                 panic("[pmc,%d] Cannot find init", __LINE__);
1797
1798         PROC_UNLOCK(p);
1799
1800         sx_slock(&proctree_lock);
1801
1802         top = p;
1803
1804         for (;;) {
1805                 pmc_log_process_mappings(po, p);
1806                 if (!LIST_EMPTY(&p->p_children))
1807                         p = LIST_FIRST(&p->p_children);
1808                 else for (;;) {
1809                         if (p == top)
1810                                 goto done;
1811                         if (LIST_NEXT(p, p_sibling)) {
1812                                 p = LIST_NEXT(p, p_sibling);
1813                                 break;
1814                         }
1815                         p = p->p_pptr;
1816                 }
1817         }
1818  done:
1819         sx_sunlock(&proctree_lock);
1820 }
1821
1822 /*
1823  * The 'hook' invoked from the kernel proper
1824  */
1825
1826
1827 #ifdef  HWPMC_DEBUG
1828 const char *pmc_hooknames[] = {
1829         /* these strings correspond to PMC_FN_* in <sys/pmckern.h> */
1830         "",
1831         "EXEC",
1832         "CSW-IN",
1833         "CSW-OUT",
1834         "SAMPLE",
1835         "UNUSED1",
1836         "UNUSED2",
1837         "MMAP",
1838         "MUNMAP",
1839         "CALLCHAIN-NMI",
1840         "CALLCHAIN-SOFT",
1841         "SOFTSAMPLING"
1842 };
1843 #endif
1844
1845 static int
1846 pmc_hook_handler(struct thread *td, int function, void *arg)
1847 {
1848
1849         PMCDBG4(MOD,PMH,1, "hook td=%p func=%d \"%s\" arg=%p", td, function,
1850             pmc_hooknames[function], arg);
1851
1852         switch (function)
1853         {
1854
1855         /*
1856          * Process exec()
1857          */
1858
1859         case PMC_FN_PROCESS_EXEC:
1860         {
1861                 char *fullpath, *freepath;
1862                 unsigned int ri;
1863                 int is_using_hwpmcs;
1864                 struct pmc *pm;
1865                 struct proc *p;
1866                 struct pmc_owner *po;
1867                 struct pmc_process *pp;
1868                 struct pmckern_procexec *pk;
1869
1870                 sx_assert(&pmc_sx, SX_XLOCKED);
1871
1872                 p = td->td_proc;
1873                 pmc_getfilename(p->p_textvp, &fullpath, &freepath);
1874
1875                 pk = (struct pmckern_procexec *) arg;
1876
1877                 /* Inform owners of SS mode PMCs of the exec event. */
1878                 LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1879                     if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1880                             pmclog_process_procexec(po, PMC_ID_INVALID,
1881                                 p->p_pid, pk->pm_entryaddr, fullpath);
1882
1883                 PROC_LOCK(p);
1884                 is_using_hwpmcs = p->p_flag & P_HWPMC;
1885                 PROC_UNLOCK(p);
1886
1887                 if (!is_using_hwpmcs) {
1888                         if (freepath)
1889                                 free(freepath, M_TEMP);
1890                         break;
1891                 }
1892
1893                 /*
1894                  * PMCs are not inherited across an exec():  remove any
1895                  * PMCs that this process is the owner of.
1896                  */
1897
1898                 if ((po = pmc_find_owner_descriptor(p)) != NULL) {
1899                         pmc_remove_owner(po);
1900                         pmc_destroy_owner_descriptor(po);
1901                 }
1902
1903                 /*
1904                  * If the process being exec'ed is not the target of any
1905                  * PMC, we are done.
1906                  */
1907                 if ((pp = pmc_find_process_descriptor(p, 0)) == NULL) {
1908                         if (freepath)
1909                                 free(freepath, M_TEMP);
1910                         break;
1911                 }
1912
1913                 /*
1914                  * Log the exec event to all monitoring owners.  Skip
1915                  * owners who have already recieved the event because
1916                  * they had system sampling PMCs active.
1917                  */
1918                 for (ri = 0; ri < md->pmd_npmc; ri++)
1919                         if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) {
1920                                 po = pm->pm_owner;
1921                                 if (po->po_sscount == 0 &&
1922                                     po->po_flags & PMC_PO_OWNS_LOGFILE)
1923                                         pmclog_process_procexec(po, pm->pm_id,
1924                                             p->p_pid, pk->pm_entryaddr,
1925                                             fullpath);
1926                         }
1927
1928                 if (freepath)
1929                         free(freepath, M_TEMP);
1930
1931
1932                 PMCDBG4(PRC,EXC,1, "exec proc=%p (%d, %s) cred-changed=%d",
1933                     p, p->p_pid, p->p_comm, pk->pm_credentialschanged);
1934
1935                 if (pk->pm_credentialschanged == 0) /* no change */
1936                         break;
1937
1938                 /*
1939                  * If the newly exec()'ed process has a different credential
1940                  * than before, allow it to be the target of a PMC only if
1941                  * the PMC's owner has sufficient priviledge.
1942                  */
1943
1944                 for (ri = 0; ri < md->pmd_npmc; ri++)
1945                         if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL)
1946                                 if (pmc_can_attach(pm, td->td_proc) != 0)
1947                                         pmc_detach_one_process(td->td_proc,
1948                                             pm, PMC_FLAG_NONE);
1949
1950                 KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= (int) md->pmd_npmc,
1951                     ("[pmc,%d] Illegal ref count %d on pp %p", __LINE__,
1952                         pp->pp_refcnt, pp));
1953
1954                 /*
1955                  * If this process is no longer the target of any
1956                  * PMCs, we can remove the process entry and free
1957                  * up space.
1958                  */
1959
1960                 if (pp->pp_refcnt == 0) {
1961                         pmc_remove_process_descriptor(pp);
1962                         free(pp, M_PMC);
1963                         break;
1964                 }
1965
1966         }
1967         break;
1968
1969         case PMC_FN_CSW_IN:
1970                 pmc_process_csw_in(td);
1971                 break;
1972
1973         case PMC_FN_CSW_OUT:
1974                 pmc_process_csw_out(td);
1975                 break;
1976
1977         /*
1978          * Process accumulated PC samples.
1979          *
1980          * This function is expected to be called by hardclock() for
1981          * each CPU that has accumulated PC samples.
1982          *
1983          * This function is to be executed on the CPU whose samples
1984          * are being processed.
1985          */
1986         case PMC_FN_DO_SAMPLES:
1987
1988                 /*
1989                  * Clear the cpu specific bit in the CPU mask before
1990                  * do the rest of the processing.  If the NMI handler
1991                  * gets invoked after the "atomic_clear_int()" call
1992                  * below but before "pmc_process_samples()" gets
1993                  * around to processing the interrupt, then we will
1994                  * come back here at the next hardclock() tick (and
1995                  * may find nothing to do if "pmc_process_samples()"
1996                  * had already processed the interrupt).  We don't
1997                  * lose the interrupt sample.
1998                  */
1999                 CPU_CLR_ATOMIC(PCPU_GET(cpuid), &pmc_cpumask);
2000                 pmc_process_samples(PCPU_GET(cpuid), PMC_HR);
2001                 pmc_process_samples(PCPU_GET(cpuid), PMC_SR);
2002                 break;
2003
2004         case PMC_FN_MMAP:
2005                 sx_assert(&pmc_sx, SX_LOCKED);
2006                 pmc_process_mmap(td, (struct pmckern_map_in *) arg);
2007                 break;
2008
2009         case PMC_FN_MUNMAP:
2010                 sx_assert(&pmc_sx, SX_LOCKED);
2011                 pmc_process_munmap(td, (struct pmckern_map_out *) arg);
2012                 break;
2013
2014         case PMC_FN_USER_CALLCHAIN:
2015                 /*
2016                  * Record a call chain.
2017                  */
2018                 KASSERT(td == curthread, ("[pmc,%d] td != curthread",
2019                     __LINE__));
2020
2021                 pmc_capture_user_callchain(PCPU_GET(cpuid), PMC_HR,
2022                     (struct trapframe *) arg);
2023                 td->td_pflags &= ~TDP_CALLCHAIN;
2024                 break;
2025
2026         case PMC_FN_USER_CALLCHAIN_SOFT:
2027                 /*
2028                  * Record a call chain.
2029                  */
2030                 KASSERT(td == curthread, ("[pmc,%d] td != curthread",
2031                     __LINE__));
2032                 pmc_capture_user_callchain(PCPU_GET(cpuid), PMC_SR,
2033                     (struct trapframe *) arg);
2034                 td->td_pflags &= ~TDP_CALLCHAIN;
2035                 break;
2036
2037         case PMC_FN_SOFT_SAMPLING:
2038                 /*
2039                  * Call soft PMC sampling intr.
2040                  */
2041                 pmc_soft_intr((struct pmckern_soft *) arg);
2042                 break;
2043
2044         default:
2045 #ifdef  HWPMC_DEBUG
2046                 KASSERT(0, ("[pmc,%d] unknown hook %d\n", __LINE__, function));
2047 #endif
2048                 break;
2049
2050         }
2051
2052         return 0;
2053 }
2054
2055 /*
2056  * allocate a 'struct pmc_owner' descriptor in the owner hash table.
2057  */
2058
2059 static struct pmc_owner *
2060 pmc_allocate_owner_descriptor(struct proc *p)
2061 {
2062         uint32_t hindex;
2063         struct pmc_owner *po;
2064         struct pmc_ownerhash *poh;
2065
2066         hindex = PMC_HASH_PTR(p, pmc_ownerhashmask);
2067         poh = &pmc_ownerhash[hindex];
2068
2069         /* allocate space for N pointers and one descriptor struct */
2070         po = malloc(sizeof(struct pmc_owner), M_PMC, M_WAITOK|M_ZERO);
2071         po->po_owner = p;
2072         LIST_INSERT_HEAD(poh, po, po_next); /* insert into hash table */
2073
2074         TAILQ_INIT(&po->po_logbuffers);
2075         mtx_init(&po->po_mtx, "pmc-owner-mtx", "pmc-per-proc", MTX_SPIN);
2076
2077         PMCDBG4(OWN,ALL,1, "allocate-owner proc=%p (%d, %s) pmc-owner=%p",
2078             p, p->p_pid, p->p_comm, po);
2079
2080         return po;
2081 }
2082
2083 static void
2084 pmc_destroy_owner_descriptor(struct pmc_owner *po)
2085 {
2086
2087         PMCDBG4(OWN,REL,1, "destroy-owner po=%p proc=%p (%d, %s)",
2088             po, po->po_owner, po->po_owner->p_pid, po->po_owner->p_comm);
2089
2090         mtx_destroy(&po->po_mtx);
2091         free(po, M_PMC);
2092 }
2093
2094 /*
2095  * find the descriptor corresponding to process 'p', adding or removing it
2096  * as specified by 'mode'.
2097  */
2098
2099 static struct pmc_process *
2100 pmc_find_process_descriptor(struct proc *p, uint32_t mode)
2101 {
2102         uint32_t hindex;
2103         struct pmc_process *pp, *ppnew;
2104         struct pmc_processhash *pph;
2105
2106         hindex = PMC_HASH_PTR(p, pmc_processhashmask);
2107         pph = &pmc_processhash[hindex];
2108
2109         ppnew = NULL;
2110
2111         /*
2112          * Pre-allocate memory in the FIND_ALLOCATE case since we
2113          * cannot call malloc(9) once we hold a spin lock.
2114          */
2115         if (mode & PMC_FLAG_ALLOCATE)
2116                 ppnew = malloc(sizeof(struct pmc_process) + md->pmd_npmc *
2117                     sizeof(struct pmc_targetstate), M_PMC, M_WAITOK|M_ZERO);
2118
2119         mtx_lock_spin(&pmc_processhash_mtx);
2120         LIST_FOREACH(pp, pph, pp_next)
2121             if (pp->pp_proc == p)
2122                     break;
2123
2124         if ((mode & PMC_FLAG_REMOVE) && pp != NULL)
2125                 LIST_REMOVE(pp, pp_next);
2126
2127         if ((mode & PMC_FLAG_ALLOCATE) && pp == NULL &&
2128             ppnew != NULL) {
2129                 ppnew->pp_proc = p;
2130                 LIST_INSERT_HEAD(pph, ppnew, pp_next);
2131                 pp = ppnew;
2132                 ppnew = NULL;
2133         }
2134         mtx_unlock_spin(&pmc_processhash_mtx);
2135
2136         if (pp != NULL && ppnew != NULL)
2137                 free(ppnew, M_PMC);
2138
2139         return pp;
2140 }
2141
2142 /*
2143  * remove a process descriptor from the process hash table.
2144  */
2145
2146 static void
2147 pmc_remove_process_descriptor(struct pmc_process *pp)
2148 {
2149         KASSERT(pp->pp_refcnt == 0,
2150             ("[pmc,%d] Removing process descriptor %p with count %d",
2151                 __LINE__, pp, pp->pp_refcnt));
2152
2153         mtx_lock_spin(&pmc_processhash_mtx);
2154         LIST_REMOVE(pp, pp_next);
2155         mtx_unlock_spin(&pmc_processhash_mtx);
2156 }
2157
2158
2159 /*
2160  * find an owner descriptor corresponding to proc 'p'
2161  */
2162
2163 static struct pmc_owner *
2164 pmc_find_owner_descriptor(struct proc *p)
2165 {
2166         uint32_t hindex;
2167         struct pmc_owner *po;
2168         struct pmc_ownerhash *poh;
2169
2170         hindex = PMC_HASH_PTR(p, pmc_ownerhashmask);
2171         poh = &pmc_ownerhash[hindex];
2172
2173         po = NULL;
2174         LIST_FOREACH(po, poh, po_next)
2175             if (po->po_owner == p)
2176                     break;
2177
2178         PMCDBG5(OWN,FND,1, "find-owner proc=%p (%d, %s) hindex=0x%x -> "
2179             "pmc-owner=%p", p, p->p_pid, p->p_comm, hindex, po);
2180
2181         return po;
2182 }
2183
2184 /*
2185  * pmc_allocate_pmc_descriptor
2186  *
2187  * Allocate a pmc descriptor and initialize its
2188  * fields.
2189  */
2190
2191 static struct pmc *
2192 pmc_allocate_pmc_descriptor(void)
2193 {
2194         struct pmc *pmc;
2195
2196         pmc = malloc(sizeof(struct pmc), M_PMC, M_WAITOK|M_ZERO);
2197
2198         PMCDBG1(PMC,ALL,1, "allocate-pmc -> pmc=%p", pmc);
2199
2200         return pmc;
2201 }
2202
2203 /*
2204  * Destroy a pmc descriptor.
2205  */
2206
2207 static void
2208 pmc_destroy_pmc_descriptor(struct pmc *pm)
2209 {
2210
2211         KASSERT(pm->pm_state == PMC_STATE_DELETED ||
2212             pm->pm_state == PMC_STATE_FREE,
2213             ("[pmc,%d] destroying non-deleted PMC", __LINE__));
2214         KASSERT(LIST_EMPTY(&pm->pm_targets),
2215             ("[pmc,%d] destroying pmc with targets", __LINE__));
2216         KASSERT(pm->pm_owner == NULL,
2217             ("[pmc,%d] destroying pmc attached to an owner", __LINE__));
2218         KASSERT(pm->pm_runcount == 0,
2219             ("[pmc,%d] pmc has non-zero run count %d", __LINE__,
2220                 pm->pm_runcount));
2221
2222         free(pm, M_PMC);
2223 }
2224
2225 static void
2226 pmc_wait_for_pmc_idle(struct pmc *pm)
2227 {
2228 #ifdef HWPMC_DEBUG
2229         volatile int maxloop;
2230
2231         maxloop = 100 * pmc_cpu_max();
2232 #endif
2233         /*
2234          * Loop (with a forced context switch) till the PMC's runcount
2235          * comes down to zero.
2236          */
2237         while (atomic_load_acq_32(&pm->pm_runcount) > 0) {
2238 #ifdef HWPMC_DEBUG
2239                 maxloop--;
2240                 KASSERT(maxloop > 0,
2241                     ("[pmc,%d] (ri%d, rc%d) waiting too long for "
2242                         "pmc to be free", __LINE__,
2243                         PMC_TO_ROWINDEX(pm), pm->pm_runcount));
2244 #endif
2245                 pmc_force_context_switch();
2246         }
2247 }
2248
2249 /*
2250  * This function does the following things:
2251  *
2252  *  - detaches the PMC from hardware
2253  *  - unlinks all target threads that were attached to it
2254  *  - removes the PMC from its owner's list
2255  *  - destroys the PMC private mutex
2256  *
2257  * Once this function completes, the given pmc pointer can be freed by
2258  * calling pmc_destroy_pmc_descriptor().
2259  */
2260
2261 static void
2262 pmc_release_pmc_descriptor(struct pmc *pm)
2263 {
2264         enum pmc_mode mode;
2265         struct pmc_hw *phw;
2266         u_int adjri, ri, cpu;
2267         struct pmc_owner *po;
2268         struct pmc_binding pb;
2269         struct pmc_process *pp;
2270         struct pmc_classdep *pcd;
2271         struct pmc_target *ptgt, *tmp;
2272
2273         sx_assert(&pmc_sx, SX_XLOCKED);
2274
2275         KASSERT(pm, ("[pmc,%d] null pmc", __LINE__));
2276
2277         ri   = PMC_TO_ROWINDEX(pm);
2278         pcd  = pmc_ri_to_classdep(md, ri, &adjri);
2279         mode = PMC_TO_MODE(pm);
2280
2281         PMCDBG3(PMC,REL,1, "release-pmc pmc=%p ri=%d mode=%d", pm, ri,
2282             mode);
2283
2284         /*
2285          * First, we take the PMC off hardware.
2286          */
2287         cpu = 0;
2288         if (PMC_IS_SYSTEM_MODE(mode)) {
2289
2290                 /*
2291                  * A system mode PMC runs on a specific CPU.  Switch
2292                  * to this CPU and turn hardware off.
2293                  */
2294                 pmc_save_cpu_binding(&pb);
2295
2296                 cpu = PMC_TO_CPU(pm);
2297
2298                 pmc_select_cpu(cpu);
2299
2300                 /* switch off non-stalled CPUs */
2301                 CPU_CLR_ATOMIC(cpu, &pm->pm_cpustate);
2302                 if (pm->pm_state == PMC_STATE_RUNNING &&
2303                     !CPU_ISSET(cpu, &pm->pm_stalled)) {
2304
2305                         phw = pmc_pcpu[cpu]->pc_hwpmcs[ri];
2306
2307                         KASSERT(phw->phw_pmc == pm,
2308                             ("[pmc, %d] pmc ptr ri(%d) hw(%p) pm(%p)",
2309                                 __LINE__, ri, phw->phw_pmc, pm));
2310                         PMCDBG2(PMC,REL,2, "stopping cpu=%d ri=%d", cpu, ri);
2311
2312                         critical_enter();
2313                         pcd->pcd_stop_pmc(cpu, adjri);
2314                         critical_exit();
2315                 }
2316
2317                 PMCDBG2(PMC,REL,2, "decfg cpu=%d ri=%d", cpu, ri);
2318
2319                 critical_enter();
2320                 pcd->pcd_config_pmc(cpu, adjri, NULL);
2321                 critical_exit();
2322
2323                 /* adjust the global and process count of SS mode PMCs */
2324                 if (mode == PMC_MODE_SS && pm->pm_state == PMC_STATE_RUNNING) {
2325                         po = pm->pm_owner;
2326                         po->po_sscount--;
2327                         if (po->po_sscount == 0) {
2328                                 atomic_subtract_rel_int(&pmc_ss_count, 1);
2329                                 LIST_REMOVE(po, po_ssnext);
2330                         }
2331                 }
2332
2333                 pm->pm_state = PMC_STATE_DELETED;
2334
2335                 pmc_restore_cpu_binding(&pb);
2336
2337                 /*
2338                  * We could have references to this PMC structure in
2339                  * the per-cpu sample queues.  Wait for the queue to
2340                  * drain.
2341                  */
2342                 pmc_wait_for_pmc_idle(pm);
2343
2344         } else if (PMC_IS_VIRTUAL_MODE(mode)) {
2345
2346                 /*
2347                  * A virtual PMC could be running on multiple CPUs at
2348                  * a given instant.
2349                  *
2350                  * By marking its state as DELETED, we ensure that
2351                  * this PMC is never further scheduled on hardware.
2352                  *
2353                  * Then we wait till all CPUs are done with this PMC.
2354                  */
2355                 pm->pm_state = PMC_STATE_DELETED;
2356
2357
2358                 /* Wait for the PMCs runcount to come to zero. */
2359                 pmc_wait_for_pmc_idle(pm);
2360
2361                 /*
2362                  * At this point the PMC is off all CPUs and cannot be
2363                  * freshly scheduled onto a CPU.  It is now safe to
2364                  * unlink all targets from this PMC.  If a
2365                  * process-record's refcount falls to zero, we remove
2366                  * it from the hash table.  The module-wide SX lock
2367                  * protects us from races.
2368                  */
2369                 LIST_FOREACH_SAFE(ptgt, &pm->pm_targets, pt_next, tmp) {
2370                         pp = ptgt->pt_process;
2371                         pmc_unlink_target_process(pm, pp); /* frees 'ptgt' */
2372
2373                         PMCDBG1(PMC,REL,3, "pp->refcnt=%d", pp->pp_refcnt);
2374
2375                         /*
2376                          * If the target process record shows that no
2377                          * PMCs are attached to it, reclaim its space.
2378                          */
2379
2380                         if (pp->pp_refcnt == 0) {
2381                                 pmc_remove_process_descriptor(pp);
2382                                 free(pp, M_PMC);
2383                         }
2384                 }
2385
2386                 cpu = curthread->td_oncpu; /* setup cpu for pmd_release() */
2387
2388         }
2389
2390         /*
2391          * Release any MD resources
2392          */
2393         (void) pcd->pcd_release_pmc(cpu, adjri, pm);
2394
2395         /*
2396          * Update row disposition
2397          */
2398
2399         if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm)))
2400                 PMC_UNMARK_ROW_STANDALONE(ri);
2401         else
2402                 PMC_UNMARK_ROW_THREAD(ri);
2403
2404         /* unlink from the owner's list */
2405         if (pm->pm_owner) {
2406                 LIST_REMOVE(pm, pm_next);
2407                 pm->pm_owner = NULL;
2408         }
2409 }
2410
2411 /*
2412  * Register an owner and a pmc.
2413  */
2414
2415 static int
2416 pmc_register_owner(struct proc *p, struct pmc *pmc)
2417 {
2418         struct pmc_owner *po;
2419
2420         sx_assert(&pmc_sx, SX_XLOCKED);
2421
2422         if ((po = pmc_find_owner_descriptor(p)) == NULL)
2423                 if ((po = pmc_allocate_owner_descriptor(p)) == NULL)
2424                         return ENOMEM;
2425
2426         KASSERT(pmc->pm_owner == NULL,
2427             ("[pmc,%d] attempting to own an initialized PMC", __LINE__));
2428         pmc->pm_owner  = po;
2429
2430         LIST_INSERT_HEAD(&po->po_pmcs, pmc, pm_next);
2431
2432         PROC_LOCK(p);
2433         p->p_flag |= P_HWPMC;
2434         PROC_UNLOCK(p);
2435
2436         if (po->po_flags & PMC_PO_OWNS_LOGFILE)
2437                 pmclog_process_pmcallocate(pmc);
2438
2439         PMCDBG2(PMC,REG,1, "register-owner pmc-owner=%p pmc=%p",
2440             po, pmc);
2441
2442         return 0;
2443 }
2444
2445 /*
2446  * Return the current row disposition:
2447  * == 0 => FREE
2448  *  > 0 => PROCESS MODE
2449  *  < 0 => SYSTEM MODE
2450  */
2451
2452 int
2453 pmc_getrowdisp(int ri)
2454 {
2455         return pmc_pmcdisp[ri];
2456 }
2457
2458 /*
2459  * Check if a PMC at row index 'ri' can be allocated to the current
2460  * process.
2461  *
2462  * Allocation can fail if:
2463  *   - the current process is already being profiled by a PMC at index 'ri',
2464  *     attached to it via OP_PMCATTACH.
2465  *   - the current process has already allocated a PMC at index 'ri'
2466  *     via OP_ALLOCATE.
2467  */
2468
2469 static int
2470 pmc_can_allocate_rowindex(struct proc *p, unsigned int ri, int cpu)
2471 {
2472         enum pmc_mode mode;
2473         struct pmc *pm;
2474         struct pmc_owner *po;
2475         struct pmc_process *pp;
2476
2477         PMCDBG5(PMC,ALR,1, "can-allocate-rowindex proc=%p (%d, %s) ri=%d "
2478             "cpu=%d", p, p->p_pid, p->p_comm, ri, cpu);
2479
2480         /*
2481          * We shouldn't have already allocated a process-mode PMC at
2482          * row index 'ri'.
2483          *
2484          * We shouldn't have allocated a system-wide PMC on the same
2485          * CPU and same RI.
2486          */
2487         if ((po = pmc_find_owner_descriptor(p)) != NULL)
2488                 LIST_FOREACH(pm, &po->po_pmcs, pm_next) {
2489                     if (PMC_TO_ROWINDEX(pm) == ri) {
2490                             mode = PMC_TO_MODE(pm);
2491                             if (PMC_IS_VIRTUAL_MODE(mode))
2492                                     return EEXIST;
2493                             if (PMC_IS_SYSTEM_MODE(mode) &&
2494                                 (int) PMC_TO_CPU(pm) == cpu)
2495                                     return EEXIST;
2496                     }
2497                 }
2498
2499         /*
2500          * We also shouldn't be the target of any PMC at this index
2501          * since otherwise a PMC_ATTACH to ourselves will fail.
2502          */
2503         if ((pp = pmc_find_process_descriptor(p, 0)) != NULL)
2504                 if (pp->pp_pmcs[ri].pp_pmc)
2505                         return EEXIST;
2506
2507         PMCDBG4(PMC,ALR,2, "can-allocate-rowindex proc=%p (%d, %s) ri=%d ok",
2508             p, p->p_pid, p->p_comm, ri);
2509
2510         return 0;
2511 }
2512
2513 /*
2514  * Check if a given PMC at row index 'ri' can be currently used in
2515  * mode 'mode'.
2516  */
2517
2518 static int
2519 pmc_can_allocate_row(int ri, enum pmc_mode mode)
2520 {
2521         enum pmc_disp   disp;
2522
2523         sx_assert(&pmc_sx, SX_XLOCKED);
2524
2525         PMCDBG2(PMC,ALR,1, "can-allocate-row ri=%d mode=%d", ri, mode);
2526
2527         if (PMC_IS_SYSTEM_MODE(mode))
2528                 disp = PMC_DISP_STANDALONE;
2529         else
2530                 disp = PMC_DISP_THREAD;
2531
2532         /*
2533          * check disposition for PMC row 'ri':
2534          *
2535          * Expected disposition         Row-disposition         Result
2536          *
2537          * STANDALONE                   STANDALONE or FREE      proceed
2538          * STANDALONE                   THREAD                  fail
2539          * THREAD                       THREAD or FREE          proceed
2540          * THREAD                       STANDALONE              fail
2541          */
2542
2543         if (!PMC_ROW_DISP_IS_FREE(ri) &&
2544             !(disp == PMC_DISP_THREAD && PMC_ROW_DISP_IS_THREAD(ri)) &&
2545             !(disp == PMC_DISP_STANDALONE && PMC_ROW_DISP_IS_STANDALONE(ri)))
2546                 return EBUSY;
2547
2548         /*
2549          * All OK
2550          */
2551
2552         PMCDBG2(PMC,ALR,2, "can-allocate-row ri=%d mode=%d ok", ri, mode);
2553
2554         return 0;
2555
2556 }
2557
2558 /*
2559  * Find a PMC descriptor with user handle 'pmcid' for thread 'td'.
2560  */
2561
2562 static struct pmc *
2563 pmc_find_pmc_descriptor_in_process(struct pmc_owner *po, pmc_id_t pmcid)
2564 {
2565         struct pmc *pm;
2566
2567         KASSERT(PMC_ID_TO_ROWINDEX(pmcid) < md->pmd_npmc,
2568             ("[pmc,%d] Illegal pmc index %d (max %d)", __LINE__,
2569                 PMC_ID_TO_ROWINDEX(pmcid), md->pmd_npmc));
2570
2571         LIST_FOREACH(pm, &po->po_pmcs, pm_next)
2572             if (pm->pm_id == pmcid)
2573                     return pm;
2574
2575         return NULL;
2576 }
2577
2578 static int
2579 pmc_find_pmc(pmc_id_t pmcid, struct pmc **pmc)
2580 {
2581
2582         struct pmc *pm;
2583         struct pmc_owner *po;
2584
2585         PMCDBG1(PMC,FND,1, "find-pmc id=%d", pmcid);
2586
2587         if ((po = pmc_find_owner_descriptor(curthread->td_proc)) == NULL)
2588                 return ESRCH;
2589
2590         if ((pm = pmc_find_pmc_descriptor_in_process(po, pmcid)) == NULL)
2591                 return EINVAL;
2592
2593         PMCDBG2(PMC,FND,2, "find-pmc id=%d -> pmc=%p", pmcid, pm);
2594
2595         *pmc = pm;
2596         return 0;
2597 }
2598
2599 /*
2600  * Start a PMC.
2601  */
2602
2603 static int
2604 pmc_start(struct pmc *pm)
2605 {
2606         enum pmc_mode mode;
2607         struct pmc_owner *po;
2608         struct pmc_binding pb;
2609         struct pmc_classdep *pcd;
2610         int adjri, error, cpu, ri;
2611
2612         KASSERT(pm != NULL,
2613             ("[pmc,%d] null pm", __LINE__));
2614
2615         mode = PMC_TO_MODE(pm);
2616         ri   = PMC_TO_ROWINDEX(pm);
2617         pcd  = pmc_ri_to_classdep(md, ri, &adjri);
2618
2619         error = 0;
2620
2621         PMCDBG3(PMC,OPS,1, "start pmc=%p mode=%d ri=%d", pm, mode, ri);
2622
2623         po = pm->pm_owner;
2624
2625         /*
2626          * Disallow PMCSTART if a logfile is required but has not been
2627          * configured yet.
2628          */
2629         if ((pm->pm_flags & PMC_F_NEEDS_LOGFILE) &&
2630             (po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
2631                 return (EDOOFUS);       /* programming error */
2632
2633         /*
2634          * If this is a sampling mode PMC, log mapping information for
2635          * the kernel modules that are currently loaded.
2636          */
2637         if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
2638             pmc_log_kernel_mappings(pm);
2639
2640         if (PMC_IS_VIRTUAL_MODE(mode)) {
2641
2642                 /*
2643                  * If a PMCATTACH has never been done on this PMC,
2644                  * attach it to its owner process.
2645                  */
2646
2647                 if (LIST_EMPTY(&pm->pm_targets))
2648                         error = (pm->pm_flags & PMC_F_ATTACH_DONE) ? ESRCH :
2649                             pmc_attach_process(po->po_owner, pm);
2650
2651                 /*
2652                  * If the PMC is attached to its owner, then force a context
2653                  * switch to ensure that the MD state gets set correctly.
2654                  */
2655
2656                 if (error == 0) {
2657                         pm->pm_state = PMC_STATE_RUNNING;
2658                         if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER)
2659                                 pmc_force_context_switch();
2660                 }
2661
2662                 return (error);
2663         }
2664
2665
2666         /*
2667          * A system-wide PMC.
2668          *
2669          * Add the owner to the global list if this is a system-wide
2670          * sampling PMC.
2671          */
2672
2673         if (mode == PMC_MODE_SS) {
2674                 if (po->po_sscount == 0) {
2675                         LIST_INSERT_HEAD(&pmc_ss_owners, po, po_ssnext);
2676                         atomic_add_rel_int(&pmc_ss_count, 1);
2677                         PMCDBG1(PMC,OPS,1, "po=%p in global list", po);
2678                 }
2679                 po->po_sscount++;
2680
2681                 /*
2682                  * Log mapping information for all existing processes in the
2683                  * system.  Subsequent mappings are logged as they happen;
2684                  * see pmc_process_mmap().
2685                  */
2686                 if (po->po_logprocmaps == 0) {
2687                         pmc_log_all_process_mappings(po);
2688                         po->po_logprocmaps = 1;
2689                 }
2690         }
2691
2692         /*
2693          * Move to the CPU associated with this
2694          * PMC, and start the hardware.
2695          */
2696
2697         pmc_save_cpu_binding(&pb);
2698
2699         cpu = PMC_TO_CPU(pm);
2700
2701         if (!pmc_cpu_is_active(cpu))
2702                 return (ENXIO);
2703
2704         pmc_select_cpu(cpu);
2705
2706         /*
2707          * global PMCs are configured at allocation time
2708          * so write out the initial value and start the PMC.
2709          */
2710
2711         pm->pm_state = PMC_STATE_RUNNING;
2712
2713         critical_enter();
2714         if ((error = pcd->pcd_write_pmc(cpu, adjri,
2715                  PMC_IS_SAMPLING_MODE(mode) ?
2716                  pm->pm_sc.pm_reloadcount :
2717                  pm->pm_sc.pm_initial)) == 0) {
2718                 /* If a sampling mode PMC, reset stalled state. */
2719                 if (PMC_IS_SAMPLING_MODE(mode))
2720                         CPU_CLR_ATOMIC(cpu, &pm->pm_stalled);
2721
2722                 /* Indicate that we desire this to run. Start it. */
2723                 CPU_SET_ATOMIC(cpu, &pm->pm_cpustate);
2724                 error = pcd->pcd_start_pmc(cpu, adjri);
2725         }
2726         critical_exit();
2727
2728         pmc_restore_cpu_binding(&pb);
2729
2730         return (error);
2731 }
2732
2733 /*
2734  * Stop a PMC.
2735  */
2736
2737 static int
2738 pmc_stop(struct pmc *pm)
2739 {
2740         struct pmc_owner *po;
2741         struct pmc_binding pb;
2742         struct pmc_classdep *pcd;
2743         int adjri, cpu, error, ri;
2744
2745         KASSERT(pm != NULL, ("[pmc,%d] null pmc", __LINE__));
2746
2747         PMCDBG3(PMC,OPS,1, "stop pmc=%p mode=%d ri=%d", pm,
2748             PMC_TO_MODE(pm), PMC_TO_ROWINDEX(pm));
2749
2750         pm->pm_state = PMC_STATE_STOPPED;
2751
2752         /*
2753          * If the PMC is a virtual mode one, changing the state to
2754          * non-RUNNING is enough to ensure that the PMC never gets
2755          * scheduled.
2756          *
2757          * If this PMC is current running on a CPU, then it will
2758          * handled correctly at the time its target process is context
2759          * switched out.
2760          */
2761
2762         if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)))
2763                 return 0;
2764
2765         /*
2766          * A system-mode PMC.  Move to the CPU associated with
2767          * this PMC, and stop the hardware.  We update the
2768          * 'initial count' so that a subsequent PMCSTART will
2769          * resume counting from the current hardware count.
2770          */
2771
2772         pmc_save_cpu_binding(&pb);
2773
2774         cpu = PMC_TO_CPU(pm);
2775
2776         KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
2777             ("[pmc,%d] illegal cpu=%d", __LINE__, cpu));
2778
2779         if (!pmc_cpu_is_active(cpu))
2780                 return ENXIO;
2781
2782         pmc_select_cpu(cpu);
2783
2784         ri = PMC_TO_ROWINDEX(pm);
2785         pcd = pmc_ri_to_classdep(md, ri, &adjri);
2786
2787         CPU_CLR_ATOMIC(cpu, &pm->pm_cpustate);
2788         critical_enter();
2789         if ((error = pcd->pcd_stop_pmc(cpu, adjri)) == 0)
2790                 error = pcd->pcd_read_pmc(cpu, adjri, &pm->pm_sc.pm_initial);
2791         critical_exit();
2792
2793         pmc_restore_cpu_binding(&pb);
2794
2795         po = pm->pm_owner;
2796
2797         /* remove this owner from the global list of SS PMC owners */
2798         if (PMC_TO_MODE(pm) == PMC_MODE_SS) {
2799                 po->po_sscount--;
2800                 if (po->po_sscount == 0) {
2801                         atomic_subtract_rel_int(&pmc_ss_count, 1);
2802                         LIST_REMOVE(po, po_ssnext);
2803                         PMCDBG1(PMC,OPS,2,"po=%p removed from global list", po);
2804                 }
2805         }
2806
2807         return (error);
2808 }
2809
2810
2811 #ifdef  HWPMC_DEBUG
2812 static const char *pmc_op_to_name[] = {
2813 #undef  __PMC_OP
2814 #define __PMC_OP(N, D)  #N ,
2815         __PMC_OPS()
2816         NULL
2817 };
2818 #endif
2819
2820 /*
2821  * The syscall interface
2822  */
2823
2824 #define PMC_GET_SX_XLOCK(...) do {              \
2825         sx_xlock(&pmc_sx);                      \
2826         if (pmc_hook == NULL) {                 \
2827                 sx_xunlock(&pmc_sx);            \
2828                 return __VA_ARGS__;             \
2829         }                                       \
2830 } while (0)
2831
2832 #define PMC_DOWNGRADE_SX() do {                 \
2833         sx_downgrade(&pmc_sx);                  \
2834         is_sx_downgraded = 1;                   \
2835 } while (0)
2836
2837 static int
2838 pmc_syscall_handler(struct thread *td, void *syscall_args)
2839 {
2840         int error, is_sx_downgraded, is_sx_locked, op;
2841         struct pmc_syscall_args *c;
2842         void *arg;
2843
2844         PMC_GET_SX_XLOCK(ENOSYS);
2845
2846         DROP_GIANT();
2847
2848         is_sx_downgraded = 0;
2849         is_sx_locked = 1;
2850
2851         c = (struct pmc_syscall_args *) syscall_args;
2852
2853         op = c->pmop_code;
2854         arg = c->pmop_data;
2855
2856         PMCDBG3(MOD,PMS,1, "syscall op=%d \"%s\" arg=%p", op,
2857             pmc_op_to_name[op], arg);
2858
2859         error = 0;
2860         atomic_add_int(&pmc_stats.pm_syscalls, 1);
2861
2862         switch(op)
2863         {
2864
2865
2866         /*
2867          * Configure a log file.
2868          *
2869          * XXX This OP will be reworked.
2870          */
2871
2872         case PMC_OP_CONFIGURELOG:
2873         {
2874                 struct proc *p;
2875                 struct pmc *pm;
2876                 struct pmc_owner *po;
2877                 struct pmc_op_configurelog cl;
2878
2879                 sx_assert(&pmc_sx, SX_XLOCKED);
2880
2881                 if ((error = copyin(arg, &cl, sizeof(cl))) != 0)
2882                         break;
2883
2884                 /* mark this process as owning a log file */
2885                 p = td->td_proc;
2886                 if ((po = pmc_find_owner_descriptor(p)) == NULL)
2887                         if ((po = pmc_allocate_owner_descriptor(p)) == NULL) {
2888                                 error = ENOMEM;
2889                                 break;
2890                         }
2891
2892                 /*
2893                  * If a valid fd was passed in, try to configure that,
2894                  * otherwise if 'fd' was less than zero and there was
2895                  * a log file configured, flush its buffers and
2896                  * de-configure it.
2897                  */
2898                 if (cl.pm_logfd >= 0) {
2899                         sx_xunlock(&pmc_sx);
2900                         is_sx_locked = 0;
2901                         error = pmclog_configure_log(md, po, cl.pm_logfd);
2902                 } else if (po->po_flags & PMC_PO_OWNS_LOGFILE) {
2903                         pmclog_process_closelog(po);
2904                         error = pmclog_close(po);
2905                         if (error == 0) {
2906                                 LIST_FOREACH(pm, &po->po_pmcs, pm_next)
2907                                     if (pm->pm_flags & PMC_F_NEEDS_LOGFILE &&
2908                                         pm->pm_state == PMC_STATE_RUNNING)
2909                                             pmc_stop(pm);
2910                                 error = pmclog_deconfigure_log(po);
2911                         }
2912                 } else
2913                         error = EINVAL;
2914
2915                 if (error)
2916                         break;
2917         }
2918         break;
2919
2920         /*
2921          * Flush a log file.
2922          */
2923
2924         case PMC_OP_FLUSHLOG:
2925         {
2926                 struct pmc_owner *po;
2927
2928                 sx_assert(&pmc_sx, SX_XLOCKED);
2929
2930                 if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) {
2931                         error = EINVAL;
2932                         break;
2933                 }
2934
2935                 error = pmclog_flush(po);
2936         }
2937         break;
2938
2939         /*
2940          * Close a log file.
2941          */
2942
2943         case PMC_OP_CLOSELOG:
2944         {
2945                 struct pmc_owner *po;
2946
2947                 sx_assert(&pmc_sx, SX_XLOCKED);
2948
2949                 if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) {
2950                         error = EINVAL;
2951                         break;
2952                 }
2953
2954                 error = pmclog_close(po);
2955         }
2956         break;
2957
2958         /*
2959          * Retrieve hardware configuration.
2960          */
2961
2962         case PMC_OP_GETCPUINFO: /* CPU information */
2963         {
2964                 struct pmc_op_getcpuinfo gci;
2965                 struct pmc_classinfo *pci;
2966                 struct pmc_classdep *pcd;
2967                 int cl;
2968
2969                 gci.pm_cputype = md->pmd_cputype;
2970                 gci.pm_ncpu    = pmc_cpu_max();
2971                 gci.pm_npmc    = md->pmd_npmc;
2972                 gci.pm_nclass  = md->pmd_nclass;
2973                 pci = gci.pm_classes;
2974                 pcd = md->pmd_classdep;
2975                 for (cl = 0; cl < md->pmd_nclass; cl++, pci++, pcd++) {
2976                         pci->pm_caps  = pcd->pcd_caps;
2977                         pci->pm_class = pcd->pcd_class;
2978                         pci->pm_width = pcd->pcd_width;
2979                         pci->pm_num   = pcd->pcd_num;
2980                 }
2981                 error = copyout(&gci, arg, sizeof(gci));
2982         }
2983         break;
2984
2985         /*
2986          * Retrieve soft events list.
2987          */
2988         case PMC_OP_GETDYNEVENTINFO:
2989         {
2990                 enum pmc_class                  cl;
2991                 enum pmc_event                  ev;
2992                 struct pmc_op_getdyneventinfo   *gei;
2993                 struct pmc_dyn_event_descr      dev;
2994                 struct pmc_soft                 *ps;
2995                 uint32_t                        nevent;
2996
2997                 sx_assert(&pmc_sx, SX_LOCKED);
2998
2999                 gei = (struct pmc_op_getdyneventinfo *) arg;
3000
3001                 if ((error = copyin(&gei->pm_class, &cl, sizeof(cl))) != 0)
3002                         break;
3003
3004                 /* Only SOFT class is dynamic. */
3005                 if (cl != PMC_CLASS_SOFT) {
3006                         error = EINVAL;
3007                         break;
3008                 }
3009
3010                 nevent = 0;
3011                 for (ev = PMC_EV_SOFT_FIRST; (int)ev <= PMC_EV_SOFT_LAST; ev++) {
3012                         ps = pmc_soft_ev_acquire(ev);
3013                         if (ps == NULL)
3014                                 continue;
3015                         bcopy(&ps->ps_ev, &dev, sizeof(dev));
3016                         pmc_soft_ev_release(ps);
3017
3018                         error = copyout(&dev,
3019                             &gei->pm_events[nevent],
3020                             sizeof(struct pmc_dyn_event_descr));
3021                         if (error != 0)
3022                                 break;
3023                         nevent++;
3024                 }
3025                 if (error != 0)
3026                         break;
3027
3028                 error = copyout(&nevent, &gei->pm_nevent,
3029                     sizeof(nevent));
3030         }
3031         break;
3032
3033         /*
3034          * Get module statistics
3035          */
3036
3037         case PMC_OP_GETDRIVERSTATS:
3038         {
3039                 struct pmc_op_getdriverstats gms;
3040
3041                 bcopy(&pmc_stats, &gms, sizeof(gms));
3042                 error = copyout(&gms, arg, sizeof(gms));
3043         }
3044         break;
3045
3046
3047         /*
3048          * Retrieve module version number
3049          */
3050
3051         case PMC_OP_GETMODULEVERSION:
3052         {
3053                 uint32_t cv, modv;
3054
3055                 /* retrieve the client's idea of the ABI version */
3056                 if ((error = copyin(arg, &cv, sizeof(uint32_t))) != 0)
3057                         break;
3058                 /* don't service clients newer than our driver */
3059                 modv = PMC_VERSION;
3060                 if ((cv & 0xFFFF0000) > (modv & 0xFFFF0000)) {
3061                         error = EPROGMISMATCH;
3062                         break;
3063                 }
3064                 error = copyout(&modv, arg, sizeof(int));
3065         }
3066         break;
3067
3068
3069         /*
3070          * Retrieve the state of all the PMCs on a given
3071          * CPU.
3072          */
3073
3074         case PMC_OP_GETPMCINFO:
3075         {
3076                 int ari;
3077                 struct pmc *pm;
3078                 size_t pmcinfo_size;
3079                 uint32_t cpu, n, npmc;
3080                 struct pmc_owner *po;
3081                 struct pmc_binding pb;
3082                 struct pmc_classdep *pcd;
3083                 struct pmc_info *p, *pmcinfo;
3084                 struct pmc_op_getpmcinfo *gpi;
3085
3086                 PMC_DOWNGRADE_SX();
3087
3088                 gpi = (struct pmc_op_getpmcinfo *) arg;
3089
3090                 if ((error = copyin(&gpi->pm_cpu, &cpu, sizeof(cpu))) != 0)
3091                         break;
3092
3093                 if (cpu >= pmc_cpu_max()) {
3094                         error = EINVAL;
3095                         break;
3096                 }
3097
3098                 if (!pmc_cpu_is_active(cpu)) {
3099                         error = ENXIO;
3100                         break;
3101                 }
3102
3103                 /* switch to CPU 'cpu' */
3104                 pmc_save_cpu_binding(&pb);
3105                 pmc_select_cpu(cpu);
3106
3107                 npmc = md->pmd_npmc;
3108
3109                 pmcinfo_size = npmc * sizeof(struct pmc_info);
3110                 pmcinfo = malloc(pmcinfo_size, M_PMC, M_WAITOK);
3111
3112                 p = pmcinfo;
3113
3114                 for (n = 0; n < md->pmd_npmc; n++, p++) {
3115
3116                         pcd = pmc_ri_to_classdep(md, n, &ari);
3117
3118                         KASSERT(pcd != NULL,
3119                             ("[pmc,%d] null pcd ri=%d", __LINE__, n));
3120
3121                         if ((error = pcd->pcd_describe(cpu, ari, p, &pm)) != 0)
3122                                 break;
3123
3124                         if (PMC_ROW_DISP_IS_STANDALONE(n))
3125                                 p->pm_rowdisp = PMC_DISP_STANDALONE;
3126                         else if (PMC_ROW_DISP_IS_THREAD(n))
3127                                 p->pm_rowdisp = PMC_DISP_THREAD;
3128                         else
3129                                 p->pm_rowdisp = PMC_DISP_FREE;
3130
3131                         p->pm_ownerpid = -1;
3132
3133                         if (pm == NULL) /* no PMC associated */
3134                                 continue;
3135
3136                         po = pm->pm_owner;
3137
3138                         KASSERT(po->po_owner != NULL,
3139                             ("[pmc,%d] pmc_owner had a null proc pointer",
3140                                 __LINE__));
3141
3142                         p->pm_ownerpid = po->po_owner->p_pid;
3143                         p->pm_mode     = PMC_TO_MODE(pm);
3144                         p->pm_event    = pm->pm_event;
3145                         p->pm_flags    = pm->pm_flags;
3146
3147                         if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
3148                                 p->pm_reloadcount =
3149                                     pm->pm_sc.pm_reloadcount;
3150                 }
3151
3152                 pmc_restore_cpu_binding(&pb);
3153
3154                 /* now copy out the PMC info collected */
3155                 if (error == 0)
3156                         error = copyout(pmcinfo, &gpi->pm_pmcs, pmcinfo_size);
3157
3158                 free(pmcinfo, M_PMC);
3159         }
3160         break;
3161
3162
3163         /*
3164          * Set the administrative state of a PMC.  I.e. whether
3165          * the PMC is to be used or not.
3166          */
3167
3168         case PMC_OP_PMCADMIN:
3169         {
3170                 int cpu, ri;
3171                 enum pmc_state request;
3172                 struct pmc_cpu *pc;
3173                 struct pmc_hw *phw;
3174                 struct pmc_op_pmcadmin pma;
3175                 struct pmc_binding pb;
3176
3177                 sx_assert(&pmc_sx, SX_XLOCKED);
3178
3179                 KASSERT(td == curthread,
3180                     ("[pmc,%d] td != curthread", __LINE__));
3181
3182                 error = priv_check(td, PRIV_PMC_MANAGE);
3183                 if (error)
3184                         break;
3185
3186                 if ((error = copyin(arg, &pma, sizeof(pma))) != 0)
3187                         break;
3188
3189                 cpu = pma.pm_cpu;
3190
3191                 if (cpu < 0 || cpu >= (int) pmc_cpu_max()) {
3192                         error = EINVAL;
3193                         break;
3194                 }
3195
3196                 if (!pmc_cpu_is_active(cpu)) {
3197                         error = ENXIO;
3198                         break;
3199                 }
3200
3201                 request = pma.pm_state;
3202
3203                 if (request != PMC_STATE_DISABLED &&
3204                     request != PMC_STATE_FREE) {
3205                         error = EINVAL;
3206                         break;
3207                 }
3208
3209                 ri = pma.pm_pmc; /* pmc id == row index */
3210                 if (ri < 0 || ri >= (int) md->pmd_npmc) {
3211                         error = EINVAL;
3212                         break;
3213                 }
3214
3215                 /*
3216                  * We can't disable a PMC with a row-index allocated
3217                  * for process virtual PMCs.
3218                  */
3219
3220                 if (PMC_ROW_DISP_IS_THREAD(ri) &&
3221                     request == PMC_STATE_DISABLED) {
3222                         error = EBUSY;
3223                         break;
3224                 }
3225
3226                 /*
3227                  * otherwise, this PMC on this CPU is either free or
3228                  * in system-wide mode.
3229                  */
3230
3231                 pmc_save_cpu_binding(&pb);
3232                 pmc_select_cpu(cpu);
3233
3234                 pc  = pmc_pcpu[cpu];
3235                 phw = pc->pc_hwpmcs[ri];
3236
3237                 /*
3238                  * XXX do we need some kind of 'forced' disable?
3239                  */
3240
3241                 if (phw->phw_pmc == NULL) {
3242                         if (request == PMC_STATE_DISABLED &&
3243                             (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED)) {
3244                                 phw->phw_state &= ~PMC_PHW_FLAG_IS_ENABLED;
3245                                 PMC_MARK_ROW_STANDALONE(ri);
3246                         } else if (request == PMC_STATE_FREE &&
3247                             (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0) {
3248                                 phw->phw_state |=  PMC_PHW_FLAG_IS_ENABLED;
3249                                 PMC_UNMARK_ROW_STANDALONE(ri);
3250                         }
3251                         /* other cases are a no-op */
3252                 } else
3253                         error = EBUSY;
3254
3255                 pmc_restore_cpu_binding(&pb);
3256         }
3257         break;
3258
3259
3260         /*
3261          * Allocate a PMC.
3262          */
3263
3264         case PMC_OP_PMCALLOCATE:
3265         {
3266                 int adjri, n;
3267                 u_int cpu;
3268                 uint32_t caps;
3269                 struct pmc *pmc;
3270                 enum pmc_mode mode;
3271                 struct pmc_hw *phw;
3272                 struct pmc_binding pb;
3273                 struct pmc_classdep *pcd;
3274                 struct pmc_op_pmcallocate pa;
3275
3276                 if ((error = copyin(arg, &pa, sizeof(pa))) != 0)
3277                         break;
3278
3279                 caps = pa.pm_caps;
3280                 mode = pa.pm_mode;
3281                 cpu  = pa.pm_cpu;
3282
3283                 if ((mode != PMC_MODE_SS  &&  mode != PMC_MODE_SC  &&
3284                      mode != PMC_MODE_TS  &&  mode != PMC_MODE_TC) ||
3285                     (cpu != (u_int) PMC_CPU_ANY && cpu >= pmc_cpu_max())) {
3286                         error = EINVAL;
3287                         break;
3288                 }
3289
3290                 /*
3291                  * Virtual PMCs should only ask for a default CPU.
3292                  * System mode PMCs need to specify a non-default CPU.
3293                  */
3294
3295                 if ((PMC_IS_VIRTUAL_MODE(mode) && cpu != (u_int) PMC_CPU_ANY) ||
3296                     (PMC_IS_SYSTEM_MODE(mode) && cpu == (u_int) PMC_CPU_ANY)) {
3297                         error = EINVAL;
3298                         break;
3299                 }
3300
3301                 /*
3302                  * Check that an inactive CPU is not being asked for.
3303                  */
3304
3305                 if (PMC_IS_SYSTEM_MODE(mode) && !pmc_cpu_is_active(cpu)) {
3306                         error = ENXIO;
3307                         break;
3308                 }
3309
3310                 /*
3311                  * Refuse an allocation for a system-wide PMC if this
3312                  * process has been jailed, or if this process lacks
3313                  * super-user credentials and the sysctl tunable
3314                  * 'security.bsd.unprivileged_syspmcs' is zero.
3315                  */
3316
3317                 if (PMC_IS_SYSTEM_MODE(mode)) {
3318                         if (jailed(curthread->td_ucred)) {
3319                                 error = EPERM;
3320                                 break;
3321                         }
3322                         if (!pmc_unprivileged_syspmcs) {
3323                                 error = priv_check(curthread,
3324                                     PRIV_PMC_SYSTEM);
3325                                 if (error)
3326                                         break;
3327                         }
3328                 }
3329
3330                 /*
3331                  * Look for valid values for 'pm_flags'
3332                  */
3333
3334                 if ((pa.pm_flags & ~(PMC_F_DESCENDANTS | PMC_F_LOG_PROCCSW |
3335                     PMC_F_LOG_PROCEXIT | PMC_F_CALLCHAIN)) != 0) {
3336                         error = EINVAL;
3337                         break;
3338                 }
3339
3340                 /* process logging options are not allowed for system PMCs */
3341                 if (PMC_IS_SYSTEM_MODE(mode) && (pa.pm_flags &
3342                     (PMC_F_LOG_PROCCSW | PMC_F_LOG_PROCEXIT))) {
3343                         error = EINVAL;
3344                         break;
3345                 }
3346
3347                 /*
3348                  * All sampling mode PMCs need to be able to interrupt the
3349                  * CPU.
3350                  */
3351                 if (PMC_IS_SAMPLING_MODE(mode))
3352                         caps |= PMC_CAP_INTERRUPT;
3353
3354                 /* A valid class specifier should have been passed in. */
3355                 for (n = 0; n < md->pmd_nclass; n++)
3356                         if (md->pmd_classdep[n].pcd_class == pa.pm_class)
3357                                 break;
3358                 if (n == md->pmd_nclass) {
3359                         error = EINVAL;
3360                         break;
3361                 }
3362
3363                 /* The requested PMC capabilities should be feasible. */
3364                 if ((md->pmd_classdep[n].pcd_caps & caps) != caps) {
3365                         error = EOPNOTSUPP;
3366                         break;
3367                 }
3368
3369                 PMCDBG4(PMC,ALL,2, "event=%d caps=0x%x mode=%d cpu=%d",
3370                     pa.pm_ev, caps, mode, cpu);
3371
3372                 pmc = pmc_allocate_pmc_descriptor();
3373                 pmc->pm_id    = PMC_ID_MAKE_ID(cpu,pa.pm_mode,pa.pm_class,
3374                     PMC_ID_INVALID);
3375                 pmc->pm_event = pa.pm_ev;
3376                 pmc->pm_state = PMC_STATE_FREE;
3377                 pmc->pm_caps  = caps;
3378                 pmc->pm_flags = pa.pm_flags;
3379
3380                 /* switch thread to CPU 'cpu' */
3381                 pmc_save_cpu_binding(&pb);
3382
3383 #define PMC_IS_SHAREABLE_PMC(cpu, n)                            \
3384         (pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_state &           \
3385          PMC_PHW_FLAG_IS_SHAREABLE)
3386 #define PMC_IS_UNALLOCATED(cpu, n)                              \
3387         (pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_pmc == NULL)
3388
3389                 if (PMC_IS_SYSTEM_MODE(mode)) {
3390                         pmc_select_cpu(cpu);
3391                         for (n = 0; n < (int) md->pmd_npmc; n++) {
3392                                 pcd = pmc_ri_to_classdep(md, n, &adjri);
3393                                 if (pmc_can_allocate_row(n, mode) == 0 &&
3394                                     pmc_can_allocate_rowindex(
3395                                             curthread->td_proc, n, cpu) == 0 &&
3396                                     (PMC_IS_UNALLOCATED(cpu, n) ||
3397                                      PMC_IS_SHAREABLE_PMC(cpu, n)) &&
3398                                     pcd->pcd_allocate_pmc(cpu, adjri, pmc,
3399                                         &pa) == 0)
3400                                         break;
3401                         }
3402                 } else {
3403                         /* Process virtual mode */
3404                         for (n = 0; n < (int) md->pmd_npmc; n++) {
3405                                 pcd = pmc_ri_to_classdep(md, n, &adjri);
3406                                 if (pmc_can_allocate_row(n, mode) == 0 &&
3407                                     pmc_can_allocate_rowindex(
3408                                             curthread->td_proc, n,
3409                                             PMC_CPU_ANY) == 0 &&
3410                                     pcd->pcd_allocate_pmc(curthread->td_oncpu,
3411                                         adjri, pmc, &pa) == 0)
3412                                         break;
3413                         }
3414                 }
3415
3416 #undef  PMC_IS_UNALLOCATED
3417 #undef  PMC_IS_SHAREABLE_PMC
3418
3419                 pmc_restore_cpu_binding(&pb);
3420
3421                 if (n == (int) md->pmd_npmc) {
3422                         pmc_destroy_pmc_descriptor(pmc);
3423                         pmc = NULL;
3424                         error = EINVAL;
3425                         break;
3426                 }
3427
3428                 /* Fill in the correct value in the ID field */
3429                 pmc->pm_id = PMC_ID_MAKE_ID(cpu,mode,pa.pm_class,n);
3430
3431                 PMCDBG5(PMC,ALL,2, "ev=%d class=%d mode=%d n=%d -> pmcid=%x",
3432                     pmc->pm_event, pa.pm_class, mode, n, pmc->pm_id);
3433
3434                 /* Process mode PMCs with logging enabled need log files */
3435                 if (pmc->pm_flags & (PMC_F_LOG_PROCEXIT | PMC_F_LOG_PROCCSW))
3436                         pmc->pm_flags |= PMC_F_NEEDS_LOGFILE;
3437
3438                 /* All system mode sampling PMCs require a log file */
3439                 if (PMC_IS_SAMPLING_MODE(mode) && PMC_IS_SYSTEM_MODE(mode))
3440                         pmc->pm_flags |= PMC_F_NEEDS_LOGFILE;
3441
3442                 /*
3443                  * Configure global pmc's immediately
3444                  */
3445
3446                 if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pmc))) {
3447
3448                         pmc_save_cpu_binding(&pb);
3449                         pmc_select_cpu(cpu);
3450
3451                         phw = pmc_pcpu[cpu]->pc_hwpmcs[n];
3452                         pcd = pmc_ri_to_classdep(md, n, &adjri);
3453
3454                         if ((phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0 ||
3455                             (error = pcd->pcd_config_pmc(cpu, adjri, pmc)) != 0) {
3456                                 (void) pcd->pcd_release_pmc(cpu, adjri, pmc);
3457                                 pmc_destroy_pmc_descriptor(pmc);
3458                                 pmc = NULL;
3459                                 pmc_restore_cpu_binding(&pb);
3460                                 error = EPERM;
3461                                 break;
3462                         }
3463
3464                         pmc_restore_cpu_binding(&pb);
3465                 }
3466
3467                 pmc->pm_state    = PMC_STATE_ALLOCATED;
3468
3469                 /*
3470                  * mark row disposition
3471                  */
3472
3473                 if (PMC_IS_SYSTEM_MODE(mode))
3474                         PMC_MARK_ROW_STANDALONE(n);
3475                 else
3476                         PMC_MARK_ROW_THREAD(n);
3477
3478                 /*
3479                  * Register this PMC with the current thread as its owner.
3480                  */
3481
3482                 if ((error =
3483                     pmc_register_owner(curthread->td_proc, pmc)) != 0) {
3484                         pmc_release_pmc_descriptor(pmc);
3485                         pmc_destroy_pmc_descriptor(pmc);
3486                         pmc = NULL;
3487                         break;
3488                 }
3489
3490                 /*
3491                  * Return the allocated index.
3492                  */
3493
3494                 pa.pm_pmcid = pmc->pm_id;
3495
3496                 error = copyout(&pa, arg, sizeof(pa));
3497         }
3498         break;
3499
3500
3501         /*
3502          * Attach a PMC to a process.
3503          */
3504
3505         case PMC_OP_PMCATTACH:
3506         {
3507                 struct pmc *pm;
3508                 struct proc *p;
3509                 struct pmc_op_pmcattach a;
3510
3511                 sx_assert(&pmc_sx, SX_XLOCKED);
3512
3513                 if ((error = copyin(arg, &a, sizeof(a))) != 0)
3514                         break;
3515
3516                 if (a.pm_pid < 0) {
3517                         error = EINVAL;
3518                         break;
3519                 } else if (a.pm_pid == 0)
3520                         a.pm_pid = td->td_proc->p_pid;
3521
3522                 if ((error = pmc_find_pmc(a.pm_pmc, &pm)) != 0)
3523                         break;
3524
3525                 if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm))) {
3526                         error = EINVAL;
3527                         break;
3528                 }
3529
3530                 /* PMCs may be (re)attached only when allocated or stopped */
3531                 if (pm->pm_state == PMC_STATE_RUNNING) {
3532                         error = EBUSY;
3533                         break;
3534                 } else if (pm->pm_state != PMC_STATE_ALLOCATED &&
3535                     pm->pm_state != PMC_STATE_STOPPED) {
3536                         error = EINVAL;
3537                         break;
3538                 }
3539
3540                 /* lookup pid */
3541                 if ((p = pfind(a.pm_pid)) == NULL) {
3542                         error = ESRCH;
3543                         break;
3544                 }
3545
3546                 /*
3547                  * Ignore processes that are working on exiting.
3548                  */
3549                 if (p->p_flag & P_WEXIT) {
3550                         error = ESRCH;
3551                         PROC_UNLOCK(p); /* pfind() returns a locked process */
3552                         break;
3553                 }
3554
3555                 /*
3556                  * we are allowed to attach a PMC to a process if
3557                  * we can debug it.
3558                  */
3559                 error = p_candebug(curthread, p);
3560
3561                 PROC_UNLOCK(p);
3562
3563                 if (error == 0)
3564                         error = pmc_attach_process(p, pm);
3565         }
3566         break;
3567
3568
3569         /*
3570          * Detach an attached PMC from a process.
3571          */
3572
3573         case PMC_OP_PMCDETACH:
3574         {
3575                 struct pmc *pm;
3576                 struct proc *p;
3577                 struct pmc_op_pmcattach a;
3578
3579                 if ((error = copyin(arg, &a, sizeof(a))) != 0)
3580                         break;
3581
3582                 if (a.pm_pid < 0) {
3583                         error = EINVAL;
3584                         break;
3585                 } else if (a.pm_pid == 0)
3586                         a.pm_pid = td->td_proc->p_pid;
3587
3588                 if ((error = pmc_find_pmc(a.pm_pmc, &pm)) != 0)
3589                         break;
3590
3591                 if ((p = pfind(a.pm_pid)) == NULL) {
3592                         error = ESRCH;
3593                         break;
3594                 }
3595
3596                 /*
3597                  * Treat processes that are in the process of exiting
3598                  * as if they were not present.
3599                  */
3600
3601                 if (p->p_flag & P_WEXIT)
3602                         error = ESRCH;
3603
3604                 PROC_UNLOCK(p); /* pfind() returns a locked process */
3605
3606                 if (error == 0)
3607                         error = pmc_detach_process(p, pm);
3608         }
3609         break;
3610
3611
3612         /*
3613          * Retrieve the MSR number associated with the counter
3614          * 'pmc_id'.  This allows processes to directly use RDPMC
3615          * instructions to read their PMCs, without the overhead of a
3616          * system call.
3617          */
3618
3619         case PMC_OP_PMCGETMSR:
3620         {
3621                 int adjri, ri;
3622                 struct pmc *pm;
3623                 struct pmc_target *pt;
3624                 struct pmc_op_getmsr gm;
3625                 struct pmc_classdep *pcd;
3626
3627                 PMC_DOWNGRADE_SX();
3628
3629                 if ((error = copyin(arg, &gm, sizeof(gm))) != 0)
3630                         break;
3631
3632                 if ((error = pmc_find_pmc(gm.pm_pmcid, &pm)) != 0)
3633                         break;
3634
3635                 /*
3636                  * The allocated PMC has to be a process virtual PMC,
3637                  * i.e., of type MODE_T[CS].  Global PMCs can only be
3638                  * read using the PMCREAD operation since they may be
3639                  * allocated on a different CPU than the one we could
3640                  * be running on at the time of the RDPMC instruction.
3641                  *
3642                  * The GETMSR operation is not allowed for PMCs that
3643                  * are inherited across processes.
3644                  */
3645
3646                 if (!PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)) ||
3647                     (pm->pm_flags & PMC_F_DESCENDANTS)) {
3648                         error = EINVAL;
3649                         break;
3650                 }
3651
3652                 /*
3653                  * It only makes sense to use a RDPMC (or its
3654                  * equivalent instruction on non-x86 architectures) on
3655                  * a process that has allocated and attached a PMC to
3656                  * itself.  Conversely the PMC is only allowed to have
3657                  * one process attached to it -- its owner.
3658                  */
3659
3660                 if ((pt = LIST_FIRST(&pm->pm_targets)) == NULL ||
3661                     LIST_NEXT(pt, pt_next) != NULL ||
3662                     pt->pt_process->pp_proc != pm->pm_owner->po_owner) {
3663                         error = EINVAL;
3664                         break;
3665                 }
3666
3667                 ri = PMC_TO_ROWINDEX(pm);
3668                 pcd = pmc_ri_to_classdep(md, ri, &adjri);
3669
3670                 /* PMC class has no 'GETMSR' support */
3671                 if (pcd->pcd_get_msr == NULL) {
3672                         error = ENOSYS;
3673                         break;
3674                 }
3675
3676                 if ((error = (*pcd->pcd_get_msr)(adjri, &gm.pm_msr)) < 0)
3677                         break;
3678
3679                 if ((error = copyout(&gm, arg, sizeof(gm))) < 0)
3680                         break;
3681
3682                 /*
3683                  * Mark our process as using MSRs.  Update machine
3684                  * state using a forced context switch.
3685                  */
3686
3687                 pt->pt_process->pp_flags |= PMC_PP_ENABLE_MSR_ACCESS;
3688                 pmc_force_context_switch();
3689
3690         }
3691         break;
3692
3693         /*
3694          * Release an allocated PMC
3695          */
3696
3697         case PMC_OP_PMCRELEASE:
3698         {
3699                 pmc_id_t pmcid;
3700                 struct pmc *pm;
3701                 struct pmc_owner *po;
3702                 struct pmc_op_simple sp;
3703
3704                 /*
3705                  * Find PMC pointer for the named PMC.
3706                  *
3707                  * Use pmc_release_pmc_descriptor() to switch off the
3708                  * PMC, remove all its target threads, and remove the
3709                  * PMC from its owner's list.
3710                  *
3711                  * Remove the owner record if this is the last PMC
3712                  * owned.
3713                  *
3714                  * Free up space.
3715                  */
3716
3717                 if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
3718                         break;
3719
3720                 pmcid = sp.pm_pmcid;
3721
3722                 if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
3723                         break;
3724
3725                 po = pm->pm_owner;
3726                 pmc_release_pmc_descriptor(pm);
3727                 pmc_maybe_remove_owner(po);
3728                 pmc_destroy_pmc_descriptor(pm);
3729         }
3730         break;
3731
3732
3733         /*
3734          * Read and/or write a PMC.
3735          */
3736
3737         case PMC_OP_PMCRW:
3738         {
3739                 int adjri;
3740                 struct pmc *pm;
3741                 uint32_t cpu, ri;
3742                 pmc_value_t oldvalue;
3743                 struct pmc_binding pb;
3744                 struct pmc_op_pmcrw prw;
3745                 struct pmc_classdep *pcd;
3746                 struct pmc_op_pmcrw *pprw;
3747
3748                 PMC_DOWNGRADE_SX();
3749
3750                 if ((error = copyin(arg, &prw, sizeof(prw))) != 0)
3751                         break;
3752
3753                 ri = 0;
3754                 PMCDBG2(PMC,OPS,1, "rw id=%d flags=0x%x", prw.pm_pmcid,
3755                     prw.pm_flags);
3756
3757                 /* must have at least one flag set */
3758                 if ((prw.pm_flags & (PMC_F_OLDVALUE|PMC_F_NEWVALUE)) == 0) {
3759                         error = EINVAL;
3760                         break;
3761                 }
3762
3763                 /* locate pmc descriptor */
3764                 if ((error = pmc_find_pmc(prw.pm_pmcid, &pm)) != 0)
3765                         break;
3766
3767                 /* Can't read a PMC that hasn't been started. */
3768                 if (pm->pm_state != PMC_STATE_ALLOCATED &&
3769                     pm->pm_state != PMC_STATE_STOPPED &&
3770                     pm->pm_state != PMC_STATE_RUNNING) {
3771                         error = EINVAL;
3772                         break;
3773                 }
3774
3775                 /* writing a new value is allowed only for 'STOPPED' pmcs */
3776                 if (pm->pm_state == PMC_STATE_RUNNING &&
3777                     (prw.pm_flags & PMC_F_NEWVALUE)) {
3778                         error = EBUSY;
3779                         break;
3780                 }
3781
3782                 if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) {
3783
3784                         /*
3785                          * If this PMC is attached to its owner (i.e.,
3786                          * the process requesting this operation) and
3787                          * is running, then attempt to get an
3788                          * upto-date reading from hardware for a READ.
3789                          * Writes are only allowed when the PMC is
3790                          * stopped, so only update the saved value
3791                          * field.
3792                          *
3793                          * If the PMC is not running, or is not
3794                          * attached to its owner, read/write to the
3795                          * savedvalue field.
3796                          */
3797
3798                         ri = PMC_TO_ROWINDEX(pm);
3799                         pcd = pmc_ri_to_classdep(md, ri, &adjri);
3800
3801                         mtx_pool_lock_spin(pmc_mtxpool, pm);
3802                         cpu = curthread->td_oncpu;
3803
3804                         if (prw.pm_flags & PMC_F_OLDVALUE) {
3805                                 if ((pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) &&
3806                                     (pm->pm_state == PMC_STATE_RUNNING))
3807                                         error = (*pcd->pcd_read_pmc)(cpu, adjri,
3808                                             &oldvalue);
3809                                 else
3810                                         oldvalue = pm->pm_gv.pm_savedvalue;
3811                         }
3812                         if (prw.pm_flags & PMC_F_NEWVALUE)
3813                                 pm->pm_gv.pm_savedvalue = prw.pm_value;
3814
3815                         mtx_pool_unlock_spin(pmc_mtxpool, pm);
3816
3817                 } else { /* System mode PMCs */
3818                         cpu = PMC_TO_CPU(pm);
3819                         ri  = PMC_TO_ROWINDEX(pm);
3820                         pcd = pmc_ri_to_classdep(md, ri, &adjri);
3821
3822                         if (!pmc_cpu_is_active(cpu)) {
3823                                 error = ENXIO;
3824                                 break;
3825                         }
3826
3827                         /* move this thread to CPU 'cpu' */
3828                         pmc_save_cpu_binding(&pb);
3829                         pmc_select_cpu(cpu);
3830
3831                         critical_enter();
3832                         /* save old value */
3833                         if (prw.pm_flags & PMC_F_OLDVALUE)
3834                                 if ((error = (*pcd->pcd_read_pmc)(cpu, adjri,
3835                                          &oldvalue)))
3836                                         goto error;
3837                         /* write out new value */
3838                         if (prw.pm_flags & PMC_F_NEWVALUE)
3839                                 error = (*pcd->pcd_write_pmc)(cpu, adjri,
3840                                     prw.pm_value);
3841                 error:
3842                         critical_exit();
3843                         pmc_restore_cpu_binding(&pb);
3844                         if (error)
3845                                 break;
3846                 }
3847
3848                 pprw = (struct pmc_op_pmcrw *) arg;
3849
3850 #ifdef  HWPMC_DEBUG
3851                 if (prw.pm_flags & PMC_F_NEWVALUE)
3852                         PMCDBG3(PMC,OPS,2, "rw id=%d new %jx -> old %jx",
3853                             ri, prw.pm_value, oldvalue);
3854                 else if (prw.pm_flags & PMC_F_OLDVALUE)
3855                         PMCDBG2(PMC,OPS,2, "rw id=%d -> old %jx", ri, oldvalue);
3856 #endif
3857
3858                 /* return old value if requested */
3859                 if (prw.pm_flags & PMC_F_OLDVALUE)
3860                         if ((error = copyout(&oldvalue, &pprw->pm_value,
3861                                  sizeof(prw.pm_value))))
3862                                 break;
3863
3864         }
3865         break;
3866
3867
3868         /*
3869          * Set the sampling rate for a sampling mode PMC and the
3870          * initial count for a counting mode PMC.
3871          */
3872
3873         case PMC_OP_PMCSETCOUNT:
3874         {
3875                 struct pmc *pm;
3876                 struct pmc_op_pmcsetcount sc;
3877
3878                 PMC_DOWNGRADE_SX();
3879
3880                 if ((error = copyin(arg, &sc, sizeof(sc))) != 0)
3881                         break;
3882
3883                 if ((error = pmc_find_pmc(sc.pm_pmcid, &pm)) != 0)
3884                         break;
3885
3886                 if (pm->pm_state == PMC_STATE_RUNNING) {
3887                         error = EBUSY;
3888                         break;
3889                 }
3890
3891                 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
3892                         pm->pm_sc.pm_reloadcount = sc.pm_count;
3893                 else
3894                         pm->pm_sc.pm_initial = sc.pm_count;
3895         }
3896         break;
3897
3898
3899         /*
3900          * Start a PMC.
3901          */
3902
3903         case PMC_OP_PMCSTART:
3904         {
3905                 pmc_id_t pmcid;
3906                 struct pmc *pm;
3907                 struct pmc_op_simple sp;
3908
3909                 sx_assert(&pmc_sx, SX_XLOCKED);
3910
3911                 if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
3912                         break;
3913
3914                 pmcid = sp.pm_pmcid;
3915
3916                 if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
3917                         break;
3918
3919                 KASSERT(pmcid == pm->pm_id,
3920                     ("[pmc,%d] pmcid %x != id %x", __LINE__,
3921                         pm->pm_id, pmcid));
3922
3923                 if (pm->pm_state == PMC_STATE_RUNNING) /* already running */
3924                         break;
3925                 else if (pm->pm_state != PMC_STATE_STOPPED &&
3926                     pm->pm_state != PMC_STATE_ALLOCATED) {
3927                         error = EINVAL;
3928                         break;
3929                 }
3930
3931                 error = pmc_start(pm);
3932         }
3933         break;
3934
3935
3936         /*
3937          * Stop a PMC.
3938          */
3939
3940         case PMC_OP_PMCSTOP:
3941         {
3942                 pmc_id_t pmcid;
3943                 struct pmc *pm;
3944                 struct pmc_op_simple sp;
3945
3946                 PMC_DOWNGRADE_SX();
3947
3948                 if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
3949                         break;
3950
3951                 pmcid = sp.pm_pmcid;
3952
3953                 /*
3954                  * Mark the PMC as inactive and invoke the MD stop
3955                  * routines if needed.
3956                  */
3957
3958                 if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
3959                         break;
3960
3961                 KASSERT(pmcid == pm->pm_id,
3962                     ("[pmc,%d] pmc id %x != pmcid %x", __LINE__,
3963                         pm->pm_id, pmcid));
3964
3965                 if (pm->pm_state == PMC_STATE_STOPPED) /* already stopped */
3966                         break;
3967                 else if (pm->pm_state != PMC_STATE_RUNNING) {
3968                         error = EINVAL;
3969                         break;
3970                 }
3971
3972                 error = pmc_stop(pm);
3973         }
3974         break;
3975
3976
3977         /*
3978          * Write a user supplied value to the log file.
3979          */
3980
3981         case PMC_OP_WRITELOG:
3982         {
3983                 struct pmc_op_writelog wl;
3984                 struct pmc_owner *po;
3985
3986                 PMC_DOWNGRADE_SX();
3987
3988                 if ((error = copyin(arg, &wl, sizeof(wl))) != 0)
3989                         break;
3990
3991                 if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) {
3992                         error = EINVAL;
3993                         break;
3994                 }
3995
3996                 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) {
3997                         error = EINVAL;
3998                         break;
3999                 }
4000
4001                 error = pmclog_process_userlog(po, &wl);
4002         }
4003         break;
4004
4005
4006         default:
4007                 error = EINVAL;
4008                 break;
4009         }
4010
4011         if (is_sx_locked != 0) {
4012                 if (is_sx_downgraded)
4013                         sx_sunlock(&pmc_sx);
4014                 else
4015                         sx_xunlock(&pmc_sx);
4016         }
4017
4018         if (error)
4019                 atomic_add_int(&pmc_stats.pm_syscall_errors, 1);
4020
4021         PICKUP_GIANT();
4022
4023         return error;
4024 }
4025
4026 /*
4027  * Helper functions
4028  */
4029
4030
4031 /*
4032  * Mark the thread as needing callchain capture and post an AST.  The
4033  * actual callchain capture will be done in a context where it is safe
4034  * to take page faults.
4035  */
4036
4037 static void
4038 pmc_post_callchain_callback(void)
4039 {
4040         struct thread *td;
4041
4042         td = curthread;
4043
4044         /*
4045          * If there is multiple PMCs for the same interrupt ignore new post
4046          */
4047         if (td->td_pflags & TDP_CALLCHAIN)
4048                 return;
4049
4050         /*
4051          * Mark this thread as needing callchain capture.
4052          * `td->td_pflags' will be safe to touch because this thread
4053          * was in user space when it was interrupted.
4054          */
4055         td->td_pflags |= TDP_CALLCHAIN;
4056
4057         /*
4058          * Don't let this thread migrate between CPUs until callchain
4059          * capture completes.
4060          */
4061         sched_pin();
4062
4063         return;
4064 }
4065
4066 /*
4067  * Interrupt processing.
4068  *
4069  * Find a free slot in the per-cpu array of samples and capture the
4070  * current callchain there.  If a sample was successfully added, a bit
4071  * is set in mask 'pmc_cpumask' denoting that the DO_SAMPLES hook
4072  * needs to be invoked from the clock handler.
4073  *
4074  * This function is meant to be called from an NMI handler.  It cannot
4075  * use any of the locking primitives supplied by the OS.
4076  */
4077
4078 int
4079 pmc_process_interrupt(int cpu, int ring, struct pmc *pm, struct trapframe *tf,
4080     int inuserspace)
4081 {
4082         int error, callchaindepth;
4083         struct thread *td;
4084         struct pmc_sample *ps;
4085         struct pmc_samplebuffer *psb;
4086
4087         error = 0;
4088
4089         /*
4090          * Allocate space for a sample buffer.
4091          */
4092         psb = pmc_pcpu[cpu]->pc_sb[ring];
4093
4094         ps = psb->ps_write;
4095         if (ps->ps_nsamples) {  /* in use, reader hasn't caught up */
4096                 CPU_SET_ATOMIC(cpu, &pm->pm_stalled);
4097                 atomic_add_int(&pmc_stats.pm_intr_bufferfull, 1);
4098                 PMCDBG6(SAM,INT,1,"(spc) cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d",
4099                     cpu, pm, (void *) tf, inuserspace,
4100                     (int) (psb->ps_write - psb->ps_samples),
4101                     (int) (psb->ps_read - psb->ps_samples));
4102                 callchaindepth = 1;
4103                 error = ENOMEM;
4104                 goto done;
4105         }
4106
4107
4108         /* Fill in entry. */
4109         PMCDBG6(SAM,INT,1,"cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d", cpu, pm,
4110             (void *) tf, inuserspace,
4111             (int) (psb->ps_write - psb->ps_samples),
4112             (int) (psb->ps_read - psb->ps_samples));
4113
4114         KASSERT(pm->pm_runcount >= 0,
4115             ("[pmc,%d] pm=%p runcount %d", __LINE__, (void *) pm,
4116                 pm->pm_runcount));
4117
4118         atomic_add_rel_int(&pm->pm_runcount, 1);        /* hold onto PMC */
4119
4120         ps->ps_pmc = pm;
4121         if ((td = curthread) && td->td_proc)
4122                 ps->ps_pid = td->td_proc->p_pid;
4123         else
4124                 ps->ps_pid = -1;
4125         ps->ps_cpu = cpu;
4126         ps->ps_td = td;
4127         ps->ps_flags = inuserspace ? PMC_CC_F_USERSPACE : 0;
4128
4129         callchaindepth = (pm->pm_flags & PMC_F_CALLCHAIN) ?
4130             pmc_callchaindepth : 1;
4131
4132         if (callchaindepth == 1)
4133                 ps->ps_pc[0] = PMC_TRAPFRAME_TO_PC(tf);
4134         else {
4135                 /*
4136                  * Kernel stack traversals can be done immediately,
4137                  * while we defer to an AST for user space traversals.
4138                  */
4139                 if (!inuserspace) {
4140                         callchaindepth =
4141                             pmc_save_kernel_callchain(ps->ps_pc,
4142                                 callchaindepth, tf);
4143                 } else {
4144                         pmc_post_callchain_callback();
4145                         callchaindepth = PMC_SAMPLE_INUSE;
4146                 }
4147         }
4148
4149         ps->ps_nsamples = callchaindepth;       /* mark entry as in use */
4150
4151         /* increment write pointer, modulo ring buffer size */
4152         ps++;
4153         if (ps == psb->ps_fence)
4154                 psb->ps_write = psb->ps_samples;
4155         else
4156                 psb->ps_write = ps;
4157
4158  done:
4159         /* mark CPU as needing processing */
4160         if (callchaindepth != PMC_SAMPLE_INUSE)
4161                 CPU_SET_ATOMIC(cpu, &pmc_cpumask);
4162
4163         return (error);
4164 }
4165
4166 /*
4167  * Capture a user call chain.  This function will be called from ast()
4168  * before control returns to userland and before the process gets
4169  * rescheduled.
4170  */
4171
4172 static void
4173 pmc_capture_user_callchain(int cpu, int ring, struct trapframe *tf)
4174 {
4175         struct pmc *pm;
4176         struct thread *td;
4177         struct pmc_sample *ps, *ps_end;
4178         struct pmc_samplebuffer *psb;
4179 #ifdef  INVARIANTS
4180         int ncallchains;
4181 #endif
4182
4183         psb = pmc_pcpu[cpu]->pc_sb[ring];
4184         td = curthread;
4185
4186         KASSERT(td->td_pflags & TDP_CALLCHAIN,
4187             ("[pmc,%d] Retrieving callchain for thread that doesn't want it",
4188                 __LINE__));
4189
4190 #ifdef  INVARIANTS
4191         ncallchains = 0;
4192 #endif
4193
4194         /*
4195          * Iterate through all deferred callchain requests.
4196          * Walk from the current read pointer to the current
4197          * write pointer.
4198          */
4199
4200         ps = psb->ps_read;
4201         ps_end = psb->ps_write;
4202         do {
4203                 if (ps->ps_nsamples != PMC_SAMPLE_INUSE)
4204                         goto next;
4205                 if (ps->ps_td != td)
4206                         goto next;
4207
4208                 KASSERT(ps->ps_cpu == cpu,
4209                     ("[pmc,%d] cpu mismatch ps_cpu=%d pcpu=%d", __LINE__,
4210                         ps->ps_cpu, PCPU_GET(cpuid)));
4211
4212                 pm = ps->ps_pmc;
4213
4214                 KASSERT(pm->pm_flags & PMC_F_CALLCHAIN,
4215                     ("[pmc,%d] Retrieving callchain for PMC that doesn't "
4216                         "want it", __LINE__));
4217
4218                 KASSERT(pm->pm_runcount > 0,
4219                     ("[pmc,%d] runcount %d", __LINE__, pm->pm_runcount));
4220
4221                 /*
4222                  * Retrieve the callchain and mark the sample buffer
4223                  * as 'processable' by the timer tick sweep code.
4224                  */
4225                 ps->ps_nsamples = pmc_save_user_callchain(ps->ps_pc,
4226                     pmc_callchaindepth, tf);
4227
4228 #ifdef  INVARIANTS
4229                 ncallchains++;
4230 #endif
4231
4232 next:
4233                 /* increment the pointer, modulo sample ring size */
4234                 if (++ps == psb->ps_fence)
4235                         ps = psb->ps_samples;
4236         } while (ps != ps_end);
4237
4238         KASSERT(ncallchains > 0,
4239             ("[pmc,%d] cpu %d didn't find a sample to collect", __LINE__,
4240                 cpu));
4241
4242         KASSERT(td->td_pinned == 1,
4243             ("[pmc,%d] invalid td_pinned value", __LINE__));
4244         sched_unpin();  /* Can migrate safely now. */
4245
4246         /* mark CPU as needing processing */
4247         CPU_SET_ATOMIC(cpu, &pmc_cpumask);
4248
4249         return;
4250 }
4251
4252 /*
4253  * Process saved PC samples.
4254  */
4255
4256 static void
4257 pmc_process_samples(int cpu, int ring)
4258 {
4259         struct pmc *pm;
4260         int adjri, n;
4261         struct thread *td;
4262         struct pmc_owner *po;
4263         struct pmc_sample *ps;
4264         struct pmc_classdep *pcd;
4265         struct pmc_samplebuffer *psb;
4266
4267         KASSERT(PCPU_GET(cpuid) == cpu,
4268             ("[pmc,%d] not on the correct CPU pcpu=%d cpu=%d", __LINE__,
4269                 PCPU_GET(cpuid), cpu));
4270
4271         psb = pmc_pcpu[cpu]->pc_sb[ring];
4272
4273         for (n = 0; n < pmc_nsamples; n++) { /* bound on #iterations */
4274
4275                 ps = psb->ps_read;
4276                 if (ps->ps_nsamples == PMC_SAMPLE_FREE)
4277                         break;
4278
4279                 pm = ps->ps_pmc;
4280
4281                 KASSERT(pm->pm_runcount > 0,
4282                     ("[pmc,%d] pm=%p runcount %d", __LINE__, (void *) pm,
4283                         pm->pm_runcount));
4284
4285                 po = pm->pm_owner;
4286
4287                 KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)),
4288                     ("[pmc,%d] pmc=%p non-sampling mode=%d", __LINE__,
4289                         pm, PMC_TO_MODE(pm)));
4290
4291                 /* Ignore PMCs that have been switched off */
4292                 if (pm->pm_state != PMC_STATE_RUNNING)
4293                         goto entrydone;
4294
4295                 /* If there is a pending AST wait for completion */
4296                 if (ps->ps_nsamples == PMC_SAMPLE_INUSE) {
4297                         /* Need a rescan at a later time. */
4298                         CPU_SET_ATOMIC(cpu, &pmc_cpumask);
4299                         break;
4300                 }
4301
4302                 PMCDBG6(SAM,OPS,1,"cpu=%d pm=%p n=%d fl=%x wr=%d rd=%d", cpu,
4303                     pm, ps->ps_nsamples, ps->ps_flags,
4304                     (int) (psb->ps_write - psb->ps_samples),
4305                     (int) (psb->ps_read - psb->ps_samples));
4306
4307                 /*
4308                  * If this is a process-mode PMC that is attached to
4309                  * its owner, and if the PC is in user mode, update
4310                  * profiling statistics like timer-based profiling
4311                  * would have done.
4312                  */
4313                 if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) {
4314                         if (ps->ps_flags & PMC_CC_F_USERSPACE) {
4315                                 td = FIRST_THREAD_IN_PROC(po->po_owner);
4316                                 addupc_intr(td, ps->ps_pc[0], 1);
4317                         }
4318                         goto entrydone;
4319                 }
4320
4321                 /*
4322                  * Otherwise, this is either a sampling mode PMC that
4323                  * is attached to a different process than its owner,
4324                  * or a system-wide sampling PMC.  Dispatch a log
4325                  * entry to the PMC's owner process.
4326                  */
4327                 pmclog_process_callchain(pm, ps);
4328
4329         entrydone:
4330                 ps->ps_nsamples = 0; /* mark entry as free */
4331                 atomic_subtract_rel_int(&pm->pm_runcount, 1);
4332
4333                 /* increment read pointer, modulo sample size */
4334                 if (++ps == psb->ps_fence)
4335                         psb->ps_read = psb->ps_samples;
4336                 else
4337                         psb->ps_read = ps;
4338         }
4339
4340         atomic_add_int(&pmc_stats.pm_log_sweeps, 1);
4341
4342         /* Do not re-enable stalled PMCs if we failed to process any samples */
4343         if (n == 0)
4344                 return;
4345
4346         /*
4347          * Restart any stalled sampling PMCs on this CPU.
4348          *
4349          * If the NMI handler sets the pm_stalled field of a PMC after
4350          * the check below, we'll end up processing the stalled PMC at
4351          * the next hardclock tick.
4352          */
4353         for (n = 0; n < md->pmd_npmc; n++) {
4354                 pcd = pmc_ri_to_classdep(md, n, &adjri);
4355                 KASSERT(pcd != NULL,
4356                     ("[pmc,%d] null pcd ri=%d", __LINE__, n));
4357                 (void) (*pcd->pcd_get_config)(cpu,adjri,&pm);
4358
4359                 if (pm == NULL ||                        /* !cfg'ed */
4360                     pm->pm_state != PMC_STATE_RUNNING || /* !active */
4361                     !PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) || /* !sampling */
4362                     !CPU_ISSET(cpu, &pm->pm_cpustate) || /* !desired */
4363                     !CPU_ISSET(cpu, &pm->pm_stalled)) /* !stalled */
4364                         continue;
4365
4366                 CPU_CLR_ATOMIC(cpu, &pm->pm_stalled);
4367                 (*pcd->pcd_start_pmc)(cpu, adjri);
4368         }
4369 }
4370
4371 /*
4372  * Event handlers.
4373  */
4374
4375 /*
4376  * Handle a process exit.
4377  *
4378  * Remove this process from all hash tables.  If this process
4379  * owned any PMCs, turn off those PMCs and deallocate them,
4380  * removing any associations with target processes.
4381  *
4382  * This function will be called by the last 'thread' of a
4383  * process.
4384  *
4385  * XXX This eventhandler gets called early in the exit process.
4386  * Consider using a 'hook' invocation from thread_exit() or equivalent
4387  * spot.  Another negative is that kse_exit doesn't seem to call
4388  * exit1() [??].
4389  *
4390  */
4391
4392 static void
4393 pmc_process_exit(void *arg __unused, struct proc *p)
4394 {
4395         struct pmc *pm;
4396         int adjri, cpu;
4397         unsigned int ri;
4398         int is_using_hwpmcs;
4399         struct pmc_owner *po;
4400         struct pmc_process *pp;
4401         struct pmc_classdep *pcd;
4402         pmc_value_t newvalue, tmp;
4403
4404         PROC_LOCK(p);
4405         is_using_hwpmcs = p->p_flag & P_HWPMC;
4406         PROC_UNLOCK(p);
4407
4408         /*
4409          * Log a sysexit event to all SS PMC owners.
4410          */
4411         LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
4412             if (po->po_flags & PMC_PO_OWNS_LOGFILE)
4413                     pmclog_process_sysexit(po, p->p_pid);
4414
4415         if (!is_using_hwpmcs)
4416                 return;
4417
4418         PMC_GET_SX_XLOCK();
4419         PMCDBG3(PRC,EXT,1,"process-exit proc=%p (%d, %s)", p, p->p_pid,
4420             p->p_comm);
4421
4422         /*
4423          * Since this code is invoked by the last thread in an exiting
4424          * process, we would have context switched IN at some prior
4425          * point.  However, with PREEMPTION, kernel mode context
4426          * switches may happen any time, so we want to disable a
4427          * context switch OUT till we get any PMCs targetting this
4428          * process off the hardware.
4429          *
4430          * We also need to atomically remove this process'
4431          * entry from our target process hash table, using
4432          * PMC_FLAG_REMOVE.
4433          */
4434         PMCDBG3(PRC,EXT,1, "process-exit proc=%p (%d, %s)", p, p->p_pid,
4435             p->p_comm);
4436
4437         critical_enter(); /* no preemption */
4438
4439         cpu = curthread->td_oncpu;
4440
4441         if ((pp = pmc_find_process_descriptor(p,
4442                  PMC_FLAG_REMOVE)) != NULL) {
4443
4444                 PMCDBG2(PRC,EXT,2,
4445                     "process-exit proc=%p pmc-process=%p", p, pp);
4446
4447                 /*
4448                  * The exiting process could the target of
4449                  * some PMCs which will be running on
4450                  * currently executing CPU.
4451                  *
4452                  * We need to turn these PMCs off like we
4453                  * would do at context switch OUT time.
4454                  */
4455                 for (ri = 0; ri < md->pmd_npmc; ri++) {
4456
4457                         /*
4458                          * Pick up the pmc pointer from hardware
4459                          * state similar to the CSW_OUT code.
4460                          */
4461                         pm = NULL;
4462
4463                         pcd = pmc_ri_to_classdep(md, ri, &adjri);
4464
4465                         (void) (*pcd->pcd_get_config)(cpu, adjri, &pm);
4466
4467                         PMCDBG2(PRC,EXT,2, "ri=%d pm=%p", ri, pm);
4468
4469                         if (pm == NULL ||
4470                             !PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)))
4471                                 continue;
4472
4473                         PMCDBG4(PRC,EXT,2, "ppmcs[%d]=%p pm=%p "
4474                             "state=%d", ri, pp->pp_pmcs[ri].pp_pmc,
4475                             pm, pm->pm_state);
4476
4477                         KASSERT(PMC_TO_ROWINDEX(pm) == ri,
4478                             ("[pmc,%d] ri mismatch pmc(%d) ri(%d)",
4479                                 __LINE__, PMC_TO_ROWINDEX(pm), ri));
4480
4481                         KASSERT(pm == pp->pp_pmcs[ri].pp_pmc,
4482                             ("[pmc,%d] pm %p != pp_pmcs[%d] %p",
4483                                 __LINE__, pm, ri, pp->pp_pmcs[ri].pp_pmc));
4484
4485                         KASSERT(pm->pm_runcount > 0,
4486                             ("[pmc,%d] bad runcount ri %d rc %d",
4487                                 __LINE__, ri, pm->pm_runcount));
4488
4489                         /*
4490                          * Change desired state, and then stop if not
4491                          * stalled. This two-step dance should avoid
4492                          * race conditions where an interrupt re-enables
4493                          * the PMC after this code has already checked
4494                          * the pm_stalled flag.
4495                          */
4496                         if (CPU_ISSET(cpu, &pm->pm_cpustate)) {
4497                                 CPU_CLR_ATOMIC(cpu, &pm->pm_cpustate);
4498                                 if (!CPU_ISSET(cpu, &pm->pm_stalled)) {
4499                                         (void) pcd->pcd_stop_pmc(cpu, adjri);
4500                                         pcd->pcd_read_pmc(cpu, adjri,
4501                                             &newvalue);
4502                                         tmp = newvalue -
4503                                             PMC_PCPU_SAVED(cpu,ri);
4504
4505                                         mtx_pool_lock_spin(pmc_mtxpool, pm);
4506                                         pm->pm_gv.pm_savedvalue += tmp;
4507                                         pp->pp_pmcs[ri].pp_pmcval += tmp;
4508                                         mtx_pool_unlock_spin(pmc_mtxpool, pm);
4509                                 }
4510                         }
4511
4512                         atomic_subtract_rel_int(&pm->pm_runcount,1);
4513
4514                         KASSERT((int) pm->pm_runcount >= 0,
4515                             ("[pmc,%d] runcount is %d", __LINE__, ri));
4516
4517                         (void) pcd->pcd_config_pmc(cpu, adjri, NULL);
4518                 }
4519
4520                 /*
4521                  * Inform the MD layer of this pseudo "context switch
4522                  * out"
4523                  */
4524                 (void) md->pmd_switch_out(pmc_pcpu[cpu], pp);
4525
4526                 critical_exit(); /* ok to be pre-empted now */
4527
4528                 /*
4529                  * Unlink this process from the PMCs that are
4530                  * targetting it.  This will send a signal to
4531                  * all PMC owner's whose PMCs are orphaned.
4532                  *
4533                  * Log PMC value at exit time if requested.
4534                  */
4535                 for (ri = 0; ri < md->pmd_npmc; ri++)
4536                         if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) {
4537                                 if (pm->pm_flags & PMC_F_NEEDS_LOGFILE &&
4538                                     PMC_IS_COUNTING_MODE(PMC_TO_MODE(pm)))
4539                                         pmclog_process_procexit(pm, pp);
4540                                 pmc_unlink_target_process(pm, pp);
4541                         }
4542                 free(pp, M_PMC);
4543
4544         } else
4545                 critical_exit(); /* pp == NULL */
4546
4547
4548         /*
4549          * If the process owned PMCs, free them up and free up
4550          * memory.
4551          */
4552         if ((po = pmc_find_owner_descriptor(p)) != NULL) {
4553                 pmc_remove_owner(po);
4554                 pmc_destroy_owner_descriptor(po);
4555         }
4556
4557         sx_xunlock(&pmc_sx);
4558 }
4559
4560 /*
4561  * Handle a process fork.
4562  *
4563  * If the parent process 'p1' is under HWPMC monitoring, then copy
4564  * over any attached PMCs that have 'do_descendants' semantics.
4565  */
4566
4567 static void
4568 pmc_process_fork(void *arg __unused, struct proc *p1, struct proc *newproc,
4569     int flags)
4570 {
4571         int is_using_hwpmcs;
4572         unsigned int ri;
4573         uint32_t do_descendants;
4574         struct pmc *pm;
4575         struct pmc_owner *po;
4576         struct pmc_process *ppnew, *ppold;
4577
4578         (void) flags;           /* unused parameter */
4579
4580         PROC_LOCK(p1);
4581         is_using_hwpmcs = p1->p_flag & P_HWPMC;
4582         PROC_UNLOCK(p1);
4583
4584         /*
4585          * If there are system-wide sampling PMCs active, we need to
4586          * log all fork events to their owner's logs.
4587          */
4588
4589         LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
4590             if (po->po_flags & PMC_PO_OWNS_LOGFILE)
4591                     pmclog_process_procfork(po, p1->p_pid, newproc->p_pid);
4592
4593         if (!is_using_hwpmcs)
4594                 return;
4595
4596         PMC_GET_SX_XLOCK();
4597         PMCDBG4(PMC,FRK,1, "process-fork proc=%p (%d, %s) -> %p", p1,
4598             p1->p_pid, p1->p_comm, newproc);
4599
4600         /*
4601          * If the parent process (curthread->td_proc) is a
4602          * target of any PMCs, look for PMCs that are to be
4603          * inherited, and link these into the new process
4604          * descriptor.
4605          */
4606         if ((ppold = pmc_find_process_descriptor(curthread->td_proc,
4607                  PMC_FLAG_NONE)) == NULL)
4608                 goto done;              /* nothing to do */
4609
4610         do_descendants = 0;
4611         for (ri = 0; ri < md->pmd_npmc; ri++)
4612                 if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL)
4613                         do_descendants |= pm->pm_flags & PMC_F_DESCENDANTS;
4614         if (do_descendants == 0) /* nothing to do */
4615                 goto done;
4616
4617         /* allocate a descriptor for the new process  */
4618         if ((ppnew = pmc_find_process_descriptor(newproc,
4619                  PMC_FLAG_ALLOCATE)) == NULL)
4620                 goto done;
4621
4622         /*
4623          * Run through all PMCs that were targeting the old process
4624          * and which specified F_DESCENDANTS and attach them to the
4625          * new process.
4626          *
4627          * Log the fork event to all owners of PMCs attached to this
4628          * process, if not already logged.
4629          */
4630         for (ri = 0; ri < md->pmd_npmc; ri++)
4631                 if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL &&
4632                     (pm->pm_flags & PMC_F_DESCENDANTS)) {
4633                         pmc_link_target_process(pm, ppnew);
4634                         po = pm->pm_owner;
4635                         if (po->po_sscount == 0 &&
4636                             po->po_flags & PMC_PO_OWNS_LOGFILE)
4637                                 pmclog_process_procfork(po, p1->p_pid,
4638                                     newproc->p_pid);
4639                 }
4640
4641         /*
4642          * Now mark the new process as being tracked by this driver.
4643          */
4644         PROC_LOCK(newproc);
4645         newproc->p_flag |= P_HWPMC;
4646         PROC_UNLOCK(newproc);
4647
4648  done:
4649         sx_xunlock(&pmc_sx);
4650 }
4651
4652 static void
4653 pmc_kld_load(void *arg __unused, linker_file_t lf)
4654 {
4655         struct pmc_owner *po;
4656
4657         sx_slock(&pmc_sx);
4658
4659         /*
4660          * Notify owners of system sampling PMCs about KLD operations.
4661          */
4662         LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
4663                 if (po->po_flags & PMC_PO_OWNS_LOGFILE)
4664                         pmclog_process_map_in(po, (pid_t) -1,
4665                             (uintfptr_t) lf->address, lf->filename);
4666
4667         /*
4668          * TODO: Notify owners of (all) process-sampling PMCs too.
4669          */
4670
4671         sx_sunlock(&pmc_sx);
4672 }
4673
4674 static void
4675 pmc_kld_unload(void *arg __unused, const char *filename __unused,
4676     caddr_t address, size_t size)
4677 {
4678         struct pmc_owner *po;
4679
4680         sx_slock(&pmc_sx);
4681
4682         LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
4683                 if (po->po_flags & PMC_PO_OWNS_LOGFILE)
4684                         pmclog_process_map_out(po, (pid_t) -1,
4685                             (uintfptr_t) address, (uintfptr_t) address + size);
4686
4687         /*
4688          * TODO: Notify owners of process-sampling PMCs.
4689          */
4690
4691         sx_sunlock(&pmc_sx);
4692 }
4693
4694 /*
4695  * initialization
4696  */
4697
4698 static const char *pmc_name_of_pmcclass[] = {
4699 #undef  __PMC_CLASS
4700 #define __PMC_CLASS(N) #N ,
4701         __PMC_CLASSES()
4702 };
4703
4704 /*
4705  * Base class initializer: allocate structure and set default classes.
4706  */
4707 struct pmc_mdep *
4708 pmc_mdep_alloc(int nclasses)
4709 {
4710         struct pmc_mdep *md;
4711         int     n;
4712
4713         /* SOFT + md classes */
4714         n = 1 + nclasses;
4715         md = malloc(sizeof(struct pmc_mdep) + n *
4716             sizeof(struct pmc_classdep), M_PMC, M_WAITOK|M_ZERO);
4717         md->pmd_nclass = n;
4718
4719         /* Add base class. */
4720         pmc_soft_initialize(md);
4721         return md;
4722 }
4723
4724 void
4725 pmc_mdep_free(struct pmc_mdep *md)
4726 {
4727         pmc_soft_finalize(md);
4728         free(md, M_PMC);
4729 }
4730
4731 static int
4732 generic_switch_in(struct pmc_cpu *pc, struct pmc_process *pp)
4733 {
4734         (void) pc; (void) pp;
4735
4736         return (0);
4737 }
4738
4739 static int
4740 generic_switch_out(struct pmc_cpu *pc, struct pmc_process *pp)
4741 {
4742         (void) pc; (void) pp;
4743
4744         return (0);
4745 }
4746
4747 static struct pmc_mdep *
4748 pmc_generic_cpu_initialize(void)
4749 {
4750         struct pmc_mdep *md;
4751
4752         md = pmc_mdep_alloc(0);
4753
4754         md->pmd_cputype    = PMC_CPU_GENERIC;
4755
4756         md->pmd_pcpu_init  = NULL;
4757         md->pmd_pcpu_fini  = NULL;
4758         md->pmd_switch_in  = generic_switch_in;
4759         md->pmd_switch_out = generic_switch_out;
4760
4761         return (md);
4762 }
4763
4764 static void
4765 pmc_generic_cpu_finalize(struct pmc_mdep *md)
4766 {
4767         (void) md;
4768 }
4769
4770
4771 static int
4772 pmc_initialize(void)
4773 {
4774         int c, cpu, error, n, ri;
4775         unsigned int maxcpu;
4776         struct pmc_binding pb;
4777         struct pmc_sample *ps;
4778         struct pmc_classdep *pcd;
4779         struct pmc_samplebuffer *sb;
4780
4781         md = NULL;
4782         error = 0;
4783
4784 #ifdef  HWPMC_DEBUG
4785         /* parse debug flags first */
4786         if (TUNABLE_STR_FETCH(PMC_SYSCTL_NAME_PREFIX "debugflags",
4787                 pmc_debugstr, sizeof(pmc_debugstr)))
4788                 pmc_debugflags_parse(pmc_debugstr,
4789                     pmc_debugstr+strlen(pmc_debugstr));
4790 #endif
4791
4792         PMCDBG1(MOD,INI,0, "PMC Initialize (version %x)", PMC_VERSION);
4793
4794         /* check kernel version */
4795         if (pmc_kernel_version != PMC_VERSION) {
4796                 if (pmc_kernel_version == 0)
4797                         printf("hwpmc: this kernel has not been compiled with "
4798                             "'options HWPMC_HOOKS'.\n");
4799                 else
4800                         printf("hwpmc: kernel version (0x%x) does not match "
4801                             "module version (0x%x).\n", pmc_kernel_version,
4802                             PMC_VERSION);
4803                 return EPROGMISMATCH;
4804         }
4805
4806         /*
4807          * check sysctl parameters
4808          */
4809
4810         if (pmc_hashsize <= 0) {
4811                 (void) printf("hwpmc: tunable \"hashsize\"=%d must be "
4812                     "greater than zero.\n", pmc_hashsize);
4813                 pmc_hashsize = PMC_HASH_SIZE;
4814         }
4815
4816         if (pmc_nsamples <= 0 || pmc_nsamples > 65535) {
4817                 (void) printf("hwpmc: tunable \"nsamples\"=%d out of "
4818                     "range.\n", pmc_nsamples);
4819                 pmc_nsamples = PMC_NSAMPLES;
4820         }
4821
4822         if (pmc_callchaindepth <= 0 ||
4823             pmc_callchaindepth > PMC_CALLCHAIN_DEPTH_MAX) {
4824                 (void) printf("hwpmc: tunable \"callchaindepth\"=%d out of "
4825                     "range - using %d.\n", pmc_callchaindepth,
4826                     PMC_CALLCHAIN_DEPTH_MAX);
4827                 pmc_callchaindepth = PMC_CALLCHAIN_DEPTH_MAX;
4828         }
4829
4830         md = pmc_md_initialize();
4831         if (md == NULL) {
4832                 /* Default to generic CPU. */
4833                 md = pmc_generic_cpu_initialize();
4834                 if (md == NULL)
4835                         return (ENOSYS);
4836         }
4837
4838         KASSERT(md->pmd_nclass >= 1 && md->pmd_npmc >= 1,
4839             ("[pmc,%d] no classes or pmcs", __LINE__));
4840
4841         /* Compute the map from row-indices to classdep pointers. */
4842         pmc_rowindex_to_classdep = malloc(sizeof(struct pmc_classdep *) *
4843             md->pmd_npmc, M_PMC, M_WAITOK|M_ZERO);
4844
4845         for (n = 0; n < md->pmd_npmc; n++)
4846                 pmc_rowindex_to_classdep[n] = NULL;
4847         for (ri = c = 0; c < md->pmd_nclass; c++) {
4848                 pcd = &md->pmd_classdep[c];
4849                 for (n = 0; n < pcd->pcd_num; n++, ri++)
4850                         pmc_rowindex_to_classdep[ri] = pcd;
4851         }
4852
4853         KASSERT(ri == md->pmd_npmc,
4854             ("[pmc,%d] npmc miscomputed: ri=%d, md->npmc=%d", __LINE__,
4855             ri, md->pmd_npmc));
4856
4857         maxcpu = pmc_cpu_max();
4858
4859         /* allocate space for the per-cpu array */
4860         pmc_pcpu = malloc(maxcpu * sizeof(struct pmc_cpu *), M_PMC,
4861             M_WAITOK|M_ZERO);
4862
4863         /* per-cpu 'saved values' for managing process-mode PMCs */
4864         pmc_pcpu_saved = malloc(sizeof(pmc_value_t) * maxcpu * md->pmd_npmc,
4865             M_PMC, M_WAITOK);
4866
4867         /* Perform CPU-dependent initialization. */
4868         pmc_save_cpu_binding(&pb);
4869         error = 0;
4870         for (cpu = 0; error == 0 && cpu < maxcpu; cpu++) {
4871                 if (!pmc_cpu_is_active(cpu))
4872                         continue;
4873                 pmc_select_cpu(cpu);
4874                 pmc_pcpu[cpu] = malloc(sizeof(struct pmc_cpu) +
4875                     md->pmd_npmc * sizeof(struct pmc_hw *), M_PMC,
4876                     M_WAITOK|M_ZERO);
4877                 if (md->pmd_pcpu_init)
4878                         error = md->pmd_pcpu_init(md, cpu);
4879                 for (n = 0; error == 0 && n < md->pmd_nclass; n++)
4880                         error = md->pmd_classdep[n].pcd_pcpu_init(md, cpu);
4881         }
4882         pmc_restore_cpu_binding(&pb);
4883
4884         if (error)
4885                 return (error);
4886
4887         /* allocate space for the sample array */
4888         for (cpu = 0; cpu < maxcpu; cpu++) {
4889                 if (!pmc_cpu_is_active(cpu))
4890                         continue;
4891
4892                 sb = malloc(sizeof(struct pmc_samplebuffer) +
4893                     pmc_nsamples * sizeof(struct pmc_sample), M_PMC,
4894                     M_WAITOK|M_ZERO);
4895                 sb->ps_read = sb->ps_write = sb->ps_samples;
4896                 sb->ps_fence = sb->ps_samples + pmc_nsamples;
4897
4898                 KASSERT(pmc_pcpu[cpu] != NULL,
4899                     ("[pmc,%d] cpu=%d Null per-cpu data", __LINE__, cpu));
4900
4901                 sb->ps_callchains = malloc(pmc_callchaindepth * pmc_nsamples *
4902                     sizeof(uintptr_t), M_PMC, M_WAITOK|M_ZERO);
4903
4904                 for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++)
4905                         ps->ps_pc = sb->ps_callchains +
4906                             (n * pmc_callchaindepth);
4907
4908                 pmc_pcpu[cpu]->pc_sb[PMC_HR] = sb;
4909
4910                 sb = malloc(sizeof(struct pmc_samplebuffer) +
4911                     pmc_nsamples * sizeof(struct pmc_sample), M_PMC,
4912                     M_WAITOK|M_ZERO);
4913                 sb->ps_read = sb->ps_write = sb->ps_samples;
4914                 sb->ps_fence = sb->ps_samples + pmc_nsamples;
4915
4916                 KASSERT(pmc_pcpu[cpu] != NULL,
4917                     ("[pmc,%d] cpu=%d Null per-cpu data", __LINE__, cpu));
4918
4919                 sb->ps_callchains = malloc(pmc_callchaindepth * pmc_nsamples *
4920                     sizeof(uintptr_t), M_PMC, M_WAITOK|M_ZERO);
4921
4922                 for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++)
4923                         ps->ps_pc = sb->ps_callchains +
4924                             (n * pmc_callchaindepth);
4925
4926                 pmc_pcpu[cpu]->pc_sb[PMC_SR] = sb;
4927         }
4928
4929         /* allocate space for the row disposition array */
4930         pmc_pmcdisp = malloc(sizeof(enum pmc_mode) * md->pmd_npmc,
4931             M_PMC, M_WAITOK|M_ZERO);
4932
4933         /* mark all PMCs as available */
4934         for (n = 0; n < (int) md->pmd_npmc; n++)
4935                 PMC_MARK_ROW_FREE(n);
4936
4937         /* allocate thread hash tables */
4938         pmc_ownerhash = hashinit(pmc_hashsize, M_PMC,
4939             &pmc_ownerhashmask);
4940
4941         pmc_processhash = hashinit(pmc_hashsize, M_PMC,
4942             &pmc_processhashmask);
4943         mtx_init(&pmc_processhash_mtx, "pmc-process-hash", "pmc-leaf",
4944             MTX_SPIN);
4945
4946         LIST_INIT(&pmc_ss_owners);
4947         pmc_ss_count = 0;
4948
4949         /* allocate a pool of spin mutexes */
4950         pmc_mtxpool = mtx_pool_create("pmc-leaf", pmc_mtxpool_size,
4951             MTX_SPIN);
4952
4953         PMCDBG4(MOD,INI,1, "pmc_ownerhash=%p, mask=0x%lx "
4954             "targethash=%p mask=0x%lx", pmc_ownerhash, pmc_ownerhashmask,
4955             pmc_processhash, pmc_processhashmask);
4956
4957         /* register process {exit,fork,exec} handlers */
4958         pmc_exit_tag = EVENTHANDLER_REGISTER(process_exit,
4959             pmc_process_exit, NULL, EVENTHANDLER_PRI_ANY);
4960         pmc_fork_tag = EVENTHANDLER_REGISTER(process_fork,
4961             pmc_process_fork, NULL, EVENTHANDLER_PRI_ANY);
4962
4963         /* register kld event handlers */
4964         pmc_kld_load_tag = EVENTHANDLER_REGISTER(kld_load, pmc_kld_load,
4965             NULL, EVENTHANDLER_PRI_ANY);
4966         pmc_kld_unload_tag = EVENTHANDLER_REGISTER(kld_unload, pmc_kld_unload,
4967             NULL, EVENTHANDLER_PRI_ANY);
4968
4969         /* initialize logging */
4970         pmclog_initialize();
4971
4972         /* set hook functions */
4973         pmc_intr = md->pmd_intr;
4974         pmc_hook = pmc_hook_handler;
4975
4976         if (error == 0) {
4977                 printf(PMC_MODULE_NAME ":");
4978                 for (n = 0; n < (int) md->pmd_nclass; n++) {
4979                         pcd = &md->pmd_classdep[n];
4980                         printf(" %s/%d/%d/0x%b",
4981                             pmc_name_of_pmcclass[pcd->pcd_class],
4982                             pcd->pcd_num,
4983                             pcd->pcd_width,
4984                             pcd->pcd_caps,
4985                             "\20"
4986                             "\1INT\2USR\3SYS\4EDG\5THR"
4987                             "\6REA\7WRI\10INV\11QUA\12PRC"
4988                             "\13TAG\14CSC");
4989                 }
4990                 printf("\n");
4991         }
4992
4993         return (error);
4994 }
4995
4996 /* prepare to be unloaded */
4997 static void
4998 pmc_cleanup(void)
4999 {
5000         int c, cpu;
5001         unsigned int maxcpu;
5002         struct pmc_ownerhash *ph;
5003         struct pmc_owner *po, *tmp;
5004         struct pmc_binding pb;
5005 #ifdef  HWPMC_DEBUG
5006         struct pmc_processhash *prh;
5007 #endif
5008
5009         PMCDBG0(MOD,INI,0, "cleanup");
5010
5011         /* switch off sampling */
5012         CPU_ZERO(&pmc_cpumask);
5013         pmc_intr = NULL;
5014
5015         sx_xlock(&pmc_sx);
5016         if (pmc_hook == NULL) { /* being unloaded already */
5017                 sx_xunlock(&pmc_sx);
5018                 return;
5019         }
5020
5021         pmc_hook = NULL; /* prevent new threads from entering module */
5022
5023         /* deregister event handlers */
5024         EVENTHANDLER_DEREGISTER(process_fork, pmc_fork_tag);
5025         EVENTHANDLER_DEREGISTER(process_exit, pmc_exit_tag);
5026         EVENTHANDLER_DEREGISTER(kld_load, pmc_kld_load_tag);
5027         EVENTHANDLER_DEREGISTER(kld_unload, pmc_kld_unload_tag);
5028
5029         /* send SIGBUS to all owner threads, free up allocations */
5030         if (pmc_ownerhash)
5031                 for (ph = pmc_ownerhash;
5032                      ph <= &pmc_ownerhash[pmc_ownerhashmask];
5033                      ph++) {
5034                         LIST_FOREACH_SAFE(po, ph, po_next, tmp) {
5035                                 pmc_remove_owner(po);
5036
5037                                 /* send SIGBUS to owner processes */
5038                                 PMCDBG3(MOD,INI,2, "cleanup signal proc=%p "
5039                                     "(%d, %s)", po->po_owner,
5040                                     po->po_owner->p_pid,
5041                                     po->po_owner->p_comm);
5042
5043                                 PROC_LOCK(po->po_owner);
5044                                 kern_psignal(po->po_owner, SIGBUS);
5045                                 PROC_UNLOCK(po->po_owner);
5046
5047                                 pmc_destroy_owner_descriptor(po);
5048                         }
5049                 }
5050
5051         /* reclaim allocated data structures */
5052         if (pmc_mtxpool)
5053                 mtx_pool_destroy(&pmc_mtxpool);
5054
5055         mtx_destroy(&pmc_processhash_mtx);
5056         if (pmc_processhash) {
5057 #ifdef  HWPMC_DEBUG
5058                 struct pmc_process *pp;
5059
5060                 PMCDBG0(MOD,INI,3, "destroy process hash");
5061                 for (prh = pmc_processhash;
5062                      prh <= &pmc_processhash[pmc_processhashmask];
5063                      prh++)
5064                         LIST_FOREACH(pp, prh, pp_next)
5065                             PMCDBG1(MOD,INI,3, "pid=%d", pp->pp_proc->p_pid);
5066 #endif
5067
5068                 hashdestroy(pmc_processhash, M_PMC, pmc_processhashmask);
5069                 pmc_processhash = NULL;
5070         }
5071
5072         if (pmc_ownerhash) {
5073                 PMCDBG0(MOD,INI,3, "destroy owner hash");
5074                 hashdestroy(pmc_ownerhash, M_PMC, pmc_ownerhashmask);
5075                 pmc_ownerhash = NULL;
5076         }
5077
5078         KASSERT(LIST_EMPTY(&pmc_ss_owners),
5079             ("[pmc,%d] Global SS owner list not empty", __LINE__));
5080         KASSERT(pmc_ss_count == 0,
5081             ("[pmc,%d] Global SS count not empty", __LINE__));
5082
5083         /* do processor and pmc-class dependent cleanup */
5084         maxcpu = pmc_cpu_max();
5085
5086         PMCDBG0(MOD,INI,3, "md cleanup");
5087         if (md) {
5088                 pmc_save_cpu_binding(&pb);
5089                 for (cpu = 0; cpu < maxcpu; cpu++) {
5090                         PMCDBG2(MOD,INI,1,"pmc-cleanup cpu=%d pcs=%p",
5091                             cpu, pmc_pcpu[cpu]);
5092                         if (!pmc_cpu_is_active(cpu) || pmc_pcpu[cpu] == NULL)
5093                                 continue;
5094                         pmc_select_cpu(cpu);
5095                         for (c = 0; c < md->pmd_nclass; c++)
5096                                 md->pmd_classdep[c].pcd_pcpu_fini(md, cpu);
5097                         if (md->pmd_pcpu_fini)
5098                                 md->pmd_pcpu_fini(md, cpu);
5099                 }
5100
5101                 if (md->pmd_cputype == PMC_CPU_GENERIC)
5102                         pmc_generic_cpu_finalize(md);
5103                 else
5104                         pmc_md_finalize(md);
5105
5106                 pmc_mdep_free(md);
5107                 md = NULL;
5108                 pmc_restore_cpu_binding(&pb);
5109         }
5110
5111         /* Free per-cpu descriptors. */
5112         for (cpu = 0; cpu < maxcpu; cpu++) {
5113                 if (!pmc_cpu_is_active(cpu))
5114                         continue;
5115                 KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_HR] != NULL,
5116                     ("[pmc,%d] Null hw cpu sample buffer cpu=%d", __LINE__,
5117                         cpu));
5118                 KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_SR] != NULL,
5119                     ("[pmc,%d] Null sw cpu sample buffer cpu=%d", __LINE__,
5120                         cpu));
5121                 free(pmc_pcpu[cpu]->pc_sb[PMC_HR]->ps_callchains, M_PMC);
5122                 free(pmc_pcpu[cpu]->pc_sb[PMC_HR], M_PMC);
5123                 free(pmc_pcpu[cpu]->pc_sb[PMC_SR]->ps_callchains, M_PMC);
5124                 free(pmc_pcpu[cpu]->pc_sb[PMC_SR], M_PMC);
5125                 free(pmc_pcpu[cpu], M_PMC);
5126         }
5127
5128         free(pmc_pcpu, M_PMC);
5129         pmc_pcpu = NULL;
5130
5131         free(pmc_pcpu_saved, M_PMC);
5132         pmc_pcpu_saved = NULL;
5133
5134         if (pmc_pmcdisp) {
5135                 free(pmc_pmcdisp, M_PMC);
5136                 pmc_pmcdisp = NULL;
5137         }
5138
5139         if (pmc_rowindex_to_classdep) {
5140                 free(pmc_rowindex_to_classdep, M_PMC);
5141                 pmc_rowindex_to_classdep = NULL;
5142         }
5143
5144         pmclog_shutdown();
5145
5146         sx_xunlock(&pmc_sx);    /* we are done */
5147 }
5148
5149 /*
5150  * The function called at load/unload.
5151  */
5152
5153 static int
5154 load (struct module *module __unused, int cmd, void *arg __unused)
5155 {
5156         int error;
5157
5158         error = 0;
5159
5160         switch (cmd) {
5161         case MOD_LOAD :
5162                 /* initialize the subsystem */
5163                 error = pmc_initialize();
5164                 if (error != 0)
5165                         break;
5166                 PMCDBG2(MOD,INI,1, "syscall=%d maxcpu=%d",
5167                     pmc_syscall_num, pmc_cpu_max());
5168                 break;
5169
5170
5171         case MOD_UNLOAD :
5172         case MOD_SHUTDOWN:
5173                 pmc_cleanup();
5174                 PMCDBG0(MOD,INI,1, "unloaded");
5175                 break;
5176
5177         default :
5178                 error = EINVAL; /* XXX should panic(9) */
5179                 break;
5180         }
5181
5182         return error;
5183 }
5184
5185 /* memory pool */
5186 MALLOC_DEFINE(M_PMC, "pmc", "Memory space for the PMC module");