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 SYSCTL_INT(_kern_hwpmc, OID_AUTO, callchaindepth, CTLFLAG_RDTUN,
 239     &pmc_callchaindepth, 0, "depth of call chain records");
 240
 241 #ifdef  HWPMC_DEBUG
 242 struct pmc_debugflags pmc_debugflags = PMC_DEBUG_DEFAULT_FLAGS;
 243 char    pmc_debugstr[PMC_DEBUG_STRSIZE];
 244 TUNABLE_STR(PMC_SYSCTL_NAME_PREFIX "debugflags", pmc_debugstr,
 245     sizeof(pmc_debugstr));
 246 SYSCTL_PROC(_kern_hwpmc, OID_AUTO, debugflags,
 247     CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NOFETCH,
 248     0, 0, pmc_debugflags_sysctl_handler, "A", "debug flags");
 249 #endif
 250
 251 /*
 252  * kern.hwpmc.hashrows -- determines the number of rows in the
 253  * of the hash table used to look up threads
 254  */
 255
 256 static int pmc_hashsize = PMC_HASH_SIZE;
 257 SYSCTL_INT(_kern_hwpmc, OID_AUTO, hashsize, CTLFLAG_RDTUN,
 258     &pmc_hashsize, 0, "rows in hash tables");
 259
 260 /*
 261  * kern.hwpmc.nsamples --- number of PC samples/callchain stacks per CPU
 262  */
 263
 264 static int pmc_nsamples = PMC_NSAMPLES;
 265 SYSCTL_INT(_kern_hwpmc, OID_AUTO, nsamples, CTLFLAG_RDTUN,
 266     &pmc_nsamples, 0, "number of PC samples per CPU");
 267
 268
 269 /*
 270  * kern.hwpmc.mtxpoolsize -- number of mutexes in the mutex pool.
 271  */
 272
 273 static int pmc_mtxpool_size = PMC_MTXPOOL_SIZE;
 274 SYSCTL_INT(_kern_hwpmc, OID_AUTO, mtxpoolsize, CTLFLAG_RDTUN,
 275     &pmc_mtxpool_size, 0, "size of spin mutex pool");
 276
 277
 278 /*
 279  * security.bsd.unprivileged_syspmcs -- allow non-root processes to
 280  * allocate system-wide PMCs.
 281  *
 282  * Allowing unprivileged processes to allocate system PMCs is convenient
 283  * if system-wide measurements need to be taken concurrently with other
 284  * per-process measurements.  This feature is turned off by default.
 285  */
 286
 287 static int pmc_unprivileged_syspmcs = 0;
 288 SYSCTL_INT(_security_bsd, OID_AUTO, unprivileged_syspmcs, CTLFLAG_RWTUN,
 289     &pmc_unprivileged_syspmcs, 0,
 290     "allow unprivileged process to allocate system PMCs");
 291
 292 /*
 293  * Hash function.  Discard the lower 2 bits of the pointer since
 294  * these are always zero for our uses.  The hash multiplier is
 295  * round((2^LONG_BIT) * ((sqrt(5)-1)/2)).
 296  */
 297
 298 #if     LONG_BIT == 64
 299 #define _PMC_HM         11400714819323198486u
 300 #elif   LONG_BIT == 32
 301 #define _PMC_HM         2654435769u
 302 #else
 303 #error  Must know the size of 'long' to compile
 304 #endif
 305
 306 #define PMC_HASH_PTR(P,M)       ((((unsigned long) (P) >> 2) * _PMC_HM) & (M))
 307
 308 /*
 309  * Syscall structures
 310  */
 311
 312 /* The `sysent' for the new syscall */
 313 static struct sysent pmc_sysent = {
 314         2,                      /* sy_narg */
 315         pmc_syscall_handler     /* sy_call */
 316 };
 317
 318 static struct syscall_module_data pmc_syscall_mod = {
 319         load,
 320         NULL,
 321         &pmc_syscall_num,
 322         &pmc_sysent,
 323 #if (__FreeBSD_version >= 1100000)
 324         { 0, NULL },
 325         SY_THR_STATIC_KLD,
 326 #else
 327         { 0, NULL }
 328 #endif
 329 };
 330
 331 static moduledata_t pmc_mod = {
 332         PMC_MODULE_NAME,
 333         syscall_module_handler,
 334         &pmc_syscall_mod
 335 };
 336
 337 DECLARE_MODULE(pmc, pmc_mod, SI_SUB_SMP, SI_ORDER_ANY);
 338 MODULE_VERSION(pmc, PMC_VERSION);
 339
 340 #ifdef  HWPMC_DEBUG
 341 enum pmc_dbgparse_state {
 342         PMCDS_WS,               /* in whitespace */
 343         PMCDS_MAJOR,            /* seen a major keyword */
 344         PMCDS_MINOR
 345 };
 346
 347 static int
 348 pmc_debugflags_parse(char *newstr, char *fence)
 349 {
 350         char c, *p, *q;
 351         struct pmc_debugflags *tmpflags;
 352         int error, found, *newbits, tmp;
 353         size_t kwlen;
 354
 355         tmpflags = malloc(sizeof(*tmpflags), M_PMC, M_WAITOK|M_ZERO);
 356
 357         p = newstr;
 358         error = 0;
 359
 360         for (; p < fence && (c = *p); p++) {
 361
 362                 /* skip white space */
 363                 if (c == ' ' || c == '\t')
 364                         continue;
 365
 366                 /* look for a keyword followed by "=" */
 367                 for (q = p; p < fence && (c = *p) && c != '='; p++)
 368                         ;
 369                 if (c != '=') {
 370                         error = EINVAL;
 371                         goto done;
 372                 }
 373
 374                 kwlen = p - q;
 375                 newbits = NULL;
 376
 377                 /* lookup flag group name */
 378 #define DBG_SET_FLAG_MAJ(S,F)                                           \
 379                 if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0)  \
 380                         newbits = &tmpflags->pdb_ ## F;
 381
 382                 DBG_SET_FLAG_MAJ("cpu",         CPU);
 383                 DBG_SET_FLAG_MAJ("csw",         CSW);
 384                 DBG_SET_FLAG_MAJ("logging",     LOG);
 385                 DBG_SET_FLAG_MAJ("module",      MOD);
 386                 DBG_SET_FLAG_MAJ("md",          MDP);
 387                 DBG_SET_FLAG_MAJ("owner",       OWN);
 388                 DBG_SET_FLAG_MAJ("pmc",         PMC);
 389                 DBG_SET_FLAG_MAJ("process",     PRC);
 390                 DBG_SET_FLAG_MAJ("sampling",    SAM);
 391
 392                 if (newbits == NULL) {
 393                         error = EINVAL;
 394                         goto done;
 395                 }
 396
 397                 p++;            /* skip the '=' */
 398
 399                 /* Now parse the individual flags */
 400                 tmp = 0;
 401         newflag:
 402                 for (q = p; p < fence && (c = *p); p++)
 403                         if (c == ' ' || c == '\t' || c == ',')
 404                                 break;
 405
 406                 /* p == fence or c == ws or c == "," or c == 0 */
 407
 408                 if ((kwlen = p - q) == 0) {
 409                         *newbits = tmp;
 410                         continue;
 411                 }
 412
 413                 found = 0;
 414 #define DBG_SET_FLAG_MIN(S,F)                                           \
 415                 if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0)  \
 416                         tmp |= found = (1 << PMC_DEBUG_MIN_ ## F)
 417
 418                 /* a '*' denotes all possible flags in the group */
 419                 if (kwlen == 1 && *q == '*')
 420                         tmp = found = ~0;
 421                 /* look for individual flag names */
 422                 DBG_SET_FLAG_MIN("allocaterow", ALR);
 423                 DBG_SET_FLAG_MIN("allocate",    ALL);
 424                 DBG_SET_FLAG_MIN("attach",      ATT);
 425                 DBG_SET_FLAG_MIN("bind",        BND);
 426                 DBG_SET_FLAG_MIN("config",      CFG);
 427                 DBG_SET_FLAG_MIN("exec",        EXC);
 428                 DBG_SET_FLAG_MIN("exit",        EXT);
 429                 DBG_SET_FLAG_MIN("find",        FND);
 430                 DBG_SET_FLAG_MIN("flush",       FLS);
 431                 DBG_SET_FLAG_MIN("fork",        FRK);
 432                 DBG_SET_FLAG_MIN("getbuf",      GTB);
 433                 DBG_SET_FLAG_MIN("hook",        PMH);
 434                 DBG_SET_FLAG_MIN("init",        INI);
 435                 DBG_SET_FLAG_MIN("intr",        INT);
 436                 DBG_SET_FLAG_MIN("linktarget",  TLK);
 437                 DBG_SET_FLAG_MIN("mayberemove", OMR);
 438                 DBG_SET_FLAG_MIN("ops",         OPS);
 439                 DBG_SET_FLAG_MIN("read",        REA);
 440                 DBG_SET_FLAG_MIN("register",    REG);
 441                 DBG_SET_FLAG_MIN("release",     REL);
 442                 DBG_SET_FLAG_MIN("remove",      ORM);
 443                 DBG_SET_FLAG_MIN("sample",      SAM);
 444                 DBG_SET_FLAG_MIN("scheduleio",  SIO);
 445                 DBG_SET_FLAG_MIN("select",      SEL);
 446                 DBG_SET_FLAG_MIN("signal",      SIG);
 447                 DBG_SET_FLAG_MIN("swi",         SWI);
 448                 DBG_SET_FLAG_MIN("swo",         SWO);
 449                 DBG_SET_FLAG_MIN("start",       STA);
 450                 DBG_SET_FLAG_MIN("stop",        STO);
 451                 DBG_SET_FLAG_MIN("syscall",     PMS);
 452                 DBG_SET_FLAG_MIN("unlinktarget", TUL);
 453                 DBG_SET_FLAG_MIN("write",       WRI);
 454                 if (found == 0) {
 455                         /* unrecognized flag name */
 456                         error = EINVAL;
 457                         goto done;
 458                 }
 459
 460                 if (c == 0 || c == ' ' || c == '\t') {  /* end of flag group */
 461                         *newbits = tmp;
 462                         continue;
 463                 }
 464
 465                 p++;
 466                 goto newflag;
 467         }
 468
 469         /* save the new flag set */
 470         bcopy(tmpflags, &pmc_debugflags, sizeof(pmc_debugflags));
 471
 472  done:
 473         free(tmpflags, M_PMC);
 474         return error;
 475 }
 476
 477 static int
 478 pmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS)
 479 {
 480         char *fence, *newstr;
 481         int error;
 482         unsigned int n;
 483
 484         (void) arg1; (void) arg2; /* unused parameters */
 485
 486         n = sizeof(pmc_debugstr);
 487         newstr = malloc(n, M_PMC, M_WAITOK|M_ZERO);
 488         (void) strlcpy(newstr, pmc_debugstr, n);
 489
 490         error = sysctl_handle_string(oidp, newstr, n, req);
 491
 492         /* if there is a new string, parse and copy it */
 493         if (error == 0 && req->newptr != NULL) {
 494                 fence = newstr + (n < req->newlen ? n : req->newlen + 1);
 495                 if ((error = pmc_debugflags_parse(newstr, fence)) == 0)
 496                         (void) strlcpy(pmc_debugstr, newstr,
 497                             sizeof(pmc_debugstr));
 498         }
 499
 500         free(newstr, M_PMC);
 501
 502         return error;
 503 }
 504 #endif
 505
 506 /*
 507  * Map a row index to a classdep structure and return the adjusted row
 508  * index for the PMC class index.
 509  */
 510 static struct pmc_classdep *
 511 pmc_ri_to_classdep(struct pmc_mdep *md, int ri, int *adjri)
 512 {
 513         struct pmc_classdep *pcd;
 514
 515         (void) md;
 516
 517         KASSERT(ri >= 0 && ri < md->pmd_npmc,
 518             ("[pmc,%d] illegal row-index %d", __LINE__, ri));
 519
 520         pcd = pmc_rowindex_to_classdep[ri];
 521
 522         KASSERT(pcd != NULL,
 523             ("[pmc,%d] ri %d null pcd", __LINE__, ri));
 524
 525         *adjri = ri - pcd->pcd_ri;
 526
 527         KASSERT(*adjri >= 0 && *adjri < pcd->pcd_num,
 528             ("[pmc,%d] adjusted row-index %d", __LINE__, *adjri));
 529
 530         return (pcd);
 531 }
 532
 533 /*
 534  * Concurrency Control
 535  *
 536  * The driver manages the following data structures:
 537  *
 538  *   - target process descriptors, one per target process
 539  *   - owner process descriptors (and attached lists), one per owner process
 540  *   - lookup hash tables for owner and target processes
 541  *   - PMC descriptors (and attached lists)
 542  *   - per-cpu hardware state
 543  *   - the 'hook' variable through which the kernel calls into
 544  *     this module
 545  *   - the machine hardware state (managed by the MD layer)
 546  *
 547  * These data structures are accessed from:
 548  *
 549  * - thread context-switch code
 550  * - interrupt handlers (possibly on multiple cpus)
 551  * - kernel threads on multiple cpus running on behalf of user
 552  *   processes doing system calls
 553  * - this driver's private kernel threads
 554  *
 555  * = Locks and Locking strategy =
 556  *
 557  * The driver uses four locking strategies for its operation:
 558  *
 559  * - The global SX lock "pmc_sx" is used to protect internal
 560  *   data structures.
 561  *
 562  *   Calls into the module by syscall() start with this lock being
 563  *   held in exclusive mode.  Depending on the requested operation,
 564  *   the lock may be downgraded to 'shared' mode to allow more
 565  *   concurrent readers into the module.  Calls into the module from
 566  *   other parts of the kernel acquire the lock in shared mode.
 567  *
 568  *   This SX lock is held in exclusive mode for any operations that
 569  *   modify the linkages between the driver's internal data structures.
 570  *
 571  *   The 'pmc_hook' function pointer is also protected by this lock.
 572  *   It is only examined with the sx lock held in exclusive mode.  The
 573  *   kernel module is allowed to be unloaded only with the sx lock held
 574  *   in exclusive mode.  In normal syscall handling, after acquiring the
 575  *   pmc_sx lock we first check that 'pmc_hook' is non-null before
 576  *   proceeding.  This prevents races between the thread unloading the module
 577  *   and other threads seeking to use the module.
 578  *
 579  * - Lookups of target process structures and owner process structures
 580  *   cannot use the global "pmc_sx" SX lock because these lookups need
 581  *   to happen during context switches and in other critical sections
 582  *   where sleeping is not allowed.  We protect these lookup tables
 583  *   with their own private spin-mutexes, "pmc_processhash_mtx" and
 584  *   "pmc_ownerhash_mtx".
 585  *
 586  * - Interrupt handlers work in a lock free manner.  At interrupt
 587  *   time, handlers look at the PMC pointer (phw->phw_pmc) configured
 588  *   when the PMC was started.  If this pointer is NULL, the interrupt
 589  *   is ignored after updating driver statistics.  We ensure that this
 590  *   pointer is set (using an atomic operation if necessary) before the
 591  *   PMC hardware is started.  Conversely, this pointer is unset atomically
 592  *   only after the PMC hardware is stopped.
 593  *
 594  *   We ensure that everything needed for the operation of an
 595  *   interrupt handler is available without it needing to acquire any
 596  *   locks.  We also ensure that a PMC's software state is destroyed only
 597  *   after the PMC is taken off hardware (on all CPUs).
 598  *
 599  * - Context-switch handling with process-private PMCs needs more
 600  *   care.
 601  *
 602  *   A given process may be the target of multiple PMCs.  For example,
 603  *   PMCATTACH and PMCDETACH may be requested by a process on one CPU
 604  *   while the target process is running on another.  A PMC could also
 605  *   be getting released because its owner is exiting.  We tackle
 606  *   these situations in the following manner:
 607  *
 608  *   - each target process structure 'pmc_process' has an array
 609  *     of 'struct pmc *' pointers, one for each hardware PMC.
 610  *
 611  *   - At context switch IN time, each "target" PMC in RUNNING state
 612  *     gets started on hardware and a pointer to each PMC is copied into
 613  *     the per-cpu phw array.  The 'runcount' for the PMC is
 614  *     incremented.
 615  *
 616  *   - At context switch OUT time, all process-virtual PMCs are stopped
 617  *     on hardware.  The saved value is added to the PMCs value field
 618  *     only if the PMC is in a non-deleted state (the PMCs state could
 619  *     have changed during the current time slice).
 620  *
 621  *     Note that since in-between a switch IN on a processor and a switch
 622  *     OUT, the PMC could have been released on another CPU.  Therefore
 623  *     context switch OUT always looks at the hardware state to turn
 624  *     OFF PMCs and will update a PMC's saved value only if reachable
 625  *     from the target process record.
 626  *
 627  *   - OP PMCRELEASE could be called on a PMC at any time (the PMC could
 628  *     be attached to many processes at the time of the call and could
 629  *     be active on multiple CPUs).
 630  *
 631  *     We prevent further scheduling of the PMC by marking it as in
 632  *     state 'DELETED'.  If the runcount of the PMC is non-zero then
 633  *     this PMC is currently running on a CPU somewhere.  The thread
 634  *     doing the PMCRELEASE operation waits by repeatedly doing a
 635  *     pause() till the runcount comes to zero.
 636  *
 637  * The contents of a PMC descriptor (struct pmc) are protected using
 638  * a spin-mutex.  In order to save space, we use a mutex pool.
 639  *
 640  * In terms of lock types used by witness(4), we use:
 641  * - Type "pmc-sx", used by the global SX lock.
 642  * - Type "pmc-sleep", for sleep mutexes used by logger threads.
 643  * - Type "pmc-per-proc", for protecting PMC owner descriptors.
 644  * - Type "pmc-leaf", used for all other spin mutexes.
 645  */
 646
 647 /*
 648  * save the cpu binding of the current kthread
 649  */
 650
 651 static void
 652 pmc_save_cpu_binding(struct pmc_binding *pb)
 653 {
 654         PMCDBG0(CPU,BND,2, "save-cpu");
 655         thread_lock(curthread);
 656         pb->pb_bound = sched_is_bound(curthread);
 657         pb->pb_cpu   = curthread->td_oncpu;
 658         thread_unlock(curthread);
 659         PMCDBG1(CPU,BND,2, "save-cpu cpu=%d", pb->pb_cpu);
 660 }
 661
 662 /*
 663  * restore the cpu binding of the current thread
 664  */
 665
 666 static void
 667 pmc_restore_cpu_binding(struct pmc_binding *pb)
 668 {
 669         PMCDBG2(CPU,BND,2, "restore-cpu curcpu=%d restore=%d",
 670             curthread->td_oncpu, pb->pb_cpu);
 671         thread_lock(curthread);
 672         if (pb->pb_bound)
 673                 sched_bind(curthread, pb->pb_cpu);
 674         else
 675                 sched_unbind(curthread);
 676         thread_unlock(curthread);
 677         PMCDBG0(CPU,BND,2, "restore-cpu done");
 678 }
 679
 680 /*
 681  * move execution over the specified cpu and bind it there.
 682  */
 683
 684 static void
 685 pmc_select_cpu(int cpu)
 686 {
 687         KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
 688             ("[pmc,%d] bad cpu number %d", __LINE__, cpu));
 689
 690         /* Never move to an inactive CPU. */
 691         KASSERT(pmc_cpu_is_active(cpu), ("[pmc,%d] selecting inactive "
 692             "CPU %d", __LINE__, cpu));
 693
 694         PMCDBG1(CPU,SEL,2, "select-cpu cpu=%d", cpu);
 695         thread_lock(curthread);
 696         sched_bind(curthread, cpu);
 697         thread_unlock(curthread);
 698
 699         KASSERT(curthread->td_oncpu == cpu,
 700             ("[pmc,%d] CPU not bound [cpu=%d, curr=%d]", __LINE__,
 701                 cpu, curthread->td_oncpu));
 702
 703         PMCDBG1(CPU,SEL,2, "select-cpu cpu=%d ok", cpu);
 704 }
 705
 706 /*
 707  * Force a context switch.
 708  *
 709  * We do this by pause'ing for 1 tick -- invoking mi_switch() is not
 710  * guaranteed to force a context switch.
 711  */
 712
 713 static void
 714 pmc_force_context_switch(void)
 715 {
 716
 717         pause("pmcctx", 1);
 718 }
 719
 720 /*
 721  * Get the file name for an executable.  This is a simple wrapper
 722  * around vn_fullpath(9).
 723  */
 724
 725 static void
 726 pmc_getfilename(struct vnode *v, char **fullpath, char **freepath)
 727 {
 728
 729         *fullpath = "unknown";
 730         *freepath = NULL;
 731         vn_fullpath(curthread, v, fullpath, freepath);
 732 }
 733
 734 /*
 735  * remove an process owning PMCs
 736  */
 737
 738 void
 739 pmc_remove_owner(struct pmc_owner *po)
 740 {
 741         struct pmc *pm, *tmp;
 742
 743         sx_assert(&pmc_sx, SX_XLOCKED);
 744
 745         PMCDBG1(OWN,ORM,1, "remove-owner po=%p", po);
 746
 747         /* Remove descriptor from the owner hash table */
 748         LIST_REMOVE(po, po_next);
 749
 750         /* release all owned PMC descriptors */
 751         LIST_FOREACH_SAFE(pm, &po->po_pmcs, pm_next, tmp) {
 752                 PMCDBG1(OWN,ORM,2, "pmc=%p", pm);
 753                 KASSERT(pm->pm_owner == po,
 754                     ("[pmc,%d] owner %p != po %p", __LINE__, pm->pm_owner, po));
 755
 756                 pmc_release_pmc_descriptor(pm); /* will unlink from the list */
 757                 pmc_destroy_pmc_descriptor(pm);
 758         }
 759
 760         KASSERT(po->po_sscount == 0,
 761             ("[pmc,%d] SS count not zero", __LINE__));
 762         KASSERT(LIST_EMPTY(&po->po_pmcs),
 763             ("[pmc,%d] PMC list not empty", __LINE__));
 764
 765         /* de-configure the log file if present */
 766         if (po->po_flags & PMC_PO_OWNS_LOGFILE)
 767                 pmclog_deconfigure_log(po);
 768 }
 769
 770 /*
 771  * remove an owner process record if all conditions are met.
 772  */
 773
 774 static void
 775 pmc_maybe_remove_owner(struct pmc_owner *po)
 776 {
 777
 778         PMCDBG1(OWN,OMR,1, "maybe-remove-owner po=%p", po);
 779
 780         /*
 781          * Remove owner record if
 782          * - this process does not own any PMCs
 783          * - this process has not allocated a system-wide sampling buffer
 784          */
 785
 786         if (LIST_EMPTY(&po->po_pmcs) &&
 787             ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)) {
 788                 pmc_remove_owner(po);
 789                 pmc_destroy_owner_descriptor(po);
 790         }
 791 }
 792
 793 /*
 794  * Add an association between a target process and a PMC.
 795  */
 796
 797 static void
 798 pmc_link_target_process(struct pmc *pm, struct pmc_process *pp)
 799 {
 800         int ri;
 801         struct pmc_target *pt;
 802
 803         sx_assert(&pmc_sx, SX_XLOCKED);
 804
 805         KASSERT(pm != NULL && pp != NULL,
 806             ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp));
 807         KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)),
 808             ("[pmc,%d] Attaching a non-process-virtual pmc=%p to pid=%d",
 809                 __LINE__, pm, pp->pp_proc->p_pid));
 810         KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= ((int) md->pmd_npmc - 1),
 811             ("[pmc,%d] Illegal reference count %d for process record %p",
 812                 __LINE__, pp->pp_refcnt, (void *) pp));
 813
 814         ri = PMC_TO_ROWINDEX(pm);
 815
 816         PMCDBG3(PRC,TLK,1, "link-target pmc=%p ri=%d pmc-process=%p",
 817             pm, ri, pp);
 818
 819 #ifdef  HWPMC_DEBUG
 820         LIST_FOREACH(pt, &pm->pm_targets, pt_next)
 821             if (pt->pt_process == pp)
 822                     KASSERT(0, ("[pmc,%d] pp %p already in pmc %p targets",
 823                                 __LINE__, pp, pm));
 824 #endif
 825
 826         pt = malloc(sizeof(struct pmc_target), M_PMC, M_WAITOK|M_ZERO);
 827         pt->pt_process = pp;
 828
 829         LIST_INSERT_HEAD(&pm->pm_targets, pt, pt_next);
 830
 831         atomic_store_rel_ptr((uintptr_t *)&pp->pp_pmcs[ri].pp_pmc,
 832             (uintptr_t)pm);
 833
 834         if (pm->pm_owner->po_owner == pp->pp_proc)
 835                 pm->pm_flags |= PMC_F_ATTACHED_TO_OWNER;
 836
 837         /*
 838          * Initialize the per-process values at this row index.
 839          */
 840         pp->pp_pmcs[ri].pp_pmcval = PMC_TO_MODE(pm) == PMC_MODE_TS ?
 841             pm->pm_sc.pm_reloadcount : 0;
 842
 843         pp->pp_refcnt++;
 844
 845 }
 846
 847 /*
 848  * Removes the association between a target process and a PMC.
 849  */
 850
 851 static void
 852 pmc_unlink_target_process(struct pmc *pm, struct pmc_process *pp)
 853 {
 854         int ri;
 855         struct proc *p;
 856         struct pmc_target *ptgt;
 857
 858         sx_assert(&pmc_sx, SX_XLOCKED);
 859
 860         KASSERT(pm != NULL && pp != NULL,
 861             ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp));
 862
 863         KASSERT(pp->pp_refcnt >= 1 && pp->pp_refcnt <= (int) md->pmd_npmc,
 864             ("[pmc,%d] Illegal ref count %d on process record %p",
 865                 __LINE__, pp->pp_refcnt, (void *) pp));
 866
 867         ri = PMC_TO_ROWINDEX(pm);
 868
 869         PMCDBG3(PRC,TUL,1, "unlink-target pmc=%p ri=%d pmc-process=%p",
 870             pm, ri, pp);
 871
 872         KASSERT(pp->pp_pmcs[ri].pp_pmc == pm,
 873             ("[pmc,%d] PMC ri %d mismatch pmc %p pp->[ri] %p", __LINE__,
 874                 ri, pm, pp->pp_pmcs[ri].pp_pmc));
 875
 876         pp->pp_pmcs[ri].pp_pmc = NULL;
 877         pp->pp_pmcs[ri].pp_pmcval = (pmc_value_t) 0;
 878
 879         /* Remove owner-specific flags */
 880         if (pm->pm_owner->po_owner == pp->pp_proc) {
 881                 pp->pp_flags &= ~PMC_PP_ENABLE_MSR_ACCESS;
 882                 pm->pm_flags &= ~PMC_F_ATTACHED_TO_OWNER;
 883         }
 884
 885         pp->pp_refcnt--;
 886
 887         /* Remove the target process from the PMC structure */
 888         LIST_FOREACH(ptgt, &pm->pm_targets, pt_next)
 889                 if (ptgt->pt_process == pp)
 890                         break;
 891
 892         KASSERT(ptgt != NULL, ("[pmc,%d] process %p (pp: %p) not found "
 893                     "in pmc %p", __LINE__, pp->pp_proc, pp, pm));
 894
 895         LIST_REMOVE(ptgt, pt_next);
 896         free(ptgt, M_PMC);
 897
 898         /* if the PMC now lacks targets, send the owner a SIGIO */
 899         if (LIST_EMPTY(&pm->pm_targets)) {
 900                 p = pm->pm_owner->po_owner;
 901                 PROC_LOCK(p);
 902                 kern_psignal(p, SIGIO);
 903                 PROC_UNLOCK(p);
 904
 905                 PMCDBG2(PRC,SIG,2, "signalling proc=%p signal=%d", p,
 906                     SIGIO);
 907         }
 908 }
 909
 910 /*
 911  * Check if PMC 'pm' may be attached to target process 't'.
 912  */
 913
 914 static int
 915 pmc_can_attach(struct pmc *pm, struct proc *t)
 916 {
 917         struct proc *o;         /* pmc owner */
 918         struct ucred *oc, *tc;  /* owner, target credentials */
 919         int decline_attach, i;
 920
 921         /*
 922          * A PMC's owner can always attach that PMC to itself.
 923          */
 924
 925         if ((o = pm->pm_owner->po_owner) == t)
 926                 return 0;
 927
 928         PROC_LOCK(o);
 929         oc = o->p_ucred;
 930         crhold(oc);
 931         PROC_UNLOCK(o);
 932
 933         PROC_LOCK(t);
 934         tc = t->p_ucred;
 935         crhold(tc);
 936         PROC_UNLOCK(t);
 937
 938         /*
 939          * The effective uid of the PMC owner should match at least one
 940          * of the {effective,real,saved} uids of the target process.
 941          */
 942
 943         decline_attach = oc->cr_uid != tc->cr_uid &&
 944             oc->cr_uid != tc->cr_svuid &&
 945             oc->cr_uid != tc->cr_ruid;
 946
 947         /*
 948          * Every one of the target's group ids, must be in the owner's
 949          * group list.
 950          */
 951         for (i = 0; !decline_attach && i < tc->cr_ngroups; i++)
 952                 decline_attach = !groupmember(tc->cr_groups[i], oc);
 953
 954         /* check the read and saved gids too */
 955         if (decline_attach == 0)
 956                 decline_attach = !groupmember(tc->cr_rgid, oc) ||
 957                     !groupmember(tc->cr_svgid, oc);
 958
 959         crfree(tc);
 960         crfree(oc);
 961
 962         return !decline_attach;
 963 }
 964
 965 /*
 966  * Attach a process to a PMC.
 967  */
 968
 969 static int
 970 pmc_attach_one_process(struct proc *p, struct pmc *pm)
 971 {
 972         int ri;
 973         char *fullpath, *freepath;
 974         struct pmc_process      *pp;
 975
 976         sx_assert(&pmc_sx, SX_XLOCKED);
 977
 978         PMCDBG5(PRC,ATT,2, "attach-one pm=%p ri=%d proc=%p (%d, %s)", pm,
 979             PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm);
 980
 981         /*
 982          * Locate the process descriptor corresponding to process 'p',
 983          * allocating space as needed.
 984          *
 985          * Verify that rowindex 'pm_rowindex' is free in the process
 986          * descriptor.
 987          *
 988          * If not, allocate space for a descriptor and link the
 989          * process descriptor and PMC.
 990          */
 991         ri = PMC_TO_ROWINDEX(pm);
 992
 993         if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_ALLOCATE)) == NULL)
 994                 return ENOMEM;
 995
 996         if (pp->pp_pmcs[ri].pp_pmc == pm) /* already present at slot [ri] */
 997                 return EEXIST;
 998
 999         if (pp->pp_pmcs[ri].pp_pmc != NULL)
1000                 return EBUSY;
1001
1002         pmc_link_target_process(pm, pp);
1003
1004         if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) &&
1005             (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) == 0)
1006                 pm->pm_flags |= PMC_F_NEEDS_LOGFILE;
1007
1008         pm->pm_flags |= PMC_F_ATTACH_DONE; /* mark as attached */
1009
1010         /* issue an attach event to a configured log file */
1011         if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE) {
1012                 pmc_getfilename(p->p_textvp, &fullpath, &freepath);
1013                 if (p->p_flag & P_KTHREAD) {
1014                         fullpath = kernelname;
1015                         freepath = NULL;
1016                 } else
1017                         pmclog_process_pmcattach(pm, p->p_pid, fullpath);
1018                 if (freepath)
1019                         free(freepath, M_TEMP);
1020                 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1021                         pmc_log_process_mappings(pm->pm_owner, p);
1022         }
1023         /* mark process as using HWPMCs */
1024         PROC_LOCK(p);
1025         p->p_flag |= P_HWPMC;
1026         PROC_UNLOCK(p);
1027
1028         return 0;
1029 }
1030
1031 /*
1032  * Attach a process and optionally its children
1033  */
1034
1035 static int
1036 pmc_attach_process(struct proc *p, struct pmc *pm)
1037 {
1038         int error;
1039         struct proc *top;
1040
1041         sx_assert(&pmc_sx, SX_XLOCKED);
1042
1043         PMCDBG5(PRC,ATT,1, "attach pm=%p ri=%d proc=%p (%d, %s)", pm,
1044             PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm);
1045
1046
1047         /*
1048          * If this PMC successfully allowed a GETMSR operation
1049          * in the past, disallow further ATTACHes.
1050          */
1051
1052         if ((pm->pm_flags & PMC_PP_ENABLE_MSR_ACCESS) != 0)
1053                 return EPERM;
1054
1055         if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0)
1056                 return pmc_attach_one_process(p, pm);
1057
1058         /*
1059          * Traverse all child processes, attaching them to
1060          * this PMC.
1061          */
1062
1063         sx_slock(&proctree_lock);
1064
1065         top = p;
1066
1067         for (;;) {
1068                 if ((error = pmc_attach_one_process(p, pm)) != 0)
1069                         break;
1070                 if (!LIST_EMPTY(&p->p_children))
1071                         p = LIST_FIRST(&p->p_children);
1072                 else for (;;) {
1073                         if (p == top)
1074                                 goto done;
1075                         if (LIST_NEXT(p, p_sibling)) {
1076                                 p = LIST_NEXT(p, p_sibling);
1077                                 break;
1078                         }
1079                         p = p->p_pptr;
1080                 }
1081         }
1082
1083         if (error)
1084                 (void) pmc_detach_process(top, pm);
1085
1086  done:
1087         sx_sunlock(&proctree_lock);
1088         return error;
1089 }
1090
1091 /*
1092  * Detach a process from a PMC.  If there are no other PMCs tracking
1093  * this process, remove the process structure from its hash table.  If
1094  * 'flags' contains PMC_FLAG_REMOVE, then free the process structure.
1095  */
1096
1097 static int
1098 pmc_detach_one_process(struct proc *p, struct pmc *pm, int flags)
1099 {
1100         int ri;
1101         struct pmc_process *pp;
1102
1103         sx_assert(&pmc_sx, SX_XLOCKED);
1104
1105         KASSERT(pm != NULL,
1106             ("[pmc,%d] null pm pointer", __LINE__));
1107
1108         ri = PMC_TO_ROWINDEX(pm);
1109
1110         PMCDBG6(PRC,ATT,2, "detach-one pm=%p ri=%d proc=%p (%d, %s) flags=0x%x",
1111             pm, ri, p, p->p_pid, p->p_comm, flags);
1112
1113         if ((pp = pmc_find_process_descriptor(p, 0)) == NULL)
1114                 return ESRCH;
1115
1116         if (pp->pp_pmcs[ri].pp_pmc != pm)
1117                 return EINVAL;
1118
1119         pmc_unlink_target_process(pm, pp);
1120
1121         /* Issue a detach entry if a log file is configured */
1122         if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE)
1123                 pmclog_process_pmcdetach(pm, p->p_pid);
1124
1125         /*
1126          * If there are no PMCs targetting this process, we remove its
1127          * descriptor from the target hash table and unset the P_HWPMC
1128          * flag in the struct proc.
1129          */
1130         KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= (int) md->pmd_npmc,
1131             ("[pmc,%d] Illegal refcnt %d for process struct %p",
1132                 __LINE__, pp->pp_refcnt, pp));
1133
1134         if (pp->pp_refcnt != 0) /* still a target of some PMC */
1135                 return 0;
1136
1137         pmc_remove_process_descriptor(pp);
1138
1139         if (flags & PMC_FLAG_REMOVE)
1140                 free(pp, M_PMC);
1141
1142         PROC_LOCK(p);
1143         p->p_flag &= ~P_HWPMC;
1144         PROC_UNLOCK(p);
1145
1146         return 0;
1147 }
1148
1149 /*
1150  * Detach a process and optionally its descendants from a PMC.
1151  */
1152
1153 static int
1154 pmc_detach_process(struct proc *p, struct pmc *pm)
1155 {
1156         struct proc *top;
1157
1158         sx_assert(&pmc_sx, SX_XLOCKED);
1159
1160         PMCDBG5(PRC,ATT,1, "detach pm=%p ri=%d proc=%p (%d, %s)", pm,
1161             PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm);
1162
1163         if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0)
1164                 return pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE);
1165
1166         /*
1167          * Traverse all children, detaching them from this PMC.  We
1168          * ignore errors since we could be detaching a PMC from a
1169          * partially attached proc tree.
1170          */
1171
1172         sx_slock(&proctree_lock);
1173
1174         top = p;
1175
1176         for (;;) {
1177                 (void) pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE);
1178
1179                 if (!LIST_EMPTY(&p->p_children))
1180                         p = LIST_FIRST(&p->p_children);
1181                 else for (;;) {
1182                         if (p == top)
1183                                 goto done;
1184                         if (LIST_NEXT(p, p_sibling)) {
1185                                 p = LIST_NEXT(p, p_sibling);
1186                                 break;
1187                         }
1188                         p = p->p_pptr;
1189                 }
1190         }
1191
1192  done:
1193         sx_sunlock(&proctree_lock);
1194
1195         if (LIST_EMPTY(&pm->pm_targets))
1196                 pm->pm_flags &= ~PMC_F_ATTACH_DONE;
1197
1198         return 0;
1199 }
1200
1201
1202 /*
1203  * Thread context switch IN
1204  */
1205
1206 static void
1207 pmc_process_csw_in(struct thread *td)
1208 {
1209         int cpu;
1210         unsigned int adjri, ri;
1211         struct pmc *pm;
1212         struct proc *p;
1213         struct pmc_cpu *pc;
1214         struct pmc_hw *phw;
1215         pmc_value_t newvalue;
1216         struct pmc_process *pp;
1217         struct pmc_classdep *pcd;
1218
1219         p = td->td_proc;
1220
1221         if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE)) == NULL)
1222                 return;
1223
1224         KASSERT(pp->pp_proc == td->td_proc,
1225             ("[pmc,%d] not my thread state", __LINE__));
1226
1227         critical_enter(); /* no preemption from this point */
1228
1229         cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */
1230
1231         PMCDBG5(CSW,SWI,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p,
1232             p->p_pid, p->p_comm, pp);
1233
1234         KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
1235             ("[pmc,%d] wierd CPU id %d", __LINE__, cpu));
1236
1237         pc = pmc_pcpu[cpu];
1238
1239         for (ri = 0; ri < md->pmd_npmc; ri++) {
1240
1241                 if ((pm = pp->pp_pmcs[ri].pp_pmc) == NULL)
1242                         continue;
1243
1244                 KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)),
1245                     ("[pmc,%d] Target PMC in non-virtual mode (%d)",
1246                         __LINE__, PMC_TO_MODE(pm)));
1247
1248                 KASSERT(PMC_TO_ROWINDEX(pm) == ri,
1249                     ("[pmc,%d] Row index mismatch pmc %d != ri %d",
1250                         __LINE__, PMC_TO_ROWINDEX(pm), ri));
1251
1252                 /*
1253                  * Only PMCs that are marked as 'RUNNING' need
1254                  * be placed on hardware.
1255                  */
1256
1257                 if (pm->pm_state != PMC_STATE_RUNNING)
1258                         continue;
1259
1260                 /* increment PMC runcount */
1261                 atomic_add_rel_int(&pm->pm_runcount, 1);
1262
1263                 /* configure the HWPMC we are going to use. */
1264                 pcd = pmc_ri_to_classdep(md, ri, &adjri);
1265                 pcd->pcd_config_pmc(cpu, adjri, pm);
1266
1267                 phw = pc->pc_hwpmcs[ri];
1268
1269                 KASSERT(phw != NULL,
1270                     ("[pmc,%d] null hw pointer", __LINE__));
1271
1272                 KASSERT(phw->phw_pmc == pm,
1273                     ("[pmc,%d] hw->pmc %p != pmc %p", __LINE__,
1274                         phw->phw_pmc, pm));
1275
1276                 /*
1277                  * Write out saved value and start the PMC.
1278                  *
1279                  * Sampling PMCs use a per-process value, while
1280                  * counting mode PMCs use a per-pmc value that is
1281                  * inherited across descendants.
1282                  */
1283                 if (PMC_TO_MODE(pm) == PMC_MODE_TS) {
1284                         mtx_pool_lock_spin(pmc_mtxpool, pm);
1285
1286                         /*
1287                          * Use the saved value calculated after the most recent
1288                          * thread switch out to start this counter.  Reset
1289                          * the saved count in case another thread from this
1290                          * process switches in before any threads switch out.
1291                          */
1292                         newvalue = PMC_PCPU_SAVED(cpu,ri) =
1293                             pp->pp_pmcs[ri].pp_pmcval;
1294                         pp->pp_pmcs[ri].pp_pmcval = pm->pm_sc.pm_reloadcount;
1295                         mtx_pool_unlock_spin(pmc_mtxpool, pm);
1296                 } else {
1297                         KASSERT(PMC_TO_MODE(pm) == PMC_MODE_TC,
1298                             ("[pmc,%d] illegal mode=%d", __LINE__,
1299                             PMC_TO_MODE(pm)));
1300                         mtx_pool_lock_spin(pmc_mtxpool, pm);
1301                         newvalue = PMC_PCPU_SAVED(cpu, ri) =
1302                             pm->pm_gv.pm_savedvalue;
1303                         mtx_pool_unlock_spin(pmc_mtxpool, pm);
1304                 }
1305
1306                 PMCDBG3(CSW,SWI,1,"cpu=%d ri=%d new=%jd", cpu, ri, newvalue);
1307
1308                 pcd->pcd_write_pmc(cpu, adjri, newvalue);
1309
1310                 /* If a sampling mode PMC, reset stalled state. */
1311                 if (PMC_TO_MODE(pm) == PMC_MODE_TS)
1312                         CPU_CLR_ATOMIC(cpu, &pm->pm_stalled);
1313
1314                 /* Indicate that we desire this to run. */
1315                 CPU_SET_ATOMIC(cpu, &pm->pm_cpustate);
1316
1317                 /* Start the PMC. */
1318                 pcd->pcd_start_pmc(cpu, adjri);
1319         }
1320
1321         /*
1322          * perform any other architecture/cpu dependent thread
1323          * switch-in actions.
1324          */
1325
1326         (void) (*md->pmd_switch_in)(pc, pp);
1327
1328         critical_exit();
1329
1330 }
1331
1332 /*
1333  * Thread context switch OUT.
1334  */
1335
1336 static void
1337 pmc_process_csw_out(struct thread *td)
1338 {
1339         int cpu;
1340         int64_t tmp;
1341         struct pmc *pm;
1342         struct proc *p;
1343         enum pmc_mode mode;
1344         struct pmc_cpu *pc;
1345         pmc_value_t newvalue;
1346         unsigned int adjri, ri;
1347         struct pmc_process *pp;
1348         struct pmc_classdep *pcd;
1349
1350
1351         /*
1352          * Locate our process descriptor; this may be NULL if
1353          * this process is exiting and we have already removed
1354          * the process from the target process table.
1355          *
1356          * Note that due to kernel preemption, multiple
1357          * context switches may happen while the process is
1358          * exiting.
1359          *
1360          * Note also that if the target process cannot be
1361          * found we still need to deconfigure any PMCs that
1362          * are currently running on hardware.
1363          */
1364
1365         p = td->td_proc;
1366         pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE);
1367
1368         /*
1369          * save PMCs
1370          */
1371
1372         critical_enter();
1373
1374         cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */
1375
1376         PMCDBG5(CSW,SWO,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p,
1377             p->p_pid, p->p_comm, pp);
1378
1379         KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
1380             ("[pmc,%d wierd CPU id %d", __LINE__, cpu));
1381
1382         pc = pmc_pcpu[cpu];
1383
1384         /*
1385          * When a PMC gets unlinked from a target PMC, it will
1386          * be removed from the target's pp_pmc[] array.
1387          *
1388          * However, on a MP system, the target could have been
1389          * executing on another CPU at the time of the unlink.
1390          * So, at context switch OUT time, we need to look at
1391          * the hardware to determine if a PMC is scheduled on
1392          * it.
1393          */
1394
1395         for (ri = 0; ri < md->pmd_npmc; ri++) {
1396
1397                 pcd = pmc_ri_to_classdep(md, ri, &adjri);
1398                 pm  = NULL;
1399                 (void) (*pcd->pcd_get_config)(cpu, adjri, &pm);
1400
1401                 if (pm == NULL) /* nothing at this row index */
1402                         continue;
1403
1404                 mode = PMC_TO_MODE(pm);
1405                 if (!PMC_IS_VIRTUAL_MODE(mode))
1406                         continue; /* not a process virtual PMC */
1407
1408                 KASSERT(PMC_TO_ROWINDEX(pm) == ri,
1409                     ("[pmc,%d] ri mismatch pmc(%d) ri(%d)",
1410                         __LINE__, PMC_TO_ROWINDEX(pm), ri));
1411
1412                 /*
1413                  * Change desired state, and then stop if not stalled.
1414                  * This two-step dance should avoid race conditions where
1415                  * an interrupt re-enables the PMC after this code has
1416                  * already checked the pm_stalled flag.
1417                  */
1418                 CPU_CLR_ATOMIC(cpu, &pm->pm_cpustate);
1419                 if (!CPU_ISSET(cpu, &pm->pm_stalled))
1420                         pcd->pcd_stop_pmc(cpu, adjri);
1421
1422                 /* reduce this PMC's runcount */
1423                 atomic_subtract_rel_int(&pm->pm_runcount, 1);
1424
1425                 /*
1426                  * If this PMC is associated with this process,
1427                  * save the reading.
1428                  */
1429
1430                 if (pp != NULL && pp->pp_pmcs[ri].pp_pmc != NULL) {
1431
1432                         KASSERT(pm == pp->pp_pmcs[ri].pp_pmc,
1433                             ("[pmc,%d] pm %p != pp_pmcs[%d] %p", __LINE__,
1434                                 pm, ri, pp->pp_pmcs[ri].pp_pmc));
1435
1436                         KASSERT(pp->pp_refcnt > 0,
1437                             ("[pmc,%d] pp refcnt = %d", __LINE__,
1438                                 pp->pp_refcnt));
1439
1440                         pcd->pcd_read_pmc(cpu, adjri, &newvalue);
1441
1442                         if (mode == PMC_MODE_TS) {
1443                                 PMCDBG3(CSW,SWO,1,"cpu=%d ri=%d tmp=%jd (samp)",
1444                                     cpu, ri, PMC_PCPU_SAVED(cpu,ri) - newvalue);
1445
1446                                 /*
1447                                  * For sampling process-virtual PMCs,
1448                                  * newvalue is the number of events to be seen
1449                                  * until the next sampling interrupt.
1450                                  * We can just add the events left from this
1451                                  * invocation to the counter, then adjust
1452                                  * in case we overflow our range.
1453                                  *
1454                                  * (Recall that we reload the counter every
1455                                  * time we use it.)
1456                                  */
1457                                 mtx_pool_lock_spin(pmc_mtxpool, pm);
1458
1459                                 pp->pp_pmcs[ri].pp_pmcval += newvalue;
1460                                 if (pp->pp_pmcs[ri].pp_pmcval >
1461                                     pm->pm_sc.pm_reloadcount)
1462                                         pp->pp_pmcs[ri].pp_pmcval -=
1463                                             pm->pm_sc.pm_reloadcount;
1464                                 KASSERT(pp->pp_pmcs[ri].pp_pmcval > 0 &&
1465                                     pp->pp_pmcs[ri].pp_pmcval <=
1466                                     pm->pm_sc.pm_reloadcount,
1467                                     ("[pmc,%d] pp_pmcval outside of expected "
1468                                     "range cpu=%d ri=%d pp_pmcval=%jx "
1469                                     "pm_reloadcount=%jx", __LINE__, cpu, ri,
1470                                     pp->pp_pmcs[ri].pp_pmcval,
1471                                     pm->pm_sc.pm_reloadcount));
1472                                 mtx_pool_unlock_spin(pmc_mtxpool, pm);
1473
1474                         } else {
1475                                 tmp = newvalue - PMC_PCPU_SAVED(cpu,ri);
1476
1477                                 PMCDBG3(CSW,SWO,1,"cpu=%d ri=%d tmp=%jd (count)",
1478                                     cpu, ri, tmp);
1479
1480                                 /*
1481                                  * For counting process-virtual PMCs,
1482                                  * we expect the count to be
1483                                  * increasing monotonically, modulo a 64
1484                                  * bit wraparound.
1485                                  */
1486                                 KASSERT((int64_t) tmp >= 0,
1487                                     ("[pmc,%d] negative increment cpu=%d "
1488                                      "ri=%d newvalue=%jx saved=%jx "
1489                                      "incr=%jx", __LINE__, cpu, ri,
1490                                      newvalue, PMC_PCPU_SAVED(cpu,ri), tmp));
1491
1492                                 mtx_pool_lock_spin(pmc_mtxpool, pm);
1493                                 pm->pm_gv.pm_savedvalue += tmp;
1494                                 pp->pp_pmcs[ri].pp_pmcval += tmp;
1495                                 mtx_pool_unlock_spin(pmc_mtxpool, pm);
1496
1497                                 if (pm->pm_flags & PMC_F_LOG_PROCCSW)
1498                                         pmclog_process_proccsw(pm, pp, tmp);
1499                         }
1500                 }
1501
1502                 /* mark hardware as free */
1503                 pcd->pcd_config_pmc(cpu, adjri, NULL);
1504         }
1505
1506         /*
1507          * perform any other architecture/cpu dependent thread
1508          * switch out functions.
1509          */
1510
1511         (void) (*md->pmd_switch_out)(pc, pp);
1512
1513         critical_exit();
1514 }
1515
1516 /*
1517  * A mapping change for a process.
1518  */
1519
1520 static void
1521 pmc_process_mmap(struct thread *td, struct pmckern_map_in *pkm)
1522 {
1523         int ri;
1524         pid_t pid;
1525         char *fullpath, *freepath;
1526         const struct pmc *pm;
1527         struct pmc_owner *po;
1528         const struct pmc_process *pp;
1529
1530         freepath = fullpath = NULL;
1531         pmc_getfilename((struct vnode *) pkm->pm_file, &fullpath, &freepath);
1532
1533         pid = td->td_proc->p_pid;
1534
1535         /* Inform owners of all system-wide sampling PMCs. */
1536         LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1537             if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1538                 pmclog_process_map_in(po, pid, pkm->pm_address, fullpath);
1539
1540         if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL)
1541                 goto done;
1542
1543         /*
1544          * Inform sampling PMC owners tracking this process.
1545          */
1546         for (ri = 0; ri < md->pmd_npmc; ri++)
1547                 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL &&
1548                     PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1549                         pmclog_process_map_in(pm->pm_owner,
1550                             pid, pkm->pm_address, fullpath);
1551
1552   done:
1553         if (freepath)
1554                 free(freepath, M_TEMP);
1555 }
1556
1557
1558 /*
1559  * Log an munmap request.
1560  */
1561
1562 static void
1563 pmc_process_munmap(struct thread *td, struct pmckern_map_out *pkm)
1564 {
1565         int ri;
1566         pid_t pid;
1567         struct pmc_owner *po;
1568         const struct pmc *pm;
1569         const struct pmc_process *pp;
1570
1571         pid = td->td_proc->p_pid;
1572
1573         LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1574             if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1575                 pmclog_process_map_out(po, pid, pkm->pm_address,
1576                     pkm->pm_address + pkm->pm_size);
1577
1578         if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL)
1579                 return;
1580
1581         for (ri = 0; ri < md->pmd_npmc; ri++)
1582                 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL &&
1583                     PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1584                         pmclog_process_map_out(pm->pm_owner, pid,
1585                             pkm->pm_address, pkm->pm_address + pkm->pm_size);
1586 }
1587
1588 /*
1589  * Log mapping information about the kernel.
1590  */
1591
1592 static void
1593 pmc_log_kernel_mappings(struct pmc *pm)
1594 {
1595         struct pmc_owner *po;
1596         struct pmckern_map_in *km, *kmbase;
1597
1598         sx_assert(&pmc_sx, SX_LOCKED);
1599         KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)),
1600             ("[pmc,%d] non-sampling PMC (%p) desires mapping information",
1601                 __LINE__, (void *) pm));
1602
1603         po = pm->pm_owner;
1604
1605         if (po->po_flags & PMC_PO_INITIAL_MAPPINGS_DONE)
1606                 return;
1607
1608         /*
1609          * Log the current set of kernel modules.
1610          */
1611         kmbase = linker_hwpmc_list_objects();
1612         for (km = kmbase; km->pm_file != NULL; km++) {
1613                 PMCDBG2(LOG,REG,1,"%s %p", (char *) km->pm_file,
1614                     (void *) km->pm_address);
1615                 pmclog_process_map_in(po, (pid_t) -1, km->pm_address,
1616                     km->pm_file);
1617         }
1618         free(kmbase, M_LINKER);
1619
1620         po->po_flags |= PMC_PO_INITIAL_MAPPINGS_DONE;
1621 }
1622
1623 /*
1624  * Log the mappings for a single process.
1625  */
1626
1627 static void
1628 pmc_log_process_mappings(struct pmc_owner *po, struct proc *p)
1629 {
1630         vm_map_t map;
1631         struct vnode *vp;
1632         struct vmspace *vm;
1633         vm_map_entry_t entry;
1634         vm_offset_t last_end;
1635         u_int last_timestamp;
1636         struct vnode *last_vp;
1637         vm_offset_t start_addr;
1638         vm_object_t obj, lobj, tobj;
1639         char *fullpath, *freepath;
1640
1641         last_vp = NULL;
1642         last_end = (vm_offset_t) 0;
1643         fullpath = freepath = NULL;
1644
1645         if ((vm = vmspace_acquire_ref(p)) == NULL)
1646                 return;
1647
1648         map = &vm->vm_map;
1649         vm_map_lock_read(map);
1650
1651         for (entry = map->header.next; entry != &map->header; entry = entry->next) {
1652
1653                 if (entry == NULL) {
1654                         PMCDBG2(LOG,OPS,2, "hwpmc: vm_map entry unexpectedly "
1655                             "NULL! pid=%d vm_map=%p\n", p->p_pid, map);
1656                         break;
1657                 }
1658
1659                 /*
1660                  * We only care about executable map entries.
1661                  */
1662                 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) ||
1663                     !(entry->protection & VM_PROT_EXECUTE) ||
1664                     (entry->object.vm_object == NULL)) {
1665                         continue;
1666                 }
1667
1668                 obj = entry->object.vm_object;
1669                 VM_OBJECT_RLOCK(obj);
1670
1671                 /*
1672                  * Walk the backing_object list to find the base
1673                  * (non-shadowed) vm_object.
1674                  */
1675                 for (lobj = tobj = obj; tobj != NULL; tobj = tobj->backing_object) {
1676                         if (tobj != obj)
1677                                 VM_OBJECT_RLOCK(tobj);
1678                         if (lobj != obj)
1679                                 VM_OBJECT_RUNLOCK(lobj);
1680                         lobj = tobj;
1681                 }
1682
1683                 /*
1684                  * At this point lobj is the base vm_object and it is locked.
1685                  */
1686                 if (lobj == NULL) {
1687                         PMCDBG3(LOG,OPS,2, "hwpmc: lobj unexpectedly NULL! pid=%d "
1688                             "vm_map=%p vm_obj=%p\n", p->p_pid, map, obj);
1689                         VM_OBJECT_RUNLOCK(obj);
1690                         continue;
1691                 }
1692
1693                 vp = vm_object_vnode(lobj);
1694                 if (vp == NULL) {
1695                         if (lobj != obj)
1696                                 VM_OBJECT_RUNLOCK(lobj);
1697                         VM_OBJECT_RUNLOCK(obj);
1698                         continue;
1699                 }
1700
1701                 /*
1702                  * Skip contiguous regions that point to the same
1703                  * vnode, so we don't emit redundant MAP-IN
1704                  * directives.
1705                  */
1706                 if (entry->start == last_end && vp == last_vp) {
1707                         last_end = entry->end;
1708                         if (lobj != obj)
1709                                 VM_OBJECT_RUNLOCK(lobj);
1710                         VM_OBJECT_RUNLOCK(obj);
1711                         continue;
1712                 }
1713
1714                 /*
1715                  * We don't want to keep the proc's vm_map or this
1716                  * vm_object locked while we walk the pathname, since
1717                  * vn_fullpath() can sleep.  However, if we drop the
1718                  * lock, it's possible for concurrent activity to
1719                  * modify the vm_map list.  To protect against this,
1720                  * we save the vm_map timestamp before we release the
1721                  * lock, and check it after we reacquire the lock
1722                  * below.
1723                  */
1724                 start_addr = entry->start;
1725                 last_end = entry->end;
1726                 last_timestamp = map->timestamp;
1727                 vm_map_unlock_read(map);
1728
1729                 vref(vp);
1730                 if (lobj != obj)
1731                         VM_OBJECT_RUNLOCK(lobj);
1732
1733                 VM_OBJECT_RUNLOCK(obj);
1734
1735                 freepath = NULL;
1736                 pmc_getfilename(vp, &fullpath, &freepath);
1737                 last_vp = vp;
1738
1739                 vrele(vp);
1740
1741                 vp = NULL;
1742                 pmclog_process_map_in(po, p->p_pid, start_addr, fullpath);
1743                 if (freepath)
1744                         free(freepath, M_TEMP);
1745
1746                 vm_map_lock_read(map);
1747
1748                 /*
1749                  * If our saved timestamp doesn't match, this means
1750                  * that the vm_map was modified out from under us and
1751                  * we can't trust our current "entry" pointer.  Do a
1752                  * new lookup for this entry.  If there is no entry
1753                  * for this address range, vm_map_lookup_entry() will
1754                  * return the previous one, so we always want to go to
1755                  * entry->next on the next loop iteration.
1756                  *
1757                  * There is an edge condition here that can occur if
1758                  * there is no entry at or before this address.  In
1759                  * this situation, vm_map_lookup_entry returns
1760                  * &map->header, which would cause our loop to abort
1761                  * without processing the rest of the map.  However,
1762                  * in practice this will never happen for process
1763                  * vm_map.  This is because the executable's text
1764                  * segment is the first mapping in the proc's address
1765                  * space, and this mapping is never removed until the
1766                  * process exits, so there will always be a non-header
1767                  * entry at or before the requested address for
1768                  * vm_map_lookup_entry to return.
1769                  */
1770                 if (map->timestamp != last_timestamp)
1771                         vm_map_lookup_entry(map, last_end - 1, &entry);
1772         }
1773
1774         vm_map_unlock_read(map);
1775         vmspace_free(vm);
1776         return;
1777 }
1778
1779 /*
1780  * Log mappings for all processes in the system.
1781  */
1782
1783 static void
1784 pmc_log_all_process_mappings(struct pmc_owner *po)
1785 {
1786         struct proc *p, *top;
1787
1788         sx_assert(&pmc_sx, SX_XLOCKED);
1789
1790         if ((p = pfind(1)) == NULL)
1791                 panic("[pmc,%d] Cannot find init", __LINE__);
1792
1793         PROC_UNLOCK(p);
1794
1795         sx_slock(&proctree_lock);
1796
1797         top = p;
1798
1799         for (;;) {
1800                 pmc_log_process_mappings(po, p);
1801                 if (!LIST_EMPTY(&p->p_children))
1802                         p = LIST_FIRST(&p->p_children);
1803                 else for (;;) {
1804                         if (p == top)
1805                                 goto done;
1806                         if (LIST_NEXT(p, p_sibling)) {
1807                                 p = LIST_NEXT(p, p_sibling);
1808                                 break;
1809                         }
1810                         p = p->p_pptr;
1811                 }
1812         }
1813  done:
1814         sx_sunlock(&proctree_lock);
1815 }
1816
1817 /*
1818  * The 'hook' invoked from the kernel proper
1819  */
1820
1821
1822 #ifdef  HWPMC_DEBUG
1823 const char *pmc_hooknames[] = {
1824         /* these strings correspond to PMC_FN_* in <sys/pmckern.h> */
1825         "",
1826         "EXEC",
1827         "CSW-IN",
1828         "CSW-OUT",
1829         "SAMPLE",
1830         "UNUSED1",
1831         "UNUSED2",
1832         "MMAP",
1833         "MUNMAP",
1834         "CALLCHAIN-NMI",
1835         "CALLCHAIN-SOFT",
1836         "SOFTSAMPLING"
1837 };
1838 #endif
1839
1840 static int
1841 pmc_hook_handler(struct thread *td, int function, void *arg)
1842 {
1843
1844         PMCDBG4(MOD,PMH,1, "hook td=%p func=%d \"%s\" arg=%p", td, function,
1845             pmc_hooknames[function], arg);
1846
1847         switch (function)
1848         {
1849
1850         /*
1851          * Process exec()
1852          */
1853
1854         case PMC_FN_PROCESS_EXEC:
1855         {
1856                 char *fullpath, *freepath;
1857                 unsigned int ri;
1858                 int is_using_hwpmcs;
1859                 struct pmc *pm;
1860                 struct proc *p;
1861                 struct pmc_owner *po;
1862                 struct pmc_process *pp;
1863                 struct pmckern_procexec *pk;
1864
1865                 sx_assert(&pmc_sx, SX_XLOCKED);
1866
1867                 p = td->td_proc;
1868                 pmc_getfilename(p->p_textvp, &fullpath, &freepath);
1869
1870                 pk = (struct pmckern_procexec *) arg;
1871
1872                 /* Inform owners of SS mode PMCs of the exec event. */
1873                 LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1874                     if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1875                             pmclog_process_procexec(po, PMC_ID_INVALID,
1876                                 p->p_pid, pk->pm_entryaddr, fullpath);
1877
1878                 PROC_LOCK(p);
1879                 is_using_hwpmcs = p->p_flag & P_HWPMC;
1880                 PROC_UNLOCK(p);
1881
1882                 if (!is_using_hwpmcs) {
1883                         if (freepath)
1884                                 free(freepath, M_TEMP);
1885                         break;
1886                 }
1887
1888                 /*
1889                  * PMCs are not inherited across an exec():  remove any
1890                  * PMCs that this process is the owner of.
1891                  */
1892
1893                 if ((po = pmc_find_owner_descriptor(p)) != NULL) {
1894                         pmc_remove_owner(po);
1895                         pmc_destroy_owner_descriptor(po);
1896                 }
1897
1898                 /*
1899                  * If the process being exec'ed is not the target of any
1900                  * PMC, we are done.
1901                  */
1902                 if ((pp = pmc_find_process_descriptor(p, 0)) == NULL) {
1903                         if (freepath)
1904                                 free(freepath, M_TEMP);
1905                         break;
1906                 }
1907
1908                 /*
1909                  * Log the exec event to all monitoring owners.  Skip
1910                  * owners who have already recieved the event because
1911                  * they had system sampling PMCs active.
1912                  */
1913                 for (ri = 0; ri < md->pmd_npmc; ri++)
1914                         if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) {
1915                                 po = pm->pm_owner;
1916                                 if (po->po_sscount == 0 &&
1917                                     po->po_flags & PMC_PO_OWNS_LOGFILE)
1918                                         pmclog_process_procexec(po, pm->pm_id,
1919                                             p->p_pid, pk->pm_entryaddr,
1920                                             fullpath);
1921                         }
1922
1923                 if (freepath)
1924                         free(freepath, M_TEMP);
1925
1926
1927                 PMCDBG4(PRC,EXC,1, "exec proc=%p (%d, %s) cred-changed=%d",
1928                     p, p->p_pid, p->p_comm, pk->pm_credentialschanged);
1929
1930                 if (pk->pm_credentialschanged == 0) /* no change */
1931                         break;
1932
1933                 /*
1934                  * If the newly exec()'ed process has a different credential
1935                  * than before, allow it to be the target of a PMC only if
1936                  * the PMC's owner has sufficient priviledge.
1937                  */
1938
1939                 for (ri = 0; ri < md->pmd_npmc; ri++)
1940                         if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL)
1941                                 if (pmc_can_attach(pm, td->td_proc) != 0)
1942                                         pmc_detach_one_process(td->td_proc,
1943                                             pm, PMC_FLAG_NONE);
1944
1945                 KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= (int) md->pmd_npmc,
1946                     ("[pmc,%d] Illegal ref count %d on pp %p", __LINE__,
1947                         pp->pp_refcnt, pp));
1948
1949                 /*
1950                  * If this process is no longer the target of any
1951                  * PMCs, we can remove the process entry and free
1952                  * up space.
1953                  */
1954
1955                 if (pp->pp_refcnt == 0) {
1956                         pmc_remove_process_descriptor(pp);
1957                         free(pp, M_PMC);
1958                         break;
1959                 }
1960
1961         }
1962         break;
1963
1964         case PMC_FN_CSW_IN:
1965                 pmc_process_csw_in(td);
1966                 break;
1967
1968         case PMC_FN_CSW_OUT:
1969                 pmc_process_csw_out(td);
1970                 break;
1971
1972         /*
1973          * Process accumulated PC samples.
1974          *
1975          * This function is expected to be called by hardclock() for
1976          * each CPU that has accumulated PC samples.
1977          *
1978          * This function is to be executed on the CPU whose samples
1979          * are being processed.
1980          */
1981         case PMC_FN_DO_SAMPLES:
1982
1983                 /*
1984                  * Clear the cpu specific bit in the CPU mask before
1985                  * do the rest of the processing.  If the NMI handler
1986                  * gets invoked after the "atomic_clear_int()" call
1987                  * below but before "pmc_process_samples()" gets
1988                  * around to processing the interrupt, then we will
1989                  * come back here at the next hardclock() tick (and
1990                  * may find nothing to do if "pmc_process_samples()"
1991                  * had already processed the interrupt).  We don't
1992                  * lose the interrupt sample.
1993                  */
1994                 CPU_CLR_ATOMIC(PCPU_GET(cpuid), &pmc_cpumask);
1995                 pmc_process_samples(PCPU_GET(cpuid), PMC_HR);
1996                 pmc_process_samples(PCPU_GET(cpuid), PMC_SR);
1997                 break;
1998
1999         case PMC_FN_MMAP:
2000                 sx_assert(&pmc_sx, SX_LOCKED);
2001                 pmc_process_mmap(td, (struct pmckern_map_in *) arg);
2002                 break;
2003
2004         case PMC_FN_MUNMAP:
2005                 sx_assert(&pmc_sx, SX_LOCKED);
2006                 pmc_process_munmap(td, (struct pmckern_map_out *) arg);
2007                 break;
2008
2009         case PMC_FN_USER_CALLCHAIN:
2010                 /*
2011                  * Record a call chain.
2012                  */
2013                 KASSERT(td == curthread, ("[pmc,%d] td != curthread",
2014                     __LINE__));
2015
2016                 pmc_capture_user_callchain(PCPU_GET(cpuid), PMC_HR,
2017                     (struct trapframe *) arg);
2018                 td->td_pflags &= ~TDP_CALLCHAIN;
2019                 break;
2020
2021         case PMC_FN_USER_CALLCHAIN_SOFT:
2022                 /*
2023                  * Record a call chain.
2024                  */
2025                 KASSERT(td == curthread, ("[pmc,%d] td != curthread",
2026                     __LINE__));
2027                 pmc_capture_user_callchain(PCPU_GET(cpuid), PMC_SR,
2028                     (struct trapframe *) arg);
2029                 td->td_pflags &= ~TDP_CALLCHAIN;
2030                 break;
2031
2032         case PMC_FN_SOFT_SAMPLING:
2033                 /*
2034                  * Call soft PMC sampling intr.
2035                  */
2036                 pmc_soft_intr((struct pmckern_soft *) arg);
2037                 break;
2038
2039         default:
2040 #ifdef  HWPMC_DEBUG
2041                 KASSERT(0, ("[pmc,%d] unknown hook %d\n", __LINE__, function));
2042 #endif
2043                 break;
2044
2045         }
2046
2047         return 0;
2048 }
2049
2050 /*
2051  * allocate a 'struct pmc_owner' descriptor in the owner hash table.
2052  */
2053
2054 static struct pmc_owner *
2055 pmc_allocate_owner_descriptor(struct proc *p)
2056 {
2057         uint32_t hindex;
2058         struct pmc_owner *po;
2059         struct pmc_ownerhash *poh;
2060
2061         hindex = PMC_HASH_PTR(p, pmc_ownerhashmask);
2062         poh = &pmc_ownerhash[hindex];
2063
2064         /* allocate space for N pointers and one descriptor struct */
2065         po = malloc(sizeof(struct pmc_owner), M_PMC, M_WAITOK|M_ZERO);
2066         po->po_owner = p;
2067         LIST_INSERT_HEAD(poh, po, po_next); /* insert into hash table */
2068
2069         TAILQ_INIT(&po->po_logbuffers);
2070         mtx_init(&po->po_mtx, "pmc-owner-mtx", "pmc-per-proc", MTX_SPIN);
2071
2072         PMCDBG4(OWN,ALL,1, "allocate-owner proc=%p (%d, %s) pmc-owner=%p",
2073             p, p->p_pid, p->p_comm, po);
2074
2075         return po;
2076 }
2077
2078 static void
2079 pmc_destroy_owner_descriptor(struct pmc_owner *po)
2080 {
2081
2082         PMCDBG4(OWN,REL,1, "destroy-owner po=%p proc=%p (%d, %s)",
2083             po, po->po_owner, po->po_owner->p_pid, po->po_owner->p_comm);
2084
2085         mtx_destroy(&po->po_mtx);
2086         free(po, M_PMC);
2087 }
2088
2089 /*
2090  * find the descriptor corresponding to process 'p', adding or removing it
2091  * as specified by 'mode'.
2092  */
2093
2094 static struct pmc_process *
2095 pmc_find_process_descriptor(struct proc *p, uint32_t mode)
2096 {
2097         uint32_t hindex;
2098         struct pmc_process *pp, *ppnew;
2099         struct pmc_processhash *pph;
2100
2101         hindex = PMC_HASH_PTR(p, pmc_processhashmask);
2102         pph = &pmc_processhash[hindex];
2103
2104         ppnew = NULL;
2105
2106         /*
2107          * Pre-allocate memory in the FIND_ALLOCATE case since we
2108          * cannot call malloc(9) once we hold a spin lock.
2109          */
2110         if (mode & PMC_FLAG_ALLOCATE)
2111                 ppnew = malloc(sizeof(struct pmc_process) + md->pmd_npmc *
2112                     sizeof(struct pmc_targetstate), M_PMC, M_WAITOK|M_ZERO);
2113
2114         mtx_lock_spin(&pmc_processhash_mtx);
2115         LIST_FOREACH(pp, pph, pp_next)
2116             if (pp->pp_proc == p)
2117                     break;
2118
2119         if ((mode & PMC_FLAG_REMOVE) && pp != NULL)
2120                 LIST_REMOVE(pp, pp_next);
2121
2122         if ((mode & PMC_FLAG_ALLOCATE) && pp == NULL &&
2123             ppnew != NULL) {
2124                 ppnew->pp_proc = p;
2125                 LIST_INSERT_HEAD(pph, ppnew, pp_next);
2126                 pp = ppnew;
2127                 ppnew = NULL;
2128         }
2129         mtx_unlock_spin(&pmc_processhash_mtx);
2130
2131         if (pp != NULL && ppnew != NULL)
2132                 free(ppnew, M_PMC);
2133
2134         return pp;
2135 }
2136
2137 /*
2138  * remove a process descriptor from the process hash table.
2139  */
2140
2141 static void
2142 pmc_remove_process_descriptor(struct pmc_process *pp)
2143 {
2144         KASSERT(pp->pp_refcnt == 0,
2145             ("[pmc,%d] Removing process descriptor %p with count %d",
2146                 __LINE__, pp, pp->pp_refcnt));
2147
2148         mtx_lock_spin(&pmc_processhash_mtx);
2149         LIST_REMOVE(pp, pp_next);
2150         mtx_unlock_spin(&pmc_processhash_mtx);
2151 }
2152
2153
2154 /*
2155  * find an owner descriptor corresponding to proc 'p'
2156  */
2157
2158 static struct pmc_owner *
2159 pmc_find_owner_descriptor(struct proc *p)
2160 {
2161         uint32_t hindex;
2162         struct pmc_owner *po;
2163         struct pmc_ownerhash *poh;
2164
2165         hindex = PMC_HASH_PTR(p, pmc_ownerhashmask);
2166         poh = &pmc_ownerhash[hindex];
2167
2168         po = NULL;
2169         LIST_FOREACH(po, poh, po_next)
2170             if (po->po_owner == p)
2171                     break;
2172
2173         PMCDBG5(OWN,FND,1, "find-owner proc=%p (%d, %s) hindex=0x%x -> "
2174             "pmc-owner=%p", p, p->p_pid, p->p_comm, hindex, po);
2175
2176         return po;
2177 }
2178
2179 /*
2180  * pmc_allocate_pmc_descriptor
2181  *
2182  * Allocate a pmc descriptor and initialize its
2183  * fields.
2184  */
2185
2186 static struct pmc *
2187 pmc_allocate_pmc_descriptor(void)
2188 {
2189         struct pmc *pmc;
2190
2191         pmc = malloc(sizeof(struct pmc), M_PMC, M_WAITOK|M_ZERO);
2192
2193         PMCDBG1(PMC,ALL,1, "allocate-pmc -> pmc=%p", pmc);
2194
2195         return pmc;
2196 }
2197
2198 /*
2199  * Destroy a pmc descriptor.
2200  */
2201
2202 static void
2203 pmc_destroy_pmc_descriptor(struct pmc *pm)
2204 {
2205
2206         KASSERT(pm->pm_state == PMC_STATE_DELETED ||
2207             pm->pm_state == PMC_STATE_FREE,
2208             ("[pmc,%d] destroying non-deleted PMC", __LINE__));
2209         KASSERT(LIST_EMPTY(&pm->pm_targets),
2210             ("[pmc,%d] destroying pmc with targets", __LINE__));
2211         KASSERT(pm->pm_owner == NULL,
2212             ("[pmc,%d] destroying pmc attached to an owner", __LINE__));
2213         KASSERT(pm->pm_runcount == 0,
2214             ("[pmc,%d] pmc has non-zero run count %d", __LINE__,
2215                 pm->pm_runcount));
2216
2217         free(pm, M_PMC);
2218 }
2219
2220 static void
2221 pmc_wait_for_pmc_idle(struct pmc *pm)
2222 {
2223 #ifdef HWPMC_DEBUG
2224         volatile int maxloop;
2225
2226         maxloop = 100 * pmc_cpu_max();
2227 #endif
2228         /*
2229          * Loop (with a forced context switch) till the PMC's runcount
2230          * comes down to zero.
2231          */
2232         while (atomic_load_acq_32(&pm->pm_runcount) > 0) {
2233 #ifdef HWPMC_DEBUG
2234                 maxloop--;
2235                 KASSERT(maxloop > 0,
2236                     ("[pmc,%d] (ri%d, rc%d) waiting too long for "
2237                         "pmc to be free", __LINE__,
2238                         PMC_TO_ROWINDEX(pm), pm->pm_runcount));
2239 #endif
2240                 pmc_force_context_switch();
2241         }
2242 }
2243
2244 /*
2245  * This function does the following things:
2246  *
2247  *  - detaches the PMC from hardware
2248  *  - unlinks all target threads that were attached to it
2249  *  - removes the PMC from its owner's list
2250  *  - destroys the PMC private mutex
2251  *
2252  * Once this function completes, the given pmc pointer can be freed by
2253  * calling pmc_destroy_pmc_descriptor().
2254  */
2255
2256 static void
2257 pmc_release_pmc_descriptor(struct pmc *pm)
2258 {
2259         enum pmc_mode mode;
2260         struct pmc_hw *phw;
2261         u_int adjri, ri, cpu;
2262         struct pmc_owner *po;
2263         struct pmc_binding pb;
2264         struct pmc_process *pp;
2265         struct pmc_classdep *pcd;
2266         struct pmc_target *ptgt, *tmp;
2267
2268         sx_assert(&pmc_sx, SX_XLOCKED);
2269
2270         KASSERT(pm, ("[pmc,%d] null pmc", __LINE__));
2271
2272         ri   = PMC_TO_ROWINDEX(pm);
2273         pcd  = pmc_ri_to_classdep(md, ri, &adjri);
2274         mode = PMC_TO_MODE(pm);
2275
2276         PMCDBG3(PMC,REL,1, "release-pmc pmc=%p ri=%d mode=%d", pm, ri,
2277             mode);
2278
2279         /*
2280          * First, we take the PMC off hardware.
2281          */
2282         cpu = 0;
2283         if (PMC_IS_SYSTEM_MODE(mode)) {
2284
2285                 /*
2286                  * A system mode PMC runs on a specific CPU.  Switch
2287                  * to this CPU and turn hardware off.
2288                  */
2289                 pmc_save_cpu_binding(&pb);
2290
2291                 cpu = PMC_TO_CPU(pm);
2292
2293                 pmc_select_cpu(cpu);
2294
2295                 /* switch off non-stalled CPUs */
2296                 CPU_CLR_ATOMIC(cpu, &pm->pm_cpustate);
2297                 if (pm->pm_state == PMC_STATE_RUNNING &&
2298                     !CPU_ISSET(cpu, &pm->pm_stalled)) {
2299
2300                         phw = pmc_pcpu[cpu]->pc_hwpmcs[ri];
2301
2302                         KASSERT(phw->phw_pmc == pm,
2303                             ("[pmc, %d] pmc ptr ri(%d) hw(%p) pm(%p)",
2304                                 __LINE__, ri, phw->phw_pmc, pm));
2305                         PMCDBG2(PMC,REL,2, "stopping cpu=%d ri=%d", cpu, ri);
2306
2307                         critical_enter();
2308                         pcd->pcd_stop_pmc(cpu, adjri);
2309                         critical_exit();
2310                 }
2311
2312                 PMCDBG2(PMC,REL,2, "decfg cpu=%d ri=%d", cpu, ri);
2313
2314                 critical_enter();
2315                 pcd->pcd_config_pmc(cpu, adjri, NULL);
2316                 critical_exit();
2317
2318                 /* adjust the global and process count of SS mode PMCs */
2319                 if (mode == PMC_MODE_SS && pm->pm_state == PMC_STATE_RUNNING) {
2320                         po = pm->pm_owner;
2321                         po->po_sscount--;
2322                         if (po->po_sscount == 0) {
2323                                 atomic_subtract_rel_int(&pmc_ss_count, 1);
2324                                 LIST_REMOVE(po, po_ssnext);
2325                         }
2326                 }
2327
2328                 pm->pm_state = PMC_STATE_DELETED;
2329
2330                 pmc_restore_cpu_binding(&pb);
2331
2332                 /*
2333                  * We could have references to this PMC structure in
2334                  * the per-cpu sample queues.  Wait for the queue to
2335                  * drain.
2336                  */
2337                 pmc_wait_for_pmc_idle(pm);
2338
2339         } else if (PMC_IS_VIRTUAL_MODE(mode)) {
2340
2341                 /*
2342                  * A virtual PMC could be running on multiple CPUs at
2343                  * a given instant.
2344                  *
2345                  * By marking its state as DELETED, we ensure that
2346                  * this PMC is never further scheduled on hardware.
2347                  *
2348                  * Then we wait till all CPUs are done with this PMC.
2349                  */
2350                 pm->pm_state = PMC_STATE_DELETED;
2351
2352
2353                 /* Wait for the PMCs runcount to come to zero. */
2354                 pmc_wait_for_pmc_idle(pm);
2355
2356                 /*
2357                  * At this point the PMC is off all CPUs and cannot be
2358                  * freshly scheduled onto a CPU.  It is now safe to
2359                  * unlink all targets from this PMC.  If a
2360                  * process-record's refcount falls to zero, we remove
2361                  * it from the hash table.  The module-wide SX lock
2362                  * protects us from races.
2363                  */
2364                 LIST_FOREACH_SAFE(ptgt, &pm->pm_targets, pt_next, tmp) {
2365                         pp = ptgt->pt_process;
2366                         pmc_unlink_target_process(pm, pp); /* frees 'ptgt' */
2367
2368                         PMCDBG1(PMC,REL,3, "pp->refcnt=%d", pp->pp_refcnt);
2369
2370                         /*
2371                          * If the target process record shows that no
2372                          * PMCs are attached to it, reclaim its space.
2373                          */
2374
2375                         if (pp->pp_refcnt == 0) {
2376                                 pmc_remove_process_descriptor(pp);
2377                                 free(pp, M_PMC);
2378                         }
2379                 }
2380
2381                 cpu = curthread->td_oncpu; /* setup cpu for pmd_release() */
2382
2383         }
2384
2385         /*
2386          * Release any MD resources
2387          */
2388         (void) pcd->pcd_release_pmc(cpu, adjri, pm);
2389
2390         /*
2391          * Update row disposition
2392          */
2393
2394         if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm)))
2395                 PMC_UNMARK_ROW_STANDALONE(ri);
2396         else
2397                 PMC_UNMARK_ROW_THREAD(ri);
2398
2399         /* unlink from the owner's list */
2400         if (pm->pm_owner) {
2401                 LIST_REMOVE(pm, pm_next);
2402                 pm->pm_owner = NULL;
2403         }
2404 }
2405
2406 /*
2407  * Register an owner and a pmc.
2408  */
2409
2410 static int
2411 pmc_register_owner(struct proc *p, struct pmc *pmc)
2412 {
2413         struct pmc_owner *po;
2414
2415         sx_assert(&pmc_sx, SX_XLOCKED);
2416
2417         if ((po = pmc_find_owner_descriptor(p)) == NULL)
2418                 if ((po = pmc_allocate_owner_descriptor(p)) == NULL)
2419                         return ENOMEM;
2420
2421         KASSERT(pmc->pm_owner == NULL,
2422             ("[pmc,%d] attempting to own an initialized PMC", __LINE__));
2423         pmc->pm_owner  = po;
2424
2425         LIST_INSERT_HEAD(&po->po_pmcs, pmc, pm_next);
2426
2427         PROC_LOCK(p);
2428         p->p_flag |= P_HWPMC;
2429         PROC_UNLOCK(p);
2430
2431         if (po->po_flags & PMC_PO_OWNS_LOGFILE)
2432                 pmclog_process_pmcallocate(pmc);
2433
2434         PMCDBG2(PMC,REG,1, "register-owner pmc-owner=%p pmc=%p",
2435             po, pmc);
2436
2437         return 0;
2438 }
2439
2440 /*
2441  * Return the current row disposition:
2442  * == 0 => FREE
2443  *  > 0 => PROCESS MODE
2444  *  < 0 => SYSTEM MODE
2445  */
2446
2447 int
2448 pmc_getrowdisp(int ri)
2449 {
2450         return pmc_pmcdisp[ri];
2451 }
2452
2453 /*
2454  * Check if a PMC at row index 'ri' can be allocated to the current
2455  * process.
2456  *
2457  * Allocation can fail if:
2458  *   - the current process is already being profiled by a PMC at index 'ri',
2459  *     attached to it via OP_PMCATTACH.
2460  *   - the current process has already allocated a PMC at index 'ri'
2461  *     via OP_ALLOCATE.
2462  */
2463
2464 static int
2465 pmc_can_allocate_rowindex(struct proc *p, unsigned int ri, int cpu)
2466 {
2467         enum pmc_mode mode;
2468         struct pmc *pm;
2469         struct pmc_owner *po;
2470         struct pmc_process *pp;
2471
2472         PMCDBG5(PMC,ALR,1, "can-allocate-rowindex proc=%p (%d, %s) ri=%d "
2473             "cpu=%d", p, p->p_pid, p->p_comm, ri, cpu);
2474
2475         /*
2476          * We shouldn't have already allocated a process-mode PMC at
2477          * row index 'ri'.
2478          *
2479          * We shouldn't have allocated a system-wide PMC on the same
2480          * CPU and same RI.
2481          */
2482         if ((po = pmc_find_owner_descriptor(p)) != NULL)
2483                 LIST_FOREACH(pm, &po->po_pmcs, pm_next) {
2484                     if (PMC_TO_ROWINDEX(pm) == ri) {
2485                             mode = PMC_TO_MODE(pm);
2486                             if (PMC_IS_VIRTUAL_MODE(mode))
2487                                     return EEXIST;
2488                             if (PMC_IS_SYSTEM_MODE(mode) &&
2489                                 (int) PMC_TO_CPU(pm) == cpu)
2490                                     return EEXIST;
2491                     }
2492                 }
2493
2494         /*
2495          * We also shouldn't be the target of any PMC at this index
2496          * since otherwise a PMC_ATTACH to ourselves will fail.
2497          */
2498         if ((pp = pmc_find_process_descriptor(p, 0)) != NULL)
2499                 if (pp->pp_pmcs[ri].pp_pmc)
2500                         return EEXIST;
2501
2502         PMCDBG4(PMC,ALR,2, "can-allocate-rowindex proc=%p (%d, %s) ri=%d ok",
2503             p, p->p_pid, p->p_comm, ri);
2504
2505         return 0;
2506 }
2507
2508 /*
2509  * Check if a given PMC at row index 'ri' can be currently used in
2510  * mode 'mode'.
2511  */
2512
2513 static int
2514 pmc_can_allocate_row(int ri, enum pmc_mode mode)
2515 {
2516         enum pmc_disp   disp;
2517
2518         sx_assert(&pmc_sx, SX_XLOCKED);
2519
2520         PMCDBG2(PMC,ALR,1, "can-allocate-row ri=%d mode=%d", ri, mode);
2521
2522         if (PMC_IS_SYSTEM_MODE(mode))
2523                 disp = PMC_DISP_STANDALONE;
2524         else
2525                 disp = PMC_DISP_THREAD;
2526
2527         /*
2528          * check disposition for PMC row 'ri':
2529          *
2530          * Expected disposition         Row-disposition         Result
2531          *
2532          * STANDALONE                   STANDALONE or FREE      proceed
2533          * STANDALONE                   THREAD                  fail
2534          * THREAD                       THREAD or FREE          proceed
2535          * THREAD                       STANDALONE              fail
2536          */
2537
2538         if (!PMC_ROW_DISP_IS_FREE(ri) &&
2539             !(disp == PMC_DISP_THREAD && PMC_ROW_DISP_IS_THREAD(ri)) &&
2540             !(disp == PMC_DISP_STANDALONE && PMC_ROW_DISP_IS_STANDALONE(ri)))
2541                 return EBUSY;
2542
2543         /*
2544          * All OK
2545          */
2546
2547         PMCDBG2(PMC,ALR,2, "can-allocate-row ri=%d mode=%d ok", ri, mode);
2548
2549         return 0;
2550
2551 }
2552
2553 /*
2554  * Find a PMC descriptor with user handle 'pmcid' for thread 'td'.
2555  */
2556
2557 static struct pmc *
2558 pmc_find_pmc_descriptor_in_process(struct pmc_owner *po, pmc_id_t pmcid)
2559 {
2560         struct pmc *pm;
2561
2562         KASSERT(PMC_ID_TO_ROWINDEX(pmcid) < md->pmd_npmc,
2563             ("[pmc,%d] Illegal pmc index %d (max %d)", __LINE__,
2564                 PMC_ID_TO_ROWINDEX(pmcid), md->pmd_npmc));
2565
2566         LIST_FOREACH(pm, &po->po_pmcs, pm_next)
2567             if (pm->pm_id == pmcid)
2568                     return pm;
2569
2570         return NULL;
2571 }
2572
2573 static int
2574 pmc_find_pmc(pmc_id_t pmcid, struct pmc **pmc)
2575 {
2576
2577         struct pmc *pm, *opm;
2578         struct pmc_owner *po;
2579         struct pmc_process *pp;
2580
2581         KASSERT(PMC_ID_TO_ROWINDEX(pmcid) < md->pmd_npmc,
2582             ("[pmc,%d] Illegal pmc index %d (max %d)", __LINE__,
2583                 PMC_ID_TO_ROWINDEX(pmcid), md->pmd_npmc));
2584         PMCDBG1(PMC,FND,1, "find-pmc id=%d", pmcid);
2585
2586         if ((po = pmc_find_owner_descriptor(curthread->td_proc)) == NULL) {
2587                 /*
2588                  * In case of PMC_F_DESCENDANTS child processes we will not find
2589                  * the current process in the owners hash list.  Find the owner
2590                  * process first and from there lookup the po.
2591                  */
2592                 if ((pp = pmc_find_process_descriptor(curthread->td_proc,
2593                     PMC_FLAG_NONE)) == NULL) {
2594                         return ESRCH;
2595                 } else {
2596                         opm = pp->pp_pmcs[PMC_ID_TO_ROWINDEX(pmcid)].pp_pmc;
2597                         if (opm == NULL)
2598                                 return ESRCH;
2599                         if ((opm->pm_flags & (PMC_F_ATTACHED_TO_OWNER|
2600                             PMC_F_DESCENDANTS)) != (PMC_F_ATTACHED_TO_OWNER|
2601                             PMC_F_DESCENDANTS))
2602                                 return ESRCH;
2603                         po = opm->pm_owner;
2604                 }
2605         }
2606
2607         if ((pm = pmc_find_pmc_descriptor_in_process(po, pmcid)) == NULL)
2608                 return EINVAL;
2609
2610         PMCDBG2(PMC,FND,2, "find-pmc id=%d -> pmc=%p", pmcid, pm);
2611
2612         *pmc = pm;
2613         return 0;
2614 }
2615
2616 /*
2617  * Start a PMC.
2618  */
2619
2620 static int
2621 pmc_start(struct pmc *pm)
2622 {
2623         enum pmc_mode mode;
2624         struct pmc_owner *po;
2625         struct pmc_binding pb;
2626         struct pmc_classdep *pcd;
2627         int adjri, error, cpu, ri;
2628
2629         KASSERT(pm != NULL,
2630             ("[pmc,%d] null pm", __LINE__));
2631
2632         mode = PMC_TO_MODE(pm);
2633         ri   = PMC_TO_ROWINDEX(pm);
2634         pcd  = pmc_ri_to_classdep(md, ri, &adjri);
2635
2636         error = 0;
2637
2638         PMCDBG3(PMC,OPS,1, "start pmc=%p mode=%d ri=%d", pm, mode, ri);
2639
2640         po = pm->pm_owner;
2641
2642         /*
2643          * Disallow PMCSTART if a logfile is required but has not been
2644          * configured yet.
2645          */
2646         if ((pm->pm_flags & PMC_F_NEEDS_LOGFILE) &&
2647             (po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
2648                 return (EDOOFUS);       /* programming error */
2649
2650         /*
2651          * If this is a sampling mode PMC, log mapping information for
2652          * the kernel modules that are currently loaded.
2653          */
2654         if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
2655             pmc_log_kernel_mappings(pm);
2656
2657         if (PMC_IS_VIRTUAL_MODE(mode)) {
2658
2659                 /*
2660                  * If a PMCATTACH has never been done on this PMC,
2661                  * attach it to its owner process.
2662                  */
2663
2664                 if (LIST_EMPTY(&pm->pm_targets))
2665                         error = (pm->pm_flags & PMC_F_ATTACH_DONE) ? ESRCH :
2666                             pmc_attach_process(po->po_owner, pm);
2667
2668                 /*
2669                  * If the PMC is attached to its owner, then force a context
2670                  * switch to ensure that the MD state gets set correctly.
2671                  */
2672
2673                 if (error == 0) {
2674                         pm->pm_state = PMC_STATE_RUNNING;
2675                         if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER)
2676                                 pmc_force_context_switch();
2677                 }
2678
2679                 return (error);
2680         }
2681
2682
2683         /*
2684          * A system-wide PMC.
2685          *
2686          * Add the owner to the global list if this is a system-wide
2687          * sampling PMC.
2688          */
2689
2690         if (mode == PMC_MODE_SS) {
2691                 if (po->po_sscount == 0) {
2692                         LIST_INSERT_HEAD(&pmc_ss_owners, po, po_ssnext);
2693                         atomic_add_rel_int(&pmc_ss_count, 1);
2694                         PMCDBG1(PMC,OPS,1, "po=%p in global list", po);
2695                 }
2696                 po->po_sscount++;
2697
2698                 /*
2699                  * Log mapping information for all existing processes in the
2700                  * system.  Subsequent mappings are logged as they happen;
2701                  * see pmc_process_mmap().
2702                  */
2703                 if (po->po_logprocmaps == 0) {
2704                         pmc_log_all_process_mappings(po);
2705                         po->po_logprocmaps = 1;
2706                 }
2707         }
2708
2709         /*
2710          * Move to the CPU associated with this
2711          * PMC, and start the hardware.
2712          */
2713
2714         pmc_save_cpu_binding(&pb);
2715
2716         cpu = PMC_TO_CPU(pm);
2717
2718         if (!pmc_cpu_is_active(cpu))
2719                 return (ENXIO);
2720
2721         pmc_select_cpu(cpu);
2722
2723         /*
2724          * global PMCs are configured at allocation time
2725          * so write out the initial value and start the PMC.
2726          */
2727
2728         pm->pm_state = PMC_STATE_RUNNING;
2729
2730         critical_enter();
2731         if ((error = pcd->pcd_write_pmc(cpu, adjri,
2732                  PMC_IS_SAMPLING_MODE(mode) ?
2733                  pm->pm_sc.pm_reloadcount :
2734                  pm->pm_sc.pm_initial)) == 0) {
2735                 /* If a sampling mode PMC, reset stalled state. */
2736                 if (PMC_IS_SAMPLING_MODE(mode))
2737                         CPU_CLR_ATOMIC(cpu, &pm->pm_stalled);
2738
2739                 /* Indicate that we desire this to run. Start it. */
2740                 CPU_SET_ATOMIC(cpu, &pm->pm_cpustate);
2741                 error = pcd->pcd_start_pmc(cpu, adjri);
2742         }
2743         critical_exit();
2744
2745         pmc_restore_cpu_binding(&pb);
2746
2747         return (error);
2748 }
2749
2750 /*
2751  * Stop a PMC.
2752  */
2753
2754 static int
2755 pmc_stop(struct pmc *pm)
2756 {
2757         struct pmc_owner *po;
2758         struct pmc_binding pb;
2759         struct pmc_classdep *pcd;
2760         int adjri, cpu, error, ri;
2761
2762         KASSERT(pm != NULL, ("[pmc,%d] null pmc", __LINE__));
2763
2764         PMCDBG3(PMC,OPS,1, "stop pmc=%p mode=%d ri=%d", pm,
2765             PMC_TO_MODE(pm), PMC_TO_ROWINDEX(pm));
2766
2767         pm->pm_state = PMC_STATE_STOPPED;
2768
2769         /*
2770          * If the PMC is a virtual mode one, changing the state to
2771          * non-RUNNING is enough to ensure that the PMC never gets
2772          * scheduled.
2773          *
2774          * If this PMC is current running on a CPU, then it will
2775          * handled correctly at the time its target process is context
2776          * switched out.
2777          */
2778
2779         if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)))
2780                 return 0;
2781
2782         /*
2783          * A system-mode PMC.  Move to the CPU associated with
2784          * this PMC, and stop the hardware.  We update the
2785          * 'initial count' so that a subsequent PMCSTART will
2786          * resume counting from the current hardware count.
2787          */
2788
2789         pmc_save_cpu_binding(&pb);
2790
2791         cpu = PMC_TO_CPU(pm);
2792
2793         KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
2794             ("[pmc,%d] illegal cpu=%d", __LINE__, cpu));
2795
2796         if (!pmc_cpu_is_active(cpu))
2797                 return ENXIO;
2798
2799         pmc_select_cpu(cpu);
2800
2801         ri = PMC_TO_ROWINDEX(pm);
2802         pcd = pmc_ri_to_classdep(md, ri, &adjri);
2803
2804         CPU_CLR_ATOMIC(cpu, &pm->pm_cpustate);
2805         critical_enter();
2806         if ((error = pcd->pcd_stop_pmc(cpu, adjri)) == 0)
2807                 error = pcd->pcd_read_pmc(cpu, adjri, &pm->pm_sc.pm_initial);
2808         critical_exit();
2809
2810         pmc_restore_cpu_binding(&pb);
2811
2812         po = pm->pm_owner;
2813
2814         /* remove this owner from the global list of SS PMC owners */
2815         if (PMC_TO_MODE(pm) == PMC_MODE_SS) {
2816                 po->po_sscount--;
2817                 if (po->po_sscount == 0) {
2818                         atomic_subtract_rel_int(&pmc_ss_count, 1);
2819                         LIST_REMOVE(po, po_ssnext);
2820                         PMCDBG1(PMC,OPS,2,"po=%p removed from global list", po);
2821                 }
2822         }
2823
2824         return (error);
2825 }
2826
2827
2828 #ifdef  HWPMC_DEBUG
2829 static const char *pmc_op_to_name[] = {
2830 #undef  __PMC_OP
2831 #define __PMC_OP(N, D)  #N ,
2832         __PMC_OPS()
2833         NULL
2834 };
2835 #endif
2836
2837 /*
2838  * The syscall interface
2839  */
2840
2841 #define PMC_GET_SX_XLOCK(...) do {              \
2842         sx_xlock(&pmc_sx);                      \
2843         if (pmc_hook == NULL) {                 \
2844                 sx_xunlock(&pmc_sx);            \
2845                 return __VA_ARGS__;             \
2846         }                                       \
2847 } while (0)
2848
2849 #define PMC_DOWNGRADE_SX() do {                 \
2850         sx_downgrade(&pmc_sx);                  \
2851         is_sx_downgraded = 1;                   \
2852 } while (0)
2853
2854 static int
2855 pmc_syscall_handler(struct thread *td, void *syscall_args)
2856 {
2857         int error, is_sx_downgraded, is_sx_locked, op;
2858         struct pmc_syscall_args *c;
2859         void *arg;
2860
2861         PMC_GET_SX_XLOCK(ENOSYS);
2862
2863         DROP_GIANT();
2864
2865         is_sx_downgraded = 0;
2866         is_sx_locked = 1;
2867
2868         c = (struct pmc_syscall_args *) syscall_args;
2869
2870         op = c->pmop_code;
2871         arg = c->pmop_data;
2872
2873         PMCDBG3(MOD,PMS,1, "syscall op=%d \"%s\" arg=%p", op,
2874             pmc_op_to_name[op], arg);
2875
2876         error = 0;
2877         atomic_add_int(&pmc_stats.pm_syscalls, 1);
2878
2879         switch(op)
2880         {
2881
2882
2883         /*
2884          * Configure a log file.
2885          *
2886          * XXX This OP will be reworked.
2887          */
2888
2889         case PMC_OP_CONFIGURELOG:
2890         {
2891                 struct proc *p;
2892                 struct pmc *pm;
2893                 struct pmc_owner *po;
2894                 struct pmc_op_configurelog cl;
2895
2896                 sx_assert(&pmc_sx, SX_XLOCKED);
2897
2898                 if ((error = copyin(arg, &cl, sizeof(cl))) != 0)
2899                         break;
2900
2901                 /* mark this process as owning a log file */
2902                 p = td->td_proc;
2903                 if ((po = pmc_find_owner_descriptor(p)) == NULL)
2904                         if ((po = pmc_allocate_owner_descriptor(p)) == NULL) {
2905                                 error = ENOMEM;
2906                                 break;
2907                         }
2908
2909                 /*
2910                  * If a valid fd was passed in, try to configure that,
2911                  * otherwise if 'fd' was less than zero and there was
2912                  * a log file configured, flush its buffers and
2913                  * de-configure it.
2914                  */
2915                 if (cl.pm_logfd >= 0) {
2916                         sx_xunlock(&pmc_sx);
2917                         is_sx_locked = 0;
2918                         error = pmclog_configure_log(md, po, cl.pm_logfd);
2919                 } else if (po->po_flags & PMC_PO_OWNS_LOGFILE) {
2920                         pmclog_process_closelog(po);
2921                         error = pmclog_close(po);
2922                         if (error == 0) {
2923                                 LIST_FOREACH(pm, &po->po_pmcs, pm_next)
2924                                     if (pm->pm_flags & PMC_F_NEEDS_LOGFILE &&
2925                                         pm->pm_state == PMC_STATE_RUNNING)
2926                                             pmc_stop(pm);
2927                                 error = pmclog_deconfigure_log(po);
2928                         }
2929                 } else
2930                         error = EINVAL;
2931
2932                 if (error)
2933                         break;
2934         }
2935         break;
2936
2937         /*
2938          * Flush a log file.
2939          */
2940
2941         case PMC_OP_FLUSHLOG:
2942         {
2943                 struct pmc_owner *po;
2944
2945                 sx_assert(&pmc_sx, SX_XLOCKED);
2946
2947                 if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) {
2948                         error = EINVAL;
2949                         break;
2950                 }
2951
2952                 error = pmclog_flush(po);
2953         }
2954         break;
2955
2956         /*
2957          * Close a log file.
2958          */
2959
2960         case PMC_OP_CLOSELOG:
2961         {
2962                 struct pmc_owner *po;
2963
2964                 sx_assert(&pmc_sx, SX_XLOCKED);
2965
2966                 if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) {
2967                         error = EINVAL;
2968                         break;
2969                 }
2970
2971                 error = pmclog_close(po);
2972         }
2973         break;
2974
2975         /*
2976          * Retrieve hardware configuration.
2977          */
2978
2979         case PMC_OP_GETCPUINFO: /* CPU information */
2980         {
2981                 struct pmc_op_getcpuinfo gci;
2982                 struct pmc_classinfo *pci;
2983                 struct pmc_classdep *pcd;
2984                 int cl;
2985
2986                 gci.pm_cputype = md->pmd_cputype;
2987                 gci.pm_ncpu    = pmc_cpu_max();
2988                 gci.pm_npmc    = md->pmd_npmc;
2989                 gci.pm_nclass  = md->pmd_nclass;
2990                 pci = gci.pm_classes;
2991                 pcd = md->pmd_classdep;
2992                 for (cl = 0; cl < md->pmd_nclass; cl++, pci++, pcd++) {
2993                         pci->pm_caps  = pcd->pcd_caps;
2994                         pci->pm_class = pcd->pcd_class;
2995                         pci->pm_width = pcd->pcd_width;
2996                         pci->pm_num   = pcd->pcd_num;
2997                 }
2998                 error = copyout(&gci, arg, sizeof(gci));
2999         }
3000         break;
3001
3002         /*
3003          * Retrieve soft events list.
3004          */
3005         case PMC_OP_GETDYNEVENTINFO:
3006         {
3007                 enum pmc_class                  cl;
3008                 enum pmc_event                  ev;
3009                 struct pmc_op_getdyneventinfo   *gei;
3010                 struct pmc_dyn_event_descr      dev;
3011                 struct pmc_soft                 *ps;
3012                 uint32_t                        nevent;
3013
3014                 sx_assert(&pmc_sx, SX_LOCKED);
3015
3016                 gei = (struct pmc_op_getdyneventinfo *) arg;
3017
3018                 if ((error = copyin(&gei->pm_class, &cl, sizeof(cl))) != 0)
3019                         break;
3020
3021                 /* Only SOFT class is dynamic. */
3022                 if (cl != PMC_CLASS_SOFT) {
3023                         error = EINVAL;
3024                         break;
3025                 }
3026
3027                 nevent = 0;
3028                 for (ev = PMC_EV_SOFT_FIRST; (int)ev <= PMC_EV_SOFT_LAST; ev++) {
3029                         ps = pmc_soft_ev_acquire(ev);
3030                         if (ps == NULL)
3031                                 continue;
3032                         bcopy(&ps->ps_ev, &dev, sizeof(dev));
3033                         pmc_soft_ev_release(ps);
3034
3035                         error = copyout(&dev,
3036                             &gei->pm_events[nevent],
3037                             sizeof(struct pmc_dyn_event_descr));
3038                         if (error != 0)
3039                                 break;
3040                         nevent++;
3041                 }
3042                 if (error != 0)
3043                         break;
3044
3045                 error = copyout(&nevent, &gei->pm_nevent,
3046                     sizeof(nevent));
3047         }
3048         break;
3049
3050         /*
3051          * Get module statistics
3052          */
3053
3054         case PMC_OP_GETDRIVERSTATS:
3055         {
3056                 struct pmc_op_getdriverstats gms;
3057
3058                 bcopy(&pmc_stats, &gms, sizeof(gms));
3059                 error = copyout(&gms, arg, sizeof(gms));
3060         }
3061         break;
3062
3063
3064         /*
3065          * Retrieve module version number
3066          */
3067
3068         case PMC_OP_GETMODULEVERSION:
3069         {
3070                 uint32_t cv, modv;
3071
3072                 /* retrieve the client's idea of the ABI version */
3073                 if ((error = copyin(arg, &cv, sizeof(uint32_t))) != 0)
3074                         break;
3075                 /* don't service clients newer than our driver */
3076                 modv = PMC_VERSION;
3077                 if ((cv & 0xFFFF0000) > (modv & 0xFFFF0000)) {
3078                         error = EPROGMISMATCH;
3079                         break;
3080                 }
3081                 error = copyout(&modv, arg, sizeof(int));
3082         }
3083         break;
3084
3085
3086         /*
3087          * Retrieve the state of all the PMCs on a given
3088          * CPU.
3089          */
3090
3091         case PMC_OP_GETPMCINFO:
3092         {
3093                 int ari;
3094                 struct pmc *pm;
3095                 size_t pmcinfo_size;
3096                 uint32_t cpu, n, npmc;
3097                 struct pmc_owner *po;
3098                 struct pmc_binding pb;
3099                 struct pmc_classdep *pcd;
3100                 struct pmc_info *p, *pmcinfo;
3101                 struct pmc_op_getpmcinfo *gpi;
3102
3103                 PMC_DOWNGRADE_SX();
3104
3105                 gpi = (struct pmc_op_getpmcinfo *) arg;
3106
3107                 if ((error = copyin(&gpi->pm_cpu, &cpu, sizeof(cpu))) != 0)
3108                         break;
3109
3110                 if (cpu >= pmc_cpu_max()) {
3111                         error = EINVAL;
3112                         break;
3113                 }
3114
3115                 if (!pmc_cpu_is_active(cpu)) {
3116                         error = ENXIO;
3117                         break;
3118                 }
3119
3120                 /* switch to CPU 'cpu' */
3121                 pmc_save_cpu_binding(&pb);
3122                 pmc_select_cpu(cpu);
3123
3124                 npmc = md->pmd_npmc;
3125
3126                 pmcinfo_size = npmc * sizeof(struct pmc_info);
3127                 pmcinfo = malloc(pmcinfo_size, M_PMC, M_WAITOK);
3128
3129                 p = pmcinfo;
3130
3131                 for (n = 0; n < md->pmd_npmc; n++, p++) {
3132
3133                         pcd = pmc_ri_to_classdep(md, n, &ari);
3134
3135                         KASSERT(pcd != NULL,
3136                             ("[pmc,%d] null pcd ri=%d", __LINE__, n));
3137
3138                         if ((error = pcd->pcd_describe(cpu, ari, p, &pm)) != 0)
3139                                 break;
3140
3141                         if (PMC_ROW_DISP_IS_STANDALONE(n))
3142                                 p->pm_rowdisp = PMC_DISP_STANDALONE;
3143                         else if (PMC_ROW_DISP_IS_THREAD(n))
3144                                 p->pm_rowdisp = PMC_DISP_THREAD;
3145                         else
3146                                 p->pm_rowdisp = PMC_DISP_FREE;
3147
3148                         p->pm_ownerpid = -1;
3149
3150                         if (pm == NULL) /* no PMC associated */
3151                                 continue;
3152
3153                         po = pm->pm_owner;
3154
3155                         KASSERT(po->po_owner != NULL,
3156                             ("[pmc,%d] pmc_owner had a null proc pointer",
3157                                 __LINE__));
3158
3159                         p->pm_ownerpid = po->po_owner->p_pid;
3160                         p->pm_mode     = PMC_TO_MODE(pm);
3161                         p->pm_event    = pm->pm_event;
3162                         p->pm_flags    = pm->pm_flags;
3163
3164                         if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
3165                                 p->pm_reloadcount =
3166                                     pm->pm_sc.pm_reloadcount;
3167                 }
3168
3169                 pmc_restore_cpu_binding(&pb);
3170
3171                 /* now copy out the PMC info collected */
3172                 if (error == 0)
3173                         error = copyout(pmcinfo, &gpi->pm_pmcs, pmcinfo_size);
3174
3175                 free(pmcinfo, M_PMC);
3176         }
3177         break;
3178
3179
3180         /*
3181          * Set the administrative state of a PMC.  I.e. whether
3182          * the PMC is to be used or not.
3183          */
3184
3185         case PMC_OP_PMCADMIN:
3186         {
3187                 int cpu, ri;
3188                 enum pmc_state request;
3189                 struct pmc_cpu *pc;
3190                 struct pmc_hw *phw;
3191                 struct pmc_op_pmcadmin pma;
3192                 struct pmc_binding pb;
3193
3194                 sx_assert(&pmc_sx, SX_XLOCKED);
3195
3196                 KASSERT(td == curthread,
3197                     ("[pmc,%d] td != curthread", __LINE__));
3198
3199                 error = priv_check(td, PRIV_PMC_MANAGE);
3200                 if (error)
3201                         break;
3202
3203                 if ((error = copyin(arg, &pma, sizeof(pma))) != 0)
3204                         break;
3205
3206                 cpu = pma.pm_cpu;
3207
3208                 if (cpu < 0 || cpu >= (int) pmc_cpu_max()) {
3209                         error = EINVAL;
3210                         break;
3211                 }
3212
3213                 if (!pmc_cpu_is_active(cpu)) {
3214                         error = ENXIO;
3215                         break;
3216                 }
3217
3218                 request = pma.pm_state;
3219
3220                 if (request != PMC_STATE_DISABLED &&
3221                     request != PMC_STATE_FREE) {
3222                         error = EINVAL;
3223                         break;
3224                 }
3225
3226                 ri = pma.pm_pmc; /* pmc id == row index */
3227                 if (ri < 0 || ri >= (int) md->pmd_npmc) {
3228                         error = EINVAL;
3229                         break;
3230                 }
3231
3232                 /*
3233                  * We can't disable a PMC with a row-index allocated
3234                  * for process virtual PMCs.
3235                  */
3236
3237                 if (PMC_ROW_DISP_IS_THREAD(ri) &&
3238                     request == PMC_STATE_DISABLED) {
3239                         error = EBUSY;
3240                         break;
3241                 }
3242
3243                 /*
3244                  * otherwise, this PMC on this CPU is either free or
3245                  * in system-wide mode.
3246                  */
3247
3248                 pmc_save_cpu_binding(&pb);
3249                 pmc_select_cpu(cpu);
3250
3251                 pc  = pmc_pcpu[cpu];
3252                 phw = pc->pc_hwpmcs[ri];
3253
3254                 /*
3255                  * XXX do we need some kind of 'forced' disable?
3256                  */
3257
3258                 if (phw->phw_pmc == NULL) {
3259                         if (request == PMC_STATE_DISABLED &&
3260                             (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED)) {
3261                                 phw->phw_state &= ~PMC_PHW_FLAG_IS_ENABLED;
3262                                 PMC_MARK_ROW_STANDALONE(ri);
3263                         } else if (request == PMC_STATE_FREE &&
3264                             (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0) {
3265                                 phw->phw_state |=  PMC_PHW_FLAG_IS_ENABLED;
3266                                 PMC_UNMARK_ROW_STANDALONE(ri);
3267                         }
3268                         /* other cases are a no-op */
3269                 } else
3270                         error = EBUSY;
3271
3272                 pmc_restore_cpu_binding(&pb);
3273         }
3274         break;
3275
3276
3277         /*
3278          * Allocate a PMC.
3279          */
3280
3281         case PMC_OP_PMCALLOCATE:
3282         {
3283                 int adjri, n;
3284                 u_int cpu;
3285                 uint32_t caps;
3286                 struct pmc *pmc;
3287                 enum pmc_mode mode;
3288                 struct pmc_hw *phw;
3289                 struct pmc_binding pb;
3290                 struct pmc_classdep *pcd;
3291                 struct pmc_op_pmcallocate pa;
3292
3293                 if ((error = copyin(arg, &pa, sizeof(pa))) != 0)
3294                         break;
3295
3296                 caps = pa.pm_caps;
3297                 mode = pa.pm_mode;
3298                 cpu  = pa.pm_cpu;
3299
3300                 if ((mode != PMC_MODE_SS  &&  mode != PMC_MODE_SC  &&
3301                      mode != PMC_MODE_TS  &&  mode != PMC_MODE_TC) ||
3302                     (cpu != (u_int) PMC_CPU_ANY && cpu >= pmc_cpu_max())) {
3303                         error = EINVAL;
3304                         break;
3305                 }
3306
3307                 /*
3308                  * Virtual PMCs should only ask for a default CPU.
3309                  * System mode PMCs need to specify a non-default CPU.
3310                  */
3311
3312                 if ((PMC_IS_VIRTUAL_MODE(mode) && cpu != (u_int) PMC_CPU_ANY) ||
3313                     (PMC_IS_SYSTEM_MODE(mode) && cpu == (u_int) PMC_CPU_ANY)) {
3314                         error = EINVAL;
3315                         break;
3316                 }
3317
3318                 /*
3319                  * Check that an inactive CPU is not being asked for.
3320                  */
3321
3322                 if (PMC_IS_SYSTEM_MODE(mode) && !pmc_cpu_is_active(cpu)) {
3323                         error = ENXIO;
3324                         break;
3325                 }
3326
3327                 /*
3328                  * Refuse an allocation for a system-wide PMC if this
3329                  * process has been jailed, or if this process lacks
3330                  * super-user credentials and the sysctl tunable
3331                  * 'security.bsd.unprivileged_syspmcs' is zero.
3332                  */
3333
3334                 if (PMC_IS_SYSTEM_MODE(mode)) {
3335                         if (jailed(curthread->td_ucred)) {
3336                                 error = EPERM;
3337                                 break;
3338                         }
3339                         if (!pmc_unprivileged_syspmcs) {
3340                                 error = priv_check(curthread,
3341                                     PRIV_PMC_SYSTEM);
3342                                 if (error)
3343                                         break;
3344                         }
3345                 }
3346
3347                 /*
3348                  * Look for valid values for 'pm_flags'
3349                  */
3350
3351                 if ((pa.pm_flags & ~(PMC_F_DESCENDANTS | PMC_F_LOG_PROCCSW |
3352                     PMC_F_LOG_PROCEXIT | PMC_F_CALLCHAIN)) != 0) {
3353                         error = EINVAL;
3354                         break;
3355                 }
3356
3357                 /* process logging options are not allowed for system PMCs */
3358                 if (PMC_IS_SYSTEM_MODE(mode) && (pa.pm_flags &
3359                     (PMC_F_LOG_PROCCSW | PMC_F_LOG_PROCEXIT))) {
3360                         error = EINVAL;
3361                         break;
3362                 }
3363
3364                 /*
3365                  * All sampling mode PMCs need to be able to interrupt the
3366                  * CPU.
3367                  */
3368                 if (PMC_IS_SAMPLING_MODE(mode))
3369                         caps |= PMC_CAP_INTERRUPT;
3370
3371                 /* A valid class specifier should have been passed in. */
3372                 for (n = 0; n < md->pmd_nclass; n++)
3373                         if (md->pmd_classdep[n].pcd_class == pa.pm_class)
3374                                 break;
3375                 if (n == md->pmd_nclass) {
3376                         error = EINVAL;
3377                         break;
3378                 }
3379
3380                 /* The requested PMC capabilities should be feasible. */
3381                 if ((md->pmd_classdep[n].pcd_caps & caps) != caps) {
3382                         error = EOPNOTSUPP;
3383                         break;
3384                 }
3385
3386                 PMCDBG4(PMC,ALL,2, "event=%d caps=0x%x mode=%d cpu=%d",
3387                     pa.pm_ev, caps, mode, cpu);
3388
3389                 pmc = pmc_allocate_pmc_descriptor();
3390                 pmc->pm_id    = PMC_ID_MAKE_ID(cpu,pa.pm_mode,pa.pm_class,
3391                     PMC_ID_INVALID);
3392                 pmc->pm_event = pa.pm_ev;
3393                 pmc->pm_state = PMC_STATE_FREE;
3394                 pmc->pm_caps  = caps;
3395                 pmc->pm_flags = pa.pm_flags;
3396
3397                 /* switch thread to CPU 'cpu' */
3398                 pmc_save_cpu_binding(&pb);
3399
3400 #define PMC_IS_SHAREABLE_PMC(cpu, n)                            \
3401         (pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_state &           \
3402          PMC_PHW_FLAG_IS_SHAREABLE)
3403 #define PMC_IS_UNALLOCATED(cpu, n)                              \
3404         (pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_pmc == NULL)
3405
3406                 if (PMC_IS_SYSTEM_MODE(mode)) {
3407                         pmc_select_cpu(cpu);
3408                         for (n = 0; n < (int) md->pmd_npmc; n++) {
3409                                 pcd = pmc_ri_to_classdep(md, n, &adjri);
3410                                 if (pmc_can_allocate_row(n, mode) == 0 &&
3411                                     pmc_can_allocate_rowindex(
3412                                             curthread->td_proc, n, cpu) == 0 &&
3413                                     (PMC_IS_UNALLOCATED(cpu, n) ||
3414                                      PMC_IS_SHAREABLE_PMC(cpu, n)) &&
3415                                     pcd->pcd_allocate_pmc(cpu, adjri, pmc,
3416                                         &pa) == 0)
3417                                         break;
3418                         }
3419                 } else {
3420                         /* Process virtual mode */
3421                         for (n = 0; n < (int) md->pmd_npmc; n++) {
3422                                 pcd = pmc_ri_to_classdep(md, n, &adjri);
3423                                 if (pmc_can_allocate_row(n, mode) == 0 &&
3424                                     pmc_can_allocate_rowindex(
3425                                             curthread->td_proc, n,
3426                                             PMC_CPU_ANY) == 0 &&
3427                                     pcd->pcd_allocate_pmc(curthread->td_oncpu,
3428                                         adjri, pmc, &pa) == 0)
3429                                         break;
3430                         }
3431                 }
3432
3433 #undef  PMC_IS_UNALLOCATED
3434 #undef  PMC_IS_SHAREABLE_PMC
3435
3436                 pmc_restore_cpu_binding(&pb);
3437
3438                 if (n == (int) md->pmd_npmc) {
3439                         pmc_destroy_pmc_descriptor(pmc);
3440                         pmc = NULL;
3441                         error = EINVAL;
3442                         break;
3443                 }
3444
3445                 /* Fill in the correct value in the ID field */
3446                 pmc->pm_id = PMC_ID_MAKE_ID(cpu,mode,pa.pm_class,n);
3447
3448                 PMCDBG5(PMC,ALL,2, "ev=%d class=%d mode=%d n=%d -> pmcid=%x",
3449                     pmc->pm_event, pa.pm_class, mode, n, pmc->pm_id);
3450
3451                 /* Process mode PMCs with logging enabled need log files */
3452                 if (pmc->pm_flags & (PMC_F_LOG_PROCEXIT | PMC_F_LOG_PROCCSW))
3453                         pmc->pm_flags |= PMC_F_NEEDS_LOGFILE;
3454
3455                 /* All system mode sampling PMCs require a log file */
3456                 if (PMC_IS_SAMPLING_MODE(mode) && PMC_IS_SYSTEM_MODE(mode))
3457                         pmc->pm_flags |= PMC_F_NEEDS_LOGFILE;
3458
3459                 /*
3460                  * Configure global pmc's immediately
3461                  */
3462
3463                 if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pmc))) {
3464
3465                         pmc_save_cpu_binding(&pb);
3466                         pmc_select_cpu(cpu);
3467
3468                         phw = pmc_pcpu[cpu]->pc_hwpmcs[n];
3469                         pcd = pmc_ri_to_classdep(md, n, &adjri);
3470
3471                         if ((phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0 ||
3472                             (error = pcd->pcd_config_pmc(cpu, adjri, pmc)) != 0) {
3473                                 (void) pcd->pcd_release_pmc(cpu, adjri, pmc);
3474                                 pmc_destroy_pmc_descriptor(pmc);
3475                                 pmc = NULL;
3476                                 pmc_restore_cpu_binding(&pb);
3477                                 error = EPERM;
3478                                 break;
3479                         }
3480
3481                         pmc_restore_cpu_binding(&pb);
3482                 }
3483
3484                 pmc->pm_state    = PMC_STATE_ALLOCATED;
3485
3486                 /*
3487                  * mark row disposition
3488                  */
3489
3490                 if (PMC_IS_SYSTEM_MODE(mode))
3491                         PMC_MARK_ROW_STANDALONE(n);
3492                 else
3493                         PMC_MARK_ROW_THREAD(n);
3494
3495                 /*
3496                  * Register this PMC with the current thread as its owner.
3497                  */
3498
3499                 if ((error =
3500                     pmc_register_owner(curthread->td_proc, pmc)) != 0) {
3501                         pmc_release_pmc_descriptor(pmc);
3502                         pmc_destroy_pmc_descriptor(pmc);
3503                         pmc = NULL;
3504                         break;
3505                 }
3506
3507                 /*
3508                  * Return the allocated index.
3509                  */
3510
3511                 pa.pm_pmcid = pmc->pm_id;
3512
3513                 error = copyout(&pa, arg, sizeof(pa));
3514         }
3515         break;
3516
3517
3518         /*
3519          * Attach a PMC to a process.
3520          */
3521
3522         case PMC_OP_PMCATTACH:
3523         {
3524                 struct pmc *pm;
3525                 struct proc *p;
3526                 struct pmc_op_pmcattach a;
3527
3528                 sx_assert(&pmc_sx, SX_XLOCKED);
3529
3530                 if ((error = copyin(arg, &a, sizeof(a))) != 0)
3531                         break;
3532
3533                 if (a.pm_pid < 0) {
3534                         error = EINVAL;
3535                         break;
3536                 } else if (a.pm_pid == 0)
3537                         a.pm_pid = td->td_proc->p_pid;
3538
3539                 if ((error = pmc_find_pmc(a.pm_pmc, &pm)) != 0)
3540                         break;
3541
3542                 if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm))) {
3543                         error = EINVAL;
3544                         break;
3545                 }
3546
3547                 /* PMCs may be (re)attached only when allocated or stopped */
3548                 if (pm->pm_state == PMC_STATE_RUNNING) {
3549                         error = EBUSY;
3550                         break;
3551                 } else if (pm->pm_state != PMC_STATE_ALLOCATED &&
3552                     pm->pm_state != PMC_STATE_STOPPED) {
3553                         error = EINVAL;
3554                         break;
3555                 }
3556
3557                 /* lookup pid */
3558                 if ((p = pfind(a.pm_pid)) == NULL) {
3559                         error = ESRCH;
3560                         break;
3561                 }
3562
3563                 /*
3564                  * Ignore processes that are working on exiting.
3565                  */
3566                 if (p->p_flag & P_WEXIT) {
3567                         error = ESRCH;
3568                         PROC_UNLOCK(p); /* pfind() returns a locked process */
3569                         break;
3570                 }
3571
3572                 /*
3573                  * we are allowed to attach a PMC to a process if
3574                  * we can debug it.
3575                  */
3576                 error = p_candebug(curthread, p);
3577
3578                 PROC_UNLOCK(p);
3579
3580                 if (error == 0)
3581                         error = pmc_attach_process(p, pm);
3582         }
3583         break;
3584
3585
3586         /*
3587          * Detach an attached PMC from a process.
3588          */
3589
3590         case PMC_OP_PMCDETACH:
3591         {
3592                 struct pmc *pm;
3593                 struct proc *p;
3594                 struct pmc_op_pmcattach a;
3595
3596                 if ((error = copyin(arg, &a, sizeof(a))) != 0)
3597                         break;
3598
3599                 if (a.pm_pid < 0) {
3600                         error = EINVAL;
3601                         break;
3602                 } else if (a.pm_pid == 0)
3603                         a.pm_pid = td->td_proc->p_pid;
3604
3605                 if ((error = pmc_find_pmc(a.pm_pmc, &pm)) != 0)
3606                         break;
3607
3608                 if ((p = pfind(a.pm_pid)) == NULL) {
3609                         error = ESRCH;
3610                         break;
3611                 }
3612
3613                 /*
3614                  * Treat processes that are in the process of exiting
3615                  * as if they were not present.
3616                  */
3617
3618                 if (p->p_flag & P_WEXIT)
3619                         error = ESRCH;
3620
3621                 PROC_UNLOCK(p); /* pfind() returns a locked process */
3622
3623                 if (error == 0)
3624                         error = pmc_detach_process(p, pm);
3625         }
3626         break;
3627
3628
3629         /*
3630          * Retrieve the MSR number associated with the counter
3631          * 'pmc_id'.  This allows processes to directly use RDPMC
3632          * instructions to read their PMCs, without the overhead of a
3633          * system call.
3634          */
3635
3636         case PMC_OP_PMCGETMSR:
3637         {
3638                 int adjri, ri;
3639                 struct pmc *pm;
3640                 struct pmc_target *pt;
3641                 struct pmc_op_getmsr gm;
3642                 struct pmc_classdep *pcd;
3643
3644                 PMC_DOWNGRADE_SX();
3645
3646                 if ((error = copyin(arg, &gm, sizeof(gm))) != 0)
3647                         break;
3648
3649                 if ((error = pmc_find_pmc(gm.pm_pmcid, &pm)) != 0)
3650                         break;
3651
3652                 /*
3653                  * The allocated PMC has to be a process virtual PMC,
3654                  * i.e., of type MODE_T[CS].  Global PMCs can only be
3655                  * read using the PMCREAD operation since they may be
3656                  * allocated on a different CPU than the one we could
3657                  * be running on at the time of the RDPMC instruction.
3658                  *
3659                  * The GETMSR operation is not allowed for PMCs that
3660                  * are inherited across processes.
3661                  */
3662
3663                 if (!PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)) ||
3664                     (pm->pm_flags & PMC_F_DESCENDANTS)) {
3665                         error = EINVAL;
3666                         break;
3667                 }
3668
3669                 /*
3670                  * It only makes sense to use a RDPMC (or its
3671                  * equivalent instruction on non-x86 architectures) on
3672                  * a process that has allocated and attached a PMC to
3673                  * itself.  Conversely the PMC is only allowed to have
3674                  * one process attached to it -- its owner.
3675                  */
3676
3677                 if ((pt = LIST_FIRST(&pm->pm_targets)) == NULL ||
3678                     LIST_NEXT(pt, pt_next) != NULL ||
3679                     pt->pt_process->pp_proc != pm->pm_owner->po_owner) {
3680                         error = EINVAL;
3681                         break;
3682                 }
3683
3684                 ri = PMC_TO_ROWINDEX(pm);
3685                 pcd = pmc_ri_to_classdep(md, ri, &adjri);
3686
3687                 /* PMC class has no 'GETMSR' support */
3688                 if (pcd->pcd_get_msr == NULL) {
3689                         error = ENOSYS;
3690                         break;
3691                 }
3692
3693                 if ((error = (*pcd->pcd_get_msr)(adjri, &gm.pm_msr)) < 0)
3694                         break;
3695
3696                 if ((error = copyout(&gm, arg, sizeof(gm))) < 0)
3697                         break;
3698
3699                 /*
3700                  * Mark our process as using MSRs.  Update machine
3701                  * state using a forced context switch.
3702                  */
3703
3704                 pt->pt_process->pp_flags |= PMC_PP_ENABLE_MSR_ACCESS;
3705                 pmc_force_context_switch();
3706
3707         }
3708         break;
3709
3710         /*
3711          * Release an allocated PMC
3712          */
3713
3714         case PMC_OP_PMCRELEASE:
3715         {
3716                 pmc_id_t pmcid;
3717                 struct pmc *pm;
3718                 struct pmc_owner *po;
3719                 struct pmc_op_simple sp;
3720
3721                 /*
3722                  * Find PMC pointer for the named PMC.
3723                  *
3724                  * Use pmc_release_pmc_descriptor() to switch off the
3725                  * PMC, remove all its target threads, and remove the
3726                  * PMC from its owner's list.
3727                  *
3728                  * Remove the owner record if this is the last PMC
3729                  * owned.
3730                  *
3731                  * Free up space.
3732                  */
3733
3734                 if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
3735                         break;
3736
3737                 pmcid = sp.pm_pmcid;
3738
3739                 if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
3740                         break;
3741
3742                 po = pm->pm_owner;
3743                 pmc_release_pmc_descriptor(pm);
3744                 pmc_maybe_remove_owner(po);
3745                 pmc_destroy_pmc_descriptor(pm);
3746         }
3747         break;
3748
3749
3750         /*
3751          * Read and/or write a PMC.
3752          */
3753
3754         case PMC_OP_PMCRW:
3755         {
3756                 int adjri;
3757                 struct pmc *pm;
3758                 uint32_t cpu, ri;
3759                 pmc_value_t oldvalue;
3760                 struct pmc_binding pb;
3761                 struct pmc_op_pmcrw prw;
3762                 struct pmc_classdep *pcd;
3763                 struct pmc_op_pmcrw *pprw;
3764
3765                 PMC_DOWNGRADE_SX();
3766
3767                 if ((error = copyin(arg, &prw, sizeof(prw))) != 0)
3768                         break;
3769
3770                 ri = 0;
3771                 PMCDBG2(PMC,OPS,1, "rw id=%d flags=0x%x", prw.pm_pmcid,
3772                     prw.pm_flags);
3773
3774                 /* must have at least one flag set */
3775                 if ((prw.pm_flags & (PMC_F_OLDVALUE|PMC_F_NEWVALUE)) == 0) {
3776                         error = EINVAL;
3777                         break;
3778                 }
3779
3780                 /* locate pmc descriptor */
3781                 if ((error = pmc_find_pmc(prw.pm_pmcid, &pm)) != 0)
3782                         break;
3783
3784                 /* Can't read a PMC that hasn't been started. */
3785                 if (pm->pm_state != PMC_STATE_ALLOCATED &&
3786                     pm->pm_state != PMC_STATE_STOPPED &&
3787                     pm->pm_state != PMC_STATE_RUNNING) {
3788                         error = EINVAL;
3789                         break;
3790                 }
3791
3792                 /* writing a new value is allowed only for 'STOPPED' pmcs */
3793                 if (pm->pm_state == PMC_STATE_RUNNING &&
3794                     (prw.pm_flags & PMC_F_NEWVALUE)) {
3795                         error = EBUSY;
3796                         break;
3797                 }
3798
3799                 if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) {
3800
3801                         /*
3802                          * If this PMC is attached to its owner (i.e.,
3803                          * the process requesting this operation) and
3804                          * is running, then attempt to get an
3805                          * upto-date reading from hardware for a READ.
3806                          * Writes are only allowed when the PMC is
3807                          * stopped, so only update the saved value
3808                          * field.
3809                          *
3810                          * If the PMC is not running, or is not
3811                          * attached to its owner, read/write to the
3812                          * savedvalue field.
3813                          */
3814
3815                         ri = PMC_TO_ROWINDEX(pm);
3816                         pcd = pmc_ri_to_classdep(md, ri, &adjri);
3817
3818                         mtx_pool_lock_spin(pmc_mtxpool, pm);
3819                         cpu = curthread->td_oncpu;
3820
3821                         if (prw.pm_flags & PMC_F_OLDVALUE) {
3822                                 if ((pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) &&
3823                                     (pm->pm_state == PMC_STATE_RUNNING))
3824                                         error = (*pcd->pcd_read_pmc)(cpu, adjri,
3825                                             &oldvalue);
3826                                 else
3827                                         oldvalue = pm->pm_gv.pm_savedvalue;
3828                         }
3829                         if (prw.pm_flags & PMC_F_NEWVALUE)
3830                                 pm->pm_gv.pm_savedvalue = prw.pm_value;
3831
3832                         mtx_pool_unlock_spin(pmc_mtxpool, pm);
3833
3834                 } else { /* System mode PMCs */
3835                         cpu = PMC_TO_CPU(pm);
3836                         ri  = PMC_TO_ROWINDEX(pm);
3837                         pcd = pmc_ri_to_classdep(md, ri, &adjri);
3838
3839                         if (!pmc_cpu_is_active(cpu)) {
3840                                 error = ENXIO;
3841                                 break;
3842                         }
3843
3844                         /* move this thread to CPU 'cpu' */
3845                         pmc_save_cpu_binding(&pb);
3846                         pmc_select_cpu(cpu);
3847
3848                         critical_enter();
3849                         /* save old value */
3850                         if (prw.pm_flags & PMC_F_OLDVALUE)
3851                                 if ((error = (*pcd->pcd_read_pmc)(cpu, adjri,
3852                                          &oldvalue)))
3853                                         goto error;
3854                         /* write out new value */
3855                         if (prw.pm_flags & PMC_F_NEWVALUE)
3856                                 error = (*pcd->pcd_write_pmc)(cpu, adjri,
3857                                     prw.pm_value);
3858                 error:
3859                         critical_exit();
3860                         pmc_restore_cpu_binding(&pb);
3861                         if (error)
3862                                 break;
3863                 }
3864
3865                 pprw = (struct pmc_op_pmcrw *) arg;
3866
3867 #ifdef  HWPMC_DEBUG
3868                 if (prw.pm_flags & PMC_F_NEWVALUE)
3869                         PMCDBG3(PMC,OPS,2, "rw id=%d new %jx -> old %jx",
3870                             ri, prw.pm_value, oldvalue);
3871                 else if (prw.pm_flags & PMC_F_OLDVALUE)
3872                         PMCDBG2(PMC,OPS,2, "rw id=%d -> old %jx", ri, oldvalue);
3873 #endif
3874
3875                 /* return old value if requested */
3876                 if (prw.pm_flags & PMC_F_OLDVALUE)
3877                         if ((error = copyout(&oldvalue, &pprw->pm_value,
3878                                  sizeof(prw.pm_value))))
3879                                 break;
3880
3881         }
3882         break;
3883
3884
3885         /*
3886          * Set the sampling rate for a sampling mode PMC and the
3887          * initial count for a counting mode PMC.
3888          */
3889
3890         case PMC_OP_PMCSETCOUNT:
3891         {
3892                 struct pmc *pm;
3893                 struct pmc_op_pmcsetcount sc;
3894
3895                 PMC_DOWNGRADE_SX();
3896
3897                 if ((error = copyin(arg, &sc, sizeof(sc))) != 0)
3898                         break;
3899
3900                 if ((error = pmc_find_pmc(sc.pm_pmcid, &pm)) != 0)
3901                         break;
3902
3903                 if (pm->pm_state == PMC_STATE_RUNNING) {
3904                         error = EBUSY;
3905                         break;
3906                 }
3907
3908                 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
3909                         pm->pm_sc.pm_reloadcount = sc.pm_count;
3910                 else
3911                         pm->pm_sc.pm_initial = sc.pm_count;
3912         }
3913         break;
3914
3915
3916         /*
3917          * Start a PMC.
3918          */
3919
3920         case PMC_OP_PMCSTART:
3921         {
3922                 pmc_id_t pmcid;
3923                 struct pmc *pm;
3924                 struct pmc_op_simple sp;
3925
3926                 sx_assert(&pmc_sx, SX_XLOCKED);
3927
3928                 if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
3929                         break;
3930
3931                 pmcid = sp.pm_pmcid;
3932
3933                 if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
3934                         break;
3935
3936                 KASSERT(pmcid == pm->pm_id,
3937                     ("[pmc,%d] pmcid %x != id %x", __LINE__,
3938                         pm->pm_id, pmcid));
3939
3940                 if (pm->pm_state == PMC_STATE_RUNNING) /* already running */
3941                         break;
3942                 else if (pm->pm_state != PMC_STATE_STOPPED &&
3943                     pm->pm_state != PMC_STATE_ALLOCATED) {
3944                         error = EINVAL;
3945                         break;
3946                 }
3947
3948                 error = pmc_start(pm);
3949         }
3950         break;
3951
3952
3953         /*
3954          * Stop a PMC.
3955          */
3956
3957         case PMC_OP_PMCSTOP:
3958         {
3959                 pmc_id_t pmcid;
3960                 struct pmc *pm;
3961                 struct pmc_op_simple sp;
3962
3963                 PMC_DOWNGRADE_SX();
3964
3965                 if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
3966                         break;
3967
3968                 pmcid = sp.pm_pmcid;
3969
3970                 /*
3971                  * Mark the PMC as inactive and invoke the MD stop
3972                  * routines if needed.
3973                  */
3974
3975                 if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
3976                         break;
3977
3978                 KASSERT(pmcid == pm->pm_id,
3979                     ("[pmc,%d] pmc id %x != pmcid %x", __LINE__,
3980                         pm->pm_id, pmcid));
3981
3982                 if (pm->pm_state == PMC_STATE_STOPPED) /* already stopped */
3983                         break;
3984                 else if (pm->pm_state != PMC_STATE_RUNNING) {
3985                         error = EINVAL;
3986                         break;
3987                 }
3988
3989                 error = pmc_stop(pm);
3990         }
3991         break;
3992
3993
3994         /*
3995          * Write a user supplied value to the log file.
3996          */
3997
3998         case PMC_OP_WRITELOG:
3999         {
4000                 struct pmc_op_writelog wl;
4001                 struct pmc_owner *po;
4002
4003                 PMC_DOWNGRADE_SX();
4004
4005                 if ((error = copyin(arg, &wl, sizeof(wl))) != 0)
4006                         break;
4007
4008                 if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) {
4009                         error = EINVAL;
4010                         break;
4011                 }
4012
4013                 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) {
4014                         error = EINVAL;
4015                         break;
4016                 }
4017
4018                 error = pmclog_process_userlog(po, &wl);
4019         }
4020         break;
4021
4022
4023         default:
4024                 error = EINVAL;
4025                 break;
4026         }
4027
4028         if (is_sx_locked != 0) {
4029                 if (is_sx_downgraded)
4030                         sx_sunlock(&pmc_sx);
4031                 else
4032                         sx_xunlock(&pmc_sx);
4033         }
4034
4035         if (error)
4036                 atomic_add_int(&pmc_stats.pm_syscall_errors, 1);
4037
4038         PICKUP_GIANT();
4039
4040         return error;
4041 }
4042
4043 /*
4044  * Helper functions
4045  */
4046
4047
4048 /*
4049  * Mark the thread as needing callchain capture and post an AST.  The
4050  * actual callchain capture will be done in a context where it is safe
4051  * to take page faults.
4052  */
4053
4054 static void
4055 pmc_post_callchain_callback(void)
4056 {
4057         struct thread *td;
4058
4059         td = curthread;
4060
4061         /*
4062          * If there is multiple PMCs for the same interrupt ignore new post
4063          */
4064         if (td->td_pflags & TDP_CALLCHAIN)
4065                 return;
4066
4067         /*
4068          * Mark this thread as needing callchain capture.
4069          * `td->td_pflags' will be safe to touch because this thread
4070          * was in user space when it was interrupted.
4071          */
4072         td->td_pflags |= TDP_CALLCHAIN;
4073
4074         /*
4075          * Don't let this thread migrate between CPUs until callchain
4076          * capture completes.
4077          */
4078         sched_pin();
4079
4080         return;
4081 }
4082
4083 /*
4084  * Interrupt processing.
4085  *
4086  * Find a free slot in the per-cpu array of samples and capture the
4087  * current callchain there.  If a sample was successfully added, a bit
4088  * is set in mask 'pmc_cpumask' denoting that the DO_SAMPLES hook
4089  * needs to be invoked from the clock handler.
4090  *
4091  * This function is meant to be called from an NMI handler.  It cannot
4092  * use any of the locking primitives supplied by the OS.
4093  */
4094
4095 int
4096 pmc_process_interrupt(int cpu, int ring, struct pmc *pm, struct trapframe *tf,
4097     int inuserspace)
4098 {
4099         int error, callchaindepth;
4100         struct thread *td;
4101         struct pmc_sample *ps;
4102         struct pmc_samplebuffer *psb;
4103
4104         error = 0;
4105
4106         /*
4107          * Allocate space for a sample buffer.
4108          */
4109         psb = pmc_pcpu[cpu]->pc_sb[ring];
4110
4111         ps = psb->ps_write;
4112         if (ps->ps_nsamples) {  /* in use, reader hasn't caught up */
4113                 CPU_SET_ATOMIC(cpu, &pm->pm_stalled);
4114                 atomic_add_int(&pmc_stats.pm_intr_bufferfull, 1);
4115                 PMCDBG6(SAM,INT,1,"(spc) cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d",
4116                     cpu, pm, (void *) tf, inuserspace,
4117                     (int) (psb->ps_write - psb->ps_samples),
4118                     (int) (psb->ps_read - psb->ps_samples));
4119                 callchaindepth = 1;
4120                 error = ENOMEM;
4121                 goto done;
4122         }
4123
4124
4125         /* Fill in entry. */
4126         PMCDBG6(SAM,INT,1,"cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d", cpu, pm,
4127             (void *) tf, inuserspace,
4128             (int) (psb->ps_write - psb->ps_samples),
4129             (int) (psb->ps_read - psb->ps_samples));
4130
4131         KASSERT(pm->pm_runcount >= 0,
4132             ("[pmc,%d] pm=%p runcount %d", __LINE__, (void *) pm,
4133                 pm->pm_runcount));
4134
4135         atomic_add_rel_int(&pm->pm_runcount, 1);        /* hold onto PMC */
4136
4137         ps->ps_pmc = pm;
4138         if ((td = curthread) && td->td_proc)
4139                 ps->ps_pid = td->td_proc->p_pid;
4140         else
4141                 ps->ps_pid = -1;
4142         ps->ps_cpu = cpu;
4143         ps->ps_td = td;
4144         ps->ps_flags = inuserspace ? PMC_CC_F_USERSPACE : 0;
4145
4146         callchaindepth = (pm->pm_flags & PMC_F_CALLCHAIN) ?
4147             pmc_callchaindepth : 1;
4148
4149         if (callchaindepth == 1)
4150                 ps->ps_pc[0] = PMC_TRAPFRAME_TO_PC(tf);
4151         else {
4152                 /*
4153                  * Kernel stack traversals can be done immediately,
4154                  * while we defer to an AST for user space traversals.
4155                  */
4156                 if (!inuserspace) {
4157                         callchaindepth =
4158                             pmc_save_kernel_callchain(ps->ps_pc,
4159                                 callchaindepth, tf);
4160                 } else {
4161                         pmc_post_callchain_callback();
4162                         callchaindepth = PMC_SAMPLE_INUSE;
4163                 }
4164         }
4165
4166         ps->ps_nsamples = callchaindepth;       /* mark entry as in use */
4167
4168         /* increment write pointer, modulo ring buffer size */
4169         ps++;
4170         if (ps == psb->ps_fence)
4171                 psb->ps_write = psb->ps_samples;
4172         else
4173                 psb->ps_write = ps;
4174
4175  done:
4176         /* mark CPU as needing processing */
4177         if (callchaindepth != PMC_SAMPLE_INUSE)
4178                 CPU_SET_ATOMIC(cpu, &pmc_cpumask);
4179
4180         return (error);
4181 }
4182
4183 /*
4184  * Capture a user call chain.  This function will be called from ast()
4185  * before control returns to userland and before the process gets
4186  * rescheduled.
4187  */
4188
4189 static void
4190 pmc_capture_user_callchain(int cpu, int ring, struct trapframe *tf)
4191 {
4192         struct pmc *pm;
4193         struct thread *td;
4194         struct pmc_sample *ps, *ps_end;
4195         struct pmc_samplebuffer *psb;
4196 #ifdef  INVARIANTS
4197         int ncallchains;
4198 #endif
4199
4200         psb = pmc_pcpu[cpu]->pc_sb[ring];
4201         td = curthread;
4202
4203         KASSERT(td->td_pflags & TDP_CALLCHAIN,
4204             ("[pmc,%d] Retrieving callchain for thread that doesn't want it",
4205                 __LINE__));
4206
4207 #ifdef  INVARIANTS
4208         ncallchains = 0;
4209 #endif
4210
4211         /*
4212          * Iterate through all deferred callchain requests.
4213          * Walk from the current read pointer to the current
4214          * write pointer.
4215          */
4216
4217         ps = psb->ps_read;
4218         ps_end = psb->ps_write;
4219         do {
4220                 if (ps->ps_nsamples != PMC_SAMPLE_INUSE)
4221                         goto next;
4222                 if (ps->ps_td != td)
4223                         goto next;
4224
4225                 KASSERT(ps->ps_cpu == cpu,
4226                     ("[pmc,%d] cpu mismatch ps_cpu=%d pcpu=%d", __LINE__,
4227                         ps->ps_cpu, PCPU_GET(cpuid)));
4228
4229                 pm = ps->ps_pmc;
4230
4231                 KASSERT(pm->pm_flags & PMC_F_CALLCHAIN,
4232                     ("[pmc,%d] Retrieving callchain for PMC that doesn't "
4233                         "want it", __LINE__));
4234
4235                 KASSERT(pm->pm_runcount > 0,
4236                     ("[pmc,%d] runcount %d", __LINE__, pm->pm_runcount));
4237
4238                 /*
4239                  * Retrieve the callchain and mark the sample buffer
4240                  * as 'processable' by the timer tick sweep code.
4241                  */
4242                 ps->ps_nsamples = pmc_save_user_callchain(ps->ps_pc,
4243                     pmc_callchaindepth, tf);
4244
4245 #ifdef  INVARIANTS
4246                 ncallchains++;
4247 #endif
4248
4249 next:
4250                 /* increment the pointer, modulo sample ring size */
4251                 if (++ps == psb->ps_fence)
4252                         ps = psb->ps_samples;
4253         } while (ps != ps_end);
4254
4255         KASSERT(ncallchains > 0,
4256             ("[pmc,%d] cpu %d didn't find a sample to collect", __LINE__,
4257                 cpu));
4258
4259         KASSERT(td->td_pinned == 1,
4260             ("[pmc,%d] invalid td_pinned value", __LINE__));
4261         sched_unpin();  /* Can migrate safely now. */
4262
4263         /* mark CPU as needing processing */
4264         CPU_SET_ATOMIC(cpu, &pmc_cpumask);
4265
4266         return;
4267 }
4268
4269 /*
4270  * Process saved PC samples.
4271  */
4272
4273 static void
4274 pmc_process_samples(int cpu, int ring)
4275 {
4276         struct pmc *pm;
4277         int adjri, n;
4278         struct thread *td;
4279         struct pmc_owner *po;
4280         struct pmc_sample *ps;
4281         struct pmc_classdep *pcd;
4282         struct pmc_samplebuffer *psb;
4283
4284         KASSERT(PCPU_GET(cpuid) == cpu,
4285             ("[pmc,%d] not on the correct CPU pcpu=%d cpu=%d", __LINE__,
4286                 PCPU_GET(cpuid), cpu));
4287
4288         psb = pmc_pcpu[cpu]->pc_sb[ring];
4289
4290         for (n = 0; n < pmc_nsamples; n++) { /* bound on #iterations */
4291
4292                 ps = psb->ps_read;
4293                 if (ps->ps_nsamples == PMC_SAMPLE_FREE)
4294                         break;
4295
4296                 pm = ps->ps_pmc;
4297
4298                 KASSERT(pm->pm_runcount > 0,
4299                     ("[pmc,%d] pm=%p runcount %d", __LINE__, (void *) pm,
4300                         pm->pm_runcount));
4301
4302                 po = pm->pm_owner;
4303
4304                 KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)),
4305                     ("[pmc,%d] pmc=%p non-sampling mode=%d", __LINE__,
4306                         pm, PMC_TO_MODE(pm)));
4307
4308                 /* Ignore PMCs that have been switched off */
4309                 if (pm->pm_state != PMC_STATE_RUNNING)
4310                         goto entrydone;
4311
4312                 /* If there is a pending AST wait for completion */
4313                 if (ps->ps_nsamples == PMC_SAMPLE_INUSE) {
4314                         /* Need a rescan at a later time. */
4315                         CPU_SET_ATOMIC(cpu, &pmc_cpumask);
4316                         break;
4317                 }
4318
4319                 PMCDBG6(SAM,OPS,1,"cpu=%d pm=%p n=%d fl=%x wr=%d rd=%d", cpu,
4320                     pm, ps->ps_nsamples, ps->ps_flags,
4321                     (int) (psb->ps_write - psb->ps_samples),
4322                     (int) (psb->ps_read - psb->ps_samples));
4323
4324                 /*
4325                  * If this is a process-mode PMC that is attached to
4326                  * its owner, and if the PC is in user mode, update
4327                  * profiling statistics like timer-based profiling
4328                  * would have done.
4329                  */
4330                 if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) {
4331                         if (ps->ps_flags & PMC_CC_F_USERSPACE) {
4332                                 td = FIRST_THREAD_IN_PROC(po->po_owner);
4333                                 addupc_intr(td, ps->ps_pc[0], 1);
4334                         }
4335                         goto entrydone;
4336                 }
4337
4338                 /*
4339                  * Otherwise, this is either a sampling mode PMC that
4340                  * is attached to a different process than its owner,
4341                  * or a system-wide sampling PMC.  Dispatch a log
4342                  * entry to the PMC's owner process.
4343                  */
4344                 pmclog_process_callchain(pm, ps);
4345
4346         entrydone:
4347                 ps->ps_nsamples = 0; /* mark entry as free */
4348                 atomic_subtract_rel_int(&pm->pm_runcount, 1);
4349
4350                 /* increment read pointer, modulo sample size */
4351                 if (++ps == psb->ps_fence)
4352                         psb->ps_read = psb->ps_samples;
4353                 else
4354                         psb->ps_read = ps;
4355         }
4356
4357         atomic_add_int(&pmc_stats.pm_log_sweeps, 1);
4358
4359         /* Do not re-enable stalled PMCs if we failed to process any samples */
4360         if (n == 0)
4361                 return;
4362
4363         /*
4364          * Restart any stalled sampling PMCs on this CPU.
4365          *
4366          * If the NMI handler sets the pm_stalled field of a PMC after
4367          * the check below, we'll end up processing the stalled PMC at
4368          * the next hardclock tick.
4369          */
4370         for (n = 0; n < md->pmd_npmc; n++) {
4371                 pcd = pmc_ri_to_classdep(md, n, &adjri);
4372                 KASSERT(pcd != NULL,
4373                     ("[pmc,%d] null pcd ri=%d", __LINE__, n));
4374                 (void) (*pcd->pcd_get_config)(cpu,adjri,&pm);
4375
4376                 if (pm == NULL ||                        /* !cfg'ed */
4377                     pm->pm_state != PMC_STATE_RUNNING || /* !active */
4378                     !PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) || /* !sampling */
4379                     !CPU_ISSET(cpu, &pm->pm_cpustate) || /* !desired */
4380                     !CPU_ISSET(cpu, &pm->pm_stalled)) /* !stalled */
4381                         continue;
4382
4383                 CPU_CLR_ATOMIC(cpu, &pm->pm_stalled);
4384                 (*pcd->pcd_start_pmc)(cpu, adjri);
4385         }
4386 }
4387
4388 /*
4389  * Event handlers.
4390  */
4391
4392 /*
4393  * Handle a process exit.
4394  *
4395  * Remove this process from all hash tables.  If this process
4396  * owned any PMCs, turn off those PMCs and deallocate them,
4397  * removing any associations with target processes.
4398  *
4399  * This function will be called by the last 'thread' of a
4400  * process.
4401  *
4402  * XXX This eventhandler gets called early in the exit process.
4403  * Consider using a 'hook' invocation from thread_exit() or equivalent
4404  * spot.  Another negative is that kse_exit doesn't seem to call
4405  * exit1() [??].
4406  *
4407  */
4408
4409 static void
4410 pmc_process_exit(void *arg __unused, struct proc *p)
4411 {
4412         struct pmc *pm;
4413         int adjri, cpu;
4414         unsigned int ri;
4415         int is_using_hwpmcs;
4416         struct pmc_owner *po;
4417         struct pmc_process *pp;
4418         struct pmc_classdep *pcd;
4419         pmc_value_t newvalue, tmp;
4420
4421         PROC_LOCK(p);
4422         is_using_hwpmcs = p->p_flag & P_HWPMC;
4423         PROC_UNLOCK(p);
4424
4425         /*
4426          * Log a sysexit event to all SS PMC owners.
4427          */
4428         LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
4429             if (po->po_flags & PMC_PO_OWNS_LOGFILE)
4430                     pmclog_process_sysexit(po, p->p_pid);
4431
4432         if (!is_using_hwpmcs)
4433                 return;
4434
4435         PMC_GET_SX_XLOCK();
4436         PMCDBG3(PRC,EXT,1,"process-exit proc=%p (%d, %s)", p, p->p_pid,
4437             p->p_comm);
4438
4439         /*
4440          * Since this code is invoked by the last thread in an exiting
4441          * process, we would have context switched IN at some prior
4442          * point.  However, with PREEMPTION, kernel mode context
4443          * switches may happen any time, so we want to disable a
4444          * context switch OUT till we get any PMCs targetting this
4445          * process off the hardware.
4446          *
4447          * We also need to atomically remove this process'
4448          * entry from our target process hash table, using
4449          * PMC_FLAG_REMOVE.
4450          */
4451         PMCDBG3(PRC,EXT,1, "process-exit proc=%p (%d, %s)", p, p->p_pid,
4452             p->p_comm);
4453
4454         critical_enter(); /* no preemption */
4455
4456         cpu = curthread->td_oncpu;
4457
4458         if ((pp = pmc_find_process_descriptor(p,
4459                  PMC_FLAG_REMOVE)) != NULL) {
4460
4461                 PMCDBG2(PRC,EXT,2,
4462                     "process-exit proc=%p pmc-process=%p", p, pp);
4463
4464                 /*
4465                  * The exiting process could the target of
4466                  * some PMCs which will be running on
4467                  * currently executing CPU.
4468                  *
4469                  * We need to turn these PMCs off like we
4470                  * would do at context switch OUT time.
4471                  */
4472                 for (ri = 0; ri < md->pmd_npmc; ri++) {
4473
4474                         /*
4475                          * Pick up the pmc pointer from hardware
4476                          * state similar to the CSW_OUT code.
4477                          */
4478                         pm = NULL;
4479
4480                         pcd = pmc_ri_to_classdep(md, ri, &adjri);
4481
4482                         (void) (*pcd->pcd_get_config)(cpu, adjri, &pm);
4483
4484                         PMCDBG2(PRC,EXT,2, "ri=%d pm=%p", ri, pm);
4485
4486                         if (pm == NULL ||
4487                             !PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)))
4488                                 continue;
4489
4490                         PMCDBG4(PRC,EXT,2, "ppmcs[%d]=%p pm=%p "
4491                             "state=%d", ri, pp->pp_pmcs[ri].pp_pmc,
4492                             pm, pm->pm_state);
4493
4494                         KASSERT(PMC_TO_ROWINDEX(pm) == ri,
4495                             ("[pmc,%d] ri mismatch pmc(%d) ri(%d)",
4496                                 __LINE__, PMC_TO_ROWINDEX(pm), ri));
4497
4498                         KASSERT(pm == pp->pp_pmcs[ri].pp_pmc,
4499                             ("[pmc,%d] pm %p != pp_pmcs[%d] %p",
4500                                 __LINE__, pm, ri, pp->pp_pmcs[ri].pp_pmc));
4501
4502                         KASSERT(pm->pm_runcount > 0,
4503                             ("[pmc,%d] bad runcount ri %d rc %d",
4504                                 __LINE__, ri, pm->pm_runcount));
4505
4506                         /*
4507                          * Change desired state, and then stop if not
4508                          * stalled. This two-step dance should avoid
4509                          * race conditions where an interrupt re-enables
4510                          * the PMC after this code has already checked
4511                          * the pm_stalled flag.
4512                          */
4513                         if (CPU_ISSET(cpu, &pm->pm_cpustate)) {
4514                                 CPU_CLR_ATOMIC(cpu, &pm->pm_cpustate);
4515                                 if (!CPU_ISSET(cpu, &pm->pm_stalled)) {
4516                                         (void) pcd->pcd_stop_pmc(cpu, adjri);
4517                                         pcd->pcd_read_pmc(cpu, adjri,
4518                                             &newvalue);
4519                                         tmp = newvalue -
4520                                             PMC_PCPU_SAVED(cpu,ri);
4521
4522                                         mtx_pool_lock_spin(pmc_mtxpool, pm);
4523                                         pm->pm_gv.pm_savedvalue += tmp;
4524                                         pp->pp_pmcs[ri].pp_pmcval += tmp;
4525                                         mtx_pool_unlock_spin(pmc_mtxpool, pm);
4526                                 }
4527                         }
4528
4529                         atomic_subtract_rel_int(&pm->pm_runcount,1);
4530
4531                         KASSERT((int) pm->pm_runcount >= 0,
4532                             ("[pmc,%d] runcount is %d", __LINE__, ri));
4533
4534                         (void) pcd->pcd_config_pmc(cpu, adjri, NULL);
4535                 }
4536
4537                 /*
4538                  * Inform the MD layer of this pseudo "context switch
4539                  * out"
4540                  */
4541                 (void) md->pmd_switch_out(pmc_pcpu[cpu], pp);
4542
4543                 critical_exit(); /* ok to be pre-empted now */
4544
4545                 /*
4546                  * Unlink this process from the PMCs that are
4547                  * targetting it.  This will send a signal to
4548                  * all PMC owner's whose PMCs are orphaned.
4549                  *
4550                  * Log PMC value at exit time if requested.
4551                  */
4552                 for (ri = 0; ri < md->pmd_npmc; ri++)
4553                         if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) {
4554                                 if (pm->pm_flags & PMC_F_NEEDS_LOGFILE &&
4555                                     PMC_IS_COUNTING_MODE(PMC_TO_MODE(pm)))
4556                                         pmclog_process_procexit(pm, pp);
4557                                 pmc_unlink_target_process(pm, pp);
4558                         }
4559                 free(pp, M_PMC);
4560
4561         } else
4562                 critical_exit(); /* pp == NULL */
4563
4564
4565         /*
4566          * If the process owned PMCs, free them up and free up
4567          * memory.
4568          */
4569         if ((po = pmc_find_owner_descriptor(p)) != NULL) {
4570                 pmc_remove_owner(po);
4571                 pmc_destroy_owner_descriptor(po);
4572         }
4573
4574         sx_xunlock(&pmc_sx);
4575 }
4576
4577 /*
4578  * Handle a process fork.
4579  *
4580  * If the parent process 'p1' is under HWPMC monitoring, then copy
4581  * over any attached PMCs that have 'do_descendants' semantics.
4582  */
4583
4584 static void
4585 pmc_process_fork(void *arg __unused, struct proc *p1, struct proc *newproc,
4586     int flags)
4587 {
4588         int is_using_hwpmcs;
4589         unsigned int ri;
4590         uint32_t do_descendants;
4591         struct pmc *pm;
4592         struct pmc_owner *po;
4593         struct pmc_process *ppnew, *ppold;
4594
4595         (void) flags;           /* unused parameter */
4596
4597         PROC_LOCK(p1);
4598         is_using_hwpmcs = p1->p_flag & P_HWPMC;
4599         PROC_UNLOCK(p1);
4600
4601         /*
4602          * If there are system-wide sampling PMCs active, we need to
4603          * log all fork events to their owner's logs.
4604          */
4605
4606         LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
4607             if (po->po_flags & PMC_PO_OWNS_LOGFILE)
4608                     pmclog_process_procfork(po, p1->p_pid, newproc->p_pid);
4609
4610         if (!is_using_hwpmcs)
4611                 return;
4612
4613         PMC_GET_SX_XLOCK();
4614         PMCDBG4(PMC,FRK,1, "process-fork proc=%p (%d, %s) -> %p", p1,
4615             p1->p_pid, p1->p_comm, newproc);
4616
4617         /*
4618          * If the parent process (curthread->td_proc) is a
4619          * target of any PMCs, look for PMCs that are to be
4620          * inherited, and link these into the new process
4621          * descriptor.
4622          */
4623         if ((ppold = pmc_find_process_descriptor(curthread->td_proc,
4624                  PMC_FLAG_NONE)) == NULL)
4625                 goto done;              /* nothing to do */
4626
4627         do_descendants = 0;
4628         for (ri = 0; ri < md->pmd_npmc; ri++)
4629                 if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL)
4630                         do_descendants |= pm->pm_flags & PMC_F_DESCENDANTS;
4631         if (do_descendants == 0) /* nothing to do */
4632                 goto done;
4633
4634         /* allocate a descriptor for the new process  */
4635         if ((ppnew = pmc_find_process_descriptor(newproc,
4636                  PMC_FLAG_ALLOCATE)) == NULL)
4637                 goto done;
4638
4639         /*
4640          * Run through all PMCs that were targeting the old process
4641          * and which specified F_DESCENDANTS and attach them to the
4642          * new process.
4643          *
4644          * Log the fork event to all owners of PMCs attached to this
4645          * process, if not already logged.
4646          */
4647         for (ri = 0; ri < md->pmd_npmc; ri++)
4648                 if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL &&
4649                     (pm->pm_flags & PMC_F_DESCENDANTS)) {
4650                         pmc_link_target_process(pm, ppnew);
4651                         po = pm->pm_owner;
4652                         if (po->po_sscount == 0 &&
4653                             po->po_flags & PMC_PO_OWNS_LOGFILE)
4654                                 pmclog_process_procfork(po, p1->p_pid,
4655                                     newproc->p_pid);
4656                 }
4657
4658         /*
4659          * Now mark the new process as being tracked by this driver.
4660          */
4661         PROC_LOCK(newproc);
4662         newproc->p_flag |= P_HWPMC;
4663         PROC_UNLOCK(newproc);
4664
4665  done:
4666         sx_xunlock(&pmc_sx);
4667 }
4668
4669 static void
4670 pmc_kld_load(void *arg __unused, linker_file_t lf)
4671 {
4672         struct pmc_owner *po;
4673
4674         sx_slock(&pmc_sx);
4675
4676         /*
4677          * Notify owners of system sampling PMCs about KLD operations.
4678          */
4679         LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
4680                 if (po->po_flags & PMC_PO_OWNS_LOGFILE)
4681                         pmclog_process_map_in(po, (pid_t) -1,
4682                             (uintfptr_t) lf->address, lf->filename);
4683
4684         /*
4685          * TODO: Notify owners of (all) process-sampling PMCs too.
4686          */
4687
4688         sx_sunlock(&pmc_sx);
4689 }
4690
4691 static void
4692 pmc_kld_unload(void *arg __unused, const char *filename __unused,
4693     caddr_t address, size_t size)
4694 {
4695         struct pmc_owner *po;
4696
4697         sx_slock(&pmc_sx);
4698
4699         LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
4700                 if (po->po_flags & PMC_PO_OWNS_LOGFILE)
4701                         pmclog_process_map_out(po, (pid_t) -1,
4702                             (uintfptr_t) address, (uintfptr_t) address + size);
4703
4704         /*
4705          * TODO: Notify owners of process-sampling PMCs.
4706          */
4707
4708         sx_sunlock(&pmc_sx);
4709 }
4710
4711 /*
4712  * initialization
4713  */
4714 static const char *
4715 pmc_name_of_pmcclass(enum pmc_class class)
4716 {
4717
4718         switch (class) {
4719 #undef  __PMC_CLASS
4720 #define __PMC_CLASS(S,V,D)                                              \
4721         case PMC_CLASS_##S:                                             \
4722                 return #S;
4723         __PMC_CLASSES();
4724         default:
4725                 return ("<unknown>");
4726         }
4727 }
4728
4729 /*
4730  * Base class initializer: allocate structure and set default classes.
4731  */
4732 struct pmc_mdep *
4733 pmc_mdep_alloc(int nclasses)
4734 {
4735         struct pmc_mdep *md;
4736         int     n;
4737
4738         /* SOFT + md classes */
4739         n = 1 + nclasses;
4740         md = malloc(sizeof(struct pmc_mdep) + n *
4741             sizeof(struct pmc_classdep), M_PMC, M_WAITOK|M_ZERO);
4742         md->pmd_nclass = n;
4743
4744         /* Add base class. */
4745         pmc_soft_initialize(md);
4746         return md;
4747 }
4748
4749 void
4750 pmc_mdep_free(struct pmc_mdep *md)
4751 {
4752         pmc_soft_finalize(md);
4753         free(md, M_PMC);
4754 }
4755
4756 static int
4757 generic_switch_in(struct pmc_cpu *pc, struct pmc_process *pp)
4758 {
4759         (void) pc; (void) pp;
4760
4761         return (0);
4762 }
4763
4764 static int
4765 generic_switch_out(struct pmc_cpu *pc, struct pmc_process *pp)
4766 {
4767         (void) pc; (void) pp;
4768
4769         return (0);
4770 }
4771
4772 static struct pmc_mdep *
4773 pmc_generic_cpu_initialize(void)
4774 {
4775         struct pmc_mdep *md;
4776
4777         md = pmc_mdep_alloc(0);
4778
4779         md->pmd_cputype    = PMC_CPU_GENERIC;
4780
4781         md->pmd_pcpu_init  = NULL;
4782         md->pmd_pcpu_fini  = NULL;
4783         md->pmd_switch_in  = generic_switch_in;
4784         md->pmd_switch_out = generic_switch_out;
4785
4786         return (md);
4787 }
4788
4789 static void
4790 pmc_generic_cpu_finalize(struct pmc_mdep *md)
4791 {
4792         (void) md;
4793 }
4794
4795
4796 static int
4797 pmc_initialize(void)
4798 {
4799         int c, cpu, error, n, ri;
4800         unsigned int maxcpu;
4801         struct pmc_binding pb;
4802         struct pmc_sample *ps;
4803         struct pmc_classdep *pcd;
4804         struct pmc_samplebuffer *sb;
4805
4806         md = NULL;
4807         error = 0;
4808
4809 #ifdef  HWPMC_DEBUG
4810         /* parse debug flags first */
4811         if (TUNABLE_STR_FETCH(PMC_SYSCTL_NAME_PREFIX "debugflags",
4812                 pmc_debugstr, sizeof(pmc_debugstr)))
4813                 pmc_debugflags_parse(pmc_debugstr,
4814                     pmc_debugstr+strlen(pmc_debugstr));
4815 #endif
4816
4817         PMCDBG1(MOD,INI,0, "PMC Initialize (version %x)", PMC_VERSION);
4818
4819         /* check kernel version */
4820         if (pmc_kernel_version != PMC_VERSION) {
4821                 if (pmc_kernel_version == 0)
4822                         printf("hwpmc: this kernel has not been compiled with "
4823                             "'options HWPMC_HOOKS'.\n");
4824                 else
4825                         printf("hwpmc: kernel version (0x%x) does not match "
4826                             "module version (0x%x).\n", pmc_kernel_version,
4827                             PMC_VERSION);
4828                 return EPROGMISMATCH;
4829         }
4830
4831         /*
4832          * check sysctl parameters
4833          */
4834
4835         if (pmc_hashsize <= 0) {
4836                 (void) printf("hwpmc: tunable \"hashsize\"=%d must be "
4837                     "greater than zero.\n", pmc_hashsize);
4838                 pmc_hashsize = PMC_HASH_SIZE;
4839         }
4840
4841         if (pmc_nsamples <= 0 || pmc_nsamples > 65535) {
4842                 (void) printf("hwpmc: tunable \"nsamples\"=%d out of "
4843                     "range.\n", pmc_nsamples);
4844                 pmc_nsamples = PMC_NSAMPLES;
4845         }
4846
4847         if (pmc_callchaindepth <= 0 ||
4848             pmc_callchaindepth > PMC_CALLCHAIN_DEPTH_MAX) {
4849                 (void) printf("hwpmc: tunable \"callchaindepth\"=%d out of "
4850                     "range - using %d.\n", pmc_callchaindepth,
4851                     PMC_CALLCHAIN_DEPTH_MAX);
4852                 pmc_callchaindepth = PMC_CALLCHAIN_DEPTH_MAX;
4853         }
4854
4855         md = pmc_md_initialize();
4856         if (md == NULL) {
4857                 /* Default to generic CPU. */
4858                 md = pmc_generic_cpu_initialize();
4859                 if (md == NULL)
4860                         return (ENOSYS);
4861         }
4862
4863         KASSERT(md->pmd_nclass >= 1 && md->pmd_npmc >= 1,
4864             ("[pmc,%d] no classes or pmcs", __LINE__));
4865
4866         /* Compute the map from row-indices to classdep pointers. */
4867         pmc_rowindex_to_classdep = malloc(sizeof(struct pmc_classdep *) *
4868             md->pmd_npmc, M_PMC, M_WAITOK|M_ZERO);
4869
4870         for (n = 0; n < md->pmd_npmc; n++)
4871                 pmc_rowindex_to_classdep[n] = NULL;
4872         for (ri = c = 0; c < md->pmd_nclass; c++) {
4873                 pcd = &md->pmd_classdep[c];
4874                 for (n = 0; n < pcd->pcd_num; n++, ri++)
4875                         pmc_rowindex_to_classdep[ri] = pcd;
4876         }
4877
4878         KASSERT(ri == md->pmd_npmc,
4879             ("[pmc,%d] npmc miscomputed: ri=%d, md->npmc=%d", __LINE__,
4880             ri, md->pmd_npmc));
4881
4882         maxcpu = pmc_cpu_max();
4883
4884         /* allocate space for the per-cpu array */
4885         pmc_pcpu = malloc(maxcpu * sizeof(struct pmc_cpu *), M_PMC,
4886             M_WAITOK|M_ZERO);
4887
4888         /* per-cpu 'saved values' for managing process-mode PMCs */
4889         pmc_pcpu_saved = malloc(sizeof(pmc_value_t) * maxcpu * md->pmd_npmc,
4890             M_PMC, M_WAITOK);
4891
4892         /* Perform CPU-dependent initialization. */
4893         pmc_save_cpu_binding(&pb);
4894         error = 0;
4895         for (cpu = 0; error == 0 && cpu < maxcpu; cpu++) {
4896                 if (!pmc_cpu_is_active(cpu))
4897                         continue;
4898                 pmc_select_cpu(cpu);
4899                 pmc_pcpu[cpu] = malloc(sizeof(struct pmc_cpu) +
4900                     md->pmd_npmc * sizeof(struct pmc_hw *), M_PMC,
4901                     M_WAITOK|M_ZERO);
4902                 if (md->pmd_pcpu_init)
4903                         error = md->pmd_pcpu_init(md, cpu);
4904                 for (n = 0; error == 0 && n < md->pmd_nclass; n++)
4905                         error = md->pmd_classdep[n].pcd_pcpu_init(md, cpu);
4906         }
4907         pmc_restore_cpu_binding(&pb);
4908
4909         if (error)
4910                 return (error);
4911
4912         /* allocate space for the sample array */
4913         for (cpu = 0; cpu < maxcpu; cpu++) {
4914                 if (!pmc_cpu_is_active(cpu))
4915                         continue;
4916
4917                 sb = malloc(sizeof(struct pmc_samplebuffer) +
4918                     pmc_nsamples * sizeof(struct pmc_sample), M_PMC,
4919                     M_WAITOK|M_ZERO);
4920                 sb->ps_read = sb->ps_write = sb->ps_samples;
4921                 sb->ps_fence = sb->ps_samples + pmc_nsamples;
4922
4923                 KASSERT(pmc_pcpu[cpu] != NULL,
4924                     ("[pmc,%d] cpu=%d Null per-cpu data", __LINE__, cpu));
4925
4926                 sb->ps_callchains = malloc(pmc_callchaindepth * pmc_nsamples *
4927                     sizeof(uintptr_t), M_PMC, M_WAITOK|M_ZERO);
4928
4929                 for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++)
4930                         ps->ps_pc = sb->ps_callchains +
4931                             (n * pmc_callchaindepth);
4932
4933                 pmc_pcpu[cpu]->pc_sb[PMC_HR] = sb;
4934
4935                 sb = malloc(sizeof(struct pmc_samplebuffer) +
4936                     pmc_nsamples * sizeof(struct pmc_sample), M_PMC,
4937                     M_WAITOK|M_ZERO);
4938                 sb->ps_read = sb->ps_write = sb->ps_samples;
4939                 sb->ps_fence = sb->ps_samples + pmc_nsamples;
4940
4941                 KASSERT(pmc_pcpu[cpu] != NULL,
4942                     ("[pmc,%d] cpu=%d Null per-cpu data", __LINE__, cpu));
4943
4944                 sb->ps_callchains = malloc(pmc_callchaindepth * pmc_nsamples *
4945                     sizeof(uintptr_t), M_PMC, M_WAITOK|M_ZERO);
4946
4947                 for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++)
4948                         ps->ps_pc = sb->ps_callchains +
4949                             (n * pmc_callchaindepth);
4950
4951                 pmc_pcpu[cpu]->pc_sb[PMC_SR] = sb;
4952         }
4953
4954         /* allocate space for the row disposition array */
4955         pmc_pmcdisp = malloc(sizeof(enum pmc_mode) * md->pmd_npmc,
4956             M_PMC, M_WAITOK|M_ZERO);
4957
4958         /* mark all PMCs as available */
4959         for (n = 0; n < (int) md->pmd_npmc; n++)
4960                 PMC_MARK_ROW_FREE(n);
4961
4962         /* allocate thread hash tables */
4963         pmc_ownerhash = hashinit(pmc_hashsize, M_PMC,
4964             &pmc_ownerhashmask);
4965
4966         pmc_processhash = hashinit(pmc_hashsize, M_PMC,
4967             &pmc_processhashmask);
4968         mtx_init(&pmc_processhash_mtx, "pmc-process-hash", "pmc-leaf",
4969             MTX_SPIN);
4970
4971         LIST_INIT(&pmc_ss_owners);
4972         pmc_ss_count = 0;
4973
4974         /* allocate a pool of spin mutexes */
4975         pmc_mtxpool = mtx_pool_create("pmc-leaf", pmc_mtxpool_size,
4976             MTX_SPIN);
4977
4978         PMCDBG4(MOD,INI,1, "pmc_ownerhash=%p, mask=0x%lx "
4979             "targethash=%p mask=0x%lx", pmc_ownerhash, pmc_ownerhashmask,
4980             pmc_processhash, pmc_processhashmask);
4981
4982         /* register process {exit,fork,exec} handlers */
4983         pmc_exit_tag = EVENTHANDLER_REGISTER(process_exit,
4984             pmc_process_exit, NULL, EVENTHANDLER_PRI_ANY);
4985         pmc_fork_tag = EVENTHANDLER_REGISTER(process_fork,
4986             pmc_process_fork, NULL, EVENTHANDLER_PRI_ANY);
4987
4988         /* register kld event handlers */
4989         pmc_kld_load_tag = EVENTHANDLER_REGISTER(kld_load, pmc_kld_load,
4990             NULL, EVENTHANDLER_PRI_ANY);
4991         pmc_kld_unload_tag = EVENTHANDLER_REGISTER(kld_unload, pmc_kld_unload,
4992             NULL, EVENTHANDLER_PRI_ANY);
4993
4994         /* initialize logging */
4995         pmclog_initialize();
4996
4997         /* set hook functions */
4998         pmc_intr = md->pmd_intr;
4999         pmc_hook = pmc_hook_handler;
5000
5001         if (error == 0) {
5002                 printf(PMC_MODULE_NAME ":");
5003                 for (n = 0; n < (int) md->pmd_nclass; n++) {
5004                         pcd = &md->pmd_classdep[n];
5005                         printf(" %s/%d/%d/0x%b",
5006                             pmc_name_of_pmcclass(pcd->pcd_class),
5007                             pcd->pcd_num,
5008                             pcd->pcd_width,
5009                             pcd->pcd_caps,
5010                             "\20"
5011                             "\1INT\2USR\3SYS\4EDG\5THR"
5012                             "\6REA\7WRI\10INV\11QUA\12PRC"
5013                             "\13TAG\14CSC");
5014                 }
5015                 printf("\n");
5016         }
5017
5018         return (error);
5019 }
5020
5021 /* prepare to be unloaded */
5022 static void
5023 pmc_cleanup(void)
5024 {
5025         int c, cpu;
5026         unsigned int maxcpu;
5027         struct pmc_ownerhash *ph;
5028         struct pmc_owner *po, *tmp;
5029         struct pmc_binding pb;
5030 #ifdef  HWPMC_DEBUG
5031         struct pmc_processhash *prh;
5032 #endif
5033
5034         PMCDBG0(MOD,INI,0, "cleanup");
5035
5036         /* switch off sampling */
5037         CPU_ZERO(&pmc_cpumask);
5038         pmc_intr = NULL;
5039
5040         sx_xlock(&pmc_sx);
5041         if (pmc_hook == NULL) { /* being unloaded already */
5042                 sx_xunlock(&pmc_sx);
5043                 return;
5044         }
5045
5046         pmc_hook = NULL; /* prevent new threads from entering module */
5047
5048         /* deregister event handlers */
5049         EVENTHANDLER_DEREGISTER(process_fork, pmc_fork_tag);
5050         EVENTHANDLER_DEREGISTER(process_exit, pmc_exit_tag);
5051         EVENTHANDLER_DEREGISTER(kld_load, pmc_kld_load_tag);
5052         EVENTHANDLER_DEREGISTER(kld_unload, pmc_kld_unload_tag);
5053
5054         /* send SIGBUS to all owner threads, free up allocations */
5055         if (pmc_ownerhash)
5056                 for (ph = pmc_ownerhash;
5057                      ph <= &pmc_ownerhash[pmc_ownerhashmask];
5058                      ph++) {
5059                         LIST_FOREACH_SAFE(po, ph, po_next, tmp) {
5060                                 pmc_remove_owner(po);
5061
5062                                 /* send SIGBUS to owner processes */
5063                                 PMCDBG3(MOD,INI,2, "cleanup signal proc=%p "
5064                                     "(%d, %s)", po->po_owner,
5065                                     po->po_owner->p_pid,
5066                                     po->po_owner->p_comm);
5067
5068                                 PROC_LOCK(po->po_owner);
5069                                 kern_psignal(po->po_owner, SIGBUS);
5070                                 PROC_UNLOCK(po->po_owner);
5071
5072                                 pmc_destroy_owner_descriptor(po);
5073                         }
5074                 }
5075
5076         /* reclaim allocated data structures */
5077         if (pmc_mtxpool)
5078                 mtx_pool_destroy(&pmc_mtxpool);
5079
5080         mtx_destroy(&pmc_processhash_mtx);
5081         if (pmc_processhash) {
5082 #ifdef  HWPMC_DEBUG
5083                 struct pmc_process *pp;
5084
5085                 PMCDBG0(MOD,INI,3, "destroy process hash");
5086                 for (prh = pmc_processhash;
5087                      prh <= &pmc_processhash[pmc_processhashmask];
5088                      prh++)
5089                         LIST_FOREACH(pp, prh, pp_next)
5090                             PMCDBG1(MOD,INI,3, "pid=%d", pp->pp_proc->p_pid);
5091 #endif
5092
5093                 hashdestroy(pmc_processhash, M_PMC, pmc_processhashmask);
5094                 pmc_processhash = NULL;
5095         }
5096
5097         if (pmc_ownerhash) {
5098                 PMCDBG0(MOD,INI,3, "destroy owner hash");
5099                 hashdestroy(pmc_ownerhash, M_PMC, pmc_ownerhashmask);
5100                 pmc_ownerhash = NULL;
5101         }
5102
5103         KASSERT(LIST_EMPTY(&pmc_ss_owners),
5104             ("[pmc,%d] Global SS owner list not empty", __LINE__));
5105         KASSERT(pmc_ss_count == 0,
5106             ("[pmc,%d] Global SS count not empty", __LINE__));
5107
5108         /* do processor and pmc-class dependent cleanup */
5109         maxcpu = pmc_cpu_max();
5110
5111         PMCDBG0(MOD,INI,3, "md cleanup");
5112         if (md) {
5113                 pmc_save_cpu_binding(&pb);
5114                 for (cpu = 0; cpu < maxcpu; cpu++) {
5115                         PMCDBG2(MOD,INI,1,"pmc-cleanup cpu=%d pcs=%p",
5116                             cpu, pmc_pcpu[cpu]);
5117                         if (!pmc_cpu_is_active(cpu) || pmc_pcpu[cpu] == NULL)
5118                                 continue;
5119                         pmc_select_cpu(cpu);
5120                         for (c = 0; c < md->pmd_nclass; c++)
5121                                 md->pmd_classdep[c].pcd_pcpu_fini(md, cpu);
5122                         if (md->pmd_pcpu_fini)
5123                                 md->pmd_pcpu_fini(md, cpu);
5124                 }
5125
5126                 if (md->pmd_cputype == PMC_CPU_GENERIC)
5127                         pmc_generic_cpu_finalize(md);
5128                 else
5129                         pmc_md_finalize(md);
5130
5131                 pmc_mdep_free(md);
5132                 md = NULL;
5133                 pmc_restore_cpu_binding(&pb);
5134         }
5135
5136         /* Free per-cpu descriptors. */
5137         for (cpu = 0; cpu < maxcpu; cpu++) {
5138                 if (!pmc_cpu_is_active(cpu))
5139                         continue;
5140                 KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_HR] != NULL,
5141                     ("[pmc,%d] Null hw cpu sample buffer cpu=%d", __LINE__,
5142                         cpu));
5143                 KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_SR] != NULL,
5144                     ("[pmc,%d] Null sw cpu sample buffer cpu=%d", __LINE__,
5145                         cpu));
5146                 free(pmc_pcpu[cpu]->pc_sb[PMC_HR]->ps_callchains, M_PMC);
5147                 free(pmc_pcpu[cpu]->pc_sb[PMC_HR], M_PMC);
5148                 free(pmc_pcpu[cpu]->pc_sb[PMC_SR]->ps_callchains, M_PMC);
5149                 free(pmc_pcpu[cpu]->pc_sb[PMC_SR], M_PMC);
5150                 free(pmc_pcpu[cpu], M_PMC);
5151         }
5152
5153         free(pmc_pcpu, M_PMC);
5154         pmc_pcpu = NULL;
5155
5156         free(pmc_pcpu_saved, M_PMC);
5157         pmc_pcpu_saved = NULL;
5158
5159         if (pmc_pmcdisp) {
5160                 free(pmc_pmcdisp, M_PMC);
5161                 pmc_pmcdisp = NULL;
5162         }
5163
5164         if (pmc_rowindex_to_classdep) {
5165                 free(pmc_rowindex_to_classdep, M_PMC);
5166                 pmc_rowindex_to_classdep = NULL;
5167         }
5168
5169         pmclog_shutdown();
5170
5171         sx_xunlock(&pmc_sx);    /* we are done */
5172 }
5173
5174 /*
5175  * The function called at load/unload.
5176  */
5177
5178 static int
5179 load (struct module *module __unused, int cmd, void *arg __unused)
5180 {
5181         int error;
5182
5183         error = 0;
5184
5185         switch (cmd) {
5186         case MOD_LOAD :
5187                 /* initialize the subsystem */
5188                 error = pmc_initialize();
5189                 if (error != 0)
5190                         break;
5191                 PMCDBG2(MOD,INI,1, "syscall=%d maxcpu=%d",
5192                     pmc_syscall_num, pmc_cpu_max());
5193                 break;
5194
5195
5196         case MOD_UNLOAD :
5197         case MOD_SHUTDOWN:
5198                 pmc_cleanup();
5199                 PMCDBG0(MOD,INI,1, "unloaded");
5200                 break;
5201
5202         default :
5203                 error = EINVAL; /* XXX should panic(9) */
5204                 break;
5205         }
5206
5207         return error;
5208 }
5209
5210 /* memory pool */
5211 MALLOC_DEFINE(M_PMC, "pmc", "Memory space for the PMC module");