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/sched.h>
  54 #include <sys/signalvar.h>
  55 #include <sys/smp.h>
  56 #include <sys/sx.h>
  57 #include <sys/sysctl.h>
  58 #include <sys/sysent.h>
  59 #include <sys/systm.h>
  60 #include <sys/vnode.h>
  61
  62 #include <sys/linker.h>         /* needs to be after <sys/malloc.h> */
  63
  64 #include <machine/atomic.h>
  65 #include <machine/md_var.h>
  66
  67 #include <vm/vm.h>
  68 #include <vm/vm_extern.h>
  69 #include <vm/pmap.h>
  70 #include <vm/vm_map.h>
  71 #include <vm/vm_object.h>
  72
  73 #include "hwpmc_soft.h"
  74
  75 /*
  76  * Types
  77  */
  78
  79 enum pmc_flags {
  80         PMC_FLAG_NONE     = 0x00, /* do nothing */
  81         PMC_FLAG_REMOVE   = 0x01, /* atomically remove entry from hash */
  82         PMC_FLAG_ALLOCATE = 0x02, /* add entry to hash if not found */
  83 };
  84
  85 /*
  86  * The offset in sysent where the syscall is allocated.
  87  */
  88
  89 static int pmc_syscall_num = NO_SYSCALL;
  90 struct pmc_cpu          **pmc_pcpu;      /* per-cpu state */
  91 pmc_value_t             *pmc_pcpu_saved; /* saved PMC values: CSW handling */
  92
  93 #define PMC_PCPU_SAVED(C,R)     pmc_pcpu_saved[(R) + md->pmd_npmc*(C)]
  94
  95 struct mtx_pool         *pmc_mtxpool;
  96 static int              *pmc_pmcdisp;    /* PMC row dispositions */
  97
  98 #define PMC_ROW_DISP_IS_FREE(R)         (pmc_pmcdisp[(R)] == 0)
  99 #define PMC_ROW_DISP_IS_THREAD(R)       (pmc_pmcdisp[(R)] > 0)
 100 #define PMC_ROW_DISP_IS_STANDALONE(R)   (pmc_pmcdisp[(R)] < 0)
 101
 102 #define PMC_MARK_ROW_FREE(R) do {                                         \
 103         pmc_pmcdisp[(R)] = 0;                                             \
 104 } while (0)
 105
 106 #define PMC_MARK_ROW_STANDALONE(R) do {                                   \
 107         KASSERT(pmc_pmcdisp[(R)] <= 0, ("[pmc,%d] row disposition error", \
 108                     __LINE__));                                           \
 109         atomic_add_int(&pmc_pmcdisp[(R)], -1);                            \
 110         KASSERT(pmc_pmcdisp[(R)] >= (-pmc_cpu_max_active()),              \
 111                 ("[pmc,%d] row disposition error", __LINE__));            \
 112 } while (0)
 113
 114 #define PMC_UNMARK_ROW_STANDALONE(R) do {                                 \
 115         atomic_add_int(&pmc_pmcdisp[(R)], 1);                             \
 116         KASSERT(pmc_pmcdisp[(R)] <= 0, ("[pmc,%d] row disposition error", \
 117                     __LINE__));                                           \
 118 } while (0)
 119
 120 #define PMC_MARK_ROW_THREAD(R) do {                                       \
 121         KASSERT(pmc_pmcdisp[(R)] >= 0, ("[pmc,%d] row disposition error", \
 122                     __LINE__));                                           \
 123         atomic_add_int(&pmc_pmcdisp[(R)], 1);                             \
 124 } while (0)
 125
 126 #define PMC_UNMARK_ROW_THREAD(R) do {                                     \
 127         atomic_add_int(&pmc_pmcdisp[(R)], -1);                            \
 128         KASSERT(pmc_pmcdisp[(R)] >= 0, ("[pmc,%d] row disposition error", \
 129                     __LINE__));                                           \
 130 } while (0)
 131
 132
 133 /* various event handlers */
 134 static eventhandler_tag pmc_exit_tag, pmc_fork_tag, pmc_kld_load_tag,
 135     pmc_kld_unload_tag;
 136
 137 /* Module statistics */
 138 struct pmc_op_getdriverstats pmc_stats;
 139
 140 /* Machine/processor dependent operations */
 141 static struct pmc_mdep  *md;
 142
 143 /*
 144  * Hash tables mapping owner processes and target threads to PMCs.
 145  */
 146
 147 struct mtx pmc_processhash_mtx;         /* spin mutex */
 148 static u_long pmc_processhashmask;
 149 static LIST_HEAD(pmc_processhash, pmc_process)  *pmc_processhash;
 150
 151 /*
 152  * Hash table of PMC owner descriptors.  This table is protected by
 153  * the shared PMC "sx" lock.
 154  */
 155
 156 static u_long pmc_ownerhashmask;
 157 static LIST_HEAD(pmc_ownerhash, pmc_owner)      *pmc_ownerhash;
 158
 159 /*
 160  * List of PMC owners with system-wide sampling PMCs.
 161  */
 162
 163 static LIST_HEAD(, pmc_owner)                   pmc_ss_owners;
 164
 165
 166 /*
 167  * A map of row indices to classdep structures.
 168  */
 169 static struct pmc_classdep **pmc_rowindex_to_classdep;
 170
 171 /*
 172  * Prototypes
 173  */
 174
 175 #ifdef  DEBUG
 176 static int      pmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS);
 177 static int      pmc_debugflags_parse(char *newstr, char *fence);
 178 #endif
 179
 180 static int      load(struct module *module, int cmd, void *arg);
 181 static int      pmc_attach_process(struct proc *p, struct pmc *pm);
 182 static struct pmc *pmc_allocate_pmc_descriptor(void);
 183 static struct pmc_owner *pmc_allocate_owner_descriptor(struct proc *p);
 184 static int      pmc_attach_one_process(struct proc *p, struct pmc *pm);
 185 static int      pmc_can_allocate_rowindex(struct proc *p, unsigned int ri,
 186     int cpu);
 187 static int      pmc_can_attach(struct pmc *pm, struct proc *p);
 188 static void     pmc_capture_user_callchain(int cpu, int soft, struct trapframe *tf);
 189 static void     pmc_cleanup(void);
 190 static int      pmc_detach_process(struct proc *p, struct pmc *pm);
 191 static int      pmc_detach_one_process(struct proc *p, struct pmc *pm,
 192     int flags);
 193 static void     pmc_destroy_owner_descriptor(struct pmc_owner *po);
 194 static struct pmc_owner *pmc_find_owner_descriptor(struct proc *p);
 195 static int      pmc_find_pmc(pmc_id_t pmcid, struct pmc **pm);
 196 static struct pmc *pmc_find_pmc_descriptor_in_process(struct pmc_owner *po,
 197     pmc_id_t pmc);
 198 static struct pmc_process *pmc_find_process_descriptor(struct proc *p,
 199     uint32_t mode);
 200 static void     pmc_force_context_switch(void);
 201 static void     pmc_link_target_process(struct pmc *pm,
 202     struct pmc_process *pp);
 203 static void     pmc_log_all_process_mappings(struct pmc_owner *po);
 204 static void     pmc_log_kernel_mappings(struct pmc *pm);
 205 static void     pmc_log_process_mappings(struct pmc_owner *po, struct proc *p);
 206 static void     pmc_maybe_remove_owner(struct pmc_owner *po);
 207 static void     pmc_process_csw_in(struct thread *td);
 208 static void     pmc_process_csw_out(struct thread *td);
 209 static void     pmc_process_exit(void *arg, struct proc *p);
 210 static void     pmc_process_fork(void *arg, struct proc *p1,
 211     struct proc *p2, int n);
 212 static void     pmc_process_samples(int cpu, int soft);
 213 static void     pmc_release_pmc_descriptor(struct pmc *pmc);
 214 static void     pmc_remove_owner(struct pmc_owner *po);
 215 static void     pmc_remove_process_descriptor(struct pmc_process *pp);
 216 static void     pmc_restore_cpu_binding(struct pmc_binding *pb);
 217 static void     pmc_save_cpu_binding(struct pmc_binding *pb);
 218 static void     pmc_select_cpu(int cpu);
 219 static int      pmc_start(struct pmc *pm);
 220 static int      pmc_stop(struct pmc *pm);
 221 static int      pmc_syscall_handler(struct thread *td, void *syscall_args);
 222 static void     pmc_unlink_target_process(struct pmc *pmc,
 223     struct pmc_process *pp);
 224 static int generic_switch_in(struct pmc_cpu *pc, struct pmc_process *pp);
 225 static int generic_switch_out(struct pmc_cpu *pc, struct pmc_process *pp);
 226 static struct pmc_mdep *pmc_generic_cpu_initialize(void);
 227 static void pmc_generic_cpu_finalize(struct pmc_mdep *md);
 228
 229 /*
 230  * Kernel tunables and sysctl(8) interface.
 231  */
 232
 233 SYSCTL_DECL(_kern_hwpmc);
 234
 235 static int pmc_callchaindepth = PMC_CALLCHAIN_DEPTH;
 236 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "callchaindepth", &pmc_callchaindepth);
 237 SYSCTL_INT(_kern_hwpmc, OID_AUTO, callchaindepth, CTLFLAG_TUN|CTLFLAG_RD,
 238     &pmc_callchaindepth, 0, "depth of call chain records");
 239
 240 #ifdef  DEBUG
 241 struct pmc_debugflags pmc_debugflags = PMC_DEBUG_DEFAULT_FLAGS;
 242 char    pmc_debugstr[PMC_DEBUG_STRSIZE];
 243 TUNABLE_STR(PMC_SYSCTL_NAME_PREFIX "debugflags", pmc_debugstr,
 244     sizeof(pmc_debugstr));
 245 SYSCTL_PROC(_kern_hwpmc, OID_AUTO, debugflags,
 246     CTLTYPE_STRING|CTLFLAG_RW|CTLFLAG_TUN,
 247     0, 0, pmc_debugflags_sysctl_handler, "A", "debug flags");
 248 #endif
 249
 250 /*
 251  * kern.hwpmc.hashrows -- determines the number of rows in the
 252  * of the hash table used to look up threads
 253  */
 254
 255 static int pmc_hashsize = PMC_HASH_SIZE;
 256 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "hashsize", &pmc_hashsize);
 257 SYSCTL_INT(_kern_hwpmc, OID_AUTO, hashsize, CTLFLAG_TUN|CTLFLAG_RD,
 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 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "nsamples", &pmc_nsamples);
 266 SYSCTL_INT(_kern_hwpmc, OID_AUTO, nsamples, CTLFLAG_TUN|CTLFLAG_RD,
 267     &pmc_nsamples, 0, "number of PC samples per CPU");
 268
 269
 270 /*
 271  * kern.hwpmc.mtxpoolsize -- number of mutexes in the mutex pool.
 272  */
 273
 274 static int pmc_mtxpool_size = PMC_MTXPOOL_SIZE;
 275 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "mtxpoolsize", &pmc_mtxpool_size);
 276 SYSCTL_INT(_kern_hwpmc, OID_AUTO, mtxpoolsize, CTLFLAG_TUN|CTLFLAG_RD,
 277     &pmc_mtxpool_size, 0, "size of spin mutex pool");
 278
 279
 280 /*
 281  * security.bsd.unprivileged_syspmcs -- allow non-root processes to
 282  * allocate system-wide PMCs.
 283  *
 284  * Allowing unprivileged processes to allocate system PMCs is convenient
 285  * if system-wide measurements need to be taken concurrently with other
 286  * per-process measurements.  This feature is turned off by default.
 287  */
 288
 289 static int pmc_unprivileged_syspmcs = 0;
 290 TUNABLE_INT("security.bsd.unprivileged_syspmcs", &pmc_unprivileged_syspmcs);
 291 SYSCTL_INT(_security_bsd, OID_AUTO, unprivileged_syspmcs, CTLFLAG_RW,
 292     &pmc_unprivileged_syspmcs, 0,
 293     "allow unprivileged process to allocate system PMCs");
 294
 295 /*
 296  * Hash function.  Discard the lower 2 bits of the pointer since
 297  * these are always zero for our uses.  The hash multiplier is
 298  * round((2^LONG_BIT) * ((sqrt(5)-1)/2)).
 299  */
 300
 301 #if     LONG_BIT == 64
 302 #define _PMC_HM         11400714819323198486u
 303 #elif   LONG_BIT == 32
 304 #define _PMC_HM         2654435769u
 305 #else
 306 #error  Must know the size of 'long' to compile
 307 #endif
 308
 309 #define PMC_HASH_PTR(P,M)       ((((unsigned long) (P) >> 2) * _PMC_HM) & (M))
 310
 311 /*
 312  * Syscall structures
 313  */
 314
 315 /* The `sysent' for the new syscall */
 316 static struct sysent pmc_sysent = {
 317         2,                      /* sy_narg */
 318         pmc_syscall_handler     /* sy_call */
 319 };
 320
 321 static struct syscall_module_data pmc_syscall_mod = {
 322         load,
 323         NULL,
 324         &pmc_syscall_num,
 325         &pmc_sysent,
 326         { 0, NULL }
 327 };
 328
 329 static moduledata_t pmc_mod = {
 330         PMC_MODULE_NAME,
 331         syscall_module_handler,
 332         &pmc_syscall_mod
 333 };
 334
 335 DECLARE_MODULE(pmc, pmc_mod, SI_SUB_SMP, SI_ORDER_ANY);
 336 MODULE_VERSION(pmc, PMC_VERSION);
 337
 338 #ifdef  DEBUG
 339 enum pmc_dbgparse_state {
 340         PMCDS_WS,               /* in whitespace */
 341         PMCDS_MAJOR,            /* seen a major keyword */
 342         PMCDS_MINOR
 343 };
 344
 345 static int
 346 pmc_debugflags_parse(char *newstr, char *fence)
 347 {
 348         char c, *p, *q;
 349         struct pmc_debugflags *tmpflags;
 350         int error, found, *newbits, tmp;
 351         size_t kwlen;
 352
 353         tmpflags = malloc(sizeof(*tmpflags), M_PMC, M_WAITOK|M_ZERO);
 354
 355         p = newstr;
 356         error = 0;
 357
 358         for (; p < fence && (c = *p); p++) {
 359
 360                 /* skip white space */
 361                 if (c == ' ' || c == '\t')
 362                         continue;
 363
 364                 /* look for a keyword followed by "=" */
 365                 for (q = p; p < fence && (c = *p) && c != '='; p++)
 366                         ;
 367                 if (c != '=') {
 368                         error = EINVAL;
 369                         goto done;
 370                 }
 371
 372                 kwlen = p - q;
 373                 newbits = NULL;
 374
 375                 /* lookup flag group name */
 376 #define DBG_SET_FLAG_MAJ(S,F)                                           \
 377                 if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0)  \
 378                         newbits = &tmpflags->pdb_ ## F;
 379
 380                 DBG_SET_FLAG_MAJ("cpu",         CPU);
 381                 DBG_SET_FLAG_MAJ("csw",         CSW);
 382                 DBG_SET_FLAG_MAJ("logging",     LOG);
 383                 DBG_SET_FLAG_MAJ("module",      MOD);
 384                 DBG_SET_FLAG_MAJ("md",          MDP);
 385                 DBG_SET_FLAG_MAJ("owner",       OWN);
 386                 DBG_SET_FLAG_MAJ("pmc",         PMC);
 387                 DBG_SET_FLAG_MAJ("process",     PRC);
 388                 DBG_SET_FLAG_MAJ("sampling",    SAM);
 389
 390                 if (newbits == NULL) {
 391                         error = EINVAL;
 392                         goto done;
 393                 }
 394
 395                 p++;            /* skip the '=' */
 396
 397                 /* Now parse the individual flags */
 398                 tmp = 0;
 399         newflag:
 400                 for (q = p; p < fence && (c = *p); p++)
 401                         if (c == ' ' || c == '\t' || c == ',')
 402                                 break;
 403
 404                 /* p == fence or c == ws or c == "," or c == 0 */
 405
 406                 if ((kwlen = p - q) == 0) {
 407                         *newbits = tmp;
 408                         continue;
 409                 }
 410
 411                 found = 0;
 412 #define DBG_SET_FLAG_MIN(S,F)                                           \
 413                 if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0)  \
 414                         tmp |= found = (1 << PMC_DEBUG_MIN_ ## F)
 415
 416                 /* a '*' denotes all possible flags in the group */
 417                 if (kwlen == 1 && *q == '*')
 418                         tmp = found = ~0;
 419                 /* look for individual flag names */
 420                 DBG_SET_FLAG_MIN("allocaterow", ALR);
 421                 DBG_SET_FLAG_MIN("allocate",    ALL);
 422                 DBG_SET_FLAG_MIN("attach",      ATT);
 423                 DBG_SET_FLAG_MIN("bind",        BND);
 424                 DBG_SET_FLAG_MIN("config",      CFG);
 425                 DBG_SET_FLAG_MIN("exec",        EXC);
 426                 DBG_SET_FLAG_MIN("exit",        EXT);
 427                 DBG_SET_FLAG_MIN("find",        FND);
 428                 DBG_SET_FLAG_MIN("flush",       FLS);
 429                 DBG_SET_FLAG_MIN("fork",        FRK);
 430                 DBG_SET_FLAG_MIN("getbuf",      GTB);
 431                 DBG_SET_FLAG_MIN("hook",        PMH);
 432                 DBG_SET_FLAG_MIN("init",        INI);
 433                 DBG_SET_FLAG_MIN("intr",        INT);
 434                 DBG_SET_FLAG_MIN("linktarget",  TLK);
 435                 DBG_SET_FLAG_MIN("mayberemove", OMR);
 436                 DBG_SET_FLAG_MIN("ops",         OPS);
 437                 DBG_SET_FLAG_MIN("read",        REA);
 438                 DBG_SET_FLAG_MIN("register",    REG);
 439                 DBG_SET_FLAG_MIN("release",     REL);
 440                 DBG_SET_FLAG_MIN("remove",      ORM);
 441                 DBG_SET_FLAG_MIN("sample",      SAM);
 442                 DBG_SET_FLAG_MIN("scheduleio",  SIO);
 443                 DBG_SET_FLAG_MIN("select",      SEL);
 444                 DBG_SET_FLAG_MIN("signal",      SIG);
 445                 DBG_SET_FLAG_MIN("swi",         SWI);
 446                 DBG_SET_FLAG_MIN("swo",         SWO);
 447                 DBG_SET_FLAG_MIN("start",       STA);
 448                 DBG_SET_FLAG_MIN("stop",        STO);
 449                 DBG_SET_FLAG_MIN("syscall",     PMS);
 450                 DBG_SET_FLAG_MIN("unlinktarget", TUL);
 451                 DBG_SET_FLAG_MIN("write",       WRI);
 452                 if (found == 0) {
 453                         /* unrecognized flag name */
 454                         error = EINVAL;
 455                         goto done;
 456                 }
 457
 458                 if (c == 0 || c == ' ' || c == '\t') {  /* end of flag group */
 459                         *newbits = tmp;
 460                         continue;
 461                 }
 462
 463                 p++;
 464                 goto newflag;
 465         }
 466
 467         /* save the new flag set */
 468         bcopy(tmpflags, &pmc_debugflags, sizeof(pmc_debugflags));
 469
 470  done:
 471         free(tmpflags, M_PMC);
 472         return error;
 473 }
 474
 475 static int
 476 pmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS)
 477 {
 478         char *fence, *newstr;
 479         int error;
 480         unsigned int n;
 481
 482         (void) arg1; (void) arg2; /* unused parameters */
 483
 484         n = sizeof(pmc_debugstr);
 485         newstr = malloc(n, M_PMC, M_WAITOK|M_ZERO);
 486         (void) strlcpy(newstr, pmc_debugstr, n);
 487
 488         error = sysctl_handle_string(oidp, newstr, n, req);
 489
 490         /* if there is a new string, parse and copy it */
 491         if (error == 0 && req->newptr != NULL) {
 492                 fence = newstr + (n < req->newlen ? n : req->newlen + 1);
 493                 if ((error = pmc_debugflags_parse(newstr, fence)) == 0)
 494                         (void) strlcpy(pmc_debugstr, newstr,
 495                             sizeof(pmc_debugstr));
 496         }
 497
 498         free(newstr, M_PMC);
 499
 500         return error;
 501 }
 502 #endif
 503
 504 /*
 505  * Map a row index to a classdep structure and return the adjusted row
 506  * index for the PMC class index.
 507  */
 508 static struct pmc_classdep *
 509 pmc_ri_to_classdep(struct pmc_mdep *md, int ri, int *adjri)
 510 {
 511         struct pmc_classdep *pcd;
 512
 513         (void) md;
 514
 515         KASSERT(ri >= 0 && ri < md->pmd_npmc,
 516             ("[pmc,%d] illegal row-index %d", __LINE__, ri));
 517
 518         pcd = pmc_rowindex_to_classdep[ri];
 519
 520         KASSERT(pcd != NULL,
 521             ("[pmc,%d] ri %d null pcd", __LINE__, ri));
 522
 523         *adjri = ri - pcd->pcd_ri;
 524
 525         KASSERT(*adjri >= 0 && *adjri < pcd->pcd_num,
 526             ("[pmc,%d] adjusted row-index %d", __LINE__, *adjri));
 527
 528         return (pcd);
 529 }
 530
 531 /*
 532  * Concurrency Control
 533  *
 534  * The driver manages the following data structures:
 535  *
 536  *   - target process descriptors, one per target process
 537  *   - owner process descriptors (and attached lists), one per owner process
 538  *   - lookup hash tables for owner and target processes
 539  *   - PMC descriptors (and attached lists)
 540  *   - per-cpu hardware state
 541  *   - the 'hook' variable through which the kernel calls into
 542  *     this module
 543  *   - the machine hardware state (managed by the MD layer)
 544  *
 545  * These data structures are accessed from:
 546  *
 547  * - thread context-switch code
 548  * - interrupt handlers (possibly on multiple cpus)
 549  * - kernel threads on multiple cpus running on behalf of user
 550  *   processes doing system calls
 551  * - this driver's private kernel threads
 552  *
 553  * = Locks and Locking strategy =
 554  *
 555  * The driver uses four locking strategies for its operation:
 556  *
 557  * - The global SX lock "pmc_sx" is used to protect internal
 558  *   data structures.
 559  *
 560  *   Calls into the module by syscall() start with this lock being
 561  *   held in exclusive mode.  Depending on the requested operation,
 562  *   the lock may be downgraded to 'shared' mode to allow more
 563  *   concurrent readers into the module.  Calls into the module from
 564  *   other parts of the kernel acquire the lock in shared mode.
 565  *
 566  *   This SX lock is held in exclusive mode for any operations that
 567  *   modify the linkages between the driver's internal data structures.
 568  *
 569  *   The 'pmc_hook' function pointer is also protected by this lock.
 570  *   It is only examined with the sx lock held in exclusive mode.  The
 571  *   kernel module is allowed to be unloaded only with the sx lock held
 572  *   in exclusive mode.  In normal syscall handling, after acquiring the
 573  *   pmc_sx lock we first check that 'pmc_hook' is non-null before
 574  *   proceeding.  This prevents races between the thread unloading the module
 575  *   and other threads seeking to use the module.
 576  *
 577  * - Lookups of target process structures and owner process structures
 578  *   cannot use the global "pmc_sx" SX lock because these lookups need
 579  *   to happen during context switches and in other critical sections
 580  *   where sleeping is not allowed.  We protect these lookup tables
 581  *   with their own private spin-mutexes, "pmc_processhash_mtx" and
 582  *   "pmc_ownerhash_mtx".
 583  *
 584  * - Interrupt handlers work in a lock free manner.  At interrupt
 585  *   time, handlers look at the PMC pointer (phw->phw_pmc) configured
 586  *   when the PMC was started.  If this pointer is NULL, the interrupt
 587  *   is ignored after updating driver statistics.  We ensure that this
 588  *   pointer is set (using an atomic operation if necessary) before the
 589  *   PMC hardware is started.  Conversely, this pointer is unset atomically
 590  *   only after the PMC hardware is stopped.
 591  *
 592  *   We ensure that everything needed for the operation of an
 593  *   interrupt handler is available without it needing to acquire any
 594  *   locks.  We also ensure that a PMC's software state is destroyed only
 595  *   after the PMC is taken off hardware (on all CPUs).
 596  *
 597  * - Context-switch handling with process-private PMCs needs more
 598  *   care.
 599  *
 600  *   A given process may be the target of multiple PMCs.  For example,
 601  *   PMCATTACH and PMCDETACH may be requested by a process on one CPU
 602  *   while the target process is running on another.  A PMC could also
 603  *   be getting released because its owner is exiting.  We tackle
 604  *   these situations in the following manner:
 605  *
 606  *   - each target process structure 'pmc_process' has an array
 607  *     of 'struct pmc *' pointers, one for each hardware PMC.
 608  *
 609  *   - At context switch IN time, each "target" PMC in RUNNING state
 610  *     gets started on hardware and a pointer to each PMC is copied into
 611  *     the per-cpu phw array.  The 'runcount' for the PMC is
 612  *     incremented.
 613  *
 614  *   - At context switch OUT time, all process-virtual PMCs are stopped
 615  *     on hardware.  The saved value is added to the PMCs value field
 616  *     only if the PMC is in a non-deleted state (the PMCs state could
 617  *     have changed during the current time slice).
 618  *
 619  *     Note that since in-between a switch IN on a processor and a switch
 620  *     OUT, the PMC could have been released on another CPU.  Therefore
 621  *     context switch OUT always looks at the hardware state to turn
 622  *     OFF PMCs and will update a PMC's saved value only if reachable
 623  *     from the target process record.
 624  *
 625  *   - OP PMCRELEASE could be called on a PMC at any time (the PMC could
 626  *     be attached to many processes at the time of the call and could
 627  *     be active on multiple CPUs).
 628  *
 629  *     We prevent further scheduling of the PMC by marking it as in
 630  *     state 'DELETED'.  If the runcount of the PMC is non-zero then
 631  *     this PMC is currently running on a CPU somewhere.  The thread
 632  *     doing the PMCRELEASE operation waits by repeatedly doing a
 633  *     pause() till the runcount comes to zero.
 634  *
 635  * The contents of a PMC descriptor (struct pmc) are protected using
 636  * a spin-mutex.  In order to save space, we use a mutex pool.
 637  *
 638  * In terms of lock types used by witness(4), we use:
 639  * - Type "pmc-sx", used by the global SX lock.
 640  * - Type "pmc-sleep", for sleep mutexes used by logger threads.
 641  * - Type "pmc-per-proc", for protecting PMC owner descriptors.
 642  * - Type "pmc-leaf", used for all other spin mutexes.
 643  */
 644
 645 /*
 646  * save the cpu binding of the current kthread
 647  */
 648
 649 static void
 650 pmc_save_cpu_binding(struct pmc_binding *pb)
 651 {
 652         PMCDBG(CPU,BND,2, "%s", "save-cpu");
 653         thread_lock(curthread);
 654         pb->pb_bound = sched_is_bound(curthread);
 655         pb->pb_cpu   = curthread->td_oncpu;
 656         thread_unlock(curthread);
 657         PMCDBG(CPU,BND,2, "save-cpu cpu=%d", pb->pb_cpu);
 658 }
 659
 660 /*
 661  * restore the cpu binding of the current thread
 662  */
 663
 664 static void
 665 pmc_restore_cpu_binding(struct pmc_binding *pb)
 666 {
 667         PMCDBG(CPU,BND,2, "restore-cpu curcpu=%d restore=%d",
 668             curthread->td_oncpu, pb->pb_cpu);
 669         thread_lock(curthread);
 670         if (pb->pb_bound)
 671                 sched_bind(curthread, pb->pb_cpu);
 672         else
 673                 sched_unbind(curthread);
 674         thread_unlock(curthread);
 675         PMCDBG(CPU,BND,2, "%s", "restore-cpu done");
 676 }
 677
 678 /*
 679  * move execution over the specified cpu and bind it there.
 680  */
 681
 682 static void
 683 pmc_select_cpu(int cpu)
 684 {
 685         KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
 686             ("[pmc,%d] bad cpu number %d", __LINE__, cpu));
 687
 688         /* Never move to an inactive CPU. */
 689         KASSERT(pmc_cpu_is_active(cpu), ("[pmc,%d] selecting inactive "
 690             "CPU %d", __LINE__, cpu));
 691
 692         PMCDBG(CPU,SEL,2, "select-cpu cpu=%d", cpu);
 693         thread_lock(curthread);
 694         sched_bind(curthread, cpu);
 695         thread_unlock(curthread);
 696
 697         KASSERT(curthread->td_oncpu == cpu,
 698             ("[pmc,%d] CPU not bound [cpu=%d, curr=%d]", __LINE__,
 699                 cpu, curthread->td_oncpu));
 700
 701         PMCDBG(CPU,SEL,2, "select-cpu cpu=%d ok", cpu);
 702 }
 703
 704 /*
 705  * Force a context switch.
 706  *
 707  * We do this by pause'ing for 1 tick -- invoking mi_switch() is not
 708  * guaranteed to force a context switch.
 709  */
 710
 711 static void
 712 pmc_force_context_switch(void)
 713 {
 714
 715         pause("pmcctx", 1);
 716 }
 717
 718 /*
 719  * Get the file name for an executable.  This is a simple wrapper
 720  * around vn_fullpath(9).
 721  */
 722
 723 static void
 724 pmc_getfilename(struct vnode *v, char **fullpath, char **freepath)
 725 {
 726
 727         *fullpath = "unknown";
 728         *freepath = NULL;
 729         vn_fullpath(curthread, v, fullpath, freepath);
 730 }
 731
 732 /*
 733  * remove an process owning PMCs
 734  */
 735
 736 void
 737 pmc_remove_owner(struct pmc_owner *po)
 738 {
 739         struct pmc *pm, *tmp;
 740
 741         sx_assert(&pmc_sx, SX_XLOCKED);
 742
 743         PMCDBG(OWN,ORM,1, "remove-owner po=%p", po);
 744
 745         /* Remove descriptor from the owner hash table */
 746         LIST_REMOVE(po, po_next);
 747
 748         /* release all owned PMC descriptors */
 749         LIST_FOREACH_SAFE(pm, &po->po_pmcs, pm_next, tmp) {
 750                 PMCDBG(OWN,ORM,2, "pmc=%p", pm);
 751                 KASSERT(pm->pm_owner == po,
 752                     ("[pmc,%d] owner %p != po %p", __LINE__, pm->pm_owner, po));
 753
 754                 pmc_release_pmc_descriptor(pm); /* will unlink from the list */
 755         }
 756
 757         KASSERT(po->po_sscount == 0,
 758             ("[pmc,%d] SS count not zero", __LINE__));
 759         KASSERT(LIST_EMPTY(&po->po_pmcs),
 760             ("[pmc,%d] PMC list not empty", __LINE__));
 761
 762         /* de-configure the log file if present */
 763         if (po->po_flags & PMC_PO_OWNS_LOGFILE)
 764                 pmclog_deconfigure_log(po);
 765 }
 766
 767 /*
 768  * remove an owner process record if all conditions are met.
 769  */
 770
 771 static void
 772 pmc_maybe_remove_owner(struct pmc_owner *po)
 773 {
 774
 775         PMCDBG(OWN,OMR,1, "maybe-remove-owner po=%p", po);
 776
 777         /*
 778          * Remove owner record if
 779          * - this process does not own any PMCs
 780          * - this process has not allocated a system-wide sampling buffer
 781          */
 782
 783         if (LIST_EMPTY(&po->po_pmcs) &&
 784             ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)) {
 785                 pmc_remove_owner(po);
 786                 pmc_destroy_owner_descriptor(po);
 787         }
 788 }
 789
 790 /*
 791  * Add an association between a target process and a PMC.
 792  */
 793
 794 static void
 795 pmc_link_target_process(struct pmc *pm, struct pmc_process *pp)
 796 {
 797         int ri;
 798         struct pmc_target *pt;
 799
 800         sx_assert(&pmc_sx, SX_XLOCKED);
 801
 802         KASSERT(pm != NULL && pp != NULL,
 803             ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp));
 804         KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)),
 805             ("[pmc,%d] Attaching a non-process-virtual pmc=%p to pid=%d",
 806                 __LINE__, pm, pp->pp_proc->p_pid));
 807         KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= ((int) md->pmd_npmc - 1),
 808             ("[pmc,%d] Illegal reference count %d for process record %p",
 809                 __LINE__, pp->pp_refcnt, (void *) pp));
 810
 811         ri = PMC_TO_ROWINDEX(pm);
 812
 813         PMCDBG(PRC,TLK,1, "link-target pmc=%p ri=%d pmc-process=%p",
 814             pm, ri, pp);
 815
 816 #ifdef  DEBUG
 817         LIST_FOREACH(pt, &pm->pm_targets, pt_next)
 818             if (pt->pt_process == pp)
 819                     KASSERT(0, ("[pmc,%d] pp %p already in pmc %p targets",
 820                                 __LINE__, pp, pm));
 821 #endif
 822
 823         pt = malloc(sizeof(struct pmc_target), M_PMC, M_WAITOK|M_ZERO);
 824         pt->pt_process = pp;
 825
 826         LIST_INSERT_HEAD(&pm->pm_targets, pt, pt_next);
 827
 828         atomic_store_rel_ptr((uintptr_t *)&pp->pp_pmcs[ri].pp_pmc,
 829             (uintptr_t)pm);
 830
 831         if (pm->pm_owner->po_owner == pp->pp_proc)
 832                 pm->pm_flags |= PMC_F_ATTACHED_TO_OWNER;
 833
 834         /*
 835          * Initialize the per-process values at this row index.
 836          */
 837         pp->pp_pmcs[ri].pp_pmcval = PMC_TO_MODE(pm) == PMC_MODE_TS ?
 838             pm->pm_sc.pm_reloadcount : 0;
 839
 840         pp->pp_refcnt++;
 841
 842 }
 843
 844 /*
 845  * Removes the association between a target process and a PMC.
 846  */
 847
 848 static void
 849 pmc_unlink_target_process(struct pmc *pm, struct pmc_process *pp)
 850 {
 851         int ri;
 852         struct proc *p;
 853         struct pmc_target *ptgt;
 854
 855         sx_assert(&pmc_sx, SX_XLOCKED);
 856
 857         KASSERT(pm != NULL && pp != NULL,
 858             ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp));
 859
 860         KASSERT(pp->pp_refcnt >= 1 && pp->pp_refcnt <= (int) md->pmd_npmc,
 861             ("[pmc,%d] Illegal ref count %d on process record %p",
 862                 __LINE__, pp->pp_refcnt, (void *) pp));
 863
 864         ri = PMC_TO_ROWINDEX(pm);
 865
 866         PMCDBG(PRC,TUL,1, "unlink-target pmc=%p ri=%d pmc-process=%p",
 867             pm, ri, pp);
 868
 869         KASSERT(pp->pp_pmcs[ri].pp_pmc == pm,
 870             ("[pmc,%d] PMC ri %d mismatch pmc %p pp->[ri] %p", __LINE__,
 871                 ri, pm, pp->pp_pmcs[ri].pp_pmc));
 872
 873         pp->pp_pmcs[ri].pp_pmc = NULL;
 874         pp->pp_pmcs[ri].pp_pmcval = (pmc_value_t) 0;
 875
 876         /* Remove owner-specific flags */
 877         if (pm->pm_owner->po_owner == pp->pp_proc) {
 878                 pp->pp_flags &= ~PMC_PP_ENABLE_MSR_ACCESS;
 879                 pm->pm_flags &= ~PMC_F_ATTACHED_TO_OWNER;
 880         }
 881
 882         pp->pp_refcnt--;
 883
 884         /* Remove the target process from the PMC structure */
 885         LIST_FOREACH(ptgt, &pm->pm_targets, pt_next)
 886                 if (ptgt->pt_process == pp)
 887                         break;
 888
 889         KASSERT(ptgt != NULL, ("[pmc,%d] process %p (pp: %p) not found "
 890                     "in pmc %p", __LINE__, pp->pp_proc, pp, pm));
 891
 892         LIST_REMOVE(ptgt, pt_next);
 893         free(ptgt, M_PMC);
 894
 895         /* if the PMC now lacks targets, send the owner a SIGIO */
 896         if (LIST_EMPTY(&pm->pm_targets)) {
 897                 p = pm->pm_owner->po_owner;
 898                 PROC_LOCK(p);
 899                 kern_psignal(p, SIGIO);
 900                 PROC_UNLOCK(p);
 901
 902                 PMCDBG(PRC,SIG,2, "signalling proc=%p signal=%d", p,
 903                     SIGIO);
 904         }
 905 }
 906
 907 /*
 908  * Check if PMC 'pm' may be attached to target process 't'.
 909  */
 910
 911 static int
 912 pmc_can_attach(struct pmc *pm, struct proc *t)
 913 {
 914         struct proc *o;         /* pmc owner */
 915         struct ucred *oc, *tc;  /* owner, target credentials */
 916         int decline_attach, i;
 917
 918         /*
 919          * A PMC's owner can always attach that PMC to itself.
 920          */
 921
 922         if ((o = pm->pm_owner->po_owner) == t)
 923                 return 0;
 924
 925         PROC_LOCK(o);
 926         oc = o->p_ucred;
 927         crhold(oc);
 928         PROC_UNLOCK(o);
 929
 930         PROC_LOCK(t);
 931         tc = t->p_ucred;
 932         crhold(tc);
 933         PROC_UNLOCK(t);
 934
 935         /*
 936          * The effective uid of the PMC owner should match at least one
 937          * of the {effective,real,saved} uids of the target process.
 938          */
 939
 940         decline_attach = oc->cr_uid != tc->cr_uid &&
 941             oc->cr_uid != tc->cr_svuid &&
 942             oc->cr_uid != tc->cr_ruid;
 943
 944         /*
 945          * Every one of the target's group ids, must be in the owner's
 946          * group list.
 947          */
 948         for (i = 0; !decline_attach && i < tc->cr_ngroups; i++)
 949                 decline_attach = !groupmember(tc->cr_groups[i], oc);
 950
 951         /* check the read and saved gids too */
 952         if (decline_attach == 0)
 953                 decline_attach = !groupmember(tc->cr_rgid, oc) ||
 954                     !groupmember(tc->cr_svgid, oc);
 955
 956         crfree(tc);
 957         crfree(oc);
 958
 959         return !decline_attach;
 960 }
 961
 962 /*
 963  * Attach a process to a PMC.
 964  */
 965
 966 static int
 967 pmc_attach_one_process(struct proc *p, struct pmc *pm)
 968 {
 969         int ri;
 970         char *fullpath, *freepath;
 971         struct pmc_process      *pp;
 972
 973         sx_assert(&pmc_sx, SX_XLOCKED);
 974
 975         PMCDBG(PRC,ATT,2, "attach-one pm=%p ri=%d proc=%p (%d, %s)", pm,
 976             PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm);
 977
 978         /*
 979          * Locate the process descriptor corresponding to process 'p',
 980          * allocating space as needed.
 981          *
 982          * Verify that rowindex 'pm_rowindex' is free in the process
 983          * descriptor.
 984          *
 985          * If not, allocate space for a descriptor and link the
 986          * process descriptor and PMC.
 987          */
 988         ri = PMC_TO_ROWINDEX(pm);
 989
 990         if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_ALLOCATE)) == NULL)
 991                 return ENOMEM;
 992
 993         if (pp->pp_pmcs[ri].pp_pmc == pm) /* already present at slot [ri] */
 994                 return EEXIST;
 995
 996         if (pp->pp_pmcs[ri].pp_pmc != NULL)
 997                 return EBUSY;
 998
 999         pmc_link_target_process(pm, pp);
1000
1001         if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) &&
1002             (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) == 0)
1003                 pm->pm_flags |= PMC_F_NEEDS_LOGFILE;
1004
1005         pm->pm_flags |= PMC_F_ATTACH_DONE; /* mark as attached */
1006
1007         /* issue an attach event to a configured log file */
1008         if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE) {
1009                 pmc_getfilename(p->p_textvp, &fullpath, &freepath);
1010                 if (p->p_flag & P_KTHREAD) {
1011                         fullpath = kernelname;
1012                         freepath = NULL;
1013                 } else
1014                         pmclog_process_pmcattach(pm, p->p_pid, fullpath);
1015                 if (freepath)
1016                         free(freepath, M_TEMP);
1017                 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1018                         pmc_log_process_mappings(pm->pm_owner, p);
1019         }
1020         /* mark process as using HWPMCs */
1021         PROC_LOCK(p);
1022         p->p_flag |= P_HWPMC;
1023         PROC_UNLOCK(p);
1024
1025         return 0;
1026 }
1027
1028 /*
1029  * Attach a process and optionally its children
1030  */
1031
1032 static int
1033 pmc_attach_process(struct proc *p, struct pmc *pm)
1034 {
1035         int error;
1036         struct proc *top;
1037
1038         sx_assert(&pmc_sx, SX_XLOCKED);
1039
1040         PMCDBG(PRC,ATT,1, "attach pm=%p ri=%d proc=%p (%d, %s)", pm,
1041             PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm);
1042
1043
1044         /*
1045          * If this PMC successfully allowed a GETMSR operation
1046          * in the past, disallow further ATTACHes.
1047          */
1048
1049         if ((pm->pm_flags & PMC_PP_ENABLE_MSR_ACCESS) != 0)
1050                 return EPERM;
1051
1052         if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0)
1053                 return pmc_attach_one_process(p, pm);
1054
1055         /*
1056          * Traverse all child processes, attaching them to
1057          * this PMC.
1058          */
1059
1060         sx_slock(&proctree_lock);
1061
1062         top = p;
1063
1064         for (;;) {
1065                 if ((error = pmc_attach_one_process(p, pm)) != 0)
1066                         break;
1067                 if (!LIST_EMPTY(&p->p_children))
1068                         p = LIST_FIRST(&p->p_children);
1069                 else for (;;) {
1070                         if (p == top)
1071                                 goto done;
1072                         if (LIST_NEXT(p, p_sibling)) {
1073                                 p = LIST_NEXT(p, p_sibling);
1074                                 break;
1075                         }
1076                         p = p->p_pptr;
1077                 }
1078         }
1079
1080         if (error)
1081                 (void) pmc_detach_process(top, pm);
1082
1083  done:
1084         sx_sunlock(&proctree_lock);
1085         return error;
1086 }
1087
1088 /*
1089  * Detach a process from a PMC.  If there are no other PMCs tracking
1090  * this process, remove the process structure from its hash table.  If
1091  * 'flags' contains PMC_FLAG_REMOVE, then free the process structure.
1092  */
1093
1094 static int
1095 pmc_detach_one_process(struct proc *p, struct pmc *pm, int flags)
1096 {
1097         int ri;
1098         struct pmc_process *pp;
1099
1100         sx_assert(&pmc_sx, SX_XLOCKED);
1101
1102         KASSERT(pm != NULL,
1103             ("[pmc,%d] null pm pointer", __LINE__));
1104
1105         ri = PMC_TO_ROWINDEX(pm);
1106
1107         PMCDBG(PRC,ATT,2, "detach-one pm=%p ri=%d proc=%p (%d, %s) flags=0x%x",
1108             pm, ri, p, p->p_pid, p->p_comm, flags);
1109
1110         if ((pp = pmc_find_process_descriptor(p, 0)) == NULL)
1111                 return ESRCH;
1112
1113         if (pp->pp_pmcs[ri].pp_pmc != pm)
1114                 return EINVAL;
1115
1116         pmc_unlink_target_process(pm, pp);
1117
1118         /* Issue a detach entry if a log file is configured */
1119         if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE)
1120                 pmclog_process_pmcdetach(pm, p->p_pid);
1121
1122         /*
1123          * If there are no PMCs targetting this process, we remove its
1124          * descriptor from the target hash table and unset the P_HWPMC
1125          * flag in the struct proc.
1126          */
1127         KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= (int) md->pmd_npmc,
1128             ("[pmc,%d] Illegal refcnt %d for process struct %p",
1129                 __LINE__, pp->pp_refcnt, pp));
1130
1131         if (pp->pp_refcnt != 0) /* still a target of some PMC */
1132                 return 0;
1133
1134         pmc_remove_process_descriptor(pp);
1135
1136         if (flags & PMC_FLAG_REMOVE)
1137                 free(pp, M_PMC);
1138
1139         PROC_LOCK(p);
1140         p->p_flag &= ~P_HWPMC;
1141         PROC_UNLOCK(p);
1142
1143         return 0;
1144 }
1145
1146 /*
1147  * Detach a process and optionally its descendants from a PMC.
1148  */
1149
1150 static int
1151 pmc_detach_process(struct proc *p, struct pmc *pm)
1152 {
1153         struct proc *top;
1154
1155         sx_assert(&pmc_sx, SX_XLOCKED);
1156
1157         PMCDBG(PRC,ATT,1, "detach pm=%p ri=%d proc=%p (%d, %s)", pm,
1158             PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm);
1159
1160         if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0)
1161                 return pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE);
1162
1163         /*
1164          * Traverse all children, detaching them from this PMC.  We
1165          * ignore errors since we could be detaching a PMC from a
1166          * partially attached proc tree.
1167          */
1168
1169         sx_slock(&proctree_lock);
1170
1171         top = p;
1172
1173         for (;;) {
1174                 (void) pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE);
1175
1176                 if (!LIST_EMPTY(&p->p_children))
1177                         p = LIST_FIRST(&p->p_children);
1178                 else for (;;) {
1179                         if (p == top)
1180                                 goto done;
1181                         if (LIST_NEXT(p, p_sibling)) {
1182                                 p = LIST_NEXT(p, p_sibling);
1183                                 break;
1184                         }
1185                         p = p->p_pptr;
1186                 }
1187         }
1188
1189  done:
1190         sx_sunlock(&proctree_lock);
1191
1192         if (LIST_EMPTY(&pm->pm_targets))
1193                 pm->pm_flags &= ~PMC_F_ATTACH_DONE;
1194
1195         return 0;
1196 }
1197
1198
1199 /*
1200  * Thread context switch IN
1201  */
1202
1203 static void
1204 pmc_process_csw_in(struct thread *td)
1205 {
1206         int cpu;
1207         unsigned int adjri, ri;
1208         struct pmc *pm;
1209         struct proc *p;
1210         struct pmc_cpu *pc;
1211         struct pmc_hw *phw;
1212         pmc_value_t newvalue;
1213         struct pmc_process *pp;
1214         struct pmc_classdep *pcd;
1215
1216         p = td->td_proc;
1217
1218         if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE)) == NULL)
1219                 return;
1220
1221         KASSERT(pp->pp_proc == td->td_proc,
1222             ("[pmc,%d] not my thread state", __LINE__));
1223
1224         critical_enter(); /* no preemption from this point */
1225
1226         cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */
1227
1228         PMCDBG(CSW,SWI,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p,
1229             p->p_pid, p->p_comm, pp);
1230
1231         KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
1232             ("[pmc,%d] wierd CPU id %d", __LINE__, cpu));
1233
1234         pc = pmc_pcpu[cpu];
1235
1236         for (ri = 0; ri < md->pmd_npmc; ri++) {
1237
1238                 if ((pm = pp->pp_pmcs[ri].pp_pmc) == NULL)
1239                         continue;
1240
1241                 KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)),
1242                     ("[pmc,%d] Target PMC in non-virtual mode (%d)",
1243                         __LINE__, PMC_TO_MODE(pm)));
1244
1245                 KASSERT(PMC_TO_ROWINDEX(pm) == ri,
1246                     ("[pmc,%d] Row index mismatch pmc %d != ri %d",
1247                         __LINE__, PMC_TO_ROWINDEX(pm), ri));
1248
1249                 /*
1250                  * Only PMCs that are marked as 'RUNNING' need
1251                  * be placed on hardware.
1252                  */
1253
1254                 if (pm->pm_state != PMC_STATE_RUNNING)
1255                         continue;
1256
1257                 /* increment PMC runcount */
1258                 atomic_add_rel_int(&pm->pm_runcount, 1);
1259
1260                 /* configure the HWPMC we are going to use. */
1261                 pcd = pmc_ri_to_classdep(md, ri, &adjri);
1262                 pcd->pcd_config_pmc(cpu, adjri, pm);
1263
1264                 phw = pc->pc_hwpmcs[ri];
1265
1266                 KASSERT(phw != NULL,
1267                     ("[pmc,%d] null hw pointer", __LINE__));
1268
1269                 KASSERT(phw->phw_pmc == pm,
1270                     ("[pmc,%d] hw->pmc %p != pmc %p", __LINE__,
1271                         phw->phw_pmc, pm));
1272
1273                 /*
1274                  * Write out saved value and start the PMC.
1275                  *
1276                  * Sampling PMCs use a per-process value, while
1277                  * counting mode PMCs use a per-pmc value that is
1278                  * inherited across descendants.
1279                  */
1280                 if (PMC_TO_MODE(pm) == PMC_MODE_TS) {
1281                         mtx_pool_lock_spin(pmc_mtxpool, pm);
1282                         newvalue = PMC_PCPU_SAVED(cpu,ri) =
1283                             pp->pp_pmcs[ri].pp_pmcval;
1284                         mtx_pool_unlock_spin(pmc_mtxpool, pm);
1285                 } else {
1286                         KASSERT(PMC_TO_MODE(pm) == PMC_MODE_TC,
1287                             ("[pmc,%d] illegal mode=%d", __LINE__,
1288                             PMC_TO_MODE(pm)));
1289                         mtx_pool_lock_spin(pmc_mtxpool, pm);
1290                         newvalue = PMC_PCPU_SAVED(cpu, ri) =
1291                             pm->pm_gv.pm_savedvalue;
1292                         mtx_pool_unlock_spin(pmc_mtxpool, pm);
1293                 }
1294
1295                 PMCDBG(CSW,SWI,1,"cpu=%d ri=%d new=%jd", cpu, ri, newvalue);
1296
1297                 pcd->pcd_write_pmc(cpu, adjri, newvalue);
1298                 pcd->pcd_start_pmc(cpu, adjri);
1299         }
1300
1301         /*
1302          * perform any other architecture/cpu dependent thread
1303          * switch-in actions.
1304          */
1305
1306         (void) (*md->pmd_switch_in)(pc, pp);
1307
1308         critical_exit();
1309
1310 }
1311
1312 /*
1313  * Thread context switch OUT.
1314  */
1315
1316 static void
1317 pmc_process_csw_out(struct thread *td)
1318 {
1319         int cpu;
1320         int64_t tmp;
1321         struct pmc *pm;
1322         struct proc *p;
1323         enum pmc_mode mode;
1324         struct pmc_cpu *pc;
1325         pmc_value_t newvalue;
1326         unsigned int adjri, ri;
1327         struct pmc_process *pp;
1328         struct pmc_classdep *pcd;
1329
1330
1331         /*
1332          * Locate our process descriptor; this may be NULL if
1333          * this process is exiting and we have already removed
1334          * the process from the target process table.
1335          *
1336          * Note that due to kernel preemption, multiple
1337          * context switches may happen while the process is
1338          * exiting.
1339          *
1340          * Note also that if the target process cannot be
1341          * found we still need to deconfigure any PMCs that
1342          * are currently running on hardware.
1343          */
1344
1345         p = td->td_proc;
1346         pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE);
1347
1348         /*
1349          * save PMCs
1350          */
1351
1352         critical_enter();
1353
1354         cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */
1355
1356         PMCDBG(CSW,SWO,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p,
1357             p->p_pid, p->p_comm, pp);
1358
1359         KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
1360             ("[pmc,%d wierd CPU id %d", __LINE__, cpu));
1361
1362         pc = pmc_pcpu[cpu];
1363
1364         /*
1365          * When a PMC gets unlinked from a target PMC, it will
1366          * be removed from the target's pp_pmc[] array.
1367          *
1368          * However, on a MP system, the target could have been
1369          * executing on another CPU at the time of the unlink.
1370          * So, at context switch OUT time, we need to look at
1371          * the hardware to determine if a PMC is scheduled on
1372          * it.
1373          */
1374
1375         for (ri = 0; ri < md->pmd_npmc; ri++) {
1376
1377                 pcd = pmc_ri_to_classdep(md, ri, &adjri);
1378                 pm  = NULL;
1379                 (void) (*pcd->pcd_get_config)(cpu, adjri, &pm);
1380
1381                 if (pm == NULL) /* nothing at this row index */
1382                         continue;
1383
1384                 mode = PMC_TO_MODE(pm);
1385                 if (!PMC_IS_VIRTUAL_MODE(mode))
1386                         continue; /* not a process virtual PMC */
1387
1388                 KASSERT(PMC_TO_ROWINDEX(pm) == ri,
1389                     ("[pmc,%d] ri mismatch pmc(%d) ri(%d)",
1390                         __LINE__, PMC_TO_ROWINDEX(pm), ri));
1391
1392                 /* Stop hardware if not already stopped */
1393                 if (pm->pm_stalled == 0)
1394                         pcd->pcd_stop_pmc(cpu, adjri);
1395
1396                 /* reduce this PMC's runcount */
1397                 atomic_subtract_rel_int(&pm->pm_runcount, 1);
1398
1399                 /*
1400                  * If this PMC is associated with this process,
1401                  * save the reading.
1402                  */
1403
1404                 if (pp != NULL && pp->pp_pmcs[ri].pp_pmc != NULL) {
1405
1406                         KASSERT(pm == pp->pp_pmcs[ri].pp_pmc,
1407                             ("[pmc,%d] pm %p != pp_pmcs[%d] %p", __LINE__,
1408                                 pm, ri, pp->pp_pmcs[ri].pp_pmc));
1409
1410                         KASSERT(pp->pp_refcnt > 0,
1411                             ("[pmc,%d] pp refcnt = %d", __LINE__,
1412                                 pp->pp_refcnt));
1413
1414                         pcd->pcd_read_pmc(cpu, adjri, &newvalue);
1415
1416                         tmp = newvalue - PMC_PCPU_SAVED(cpu,ri);
1417
1418                         PMCDBG(CSW,SWO,1,"cpu=%d ri=%d tmp=%jd", cpu, ri,
1419                             tmp);
1420
1421                         if (mode == PMC_MODE_TS) {
1422
1423                                 /*
1424                                  * For sampling process-virtual PMCs,
1425                                  * we expect the count to be
1426                                  * decreasing as the 'value'
1427                                  * programmed into the PMC is the
1428                                  * number of events to be seen till
1429                                  * the next sampling interrupt.
1430                                  */
1431                                 if (tmp < 0)
1432                                         tmp += pm->pm_sc.pm_reloadcount;
1433                                 mtx_pool_lock_spin(pmc_mtxpool, pm);
1434                                 pp->pp_pmcs[ri].pp_pmcval -= tmp;
1435                                 if ((int64_t) pp->pp_pmcs[ri].pp_pmcval < 0)
1436                                         pp->pp_pmcs[ri].pp_pmcval +=
1437                                             pm->pm_sc.pm_reloadcount;
1438                                 mtx_pool_unlock_spin(pmc_mtxpool, pm);
1439
1440                         } else {
1441
1442                                 /*
1443                                  * For counting process-virtual PMCs,
1444                                  * we expect the count to be
1445                                  * increasing monotonically, modulo a 64
1446                                  * bit wraparound.
1447                                  */
1448                                 KASSERT((int64_t) tmp >= 0,
1449                                     ("[pmc,%d] negative increment cpu=%d "
1450                                      "ri=%d newvalue=%jx saved=%jx "
1451                                      "incr=%jx", __LINE__, cpu, ri,
1452                                      newvalue, PMC_PCPU_SAVED(cpu,ri), tmp));
1453
1454                                 mtx_pool_lock_spin(pmc_mtxpool, pm);
1455                                 pm->pm_gv.pm_savedvalue += tmp;
1456                                 pp->pp_pmcs[ri].pp_pmcval += tmp;
1457                                 mtx_pool_unlock_spin(pmc_mtxpool, pm);
1458
1459                                 if (pm->pm_flags & PMC_F_LOG_PROCCSW)
1460                                         pmclog_process_proccsw(pm, pp, tmp);
1461                         }
1462                 }
1463
1464                 /* mark hardware as free */
1465                 pcd->pcd_config_pmc(cpu, adjri, NULL);
1466         }
1467
1468         /*
1469          * perform any other architecture/cpu dependent thread
1470          * switch out functions.
1471          */
1472
1473         (void) (*md->pmd_switch_out)(pc, pp);
1474
1475         critical_exit();
1476 }
1477
1478 /*
1479  * A mapping change for a process.
1480  */
1481
1482 static void
1483 pmc_process_mmap(struct thread *td, struct pmckern_map_in *pkm)
1484 {
1485         int ri;
1486         pid_t pid;
1487         char *fullpath, *freepath;
1488         const struct pmc *pm;
1489         struct pmc_owner *po;
1490         const struct pmc_process *pp;
1491
1492         freepath = fullpath = NULL;
1493         pmc_getfilename((struct vnode *) pkm->pm_file, &fullpath, &freepath);
1494
1495         pid = td->td_proc->p_pid;
1496
1497         /* Inform owners of all system-wide sampling PMCs. */
1498         LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1499             if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1500                 pmclog_process_map_in(po, pid, pkm->pm_address, fullpath);
1501
1502         if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL)
1503                 goto done;
1504
1505         /*
1506          * Inform sampling PMC owners tracking this process.
1507          */
1508         for (ri = 0; ri < md->pmd_npmc; ri++)
1509                 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL &&
1510                     PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1511                         pmclog_process_map_in(pm->pm_owner,
1512                             pid, pkm->pm_address, fullpath);
1513
1514   done:
1515         if (freepath)
1516                 free(freepath, M_TEMP);
1517 }
1518
1519
1520 /*
1521  * Log an munmap request.
1522  */
1523
1524 static void
1525 pmc_process_munmap(struct thread *td, struct pmckern_map_out *pkm)
1526 {
1527         int ri;
1528         pid_t pid;
1529         struct pmc_owner *po;
1530         const struct pmc *pm;
1531         const struct pmc_process *pp;
1532
1533         pid = td->td_proc->p_pid;
1534
1535         LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1536             if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1537                 pmclog_process_map_out(po, pid, pkm->pm_address,
1538                     pkm->pm_address + pkm->pm_size);
1539
1540         if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL)
1541                 return;
1542
1543         for (ri = 0; ri < md->pmd_npmc; ri++)
1544                 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL &&
1545                     PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1546                         pmclog_process_map_out(pm->pm_owner, pid,
1547                             pkm->pm_address, pkm->pm_address + pkm->pm_size);
1548 }
1549
1550 /*
1551  * Log mapping information about the kernel.
1552  */
1553
1554 static void
1555 pmc_log_kernel_mappings(struct pmc *pm)
1556 {
1557         struct pmc_owner *po;
1558         struct pmckern_map_in *km, *kmbase;
1559
1560         sx_assert(&pmc_sx, SX_LOCKED);
1561         KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)),
1562             ("[pmc,%d] non-sampling PMC (%p) desires mapping information",
1563                 __LINE__, (void *) pm));
1564
1565         po = pm->pm_owner;
1566
1567         if (po->po_flags & PMC_PO_INITIAL_MAPPINGS_DONE)
1568                 return;
1569
1570         /*
1571          * Log the current set of kernel modules.
1572          */
1573         kmbase = linker_hwpmc_list_objects();
1574         for (km = kmbase; km->pm_file != NULL; km++) {
1575                 PMCDBG(LOG,REG,1,"%s %p", (char *) km->pm_file,
1576                     (void *) km->pm_address);
1577                 pmclog_process_map_in(po, (pid_t) -1, km->pm_address,
1578                     km->pm_file);
1579         }
1580         free(kmbase, M_LINKER);
1581
1582         po->po_flags |= PMC_PO_INITIAL_MAPPINGS_DONE;
1583 }
1584
1585 /*
1586  * Log the mappings for a single process.
1587  */
1588
1589 static void
1590 pmc_log_process_mappings(struct pmc_owner *po, struct proc *p)
1591 {
1592         int locked;
1593         vm_map_t map;
1594         struct vnode *vp;
1595         struct vmspace *vm;
1596         vm_map_entry_t entry;
1597         vm_offset_t last_end;
1598         u_int last_timestamp;
1599         struct vnode *last_vp;
1600         vm_offset_t start_addr;
1601         vm_object_t obj, lobj, tobj;
1602         char *fullpath, *freepath;
1603
1604         last_vp = NULL;
1605         last_end = (vm_offset_t) 0;
1606         fullpath = freepath = NULL;
1607
1608         if ((vm = vmspace_acquire_ref(p)) == NULL)
1609                 return;
1610
1611         map = &vm->vm_map;
1612         vm_map_lock_read(map);
1613
1614         for (entry = map->header.next; entry != &map->header; entry = entry->next) {
1615
1616                 if (entry == NULL) {
1617                         PMCDBG(LOG,OPS,2, "hwpmc: vm_map entry unexpectedly "
1618                             "NULL! pid=%d vm_map=%p\n", p->p_pid, map);
1619                         break;
1620                 }
1621
1622                 /*
1623                  * We only care about executable map entries.
1624                  */
1625                 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) ||
1626                     !(entry->protection & VM_PROT_EXECUTE) ||
1627                     (entry->object.vm_object == NULL)) {
1628                         continue;
1629                 }
1630
1631                 obj = entry->object.vm_object;
1632                 VM_OBJECT_LOCK(obj);
1633
1634                 /*
1635                  * Walk the backing_object list to find the base
1636                  * (non-shadowed) vm_object.
1637                  */
1638                 for (lobj = tobj = obj; tobj != NULL; tobj = tobj->backing_object) {
1639                         if (tobj != obj)
1640                                 VM_OBJECT_LOCK(tobj);
1641                         if (lobj != obj)
1642                                 VM_OBJECT_UNLOCK(lobj);
1643                         lobj = tobj;
1644                 }
1645
1646                 /*
1647                  * At this point lobj is the base vm_object and it is locked.
1648                  */
1649                 if (lobj == NULL) {
1650                         PMCDBG(LOG,OPS,2, "hwpmc: lobj unexpectedly NULL! pid=%d "
1651                             "vm_map=%p vm_obj=%p\n", p->p_pid, map, obj);
1652                         VM_OBJECT_UNLOCK(obj);
1653                         continue;
1654                 }
1655
1656                 if (lobj->type != OBJT_VNODE || lobj->handle == NULL) {
1657                         if (lobj != obj)
1658                                 VM_OBJECT_UNLOCK(lobj);
1659                         VM_OBJECT_UNLOCK(obj);
1660                         continue;
1661                 }
1662
1663                 /*
1664                  * Skip contiguous regions that point to the same
1665                  * vnode, so we don't emit redundant MAP-IN
1666                  * directives.
1667                  */
1668                 if (entry->start == last_end && lobj->handle == last_vp) {
1669                         last_end = entry->end;
1670                         if (lobj != obj)
1671                                 VM_OBJECT_UNLOCK(lobj);
1672                         VM_OBJECT_UNLOCK(obj);
1673                         continue;
1674                 }
1675
1676                 /*
1677                  * We don't want to keep the proc's vm_map or this
1678                  * vm_object locked while we walk the pathname, since
1679                  * vn_fullpath() can sleep.  However, if we drop the
1680                  * lock, it's possible for concurrent activity to
1681                  * modify the vm_map list.  To protect against this,
1682                  * we save the vm_map timestamp before we release the
1683                  * lock, and check it after we reacquire the lock
1684                  * below.
1685                  */
1686                 start_addr = entry->start;
1687                 last_end = entry->end;
1688                 last_timestamp = map->timestamp;
1689                 vm_map_unlock_read(map);
1690
1691                 vp = lobj->handle;
1692                 vref(vp);
1693                 if (lobj != obj)
1694                         VM_OBJECT_UNLOCK(lobj);
1695
1696                 VM_OBJECT_UNLOCK(obj);
1697
1698                 freepath = NULL;
1699                 pmc_getfilename(vp, &fullpath, &freepath);
1700                 last_vp = vp;
1701
1702                 locked = VFS_LOCK_GIANT(vp->v_mount);
1703                 vrele(vp);
1704                 VFS_UNLOCK_GIANT(locked);
1705
1706                 vp = NULL;
1707                 pmclog_process_map_in(po, p->p_pid, start_addr, fullpath);
1708                 if (freepath)
1709                         free(freepath, M_TEMP);
1710
1711                 vm_map_lock_read(map);
1712
1713                 /*
1714                  * If our saved timestamp doesn't match, this means
1715                  * that the vm_map was modified out from under us and
1716                  * we can't trust our current "entry" pointer.  Do a
1717                  * new lookup for this entry.  If there is no entry
1718                  * for this address range, vm_map_lookup_entry() will
1719                  * return the previous one, so we always want to go to
1720                  * entry->next on the next loop iteration.
1721                  *
1722                  * There is an edge condition here that can occur if
1723                  * there is no entry at or before this address.  In
1724                  * this situation, vm_map_lookup_entry returns
1725                  * &map->header, which would cause our loop to abort
1726                  * without processing the rest of the map.  However,
1727                  * in practice this will never happen for process
1728                  * vm_map.  This is because the executable's text
1729                  * segment is the first mapping in the proc's address
1730                  * space, and this mapping is never removed until the
1731                  * process exits, so there will always be a non-header
1732                  * entry at or before the requested address for
1733                  * vm_map_lookup_entry to return.
1734                  */
1735                 if (map->timestamp != last_timestamp)
1736                         vm_map_lookup_entry(map, last_end - 1, &entry);
1737         }
1738
1739         vm_map_unlock_read(map);
1740         vmspace_free(vm);
1741         return;
1742 }
1743
1744 /*
1745  * Log mappings for all processes in the system.
1746  */
1747
1748 static void
1749 pmc_log_all_process_mappings(struct pmc_owner *po)
1750 {
1751         struct proc *p, *top;
1752
1753         sx_assert(&pmc_sx, SX_XLOCKED);
1754
1755         if ((p = pfind(1)) == NULL)
1756                 panic("[pmc,%d] Cannot find init", __LINE__);
1757
1758         PROC_UNLOCK(p);
1759
1760         sx_slock(&proctree_lock);
1761
1762         top = p;
1763
1764         for (;;) {
1765                 pmc_log_process_mappings(po, p);
1766                 if (!LIST_EMPTY(&p->p_children))
1767                         p = LIST_FIRST(&p->p_children);
1768                 else for (;;) {
1769                         if (p == top)
1770                                 goto done;
1771                         if (LIST_NEXT(p, p_sibling)) {
1772                                 p = LIST_NEXT(p, p_sibling);
1773                                 break;
1774                         }
1775                         p = p->p_pptr;
1776                 }
1777         }
1778  done:
1779         sx_sunlock(&proctree_lock);
1780 }
1781
1782 /*
1783  * The 'hook' invoked from the kernel proper
1784  */
1785
1786
1787 #ifdef  DEBUG
1788 const char *pmc_hooknames[] = {
1789         /* these strings correspond to PMC_FN_* in <sys/pmckern.h> */
1790         "",
1791         "EXEC",
1792         "CSW-IN",
1793         "CSW-OUT",
1794         "SAMPLE",
1795         "UNUSED1",
1796         "UNUSED2",
1797         "MMAP",
1798         "MUNMAP",
1799         "CALLCHAIN-NMI",
1800         "CALLCHAIN-SOFT",
1801         "SOFTSAMPLING"
1802 };
1803 #endif
1804
1805 static int
1806 pmc_hook_handler(struct thread *td, int function, void *arg)
1807 {
1808
1809         PMCDBG(MOD,PMH,1, "hook td=%p func=%d \"%s\" arg=%p", td, function,
1810             pmc_hooknames[function], arg);
1811
1812         switch (function)
1813         {
1814
1815         /*
1816          * Process exec()
1817          */
1818
1819         case PMC_FN_PROCESS_EXEC:
1820         {
1821                 char *fullpath, *freepath;
1822                 unsigned int ri;
1823                 int is_using_hwpmcs;
1824                 struct pmc *pm;
1825                 struct proc *p;
1826                 struct pmc_owner *po;
1827                 struct pmc_process *pp;
1828                 struct pmckern_procexec *pk;
1829
1830                 sx_assert(&pmc_sx, SX_XLOCKED);
1831
1832                 p = td->td_proc;
1833                 pmc_getfilename(p->p_textvp, &fullpath, &freepath);
1834
1835                 pk = (struct pmckern_procexec *) arg;
1836
1837                 /* Inform owners of SS mode PMCs of the exec event. */
1838                 LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1839                     if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1840                             pmclog_process_procexec(po, PMC_ID_INVALID,
1841                                 p->p_pid, pk->pm_entryaddr, fullpath);
1842
1843                 PROC_LOCK(p);
1844                 is_using_hwpmcs = p->p_flag & P_HWPMC;
1845                 PROC_UNLOCK(p);
1846
1847                 if (!is_using_hwpmcs) {
1848                         if (freepath)
1849                                 free(freepath, M_TEMP);
1850                         break;
1851                 }
1852
1853                 /*
1854                  * PMCs are not inherited across an exec():  remove any
1855                  * PMCs that this process is the owner of.
1856                  */
1857
1858                 if ((po = pmc_find_owner_descriptor(p)) != NULL) {
1859                         pmc_remove_owner(po);
1860                         pmc_destroy_owner_descriptor(po);
1861                 }
1862
1863                 /*
1864                  * If the process being exec'ed is not the target of any
1865                  * PMC, we are done.
1866                  */
1867                 if ((pp = pmc_find_process_descriptor(p, 0)) == NULL) {
1868                         if (freepath)
1869                                 free(freepath, M_TEMP);
1870                         break;
1871                 }
1872
1873                 /*
1874                  * Log the exec event to all monitoring owners.  Skip
1875                  * owners who have already recieved the event because
1876                  * they had system sampling PMCs active.
1877                  */
1878                 for (ri = 0; ri < md->pmd_npmc; ri++)
1879                         if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) {
1880                                 po = pm->pm_owner;
1881                                 if (po->po_sscount == 0 &&
1882                                     po->po_flags & PMC_PO_OWNS_LOGFILE)
1883                                         pmclog_process_procexec(po, pm->pm_id,
1884                                             p->p_pid, pk->pm_entryaddr,
1885                                             fullpath);
1886                         }
1887
1888                 if (freepath)
1889                         free(freepath, M_TEMP);
1890
1891
1892                 PMCDBG(PRC,EXC,1, "exec proc=%p (%d, %s) cred-changed=%d",
1893                     p, p->p_pid, p->p_comm, pk->pm_credentialschanged);
1894
1895                 if (pk->pm_credentialschanged == 0) /* no change */
1896                         break;
1897
1898                 /*
1899                  * If the newly exec()'ed process has a different credential
1900                  * than before, allow it to be the target of a PMC only if
1901                  * the PMC's owner has sufficient priviledge.
1902                  */
1903
1904                 for (ri = 0; ri < md->pmd_npmc; ri++)
1905                         if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL)
1906                                 if (pmc_can_attach(pm, td->td_proc) != 0)
1907                                         pmc_detach_one_process(td->td_proc,
1908                                             pm, PMC_FLAG_NONE);
1909
1910                 KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= (int) md->pmd_npmc,
1911                     ("[pmc,%d] Illegal ref count %d on pp %p", __LINE__,
1912                         pp->pp_refcnt, pp));
1913
1914                 /*
1915                  * If this process is no longer the target of any
1916                  * PMCs, we can remove the process entry and free
1917                  * up space.
1918                  */
1919
1920                 if (pp->pp_refcnt == 0) {
1921                         pmc_remove_process_descriptor(pp);
1922                         free(pp, M_PMC);
1923                         break;
1924                 }
1925
1926         }
1927         break;
1928
1929         case PMC_FN_CSW_IN:
1930                 pmc_process_csw_in(td);
1931                 break;
1932
1933         case PMC_FN_CSW_OUT:
1934                 pmc_process_csw_out(td);
1935                 break;
1936
1937         /*
1938          * Process accumulated PC samples.
1939          *
1940          * This function is expected to be called by hardclock() for
1941          * each CPU that has accumulated PC samples.
1942          *
1943          * This function is to be executed on the CPU whose samples
1944          * are being processed.
1945          */
1946         case PMC_FN_DO_SAMPLES:
1947
1948                 /*
1949                  * Clear the cpu specific bit in the CPU mask before
1950                  * do the rest of the processing.  If the NMI handler
1951                  * gets invoked after the "atomic_clear_int()" call
1952                  * below but before "pmc_process_samples()" gets
1953                  * around to processing the interrupt, then we will
1954                  * come back here at the next hardclock() tick (and
1955                  * may find nothing to do if "pmc_process_samples()"
1956                  * had already processed the interrupt).  We don't
1957                  * lose the interrupt sample.
1958                  */
1959                 CPU_CLR_ATOMIC(PCPU_GET(cpuid), &pmc_cpumask);
1960                 pmc_process_samples(PCPU_GET(cpuid), PMC_HR);
1961                 pmc_process_samples(PCPU_GET(cpuid), PMC_SR);
1962                 break;
1963
1964         case PMC_FN_MMAP:
1965                 sx_assert(&pmc_sx, SX_LOCKED);
1966                 pmc_process_mmap(td, (struct pmckern_map_in *) arg);
1967                 break;
1968
1969         case PMC_FN_MUNMAP:
1970                 sx_assert(&pmc_sx, SX_LOCKED);
1971                 pmc_process_munmap(td, (struct pmckern_map_out *) arg);
1972                 break;
1973
1974         case PMC_FN_USER_CALLCHAIN:
1975                 /*
1976                  * Record a call chain.
1977                  */
1978                 KASSERT(td == curthread, ("[pmc,%d] td != curthread",
1979                     __LINE__));
1980
1981                 pmc_capture_user_callchain(PCPU_GET(cpuid), PMC_HR,
1982                     (struct trapframe *) arg);
1983                 td->td_pflags &= ~TDP_CALLCHAIN;
1984                 break;
1985
1986         case PMC_FN_USER_CALLCHAIN_SOFT:
1987                 /*
1988                  * Record a call chain.
1989                  */
1990                 KASSERT(td == curthread, ("[pmc,%d] td != curthread",
1991                     __LINE__));
1992                 pmc_capture_user_callchain(PCPU_GET(cpuid), PMC_SR,
1993                     (struct trapframe *) arg);
1994                 td->td_pflags &= ~TDP_CALLCHAIN;
1995                 break;
1996
1997         case PMC_FN_SOFT_SAMPLING:
1998                 /*
1999                  * Call soft PMC sampling intr.
2000                  */
2001                 pmc_soft_intr((struct pmckern_soft *) arg);
2002                 break;
2003
2004         default:
2005 #ifdef  DEBUG
2006                 KASSERT(0, ("[pmc,%d] unknown hook %d\n", __LINE__, function));
2007 #endif
2008                 break;
2009
2010         }
2011
2012         return 0;
2013 }
2014
2015 /*
2016  * allocate a 'struct pmc_owner' descriptor in the owner hash table.
2017  */
2018
2019 static struct pmc_owner *
2020 pmc_allocate_owner_descriptor(struct proc *p)
2021 {
2022         uint32_t hindex;
2023         struct pmc_owner *po;
2024         struct pmc_ownerhash *poh;
2025
2026         hindex = PMC_HASH_PTR(p, pmc_ownerhashmask);
2027         poh = &pmc_ownerhash[hindex];
2028
2029         /* allocate space for N pointers and one descriptor struct */
2030         po = malloc(sizeof(struct pmc_owner), M_PMC, M_WAITOK|M_ZERO);
2031         po->po_sscount = po->po_error = po->po_flags = po->po_logprocmaps = 0;
2032         po->po_file  = NULL;
2033         po->po_owner = p;
2034         po->po_kthread = NULL;
2035         LIST_INIT(&po->po_pmcs);
2036         LIST_INSERT_HEAD(poh, po, po_next); /* insert into hash table */
2037
2038         TAILQ_INIT(&po->po_logbuffers);
2039         mtx_init(&po->po_mtx, "pmc-owner-mtx", "pmc-per-proc", MTX_SPIN);
2040
2041         PMCDBG(OWN,ALL,1, "allocate-owner proc=%p (%d, %s) pmc-owner=%p",
2042             p, p->p_pid, p->p_comm, po);
2043
2044         return po;
2045 }
2046
2047 static void
2048 pmc_destroy_owner_descriptor(struct pmc_owner *po)
2049 {
2050
2051         PMCDBG(OWN,REL,1, "destroy-owner po=%p proc=%p (%d, %s)",
2052             po, po->po_owner, po->po_owner->p_pid, po->po_owner->p_comm);
2053
2054         mtx_destroy(&po->po_mtx);
2055         free(po, M_PMC);
2056 }
2057
2058 /*
2059  * find the descriptor corresponding to process 'p', adding or removing it
2060  * as specified by 'mode'.
2061  */
2062
2063 static struct pmc_process *
2064 pmc_find_process_descriptor(struct proc *p, uint32_t mode)
2065 {
2066         uint32_t hindex;
2067         struct pmc_process *pp, *ppnew;
2068         struct pmc_processhash *pph;
2069
2070         hindex = PMC_HASH_PTR(p, pmc_processhashmask);
2071         pph = &pmc_processhash[hindex];
2072
2073         ppnew = NULL;
2074
2075         /*
2076          * Pre-allocate memory in the FIND_ALLOCATE case since we
2077          * cannot call malloc(9) once we hold a spin lock.
2078          */
2079         if (mode & PMC_FLAG_ALLOCATE)
2080                 ppnew = malloc(sizeof(struct pmc_process) + md->pmd_npmc *
2081                     sizeof(struct pmc_targetstate), M_PMC, M_WAITOK|M_ZERO);
2082
2083         mtx_lock_spin(&pmc_processhash_mtx);
2084         LIST_FOREACH(pp, pph, pp_next)
2085             if (pp->pp_proc == p)
2086                     break;
2087
2088         if ((mode & PMC_FLAG_REMOVE) && pp != NULL)
2089                 LIST_REMOVE(pp, pp_next);
2090
2091         if ((mode & PMC_FLAG_ALLOCATE) && pp == NULL &&
2092             ppnew != NULL) {
2093                 ppnew->pp_proc = p;
2094                 LIST_INSERT_HEAD(pph, ppnew, pp_next);
2095                 pp = ppnew;
2096                 ppnew = NULL;
2097         }
2098         mtx_unlock_spin(&pmc_processhash_mtx);
2099
2100         if (pp != NULL && ppnew != NULL)
2101                 free(ppnew, M_PMC);
2102
2103         return pp;
2104 }
2105
2106 /*
2107  * remove a process descriptor from the process hash table.
2108  */
2109
2110 static void
2111 pmc_remove_process_descriptor(struct pmc_process *pp)
2112 {
2113         KASSERT(pp->pp_refcnt == 0,
2114             ("[pmc,%d] Removing process descriptor %p with count %d",
2115                 __LINE__, pp, pp->pp_refcnt));
2116
2117         mtx_lock_spin(&pmc_processhash_mtx);
2118         LIST_REMOVE(pp, pp_next);
2119         mtx_unlock_spin(&pmc_processhash_mtx);
2120 }
2121
2122
2123 /*
2124  * find an owner descriptor corresponding to proc 'p'
2125  */
2126
2127 static struct pmc_owner *
2128 pmc_find_owner_descriptor(struct proc *p)
2129 {
2130         uint32_t hindex;
2131         struct pmc_owner *po;
2132         struct pmc_ownerhash *poh;
2133
2134         hindex = PMC_HASH_PTR(p, pmc_ownerhashmask);
2135         poh = &pmc_ownerhash[hindex];
2136
2137         po = NULL;
2138         LIST_FOREACH(po, poh, po_next)
2139             if (po->po_owner == p)
2140                     break;
2141
2142         PMCDBG(OWN,FND,1, "find-owner proc=%p (%d, %s) hindex=0x%x -> "
2143             "pmc-owner=%p", p, p->p_pid, p->p_comm, hindex, po);
2144
2145         return po;
2146 }
2147
2148 /*
2149  * pmc_allocate_pmc_descriptor
2150  *
2151  * Allocate a pmc descriptor and initialize its
2152  * fields.
2153  */
2154
2155 static struct pmc *
2156 pmc_allocate_pmc_descriptor(void)
2157 {
2158         struct pmc *pmc;
2159
2160         pmc = malloc(sizeof(struct pmc), M_PMC, M_WAITOK|M_ZERO);
2161
2162         if (pmc != NULL) {
2163                 pmc->pm_owner = NULL;
2164                 LIST_INIT(&pmc->pm_targets);
2165         }
2166
2167         PMCDBG(PMC,ALL,1, "allocate-pmc -> pmc=%p", pmc);
2168
2169         return pmc;
2170 }
2171
2172 /*
2173  * Destroy a pmc descriptor.
2174  */
2175
2176 static void
2177 pmc_destroy_pmc_descriptor(struct pmc *pm)
2178 {
2179         (void) pm;
2180
2181 #ifdef  DEBUG
2182         KASSERT(pm->pm_state == PMC_STATE_DELETED ||
2183             pm->pm_state == PMC_STATE_FREE,
2184             ("[pmc,%d] destroying non-deleted PMC", __LINE__));
2185         KASSERT(LIST_EMPTY(&pm->pm_targets),
2186             ("[pmc,%d] destroying pmc with targets", __LINE__));
2187         KASSERT(pm->pm_owner == NULL,
2188             ("[pmc,%d] destroying pmc attached to an owner", __LINE__));
2189         KASSERT(pm->pm_runcount == 0,
2190             ("[pmc,%d] pmc has non-zero run count %d", __LINE__,
2191                 pm->pm_runcount));
2192 #endif
2193 }
2194
2195 static void
2196 pmc_wait_for_pmc_idle(struct pmc *pm)
2197 {
2198 #ifdef DEBUG
2199         volatile int maxloop;
2200
2201         maxloop = 100 * pmc_cpu_max();
2202 #endif
2203         /*
2204          * Loop (with a forced context switch) till the PMC's runcount
2205          * comes down to zero.
2206          */
2207         while (atomic_load_acq_32(&pm->pm_runcount) > 0) {
2208 #ifdef DEBUG
2209                 maxloop--;
2210                 KASSERT(maxloop > 0,
2211                     ("[pmc,%d] (ri%d, rc%d) waiting too long for "
2212                         "pmc to be free", __LINE__,
2213                         PMC_TO_ROWINDEX(pm), pm->pm_runcount));
2214 #endif
2215                 pmc_force_context_switch();
2216         }
2217 }
2218
2219 /*
2220  * This function does the following things:
2221  *
2222  *  - detaches the PMC from hardware
2223  *  - unlinks all target threads that were attached to it
2224  *  - removes the PMC from its owner's list
2225  *  - destroy's the PMC private mutex
2226  *
2227  * Once this function completes, the given pmc pointer can be safely
2228  * FREE'd by the caller.
2229  */
2230
2231 static void
2232 pmc_release_pmc_descriptor(struct pmc *pm)
2233 {
2234         enum pmc_mode mode;
2235         struct pmc_hw *phw;
2236         u_int adjri, ri, cpu;
2237         struct pmc_owner *po;
2238         struct pmc_binding pb;
2239         struct pmc_process *pp;
2240         struct pmc_classdep *pcd;
2241         struct pmc_target *ptgt, *tmp;
2242
2243         sx_assert(&pmc_sx, SX_XLOCKED);
2244
2245         KASSERT(pm, ("[pmc,%d] null pmc", __LINE__));
2246
2247         ri   = PMC_TO_ROWINDEX(pm);
2248         pcd  = pmc_ri_to_classdep(md, ri, &adjri);
2249         mode = PMC_TO_MODE(pm);
2250
2251         PMCDBG(PMC,REL,1, "release-pmc pmc=%p ri=%d mode=%d", pm, ri,
2252             mode);
2253
2254         /*
2255          * First, we take the PMC off hardware.
2256          */
2257         cpu = 0;
2258         if (PMC_IS_SYSTEM_MODE(mode)) {
2259
2260                 /*
2261                  * A system mode PMC runs on a specific CPU.  Switch
2262                  * to this CPU and turn hardware off.
2263                  */
2264                 pmc_save_cpu_binding(&pb);
2265
2266                 cpu = PMC_TO_CPU(pm);
2267
2268                 pmc_select_cpu(cpu);
2269
2270                 /* switch off non-stalled CPUs */
2271                 if (pm->pm_state == PMC_STATE_RUNNING &&
2272                     pm->pm_stalled == 0) {
2273
2274                         phw = pmc_pcpu[cpu]->pc_hwpmcs[ri];
2275
2276                         KASSERT(phw->phw_pmc == pm,
2277                             ("[pmc, %d] pmc ptr ri(%d) hw(%p) pm(%p)",
2278                                 __LINE__, ri, phw->phw_pmc, pm));
2279                         PMCDBG(PMC,REL,2, "stopping cpu=%d ri=%d", cpu, ri);
2280
2281                         critical_enter();
2282                         pcd->pcd_stop_pmc(cpu, adjri);
2283                         critical_exit();
2284                 }
2285
2286                 PMCDBG(PMC,REL,2, "decfg cpu=%d ri=%d", cpu, ri);
2287
2288                 critical_enter();
2289                 pcd->pcd_config_pmc(cpu, adjri, NULL);
2290                 critical_exit();
2291
2292                 /* adjust the global and process count of SS mode PMCs */
2293                 if (mode == PMC_MODE_SS && pm->pm_state == PMC_STATE_RUNNING) {
2294                         po = pm->pm_owner;
2295                         po->po_sscount--;
2296                         if (po->po_sscount == 0) {
2297                                 atomic_subtract_rel_int(&pmc_ss_count, 1);
2298                                 LIST_REMOVE(po, po_ssnext);
2299                         }
2300                 }
2301
2302                 pm->pm_state = PMC_STATE_DELETED;
2303
2304                 pmc_restore_cpu_binding(&pb);
2305
2306                 /*
2307                  * We could have references to this PMC structure in
2308                  * the per-cpu sample queues.  Wait for the queue to
2309                  * drain.
2310                  */
2311                 pmc_wait_for_pmc_idle(pm);
2312
2313         } else if (PMC_IS_VIRTUAL_MODE(mode)) {
2314
2315                 /*
2316                  * A virtual PMC could be running on multiple CPUs at
2317                  * a given instant.
2318                  *
2319                  * By marking its state as DELETED, we ensure that
2320                  * this PMC is never further scheduled on hardware.
2321                  *
2322                  * Then we wait till all CPUs are done with this PMC.
2323                  */
2324                 pm->pm_state = PMC_STATE_DELETED;
2325
2326
2327                 /* Wait for the PMCs runcount to come to zero. */
2328                 pmc_wait_for_pmc_idle(pm);
2329
2330                 /*
2331                  * At this point the PMC is off all CPUs and cannot be
2332                  * freshly scheduled onto a CPU.  It is now safe to
2333                  * unlink all targets from this PMC.  If a
2334                  * process-record's refcount falls to zero, we remove
2335                  * it from the hash table.  The module-wide SX lock
2336                  * protects us from races.
2337                  */
2338                 LIST_FOREACH_SAFE(ptgt, &pm->pm_targets, pt_next, tmp) {
2339                         pp = ptgt->pt_process;
2340                         pmc_unlink_target_process(pm, pp); /* frees 'ptgt' */
2341
2342                         PMCDBG(PMC,REL,3, "pp->refcnt=%d", pp->pp_refcnt);
2343
2344                         /*
2345                          * If the target process record shows that no
2346                          * PMCs are attached to it, reclaim its space.
2347                          */
2348
2349                         if (pp->pp_refcnt == 0) {
2350                                 pmc_remove_process_descriptor(pp);
2351                                 free(pp, M_PMC);
2352                         }
2353                 }
2354
2355                 cpu = curthread->td_oncpu; /* setup cpu for pmd_release() */
2356
2357         }
2358
2359         /*
2360          * Release any MD resources
2361          */
2362         (void) pcd->pcd_release_pmc(cpu, adjri, pm);
2363
2364         /*
2365          * Update row disposition
2366          */
2367
2368         if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm)))
2369                 PMC_UNMARK_ROW_STANDALONE(ri);
2370         else
2371                 PMC_UNMARK_ROW_THREAD(ri);
2372
2373         /* unlink from the owner's list */
2374         if (pm->pm_owner) {
2375                 LIST_REMOVE(pm, pm_next);
2376                 pm->pm_owner = NULL;
2377         }
2378
2379         pmc_destroy_pmc_descriptor(pm);
2380 }
2381
2382 /*
2383  * Register an owner and a pmc.
2384  */
2385
2386 static int
2387 pmc_register_owner(struct proc *p, struct pmc *pmc)
2388 {
2389         struct pmc_owner *po;
2390
2391         sx_assert(&pmc_sx, SX_XLOCKED);
2392
2393         if ((po = pmc_find_owner_descriptor(p)) == NULL)
2394                 if ((po = pmc_allocate_owner_descriptor(p)) == NULL)
2395                         return ENOMEM;
2396
2397         KASSERT(pmc->pm_owner == NULL,
2398             ("[pmc,%d] attempting to own an initialized PMC", __LINE__));
2399         pmc->pm_owner  = po;
2400
2401         LIST_INSERT_HEAD(&po->po_pmcs, pmc, pm_next);
2402
2403         PROC_LOCK(p);
2404         p->p_flag |= P_HWPMC;
2405         PROC_UNLOCK(p);
2406
2407         if (po->po_flags & PMC_PO_OWNS_LOGFILE)
2408                 pmclog_process_pmcallocate(pmc);
2409
2410         PMCDBG(PMC,REG,1, "register-owner pmc-owner=%p pmc=%p",
2411             po, pmc);
2412
2413         return 0;
2414 }
2415
2416 /*
2417  * Return the current row disposition:
2418  * == 0 => FREE
2419  *  > 0 => PROCESS MODE
2420  *  < 0 => SYSTEM MODE
2421  */
2422
2423 int
2424 pmc_getrowdisp(int ri)
2425 {
2426         return pmc_pmcdisp[ri];
2427 }
2428
2429 /*
2430  * Check if a PMC at row index 'ri' can be allocated to the current
2431  * process.
2432  *
2433  * Allocation can fail if:
2434  *   - the current process is already being profiled by a PMC at index 'ri',
2435  *     attached to it via OP_PMCATTACH.
2436  *   - the current process has already allocated a PMC at index 'ri'
2437  *     via OP_ALLOCATE.
2438  */
2439
2440 static int
2441 pmc_can_allocate_rowindex(struct proc *p, unsigned int ri, int cpu)
2442 {
2443         enum pmc_mode mode;
2444         struct pmc *pm;
2445         struct pmc_owner *po;
2446         struct pmc_process *pp;
2447
2448         PMCDBG(PMC,ALR,1, "can-allocate-rowindex proc=%p (%d, %s) ri=%d "
2449             "cpu=%d", p, p->p_pid, p->p_comm, ri, cpu);
2450
2451         /*
2452          * We shouldn't have already allocated a process-mode PMC at
2453          * row index 'ri'.
2454          *
2455          * We shouldn't have allocated a system-wide PMC on the same
2456          * CPU and same RI.
2457          */
2458         if ((po = pmc_find_owner_descriptor(p)) != NULL)
2459                 LIST_FOREACH(pm, &po->po_pmcs, pm_next) {
2460                     if (PMC_TO_ROWINDEX(pm) == ri) {
2461                             mode = PMC_TO_MODE(pm);
2462                             if (PMC_IS_VIRTUAL_MODE(mode))
2463                                     return EEXIST;
2464                             if (PMC_IS_SYSTEM_MODE(mode) &&
2465                                 (int) PMC_TO_CPU(pm) == cpu)
2466                                     return EEXIST;
2467                     }
2468                 }
2469
2470         /*
2471          * We also shouldn't be the target of any PMC at this index
2472          * since otherwise a PMC_ATTACH to ourselves will fail.
2473          */
2474         if ((pp = pmc_find_process_descriptor(p, 0)) != NULL)
2475                 if (pp->pp_pmcs[ri].pp_pmc)
2476                         return EEXIST;
2477
2478         PMCDBG(PMC,ALR,2, "can-allocate-rowindex proc=%p (%d, %s) ri=%d ok",
2479             p, p->p_pid, p->p_comm, ri);
2480
2481         return 0;
2482 }
2483
2484 /*
2485  * Check if a given PMC at row index 'ri' can be currently used in
2486  * mode 'mode'.
2487  */
2488
2489 static int
2490 pmc_can_allocate_row(int ri, enum pmc_mode mode)
2491 {
2492         enum pmc_disp   disp;
2493
2494         sx_assert(&pmc_sx, SX_XLOCKED);
2495
2496         PMCDBG(PMC,ALR,1, "can-allocate-row ri=%d mode=%d", ri, mode);
2497
2498         if (PMC_IS_SYSTEM_MODE(mode))
2499                 disp = PMC_DISP_STANDALONE;
2500         else
2501                 disp = PMC_DISP_THREAD;
2502
2503         /*
2504          * check disposition for PMC row 'ri':
2505          *
2506          * Expected disposition         Row-disposition         Result
2507          *
2508          * STANDALONE                   STANDALONE or FREE      proceed
2509          * STANDALONE                   THREAD                  fail
2510          * THREAD                       THREAD or FREE          proceed
2511          * THREAD                       STANDALONE              fail
2512          */
2513
2514         if (!PMC_ROW_DISP_IS_FREE(ri) &&
2515             !(disp == PMC_DISP_THREAD && PMC_ROW_DISP_IS_THREAD(ri)) &&
2516             !(disp == PMC_DISP_STANDALONE && PMC_ROW_DISP_IS_STANDALONE(ri)))
2517                 return EBUSY;
2518
2519         /*
2520          * All OK
2521          */
2522
2523         PMCDBG(PMC,ALR,2, "can-allocate-row ri=%d mode=%d ok", ri, mode);
2524
2525         return 0;
2526
2527 }
2528
2529 /*
2530  * Find a PMC descriptor with user handle 'pmcid' for thread 'td'.
2531  */
2532
2533 static struct pmc *
2534 pmc_find_pmc_descriptor_in_process(struct pmc_owner *po, pmc_id_t pmcid)
2535 {
2536         struct pmc *pm;
2537
2538         KASSERT(PMC_ID_TO_ROWINDEX(pmcid) < md->pmd_npmc,
2539             ("[pmc,%d] Illegal pmc index %d (max %d)", __LINE__,
2540                 PMC_ID_TO_ROWINDEX(pmcid), md->pmd_npmc));
2541
2542         LIST_FOREACH(pm, &po->po_pmcs, pm_next)
2543             if (pm->pm_id == pmcid)
2544                     return pm;
2545
2546         return NULL;
2547 }
2548
2549 static int
2550 pmc_find_pmc(pmc_id_t pmcid, struct pmc **pmc)
2551 {
2552
2553         struct pmc *pm;
2554         struct pmc_owner *po;
2555
2556         PMCDBG(PMC,FND,1, "find-pmc id=%d", pmcid);
2557
2558         if ((po = pmc_find_owner_descriptor(curthread->td_proc)) == NULL)
2559                 return ESRCH;
2560
2561         if ((pm = pmc_find_pmc_descriptor_in_process(po, pmcid)) == NULL)
2562                 return EINVAL;
2563
2564         PMCDBG(PMC,FND,2, "find-pmc id=%d -> pmc=%p", pmcid, pm);
2565
2566         *pmc = pm;
2567         return 0;
2568 }
2569
2570 /*
2571  * Start a PMC.
2572  */
2573
2574 static int
2575 pmc_start(struct pmc *pm)
2576 {
2577         enum pmc_mode mode;
2578         struct pmc_owner *po;
2579         struct pmc_binding pb;
2580         struct pmc_classdep *pcd;
2581         int adjri, error, cpu, ri;
2582
2583         KASSERT(pm != NULL,
2584             ("[pmc,%d] null pm", __LINE__));
2585
2586         mode = PMC_TO_MODE(pm);
2587         ri   = PMC_TO_ROWINDEX(pm);
2588         pcd  = pmc_ri_to_classdep(md, ri, &adjri);
2589
2590         error = 0;
2591
2592         PMCDBG(PMC,OPS,1, "start pmc=%p mode=%d ri=%d", pm, mode, ri);
2593
2594         po = pm->pm_owner;
2595
2596         /*
2597          * Disallow PMCSTART if a logfile is required but has not been
2598          * configured yet.
2599          */
2600         if ((pm->pm_flags & PMC_F_NEEDS_LOGFILE) &&
2601             (po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
2602                 return (EDOOFUS);       /* programming error */
2603
2604         /*
2605          * If this is a sampling mode PMC, log mapping information for
2606          * the kernel modules that are currently loaded.
2607          */
2608         if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
2609             pmc_log_kernel_mappings(pm);
2610
2611         if (PMC_IS_VIRTUAL_MODE(mode)) {
2612
2613                 /*
2614                  * If a PMCATTACH has never been done on this PMC,
2615                  * attach it to its owner process.
2616                  */
2617
2618                 if (LIST_EMPTY(&pm->pm_targets))
2619                         error = (pm->pm_flags & PMC_F_ATTACH_DONE) ? ESRCH :
2620                             pmc_attach_process(po->po_owner, pm);
2621
2622                 /*
2623                  * If the PMC is attached to its owner, then force a context
2624                  * switch to ensure that the MD state gets set correctly.
2625                  */
2626
2627                 if (error == 0) {
2628                         pm->pm_state = PMC_STATE_RUNNING;
2629                         if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER)
2630                                 pmc_force_context_switch();
2631                 }
2632
2633                 return (error);
2634         }
2635
2636
2637         /*
2638          * A system-wide PMC.
2639          *
2640          * Add the owner to the global list if this is a system-wide
2641          * sampling PMC.
2642          */
2643
2644         if (mode == PMC_MODE_SS) {
2645                 if (po->po_sscount == 0) {
2646                         LIST_INSERT_HEAD(&pmc_ss_owners, po, po_ssnext);
2647                         atomic_add_rel_int(&pmc_ss_count, 1);
2648                         PMCDBG(PMC,OPS,1, "po=%p in global list", po);
2649                 }
2650                 po->po_sscount++;
2651
2652                 /*
2653                  * Log mapping information for all existing processes in the
2654                  * system.  Subsequent mappings are logged as they happen;
2655                  * see pmc_process_mmap().
2656                  */
2657                 if (po->po_logprocmaps == 0) {
2658                         pmc_log_all_process_mappings(po);
2659                         po->po_logprocmaps = 1;
2660                 }
2661         }
2662
2663         /*
2664          * Move to the CPU associated with this
2665          * PMC, and start the hardware.
2666          */
2667
2668         pmc_save_cpu_binding(&pb);
2669
2670         cpu = PMC_TO_CPU(pm);
2671
2672         if (!pmc_cpu_is_active(cpu))
2673                 return (ENXIO);
2674
2675         pmc_select_cpu(cpu);
2676
2677         /*
2678          * global PMCs are configured at allocation time
2679          * so write out the initial value and start the PMC.
2680          */
2681
2682         pm->pm_state = PMC_STATE_RUNNING;
2683
2684         critical_enter();
2685         if ((error = pcd->pcd_write_pmc(cpu, adjri,
2686                  PMC_IS_SAMPLING_MODE(mode) ?
2687                  pm->pm_sc.pm_reloadcount :
2688                  pm->pm_sc.pm_initial)) == 0)
2689                 error = pcd->pcd_start_pmc(cpu, adjri);
2690         critical_exit();
2691
2692         pmc_restore_cpu_binding(&pb);
2693
2694         return (error);
2695 }
2696
2697 /*
2698  * Stop a PMC.
2699  */
2700
2701 static int
2702 pmc_stop(struct pmc *pm)
2703 {
2704         struct pmc_owner *po;
2705         struct pmc_binding pb;
2706         struct pmc_classdep *pcd;
2707         int adjri, cpu, error, ri;
2708
2709         KASSERT(pm != NULL, ("[pmc,%d] null pmc", __LINE__));
2710
2711         PMCDBG(PMC,OPS,1, "stop pmc=%p mode=%d ri=%d", pm,
2712             PMC_TO_MODE(pm), PMC_TO_ROWINDEX(pm));
2713
2714         pm->pm_state = PMC_STATE_STOPPED;
2715
2716         /*
2717          * If the PMC is a virtual mode one, changing the state to
2718          * non-RUNNING is enough to ensure that the PMC never gets
2719          * scheduled.
2720          *
2721          * If this PMC is current running on a CPU, then it will
2722          * handled correctly at the time its target process is context
2723          * switched out.
2724          */
2725
2726         if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)))
2727                 return 0;
2728
2729         /*
2730          * A system-mode PMC.  Move to the CPU associated with
2731          * this PMC, and stop the hardware.  We update the
2732          * 'initial count' so that a subsequent PMCSTART will
2733          * resume counting from the current hardware count.
2734          */
2735
2736         pmc_save_cpu_binding(&pb);
2737
2738         cpu = PMC_TO_CPU(pm);
2739
2740         KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
2741             ("[pmc,%d] illegal cpu=%d", __LINE__, cpu));
2742
2743         if (!pmc_cpu_is_active(cpu))
2744                 return ENXIO;
2745
2746         pmc_select_cpu(cpu);
2747
2748         ri = PMC_TO_ROWINDEX(pm);
2749         pcd = pmc_ri_to_classdep(md, ri, &adjri);
2750
2751         critical_enter();
2752         if ((error = pcd->pcd_stop_pmc(cpu, adjri)) == 0)
2753                 error = pcd->pcd_read_pmc(cpu, adjri, &pm->pm_sc.pm_initial);
2754         critical_exit();
2755
2756         pmc_restore_cpu_binding(&pb);
2757
2758         po = pm->pm_owner;
2759
2760         /* remove this owner from the global list of SS PMC owners */
2761         if (PMC_TO_MODE(pm) == PMC_MODE_SS) {
2762                 po->po_sscount--;
2763                 if (po->po_sscount == 0) {
2764                         atomic_subtract_rel_int(&pmc_ss_count, 1);
2765                         LIST_REMOVE(po, po_ssnext);
2766                         PMCDBG(PMC,OPS,2,"po=%p removed from global list", po);
2767                 }
2768         }
2769
2770         return (error);
2771 }
2772
2773
2774 #ifdef  DEBUG
2775 static const char *pmc_op_to_name[] = {
2776 #undef  __PMC_OP
2777 #define __PMC_OP(N, D)  #N ,
2778         __PMC_OPS()
2779         NULL
2780 };
2781 #endif
2782
2783 /*
2784  * The syscall interface
2785  */
2786
2787 #define PMC_GET_SX_XLOCK(...) do {              \
2788         sx_xlock(&pmc_sx);                      \
2789         if (pmc_hook == NULL) {                 \
2790                 sx_xunlock(&pmc_sx);            \
2791                 return __VA_ARGS__;             \
2792         }                                       \
2793 } while (0)
2794
2795 #define PMC_DOWNGRADE_SX() do {                 \
2796         sx_downgrade(&pmc_sx);                  \
2797         is_sx_downgraded = 1;                   \
2798 } while (0)
2799
2800 static int
2801 pmc_syscall_handler(struct thread *td, void *syscall_args)
2802 {
2803         int error, is_sx_downgraded, is_sx_locked, op;
2804         struct pmc_syscall_args *c;
2805         void *arg;
2806
2807         PMC_GET_SX_XLOCK(ENOSYS);
2808
2809         DROP_GIANT();
2810
2811         is_sx_downgraded = 0;
2812         is_sx_locked = 1;
2813
2814         c = (struct pmc_syscall_args *) syscall_args;
2815
2816         op = c->pmop_code;
2817         arg = c->pmop_data;
2818
2819         PMCDBG(MOD,PMS,1, "syscall op=%d \"%s\" arg=%p", op,
2820             pmc_op_to_name[op], arg);
2821
2822         error = 0;
2823         atomic_add_int(&pmc_stats.pm_syscalls, 1);
2824
2825         switch(op)
2826         {
2827
2828
2829         /*
2830          * Configure a log file.
2831          *
2832          * XXX This OP will be reworked.
2833          */
2834
2835         case PMC_OP_CONFIGURELOG:
2836         {
2837                 struct proc *p;
2838                 struct pmc *pm;
2839                 struct pmc_owner *po;
2840                 struct pmc_op_configurelog cl;
2841
2842                 sx_assert(&pmc_sx, SX_XLOCKED);
2843
2844                 if ((error = copyin(arg, &cl, sizeof(cl))) != 0)
2845                         break;
2846
2847                 /* mark this process as owning a log file */
2848                 p = td->td_proc;
2849                 if ((po = pmc_find_owner_descriptor(p)) == NULL)
2850                         if ((po = pmc_allocate_owner_descriptor(p)) == NULL) {
2851                                 error = ENOMEM;
2852                                 break;
2853                         }
2854
2855                 /*
2856                  * If a valid fd was passed in, try to configure that,
2857                  * otherwise if 'fd' was less than zero and there was
2858                  * a log file configured, flush its buffers and
2859                  * de-configure it.
2860                  */
2861                 if (cl.pm_logfd >= 0) {
2862                         sx_xunlock(&pmc_sx);
2863                         is_sx_locked = 0;
2864                         error = pmclog_configure_log(md, po, cl.pm_logfd);
2865                 } else if (po->po_flags & PMC_PO_OWNS_LOGFILE) {
2866                         pmclog_process_closelog(po);
2867                         error = pmclog_close(po);
2868                         if (error == 0) {
2869                                 LIST_FOREACH(pm, &po->po_pmcs, pm_next)
2870                                     if (pm->pm_flags & PMC_F_NEEDS_LOGFILE &&
2871                                         pm->pm_state == PMC_STATE_RUNNING)
2872                                             pmc_stop(pm);
2873                                 error = pmclog_deconfigure_log(po);
2874                         }
2875                 } else
2876                         error = EINVAL;
2877
2878                 if (error)
2879                         break;
2880         }
2881         break;
2882
2883         /*
2884          * Flush a log file.
2885          */
2886
2887         case PMC_OP_FLUSHLOG:
2888         {
2889                 struct pmc_owner *po;
2890
2891                 sx_assert(&pmc_sx, SX_XLOCKED);
2892
2893                 if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) {
2894                         error = EINVAL;
2895                         break;
2896                 }
2897
2898                 error = pmclog_flush(po);
2899         }
2900         break;
2901
2902         /*
2903          * Close a log file.
2904          */
2905
2906         case PMC_OP_CLOSELOG:
2907         {
2908                 struct pmc_owner *po;
2909
2910                 sx_assert(&pmc_sx, SX_XLOCKED);
2911
2912                 if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) {
2913                         error = EINVAL;
2914                         break;
2915                 }
2916
2917                 error = pmclog_close(po);
2918         }
2919         break;
2920
2921         /*
2922          * Retrieve hardware configuration.
2923          */
2924
2925         case PMC_OP_GETCPUINFO: /* CPU information */
2926         {
2927                 struct pmc_op_getcpuinfo gci;
2928                 struct pmc_classinfo *pci;
2929                 struct pmc_classdep *pcd;
2930                 int cl;
2931
2932                 gci.pm_cputype = md->pmd_cputype;
2933                 gci.pm_ncpu    = pmc_cpu_max();
2934                 gci.pm_npmc    = md->pmd_npmc;
2935                 gci.pm_nclass  = md->pmd_nclass;
2936                 pci = gci.pm_classes;
2937                 pcd = md->pmd_classdep;
2938                 for (cl = 0; cl < md->pmd_nclass; cl++, pci++, pcd++) {
2939                         pci->pm_caps  = pcd->pcd_caps;
2940                         pci->pm_class = pcd->pcd_class;
2941                         pci->pm_width = pcd->pcd_width;
2942                         pci->pm_num   = pcd->pcd_num;
2943                 }
2944                 error = copyout(&gci, arg, sizeof(gci));
2945         }
2946         break;
2947
2948         /*
2949          * Retrieve soft events list.
2950          */
2951         case PMC_OP_GETDYNEVENTINFO:
2952         {
2953                 enum pmc_class                  cl;
2954                 enum pmc_event                  ev;
2955                 struct pmc_op_getdyneventinfo   *gei;
2956                 struct pmc_dyn_event_descr      dev;
2957                 struct pmc_soft                 *ps;
2958                 uint32_t                        nevent;
2959
2960                 sx_assert(&pmc_sx, SX_LOCKED);
2961
2962                 gei = (struct pmc_op_getdyneventinfo *) arg;
2963
2964                 if ((error = copyin(&gei->pm_class, &cl, sizeof(cl))) != 0)
2965                         break;
2966
2967                 /* Only SOFT class is dynamic. */
2968                 if (cl != PMC_CLASS_SOFT) {
2969                         error = EINVAL;
2970                         break;
2971                 }
2972
2973                 nevent = 0;
2974                 for (ev = PMC_EV_SOFT_FIRST; ev <= PMC_EV_SOFT_LAST; ev++) {
2975                         ps = pmc_soft_ev_acquire(ev);
2976                         if (ps == NULL)
2977                                 continue;
2978                         bcopy(&ps->ps_ev, &dev, sizeof(dev));
2979                         pmc_soft_ev_release(ps);
2980
2981                         error = copyout(&dev,
2982                             &gei->pm_events[nevent],
2983                             sizeof(struct pmc_dyn_event_descr));
2984                         if (error != 0)
2985                                 break;
2986                         nevent++;
2987                 }
2988                 if (error != 0)
2989                         break;
2990
2991                 error = copyout(&nevent, &gei->pm_nevent,
2992                     sizeof(nevent));
2993         }
2994         break;
2995
2996         /*
2997          * Get module statistics
2998          */
2999
3000         case PMC_OP_GETDRIVERSTATS:
3001         {
3002                 struct pmc_op_getdriverstats gms;
3003
3004                 bcopy(&pmc_stats, &gms, sizeof(gms));
3005                 error = copyout(&gms, arg, sizeof(gms));
3006         }
3007         break;
3008
3009
3010         /*
3011          * Retrieve module version number
3012          */
3013
3014         case PMC_OP_GETMODULEVERSION:
3015         {
3016                 uint32_t cv, modv;
3017
3018                 /* retrieve the client's idea of the ABI version */
3019                 if ((error = copyin(arg, &cv, sizeof(uint32_t))) != 0)
3020                         break;
3021                 /* don't service clients newer than our driver */
3022                 modv = PMC_VERSION;
3023                 if ((cv & 0xFFFF0000) > (modv & 0xFFFF0000)) {
3024                         error = EPROGMISMATCH;
3025                         break;
3026                 }
3027                 error = copyout(&modv, arg, sizeof(int));
3028         }
3029         break;
3030
3031
3032         /*
3033          * Retrieve the state of all the PMCs on a given
3034          * CPU.
3035          */
3036
3037         case PMC_OP_GETPMCINFO:
3038         {
3039                 int ari;
3040                 struct pmc *pm;
3041                 size_t pmcinfo_size;
3042                 uint32_t cpu, n, npmc;
3043                 struct pmc_owner *po;
3044                 struct pmc_binding pb;
3045                 struct pmc_classdep *pcd;
3046                 struct pmc_info *p, *pmcinfo;
3047                 struct pmc_op_getpmcinfo *gpi;
3048
3049                 PMC_DOWNGRADE_SX();
3050
3051                 gpi = (struct pmc_op_getpmcinfo *) arg;
3052
3053                 if ((error = copyin(&gpi->pm_cpu, &cpu, sizeof(cpu))) != 0)
3054                         break;
3055
3056                 if (cpu >= pmc_cpu_max()) {
3057                         error = EINVAL;
3058                         break;
3059                 }
3060
3061                 if (!pmc_cpu_is_active(cpu)) {
3062                         error = ENXIO;
3063                         break;
3064                 }
3065
3066                 /* switch to CPU 'cpu' */
3067                 pmc_save_cpu_binding(&pb);
3068                 pmc_select_cpu(cpu);
3069
3070                 npmc = md->pmd_npmc;
3071
3072                 pmcinfo_size = npmc * sizeof(struct pmc_info);
3073                 pmcinfo = malloc(pmcinfo_size, M_PMC, M_WAITOK);
3074
3075                 p = pmcinfo;
3076
3077                 for (n = 0; n < md->pmd_npmc; n++, p++) {
3078
3079                         pcd = pmc_ri_to_classdep(md, n, &ari);
3080
3081                         KASSERT(pcd != NULL,
3082                             ("[pmc,%d] null pcd ri=%d", __LINE__, n));
3083
3084                         if ((error = pcd->pcd_describe(cpu, ari, p, &pm)) != 0)
3085                                 break;
3086
3087                         if (PMC_ROW_DISP_IS_STANDALONE(n))
3088                                 p->pm_rowdisp = PMC_DISP_STANDALONE;
3089                         else if (PMC_ROW_DISP_IS_THREAD(n))
3090                                 p->pm_rowdisp = PMC_DISP_THREAD;
3091                         else
3092                                 p->pm_rowdisp = PMC_DISP_FREE;
3093
3094                         p->pm_ownerpid = -1;
3095
3096                         if (pm == NULL) /* no PMC associated */
3097                                 continue;
3098
3099                         po = pm->pm_owner;
3100
3101                         KASSERT(po->po_owner != NULL,
3102                             ("[pmc,%d] pmc_owner had a null proc pointer",
3103                                 __LINE__));
3104
3105                         p->pm_ownerpid = po->po_owner->p_pid;
3106                         p->pm_mode     = PMC_TO_MODE(pm);
3107                         p->pm_event    = pm->pm_event;
3108                         p->pm_flags    = pm->pm_flags;
3109
3110                         if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
3111                                 p->pm_reloadcount =
3112                                     pm->pm_sc.pm_reloadcount;
3113                 }
3114
3115                 pmc_restore_cpu_binding(&pb);
3116
3117                 /* now copy out the PMC info collected */
3118                 if (error == 0)
3119                         error = copyout(pmcinfo, &gpi->pm_pmcs, pmcinfo_size);
3120
3121                 free(pmcinfo, M_PMC);
3122         }
3123         break;
3124
3125
3126         /*
3127          * Set the administrative state of a PMC.  I.e. whether
3128          * the PMC is to be used or not.
3129          */
3130
3131         case PMC_OP_PMCADMIN:
3132         {
3133                 int cpu, ri;
3134                 enum pmc_state request;
3135                 struct pmc_cpu *pc;
3136                 struct pmc_hw *phw;
3137                 struct pmc_op_pmcadmin pma;
3138                 struct pmc_binding pb;
3139
3140                 sx_assert(&pmc_sx, SX_XLOCKED);
3141
3142                 KASSERT(td == curthread,
3143                     ("[pmc,%d] td != curthread", __LINE__));
3144
3145                 error = priv_check(td, PRIV_PMC_MANAGE);
3146                 if (error)
3147                         break;
3148
3149                 if ((error = copyin(arg, &pma, sizeof(pma))) != 0)
3150                         break;
3151
3152                 cpu = pma.pm_cpu;
3153
3154                 if (cpu < 0 || cpu >= (int) pmc_cpu_max()) {
3155                         error = EINVAL;
3156                         break;
3157                 }
3158
3159                 if (!pmc_cpu_is_active(cpu)) {
3160                         error = ENXIO;
3161                         break;
3162                 }
3163
3164                 request = pma.pm_state;
3165
3166                 if (request != PMC_STATE_DISABLED &&
3167                     request != PMC_STATE_FREE) {
3168                         error = EINVAL;
3169                         break;
3170                 }
3171
3172                 ri = pma.pm_pmc; /* pmc id == row index */
3173                 if (ri < 0 || ri >= (int) md->pmd_npmc) {
3174                         error = EINVAL;
3175                         break;
3176                 }
3177
3178                 /*
3179                  * We can't disable a PMC with a row-index allocated
3180                  * for process virtual PMCs.
3181                  */
3182
3183                 if (PMC_ROW_DISP_IS_THREAD(ri) &&
3184                     request == PMC_STATE_DISABLED) {
3185                         error = EBUSY;
3186                         break;
3187                 }
3188
3189                 /*
3190                  * otherwise, this PMC on this CPU is either free or
3191                  * in system-wide mode.
3192                  */
3193
3194                 pmc_save_cpu_binding(&pb);
3195                 pmc_select_cpu(cpu);
3196
3197                 pc  = pmc_pcpu[cpu];
3198                 phw = pc->pc_hwpmcs[ri];
3199
3200                 /*
3201                  * XXX do we need some kind of 'forced' disable?
3202                  */
3203
3204                 if (phw->phw_pmc == NULL) {
3205                         if (request == PMC_STATE_DISABLED &&
3206                             (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED)) {
3207                                 phw->phw_state &= ~PMC_PHW_FLAG_IS_ENABLED;
3208                                 PMC_MARK_ROW_STANDALONE(ri);
3209                         } else if (request == PMC_STATE_FREE &&
3210                             (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0) {
3211                                 phw->phw_state |=  PMC_PHW_FLAG_IS_ENABLED;
3212                                 PMC_UNMARK_ROW_STANDALONE(ri);
3213                         }
3214                         /* other cases are a no-op */
3215                 } else
3216                         error = EBUSY;
3217
3218                 pmc_restore_cpu_binding(&pb);
3219         }
3220         break;
3221
3222
3223         /*
3224          * Allocate a PMC.
3225          */
3226
3227         case PMC_OP_PMCALLOCATE:
3228         {
3229                 int adjri, n;
3230                 u_int cpu;
3231                 uint32_t caps;
3232                 struct pmc *pmc;
3233                 enum pmc_mode mode;
3234                 struct pmc_hw *phw;
3235                 struct pmc_binding pb;
3236                 struct pmc_classdep *pcd;
3237                 struct pmc_op_pmcallocate pa;
3238
3239                 if ((error = copyin(arg, &pa, sizeof(pa))) != 0)
3240                         break;
3241
3242                 caps = pa.pm_caps;
3243                 mode = pa.pm_mode;
3244                 cpu  = pa.pm_cpu;
3245
3246                 if ((mode != PMC_MODE_SS  &&  mode != PMC_MODE_SC  &&
3247                      mode != PMC_MODE_TS  &&  mode != PMC_MODE_TC) ||
3248                     (cpu != (u_int) PMC_CPU_ANY && cpu >= pmc_cpu_max())) {
3249                         error = EINVAL;
3250                         break;
3251                 }
3252
3253                 /*
3254                  * Virtual PMCs should only ask for a default CPU.
3255                  * System mode PMCs need to specify a non-default CPU.
3256                  */
3257
3258                 if ((PMC_IS_VIRTUAL_MODE(mode) && cpu != (u_int) PMC_CPU_ANY) ||
3259                     (PMC_IS_SYSTEM_MODE(mode) && cpu == (u_int) PMC_CPU_ANY)) {
3260                         error = EINVAL;
3261                         break;
3262                 }
3263
3264                 /*
3265                  * Check that an inactive CPU is not being asked for.
3266                  */
3267
3268                 if (PMC_IS_SYSTEM_MODE(mode) && !pmc_cpu_is_active(cpu)) {
3269                         error = ENXIO;
3270                         break;
3271                 }
3272
3273                 /*
3274                  * Refuse an allocation for a system-wide PMC if this
3275                  * process has been jailed, or if this process lacks
3276                  * super-user credentials and the sysctl tunable
3277                  * 'security.bsd.unprivileged_syspmcs' is zero.
3278                  */
3279
3280                 if (PMC_IS_SYSTEM_MODE(mode)) {
3281                         if (jailed(curthread->td_ucred)) {
3282                                 error = EPERM;
3283                                 break;
3284                         }
3285                         if (!pmc_unprivileged_syspmcs) {
3286                                 error = priv_check(curthread,
3287                                     PRIV_PMC_SYSTEM);
3288                                 if (error)
3289                                         break;
3290                         }
3291                 }
3292
3293                 /*
3294                  * Look for valid values for 'pm_flags'
3295                  */
3296
3297                 if ((pa.pm_flags & ~(PMC_F_DESCENDANTS | PMC_F_LOG_PROCCSW |
3298                     PMC_F_LOG_PROCEXIT | PMC_F_CALLCHAIN)) != 0) {
3299                         error = EINVAL;
3300                         break;
3301                 }
3302
3303                 /* process logging options are not allowed for system PMCs */
3304                 if (PMC_IS_SYSTEM_MODE(mode) && (pa.pm_flags &
3305                     (PMC_F_LOG_PROCCSW | PMC_F_LOG_PROCEXIT))) {
3306                         error = EINVAL;
3307                         break;
3308                 }
3309
3310                 /*
3311                  * All sampling mode PMCs need to be able to interrupt the
3312                  * CPU.
3313                  */
3314                 if (PMC_IS_SAMPLING_MODE(mode))
3315                         caps |= PMC_CAP_INTERRUPT;
3316
3317                 /* A valid class specifier should have been passed in. */
3318                 for (n = 0; n < md->pmd_nclass; n++)
3319                         if (md->pmd_classdep[n].pcd_class == pa.pm_class)
3320                                 break;
3321                 if (n == md->pmd_nclass) {
3322                         error = EINVAL;
3323                         break;
3324                 }
3325
3326                 /* The requested PMC capabilities should be feasible. */
3327                 if ((md->pmd_classdep[n].pcd_caps & caps) != caps) {
3328                         error = EOPNOTSUPP;
3329                         break;
3330                 }
3331
3332                 PMCDBG(PMC,ALL,2, "event=%d caps=0x%x mode=%d cpu=%d",
3333                     pa.pm_ev, caps, mode, cpu);
3334
3335                 pmc = pmc_allocate_pmc_descriptor();
3336                 pmc->pm_id    = PMC_ID_MAKE_ID(cpu,pa.pm_mode,pa.pm_class,
3337                     PMC_ID_INVALID);
3338                 pmc->pm_event = pa.pm_ev;
3339                 pmc->pm_state = PMC_STATE_FREE;
3340                 pmc->pm_caps  = caps;
3341                 pmc->pm_flags = pa.pm_flags;
3342
3343                 /* switch thread to CPU 'cpu' */
3344                 pmc_save_cpu_binding(&pb);
3345
3346 #define PMC_IS_SHAREABLE_PMC(cpu, n)                            \
3347         (pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_state &           \
3348          PMC_PHW_FLAG_IS_SHAREABLE)
3349 #define PMC_IS_UNALLOCATED(cpu, n)                              \
3350         (pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_pmc == NULL)
3351
3352                 if (PMC_IS_SYSTEM_MODE(mode)) {
3353                         pmc_select_cpu(cpu);
3354                         for (n = 0; n < (int) md->pmd_npmc; n++) {
3355                                 pcd = pmc_ri_to_classdep(md, n, &adjri);
3356                                 if (pmc_can_allocate_row(n, mode) == 0 &&
3357                                     pmc_can_allocate_rowindex(
3358                                             curthread->td_proc, n, cpu) == 0 &&
3359                                     (PMC_IS_UNALLOCATED(cpu, n) ||
3360                                      PMC_IS_SHAREABLE_PMC(cpu, n)) &&
3361                                     pcd->pcd_allocate_pmc(cpu, adjri, pmc,
3362                                         &pa) == 0)
3363                                         break;
3364                         }
3365                 } else {
3366                         /* Process virtual mode */
3367                         for (n = 0; n < (int) md->pmd_npmc; n++) {
3368                                 pcd = pmc_ri_to_classdep(md, n, &adjri);
3369                                 if (pmc_can_allocate_row(n, mode) == 0 &&
3370                                     pmc_can_allocate_rowindex(
3371                                             curthread->td_proc, n,
3372                                             PMC_CPU_ANY) == 0 &&
3373                                     pcd->pcd_allocate_pmc(curthread->td_oncpu,
3374                                         adjri, pmc, &pa) == 0)
3375                                         break;
3376                         }
3377                 }
3378
3379 #undef  PMC_IS_UNALLOCATED
3380 #undef  PMC_IS_SHAREABLE_PMC
3381
3382                 pmc_restore_cpu_binding(&pb);
3383
3384                 if (n == (int) md->pmd_npmc) {
3385                         pmc_destroy_pmc_descriptor(pmc);
3386                         free(pmc, M_PMC);
3387                         pmc = NULL;
3388                         error = EINVAL;
3389                         break;
3390                 }
3391
3392                 /* Fill in the correct value in the ID field */
3393                 pmc->pm_id = PMC_ID_MAKE_ID(cpu,mode,pa.pm_class,n);
3394
3395                 PMCDBG(PMC,ALL,2, "ev=%d class=%d mode=%d n=%d -> pmcid=%x",
3396                     pmc->pm_event, pa.pm_class, mode, n, pmc->pm_id);
3397
3398                 /* Process mode PMCs with logging enabled need log files */
3399                 if (pmc->pm_flags & (PMC_F_LOG_PROCEXIT | PMC_F_LOG_PROCCSW))
3400                         pmc->pm_flags |= PMC_F_NEEDS_LOGFILE;
3401
3402                 /* All system mode sampling PMCs require a log file */
3403                 if (PMC_IS_SAMPLING_MODE(mode) && PMC_IS_SYSTEM_MODE(mode))
3404                         pmc->pm_flags |= PMC_F_NEEDS_LOGFILE;
3405
3406                 /*
3407                  * Configure global pmc's immediately
3408                  */
3409
3410                 if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pmc))) {
3411
3412                         pmc_save_cpu_binding(&pb);
3413                         pmc_select_cpu(cpu);
3414
3415                         phw = pmc_pcpu[cpu]->pc_hwpmcs[n];
3416                         pcd = pmc_ri_to_classdep(md, n, &adjri);
3417
3418                         if ((phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0 ||
3419                             (error = pcd->pcd_config_pmc(cpu, adjri, pmc)) != 0) {
3420                                 (void) pcd->pcd_release_pmc(cpu, adjri, pmc);
3421                                 pmc_destroy_pmc_descriptor(pmc);
3422                                 free(pmc, M_PMC);
3423                                 pmc = NULL;
3424                                 pmc_restore_cpu_binding(&pb);
3425                                 error = EPERM;
3426                                 break;
3427                         }
3428
3429                         pmc_restore_cpu_binding(&pb);
3430                 }
3431
3432                 pmc->pm_state    = PMC_STATE_ALLOCATED;
3433
3434                 /*
3435                  * mark row disposition
3436                  */
3437
3438                 if (PMC_IS_SYSTEM_MODE(mode))
3439                         PMC_MARK_ROW_STANDALONE(n);
3440                 else
3441                         PMC_MARK_ROW_THREAD(n);
3442
3443                 /*
3444                  * Register this PMC with the current thread as its owner.
3445                  */
3446
3447                 if ((error =
3448                     pmc_register_owner(curthread->td_proc, pmc)) != 0) {
3449                         pmc_release_pmc_descriptor(pmc);
3450                         free(pmc, M_PMC);
3451                         pmc = NULL;
3452                         break;
3453                 }
3454
3455                 /*
3456                  * Return the allocated index.
3457                  */
3458
3459                 pa.pm_pmcid = pmc->pm_id;
3460
3461                 error = copyout(&pa, arg, sizeof(pa));
3462         }
3463         break;
3464
3465
3466         /*
3467          * Attach a PMC to a process.
3468          */
3469
3470         case PMC_OP_PMCATTACH:
3471         {
3472                 struct pmc *pm;
3473                 struct proc *p;
3474                 struct pmc_op_pmcattach a;
3475
3476                 sx_assert(&pmc_sx, SX_XLOCKED);
3477
3478                 if ((error = copyin(arg, &a, sizeof(a))) != 0)
3479                         break;
3480
3481                 if (a.pm_pid < 0) {
3482                         error = EINVAL;
3483                         break;
3484                 } else if (a.pm_pid == 0)
3485                         a.pm_pid = td->td_proc->p_pid;
3486
3487                 if ((error = pmc_find_pmc(a.pm_pmc, &pm)) != 0)
3488                         break;
3489
3490                 if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm))) {
3491                         error = EINVAL;
3492                         break;
3493                 }
3494
3495                 /* PMCs may be (re)attached only when allocated or stopped */
3496                 if (pm->pm_state == PMC_STATE_RUNNING) {
3497                         error = EBUSY;
3498                         break;
3499                 } else if (pm->pm_state != PMC_STATE_ALLOCATED &&
3500                     pm->pm_state != PMC_STATE_STOPPED) {
3501                         error = EINVAL;
3502                         break;
3503                 }
3504
3505                 /* lookup pid */
3506                 if ((p = pfind(a.pm_pid)) == NULL) {
3507                         error = ESRCH;
3508                         break;
3509                 }
3510
3511                 /*
3512                  * Ignore processes that are working on exiting.
3513                  */
3514                 if (p->p_flag & P_WEXIT) {
3515                         error = ESRCH;
3516                         PROC_UNLOCK(p); /* pfind() returns a locked process */
3517                         break;
3518                 }
3519
3520                 /*
3521                  * we are allowed to attach a PMC to a process if
3522                  * we can debug it.
3523                  */
3524                 error = p_candebug(curthread, p);
3525
3526                 PROC_UNLOCK(p);
3527
3528                 if (error == 0)
3529                         error = pmc_attach_process(p, pm);
3530         }
3531         break;
3532
3533
3534         /*
3535          * Detach an attached PMC from a process.
3536          */
3537
3538         case PMC_OP_PMCDETACH:
3539         {
3540                 struct pmc *pm;
3541                 struct proc *p;
3542                 struct pmc_op_pmcattach a;
3543
3544                 if ((error = copyin(arg, &a, sizeof(a))) != 0)
3545                         break;
3546
3547                 if (a.pm_pid < 0) {
3548                         error = EINVAL;
3549                         break;
3550                 } else if (a.pm_pid == 0)
3551                         a.pm_pid = td->td_proc->p_pid;
3552
3553                 if ((error = pmc_find_pmc(a.pm_pmc, &pm)) != 0)
3554                         break;
3555
3556                 if ((p = pfind(a.pm_pid)) == NULL) {
3557                         error = ESRCH;
3558                         break;
3559                 }
3560
3561                 /*
3562                  * Treat processes that are in the process of exiting
3563                  * as if they were not present.
3564                  */
3565
3566                 if (p->p_flag & P_WEXIT)
3567                         error = ESRCH;
3568
3569                 PROC_UNLOCK(p); /* pfind() returns a locked process */
3570
3571                 if (error == 0)
3572                         error = pmc_detach_process(p, pm);
3573         }
3574         break;
3575
3576
3577         /*
3578          * Retrieve the MSR number associated with the counter
3579          * 'pmc_id'.  This allows processes to directly use RDPMC
3580          * instructions to read their PMCs, without the overhead of a
3581          * system call.
3582          */
3583
3584         case PMC_OP_PMCGETMSR:
3585         {
3586                 int adjri, ri;
3587                 struct pmc *pm;
3588                 struct pmc_target *pt;
3589                 struct pmc_op_getmsr gm;
3590                 struct pmc_classdep *pcd;
3591
3592                 PMC_DOWNGRADE_SX();
3593
3594                 if ((error = copyin(arg, &gm, sizeof(gm))) != 0)
3595                         break;
3596
3597                 if ((error = pmc_find_pmc(gm.pm_pmcid, &pm)) != 0)
3598                         break;
3599
3600                 /*
3601                  * The allocated PMC has to be a process virtual PMC,
3602                  * i.e., of type MODE_T[CS].  Global PMCs can only be
3603                  * read using the PMCREAD operation since they may be
3604                  * allocated on a different CPU than the one we could
3605                  * be running on at the time of the RDPMC instruction.
3606                  *
3607                  * The GETMSR operation is not allowed for PMCs that
3608                  * are inherited across processes.
3609                  */
3610
3611                 if (!PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)) ||
3612                     (pm->pm_flags & PMC_F_DESCENDANTS)) {
3613                         error = EINVAL;
3614                         break;
3615                 }
3616
3617                 /*
3618                  * It only makes sense to use a RDPMC (or its
3619                  * equivalent instruction on non-x86 architectures) on
3620                  * a process that has allocated and attached a PMC to
3621                  * itself.  Conversely the PMC is only allowed to have
3622                  * one process attached to it -- its owner.
3623                  */
3624
3625                 if ((pt = LIST_FIRST(&pm->pm_targets)) == NULL ||
3626                     LIST_NEXT(pt, pt_next) != NULL ||
3627                     pt->pt_process->pp_proc != pm->pm_owner->po_owner) {
3628                         error = EINVAL;
3629                         break;
3630                 }
3631
3632                 ri = PMC_TO_ROWINDEX(pm);
3633                 pcd = pmc_ri_to_classdep(md, ri, &adjri);
3634
3635                 /* PMC class has no 'GETMSR' support */
3636                 if (pcd->pcd_get_msr == NULL) {
3637                         error = ENOSYS;
3638                         break;
3639                 }
3640
3641                 if ((error = (*pcd->pcd_get_msr)(adjri, &gm.pm_msr)) < 0)
3642                         break;
3643
3644                 if ((error = copyout(&gm, arg, sizeof(gm))) < 0)
3645                         break;
3646
3647                 /*
3648                  * Mark our process as using MSRs.  Update machine
3649                  * state using a forced context switch.
3650                  */
3651
3652                 pt->pt_process->pp_flags |= PMC_PP_ENABLE_MSR_ACCESS;
3653                 pmc_force_context_switch();
3654
3655         }
3656         break;
3657
3658         /*
3659          * Release an allocated PMC
3660          */
3661
3662         case PMC_OP_PMCRELEASE:
3663         {
3664                 pmc_id_t pmcid;
3665                 struct pmc *pm;
3666                 struct pmc_owner *po;
3667                 struct pmc_op_simple sp;
3668
3669                 /*
3670                  * Find PMC pointer for the named PMC.
3671                  *
3672                  * Use pmc_release_pmc_descriptor() to switch off the
3673                  * PMC, remove all its target threads, and remove the
3674                  * PMC from its owner's list.
3675                  *
3676                  * Remove the owner record if this is the last PMC
3677                  * owned.
3678                  *
3679                  * Free up space.
3680                  */
3681
3682                 if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
3683                         break;
3684
3685                 pmcid = sp.pm_pmcid;
3686
3687                 if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
3688                         break;
3689
3690                 po = pm->pm_owner;
3691                 pmc_release_pmc_descriptor(pm);
3692                 pmc_maybe_remove_owner(po);
3693
3694                 free(pm, M_PMC);
3695         }
3696         break;
3697
3698
3699         /*
3700          * Read and/or write a PMC.
3701          */
3702
3703         case PMC_OP_PMCRW:
3704         {
3705                 int adjri;
3706                 struct pmc *pm;
3707                 uint32_t cpu, ri;
3708                 pmc_value_t oldvalue;
3709                 struct pmc_binding pb;
3710                 struct pmc_op_pmcrw prw;
3711                 struct pmc_classdep *pcd;
3712                 struct pmc_op_pmcrw *pprw;
3713
3714                 PMC_DOWNGRADE_SX();
3715
3716                 if ((error = copyin(arg, &prw, sizeof(prw))) != 0)
3717                         break;
3718
3719                 ri = 0;
3720                 PMCDBG(PMC,OPS,1, "rw id=%d flags=0x%x", prw.pm_pmcid,
3721                     prw.pm_flags);
3722
3723                 /* must have at least one flag set */
3724                 if ((prw.pm_flags & (PMC_F_OLDVALUE|PMC_F_NEWVALUE)) == 0) {
3725                         error = EINVAL;
3726                         break;
3727                 }
3728
3729                 /* locate pmc descriptor */
3730                 if ((error = pmc_find_pmc(prw.pm_pmcid, &pm)) != 0)
3731                         break;
3732
3733                 /* Can't read a PMC that hasn't been started. */
3734                 if (pm->pm_state != PMC_STATE_ALLOCATED &&
3735                     pm->pm_state != PMC_STATE_STOPPED &&
3736                     pm->pm_state != PMC_STATE_RUNNING) {
3737                         error = EINVAL;
3738                         break;
3739                 }
3740
3741                 /* writing a new value is allowed only for 'STOPPED' pmcs */
3742                 if (pm->pm_state == PMC_STATE_RUNNING &&
3743                     (prw.pm_flags & PMC_F_NEWVALUE)) {
3744                         error = EBUSY;
3745                         break;
3746                 }
3747
3748                 if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) {
3749
3750                         /*
3751                          * If this PMC is attached to its owner (i.e.,
3752                          * the process requesting this operation) and
3753                          * is running, then attempt to get an
3754                          * upto-date reading from hardware for a READ.
3755                          * Writes are only allowed when the PMC is
3756                          * stopped, so only update the saved value
3757                          * field.
3758                          *
3759                          * If the PMC is not running, or is not
3760                          * attached to its owner, read/write to the
3761                          * savedvalue field.
3762                          */
3763
3764                         ri = PMC_TO_ROWINDEX(pm);
3765                         pcd = pmc_ri_to_classdep(md, ri, &adjri);
3766
3767                         mtx_pool_lock_spin(pmc_mtxpool, pm);
3768                         cpu = curthread->td_oncpu;
3769
3770                         if (prw.pm_flags & PMC_F_OLDVALUE) {
3771                                 if ((pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) &&
3772                                     (pm->pm_state == PMC_STATE_RUNNING))
3773                                         error = (*pcd->pcd_read_pmc)(cpu, adjri,
3774                                             &oldvalue);
3775                                 else
3776                                         oldvalue = pm->pm_gv.pm_savedvalue;
3777                         }
3778                         if (prw.pm_flags & PMC_F_NEWVALUE)
3779                                 pm->pm_gv.pm_savedvalue = prw.pm_value;
3780
3781                         mtx_pool_unlock_spin(pmc_mtxpool, pm);
3782
3783                 } else { /* System mode PMCs */
3784                         cpu = PMC_TO_CPU(pm);
3785                         ri  = PMC_TO_ROWINDEX(pm);
3786                         pcd = pmc_ri_to_classdep(md, ri, &adjri);
3787
3788                         if (!pmc_cpu_is_active(cpu)) {
3789                                 error = ENXIO;
3790                                 break;
3791                         }
3792
3793                         /* move this thread to CPU 'cpu' */
3794                         pmc_save_cpu_binding(&pb);
3795                         pmc_select_cpu(cpu);
3796
3797                         critical_enter();
3798                         /* save old value */
3799                         if (prw.pm_flags & PMC_F_OLDVALUE)
3800                                 if ((error = (*pcd->pcd_read_pmc)(cpu, adjri,
3801                                          &oldvalue)))
3802                                         goto error;
3803                         /* write out new value */
3804                         if (prw.pm_flags & PMC_F_NEWVALUE)
3805                                 error = (*pcd->pcd_write_pmc)(cpu, adjri,
3806                                     prw.pm_value);
3807                 error:
3808                         critical_exit();
3809                         pmc_restore_cpu_binding(&pb);
3810                         if (error)
3811                                 break;
3812                 }
3813
3814                 pprw = (struct pmc_op_pmcrw *) arg;
3815
3816 #ifdef  DEBUG
3817                 if (prw.pm_flags & PMC_F_NEWVALUE)
3818                         PMCDBG(PMC,OPS,2, "rw id=%d new %jx -> old %jx",
3819                             ri, prw.pm_value, oldvalue);
3820                 else if (prw.pm_flags & PMC_F_OLDVALUE)
3821                         PMCDBG(PMC,OPS,2, "rw id=%d -> old %jx", ri, oldvalue);
3822 #endif
3823
3824                 /* return old value if requested */
3825                 if (prw.pm_flags & PMC_F_OLDVALUE)
3826                         if ((error = copyout(&oldvalue, &pprw->pm_value,
3827                                  sizeof(prw.pm_value))))
3828                                 break;
3829
3830         }
3831         break;
3832
3833
3834         /*
3835          * Set the sampling rate for a sampling mode PMC and the
3836          * initial count for a counting mode PMC.
3837          */
3838
3839         case PMC_OP_PMCSETCOUNT:
3840         {
3841                 struct pmc *pm;
3842                 struct pmc_op_pmcsetcount sc;
3843
3844                 PMC_DOWNGRADE_SX();
3845
3846                 if ((error = copyin(arg, &sc, sizeof(sc))) != 0)
3847                         break;
3848
3849                 if ((error = pmc_find_pmc(sc.pm_pmcid, &pm)) != 0)
3850                         break;
3851
3852                 if (pm->pm_state == PMC_STATE_RUNNING) {
3853                         error = EBUSY;
3854                         break;
3855                 }
3856
3857                 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
3858                         pm->pm_sc.pm_reloadcount = sc.pm_count;
3859                 else
3860                         pm->pm_sc.pm_initial = sc.pm_count;
3861         }
3862         break;
3863
3864
3865         /*
3866          * Start a PMC.
3867          */
3868
3869         case PMC_OP_PMCSTART:
3870         {
3871                 pmc_id_t pmcid;
3872                 struct pmc *pm;
3873                 struct pmc_op_simple sp;
3874
3875                 sx_assert(&pmc_sx, SX_XLOCKED);
3876
3877                 if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
3878                         break;
3879
3880                 pmcid = sp.pm_pmcid;
3881
3882                 if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
3883                         break;
3884
3885                 KASSERT(pmcid == pm->pm_id,
3886                     ("[pmc,%d] pmcid %x != id %x", __LINE__,
3887                         pm->pm_id, pmcid));
3888
3889                 if (pm->pm_state == PMC_STATE_RUNNING) /* already running */
3890                         break;
3891                 else if (pm->pm_state != PMC_STATE_STOPPED &&
3892                     pm->pm_state != PMC_STATE_ALLOCATED) {
3893                         error = EINVAL;
3894                         break;
3895                 }
3896
3897                 error = pmc_start(pm);
3898         }
3899         break;
3900
3901
3902         /*
3903          * Stop a PMC.
3904          */
3905
3906         case PMC_OP_PMCSTOP:
3907         {
3908                 pmc_id_t pmcid;
3909                 struct pmc *pm;
3910                 struct pmc_op_simple sp;
3911
3912                 PMC_DOWNGRADE_SX();
3913
3914                 if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
3915                         break;
3916
3917                 pmcid = sp.pm_pmcid;
3918
3919                 /*
3920                  * Mark the PMC as inactive and invoke the MD stop
3921                  * routines if needed.
3922                  */
3923
3924                 if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
3925                         break;
3926
3927                 KASSERT(pmcid == pm->pm_id,
3928                     ("[pmc,%d] pmc id %x != pmcid %x", __LINE__,
3929                         pm->pm_id, pmcid));
3930
3931                 if (pm->pm_state == PMC_STATE_STOPPED) /* already stopped */
3932                         break;
3933                 else if (pm->pm_state != PMC_STATE_RUNNING) {
3934                         error = EINVAL;
3935                         break;
3936                 }
3937
3938                 error = pmc_stop(pm);
3939         }
3940         break;
3941
3942
3943         /*
3944          * Write a user supplied value to the log file.
3945          */
3946
3947         case PMC_OP_WRITELOG:
3948         {
3949                 struct pmc_op_writelog wl;
3950                 struct pmc_owner *po;
3951
3952                 PMC_DOWNGRADE_SX();
3953
3954                 if ((error = copyin(arg, &wl, sizeof(wl))) != 0)
3955                         break;
3956
3957                 if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) {
3958                         error = EINVAL;
3959                         break;
3960                 }
3961
3962                 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) {
3963                         error = EINVAL;
3964                         break;
3965                 }
3966
3967                 error = pmclog_process_userlog(po, &wl);
3968         }
3969         break;
3970
3971
3972         default:
3973                 error = EINVAL;
3974                 break;
3975         }
3976
3977         if (is_sx_locked != 0) {
3978                 if (is_sx_downgraded)
3979                         sx_sunlock(&pmc_sx);
3980                 else
3981                         sx_xunlock(&pmc_sx);
3982         }
3983
3984         if (error)
3985                 atomic_add_int(&pmc_stats.pm_syscall_errors, 1);
3986
3987         PICKUP_GIANT();
3988
3989         return error;
3990 }
3991
3992 /*
3993  * Helper functions
3994  */
3995
3996
3997 /*
3998  * Mark the thread as needing callchain capture and post an AST.  The
3999  * actual callchain capture will be done in a context where it is safe
4000  * to take page faults.
4001  */
4002
4003 static void
4004 pmc_post_callchain_callback(void)
4005 {
4006         struct thread *td;
4007
4008         td = curthread;
4009
4010         /*
4011          * If there is multiple PMCs for the same interrupt ignore new post
4012          */
4013         if (td->td_pflags & TDP_CALLCHAIN)
4014                 return;
4015
4016         /*
4017          * Mark this thread as needing callchain capture.
4018          * `td->td_pflags' will be safe to touch because this thread
4019          * was in user space when it was interrupted.
4020          */
4021         td->td_pflags |= TDP_CALLCHAIN;
4022
4023         /*
4024          * Don't let this thread migrate between CPUs until callchain
4025          * capture completes.
4026          */
4027         sched_pin();
4028
4029         return;
4030 }
4031
4032 /*
4033  * Interrupt processing.
4034  *
4035  * Find a free slot in the per-cpu array of samples and capture the
4036  * current callchain there.  If a sample was successfully added, a bit
4037  * is set in mask 'pmc_cpumask' denoting that the DO_SAMPLES hook
4038  * needs to be invoked from the clock handler.
4039  *
4040  * This function is meant to be called from an NMI handler.  It cannot
4041  * use any of the locking primitives supplied by the OS.
4042  */
4043
4044 int
4045 pmc_process_interrupt(int cpu, int ring, struct pmc *pm, struct trapframe *tf,
4046     int inuserspace)
4047 {
4048         int error, callchaindepth;
4049         struct thread *td;
4050         struct pmc_sample *ps;
4051         struct pmc_samplebuffer *psb;
4052
4053         error = 0;
4054
4055         /*
4056          * Allocate space for a sample buffer.
4057          */
4058         psb = pmc_pcpu[cpu]->pc_sb[ring];
4059
4060         ps = psb->ps_write;
4061         if (ps->ps_nsamples) {  /* in use, reader hasn't caught up */
4062                 pm->pm_stalled = 1;
4063                 atomic_add_int(&pmc_stats.pm_intr_bufferfull, 1);
4064                 PMCDBG(SAM,INT,1,"(spc) cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d",
4065                     cpu, pm, (void *) tf, inuserspace,
4066                     (int) (psb->ps_write - psb->ps_samples),
4067                     (int) (psb->ps_read - psb->ps_samples));
4068                 error = ENOMEM;
4069                 goto done;
4070         }
4071
4072
4073         /* Fill in entry. */
4074         PMCDBG(SAM,INT,1,"cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d", cpu, pm,
4075             (void *) tf, inuserspace,
4076             (int) (psb->ps_write - psb->ps_samples),
4077             (int) (psb->ps_read - psb->ps_samples));
4078
4079         KASSERT(pm->pm_runcount >= 0,
4080             ("[pmc,%d] pm=%p runcount %d", __LINE__, (void *) pm,
4081                 pm->pm_runcount));
4082
4083         atomic_add_rel_int(&pm->pm_runcount, 1);        /* hold onto PMC */
4084
4085         ps->ps_pmc = pm;
4086         if ((td = curthread) && td->td_proc)
4087                 ps->ps_pid = td->td_proc->p_pid;
4088         else
4089                 ps->ps_pid = -1;
4090         ps->ps_cpu = cpu;
4091         ps->ps_td = td;
4092         ps->ps_flags = inuserspace ? PMC_CC_F_USERSPACE : 0;
4093
4094         callchaindepth = (pm->pm_flags & PMC_F_CALLCHAIN) ?
4095             pmc_callchaindepth : 1;
4096
4097         if (callchaindepth == 1)
4098                 ps->ps_pc[0] = PMC_TRAPFRAME_TO_PC(tf);
4099         else {
4100                 /*
4101                  * Kernel stack traversals can be done immediately,
4102                  * while we defer to an AST for user space traversals.
4103                  */
4104                 if (!inuserspace) {
4105                         callchaindepth =
4106                             pmc_save_kernel_callchain(ps->ps_pc,
4107                                 callchaindepth, tf);
4108                 } else {
4109                         pmc_post_callchain_callback();
4110                         callchaindepth = PMC_SAMPLE_INUSE;
4111                 }
4112         }
4113
4114         ps->ps_nsamples = callchaindepth;       /* mark entry as in use */
4115
4116         /* increment write pointer, modulo ring buffer size */
4117         ps++;
4118         if (ps == psb->ps_fence)
4119                 psb->ps_write = psb->ps_samples;
4120         else
4121                 psb->ps_write = ps;
4122
4123  done:
4124         /* mark CPU as needing processing */
4125         CPU_SET_ATOMIC(cpu, &pmc_cpumask);
4126
4127         return (error);
4128 }
4129
4130 /*
4131  * Capture a user call chain.  This function will be called from ast()
4132  * before control returns to userland and before the process gets
4133  * rescheduled.
4134  */
4135
4136 static void
4137 pmc_capture_user_callchain(int cpu, int ring, struct trapframe *tf)
4138 {
4139         int i;
4140         struct pmc *pm;
4141         struct thread *td;
4142         struct pmc_sample *ps;
4143         struct pmc_samplebuffer *psb;
4144 #ifdef  INVARIANTS
4145         int ncallchains;
4146 #endif
4147
4148         psb = pmc_pcpu[cpu]->pc_sb[ring];
4149         td = curthread;
4150
4151         KASSERT(td->td_pflags & TDP_CALLCHAIN,
4152             ("[pmc,%d] Retrieving callchain for thread that doesn't want it",
4153                 __LINE__));
4154
4155 #ifdef  INVARIANTS
4156         ncallchains = 0;
4157 #endif
4158
4159         /*
4160          * Iterate through all deferred callchain requests.
4161          */
4162
4163         ps = psb->ps_samples;
4164         for (i = 0; i < pmc_nsamples; i++, ps++) {
4165
4166                 if (ps->ps_nsamples != PMC_SAMPLE_INUSE)
4167                         continue;
4168                 if (ps->ps_td != td)
4169                         continue;
4170
4171                 KASSERT(ps->ps_cpu == cpu,
4172                     ("[pmc,%d] cpu mismatch ps_cpu=%d pcpu=%d", __LINE__,
4173                         ps->ps_cpu, PCPU_GET(cpuid)));
4174
4175                 pm = ps->ps_pmc;
4176
4177                 KASSERT(pm->pm_flags & PMC_F_CALLCHAIN,
4178                     ("[pmc,%d] Retrieving callchain for PMC that doesn't "
4179                         "want it", __LINE__));
4180
4181                 KASSERT(pm->pm_runcount > 0,
4182                     ("[pmc,%d] runcount %d", __LINE__, pm->pm_runcount));
4183
4184                 /*
4185                  * Retrieve the callchain and mark the sample buffer
4186                  * as 'processable' by the timer tick sweep code.
4187                  */
4188                 ps->ps_nsamples = pmc_save_user_callchain(ps->ps_pc,
4189                     pmc_callchaindepth, tf);
4190
4191 #ifdef  INVARIANTS
4192                 ncallchains++;
4193 #endif
4194         }
4195
4196         KASSERT(ncallchains > 0,
4197             ("[pmc,%d] cpu %d didn't find a sample to collect", __LINE__,
4198                 cpu));
4199
4200         KASSERT(td->td_pinned > 0,
4201             ("[pmc,%d] invalid td_pinned value", __LINE__));
4202         sched_unpin();  /* Can migrate safely now. */
4203
4204         return;
4205 }
4206
4207 /*
4208  * Process saved PC samples.
4209  */
4210
4211 static void
4212 pmc_process_samples(int cpu, int ring)
4213 {
4214         struct pmc *pm;
4215         int adjri, n;
4216         struct thread *td;
4217         struct pmc_owner *po;
4218         struct pmc_sample *ps;
4219         struct pmc_classdep *pcd;
4220         struct pmc_samplebuffer *psb;
4221
4222         KASSERT(PCPU_GET(cpuid) == cpu,
4223             ("[pmc,%d] not on the correct CPU pcpu=%d cpu=%d", __LINE__,
4224                 PCPU_GET(cpuid), cpu));
4225
4226         psb = pmc_pcpu[cpu]->pc_sb[ring];
4227
4228         for (n = 0; n < pmc_nsamples; n++) { /* bound on #iterations */
4229
4230                 ps = psb->ps_read;
4231                 if (ps->ps_nsamples == PMC_SAMPLE_FREE)
4232                         break;
4233
4234                 pm = ps->ps_pmc;
4235
4236                 KASSERT(pm->pm_runcount > 0,
4237                     ("[pmc,%d] pm=%p runcount %d", __LINE__, (void *) pm,
4238                         pm->pm_runcount));
4239
4240                 po = pm->pm_owner;
4241
4242                 KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)),
4243                     ("[pmc,%d] pmc=%p non-sampling mode=%d", __LINE__,
4244                         pm, PMC_TO_MODE(pm)));
4245
4246                 /* Ignore PMCs that have been switched off */
4247                 if (pm->pm_state != PMC_STATE_RUNNING)
4248                         goto entrydone;
4249
4250                 /* If there is a pending AST wait for completion */
4251                 if (ps->ps_nsamples == PMC_SAMPLE_INUSE) {
4252                         /* Need a rescan at a later time. */
4253                         CPU_SET_ATOMIC(cpu, &pmc_cpumask);
4254                         break;
4255                 }
4256
4257                 PMCDBG(SAM,OPS,1,"cpu=%d pm=%p n=%d fl=%x wr=%d rd=%d", cpu,
4258                     pm, ps->ps_nsamples, ps->ps_flags,
4259                     (int) (psb->ps_write - psb->ps_samples),
4260                     (int) (psb->ps_read - psb->ps_samples));
4261
4262                 /*
4263                  * If this is a process-mode PMC that is attached to
4264                  * its owner, and if the PC is in user mode, update
4265                  * profiling statistics like timer-based profiling
4266                  * would have done.
4267                  */
4268                 if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) {
4269                         if (ps->ps_flags & PMC_CC_F_USERSPACE) {
4270                                 td = FIRST_THREAD_IN_PROC(po->po_owner);
4271                                 addupc_intr(td, ps->ps_pc[0], 1);
4272                         }
4273                         goto entrydone;
4274                 }
4275
4276                 /*
4277                  * Otherwise, this is either a sampling mode PMC that
4278                  * is attached to a different process than its owner,
4279                  * or a system-wide sampling PMC.  Dispatch a log
4280                  * entry to the PMC's owner process.
4281                  */
4282                 pmclog_process_callchain(pm, ps);
4283
4284         entrydone:
4285                 ps->ps_nsamples = 0; /* mark entry as free */
4286                 atomic_subtract_rel_int(&pm->pm_runcount, 1);
4287
4288                 /* increment read pointer, modulo sample size */
4289                 if (++ps == psb->ps_fence)
4290                         psb->ps_read = psb->ps_samples;
4291                 else
4292                         psb->ps_read = ps;
4293         }
4294
4295         atomic_add_int(&pmc_stats.pm_log_sweeps, 1);
4296
4297         /* Do not re-enable stalled PMCs if we failed to process any samples */
4298         if (n == 0)
4299                 return;
4300
4301         /*
4302          * Restart any stalled sampling PMCs on this CPU.
4303          *
4304          * If the NMI handler sets the pm_stalled field of a PMC after
4305          * the check below, we'll end up processing the stalled PMC at
4306          * the next hardclock tick.
4307          */
4308         for (n = 0; n < md->pmd_npmc; n++) {
4309                 pcd = pmc_ri_to_classdep(md, n, &adjri);
4310                 KASSERT(pcd != NULL,
4311                     ("[pmc,%d] null pcd ri=%d", __LINE__, n));
4312                 (void) (*pcd->pcd_get_config)(cpu,adjri,&pm);
4313
4314                 if (pm == NULL ||                        /* !cfg'ed */
4315                     pm->pm_state != PMC_STATE_RUNNING || /* !active */
4316                     !PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) || /* !sampling */
4317                     pm->pm_stalled == 0) /* !stalled */
4318                         continue;
4319
4320                 pm->pm_stalled = 0;
4321                 (*pcd->pcd_start_pmc)(cpu, adjri);
4322         }
4323 }
4324
4325 /*
4326  * Event handlers.
4327  */
4328
4329 /*
4330  * Handle a process exit.
4331  *
4332  * Remove this process from all hash tables.  If this process
4333  * owned any PMCs, turn off those PMCs and deallocate them,
4334  * removing any associations with target processes.
4335  *
4336  * This function will be called by the last 'thread' of a
4337  * process.
4338  *
4339  * XXX This eventhandler gets called early in the exit process.
4340  * Consider using a 'hook' invocation from thread_exit() or equivalent
4341  * spot.  Another negative is that kse_exit doesn't seem to call
4342  * exit1() [??].
4343  *
4344  */
4345
4346 static void
4347 pmc_process_exit(void *arg __unused, struct proc *p)
4348 {
4349         struct pmc *pm;
4350         int adjri, cpu;
4351         unsigned int ri;
4352         int is_using_hwpmcs;
4353         struct pmc_owner *po;
4354         struct pmc_process *pp;
4355         struct pmc_classdep *pcd;
4356         pmc_value_t newvalue, tmp;
4357
4358         PROC_LOCK(p);
4359         is_using_hwpmcs = p->p_flag & P_HWPMC;
4360         PROC_UNLOCK(p);
4361
4362         /*
4363          * Log a sysexit event to all SS PMC owners.
4364          */
4365         LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
4366             if (po->po_flags & PMC_PO_OWNS_LOGFILE)
4367                     pmclog_process_sysexit(po, p->p_pid);
4368
4369         if (!is_using_hwpmcs)
4370                 return;
4371
4372         PMC_GET_SX_XLOCK();
4373         PMCDBG(PRC,EXT,1,"process-exit proc=%p (%d, %s)", p, p->p_pid,
4374             p->p_comm);
4375
4376         /*
4377          * Since this code is invoked by the last thread in an exiting
4378          * process, we would have context switched IN at some prior
4379          * point.  However, with PREEMPTION, kernel mode context
4380          * switches may happen any time, so we want to disable a
4381          * context switch OUT till we get any PMCs targetting this
4382          * process off the hardware.
4383          *
4384          * We also need to atomically remove this process'
4385          * entry from our target process hash table, using
4386          * PMC_FLAG_REMOVE.
4387          */
4388         PMCDBG(PRC,EXT,1, "process-exit proc=%p (%d, %s)", p, p->p_pid,
4389             p->p_comm);
4390
4391         critical_enter(); /* no preemption */
4392
4393         cpu = curthread->td_oncpu;
4394
4395         if ((pp = pmc_find_process_descriptor(p,
4396                  PMC_FLAG_REMOVE)) != NULL) {
4397
4398                 PMCDBG(PRC,EXT,2,
4399                     "process-exit proc=%p pmc-process=%p", p, pp);
4400
4401                 /*
4402                  * The exiting process could the target of
4403                  * some PMCs which will be running on
4404                  * currently executing CPU.
4405                  *
4406                  * We need to turn these PMCs off like we
4407                  * would do at context switch OUT time.
4408                  */
4409                 for (ri = 0; ri < md->pmd_npmc; ri++) {
4410
4411                         /*
4412                          * Pick up the pmc pointer from hardware
4413                          * state similar to the CSW_OUT code.
4414                          */
4415                         pm = NULL;
4416
4417                         pcd = pmc_ri_to_classdep(md, ri, &adjri);
4418
4419                         (void) (*pcd->pcd_get_config)(cpu, adjri, &pm);
4420
4421                         PMCDBG(PRC,EXT,2, "ri=%d pm=%p", ri, pm);
4422
4423                         if (pm == NULL ||
4424                             !PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)))
4425                                 continue;
4426
4427                         PMCDBG(PRC,EXT,2, "ppmcs[%d]=%p pm=%p "
4428                             "state=%d", ri, pp->pp_pmcs[ri].pp_pmc,
4429                             pm, pm->pm_state);
4430
4431                         KASSERT(PMC_TO_ROWINDEX(pm) == ri,
4432                             ("[pmc,%d] ri mismatch pmc(%d) ri(%d)",
4433                                 __LINE__, PMC_TO_ROWINDEX(pm), ri));
4434
4435                         KASSERT(pm == pp->pp_pmcs[ri].pp_pmc,
4436                             ("[pmc,%d] pm %p != pp_pmcs[%d] %p",
4437                                 __LINE__, pm, ri, pp->pp_pmcs[ri].pp_pmc));
4438
4439                         (void) pcd->pcd_stop_pmc(cpu, adjri);
4440
4441                         KASSERT(pm->pm_runcount > 0,
4442                             ("[pmc,%d] bad runcount ri %d rc %d",
4443                                 __LINE__, ri, pm->pm_runcount));
4444
4445                         /* Stop hardware only if it is actually running */
4446                         if (pm->pm_state == PMC_STATE_RUNNING &&
4447                             pm->pm_stalled == 0) {
4448                                 pcd->pcd_read_pmc(cpu, adjri, &newvalue);
4449                                 tmp = newvalue -
4450                                     PMC_PCPU_SAVED(cpu,ri);
4451
4452                                 mtx_pool_lock_spin(pmc_mtxpool, pm);
4453                                 pm->pm_gv.pm_savedvalue += tmp;
4454                                 pp->pp_pmcs[ri].pp_pmcval += tmp;
4455                                 mtx_pool_unlock_spin(pmc_mtxpool, pm);
4456                         }
4457
4458                         atomic_subtract_rel_int(&pm->pm_runcount,1);
4459
4460                         KASSERT((int) pm->pm_runcount >= 0,
4461                             ("[pmc,%d] runcount is %d", __LINE__, ri));
4462
4463                         (void) pcd->pcd_config_pmc(cpu, adjri, NULL);
4464                 }
4465
4466                 /*
4467                  * Inform the MD layer of this pseudo "context switch
4468                  * out"
4469                  */
4470                 (void) md->pmd_switch_out(pmc_pcpu[cpu], pp);
4471
4472                 critical_exit(); /* ok to be pre-empted now */
4473
4474                 /*
4475                  * Unlink this process from the PMCs that are
4476                  * targetting it.  This will send a signal to
4477                  * all PMC owner's whose PMCs are orphaned.
4478                  *
4479                  * Log PMC value at exit time if requested.
4480                  */
4481                 for (ri = 0; ri < md->pmd_npmc; ri++)
4482                         if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) {
4483                                 if (pm->pm_flags & PMC_F_NEEDS_LOGFILE &&
4484                                     PMC_IS_COUNTING_MODE(PMC_TO_MODE(pm)))
4485                                         pmclog_process_procexit(pm, pp);
4486                                 pmc_unlink_target_process(pm, pp);
4487                         }
4488                 free(pp, M_PMC);
4489
4490         } else
4491                 critical_exit(); /* pp == NULL */
4492
4493
4494         /*
4495          * If the process owned PMCs, free them up and free up
4496          * memory.
4497          */
4498         if ((po = pmc_find_owner_descriptor(p)) != NULL) {
4499                 pmc_remove_owner(po);
4500                 pmc_destroy_owner_descriptor(po);
4501         }
4502
4503         sx_xunlock(&pmc_sx);
4504 }
4505
4506 /*
4507  * Handle a process fork.
4508  *
4509  * If the parent process 'p1' is under HWPMC monitoring, then copy
4510  * over any attached PMCs that have 'do_descendants' semantics.
4511  */
4512
4513 static void
4514 pmc_process_fork(void *arg __unused, struct proc *p1, struct proc *newproc,
4515     int flags)
4516 {
4517         int is_using_hwpmcs;
4518         unsigned int ri;
4519         uint32_t do_descendants;
4520         struct pmc *pm;
4521         struct pmc_owner *po;
4522         struct pmc_process *ppnew, *ppold;
4523
4524         (void) flags;           /* unused parameter */
4525
4526         PROC_LOCK(p1);
4527         is_using_hwpmcs = p1->p_flag & P_HWPMC;
4528         PROC_UNLOCK(p1);
4529
4530         /*
4531          * If there are system-wide sampling PMCs active, we need to
4532          * log all fork events to their owner's logs.
4533          */
4534
4535         LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
4536             if (po->po_flags & PMC_PO_OWNS_LOGFILE)
4537                     pmclog_process_procfork(po, p1->p_pid, newproc->p_pid);
4538
4539         if (!is_using_hwpmcs)
4540                 return;
4541
4542         PMC_GET_SX_XLOCK();
4543         PMCDBG(PMC,FRK,1, "process-fork proc=%p (%d, %s) -> %p", p1,
4544             p1->p_pid, p1->p_comm, newproc);
4545
4546         /*
4547          * If the parent process (curthread->td_proc) is a
4548          * target of any PMCs, look for PMCs that are to be
4549          * inherited, and link these into the new process
4550          * descriptor.
4551          */
4552         if ((ppold = pmc_find_process_descriptor(curthread->td_proc,
4553                  PMC_FLAG_NONE)) == NULL)
4554                 goto done;              /* nothing to do */
4555
4556         do_descendants = 0;
4557         for (ri = 0; ri < md->pmd_npmc; ri++)
4558                 if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL)
4559                         do_descendants |= pm->pm_flags & PMC_F_DESCENDANTS;
4560         if (do_descendants == 0) /* nothing to do */
4561                 goto done;
4562
4563         /* allocate a descriptor for the new process  */
4564         if ((ppnew = pmc_find_process_descriptor(newproc,
4565                  PMC_FLAG_ALLOCATE)) == NULL)
4566                 goto done;
4567
4568         /*
4569          * Run through all PMCs that were targeting the old process
4570          * and which specified F_DESCENDANTS and attach them to the
4571          * new process.
4572          *
4573          * Log the fork event to all owners of PMCs attached to this
4574          * process, if not already logged.
4575          */
4576         for (ri = 0; ri < md->pmd_npmc; ri++)
4577                 if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL &&
4578                     (pm->pm_flags & PMC_F_DESCENDANTS)) {
4579                         pmc_link_target_process(pm, ppnew);
4580                         po = pm->pm_owner;
4581                         if (po->po_sscount == 0 &&
4582                             po->po_flags & PMC_PO_OWNS_LOGFILE)
4583                                 pmclog_process_procfork(po, p1->p_pid,
4584                                     newproc->p_pid);
4585                 }
4586
4587         /*
4588          * Now mark the new process as being tracked by this driver.
4589          */
4590         PROC_LOCK(newproc);
4591         newproc->p_flag |= P_HWPMC;
4592         PROC_UNLOCK(newproc);
4593
4594  done:
4595         sx_xunlock(&pmc_sx);
4596 }
4597
4598 static void
4599 pmc_kld_load(void *arg __unused, linker_file_t lf)
4600 {
4601         struct pmc_owner *po;
4602
4603         sx_slock(&pmc_sx);
4604
4605         /*
4606          * Notify owners of system sampling PMCs about KLD operations.
4607          */
4608         LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
4609                 if (po->po_flags & PMC_PO_OWNS_LOGFILE)
4610                         pmclog_process_map_in(po, (pid_t) -1,
4611                             (uintfptr_t) lf->address, lf->filename);
4612
4613         /*
4614          * TODO: Notify owners of (all) process-sampling PMCs too.
4615          */
4616
4617         sx_sunlock(&pmc_sx);
4618 }
4619
4620 static void
4621 pmc_kld_unload(void *arg __unused, const char *filename __unused,
4622     caddr_t address, size_t size)
4623 {
4624         struct pmc_owner *po;
4625
4626         sx_slock(&pmc_sx);
4627
4628         LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
4629                 if (po->po_flags & PMC_PO_OWNS_LOGFILE)
4630                         pmclog_process_map_out(po, (pid_t) -1,
4631                             (uintfptr_t) address, (uintfptr_t) address + size);
4632
4633         /*
4634          * TODO: Notify owners of process-sampling PMCs.
4635          */
4636
4637         sx_sunlock(&pmc_sx);
4638 }
4639
4640 /*
4641  * initialization
4642  */
4643
4644 static const char *pmc_name_of_pmcclass[] = {
4645 #undef  __PMC_CLASS
4646 #define __PMC_CLASS(N) #N ,
4647         __PMC_CLASSES()
4648 };
4649
4650 /*
4651  * Base class initializer: allocate structure and set default classes.
4652  */
4653 struct pmc_mdep *
4654 pmc_mdep_alloc(int nclasses)
4655 {
4656         struct pmc_mdep *md;
4657         int     n;
4658
4659         /* SOFT + md classes */
4660         n = 1 + nclasses;
4661         md = malloc(sizeof(struct pmc_mdep) + n *
4662             sizeof(struct pmc_classdep), M_PMC, M_WAITOK|M_ZERO);
4663         if (md != NULL) {
4664                 md->pmd_nclass = n;
4665
4666                 /* Add base class. */
4667                 pmc_soft_initialize(md);
4668         }
4669
4670         return md;
4671 }
4672
4673 void
4674 pmc_mdep_free(struct pmc_mdep *md)
4675 {
4676         pmc_soft_finalize(md);
4677         free(md, M_PMC);
4678 }
4679
4680 static int
4681 generic_switch_in(struct pmc_cpu *pc, struct pmc_process *pp)
4682 {
4683         (void) pc; (void) pp;
4684
4685         return (0);
4686 }
4687
4688 static int
4689 generic_switch_out(struct pmc_cpu *pc, struct pmc_process *pp)
4690 {
4691         (void) pc; (void) pp;
4692
4693         return (0);
4694 }
4695
4696 static struct pmc_mdep *
4697 pmc_generic_cpu_initialize(void)
4698 {
4699         struct pmc_mdep *md;
4700
4701         md = pmc_mdep_alloc(0);
4702
4703         md->pmd_cputype    = PMC_CPU_GENERIC;
4704
4705         md->pmd_pcpu_init  = NULL;
4706         md->pmd_pcpu_fini  = NULL;
4707         md->pmd_switch_in  = generic_switch_in;
4708         md->pmd_switch_out = generic_switch_out;
4709
4710         return (md);
4711 }
4712
4713 static void
4714 pmc_generic_cpu_finalize(struct pmc_mdep *md)
4715 {
4716         (void) md;
4717 }
4718
4719
4720 static int
4721 pmc_initialize(void)
4722 {
4723         int c, cpu, error, n, ri;
4724         unsigned int maxcpu;
4725         struct pmc_binding pb;
4726         struct pmc_sample *ps;
4727         struct pmc_classdep *pcd;
4728         struct pmc_samplebuffer *sb;
4729
4730         md = NULL;
4731         error = 0;
4732
4733 #ifdef  DEBUG
4734         /* parse debug flags first */
4735         if (TUNABLE_STR_FETCH(PMC_SYSCTL_NAME_PREFIX "debugflags",
4736                 pmc_debugstr, sizeof(pmc_debugstr)))
4737                 pmc_debugflags_parse(pmc_debugstr,
4738                     pmc_debugstr+strlen(pmc_debugstr));
4739 #endif
4740
4741         PMCDBG(MOD,INI,0, "PMC Initialize (version %x)", PMC_VERSION);
4742
4743         /* check kernel version */
4744         if (pmc_kernel_version != PMC_VERSION) {
4745                 if (pmc_kernel_version == 0)
4746                         printf("hwpmc: this kernel has not been compiled with "
4747                             "'options HWPMC_HOOKS'.\n");
4748                 else
4749                         printf("hwpmc: kernel version (0x%x) does not match "
4750                             "module version (0x%x).\n", pmc_kernel_version,
4751                             PMC_VERSION);
4752                 return EPROGMISMATCH;
4753         }
4754
4755         /*
4756          * check sysctl parameters
4757          */
4758
4759         if (pmc_hashsize <= 0) {
4760                 (void) printf("hwpmc: tunable \"hashsize\"=%d must be "
4761                     "greater than zero.\n", pmc_hashsize);
4762                 pmc_hashsize = PMC_HASH_SIZE;
4763         }
4764
4765         if (pmc_nsamples <= 0 || pmc_nsamples > 65535) {
4766                 (void) printf("hwpmc: tunable \"nsamples\"=%d out of "
4767                     "range.\n", pmc_nsamples);
4768                 pmc_nsamples = PMC_NSAMPLES;
4769         }
4770
4771         if (pmc_callchaindepth <= 0 ||
4772             pmc_callchaindepth > PMC_CALLCHAIN_DEPTH_MAX) {
4773                 (void) printf("hwpmc: tunable \"callchaindepth\"=%d out of "
4774                     "range.\n", pmc_callchaindepth);
4775                 pmc_callchaindepth = PMC_CALLCHAIN_DEPTH;
4776         }
4777
4778         md = pmc_md_initialize();
4779         if (md == NULL) {
4780                 /* Default to generic CPU. */
4781                 md = pmc_generic_cpu_initialize();
4782                 if (md == NULL)
4783                         return (ENOSYS);
4784         }
4785
4786         KASSERT(md->pmd_nclass >= 1 && md->pmd_npmc >= 1,
4787             ("[pmc,%d] no classes or pmcs", __LINE__));
4788
4789         /* Compute the map from row-indices to classdep pointers. */
4790         pmc_rowindex_to_classdep = malloc(sizeof(struct pmc_classdep *) *
4791             md->pmd_npmc, M_PMC, M_WAITOK|M_ZERO);
4792
4793         for (n = 0; n < md->pmd_npmc; n++)
4794                 pmc_rowindex_to_classdep[n] = NULL;
4795         for (ri = c = 0; c < md->pmd_nclass; c++) {
4796                 pcd = &md->pmd_classdep[c];
4797                 for (n = 0; n < pcd->pcd_num; n++, ri++)
4798                         pmc_rowindex_to_classdep[ri] = pcd;
4799         }
4800
4801         KASSERT(ri == md->pmd_npmc,
4802             ("[pmc,%d] npmc miscomputed: ri=%d, md->npmc=%d", __LINE__,
4803             ri, md->pmd_npmc));
4804
4805         maxcpu = pmc_cpu_max();
4806
4807         /* allocate space for the per-cpu array */
4808         pmc_pcpu = malloc(maxcpu * sizeof(struct pmc_cpu *), M_PMC,
4809             M_WAITOK|M_ZERO);
4810
4811         /* per-cpu 'saved values' for managing process-mode PMCs */
4812         pmc_pcpu_saved = malloc(sizeof(pmc_value_t) * maxcpu * md->pmd_npmc,
4813             M_PMC, M_WAITOK);
4814
4815         /* Perform CPU-dependent initialization. */
4816         pmc_save_cpu_binding(&pb);
4817         error = 0;
4818         for (cpu = 0; error == 0 && cpu < maxcpu; cpu++) {
4819                 if (!pmc_cpu_is_active(cpu))
4820                         continue;
4821                 pmc_select_cpu(cpu);
4822                 pmc_pcpu[cpu] = malloc(sizeof(struct pmc_cpu) +
4823                     md->pmd_npmc * sizeof(struct pmc_hw *), M_PMC,
4824                     M_WAITOK|M_ZERO);
4825                 if (md->pmd_pcpu_init)
4826                         error = md->pmd_pcpu_init(md, cpu);
4827                 for (n = 0; error == 0 && n < md->pmd_nclass; n++)
4828                         error = md->pmd_classdep[n].pcd_pcpu_init(md, cpu);
4829         }
4830         pmc_restore_cpu_binding(&pb);
4831
4832         if (error)
4833                 return (error);
4834
4835         /* allocate space for the sample array */
4836         for (cpu = 0; cpu < maxcpu; cpu++) {
4837                 if (!pmc_cpu_is_active(cpu))
4838                         continue;
4839
4840                 sb = malloc(sizeof(struct pmc_samplebuffer) +
4841                     pmc_nsamples * sizeof(struct pmc_sample), M_PMC,
4842                     M_WAITOK|M_ZERO);
4843                 sb->ps_read = sb->ps_write = sb->ps_samples;
4844                 sb->ps_fence = sb->ps_samples + pmc_nsamples;
4845
4846                 KASSERT(pmc_pcpu[cpu] != NULL,
4847                     ("[pmc,%d] cpu=%d Null per-cpu data", __LINE__, cpu));
4848
4849                 sb->ps_callchains = malloc(pmc_callchaindepth * pmc_nsamples *
4850                     sizeof(uintptr_t), M_PMC, M_WAITOK|M_ZERO);
4851
4852                 for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++)
4853                         ps->ps_pc = sb->ps_callchains +
4854                             (n * pmc_callchaindepth);
4855
4856                 pmc_pcpu[cpu]->pc_sb[PMC_HR] = sb;
4857
4858                 sb = malloc(sizeof(struct pmc_samplebuffer) +
4859                     pmc_nsamples * sizeof(struct pmc_sample), M_PMC,
4860                     M_WAITOK|M_ZERO);
4861                 sb->ps_read = sb->ps_write = sb->ps_samples;
4862                 sb->ps_fence = sb->ps_samples + pmc_nsamples;
4863
4864                 KASSERT(pmc_pcpu[cpu] != NULL,
4865                     ("[pmc,%d] cpu=%d Null per-cpu data", __LINE__, cpu));
4866
4867                 sb->ps_callchains = malloc(pmc_callchaindepth * pmc_nsamples *
4868                     sizeof(uintptr_t), M_PMC, M_WAITOK|M_ZERO);
4869
4870                 for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++)
4871                         ps->ps_pc = sb->ps_callchains +
4872                             (n * pmc_callchaindepth);
4873
4874                 pmc_pcpu[cpu]->pc_sb[PMC_SR] = sb;
4875         }
4876
4877         /* allocate space for the row disposition array */
4878         pmc_pmcdisp = malloc(sizeof(enum pmc_mode) * md->pmd_npmc,
4879             M_PMC, M_WAITOK|M_ZERO);
4880
4881         KASSERT(pmc_pmcdisp != NULL,
4882             ("[pmc,%d] pmcdisp allocation returned NULL", __LINE__));
4883
4884         /* mark all PMCs as available */
4885         for (n = 0; n < (int) md->pmd_npmc; n++)
4886                 PMC_MARK_ROW_FREE(n);
4887
4888         /* allocate thread hash tables */
4889         pmc_ownerhash = hashinit(pmc_hashsize, M_PMC,
4890             &pmc_ownerhashmask);
4891
4892         pmc_processhash = hashinit(pmc_hashsize, M_PMC,
4893             &pmc_processhashmask);
4894         mtx_init(&pmc_processhash_mtx, "pmc-process-hash", "pmc-leaf",
4895             MTX_SPIN);
4896
4897         LIST_INIT(&pmc_ss_owners);
4898         pmc_ss_count = 0;
4899
4900         /* allocate a pool of spin mutexes */
4901         pmc_mtxpool = mtx_pool_create("pmc-leaf", pmc_mtxpool_size,
4902             MTX_SPIN);
4903
4904         PMCDBG(MOD,INI,1, "pmc_ownerhash=%p, mask=0x%lx "
4905             "targethash=%p mask=0x%lx", pmc_ownerhash, pmc_ownerhashmask,
4906             pmc_processhash, pmc_processhashmask);
4907
4908         /* register process {exit,fork,exec} handlers */
4909         pmc_exit_tag = EVENTHANDLER_REGISTER(process_exit,
4910             pmc_process_exit, NULL, EVENTHANDLER_PRI_ANY);
4911         pmc_fork_tag = EVENTHANDLER_REGISTER(process_fork,
4912             pmc_process_fork, NULL, EVENTHANDLER_PRI_ANY);
4913
4914         /* register kld event handlers */
4915         pmc_kld_load_tag = EVENTHANDLER_REGISTER(kld_load, pmc_kld_load,
4916             NULL, EVENTHANDLER_PRI_ANY);
4917         pmc_kld_unload_tag = EVENTHANDLER_REGISTER(kld_unload, pmc_kld_unload,
4918             NULL, EVENTHANDLER_PRI_ANY);
4919
4920         /* initialize logging */
4921         pmclog_initialize();
4922
4923         /* set hook functions */
4924         pmc_intr = md->pmd_intr;
4925         pmc_hook = pmc_hook_handler;
4926
4927         if (error == 0) {
4928                 printf(PMC_MODULE_NAME ":");
4929                 for (n = 0; n < (int) md->pmd_nclass; n++) {
4930                         pcd = &md->pmd_classdep[n];
4931                         printf(" %s/%d/%d/0x%b",
4932                             pmc_name_of_pmcclass[pcd->pcd_class],
4933                             pcd->pcd_num,
4934                             pcd->pcd_width,
4935                             pcd->pcd_caps,
4936                             "\20"
4937                             "\1INT\2USR\3SYS\4EDG\5THR"
4938                             "\6REA\7WRI\10INV\11QUA\12PRC"
4939                             "\13TAG\14CSC");
4940                 }
4941                 printf("\n");
4942         }
4943
4944         return (error);
4945 }
4946
4947 /* prepare to be unloaded */
4948 static void
4949 pmc_cleanup(void)
4950 {
4951         int c, cpu;
4952         unsigned int maxcpu;
4953         struct pmc_ownerhash *ph;
4954         struct pmc_owner *po, *tmp;
4955         struct pmc_binding pb;
4956 #ifdef  DEBUG
4957         struct pmc_processhash *prh;
4958 #endif
4959
4960         PMCDBG(MOD,INI,0, "%s", "cleanup");
4961
4962         /* switch off sampling */
4963         CPU_ZERO(&pmc_cpumask);
4964         pmc_intr = NULL;
4965
4966         sx_xlock(&pmc_sx);
4967         if (pmc_hook == NULL) { /* being unloaded already */
4968                 sx_xunlock(&pmc_sx);
4969                 return;
4970         }
4971
4972         pmc_hook = NULL; /* prevent new threads from entering module */
4973
4974         /* deregister event handlers */
4975         EVENTHANDLER_DEREGISTER(process_fork, pmc_fork_tag);
4976         EVENTHANDLER_DEREGISTER(process_exit, pmc_exit_tag);
4977         EVENTHANDLER_DEREGISTER(kld_load, pmc_kld_load_tag);
4978         EVENTHANDLER_DEREGISTER(kld_unload, pmc_kld_unload_tag);
4979
4980         /* send SIGBUS to all owner threads, free up allocations */
4981         if (pmc_ownerhash)
4982                 for (ph = pmc_ownerhash;
4983                      ph <= &pmc_ownerhash[pmc_ownerhashmask];
4984                      ph++) {
4985                         LIST_FOREACH_SAFE(po, ph, po_next, tmp) {
4986                                 pmc_remove_owner(po);
4987
4988                                 /* send SIGBUS to owner processes */
4989                                 PMCDBG(MOD,INI,2, "cleanup signal proc=%p "
4990                                     "(%d, %s)", po->po_owner,
4991                                     po->po_owner->p_pid,
4992                                     po->po_owner->p_comm);
4993
4994                                 PROC_LOCK(po->po_owner);
4995                                 kern_psignal(po->po_owner, SIGBUS);
4996                                 PROC_UNLOCK(po->po_owner);
4997
4998                                 pmc_destroy_owner_descriptor(po);
4999                         }
5000                 }
5001
5002         /* reclaim allocated data structures */
5003         if (pmc_mtxpool)
5004                 mtx_pool_destroy(&pmc_mtxpool);
5005
5006         mtx_destroy(&pmc_processhash_mtx);
5007         if (pmc_processhash) {
5008 #ifdef  DEBUG
5009                 struct pmc_process *pp;
5010
5011                 PMCDBG(MOD,INI,3, "%s", "destroy process hash");
5012                 for (prh = pmc_processhash;
5013                      prh <= &pmc_processhash[pmc_processhashmask];
5014                      prh++)
5015                         LIST_FOREACH(pp, prh, pp_next)
5016                             PMCDBG(MOD,INI,3, "pid=%d", pp->pp_proc->p_pid);
5017 #endif
5018
5019                 hashdestroy(pmc_processhash, M_PMC, pmc_processhashmask);
5020                 pmc_processhash = NULL;
5021         }
5022
5023         if (pmc_ownerhash) {
5024                 PMCDBG(MOD,INI,3, "%s", "destroy owner hash");
5025                 hashdestroy(pmc_ownerhash, M_PMC, pmc_ownerhashmask);
5026                 pmc_ownerhash = NULL;
5027         }
5028
5029         KASSERT(LIST_EMPTY(&pmc_ss_owners),
5030             ("[pmc,%d] Global SS owner list not empty", __LINE__));
5031         KASSERT(pmc_ss_count == 0,
5032             ("[pmc,%d] Global SS count not empty", __LINE__));
5033
5034         /* do processor and pmc-class dependent cleanup */
5035         maxcpu = pmc_cpu_max();
5036
5037         PMCDBG(MOD,INI,3, "%s", "md cleanup");
5038         if (md) {
5039                 pmc_save_cpu_binding(&pb);
5040                 for (cpu = 0; cpu < maxcpu; cpu++) {
5041                         PMCDBG(MOD,INI,1,"pmc-cleanup cpu=%d pcs=%p",
5042                             cpu, pmc_pcpu[cpu]);
5043                         if (!pmc_cpu_is_active(cpu) || pmc_pcpu[cpu] == NULL)
5044                                 continue;
5045                         pmc_select_cpu(cpu);
5046                         for (c = 0; c < md->pmd_nclass; c++)
5047                                 md->pmd_classdep[c].pcd_pcpu_fini(md, cpu);
5048                         if (md->pmd_pcpu_fini)
5049                                 md->pmd_pcpu_fini(md, cpu);
5050                 }
5051
5052                 if (md->pmd_cputype == PMC_CPU_GENERIC)
5053                         pmc_generic_cpu_finalize(md);
5054                 else
5055                         pmc_md_finalize(md);
5056
5057                 pmc_mdep_free(md);
5058                 md = NULL;
5059                 pmc_restore_cpu_binding(&pb);
5060         }
5061
5062         /* Free per-cpu descriptors. */
5063         for (cpu = 0; cpu < maxcpu; cpu++) {
5064                 if (!pmc_cpu_is_active(cpu))
5065                         continue;
5066                 KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_HR] != NULL,
5067                     ("[pmc,%d] Null hw cpu sample buffer cpu=%d", __LINE__,
5068                         cpu));
5069                 KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_SR] != NULL,
5070                     ("[pmc,%d] Null sw cpu sample buffer cpu=%d", __LINE__,
5071                         cpu));
5072                 free(pmc_pcpu[cpu]->pc_sb[PMC_HR]->ps_callchains, M_PMC);
5073                 free(pmc_pcpu[cpu]->pc_sb[PMC_HR], M_PMC);
5074                 free(pmc_pcpu[cpu]->pc_sb[PMC_SR]->ps_callchains, M_PMC);
5075                 free(pmc_pcpu[cpu]->pc_sb[PMC_SR], M_PMC);
5076                 free(pmc_pcpu[cpu], M_PMC);
5077         }
5078
5079         free(pmc_pcpu, M_PMC);
5080         pmc_pcpu = NULL;
5081
5082         free(pmc_pcpu_saved, M_PMC);
5083         pmc_pcpu_saved = NULL;
5084
5085         if (pmc_pmcdisp) {
5086                 free(pmc_pmcdisp, M_PMC);
5087                 pmc_pmcdisp = NULL;
5088         }
5089
5090         if (pmc_rowindex_to_classdep) {
5091                 free(pmc_rowindex_to_classdep, M_PMC);
5092                 pmc_rowindex_to_classdep = NULL;
5093         }
5094
5095         pmclog_shutdown();
5096
5097         sx_xunlock(&pmc_sx);    /* we are done */
5098 }
5099
5100 /*
5101  * The function called at load/unload.
5102  */
5103
5104 static int
5105 load (struct module *module __unused, int cmd, void *arg __unused)
5106 {
5107         int error;
5108
5109         error = 0;
5110
5111         switch (cmd) {
5112         case MOD_LOAD :
5113                 /* initialize the subsystem */
5114                 error = pmc_initialize();
5115                 if (error != 0)
5116                         break;
5117                 PMCDBG(MOD,INI,1, "syscall=%d maxcpu=%d",
5118                     pmc_syscall_num, pmc_cpu_max());
5119                 break;
5120
5121
5122         case MOD_UNLOAD :
5123         case MOD_SHUTDOWN:
5124                 pmc_cleanup();
5125                 PMCDBG(MOD,INI,1, "%s", "unloaded");
5126                 break;
5127
5128         default :
5129                 error = EINVAL; /* XXX should panic(9) */
5130                 break;
5131         }
5132
5133         return error;
5134 }
5135
5136 /* memory pool */
5137 MALLOC_DEFINE(M_PMC, "pmc", "Memory space for the PMC module");