sys/dev/hwpmc/hwpmc_mod.c

   1 /*-
   2  * Copyright (c) 2003-2008 Joseph Koshy
   3  * Copyright (c) 2007 The FreeBSD Foundation
   4  * All rights reserved.
   5  *
   6  * Portions of this software were developed by A. Joseph Koshy under
   7  * sponsorship from the FreeBSD Foundation and Google, Inc.
   8  *
   9  * Redistribution and use in source and binary forms, with or without
  10  * modification, are permitted provided that the following conditions
  11  * are met:
  12  * 1. Redistributions of source code must retain the above copyright
  13  *    notice, this list of conditions and the following disclaimer.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice, this list of conditions and the following disclaimer in the
  16  *    documentation and/or other materials provided with the distribution.
  17  *
  18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  21  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
  22  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  28  * SUCH DAMAGE.
  29  *
  30  */
  31
  32 #include <sys/cdefs.h>
  33 __FBSDID("$FreeBSD$");
  34
  35 #include <sys/param.h>
  36 #include <sys/eventhandler.h>
  37 #include <sys/jail.h>
  38 #include <sys/kernel.h>
  39 #include <sys/kthread.h>
  40 #include <sys/limits.h>
  41 #include <sys/lock.h>
  42 #include <sys/malloc.h>
  43 #include <sys/module.h>
  44 #include <sys/mount.h>
  45 #include <sys/mutex.h>
  46 #include <sys/pmc.h>
  47 #include <sys/pmckern.h>
  48 #include <sys/pmclog.h>
  49 #include <sys/priv.h>
  50 #include <sys/proc.h>
  51 #include <sys/queue.h>
  52 #include <sys/resourcevar.h>
  53 #include <sys/rwlock.h>
  54 #include <sys/sched.h>
  55 #include <sys/signalvar.h>
  56 #include <sys/smp.h>
  57 #include <sys/sx.h>
  58 #include <sys/sysctl.h>
  59 #include <sys/sysent.h>
  60 #include <sys/systm.h>
  61 #include <sys/vnode.h>
  62
  63 #include <sys/linker.h>         /* needs to be after <sys/malloc.h> */
  64
  65 #include <machine/atomic.h>
  66 #include <machine/md_var.h>
  67
  68 #include <vm/vm.h>
  69 #include <vm/vm_extern.h>
  70 #include <vm/pmap.h>
  71 #include <vm/vm_map.h>
  72 #include <vm/vm_object.h>
  73
  74 #include "hwpmc_soft.h"
  75
  76 /*
  77  * Types
  78  */
  79
  80 enum pmc_flags {
  81         PMC_FLAG_NONE     = 0x00, /* do nothing */
  82         PMC_FLAG_REMOVE   = 0x01, /* atomically remove entry from hash */
  83         PMC_FLAG_ALLOCATE = 0x02, /* add entry to hash if not found */
  84 };
  85
  86 /*
  87  * The offset in sysent where the syscall is allocated.
  88  */
  89
  90 static int pmc_syscall_num = NO_SYSCALL;
  91 struct pmc_cpu          **pmc_pcpu;      /* per-cpu state */
  92 pmc_value_t             *pmc_pcpu_saved; /* saved PMC values: CSW handling */
  93
  94 #define PMC_PCPU_SAVED(C,R)     pmc_pcpu_saved[(R) + md->pmd_npmc*(C)]
  95
  96 struct mtx_pool         *pmc_mtxpool;
  97 static int              *pmc_pmcdisp;    /* PMC row dispositions */
  98
  99 #define PMC_ROW_DISP_IS_FREE(R)         (pmc_pmcdisp[(R)] == 0)
 100 #define PMC_ROW_DISP_IS_THREAD(R)       (pmc_pmcdisp[(R)] > 0)
 101 #define PMC_ROW_DISP_IS_STANDALONE(R)   (pmc_pmcdisp[(R)] < 0)
 102
 103 #define PMC_MARK_ROW_FREE(R) do {                                         \
 104         pmc_pmcdisp[(R)] = 0;                                             \
 105 } while (0)
 106
 107 #define PMC_MARK_ROW_STANDALONE(R) do {                                   \
 108         KASSERT(pmc_pmcdisp[(R)] <= 0, ("[pmc,%d] row disposition error", \
 109                     __LINE__));                                           \
 110         atomic_add_int(&pmc_pmcdisp[(R)], -1);                            \
 111         KASSERT(pmc_pmcdisp[(R)] >= (-pmc_cpu_max_active()),              \
 112                 ("[pmc,%d] row disposition error", __LINE__));            \
 113 } while (0)
 114
 115 #define PMC_UNMARK_ROW_STANDALONE(R) do {                                 \
 116         atomic_add_int(&pmc_pmcdisp[(R)], 1);                             \
 117         KASSERT(pmc_pmcdisp[(R)] <= 0, ("[pmc,%d] row disposition error", \
 118                     __LINE__));                                           \
 119 } while (0)
 120
 121 #define PMC_MARK_ROW_THREAD(R) do {                                       \
 122         KASSERT(pmc_pmcdisp[(R)] >= 0, ("[pmc,%d] row disposition error", \
 123                     __LINE__));                                           \
 124         atomic_add_int(&pmc_pmcdisp[(R)], 1);                             \
 125 } while (0)
 126
 127 #define PMC_UNMARK_ROW_THREAD(R) do {                                     \
 128         atomic_add_int(&pmc_pmcdisp[(R)], -1);                            \
 129         KASSERT(pmc_pmcdisp[(R)] >= 0, ("[pmc,%d] row disposition error", \
 130                     __LINE__));                                           \
 131 } while (0)
 132
 133
 134 /* various event handlers */
 135 static eventhandler_tag pmc_exit_tag, pmc_fork_tag, pmc_kld_load_tag,
 136     pmc_kld_unload_tag;
 137
 138 /* Module statistics */
 139 struct pmc_op_getdriverstats pmc_stats;
 140
 141 /* Machine/processor dependent operations */
 142 static struct pmc_mdep  *md;
 143
 144 /*
 145  * Hash tables mapping owner processes and target threads to PMCs.
 146  */
 147
 148 struct mtx pmc_processhash_mtx;         /* spin mutex */
 149 static u_long pmc_processhashmask;
 150 static LIST_HEAD(pmc_processhash, pmc_process)  *pmc_processhash;
 151
 152 /*
 153  * Hash table of PMC owner descriptors.  This table is protected by
 154  * the shared PMC "sx" lock.
 155  */
 156
 157 static u_long pmc_ownerhashmask;
 158 static LIST_HEAD(pmc_ownerhash, pmc_owner)      *pmc_ownerhash;
 159
 160 /*
 161  * List of PMC owners with system-wide sampling PMCs.
 162  */
 163
 164 static LIST_HEAD(, pmc_owner)                   pmc_ss_owners;
 165
 166
 167 /*
 168  * A map of row indices to classdep structures.
 169  */
 170 static struct pmc_classdep **pmc_rowindex_to_classdep;
 171
 172 /*
 173  * Prototypes
 174  */
 175
 176 #ifdef  HWPMC_DEBUG
 177 static int      pmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS);
 178 static int      pmc_debugflags_parse(char *newstr, char *fence);
 179 #endif
 180
 181 static int      load(struct module *module, int cmd, void *arg);
 182 static int      pmc_attach_process(struct proc *p, struct pmc *pm);
 183 static struct pmc *pmc_allocate_pmc_descriptor(void);
 184 static struct pmc_owner *pmc_allocate_owner_descriptor(struct proc *p);
 185 static int      pmc_attach_one_process(struct proc *p, struct pmc *pm);
 186 static int      pmc_can_allocate_rowindex(struct proc *p, unsigned int ri,
 187     int cpu);
 188 static int      pmc_can_attach(struct pmc *pm, struct proc *p);
 189 static void     pmc_capture_user_callchain(int cpu, int soft, struct trapframe *tf);
 190 static void     pmc_cleanup(void);
 191 static int      pmc_detach_process(struct proc *p, struct pmc *pm);
 192 static int      pmc_detach_one_process(struct proc *p, struct pmc *pm,
 193     int flags);
 194 static void     pmc_destroy_owner_descriptor(struct pmc_owner *po);
 195 static void     pmc_destroy_pmc_descriptor(struct pmc *pm);
 196 static struct pmc_owner *pmc_find_owner_descriptor(struct proc *p);
 197 static int      pmc_find_pmc(pmc_id_t pmcid, struct pmc **pm);
 198 static struct pmc *pmc_find_pmc_descriptor_in_process(struct pmc_owner *po,
 199     pmc_id_t pmc);
 200 static struct pmc_process *pmc_find_process_descriptor(struct proc *p,
 201     uint32_t mode);
 202 static void     pmc_force_context_switch(void);
 203 static void     pmc_link_target_process(struct pmc *pm,
 204     struct pmc_process *pp);
 205 static void     pmc_log_all_process_mappings(struct pmc_owner *po);
 206 static void     pmc_log_kernel_mappings(struct pmc *pm);
 207 static void     pmc_log_process_mappings(struct pmc_owner *po, struct proc *p);
 208 static void     pmc_maybe_remove_owner(struct pmc_owner *po);
 209 static void     pmc_process_csw_in(struct thread *td);
 210 static void     pmc_process_csw_out(struct thread *td);
 211 static void     pmc_process_exit(void *arg, struct proc *p);
 212 static void     pmc_process_fork(void *arg, struct proc *p1,
 213     struct proc *p2, int n);
 214 static void     pmc_process_samples(int cpu, int soft);
 215 static void     pmc_release_pmc_descriptor(struct pmc *pmc);
 216 static void     pmc_remove_owner(struct pmc_owner *po);
 217 static void     pmc_remove_process_descriptor(struct pmc_process *pp);
 218 static void     pmc_restore_cpu_binding(struct pmc_binding *pb);
 219 static void     pmc_save_cpu_binding(struct pmc_binding *pb);
 220 static void     pmc_select_cpu(int cpu);
 221 static int      pmc_start(struct pmc *pm);
 222 static int      pmc_stop(struct pmc *pm);
 223 static int      pmc_syscall_handler(struct thread *td, void *syscall_args);
 224 static void     pmc_unlink_target_process(struct pmc *pmc,
 225     struct pmc_process *pp);
 226 static int generic_switch_in(struct pmc_cpu *pc, struct pmc_process *pp);
 227 static int generic_switch_out(struct pmc_cpu *pc, struct pmc_process *pp);
 228 static struct pmc_mdep *pmc_generic_cpu_initialize(void);
 229 static void pmc_generic_cpu_finalize(struct pmc_mdep *md);
 230
 231 /*
 232  * Kernel tunables and sysctl(8) interface.
 233  */
 234
 235 SYSCTL_DECL(_kern_hwpmc);
 236
 237 static int pmc_callchaindepth = PMC_CALLCHAIN_DEPTH;
 238 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "callchaindepth", &pmc_callchaindepth);
 239 SYSCTL_INT(_kern_hwpmc, OID_AUTO, callchaindepth, CTLFLAG_TUN|CTLFLAG_RD,
 240     &pmc_callchaindepth, 0, "depth of call chain records");
 241
 242 #ifdef  HWPMC_DEBUG
 243 struct pmc_debugflags pmc_debugflags = PMC_DEBUG_DEFAULT_FLAGS;
 244 char    pmc_debugstr[PMC_DEBUG_STRSIZE];
 245 TUNABLE_STR(PMC_SYSCTL_NAME_PREFIX "debugflags", pmc_debugstr,
 246     sizeof(pmc_debugstr));
 247 SYSCTL_PROC(_kern_hwpmc, OID_AUTO, debugflags,
 248     CTLTYPE_STRING|CTLFLAG_RW|CTLFLAG_TUN,
 249     0, 0, pmc_debugflags_sysctl_handler, "A", "debug flags");
 250 #endif
 251
 252 /*
 253  * kern.hwpmc.hashrows -- determines the number of rows in the
 254  * of the hash table used to look up threads
 255  */
 256
 257 static int pmc_hashsize = PMC_HASH_SIZE;
 258 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "hashsize", &pmc_hashsize);
 259 SYSCTL_INT(_kern_hwpmc, OID_AUTO, hashsize, CTLFLAG_TUN|CTLFLAG_RD,
 260     &pmc_hashsize, 0, "rows in hash tables");
 261
 262 /*
 263  * kern.hwpmc.nsamples --- number of PC samples/callchain stacks per CPU
 264  */
 265
 266 static int pmc_nsamples = PMC_NSAMPLES;
 267 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "nsamples", &pmc_nsamples);
 268 SYSCTL_INT(_kern_hwpmc, OID_AUTO, nsamples, CTLFLAG_TUN|CTLFLAG_RD,
 269     &pmc_nsamples, 0, "number of PC samples per CPU");
 270
 271
 272 /*
 273  * kern.hwpmc.mtxpoolsize -- number of mutexes in the mutex pool.
 274  */
 275
 276 static int pmc_mtxpool_size = PMC_MTXPOOL_SIZE;
 277 TUNABLE_INT(PMC_SYSCTL_NAME_PREFIX "mtxpoolsize", &pmc_mtxpool_size);
 278 SYSCTL_INT(_kern_hwpmc, OID_AUTO, mtxpoolsize, CTLFLAG_TUN|CTLFLAG_RD,
 279     &pmc_mtxpool_size, 0, "size of spin mutex pool");
 280
 281
 282 /*
 283  * security.bsd.unprivileged_syspmcs -- allow non-root processes to
 284  * allocate system-wide PMCs.
 285  *
 286  * Allowing unprivileged processes to allocate system PMCs is convenient
 287  * if system-wide measurements need to be taken concurrently with other
 288  * per-process measurements.  This feature is turned off by default.
 289  */
 290
 291 static int pmc_unprivileged_syspmcs = 0;
 292 TUNABLE_INT("security.bsd.unprivileged_syspmcs", &pmc_unprivileged_syspmcs);
 293 SYSCTL_INT(_security_bsd, OID_AUTO, unprivileged_syspmcs, CTLFLAG_RW,
 294     &pmc_unprivileged_syspmcs, 0,
 295     "allow unprivileged process to allocate system PMCs");
 296
 297 /*
 298  * Hash function.  Discard the lower 2 bits of the pointer since
 299  * these are always zero for our uses.  The hash multiplier is
 300  * round((2^LONG_BIT) * ((sqrt(5)-1)/2)).
 301  */
 302
 303 #if     LONG_BIT == 64
 304 #define _PMC_HM         11400714819323198486u
 305 #elif   LONG_BIT == 32
 306 #define _PMC_HM         2654435769u
 307 #else
 308 #error  Must know the size of 'long' to compile
 309 #endif
 310
 311 #define PMC_HASH_PTR(P,M)       ((((unsigned long) (P) >> 2) * _PMC_HM) & (M))
 312
 313 /*
 314  * Syscall structures
 315  */
 316
 317 /* The `sysent' for the new syscall */
 318 static struct sysent pmc_sysent = {
 319         2,                      /* sy_narg */
 320         pmc_syscall_handler     /* sy_call */
 321 };
 322
 323 static struct syscall_module_data pmc_syscall_mod = {
 324         load,
 325         NULL,
 326         &pmc_syscall_num,
 327         &pmc_sysent,
 328 #if (__FreeBSD_version >= 1100000)
 329         { 0, NULL },
 330         SY_THR_STATIC_KLD,
 331 #else
 332         { 0, NULL }
 333 #endif
 334 };
 335
 336 static moduledata_t pmc_mod = {
 337         PMC_MODULE_NAME,
 338         syscall_module_handler,
 339         &pmc_syscall_mod
 340 };
 341
 342 DECLARE_MODULE(pmc, pmc_mod, SI_SUB_SMP, SI_ORDER_ANY);
 343 MODULE_VERSION(pmc, PMC_VERSION);
 344
 345 #ifdef  HWPMC_DEBUG
 346 enum pmc_dbgparse_state {
 347         PMCDS_WS,               /* in whitespace */
 348         PMCDS_MAJOR,            /* seen a major keyword */
 349         PMCDS_MINOR
 350 };
 351
 352 static int
 353 pmc_debugflags_parse(char *newstr, char *fence)
 354 {
 355         char c, *p, *q;
 356         struct pmc_debugflags *tmpflags;
 357         int error, found, *newbits, tmp;
 358         size_t kwlen;
 359
 360         tmpflags = malloc(sizeof(*tmpflags), M_PMC, M_WAITOK|M_ZERO);
 361
 362         p = newstr;
 363         error = 0;
 364
 365         for (; p < fence && (c = *p); p++) {
 366
 367                 /* skip white space */
 368                 if (c == ' ' || c == '\t')
 369                         continue;
 370
 371                 /* look for a keyword followed by "=" */
 372                 for (q = p; p < fence && (c = *p) && c != '='; p++)
 373                         ;
 374                 if (c != '=') {
 375                         error = EINVAL;
 376                         goto done;
 377                 }
 378
 379                 kwlen = p - q;
 380                 newbits = NULL;
 381
 382                 /* lookup flag group name */
 383 #define DBG_SET_FLAG_MAJ(S,F)                                           \
 384                 if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0)  \
 385                         newbits = &tmpflags->pdb_ ## F;
 386
 387                 DBG_SET_FLAG_MAJ("cpu",         CPU);
 388                 DBG_SET_FLAG_MAJ("csw",         CSW);
 389                 DBG_SET_FLAG_MAJ("logging",     LOG);
 390                 DBG_SET_FLAG_MAJ("module",      MOD);
 391                 DBG_SET_FLAG_MAJ("md",          MDP);
 392                 DBG_SET_FLAG_MAJ("owner",       OWN);
 393                 DBG_SET_FLAG_MAJ("pmc",         PMC);
 394                 DBG_SET_FLAG_MAJ("process",     PRC);
 395                 DBG_SET_FLAG_MAJ("sampling",    SAM);
 396
 397                 if (newbits == NULL) {
 398                         error = EINVAL;
 399                         goto done;
 400                 }
 401
 402                 p++;            /* skip the '=' */
 403
 404                 /* Now parse the individual flags */
 405                 tmp = 0;
 406         newflag:
 407                 for (q = p; p < fence && (c = *p); p++)
 408                         if (c == ' ' || c == '\t' || c == ',')
 409                                 break;
 410
 411                 /* p == fence or c == ws or c == "," or c == 0 */
 412
 413                 if ((kwlen = p - q) == 0) {
 414                         *newbits = tmp;
 415                         continue;
 416                 }
 417
 418                 found = 0;
 419 #define DBG_SET_FLAG_MIN(S,F)                                           \
 420                 if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0)  \
 421                         tmp |= found = (1 << PMC_DEBUG_MIN_ ## F)
 422
 423                 /* a '*' denotes all possible flags in the group */
 424                 if (kwlen == 1 && *q == '*')
 425                         tmp = found = ~0;
 426                 /* look for individual flag names */
 427                 DBG_SET_FLAG_MIN("allocaterow", ALR);
 428                 DBG_SET_FLAG_MIN("allocate",    ALL);
 429                 DBG_SET_FLAG_MIN("attach",      ATT);
 430                 DBG_SET_FLAG_MIN("bind",        BND);
 431                 DBG_SET_FLAG_MIN("config",      CFG);
 432                 DBG_SET_FLAG_MIN("exec",        EXC);
 433                 DBG_SET_FLAG_MIN("exit",        EXT);
 434                 DBG_SET_FLAG_MIN("find",        FND);
 435                 DBG_SET_FLAG_MIN("flush",       FLS);
 436                 DBG_SET_FLAG_MIN("fork",        FRK);
 437                 DBG_SET_FLAG_MIN("getbuf",      GTB);
 438                 DBG_SET_FLAG_MIN("hook",        PMH);
 439                 DBG_SET_FLAG_MIN("init",        INI);
 440                 DBG_SET_FLAG_MIN("intr",        INT);
 441                 DBG_SET_FLAG_MIN("linktarget",  TLK);
 442                 DBG_SET_FLAG_MIN("mayberemove", OMR);
 443                 DBG_SET_FLAG_MIN("ops",         OPS);
 444                 DBG_SET_FLAG_MIN("read",        REA);
 445                 DBG_SET_FLAG_MIN("register",    REG);
 446                 DBG_SET_FLAG_MIN("release",     REL);
 447                 DBG_SET_FLAG_MIN("remove",      ORM);
 448                 DBG_SET_FLAG_MIN("sample",      SAM);
 449                 DBG_SET_FLAG_MIN("scheduleio",  SIO);
 450                 DBG_SET_FLAG_MIN("select",      SEL);
 451                 DBG_SET_FLAG_MIN("signal",      SIG);
 452                 DBG_SET_FLAG_MIN("swi",         SWI);
 453                 DBG_SET_FLAG_MIN("swo",         SWO);
 454                 DBG_SET_FLAG_MIN("start",       STA);
 455                 DBG_SET_FLAG_MIN("stop",        STO);
 456                 DBG_SET_FLAG_MIN("syscall",     PMS);
 457                 DBG_SET_FLAG_MIN("unlinktarget", TUL);
 458                 DBG_SET_FLAG_MIN("write",       WRI);
 459                 if (found == 0) {
 460                         /* unrecognized flag name */
 461                         error = EINVAL;
 462                         goto done;
 463                 }
 464
 465                 if (c == 0 || c == ' ' || c == '\t') {  /* end of flag group */
 466                         *newbits = tmp;
 467                         continue;
 468                 }
 469
 470                 p++;
 471                 goto newflag;
 472         }
 473
 474         /* save the new flag set */
 475         bcopy(tmpflags, &pmc_debugflags, sizeof(pmc_debugflags));
 476
 477  done:
 478         free(tmpflags, M_PMC);
 479         return error;
 480 }
 481
 482 static int
 483 pmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS)
 484 {
 485         char *fence, *newstr;
 486         int error;
 487         unsigned int n;
 488
 489         (void) arg1; (void) arg2; /* unused parameters */
 490
 491         n = sizeof(pmc_debugstr);
 492         newstr = malloc(n, M_PMC, M_WAITOK|M_ZERO);
 493         (void) strlcpy(newstr, pmc_debugstr, n);
 494
 495         error = sysctl_handle_string(oidp, newstr, n, req);
 496
 497         /* if there is a new string, parse and copy it */
 498         if (error == 0 && req->newptr != NULL) {
 499                 fence = newstr + (n < req->newlen ? n : req->newlen + 1);
 500                 if ((error = pmc_debugflags_parse(newstr, fence)) == 0)
 501                         (void) strlcpy(pmc_debugstr, newstr,
 502                             sizeof(pmc_debugstr));
 503         }
 504
 505         free(newstr, M_PMC);
 506
 507         return error;
 508 }
 509 #endif
 510
 511 /*
 512  * Map a row index to a classdep structure and return the adjusted row
 513  * index for the PMC class index.
 514  */
 515 static struct pmc_classdep *
 516 pmc_ri_to_classdep(struct pmc_mdep *md, int ri, int *adjri)
 517 {
 518         struct pmc_classdep *pcd;
 519
 520         (void) md;
 521
 522         KASSERT(ri >= 0 && ri < md->pmd_npmc,
 523             ("[pmc,%d] illegal row-index %d", __LINE__, ri));
 524
 525         pcd = pmc_rowindex_to_classdep[ri];
 526
 527         KASSERT(pcd != NULL,
 528             ("[pmc,%d] ri %d null pcd", __LINE__, ri));
 529
 530         *adjri = ri - pcd->pcd_ri;
 531
 532         KASSERT(*adjri >= 0 && *adjri < pcd->pcd_num,
 533             ("[pmc,%d] adjusted row-index %d", __LINE__, *adjri));
 534
 535         return (pcd);
 536 }
 537
 538 /*
 539  * Concurrency Control
 540  *
 541  * The driver manages the following data structures:
 542  *
 543  *   - target process descriptors, one per target process
 544  *   - owner process descriptors (and attached lists), one per owner process
 545  *   - lookup hash tables for owner and target processes
 546  *   - PMC descriptors (and attached lists)
 547  *   - per-cpu hardware state
 548  *   - the 'hook' variable through which the kernel calls into
 549  *     this module
 550  *   - the machine hardware state (managed by the MD layer)
 551  *
 552  * These data structures are accessed from:
 553  *
 554  * - thread context-switch code
 555  * - interrupt handlers (possibly on multiple cpus)
 556  * - kernel threads on multiple cpus running on behalf of user
 557  *   processes doing system calls
 558  * - this driver's private kernel threads
 559  *
 560  * = Locks and Locking strategy =
 561  *
 562  * The driver uses four locking strategies for its operation:
 563  *
 564  * - The global SX lock "pmc_sx" is used to protect internal
 565  *   data structures.
 566  *
 567  *   Calls into the module by syscall() start with this lock being
 568  *   held in exclusive mode.  Depending on the requested operation,
 569  *   the lock may be downgraded to 'shared' mode to allow more
 570  *   concurrent readers into the module.  Calls into the module from
 571  *   other parts of the kernel acquire the lock in shared mode.
 572  *
 573  *   This SX lock is held in exclusive mode for any operations that
 574  *   modify the linkages between the driver's internal data structures.
 575  *
 576  *   The 'pmc_hook' function pointer is also protected by this lock.
 577  *   It is only examined with the sx lock held in exclusive mode.  The
 578  *   kernel module is allowed to be unloaded only with the sx lock held
 579  *   in exclusive mode.  In normal syscall handling, after acquiring the
 580  *   pmc_sx lock we first check that 'pmc_hook' is non-null before
 581  *   proceeding.  This prevents races between the thread unloading the module
 582  *   and other threads seeking to use the module.
 583  *
 584  * - Lookups of target process structures and owner process structures
 585  *   cannot use the global "pmc_sx" SX lock because these lookups need
 586  *   to happen during context switches and in other critical sections
 587  *   where sleeping is not allowed.  We protect these lookup tables
 588  *   with their own private spin-mutexes, "pmc_processhash_mtx" and
 589  *   "pmc_ownerhash_mtx".
 590  *
 591  * - Interrupt handlers work in a lock free manner.  At interrupt
 592  *   time, handlers look at the PMC pointer (phw->phw_pmc) configured
 593  *   when the PMC was started.  If this pointer is NULL, the interrupt
 594  *   is ignored after updating driver statistics.  We ensure that this
 595  *   pointer is set (using an atomic operation if necessary) before the
 596  *   PMC hardware is started.  Conversely, this pointer is unset atomically
 597  *   only after the PMC hardware is stopped.
 598  *
 599  *   We ensure that everything needed for the operation of an
 600  *   interrupt handler is available without it needing to acquire any
 601  *   locks.  We also ensure that a PMC's software state is destroyed only
 602  *   after the PMC is taken off hardware (on all CPUs).
 603  *
 604  * - Context-switch handling with process-private PMCs needs more
 605  *   care.
 606  *
 607  *   A given process may be the target of multiple PMCs.  For example,
 608  *   PMCATTACH and PMCDETACH may be requested by a process on one CPU
 609  *   while the target process is running on another.  A PMC could also
 610  *   be getting released because its owner is exiting.  We tackle
 611  *   these situations in the following manner:
 612  *
 613  *   - each target process structure 'pmc_process' has an array
 614  *     of 'struct pmc *' pointers, one for each hardware PMC.
 615  *
 616  *   - At context switch IN time, each "target" PMC in RUNNING state
 617  *     gets started on hardware and a pointer to each PMC is copied into
 618  *     the per-cpu phw array.  The 'runcount' for the PMC is
 619  *     incremented.
 620  *
 621  *   - At context switch OUT time, all process-virtual PMCs are stopped
 622  *     on hardware.  The saved value is added to the PMCs value field
 623  *     only if the PMC is in a non-deleted state (the PMCs state could
 624  *     have changed during the current time slice).
 625  *
 626  *     Note that since in-between a switch IN on a processor and a switch
 627  *     OUT, the PMC could have been released on another CPU.  Therefore
 628  *     context switch OUT always looks at the hardware state to turn
 629  *     OFF PMCs and will update a PMC's saved value only if reachable
 630  *     from the target process record.
 631  *
 632  *   - OP PMCRELEASE could be called on a PMC at any time (the PMC could
 633  *     be attached to many processes at the time of the call and could
 634  *     be active on multiple CPUs).
 635  *
 636  *     We prevent further scheduling of the PMC by marking it as in
 637  *     state 'DELETED'.  If the runcount of the PMC is non-zero then
 638  *     this PMC is currently running on a CPU somewhere.  The thread
 639  *     doing the PMCRELEASE operation waits by repeatedly doing a
 640  *     pause() till the runcount comes to zero.
 641  *
 642  * The contents of a PMC descriptor (struct pmc) are protected using
 643  * a spin-mutex.  In order to save space, we use a mutex pool.
 644  *
 645  * In terms of lock types used by witness(4), we use:
 646  * - Type "pmc-sx", used by the global SX lock.
 647  * - Type "pmc-sleep", for sleep mutexes used by logger threads.
 648  * - Type "pmc-per-proc", for protecting PMC owner descriptors.
 649  * - Type "pmc-leaf", used for all other spin mutexes.
 650  */
 651
 652 /*
 653  * save the cpu binding of the current kthread
 654  */
 655
 656 static void
 657 pmc_save_cpu_binding(struct pmc_binding *pb)
 658 {
 659         PMCDBG0(CPU,BND,2, "save-cpu");
 660         thread_lock(curthread);
 661         pb->pb_bound = sched_is_bound(curthread);
 662         pb->pb_cpu   = curthread->td_oncpu;
 663         thread_unlock(curthread);
 664         PMCDBG1(CPU,BND,2, "save-cpu cpu=%d", pb->pb_cpu);
 665 }
 666
 667 /*
 668  * restore the cpu binding of the current thread
 669  */
 670
 671 static void
 672 pmc_restore_cpu_binding(struct pmc_binding *pb)
 673 {
 674         PMCDBG2(CPU,BND,2, "restore-cpu curcpu=%d restore=%d",
 675             curthread->td_oncpu, pb->pb_cpu);
 676         thread_lock(curthread);
 677         if (pb->pb_bound)
 678                 sched_bind(curthread, pb->pb_cpu);
 679         else
 680                 sched_unbind(curthread);
 681         thread_unlock(curthread);
 682         PMCDBG0(CPU,BND,2, "restore-cpu done");
 683 }
 684
 685 /*
 686  * move execution over the specified cpu and bind it there.
 687  */
 688
 689 static void
 690 pmc_select_cpu(int cpu)
 691 {
 692         KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
 693             ("[pmc,%d] bad cpu number %d", __LINE__, cpu));
 694
 695         /* Never move to an inactive CPU. */
 696         KASSERT(pmc_cpu_is_active(cpu), ("[pmc,%d] selecting inactive "
 697             "CPU %d", __LINE__, cpu));
 698
 699         PMCDBG1(CPU,SEL,2, "select-cpu cpu=%d", cpu);
 700         thread_lock(curthread);
 701         sched_bind(curthread, cpu);
 702         thread_unlock(curthread);
 703
 704         KASSERT(curthread->td_oncpu == cpu,
 705             ("[pmc,%d] CPU not bound [cpu=%d, curr=%d]", __LINE__,
 706                 cpu, curthread->td_oncpu));
 707
 708         PMCDBG1(CPU,SEL,2, "select-cpu cpu=%d ok", cpu);
 709 }
 710
 711 /*
 712  * Force a context switch.
 713  *
 714  * We do this by pause'ing for 1 tick -- invoking mi_switch() is not
 715  * guaranteed to force a context switch.
 716  */
 717
 718 static void
 719 pmc_force_context_switch(void)
 720 {
 721
 722         pause("pmcctx", 1);
 723 }
 724
 725 /*
 726  * Get the file name for an executable.  This is a simple wrapper
 727  * around vn_fullpath(9).
 728  */
 729
 730 static void
 731 pmc_getfilename(struct vnode *v, char **fullpath, char **freepath)
 732 {
 733
 734         *fullpath = "unknown";
 735         *freepath = NULL;
 736         vn_fullpath(curthread, v, fullpath, freepath);
 737 }
 738
 739 /*
 740  * remove an process owning PMCs
 741  */
 742
 743 void
 744 pmc_remove_owner(struct pmc_owner *po)
 745 {
 746         struct pmc *pm, *tmp;
 747
 748         sx_assert(&pmc_sx, SX_XLOCKED);
 749
 750         PMCDBG1(OWN,ORM,1, "remove-owner po=%p", po);
 751
 752         /* Remove descriptor from the owner hash table */
 753         LIST_REMOVE(po, po_next);
 754
 755         /* release all owned PMC descriptors */
 756         LIST_FOREACH_SAFE(pm, &po->po_pmcs, pm_next, tmp) {
 757                 PMCDBG1(OWN,ORM,2, "pmc=%p", pm);
 758                 KASSERT(pm->pm_owner == po,
 759                     ("[pmc,%d] owner %p != po %p", __LINE__, pm->pm_owner, po));
 760
 761                 pmc_release_pmc_descriptor(pm); /* will unlink from the list */
 762                 pmc_destroy_pmc_descriptor(pm);
 763         }
 764
 765         KASSERT(po->po_sscount == 0,
 766             ("[pmc,%d] SS count not zero", __LINE__));
 767         KASSERT(LIST_EMPTY(&po->po_pmcs),
 768             ("[pmc,%d] PMC list not empty", __LINE__));
 769
 770         /* de-configure the log file if present */
 771         if (po->po_flags & PMC_PO_OWNS_LOGFILE)
 772                 pmclog_deconfigure_log(po);
 773 }
 774
 775 /*
 776  * remove an owner process record if all conditions are met.
 777  */
 778
 779 static void
 780 pmc_maybe_remove_owner(struct pmc_owner *po)
 781 {
 782
 783         PMCDBG1(OWN,OMR,1, "maybe-remove-owner po=%p", po);
 784
 785         /*
 786          * Remove owner record if
 787          * - this process does not own any PMCs
 788          * - this process has not allocated a system-wide sampling buffer
 789          */
 790
 791         if (LIST_EMPTY(&po->po_pmcs) &&
 792             ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)) {
 793                 pmc_remove_owner(po);
 794                 pmc_destroy_owner_descriptor(po);
 795         }
 796 }
 797
 798 /*
 799  * Add an association between a target process and a PMC.
 800  */
 801
 802 static void
 803 pmc_link_target_process(struct pmc *pm, struct pmc_process *pp)
 804 {
 805         int ri;
 806         struct pmc_target *pt;
 807
 808         sx_assert(&pmc_sx, SX_XLOCKED);
 809
 810         KASSERT(pm != NULL && pp != NULL,
 811             ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp));
 812         KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)),
 813             ("[pmc,%d] Attaching a non-process-virtual pmc=%p to pid=%d",
 814                 __LINE__, pm, pp->pp_proc->p_pid));
 815         KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= ((int) md->pmd_npmc - 1),
 816             ("[pmc,%d] Illegal reference count %d for process record %p",
 817                 __LINE__, pp->pp_refcnt, (void *) pp));
 818
 819         ri = PMC_TO_ROWINDEX(pm);
 820
 821         PMCDBG3(PRC,TLK,1, "link-target pmc=%p ri=%d pmc-process=%p",
 822             pm, ri, pp);
 823
 824 #ifdef  HWPMC_DEBUG
 825         LIST_FOREACH(pt, &pm->pm_targets, pt_next)
 826             if (pt->pt_process == pp)
 827                     KASSERT(0, ("[pmc,%d] pp %p already in pmc %p targets",
 828                                 __LINE__, pp, pm));
 829 #endif
 830
 831         pt = malloc(sizeof(struct pmc_target), M_PMC, M_WAITOK|M_ZERO);
 832         pt->pt_process = pp;
 833
 834         LIST_INSERT_HEAD(&pm->pm_targets, pt, pt_next);
 835
 836         atomic_store_rel_ptr((uintptr_t *)&pp->pp_pmcs[ri].pp_pmc,
 837             (uintptr_t)pm);
 838
 839         if (pm->pm_owner->po_owner == pp->pp_proc)
 840                 pm->pm_flags |= PMC_F_ATTACHED_TO_OWNER;
 841
 842         /*
 843          * Initialize the per-process values at this row index.
 844          */
 845         pp->pp_pmcs[ri].pp_pmcval = PMC_TO_MODE(pm) == PMC_MODE_TS ?
 846             pm->pm_sc.pm_reloadcount : 0;
 847
 848         pp->pp_refcnt++;
 849
 850 }
 851
 852 /*
 853  * Removes the association between a target process and a PMC.
 854  */
 855
 856 static void
 857 pmc_unlink_target_process(struct pmc *pm, struct pmc_process *pp)
 858 {
 859         int ri;
 860         struct proc *p;
 861         struct pmc_target *ptgt;
 862
 863         sx_assert(&pmc_sx, SX_XLOCKED);
 864
 865         KASSERT(pm != NULL && pp != NULL,
 866             ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp));
 867
 868         KASSERT(pp->pp_refcnt >= 1 && pp->pp_refcnt <= (int) md->pmd_npmc,
 869             ("[pmc,%d] Illegal ref count %d on process record %p",
 870                 __LINE__, pp->pp_refcnt, (void *) pp));
 871
 872         ri = PMC_TO_ROWINDEX(pm);
 873
 874         PMCDBG3(PRC,TUL,1, "unlink-target pmc=%p ri=%d pmc-process=%p",
 875             pm, ri, pp);
 876
 877         KASSERT(pp->pp_pmcs[ri].pp_pmc == pm,
 878             ("[pmc,%d] PMC ri %d mismatch pmc %p pp->[ri] %p", __LINE__,
 879                 ri, pm, pp->pp_pmcs[ri].pp_pmc));
 880
 881         pp->pp_pmcs[ri].pp_pmc = NULL;
 882         pp->pp_pmcs[ri].pp_pmcval = (pmc_value_t) 0;
 883
 884         /* Remove owner-specific flags */
 885         if (pm->pm_owner->po_owner == pp->pp_proc) {
 886                 pp->pp_flags &= ~PMC_PP_ENABLE_MSR_ACCESS;
 887                 pm->pm_flags &= ~PMC_F_ATTACHED_TO_OWNER;
 888         }
 889
 890         pp->pp_refcnt--;
 891
 892         /* Remove the target process from the PMC structure */
 893         LIST_FOREACH(ptgt, &pm->pm_targets, pt_next)
 894                 if (ptgt->pt_process == pp)
 895                         break;
 896
 897         KASSERT(ptgt != NULL, ("[pmc,%d] process %p (pp: %p) not found "
 898                     "in pmc %p", __LINE__, pp->pp_proc, pp, pm));
 899
 900         LIST_REMOVE(ptgt, pt_next);
 901         free(ptgt, M_PMC);
 902
 903         /* if the PMC now lacks targets, send the owner a SIGIO */
 904         if (LIST_EMPTY(&pm->pm_targets)) {
 905                 p = pm->pm_owner->po_owner;
 906                 PROC_LOCK(p);
 907                 kern_psignal(p, SIGIO);
 908                 PROC_UNLOCK(p);
 909
 910                 PMCDBG2(PRC,SIG,2, "signalling proc=%p signal=%d", p,
 911                     SIGIO);
 912         }
 913 }
 914
 915 /*
 916  * Check if PMC 'pm' may be attached to target process 't'.
 917  */
 918
 919 static int
 920 pmc_can_attach(struct pmc *pm, struct proc *t)
 921 {
 922         struct proc *o;         /* pmc owner */
 923         struct ucred *oc, *tc;  /* owner, target credentials */
 924         int decline_attach, i;
 925
 926         /*
 927          * A PMC's owner can always attach that PMC to itself.
 928          */
 929
 930         if ((o = pm->pm_owner->po_owner) == t)
 931                 return 0;
 932
 933         PROC_LOCK(o);
 934         oc = o->p_ucred;
 935         crhold(oc);
 936         PROC_UNLOCK(o);
 937
 938         PROC_LOCK(t);
 939         tc = t->p_ucred;
 940         crhold(tc);
 941         PROC_UNLOCK(t);
 942
 943         /*
 944          * The effective uid of the PMC owner should match at least one
 945          * of the {effective,real,saved} uids of the target process.
 946          */
 947
 948         decline_attach = oc->cr_uid != tc->cr_uid &&
 949             oc->cr_uid != tc->cr_svuid &&
 950             oc->cr_uid != tc->cr_ruid;
 951
 952         /*
 953          * Every one of the target's group ids, must be in the owner's
 954          * group list.
 955          */
 956         for (i = 0; !decline_attach && i < tc->cr_ngroups; i++)
 957                 decline_attach = !groupmember(tc->cr_groups[i], oc);
 958
 959         /* check the read and saved gids too */
 960         if (decline_attach == 0)
 961                 decline_attach = !groupmember(tc->cr_rgid, oc) ||
 962                     !groupmember(tc->cr_svgid, oc);
 963
 964         crfree(tc);
 965         crfree(oc);
 966
 967         return !decline_attach;
 968 }
 969
 970 /*
 971  * Attach a process to a PMC.
 972  */
 973
 974 static int
 975 pmc_attach_one_process(struct proc *p, struct pmc *pm)
 976 {
 977         int ri;
 978         char *fullpath, *freepath;
 979         struct pmc_process      *pp;
 980
 981         sx_assert(&pmc_sx, SX_XLOCKED);
 982
 983         PMCDBG5(PRC,ATT,2, "attach-one pm=%p ri=%d proc=%p (%d, %s)", pm,
 984             PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm);
 985
 986         /*
 987          * Locate the process descriptor corresponding to process 'p',
 988          * allocating space as needed.
 989          *
 990          * Verify that rowindex 'pm_rowindex' is free in the process
 991          * descriptor.
 992          *
 993          * If not, allocate space for a descriptor and link the
 994          * process descriptor and PMC.
 995          */
 996         ri = PMC_TO_ROWINDEX(pm);
 997
 998         if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_ALLOCATE)) == NULL)
 999                 return ENOMEM;
1000
1001         if (pp->pp_pmcs[ri].pp_pmc == pm) /* already present at slot [ri] */
1002                 return EEXIST;
1003
1004         if (pp->pp_pmcs[ri].pp_pmc != NULL)
1005                 return EBUSY;
1006
1007         pmc_link_target_process(pm, pp);
1008
1009         if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) &&
1010             (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) == 0)
1011                 pm->pm_flags |= PMC_F_NEEDS_LOGFILE;
1012
1013         pm->pm_flags |= PMC_F_ATTACH_DONE; /* mark as attached */
1014
1015         /* issue an attach event to a configured log file */
1016         if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE) {
1017                 pmc_getfilename(p->p_textvp, &fullpath, &freepath);
1018                 if (p->p_flag & P_KTHREAD) {
1019                         fullpath = kernelname;
1020                         freepath = NULL;
1021                 } else
1022                         pmclog_process_pmcattach(pm, p->p_pid, fullpath);
1023                 if (freepath)
1024                         free(freepath, M_TEMP);
1025                 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1026                         pmc_log_process_mappings(pm->pm_owner, p);
1027         }
1028         /* mark process as using HWPMCs */
1029         PROC_LOCK(p);
1030         p->p_flag |= P_HWPMC;
1031         PROC_UNLOCK(p);
1032
1033         return 0;
1034 }
1035
1036 /*
1037  * Attach a process and optionally its children
1038  */
1039
1040 static int
1041 pmc_attach_process(struct proc *p, struct pmc *pm)
1042 {
1043         int error;
1044         struct proc *top;
1045
1046         sx_assert(&pmc_sx, SX_XLOCKED);
1047
1048         PMCDBG5(PRC,ATT,1, "attach pm=%p ri=%d proc=%p (%d, %s)", pm,
1049             PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm);
1050
1051
1052         /*
1053          * If this PMC successfully allowed a GETMSR operation
1054          * in the past, disallow further ATTACHes.
1055          */
1056
1057         if ((pm->pm_flags & PMC_PP_ENABLE_MSR_ACCESS) != 0)
1058                 return EPERM;
1059
1060         if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0)
1061                 return pmc_attach_one_process(p, pm);
1062
1063         /*
1064          * Traverse all child processes, attaching them to
1065          * this PMC.
1066          */
1067
1068         sx_slock(&proctree_lock);
1069
1070         top = p;
1071
1072         for (;;) {
1073                 if ((error = pmc_attach_one_process(p, pm)) != 0)
1074                         break;
1075                 if (!LIST_EMPTY(&p->p_children))
1076                         p = LIST_FIRST(&p->p_children);
1077                 else for (;;) {
1078                         if (p == top)
1079                                 goto done;
1080                         if (LIST_NEXT(p, p_sibling)) {
1081                                 p = LIST_NEXT(p, p_sibling);
1082                                 break;
1083                         }
1084                         p = p->p_pptr;
1085                 }
1086         }
1087
1088         if (error)
1089                 (void) pmc_detach_process(top, pm);
1090
1091  done:
1092         sx_sunlock(&proctree_lock);
1093         return error;
1094 }
1095
1096 /*
1097  * Detach a process from a PMC.  If there are no other PMCs tracking
1098  * this process, remove the process structure from its hash table.  If
1099  * 'flags' contains PMC_FLAG_REMOVE, then free the process structure.
1100  */
1101
1102 static int
1103 pmc_detach_one_process(struct proc *p, struct pmc *pm, int flags)
1104 {
1105         int ri;
1106         struct pmc_process *pp;
1107
1108         sx_assert(&pmc_sx, SX_XLOCKED);
1109
1110         KASSERT(pm != NULL,
1111             ("[pmc,%d] null pm pointer", __LINE__));
1112
1113         ri = PMC_TO_ROWINDEX(pm);
1114
1115         PMCDBG6(PRC,ATT,2, "detach-one pm=%p ri=%d proc=%p (%d, %s) flags=0x%x",
1116             pm, ri, p, p->p_pid, p->p_comm, flags);
1117
1118         if ((pp = pmc_find_process_descriptor(p, 0)) == NULL)
1119                 return ESRCH;
1120
1121         if (pp->pp_pmcs[ri].pp_pmc != pm)
1122                 return EINVAL;
1123
1124         pmc_unlink_target_process(pm, pp);
1125
1126         /* Issue a detach entry if a log file is configured */
1127         if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE)
1128                 pmclog_process_pmcdetach(pm, p->p_pid);
1129
1130         /*
1131          * If there are no PMCs targetting this process, we remove its
1132          * descriptor from the target hash table and unset the P_HWPMC
1133          * flag in the struct proc.
1134          */
1135         KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= (int) md->pmd_npmc,
1136             ("[pmc,%d] Illegal refcnt %d for process struct %p",
1137                 __LINE__, pp->pp_refcnt, pp));
1138
1139         if (pp->pp_refcnt != 0) /* still a target of some PMC */
1140                 return 0;
1141
1142         pmc_remove_process_descriptor(pp);
1143
1144         if (flags & PMC_FLAG_REMOVE)
1145                 free(pp, M_PMC);
1146
1147         PROC_LOCK(p);
1148         p->p_flag &= ~P_HWPMC;
1149         PROC_UNLOCK(p);
1150
1151         return 0;
1152 }
1153
1154 /*
1155  * Detach a process and optionally its descendants from a PMC.
1156  */
1157
1158 static int
1159 pmc_detach_process(struct proc *p, struct pmc *pm)
1160 {
1161         struct proc *top;
1162
1163         sx_assert(&pmc_sx, SX_XLOCKED);
1164
1165         PMCDBG5(PRC,ATT,1, "detach pm=%p ri=%d proc=%p (%d, %s)", pm,
1166             PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm);
1167
1168         if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0)
1169                 return pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE);
1170
1171         /*
1172          * Traverse all children, detaching them from this PMC.  We
1173          * ignore errors since we could be detaching a PMC from a
1174          * partially attached proc tree.
1175          */
1176
1177         sx_slock(&proctree_lock);
1178
1179         top = p;
1180
1181         for (;;) {
1182                 (void) pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE);
1183
1184                 if (!LIST_EMPTY(&p->p_children))
1185                         p = LIST_FIRST(&p->p_children);
1186                 else for (;;) {
1187                         if (p == top)
1188                                 goto done;
1189                         if (LIST_NEXT(p, p_sibling)) {
1190                                 p = LIST_NEXT(p, p_sibling);
1191                                 break;
1192                         }
1193                         p = p->p_pptr;
1194                 }
1195         }
1196
1197  done:
1198         sx_sunlock(&proctree_lock);
1199
1200         if (LIST_EMPTY(&pm->pm_targets))
1201                 pm->pm_flags &= ~PMC_F_ATTACH_DONE;
1202
1203         return 0;
1204 }
1205
1206
1207 /*
1208  * Thread context switch IN
1209  */
1210
1211 static void
1212 pmc_process_csw_in(struct thread *td)
1213 {
1214         int cpu;
1215         unsigned int adjri, ri;
1216         struct pmc *pm;
1217         struct proc *p;
1218         struct pmc_cpu *pc;
1219         struct pmc_hw *phw;
1220         pmc_value_t newvalue;
1221         struct pmc_process *pp;
1222         struct pmc_classdep *pcd;
1223
1224         p = td->td_proc;
1225
1226         if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE)) == NULL)
1227                 return;
1228
1229         KASSERT(pp->pp_proc == td->td_proc,
1230             ("[pmc,%d] not my thread state", __LINE__));
1231
1232         critical_enter(); /* no preemption from this point */
1233
1234         cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */
1235
1236         PMCDBG5(CSW,SWI,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p,
1237             p->p_pid, p->p_comm, pp);
1238
1239         KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
1240             ("[pmc,%d] wierd CPU id %d", __LINE__, cpu));
1241
1242         pc = pmc_pcpu[cpu];
1243
1244         for (ri = 0; ri < md->pmd_npmc; ri++) {
1245
1246                 if ((pm = pp->pp_pmcs[ri].pp_pmc) == NULL)
1247                         continue;
1248
1249                 KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)),
1250                     ("[pmc,%d] Target PMC in non-virtual mode (%d)",
1251                         __LINE__, PMC_TO_MODE(pm)));
1252
1253                 KASSERT(PMC_TO_ROWINDEX(pm) == ri,
1254                     ("[pmc,%d] Row index mismatch pmc %d != ri %d",
1255                         __LINE__, PMC_TO_ROWINDEX(pm), ri));
1256
1257                 /*
1258                  * Only PMCs that are marked as 'RUNNING' need
1259                  * be placed on hardware.
1260                  */
1261
1262                 if (pm->pm_state != PMC_STATE_RUNNING)
1263                         continue;
1264
1265                 /* increment PMC runcount */
1266                 atomic_add_rel_int(&pm->pm_runcount, 1);
1267
1268                 /* configure the HWPMC we are going to use. */
1269                 pcd = pmc_ri_to_classdep(md, ri, &adjri);
1270                 pcd->pcd_config_pmc(cpu, adjri, pm);
1271
1272                 phw = pc->pc_hwpmcs[ri];
1273
1274                 KASSERT(phw != NULL,
1275                     ("[pmc,%d] null hw pointer", __LINE__));
1276
1277                 KASSERT(phw->phw_pmc == pm,
1278                     ("[pmc,%d] hw->pmc %p != pmc %p", __LINE__,
1279                         phw->phw_pmc, pm));
1280
1281                 /*
1282                  * Write out saved value and start the PMC.
1283                  *
1284                  * Sampling PMCs use a per-process value, while
1285                  * counting mode PMCs use a per-pmc value that is
1286                  * inherited across descendants.
1287                  */
1288                 if (PMC_TO_MODE(pm) == PMC_MODE_TS) {
1289                         mtx_pool_lock_spin(pmc_mtxpool, pm);
1290                         newvalue = PMC_PCPU_SAVED(cpu,ri) =
1291                             pp->pp_pmcs[ri].pp_pmcval;
1292                         mtx_pool_unlock_spin(pmc_mtxpool, pm);
1293                 } else {
1294                         KASSERT(PMC_TO_MODE(pm) == PMC_MODE_TC,
1295                             ("[pmc,%d] illegal mode=%d", __LINE__,
1296                             PMC_TO_MODE(pm)));
1297                         mtx_pool_lock_spin(pmc_mtxpool, pm);
1298                         newvalue = PMC_PCPU_SAVED(cpu, ri) =
1299                             pm->pm_gv.pm_savedvalue;
1300                         mtx_pool_unlock_spin(pmc_mtxpool, pm);
1301                 }
1302
1303                 PMCDBG3(CSW,SWI,1,"cpu=%d ri=%d new=%jd", cpu, ri, newvalue);
1304
1305                 pcd->pcd_write_pmc(cpu, adjri, newvalue);
1306                 pcd->pcd_start_pmc(cpu, adjri);
1307         }
1308
1309         /*
1310          * perform any other architecture/cpu dependent thread
1311          * switch-in actions.
1312          */
1313
1314         (void) (*md->pmd_switch_in)(pc, pp);
1315
1316         critical_exit();
1317
1318 }
1319
1320 /*
1321  * Thread context switch OUT.
1322  */
1323
1324 static void
1325 pmc_process_csw_out(struct thread *td)
1326 {
1327         int cpu;
1328         int64_t tmp;
1329         struct pmc *pm;
1330         struct proc *p;
1331         enum pmc_mode mode;
1332         struct pmc_cpu *pc;
1333         pmc_value_t newvalue;
1334         unsigned int adjri, ri;
1335         struct pmc_process *pp;
1336         struct pmc_classdep *pcd;
1337
1338
1339         /*
1340          * Locate our process descriptor; this may be NULL if
1341          * this process is exiting and we have already removed
1342          * the process from the target process table.
1343          *
1344          * Note that due to kernel preemption, multiple
1345          * context switches may happen while the process is
1346          * exiting.
1347          *
1348          * Note also that if the target process cannot be
1349          * found we still need to deconfigure any PMCs that
1350          * are currently running on hardware.
1351          */
1352
1353         p = td->td_proc;
1354         pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE);
1355
1356         /*
1357          * save PMCs
1358          */
1359
1360         critical_enter();
1361
1362         cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */
1363
1364         PMCDBG5(CSW,SWO,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p,
1365             p->p_pid, p->p_comm, pp);
1366
1367         KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
1368             ("[pmc,%d wierd CPU id %d", __LINE__, cpu));
1369
1370         pc = pmc_pcpu[cpu];
1371
1372         /*
1373          * When a PMC gets unlinked from a target PMC, it will
1374          * be removed from the target's pp_pmc[] array.
1375          *
1376          * However, on a MP system, the target could have been
1377          * executing on another CPU at the time of the unlink.
1378          * So, at context switch OUT time, we need to look at
1379          * the hardware to determine if a PMC is scheduled on
1380          * it.
1381          */
1382
1383         for (ri = 0; ri < md->pmd_npmc; ri++) {
1384
1385                 pcd = pmc_ri_to_classdep(md, ri, &adjri);
1386                 pm  = NULL;
1387                 (void) (*pcd->pcd_get_config)(cpu, adjri, &pm);
1388
1389                 if (pm == NULL) /* nothing at this row index */
1390                         continue;
1391
1392                 mode = PMC_TO_MODE(pm);
1393                 if (!PMC_IS_VIRTUAL_MODE(mode))
1394                         continue; /* not a process virtual PMC */
1395
1396                 KASSERT(PMC_TO_ROWINDEX(pm) == ri,
1397                     ("[pmc,%d] ri mismatch pmc(%d) ri(%d)",
1398                         __LINE__, PMC_TO_ROWINDEX(pm), ri));
1399
1400                 /* Stop hardware if not already stopped */
1401                 if (pm->pm_stalled == 0)
1402                         pcd->pcd_stop_pmc(cpu, adjri);
1403
1404                 /* reduce this PMC's runcount */
1405                 atomic_subtract_rel_int(&pm->pm_runcount, 1);
1406
1407                 /*
1408                  * If this PMC is associated with this process,
1409                  * save the reading.
1410                  */
1411
1412                 if (pp != NULL && pp->pp_pmcs[ri].pp_pmc != NULL) {
1413
1414                         KASSERT(pm == pp->pp_pmcs[ri].pp_pmc,
1415                             ("[pmc,%d] pm %p != pp_pmcs[%d] %p", __LINE__,
1416                                 pm, ri, pp->pp_pmcs[ri].pp_pmc));
1417
1418                         KASSERT(pp->pp_refcnt > 0,
1419                             ("[pmc,%d] pp refcnt = %d", __LINE__,
1420                                 pp->pp_refcnt));
1421
1422                         pcd->pcd_read_pmc(cpu, adjri, &newvalue);
1423
1424                         tmp = newvalue - PMC_PCPU_SAVED(cpu,ri);
1425
1426                         PMCDBG3(CSW,SWO,1,"cpu=%d ri=%d tmp=%jd", cpu, ri,
1427                             tmp);
1428
1429                         if (mode == PMC_MODE_TS) {
1430
1431                                 /*
1432                                  * For sampling process-virtual PMCs,
1433                                  * we expect the count to be
1434                                  * decreasing as the 'value'
1435                                  * programmed into the PMC is the
1436                                  * number of events to be seen till
1437                                  * the next sampling interrupt.
1438                                  */
1439                                 if (tmp < 0)
1440                                         tmp += pm->pm_sc.pm_reloadcount;
1441                                 mtx_pool_lock_spin(pmc_mtxpool, pm);
1442                                 pp->pp_pmcs[ri].pp_pmcval -= tmp;
1443                                 if ((int64_t) pp->pp_pmcs[ri].pp_pmcval <= 0)
1444                                         pp->pp_pmcs[ri].pp_pmcval +=
1445                                             pm->pm_sc.pm_reloadcount;
1446                                 mtx_pool_unlock_spin(pmc_mtxpool, pm);
1447
1448                         } else {
1449
1450                                 /*
1451                                  * For counting process-virtual PMCs,
1452                                  * we expect the count to be
1453                                  * increasing monotonically, modulo a 64
1454                                  * bit wraparound.
1455                                  */
1456                                 KASSERT((int64_t) tmp >= 0,
1457                                     ("[pmc,%d] negative increment cpu=%d "
1458                                      "ri=%d newvalue=%jx saved=%jx "
1459                                      "incr=%jx", __LINE__, cpu, ri,
1460                                      newvalue, PMC_PCPU_SAVED(cpu,ri), tmp));
1461
1462                                 mtx_pool_lock_spin(pmc_mtxpool, pm);
1463                                 pm->pm_gv.pm_savedvalue += tmp;
1464                                 pp->pp_pmcs[ri].pp_pmcval += tmp;
1465                                 mtx_pool_unlock_spin(pmc_mtxpool, pm);
1466
1467                                 if (pm->pm_flags & PMC_F_LOG_PROCCSW)
1468                                         pmclog_process_proccsw(pm, pp, tmp);
1469                         }
1470                 }
1471
1472                 /* mark hardware as free */
1473                 pcd->pcd_config_pmc(cpu, adjri, NULL);
1474         }
1475
1476         /*
1477          * perform any other architecture/cpu dependent thread
1478          * switch out functions.
1479          */
1480
1481         (void) (*md->pmd_switch_out)(pc, pp);
1482
1483         critical_exit();
1484 }
1485
1486 /*
1487  * A mapping change for a process.
1488  */
1489
1490 static void
1491 pmc_process_mmap(struct thread *td, struct pmckern_map_in *pkm)
1492 {
1493         int ri;
1494         pid_t pid;
1495         char *fullpath, *freepath;
1496         const struct pmc *pm;
1497         struct pmc_owner *po;
1498         const struct pmc_process *pp;
1499
1500         freepath = fullpath = NULL;
1501         pmc_getfilename((struct vnode *) pkm->pm_file, &fullpath, &freepath);
1502
1503         pid = td->td_proc->p_pid;
1504
1505         /* Inform owners of all system-wide sampling PMCs. */
1506         LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1507             if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1508                 pmclog_process_map_in(po, pid, pkm->pm_address, fullpath);
1509
1510         if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL)
1511                 goto done;
1512
1513         /*
1514          * Inform sampling PMC owners tracking this process.
1515          */
1516         for (ri = 0; ri < md->pmd_npmc; ri++)
1517                 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL &&
1518                     PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1519                         pmclog_process_map_in(pm->pm_owner,
1520                             pid, pkm->pm_address, fullpath);
1521
1522   done:
1523         if (freepath)
1524                 free(freepath, M_TEMP);
1525 }
1526
1527
1528 /*
1529  * Log an munmap request.
1530  */
1531
1532 static void
1533 pmc_process_munmap(struct thread *td, struct pmckern_map_out *pkm)
1534 {
1535         int ri;
1536         pid_t pid;
1537         struct pmc_owner *po;
1538         const struct pmc *pm;
1539         const struct pmc_process *pp;
1540
1541         pid = td->td_proc->p_pid;
1542
1543         LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1544             if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1545                 pmclog_process_map_out(po, pid, pkm->pm_address,
1546                     pkm->pm_address + pkm->pm_size);
1547
1548         if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL)
1549                 return;
1550
1551         for (ri = 0; ri < md->pmd_npmc; ri++)
1552                 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL &&
1553                     PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1554                         pmclog_process_map_out(pm->pm_owner, pid,
1555                             pkm->pm_address, pkm->pm_address + pkm->pm_size);
1556 }
1557
1558 /*
1559  * Log mapping information about the kernel.
1560  */
1561
1562 static void
1563 pmc_log_kernel_mappings(struct pmc *pm)
1564 {
1565         struct pmc_owner *po;
1566         struct pmckern_map_in *km, *kmbase;
1567
1568         sx_assert(&pmc_sx, SX_LOCKED);
1569         KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)),
1570             ("[pmc,%d] non-sampling PMC (%p) desires mapping information",
1571                 __LINE__, (void *) pm));
1572
1573         po = pm->pm_owner;
1574
1575         if (po->po_flags & PMC_PO_INITIAL_MAPPINGS_DONE)
1576                 return;
1577
1578         /*
1579          * Log the current set of kernel modules.
1580          */
1581         kmbase = linker_hwpmc_list_objects();
1582         for (km = kmbase; km->pm_file != NULL; km++) {
1583                 PMCDBG2(LOG,REG,1,"%s %p", (char *) km->pm_file,
1584                     (void *) km->pm_address);
1585                 pmclog_process_map_in(po, (pid_t) -1, km->pm_address,
1586                     km->pm_file);
1587         }
1588         free(kmbase, M_LINKER);
1589
1590         po->po_flags |= PMC_PO_INITIAL_MAPPINGS_DONE;
1591 }
1592
1593 /*
1594  * Log the mappings for a single process.
1595  */
1596
1597 static void
1598 pmc_log_process_mappings(struct pmc_owner *po, struct proc *p)
1599 {
1600         vm_map_t map;
1601         struct vnode *vp;
1602         struct vmspace *vm;
1603         vm_map_entry_t entry;
1604         vm_offset_t last_end;
1605         u_int last_timestamp;
1606         struct vnode *last_vp;
1607         vm_offset_t start_addr;
1608         vm_object_t obj, lobj, tobj;
1609         char *fullpath, *freepath;
1610
1611         last_vp = NULL;
1612         last_end = (vm_offset_t) 0;
1613         fullpath = freepath = NULL;
1614
1615         if ((vm = vmspace_acquire_ref(p)) == NULL)
1616                 return;
1617
1618         map = &vm->vm_map;
1619         vm_map_lock_read(map);
1620
1621         for (entry = map->header.next; entry != &map->header; entry = entry->next) {
1622
1623                 if (entry == NULL) {
1624                         PMCDBG2(LOG,OPS,2, "hwpmc: vm_map entry unexpectedly "
1625                             "NULL! pid=%d vm_map=%p\n", p->p_pid, map);
1626                         break;
1627                 }
1628
1629                 /*
1630                  * We only care about executable map entries.
1631                  */
1632                 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) ||
1633                     !(entry->protection & VM_PROT_EXECUTE) ||
1634                     (entry->object.vm_object == NULL)) {
1635                         continue;
1636                 }
1637
1638                 obj = entry->object.vm_object;
1639                 VM_OBJECT_RLOCK(obj);
1640
1641                 /*
1642                  * Walk the backing_object list to find the base
1643                  * (non-shadowed) vm_object.
1644                  */
1645                 for (lobj = tobj = obj; tobj != NULL; tobj = tobj->backing_object) {
1646                         if (tobj != obj)
1647                                 VM_OBJECT_RLOCK(tobj);
1648                         if (lobj != obj)
1649                                 VM_OBJECT_RUNLOCK(lobj);
1650                         lobj = tobj;
1651                 }
1652
1653                 /*
1654                  * At this point lobj is the base vm_object and it is locked.
1655                  */
1656                 if (lobj == NULL) {
1657                         PMCDBG3(LOG,OPS,2, "hwpmc: lobj unexpectedly NULL! pid=%d "
1658                             "vm_map=%p vm_obj=%p\n", p->p_pid, map, obj);
1659                         VM_OBJECT_RUNLOCK(obj);
1660                         continue;
1661                 }
1662
1663                 if (lobj->type != OBJT_VNODE || lobj->handle == NULL) {
1664                         if (lobj != obj)
1665                                 VM_OBJECT_RUNLOCK(lobj);
1666                         VM_OBJECT_RUNLOCK(obj);
1667                         continue;
1668                 }
1669
1670                 /*
1671                  * Skip contiguous regions that point to the same
1672                  * vnode, so we don't emit redundant MAP-IN
1673                  * directives.
1674                  */
1675                 if (entry->start == last_end && lobj->handle == last_vp) {
1676                         last_end = entry->end;
1677                         if (lobj != obj)
1678                                 VM_OBJECT_RUNLOCK(lobj);
1679                         VM_OBJECT_RUNLOCK(obj);
1680                         continue;
1681                 }
1682
1683                 /*
1684                  * We don't want to keep the proc's vm_map or this
1685                  * vm_object locked while we walk the pathname, since
1686                  * vn_fullpath() can sleep.  However, if we drop the
1687                  * lock, it's possible for concurrent activity to
1688                  * modify the vm_map list.  To protect against this,
1689                  * we save the vm_map timestamp before we release the
1690                  * lock, and check it after we reacquire the lock
1691                  * below.
1692                  */
1693                 start_addr = entry->start;
1694                 last_end = entry->end;
1695                 last_timestamp = map->timestamp;
1696                 vm_map_unlock_read(map);
1697
1698                 vp = lobj->handle;
1699                 vref(vp);
1700                 if (lobj != obj)
1701                         VM_OBJECT_RUNLOCK(lobj);
1702
1703                 VM_OBJECT_RUNLOCK(obj);
1704
1705                 freepath = NULL;
1706                 pmc_getfilename(vp, &fullpath, &freepath);
1707                 last_vp = vp;
1708
1709                 vrele(vp);
1710
1711                 vp = NULL;
1712                 pmclog_process_map_in(po, p->p_pid, start_addr, fullpath);
1713                 if (freepath)
1714                         free(freepath, M_TEMP);
1715
1716                 vm_map_lock_read(map);
1717
1718                 /*
1719                  * If our saved timestamp doesn't match, this means
1720                  * that the vm_map was modified out from under us and
1721                  * we can't trust our current "entry" pointer.  Do a
1722                  * new lookup for this entry.  If there is no entry
1723                  * for this address range, vm_map_lookup_entry() will
1724                  * return the previous one, so we always want to go to
1725                  * entry->next on the next loop iteration.
1726                  *
1727                  * There is an edge condition here that can occur if
1728                  * there is no entry at or before this address.  In
1729                  * this situation, vm_map_lookup_entry returns
1730                  * &map->header, which would cause our loop to abort
1731                  * without processing the rest of the map.  However,
1732                  * in practice this will never happen for process
1733                  * vm_map.  This is because the executable's text
1734                  * segment is the first mapping in the proc's address
1735                  * space, and this mapping is never removed until the
1736                  * process exits, so there will always be a non-header
1737                  * entry at or before the requested address for
1738                  * vm_map_lookup_entry to return.
1739                  */
1740                 if (map->timestamp != last_timestamp)
1741                         vm_map_lookup_entry(map, last_end - 1, &entry);
1742         }
1743
1744         vm_map_unlock_read(map);
1745         vmspace_free(vm);
1746         return;
1747 }
1748
1749 /*
1750  * Log mappings for all processes in the system.
1751  */
1752
1753 static void
1754 pmc_log_all_process_mappings(struct pmc_owner *po)
1755 {
1756         struct proc *p, *top;
1757
1758         sx_assert(&pmc_sx, SX_XLOCKED);
1759
1760         if ((p = pfind(1)) == NULL)
1761                 panic("[pmc,%d] Cannot find init", __LINE__);
1762
1763         PROC_UNLOCK(p);
1764
1765         sx_slock(&proctree_lock);
1766
1767         top = p;
1768
1769         for (;;) {
1770                 pmc_log_process_mappings(po, p);
1771                 if (!LIST_EMPTY(&p->p_children))
1772                         p = LIST_FIRST(&p->p_children);
1773                 else for (;;) {
1774                         if (p == top)
1775                                 goto done;
1776                         if (LIST_NEXT(p, p_sibling)) {
1777                                 p = LIST_NEXT(p, p_sibling);
1778                                 break;
1779                         }
1780                         p = p->p_pptr;
1781                 }
1782         }
1783  done:
1784         sx_sunlock(&proctree_lock);
1785 }
1786
1787 /*
1788  * The 'hook' invoked from the kernel proper
1789  */
1790
1791
1792 #ifdef  HWPMC_DEBUG
1793 const char *pmc_hooknames[] = {
1794         /* these strings correspond to PMC_FN_* in <sys/pmckern.h> */
1795         "",
1796         "EXEC",
1797         "CSW-IN",
1798         "CSW-OUT",
1799         "SAMPLE",
1800         "UNUSED1",
1801         "UNUSED2",
1802         "MMAP",
1803         "MUNMAP",
1804         "CALLCHAIN-NMI",
1805         "CALLCHAIN-SOFT",
1806         "SOFTSAMPLING"
1807 };
1808 #endif
1809
1810 static int
1811 pmc_hook_handler(struct thread *td, int function, void *arg)
1812 {
1813
1814         PMCDBG4(MOD,PMH,1, "hook td=%p func=%d \"%s\" arg=%p", td, function,
1815             pmc_hooknames[function], arg);
1816
1817         switch (function)
1818         {
1819
1820         /*
1821          * Process exec()
1822          */
1823
1824         case PMC_FN_PROCESS_EXEC:
1825         {
1826                 char *fullpath, *freepath;
1827                 unsigned int ri;
1828                 int is_using_hwpmcs;
1829                 struct pmc *pm;
1830                 struct proc *p;
1831                 struct pmc_owner *po;
1832                 struct pmc_process *pp;
1833                 struct pmckern_procexec *pk;
1834
1835                 sx_assert(&pmc_sx, SX_XLOCKED);
1836
1837                 p = td->td_proc;
1838                 pmc_getfilename(p->p_textvp, &fullpath, &freepath);
1839
1840                 pk = (struct pmckern_procexec *) arg;
1841
1842                 /* Inform owners of SS mode PMCs of the exec event. */
1843                 LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1844                     if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1845                             pmclog_process_procexec(po, PMC_ID_INVALID,
1846                                 p->p_pid, pk->pm_entryaddr, fullpath);
1847
1848                 PROC_LOCK(p);
1849                 is_using_hwpmcs = p->p_flag & P_HWPMC;
1850                 PROC_UNLOCK(p);
1851
1852                 if (!is_using_hwpmcs) {
1853                         if (freepath)
1854                                 free(freepath, M_TEMP);
1855                         break;
1856                 }
1857
1858                 /*
1859                  * PMCs are not inherited across an exec():  remove any
1860                  * PMCs that this process is the owner of.
1861                  */
1862
1863                 if ((po = pmc_find_owner_descriptor(p)) != NULL) {
1864                         pmc_remove_owner(po);
1865                         pmc_destroy_owner_descriptor(po);
1866                 }
1867
1868                 /*
1869                  * If the process being exec'ed is not the target of any
1870                  * PMC, we are done.
1871                  */
1872                 if ((pp = pmc_find_process_descriptor(p, 0)) == NULL) {
1873                         if (freepath)
1874                                 free(freepath, M_TEMP);
1875                         break;
1876                 }
1877
1878                 /*
1879                  * Log the exec event to all monitoring owners.  Skip
1880                  * owners who have already recieved the event because
1881                  * they had system sampling PMCs active.
1882                  */
1883                 for (ri = 0; ri < md->pmd_npmc; ri++)
1884                         if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) {
1885                                 po = pm->pm_owner;
1886                                 if (po->po_sscount == 0 &&
1887                                     po->po_flags & PMC_PO_OWNS_LOGFILE)
1888                                         pmclog_process_procexec(po, pm->pm_id,
1889                                             p->p_pid, pk->pm_entryaddr,
1890                                             fullpath);
1891                         }
1892
1893                 if (freepath)
1894                         free(freepath, M_TEMP);
1895
1896
1897                 PMCDBG4(PRC,EXC,1, "exec proc=%p (%d, %s) cred-changed=%d",
1898                     p, p->p_pid, p->p_comm, pk->pm_credentialschanged);
1899
1900                 if (pk->pm_credentialschanged == 0) /* no change */
1901                         break;
1902
1903                 /*
1904                  * If the newly exec()'ed process has a different credential
1905                  * than before, allow it to be the target of a PMC only if
1906                  * the PMC's owner has sufficient priviledge.
1907                  */
1908
1909                 for (ri = 0; ri < md->pmd_npmc; ri++)
1910                         if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL)
1911                                 if (pmc_can_attach(pm, td->td_proc) != 0)
1912                                         pmc_detach_one_process(td->td_proc,
1913                                             pm, PMC_FLAG_NONE);
1914
1915                 KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= (int) md->pmd_npmc,
1916                     ("[pmc,%d] Illegal ref count %d on pp %p", __LINE__,
1917                         pp->pp_refcnt, pp));
1918
1919                 /*
1920                  * If this process is no longer the target of any
1921                  * PMCs, we can remove the process entry and free
1922                  * up space.
1923                  */
1924
1925                 if (pp->pp_refcnt == 0) {
1926                         pmc_remove_process_descriptor(pp);
1927                         free(pp, M_PMC);
1928                         break;
1929                 }
1930
1931         }
1932         break;
1933
1934         case PMC_FN_CSW_IN:
1935                 pmc_process_csw_in(td);
1936                 break;
1937
1938         case PMC_FN_CSW_OUT:
1939                 pmc_process_csw_out(td);
1940                 break;
1941
1942         /*
1943          * Process accumulated PC samples.
1944          *
1945          * This function is expected to be called by hardclock() for
1946          * each CPU that has accumulated PC samples.
1947          *
1948          * This function is to be executed on the CPU whose samples
1949          * are being processed.
1950          */
1951         case PMC_FN_DO_SAMPLES:
1952
1953                 /*
1954                  * Clear the cpu specific bit in the CPU mask before
1955                  * do the rest of the processing.  If the NMI handler
1956                  * gets invoked after the "atomic_clear_int()" call
1957                  * below but before "pmc_process_samples()" gets
1958                  * around to processing the interrupt, then we will
1959                  * come back here at the next hardclock() tick (and
1960                  * may find nothing to do if "pmc_process_samples()"
1961                  * had already processed the interrupt).  We don't
1962                  * lose the interrupt sample.
1963                  */
1964                 CPU_CLR_ATOMIC(PCPU_GET(cpuid), &pmc_cpumask);
1965                 pmc_process_samples(PCPU_GET(cpuid), PMC_HR);
1966                 pmc_process_samples(PCPU_GET(cpuid), PMC_SR);
1967                 break;
1968
1969         case PMC_FN_MMAP:
1970                 sx_assert(&pmc_sx, SX_LOCKED);
1971                 pmc_process_mmap(td, (struct pmckern_map_in *) arg);
1972                 break;
1973
1974         case PMC_FN_MUNMAP:
1975                 sx_assert(&pmc_sx, SX_LOCKED);
1976                 pmc_process_munmap(td, (struct pmckern_map_out *) arg);
1977                 break;
1978
1979         case PMC_FN_USER_CALLCHAIN:
1980                 /*
1981                  * Record a call chain.
1982                  */
1983                 KASSERT(td == curthread, ("[pmc,%d] td != curthread",
1984                     __LINE__));
1985
1986                 pmc_capture_user_callchain(PCPU_GET(cpuid), PMC_HR,
1987                     (struct trapframe *) arg);
1988                 td->td_pflags &= ~TDP_CALLCHAIN;
1989                 break;
1990
1991         case PMC_FN_USER_CALLCHAIN_SOFT:
1992                 /*
1993                  * Record a call chain.
1994                  */
1995                 KASSERT(td == curthread, ("[pmc,%d] td != curthread",
1996                     __LINE__));
1997                 pmc_capture_user_callchain(PCPU_GET(cpuid), PMC_SR,
1998                     (struct trapframe *) arg);
1999                 td->td_pflags &= ~TDP_CALLCHAIN;
2000                 break;
2001
2002         case PMC_FN_SOFT_SAMPLING:
2003                 /*
2004                  * Call soft PMC sampling intr.
2005                  */
2006                 pmc_soft_intr((struct pmckern_soft *) arg);
2007                 break;
2008
2009         default:
2010 #ifdef  HWPMC_DEBUG
2011                 KASSERT(0, ("[pmc,%d] unknown hook %d\n", __LINE__, function));
2012 #endif
2013                 break;
2014
2015         }
2016
2017         return 0;
2018 }
2019
2020 /*
2021  * allocate a 'struct pmc_owner' descriptor in the owner hash table.
2022  */
2023
2024 static struct pmc_owner *
2025 pmc_allocate_owner_descriptor(struct proc *p)
2026 {
2027         uint32_t hindex;
2028         struct pmc_owner *po;
2029         struct pmc_ownerhash *poh;
2030
2031         hindex = PMC_HASH_PTR(p, pmc_ownerhashmask);
2032         poh = &pmc_ownerhash[hindex];
2033
2034         /* allocate space for N pointers and one descriptor struct */
2035         po = malloc(sizeof(struct pmc_owner), M_PMC, M_WAITOK|M_ZERO);
2036         po->po_owner = p;
2037         LIST_INSERT_HEAD(poh, po, po_next); /* insert into hash table */
2038
2039         TAILQ_INIT(&po->po_logbuffers);
2040         mtx_init(&po->po_mtx, "pmc-owner-mtx", "pmc-per-proc", MTX_SPIN);
2041
2042         PMCDBG4(OWN,ALL,1, "allocate-owner proc=%p (%d, %s) pmc-owner=%p",
2043             p, p->p_pid, p->p_comm, po);
2044
2045         return po;
2046 }
2047
2048 static void
2049 pmc_destroy_owner_descriptor(struct pmc_owner *po)
2050 {
2051
2052         PMCDBG4(OWN,REL,1, "destroy-owner po=%p proc=%p (%d, %s)",
2053             po, po->po_owner, po->po_owner->p_pid, po->po_owner->p_comm);
2054
2055         mtx_destroy(&po->po_mtx);
2056         free(po, M_PMC);
2057 }
2058
2059 /*
2060  * find the descriptor corresponding to process 'p', adding or removing it
2061  * as specified by 'mode'.
2062  */
2063
2064 static struct pmc_process *
2065 pmc_find_process_descriptor(struct proc *p, uint32_t mode)
2066 {
2067         uint32_t hindex;
2068         struct pmc_process *pp, *ppnew;
2069         struct pmc_processhash *pph;
2070
2071         hindex = PMC_HASH_PTR(p, pmc_processhashmask);
2072         pph = &pmc_processhash[hindex];
2073
2074         ppnew = NULL;
2075
2076         /*
2077          * Pre-allocate memory in the FIND_ALLOCATE case since we
2078          * cannot call malloc(9) once we hold a spin lock.
2079          */
2080         if (mode & PMC_FLAG_ALLOCATE)
2081                 ppnew = malloc(sizeof(struct pmc_process) + md->pmd_npmc *
2082                     sizeof(struct pmc_targetstate), M_PMC, M_WAITOK|M_ZERO);
2083
2084         mtx_lock_spin(&pmc_processhash_mtx);
2085         LIST_FOREACH(pp, pph, pp_next)
2086             if (pp->pp_proc == p)
2087                     break;
2088
2089         if ((mode & PMC_FLAG_REMOVE) && pp != NULL)
2090                 LIST_REMOVE(pp, pp_next);
2091
2092         if ((mode & PMC_FLAG_ALLOCATE) && pp == NULL &&
2093             ppnew != NULL) {
2094                 ppnew->pp_proc = p;
2095                 LIST_INSERT_HEAD(pph, ppnew, pp_next);
2096                 pp = ppnew;
2097                 ppnew = NULL;
2098         }
2099         mtx_unlock_spin(&pmc_processhash_mtx);
2100
2101         if (pp != NULL && ppnew != NULL)
2102                 free(ppnew, M_PMC);
2103
2104         return pp;
2105 }
2106
2107 /*
2108  * remove a process descriptor from the process hash table.
2109  */
2110
2111 static void
2112 pmc_remove_process_descriptor(struct pmc_process *pp)
2113 {
2114         KASSERT(pp->pp_refcnt == 0,
2115             ("[pmc,%d] Removing process descriptor %p with count %d",
2116                 __LINE__, pp, pp->pp_refcnt));
2117
2118         mtx_lock_spin(&pmc_processhash_mtx);
2119         LIST_REMOVE(pp, pp_next);
2120         mtx_unlock_spin(&pmc_processhash_mtx);
2121 }
2122
2123
2124 /*
2125  * find an owner descriptor corresponding to proc 'p'
2126  */
2127
2128 static struct pmc_owner *
2129 pmc_find_owner_descriptor(struct proc *p)
2130 {
2131         uint32_t hindex;
2132         struct pmc_owner *po;
2133         struct pmc_ownerhash *poh;
2134
2135         hindex = PMC_HASH_PTR(p, pmc_ownerhashmask);
2136         poh = &pmc_ownerhash[hindex];
2137
2138         po = NULL;
2139         LIST_FOREACH(po, poh, po_next)
2140             if (po->po_owner == p)
2141                     break;
2142
2143         PMCDBG5(OWN,FND,1, "find-owner proc=%p (%d, %s) hindex=0x%x -> "
2144             "pmc-owner=%p", p, p->p_pid, p->p_comm, hindex, po);
2145
2146         return po;
2147 }
2148
2149 /*
2150  * pmc_allocate_pmc_descriptor
2151  *
2152  * Allocate a pmc descriptor and initialize its
2153  * fields.
2154  */
2155
2156 static struct pmc *
2157 pmc_allocate_pmc_descriptor(void)
2158 {
2159         struct pmc *pmc;
2160
2161         pmc = malloc(sizeof(struct pmc), M_PMC, M_WAITOK|M_ZERO);
2162
2163         PMCDBG1(PMC,ALL,1, "allocate-pmc -> pmc=%p", pmc);
2164
2165         return pmc;
2166 }
2167
2168 /*
2169  * Destroy a pmc descriptor.
2170  */
2171
2172 static void
2173 pmc_destroy_pmc_descriptor(struct pmc *pm)
2174 {
2175
2176         KASSERT(pm->pm_state == PMC_STATE_DELETED ||
2177             pm->pm_state == PMC_STATE_FREE,
2178             ("[pmc,%d] destroying non-deleted PMC", __LINE__));
2179         KASSERT(LIST_EMPTY(&pm->pm_targets),
2180             ("[pmc,%d] destroying pmc with targets", __LINE__));
2181         KASSERT(pm->pm_owner == NULL,
2182             ("[pmc,%d] destroying pmc attached to an owner", __LINE__));
2183         KASSERT(pm->pm_runcount == 0,
2184             ("[pmc,%d] pmc has non-zero run count %d", __LINE__,
2185                 pm->pm_runcount));
2186
2187         free(pm, M_PMC);
2188 }
2189
2190 static void
2191 pmc_wait_for_pmc_idle(struct pmc *pm)
2192 {
2193 #ifdef HWPMC_DEBUG
2194         volatile int maxloop;
2195
2196         maxloop = 100 * pmc_cpu_max();
2197 #endif
2198         /*
2199          * Loop (with a forced context switch) till the PMC's runcount
2200          * comes down to zero.
2201          */
2202         while (atomic_load_acq_32(&pm->pm_runcount) > 0) {
2203 #ifdef HWPMC_DEBUG
2204                 maxloop--;
2205                 KASSERT(maxloop > 0,
2206                     ("[pmc,%d] (ri%d, rc%d) waiting too long for "
2207                         "pmc to be free", __LINE__,
2208                         PMC_TO_ROWINDEX(pm), pm->pm_runcount));
2209 #endif
2210                 pmc_force_context_switch();
2211         }
2212 }
2213
2214 /*
2215  * This function does the following things:
2216  *
2217  *  - detaches the PMC from hardware
2218  *  - unlinks all target threads that were attached to it
2219  *  - removes the PMC from its owner's list
2220  *  - destroys the PMC private mutex
2221  *
2222  * Once this function completes, the given pmc pointer can be freed by
2223  * calling pmc_destroy_pmc_descriptor().
2224  */
2225
2226 static void
2227 pmc_release_pmc_descriptor(struct pmc *pm)
2228 {
2229         enum pmc_mode mode;
2230         struct pmc_hw *phw;
2231         u_int adjri, ri, cpu;
2232         struct pmc_owner *po;
2233         struct pmc_binding pb;
2234         struct pmc_process *pp;
2235         struct pmc_classdep *pcd;
2236         struct pmc_target *ptgt, *tmp;
2237
2238         sx_assert(&pmc_sx, SX_XLOCKED);
2239
2240         KASSERT(pm, ("[pmc,%d] null pmc", __LINE__));
2241
2242         ri   = PMC_TO_ROWINDEX(pm);
2243         pcd  = pmc_ri_to_classdep(md, ri, &adjri);
2244         mode = PMC_TO_MODE(pm);
2245
2246         PMCDBG3(PMC,REL,1, "release-pmc pmc=%p ri=%d mode=%d", pm, ri,
2247             mode);
2248
2249         /*
2250          * First, we take the PMC off hardware.
2251          */
2252         cpu = 0;
2253         if (PMC_IS_SYSTEM_MODE(mode)) {
2254
2255                 /*
2256                  * A system mode PMC runs on a specific CPU.  Switch
2257                  * to this CPU and turn hardware off.
2258                  */
2259                 pmc_save_cpu_binding(&pb);
2260
2261                 cpu = PMC_TO_CPU(pm);
2262
2263                 pmc_select_cpu(cpu);
2264
2265                 /* switch off non-stalled CPUs */
2266                 if (pm->pm_state == PMC_STATE_RUNNING &&
2267                     pm->pm_stalled == 0) {
2268
2269                         phw = pmc_pcpu[cpu]->pc_hwpmcs[ri];
2270
2271                         KASSERT(phw->phw_pmc == pm,
2272                             ("[pmc, %d] pmc ptr ri(%d) hw(%p) pm(%p)",
2273                                 __LINE__, ri, phw->phw_pmc, pm));
2274                         PMCDBG2(PMC,REL,2, "stopping cpu=%d ri=%d", cpu, ri);
2275
2276                         critical_enter();
2277                         pcd->pcd_stop_pmc(cpu, adjri);
2278                         critical_exit();
2279                 }
2280
2281                 PMCDBG2(PMC,REL,2, "decfg cpu=%d ri=%d", cpu, ri);
2282
2283                 critical_enter();
2284                 pcd->pcd_config_pmc(cpu, adjri, NULL);
2285                 critical_exit();
2286
2287                 /* adjust the global and process count of SS mode PMCs */
2288                 if (mode == PMC_MODE_SS && pm->pm_state == PMC_STATE_RUNNING) {
2289                         po = pm->pm_owner;
2290                         po->po_sscount--;
2291                         if (po->po_sscount == 0) {
2292                                 atomic_subtract_rel_int(&pmc_ss_count, 1);
2293                                 LIST_REMOVE(po, po_ssnext);
2294                         }
2295                 }
2296
2297                 pm->pm_state = PMC_STATE_DELETED;
2298
2299                 pmc_restore_cpu_binding(&pb);
2300
2301                 /*
2302                  * We could have references to this PMC structure in
2303                  * the per-cpu sample queues.  Wait for the queue to
2304                  * drain.
2305                  */
2306                 pmc_wait_for_pmc_idle(pm);
2307
2308         } else if (PMC_IS_VIRTUAL_MODE(mode)) {
2309
2310                 /*
2311                  * A virtual PMC could be running on multiple CPUs at
2312                  * a given instant.
2313                  *
2314                  * By marking its state as DELETED, we ensure that
2315                  * this PMC is never further scheduled on hardware.
2316                  *
2317                  * Then we wait till all CPUs are done with this PMC.
2318                  */
2319                 pm->pm_state = PMC_STATE_DELETED;
2320
2321
2322                 /* Wait for the PMCs runcount to come to zero. */
2323                 pmc_wait_for_pmc_idle(pm);
2324
2325                 /*
2326                  * At this point the PMC is off all CPUs and cannot be
2327                  * freshly scheduled onto a CPU.  It is now safe to
2328                  * unlink all targets from this PMC.  If a
2329                  * process-record's refcount falls to zero, we remove
2330                  * it from the hash table.  The module-wide SX lock
2331                  * protects us from races.
2332                  */
2333                 LIST_FOREACH_SAFE(ptgt, &pm->pm_targets, pt_next, tmp) {
2334                         pp = ptgt->pt_process;
2335                         pmc_unlink_target_process(pm, pp); /* frees 'ptgt' */
2336
2337                         PMCDBG1(PMC,REL,3, "pp->refcnt=%d", pp->pp_refcnt);
2338
2339                         /*
2340                          * If the target process record shows that no
2341                          * PMCs are attached to it, reclaim its space.
2342                          */
2343
2344                         if (pp->pp_refcnt == 0) {
2345                                 pmc_remove_process_descriptor(pp);
2346                                 free(pp, M_PMC);
2347                         }
2348                 }
2349
2350                 cpu = curthread->td_oncpu; /* setup cpu for pmd_release() */
2351
2352         }
2353
2354         /*
2355          * Release any MD resources
2356          */
2357         (void) pcd->pcd_release_pmc(cpu, adjri, pm);
2358
2359         /*
2360          * Update row disposition
2361          */
2362
2363         if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm)))
2364                 PMC_UNMARK_ROW_STANDALONE(ri);
2365         else
2366                 PMC_UNMARK_ROW_THREAD(ri);
2367
2368         /* unlink from the owner's list */
2369         if (pm->pm_owner) {
2370                 LIST_REMOVE(pm, pm_next);
2371                 pm->pm_owner = NULL;
2372         }
2373 }
2374
2375 /*
2376  * Register an owner and a pmc.
2377  */
2378
2379 static int
2380 pmc_register_owner(struct proc *p, struct pmc *pmc)
2381 {
2382         struct pmc_owner *po;
2383
2384         sx_assert(&pmc_sx, SX_XLOCKED);
2385
2386         if ((po = pmc_find_owner_descriptor(p)) == NULL)
2387                 if ((po = pmc_allocate_owner_descriptor(p)) == NULL)
2388                         return ENOMEM;
2389
2390         KASSERT(pmc->pm_owner == NULL,
2391             ("[pmc,%d] attempting to own an initialized PMC", __LINE__));
2392         pmc->pm_owner  = po;
2393
2394         LIST_INSERT_HEAD(&po->po_pmcs, pmc, pm_next);
2395
2396         PROC_LOCK(p);
2397         p->p_flag |= P_HWPMC;
2398         PROC_UNLOCK(p);
2399
2400         if (po->po_flags & PMC_PO_OWNS_LOGFILE)
2401                 pmclog_process_pmcallocate(pmc);
2402
2403         PMCDBG2(PMC,REG,1, "register-owner pmc-owner=%p pmc=%p",
2404             po, pmc);
2405
2406         return 0;
2407 }
2408
2409 /*
2410  * Return the current row disposition:
2411  * == 0 => FREE
2412  *  > 0 => PROCESS MODE
2413  *  < 0 => SYSTEM MODE
2414  */
2415
2416 int
2417 pmc_getrowdisp(int ri)
2418 {
2419         return pmc_pmcdisp[ri];
2420 }
2421
2422 /*
2423  * Check if a PMC at row index 'ri' can be allocated to the current
2424  * process.
2425  *
2426  * Allocation can fail if:
2427  *   - the current process is already being profiled by a PMC at index 'ri',
2428  *     attached to it via OP_PMCATTACH.
2429  *   - the current process has already allocated a PMC at index 'ri'
2430  *     via OP_ALLOCATE.
2431  */
2432
2433 static int
2434 pmc_can_allocate_rowindex(struct proc *p, unsigned int ri, int cpu)
2435 {
2436         enum pmc_mode mode;
2437         struct pmc *pm;
2438         struct pmc_owner *po;
2439         struct pmc_process *pp;
2440
2441         PMCDBG5(PMC,ALR,1, "can-allocate-rowindex proc=%p (%d, %s) ri=%d "
2442             "cpu=%d", p, p->p_pid, p->p_comm, ri, cpu);
2443
2444         /*
2445          * We shouldn't have already allocated a process-mode PMC at
2446          * row index 'ri'.
2447          *
2448          * We shouldn't have allocated a system-wide PMC on the same
2449          * CPU and same RI.
2450          */
2451         if ((po = pmc_find_owner_descriptor(p)) != NULL)
2452                 LIST_FOREACH(pm, &po->po_pmcs, pm_next) {
2453                     if (PMC_TO_ROWINDEX(pm) == ri) {
2454                             mode = PMC_TO_MODE(pm);
2455                             if (PMC_IS_VIRTUAL_MODE(mode))
2456                                     return EEXIST;
2457                             if (PMC_IS_SYSTEM_MODE(mode) &&
2458                                 (int) PMC_TO_CPU(pm) == cpu)
2459                                     return EEXIST;
2460                     }
2461                 }
2462
2463         /*
2464          * We also shouldn't be the target of any PMC at this index
2465          * since otherwise a PMC_ATTACH to ourselves will fail.
2466          */
2467         if ((pp = pmc_find_process_descriptor(p, 0)) != NULL)
2468                 if (pp->pp_pmcs[ri].pp_pmc)
2469                         return EEXIST;
2470
2471         PMCDBG4(PMC,ALR,2, "can-allocate-rowindex proc=%p (%d, %s) ri=%d ok",
2472             p, p->p_pid, p->p_comm, ri);
2473
2474         return 0;
2475 }
2476
2477 /*
2478  * Check if a given PMC at row index 'ri' can be currently used in
2479  * mode 'mode'.
2480  */
2481
2482 static int
2483 pmc_can_allocate_row(int ri, enum pmc_mode mode)
2484 {
2485         enum pmc_disp   disp;
2486
2487         sx_assert(&pmc_sx, SX_XLOCKED);
2488
2489         PMCDBG2(PMC,ALR,1, "can-allocate-row ri=%d mode=%d", ri, mode);
2490
2491         if (PMC_IS_SYSTEM_MODE(mode))
2492                 disp = PMC_DISP_STANDALONE;
2493         else
2494                 disp = PMC_DISP_THREAD;
2495
2496         /*
2497          * check disposition for PMC row 'ri':
2498          *
2499          * Expected disposition         Row-disposition         Result
2500          *
2501          * STANDALONE                   STANDALONE or FREE      proceed
2502          * STANDALONE                   THREAD                  fail
2503          * THREAD                       THREAD or FREE          proceed
2504          * THREAD                       STANDALONE              fail
2505          */
2506
2507         if (!PMC_ROW_DISP_IS_FREE(ri) &&
2508             !(disp == PMC_DISP_THREAD && PMC_ROW_DISP_IS_THREAD(ri)) &&
2509             !(disp == PMC_DISP_STANDALONE && PMC_ROW_DISP_IS_STANDALONE(ri)))
2510                 return EBUSY;
2511
2512         /*
2513          * All OK
2514          */
2515
2516         PMCDBG2(PMC,ALR,2, "can-allocate-row ri=%d mode=%d ok", ri, mode);
2517
2518         return 0;
2519
2520 }
2521
2522 /*
2523  * Find a PMC descriptor with user handle 'pmcid' for thread 'td'.
2524  */
2525
2526 static struct pmc *
2527 pmc_find_pmc_descriptor_in_process(struct pmc_owner *po, pmc_id_t pmcid)
2528 {
2529         struct pmc *pm;
2530
2531         KASSERT(PMC_ID_TO_ROWINDEX(pmcid) < md->pmd_npmc,
2532             ("[pmc,%d] Illegal pmc index %d (max %d)", __LINE__,
2533                 PMC_ID_TO_ROWINDEX(pmcid), md->pmd_npmc));
2534
2535         LIST_FOREACH(pm, &po->po_pmcs, pm_next)
2536             if (pm->pm_id == pmcid)
2537                     return pm;
2538
2539         return NULL;
2540 }
2541
2542 static int
2543 pmc_find_pmc(pmc_id_t pmcid, struct pmc **pmc)
2544 {
2545
2546         struct pmc *pm;
2547         struct pmc_owner *po;
2548
2549         PMCDBG1(PMC,FND,1, "find-pmc id=%d", pmcid);
2550
2551         if ((po = pmc_find_owner_descriptor(curthread->td_proc)) == NULL)
2552                 return ESRCH;
2553
2554         if ((pm = pmc_find_pmc_descriptor_in_process(po, pmcid)) == NULL)
2555                 return EINVAL;
2556
2557         PMCDBG2(PMC,FND,2, "find-pmc id=%d -> pmc=%p", pmcid, pm);
2558
2559         *pmc = pm;
2560         return 0;
2561 }
2562
2563 /*
2564  * Start a PMC.
2565  */
2566
2567 static int
2568 pmc_start(struct pmc *pm)
2569 {
2570         enum pmc_mode mode;
2571         struct pmc_owner *po;
2572         struct pmc_binding pb;
2573         struct pmc_classdep *pcd;
2574         int adjri, error, cpu, ri;
2575
2576         KASSERT(pm != NULL,
2577             ("[pmc,%d] null pm", __LINE__));
2578
2579         mode = PMC_TO_MODE(pm);
2580         ri   = PMC_TO_ROWINDEX(pm);
2581         pcd  = pmc_ri_to_classdep(md, ri, &adjri);
2582
2583         error = 0;
2584
2585         PMCDBG3(PMC,OPS,1, "start pmc=%p mode=%d ri=%d", pm, mode, ri);
2586
2587         po = pm->pm_owner;
2588
2589         /*
2590          * Disallow PMCSTART if a logfile is required but has not been
2591          * configured yet.
2592          */
2593         if ((pm->pm_flags & PMC_F_NEEDS_LOGFILE) &&
2594             (po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
2595                 return (EDOOFUS);       /* programming error */
2596
2597         /*
2598          * If this is a sampling mode PMC, log mapping information for
2599          * the kernel modules that are currently loaded.
2600          */
2601         if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
2602             pmc_log_kernel_mappings(pm);
2603
2604         if (PMC_IS_VIRTUAL_MODE(mode)) {
2605
2606                 /*
2607                  * If a PMCATTACH has never been done on this PMC,
2608                  * attach it to its owner process.
2609                  */
2610
2611                 if (LIST_EMPTY(&pm->pm_targets))
2612                         error = (pm->pm_flags & PMC_F_ATTACH_DONE) ? ESRCH :
2613                             pmc_attach_process(po->po_owner, pm);
2614
2615                 /*
2616                  * If the PMC is attached to its owner, then force a context
2617                  * switch to ensure that the MD state gets set correctly.
2618                  */
2619
2620                 if (error == 0) {
2621                         pm->pm_state = PMC_STATE_RUNNING;
2622                         if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER)
2623                                 pmc_force_context_switch();
2624                 }
2625
2626                 return (error);
2627         }
2628
2629
2630         /*
2631          * A system-wide PMC.
2632          *
2633          * Add the owner to the global list if this is a system-wide
2634          * sampling PMC.
2635          */
2636
2637         if (mode == PMC_MODE_SS) {
2638                 if (po->po_sscount == 0) {
2639                         LIST_INSERT_HEAD(&pmc_ss_owners, po, po_ssnext);
2640                         atomic_add_rel_int(&pmc_ss_count, 1);
2641                         PMCDBG1(PMC,OPS,1, "po=%p in global list", po);
2642                 }
2643                 po->po_sscount++;
2644
2645                 /*
2646                  * Log mapping information for all existing processes in the
2647                  * system.  Subsequent mappings are logged as they happen;
2648                  * see pmc_process_mmap().
2649                  */
2650                 if (po->po_logprocmaps == 0) {
2651                         pmc_log_all_process_mappings(po);
2652                         po->po_logprocmaps = 1;
2653                 }
2654         }
2655
2656         /*
2657          * Move to the CPU associated with this
2658          * PMC, and start the hardware.
2659          */
2660
2661         pmc_save_cpu_binding(&pb);
2662
2663         cpu = PMC_TO_CPU(pm);
2664
2665         if (!pmc_cpu_is_active(cpu))
2666                 return (ENXIO);
2667
2668         pmc_select_cpu(cpu);
2669
2670         /*
2671          * global PMCs are configured at allocation time
2672          * so write out the initial value and start the PMC.
2673          */
2674
2675         pm->pm_state = PMC_STATE_RUNNING;
2676
2677         critical_enter();
2678         if ((error = pcd->pcd_write_pmc(cpu, adjri,
2679                  PMC_IS_SAMPLING_MODE(mode) ?
2680                  pm->pm_sc.pm_reloadcount :
2681                  pm->pm_sc.pm_initial)) == 0)
2682                 error = pcd->pcd_start_pmc(cpu, adjri);
2683         critical_exit();
2684
2685         pmc_restore_cpu_binding(&pb);
2686
2687         return (error);
2688 }
2689
2690 /*
2691  * Stop a PMC.
2692  */
2693
2694 static int
2695 pmc_stop(struct pmc *pm)
2696 {
2697         struct pmc_owner *po;
2698         struct pmc_binding pb;
2699         struct pmc_classdep *pcd;
2700         int adjri, cpu, error, ri;
2701
2702         KASSERT(pm != NULL, ("[pmc,%d] null pmc", __LINE__));
2703
2704         PMCDBG3(PMC,OPS,1, "stop pmc=%p mode=%d ri=%d", pm,
2705             PMC_TO_MODE(pm), PMC_TO_ROWINDEX(pm));
2706
2707         pm->pm_state = PMC_STATE_STOPPED;
2708
2709         /*
2710          * If the PMC is a virtual mode one, changing the state to
2711          * non-RUNNING is enough to ensure that the PMC never gets
2712          * scheduled.
2713          *
2714          * If this PMC is current running on a CPU, then it will
2715          * handled correctly at the time its target process is context
2716          * switched out.
2717          */
2718
2719         if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)))
2720                 return 0;
2721
2722         /*
2723          * A system-mode PMC.  Move to the CPU associated with
2724          * this PMC, and stop the hardware.  We update the
2725          * 'initial count' so that a subsequent PMCSTART will
2726          * resume counting from the current hardware count.
2727          */
2728
2729         pmc_save_cpu_binding(&pb);
2730
2731         cpu = PMC_TO_CPU(pm);
2732
2733         KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
2734             ("[pmc,%d] illegal cpu=%d", __LINE__, cpu));
2735
2736         if (!pmc_cpu_is_active(cpu))
2737                 return ENXIO;
2738
2739         pmc_select_cpu(cpu);
2740
2741         ri = PMC_TO_ROWINDEX(pm);
2742         pcd = pmc_ri_to_classdep(md, ri, &adjri);
2743
2744         critical_enter();
2745         if ((error = pcd->pcd_stop_pmc(cpu, adjri)) == 0)
2746                 error = pcd->pcd_read_pmc(cpu, adjri, &pm->pm_sc.pm_initial);
2747         critical_exit();
2748
2749         pmc_restore_cpu_binding(&pb);
2750
2751         po = pm->pm_owner;
2752
2753         /* remove this owner from the global list of SS PMC owners */
2754         if (PMC_TO_MODE(pm) == PMC_MODE_SS) {
2755                 po->po_sscount--;
2756                 if (po->po_sscount == 0) {
2757                         atomic_subtract_rel_int(&pmc_ss_count, 1);
2758                         LIST_REMOVE(po, po_ssnext);
2759                         PMCDBG1(PMC,OPS,2,"po=%p removed from global list", po);
2760                 }
2761         }
2762
2763         return (error);
2764 }
2765
2766
2767 #ifdef  HWPMC_DEBUG
2768 static const char *pmc_op_to_name[] = {
2769 #undef  __PMC_OP
2770 #define __PMC_OP(N, D)  #N ,
2771         __PMC_OPS()
2772         NULL
2773 };
2774 #endif
2775
2776 /*
2777  * The syscall interface
2778  */
2779
2780 #define PMC_GET_SX_XLOCK(...) do {              \
2781         sx_xlock(&pmc_sx);                      \
2782         if (pmc_hook == NULL) {                 \
2783                 sx_xunlock(&pmc_sx);            \
2784                 return __VA_ARGS__;             \
2785         }                                       \
2786 } while (0)
2787
2788 #define PMC_DOWNGRADE_SX() do {                 \
2789         sx_downgrade(&pmc_sx);                  \
2790         is_sx_downgraded = 1;                   \
2791 } while (0)
2792
2793 static int
2794 pmc_syscall_handler(struct thread *td, void *syscall_args)
2795 {
2796         int error, is_sx_downgraded, is_sx_locked, op;
2797         struct pmc_syscall_args *c;
2798         void *arg;
2799
2800         PMC_GET_SX_XLOCK(ENOSYS);
2801
2802         DROP_GIANT();
2803
2804         is_sx_downgraded = 0;
2805         is_sx_locked = 1;
2806
2807         c = (struct pmc_syscall_args *) syscall_args;
2808
2809         op = c->pmop_code;
2810         arg = c->pmop_data;
2811
2812         PMCDBG3(MOD,PMS,1, "syscall op=%d \"%s\" arg=%p", op,
2813             pmc_op_to_name[op], arg);
2814
2815         error = 0;
2816         atomic_add_int(&pmc_stats.pm_syscalls, 1);
2817
2818         switch(op)
2819         {
2820
2821
2822         /*
2823          * Configure a log file.
2824          *
2825          * XXX This OP will be reworked.
2826          */
2827
2828         case PMC_OP_CONFIGURELOG:
2829         {
2830                 struct proc *p;
2831                 struct pmc *pm;
2832                 struct pmc_owner *po;
2833                 struct pmc_op_configurelog cl;
2834
2835                 sx_assert(&pmc_sx, SX_XLOCKED);
2836
2837                 if ((error = copyin(arg, &cl, sizeof(cl))) != 0)
2838                         break;
2839
2840                 /* mark this process as owning a log file */
2841                 p = td->td_proc;
2842                 if ((po = pmc_find_owner_descriptor(p)) == NULL)
2843                         if ((po = pmc_allocate_owner_descriptor(p)) == NULL) {
2844                                 error = ENOMEM;
2845                                 break;
2846                         }
2847
2848                 /*
2849                  * If a valid fd was passed in, try to configure that,
2850                  * otherwise if 'fd' was less than zero and there was
2851                  * a log file configured, flush its buffers and
2852                  * de-configure it.
2853                  */
2854                 if (cl.pm_logfd >= 0) {
2855                         sx_xunlock(&pmc_sx);
2856                         is_sx_locked = 0;
2857                         error = pmclog_configure_log(md, po, cl.pm_logfd);
2858                 } else if (po->po_flags & PMC_PO_OWNS_LOGFILE) {
2859                         pmclog_process_closelog(po);
2860                         error = pmclog_close(po);
2861                         if (error == 0) {
2862                                 LIST_FOREACH(pm, &po->po_pmcs, pm_next)
2863                                     if (pm->pm_flags & PMC_F_NEEDS_LOGFILE &&
2864                                         pm->pm_state == PMC_STATE_RUNNING)
2865                                             pmc_stop(pm);
2866                                 error = pmclog_deconfigure_log(po);
2867                         }
2868                 } else
2869                         error = EINVAL;
2870
2871                 if (error)
2872                         break;
2873         }
2874         break;
2875
2876         /*
2877          * Flush a log file.
2878          */
2879
2880         case PMC_OP_FLUSHLOG:
2881         {
2882                 struct pmc_owner *po;
2883
2884                 sx_assert(&pmc_sx, SX_XLOCKED);
2885
2886                 if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) {
2887                         error = EINVAL;
2888                         break;
2889                 }
2890
2891                 error = pmclog_flush(po);
2892         }
2893         break;
2894
2895         /*
2896          * Close a log file.
2897          */
2898
2899         case PMC_OP_CLOSELOG:
2900         {
2901                 struct pmc_owner *po;
2902
2903                 sx_assert(&pmc_sx, SX_XLOCKED);
2904
2905                 if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) {
2906                         error = EINVAL;
2907                         break;
2908                 }
2909
2910                 error = pmclog_close(po);
2911         }
2912         break;
2913
2914         /*
2915          * Retrieve hardware configuration.
2916          */
2917
2918         case PMC_OP_GETCPUINFO: /* CPU information */
2919         {
2920                 struct pmc_op_getcpuinfo gci;
2921                 struct pmc_classinfo *pci;
2922                 struct pmc_classdep *pcd;
2923                 int cl;
2924
2925                 gci.pm_cputype = md->pmd_cputype;
2926                 gci.pm_ncpu    = pmc_cpu_max();
2927                 gci.pm_npmc    = md->pmd_npmc;
2928                 gci.pm_nclass  = md->pmd_nclass;
2929                 pci = gci.pm_classes;
2930                 pcd = md->pmd_classdep;
2931                 for (cl = 0; cl < md->pmd_nclass; cl++, pci++, pcd++) {
2932                         pci->pm_caps  = pcd->pcd_caps;
2933                         pci->pm_class = pcd->pcd_class;
2934                         pci->pm_width = pcd->pcd_width;
2935                         pci->pm_num   = pcd->pcd_num;
2936                 }
2937                 error = copyout(&gci, arg, sizeof(gci));
2938         }
2939         break;
2940
2941         /*
2942          * Retrieve soft events list.
2943          */
2944         case PMC_OP_GETDYNEVENTINFO:
2945         {
2946                 enum pmc_class                  cl;
2947                 enum pmc_event                  ev;
2948                 struct pmc_op_getdyneventinfo   *gei;
2949                 struct pmc_dyn_event_descr      dev;
2950                 struct pmc_soft                 *ps;
2951                 uint32_t                        nevent;
2952
2953                 sx_assert(&pmc_sx, SX_LOCKED);
2954
2955                 gei = (struct pmc_op_getdyneventinfo *) arg;
2956
2957                 if ((error = copyin(&gei->pm_class, &cl, sizeof(cl))) != 0)
2958                         break;
2959
2960                 /* Only SOFT class is dynamic. */
2961                 if (cl != PMC_CLASS_SOFT) {
2962                         error = EINVAL;
2963                         break;
2964                 }
2965
2966                 nevent = 0;
2967                 for (ev = PMC_EV_SOFT_FIRST; (int)ev <= PMC_EV_SOFT_LAST; ev++) {
2968                         ps = pmc_soft_ev_acquire(ev);
2969                         if (ps == NULL)
2970                                 continue;
2971                         bcopy(&ps->ps_ev, &dev, sizeof(dev));
2972                         pmc_soft_ev_release(ps);
2973
2974                         error = copyout(&dev,
2975                             &gei->pm_events[nevent],
2976                             sizeof(struct pmc_dyn_event_descr));
2977                         if (error != 0)
2978                                 break;
2979                         nevent++;
2980                 }
2981                 if (error != 0)
2982                         break;
2983
2984                 error = copyout(&nevent, &gei->pm_nevent,
2985                     sizeof(nevent));
2986         }
2987         break;
2988
2989         /*
2990          * Get module statistics
2991          */
2992
2993         case PMC_OP_GETDRIVERSTATS:
2994         {
2995                 struct pmc_op_getdriverstats gms;
2996
2997                 bcopy(&pmc_stats, &gms, sizeof(gms));
2998                 error = copyout(&gms, arg, sizeof(gms));
2999         }
3000         break;
3001
3002
3003         /*
3004          * Retrieve module version number
3005          */
3006
3007         case PMC_OP_GETMODULEVERSION:
3008         {
3009                 uint32_t cv, modv;
3010
3011                 /* retrieve the client's idea of the ABI version */
3012                 if ((error = copyin(arg, &cv, sizeof(uint32_t))) != 0)
3013                         break;
3014                 /* don't service clients newer than our driver */
3015                 modv = PMC_VERSION;
3016                 if ((cv & 0xFFFF0000) > (modv & 0xFFFF0000)) {
3017                         error = EPROGMISMATCH;
3018                         break;
3019                 }
3020                 error = copyout(&modv, arg, sizeof(int));
3021         }
3022         break;
3023
3024
3025         /*
3026          * Retrieve the state of all the PMCs on a given
3027          * CPU.
3028          */
3029
3030         case PMC_OP_GETPMCINFO:
3031         {
3032                 int ari;
3033                 struct pmc *pm;
3034                 size_t pmcinfo_size;
3035                 uint32_t cpu, n, npmc;
3036                 struct pmc_owner *po;
3037                 struct pmc_binding pb;
3038                 struct pmc_classdep *pcd;
3039                 struct pmc_info *p, *pmcinfo;
3040                 struct pmc_op_getpmcinfo *gpi;
3041
3042                 PMC_DOWNGRADE_SX();
3043
3044                 gpi = (struct pmc_op_getpmcinfo *) arg;
3045
3046                 if ((error = copyin(&gpi->pm_cpu, &cpu, sizeof(cpu))) != 0)
3047                         break;
3048
3049                 if (cpu >= pmc_cpu_max()) {
3050                         error = EINVAL;
3051                         break;
3052                 }
3053
3054                 if (!pmc_cpu_is_active(cpu)) {
3055                         error = ENXIO;
3056                         break;
3057                 }
3058
3059                 /* switch to CPU 'cpu' */
3060                 pmc_save_cpu_binding(&pb);
3061                 pmc_select_cpu(cpu);
3062
3063                 npmc = md->pmd_npmc;
3064
3065                 pmcinfo_size = npmc * sizeof(struct pmc_info);
3066                 pmcinfo = malloc(pmcinfo_size, M_PMC, M_WAITOK);
3067
3068                 p = pmcinfo;
3069
3070                 for (n = 0; n < md->pmd_npmc; n++, p++) {
3071
3072                         pcd = pmc_ri_to_classdep(md, n, &ari);
3073
3074                         KASSERT(pcd != NULL,
3075                             ("[pmc,%d] null pcd ri=%d", __LINE__, n));
3076
3077                         if ((error = pcd->pcd_describe(cpu, ari, p, &pm)) != 0)
3078                                 break;
3079
3080                         if (PMC_ROW_DISP_IS_STANDALONE(n))
3081                                 p->pm_rowdisp = PMC_DISP_STANDALONE;
3082                         else if (PMC_ROW_DISP_IS_THREAD(n))
3083                                 p->pm_rowdisp = PMC_DISP_THREAD;
3084                         else
3085                                 p->pm_rowdisp = PMC_DISP_FREE;
3086
3087                         p->pm_ownerpid = -1;
3088
3089                         if (pm == NULL) /* no PMC associated */
3090                                 continue;
3091
3092                         po = pm->pm_owner;
3093
3094                         KASSERT(po->po_owner != NULL,
3095                             ("[pmc,%d] pmc_owner had a null proc pointer",
3096                                 __LINE__));
3097
3098                         p->pm_ownerpid = po->po_owner->p_pid;
3099                         p->pm_mode     = PMC_TO_MODE(pm);
3100                         p->pm_event    = pm->pm_event;
3101                         p->pm_flags    = pm->pm_flags;
3102
3103                         if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
3104                                 p->pm_reloadcount =
3105                                     pm->pm_sc.pm_reloadcount;
3106                 }
3107
3108                 pmc_restore_cpu_binding(&pb);
3109
3110                 /* now copy out the PMC info collected */
3111                 if (error == 0)
3112                         error = copyout(pmcinfo, &gpi->pm_pmcs, pmcinfo_size);
3113
3114                 free(pmcinfo, M_PMC);
3115         }
3116         break;
3117
3118
3119         /*
3120          * Set the administrative state of a PMC.  I.e. whether
3121          * the PMC is to be used or not.
3122          */
3123
3124         case PMC_OP_PMCADMIN:
3125         {
3126                 int cpu, ri;
3127                 enum pmc_state request;
3128                 struct pmc_cpu *pc;
3129                 struct pmc_hw *phw;
3130                 struct pmc_op_pmcadmin pma;
3131                 struct pmc_binding pb;
3132
3133                 sx_assert(&pmc_sx, SX_XLOCKED);
3134
3135                 KASSERT(td == curthread,
3136                     ("[pmc,%d] td != curthread", __LINE__));
3137
3138                 error = priv_check(td, PRIV_PMC_MANAGE);
3139                 if (error)
3140                         break;
3141
3142                 if ((error = copyin(arg, &pma, sizeof(pma))) != 0)
3143                         break;
3144
3145                 cpu = pma.pm_cpu;
3146
3147                 if (cpu < 0 || cpu >= (int) pmc_cpu_max()) {
3148                         error = EINVAL;
3149                         break;
3150                 }
3151
3152                 if (!pmc_cpu_is_active(cpu)) {
3153                         error = ENXIO;
3154                         break;
3155                 }
3156
3157                 request = pma.pm_state;
3158
3159                 if (request != PMC_STATE_DISABLED &&
3160                     request != PMC_STATE_FREE) {
3161                         error = EINVAL;
3162                         break;
3163                 }
3164
3165                 ri = pma.pm_pmc; /* pmc id == row index */
3166                 if (ri < 0 || ri >= (int) md->pmd_npmc) {
3167                         error = EINVAL;
3168                         break;
3169                 }
3170
3171                 /*
3172                  * We can't disable a PMC with a row-index allocated
3173                  * for process virtual PMCs.
3174                  */
3175
3176                 if (PMC_ROW_DISP_IS_THREAD(ri) &&
3177                     request == PMC_STATE_DISABLED) {
3178                         error = EBUSY;
3179                         break;
3180                 }
3181
3182                 /*
3183                  * otherwise, this PMC on this CPU is either free or
3184                  * in system-wide mode.
3185                  */
3186
3187                 pmc_save_cpu_binding(&pb);
3188                 pmc_select_cpu(cpu);
3189
3190                 pc  = pmc_pcpu[cpu];
3191                 phw = pc->pc_hwpmcs[ri];
3192
3193                 /*
3194                  * XXX do we need some kind of 'forced' disable?
3195                  */
3196
3197                 if (phw->phw_pmc == NULL) {
3198                         if (request == PMC_STATE_DISABLED &&
3199                             (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED)) {
3200                                 phw->phw_state &= ~PMC_PHW_FLAG_IS_ENABLED;
3201                                 PMC_MARK_ROW_STANDALONE(ri);
3202                         } else if (request == PMC_STATE_FREE &&
3203                             (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0) {
3204                                 phw->phw_state |=  PMC_PHW_FLAG_IS_ENABLED;
3205                                 PMC_UNMARK_ROW_STANDALONE(ri);
3206                         }
3207                         /* other cases are a no-op */
3208                 } else
3209                         error = EBUSY;
3210
3211                 pmc_restore_cpu_binding(&pb);
3212         }
3213         break;
3214
3215
3216         /*
3217          * Allocate a PMC.
3218          */
3219
3220         case PMC_OP_PMCALLOCATE:
3221         {
3222                 int adjri, n;
3223                 u_int cpu;
3224                 uint32_t caps;
3225                 struct pmc *pmc;
3226                 enum pmc_mode mode;
3227                 struct pmc_hw *phw;
3228                 struct pmc_binding pb;
3229                 struct pmc_classdep *pcd;
3230                 struct pmc_op_pmcallocate pa;
3231
3232                 if ((error = copyin(arg, &pa, sizeof(pa))) != 0)
3233                         break;
3234
3235                 caps = pa.pm_caps;
3236                 mode = pa.pm_mode;
3237                 cpu  = pa.pm_cpu;
3238
3239                 if ((mode != PMC_MODE_SS  &&  mode != PMC_MODE_SC  &&
3240                      mode != PMC_MODE_TS  &&  mode != PMC_MODE_TC) ||
3241                     (cpu != (u_int) PMC_CPU_ANY && cpu >= pmc_cpu_max())) {
3242                         error = EINVAL;
3243                         break;
3244                 }
3245
3246                 /*
3247                  * Virtual PMCs should only ask for a default CPU.
3248                  * System mode PMCs need to specify a non-default CPU.
3249                  */
3250
3251                 if ((PMC_IS_VIRTUAL_MODE(mode) && cpu != (u_int) PMC_CPU_ANY) ||
3252                     (PMC_IS_SYSTEM_MODE(mode) && cpu == (u_int) PMC_CPU_ANY)) {
3253                         error = EINVAL;
3254                         break;
3255                 }
3256
3257                 /*
3258                  * Check that an inactive CPU is not being asked for.
3259                  */
3260
3261                 if (PMC_IS_SYSTEM_MODE(mode) && !pmc_cpu_is_active(cpu)) {
3262                         error = ENXIO;
3263                         break;
3264                 }
3265
3266                 /*
3267                  * Refuse an allocation for a system-wide PMC if this
3268                  * process has been jailed, or if this process lacks
3269                  * super-user credentials and the sysctl tunable
3270                  * 'security.bsd.unprivileged_syspmcs' is zero.
3271                  */
3272
3273                 if (PMC_IS_SYSTEM_MODE(mode)) {
3274                         if (jailed(curthread->td_ucred)) {
3275                                 error = EPERM;
3276                                 break;
3277                         }
3278                         if (!pmc_unprivileged_syspmcs) {
3279                                 error = priv_check(curthread,
3280                                     PRIV_PMC_SYSTEM);
3281                                 if (error)
3282                                         break;
3283                         }
3284                 }
3285
3286                 /*
3287                  * Look for valid values for 'pm_flags'
3288                  */
3289
3290                 if ((pa.pm_flags & ~(PMC_F_DESCENDANTS | PMC_F_LOG_PROCCSW |
3291                     PMC_F_LOG_PROCEXIT | PMC_F_CALLCHAIN)) != 0) {
3292                         error = EINVAL;
3293                         break;
3294                 }
3295
3296                 /* process logging options are not allowed for system PMCs */
3297                 if (PMC_IS_SYSTEM_MODE(mode) && (pa.pm_flags &
3298                     (PMC_F_LOG_PROCCSW | PMC_F_LOG_PROCEXIT))) {
3299                         error = EINVAL;
3300                         break;
3301                 }
3302
3303                 /*
3304                  * All sampling mode PMCs need to be able to interrupt the
3305                  * CPU.
3306                  */
3307                 if (PMC_IS_SAMPLING_MODE(mode))
3308                         caps |= PMC_CAP_INTERRUPT;
3309
3310                 /* A valid class specifier should have been passed in. */
3311                 for (n = 0; n < md->pmd_nclass; n++)
3312                         if (md->pmd_classdep[n].pcd_class == pa.pm_class)
3313                                 break;
3314                 if (n == md->pmd_nclass) {
3315                         error = EINVAL;
3316                         break;
3317                 }
3318
3319                 /* The requested PMC capabilities should be feasible. */
3320                 if ((md->pmd_classdep[n].pcd_caps & caps) != caps) {
3321                         error = EOPNOTSUPP;
3322                         break;
3323                 }
3324
3325                 PMCDBG4(PMC,ALL,2, "event=%d caps=0x%x mode=%d cpu=%d",
3326                     pa.pm_ev, caps, mode, cpu);
3327
3328                 pmc = pmc_allocate_pmc_descriptor();
3329                 pmc->pm_id    = PMC_ID_MAKE_ID(cpu,pa.pm_mode,pa.pm_class,
3330                     PMC_ID_INVALID);
3331                 pmc->pm_event = pa.pm_ev;
3332                 pmc->pm_state = PMC_STATE_FREE;
3333                 pmc->pm_caps  = caps;
3334                 pmc->pm_flags = pa.pm_flags;
3335
3336                 /* switch thread to CPU 'cpu' */
3337                 pmc_save_cpu_binding(&pb);
3338
3339 #define PMC_IS_SHAREABLE_PMC(cpu, n)                            \
3340         (pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_state &           \
3341          PMC_PHW_FLAG_IS_SHAREABLE)
3342 #define PMC_IS_UNALLOCATED(cpu, n)                              \
3343         (pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_pmc == NULL)
3344
3345                 if (PMC_IS_SYSTEM_MODE(mode)) {
3346                         pmc_select_cpu(cpu);
3347                         for (n = 0; n < (int) md->pmd_npmc; n++) {
3348                                 pcd = pmc_ri_to_classdep(md, n, &adjri);
3349                                 if (pmc_can_allocate_row(n, mode) == 0 &&
3350                                     pmc_can_allocate_rowindex(
3351                                             curthread->td_proc, n, cpu) == 0 &&
3352                                     (PMC_IS_UNALLOCATED(cpu, n) ||
3353                                      PMC_IS_SHAREABLE_PMC(cpu, n)) &&
3354                                     pcd->pcd_allocate_pmc(cpu, adjri, pmc,
3355                                         &pa) == 0)
3356                                         break;
3357                         }
3358                 } else {
3359                         /* Process virtual mode */
3360                         for (n = 0; n < (int) md->pmd_npmc; n++) {
3361                                 pcd = pmc_ri_to_classdep(md, n, &adjri);
3362                                 if (pmc_can_allocate_row(n, mode) == 0 &&
3363                                     pmc_can_allocate_rowindex(
3364                                             curthread->td_proc, n,
3365                                             PMC_CPU_ANY) == 0 &&
3366                                     pcd->pcd_allocate_pmc(curthread->td_oncpu,
3367                                         adjri, pmc, &pa) == 0)
3368                                         break;
3369                         }
3370                 }
3371
3372 #undef  PMC_IS_UNALLOCATED
3373 #undef  PMC_IS_SHAREABLE_PMC
3374
3375                 pmc_restore_cpu_binding(&pb);
3376
3377                 if (n == (int) md->pmd_npmc) {
3378                         pmc_destroy_pmc_descriptor(pmc);
3379                         pmc = NULL;
3380                         error = EINVAL;
3381                         break;
3382                 }
3383
3384                 /* Fill in the correct value in the ID field */
3385                 pmc->pm_id = PMC_ID_MAKE_ID(cpu,mode,pa.pm_class,n);
3386
3387                 PMCDBG5(PMC,ALL,2, "ev=%d class=%d mode=%d n=%d -> pmcid=%x",
3388                     pmc->pm_event, pa.pm_class, mode, n, pmc->pm_id);
3389
3390                 /* Process mode PMCs with logging enabled need log files */
3391                 if (pmc->pm_flags & (PMC_F_LOG_PROCEXIT | PMC_F_LOG_PROCCSW))
3392                         pmc->pm_flags |= PMC_F_NEEDS_LOGFILE;
3393
3394                 /* All system mode sampling PMCs require a log file */
3395                 if (PMC_IS_SAMPLING_MODE(mode) && PMC_IS_SYSTEM_MODE(mode))
3396                         pmc->pm_flags |= PMC_F_NEEDS_LOGFILE;
3397
3398                 /*
3399                  * Configure global pmc's immediately
3400                  */
3401
3402                 if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pmc))) {
3403
3404                         pmc_save_cpu_binding(&pb);
3405                         pmc_select_cpu(cpu);
3406
3407                         phw = pmc_pcpu[cpu]->pc_hwpmcs[n];
3408                         pcd = pmc_ri_to_classdep(md, n, &adjri);
3409
3410                         if ((phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0 ||
3411                             (error = pcd->pcd_config_pmc(cpu, adjri, pmc)) != 0) {
3412                                 (void) pcd->pcd_release_pmc(cpu, adjri, pmc);
3413                                 pmc_destroy_pmc_descriptor(pmc);
3414                                 pmc = NULL;
3415                                 pmc_restore_cpu_binding(&pb);
3416                                 error = EPERM;
3417                                 break;
3418                         }
3419
3420                         pmc_restore_cpu_binding(&pb);
3421                 }
3422
3423                 pmc->pm_state    = PMC_STATE_ALLOCATED;
3424
3425                 /*
3426                  * mark row disposition
3427                  */
3428
3429                 if (PMC_IS_SYSTEM_MODE(mode))
3430                         PMC_MARK_ROW_STANDALONE(n);
3431                 else
3432                         PMC_MARK_ROW_THREAD(n);
3433
3434                 /*
3435                  * Register this PMC with the current thread as its owner.
3436                  */
3437
3438                 if ((error =
3439                     pmc_register_owner(curthread->td_proc, pmc)) != 0) {
3440                         pmc_release_pmc_descriptor(pmc);
3441                         pmc_destroy_pmc_descriptor(pmc);
3442                         pmc = NULL;
3443                         break;
3444                 }
3445
3446                 /*
3447                  * Return the allocated index.
3448                  */
3449
3450                 pa.pm_pmcid = pmc->pm_id;
3451
3452                 error = copyout(&pa, arg, sizeof(pa));
3453         }
3454         break;
3455
3456
3457         /*
3458          * Attach a PMC to a process.
3459          */
3460
3461         case PMC_OP_PMCATTACH:
3462         {
3463                 struct pmc *pm;
3464                 struct proc *p;
3465                 struct pmc_op_pmcattach a;
3466
3467                 sx_assert(&pmc_sx, SX_XLOCKED);
3468
3469                 if ((error = copyin(arg, &a, sizeof(a))) != 0)
3470                         break;
3471
3472                 if (a.pm_pid < 0) {
3473                         error = EINVAL;
3474                         break;
3475                 } else if (a.pm_pid == 0)
3476                         a.pm_pid = td->td_proc->p_pid;
3477
3478                 if ((error = pmc_find_pmc(a.pm_pmc, &pm)) != 0)
3479                         break;
3480
3481                 if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm))) {
3482                         error = EINVAL;
3483                         break;
3484                 }
3485
3486                 /* PMCs may be (re)attached only when allocated or stopped */
3487                 if (pm->pm_state == PMC_STATE_RUNNING) {
3488                         error = EBUSY;
3489                         break;
3490                 } else if (pm->pm_state != PMC_STATE_ALLOCATED &&
3491                     pm->pm_state != PMC_STATE_STOPPED) {
3492                         error = EINVAL;
3493                         break;
3494                 }
3495
3496                 /* lookup pid */
3497                 if ((p = pfind(a.pm_pid)) == NULL) {
3498                         error = ESRCH;
3499                         break;
3500                 }
3501
3502                 /*
3503                  * Ignore processes that are working on exiting.
3504                  */
3505                 if (p->p_flag & P_WEXIT) {
3506                         error = ESRCH;
3507                         PROC_UNLOCK(p); /* pfind() returns a locked process */
3508                         break;
3509                 }
3510
3511                 /*
3512                  * we are allowed to attach a PMC to a process if
3513                  * we can debug it.
3514                  */
3515                 error = p_candebug(curthread, p);
3516
3517                 PROC_UNLOCK(p);
3518
3519                 if (error == 0)
3520                         error = pmc_attach_process(p, pm);
3521         }
3522         break;
3523
3524
3525         /*
3526          * Detach an attached PMC from a process.
3527          */
3528
3529         case PMC_OP_PMCDETACH:
3530         {
3531                 struct pmc *pm;
3532                 struct proc *p;
3533                 struct pmc_op_pmcattach a;
3534
3535                 if ((error = copyin(arg, &a, sizeof(a))) != 0)
3536                         break;
3537
3538                 if (a.pm_pid < 0) {
3539                         error = EINVAL;
3540                         break;
3541                 } else if (a.pm_pid == 0)
3542                         a.pm_pid = td->td_proc->p_pid;
3543
3544                 if ((error = pmc_find_pmc(a.pm_pmc, &pm)) != 0)
3545                         break;
3546
3547                 if ((p = pfind(a.pm_pid)) == NULL) {
3548                         error = ESRCH;
3549                         break;
3550                 }
3551
3552                 /*
3553                  * Treat processes that are in the process of exiting
3554                  * as if they were not present.
3555                  */
3556
3557                 if (p->p_flag & P_WEXIT)
3558                         error = ESRCH;
3559
3560                 PROC_UNLOCK(p); /* pfind() returns a locked process */
3561
3562                 if (error == 0)
3563                         error = pmc_detach_process(p, pm);
3564         }
3565         break;
3566
3567
3568         /*
3569          * Retrieve the MSR number associated with the counter
3570          * 'pmc_id'.  This allows processes to directly use RDPMC
3571          * instructions to read their PMCs, without the overhead of a
3572          * system call.
3573          */
3574
3575         case PMC_OP_PMCGETMSR:
3576         {
3577                 int adjri, ri;
3578                 struct pmc *pm;
3579                 struct pmc_target *pt;
3580                 struct pmc_op_getmsr gm;
3581                 struct pmc_classdep *pcd;
3582
3583                 PMC_DOWNGRADE_SX();
3584
3585                 if ((error = copyin(arg, &gm, sizeof(gm))) != 0)
3586                         break;
3587
3588                 if ((error = pmc_find_pmc(gm.pm_pmcid, &pm)) != 0)
3589                         break;
3590
3591                 /*
3592                  * The allocated PMC has to be a process virtual PMC,
3593                  * i.e., of type MODE_T[CS].  Global PMCs can only be
3594                  * read using the PMCREAD operation since they may be
3595                  * allocated on a different CPU than the one we could
3596                  * be running on at the time of the RDPMC instruction.
3597                  *
3598                  * The GETMSR operation is not allowed for PMCs that
3599                  * are inherited across processes.
3600                  */
3601
3602                 if (!PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)) ||
3603                     (pm->pm_flags & PMC_F_DESCENDANTS)) {
3604                         error = EINVAL;
3605                         break;
3606                 }
3607
3608                 /*
3609                  * It only makes sense to use a RDPMC (or its
3610                  * equivalent instruction on non-x86 architectures) on
3611                  * a process that has allocated and attached a PMC to
3612                  * itself.  Conversely the PMC is only allowed to have
3613                  * one process attached to it -- its owner.
3614                  */
3615
3616                 if ((pt = LIST_FIRST(&pm->pm_targets)) == NULL ||
3617                     LIST_NEXT(pt, pt_next) != NULL ||
3618                     pt->pt_process->pp_proc != pm->pm_owner->po_owner) {
3619                         error = EINVAL;
3620                         break;
3621                 }
3622
3623                 ri = PMC_TO_ROWINDEX(pm);
3624                 pcd = pmc_ri_to_classdep(md, ri, &adjri);
3625
3626                 /* PMC class has no 'GETMSR' support */
3627                 if (pcd->pcd_get_msr == NULL) {
3628                         error = ENOSYS;
3629                         break;
3630                 }
3631
3632                 if ((error = (*pcd->pcd_get_msr)(adjri, &gm.pm_msr)) < 0)
3633                         break;
3634
3635                 if ((error = copyout(&gm, arg, sizeof(gm))) < 0)
3636                         break;
3637
3638                 /*
3639                  * Mark our process as using MSRs.  Update machine
3640                  * state using a forced context switch.
3641                  */
3642
3643                 pt->pt_process->pp_flags |= PMC_PP_ENABLE_MSR_ACCESS;
3644                 pmc_force_context_switch();
3645
3646         }
3647         break;
3648
3649         /*
3650          * Release an allocated PMC
3651          */
3652
3653         case PMC_OP_PMCRELEASE:
3654         {
3655                 pmc_id_t pmcid;
3656                 struct pmc *pm;
3657                 struct pmc_owner *po;
3658                 struct pmc_op_simple sp;
3659
3660                 /*
3661                  * Find PMC pointer for the named PMC.
3662                  *
3663                  * Use pmc_release_pmc_descriptor() to switch off the
3664                  * PMC, remove all its target threads, and remove the
3665                  * PMC from its owner's list.
3666                  *
3667                  * Remove the owner record if this is the last PMC
3668                  * owned.
3669                  *
3670                  * Free up space.
3671                  */
3672
3673                 if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
3674                         break;
3675
3676                 pmcid = sp.pm_pmcid;
3677
3678                 if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
3679                         break;
3680
3681                 po = pm->pm_owner;
3682                 pmc_release_pmc_descriptor(pm);
3683                 pmc_maybe_remove_owner(po);
3684                 pmc_destroy_pmc_descriptor(pm);
3685         }
3686         break;
3687
3688
3689         /*
3690          * Read and/or write a PMC.
3691          */
3692
3693         case PMC_OP_PMCRW:
3694         {
3695                 int adjri;
3696                 struct pmc *pm;
3697                 uint32_t cpu, ri;
3698                 pmc_value_t oldvalue;
3699                 struct pmc_binding pb;
3700                 struct pmc_op_pmcrw prw;
3701                 struct pmc_classdep *pcd;
3702                 struct pmc_op_pmcrw *pprw;
3703
3704                 PMC_DOWNGRADE_SX();
3705
3706                 if ((error = copyin(arg, &prw, sizeof(prw))) != 0)
3707                         break;
3708
3709                 ri = 0;
3710                 PMCDBG2(PMC,OPS,1, "rw id=%d flags=0x%x", prw.pm_pmcid,
3711                     prw.pm_flags);
3712
3713                 /* must have at least one flag set */
3714                 if ((prw.pm_flags & (PMC_F_OLDVALUE|PMC_F_NEWVALUE)) == 0) {
3715                         error = EINVAL;
3716                         break;
3717                 }
3718
3719                 /* locate pmc descriptor */
3720                 if ((error = pmc_find_pmc(prw.pm_pmcid, &pm)) != 0)
3721                         break;
3722
3723                 /* Can't read a PMC that hasn't been started. */
3724                 if (pm->pm_state != PMC_STATE_ALLOCATED &&
3725                     pm->pm_state != PMC_STATE_STOPPED &&
3726                     pm->pm_state != PMC_STATE_RUNNING) {
3727                         error = EINVAL;
3728                         break;
3729                 }
3730
3731                 /* writing a new value is allowed only for 'STOPPED' pmcs */
3732                 if (pm->pm_state == PMC_STATE_RUNNING &&
3733                     (prw.pm_flags & PMC_F_NEWVALUE)) {
3734                         error = EBUSY;
3735                         break;
3736                 }
3737
3738                 if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) {
3739
3740                         /*
3741                          * If this PMC is attached to its owner (i.e.,
3742                          * the process requesting this operation) and
3743                          * is running, then attempt to get an
3744                          * upto-date reading from hardware for a READ.
3745                          * Writes are only allowed when the PMC is
3746                          * stopped, so only update the saved value
3747                          * field.
3748                          *
3749                          * If the PMC is not running, or is not
3750                          * attached to its owner, read/write to the
3751                          * savedvalue field.
3752                          */
3753
3754                         ri = PMC_TO_ROWINDEX(pm);
3755                         pcd = pmc_ri_to_classdep(md, ri, &adjri);
3756
3757                         mtx_pool_lock_spin(pmc_mtxpool, pm);
3758                         cpu = curthread->td_oncpu;
3759
3760                         if (prw.pm_flags & PMC_F_OLDVALUE) {
3761                                 if ((pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) &&
3762                                     (pm->pm_state == PMC_STATE_RUNNING))
3763                                         error = (*pcd->pcd_read_pmc)(cpu, adjri,
3764                                             &oldvalue);
3765                                 else
3766                                         oldvalue = pm->pm_gv.pm_savedvalue;
3767                         }
3768                         if (prw.pm_flags & PMC_F_NEWVALUE)
3769                                 pm->pm_gv.pm_savedvalue = prw.pm_value;
3770
3771                         mtx_pool_unlock_spin(pmc_mtxpool, pm);
3772
3773                 } else { /* System mode PMCs */
3774                         cpu = PMC_TO_CPU(pm);
3775                         ri  = PMC_TO_ROWINDEX(pm);
3776                         pcd = pmc_ri_to_classdep(md, ri, &adjri);
3777
3778                         if (!pmc_cpu_is_active(cpu)) {
3779                                 error = ENXIO;
3780                                 break;
3781                         }
3782
3783                         /* move this thread to CPU 'cpu' */
3784                         pmc_save_cpu_binding(&pb);
3785                         pmc_select_cpu(cpu);
3786
3787                         critical_enter();
3788                         /* save old value */
3789                         if (prw.pm_flags & PMC_F_OLDVALUE)
3790                                 if ((error = (*pcd->pcd_read_pmc)(cpu, adjri,
3791                                          &oldvalue)))
3792                                         goto error;
3793                         /* write out new value */
3794                         if (prw.pm_flags & PMC_F_NEWVALUE)
3795                                 error = (*pcd->pcd_write_pmc)(cpu, adjri,
3796                                     prw.pm_value);
3797                 error:
3798                         critical_exit();
3799                         pmc_restore_cpu_binding(&pb);
3800                         if (error)
3801                                 break;
3802                 }
3803
3804                 pprw = (struct pmc_op_pmcrw *) arg;
3805
3806 #ifdef  HWPMC_DEBUG
3807                 if (prw.pm_flags & PMC_F_NEWVALUE)
3808                         PMCDBG3(PMC,OPS,2, "rw id=%d new %jx -> old %jx",
3809                             ri, prw.pm_value, oldvalue);
3810                 else if (prw.pm_flags & PMC_F_OLDVALUE)
3811                         PMCDBG2(PMC,OPS,2, "rw id=%d -> old %jx", ri, oldvalue);
3812 #endif
3813
3814                 /* return old value if requested */
3815                 if (prw.pm_flags & PMC_F_OLDVALUE)
3816                         if ((error = copyout(&oldvalue, &pprw->pm_value,
3817                                  sizeof(prw.pm_value))))
3818                                 break;
3819
3820         }
3821         break;
3822
3823
3824         /*
3825          * Set the sampling rate for a sampling mode PMC and the
3826          * initial count for a counting mode PMC.
3827          */
3828
3829         case PMC_OP_PMCSETCOUNT:
3830         {
3831                 struct pmc *pm;
3832                 struct pmc_op_pmcsetcount sc;
3833
3834                 PMC_DOWNGRADE_SX();
3835
3836                 if ((error = copyin(arg, &sc, sizeof(sc))) != 0)
3837                         break;
3838
3839                 if ((error = pmc_find_pmc(sc.pm_pmcid, &pm)) != 0)
3840                         break;
3841
3842                 if (pm->pm_state == PMC_STATE_RUNNING) {
3843                         error = EBUSY;
3844                         break;
3845                 }
3846
3847                 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
3848                         pm->pm_sc.pm_reloadcount = sc.pm_count;
3849                 else
3850                         pm->pm_sc.pm_initial = sc.pm_count;
3851         }
3852         break;
3853
3854
3855         /*
3856          * Start a PMC.
3857          */
3858
3859         case PMC_OP_PMCSTART:
3860         {
3861                 pmc_id_t pmcid;
3862                 struct pmc *pm;
3863                 struct pmc_op_simple sp;
3864
3865                 sx_assert(&pmc_sx, SX_XLOCKED);
3866
3867                 if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
3868                         break;
3869
3870                 pmcid = sp.pm_pmcid;
3871
3872                 if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
3873                         break;
3874
3875                 KASSERT(pmcid == pm->pm_id,
3876                     ("[pmc,%d] pmcid %x != id %x", __LINE__,
3877                         pm->pm_id, pmcid));
3878
3879                 if (pm->pm_state == PMC_STATE_RUNNING) /* already running */
3880                         break;
3881                 else if (pm->pm_state != PMC_STATE_STOPPED &&
3882                     pm->pm_state != PMC_STATE_ALLOCATED) {
3883                         error = EINVAL;
3884                         break;
3885                 }
3886
3887                 error = pmc_start(pm);
3888         }
3889         break;
3890
3891
3892         /*
3893          * Stop a PMC.
3894          */
3895
3896         case PMC_OP_PMCSTOP:
3897         {
3898                 pmc_id_t pmcid;
3899                 struct pmc *pm;
3900                 struct pmc_op_simple sp;
3901
3902                 PMC_DOWNGRADE_SX();
3903
3904                 if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
3905                         break;
3906
3907                 pmcid = sp.pm_pmcid;
3908
3909                 /*
3910                  * Mark the PMC as inactive and invoke the MD stop
3911                  * routines if needed.
3912                  */
3913
3914                 if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
3915                         break;
3916
3917                 KASSERT(pmcid == pm->pm_id,
3918                     ("[pmc,%d] pmc id %x != pmcid %x", __LINE__,
3919                         pm->pm_id, pmcid));
3920
3921                 if (pm->pm_state == PMC_STATE_STOPPED) /* already stopped */
3922                         break;
3923                 else if (pm->pm_state != PMC_STATE_RUNNING) {
3924                         error = EINVAL;
3925                         break;
3926                 }
3927
3928                 error = pmc_stop(pm);
3929         }
3930         break;
3931
3932
3933         /*
3934          * Write a user supplied value to the log file.
3935          */
3936
3937         case PMC_OP_WRITELOG:
3938         {
3939                 struct pmc_op_writelog wl;
3940                 struct pmc_owner *po;
3941
3942                 PMC_DOWNGRADE_SX();
3943
3944                 if ((error = copyin(arg, &wl, sizeof(wl))) != 0)
3945                         break;
3946
3947                 if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) {
3948                         error = EINVAL;
3949                         break;
3950                 }
3951
3952                 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) {
3953                         error = EINVAL;
3954                         break;
3955                 }
3956
3957                 error = pmclog_process_userlog(po, &wl);
3958         }
3959         break;
3960
3961
3962         default:
3963                 error = EINVAL;
3964                 break;
3965         }
3966
3967         if (is_sx_locked != 0) {
3968                 if (is_sx_downgraded)
3969                         sx_sunlock(&pmc_sx);
3970                 else
3971                         sx_xunlock(&pmc_sx);
3972         }
3973
3974         if (error)
3975                 atomic_add_int(&pmc_stats.pm_syscall_errors, 1);
3976
3977         PICKUP_GIANT();
3978
3979         return error;
3980 }
3981
3982 /*
3983  * Helper functions
3984  */
3985
3986
3987 /*
3988  * Mark the thread as needing callchain capture and post an AST.  The
3989  * actual callchain capture will be done in a context where it is safe
3990  * to take page faults.
3991  */
3992
3993 static void
3994 pmc_post_callchain_callback(void)
3995 {
3996         struct thread *td;
3997
3998         td = curthread;
3999
4000         /*
4001          * If there is multiple PMCs for the same interrupt ignore new post
4002          */
4003         if (td->td_pflags & TDP_CALLCHAIN)
4004                 return;
4005
4006         /*
4007          * Mark this thread as needing callchain capture.
4008          * `td->td_pflags' will be safe to touch because this thread
4009          * was in user space when it was interrupted.
4010          */
4011         td->td_pflags |= TDP_CALLCHAIN;
4012
4013         /*
4014          * Don't let this thread migrate between CPUs until callchain
4015          * capture completes.
4016          */
4017         sched_pin();
4018
4019         return;
4020 }
4021
4022 /*
4023  * Interrupt processing.
4024  *
4025  * Find a free slot in the per-cpu array of samples and capture the
4026  * current callchain there.  If a sample was successfully added, a bit
4027  * is set in mask 'pmc_cpumask' denoting that the DO_SAMPLES hook
4028  * needs to be invoked from the clock handler.
4029  *
4030  * This function is meant to be called from an NMI handler.  It cannot
4031  * use any of the locking primitives supplied by the OS.
4032  */
4033
4034 int
4035 pmc_process_interrupt(int cpu, int ring, struct pmc *pm, struct trapframe *tf,
4036     int inuserspace)
4037 {
4038         int error, callchaindepth;
4039         struct thread *td;
4040         struct pmc_sample *ps;
4041         struct pmc_samplebuffer *psb;
4042
4043         error = 0;
4044
4045         /*
4046          * Allocate space for a sample buffer.
4047          */
4048         psb = pmc_pcpu[cpu]->pc_sb[ring];
4049
4050         ps = psb->ps_write;
4051         if (ps->ps_nsamples) {  /* in use, reader hasn't caught up */
4052                 pm->pm_stalled = 1;
4053                 atomic_add_int(&pmc_stats.pm_intr_bufferfull, 1);
4054                 PMCDBG6(SAM,INT,1,"(spc) cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d",
4055                     cpu, pm, (void *) tf, inuserspace,
4056                     (int) (psb->ps_write - psb->ps_samples),
4057                     (int) (psb->ps_read - psb->ps_samples));
4058                 error = ENOMEM;
4059                 goto done;
4060         }
4061
4062
4063         /* Fill in entry. */
4064         PMCDBG6(SAM,INT,1,"cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d", cpu, pm,
4065             (void *) tf, inuserspace,
4066             (int) (psb->ps_write - psb->ps_samples),
4067             (int) (psb->ps_read - psb->ps_samples));
4068
4069         KASSERT(pm->pm_runcount >= 0,
4070             ("[pmc,%d] pm=%p runcount %d", __LINE__, (void *) pm,
4071                 pm->pm_runcount));
4072
4073         atomic_add_rel_int(&pm->pm_runcount, 1);        /* hold onto PMC */
4074
4075         ps->ps_pmc = pm;
4076         if ((td = curthread) && td->td_proc)
4077                 ps->ps_pid = td->td_proc->p_pid;
4078         else
4079                 ps->ps_pid = -1;
4080         ps->ps_cpu = cpu;
4081         ps->ps_td = td;
4082         ps->ps_flags = inuserspace ? PMC_CC_F_USERSPACE : 0;
4083
4084         callchaindepth = (pm->pm_flags & PMC_F_CALLCHAIN) ?
4085             pmc_callchaindepth : 1;
4086
4087         if (callchaindepth == 1)
4088                 ps->ps_pc[0] = PMC_TRAPFRAME_TO_PC(tf);
4089         else {
4090                 /*
4091                  * Kernel stack traversals can be done immediately,
4092                  * while we defer to an AST for user space traversals.
4093                  */
4094                 if (!inuserspace) {
4095                         callchaindepth =
4096                             pmc_save_kernel_callchain(ps->ps_pc,
4097                                 callchaindepth, tf);
4098                 } else {
4099                         pmc_post_callchain_callback();
4100                         callchaindepth = PMC_SAMPLE_INUSE;
4101                 }
4102         }
4103
4104         ps->ps_nsamples = callchaindepth;       /* mark entry as in use */
4105
4106         /* increment write pointer, modulo ring buffer size */
4107         ps++;
4108         if (ps == psb->ps_fence)
4109                 psb->ps_write = psb->ps_samples;
4110         else
4111                 psb->ps_write = ps;
4112
4113  done:
4114         /* mark CPU as needing processing */
4115         CPU_SET_ATOMIC(cpu, &pmc_cpumask);
4116
4117         return (error);
4118 }
4119
4120 /*
4121  * Capture a user call chain.  This function will be called from ast()
4122  * before control returns to userland and before the process gets
4123  * rescheduled.
4124  */
4125
4126 static void
4127 pmc_capture_user_callchain(int cpu, int ring, struct trapframe *tf)
4128 {
4129         int i;
4130         struct pmc *pm;
4131         struct thread *td;
4132         struct pmc_sample *ps;
4133         struct pmc_samplebuffer *psb;
4134 #ifdef  INVARIANTS
4135         int ncallchains;
4136 #endif
4137
4138         psb = pmc_pcpu[cpu]->pc_sb[ring];
4139         td = curthread;
4140
4141         KASSERT(td->td_pflags & TDP_CALLCHAIN,
4142             ("[pmc,%d] Retrieving callchain for thread that doesn't want it",
4143                 __LINE__));
4144
4145 #ifdef  INVARIANTS
4146         ncallchains = 0;
4147 #endif
4148
4149         /*
4150          * Iterate through all deferred callchain requests.
4151          */
4152
4153         ps = psb->ps_samples;
4154         for (i = 0; i < pmc_nsamples; i++, ps++) {
4155
4156                 if (ps->ps_nsamples != PMC_SAMPLE_INUSE)
4157                         continue;
4158                 if (ps->ps_td != td)
4159                         continue;
4160
4161                 KASSERT(ps->ps_cpu == cpu,
4162                     ("[pmc,%d] cpu mismatch ps_cpu=%d pcpu=%d", __LINE__,
4163                         ps->ps_cpu, PCPU_GET(cpuid)));
4164
4165                 pm = ps->ps_pmc;
4166
4167                 KASSERT(pm->pm_flags & PMC_F_CALLCHAIN,
4168                     ("[pmc,%d] Retrieving callchain for PMC that doesn't "
4169                         "want it", __LINE__));
4170
4171                 KASSERT(pm->pm_runcount > 0,
4172                     ("[pmc,%d] runcount %d", __LINE__, pm->pm_runcount));
4173
4174                 /*
4175                  * Retrieve the callchain and mark the sample buffer
4176                  * as 'processable' by the timer tick sweep code.
4177                  */
4178                 ps->ps_nsamples = pmc_save_user_callchain(ps->ps_pc,
4179                     pmc_callchaindepth, tf);
4180
4181 #ifdef  INVARIANTS
4182                 ncallchains++;
4183 #endif
4184         }
4185
4186         KASSERT(ncallchains > 0,
4187             ("[pmc,%d] cpu %d didn't find a sample to collect", __LINE__,
4188                 cpu));
4189
4190         KASSERT(td->td_pinned == 1,
4191             ("[pmc,%d] invalid td_pinned value", __LINE__));
4192         sched_unpin();  /* Can migrate safely now. */
4193
4194         return;
4195 }
4196
4197 /*
4198  * Process saved PC samples.
4199  */
4200
4201 static void
4202 pmc_process_samples(int cpu, int ring)
4203 {
4204         struct pmc *pm;
4205         int adjri, n;
4206         struct thread *td;
4207         struct pmc_owner *po;
4208         struct pmc_sample *ps;
4209         struct pmc_classdep *pcd;
4210         struct pmc_samplebuffer *psb;
4211
4212         KASSERT(PCPU_GET(cpuid) == cpu,
4213             ("[pmc,%d] not on the correct CPU pcpu=%d cpu=%d", __LINE__,
4214                 PCPU_GET(cpuid), cpu));
4215
4216         psb = pmc_pcpu[cpu]->pc_sb[ring];
4217
4218         for (n = 0; n < pmc_nsamples; n++) { /* bound on #iterations */
4219
4220                 ps = psb->ps_read;
4221                 if (ps->ps_nsamples == PMC_SAMPLE_FREE)
4222                         break;
4223
4224                 pm = ps->ps_pmc;
4225
4226                 KASSERT(pm->pm_runcount > 0,
4227                     ("[pmc,%d] pm=%p runcount %d", __LINE__, (void *) pm,
4228                         pm->pm_runcount));
4229
4230                 po = pm->pm_owner;
4231
4232                 KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)),
4233                     ("[pmc,%d] pmc=%p non-sampling mode=%d", __LINE__,
4234                         pm, PMC_TO_MODE(pm)));
4235
4236                 /* Ignore PMCs that have been switched off */
4237                 if (pm->pm_state != PMC_STATE_RUNNING)
4238                         goto entrydone;
4239
4240                 /* If there is a pending AST wait for completion */
4241                 if (ps->ps_nsamples == PMC_SAMPLE_INUSE) {
4242                         /* Need a rescan at a later time. */
4243                         CPU_SET_ATOMIC(cpu, &pmc_cpumask);
4244                         break;
4245                 }
4246
4247                 PMCDBG6(SAM,OPS,1,"cpu=%d pm=%p n=%d fl=%x wr=%d rd=%d", cpu,
4248                     pm, ps->ps_nsamples, ps->ps_flags,
4249                     (int) (psb->ps_write - psb->ps_samples),
4250                     (int) (psb->ps_read - psb->ps_samples));
4251
4252                 /*
4253                  * If this is a process-mode PMC that is attached to
4254                  * its owner, and if the PC is in user mode, update
4255                  * profiling statistics like timer-based profiling
4256                  * would have done.
4257                  */
4258                 if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) {
4259                         if (ps->ps_flags & PMC_CC_F_USERSPACE) {
4260                                 td = FIRST_THREAD_IN_PROC(po->po_owner);
4261                                 addupc_intr(td, ps->ps_pc[0], 1);
4262                         }
4263                         goto entrydone;
4264                 }
4265
4266                 /*
4267                  * Otherwise, this is either a sampling mode PMC that
4268                  * is attached to a different process than its owner,
4269                  * or a system-wide sampling PMC.  Dispatch a log
4270                  * entry to the PMC's owner process.
4271                  */
4272                 pmclog_process_callchain(pm, ps);
4273
4274         entrydone:
4275                 ps->ps_nsamples = 0; /* mark entry as free */
4276                 atomic_subtract_rel_int(&pm->pm_runcount, 1);
4277
4278                 /* increment read pointer, modulo sample size */
4279                 if (++ps == psb->ps_fence)
4280                         psb->ps_read = psb->ps_samples;
4281                 else
4282                         psb->ps_read = ps;
4283         }
4284
4285         atomic_add_int(&pmc_stats.pm_log_sweeps, 1);
4286
4287         /* Do not re-enable stalled PMCs if we failed to process any samples */
4288         if (n == 0)
4289                 return;
4290
4291         /*
4292          * Restart any stalled sampling PMCs on this CPU.
4293          *
4294          * If the NMI handler sets the pm_stalled field of a PMC after
4295          * the check below, we'll end up processing the stalled PMC at
4296          * the next hardclock tick.
4297          */
4298         for (n = 0; n < md->pmd_npmc; n++) {
4299                 pcd = pmc_ri_to_classdep(md, n, &adjri);
4300                 KASSERT(pcd != NULL,
4301                     ("[pmc,%d] null pcd ri=%d", __LINE__, n));
4302                 (void) (*pcd->pcd_get_config)(cpu,adjri,&pm);
4303
4304                 if (pm == NULL ||                        /* !cfg'ed */
4305                     pm->pm_state != PMC_STATE_RUNNING || /* !active */
4306                     !PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) || /* !sampling */
4307                     pm->pm_stalled == 0) /* !stalled */
4308                         continue;
4309
4310                 pm->pm_stalled = 0;
4311                 (*pcd->pcd_start_pmc)(cpu, adjri);
4312         }
4313 }
4314
4315 /*
4316  * Event handlers.
4317  */
4318
4319 /*
4320  * Handle a process exit.
4321  *
4322  * Remove this process from all hash tables.  If this process
4323  * owned any PMCs, turn off those PMCs and deallocate them,
4324  * removing any associations with target processes.
4325  *
4326  * This function will be called by the last 'thread' of a
4327  * process.
4328  *
4329  * XXX This eventhandler gets called early in the exit process.
4330  * Consider using a 'hook' invocation from thread_exit() or equivalent
4331  * spot.  Another negative is that kse_exit doesn't seem to call
4332  * exit1() [??].
4333  *
4334  */
4335
4336 static void
4337 pmc_process_exit(void *arg __unused, struct proc *p)
4338 {
4339         struct pmc *pm;
4340         int adjri, cpu;
4341         unsigned int ri;
4342         int is_using_hwpmcs;
4343         struct pmc_owner *po;
4344         struct pmc_process *pp;
4345         struct pmc_classdep *pcd;
4346         pmc_value_t newvalue, tmp;
4347
4348         PROC_LOCK(p);
4349         is_using_hwpmcs = p->p_flag & P_HWPMC;
4350         PROC_UNLOCK(p);
4351
4352         /*
4353          * Log a sysexit event to all SS PMC owners.
4354          */
4355         LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
4356             if (po->po_flags & PMC_PO_OWNS_LOGFILE)
4357                     pmclog_process_sysexit(po, p->p_pid);
4358
4359         if (!is_using_hwpmcs)
4360                 return;
4361
4362         PMC_GET_SX_XLOCK();
4363         PMCDBG3(PRC,EXT,1,"process-exit proc=%p (%d, %s)", p, p->p_pid,
4364             p->p_comm);
4365
4366         /*
4367          * Since this code is invoked by the last thread in an exiting
4368          * process, we would have context switched IN at some prior
4369          * point.  However, with PREEMPTION, kernel mode context
4370          * switches may happen any time, so we want to disable a
4371          * context switch OUT till we get any PMCs targetting this
4372          * process off the hardware.
4373          *
4374          * We also need to atomically remove this process'
4375          * entry from our target process hash table, using
4376          * PMC_FLAG_REMOVE.
4377          */
4378         PMCDBG3(PRC,EXT,1, "process-exit proc=%p (%d, %s)", p, p->p_pid,
4379             p->p_comm);
4380
4381         critical_enter(); /* no preemption */
4382
4383         cpu = curthread->td_oncpu;
4384
4385         if ((pp = pmc_find_process_descriptor(p,
4386                  PMC_FLAG_REMOVE)) != NULL) {
4387
4388                 PMCDBG2(PRC,EXT,2,
4389                     "process-exit proc=%p pmc-process=%p", p, pp);
4390
4391                 /*
4392                  * The exiting process could the target of
4393                  * some PMCs which will be running on
4394                  * currently executing CPU.
4395                  *
4396                  * We need to turn these PMCs off like we
4397                  * would do at context switch OUT time.
4398                  */
4399                 for (ri = 0; ri < md->pmd_npmc; ri++) {
4400
4401                         /*
4402                          * Pick up the pmc pointer from hardware
4403                          * state similar to the CSW_OUT code.
4404                          */
4405                         pm = NULL;
4406
4407                         pcd = pmc_ri_to_classdep(md, ri, &adjri);
4408
4409                         (void) (*pcd->pcd_get_config)(cpu, adjri, &pm);
4410
4411                         PMCDBG2(PRC,EXT,2, "ri=%d pm=%p", ri, pm);
4412
4413                         if (pm == NULL ||
4414                             !PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)))
4415                                 continue;
4416
4417                         PMCDBG4(PRC,EXT,2, "ppmcs[%d]=%p pm=%p "
4418                             "state=%d", ri, pp->pp_pmcs[ri].pp_pmc,
4419                             pm, pm->pm_state);
4420
4421                         KASSERT(PMC_TO_ROWINDEX(pm) == ri,
4422                             ("[pmc,%d] ri mismatch pmc(%d) ri(%d)",
4423                                 __LINE__, PMC_TO_ROWINDEX(pm), ri));
4424
4425                         KASSERT(pm == pp->pp_pmcs[ri].pp_pmc,
4426                             ("[pmc,%d] pm %p != pp_pmcs[%d] %p",
4427                                 __LINE__, pm, ri, pp->pp_pmcs[ri].pp_pmc));
4428
4429                         (void) pcd->pcd_stop_pmc(cpu, adjri);
4430
4431                         KASSERT(pm->pm_runcount > 0,
4432                             ("[pmc,%d] bad runcount ri %d rc %d",
4433                                 __LINE__, ri, pm->pm_runcount));
4434
4435                         /* Stop hardware only if it is actually running */
4436                         if (pm->pm_state == PMC_STATE_RUNNING &&
4437                             pm->pm_stalled == 0) {
4438                                 pcd->pcd_read_pmc(cpu, adjri, &newvalue);
4439                                 tmp = newvalue -
4440                                     PMC_PCPU_SAVED(cpu,ri);
4441
4442                                 mtx_pool_lock_spin(pmc_mtxpool, pm);
4443                                 pm->pm_gv.pm_savedvalue += tmp;
4444                                 pp->pp_pmcs[ri].pp_pmcval += tmp;
4445                                 mtx_pool_unlock_spin(pmc_mtxpool, pm);
4446                         }
4447
4448                         atomic_subtract_rel_int(&pm->pm_runcount,1);
4449
4450                         KASSERT((int) pm->pm_runcount >= 0,
4451                             ("[pmc,%d] runcount is %d", __LINE__, ri));
4452
4453                         (void) pcd->pcd_config_pmc(cpu, adjri, NULL);
4454                 }
4455
4456                 /*
4457                  * Inform the MD layer of this pseudo "context switch
4458                  * out"
4459                  */
4460                 (void) md->pmd_switch_out(pmc_pcpu[cpu], pp);
4461
4462                 critical_exit(); /* ok to be pre-empted now */
4463
4464                 /*
4465                  * Unlink this process from the PMCs that are
4466                  * targetting it.  This will send a signal to
4467                  * all PMC owner's whose PMCs are orphaned.
4468                  *
4469                  * Log PMC value at exit time if requested.
4470                  */
4471                 for (ri = 0; ri < md->pmd_npmc; ri++)
4472                         if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) {
4473                                 if (pm->pm_flags & PMC_F_NEEDS_LOGFILE &&
4474                                     PMC_IS_COUNTING_MODE(PMC_TO_MODE(pm)))
4475                                         pmclog_process_procexit(pm, pp);
4476                                 pmc_unlink_target_process(pm, pp);
4477                         }
4478                 free(pp, M_PMC);
4479
4480         } else
4481                 critical_exit(); /* pp == NULL */
4482
4483
4484         /*
4485          * If the process owned PMCs, free them up and free up
4486          * memory.
4487          */
4488         if ((po = pmc_find_owner_descriptor(p)) != NULL) {
4489                 pmc_remove_owner(po);
4490                 pmc_destroy_owner_descriptor(po);
4491         }
4492
4493         sx_xunlock(&pmc_sx);
4494 }
4495
4496 /*
4497  * Handle a process fork.
4498  *
4499  * If the parent process 'p1' is under HWPMC monitoring, then copy
4500  * over any attached PMCs that have 'do_descendants' semantics.
4501  */
4502
4503 static void
4504 pmc_process_fork(void *arg __unused, struct proc *p1, struct proc *newproc,
4505     int flags)
4506 {
4507         int is_using_hwpmcs;
4508         unsigned int ri;
4509         uint32_t do_descendants;
4510         struct pmc *pm;
4511         struct pmc_owner *po;
4512         struct pmc_process *ppnew, *ppold;
4513
4514         (void) flags;           /* unused parameter */
4515
4516         PROC_LOCK(p1);
4517         is_using_hwpmcs = p1->p_flag & P_HWPMC;
4518         PROC_UNLOCK(p1);
4519
4520         /*
4521          * If there are system-wide sampling PMCs active, we need to
4522          * log all fork events to their owner's logs.
4523          */
4524
4525         LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
4526             if (po->po_flags & PMC_PO_OWNS_LOGFILE)
4527                     pmclog_process_procfork(po, p1->p_pid, newproc->p_pid);
4528
4529         if (!is_using_hwpmcs)
4530                 return;
4531
4532         PMC_GET_SX_XLOCK();
4533         PMCDBG4(PMC,FRK,1, "process-fork proc=%p (%d, %s) -> %p", p1,
4534             p1->p_pid, p1->p_comm, newproc);
4535
4536         /*
4537          * If the parent process (curthread->td_proc) is a
4538          * target of any PMCs, look for PMCs that are to be
4539          * inherited, and link these into the new process
4540          * descriptor.
4541          */
4542         if ((ppold = pmc_find_process_descriptor(curthread->td_proc,
4543                  PMC_FLAG_NONE)) == NULL)
4544                 goto done;              /* nothing to do */
4545
4546         do_descendants = 0;
4547         for (ri = 0; ri < md->pmd_npmc; ri++)
4548                 if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL)
4549                         do_descendants |= pm->pm_flags & PMC_F_DESCENDANTS;
4550         if (do_descendants == 0) /* nothing to do */
4551                 goto done;
4552
4553         /* allocate a descriptor for the new process  */
4554         if ((ppnew = pmc_find_process_descriptor(newproc,
4555                  PMC_FLAG_ALLOCATE)) == NULL)
4556                 goto done;
4557
4558         /*
4559          * Run through all PMCs that were targeting the old process
4560          * and which specified F_DESCENDANTS and attach them to the
4561          * new process.
4562          *
4563          * Log the fork event to all owners of PMCs attached to this
4564          * process, if not already logged.
4565          */
4566         for (ri = 0; ri < md->pmd_npmc; ri++)
4567                 if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL &&
4568                     (pm->pm_flags & PMC_F_DESCENDANTS)) {
4569                         pmc_link_target_process(pm, ppnew);
4570                         po = pm->pm_owner;
4571                         if (po->po_sscount == 0 &&
4572                             po->po_flags & PMC_PO_OWNS_LOGFILE)
4573                                 pmclog_process_procfork(po, p1->p_pid,
4574                                     newproc->p_pid);
4575                 }
4576
4577         /*
4578          * Now mark the new process as being tracked by this driver.
4579          */
4580         PROC_LOCK(newproc);
4581         newproc->p_flag |= P_HWPMC;
4582         PROC_UNLOCK(newproc);
4583
4584  done:
4585         sx_xunlock(&pmc_sx);
4586 }
4587
4588 static void
4589 pmc_kld_load(void *arg __unused, linker_file_t lf)
4590 {
4591         struct pmc_owner *po;
4592
4593         sx_slock(&pmc_sx);
4594
4595         /*
4596          * Notify owners of system sampling PMCs about KLD operations.
4597          */
4598         LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
4599                 if (po->po_flags & PMC_PO_OWNS_LOGFILE)
4600                         pmclog_process_map_in(po, (pid_t) -1,
4601                             (uintfptr_t) lf->address, lf->filename);
4602
4603         /*
4604          * TODO: Notify owners of (all) process-sampling PMCs too.
4605          */
4606
4607         sx_sunlock(&pmc_sx);
4608 }
4609
4610 static void
4611 pmc_kld_unload(void *arg __unused, const char *filename __unused,
4612     caddr_t address, size_t size)
4613 {
4614         struct pmc_owner *po;
4615
4616         sx_slock(&pmc_sx);
4617
4618         LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
4619                 if (po->po_flags & PMC_PO_OWNS_LOGFILE)
4620                         pmclog_process_map_out(po, (pid_t) -1,
4621                             (uintfptr_t) address, (uintfptr_t) address + size);
4622
4623         /*
4624          * TODO: Notify owners of process-sampling PMCs.
4625          */
4626
4627         sx_sunlock(&pmc_sx);
4628 }
4629
4630 /*
4631  * initialization
4632  */
4633
4634 static const char *pmc_name_of_pmcclass[] = {
4635 #undef  __PMC_CLASS
4636 #define __PMC_CLASS(N) #N ,
4637         __PMC_CLASSES()
4638 };
4639
4640 /*
4641  * Base class initializer: allocate structure and set default classes.
4642  */
4643 struct pmc_mdep *
4644 pmc_mdep_alloc(int nclasses)
4645 {
4646         struct pmc_mdep *md;
4647         int     n;
4648
4649         /* SOFT + md classes */
4650         n = 1 + nclasses;
4651         md = malloc(sizeof(struct pmc_mdep) + n *
4652             sizeof(struct pmc_classdep), M_PMC, M_WAITOK|M_ZERO);
4653         md->pmd_nclass = n;
4654
4655         /* Add base class. */
4656         pmc_soft_initialize(md);
4657         return md;
4658 }
4659
4660 void
4661 pmc_mdep_free(struct pmc_mdep *md)
4662 {
4663         pmc_soft_finalize(md);
4664         free(md, M_PMC);
4665 }
4666
4667 static int
4668 generic_switch_in(struct pmc_cpu *pc, struct pmc_process *pp)
4669 {
4670         (void) pc; (void) pp;
4671
4672         return (0);
4673 }
4674
4675 static int
4676 generic_switch_out(struct pmc_cpu *pc, struct pmc_process *pp)
4677 {
4678         (void) pc; (void) pp;
4679
4680         return (0);
4681 }
4682
4683 static struct pmc_mdep *
4684 pmc_generic_cpu_initialize(void)
4685 {
4686         struct pmc_mdep *md;
4687
4688         md = pmc_mdep_alloc(0);
4689
4690         md->pmd_cputype    = PMC_CPU_GENERIC;
4691
4692         md->pmd_pcpu_init  = NULL;
4693         md->pmd_pcpu_fini  = NULL;
4694         md->pmd_switch_in  = generic_switch_in;
4695         md->pmd_switch_out = generic_switch_out;
4696
4697         return (md);
4698 }
4699
4700 static void
4701 pmc_generic_cpu_finalize(struct pmc_mdep *md)
4702 {
4703         (void) md;
4704 }
4705
4706
4707 static int
4708 pmc_initialize(void)
4709 {
4710         int c, cpu, error, n, ri;
4711         unsigned int maxcpu;
4712         struct pmc_binding pb;
4713         struct pmc_sample *ps;
4714         struct pmc_classdep *pcd;
4715         struct pmc_samplebuffer *sb;
4716
4717         md = NULL;
4718         error = 0;
4719
4720 #ifdef  HWPMC_DEBUG
4721         /* parse debug flags first */
4722         if (TUNABLE_STR_FETCH(PMC_SYSCTL_NAME_PREFIX "debugflags",
4723                 pmc_debugstr, sizeof(pmc_debugstr)))
4724                 pmc_debugflags_parse(pmc_debugstr,
4725                     pmc_debugstr+strlen(pmc_debugstr));
4726 #endif
4727
4728         PMCDBG1(MOD,INI,0, "PMC Initialize (version %x)", PMC_VERSION);
4729
4730         /* check kernel version */
4731         if (pmc_kernel_version != PMC_VERSION) {
4732                 if (pmc_kernel_version == 0)
4733                         printf("hwpmc: this kernel has not been compiled with "
4734                             "'options HWPMC_HOOKS'.\n");
4735                 else
4736                         printf("hwpmc: kernel version (0x%x) does not match "
4737                             "module version (0x%x).\n", pmc_kernel_version,
4738                             PMC_VERSION);
4739                 return EPROGMISMATCH;
4740         }
4741
4742         /*
4743          * check sysctl parameters
4744          */
4745
4746         if (pmc_hashsize <= 0) {
4747                 (void) printf("hwpmc: tunable \"hashsize\"=%d must be "
4748                     "greater than zero.\n", pmc_hashsize);
4749                 pmc_hashsize = PMC_HASH_SIZE;
4750         }
4751
4752         if (pmc_nsamples <= 0 || pmc_nsamples > 65535) {
4753                 (void) printf("hwpmc: tunable \"nsamples\"=%d out of "
4754                     "range.\n", pmc_nsamples);
4755                 pmc_nsamples = PMC_NSAMPLES;
4756         }
4757
4758         if (pmc_callchaindepth <= 0 ||
4759             pmc_callchaindepth > PMC_CALLCHAIN_DEPTH_MAX) {
4760                 (void) printf("hwpmc: tunable \"callchaindepth\"=%d out of "
4761                     "range - using %d.\n", pmc_callchaindepth,
4762                     PMC_CALLCHAIN_DEPTH_MAX);
4763                 pmc_callchaindepth = PMC_CALLCHAIN_DEPTH_MAX;
4764         }
4765
4766         md = pmc_md_initialize();
4767         if (md == NULL) {
4768                 /* Default to generic CPU. */
4769                 md = pmc_generic_cpu_initialize();
4770                 if (md == NULL)
4771                         return (ENOSYS);
4772         }
4773
4774         KASSERT(md->pmd_nclass >= 1 && md->pmd_npmc >= 1,
4775             ("[pmc,%d] no classes or pmcs", __LINE__));
4776
4777         /* Compute the map from row-indices to classdep pointers. */
4778         pmc_rowindex_to_classdep = malloc(sizeof(struct pmc_classdep *) *
4779             md->pmd_npmc, M_PMC, M_WAITOK|M_ZERO);
4780
4781         for (n = 0; n < md->pmd_npmc; n++)
4782                 pmc_rowindex_to_classdep[n] = NULL;
4783         for (ri = c = 0; c < md->pmd_nclass; c++) {
4784                 pcd = &md->pmd_classdep[c];
4785                 for (n = 0; n < pcd->pcd_num; n++, ri++)
4786                         pmc_rowindex_to_classdep[ri] = pcd;
4787         }
4788
4789         KASSERT(ri == md->pmd_npmc,
4790             ("[pmc,%d] npmc miscomputed: ri=%d, md->npmc=%d", __LINE__,
4791             ri, md->pmd_npmc));
4792
4793         maxcpu = pmc_cpu_max();
4794
4795         /* allocate space for the per-cpu array */
4796         pmc_pcpu = malloc(maxcpu * sizeof(struct pmc_cpu *), M_PMC,
4797             M_WAITOK|M_ZERO);
4798
4799         /* per-cpu 'saved values' for managing process-mode PMCs */
4800         pmc_pcpu_saved = malloc(sizeof(pmc_value_t) * maxcpu * md->pmd_npmc,
4801             M_PMC, M_WAITOK);
4802
4803         /* Perform CPU-dependent initialization. */
4804         pmc_save_cpu_binding(&pb);
4805         error = 0;
4806         for (cpu = 0; error == 0 && cpu < maxcpu; cpu++) {
4807                 if (!pmc_cpu_is_active(cpu))
4808                         continue;
4809                 pmc_select_cpu(cpu);
4810                 pmc_pcpu[cpu] = malloc(sizeof(struct pmc_cpu) +
4811                     md->pmd_npmc * sizeof(struct pmc_hw *), M_PMC,
4812                     M_WAITOK|M_ZERO);
4813                 if (md->pmd_pcpu_init)
4814                         error = md->pmd_pcpu_init(md, cpu);
4815                 for (n = 0; error == 0 && n < md->pmd_nclass; n++)
4816                         error = md->pmd_classdep[n].pcd_pcpu_init(md, cpu);
4817         }
4818         pmc_restore_cpu_binding(&pb);
4819
4820         if (error)
4821                 return (error);
4822
4823         /* allocate space for the sample array */
4824         for (cpu = 0; cpu < maxcpu; cpu++) {
4825                 if (!pmc_cpu_is_active(cpu))
4826                         continue;
4827
4828                 sb = malloc(sizeof(struct pmc_samplebuffer) +
4829                     pmc_nsamples * sizeof(struct pmc_sample), M_PMC,
4830                     M_WAITOK|M_ZERO);
4831                 sb->ps_read = sb->ps_write = sb->ps_samples;
4832                 sb->ps_fence = sb->ps_samples + pmc_nsamples;
4833
4834                 KASSERT(pmc_pcpu[cpu] != NULL,
4835                     ("[pmc,%d] cpu=%d Null per-cpu data", __LINE__, cpu));
4836
4837                 sb->ps_callchains = malloc(pmc_callchaindepth * pmc_nsamples *
4838                     sizeof(uintptr_t), M_PMC, M_WAITOK|M_ZERO);
4839
4840                 for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++)
4841                         ps->ps_pc = sb->ps_callchains +
4842                             (n * pmc_callchaindepth);
4843
4844                 pmc_pcpu[cpu]->pc_sb[PMC_HR] = sb;
4845
4846                 sb = malloc(sizeof(struct pmc_samplebuffer) +
4847                     pmc_nsamples * sizeof(struct pmc_sample), M_PMC,
4848                     M_WAITOK|M_ZERO);
4849                 sb->ps_read = sb->ps_write = sb->ps_samples;
4850                 sb->ps_fence = sb->ps_samples + pmc_nsamples;
4851
4852                 KASSERT(pmc_pcpu[cpu] != NULL,
4853                     ("[pmc,%d] cpu=%d Null per-cpu data", __LINE__, cpu));
4854
4855                 sb->ps_callchains = malloc(pmc_callchaindepth * pmc_nsamples *
4856                     sizeof(uintptr_t), M_PMC, M_WAITOK|M_ZERO);
4857
4858                 for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++)
4859                         ps->ps_pc = sb->ps_callchains +
4860                             (n * pmc_callchaindepth);
4861
4862                 pmc_pcpu[cpu]->pc_sb[PMC_SR] = sb;
4863         }
4864
4865         /* allocate space for the row disposition array */
4866         pmc_pmcdisp = malloc(sizeof(enum pmc_mode) * md->pmd_npmc,
4867             M_PMC, M_WAITOK|M_ZERO);
4868
4869         /* mark all PMCs as available */
4870         for (n = 0; n < (int) md->pmd_npmc; n++)
4871                 PMC_MARK_ROW_FREE(n);
4872
4873         /* allocate thread hash tables */
4874         pmc_ownerhash = hashinit(pmc_hashsize, M_PMC,
4875             &pmc_ownerhashmask);
4876
4877         pmc_processhash = hashinit(pmc_hashsize, M_PMC,
4878             &pmc_processhashmask);
4879         mtx_init(&pmc_processhash_mtx, "pmc-process-hash", "pmc-leaf",
4880             MTX_SPIN);
4881
4882         LIST_INIT(&pmc_ss_owners);
4883         pmc_ss_count = 0;
4884
4885         /* allocate a pool of spin mutexes */
4886         pmc_mtxpool = mtx_pool_create("pmc-leaf", pmc_mtxpool_size,
4887             MTX_SPIN);
4888
4889         PMCDBG4(MOD,INI,1, "pmc_ownerhash=%p, mask=0x%lx "
4890             "targethash=%p mask=0x%lx", pmc_ownerhash, pmc_ownerhashmask,
4891             pmc_processhash, pmc_processhashmask);
4892
4893         /* register process {exit,fork,exec} handlers */
4894         pmc_exit_tag = EVENTHANDLER_REGISTER(process_exit,
4895             pmc_process_exit, NULL, EVENTHANDLER_PRI_ANY);
4896         pmc_fork_tag = EVENTHANDLER_REGISTER(process_fork,
4897             pmc_process_fork, NULL, EVENTHANDLER_PRI_ANY);
4898
4899         /* register kld event handlers */
4900         pmc_kld_load_tag = EVENTHANDLER_REGISTER(kld_load, pmc_kld_load,
4901             NULL, EVENTHANDLER_PRI_ANY);
4902         pmc_kld_unload_tag = EVENTHANDLER_REGISTER(kld_unload, pmc_kld_unload,
4903             NULL, EVENTHANDLER_PRI_ANY);
4904
4905         /* initialize logging */
4906         pmclog_initialize();
4907
4908         /* set hook functions */
4909         pmc_intr = md->pmd_intr;
4910         pmc_hook = pmc_hook_handler;
4911
4912         if (error == 0) {
4913                 printf(PMC_MODULE_NAME ":");
4914                 for (n = 0; n < (int) md->pmd_nclass; n++) {
4915                         pcd = &md->pmd_classdep[n];
4916                         printf(" %s/%d/%d/0x%b",
4917                             pmc_name_of_pmcclass[pcd->pcd_class],
4918                             pcd->pcd_num,
4919                             pcd->pcd_width,
4920                             pcd->pcd_caps,
4921                             "\20"
4922                             "\1INT\2USR\3SYS\4EDG\5THR"
4923                             "\6REA\7WRI\10INV\11QUA\12PRC"
4924                             "\13TAG\14CSC");
4925                 }
4926                 printf("\n");
4927         }
4928
4929         return (error);
4930 }
4931
4932 /* prepare to be unloaded */
4933 static void
4934 pmc_cleanup(void)
4935 {
4936         int c, cpu;
4937         unsigned int maxcpu;
4938         struct pmc_ownerhash *ph;
4939         struct pmc_owner *po, *tmp;
4940         struct pmc_binding pb;
4941 #ifdef  HWPMC_DEBUG
4942         struct pmc_processhash *prh;
4943 #endif
4944
4945         PMCDBG0(MOD,INI,0, "cleanup");
4946
4947         /* switch off sampling */
4948         CPU_ZERO(&pmc_cpumask);
4949         pmc_intr = NULL;
4950
4951         sx_xlock(&pmc_sx);
4952         if (pmc_hook == NULL) { /* being unloaded already */
4953                 sx_xunlock(&pmc_sx);
4954                 return;
4955         }
4956
4957         pmc_hook = NULL; /* prevent new threads from entering module */
4958
4959         /* deregister event handlers */
4960         EVENTHANDLER_DEREGISTER(process_fork, pmc_fork_tag);
4961         EVENTHANDLER_DEREGISTER(process_exit, pmc_exit_tag);
4962         EVENTHANDLER_DEREGISTER(kld_load, pmc_kld_load_tag);
4963         EVENTHANDLER_DEREGISTER(kld_unload, pmc_kld_unload_tag);
4964
4965         /* send SIGBUS to all owner threads, free up allocations */
4966         if (pmc_ownerhash)
4967                 for (ph = pmc_ownerhash;
4968                      ph <= &pmc_ownerhash[pmc_ownerhashmask];
4969                      ph++) {
4970                         LIST_FOREACH_SAFE(po, ph, po_next, tmp) {
4971                                 pmc_remove_owner(po);
4972
4973                                 /* send SIGBUS to owner processes */
4974                                 PMCDBG3(MOD,INI,2, "cleanup signal proc=%p "
4975                                     "(%d, %s)", po->po_owner,
4976                                     po->po_owner->p_pid,
4977                                     po->po_owner->p_comm);
4978
4979                                 PROC_LOCK(po->po_owner);
4980                                 kern_psignal(po->po_owner, SIGBUS);
4981                                 PROC_UNLOCK(po->po_owner);
4982
4983                                 pmc_destroy_owner_descriptor(po);
4984                         }
4985                 }
4986
4987         /* reclaim allocated data structures */
4988         if (pmc_mtxpool)
4989                 mtx_pool_destroy(&pmc_mtxpool);
4990
4991         mtx_destroy(&pmc_processhash_mtx);
4992         if (pmc_processhash) {
4993 #ifdef  HWPMC_DEBUG
4994                 struct pmc_process *pp;
4995
4996                 PMCDBG0(MOD,INI,3, "destroy process hash");
4997                 for (prh = pmc_processhash;
4998                      prh <= &pmc_processhash[pmc_processhashmask];
4999                      prh++)
5000                         LIST_FOREACH(pp, prh, pp_next)
5001                             PMCDBG1(MOD,INI,3, "pid=%d", pp->pp_proc->p_pid);
5002 #endif
5003
5004                 hashdestroy(pmc_processhash, M_PMC, pmc_processhashmask);
5005                 pmc_processhash = NULL;
5006         }
5007
5008         if (pmc_ownerhash) {
5009                 PMCDBG0(MOD,INI,3, "destroy owner hash");
5010                 hashdestroy(pmc_ownerhash, M_PMC, pmc_ownerhashmask);
5011                 pmc_ownerhash = NULL;
5012         }
5013
5014         KASSERT(LIST_EMPTY(&pmc_ss_owners),
5015             ("[pmc,%d] Global SS owner list not empty", __LINE__));
5016         KASSERT(pmc_ss_count == 0,
5017             ("[pmc,%d] Global SS count not empty", __LINE__));
5018
5019         /* do processor and pmc-class dependent cleanup */
5020         maxcpu = pmc_cpu_max();
5021
5022         PMCDBG0(MOD,INI,3, "md cleanup");
5023         if (md) {
5024                 pmc_save_cpu_binding(&pb);
5025                 for (cpu = 0; cpu < maxcpu; cpu++) {
5026                         PMCDBG2(MOD,INI,1,"pmc-cleanup cpu=%d pcs=%p",
5027                             cpu, pmc_pcpu[cpu]);
5028                         if (!pmc_cpu_is_active(cpu) || pmc_pcpu[cpu] == NULL)
5029                                 continue;
5030                         pmc_select_cpu(cpu);
5031                         for (c = 0; c < md->pmd_nclass; c++)
5032                                 md->pmd_classdep[c].pcd_pcpu_fini(md, cpu);
5033                         if (md->pmd_pcpu_fini)
5034                                 md->pmd_pcpu_fini(md, cpu);
5035                 }
5036
5037                 if (md->pmd_cputype == PMC_CPU_GENERIC)
5038                         pmc_generic_cpu_finalize(md);
5039                 else
5040                         pmc_md_finalize(md);
5041
5042                 pmc_mdep_free(md);
5043                 md = NULL;
5044                 pmc_restore_cpu_binding(&pb);
5045         }
5046
5047         /* Free per-cpu descriptors. */
5048         for (cpu = 0; cpu < maxcpu; cpu++) {
5049                 if (!pmc_cpu_is_active(cpu))
5050                         continue;
5051                 KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_HR] != NULL,
5052                     ("[pmc,%d] Null hw cpu sample buffer cpu=%d", __LINE__,
5053                         cpu));
5054                 KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_SR] != NULL,
5055                     ("[pmc,%d] Null sw cpu sample buffer cpu=%d", __LINE__,
5056                         cpu));
5057                 free(pmc_pcpu[cpu]->pc_sb[PMC_HR]->ps_callchains, M_PMC);
5058                 free(pmc_pcpu[cpu]->pc_sb[PMC_HR], M_PMC);
5059                 free(pmc_pcpu[cpu]->pc_sb[PMC_SR]->ps_callchains, M_PMC);
5060                 free(pmc_pcpu[cpu]->pc_sb[PMC_SR], M_PMC);
5061                 free(pmc_pcpu[cpu], M_PMC);
5062         }
5063
5064         free(pmc_pcpu, M_PMC);
5065         pmc_pcpu = NULL;
5066
5067         free(pmc_pcpu_saved, M_PMC);
5068         pmc_pcpu_saved = NULL;
5069
5070         if (pmc_pmcdisp) {
5071                 free(pmc_pmcdisp, M_PMC);
5072                 pmc_pmcdisp = NULL;
5073         }
5074
5075         if (pmc_rowindex_to_classdep) {
5076                 free(pmc_rowindex_to_classdep, M_PMC);
5077                 pmc_rowindex_to_classdep = NULL;
5078         }
5079
5080         pmclog_shutdown();
5081
5082         sx_xunlock(&pmc_sx);    /* we are done */
5083 }
5084
5085 /*
5086  * The function called at load/unload.
5087  */
5088
5089 static int
5090 load (struct module *module __unused, int cmd, void *arg __unused)
5091 {
5092         int error;
5093
5094         error = 0;
5095
5096         switch (cmd) {
5097         case MOD_LOAD :
5098                 /* initialize the subsystem */
5099                 error = pmc_initialize();
5100                 if (error != 0)
5101                         break;
5102                 PMCDBG2(MOD,INI,1, "syscall=%d maxcpu=%d",
5103                     pmc_syscall_num, pmc_cpu_max());
5104                 break;
5105
5106
5107         case MOD_UNLOAD :
5108         case MOD_SHUTDOWN:
5109                 pmc_cleanup();
5110                 PMCDBG0(MOD,INI,1, "unloaded");
5111                 break;
5112
5113         default :
5114                 error = EINVAL; /* XXX should panic(9) */
5115                 break;
5116         }
5117
5118         return error;
5119 }
5120
5121 /* memory pool */
5122 MALLOC_DEFINE(M_PMC, "pmc", "Memory space for the PMC module");