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