sys/dev/hwpmc/hwpmc_mod.c

   1 /*-
   2  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
   3  *
   4  * Copyright (c) 2003-2008 Joseph Koshy
   5  * Copyright (c) 2007 The FreeBSD Foundation
   6  * Copyright (c) 2018 Matthew Macy
   7  * All rights reserved.
   8  *
   9  * Portions of this software were developed by A. Joseph Koshy under
  10  * sponsorship from the FreeBSD Foundation and Google, Inc.
  11  *
  12  * Redistribution and use in source and binary forms, with or without
  13  * modification, are permitted provided that the following conditions
  14  * are met:
  15  * 1. Redistributions of source code must retain the above copyright
  16  *    notice, this list of conditions and the following disclaimer.
  17  * 2. Redistributions in binary form must reproduce the above copyright
  18  *    notice, this list of conditions and the following disclaimer in the
  19  *    documentation and/or other materials provided with the distribution.
  20  *
  21  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
  22  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
  23  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
  24  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
  25  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
  26  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
  27  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
  28  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
  29  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
  30  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  31  * SUCH DAMAGE.
  32  *
  33  */
  34
  35 #include <sys/cdefs.h>
  36 __FBSDID("$FreeBSD$");
  37
  38 #include <sys/param.h>
  39 #include <sys/systm.h>
  40 #include <sys/domainset.h>
  41 #include <sys/eventhandler.h>
  42 #include <sys/jail.h>
  43 #include <sys/kernel.h>
  44 #include <sys/kthread.h>
  45 #include <sys/limits.h>
  46 #include <sys/lock.h>
  47 #include <sys/malloc.h>
  48 #include <sys/module.h>
  49 #include <sys/mount.h>
  50 #include <sys/mutex.h>
  51 #include <sys/pmc.h>
  52 #include <sys/pmckern.h>
  53 #include <sys/pmclog.h>
  54 #include <sys/priv.h>
  55 #include <sys/proc.h>
  56 #include <sys/queue.h>
  57 #include <sys/resourcevar.h>
  58 #include <sys/rwlock.h>
  59 #include <sys/sched.h>
  60 #include <sys/signalvar.h>
  61 #include <sys/smp.h>
  62 #include <sys/sx.h>
  63 #include <sys/sysctl.h>
  64 #include <sys/sysent.h>
  65 #include <sys/syslog.h>
  66 #include <sys/taskqueue.h>
  67 #include <sys/vnode.h>
  68
  69 #include <sys/linker.h>         /* needs to be after <sys/malloc.h> */
  70
  71 #include <machine/atomic.h>
  72 #include <machine/md_var.h>
  73
  74 #include <vm/vm.h>
  75 #include <vm/vm_extern.h>
  76 #include <vm/pmap.h>
  77 #include <vm/vm_map.h>
  78 #include <vm/vm_object.h>
  79
  80 #include "hwpmc_soft.h"
  81
  82 #define PMC_EPOCH_ENTER() struct epoch_tracker pmc_et; epoch_enter_preempt(global_epoch_preempt, &pmc_et)
  83 #define PMC_EPOCH_EXIT() epoch_exit_preempt(global_epoch_preempt, &pmc_et)
  84
  85 /*
  86  * Types
  87  */
  88
  89 enum pmc_flags {
  90         PMC_FLAG_NONE     = 0x00, /* do nothing */
  91         PMC_FLAG_REMOVE   = 0x01, /* atomically remove entry from hash */
  92         PMC_FLAG_ALLOCATE = 0x02, /* add entry to hash if not found */
  93         PMC_FLAG_NOWAIT   = 0x04, /* do not wait for mallocs */
  94 };
  95
  96 /*
  97  * The offset in sysent where the syscall is allocated.
  98  */
  99
 100 static int pmc_syscall_num = NO_SYSCALL;
 101 struct pmc_cpu          **pmc_pcpu;      /* per-cpu state */
 102 pmc_value_t             *pmc_pcpu_saved; /* saved PMC values: CSW handling */
 103
 104 #define PMC_PCPU_SAVED(C,R)     pmc_pcpu_saved[(R) + md->pmd_npmc*(C)]
 105
 106 struct mtx_pool         *pmc_mtxpool;
 107 static int              *pmc_pmcdisp;    /* PMC row dispositions */
 108
 109 #define PMC_ROW_DISP_IS_FREE(R)         (pmc_pmcdisp[(R)] == 0)
 110 #define PMC_ROW_DISP_IS_THREAD(R)       (pmc_pmcdisp[(R)] > 0)
 111 #define PMC_ROW_DISP_IS_STANDALONE(R)   (pmc_pmcdisp[(R)] < 0)
 112
 113 #define PMC_MARK_ROW_FREE(R) do {                                         \
 114         pmc_pmcdisp[(R)] = 0;                                             \
 115 } while (0)
 116
 117 #define PMC_MARK_ROW_STANDALONE(R) do {                                   \
 118         KASSERT(pmc_pmcdisp[(R)] <= 0, ("[pmc,%d] row disposition error", \
 119                     __LINE__));                                           \
 120         atomic_add_int(&pmc_pmcdisp[(R)], -1);                            \
 121         KASSERT(pmc_pmcdisp[(R)] >= (-pmc_cpu_max_active()),              \
 122                 ("[pmc,%d] row disposition error", __LINE__));            \
 123 } while (0)
 124
 125 #define PMC_UNMARK_ROW_STANDALONE(R) do {                                 \
 126         atomic_add_int(&pmc_pmcdisp[(R)], 1);                             \
 127         KASSERT(pmc_pmcdisp[(R)] <= 0, ("[pmc,%d] row disposition error", \
 128                     __LINE__));                                           \
 129 } while (0)
 130
 131 #define PMC_MARK_ROW_THREAD(R) do {                                       \
 132         KASSERT(pmc_pmcdisp[(R)] >= 0, ("[pmc,%d] row disposition error", \
 133                     __LINE__));                                           \
 134         atomic_add_int(&pmc_pmcdisp[(R)], 1);                             \
 135 } while (0)
 136
 137 #define PMC_UNMARK_ROW_THREAD(R) do {                                     \
 138         atomic_add_int(&pmc_pmcdisp[(R)], -1);                            \
 139         KASSERT(pmc_pmcdisp[(R)] >= 0, ("[pmc,%d] row disposition error", \
 140                     __LINE__));                                           \
 141 } while (0)
 142
 143
 144 /* various event handlers */
 145 static eventhandler_tag pmc_exit_tag, pmc_fork_tag, pmc_kld_load_tag,
 146     pmc_kld_unload_tag;
 147
 148 /* Module statistics */
 149 struct pmc_driverstats pmc_stats;
 150
 151
 152 /* Machine/processor dependent operations */
 153 static struct pmc_mdep  *md;
 154
 155 /*
 156  * Hash tables mapping owner processes and target threads to PMCs.
 157  */
 158
 159 struct mtx pmc_processhash_mtx;         /* spin mutex */
 160 static u_long pmc_processhashmask;
 161 static LIST_HEAD(pmc_processhash, pmc_process)  *pmc_processhash;
 162
 163 /*
 164  * Hash table of PMC owner descriptors.  This table is protected by
 165  * the shared PMC "sx" lock.
 166  */
 167
 168 static u_long pmc_ownerhashmask;
 169 static LIST_HEAD(pmc_ownerhash, pmc_owner)      *pmc_ownerhash;
 170
 171 /*
 172  * List of PMC owners with system-wide sampling PMCs.
 173  */
 174
 175 static CK_LIST_HEAD(, pmc_owner)                        pmc_ss_owners;
 176
 177 /*
 178  * List of free thread entries. This is protected by the spin
 179  * mutex.
 180  */
 181 static struct mtx pmc_threadfreelist_mtx;       /* spin mutex */
 182 static LIST_HEAD(, pmc_thread)                  pmc_threadfreelist;
 183 static int pmc_threadfreelist_entries=0;
 184 #define THREADENTRY_SIZE                                                \
 185 (sizeof(struct pmc_thread) + (md->pmd_npmc * sizeof(struct pmc_threadpmcstate)))
 186
 187 /*
 188  * Task to free thread descriptors
 189  */
 190 static struct task free_task;
 191
 192 /*
 193  * A map of row indices to classdep structures.
 194  */
 195 static struct pmc_classdep **pmc_rowindex_to_classdep;
 196
 197 /*
 198  * Prototypes
 199  */
 200
 201 #ifdef  HWPMC_DEBUG
 202 static int      pmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS);
 203 static int      pmc_debugflags_parse(char *newstr, char *fence);
 204 #endif
 205
 206 static int      load(struct module *module, int cmd, void *arg);
 207 static int      pmc_add_sample(ring_type_t ring, struct pmc *pm, struct trapframe *tf);
 208 static void     pmc_add_thread_descriptors_from_proc(struct proc *p,
 209     struct pmc_process *pp);
 210 static int      pmc_attach_process(struct proc *p, struct pmc *pm);
 211 static struct pmc *pmc_allocate_pmc_descriptor(void);
 212 static struct pmc_owner *pmc_allocate_owner_descriptor(struct proc *p);
 213 static int      pmc_attach_one_process(struct proc *p, struct pmc *pm);
 214 static int      pmc_can_allocate_rowindex(struct proc *p, unsigned int ri,
 215     int cpu);
 216 static int      pmc_can_attach(struct pmc *pm, struct proc *p);
 217 static void     pmc_capture_user_callchain(int cpu, int soft, struct trapframe *tf);
 218 static void     pmc_cleanup(void);
 219 static int      pmc_detach_process(struct proc *p, struct pmc *pm);
 220 static int      pmc_detach_one_process(struct proc *p, struct pmc *pm,
 221     int flags);
 222 static void     pmc_destroy_owner_descriptor(struct pmc_owner *po);
 223 static void     pmc_destroy_pmc_descriptor(struct pmc *pm);
 224 static void     pmc_destroy_process_descriptor(struct pmc_process *pp);
 225 static struct pmc_owner *pmc_find_owner_descriptor(struct proc *p);
 226 static int      pmc_find_pmc(pmc_id_t pmcid, struct pmc **pm);
 227 static struct pmc *pmc_find_pmc_descriptor_in_process(struct pmc_owner *po,
 228     pmc_id_t pmc);
 229 static struct pmc_process *pmc_find_process_descriptor(struct proc *p,
 230     uint32_t mode);
 231 static struct pmc_thread *pmc_find_thread_descriptor(struct pmc_process *pp,
 232     struct thread *td, uint32_t mode);
 233 static void     pmc_force_context_switch(void);
 234 static void     pmc_link_target_process(struct pmc *pm,
 235     struct pmc_process *pp);
 236 static void     pmc_log_all_process_mappings(struct pmc_owner *po);
 237 static void     pmc_log_kernel_mappings(struct pmc *pm);
 238 static void     pmc_log_process_mappings(struct pmc_owner *po, struct proc *p);
 239 static void     pmc_maybe_remove_owner(struct pmc_owner *po);
 240 static void     pmc_process_csw_in(struct thread *td);
 241 static void     pmc_process_csw_out(struct thread *td);
 242 static void     pmc_process_exit(void *arg, struct proc *p);
 243 static void     pmc_process_fork(void *arg, struct proc *p1,
 244     struct proc *p2, int n);
 245 static void     pmc_process_samples(int cpu, ring_type_t soft);
 246 static void     pmc_release_pmc_descriptor(struct pmc *pmc);
 247 static void     pmc_process_thread_add(struct thread *td);
 248 static void     pmc_process_thread_delete(struct thread *td);
 249 static void     pmc_process_thread_userret(struct thread *td);
 250 static void     pmc_remove_owner(struct pmc_owner *po);
 251 static void     pmc_remove_process_descriptor(struct pmc_process *pp);
 252 static void     pmc_restore_cpu_binding(struct pmc_binding *pb);
 253 static void     pmc_save_cpu_binding(struct pmc_binding *pb);
 254 static void     pmc_select_cpu(int cpu);
 255 static int      pmc_start(struct pmc *pm);
 256 static int      pmc_stop(struct pmc *pm);
 257 static int      pmc_syscall_handler(struct thread *td, void *syscall_args);
 258 static struct pmc_thread *pmc_thread_descriptor_pool_alloc(void);
 259 static void     pmc_thread_descriptor_pool_drain(void);
 260 static void     pmc_thread_descriptor_pool_free(struct pmc_thread *pt);
 261 static void     pmc_unlink_target_process(struct pmc *pmc,
 262     struct pmc_process *pp);
 263 static int generic_switch_in(struct pmc_cpu *pc, struct pmc_process *pp);
 264 static int generic_switch_out(struct pmc_cpu *pc, struct pmc_process *pp);
 265 static struct pmc_mdep *pmc_generic_cpu_initialize(void);
 266 static void pmc_generic_cpu_finalize(struct pmc_mdep *md);
 267 static void pmc_post_callchain_callback(void);
 268 static void pmc_process_threadcreate(struct thread *td);
 269 static void pmc_process_threadexit(struct thread *td);
 270 static void pmc_process_proccreate(struct proc *p);
 271 static void pmc_process_allproc(struct pmc *pm);
 272
 273 /*
 274  * Kernel tunables and sysctl(8) interface.
 275  */
 276
 277 SYSCTL_DECL(_kern_hwpmc);
 278 SYSCTL_NODE(_kern_hwpmc, OID_AUTO, stats, CTLFLAG_RW, 0, "HWPMC stats");
 279
 280
 281 /* Stats. */
 282 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, intr_ignored, CTLFLAG_RW,
 283                                    &pmc_stats.pm_intr_ignored, "# of interrupts ignored");
 284 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, intr_processed, CTLFLAG_RW,
 285                                    &pmc_stats.pm_intr_processed, "# of interrupts processed");
 286 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, intr_bufferfull, CTLFLAG_RW,
 287                                    &pmc_stats.pm_intr_bufferfull, "# of interrupts where buffer was full");
 288 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, syscalls, CTLFLAG_RW,
 289                                    &pmc_stats.pm_syscalls, "# of syscalls");
 290 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, syscall_errors, CTLFLAG_RW,
 291                                    &pmc_stats.pm_syscall_errors, "# of syscall_errors");
 292 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, buffer_requests, CTLFLAG_RW,
 293                                    &pmc_stats.pm_buffer_requests, "# of buffer requests");
 294 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, buffer_requests_failed, CTLFLAG_RW,
 295                                    &pmc_stats.pm_buffer_requests_failed, "# of buffer requests which failed");
 296 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, log_sweeps, CTLFLAG_RW,
 297                                    &pmc_stats.pm_log_sweeps, "# of ?");
 298 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, merges, CTLFLAG_RW,
 299                                    &pmc_stats.pm_merges, "# of times kernel stack was found for user trace");
 300 SYSCTL_COUNTER_U64(_kern_hwpmc_stats, OID_AUTO, overwrites, CTLFLAG_RW,
 301                                    &pmc_stats.pm_overwrites, "# of times a sample was overwritten before being logged");
 302
 303 static int pmc_callchaindepth = PMC_CALLCHAIN_DEPTH;
 304 SYSCTL_INT(_kern_hwpmc, OID_AUTO, callchaindepth, CTLFLAG_RDTUN,
 305     &pmc_callchaindepth, 0, "depth of call chain records");
 306
 307 char pmc_cpuid[PMC_CPUID_LEN];
 308 SYSCTL_STRING(_kern_hwpmc, OID_AUTO, cpuid, CTLFLAG_RD,
 309         pmc_cpuid, 0, "cpu version string");
 310 #ifdef  HWPMC_DEBUG
 311 struct pmc_debugflags pmc_debugflags = PMC_DEBUG_DEFAULT_FLAGS;
 312 char    pmc_debugstr[PMC_DEBUG_STRSIZE];
 313 TUNABLE_STR(PMC_SYSCTL_NAME_PREFIX "debugflags", pmc_debugstr,
 314     sizeof(pmc_debugstr));
 315 SYSCTL_PROC(_kern_hwpmc, OID_AUTO, debugflags,
 316     CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NOFETCH,
 317     0, 0, pmc_debugflags_sysctl_handler, "A", "debug flags");
 318 #endif
 319
 320
 321 /*
 322  * kern.hwpmc.hashrows -- determines the number of rows in the
 323  * of the hash table used to look up threads
 324  */
 325
 326 static int pmc_hashsize = PMC_HASH_SIZE;
 327 SYSCTL_INT(_kern_hwpmc, OID_AUTO, hashsize, CTLFLAG_RDTUN,
 328     &pmc_hashsize, 0, "rows in hash tables");
 329
 330 /*
 331  * kern.hwpmc.nsamples --- number of PC samples/callchain stacks per CPU
 332  */
 333
 334 static int pmc_nsamples = PMC_NSAMPLES;
 335 SYSCTL_INT(_kern_hwpmc, OID_AUTO, nsamples, CTLFLAG_RDTUN,
 336     &pmc_nsamples, 0, "number of PC samples per CPU");
 337
 338 static uint64_t pmc_sample_mask = PMC_NSAMPLES-1;
 339
 340 /*
 341  * kern.hwpmc.mtxpoolsize -- number of mutexes in the mutex pool.
 342  */
 343
 344 static int pmc_mtxpool_size = PMC_MTXPOOL_SIZE;
 345 SYSCTL_INT(_kern_hwpmc, OID_AUTO, mtxpoolsize, CTLFLAG_RDTUN,
 346     &pmc_mtxpool_size, 0, "size of spin mutex pool");
 347
 348
 349 /*
 350  * kern.hwpmc.threadfreelist_entries -- number of free entries
 351  */
 352
 353 SYSCTL_INT(_kern_hwpmc, OID_AUTO, threadfreelist_entries, CTLFLAG_RD,
 354     &pmc_threadfreelist_entries, 0, "number of avalable thread entries");
 355
 356
 357 /*
 358  * kern.hwpmc.threadfreelist_max -- maximum number of free entries
 359  */
 360
 361 static int pmc_threadfreelist_max = PMC_THREADLIST_MAX;
 362 SYSCTL_INT(_kern_hwpmc, OID_AUTO, threadfreelist_max, CTLFLAG_RW,
 363     &pmc_threadfreelist_max, 0,
 364     "maximum number of available thread entries before freeing some");
 365
 366
 367 /*
 368  * security.bsd.unprivileged_syspmcs -- allow non-root processes to
 369  * allocate system-wide PMCs.
 370  *
 371  * Allowing unprivileged processes to allocate system PMCs is convenient
 372  * if system-wide measurements need to be taken concurrently with other
 373  * per-process measurements.  This feature is turned off by default.
 374  */
 375
 376 static int pmc_unprivileged_syspmcs = 0;
 377 SYSCTL_INT(_security_bsd, OID_AUTO, unprivileged_syspmcs, CTLFLAG_RWTUN,
 378     &pmc_unprivileged_syspmcs, 0,
 379     "allow unprivileged process to allocate system PMCs");
 380
 381 /*
 382  * Hash function.  Discard the lower 2 bits of the pointer since
 383  * these are always zero for our uses.  The hash multiplier is
 384  * round((2^LONG_BIT) * ((sqrt(5)-1)/2)).
 385  */
 386
 387 #if     LONG_BIT == 64
 388 #define _PMC_HM         11400714819323198486u
 389 #elif   LONG_BIT == 32
 390 #define _PMC_HM         2654435769u
 391 #else
 392 #error  Must know the size of 'long' to compile
 393 #endif
 394
 395 #define PMC_HASH_PTR(P,M)       ((((unsigned long) (P) >> 2) * _PMC_HM) & (M))
 396
 397 /*
 398  * Syscall structures
 399  */
 400
 401 /* The `sysent' for the new syscall */
 402 static struct sysent pmc_sysent = {
 403         .sy_narg =      2,
 404         .sy_call =      pmc_syscall_handler,
 405 };
 406
 407 static struct syscall_module_data pmc_syscall_mod = {
 408         .chainevh =     load,
 409         .chainarg =     NULL,
 410         .offset =       &pmc_syscall_num,
 411         .new_sysent =   &pmc_sysent,
 412         .old_sysent =   { .sy_narg = 0, .sy_call = NULL },
 413         .flags =        SY_THR_STATIC_KLD,
 414 };
 415
 416 static moduledata_t pmc_mod = {
 417         .name =         PMC_MODULE_NAME,
 418         .evhand =       syscall_module_handler,
 419         .priv =         &pmc_syscall_mod,
 420 };
 421
 422 #ifdef EARLY_AP_STARTUP
 423 DECLARE_MODULE(pmc, pmc_mod, SI_SUB_SYSCALLS, SI_ORDER_ANY);
 424 #else
 425 DECLARE_MODULE(pmc, pmc_mod, SI_SUB_SMP, SI_ORDER_ANY);
 426 #endif
 427 MODULE_VERSION(pmc, PMC_VERSION);
 428
 429 #ifdef  HWPMC_DEBUG
 430 enum pmc_dbgparse_state {
 431         PMCDS_WS,               /* in whitespace */
 432         PMCDS_MAJOR,            /* seen a major keyword */
 433         PMCDS_MINOR
 434 };
 435
 436 static int
 437 pmc_debugflags_parse(char *newstr, char *fence)
 438 {
 439         char c, *p, *q;
 440         struct pmc_debugflags *tmpflags;
 441         int error, found, *newbits, tmp;
 442         size_t kwlen;
 443
 444         tmpflags = malloc(sizeof(*tmpflags), M_PMC, M_WAITOK|M_ZERO);
 445
 446         p = newstr;
 447         error = 0;
 448
 449         for (; p < fence && (c = *p); p++) {
 450
 451                 /* skip white space */
 452                 if (c == ' ' || c == '\t')
 453                         continue;
 454
 455                 /* look for a keyword followed by "=" */
 456                 for (q = p; p < fence && (c = *p) && c != '='; p++)
 457                         ;
 458                 if (c != '=') {
 459                         error = EINVAL;
 460                         goto done;
 461                 }
 462
 463                 kwlen = p - q;
 464                 newbits = NULL;
 465
 466                 /* lookup flag group name */
 467 #define DBG_SET_FLAG_MAJ(S,F)                                           \
 468                 if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0)  \
 469                         newbits = &tmpflags->pdb_ ## F;
 470
 471                 DBG_SET_FLAG_MAJ("cpu",         CPU);
 472                 DBG_SET_FLAG_MAJ("csw",         CSW);
 473                 DBG_SET_FLAG_MAJ("logging",     LOG);
 474                 DBG_SET_FLAG_MAJ("module",      MOD);
 475                 DBG_SET_FLAG_MAJ("md",          MDP);
 476                 DBG_SET_FLAG_MAJ("owner",       OWN);
 477                 DBG_SET_FLAG_MAJ("pmc",         PMC);
 478                 DBG_SET_FLAG_MAJ("process",     PRC);
 479                 DBG_SET_FLAG_MAJ("sampling",    SAM);
 480
 481                 if (newbits == NULL) {
 482                         error = EINVAL;
 483                         goto done;
 484                 }
 485
 486                 p++;            /* skip the '=' */
 487
 488                 /* Now parse the individual flags */
 489                 tmp = 0;
 490         newflag:
 491                 for (q = p; p < fence && (c = *p); p++)
 492                         if (c == ' ' || c == '\t' || c == ',')
 493                                 break;
 494
 495                 /* p == fence or c == ws or c == "," or c == 0 */
 496
 497                 if ((kwlen = p - q) == 0) {
 498                         *newbits = tmp;
 499                         continue;
 500                 }
 501
 502                 found = 0;
 503 #define DBG_SET_FLAG_MIN(S,F)                                           \
 504                 if (kwlen == sizeof(S)-1 && strncmp(q, S, kwlen) == 0)  \
 505                         tmp |= found = (1 << PMC_DEBUG_MIN_ ## F)
 506
 507                 /* a '*' denotes all possible flags in the group */
 508                 if (kwlen == 1 && *q == '*')
 509                         tmp = found = ~0;
 510                 /* look for individual flag names */
 511                 DBG_SET_FLAG_MIN("allocaterow", ALR);
 512                 DBG_SET_FLAG_MIN("allocate",    ALL);
 513                 DBG_SET_FLAG_MIN("attach",      ATT);
 514                 DBG_SET_FLAG_MIN("bind",        BND);
 515                 DBG_SET_FLAG_MIN("config",      CFG);
 516                 DBG_SET_FLAG_MIN("exec",        EXC);
 517                 DBG_SET_FLAG_MIN("exit",        EXT);
 518                 DBG_SET_FLAG_MIN("find",        FND);
 519                 DBG_SET_FLAG_MIN("flush",       FLS);
 520                 DBG_SET_FLAG_MIN("fork",        FRK);
 521                 DBG_SET_FLAG_MIN("getbuf",      GTB);
 522                 DBG_SET_FLAG_MIN("hook",        PMH);
 523                 DBG_SET_FLAG_MIN("init",        INI);
 524                 DBG_SET_FLAG_MIN("intr",        INT);
 525                 DBG_SET_FLAG_MIN("linktarget",  TLK);
 526                 DBG_SET_FLAG_MIN("mayberemove", OMR);
 527                 DBG_SET_FLAG_MIN("ops",         OPS);
 528                 DBG_SET_FLAG_MIN("read",        REA);
 529                 DBG_SET_FLAG_MIN("register",    REG);
 530                 DBG_SET_FLAG_MIN("release",     REL);
 531                 DBG_SET_FLAG_MIN("remove",      ORM);
 532                 DBG_SET_FLAG_MIN("sample",      SAM);
 533                 DBG_SET_FLAG_MIN("scheduleio",  SIO);
 534                 DBG_SET_FLAG_MIN("select",      SEL);
 535                 DBG_SET_FLAG_MIN("signal",      SIG);
 536                 DBG_SET_FLAG_MIN("swi",         SWI);
 537                 DBG_SET_FLAG_MIN("swo",         SWO);
 538                 DBG_SET_FLAG_MIN("start",       STA);
 539                 DBG_SET_FLAG_MIN("stop",        STO);
 540                 DBG_SET_FLAG_MIN("syscall",     PMS);
 541                 DBG_SET_FLAG_MIN("unlinktarget", TUL);
 542                 DBG_SET_FLAG_MIN("write",       WRI);
 543                 if (found == 0) {
 544                         /* unrecognized flag name */
 545                         error = EINVAL;
 546                         goto done;
 547                 }
 548
 549                 if (c == 0 || c == ' ' || c == '\t') {  /* end of flag group */
 550                         *newbits = tmp;
 551                         continue;
 552                 }
 553
 554                 p++;
 555                 goto newflag;
 556         }
 557
 558         /* save the new flag set */
 559         bcopy(tmpflags, &pmc_debugflags, sizeof(pmc_debugflags));
 560
 561  done:
 562         free(tmpflags, M_PMC);
 563         return error;
 564 }
 565
 566 static int
 567 pmc_debugflags_sysctl_handler(SYSCTL_HANDLER_ARGS)
 568 {
 569         char *fence, *newstr;
 570         int error;
 571         unsigned int n;
 572
 573         (void) arg1; (void) arg2; /* unused parameters */
 574
 575         n = sizeof(pmc_debugstr);
 576         newstr = malloc(n, M_PMC, M_WAITOK|M_ZERO);
 577         (void) strlcpy(newstr, pmc_debugstr, n);
 578
 579         error = sysctl_handle_string(oidp, newstr, n, req);
 580
 581         /* if there is a new string, parse and copy it */
 582         if (error == 0 && req->newptr != NULL) {
 583                 fence = newstr + (n < req->newlen ? n : req->newlen + 1);
 584                 if ((error = pmc_debugflags_parse(newstr, fence)) == 0)
 585                         (void) strlcpy(pmc_debugstr, newstr,
 586                             sizeof(pmc_debugstr));
 587         }
 588
 589         free(newstr, M_PMC);
 590
 591         return error;
 592 }
 593 #endif
 594
 595 /*
 596  * Map a row index to a classdep structure and return the adjusted row
 597  * index for the PMC class index.
 598  */
 599 static struct pmc_classdep *
 600 pmc_ri_to_classdep(struct pmc_mdep *md, int ri, int *adjri)
 601 {
 602         struct pmc_classdep *pcd;
 603
 604         (void) md;
 605
 606         KASSERT(ri >= 0 && ri < md->pmd_npmc,
 607             ("[pmc,%d] illegal row-index %d", __LINE__, ri));
 608
 609         pcd = pmc_rowindex_to_classdep[ri];
 610
 611         KASSERT(pcd != NULL,
 612             ("[pmc,%d] ri %d null pcd", __LINE__, ri));
 613
 614         *adjri = ri - pcd->pcd_ri;
 615
 616         KASSERT(*adjri >= 0 && *adjri < pcd->pcd_num,
 617             ("[pmc,%d] adjusted row-index %d", __LINE__, *adjri));
 618
 619         return (pcd);
 620 }
 621
 622 /*
 623  * Concurrency Control
 624  *
 625  * The driver manages the following data structures:
 626  *
 627  *   - target process descriptors, one per target process
 628  *   - owner process descriptors (and attached lists), one per owner process
 629  *   - lookup hash tables for owner and target processes
 630  *   - PMC descriptors (and attached lists)
 631  *   - per-cpu hardware state
 632  *   - the 'hook' variable through which the kernel calls into
 633  *     this module
 634  *   - the machine hardware state (managed by the MD layer)
 635  *
 636  * These data structures are accessed from:
 637  *
 638  * - thread context-switch code
 639  * - interrupt handlers (possibly on multiple cpus)
 640  * - kernel threads on multiple cpus running on behalf of user
 641  *   processes doing system calls
 642  * - this driver's private kernel threads
 643  *
 644  * = Locks and Locking strategy =
 645  *
 646  * The driver uses four locking strategies for its operation:
 647  *
 648  * - The global SX lock "pmc_sx" is used to protect internal
 649  *   data structures.
 650  *
 651  *   Calls into the module by syscall() start with this lock being
 652  *   held in exclusive mode.  Depending on the requested operation,
 653  *   the lock may be downgraded to 'shared' mode to allow more
 654  *   concurrent readers into the module.  Calls into the module from
 655  *   other parts of the kernel acquire the lock in shared mode.
 656  *
 657  *   This SX lock is held in exclusive mode for any operations that
 658  *   modify the linkages between the driver's internal data structures.
 659  *
 660  *   The 'pmc_hook' function pointer is also protected by this lock.
 661  *   It is only examined with the sx lock held in exclusive mode.  The
 662  *   kernel module is allowed to be unloaded only with the sx lock held
 663  *   in exclusive mode.  In normal syscall handling, after acquiring the
 664  *   pmc_sx lock we first check that 'pmc_hook' is non-null before
 665  *   proceeding.  This prevents races between the thread unloading the module
 666  *   and other threads seeking to use the module.
 667  *
 668  * - Lookups of target process structures and owner process structures
 669  *   cannot use the global "pmc_sx" SX lock because these lookups need
 670  *   to happen during context switches and in other critical sections
 671  *   where sleeping is not allowed.  We protect these lookup tables
 672  *   with their own private spin-mutexes, "pmc_processhash_mtx" and
 673  *   "pmc_ownerhash_mtx".
 674  *
 675  * - Interrupt handlers work in a lock free manner.  At interrupt
 676  *   time, handlers look at the PMC pointer (phw->phw_pmc) configured
 677  *   when the PMC was started.  If this pointer is NULL, the interrupt
 678  *   is ignored after updating driver statistics.  We ensure that this
 679  *   pointer is set (using an atomic operation if necessary) before the
 680  *   PMC hardware is started.  Conversely, this pointer is unset atomically
 681  *   only after the PMC hardware is stopped.
 682  *
 683  *   We ensure that everything needed for the operation of an
 684  *   interrupt handler is available without it needing to acquire any
 685  *   locks.  We also ensure that a PMC's software state is destroyed only
 686  *   after the PMC is taken off hardware (on all CPUs).
 687  *
 688  * - Context-switch handling with process-private PMCs needs more
 689  *   care.
 690  *
 691  *   A given process may be the target of multiple PMCs.  For example,
 692  *   PMCATTACH and PMCDETACH may be requested by a process on one CPU
 693  *   while the target process is running on another.  A PMC could also
 694  *   be getting released because its owner is exiting.  We tackle
 695  *   these situations in the following manner:
 696  *
 697  *   - each target process structure 'pmc_process' has an array
 698  *     of 'struct pmc *' pointers, one for each hardware PMC.
 699  *
 700  *   - At context switch IN time, each "target" PMC in RUNNING state
 701  *     gets started on hardware and a pointer to each PMC is copied into
 702  *     the per-cpu phw array.  The 'runcount' for the PMC is
 703  *     incremented.
 704  *
 705  *   - At context switch OUT time, all process-virtual PMCs are stopped
 706  *     on hardware.  The saved value is added to the PMCs value field
 707  *     only if the PMC is in a non-deleted state (the PMCs state could
 708  *     have changed during the current time slice).
 709  *
 710  *     Note that since in-between a switch IN on a processor and a switch
 711  *     OUT, the PMC could have been released on another CPU.  Therefore
 712  *     context switch OUT always looks at the hardware state to turn
 713  *     OFF PMCs and will update a PMC's saved value only if reachable
 714  *     from the target process record.
 715  *
 716  *   - OP PMCRELEASE could be called on a PMC at any time (the PMC could
 717  *     be attached to many processes at the time of the call and could
 718  *     be active on multiple CPUs).
 719  *
 720  *     We prevent further scheduling of the PMC by marking it as in
 721  *     state 'DELETED'.  If the runcount of the PMC is non-zero then
 722  *     this PMC is currently running on a CPU somewhere.  The thread
 723  *     doing the PMCRELEASE operation waits by repeatedly doing a
 724  *     pause() till the runcount comes to zero.
 725  *
 726  * The contents of a PMC descriptor (struct pmc) are protected using
 727  * a spin-mutex.  In order to save space, we use a mutex pool.
 728  *
 729  * In terms of lock types used by witness(4), we use:
 730  * - Type "pmc-sx", used by the global SX lock.
 731  * - Type "pmc-sleep", for sleep mutexes used by logger threads.
 732  * - Type "pmc-per-proc", for protecting PMC owner descriptors.
 733  * - Type "pmc-leaf", used for all other spin mutexes.
 734  */
 735
 736 /*
 737  * save the cpu binding of the current kthread
 738  */
 739
 740 static void
 741 pmc_save_cpu_binding(struct pmc_binding *pb)
 742 {
 743         PMCDBG0(CPU,BND,2, "save-cpu");
 744         thread_lock(curthread);
 745         pb->pb_bound = sched_is_bound(curthread);
 746         pb->pb_cpu   = curthread->td_oncpu;
 747         thread_unlock(curthread);
 748         PMCDBG1(CPU,BND,2, "save-cpu cpu=%d", pb->pb_cpu);
 749 }
 750
 751 /*
 752  * restore the cpu binding of the current thread
 753  */
 754
 755 static void
 756 pmc_restore_cpu_binding(struct pmc_binding *pb)
 757 {
 758         PMCDBG2(CPU,BND,2, "restore-cpu curcpu=%d restore=%d",
 759             curthread->td_oncpu, pb->pb_cpu);
 760         thread_lock(curthread);
 761         if (pb->pb_bound)
 762                 sched_bind(curthread, pb->pb_cpu);
 763         else
 764                 sched_unbind(curthread);
 765         thread_unlock(curthread);
 766         PMCDBG0(CPU,BND,2, "restore-cpu done");
 767 }
 768
 769 /*
 770  * move execution over the specified cpu and bind it there.
 771  */
 772
 773 static void
 774 pmc_select_cpu(int cpu)
 775 {
 776         KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
 777             ("[pmc,%d] bad cpu number %d", __LINE__, cpu));
 778
 779         /* Never move to an inactive CPU. */
 780         KASSERT(pmc_cpu_is_active(cpu), ("[pmc,%d] selecting inactive "
 781             "CPU %d", __LINE__, cpu));
 782
 783         PMCDBG1(CPU,SEL,2, "select-cpu cpu=%d", cpu);
 784         thread_lock(curthread);
 785         sched_bind(curthread, cpu);
 786         thread_unlock(curthread);
 787
 788         KASSERT(curthread->td_oncpu == cpu,
 789             ("[pmc,%d] CPU not bound [cpu=%d, curr=%d]", __LINE__,
 790                 cpu, curthread->td_oncpu));
 791
 792         PMCDBG1(CPU,SEL,2, "select-cpu cpu=%d ok", cpu);
 793 }
 794
 795 /*
 796  * Force a context switch.
 797  *
 798  * We do this by pause'ing for 1 tick -- invoking mi_switch() is not
 799  * guaranteed to force a context switch.
 800  */
 801
 802 static void
 803 pmc_force_context_switch(void)
 804 {
 805
 806         pause("pmcctx", 1);
 807 }
 808
 809 uint64_t
 810 pmc_rdtsc(void)
 811 {
 812 #if defined(__i386__) || defined(__amd64__)
 813         if (__predict_true(amd_feature & AMDID_RDTSCP))
 814                 return rdtscp();
 815         else
 816                 return rdtsc();
 817 #else
 818         return get_cyclecount();
 819 #endif
 820 }
 821
 822 /*
 823  * Get the file name for an executable.  This is a simple wrapper
 824  * around vn_fullpath(9).
 825  */
 826
 827 static void
 828 pmc_getfilename(struct vnode *v, char **fullpath, char **freepath)
 829 {
 830
 831         *fullpath = "unknown";
 832         *freepath = NULL;
 833         vn_fullpath(curthread, v, fullpath, freepath);
 834 }
 835
 836 /*
 837  * remove an process owning PMCs
 838  */
 839
 840 void
 841 pmc_remove_owner(struct pmc_owner *po)
 842 {
 843         struct pmc *pm, *tmp;
 844
 845         sx_assert(&pmc_sx, SX_XLOCKED);
 846
 847         PMCDBG1(OWN,ORM,1, "remove-owner po=%p", po);
 848
 849         /* Remove descriptor from the owner hash table */
 850         LIST_REMOVE(po, po_next);
 851
 852         /* release all owned PMC descriptors */
 853         LIST_FOREACH_SAFE(pm, &po->po_pmcs, pm_next, tmp) {
 854                 PMCDBG1(OWN,ORM,2, "pmc=%p", pm);
 855                 KASSERT(pm->pm_owner == po,
 856                     ("[pmc,%d] owner %p != po %p", __LINE__, pm->pm_owner, po));
 857
 858                 pmc_release_pmc_descriptor(pm); /* will unlink from the list */
 859                 pmc_destroy_pmc_descriptor(pm);
 860         }
 861
 862         KASSERT(po->po_sscount == 0,
 863             ("[pmc,%d] SS count not zero", __LINE__));
 864         KASSERT(LIST_EMPTY(&po->po_pmcs),
 865             ("[pmc,%d] PMC list not empty", __LINE__));
 866
 867         /* de-configure the log file if present */
 868         if (po->po_flags & PMC_PO_OWNS_LOGFILE)
 869                 pmclog_deconfigure_log(po);
 870 }
 871
 872 /*
 873  * remove an owner process record if all conditions are met.
 874  */
 875
 876 static void
 877 pmc_maybe_remove_owner(struct pmc_owner *po)
 878 {
 879
 880         PMCDBG1(OWN,OMR,1, "maybe-remove-owner po=%p", po);
 881
 882         /*
 883          * Remove owner record if
 884          * - this process does not own any PMCs
 885          * - this process has not allocated a system-wide sampling buffer
 886          */
 887
 888         if (LIST_EMPTY(&po->po_pmcs) &&
 889             ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)) {
 890                 pmc_remove_owner(po);
 891                 pmc_destroy_owner_descriptor(po);
 892         }
 893 }
 894
 895 /*
 896  * Add an association between a target process and a PMC.
 897  */
 898
 899 static void
 900 pmc_link_target_process(struct pmc *pm, struct pmc_process *pp)
 901 {
 902         int ri;
 903         struct pmc_target *pt;
 904 #ifdef INVARIANTS
 905         struct pmc_thread *pt_td;
 906 #endif
 907
 908         sx_assert(&pmc_sx, SX_XLOCKED);
 909
 910         KASSERT(pm != NULL && pp != NULL,
 911             ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp));
 912         KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)),
 913             ("[pmc,%d] Attaching a non-process-virtual pmc=%p to pid=%d",
 914                 __LINE__, pm, pp->pp_proc->p_pid));
 915         KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= ((int) md->pmd_npmc - 1),
 916             ("[pmc,%d] Illegal reference count %d for process record %p",
 917                 __LINE__, pp->pp_refcnt, (void *) pp));
 918
 919         ri = PMC_TO_ROWINDEX(pm);
 920
 921         PMCDBG3(PRC,TLK,1, "link-target pmc=%p ri=%d pmc-process=%p",
 922             pm, ri, pp);
 923
 924 #ifdef  HWPMC_DEBUG
 925         LIST_FOREACH(pt, &pm->pm_targets, pt_next)
 926             if (pt->pt_process == pp)
 927                     KASSERT(0, ("[pmc,%d] pp %p already in pmc %p targets",
 928                                 __LINE__, pp, pm));
 929 #endif
 930
 931         pt = malloc(sizeof(struct pmc_target), M_PMC, M_WAITOK|M_ZERO);
 932         pt->pt_process = pp;
 933
 934         LIST_INSERT_HEAD(&pm->pm_targets, pt, pt_next);
 935
 936         atomic_store_rel_ptr((uintptr_t *)&pp->pp_pmcs[ri].pp_pmc,
 937             (uintptr_t)pm);
 938
 939         if (pm->pm_owner->po_owner == pp->pp_proc)
 940                 pm->pm_flags |= PMC_F_ATTACHED_TO_OWNER;
 941
 942         /*
 943          * Initialize the per-process values at this row index.
 944          */
 945         pp->pp_pmcs[ri].pp_pmcval = PMC_TO_MODE(pm) == PMC_MODE_TS ?
 946             pm->pm_sc.pm_reloadcount : 0;
 947
 948         pp->pp_refcnt++;
 949
 950 #ifdef INVARIANTS
 951         /* Confirm that the per-thread values at this row index are cleared. */
 952         if (PMC_TO_MODE(pm) == PMC_MODE_TS) {
 953                 mtx_lock_spin(pp->pp_tdslock);
 954                 LIST_FOREACH(pt_td, &pp->pp_tds, pt_next) {
 955                         KASSERT(pt_td->pt_pmcs[ri].pt_pmcval == (pmc_value_t) 0,
 956                             ("[pmc,%d] pt_pmcval not cleared for pid=%d at "
 957                             "ri=%d", __LINE__, pp->pp_proc->p_pid, ri));
 958                 }
 959                 mtx_unlock_spin(pp->pp_tdslock);
 960         }
 961 #endif
 962 }
 963
 964 /*
 965  * Removes the association between a target process and a PMC.
 966  */
 967
 968 static void
 969 pmc_unlink_target_process(struct pmc *pm, struct pmc_process *pp)
 970 {
 971         int ri;
 972         struct proc *p;
 973         struct pmc_target *ptgt;
 974         struct pmc_thread *pt;
 975
 976         sx_assert(&pmc_sx, SX_XLOCKED);
 977
 978         KASSERT(pm != NULL && pp != NULL,
 979             ("[pmc,%d] Null pm %p or pp %p", __LINE__, pm, pp));
 980
 981         KASSERT(pp->pp_refcnt >= 1 && pp->pp_refcnt <= (int) md->pmd_npmc,
 982             ("[pmc,%d] Illegal ref count %d on process record %p",
 983                 __LINE__, pp->pp_refcnt, (void *) pp));
 984
 985         ri = PMC_TO_ROWINDEX(pm);
 986
 987         PMCDBG3(PRC,TUL,1, "unlink-target pmc=%p ri=%d pmc-process=%p",
 988             pm, ri, pp);
 989
 990         KASSERT(pp->pp_pmcs[ri].pp_pmc == pm,
 991             ("[pmc,%d] PMC ri %d mismatch pmc %p pp->[ri] %p", __LINE__,
 992                 ri, pm, pp->pp_pmcs[ri].pp_pmc));
 993
 994         pp->pp_pmcs[ri].pp_pmc = NULL;
 995         pp->pp_pmcs[ri].pp_pmcval = (pmc_value_t) 0;
 996
 997         /* Clear the per-thread values at this row index. */
 998         if (PMC_TO_MODE(pm) == PMC_MODE_TS) {
 999                 mtx_lock_spin(pp->pp_tdslock);
1000                 LIST_FOREACH(pt, &pp->pp_tds, pt_next)
1001                         pt->pt_pmcs[ri].pt_pmcval = (pmc_value_t) 0;
1002                 mtx_unlock_spin(pp->pp_tdslock);
1003         }
1004
1005         /* Remove owner-specific flags */
1006         if (pm->pm_owner->po_owner == pp->pp_proc) {
1007                 pp->pp_flags &= ~PMC_PP_ENABLE_MSR_ACCESS;
1008                 pm->pm_flags &= ~PMC_F_ATTACHED_TO_OWNER;
1009         }
1010
1011         pp->pp_refcnt--;
1012
1013         /* Remove the target process from the PMC structure */
1014         LIST_FOREACH(ptgt, &pm->pm_targets, pt_next)
1015                 if (ptgt->pt_process == pp)
1016                         break;
1017
1018         KASSERT(ptgt != NULL, ("[pmc,%d] process %p (pp: %p) not found "
1019                     "in pmc %p", __LINE__, pp->pp_proc, pp, pm));
1020
1021         LIST_REMOVE(ptgt, pt_next);
1022         free(ptgt, M_PMC);
1023
1024         /* if the PMC now lacks targets, send the owner a SIGIO */
1025         if (LIST_EMPTY(&pm->pm_targets)) {
1026                 p = pm->pm_owner->po_owner;
1027                 PROC_LOCK(p);
1028                 kern_psignal(p, SIGIO);
1029                 PROC_UNLOCK(p);
1030
1031                 PMCDBG2(PRC,SIG,2, "signalling proc=%p signal=%d", p,
1032                     SIGIO);
1033         }
1034 }
1035
1036 /*
1037  * Check if PMC 'pm' may be attached to target process 't'.
1038  */
1039
1040 static int
1041 pmc_can_attach(struct pmc *pm, struct proc *t)
1042 {
1043         struct proc *o;         /* pmc owner */
1044         struct ucred *oc, *tc;  /* owner, target credentials */
1045         int decline_attach, i;
1046
1047         /*
1048          * A PMC's owner can always attach that PMC to itself.
1049          */
1050
1051         if ((o = pm->pm_owner->po_owner) == t)
1052                 return 0;
1053
1054         PROC_LOCK(o);
1055         oc = o->p_ucred;
1056         crhold(oc);
1057         PROC_UNLOCK(o);
1058
1059         PROC_LOCK(t);
1060         tc = t->p_ucred;
1061         crhold(tc);
1062         PROC_UNLOCK(t);
1063
1064         /*
1065          * The effective uid of the PMC owner should match at least one
1066          * of the {effective,real,saved} uids of the target process.
1067          */
1068
1069         decline_attach = oc->cr_uid != tc->cr_uid &&
1070             oc->cr_uid != tc->cr_svuid &&
1071             oc->cr_uid != tc->cr_ruid;
1072
1073         /*
1074          * Every one of the target's group ids, must be in the owner's
1075          * group list.
1076          */
1077         for (i = 0; !decline_attach && i < tc->cr_ngroups; i++)
1078                 decline_attach = !groupmember(tc->cr_groups[i], oc);
1079
1080         /* check the read and saved gids too */
1081         if (decline_attach == 0)
1082                 decline_attach = !groupmember(tc->cr_rgid, oc) ||
1083                     !groupmember(tc->cr_svgid, oc);
1084
1085         crfree(tc);
1086         crfree(oc);
1087
1088         return !decline_attach;
1089 }
1090
1091 /*
1092  * Attach a process to a PMC.
1093  */
1094
1095 static int
1096 pmc_attach_one_process(struct proc *p, struct pmc *pm)
1097 {
1098         int ri, error;
1099         char *fullpath, *freepath;
1100         struct pmc_process      *pp;
1101
1102         sx_assert(&pmc_sx, SX_XLOCKED);
1103
1104         PMCDBG5(PRC,ATT,2, "attach-one pm=%p ri=%d proc=%p (%d, %s)", pm,
1105             PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm);
1106
1107         /*
1108          * Locate the process descriptor corresponding to process 'p',
1109          * allocating space as needed.
1110          *
1111          * Verify that rowindex 'pm_rowindex' is free in the process
1112          * descriptor.
1113          *
1114          * If not, allocate space for a descriptor and link the
1115          * process descriptor and PMC.
1116          */
1117         ri = PMC_TO_ROWINDEX(pm);
1118
1119         /* mark process as using HWPMCs */
1120         PROC_LOCK(p);
1121         p->p_flag |= P_HWPMC;
1122         PROC_UNLOCK(p);
1123
1124         if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_ALLOCATE)) == NULL) {
1125                 error = ENOMEM;
1126                 goto fail;
1127         }
1128
1129         if (pp->pp_pmcs[ri].pp_pmc == pm) {/* already present at slot [ri] */
1130                 error = EEXIST;
1131                 goto fail;
1132         }
1133
1134         if (pp->pp_pmcs[ri].pp_pmc != NULL) {
1135                 error = EBUSY;
1136                 goto fail;
1137         }
1138
1139         pmc_link_target_process(pm, pp);
1140
1141         if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) &&
1142             (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) == 0)
1143                 pm->pm_flags |= PMC_F_NEEDS_LOGFILE;
1144
1145         pm->pm_flags |= PMC_F_ATTACH_DONE; /* mark as attached */
1146
1147         /* issue an attach event to a configured log file */
1148         if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE) {
1149                 if (p->p_flag & P_KPROC) {
1150                         fullpath = kernelname;
1151                         freepath = NULL;
1152                 } else {
1153                         pmc_getfilename(p->p_textvp, &fullpath, &freepath);
1154                         pmclog_process_pmcattach(pm, p->p_pid, fullpath);
1155                 }
1156                 free(freepath, M_TEMP);
1157                 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1158                         pmc_log_process_mappings(pm->pm_owner, p);
1159         }
1160
1161         return (0);
1162  fail:
1163         PROC_LOCK(p);
1164         p->p_flag &= ~P_HWPMC;
1165         PROC_UNLOCK(p);
1166         return (error);
1167 }
1168
1169 /*
1170  * Attach a process and optionally its children
1171  */
1172
1173 static int
1174 pmc_attach_process(struct proc *p, struct pmc *pm)
1175 {
1176         int error;
1177         struct proc *top;
1178
1179         sx_assert(&pmc_sx, SX_XLOCKED);
1180
1181         PMCDBG5(PRC,ATT,1, "attach pm=%p ri=%d proc=%p (%d, %s)", pm,
1182             PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm);
1183
1184
1185         /*
1186          * If this PMC successfully allowed a GETMSR operation
1187          * in the past, disallow further ATTACHes.
1188          */
1189
1190         if ((pm->pm_flags & PMC_PP_ENABLE_MSR_ACCESS) != 0)
1191                 return EPERM;
1192
1193         if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0)
1194                 return pmc_attach_one_process(p, pm);
1195
1196         /*
1197          * Traverse all child processes, attaching them to
1198          * this PMC.
1199          */
1200
1201         sx_slock(&proctree_lock);
1202
1203         top = p;
1204
1205         for (;;) {
1206                 if ((error = pmc_attach_one_process(p, pm)) != 0)
1207                         break;
1208                 if (!LIST_EMPTY(&p->p_children))
1209                         p = LIST_FIRST(&p->p_children);
1210                 else for (;;) {
1211                         if (p == top)
1212                                 goto done;
1213                         if (LIST_NEXT(p, p_sibling)) {
1214                                 p = LIST_NEXT(p, p_sibling);
1215                                 break;
1216                         }
1217                         p = p->p_pptr;
1218                 }
1219         }
1220
1221         if (error)
1222                 (void) pmc_detach_process(top, pm);
1223
1224  done:
1225         sx_sunlock(&proctree_lock);
1226         return error;
1227 }
1228
1229 /*
1230  * Detach a process from a PMC.  If there are no other PMCs tracking
1231  * this process, remove the process structure from its hash table.  If
1232  * 'flags' contains PMC_FLAG_REMOVE, then free the process structure.
1233  */
1234
1235 static int
1236 pmc_detach_one_process(struct proc *p, struct pmc *pm, int flags)
1237 {
1238         int ri;
1239         struct pmc_process *pp;
1240
1241         sx_assert(&pmc_sx, SX_XLOCKED);
1242
1243         KASSERT(pm != NULL,
1244             ("[pmc,%d] null pm pointer", __LINE__));
1245
1246         ri = PMC_TO_ROWINDEX(pm);
1247
1248         PMCDBG6(PRC,ATT,2, "detach-one pm=%p ri=%d proc=%p (%d, %s) flags=0x%x",
1249             pm, ri, p, p->p_pid, p->p_comm, flags);
1250
1251         if ((pp = pmc_find_process_descriptor(p, 0)) == NULL)
1252                 return ESRCH;
1253
1254         if (pp->pp_pmcs[ri].pp_pmc != pm)
1255                 return EINVAL;
1256
1257         pmc_unlink_target_process(pm, pp);
1258
1259         /* Issue a detach entry if a log file is configured */
1260         if (pm->pm_owner->po_flags & PMC_PO_OWNS_LOGFILE)
1261                 pmclog_process_pmcdetach(pm, p->p_pid);
1262
1263         /*
1264          * If there are no PMCs targeting this process, we remove its
1265          * descriptor from the target hash table and unset the P_HWPMC
1266          * flag in the struct proc.
1267          */
1268         KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= (int) md->pmd_npmc,
1269             ("[pmc,%d] Illegal refcnt %d for process struct %p",
1270                 __LINE__, pp->pp_refcnt, pp));
1271
1272         if (pp->pp_refcnt != 0) /* still a target of some PMC */
1273                 return 0;
1274
1275         pmc_remove_process_descriptor(pp);
1276
1277         if (flags & PMC_FLAG_REMOVE)
1278                 pmc_destroy_process_descriptor(pp);
1279
1280         PROC_LOCK(p);
1281         p->p_flag &= ~P_HWPMC;
1282         PROC_UNLOCK(p);
1283
1284         return 0;
1285 }
1286
1287 /*
1288  * Detach a process and optionally its descendants from a PMC.
1289  */
1290
1291 static int
1292 pmc_detach_process(struct proc *p, struct pmc *pm)
1293 {
1294         struct proc *top;
1295
1296         sx_assert(&pmc_sx, SX_XLOCKED);
1297
1298         PMCDBG5(PRC,ATT,1, "detach pm=%p ri=%d proc=%p (%d, %s)", pm,
1299             PMC_TO_ROWINDEX(pm), p, p->p_pid, p->p_comm);
1300
1301         if ((pm->pm_flags & PMC_F_DESCENDANTS) == 0)
1302                 return pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE);
1303
1304         /*
1305          * Traverse all children, detaching them from this PMC.  We
1306          * ignore errors since we could be detaching a PMC from a
1307          * partially attached proc tree.
1308          */
1309
1310         sx_slock(&proctree_lock);
1311
1312         top = p;
1313
1314         for (;;) {
1315                 (void) pmc_detach_one_process(p, pm, PMC_FLAG_REMOVE);
1316
1317                 if (!LIST_EMPTY(&p->p_children))
1318                         p = LIST_FIRST(&p->p_children);
1319                 else for (;;) {
1320                         if (p == top)
1321                                 goto done;
1322                         if (LIST_NEXT(p, p_sibling)) {
1323                                 p = LIST_NEXT(p, p_sibling);
1324                                 break;
1325                         }
1326                         p = p->p_pptr;
1327                 }
1328         }
1329
1330  done:
1331         sx_sunlock(&proctree_lock);
1332
1333         if (LIST_EMPTY(&pm->pm_targets))
1334                 pm->pm_flags &= ~PMC_F_ATTACH_DONE;
1335
1336         return 0;
1337 }
1338
1339
1340 /*
1341  * Thread context switch IN
1342  */
1343
1344 static void
1345 pmc_process_csw_in(struct thread *td)
1346 {
1347         int cpu;
1348         unsigned int adjri, ri;
1349         struct pmc *pm;
1350         struct proc *p;
1351         struct pmc_cpu *pc;
1352         struct pmc_hw *phw;
1353         pmc_value_t newvalue;
1354         struct pmc_process *pp;
1355         struct pmc_thread *pt;
1356         struct pmc_classdep *pcd;
1357
1358         p = td->td_proc;
1359         pt = NULL;
1360         if ((pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE)) == NULL)
1361                 return;
1362
1363         KASSERT(pp->pp_proc == td->td_proc,
1364             ("[pmc,%d] not my thread state", __LINE__));
1365
1366         critical_enter(); /* no preemption from this point */
1367
1368         cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */
1369
1370         PMCDBG5(CSW,SWI,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p,
1371             p->p_pid, p->p_comm, pp);
1372
1373         KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
1374             ("[pmc,%d] weird CPU id %d", __LINE__, cpu));
1375
1376         pc = pmc_pcpu[cpu];
1377
1378         for (ri = 0; ri < md->pmd_npmc; ri++) {
1379
1380                 if ((pm = pp->pp_pmcs[ri].pp_pmc) == NULL)
1381                         continue;
1382
1383                 KASSERT(PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)),
1384                     ("[pmc,%d] Target PMC in non-virtual mode (%d)",
1385                         __LINE__, PMC_TO_MODE(pm)));
1386
1387                 KASSERT(PMC_TO_ROWINDEX(pm) == ri,
1388                     ("[pmc,%d] Row index mismatch pmc %d != ri %d",
1389                         __LINE__, PMC_TO_ROWINDEX(pm), ri));
1390
1391                 /*
1392                  * Only PMCs that are marked as 'RUNNING' need
1393                  * be placed on hardware.
1394                  */
1395
1396                 if (pm->pm_state != PMC_STATE_RUNNING)
1397                         continue;
1398
1399                 KASSERT(counter_u64_fetch(pm->pm_runcount) >= 0,
1400             ("[pmc,%d] pm=%p runcount %ld", __LINE__, (void *) pm,
1401                  (unsigned long)counter_u64_fetch(pm->pm_runcount)));
1402
1403                 /* increment PMC runcount */
1404                 counter_u64_add(pm->pm_runcount, 1);
1405
1406                 /* configure the HWPMC we are going to use. */
1407                 pcd = pmc_ri_to_classdep(md, ri, &adjri);
1408                 pcd->pcd_config_pmc(cpu, adjri, pm);
1409
1410                 phw = pc->pc_hwpmcs[ri];
1411
1412                 KASSERT(phw != NULL,
1413                     ("[pmc,%d] null hw pointer", __LINE__));
1414
1415                 KASSERT(phw->phw_pmc == pm,
1416                     ("[pmc,%d] hw->pmc %p != pmc %p", __LINE__,
1417                         phw->phw_pmc, pm));
1418
1419                 /*
1420                  * Write out saved value and start the PMC.
1421                  *
1422                  * Sampling PMCs use a per-thread value, while
1423                  * counting mode PMCs use a per-pmc value that is
1424                  * inherited across descendants.
1425                  */
1426                 if (PMC_TO_MODE(pm) == PMC_MODE_TS) {
1427                         if (pt == NULL)
1428                                 pt = pmc_find_thread_descriptor(pp, td,
1429                                     PMC_FLAG_NONE);
1430
1431                         KASSERT(pt != NULL,
1432                             ("[pmc,%d] No thread found for td=%p", __LINE__,
1433                             td));
1434
1435                         mtx_pool_lock_spin(pmc_mtxpool, pm);
1436
1437                         /*
1438                          * If we have a thread descriptor, use the per-thread
1439                          * counter in the descriptor. If not, we will use
1440                          * a per-process counter.
1441                          *
1442                          * TODO: Remove the per-process "safety net" once
1443                          * we have thoroughly tested that we don't hit the
1444                          * above assert.
1445                          */
1446                         if (pt != NULL) {
1447                                 if (pt->pt_pmcs[ri].pt_pmcval > 0)
1448                                         newvalue = pt->pt_pmcs[ri].pt_pmcval;
1449                                 else
1450                                         newvalue = pm->pm_sc.pm_reloadcount;
1451                         } else {
1452                                 /*
1453                                  * Use the saved value calculated after the most
1454                                  * recent time a thread using the shared counter
1455                                  * switched out. Reset the saved count in case
1456                                  * another thread from this process switches in
1457                                  * before any threads switch out.
1458                                  */
1459
1460                                 newvalue = pp->pp_pmcs[ri].pp_pmcval;
1461                                 pp->pp_pmcs[ri].pp_pmcval =
1462                                     pm->pm_sc.pm_reloadcount;
1463                         }
1464                         mtx_pool_unlock_spin(pmc_mtxpool, pm);
1465                         KASSERT(newvalue > 0 && newvalue <=
1466                             pm->pm_sc.pm_reloadcount,
1467                             ("[pmc,%d] pmcval outside of expected range cpu=%d "
1468                             "ri=%d pmcval=%jx pm_reloadcount=%jx", __LINE__,
1469                             cpu, ri, newvalue, pm->pm_sc.pm_reloadcount));
1470                 } else {
1471                         KASSERT(PMC_TO_MODE(pm) == PMC_MODE_TC,
1472                             ("[pmc,%d] illegal mode=%d", __LINE__,
1473                             PMC_TO_MODE(pm)));
1474                         mtx_pool_lock_spin(pmc_mtxpool, pm);
1475                         newvalue = PMC_PCPU_SAVED(cpu, ri) =
1476                             pm->pm_gv.pm_savedvalue;
1477                         mtx_pool_unlock_spin(pmc_mtxpool, pm);
1478                 }
1479
1480                 PMCDBG3(CSW,SWI,1,"cpu=%d ri=%d new=%jd", cpu, ri, newvalue);
1481
1482                 pcd->pcd_write_pmc(cpu, adjri, newvalue);
1483
1484                 /* If a sampling mode PMC, reset stalled state. */
1485                 if (PMC_TO_MODE(pm) == PMC_MODE_TS)
1486                         pm->pm_pcpu_state[cpu].pps_stalled = 0;
1487
1488                 /* Indicate that we desire this to run. */
1489                 pm->pm_pcpu_state[cpu].pps_cpustate = 1;
1490
1491                 /* Start the PMC. */
1492                 pcd->pcd_start_pmc(cpu, adjri);
1493         }
1494
1495         /*
1496          * perform any other architecture/cpu dependent thread
1497          * switch-in actions.
1498          */
1499
1500         (void) (*md->pmd_switch_in)(pc, pp);
1501
1502         critical_exit();
1503
1504 }
1505
1506 /*
1507  * Thread context switch OUT.
1508  */
1509
1510 static void
1511 pmc_process_csw_out(struct thread *td)
1512 {
1513         int cpu;
1514         int64_t tmp;
1515         struct pmc *pm;
1516         struct proc *p;
1517         enum pmc_mode mode;
1518         struct pmc_cpu *pc;
1519         pmc_value_t newvalue;
1520         unsigned int adjri, ri;
1521         struct pmc_process *pp;
1522         struct pmc_thread *pt = NULL;
1523         struct pmc_classdep *pcd;
1524
1525
1526         /*
1527          * Locate our process descriptor; this may be NULL if
1528          * this process is exiting and we have already removed
1529          * the process from the target process table.
1530          *
1531          * Note that due to kernel preemption, multiple
1532          * context switches may happen while the process is
1533          * exiting.
1534          *
1535          * Note also that if the target process cannot be
1536          * found we still need to deconfigure any PMCs that
1537          * are currently running on hardware.
1538          */
1539
1540         p = td->td_proc;
1541         pp = pmc_find_process_descriptor(p, PMC_FLAG_NONE);
1542
1543         /*
1544          * save PMCs
1545          */
1546
1547         critical_enter();
1548
1549         cpu = PCPU_GET(cpuid); /* td->td_oncpu is invalid */
1550
1551         PMCDBG5(CSW,SWO,1, "cpu=%d proc=%p (%d, %s) pp=%p", cpu, p,
1552             p->p_pid, p->p_comm, pp);
1553
1554         KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
1555             ("[pmc,%d weird CPU id %d", __LINE__, cpu));
1556
1557         pc = pmc_pcpu[cpu];
1558
1559         /*
1560          * When a PMC gets unlinked from a target PMC, it will
1561          * be removed from the target's pp_pmc[] array.
1562          *
1563          * However, on a MP system, the target could have been
1564          * executing on another CPU at the time of the unlink.
1565          * So, at context switch OUT time, we need to look at
1566          * the hardware to determine if a PMC is scheduled on
1567          * it.
1568          */
1569
1570         for (ri = 0; ri < md->pmd_npmc; ri++) {
1571
1572                 pcd = pmc_ri_to_classdep(md, ri, &adjri);
1573                 pm  = NULL;
1574                 (void) (*pcd->pcd_get_config)(cpu, adjri, &pm);
1575
1576                 if (pm == NULL) /* nothing at this row index */
1577                         continue;
1578
1579                 mode = PMC_TO_MODE(pm);
1580                 if (!PMC_IS_VIRTUAL_MODE(mode))
1581                         continue; /* not a process virtual PMC */
1582
1583                 KASSERT(PMC_TO_ROWINDEX(pm) == ri,
1584                     ("[pmc,%d] ri mismatch pmc(%d) ri(%d)",
1585                         __LINE__, PMC_TO_ROWINDEX(pm), ri));
1586
1587                 /*
1588                  * Change desired state, and then stop if not stalled.
1589                  * This two-step dance should avoid race conditions where
1590                  * an interrupt re-enables the PMC after this code has
1591                  * already checked the pm_stalled flag.
1592                  */
1593                 pm->pm_pcpu_state[cpu].pps_cpustate = 0;
1594                 if (pm->pm_pcpu_state[cpu].pps_stalled == 0)
1595                         pcd->pcd_stop_pmc(cpu, adjri);
1596
1597                 KASSERT(counter_u64_fetch(pm->pm_runcount) > 0,
1598                         ("[pmc,%d] pm=%p runcount %ld", __LINE__, (void *) pm,
1599                          (unsigned long)counter_u64_fetch(pm->pm_runcount)));
1600
1601                 /* reduce this PMC's runcount */
1602                 counter_u64_add(pm->pm_runcount, -1);
1603
1604                 /*
1605                  * If this PMC is associated with this process,
1606                  * save the reading.
1607                  */
1608
1609                 if (pm->pm_state != PMC_STATE_DELETED && pp != NULL &&
1610                     pp->pp_pmcs[ri].pp_pmc != NULL) {
1611                         KASSERT(pm == pp->pp_pmcs[ri].pp_pmc,
1612                             ("[pmc,%d] pm %p != pp_pmcs[%d] %p", __LINE__,
1613                                 pm, ri, pp->pp_pmcs[ri].pp_pmc));
1614
1615                         KASSERT(pp->pp_refcnt > 0,
1616                             ("[pmc,%d] pp refcnt = %d", __LINE__,
1617                                 pp->pp_refcnt));
1618
1619                         pcd->pcd_read_pmc(cpu, adjri, &newvalue);
1620
1621                         if (mode == PMC_MODE_TS) {
1622                                 PMCDBG3(CSW,SWO,1,"cpu=%d ri=%d val=%jd (samp)",
1623                                     cpu, ri, newvalue);
1624
1625                                 if (pt == NULL)
1626                                         pt = pmc_find_thread_descriptor(pp, td,
1627                                             PMC_FLAG_NONE);
1628
1629                                 KASSERT(pt != NULL,
1630                                     ("[pmc,%d] No thread found for td=%p",
1631                                     __LINE__, td));
1632
1633                                 mtx_pool_lock_spin(pmc_mtxpool, pm);
1634
1635                                 /*
1636                                  * If we have a thread descriptor, save the
1637                                  * per-thread counter in the descriptor. If not,
1638                                  * we will update the per-process counter.
1639                                  *
1640                                  * TODO: Remove the per-process "safety net"
1641                                  * once we have thoroughly tested that we
1642                                  * don't hit the above assert.
1643                                  */
1644                                 if (pt != NULL)
1645                                         pt->pt_pmcs[ri].pt_pmcval = newvalue;
1646                                 else {
1647                                         /*
1648                                          * For sampling process-virtual PMCs,
1649                                          * newvalue is the number of events to
1650                                          * be seen until the next sampling
1651                                          * interrupt. We can just add the events
1652                                          * left from this invocation to the
1653                                          * counter, then adjust in case we
1654                                          * overflow our range.
1655                                          *
1656                                          * (Recall that we reload the counter
1657                                          * every time we use it.)
1658                                          */
1659                                         pp->pp_pmcs[ri].pp_pmcval += newvalue;
1660                                         if (pp->pp_pmcs[ri].pp_pmcval >
1661                                             pm->pm_sc.pm_reloadcount)
1662                                                 pp->pp_pmcs[ri].pp_pmcval -=
1663                                                     pm->pm_sc.pm_reloadcount;
1664                                 }
1665                                 mtx_pool_unlock_spin(pmc_mtxpool, pm);
1666                         } else {
1667                                 tmp = newvalue - PMC_PCPU_SAVED(cpu,ri);
1668
1669                                 PMCDBG3(CSW,SWO,1,"cpu=%d ri=%d tmp=%jd (count)",
1670                                     cpu, ri, tmp);
1671
1672                                 /*
1673                                  * For counting process-virtual PMCs,
1674                                  * we expect the count to be
1675                                  * increasing monotonically, modulo a 64
1676                                  * bit wraparound.
1677                                  */
1678                                 KASSERT(tmp >= 0,
1679                                     ("[pmc,%d] negative increment cpu=%d "
1680                                      "ri=%d newvalue=%jx saved=%jx "
1681                                      "incr=%jx", __LINE__, cpu, ri,
1682                                      newvalue, PMC_PCPU_SAVED(cpu,ri), tmp));
1683
1684                                 mtx_pool_lock_spin(pmc_mtxpool, pm);
1685                                 pm->pm_gv.pm_savedvalue += tmp;
1686                                 pp->pp_pmcs[ri].pp_pmcval += tmp;
1687                                 mtx_pool_unlock_spin(pmc_mtxpool, pm);
1688
1689                                 if (pm->pm_flags & PMC_F_LOG_PROCCSW)
1690                                         pmclog_process_proccsw(pm, pp, tmp, td);
1691                         }
1692                 }
1693
1694                 /* mark hardware as free */
1695                 pcd->pcd_config_pmc(cpu, adjri, NULL);
1696         }
1697
1698         /*
1699          * perform any other architecture/cpu dependent thread
1700          * switch out functions.
1701          */
1702
1703         (void) (*md->pmd_switch_out)(pc, pp);
1704
1705         critical_exit();
1706 }
1707
1708 /*
1709  * A new thread for a process.
1710  */
1711 static void
1712 pmc_process_thread_add(struct thread *td)
1713 {
1714         struct pmc_process *pmc;
1715
1716         pmc = pmc_find_process_descriptor(td->td_proc, PMC_FLAG_NONE);
1717         if (pmc != NULL)
1718                 pmc_find_thread_descriptor(pmc, td, PMC_FLAG_ALLOCATE);
1719 }
1720
1721 /*
1722  * A thread delete for a process.
1723  */
1724 static void
1725 pmc_process_thread_delete(struct thread *td)
1726 {
1727         struct pmc_process *pmc;
1728
1729         pmc = pmc_find_process_descriptor(td->td_proc, PMC_FLAG_NONE);
1730         if (pmc != NULL)
1731                 pmc_thread_descriptor_pool_free(pmc_find_thread_descriptor(pmc,
1732                     td, PMC_FLAG_REMOVE));
1733 }
1734
1735 /*
1736  * A userret() call for a thread.
1737  */
1738 static void
1739 pmc_process_thread_userret(struct thread *td)
1740 {
1741         sched_pin();
1742         pmc_capture_user_callchain(curcpu, PMC_UR, td->td_frame);
1743         sched_unpin();
1744 }
1745
1746 /*
1747  * A mapping change for a process.
1748  */
1749
1750 static void
1751 pmc_process_mmap(struct thread *td, struct pmckern_map_in *pkm)
1752 {
1753         int ri;
1754         pid_t pid;
1755         char *fullpath, *freepath;
1756         const struct pmc *pm;
1757         struct pmc_owner *po;
1758         const struct pmc_process *pp;
1759
1760         freepath = fullpath = NULL;
1761         MPASS(!in_epoch(global_epoch_preempt));
1762         pmc_getfilename((struct vnode *) pkm->pm_file, &fullpath, &freepath);
1763
1764         pid = td->td_proc->p_pid;
1765
1766         PMC_EPOCH_ENTER();
1767         /* Inform owners of all system-wide sampling PMCs. */
1768         CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1769             if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1770                         pmclog_process_map_in(po, pid, pkm->pm_address, fullpath);
1771
1772         if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL)
1773                 goto done;
1774
1775         /*
1776          * Inform sampling PMC owners tracking this process.
1777          */
1778         for (ri = 0; ri < md->pmd_npmc; ri++)
1779                 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL &&
1780                     PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1781                         pmclog_process_map_in(pm->pm_owner,
1782                             pid, pkm->pm_address, fullpath);
1783
1784   done:
1785         if (freepath)
1786                 free(freepath, M_TEMP);
1787         PMC_EPOCH_EXIT();
1788 }
1789
1790
1791 /*
1792  * Log an munmap request.
1793  */
1794
1795 static void
1796 pmc_process_munmap(struct thread *td, struct pmckern_map_out *pkm)
1797 {
1798         int ri;
1799         pid_t pid;
1800         struct pmc_owner *po;
1801         const struct pmc *pm;
1802         const struct pmc_process *pp;
1803
1804         pid = td->td_proc->p_pid;
1805
1806         PMC_EPOCH_ENTER();
1807         CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
1808             if (po->po_flags & PMC_PO_OWNS_LOGFILE)
1809                 pmclog_process_map_out(po, pid, pkm->pm_address,
1810                     pkm->pm_address + pkm->pm_size);
1811         PMC_EPOCH_EXIT();
1812
1813         if ((pp = pmc_find_process_descriptor(td->td_proc, 0)) == NULL)
1814                 return;
1815
1816         for (ri = 0; ri < md->pmd_npmc; ri++)
1817                 if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL &&
1818                     PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
1819                         pmclog_process_map_out(pm->pm_owner, pid,
1820                             pkm->pm_address, pkm->pm_address + pkm->pm_size);
1821 }
1822
1823 /*
1824  * Log mapping information about the kernel.
1825  */
1826
1827 static void
1828 pmc_log_kernel_mappings(struct pmc *pm)
1829 {
1830         struct pmc_owner *po;
1831         struct pmckern_map_in *km, *kmbase;
1832
1833         MPASS(in_epoch(global_epoch_preempt) || sx_xlocked(&pmc_sx));
1834         KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)),
1835             ("[pmc,%d] non-sampling PMC (%p) desires mapping information",
1836                 __LINE__, (void *) pm));
1837
1838         po = pm->pm_owner;
1839
1840         if (po->po_flags & PMC_PO_INITIAL_MAPPINGS_DONE)
1841                 return;
1842         if (PMC_TO_MODE(pm) == PMC_MODE_SS)
1843                 pmc_process_allproc(pm);
1844         /*
1845          * Log the current set of kernel modules.
1846          */
1847         kmbase = linker_hwpmc_list_objects();
1848         for (km = kmbase; km->pm_file != NULL; km++) {
1849                 PMCDBG2(LOG,REG,1,"%s %p", (char *) km->pm_file,
1850                     (void *) km->pm_address);
1851                 pmclog_process_map_in(po, (pid_t) -1, km->pm_address,
1852                     km->pm_file);
1853         }
1854         free(kmbase, M_LINKER);
1855
1856         po->po_flags |= PMC_PO_INITIAL_MAPPINGS_DONE;
1857 }
1858
1859 /*
1860  * Log the mappings for a single process.
1861  */
1862
1863 static void
1864 pmc_log_process_mappings(struct pmc_owner *po, struct proc *p)
1865 {
1866         vm_map_t map;
1867         struct vnode *vp;
1868         struct vmspace *vm;
1869         vm_map_entry_t entry;
1870         vm_offset_t last_end;
1871         u_int last_timestamp;
1872         struct vnode *last_vp;
1873         vm_offset_t start_addr;
1874         vm_object_t obj, lobj, tobj;
1875         char *fullpath, *freepath;
1876
1877         last_vp = NULL;
1878         last_end = (vm_offset_t) 0;
1879         fullpath = freepath = NULL;
1880
1881         if ((vm = vmspace_acquire_ref(p)) == NULL)
1882                 return;
1883
1884         map = &vm->vm_map;
1885         vm_map_lock_read(map);
1886
1887         for (entry = map->header.next; entry != &map->header; entry = entry->next) {
1888
1889                 if (entry == NULL) {
1890                         PMCDBG2(LOG,OPS,2, "hwpmc: vm_map entry unexpectedly "
1891                             "NULL! pid=%d vm_map=%p\n", p->p_pid, map);
1892                         break;
1893                 }
1894
1895                 /*
1896                  * We only care about executable map entries.
1897                  */
1898                 if ((entry->eflags & MAP_ENTRY_IS_SUB_MAP) ||
1899                     !(entry->protection & VM_PROT_EXECUTE) ||
1900                     (entry->object.vm_object == NULL)) {
1901                         continue;
1902                 }
1903
1904                 obj = entry->object.vm_object;
1905                 VM_OBJECT_RLOCK(obj);
1906
1907                 /*
1908                  * Walk the backing_object list to find the base
1909                  * (non-shadowed) vm_object.
1910                  */
1911                 for (lobj = tobj = obj; tobj != NULL; tobj = tobj->backing_object) {
1912                         if (tobj != obj)
1913                                 VM_OBJECT_RLOCK(tobj);
1914                         if (lobj != obj)
1915                                 VM_OBJECT_RUNLOCK(lobj);
1916                         lobj = tobj;
1917                 }
1918
1919                 /*
1920                  * At this point lobj is the base vm_object and it is locked.
1921                  */
1922                 if (lobj == NULL) {
1923                         PMCDBG3(LOG,OPS,2, "hwpmc: lobj unexpectedly NULL! pid=%d "
1924                             "vm_map=%p vm_obj=%p\n", p->p_pid, map, obj);
1925                         VM_OBJECT_RUNLOCK(obj);
1926                         continue;
1927                 }
1928
1929                 vp = vm_object_vnode(lobj);
1930                 if (vp == NULL) {
1931                         if (lobj != obj)
1932                                 VM_OBJECT_RUNLOCK(lobj);
1933                         VM_OBJECT_RUNLOCK(obj);
1934                         continue;
1935                 }
1936
1937                 /*
1938                  * Skip contiguous regions that point to the same
1939                  * vnode, so we don't emit redundant MAP-IN
1940                  * directives.
1941                  */
1942                 if (entry->start == last_end && vp == last_vp) {
1943                         last_end = entry->end;
1944                         if (lobj != obj)
1945                                 VM_OBJECT_RUNLOCK(lobj);
1946                         VM_OBJECT_RUNLOCK(obj);
1947                         continue;
1948                 }
1949
1950                 /*
1951                  * We don't want to keep the proc's vm_map or this
1952                  * vm_object locked while we walk the pathname, since
1953                  * vn_fullpath() can sleep.  However, if we drop the
1954                  * lock, it's possible for concurrent activity to
1955                  * modify the vm_map list.  To protect against this,
1956                  * we save the vm_map timestamp before we release the
1957                  * lock, and check it after we reacquire the lock
1958                  * below.
1959                  */
1960                 start_addr = entry->start;
1961                 last_end = entry->end;
1962                 last_timestamp = map->timestamp;
1963                 vm_map_unlock_read(map);
1964
1965                 vref(vp);
1966                 if (lobj != obj)
1967                         VM_OBJECT_RUNLOCK(lobj);
1968
1969                 VM_OBJECT_RUNLOCK(obj);
1970
1971                 freepath = NULL;
1972                 pmc_getfilename(vp, &fullpath, &freepath);
1973                 last_vp = vp;
1974
1975                 vrele(vp);
1976
1977                 vp = NULL;
1978                 pmclog_process_map_in(po, p->p_pid, start_addr, fullpath);
1979                 if (freepath)
1980                         free(freepath, M_TEMP);
1981
1982                 vm_map_lock_read(map);
1983
1984                 /*
1985                  * If our saved timestamp doesn't match, this means
1986                  * that the vm_map was modified out from under us and
1987                  * we can't trust our current "entry" pointer.  Do a
1988                  * new lookup for this entry.  If there is no entry
1989                  * for this address range, vm_map_lookup_entry() will
1990                  * return the previous one, so we always want to go to
1991                  * entry->next on the next loop iteration.
1992                  *
1993                  * There is an edge condition here that can occur if
1994                  * there is no entry at or before this address.  In
1995                  * this situation, vm_map_lookup_entry returns
1996                  * &map->header, which would cause our loop to abort
1997                  * without processing the rest of the map.  However,
1998                  * in practice this will never happen for process
1999                  * vm_map.  This is because the executable's text
2000                  * segment is the first mapping in the proc's address
2001                  * space, and this mapping is never removed until the
2002                  * process exits, so there will always be a non-header
2003                  * entry at or before the requested address for
2004                  * vm_map_lookup_entry to return.
2005                  */
2006                 if (map->timestamp != last_timestamp)
2007                         vm_map_lookup_entry(map, last_end - 1, &entry);
2008         }
2009
2010         vm_map_unlock_read(map);
2011         vmspace_free(vm);
2012         return;
2013 }
2014
2015 /*
2016  * Log mappings for all processes in the system.
2017  */
2018
2019 static void
2020 pmc_log_all_process_mappings(struct pmc_owner *po)
2021 {
2022         struct proc *p, *top;
2023
2024         sx_assert(&pmc_sx, SX_XLOCKED);
2025
2026         if ((p = pfind(1)) == NULL)
2027                 panic("[pmc,%d] Cannot find init", __LINE__);
2028
2029         PROC_UNLOCK(p);
2030
2031         sx_slock(&proctree_lock);
2032
2033         top = p;
2034
2035         for (;;) {
2036                 pmc_log_process_mappings(po, p);
2037                 if (!LIST_EMPTY(&p->p_children))
2038                         p = LIST_FIRST(&p->p_children);
2039                 else for (;;) {
2040                         if (p == top)
2041                                 goto done;
2042                         if (LIST_NEXT(p, p_sibling)) {
2043                                 p = LIST_NEXT(p, p_sibling);
2044                                 break;
2045                         }
2046                         p = p->p_pptr;
2047                 }
2048         }
2049  done:
2050         sx_sunlock(&proctree_lock);
2051 }
2052
2053 /*
2054  * The 'hook' invoked from the kernel proper
2055  */
2056
2057
2058 #ifdef  HWPMC_DEBUG
2059 const char *pmc_hooknames[] = {
2060         /* these strings correspond to PMC_FN_* in <sys/pmckern.h> */
2061         "",
2062         "EXEC",
2063         "CSW-IN",
2064         "CSW-OUT",
2065         "SAMPLE",
2066         "UNUSED1",
2067         "UNUSED2",
2068         "MMAP",
2069         "MUNMAP",
2070         "CALLCHAIN-NMI",
2071         "CALLCHAIN-SOFT",
2072         "SOFTSAMPLING",
2073         "THR-CREATE",
2074         "THR-EXIT",
2075         "THR-USERRET",
2076         "THR-CREATE-LOG",
2077         "THR-EXIT-LOG",
2078         "PROC-CREATE-LOG"
2079 };
2080 #endif
2081
2082 static int
2083 pmc_hook_handler(struct thread *td, int function, void *arg)
2084 {
2085         int cpu;
2086
2087         PMCDBG4(MOD,PMH,1, "hook td=%p func=%d \"%s\" arg=%p", td, function,
2088             pmc_hooknames[function], arg);
2089
2090         switch (function)
2091         {
2092
2093         /*
2094          * Process exec()
2095          */
2096
2097         case PMC_FN_PROCESS_EXEC:
2098         {
2099                 char *fullpath, *freepath;
2100                 unsigned int ri;
2101                 int is_using_hwpmcs;
2102                 struct pmc *pm;
2103                 struct proc *p;
2104                 struct pmc_owner *po;
2105                 struct pmc_process *pp;
2106                 struct pmckern_procexec *pk;
2107
2108                 sx_assert(&pmc_sx, SX_XLOCKED);
2109
2110                 p = td->td_proc;
2111                 pmc_getfilename(p->p_textvp, &fullpath, &freepath);
2112
2113                 pk = (struct pmckern_procexec *) arg;
2114
2115                 PMC_EPOCH_ENTER();
2116                 /* Inform owners of SS mode PMCs of the exec event. */
2117                 CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
2118                     if (po->po_flags & PMC_PO_OWNS_LOGFILE)
2119                             pmclog_process_procexec(po, PMC_ID_INVALID,
2120                                 p->p_pid, pk->pm_entryaddr, fullpath);
2121                 PMC_EPOCH_EXIT();
2122
2123                 PROC_LOCK(p);
2124                 is_using_hwpmcs = p->p_flag & P_HWPMC;
2125                 PROC_UNLOCK(p);
2126
2127                 if (!is_using_hwpmcs) {
2128                         if (freepath)
2129                                 free(freepath, M_TEMP);
2130                         break;
2131                 }
2132
2133                 /*
2134                  * PMCs are not inherited across an exec():  remove any
2135                  * PMCs that this process is the owner of.
2136                  */
2137
2138                 if ((po = pmc_find_owner_descriptor(p)) != NULL) {
2139                         pmc_remove_owner(po);
2140                         pmc_destroy_owner_descriptor(po);
2141                 }
2142
2143                 /*
2144                  * If the process being exec'ed is not the target of any
2145                  * PMC, we are done.
2146                  */
2147                 if ((pp = pmc_find_process_descriptor(p, 0)) == NULL) {
2148                         if (freepath)
2149                                 free(freepath, M_TEMP);
2150                         break;
2151                 }
2152
2153                 /*
2154                  * Log the exec event to all monitoring owners.  Skip
2155                  * owners who have already received the event because
2156                  * they had system sampling PMCs active.
2157                  */
2158                 for (ri = 0; ri < md->pmd_npmc; ri++)
2159                         if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) {
2160                                 po = pm->pm_owner;
2161                                 if (po->po_sscount == 0 &&
2162                                     po->po_flags & PMC_PO_OWNS_LOGFILE)
2163                                         pmclog_process_procexec(po, pm->pm_id,
2164                                             p->p_pid, pk->pm_entryaddr,
2165                                             fullpath);
2166                         }
2167
2168                 if (freepath)
2169                         free(freepath, M_TEMP);
2170
2171
2172                 PMCDBG4(PRC,EXC,1, "exec proc=%p (%d, %s) cred-changed=%d",
2173                     p, p->p_pid, p->p_comm, pk->pm_credentialschanged);
2174
2175                 if (pk->pm_credentialschanged == 0) /* no change */
2176                         break;
2177
2178                 /*
2179                  * If the newly exec()'ed process has a different credential
2180                  * than before, allow it to be the target of a PMC only if
2181                  * the PMC's owner has sufficient privilege.
2182                  */
2183
2184                 for (ri = 0; ri < md->pmd_npmc; ri++)
2185                         if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL)
2186                                 if (pmc_can_attach(pm, td->td_proc) != 0)
2187                                         pmc_detach_one_process(td->td_proc,
2188                                             pm, PMC_FLAG_NONE);
2189
2190                 KASSERT(pp->pp_refcnt >= 0 && pp->pp_refcnt <= (int) md->pmd_npmc,
2191                     ("[pmc,%d] Illegal ref count %d on pp %p", __LINE__,
2192                         pp->pp_refcnt, pp));
2193
2194                 /*
2195                  * If this process is no longer the target of any
2196                  * PMCs, we can remove the process entry and free
2197                  * up space.
2198                  */
2199
2200                 if (pp->pp_refcnt == 0) {
2201                         pmc_remove_process_descriptor(pp);
2202                         pmc_destroy_process_descriptor(pp);
2203                         break;
2204                 }
2205
2206         }
2207         break;
2208
2209         case PMC_FN_CSW_IN:
2210                 pmc_process_csw_in(td);
2211                 break;
2212
2213         case PMC_FN_CSW_OUT:
2214                 pmc_process_csw_out(td);
2215                 break;
2216
2217         /*
2218          * Process accumulated PC samples.
2219          *
2220          * This function is expected to be called by hardclock() for
2221          * each CPU that has accumulated PC samples.
2222          *
2223          * This function is to be executed on the CPU whose samples
2224          * are being processed.
2225          */
2226         case PMC_FN_DO_SAMPLES:
2227
2228                 /*
2229                  * Clear the cpu specific bit in the CPU mask before
2230                  * do the rest of the processing.  If the NMI handler
2231                  * gets invoked after the "atomic_clear_int()" call
2232                  * below but before "pmc_process_samples()" gets
2233                  * around to processing the interrupt, then we will
2234                  * come back here at the next hardclock() tick (and
2235                  * may find nothing to do if "pmc_process_samples()"
2236                  * had already processed the interrupt).  We don't
2237                  * lose the interrupt sample.
2238                  */
2239                 DPCPU_SET(pmc_sampled, 0);
2240                 cpu = PCPU_GET(cpuid);
2241                 pmc_process_samples(cpu, PMC_HR);
2242                 pmc_process_samples(cpu, PMC_SR);
2243                 pmc_process_samples(cpu, PMC_UR);
2244                 break;
2245
2246         case PMC_FN_MMAP:
2247                 pmc_process_mmap(td, (struct pmckern_map_in *) arg);
2248                 break;
2249
2250         case PMC_FN_MUNMAP:
2251                 MPASS(in_epoch(global_epoch_preempt) || sx_xlocked(&pmc_sx));
2252                 pmc_process_munmap(td, (struct pmckern_map_out *) arg);
2253                 break;
2254
2255         case PMC_FN_PROC_CREATE_LOG:
2256                 pmc_process_proccreate((struct proc *)arg);
2257                 break;
2258
2259         case PMC_FN_USER_CALLCHAIN:
2260                 /*
2261                  * Record a call chain.
2262                  */
2263                 KASSERT(td == curthread, ("[pmc,%d] td != curthread",
2264                     __LINE__));
2265
2266                 pmc_capture_user_callchain(PCPU_GET(cpuid), PMC_HR,
2267                     (struct trapframe *) arg);
2268
2269                 KASSERT(td->td_pinned == 1,
2270                         ("[pmc,%d] invalid td_pinned value", __LINE__));
2271                 sched_unpin();  /* Can migrate safely now. */
2272
2273                 td->td_pflags &= ~TDP_CALLCHAIN;
2274                 break;
2275
2276         case PMC_FN_USER_CALLCHAIN_SOFT:
2277                 /*
2278                  * Record a call chain.
2279                  */
2280                 KASSERT(td == curthread, ("[pmc,%d] td != curthread",
2281                     __LINE__));
2282
2283                 cpu = PCPU_GET(cpuid);
2284                 pmc_capture_user_callchain(cpu, PMC_SR,
2285                     (struct trapframe *) arg);
2286
2287                 KASSERT(td->td_pinned == 1,
2288                     ("[pmc,%d] invalid td_pinned value", __LINE__));
2289
2290                 sched_unpin();  /* Can migrate safely now. */
2291
2292                 td->td_pflags &= ~TDP_CALLCHAIN;
2293                 break;
2294
2295         case PMC_FN_SOFT_SAMPLING:
2296                 /*
2297                  * Call soft PMC sampling intr.
2298                  */
2299                 pmc_soft_intr((struct pmckern_soft *) arg);
2300                 break;
2301
2302         case PMC_FN_THR_CREATE:
2303                 pmc_process_thread_add(td);
2304                 pmc_process_threadcreate(td);
2305                 break;
2306
2307         case PMC_FN_THR_CREATE_LOG:
2308                 pmc_process_threadcreate(td);
2309                 break;
2310
2311         case PMC_FN_THR_EXIT:
2312                 KASSERT(td == curthread, ("[pmc,%d] td != curthread",
2313                     __LINE__));
2314                 pmc_process_thread_delete(td);
2315                 pmc_process_threadexit(td);
2316                 break;
2317         case PMC_FN_THR_EXIT_LOG:
2318                 pmc_process_threadexit(td);
2319                 break;
2320         case PMC_FN_THR_USERRET:
2321                 KASSERT(td == curthread, ("[pmc,%d] td != curthread",
2322                     __LINE__));
2323                 pmc_process_thread_userret(td);
2324                 break;
2325
2326         default:
2327 #ifdef  HWPMC_DEBUG
2328                 KASSERT(0, ("[pmc,%d] unknown hook %d\n", __LINE__, function));
2329 #endif
2330                 break;
2331
2332         }
2333
2334         return 0;
2335 }
2336
2337 /*
2338  * allocate a 'struct pmc_owner' descriptor in the owner hash table.
2339  */
2340
2341 static struct pmc_owner *
2342 pmc_allocate_owner_descriptor(struct proc *p)
2343 {
2344         uint32_t hindex;
2345         struct pmc_owner *po;
2346         struct pmc_ownerhash *poh;
2347
2348         hindex = PMC_HASH_PTR(p, pmc_ownerhashmask);
2349         poh = &pmc_ownerhash[hindex];
2350
2351         /* allocate space for N pointers and one descriptor struct */
2352         po = malloc(sizeof(struct pmc_owner), M_PMC, M_WAITOK|M_ZERO);
2353         po->po_owner = p;
2354         LIST_INSERT_HEAD(poh, po, po_next); /* insert into hash table */
2355
2356         TAILQ_INIT(&po->po_logbuffers);
2357         mtx_init(&po->po_mtx, "pmc-owner-mtx", "pmc-per-proc", MTX_SPIN);
2358
2359         PMCDBG4(OWN,ALL,1, "allocate-owner proc=%p (%d, %s) pmc-owner=%p",
2360             p, p->p_pid, p->p_comm, po);
2361
2362         return po;
2363 }
2364
2365 static void
2366 pmc_destroy_owner_descriptor(struct pmc_owner *po)
2367 {
2368
2369         PMCDBG4(OWN,REL,1, "destroy-owner po=%p proc=%p (%d, %s)",
2370             po, po->po_owner, po->po_owner->p_pid, po->po_owner->p_comm);
2371
2372         mtx_destroy(&po->po_mtx);
2373         free(po, M_PMC);
2374 }
2375
2376 /*
2377  * Allocate a thread descriptor from the free pool.
2378  *
2379  * NOTE: This *can* return NULL.
2380  */
2381 static struct pmc_thread *
2382 pmc_thread_descriptor_pool_alloc(void)
2383 {
2384         struct pmc_thread *pt;
2385
2386         mtx_lock_spin(&pmc_threadfreelist_mtx);
2387         if ((pt = LIST_FIRST(&pmc_threadfreelist)) != NULL) {
2388                 LIST_REMOVE(pt, pt_next);
2389                 pmc_threadfreelist_entries--;
2390         }
2391         mtx_unlock_spin(&pmc_threadfreelist_mtx);
2392
2393         return (pt);
2394 }
2395
2396 /*
2397  * Add a thread descriptor to the free pool. We use this instead of free()
2398  * to maintain a cache of free entries. Additionally, we can safely call
2399  * this function when we cannot call free(), such as in a critical section.
2400  *
2401  */
2402 static void
2403 pmc_thread_descriptor_pool_free(struct pmc_thread *pt)
2404 {
2405
2406         if (pt == NULL)
2407                 return;
2408
2409         memset(pt, 0, THREADENTRY_SIZE);
2410         mtx_lock_spin(&pmc_threadfreelist_mtx);
2411         LIST_INSERT_HEAD(&pmc_threadfreelist, pt, pt_next);
2412         pmc_threadfreelist_entries++;
2413         if (pmc_threadfreelist_entries > pmc_threadfreelist_max)
2414                 taskqueue_enqueue(taskqueue_fast, &free_task);
2415         mtx_unlock_spin(&pmc_threadfreelist_mtx);
2416 }
2417
2418 /*
2419  * An asynchronous task to manage the free list.
2420  */
2421 static void
2422 pmc_thread_descriptor_pool_free_task(void *arg __unused, int pending __unused)
2423 {
2424         struct pmc_thread *pt;
2425         LIST_HEAD(, pmc_thread) tmplist;
2426         int delta;
2427
2428         LIST_INIT(&tmplist);
2429
2430         /* Determine what changes, if any, we need to make. */
2431         mtx_lock_spin(&pmc_threadfreelist_mtx);
2432         delta = pmc_threadfreelist_entries - pmc_threadfreelist_max;
2433         while (delta > 0 && (pt = LIST_FIRST(&pmc_threadfreelist)) != NULL) {
2434                 delta--;
2435                 pmc_threadfreelist_entries--;
2436                 LIST_REMOVE(pt, pt_next);
2437                 LIST_INSERT_HEAD(&tmplist, pt, pt_next);
2438         }
2439         mtx_unlock_spin(&pmc_threadfreelist_mtx);
2440
2441         /* If there are entries to free, free them. */
2442         while (!LIST_EMPTY(&tmplist)) {
2443                 pt = LIST_FIRST(&tmplist);
2444                 LIST_REMOVE(pt, pt_next);
2445                 free(pt, M_PMC);
2446         }
2447 }
2448
2449 /*
2450  * Drain the thread free pool, freeing all allocations.
2451  */
2452 static void
2453 pmc_thread_descriptor_pool_drain()
2454 {
2455         struct pmc_thread *pt, *next;
2456
2457         LIST_FOREACH_SAFE(pt, &pmc_threadfreelist, pt_next, next) {
2458                 LIST_REMOVE(pt, pt_next);
2459                 free(pt, M_PMC);
2460         }
2461 }
2462
2463 /*
2464  * find the descriptor corresponding to thread 'td', adding or removing it
2465  * as specified by 'mode'.
2466  *
2467  * Note that this supports additional mode flags in addition to those
2468  * supported by pmc_find_process_descriptor():
2469  * PMC_FLAG_NOWAIT: Causes the function to not wait for mallocs.
2470  *     This makes it safe to call while holding certain other locks.
2471  */
2472
2473 static struct pmc_thread *
2474 pmc_find_thread_descriptor(struct pmc_process *pp, struct thread *td,
2475     uint32_t mode)
2476 {
2477         struct pmc_thread *pt = NULL, *ptnew = NULL;
2478         int wait_flag;
2479
2480         KASSERT(td != NULL, ("[pmc,%d] called to add NULL td", __LINE__));
2481
2482         /*
2483          * Pre-allocate memory in the PMC_FLAG_ALLOCATE case prior to
2484          * acquiring the lock.
2485          */
2486         if (mode & PMC_FLAG_ALLOCATE) {
2487                 if ((ptnew = pmc_thread_descriptor_pool_alloc()) == NULL) {
2488                         wait_flag = M_WAITOK;
2489                         if ((mode & PMC_FLAG_NOWAIT) || in_epoch(global_epoch_preempt))
2490                                 wait_flag = M_NOWAIT;
2491
2492                         ptnew = malloc(THREADENTRY_SIZE, M_PMC,
2493                             wait_flag|M_ZERO);
2494                 }
2495         }
2496
2497         mtx_lock_spin(pp->pp_tdslock);
2498
2499         LIST_FOREACH(pt, &pp->pp_tds, pt_next)
2500                 if (pt->pt_td == td)
2501                         break;
2502
2503         if ((mode & PMC_FLAG_REMOVE) && pt != NULL)
2504                 LIST_REMOVE(pt, pt_next);
2505
2506         if ((mode & PMC_FLAG_ALLOCATE) && pt == NULL && ptnew != NULL) {
2507                 pt = ptnew;
2508                 ptnew = NULL;
2509                 pt->pt_td = td;
2510                 LIST_INSERT_HEAD(&pp->pp_tds, pt, pt_next);
2511         }
2512
2513         mtx_unlock_spin(pp->pp_tdslock);
2514
2515         if (ptnew != NULL) {
2516                 free(ptnew, M_PMC);
2517         }
2518
2519         return pt;
2520 }
2521
2522 /*
2523  * Try to add thread descriptors for each thread in a process.
2524  */
2525
2526 static void
2527 pmc_add_thread_descriptors_from_proc(struct proc *p, struct pmc_process *pp)
2528 {
2529         struct thread *curtd;
2530         struct pmc_thread **tdlist;
2531         int i, tdcnt, tdlistsz;
2532
2533         KASSERT(!PROC_LOCKED(p), ("[pmc,%d] proc unexpectedly locked",
2534             __LINE__));
2535         tdcnt = 32;
2536  restart:
2537         tdlistsz = roundup2(tdcnt, 32);
2538
2539         tdcnt = 0;
2540         tdlist = malloc(sizeof(struct pmc_thread*) * tdlistsz, M_TEMP, M_WAITOK);
2541
2542         PROC_LOCK(p);
2543         FOREACH_THREAD_IN_PROC(p, curtd)
2544                 tdcnt++;
2545         if (tdcnt >= tdlistsz) {
2546                 PROC_UNLOCK(p);
2547                 free(tdlist, M_TEMP);
2548                 goto restart;
2549         }
2550         /*
2551          * Try to add each thread to the list without sleeping. If unable,
2552          * add to a queue to retry after dropping the process lock.
2553          */
2554         tdcnt = 0;
2555         FOREACH_THREAD_IN_PROC(p, curtd) {
2556                 tdlist[tdcnt] = pmc_find_thread_descriptor(pp, curtd,
2557                                                    PMC_FLAG_ALLOCATE|PMC_FLAG_NOWAIT);
2558                 if (tdlist[tdcnt] == NULL) {
2559                         PROC_UNLOCK(p);
2560                         for (i = 0; i <= tdcnt; i++)
2561                                 pmc_thread_descriptor_pool_free(tdlist[i]);
2562                         free(tdlist, M_TEMP);
2563                         goto restart;
2564                 }
2565                 tdcnt++;
2566         }
2567         PROC_UNLOCK(p);
2568         free(tdlist, M_TEMP);
2569 }
2570
2571 /*
2572  * find the descriptor corresponding to process 'p', adding or removing it
2573  * as specified by 'mode'.
2574  */
2575
2576 static struct pmc_process *
2577 pmc_find_process_descriptor(struct proc *p, uint32_t mode)
2578 {
2579         uint32_t hindex;
2580         struct pmc_process *pp, *ppnew;
2581         struct pmc_processhash *pph;
2582
2583         hindex = PMC_HASH_PTR(p, pmc_processhashmask);
2584         pph = &pmc_processhash[hindex];
2585
2586         ppnew = NULL;
2587
2588         /*
2589          * Pre-allocate memory in the PMC_FLAG_ALLOCATE case since we
2590          * cannot call malloc(9) once we hold a spin lock.
2591          */
2592         if (mode & PMC_FLAG_ALLOCATE)
2593                 ppnew = malloc(sizeof(struct pmc_process) + md->pmd_npmc *
2594                     sizeof(struct pmc_targetstate), M_PMC, M_WAITOK|M_ZERO);
2595
2596         mtx_lock_spin(&pmc_processhash_mtx);
2597         LIST_FOREACH(pp, pph, pp_next)
2598             if (pp->pp_proc == p)
2599                     break;
2600
2601         if ((mode & PMC_FLAG_REMOVE) && pp != NULL)
2602                 LIST_REMOVE(pp, pp_next);
2603
2604         if ((mode & PMC_FLAG_ALLOCATE) && pp == NULL &&
2605             ppnew != NULL) {
2606                 ppnew->pp_proc = p;
2607                 LIST_INIT(&ppnew->pp_tds);
2608                 ppnew->pp_tdslock = mtx_pool_find(pmc_mtxpool, ppnew);
2609                 LIST_INSERT_HEAD(pph, ppnew, pp_next);
2610                 mtx_unlock_spin(&pmc_processhash_mtx);
2611                 pp = ppnew;
2612                 ppnew = NULL;
2613
2614                 /* Add thread descriptors for this process' current threads. */
2615                 pmc_add_thread_descriptors_from_proc(p, pp);
2616         }
2617         else
2618                 mtx_unlock_spin(&pmc_processhash_mtx);
2619
2620         if (ppnew != NULL)
2621                 free(ppnew, M_PMC);
2622
2623         return pp;
2624 }
2625
2626 /*
2627  * remove a process descriptor from the process hash table.
2628  */
2629
2630 static void
2631 pmc_remove_process_descriptor(struct pmc_process *pp)
2632 {
2633         KASSERT(pp->pp_refcnt == 0,
2634             ("[pmc,%d] Removing process descriptor %p with count %d",
2635                 __LINE__, pp, pp->pp_refcnt));
2636
2637         mtx_lock_spin(&pmc_processhash_mtx);
2638         LIST_REMOVE(pp, pp_next);
2639         mtx_unlock_spin(&pmc_processhash_mtx);
2640 }
2641
2642 /*
2643  * destroy a process descriptor.
2644  */
2645
2646 static void
2647 pmc_destroy_process_descriptor(struct pmc_process *pp)
2648 {
2649         struct pmc_thread *pmc_td;
2650
2651         while ((pmc_td = LIST_FIRST(&pp->pp_tds)) != NULL) {
2652                 LIST_REMOVE(pmc_td, pt_next);
2653                 pmc_thread_descriptor_pool_free(pmc_td);
2654         }
2655         free(pp, M_PMC);
2656 }
2657
2658
2659 /*
2660  * find an owner descriptor corresponding to proc 'p'
2661  */
2662
2663 static struct pmc_owner *
2664 pmc_find_owner_descriptor(struct proc *p)
2665 {
2666         uint32_t hindex;
2667         struct pmc_owner *po;
2668         struct pmc_ownerhash *poh;
2669
2670         hindex = PMC_HASH_PTR(p, pmc_ownerhashmask);
2671         poh = &pmc_ownerhash[hindex];
2672
2673         po = NULL;
2674         LIST_FOREACH(po, poh, po_next)
2675             if (po->po_owner == p)
2676                     break;
2677
2678         PMCDBG5(OWN,FND,1, "find-owner proc=%p (%d, %s) hindex=0x%x -> "
2679             "pmc-owner=%p", p, p->p_pid, p->p_comm, hindex, po);
2680
2681         return po;
2682 }
2683
2684 /*
2685  * pmc_allocate_pmc_descriptor
2686  *
2687  * Allocate a pmc descriptor and initialize its
2688  * fields.
2689  */
2690
2691 static struct pmc *
2692 pmc_allocate_pmc_descriptor(void)
2693 {
2694         struct pmc *pmc;
2695
2696         pmc = malloc(sizeof(struct pmc), M_PMC, M_WAITOK|M_ZERO);
2697         pmc->pm_runcount = counter_u64_alloc(M_WAITOK);
2698         pmc->pm_pcpu_state = malloc(sizeof(struct pmc_pcpu_state)*mp_ncpus, M_PMC, M_WAITOK|M_ZERO);
2699         PMCDBG1(PMC,ALL,1, "allocate-pmc -> pmc=%p", pmc);
2700
2701         return pmc;
2702 }
2703
2704 /*
2705  * Destroy a pmc descriptor.
2706  */
2707
2708 static void
2709 pmc_destroy_pmc_descriptor(struct pmc *pm)
2710 {
2711
2712         KASSERT(pm->pm_state == PMC_STATE_DELETED ||
2713             pm->pm_state == PMC_STATE_FREE,
2714             ("[pmc,%d] destroying non-deleted PMC", __LINE__));
2715         KASSERT(LIST_EMPTY(&pm->pm_targets),
2716             ("[pmc,%d] destroying pmc with targets", __LINE__));
2717         KASSERT(pm->pm_owner == NULL,
2718             ("[pmc,%d] destroying pmc attached to an owner", __LINE__));
2719         KASSERT(counter_u64_fetch(pm->pm_runcount) == 0,
2720             ("[pmc,%d] pmc has non-zero run count %ld", __LINE__,
2721                  (unsigned long)counter_u64_fetch(pm->pm_runcount)));
2722
2723         counter_u64_free(pm->pm_runcount);
2724         free(pm->pm_pcpu_state, M_PMC);
2725         free(pm, M_PMC);
2726 }
2727
2728 static void
2729 pmc_wait_for_pmc_idle(struct pmc *pm)
2730 {
2731 #ifdef INVARIANTS
2732         volatile int maxloop;
2733
2734         maxloop = 100 * pmc_cpu_max();
2735 #endif
2736         /*
2737          * Loop (with a forced context switch) till the PMC's runcount
2738          * comes down to zero.
2739          */
2740         pmclog_flush(pm->pm_owner, 1);
2741         while (counter_u64_fetch(pm->pm_runcount) > 0) {
2742                 pmclog_flush(pm->pm_owner, 1);
2743 #ifdef INVARIANTS
2744                 maxloop--;
2745                 KASSERT(maxloop > 0,
2746                     ("[pmc,%d] (ri%d, rc%ld) waiting too long for "
2747                         "pmc to be free", __LINE__,
2748                          PMC_TO_ROWINDEX(pm), (unsigned long)counter_u64_fetch(pm->pm_runcount)));
2749 #endif
2750                 pmc_force_context_switch();
2751         }
2752 }
2753
2754 /*
2755  * This function does the following things:
2756  *
2757  *  - detaches the PMC from hardware
2758  *  - unlinks all target threads that were attached to it
2759  *  - removes the PMC from its owner's list
2760  *  - destroys the PMC private mutex
2761  *
2762  * Once this function completes, the given pmc pointer can be freed by
2763  * calling pmc_destroy_pmc_descriptor().
2764  */
2765
2766 static void
2767 pmc_release_pmc_descriptor(struct pmc *pm)
2768 {
2769         enum pmc_mode mode;
2770         struct pmc_hw *phw;
2771         u_int adjri, ri, cpu;
2772         struct pmc_owner *po;
2773         struct pmc_binding pb;
2774         struct pmc_process *pp;
2775         struct pmc_classdep *pcd;
2776         struct pmc_target *ptgt, *tmp;
2777
2778         sx_assert(&pmc_sx, SX_XLOCKED);
2779
2780         KASSERT(pm, ("[pmc,%d] null pmc", __LINE__));
2781
2782         ri   = PMC_TO_ROWINDEX(pm);
2783         pcd  = pmc_ri_to_classdep(md, ri, &adjri);
2784         mode = PMC_TO_MODE(pm);
2785
2786         PMCDBG3(PMC,REL,1, "release-pmc pmc=%p ri=%d mode=%d", pm, ri,
2787             mode);
2788
2789         /*
2790          * First, we take the PMC off hardware.
2791          */
2792         cpu = 0;
2793         if (PMC_IS_SYSTEM_MODE(mode)) {
2794
2795                 /*
2796                  * A system mode PMC runs on a specific CPU.  Switch
2797                  * to this CPU and turn hardware off.
2798                  */
2799                 pmc_save_cpu_binding(&pb);
2800
2801                 cpu = PMC_TO_CPU(pm);
2802
2803                 pmc_select_cpu(cpu);
2804
2805                 /* switch off non-stalled CPUs */
2806                 pm->pm_pcpu_state[cpu].pps_cpustate = 0;
2807                 if (pm->pm_state == PMC_STATE_RUNNING &&
2808                         pm->pm_pcpu_state[cpu].pps_stalled == 0) {
2809
2810                         phw = pmc_pcpu[cpu]->pc_hwpmcs[ri];
2811
2812                         KASSERT(phw->phw_pmc == pm,
2813                             ("[pmc, %d] pmc ptr ri(%d) hw(%p) pm(%p)",
2814                                 __LINE__, ri, phw->phw_pmc, pm));
2815                         PMCDBG2(PMC,REL,2, "stopping cpu=%d ri=%d", cpu, ri);
2816
2817                         critical_enter();
2818                         pcd->pcd_stop_pmc(cpu, adjri);
2819                         critical_exit();
2820                 }
2821
2822                 PMCDBG2(PMC,REL,2, "decfg cpu=%d ri=%d", cpu, ri);
2823
2824                 critical_enter();
2825                 pcd->pcd_config_pmc(cpu, adjri, NULL);
2826                 critical_exit();
2827
2828                 /* adjust the global and process count of SS mode PMCs */
2829                 if (mode == PMC_MODE_SS && pm->pm_state == PMC_STATE_RUNNING) {
2830                         po = pm->pm_owner;
2831                         po->po_sscount--;
2832                         if (po->po_sscount == 0) {
2833                                 atomic_subtract_rel_int(&pmc_ss_count, 1);
2834                                 CK_LIST_REMOVE(po, po_ssnext);
2835                                 epoch_wait_preempt(global_epoch_preempt);
2836                         }
2837                 }
2838
2839                 pm->pm_state = PMC_STATE_DELETED;
2840
2841                 pmc_restore_cpu_binding(&pb);
2842
2843                 /*
2844                  * We could have references to this PMC structure in
2845                  * the per-cpu sample queues.  Wait for the queue to
2846                  * drain.
2847                  */
2848                 pmc_wait_for_pmc_idle(pm);
2849
2850         } else if (PMC_IS_VIRTUAL_MODE(mode)) {
2851
2852                 /*
2853                  * A virtual PMC could be running on multiple CPUs at
2854                  * a given instant.
2855                  *
2856                  * By marking its state as DELETED, we ensure that
2857                  * this PMC is never further scheduled on hardware.
2858                  *
2859                  * Then we wait till all CPUs are done with this PMC.
2860                  */
2861                 pm->pm_state = PMC_STATE_DELETED;
2862
2863
2864                 /* Wait for the PMCs runcount to come to zero. */
2865                 pmc_wait_for_pmc_idle(pm);
2866
2867                 /*
2868                  * At this point the PMC is off all CPUs and cannot be
2869                  * freshly scheduled onto a CPU.  It is now safe to
2870                  * unlink all targets from this PMC.  If a
2871                  * process-record's refcount falls to zero, we remove
2872                  * it from the hash table.  The module-wide SX lock
2873                  * protects us from races.
2874                  */
2875                 LIST_FOREACH_SAFE(ptgt, &pm->pm_targets, pt_next, tmp) {
2876                         pp = ptgt->pt_process;
2877                         pmc_unlink_target_process(pm, pp); /* frees 'ptgt' */
2878
2879                         PMCDBG1(PMC,REL,3, "pp->refcnt=%d", pp->pp_refcnt);
2880
2881                         /*
2882                          * If the target process record shows that no
2883                          * PMCs are attached to it, reclaim its space.
2884                          */
2885
2886                         if (pp->pp_refcnt == 0) {
2887                                 pmc_remove_process_descriptor(pp);
2888                                 pmc_destroy_process_descriptor(pp);
2889                         }
2890                 }
2891
2892                 cpu = curthread->td_oncpu; /* setup cpu for pmd_release() */
2893
2894         }
2895
2896         /*
2897          * Release any MD resources
2898          */
2899         (void) pcd->pcd_release_pmc(cpu, adjri, pm);
2900
2901         /*
2902          * Update row disposition
2903          */
2904
2905         if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm)))
2906                 PMC_UNMARK_ROW_STANDALONE(ri);
2907         else
2908                 PMC_UNMARK_ROW_THREAD(ri);
2909
2910         /* unlink from the owner's list */
2911         if (pm->pm_owner) {
2912                 LIST_REMOVE(pm, pm_next);
2913                 pm->pm_owner = NULL;
2914         }
2915 }
2916
2917 /*
2918  * Register an owner and a pmc.
2919  */
2920
2921 static int
2922 pmc_register_owner(struct proc *p, struct pmc *pmc)
2923 {
2924         struct pmc_owner *po;
2925
2926         sx_assert(&pmc_sx, SX_XLOCKED);
2927
2928         if ((po = pmc_find_owner_descriptor(p)) == NULL)
2929                 if ((po = pmc_allocate_owner_descriptor(p)) == NULL)
2930                         return ENOMEM;
2931
2932         KASSERT(pmc->pm_owner == NULL,
2933             ("[pmc,%d] attempting to own an initialized PMC", __LINE__));
2934         pmc->pm_owner  = po;
2935
2936         LIST_INSERT_HEAD(&po->po_pmcs, pmc, pm_next);
2937
2938         PROC_LOCK(p);
2939         p->p_flag |= P_HWPMC;
2940         PROC_UNLOCK(p);
2941
2942         if (po->po_flags & PMC_PO_OWNS_LOGFILE)
2943                 pmclog_process_pmcallocate(pmc);
2944
2945         PMCDBG2(PMC,REG,1, "register-owner pmc-owner=%p pmc=%p",
2946             po, pmc);
2947
2948         return 0;
2949 }
2950
2951 /*
2952  * Return the current row disposition:
2953  * == 0 => FREE
2954  *  > 0 => PROCESS MODE
2955  *  < 0 => SYSTEM MODE
2956  */
2957
2958 int
2959 pmc_getrowdisp(int ri)
2960 {
2961         return pmc_pmcdisp[ri];
2962 }
2963
2964 /*
2965  * Check if a PMC at row index 'ri' can be allocated to the current
2966  * process.
2967  *
2968  * Allocation can fail if:
2969  *   - the current process is already being profiled by a PMC at index 'ri',
2970  *     attached to it via OP_PMCATTACH.
2971  *   - the current process has already allocated a PMC at index 'ri'
2972  *     via OP_ALLOCATE.
2973  */
2974
2975 static int
2976 pmc_can_allocate_rowindex(struct proc *p, unsigned int ri, int cpu)
2977 {
2978         enum pmc_mode mode;
2979         struct pmc *pm;
2980         struct pmc_owner *po;
2981         struct pmc_process *pp;
2982
2983         PMCDBG5(PMC,ALR,1, "can-allocate-rowindex proc=%p (%d, %s) ri=%d "
2984             "cpu=%d", p, p->p_pid, p->p_comm, ri, cpu);
2985
2986         /*
2987          * We shouldn't have already allocated a process-mode PMC at
2988          * row index 'ri'.
2989          *
2990          * We shouldn't have allocated a system-wide PMC on the same
2991          * CPU and same RI.
2992          */
2993         if ((po = pmc_find_owner_descriptor(p)) != NULL)
2994                 LIST_FOREACH(pm, &po->po_pmcs, pm_next) {
2995                     if (PMC_TO_ROWINDEX(pm) == ri) {
2996                             mode = PMC_TO_MODE(pm);
2997                             if (PMC_IS_VIRTUAL_MODE(mode))
2998                                     return EEXIST;
2999                             if (PMC_IS_SYSTEM_MODE(mode) &&
3000                                 (int) PMC_TO_CPU(pm) == cpu)
3001                                     return EEXIST;
3002                     }
3003                 }
3004
3005         /*
3006          * We also shouldn't be the target of any PMC at this index
3007          * since otherwise a PMC_ATTACH to ourselves will fail.
3008          */
3009         if ((pp = pmc_find_process_descriptor(p, 0)) != NULL)
3010                 if (pp->pp_pmcs[ri].pp_pmc)
3011                         return EEXIST;
3012
3013         PMCDBG4(PMC,ALR,2, "can-allocate-rowindex proc=%p (%d, %s) ri=%d ok",
3014             p, p->p_pid, p->p_comm, ri);
3015
3016         return 0;
3017 }
3018
3019 /*
3020  * Check if a given PMC at row index 'ri' can be currently used in
3021  * mode 'mode'.
3022  */
3023
3024 static int
3025 pmc_can_allocate_row(int ri, enum pmc_mode mode)
3026 {
3027         enum pmc_disp   disp;
3028
3029         sx_assert(&pmc_sx, SX_XLOCKED);
3030
3031         PMCDBG2(PMC,ALR,1, "can-allocate-row ri=%d mode=%d", ri, mode);
3032
3033         if (PMC_IS_SYSTEM_MODE(mode))
3034                 disp = PMC_DISP_STANDALONE;
3035         else
3036                 disp = PMC_DISP_THREAD;
3037
3038         /*
3039          * check disposition for PMC row 'ri':
3040          *
3041          * Expected disposition         Row-disposition         Result
3042          *
3043          * STANDALONE                   STANDALONE or FREE      proceed
3044          * STANDALONE                   THREAD                  fail
3045          * THREAD                       THREAD or FREE          proceed
3046          * THREAD                       STANDALONE              fail
3047          */
3048
3049         if (!PMC_ROW_DISP_IS_FREE(ri) &&
3050             !(disp == PMC_DISP_THREAD && PMC_ROW_DISP_IS_THREAD(ri)) &&
3051             !(disp == PMC_DISP_STANDALONE && PMC_ROW_DISP_IS_STANDALONE(ri)))
3052                 return EBUSY;
3053
3054         /*
3055          * All OK
3056          */
3057
3058         PMCDBG2(PMC,ALR,2, "can-allocate-row ri=%d mode=%d ok", ri, mode);
3059
3060         return 0;
3061
3062 }
3063
3064 /*
3065  * Find a PMC descriptor with user handle 'pmcid' for thread 'td'.
3066  */
3067
3068 static struct pmc *
3069 pmc_find_pmc_descriptor_in_process(struct pmc_owner *po, pmc_id_t pmcid)
3070 {
3071         struct pmc *pm;
3072
3073         KASSERT(PMC_ID_TO_ROWINDEX(pmcid) < md->pmd_npmc,
3074             ("[pmc,%d] Illegal pmc index %d (max %d)", __LINE__,
3075                 PMC_ID_TO_ROWINDEX(pmcid), md->pmd_npmc));
3076
3077         LIST_FOREACH(pm, &po->po_pmcs, pm_next)
3078             if (pm->pm_id == pmcid)
3079                     return pm;
3080
3081         return NULL;
3082 }
3083
3084 static int
3085 pmc_find_pmc(pmc_id_t pmcid, struct pmc **pmc)
3086 {
3087
3088         struct pmc *pm, *opm;
3089         struct pmc_owner *po;
3090         struct pmc_process *pp;
3091
3092         PMCDBG1(PMC,FND,1, "find-pmc id=%d", pmcid);
3093         if (PMC_ID_TO_ROWINDEX(pmcid) >= md->pmd_npmc)
3094                 return (EINVAL);
3095
3096         if ((po = pmc_find_owner_descriptor(curthread->td_proc)) == NULL) {
3097                 /*
3098                  * In case of PMC_F_DESCENDANTS child processes we will not find
3099                  * the current process in the owners hash list.  Find the owner
3100                  * process first and from there lookup the po.
3101                  */
3102                 if ((pp = pmc_find_process_descriptor(curthread->td_proc,
3103                     PMC_FLAG_NONE)) == NULL) {
3104                         return ESRCH;
3105                 } else {
3106                         opm = pp->pp_pmcs[PMC_ID_TO_ROWINDEX(pmcid)].pp_pmc;
3107                         if (opm == NULL)
3108                                 return ESRCH;
3109                         if ((opm->pm_flags & (PMC_F_ATTACHED_TO_OWNER|
3110                             PMC_F_DESCENDANTS)) != (PMC_F_ATTACHED_TO_OWNER|
3111                             PMC_F_DESCENDANTS))
3112                                 return ESRCH;
3113                         po = opm->pm_owner;
3114                 }
3115         }
3116
3117         if ((pm = pmc_find_pmc_descriptor_in_process(po, pmcid)) == NULL)
3118                 return EINVAL;
3119
3120         PMCDBG2(PMC,FND,2, "find-pmc id=%d -> pmc=%p", pmcid, pm);
3121
3122         *pmc = pm;
3123         return 0;
3124 }
3125
3126 /*
3127  * Start a PMC.
3128  */
3129
3130 static int
3131 pmc_start(struct pmc *pm)
3132 {
3133         enum pmc_mode mode;
3134         struct pmc_owner *po;
3135         struct pmc_binding pb;
3136         struct pmc_classdep *pcd;
3137         int adjri, error, cpu, ri;
3138
3139         KASSERT(pm != NULL,
3140             ("[pmc,%d] null pm", __LINE__));
3141
3142         mode = PMC_TO_MODE(pm);
3143         ri   = PMC_TO_ROWINDEX(pm);
3144         pcd  = pmc_ri_to_classdep(md, ri, &adjri);
3145
3146         error = 0;
3147
3148         PMCDBG3(PMC,OPS,1, "start pmc=%p mode=%d ri=%d", pm, mode, ri);
3149
3150         po = pm->pm_owner;
3151
3152         /*
3153          * Disallow PMCSTART if a logfile is required but has not been
3154          * configured yet.
3155          */
3156         if ((pm->pm_flags & PMC_F_NEEDS_LOGFILE) &&
3157             (po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
3158                 return (EDOOFUS);       /* programming error */
3159
3160         /*
3161          * If this is a sampling mode PMC, log mapping information for
3162          * the kernel modules that are currently loaded.
3163          */
3164         if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
3165             pmc_log_kernel_mappings(pm);
3166
3167         if (PMC_IS_VIRTUAL_MODE(mode)) {
3168
3169                 /*
3170                  * If a PMCATTACH has never been done on this PMC,
3171                  * attach it to its owner process.
3172                  */
3173
3174                 if (LIST_EMPTY(&pm->pm_targets))
3175                         error = (pm->pm_flags & PMC_F_ATTACH_DONE) ? ESRCH :
3176                             pmc_attach_process(po->po_owner, pm);
3177
3178                 /*
3179                  * If the PMC is attached to its owner, then force a context
3180                  * switch to ensure that the MD state gets set correctly.
3181                  */
3182
3183                 if (error == 0) {
3184                         pm->pm_state = PMC_STATE_RUNNING;
3185                         if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER)
3186                                 pmc_force_context_switch();
3187                 }
3188
3189                 return (error);
3190         }
3191
3192
3193         /*
3194          * A system-wide PMC.
3195          *
3196          * Add the owner to the global list if this is a system-wide
3197          * sampling PMC.
3198          */
3199
3200         if (mode == PMC_MODE_SS) {
3201                 /*
3202                  * Log mapping information for all existing processes in the
3203                  * system.  Subsequent mappings are logged as they happen;
3204                  * see pmc_process_mmap().
3205                  */
3206                 if (po->po_logprocmaps == 0) {
3207                         pmc_log_all_process_mappings(po);
3208                         po->po_logprocmaps = 1;
3209                 }
3210                 po->po_sscount++;
3211                 if (po->po_sscount == 1) {
3212                         atomic_add_rel_int(&pmc_ss_count, 1);
3213                         CK_LIST_INSERT_HEAD(&pmc_ss_owners, po, po_ssnext);
3214                         PMCDBG1(PMC,OPS,1, "po=%p in global list", po);
3215                 }
3216         }
3217
3218         /*
3219          * Move to the CPU associated with this
3220          * PMC, and start the hardware.
3221          */
3222
3223         pmc_save_cpu_binding(&pb);
3224
3225         cpu = PMC_TO_CPU(pm);
3226
3227         if (!pmc_cpu_is_active(cpu))
3228                 return (ENXIO);
3229
3230         pmc_select_cpu(cpu);
3231
3232         /*
3233          * global PMCs are configured at allocation time
3234          * so write out the initial value and start the PMC.
3235          */
3236
3237         pm->pm_state = PMC_STATE_RUNNING;
3238
3239         critical_enter();
3240         if ((error = pcd->pcd_write_pmc(cpu, adjri,
3241                  PMC_IS_SAMPLING_MODE(mode) ?
3242                  pm->pm_sc.pm_reloadcount :
3243                  pm->pm_sc.pm_initial)) == 0) {
3244                 /* If a sampling mode PMC, reset stalled state. */
3245                 if (PMC_IS_SAMPLING_MODE(mode))
3246                         pm->pm_pcpu_state[cpu].pps_stalled = 0;
3247
3248                 /* Indicate that we desire this to run. Start it. */
3249                 pm->pm_pcpu_state[cpu].pps_cpustate = 1;
3250                 error = pcd->pcd_start_pmc(cpu, adjri);
3251         }
3252         critical_exit();
3253
3254         pmc_restore_cpu_binding(&pb);
3255
3256         return (error);
3257 }
3258
3259 /*
3260  * Stop a PMC.
3261  */
3262
3263 static int
3264 pmc_stop(struct pmc *pm)
3265 {
3266         struct pmc_owner *po;
3267         struct pmc_binding pb;
3268         struct pmc_classdep *pcd;
3269         int adjri, cpu, error, ri;
3270
3271         KASSERT(pm != NULL, ("[pmc,%d] null pmc", __LINE__));
3272
3273         PMCDBG3(PMC,OPS,1, "stop pmc=%p mode=%d ri=%d", pm,
3274             PMC_TO_MODE(pm), PMC_TO_ROWINDEX(pm));
3275
3276         pm->pm_state = PMC_STATE_STOPPED;
3277
3278         /*
3279          * If the PMC is a virtual mode one, changing the state to
3280          * non-RUNNING is enough to ensure that the PMC never gets
3281          * scheduled.
3282          *
3283          * If this PMC is current running on a CPU, then it will
3284          * handled correctly at the time its target process is context
3285          * switched out.
3286          */
3287
3288         if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)))
3289                 return 0;
3290
3291         /*
3292          * A system-mode PMC.  Move to the CPU associated with
3293          * this PMC, and stop the hardware.  We update the
3294          * 'initial count' so that a subsequent PMCSTART will
3295          * resume counting from the current hardware count.
3296          */
3297
3298         pmc_save_cpu_binding(&pb);
3299
3300         cpu = PMC_TO_CPU(pm);
3301
3302         KASSERT(cpu >= 0 && cpu < pmc_cpu_max(),
3303             ("[pmc,%d] illegal cpu=%d", __LINE__, cpu));
3304
3305         if (!pmc_cpu_is_active(cpu))
3306                 return ENXIO;
3307
3308         pmc_select_cpu(cpu);
3309
3310         ri = PMC_TO_ROWINDEX(pm);
3311         pcd = pmc_ri_to_classdep(md, ri, &adjri);
3312
3313         pm->pm_pcpu_state[cpu].pps_cpustate = 0;
3314         critical_enter();
3315         if ((error = pcd->pcd_stop_pmc(cpu, adjri)) == 0)
3316                 error = pcd->pcd_read_pmc(cpu, adjri, &pm->pm_sc.pm_initial);
3317         critical_exit();
3318
3319         pmc_restore_cpu_binding(&pb);
3320
3321         po = pm->pm_owner;
3322
3323         /* remove this owner from the global list of SS PMC owners */
3324         if (PMC_TO_MODE(pm) == PMC_MODE_SS) {
3325                 po->po_sscount--;
3326                 if (po->po_sscount == 0) {
3327                         atomic_subtract_rel_int(&pmc_ss_count, 1);
3328                         CK_LIST_REMOVE(po, po_ssnext);
3329                         epoch_wait_preempt(global_epoch_preempt);
3330                         PMCDBG1(PMC,OPS,2,"po=%p removed from global list", po);
3331                 }
3332         }
3333
3334         return (error);
3335 }
3336
3337 static struct pmc_classdep *
3338 pmc_class_to_classdep(enum pmc_class class)
3339 {
3340         int n;
3341
3342         for (n = 0; n < md->pmd_nclass; n++)
3343                 if (md->pmd_classdep[n].pcd_class == class)
3344                         return (&md->pmd_classdep[n]);
3345         return (NULL);
3346 }
3347
3348 #if defined(HWPMC_DEBUG) && defined(KTR)
3349 static const char *pmc_op_to_name[] = {
3350 #undef  __PMC_OP
3351 #define __PMC_OP(N, D)  #N ,
3352         __PMC_OPS()
3353         NULL
3354 };
3355 #endif
3356
3357 /*
3358  * The syscall interface
3359  */
3360
3361 #define PMC_GET_SX_XLOCK(...) do {              \
3362         sx_xlock(&pmc_sx);                      \
3363         if (pmc_hook == NULL) {                 \
3364                 sx_xunlock(&pmc_sx);            \
3365                 return __VA_ARGS__;             \
3366         }                                       \
3367 } while (0)
3368
3369 #define PMC_DOWNGRADE_SX() do {                 \
3370         sx_downgrade(&pmc_sx);                  \
3371         is_sx_downgraded = 1;                   \
3372 } while (0)
3373
3374 static int
3375 pmc_syscall_handler(struct thread *td, void *syscall_args)
3376 {
3377         int error, is_sx_downgraded, op;
3378         struct pmc_syscall_args *c;
3379         void *pmclog_proc_handle;
3380         void *arg;
3381
3382         c = (struct pmc_syscall_args *)syscall_args;
3383         op = c->pmop_code;
3384         arg = c->pmop_data;
3385         /* PMC isn't set up yet */
3386         if (pmc_hook == NULL)
3387                 return (EINVAL);
3388         if (op == PMC_OP_CONFIGURELOG) {
3389                 /*
3390                  * We cannot create the logging process inside
3391                  * pmclog_configure_log() because there is a LOR
3392                  * between pmc_sx and process structure locks.
3393                  * Instead, pre-create the process and ignite the loop
3394                  * if everything is fine, otherwise direct the process
3395                  * to exit.
3396                  */
3397                 error = pmclog_proc_create(td, &pmclog_proc_handle);
3398                 if (error != 0)
3399                         goto done_syscall;
3400         }
3401
3402         PMC_GET_SX_XLOCK(ENOSYS);
3403         is_sx_downgraded = 0;
3404         PMCDBG3(MOD,PMS,1, "syscall op=%d \"%s\" arg=%p", op,
3405             pmc_op_to_name[op], arg);
3406
3407         error = 0;
3408         counter_u64_add(pmc_stats.pm_syscalls, 1);
3409
3410         switch (op) {
3411
3412
3413         /*
3414          * Configure a log file.
3415          *
3416          * XXX This OP will be reworked.
3417          */
3418
3419         case PMC_OP_CONFIGURELOG:
3420         {
3421                 struct proc *p;
3422                 struct pmc *pm;
3423                 struct pmc_owner *po;
3424                 struct pmc_op_configurelog cl;
3425
3426                 if ((error = copyin(arg, &cl, sizeof(cl))) != 0) {
3427                         pmclog_proc_ignite(pmclog_proc_handle, NULL);
3428                         break;
3429                 }
3430
3431                 /* mark this process as owning a log file */
3432                 p = td->td_proc;
3433                 if ((po = pmc_find_owner_descriptor(p)) == NULL)
3434                         if ((po = pmc_allocate_owner_descriptor(p)) == NULL) {
3435                                 pmclog_proc_ignite(pmclog_proc_handle, NULL);
3436                                 error = ENOMEM;
3437                                 break;
3438                         }
3439
3440                 /*
3441                  * If a valid fd was passed in, try to configure that,
3442                  * otherwise if 'fd' was less than zero and there was
3443                  * a log file configured, flush its buffers and
3444                  * de-configure it.
3445                  */
3446                 if (cl.pm_logfd >= 0) {
3447                         error = pmclog_configure_log(md, po, cl.pm_logfd);
3448                         pmclog_proc_ignite(pmclog_proc_handle, error == 0 ?
3449                             po : NULL);
3450                 } else if (po->po_flags & PMC_PO_OWNS_LOGFILE) {
3451                         pmclog_proc_ignite(pmclog_proc_handle, NULL);
3452                         error = pmclog_close(po);
3453                         if (error == 0) {
3454                                 LIST_FOREACH(pm, &po->po_pmcs, pm_next)
3455                                     if (pm->pm_flags & PMC_F_NEEDS_LOGFILE &&
3456                                         pm->pm_state == PMC_STATE_RUNNING)
3457                                             pmc_stop(pm);
3458                                 error = pmclog_deconfigure_log(po);
3459                         }
3460                 } else {
3461                         pmclog_proc_ignite(pmclog_proc_handle, NULL);
3462                         error = EINVAL;
3463                 }
3464         }
3465         break;
3466
3467         /*
3468          * Flush a log file.
3469          */
3470
3471         case PMC_OP_FLUSHLOG:
3472         {
3473                 struct pmc_owner *po;
3474
3475                 sx_assert(&pmc_sx, SX_XLOCKED);
3476
3477                 if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) {
3478                         error = EINVAL;
3479                         break;
3480                 }
3481
3482                 error = pmclog_flush(po, 0);
3483         }
3484         break;
3485
3486         /*
3487          * Close a log file.
3488          */
3489
3490         case PMC_OP_CLOSELOG:
3491         {
3492                 struct pmc_owner *po;
3493
3494                 sx_assert(&pmc_sx, SX_XLOCKED);
3495
3496                 if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) {
3497                         error = EINVAL;
3498                         break;
3499                 }
3500
3501                 error = pmclog_close(po);
3502         }
3503         break;
3504
3505         /*
3506          * Retrieve hardware configuration.
3507          */
3508
3509         case PMC_OP_GETCPUINFO: /* CPU information */
3510         {
3511                 struct pmc_op_getcpuinfo gci;
3512                 struct pmc_classinfo *pci;
3513                 struct pmc_classdep *pcd;
3514                 int cl;
3515
3516                 memset(&gci, 0, sizeof(gci));
3517                 gci.pm_cputype = md->pmd_cputype;
3518                 gci.pm_ncpu    = pmc_cpu_max();
3519                 gci.pm_npmc    = md->pmd_npmc;
3520                 gci.pm_nclass  = md->pmd_nclass;
3521                 pci = gci.pm_classes;
3522                 pcd = md->pmd_classdep;
3523                 for (cl = 0; cl < md->pmd_nclass; cl++, pci++, pcd++) {
3524                         pci->pm_caps  = pcd->pcd_caps;
3525                         pci->pm_class = pcd->pcd_class;
3526                         pci->pm_width = pcd->pcd_width;
3527                         pci->pm_num   = pcd->pcd_num;
3528                 }
3529                 error = copyout(&gci, arg, sizeof(gci));
3530         }
3531         break;
3532
3533         /*
3534          * Retrieve soft events list.
3535          */
3536         case PMC_OP_GETDYNEVENTINFO:
3537         {
3538                 enum pmc_class                  cl;
3539                 enum pmc_event                  ev;
3540                 struct pmc_op_getdyneventinfo   *gei;
3541                 struct pmc_dyn_event_descr      dev;
3542                 struct pmc_soft                 *ps;
3543                 uint32_t                        nevent;
3544
3545                 sx_assert(&pmc_sx, SX_LOCKED);
3546
3547                 gei = (struct pmc_op_getdyneventinfo *) arg;
3548
3549                 if ((error = copyin(&gei->pm_class, &cl, sizeof(cl))) != 0)
3550                         break;
3551
3552                 /* Only SOFT class is dynamic. */
3553                 if (cl != PMC_CLASS_SOFT) {
3554                         error = EINVAL;
3555                         break;
3556                 }
3557
3558                 nevent = 0;
3559                 for (ev = PMC_EV_SOFT_FIRST; (int)ev <= PMC_EV_SOFT_LAST; ev++) {
3560                         ps = pmc_soft_ev_acquire(ev);
3561                         if (ps == NULL)
3562                                 continue;
3563                         bcopy(&ps->ps_ev, &dev, sizeof(dev));
3564                         pmc_soft_ev_release(ps);
3565
3566                         error = copyout(&dev,
3567                             &gei->pm_events[nevent],
3568                             sizeof(struct pmc_dyn_event_descr));
3569                         if (error != 0)
3570                                 break;
3571                         nevent++;
3572                 }
3573                 if (error != 0)
3574                         break;
3575
3576                 error = copyout(&nevent, &gei->pm_nevent,
3577                     sizeof(nevent));
3578         }
3579         break;
3580
3581         /*
3582          * Get module statistics
3583          */
3584
3585         case PMC_OP_GETDRIVERSTATS:
3586         {
3587                 struct pmc_op_getdriverstats gms;
3588 #define CFETCH(a, b, field) a.field = counter_u64_fetch(b.field)
3589                 CFETCH(gms, pmc_stats, pm_intr_ignored);
3590                 CFETCH(gms, pmc_stats, pm_intr_processed);
3591                 CFETCH(gms, pmc_stats, pm_intr_bufferfull);
3592                 CFETCH(gms, pmc_stats, pm_syscalls);
3593                 CFETCH(gms, pmc_stats, pm_syscall_errors);
3594                 CFETCH(gms, pmc_stats, pm_buffer_requests);
3595                 CFETCH(gms, pmc_stats, pm_buffer_requests_failed);
3596                 CFETCH(gms, pmc_stats, pm_log_sweeps);
3597 #undef CFETCH
3598                 error = copyout(&gms, arg, sizeof(gms));
3599         }
3600         break;
3601
3602
3603         /*
3604          * Retrieve module version number
3605          */
3606
3607         case PMC_OP_GETMODULEVERSION:
3608         {
3609                 uint32_t cv, modv;
3610
3611                 /* retrieve the client's idea of the ABI version */
3612                 if ((error = copyin(arg, &cv, sizeof(uint32_t))) != 0)
3613                         break;
3614                 /* don't service clients newer than our driver */
3615                 modv = PMC_VERSION;
3616                 if ((cv & 0xFFFF0000) > (modv & 0xFFFF0000)) {
3617                         error = EPROGMISMATCH;
3618                         break;
3619                 }
3620                 error = copyout(&modv, arg, sizeof(int));
3621         }
3622         break;
3623
3624
3625         /*
3626          * Retrieve the state of all the PMCs on a given
3627          * CPU.
3628          */
3629
3630         case PMC_OP_GETPMCINFO:
3631         {
3632                 int ari;
3633                 struct pmc *pm;
3634                 size_t pmcinfo_size;
3635                 uint32_t cpu, n, npmc;
3636                 struct pmc_owner *po;
3637                 struct pmc_binding pb;
3638                 struct pmc_classdep *pcd;
3639                 struct pmc_info *p, *pmcinfo;
3640                 struct pmc_op_getpmcinfo *gpi;
3641
3642                 PMC_DOWNGRADE_SX();
3643
3644                 gpi = (struct pmc_op_getpmcinfo *) arg;
3645
3646                 if ((error = copyin(&gpi->pm_cpu, &cpu, sizeof(cpu))) != 0)
3647                         break;
3648
3649                 if (cpu >= pmc_cpu_max()) {
3650                         error = EINVAL;
3651                         break;
3652                 }
3653
3654                 if (!pmc_cpu_is_active(cpu)) {
3655                         error = ENXIO;
3656                         break;
3657                 }
3658
3659                 /* switch to CPU 'cpu' */
3660                 pmc_save_cpu_binding(&pb);
3661                 pmc_select_cpu(cpu);
3662
3663                 npmc = md->pmd_npmc;
3664
3665                 pmcinfo_size = npmc * sizeof(struct pmc_info);
3666                 pmcinfo = malloc(pmcinfo_size, M_PMC, M_WAITOK | M_ZERO);
3667
3668                 p = pmcinfo;
3669
3670                 for (n = 0; n < md->pmd_npmc; n++, p++) {
3671
3672                         pcd = pmc_ri_to_classdep(md, n, &ari);
3673
3674                         KASSERT(pcd != NULL,
3675                             ("[pmc,%d] null pcd ri=%d", __LINE__, n));
3676
3677                         if ((error = pcd->pcd_describe(cpu, ari, p, &pm)) != 0)
3678                                 break;
3679
3680                         if (PMC_ROW_DISP_IS_STANDALONE(n))
3681                                 p->pm_rowdisp = PMC_DISP_STANDALONE;
3682                         else if (PMC_ROW_DISP_IS_THREAD(n))
3683                                 p->pm_rowdisp = PMC_DISP_THREAD;
3684                         else
3685                                 p->pm_rowdisp = PMC_DISP_FREE;
3686
3687                         p->pm_ownerpid = -1;
3688
3689                         if (pm == NULL) /* no PMC associated */
3690                                 continue;
3691
3692                         po = pm->pm_owner;
3693
3694                         KASSERT(po->po_owner != NULL,
3695                             ("[pmc,%d] pmc_owner had a null proc pointer",
3696                                 __LINE__));
3697
3698                         p->pm_ownerpid = po->po_owner->p_pid;
3699                         p->pm_mode     = PMC_TO_MODE(pm);
3700                         p->pm_event    = pm->pm_event;
3701                         p->pm_flags    = pm->pm_flags;
3702
3703                         if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)))
3704                                 p->pm_reloadcount =
3705                                     pm->pm_sc.pm_reloadcount;
3706                 }
3707
3708                 pmc_restore_cpu_binding(&pb);
3709
3710                 /* now copy out the PMC info collected */
3711                 if (error == 0)
3712                         error = copyout(pmcinfo, &gpi->pm_pmcs, pmcinfo_size);
3713
3714                 free(pmcinfo, M_PMC);
3715         }
3716         break;
3717
3718
3719         /*
3720          * Set the administrative state of a PMC.  I.e. whether
3721          * the PMC is to be used or not.
3722          */
3723
3724         case PMC_OP_PMCADMIN:
3725         {
3726                 int cpu, ri;
3727                 enum pmc_state request;
3728                 struct pmc_cpu *pc;
3729                 struct pmc_hw *phw;
3730                 struct pmc_op_pmcadmin pma;
3731                 struct pmc_binding pb;
3732
3733                 sx_assert(&pmc_sx, SX_XLOCKED);
3734
3735                 KASSERT(td == curthread,
3736                     ("[pmc,%d] td != curthread", __LINE__));
3737
3738                 error = priv_check(td, PRIV_PMC_MANAGE);
3739                 if (error)
3740                         break;
3741
3742                 if ((error = copyin(arg, &pma, sizeof(pma))) != 0)
3743                         break;
3744
3745                 cpu = pma.pm_cpu;
3746
3747                 if (cpu < 0 || cpu >= (int) pmc_cpu_max()) {
3748                         error = EINVAL;
3749                         break;
3750                 }
3751
3752                 if (!pmc_cpu_is_active(cpu)) {
3753                         error = ENXIO;
3754                         break;
3755                 }
3756
3757                 request = pma.pm_state;
3758
3759                 if (request != PMC_STATE_DISABLED &&
3760                     request != PMC_STATE_FREE) {
3761                         error = EINVAL;
3762                         break;
3763                 }
3764
3765                 ri = pma.pm_pmc; /* pmc id == row index */
3766                 if (ri < 0 || ri >= (int) md->pmd_npmc) {
3767                         error = EINVAL;
3768                         break;
3769                 }
3770
3771                 /*
3772                  * We can't disable a PMC with a row-index allocated
3773                  * for process virtual PMCs.
3774                  */
3775
3776                 if (PMC_ROW_DISP_IS_THREAD(ri) &&
3777                     request == PMC_STATE_DISABLED) {
3778                         error = EBUSY;
3779                         break;
3780                 }
3781
3782                 /*
3783                  * otherwise, this PMC on this CPU is either free or
3784                  * in system-wide mode.
3785                  */
3786
3787                 pmc_save_cpu_binding(&pb);
3788                 pmc_select_cpu(cpu);
3789
3790                 pc  = pmc_pcpu[cpu];
3791                 phw = pc->pc_hwpmcs[ri];
3792
3793                 /*
3794                  * XXX do we need some kind of 'forced' disable?
3795                  */
3796
3797                 if (phw->phw_pmc == NULL) {
3798                         if (request == PMC_STATE_DISABLED &&
3799                             (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED)) {
3800                                 phw->phw_state &= ~PMC_PHW_FLAG_IS_ENABLED;
3801                                 PMC_MARK_ROW_STANDALONE(ri);
3802                         } else if (request == PMC_STATE_FREE &&
3803                             (phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0) {
3804                                 phw->phw_state |=  PMC_PHW_FLAG_IS_ENABLED;
3805                                 PMC_UNMARK_ROW_STANDALONE(ri);
3806                         }
3807                         /* other cases are a no-op */
3808                 } else
3809                         error = EBUSY;
3810
3811                 pmc_restore_cpu_binding(&pb);
3812         }
3813         break;
3814
3815
3816         /*
3817          * Allocate a PMC.
3818          */
3819
3820         case PMC_OP_PMCALLOCATE:
3821         {
3822                 int adjri, n;
3823                 u_int cpu;
3824                 uint32_t caps;
3825                 struct pmc *pmc;
3826                 enum pmc_mode mode;
3827                 struct pmc_hw *phw;
3828                 struct pmc_binding pb;
3829                 struct pmc_classdep *pcd;
3830                 struct pmc_op_pmcallocate pa;
3831
3832                 if ((error = copyin(arg, &pa, sizeof(pa))) != 0)
3833                         break;
3834
3835                 caps = pa.pm_caps;
3836                 mode = pa.pm_mode;
3837                 cpu  = pa.pm_cpu;
3838
3839                 if ((mode != PMC_MODE_SS  &&  mode != PMC_MODE_SC  &&
3840                      mode != PMC_MODE_TS  &&  mode != PMC_MODE_TC) ||
3841                     (cpu != (u_int) PMC_CPU_ANY && cpu >= pmc_cpu_max())) {
3842                         error = EINVAL;
3843                         break;
3844                 }
3845
3846                 /*
3847                  * Virtual PMCs should only ask for a default CPU.
3848                  * System mode PMCs need to specify a non-default CPU.
3849                  */
3850
3851                 if ((PMC_IS_VIRTUAL_MODE(mode) && cpu != (u_int) PMC_CPU_ANY) ||
3852                     (PMC_IS_SYSTEM_MODE(mode) && cpu == (u_int) PMC_CPU_ANY)) {
3853                         error = EINVAL;
3854                         break;
3855                 }
3856
3857                 /*
3858                  * Check that an inactive CPU is not being asked for.
3859                  */
3860
3861                 if (PMC_IS_SYSTEM_MODE(mode) && !pmc_cpu_is_active(cpu)) {
3862                         error = ENXIO;
3863                         break;
3864                 }
3865
3866                 /*
3867                  * Refuse an allocation for a system-wide PMC if this
3868                  * process has been jailed, or if this process lacks
3869                  * super-user credentials and the sysctl tunable
3870                  * 'security.bsd.unprivileged_syspmcs' is zero.
3871                  */
3872
3873                 if (PMC_IS_SYSTEM_MODE(mode)) {
3874                         if (jailed(curthread->td_ucred)) {
3875                                 error = EPERM;
3876                                 break;
3877                         }
3878                         if (!pmc_unprivileged_syspmcs) {
3879                                 error = priv_check(curthread,
3880                                     PRIV_PMC_SYSTEM);
3881                                 if (error)
3882                                         break;
3883                         }
3884                 }
3885
3886                 /*
3887                  * Look for valid values for 'pm_flags'
3888                  */
3889
3890                 if ((pa.pm_flags & ~(PMC_F_DESCENDANTS | PMC_F_LOG_PROCCSW |
3891                     PMC_F_LOG_PROCEXIT | PMC_F_CALLCHAIN |
3892                     PMC_F_USERCALLCHAIN)) != 0) {
3893                         error = EINVAL;
3894                         break;
3895                 }
3896
3897                 /* PMC_F_USERCALLCHAIN is only valid with PMC_F_CALLCHAIN */
3898                 if ((pa.pm_flags & (PMC_F_CALLCHAIN | PMC_F_USERCALLCHAIN)) ==
3899                     PMC_F_USERCALLCHAIN) {
3900                         error = EINVAL;
3901                         break;
3902                 }
3903
3904                 /* PMC_F_USERCALLCHAIN is only valid for sampling mode */
3905                 if (pa.pm_flags & PMC_F_USERCALLCHAIN &&
3906                         mode != PMC_MODE_TS && mode != PMC_MODE_SS) {
3907                         error = EINVAL;
3908                         break;
3909                 }
3910
3911                 /* process logging options are not allowed for system PMCs */
3912                 if (PMC_IS_SYSTEM_MODE(mode) && (pa.pm_flags &
3913                     (PMC_F_LOG_PROCCSW | PMC_F_LOG_PROCEXIT))) {
3914                         error = EINVAL;
3915                         break;
3916                 }
3917
3918                 /*
3919                  * All sampling mode PMCs need to be able to interrupt the
3920                  * CPU.
3921                  */
3922                 if (PMC_IS_SAMPLING_MODE(mode))
3923                         caps |= PMC_CAP_INTERRUPT;
3924
3925                 /* A valid class specifier should have been passed in. */
3926                 pcd = pmc_class_to_classdep(pa.pm_class);
3927                 if (pcd == NULL) {
3928                         error = EINVAL;
3929                         break;
3930                 }
3931
3932                 /* The requested PMC capabilities should be feasible. */
3933                 if ((pcd->pcd_caps & caps) != caps) {
3934                         error = EOPNOTSUPP;
3935                         break;
3936                 }
3937
3938                 PMCDBG4(PMC,ALL,2, "event=%d caps=0x%x mode=%d cpu=%d",
3939                     pa.pm_ev, caps, mode, cpu);
3940
3941                 pmc = pmc_allocate_pmc_descriptor();
3942                 pmc->pm_id    = PMC_ID_MAKE_ID(cpu,pa.pm_mode,pa.pm_class,
3943                     PMC_ID_INVALID);
3944                 pmc->pm_event = pa.pm_ev;
3945                 pmc->pm_state = PMC_STATE_FREE;
3946                 pmc->pm_caps  = caps;
3947                 pmc->pm_flags = pa.pm_flags;
3948
3949                 /* XXX set lower bound on sampling for process counters */
3950                 if (PMC_IS_SAMPLING_MODE(mode)) {
3951                         /*
3952                          * Don't permit requested sample rate to be less than 1000
3953                          */
3954                         if (pa.pm_count < 1000)
3955                                 log(LOG_WARNING,
3956                                         "pmcallocate: passed sample rate %ju - setting to 1000\n",
3957                                         (uintmax_t)pa.pm_count);
3958                         pmc->pm_sc.pm_reloadcount = MAX(1000, pa.pm_count);
3959                 } else
3960                         pmc->pm_sc.pm_initial = pa.pm_count;
3961
3962                 /* switch thread to CPU 'cpu' */
3963                 pmc_save_cpu_binding(&pb);
3964
3965 #define PMC_IS_SHAREABLE_PMC(cpu, n)                            \
3966         (pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_state &           \
3967          PMC_PHW_FLAG_IS_SHAREABLE)
3968 #define PMC_IS_UNALLOCATED(cpu, n)                              \
3969         (pmc_pcpu[(cpu)]->pc_hwpmcs[(n)]->phw_pmc == NULL)
3970
3971                 if (PMC_IS_SYSTEM_MODE(mode)) {
3972                         pmc_select_cpu(cpu);
3973                         for (n = pcd->pcd_ri; n < (int) md->pmd_npmc; n++) {
3974                                 pcd = pmc_ri_to_classdep(md, n, &adjri);
3975                                 if (pmc_can_allocate_row(n, mode) == 0 &&
3976                                     pmc_can_allocate_rowindex(
3977                                             curthread->td_proc, n, cpu) == 0 &&
3978                                     (PMC_IS_UNALLOCATED(cpu, n) ||
3979                                      PMC_IS_SHAREABLE_PMC(cpu, n)) &&
3980                                     pcd->pcd_allocate_pmc(cpu, adjri, pmc,
3981                                         &pa) == 0)
3982                                         break;
3983                         }
3984                 } else {
3985                         /* Process virtual mode */
3986                         for (n = pcd->pcd_ri; n < (int) md->pmd_npmc; n++) {
3987                                 pcd = pmc_ri_to_classdep(md, n, &adjri);
3988                                 if (pmc_can_allocate_row(n, mode) == 0 &&
3989                                     pmc_can_allocate_rowindex(
3990                                             curthread->td_proc, n,
3991                                             PMC_CPU_ANY) == 0 &&
3992                                     pcd->pcd_allocate_pmc(curthread->td_oncpu,
3993                                         adjri, pmc, &pa) == 0)
3994                                         break;
3995                         }
3996                 }
3997
3998 #undef  PMC_IS_UNALLOCATED
3999 #undef  PMC_IS_SHAREABLE_PMC
4000
4001                 pmc_restore_cpu_binding(&pb);
4002
4003                 if (n == (int) md->pmd_npmc) {
4004                         pmc_destroy_pmc_descriptor(pmc);
4005                         pmc = NULL;
4006                         error = EINVAL;
4007                         break;
4008                 }
4009
4010                 /* Fill in the correct value in the ID field */
4011                 pmc->pm_id = PMC_ID_MAKE_ID(cpu,mode,pa.pm_class,n);
4012
4013                 PMCDBG5(PMC,ALL,2, "ev=%d class=%d mode=%d n=%d -> pmcid=%x",
4014                     pmc->pm_event, pa.pm_class, mode, n, pmc->pm_id);
4015
4016                 /* Process mode PMCs with logging enabled need log files */
4017                 if (pmc->pm_flags & (PMC_F_LOG_PROCEXIT | PMC_F_LOG_PROCCSW))
4018                         pmc->pm_flags |= PMC_F_NEEDS_LOGFILE;
4019
4020                 /* All system mode sampling PMCs require a log file */
4021                 if (PMC_IS_SAMPLING_MODE(mode) && PMC_IS_SYSTEM_MODE(mode))
4022                         pmc->pm_flags |= PMC_F_NEEDS_LOGFILE;
4023
4024                 /*
4025                  * Configure global pmc's immediately
4026                  */
4027
4028                 if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pmc))) {
4029
4030                         pmc_save_cpu_binding(&pb);
4031                         pmc_select_cpu(cpu);
4032
4033                         phw = pmc_pcpu[cpu]->pc_hwpmcs[n];
4034                         pcd = pmc_ri_to_classdep(md, n, &adjri);
4035
4036                         if ((phw->phw_state & PMC_PHW_FLAG_IS_ENABLED) == 0 ||
4037                             (error = pcd->pcd_config_pmc(cpu, adjri, pmc)) != 0) {
4038                                 (void) pcd->pcd_release_pmc(cpu, adjri, pmc);
4039                                 pmc_destroy_pmc_descriptor(pmc);
4040                                 pmc = NULL;
4041                                 pmc_restore_cpu_binding(&pb);
4042                                 error = EPERM;
4043                                 break;
4044                         }
4045
4046                         pmc_restore_cpu_binding(&pb);
4047                 }
4048
4049                 pmc->pm_state    = PMC_STATE_ALLOCATED;
4050                 pmc->pm_class   = pa.pm_class;
4051
4052                 /*
4053                  * mark row disposition
4054                  */
4055
4056                 if (PMC_IS_SYSTEM_MODE(mode))
4057                         PMC_MARK_ROW_STANDALONE(n);
4058                 else
4059                         PMC_MARK_ROW_THREAD(n);
4060
4061                 /*
4062                  * Register this PMC with the current thread as its owner.
4063                  */
4064
4065                 if ((error =
4066                     pmc_register_owner(curthread->td_proc, pmc)) != 0) {
4067                         pmc_release_pmc_descriptor(pmc);
4068                         pmc_destroy_pmc_descriptor(pmc);
4069                         pmc = NULL;
4070                         break;
4071                 }
4072
4073
4074                 /*
4075                  * Return the allocated index.
4076                  */
4077
4078                 pa.pm_pmcid = pmc->pm_id;
4079
4080                 error = copyout(&pa, arg, sizeof(pa));
4081         }
4082         break;
4083
4084
4085         /*
4086          * Attach a PMC to a process.
4087          */
4088
4089         case PMC_OP_PMCATTACH:
4090         {
4091                 struct pmc *pm;
4092                 struct proc *p;
4093                 struct pmc_op_pmcattach a;
4094
4095                 sx_assert(&pmc_sx, SX_XLOCKED);
4096
4097                 if ((error = copyin(arg, &a, sizeof(a))) != 0)
4098                         break;
4099
4100                 if (a.pm_pid < 0) {
4101                         error = EINVAL;
4102                         break;
4103                 } else if (a.pm_pid == 0)
4104                         a.pm_pid = td->td_proc->p_pid;
4105
4106                 if ((error = pmc_find_pmc(a.pm_pmc, &pm)) != 0)
4107                         break;
4108
4109                 if (PMC_IS_SYSTEM_MODE(PMC_TO_MODE(pm))) {
4110                         error = EINVAL;
4111                         break;
4112                 }
4113
4114                 /* PMCs may be (re)attached only when allocated or stopped */
4115                 if (pm->pm_state == PMC_STATE_RUNNING) {
4116                         error = EBUSY;
4117                         break;
4118                 } else if (pm->pm_state != PMC_STATE_ALLOCATED &&
4119                     pm->pm_state != PMC_STATE_STOPPED) {
4120                         error = EINVAL;
4121                         break;
4122                 }
4123
4124                 /* lookup pid */
4125                 if ((p = pfind(a.pm_pid)) == NULL) {
4126                         error = ESRCH;
4127                         break;
4128                 }
4129
4130                 /*
4131                  * Ignore processes that are working on exiting.
4132                  */
4133                 if (p->p_flag & P_WEXIT) {
4134                         error = ESRCH;
4135                         PROC_UNLOCK(p); /* pfind() returns a locked process */
4136                         break;
4137                 }
4138
4139                 /*
4140                  * we are allowed to attach a PMC to a process if
4141                  * we can debug it.
4142                  */
4143                 error = p_candebug(curthread, p);
4144
4145                 PROC_UNLOCK(p);
4146
4147                 if (error == 0)
4148                         error = pmc_attach_process(p, pm);
4149         }
4150         break;
4151
4152
4153         /*
4154          * Detach an attached PMC from a process.
4155          */
4156
4157         case PMC_OP_PMCDETACH:
4158         {
4159                 struct pmc *pm;
4160                 struct proc *p;
4161                 struct pmc_op_pmcattach a;
4162
4163                 if ((error = copyin(arg, &a, sizeof(a))) != 0)
4164                         break;
4165
4166                 if (a.pm_pid < 0) {
4167                         error = EINVAL;
4168                         break;
4169                 } else if (a.pm_pid == 0)
4170                         a.pm_pid = td->td_proc->p_pid;
4171
4172                 if ((error = pmc_find_pmc(a.pm_pmc, &pm)) != 0)
4173                         break;
4174
4175                 if ((p = pfind(a.pm_pid)) == NULL) {
4176                         error = ESRCH;
4177                         break;
4178                 }
4179
4180                 /*
4181                  * Treat processes that are in the process of exiting
4182                  * as if they were not present.
4183                  */
4184
4185                 if (p->p_flag & P_WEXIT)
4186                         error = ESRCH;
4187
4188                 PROC_UNLOCK(p); /* pfind() returns a locked process */
4189
4190                 if (error == 0)
4191                         error = pmc_detach_process(p, pm);
4192         }
4193         break;
4194
4195
4196         /*
4197          * Retrieve the MSR number associated with the counter
4198          * 'pmc_id'.  This allows processes to directly use RDPMC
4199          * instructions to read their PMCs, without the overhead of a
4200          * system call.
4201          */
4202
4203         case PMC_OP_PMCGETMSR:
4204         {
4205                 int adjri, ri;
4206                 struct pmc *pm;
4207                 struct pmc_target *pt;
4208                 struct pmc_op_getmsr gm;
4209                 struct pmc_classdep *pcd;
4210
4211                 PMC_DOWNGRADE_SX();
4212
4213                 if ((error = copyin(arg, &gm, sizeof(gm))) != 0)
4214                         break;
4215
4216                 if ((error = pmc_find_pmc(gm.pm_pmcid, &pm)) != 0)
4217                         break;
4218
4219                 /*
4220                  * The allocated PMC has to be a process virtual PMC,
4221                  * i.e., of type MODE_T[CS].  Global PMCs can only be
4222                  * read using the PMCREAD operation since they may be
4223                  * allocated on a different CPU than the one we could
4224                  * be running on at the time of the RDPMC instruction.
4225                  *
4226                  * The GETMSR operation is not allowed for PMCs that
4227                  * are inherited across processes.
4228                  */
4229
4230                 if (!PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)) ||
4231                     (pm->pm_flags & PMC_F_DESCENDANTS)) {
4232                         error = EINVAL;
4233                         break;
4234                 }
4235
4236                 /*
4237                  * It only makes sense to use a RDPMC (or its
4238                  * equivalent instruction on non-x86 architectures) on
4239                  * a process that has allocated and attached a PMC to
4240                  * itself.  Conversely the PMC is only allowed to have
4241                  * one process attached to it -- its owner.
4242                  */
4243
4244                 if ((pt = LIST_FIRST(&pm->pm_targets)) == NULL ||
4245                     LIST_NEXT(pt, pt_next) != NULL ||
4246                     pt->pt_process->pp_proc != pm->pm_owner->po_owner) {
4247                         error = EINVAL;
4248                         break;
4249                 }
4250
4251                 ri = PMC_TO_ROWINDEX(pm);
4252                 pcd = pmc_ri_to_classdep(md, ri, &adjri);
4253
4254                 /* PMC class has no 'GETMSR' support */
4255                 if (pcd->pcd_get_msr == NULL) {
4256                         error = ENOSYS;
4257                         break;
4258                 }
4259
4260                 if ((error = (*pcd->pcd_get_msr)(adjri, &gm.pm_msr)) < 0)
4261                         break;
4262
4263                 if ((error = copyout(&gm, arg, sizeof(gm))) < 0)
4264                         break;
4265
4266                 /*
4267                  * Mark our process as using MSRs.  Update machine
4268                  * state using a forced context switch.
4269                  */
4270
4271                 pt->pt_process->pp_flags |= PMC_PP_ENABLE_MSR_ACCESS;
4272                 pmc_force_context_switch();
4273
4274         }
4275         break;
4276
4277         /*
4278          * Release an allocated PMC
4279          */
4280
4281         case PMC_OP_PMCRELEASE:
4282         {
4283                 pmc_id_t pmcid;
4284                 struct pmc *pm;
4285                 struct pmc_owner *po;
4286                 struct pmc_op_simple sp;
4287
4288                 /*
4289                  * Find PMC pointer for the named PMC.
4290                  *
4291                  * Use pmc_release_pmc_descriptor() to switch off the
4292                  * PMC, remove all its target threads, and remove the
4293                  * PMC from its owner's list.
4294                  *
4295                  * Remove the owner record if this is the last PMC
4296                  * owned.
4297                  *
4298                  * Free up space.
4299                  */
4300
4301                 if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
4302                         break;
4303
4304                 pmcid = sp.pm_pmcid;
4305
4306                 if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
4307                         break;
4308
4309                 po = pm->pm_owner;
4310                 pmc_release_pmc_descriptor(pm);
4311                 pmc_maybe_remove_owner(po);
4312                 pmc_destroy_pmc_descriptor(pm);
4313         }
4314         break;
4315
4316
4317         /*
4318          * Read and/or write a PMC.
4319          */
4320
4321         case PMC_OP_PMCRW:
4322         {
4323                 int adjri;
4324                 struct pmc *pm;
4325                 uint32_t cpu, ri;
4326                 pmc_value_t oldvalue;
4327                 struct pmc_binding pb;
4328                 struct pmc_op_pmcrw prw;
4329                 struct pmc_classdep *pcd;
4330                 struct pmc_op_pmcrw *pprw;
4331
4332                 PMC_DOWNGRADE_SX();
4333
4334                 if ((error = copyin(arg, &prw, sizeof(prw))) != 0)
4335                         break;
4336
4337                 ri = 0;
4338                 PMCDBG2(PMC,OPS,1, "rw id=%d flags=0x%x", prw.pm_pmcid,
4339                     prw.pm_flags);
4340
4341                 /* must have at least one flag set */
4342                 if ((prw.pm_flags & (PMC_F_OLDVALUE|PMC_F_NEWVALUE)) == 0) {
4343                         error = EINVAL;
4344                         break;
4345                 }
4346
4347                 /* locate pmc descriptor */
4348                 if ((error = pmc_find_pmc(prw.pm_pmcid, &pm)) != 0)
4349                         break;
4350
4351                 /* Can't read a PMC that hasn't been started. */
4352                 if (pm->pm_state != PMC_STATE_ALLOCATED &&
4353                     pm->pm_state != PMC_STATE_STOPPED &&
4354                     pm->pm_state != PMC_STATE_RUNNING) {
4355                         error = EINVAL;
4356                         break;
4357                 }
4358
4359                 /* writing a new value is allowed only for 'STOPPED' pmcs */
4360                 if (pm->pm_state == PMC_STATE_RUNNING &&
4361                     (prw.pm_flags & PMC_F_NEWVALUE)) {
4362                         error = EBUSY;
4363                         break;
4364                 }
4365
4366                 if (PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm))) {
4367
4368                         /*
4369                          * If this PMC is attached to its owner (i.e.,
4370                          * the process requesting this operation) and
4371                          * is running, then attempt to get an
4372                          * upto-date reading from hardware for a READ.
4373                          * Writes are only allowed when the PMC is
4374                          * stopped, so only update the saved value
4375                          * field.
4376                          *
4377                          * If the PMC is not running, or is not
4378                          * attached to its owner, read/write to the
4379                          * savedvalue field.
4380                          */
4381
4382                         ri = PMC_TO_ROWINDEX(pm);
4383                         pcd = pmc_ri_to_classdep(md, ri, &adjri);
4384
4385                         mtx_pool_lock_spin(pmc_mtxpool, pm);
4386                         cpu = curthread->td_oncpu;
4387
4388                         if (prw.pm_flags & PMC_F_OLDVALUE) {
4389                                 if ((pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) &&
4390                                     (pm->pm_state == PMC_STATE_RUNNING))
4391                                         error = (*pcd->pcd_read_pmc)(cpu, adjri,
4392                                             &oldvalue);
4393                                 else
4394                                         oldvalue = pm->pm_gv.pm_savedvalue;
4395                         }
4396                         if (prw.pm_flags & PMC_F_NEWVALUE)
4397                                 pm->pm_gv.pm_savedvalue = prw.pm_value;
4398
4399                         mtx_pool_unlock_spin(pmc_mtxpool, pm);
4400
4401                 } else { /* System mode PMCs */
4402                         cpu = PMC_TO_CPU(pm);
4403                         ri  = PMC_TO_ROWINDEX(pm);
4404                         pcd = pmc_ri_to_classdep(md, ri, &adjri);
4405
4406                         if (!pmc_cpu_is_active(cpu)) {
4407                                 error = ENXIO;
4408                                 break;
4409                         }
4410
4411                         /* move this thread to CPU 'cpu' */
4412                         pmc_save_cpu_binding(&pb);
4413                         pmc_select_cpu(cpu);
4414
4415                         critical_enter();
4416                         /* save old value */
4417                         if (prw.pm_flags & PMC_F_OLDVALUE)
4418                                 if ((error = (*pcd->pcd_read_pmc)(cpu, adjri,
4419                                          &oldvalue)))
4420                                         goto error;
4421                         /* write out new value */
4422                         if (prw.pm_flags & PMC_F_NEWVALUE)
4423                                 error = (*pcd->pcd_write_pmc)(cpu, adjri,
4424                                     prw.pm_value);
4425                 error:
4426                         critical_exit();
4427                         pmc_restore_cpu_binding(&pb);
4428                         if (error)
4429                                 break;
4430                 }
4431
4432                 pprw = (struct pmc_op_pmcrw *) arg;
4433
4434 #ifdef  HWPMC_DEBUG
4435                 if (prw.pm_flags & PMC_F_NEWVALUE)
4436                         PMCDBG3(PMC,OPS,2, "rw id=%d new %jx -> old %jx",
4437                             ri, prw.pm_value, oldvalue);
4438                 else if (prw.pm_flags & PMC_F_OLDVALUE)
4439                         PMCDBG2(PMC,OPS,2, "rw id=%d -> old %jx", ri, oldvalue);
4440 #endif
4441
4442                 /* return old value if requested */
4443                 if (prw.pm_flags & PMC_F_OLDVALUE)
4444                         if ((error = copyout(&oldvalue, &pprw->pm_value,
4445                                  sizeof(prw.pm_value))))
4446                                 break;
4447
4448         }
4449         break;
4450
4451
4452         /*
4453          * Set the sampling rate for a sampling mode PMC and the
4454          * initial count for a counting mode PMC.
4455          */
4456
4457         case PMC_OP_PMCSETCOUNT:
4458         {
4459                 struct pmc *pm;
4460                 struct pmc_op_pmcsetcount sc;
4461
4462                 PMC_DOWNGRADE_SX();
4463
4464                 if ((error = copyin(arg, &sc, sizeof(sc))) != 0)
4465                         break;
4466
4467                 if ((error = pmc_find_pmc(sc.pm_pmcid, &pm)) != 0)
4468                         break;
4469
4470                 if (pm->pm_state == PMC_STATE_RUNNING) {
4471                         error = EBUSY;
4472                         break;
4473                 }
4474
4475                 if (PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm))) {
4476                         /*
4477                          * Don't permit requested sample rate to be less than 1000
4478                          */
4479                         if (sc.pm_count < 1000)
4480                                 log(LOG_WARNING,
4481                                         "pmcsetcount: passed sample rate %ju - setting to 1000\n",
4482                                         (uintmax_t)sc.pm_count);
4483                         pm->pm_sc.pm_reloadcount = MAX(1000, sc.pm_count);
4484                 } else
4485                         pm->pm_sc.pm_initial = sc.pm_count;
4486         }
4487         break;
4488
4489
4490         /*
4491          * Start a PMC.
4492          */
4493
4494         case PMC_OP_PMCSTART:
4495         {
4496                 pmc_id_t pmcid;
4497                 struct pmc *pm;
4498                 struct pmc_op_simple sp;
4499
4500                 sx_assert(&pmc_sx, SX_XLOCKED);
4501
4502                 if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
4503                         break;
4504
4505                 pmcid = sp.pm_pmcid;
4506
4507                 if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
4508                         break;
4509
4510                 KASSERT(pmcid == pm->pm_id,
4511                     ("[pmc,%d] pmcid %x != id %x", __LINE__,
4512                         pm->pm_id, pmcid));
4513
4514                 if (pm->pm_state == PMC_STATE_RUNNING) /* already running */
4515                         break;
4516                 else if (pm->pm_state != PMC_STATE_STOPPED &&
4517                     pm->pm_state != PMC_STATE_ALLOCATED) {
4518                         error = EINVAL;
4519                         break;
4520                 }
4521
4522                 error = pmc_start(pm);
4523         }
4524         break;
4525
4526
4527         /*
4528          * Stop a PMC.
4529          */
4530
4531         case PMC_OP_PMCSTOP:
4532         {
4533                 pmc_id_t pmcid;
4534                 struct pmc *pm;
4535                 struct pmc_op_simple sp;
4536
4537                 PMC_DOWNGRADE_SX();
4538
4539                 if ((error = copyin(arg, &sp, sizeof(sp))) != 0)
4540                         break;
4541
4542                 pmcid = sp.pm_pmcid;
4543
4544                 /*
4545                  * Mark the PMC as inactive and invoke the MD stop
4546                  * routines if needed.
4547                  */
4548
4549                 if ((error = pmc_find_pmc(pmcid, &pm)) != 0)
4550                         break;
4551
4552                 KASSERT(pmcid == pm->pm_id,
4553                     ("[pmc,%d] pmc id %x != pmcid %x", __LINE__,
4554                         pm->pm_id, pmcid));
4555
4556                 if (pm->pm_state == PMC_STATE_STOPPED) /* already stopped */
4557                         break;
4558                 else if (pm->pm_state != PMC_STATE_RUNNING) {
4559                         error = EINVAL;
4560                         break;
4561                 }
4562
4563                 error = pmc_stop(pm);
4564         }
4565         break;
4566
4567
4568         /*
4569          * Write a user supplied value to the log file.
4570          */
4571
4572         case PMC_OP_WRITELOG:
4573         {
4574                 struct pmc_op_writelog wl;
4575                 struct pmc_owner *po;
4576
4577                 PMC_DOWNGRADE_SX();
4578
4579                 if ((error = copyin(arg, &wl, sizeof(wl))) != 0)
4580                         break;
4581
4582                 if ((po = pmc_find_owner_descriptor(td->td_proc)) == NULL) {
4583                         error = EINVAL;
4584                         break;
4585                 }
4586
4587                 if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0) {
4588                         error = EINVAL;
4589                         break;
4590                 }
4591
4592                 error = pmclog_process_userlog(po, &wl);
4593         }
4594         break;
4595
4596
4597         default:
4598                 error = EINVAL;
4599                 break;
4600         }
4601
4602         if (is_sx_downgraded)
4603                 sx_sunlock(&pmc_sx);
4604         else
4605                 sx_xunlock(&pmc_sx);
4606 done_syscall:
4607         if (error)
4608                 counter_u64_add(pmc_stats.pm_syscall_errors, 1);
4609
4610         return (error);
4611 }
4612
4613 /*
4614  * Helper functions
4615  */
4616
4617
4618 /*
4619  * Mark the thread as needing callchain capture and post an AST.  The
4620  * actual callchain capture will be done in a context where it is safe
4621  * to take page faults.
4622  */
4623
4624 static void
4625 pmc_post_callchain_callback(void)
4626 {
4627         struct thread *td;
4628
4629         td = curthread;
4630
4631         /*
4632          * If there is multiple PMCs for the same interrupt ignore new post
4633          */
4634         if (td->td_pflags & TDP_CALLCHAIN)
4635                 return;
4636
4637         /*
4638          * Mark this thread as needing callchain capture.
4639          * `td->td_pflags' will be safe to touch because this thread
4640          * was in user space when it was interrupted.
4641          */
4642         td->td_pflags |= TDP_CALLCHAIN;
4643
4644         /*
4645          * Don't let this thread migrate between CPUs until callchain
4646          * capture completes.
4647          */
4648         sched_pin();
4649
4650         return;
4651 }
4652
4653 /*
4654  * Find a free slot in the per-cpu array of samples and capture the
4655  * current callchain there.  If a sample was successfully added, a bit
4656  * is set in mask 'pmc_cpumask' denoting that the DO_SAMPLES hook
4657  * needs to be invoked from the clock handler.
4658  *
4659  * This function is meant to be called from an NMI handler.  It cannot
4660  * use any of the locking primitives supplied by the OS.
4661  */
4662
4663 static int
4664 pmc_add_sample(ring_type_t ring, struct pmc *pm, struct trapframe *tf)
4665 {
4666         int error, cpu, callchaindepth, inuserspace;
4667         struct thread *td;
4668         struct pmc_sample *ps;
4669         struct pmc_samplebuffer *psb;
4670
4671         error = 0;
4672
4673         /*
4674          * Allocate space for a sample buffer.
4675          */
4676         cpu = curcpu;
4677         psb = pmc_pcpu[cpu]->pc_sb[ring];
4678         inuserspace = TRAPF_USERMODE(tf);
4679         ps = PMC_PROD_SAMPLE(psb);
4680         if (psb->ps_considx != psb->ps_prodidx &&
4681                 ps->ps_nsamples) {      /* in use, reader hasn't caught up */
4682                 pm->pm_pcpu_state[cpu].pps_stalled = 1;
4683                 counter_u64_add(pmc_stats.pm_intr_bufferfull, 1);
4684                 PMCDBG6(SAM,INT,1,"(spc) cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d",
4685                     cpu, pm, (void *) tf, inuserspace,
4686                     (int) (psb->ps_prodidx & pmc_sample_mask),
4687                     (int) (psb->ps_considx & pmc_sample_mask));
4688                 callchaindepth = 1;
4689                 error = ENOMEM;
4690                 goto done;
4691         }
4692
4693         /* Fill in entry. */
4694         PMCDBG6(SAM,INT,1,"cpu=%d pm=%p tf=%p um=%d wr=%d rd=%d", cpu, pm,
4695             (void *) tf, inuserspace,
4696             (int) (psb->ps_prodidx & pmc_sample_mask),
4697             (int) (psb->ps_considx & pmc_sample_mask));
4698
4699         td = curthread;
4700         ps->ps_pmc = pm;
4701         ps->ps_td = td;
4702         ps->ps_pid = td->td_proc->p_pid;
4703         ps->ps_tid = td->td_tid;
4704         ps->ps_tsc = pmc_rdtsc();
4705         ps->ps_ticks = ticks;
4706         ps->ps_cpu = cpu;
4707         ps->ps_flags = inuserspace ? PMC_CC_F_USERSPACE : 0;
4708
4709         callchaindepth = (pm->pm_flags & PMC_F_CALLCHAIN) ?
4710             pmc_callchaindepth : 1;
4711
4712         MPASS(ps->ps_pc != NULL);
4713         if (callchaindepth == 1)
4714                 ps->ps_pc[0] = PMC_TRAPFRAME_TO_PC(tf);
4715         else {
4716                 /*
4717                  * Kernel stack traversals can be done immediately,
4718                  * while we defer to an AST for user space traversals.
4719                  */
4720                 if (!inuserspace) {
4721                         callchaindepth =
4722                             pmc_save_kernel_callchain(ps->ps_pc,
4723                                 callchaindepth, tf);
4724                 } else {
4725                         pmc_post_callchain_callback();
4726                         callchaindepth = PMC_USER_CALLCHAIN_PENDING;
4727                 }
4728         }
4729
4730         ps->ps_nsamples = callchaindepth;       /* mark entry as in use */
4731         if (ring == PMC_UR) {
4732                 ps->ps_nsamples_actual = callchaindepth;        /* mark entry as in use */
4733                 ps->ps_nsamples = PMC_USER_CALLCHAIN_PENDING;
4734         } else
4735                 ps->ps_nsamples = callchaindepth;       /* mark entry as in use */
4736
4737         KASSERT(counter_u64_fetch(pm->pm_runcount) >= 0,
4738             ("[pmc,%d] pm=%p runcount %ld", __LINE__, (void *) pm,
4739                  (unsigned long)counter_u64_fetch(pm->pm_runcount)));
4740
4741         counter_u64_add(pm->pm_runcount, 1);    /* hold onto PMC */
4742         /* increment write pointer */
4743         psb->ps_prodidx++;
4744  done:
4745         /* mark CPU as needing processing */
4746         if (callchaindepth != PMC_USER_CALLCHAIN_PENDING)
4747                 DPCPU_SET(pmc_sampled, 1);
4748
4749         return (error);
4750 }
4751
4752 /*
4753  * Interrupt processing.
4754  *
4755  * This function is meant to be called from an NMI handler.  It cannot
4756  * use any of the locking primitives supplied by the OS.
4757  */
4758
4759 int
4760 pmc_process_interrupt(int ring, struct pmc *pm, struct trapframe *tf)
4761 {
4762         struct thread *td;
4763
4764         td = curthread;
4765         if ((pm->pm_flags & PMC_F_USERCALLCHAIN) &&
4766             (td->td_proc->p_flag & P_KPROC) == 0 &&
4767             !TRAPF_USERMODE(tf)) {
4768                 atomic_add_int(&td->td_pmcpend, 1);
4769                 return (pmc_add_sample(PMC_UR, pm, tf));
4770         }
4771         return (pmc_add_sample(ring, pm, tf));
4772 }
4773
4774 /*
4775  * Capture a user call chain.  This function will be called from ast()
4776  * before control returns to userland and before the process gets
4777  * rescheduled.
4778  */
4779
4780 static void
4781 pmc_capture_user_callchain(int cpu, int ring, struct trapframe *tf)
4782 {
4783         struct pmc *pm;
4784         struct thread *td;
4785         struct pmc_sample *ps;
4786         struct pmc_samplebuffer *psb;
4787         uint64_t considx, prodidx;
4788         int nsamples, nrecords, pass, iter;
4789 #ifdef  INVARIANTS
4790         int ncallchains;
4791         int nfree;
4792         int start_ticks = ticks;
4793 #endif
4794         psb = pmc_pcpu[cpu]->pc_sb[ring];
4795         td = curthread;
4796
4797         KASSERT(td->td_pflags & TDP_CALLCHAIN,
4798             ("[pmc,%d] Retrieving callchain for thread that doesn't want it",
4799                 __LINE__));
4800
4801 #ifdef  INVARIANTS
4802         ncallchains = 0;
4803         nfree = 0;
4804 #endif
4805         nrecords = INT_MAX;
4806         pass = 0;
4807  restart:
4808         if (ring == PMC_UR)
4809                 nrecords = atomic_readandclear_32(&td->td_pmcpend);
4810
4811         for (iter = 0, considx = psb->ps_considx, prodidx = psb->ps_prodidx;
4812             considx < prodidx && iter < pmc_nsamples; considx++, iter++) {
4813                 ps = PMC_CONS_SAMPLE_OFF(psb, considx);
4814
4815         /*
4816          * Iterate through all deferred callchain requests.
4817          * Walk from the current read pointer to the current
4818          * write pointer.
4819          */
4820
4821 #ifdef  INVARIANTS
4822                 if (ps->ps_nsamples == PMC_SAMPLE_FREE) {
4823                         nfree++;
4824                         continue;
4825                 }
4826
4827                 if ((ps->ps_pmc == NULL) ||
4828                     (ps->ps_pmc->pm_state != PMC_STATE_RUNNING))
4829                         nfree++;
4830 #endif
4831                 if (ps->ps_td != td ||
4832                    ps->ps_nsamples != PMC_USER_CALLCHAIN_PENDING ||
4833                    ps->ps_pmc->pm_state != PMC_STATE_RUNNING)
4834                         continue;
4835
4836                 KASSERT(ps->ps_cpu == cpu,
4837                     ("[pmc,%d] cpu mismatch ps_cpu=%d pcpu=%d", __LINE__,
4838                         ps->ps_cpu, PCPU_GET(cpuid)));
4839
4840                 pm = ps->ps_pmc;
4841
4842                 KASSERT(pm->pm_flags & PMC_F_CALLCHAIN,
4843                     ("[pmc,%d] Retrieving callchain for PMC that doesn't "
4844                         "want it", __LINE__));
4845
4846                 KASSERT(counter_u64_fetch(pm->pm_runcount) > 0,
4847                     ("[pmc,%d] runcount %ld", __LINE__, (unsigned long)counter_u64_fetch(pm->pm_runcount)));
4848
4849                 if (ring == PMC_UR) {
4850                         nsamples = ps->ps_nsamples_actual;
4851                         counter_u64_add(pmc_stats.pm_merges, 1);
4852                 } else
4853                         nsamples = 0;
4854
4855                 /*
4856                  * Retrieve the callchain and mark the sample buffer
4857                  * as 'processable' by the timer tick sweep code.
4858                  */
4859
4860 #ifdef INVARIANTS
4861                 ncallchains++;
4862 #endif
4863
4864                 if (__predict_true(nsamples < pmc_callchaindepth - 1))
4865                         nsamples += pmc_save_user_callchain(ps->ps_pc + nsamples,
4866                        pmc_callchaindepth - nsamples - 1, tf);
4867
4868                 /*
4869                  * We have to prevent hardclock from potentially overwriting
4870                  * this sample between when we read the value and when we set
4871                  * it
4872                  */
4873                 spinlock_enter();
4874                 /*
4875                  * Verify that the sample hasn't been dropped in the meantime
4876                  */
4877                 if (ps->ps_nsamples == PMC_USER_CALLCHAIN_PENDING) {
4878                         ps->ps_nsamples = nsamples;
4879                         /*
4880                          * If we couldn't get a sample, simply drop the reference
4881                          */
4882                         if (nsamples == 0)
4883                                 counter_u64_add(pm->pm_runcount, -1);
4884                 }
4885                 spinlock_exit();
4886                 if (nrecords-- == 1)
4887                         break;
4888         }
4889         if (__predict_false(ring == PMC_UR && td->td_pmcpend)) {
4890                 if (pass == 0) {
4891                         pass = 1;
4892                         goto restart;
4893                 }
4894                 /* only collect samples for this part once */
4895                 td->td_pmcpend = 0;
4896         }
4897
4898 #ifdef INVARIANTS
4899         if ((ticks - start_ticks) > hz)
4900                 log(LOG_ERR, "%s took %d ticks\n", __func__, (ticks - start_ticks));
4901 #endif
4902
4903         /* mark CPU as needing processing */
4904         DPCPU_SET(pmc_sampled, 1);
4905 }
4906
4907 /*
4908  * Process saved PC samples.
4909  */
4910
4911 static void
4912 pmc_process_samples(int cpu, ring_type_t ring)
4913 {
4914         struct pmc *pm;
4915         int adjri, n;
4916         struct thread *td;
4917         struct pmc_owner *po;
4918         struct pmc_sample *ps;
4919         struct pmc_classdep *pcd;
4920         struct pmc_samplebuffer *psb;
4921         uint64_t delta;
4922
4923         KASSERT(PCPU_GET(cpuid) == cpu,
4924             ("[pmc,%d] not on the correct CPU pcpu=%d cpu=%d", __LINE__,
4925                 PCPU_GET(cpuid), cpu));
4926
4927         psb = pmc_pcpu[cpu]->pc_sb[ring];
4928         delta = psb->ps_prodidx - psb->ps_considx;
4929         MPASS(delta <= pmc_nsamples);
4930         MPASS(psb->ps_considx <= psb->ps_prodidx);
4931         for (n = 0; psb->ps_considx < psb->ps_prodidx; psb->ps_considx++, n++) {
4932                 ps = PMC_CONS_SAMPLE(psb);
4933
4934                 if (__predict_false(ps->ps_nsamples == PMC_SAMPLE_FREE))
4935                         continue;
4936                 pm = ps->ps_pmc;
4937                 /* skip non-running samples */
4938                 if (pm->pm_state != PMC_STATE_RUNNING)
4939                         goto entrydone;
4940
4941                 KASSERT(counter_u64_fetch(pm->pm_runcount) > 0,
4942                     ("[pmc,%d] pm=%p runcount %ld", __LINE__, (void *) pm,
4943                          (unsigned long)counter_u64_fetch(pm->pm_runcount)));
4944
4945                 po = pm->pm_owner;
4946
4947                 KASSERT(PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)),
4948                     ("[pmc,%d] pmc=%p non-sampling mode=%d", __LINE__,
4949                         pm, PMC_TO_MODE(pm)));
4950
4951
4952                 /* If there is a pending AST wait for completion */
4953                 if (ps->ps_nsamples == PMC_USER_CALLCHAIN_PENDING) {
4954                         /* if we've been waiting more than 1 tick to
4955                          * collect a callchain for this record then
4956                          * drop it and move on.
4957                          */
4958                         if (ticks - ps->ps_ticks > 1) {
4959                                 /*
4960                                  * track how often we hit this as it will
4961                                  * preferentially lose user samples
4962                                  * for long running system calls
4963                                  */
4964                                 counter_u64_add(pmc_stats.pm_overwrites, 1);
4965                                 goto entrydone;
4966                         }
4967                         /* Need a rescan at a later time. */
4968                         DPCPU_SET(pmc_sampled, 1);
4969                         break;
4970                 }
4971
4972                 PMCDBG6(SAM,OPS,1,"cpu=%d pm=%p n=%d fl=%x wr=%d rd=%d", cpu,
4973                     pm, ps->ps_nsamples, ps->ps_flags,
4974                     (int) (psb->ps_prodidx & pmc_sample_mask),
4975                     (int) (psb->ps_considx & pmc_sample_mask));
4976
4977                 /*
4978                  * If this is a process-mode PMC that is attached to
4979                  * its owner, and if the PC is in user mode, update
4980                  * profiling statistics like timer-based profiling
4981                  * would have done.
4982                  *
4983                  * Otherwise, this is either a sampling-mode PMC that
4984                  * is attached to a different process than its owner,
4985                  * or a system-wide sampling PMC. Dispatch a log
4986                  * entry to the PMC's owner process.
4987                  */
4988                 if (pm->pm_flags & PMC_F_ATTACHED_TO_OWNER) {
4989                         if (ps->ps_flags & PMC_CC_F_USERSPACE) {
4990                                 td = FIRST_THREAD_IN_PROC(po->po_owner);
4991                                 addupc_intr(td, ps->ps_pc[0], 1);
4992                         }
4993                 } else
4994                         pmclog_process_callchain(pm, ps);
4995
4996         entrydone:
4997                 ps->ps_nsamples = 0; /* mark entry as free */
4998                 KASSERT(counter_u64_fetch(pm->pm_runcount) > 0,
4999                                 ("[pmc,%d] pm=%p runcount %ld", __LINE__, (void *) pm,
5000                                  (unsigned long)counter_u64_fetch(pm->pm_runcount)));
5001
5002                 counter_u64_add(pm->pm_runcount, -1);
5003         }
5004
5005         counter_u64_add(pmc_stats.pm_log_sweeps, 1);
5006
5007         /* Do not re-enable stalled PMCs if we failed to process any samples */
5008         if (n == 0)
5009                 return;
5010
5011         /*
5012          * Restart any stalled sampling PMCs on this CPU.
5013          *
5014          * If the NMI handler sets the pm_stalled field of a PMC after
5015          * the check below, we'll end up processing the stalled PMC at
5016          * the next hardclock tick.
5017          */
5018         for (n = 0; n < md->pmd_npmc; n++) {
5019                 pcd = pmc_ri_to_classdep(md, n, &adjri);
5020                 KASSERT(pcd != NULL,
5021                     ("[pmc,%d] null pcd ri=%d", __LINE__, n));
5022                 (void) (*pcd->pcd_get_config)(cpu,adjri,&pm);
5023
5024                 if (pm == NULL ||                        /* !cfg'ed */
5025                     pm->pm_state != PMC_STATE_RUNNING || /* !active */
5026                     !PMC_IS_SAMPLING_MODE(PMC_TO_MODE(pm)) || /* !sampling */
5027                         !pm->pm_pcpu_state[cpu].pps_cpustate  || /* !desired */
5028                     !pm->pm_pcpu_state[cpu].pps_stalled) /* !stalled */
5029                         continue;
5030
5031                 pm->pm_pcpu_state[cpu].pps_stalled = 0;
5032                 (*pcd->pcd_start_pmc)(cpu, adjri);
5033         }
5034 }
5035
5036 /*
5037  * Event handlers.
5038  */
5039
5040 /*
5041  * Handle a process exit.
5042  *
5043  * Remove this process from all hash tables.  If this process
5044  * owned any PMCs, turn off those PMCs and deallocate them,
5045  * removing any associations with target processes.
5046  *
5047  * This function will be called by the last 'thread' of a
5048  * process.
5049  *
5050  * XXX This eventhandler gets called early in the exit process.
5051  * Consider using a 'hook' invocation from thread_exit() or equivalent
5052  * spot.  Another negative is that kse_exit doesn't seem to call
5053  * exit1() [??].
5054  *
5055  */
5056
5057 static void
5058 pmc_process_exit(void *arg __unused, struct proc *p)
5059 {
5060         struct pmc *pm;
5061         int adjri, cpu;
5062         unsigned int ri;
5063         int is_using_hwpmcs;
5064         struct pmc_owner *po;
5065         struct pmc_process *pp;
5066         struct pmc_classdep *pcd;
5067         pmc_value_t newvalue, tmp;
5068
5069         PROC_LOCK(p);
5070         is_using_hwpmcs = p->p_flag & P_HWPMC;
5071         PROC_UNLOCK(p);
5072
5073         /*
5074          * Log a sysexit event to all SS PMC owners.
5075          */
5076         PMC_EPOCH_ENTER();
5077         CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
5078             if (po->po_flags & PMC_PO_OWNS_LOGFILE)
5079                     pmclog_process_sysexit(po, p->p_pid);
5080         PMC_EPOCH_EXIT();
5081
5082         if (!is_using_hwpmcs)
5083                 return;
5084
5085         PMC_GET_SX_XLOCK();
5086         PMCDBG3(PRC,EXT,1,"process-exit proc=%p (%d, %s)", p, p->p_pid,
5087             p->p_comm);
5088
5089         /*
5090          * Since this code is invoked by the last thread in an exiting
5091          * process, we would have context switched IN at some prior
5092          * point.  However, with PREEMPTION, kernel mode context
5093          * switches may happen any time, so we want to disable a
5094          * context switch OUT till we get any PMCs targeting this
5095          * process off the hardware.
5096          *
5097          * We also need to atomically remove this process'
5098          * entry from our target process hash table, using
5099          * PMC_FLAG_REMOVE.
5100          */
5101         PMCDBG3(PRC,EXT,1, "process-exit proc=%p (%d, %s)", p, p->p_pid,
5102             p->p_comm);
5103
5104         critical_enter(); /* no preemption */
5105
5106         cpu = curthread->td_oncpu;
5107
5108         if ((pp = pmc_find_process_descriptor(p,
5109                  PMC_FLAG_REMOVE)) != NULL) {
5110
5111                 PMCDBG2(PRC,EXT,2,
5112                     "process-exit proc=%p pmc-process=%p", p, pp);
5113
5114                 /*
5115                  * The exiting process could the target of
5116                  * some PMCs which will be running on
5117                  * currently executing CPU.
5118                  *
5119                  * We need to turn these PMCs off like we
5120                  * would do at context switch OUT time.
5121                  */
5122                 for (ri = 0; ri < md->pmd_npmc; ri++) {
5123
5124                         /*
5125                          * Pick up the pmc pointer from hardware
5126                          * state similar to the CSW_OUT code.
5127                          */
5128                         pm = NULL;
5129
5130                         pcd = pmc_ri_to_classdep(md, ri, &adjri);
5131
5132                         (void) (*pcd->pcd_get_config)(cpu, adjri, &pm);
5133
5134                         PMCDBG2(PRC,EXT,2, "ri=%d pm=%p", ri, pm);
5135
5136                         if (pm == NULL ||
5137                             !PMC_IS_VIRTUAL_MODE(PMC_TO_MODE(pm)))
5138                                 continue;
5139
5140                         PMCDBG4(PRC,EXT,2, "ppmcs[%d]=%p pm=%p "
5141                             "state=%d", ri, pp->pp_pmcs[ri].pp_pmc,
5142                             pm, pm->pm_state);
5143
5144                         KASSERT(PMC_TO_ROWINDEX(pm) == ri,
5145                             ("[pmc,%d] ri mismatch pmc(%d) ri(%d)",
5146                                 __LINE__, PMC_TO_ROWINDEX(pm), ri));
5147
5148                         KASSERT(pm == pp->pp_pmcs[ri].pp_pmc,
5149                             ("[pmc,%d] pm %p != pp_pmcs[%d] %p",
5150                                 __LINE__, pm, ri, pp->pp_pmcs[ri].pp_pmc));
5151
5152                         KASSERT(counter_u64_fetch(pm->pm_runcount) > 0,
5153                             ("[pmc,%d] bad runcount ri %d rc %ld",
5154                                  __LINE__, ri, (unsigned long)counter_u64_fetch(pm->pm_runcount)));
5155
5156                         /*
5157                          * Change desired state, and then stop if not
5158                          * stalled. This two-step dance should avoid
5159                          * race conditions where an interrupt re-enables
5160                          * the PMC after this code has already checked
5161                          * the pm_stalled flag.
5162                          */
5163                         if (pm->pm_pcpu_state[cpu].pps_cpustate) {
5164                                 pm->pm_pcpu_state[cpu].pps_cpustate = 0;
5165                                 if (!pm->pm_pcpu_state[cpu].pps_stalled) {
5166                                         (void) pcd->pcd_stop_pmc(cpu, adjri);
5167
5168                                         if (PMC_TO_MODE(pm) == PMC_MODE_TC) {
5169                                                 pcd->pcd_read_pmc(cpu, adjri,
5170                                                     &newvalue);
5171                                                 tmp = newvalue -
5172                                                     PMC_PCPU_SAVED(cpu,ri);
5173
5174                                                 mtx_pool_lock_spin(pmc_mtxpool,
5175                                                     pm);
5176                                                 pm->pm_gv.pm_savedvalue += tmp;
5177                                                 pp->pp_pmcs[ri].pp_pmcval +=
5178                                                     tmp;
5179                                                 mtx_pool_unlock_spin(
5180                                                     pmc_mtxpool, pm);
5181                                         }
5182                                 }
5183                         }
5184
5185                         KASSERT((int64_t) counter_u64_fetch(pm->pm_runcount) > 0,
5186                             ("[pmc,%d] runcount is %d", __LINE__, ri));
5187
5188                         counter_u64_add(pm->pm_runcount, -1);
5189
5190                         (void) pcd->pcd_config_pmc(cpu, adjri, NULL);
5191                 }
5192
5193                 /*
5194                  * Inform the MD layer of this pseudo "context switch
5195                  * out"
5196                  */
5197                 (void) md->pmd_switch_out(pmc_pcpu[cpu], pp);
5198
5199                 critical_exit(); /* ok to be pre-empted now */
5200
5201                 /*
5202                  * Unlink this process from the PMCs that are
5203                  * targeting it.  This will send a signal to
5204                  * all PMC owner's whose PMCs are orphaned.
5205                  *
5206                  * Log PMC value at exit time if requested.
5207                  */
5208                 for (ri = 0; ri < md->pmd_npmc; ri++)
5209                         if ((pm = pp->pp_pmcs[ri].pp_pmc) != NULL) {
5210                                 if (pm->pm_flags & PMC_F_NEEDS_LOGFILE &&
5211                                     PMC_IS_COUNTING_MODE(PMC_TO_MODE(pm)))
5212                                         pmclog_process_procexit(pm, pp);
5213                                 pmc_unlink_target_process(pm, pp);
5214                         }
5215                 free(pp, M_PMC);
5216
5217         } else
5218                 critical_exit(); /* pp == NULL */
5219
5220
5221         /*
5222          * If the process owned PMCs, free them up and free up
5223          * memory.
5224          */
5225         if ((po = pmc_find_owner_descriptor(p)) != NULL) {
5226                 pmc_remove_owner(po);
5227                 pmc_destroy_owner_descriptor(po);
5228         }
5229
5230         sx_xunlock(&pmc_sx);
5231 }
5232
5233 /*
5234  * Handle a process fork.
5235  *
5236  * If the parent process 'p1' is under HWPMC monitoring, then copy
5237  * over any attached PMCs that have 'do_descendants' semantics.
5238  */
5239
5240 static void
5241 pmc_process_fork(void *arg __unused, struct proc *p1, struct proc *newproc,
5242     int flags)
5243 {
5244         int is_using_hwpmcs;
5245         unsigned int ri;
5246         uint32_t do_descendants;
5247         struct pmc *pm;
5248         struct pmc_owner *po;
5249         struct pmc_process *ppnew, *ppold;
5250
5251         (void) flags;           /* unused parameter */
5252
5253         PROC_LOCK(p1);
5254         is_using_hwpmcs = p1->p_flag & P_HWPMC;
5255         PROC_UNLOCK(p1);
5256
5257         /*
5258          * If there are system-wide sampling PMCs active, we need to
5259          * log all fork events to their owner's logs.
5260          */
5261         PMC_EPOCH_ENTER();
5262         CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
5263             if (po->po_flags & PMC_PO_OWNS_LOGFILE) {
5264                     pmclog_process_procfork(po, p1->p_pid, newproc->p_pid);
5265                         pmclog_process_proccreate(po, newproc, 1);
5266                 }
5267         PMC_EPOCH_EXIT();
5268
5269         if (!is_using_hwpmcs)
5270                 return;
5271
5272         PMC_GET_SX_XLOCK();
5273         PMCDBG4(PMC,FRK,1, "process-fork proc=%p (%d, %s) -> %p", p1,
5274             p1->p_pid, p1->p_comm, newproc);
5275
5276         /*
5277          * If the parent process (curthread->td_proc) is a
5278          * target of any PMCs, look for PMCs that are to be
5279          * inherited, and link these into the new process
5280          * descriptor.
5281          */
5282         if ((ppold = pmc_find_process_descriptor(curthread->td_proc,
5283                  PMC_FLAG_NONE)) == NULL)
5284                 goto done;              /* nothing to do */
5285
5286         do_descendants = 0;
5287         for (ri = 0; ri < md->pmd_npmc; ri++)
5288                 if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL)
5289                         do_descendants |= pm->pm_flags & PMC_F_DESCENDANTS;
5290         if (do_descendants == 0) /* nothing to do */
5291                 goto done;
5292
5293         /*
5294          * Now mark the new process as being tracked by this driver.
5295          */
5296         PROC_LOCK(newproc);
5297         newproc->p_flag |= P_HWPMC;
5298         PROC_UNLOCK(newproc);
5299
5300         /* allocate a descriptor for the new process  */
5301         if ((ppnew = pmc_find_process_descriptor(newproc,
5302                  PMC_FLAG_ALLOCATE)) == NULL)
5303                 goto done;
5304
5305         /*
5306          * Run through all PMCs that were targeting the old process
5307          * and which specified F_DESCENDANTS and attach them to the
5308          * new process.
5309          *
5310          * Log the fork event to all owners of PMCs attached to this
5311          * process, if not already logged.
5312          */
5313         for (ri = 0; ri < md->pmd_npmc; ri++)
5314                 if ((pm = ppold->pp_pmcs[ri].pp_pmc) != NULL &&
5315                     (pm->pm_flags & PMC_F_DESCENDANTS)) {
5316                         pmc_link_target_process(pm, ppnew);
5317                         po = pm->pm_owner;
5318                         if (po->po_sscount == 0 &&
5319                             po->po_flags & PMC_PO_OWNS_LOGFILE)
5320                                 pmclog_process_procfork(po, p1->p_pid,
5321                                     newproc->p_pid);
5322                 }
5323
5324  done:
5325         sx_xunlock(&pmc_sx);
5326 }
5327
5328 static void
5329 pmc_process_threadcreate(struct thread *td)
5330 {
5331         struct pmc_owner *po;
5332
5333         PMC_EPOCH_ENTER();
5334         CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
5335             if (po->po_flags & PMC_PO_OWNS_LOGFILE)
5336                         pmclog_process_threadcreate(po, td, 1);
5337         PMC_EPOCH_EXIT();
5338 }
5339
5340 static void
5341 pmc_process_threadexit(struct thread *td)
5342 {
5343         struct pmc_owner *po;
5344
5345         PMC_EPOCH_ENTER();
5346         CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
5347             if (po->po_flags & PMC_PO_OWNS_LOGFILE)
5348                         pmclog_process_threadexit(po, td);
5349         PMC_EPOCH_EXIT();
5350 }
5351
5352 static void
5353 pmc_process_proccreate(struct proc *p)
5354 {
5355         struct pmc_owner *po;
5356
5357         PMC_EPOCH_ENTER();
5358         CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
5359             if (po->po_flags & PMC_PO_OWNS_LOGFILE)
5360                         pmclog_process_proccreate(po, p, 1 /* sync */);
5361         PMC_EPOCH_EXIT();
5362 }
5363
5364 static void
5365 pmc_process_allproc(struct pmc *pm)
5366 {
5367         struct pmc_owner *po;
5368         struct thread *td;
5369         struct proc *p;
5370
5371         po = pm->pm_owner;
5372         if ((po->po_flags & PMC_PO_OWNS_LOGFILE) == 0)
5373                 return;
5374         sx_slock(&allproc_lock);
5375         FOREACH_PROC_IN_SYSTEM(p) {
5376                 pmclog_process_proccreate(po, p, 0 /* sync */);
5377                 PROC_LOCK(p);
5378                 FOREACH_THREAD_IN_PROC(p, td)
5379                         pmclog_process_threadcreate(po, td, 0 /* sync */);
5380                 PROC_UNLOCK(p);
5381         }
5382         sx_sunlock(&allproc_lock);
5383         pmclog_flush(po, 0);
5384 }
5385
5386 static void
5387 pmc_kld_load(void *arg __unused, linker_file_t lf)
5388 {
5389         struct pmc_owner *po;
5390
5391         /*
5392          * Notify owners of system sampling PMCs about KLD operations.
5393          */
5394         PMC_EPOCH_ENTER();
5395         CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
5396                 if (po->po_flags & PMC_PO_OWNS_LOGFILE)
5397                         pmclog_process_map_in(po, (pid_t) -1,
5398                             (uintfptr_t) lf->address, lf->filename);
5399         PMC_EPOCH_EXIT();
5400
5401         /*
5402          * TODO: Notify owners of (all) process-sampling PMCs too.
5403          */
5404 }
5405
5406 static void
5407 pmc_kld_unload(void *arg __unused, const char *filename __unused,
5408     caddr_t address, size_t size)
5409 {
5410         struct pmc_owner *po;
5411
5412         PMC_EPOCH_ENTER();
5413         CK_LIST_FOREACH(po, &pmc_ss_owners, po_ssnext)
5414                 if (po->po_flags & PMC_PO_OWNS_LOGFILE)
5415                         pmclog_process_map_out(po, (pid_t) -1,
5416                             (uintfptr_t) address, (uintfptr_t) address + size);
5417         PMC_EPOCH_EXIT();
5418
5419         /*
5420          * TODO: Notify owners of process-sampling PMCs.
5421          */
5422 }
5423
5424 /*
5425  * initialization
5426  */
5427 static const char *
5428 pmc_name_of_pmcclass(enum pmc_class class)
5429 {
5430
5431         switch (class) {
5432 #undef  __PMC_CLASS
5433 #define __PMC_CLASS(S,V,D)                                              \
5434         case PMC_CLASS_##S:                                             \
5435                 return #S;
5436         __PMC_CLASSES();
5437         default:
5438                 return ("<unknown>");
5439         }
5440 }
5441
5442 /*
5443  * Base class initializer: allocate structure and set default classes.
5444  */
5445 struct pmc_mdep *
5446 pmc_mdep_alloc(int nclasses)
5447 {
5448         struct pmc_mdep *md;
5449         int     n;
5450
5451         /* SOFT + md classes */
5452         n = 1 + nclasses;
5453         md = malloc(sizeof(struct pmc_mdep) + n *
5454             sizeof(struct pmc_classdep), M_PMC, M_WAITOK|M_ZERO);
5455         md->pmd_nclass = n;
5456
5457         /* Add base class. */
5458         pmc_soft_initialize(md);
5459         return md;
5460 }
5461
5462 void
5463 pmc_mdep_free(struct pmc_mdep *md)
5464 {
5465         pmc_soft_finalize(md);
5466         free(md, M_PMC);
5467 }
5468
5469 static int
5470 generic_switch_in(struct pmc_cpu *pc, struct pmc_process *pp)
5471 {
5472         (void) pc; (void) pp;
5473
5474         return (0);
5475 }
5476
5477 static int
5478 generic_switch_out(struct pmc_cpu *pc, struct pmc_process *pp)
5479 {
5480         (void) pc; (void) pp;
5481
5482         return (0);
5483 }
5484
5485 static struct pmc_mdep *
5486 pmc_generic_cpu_initialize(void)
5487 {
5488         struct pmc_mdep *md;
5489
5490         md = pmc_mdep_alloc(0);
5491
5492         md->pmd_cputype    = PMC_CPU_GENERIC;
5493
5494         md->pmd_pcpu_init  = NULL;
5495         md->pmd_pcpu_fini  = NULL;
5496         md->pmd_switch_in  = generic_switch_in;
5497         md->pmd_switch_out = generic_switch_out;
5498
5499         return (md);
5500 }
5501
5502 static void
5503 pmc_generic_cpu_finalize(struct pmc_mdep *md)
5504 {
5505         (void) md;
5506 }
5507
5508
5509 static int
5510 pmc_initialize(void)
5511 {
5512         int c, cpu, error, n, ri;
5513         unsigned int maxcpu, domain;
5514         struct pcpu *pc;
5515         struct pmc_binding pb;
5516         struct pmc_sample *ps;
5517         struct pmc_classdep *pcd;
5518         struct pmc_samplebuffer *sb;
5519
5520         md = NULL;
5521         error = 0;
5522
5523         pmc_stats.pm_intr_ignored = counter_u64_alloc(M_WAITOK);
5524         pmc_stats.pm_intr_processed = counter_u64_alloc(M_WAITOK);
5525         pmc_stats.pm_intr_bufferfull = counter_u64_alloc(M_WAITOK);
5526         pmc_stats.pm_syscalls = counter_u64_alloc(M_WAITOK);
5527         pmc_stats.pm_syscall_errors = counter_u64_alloc(M_WAITOK);
5528         pmc_stats.pm_buffer_requests = counter_u64_alloc(M_WAITOK);
5529         pmc_stats.pm_buffer_requests_failed = counter_u64_alloc(M_WAITOK);
5530         pmc_stats.pm_log_sweeps = counter_u64_alloc(M_WAITOK);
5531         pmc_stats.pm_merges = counter_u64_alloc(M_WAITOK);
5532         pmc_stats.pm_overwrites = counter_u64_alloc(M_WAITOK);
5533
5534 #ifdef  HWPMC_DEBUG
5535         /* parse debug flags first */
5536         if (TUNABLE_STR_FETCH(PMC_SYSCTL_NAME_PREFIX "debugflags",
5537                 pmc_debugstr, sizeof(pmc_debugstr)))
5538                 pmc_debugflags_parse(pmc_debugstr,
5539                     pmc_debugstr+strlen(pmc_debugstr));
5540 #endif
5541
5542         PMCDBG1(MOD,INI,0, "PMC Initialize (version %x)", PMC_VERSION);
5543
5544         /* check kernel version */
5545         if (pmc_kernel_version != PMC_VERSION) {
5546                 if (pmc_kernel_version == 0)
5547                         printf("hwpmc: this kernel has not been compiled with "
5548                             "'options HWPMC_HOOKS'.\n");
5549                 else
5550                         printf("hwpmc: kernel version (0x%x) does not match "
5551                             "module version (0x%x).\n", pmc_kernel_version,
5552                             PMC_VERSION);
5553                 return EPROGMISMATCH;
5554         }
5555
5556         /*
5557          * check sysctl parameters
5558          */
5559
5560         if (pmc_hashsize <= 0) {
5561                 (void) printf("hwpmc: tunable \"hashsize\"=%d must be "
5562                     "greater than zero.\n", pmc_hashsize);
5563                 pmc_hashsize = PMC_HASH_SIZE;
5564         }
5565
5566         if (pmc_nsamples <= 0 || pmc_nsamples > 65535) {
5567                 (void) printf("hwpmc: tunable \"nsamples\"=%d out of "
5568                     "range.\n", pmc_nsamples);
5569                 pmc_nsamples = PMC_NSAMPLES;
5570         }
5571         pmc_sample_mask = pmc_nsamples-1;
5572
5573         if (pmc_callchaindepth <= 0 ||
5574             pmc_callchaindepth > PMC_CALLCHAIN_DEPTH_MAX) {
5575                 (void) printf("hwpmc: tunable \"callchaindepth\"=%d out of "
5576                     "range - using %d.\n", pmc_callchaindepth,
5577                     PMC_CALLCHAIN_DEPTH_MAX);
5578                 pmc_callchaindepth = PMC_CALLCHAIN_DEPTH_MAX;
5579         }
5580
5581         md = pmc_md_initialize();
5582         if (md == NULL) {
5583                 /* Default to generic CPU. */
5584                 md = pmc_generic_cpu_initialize();
5585                 if (md == NULL)
5586                         return (ENOSYS);
5587         }
5588
5589         KASSERT(md->pmd_nclass >= 1 && md->pmd_npmc >= 1,
5590             ("[pmc,%d] no classes or pmcs", __LINE__));
5591
5592         /* Compute the map from row-indices to classdep pointers. */
5593         pmc_rowindex_to_classdep = malloc(sizeof(struct pmc_classdep *) *
5594             md->pmd_npmc, M_PMC, M_WAITOK|M_ZERO);
5595
5596         for (n = 0; n < md->pmd_npmc; n++)
5597                 pmc_rowindex_to_classdep[n] = NULL;
5598         for (ri = c = 0; c < md->pmd_nclass; c++) {
5599                 pcd = &md->pmd_classdep[c];
5600                 for (n = 0; n < pcd->pcd_num; n++, ri++)
5601                         pmc_rowindex_to_classdep[ri] = pcd;
5602         }
5603
5604         KASSERT(ri == md->pmd_npmc,
5605             ("[pmc,%d] npmc miscomputed: ri=%d, md->npmc=%d", __LINE__,
5606             ri, md->pmd_npmc));
5607
5608         maxcpu = pmc_cpu_max();
5609
5610         /* allocate space for the per-cpu array */
5611         pmc_pcpu = malloc(maxcpu * sizeof(struct pmc_cpu *), M_PMC,
5612             M_WAITOK|M_ZERO);
5613
5614         /* per-cpu 'saved values' for managing process-mode PMCs */
5615         pmc_pcpu_saved = malloc(sizeof(pmc_value_t) * maxcpu * md->pmd_npmc,
5616             M_PMC, M_WAITOK);
5617
5618         /* Perform CPU-dependent initialization. */
5619         pmc_save_cpu_binding(&pb);
5620         error = 0;
5621         for (cpu = 0; error == 0 && cpu < maxcpu; cpu++) {
5622                 if (!pmc_cpu_is_active(cpu))
5623                         continue;
5624                 pmc_select_cpu(cpu);
5625                 pmc_pcpu[cpu] = malloc(sizeof(struct pmc_cpu) +
5626                     md->pmd_npmc * sizeof(struct pmc_hw *), M_PMC,
5627                     M_WAITOK|M_ZERO);
5628                 if (md->pmd_pcpu_init)
5629                         error = md->pmd_pcpu_init(md, cpu);
5630                 for (n = 0; error == 0 && n < md->pmd_nclass; n++)
5631                         error = md->pmd_classdep[n].pcd_pcpu_init(md, cpu);
5632         }
5633         pmc_restore_cpu_binding(&pb);
5634
5635         if (error)
5636                 return (error);
5637
5638         /* allocate space for the sample array */
5639         for (cpu = 0; cpu < maxcpu; cpu++) {
5640                 if (!pmc_cpu_is_active(cpu))
5641                         continue;
5642                 pc = pcpu_find(cpu);
5643                 domain = pc->pc_domain;
5644                 sb = malloc_domainset(sizeof(struct pmc_samplebuffer) +
5645                     pmc_nsamples * sizeof(struct pmc_sample), M_PMC,
5646                     DOMAINSET_PREF(domain), M_WAITOK | M_ZERO);
5647
5648                 KASSERT(pmc_pcpu[cpu] != NULL,
5649                     ("[pmc,%d] cpu=%d Null per-cpu data", __LINE__, cpu));
5650
5651                 sb->ps_callchains = malloc_domainset(pmc_callchaindepth *
5652                     pmc_nsamples * sizeof(uintptr_t), M_PMC,
5653                     DOMAINSET_PREF(domain), M_WAITOK | M_ZERO);
5654
5655                 for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++)
5656                         ps->ps_pc = sb->ps_callchains +
5657                             (n * pmc_callchaindepth);
5658
5659                 pmc_pcpu[cpu]->pc_sb[PMC_HR] = sb;
5660
5661                 sb = malloc_domainset(sizeof(struct pmc_samplebuffer) +
5662                     pmc_nsamples * sizeof(struct pmc_sample), M_PMC,
5663                     DOMAINSET_PREF(domain), M_WAITOK | M_ZERO);
5664
5665                 sb->ps_callchains = malloc_domainset(pmc_callchaindepth *
5666                     pmc_nsamples * sizeof(uintptr_t), M_PMC,
5667                     DOMAINSET_PREF(domain), M_WAITOK | M_ZERO);
5668                 for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++)
5669                         ps->ps_pc = sb->ps_callchains +
5670                             (n * pmc_callchaindepth);
5671
5672                 pmc_pcpu[cpu]->pc_sb[PMC_SR] = sb;
5673
5674                 sb = malloc_domainset(sizeof(struct pmc_samplebuffer) +
5675                     pmc_nsamples * sizeof(struct pmc_sample), M_PMC,
5676                     DOMAINSET_PREF(domain), M_WAITOK | M_ZERO);
5677                 sb->ps_callchains = malloc_domainset(pmc_callchaindepth *
5678                     pmc_nsamples * sizeof(uintptr_t), M_PMC,
5679                     DOMAINSET_PREF(domain), M_WAITOK | M_ZERO);
5680                 for (n = 0, ps = sb->ps_samples; n < pmc_nsamples; n++, ps++)
5681                         ps->ps_pc = sb->ps_callchains + n * pmc_callchaindepth;
5682
5683                 pmc_pcpu[cpu]->pc_sb[PMC_UR] = sb;
5684         }
5685
5686         /* allocate space for the row disposition array */
5687         pmc_pmcdisp = malloc(sizeof(enum pmc_mode) * md->pmd_npmc,
5688             M_PMC, M_WAITOK|M_ZERO);
5689
5690         /* mark all PMCs as available */
5691         for (n = 0; n < (int) md->pmd_npmc; n++)
5692                 PMC_MARK_ROW_FREE(n);
5693
5694         /* allocate thread hash tables */
5695         pmc_ownerhash = hashinit(pmc_hashsize, M_PMC,
5696             &pmc_ownerhashmask);
5697
5698         pmc_processhash = hashinit(pmc_hashsize, M_PMC,
5699             &pmc_processhashmask);
5700         mtx_init(&pmc_processhash_mtx, "pmc-process-hash", "pmc-leaf",
5701             MTX_SPIN);
5702
5703         CK_LIST_INIT(&pmc_ss_owners);
5704         pmc_ss_count = 0;
5705
5706         /* allocate a pool of spin mutexes */
5707         pmc_mtxpool = mtx_pool_create("pmc-leaf", pmc_mtxpool_size,
5708             MTX_SPIN);
5709
5710         PMCDBG4(MOD,INI,1, "pmc_ownerhash=%p, mask=0x%lx "
5711             "targethash=%p mask=0x%lx", pmc_ownerhash, pmc_ownerhashmask,
5712             pmc_processhash, pmc_processhashmask);
5713
5714         /* Initialize a spin mutex for the thread free list. */
5715         mtx_init(&pmc_threadfreelist_mtx, "pmc-threadfreelist", "pmc-leaf",
5716             MTX_SPIN);
5717
5718         /* Initialize the task to prune the thread free list. */
5719         TASK_INIT(&free_task, 0, pmc_thread_descriptor_pool_free_task, NULL);
5720
5721         /* register process {exit,fork,exec} handlers */
5722         pmc_exit_tag = EVENTHANDLER_REGISTER(process_exit,
5723             pmc_process_exit, NULL, EVENTHANDLER_PRI_ANY);
5724         pmc_fork_tag = EVENTHANDLER_REGISTER(process_fork,
5725             pmc_process_fork, NULL, EVENTHANDLER_PRI_ANY);
5726
5727         /* register kld event handlers */
5728         pmc_kld_load_tag = EVENTHANDLER_REGISTER(kld_load, pmc_kld_load,
5729             NULL, EVENTHANDLER_PRI_ANY);
5730         pmc_kld_unload_tag = EVENTHANDLER_REGISTER(kld_unload, pmc_kld_unload,
5731             NULL, EVENTHANDLER_PRI_ANY);
5732
5733         /* initialize logging */
5734         pmclog_initialize();
5735
5736         /* set hook functions */
5737         pmc_intr = md->pmd_intr;
5738         wmb();
5739         pmc_hook = pmc_hook_handler;
5740
5741         if (error == 0) {
5742                 printf(PMC_MODULE_NAME ":");
5743                 for (n = 0; n < (int) md->pmd_nclass; n++) {
5744                         pcd = &md->pmd_classdep[n];
5745                         printf(" %s/%d/%d/0x%b",
5746                             pmc_name_of_pmcclass(pcd->pcd_class),
5747                             pcd->pcd_num,
5748                             pcd->pcd_width,
5749                             pcd->pcd_caps,
5750                             "\20"
5751                             "\1INT\2USR\3SYS\4EDG\5THR"
5752                             "\6REA\7WRI\10INV\11QUA\12PRC"
5753                             "\13TAG\14CSC");
5754                 }
5755                 printf("\n");
5756         }
5757
5758         return (error);
5759 }
5760
5761 /* prepare to be unloaded */
5762 static void
5763 pmc_cleanup(void)
5764 {
5765         int c, cpu;
5766         unsigned int maxcpu;
5767         struct pmc_ownerhash *ph;
5768         struct pmc_owner *po, *tmp;
5769         struct pmc_binding pb;
5770 #ifdef  HWPMC_DEBUG
5771         struct pmc_processhash *prh;
5772 #endif
5773
5774         PMCDBG0(MOD,INI,0, "cleanup");
5775
5776         /* switch off sampling */
5777         CPU_FOREACH(cpu)
5778                 DPCPU_ID_SET(cpu, pmc_sampled, 0);
5779         pmc_intr = NULL;
5780
5781         sx_xlock(&pmc_sx);
5782         if (pmc_hook == NULL) { /* being unloaded already */
5783                 sx_xunlock(&pmc_sx);
5784                 return;
5785         }
5786
5787         pmc_hook = NULL; /* prevent new threads from entering module */
5788
5789         /* deregister event handlers */
5790         EVENTHANDLER_DEREGISTER(process_fork, pmc_fork_tag);
5791         EVENTHANDLER_DEREGISTER(process_exit, pmc_exit_tag);
5792         EVENTHANDLER_DEREGISTER(kld_load, pmc_kld_load_tag);
5793         EVENTHANDLER_DEREGISTER(kld_unload, pmc_kld_unload_tag);
5794
5795         /* send SIGBUS to all owner threads, free up allocations */
5796         if (pmc_ownerhash)
5797                 for (ph = pmc_ownerhash;
5798                      ph <= &pmc_ownerhash[pmc_ownerhashmask];
5799                      ph++) {
5800                         LIST_FOREACH_SAFE(po, ph, po_next, tmp) {
5801                                 pmc_remove_owner(po);
5802
5803                                 /* send SIGBUS to owner processes */
5804                                 PMCDBG3(MOD,INI,2, "cleanup signal proc=%p "
5805                                     "(%d, %s)", po->po_owner,
5806                                     po->po_owner->p_pid,
5807                                     po->po_owner->p_comm);
5808
5809                                 PROC_LOCK(po->po_owner);
5810                                 kern_psignal(po->po_owner, SIGBUS);
5811                                 PROC_UNLOCK(po->po_owner);
5812
5813                                 pmc_destroy_owner_descriptor(po);
5814                         }
5815                 }
5816
5817         /* reclaim allocated data structures */
5818         taskqueue_drain(taskqueue_fast, &free_task);
5819         mtx_destroy(&pmc_threadfreelist_mtx);
5820         pmc_thread_descriptor_pool_drain();
5821
5822         if (pmc_mtxpool)
5823                 mtx_pool_destroy(&pmc_mtxpool);
5824
5825         mtx_destroy(&pmc_processhash_mtx);
5826         if (pmc_processhash) {
5827 #ifdef  HWPMC_DEBUG
5828                 struct pmc_process *pp;
5829
5830                 PMCDBG0(MOD,INI,3, "destroy process hash");
5831                 for (prh = pmc_processhash;
5832                      prh <= &pmc_processhash[pmc_processhashmask];
5833                      prh++)
5834                         LIST_FOREACH(pp, prh, pp_next)
5835                             PMCDBG1(MOD,INI,3, "pid=%d", pp->pp_proc->p_pid);
5836 #endif
5837
5838                 hashdestroy(pmc_processhash, M_PMC, pmc_processhashmask);
5839                 pmc_processhash = NULL;
5840         }
5841
5842         if (pmc_ownerhash) {
5843                 PMCDBG0(MOD,INI,3, "destroy owner hash");
5844                 hashdestroy(pmc_ownerhash, M_PMC, pmc_ownerhashmask);
5845                 pmc_ownerhash = NULL;
5846         }
5847
5848         KASSERT(CK_LIST_EMPTY(&pmc_ss_owners),
5849             ("[pmc,%d] Global SS owner list not empty", __LINE__));
5850         KASSERT(pmc_ss_count == 0,
5851             ("[pmc,%d] Global SS count not empty", __LINE__));
5852
5853         /* do processor and pmc-class dependent cleanup */
5854         maxcpu = pmc_cpu_max();
5855
5856         PMCDBG0(MOD,INI,3, "md cleanup");
5857         if (md) {
5858                 pmc_save_cpu_binding(&pb);
5859                 for (cpu = 0; cpu < maxcpu; cpu++) {
5860                         PMCDBG2(MOD,INI,1,"pmc-cleanup cpu=%d pcs=%p",
5861                             cpu, pmc_pcpu[cpu]);
5862                         if (!pmc_cpu_is_active(cpu) || pmc_pcpu[cpu] == NULL)
5863                                 continue;
5864                         pmc_select_cpu(cpu);
5865                         for (c = 0; c < md->pmd_nclass; c++)
5866                                 md->pmd_classdep[c].pcd_pcpu_fini(md, cpu);
5867                         if (md->pmd_pcpu_fini)
5868                                 md->pmd_pcpu_fini(md, cpu);
5869                 }
5870
5871                 if (md->pmd_cputype == PMC_CPU_GENERIC)
5872                         pmc_generic_cpu_finalize(md);
5873                 else
5874                         pmc_md_finalize(md);
5875
5876                 pmc_mdep_free(md);
5877                 md = NULL;
5878                 pmc_restore_cpu_binding(&pb);
5879         }
5880
5881         /* Free per-cpu descriptors. */
5882         for (cpu = 0; cpu < maxcpu; cpu++) {
5883                 if (!pmc_cpu_is_active(cpu))
5884                         continue;
5885                 KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_HR] != NULL,
5886                     ("[pmc,%d] Null hw cpu sample buffer cpu=%d", __LINE__,
5887                         cpu));
5888                 KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_SR] != NULL,
5889                     ("[pmc,%d] Null sw cpu sample buffer cpu=%d", __LINE__,
5890                         cpu));
5891                 KASSERT(pmc_pcpu[cpu]->pc_sb[PMC_UR] != NULL,
5892                     ("[pmc,%d] Null userret cpu sample buffer cpu=%d", __LINE__,
5893                         cpu));
5894                 free_domain(pmc_pcpu[cpu]->pc_sb[PMC_HR]->ps_callchains, M_PMC);
5895                 free_domain(pmc_pcpu[cpu]->pc_sb[PMC_HR], M_PMC);
5896                 free_domain(pmc_pcpu[cpu]->pc_sb[PMC_SR]->ps_callchains, M_PMC);
5897                 free_domain(pmc_pcpu[cpu]->pc_sb[PMC_SR], M_PMC);
5898                 free_domain(pmc_pcpu[cpu]->pc_sb[PMC_UR]->ps_callchains, M_PMC);
5899                 free_domain(pmc_pcpu[cpu]->pc_sb[PMC_UR], M_PMC);
5900                 free_domain(pmc_pcpu[cpu], M_PMC);
5901         }
5902
5903         free(pmc_pcpu, M_PMC);
5904         pmc_pcpu = NULL;
5905
5906         free(pmc_pcpu_saved, M_PMC);
5907         pmc_pcpu_saved = NULL;
5908
5909         if (pmc_pmcdisp) {
5910                 free(pmc_pmcdisp, M_PMC);
5911                 pmc_pmcdisp = NULL;
5912         }
5913
5914         if (pmc_rowindex_to_classdep) {
5915                 free(pmc_rowindex_to_classdep, M_PMC);
5916                 pmc_rowindex_to_classdep = NULL;
5917         }
5918
5919         pmclog_shutdown();
5920         counter_u64_free(pmc_stats.pm_intr_ignored);
5921         counter_u64_free(pmc_stats.pm_intr_processed);
5922         counter_u64_free(pmc_stats.pm_intr_bufferfull);
5923         counter_u64_free(pmc_stats.pm_syscalls);
5924         counter_u64_free(pmc_stats.pm_syscall_errors);
5925         counter_u64_free(pmc_stats.pm_buffer_requests);
5926         counter_u64_free(pmc_stats.pm_buffer_requests_failed);
5927         counter_u64_free(pmc_stats.pm_log_sweeps);
5928         counter_u64_free(pmc_stats.pm_merges);
5929         counter_u64_free(pmc_stats.pm_overwrites);
5930         sx_xunlock(&pmc_sx);    /* we are done */
5931 }
5932
5933 /*
5934  * The function called at load/unload.
5935  */
5936
5937 static int
5938 load (struct module *module __unused, int cmd, void *arg __unused)
5939 {
5940         int error;
5941
5942         error = 0;
5943
5944         switch (cmd) {
5945         case MOD_LOAD :
5946                 /* initialize the subsystem */
5947                 error = pmc_initialize();
5948                 if (error != 0)
5949                         break;
5950                 PMCDBG2(MOD,INI,1, "syscall=%d maxcpu=%d",
5951                     pmc_syscall_num, pmc_cpu_max());
5952                 break;
5953
5954
5955         case MOD_UNLOAD :
5956         case MOD_SHUTDOWN:
5957                 pmc_cleanup();
5958                 PMCDBG0(MOD,INI,1, "unloaded");
5959                 break;
5960
5961         default :
5962                 error = EINVAL; /* XXX should panic(9) */
5963                 break;
5964         }
5965
5966         return error;
5967 }