sys/cddl/dev/profile/profile.c

   1 /*
   2  * CDDL HEADER START
   3  *
   4  * The contents of this file are subject to the terms of the
   5  * Common Development and Distribution License (the "License").
   6  * You may not use this file except in compliance with the License.
   7  *
   8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
   9  * or http://www.opensolaris.org/os/licensing.
  10  * See the License for the specific language governing permissions
  11  * and limitations under the License.
  12  *
  13  * When distributing Covered Code, include this CDDL HEADER in each
  14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
  15  * If applicable, add the following below this CDDL HEADER, with the
  16  * fields enclosed by brackets "[]" replaced with your own identifying
  17  * information: Portions Copyright [yyyy] [name of copyright owner]
  18  *
  19  * CDDL HEADER END
  20  *
  21  * Portions Copyright 2006-2008 John Birrell jb@freebsd.org
  22  *
  23  * $FreeBSD$
  24  *
  25  */
  26
  27 /*
  28  * Copyright 2006 Sun Microsystems, Inc.  All rights reserved.
  29  * Use is subject to license terms.
  30  */
  31
  32 #include <sys/cdefs.h>
  33 #include <sys/param.h>
  34 #include <sys/systm.h>
  35 #include <sys/conf.h>
  36 #include <sys/cpuvar.h>
  37 #include <sys/fcntl.h>
  38 #include <sys/filio.h>
  39 #include <sys/kdb.h>
  40 #include <sys/kernel.h>
  41 #include <sys/kmem.h>
  42 #include <sys/kthread.h>
  43 #include <sys/limits.h>
  44 #include <sys/linker.h>
  45 #include <sys/lock.h>
  46 #include <sys/malloc.h>
  47 #include <sys/module.h>
  48 #include <sys/mutex.h>
  49 #include <sys/poll.h>
  50 #include <sys/proc.h>
  51 #include <sys/selinfo.h>
  52 #include <sys/smp.h>
  53 #include <sys/uio.h>
  54 #include <sys/unistd.h>
  55 #include <machine/stdarg.h>
  56
  57 #include <sys/cyclic.h>
  58 #include <sys/dtrace.h>
  59 #include <sys/dtrace_bsd.h>
  60
  61 #define PROF_NAMELEN            15
  62
  63 #define PROF_PROFILE            0
  64 #define PROF_TICK               1
  65 #define PROF_PREFIX_PROFILE     "profile-"
  66 #define PROF_PREFIX_TICK        "tick-"
  67
  68 /*
  69  * Regardless of platform, there are five artificial frames in the case of the
  70  * profile provider:
  71  *
  72  *      profile_fire
  73  *      cyclic_expire
  74  *      cyclic_fire
  75  *      [ cbe ]
  76  *      [ locore ]
  77  *
  78  * On amd64, there are two frames associated with locore:  one in locore, and
  79  * another in common interrupt dispatch code.  (i386 has not been modified to
  80  * use this common layer.)  Further, on i386, the interrupted instruction
  81  * appears as its own stack frame.  All of this means that we need to add one
  82  * frame for amd64, and then take one away for both amd64 and i386.
  83  *
  84  * On SPARC, the picture is further complicated because the compiler
  85  * optimizes away tail-calls -- so the following frames are optimized away:
  86  *
  87  *      profile_fire
  88  *      cyclic_expire
  89  *
  90  * This gives three frames.  However, on DEBUG kernels, the cyclic_expire
  91  * frame cannot be tail-call eliminated, yielding four frames in this case.
  92  *
  93  * All of the above constraints lead to the mess below.  Yes, the profile
  94  * provider should ideally figure this out on-the-fly by hiting one of its own
  95  * probes and then walking its own stack trace.  This is complicated, however,
  96  * and the static definition doesn't seem to be overly brittle.  Still, we
  97  * allow for a manual override in case we get it completely wrong.
  98  */
  99 #ifdef __amd64
 100 #define PROF_ARTIFICIAL_FRAMES  7
 101 #else
 102 #ifdef __i386
 103 #define PROF_ARTIFICIAL_FRAMES  6
 104 #else
 105 #ifdef __sparc
 106 #ifdef DEBUG
 107 #define PROF_ARTIFICIAL_FRAMES  4
 108 #else
 109 #define PROF_ARTIFICIAL_FRAMES  3
 110 #endif
 111 #endif
 112 #endif
 113 #endif
 114
 115 typedef struct profile_probe {
 116         char            prof_name[PROF_NAMELEN];
 117         dtrace_id_t     prof_id;
 118         int             prof_kind;
 119         hrtime_t        prof_interval;
 120         cyclic_id_t     prof_cyclic;
 121 } profile_probe_t;
 122
 123 typedef struct profile_probe_percpu {
 124         hrtime_t        profc_expected;
 125         hrtime_t        profc_interval;
 126         profile_probe_t *profc_probe;
 127 } profile_probe_percpu_t;
 128
 129 static d_open_t profile_open;
 130 static int      profile_unload(void);
 131 static void     profile_create(hrtime_t, char *, int);
 132 static void     profile_destroy(void *, dtrace_id_t, void *);
 133 static void     profile_enable(void *, dtrace_id_t, void *);
 134 static void     profile_disable(void *, dtrace_id_t, void *);
 135 static void     profile_load(void *);
 136 static void     profile_provide(void *, dtrace_probedesc_t *);
 137
 138 static int profile_rates[] = {
 139     97, 199, 499, 997, 1999,
 140     4001, 4999, 0, 0, 0,
 141     0, 0, 0, 0, 0,
 142     0, 0, 0, 0, 0
 143 };
 144
 145 static int profile_ticks[] = {
 146     1, 10, 100, 500, 1000,
 147     5000, 0, 0, 0, 0,
 148     0, 0, 0, 0, 0
 149 };
 150
 151 /*
 152  * profile_max defines the upper bound on the number of profile probes that
 153  * can exist (this is to prevent malicious or clumsy users from exhausing
 154  * system resources by creating a slew of profile probes). At mod load time,
 155  * this gets its value from PROFILE_MAX_DEFAULT or profile-max-probes if it's
 156  * present in the profile.conf file.
 157  */
 158 #define PROFILE_MAX_DEFAULT     1000    /* default max. number of probes */
 159 static uint32_t profile_max = PROFILE_MAX_DEFAULT;
 160                                         /* maximum number of profile probes */
 161 static uint32_t profile_total;          /* current number of profile probes */
 162
 163 static struct cdevsw profile_cdevsw = {
 164         .d_version      = D_VERSION,
 165         .d_open         = profile_open,
 166         .d_name         = "profile",
 167 };
 168
 169 static dtrace_pattr_t profile_attr = {
 170 { DTRACE_STABILITY_EVOLVING, DTRACE_STABILITY_EVOLVING, DTRACE_CLASS_COMMON },
 171 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_UNKNOWN },
 172 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_ISA },
 173 { DTRACE_STABILITY_EVOLVING, DTRACE_STABILITY_EVOLVING, DTRACE_CLASS_COMMON },
 174 { DTRACE_STABILITY_PRIVATE, DTRACE_STABILITY_PRIVATE, DTRACE_CLASS_ISA },
 175 };
 176
 177 static dtrace_pops_t profile_pops = {
 178         profile_provide,
 179         NULL,
 180         profile_enable,
 181         profile_disable,
 182         NULL,
 183         NULL,
 184         NULL,
 185         NULL,
 186         NULL,
 187         profile_destroy
 188 };
 189
 190 static struct cdev              *profile_cdev;
 191 static dtrace_provider_id_t     profile_id;
 192 static hrtime_t                 profile_interval_min = NANOSEC / 5000;  /* 5000 hz */
 193 static int                      profile_aframes = 0;                    /* override */
 194
 195 static void
 196 profile_fire(void *arg)
 197 {
 198         profile_probe_percpu_t *pcpu = arg;
 199         profile_probe_t *prof = pcpu->profc_probe;
 200         hrtime_t late;
 201         solaris_cpu_t *c = &solaris_cpu[curcpu];
 202
 203         late = gethrtime() - pcpu->profc_expected;
 204         pcpu->profc_expected += pcpu->profc_interval;
 205
 206         dtrace_probe(prof->prof_id, c->cpu_profile_pc,
 207             c->cpu_profile_upc, late, 0, 0);
 208 }
 209
 210 static void
 211 profile_tick(void *arg)
 212 {
 213         profile_probe_t *prof = arg;
 214         solaris_cpu_t *c = &solaris_cpu[curcpu];
 215
 216         dtrace_probe(prof->prof_id, c->cpu_profile_pc,
 217             c->cpu_profile_upc, 0, 0, 0);
 218 }
 219
 220 static void
 221 profile_create(hrtime_t interval, char *name, int kind)
 222 {
 223         profile_probe_t *prof;
 224
 225         if (interval < profile_interval_min)
 226                 return;
 227
 228         if (dtrace_probe_lookup(profile_id, NULL, NULL, name) != 0)
 229                 return;
 230
 231         atomic_add_32(&profile_total, 1);
 232         if (profile_total > profile_max) {
 233                 atomic_add_32(&profile_total, -1);
 234                 return;
 235         }
 236
 237         prof = kmem_zalloc(sizeof (profile_probe_t), KM_SLEEP);
 238         (void) strcpy(prof->prof_name, name);
 239         prof->prof_interval = interval;
 240         prof->prof_cyclic = CYCLIC_NONE;
 241         prof->prof_kind = kind;
 242         prof->prof_id = dtrace_probe_create(profile_id,
 243             NULL, NULL, name,
 244             profile_aframes ? profile_aframes : PROF_ARTIFICIAL_FRAMES, prof);
 245 }
 246
 247 /*ARGSUSED*/
 248 static void
 249 profile_provide(void *arg, dtrace_probedesc_t *desc)
 250 {
 251         int i, j, rate, kind;
 252         hrtime_t val = 0, mult = 1, len = 0;
 253         char *name, *suffix = NULL;
 254
 255         const struct {
 256                 char *prefix;
 257                 int kind;
 258         } types[] = {
 259                 { PROF_PREFIX_PROFILE, PROF_PROFILE },
 260                 { PROF_PREFIX_TICK, PROF_TICK },
 261                 { 0, 0 }
 262         };
 263
 264         const struct {
 265                 char *name;
 266                 hrtime_t mult;
 267         } suffixes[] = {
 268                 { "ns",         NANOSEC / NANOSEC },
 269                 { "nsec",       NANOSEC / NANOSEC },
 270                 { "us",         NANOSEC / MICROSEC },
 271                 { "usec",       NANOSEC / MICROSEC },
 272                 { "ms",         NANOSEC / MILLISEC },
 273                 { "msec",       NANOSEC / MILLISEC },
 274                 { "s",          NANOSEC / SEC },
 275                 { "sec",        NANOSEC / SEC },
 276                 { "m",          NANOSEC * (hrtime_t)60 },
 277                 { "min",        NANOSEC * (hrtime_t)60 },
 278                 { "h",          NANOSEC * (hrtime_t)(60 * 60) },
 279                 { "hour",       NANOSEC * (hrtime_t)(60 * 60) },
 280                 { "d",          NANOSEC * (hrtime_t)(24 * 60 * 60) },
 281                 { "day",        NANOSEC * (hrtime_t)(24 * 60 * 60) },
 282                 { "hz",         0 },
 283                 { NULL }
 284         };
 285
 286         if (desc == NULL) {
 287                 char n[PROF_NAMELEN];
 288
 289                 /*
 290                  * If no description was provided, provide all of our probes.
 291                  */
 292                 for (i = 0; i < sizeof (profile_rates) / sizeof (int); i++) {
 293                         if ((rate = profile_rates[i]) == 0)
 294                                 continue;
 295
 296                         (void) snprintf(n, PROF_NAMELEN, "%s%d",
 297                             PROF_PREFIX_PROFILE, rate);
 298                         profile_create(NANOSEC / rate, n, PROF_PROFILE);
 299                 }
 300
 301                 for (i = 0; i < sizeof (profile_ticks) / sizeof (int); i++) {
 302                         if ((rate = profile_ticks[i]) == 0)
 303                                 continue;
 304
 305                         (void) snprintf(n, PROF_NAMELEN, "%s%d",
 306                             PROF_PREFIX_TICK, rate);
 307                         profile_create(NANOSEC / rate, n, PROF_TICK);
 308                 }
 309
 310                 return;
 311         }
 312
 313         name = desc->dtpd_name;
 314
 315         for (i = 0; types[i].prefix != NULL; i++) {
 316                 len = strlen(types[i].prefix);
 317
 318                 if (strncmp(name, types[i].prefix, len) != 0)
 319                         continue;
 320                 break;
 321         }
 322
 323         if (types[i].prefix == NULL)
 324                 return;
 325
 326         kind = types[i].kind;
 327         j = strlen(name) - len;
 328
 329         /*
 330          * We need to start before any time suffix.
 331          */
 332         for (j = strlen(name); j >= len; j--) {
 333                 if (name[j] >= '0' && name[j] <= '9')
 334                         break;
 335                 suffix = &name[j];
 336         }
 337
 338         ASSERT(suffix != NULL);
 339
 340         /*
 341          * Now determine the numerical value present in the probe name.
 342          */
 343         for (; j >= len; j--) {
 344                 if (name[j] < '0' || name[j] > '9')
 345                         return;
 346
 347                 val += (name[j] - '0') * mult;
 348                 mult *= (hrtime_t)10;
 349         }
 350
 351         if (val == 0)
 352                 return;
 353
 354         /*
 355          * Look-up the suffix to determine the multiplier.
 356          */
 357         for (i = 0, mult = 0; suffixes[i].name != NULL; i++) {
 358                 if (strcasecmp(suffixes[i].name, suffix) == 0) {
 359                         mult = suffixes[i].mult;
 360                         break;
 361                 }
 362         }
 363
 364         if (suffixes[i].name == NULL && *suffix != '\0')
 365                 return;
 366
 367         if (mult == 0) {
 368                 /*
 369                  * The default is frequency-per-second.
 370                  */
 371                 val = NANOSEC / val;
 372         } else {
 373                 val *= mult;
 374         }
 375
 376         profile_create(val, name, kind);
 377 }
 378
 379 /* ARGSUSED */
 380 static void
 381 profile_destroy(void *arg, dtrace_id_t id, void *parg)
 382 {
 383         profile_probe_t *prof = parg;
 384
 385         ASSERT(prof->prof_cyclic == CYCLIC_NONE);
 386         kmem_free(prof, sizeof (profile_probe_t));
 387
 388         ASSERT(profile_total >= 1);
 389         atomic_add_32(&profile_total, -1);
 390 }
 391
 392 /*ARGSUSED*/
 393 static void
 394 profile_online(void *arg, cpu_t *cpu, cyc_handler_t *hdlr, cyc_time_t *when)
 395 {
 396         profile_probe_t *prof = arg;
 397         profile_probe_percpu_t *pcpu;
 398
 399         pcpu = kmem_zalloc(sizeof (profile_probe_percpu_t), KM_SLEEP);
 400         pcpu->profc_probe = prof;
 401
 402         hdlr->cyh_func = profile_fire;
 403         hdlr->cyh_arg = pcpu;
 404
 405         when->cyt_interval = prof->prof_interval;
 406         when->cyt_when = gethrtime() + when->cyt_interval;
 407
 408         pcpu->profc_expected = when->cyt_when;
 409         pcpu->profc_interval = when->cyt_interval;
 410 }
 411
 412 /*ARGSUSED*/
 413 static void
 414 profile_offline(void *arg, cpu_t *cpu, void *oarg)
 415 {
 416         profile_probe_percpu_t *pcpu = oarg;
 417
 418         ASSERT(pcpu->profc_probe == arg);
 419         kmem_free(pcpu, sizeof (profile_probe_percpu_t));
 420 }
 421
 422 /* ARGSUSED */
 423 static void
 424 profile_enable(void *arg, dtrace_id_t id, void *parg)
 425 {
 426         profile_probe_t *prof = parg;
 427         cyc_omni_handler_t omni;
 428         cyc_handler_t hdlr;
 429         cyc_time_t when;
 430
 431         ASSERT(prof->prof_interval != 0);
 432         ASSERT(MUTEX_HELD(&cpu_lock));
 433
 434         if (prof->prof_kind == PROF_TICK) {
 435                 hdlr.cyh_func = profile_tick;
 436                 hdlr.cyh_arg = prof;
 437
 438                 when.cyt_interval = prof->prof_interval;
 439                 when.cyt_when = gethrtime() + when.cyt_interval;
 440         } else {
 441                 ASSERT(prof->prof_kind == PROF_PROFILE);
 442                 omni.cyo_online = profile_online;
 443                 omni.cyo_offline = profile_offline;
 444                 omni.cyo_arg = prof;
 445         }
 446
 447         if (prof->prof_kind == PROF_TICK) {
 448                 prof->prof_cyclic = cyclic_add(&hdlr, &when);
 449         } else {
 450                 prof->prof_cyclic = cyclic_add_omni(&omni);
 451         }
 452 }
 453
 454 /* ARGSUSED */
 455 static void
 456 profile_disable(void *arg, dtrace_id_t id, void *parg)
 457 {
 458         profile_probe_t *prof = parg;
 459
 460         ASSERT(prof->prof_cyclic != CYCLIC_NONE);
 461         ASSERT(MUTEX_HELD(&cpu_lock));
 462
 463         cyclic_remove(prof->prof_cyclic);
 464         prof->prof_cyclic = CYCLIC_NONE;
 465 }
 466
 467 static void
 468 profile_load(void *dummy)
 469 {
 470         /* Create the /dev/dtrace/profile entry. */
 471         profile_cdev = make_dev(&profile_cdevsw, 0, UID_ROOT, GID_WHEEL, 0600,
 472             "dtrace/profile");
 473
 474         if (dtrace_register("profile", &profile_attr, DTRACE_PRIV_USER,
 475             NULL, &profile_pops, NULL, &profile_id) != 0)
 476                 return;
 477 }
 478
 479
 480 static int
 481 profile_unload()
 482 {
 483         int error = 0;
 484
 485         if ((error = dtrace_unregister(profile_id)) != 0)
 486                 return (error);
 487
 488         destroy_dev(profile_cdev);
 489
 490         return (error);
 491 }
 492
 493 /* ARGSUSED */
 494 static int
 495 profile_modevent(module_t mod __unused, int type, void *data __unused)
 496 {
 497         int error = 0;
 498
 499         switch (type) {
 500         case MOD_LOAD:
 501                 break;
 502
 503         case MOD_UNLOAD:
 504                 break;
 505
 506         case MOD_SHUTDOWN:
 507                 break;
 508
 509         default:
 510                 error = EOPNOTSUPP;
 511                 break;
 512
 513         }
 514         return (error);
 515 }
 516
 517 /* ARGSUSED */
 518 static int
 519 profile_open(struct cdev *dev __unused, int oflags __unused, int devtype __unused, struct thread *td __unused)
 520 {
 521         return (0);
 522 }
 523
 524 SYSINIT(profile_load, SI_SUB_DTRACE_PROVIDER, SI_ORDER_ANY, profile_load, NULL);
 525 SYSUNINIT(profile_unload, SI_SUB_DTRACE_PROVIDER, SI_ORDER_ANY, profile_unload, NULL);
 526
 527 DEV_MODULE(profile, profile_modevent, NULL);
 528 MODULE_VERSION(profile, 1);
 529 MODULE_DEPEND(profile, dtrace, 1, 1, 1);
 530 MODULE_DEPEND(profile, cyclic, 1, 1, 1);
 531 MODULE_DEPEND(profile, opensolaris, 1, 1, 1);