osfmk/kperf/kperf_timer.c

   1 /*
   2  * Copyright (c) 2011 Apple Computer, Inc. All rights reserved.
   3  *
   4  * @APPLE_OSREFERENCE_LICENSE_HEADER_START@
   5  *
   6  * This file contains Original Code and/or Modifications of Original Code
   7  * as defined in and that are subject to the Apple Public Source License
   8  * Version 2.0 (the 'License'). You may not use this file except in
   9  * compliance with the License. The rights granted to you under the License
  10  * may not be used to create, or enable the creation or redistribution of,
  11  * unlawful or unlicensed copies of an Apple operating system, or to
  12  * circumvent, violate, or enable the circumvention or violation of, any
  13  * terms of an Apple operating system software license agreement.
  14  *
  15  * Please obtain a copy of the License at
  16  * http://www.opensource.apple.com/apsl/ and read it before using this file.
  17  *
  18  * The Original Code and all software distributed under the License are
  19  * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER
  20  * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES,
  21  * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY,
  22  * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT.
  23  * Please see the License for the specific language governing rights and
  24  * limitations under the License.
  25  *
  26  * @APPLE_OSREFERENCE_LICENSE_HEADER_END@
  27  */
  28
  29 /*  Manage timers */
  30
  31 #include <mach/mach_types.h>
  32 #include <kern/cpu_data.h> /* current_thread() */
  33 #include <kern/kalloc.h>
  34 #include <stdatomic.h>
  35 #include <sys/errno.h>
  36 #include <sys/vm.h>
  37 #include <sys/ktrace.h>
  38
  39 #include <machine/machine_routines.h>
  40 #if defined(__x86_64__)
  41 #include <i386/mp.h>
  42 #endif /* defined(__x86_64__) */
  43
  44 #include <kperf/kperf.h>
  45 #include <kperf/buffer.h>
  46 #include <kperf/context.h>
  47 #include <kperf/action.h>
  48 #include <kperf/kperf_timer.h>
  49 #include <kperf/kperf_arch.h>
  50 #include <kperf/pet.h>
  51 #include <kperf/sample.h>
  52
  53 /* the list of timers */
  54 struct kperf_timer *kperf_timerv = NULL;
  55 unsigned int kperf_timerc = 0;
  56
  57 static unsigned int pet_timer_id = 999;
  58
  59 /* maximum number of timers we can construct */
  60 #define TIMER_MAX (16)
  61
  62 static uint64_t min_period_abstime;
  63 static uint64_t min_period_bg_abstime;
  64 static uint64_t min_period_pet_abstime;
  65 static uint64_t min_period_pet_bg_abstime;
  66
  67 static uint64_t
  68 kperf_timer_min_period_abstime(void)
  69 {
  70         if (ktrace_background_active()) {
  71                 return min_period_bg_abstime;
  72         } else {
  73                 return min_period_abstime;
  74         }
  75 }
  76
  77 static uint64_t
  78 kperf_timer_min_pet_period_abstime(void)
  79 {
  80         if (ktrace_background_active()) {
  81                 return min_period_pet_bg_abstime;
  82         } else {
  83                 return min_period_pet_abstime;
  84         }
  85 }
  86
  87 static void
  88 kperf_timer_schedule(struct kperf_timer *timer, uint64_t now)
  89 {
  90         BUF_INFO(PERF_TM_SCHED, timer->period);
  91
  92         /* if we re-programmed the timer to zero, just drop it */
  93         if (timer->period == 0) {
  94                 return;
  95         }
  96
  97         /* calculate deadline */
  98         uint64_t deadline = now + timer->period;
  99
 100         /* re-schedule the timer, making sure we don't apply slop */
 101         timer_call_enter(&timer->tcall, deadline, TIMER_CALL_SYS_CRITICAL);
 102 }
 103
 104 static void
 105 kperf_sample_cpu(struct kperf_timer *timer, bool system_sample,
 106                 bool only_system)
 107 {
 108         assert(timer != NULL);
 109
 110         /* Always cut a tracepoint to show a sample event occurred */
 111         BUF_DATA(PERF_TM_HNDLR | DBG_FUNC_START, 0);
 112
 113         int ncpu = cpu_number();
 114
 115         struct kperf_sample *intbuf = kperf_intr_sample_buffer();
 116 #if DEVELOPMENT || DEBUG
 117         intbuf->sample_time = mach_absolute_time();
 118 #endif /* DEVELOPMENT || DEBUG */
 119
 120         /* On a timer, we can see the "real" current thread */
 121         struct kperf_context ctx = {
 122                 .cur_thread = current_thread(),
 123                 .trigger_type = TRIGGER_TYPE_TIMER,
 124                 .trigger_id = (unsigned int)(timer - kperf_timerv),
 125         };
 126         ctx.cur_pid = task_pid(get_threadtask(ctx.cur_thread));
 127
 128         if (ctx.trigger_id == pet_timer_id && ncpu < machine_info.logical_cpu_max) {
 129                 kperf_tid_on_cpus[ncpu] = thread_tid(ctx.cur_thread);
 130         }
 131
 132         /* make sure sampling is on */
 133         unsigned int status = kperf_sampling_status();
 134         if (status == KPERF_SAMPLING_OFF) {
 135                 BUF_INFO(PERF_TM_HNDLR | DBG_FUNC_END, SAMPLE_OFF);
 136                 return;
 137         } else if (status == KPERF_SAMPLING_SHUTDOWN) {
 138                 BUF_INFO(PERF_TM_HNDLR | DBG_FUNC_END, SAMPLE_SHUTDOWN);
 139                 return;
 140         }
 141
 142         /* call the action -- kernel-only from interrupt, pend user */
 143         int r = kperf_sample(intbuf, &ctx, timer->actionid,
 144                         SAMPLE_FLAG_PEND_USER | (system_sample ? SAMPLE_FLAG_SYSTEM : 0) |
 145                         (only_system ? SAMPLE_FLAG_ONLY_SYSTEM : 0));
 146
 147         /* end tracepoint is informational */
 148         BUF_INFO(PERF_TM_HNDLR | DBG_FUNC_END, r);
 149
 150         (void)atomic_fetch_and_explicit(&timer->pending_cpus,
 151                         ~(UINT64_C(1) << ncpu), memory_order_relaxed);
 152 }
 153
 154 void
 155 kperf_ipi_handler(void *param)
 156 {
 157         kperf_sample_cpu((struct kperf_timer *)param, false, false);
 158 }
 159
 160 static void
 161 kperf_timer_handler(void *param0, __unused void *param1)
 162 {
 163         struct kperf_timer *timer = param0;
 164         unsigned int ntimer = (unsigned int)(timer - kperf_timerv);
 165         unsigned int ncpus  = machine_info.logical_cpu_max;
 166         bool system_only_self = true;
 167
 168         if (timer->actionid == 0) {
 169                 return;
 170         }
 171
 172         timer->active = 1;
 173 #if DEVELOPMENT || DEBUG
 174         timer->fire_time = mach_absolute_time();
 175 #endif /* DEVELOPMENT || DEBUG */
 176
 177         /* along the lines of do not ipi if we are all shutting down */
 178         if (kperf_sampling_status() == KPERF_SAMPLING_SHUTDOWN) {
 179                 goto deactivate;
 180         }
 181
 182         BUF_DATA(PERF_TM_FIRE, ntimer, ntimer == pet_timer_id, timer->period,
 183                                timer->actionid);
 184
 185         if (ntimer == pet_timer_id) {
 186                 kperf_pet_fire_before();
 187
 188                 /* clean-up the thread-on-CPUs cache */
 189                 bzero(kperf_tid_on_cpus, ncpus * sizeof(*kperf_tid_on_cpus));
 190         }
 191
 192         /*
 193          * IPI other cores only if the action has non-system samplers.
 194          */
 195         if (kperf_sample_has_non_system(timer->actionid)) {
 196                 /*
 197                  * If the core that's handling the timer is not scheduling
 198                  * threads, only run system samplers.
 199                  */
 200                 system_only_self = kperf_mp_broadcast_other_running(timer);
 201         }
 202         kperf_sample_cpu(timer, true, system_only_self);
 203
 204         /* release the pet thread? */
 205         if (ntimer == pet_timer_id) {
 206                 /* PET mode is responsible for rearming the timer */
 207                 kperf_pet_fire_after();
 208         } else {
 209                 /*
 210                   * FIXME: Get the current time from elsewhere.  The next
 211                   * timer's period now includes the time taken to reach this
 212                   * point.  This causes a bias towards longer sampling periods
 213                   * than requested.
 214                   */
 215                 kperf_timer_schedule(timer, mach_absolute_time());
 216         }
 217
 218 deactivate:
 219         timer->active = 0;
 220 }
 221
 222 /* program the timer from the PET thread */
 223 void
 224 kperf_timer_pet_rearm(uint64_t elapsed_ticks)
 225 {
 226         struct kperf_timer *timer = NULL;
 227         uint64_t period = 0;
 228         uint64_t deadline;
 229
 230         /*
 231          * If the pet_timer_id is invalid, it has been disabled, so this should
 232          * do nothing.
 233          */
 234         if (pet_timer_id >= kperf_timerc) {
 235                 return;
 236         }
 237
 238         unsigned int status = kperf_sampling_status();
 239         /* do not reprogram the timer if it has been shutdown or sampling is off */
 240         if (status == KPERF_SAMPLING_OFF) {
 241                 BUF_INFO(PERF_PET_END, SAMPLE_OFF);
 242                 return;
 243         } else if (status == KPERF_SAMPLING_SHUTDOWN) {
 244                 BUF_INFO(PERF_PET_END, SAMPLE_SHUTDOWN);
 245                 return;
 246         }
 247
 248         timer = &(kperf_timerv[pet_timer_id]);
 249
 250         /* if we re-programmed the timer to zero, just drop it */
 251         if (!timer->period) {
 252                 return;
 253         }
 254
 255         /* subtract the time the pet sample took being careful not to underflow */
 256         if (timer->period > elapsed_ticks) {
 257                 period = timer->period - elapsed_ticks;
 258         }
 259
 260         /* make sure we don't set the next PET sample to happen too soon */
 261         if (period < min_period_pet_abstime) {
 262                 period = min_period_pet_abstime;
 263         }
 264
 265         /* we probably took so long in the PET thread, it makes sense to take
 266          * the time again.
 267          */
 268         deadline = mach_absolute_time() + period;
 269
 270         BUF_INFO(PERF_PET_SCHED, timer->period, period, elapsed_ticks, deadline);
 271
 272         /* re-schedule the timer, making sure we don't apply slop */
 273         timer_call_enter(&timer->tcall, deadline, TIMER_CALL_SYS_CRITICAL);
 274
 275         return;
 276 }
 277
 278 /* turn on all the timers */
 279 void
 280 kperf_timer_go(void)
 281 {
 282         /* get the PET thread going */
 283         if (pet_timer_id < kperf_timerc) {
 284                 kperf_pet_config(kperf_timerv[pet_timer_id].actionid);
 285         }
 286
 287         uint64_t now = mach_absolute_time();
 288
 289         for (unsigned int i = 0; i < kperf_timerc; i++) {
 290                 if (kperf_timerv[i].period == 0) {
 291                         continue;
 292                 }
 293
 294                 kperf_timer_schedule(&(kperf_timerv[i]), now);
 295         }
 296 }
 297
 298 void
 299 kperf_timer_stop(void)
 300 {
 301         for (unsigned int i = 0; i < kperf_timerc; i++) {
 302                 if (kperf_timerv[i].period == 0) {
 303                         continue;
 304                 }
 305
 306                 /* wait for the timer to stop */
 307                 while (kperf_timerv[i].active);
 308
 309                 timer_call_cancel(&kperf_timerv[i].tcall);
 310         }
 311
 312         /* wait for PET to stop, too */
 313         kperf_pet_config(0);
 314 }
 315
 316 unsigned int
 317 kperf_timer_get_petid(void)
 318 {
 319         return pet_timer_id;
 320 }
 321
 322 int
 323 kperf_timer_set_petid(unsigned int timerid)
 324 {
 325         if (timerid < kperf_timerc) {
 326                 uint64_t min_period;
 327
 328                 min_period = kperf_timer_min_pet_period_abstime();
 329                 if (kperf_timerv[timerid].period < min_period) {
 330                         kperf_timerv[timerid].period = min_period;
 331                 }
 332                 kperf_pet_config(kperf_timerv[timerid].actionid);
 333         } else {
 334                 /* clear the PET trigger if it's a bogus ID */
 335                 kperf_pet_config(0);
 336         }
 337
 338         pet_timer_id = timerid;
 339
 340         return 0;
 341 }
 342
 343 int
 344 kperf_timer_get_period(unsigned int timerid, uint64_t *period_abstime)
 345 {
 346         if (timerid >= kperf_timerc) {
 347                 return EINVAL;
 348         }
 349
 350         *period_abstime = kperf_timerv[timerid].period;
 351         return 0;
 352 }
 353
 354 int
 355 kperf_timer_set_period(unsigned int timerid, uint64_t period_abstime)
 356 {
 357         uint64_t min_period;
 358
 359         if (timerid >= kperf_timerc) {
 360                 return EINVAL;
 361         }
 362
 363         if (pet_timer_id == timerid) {
 364                 min_period = kperf_timer_min_pet_period_abstime();
 365         } else {
 366                 min_period = kperf_timer_min_period_abstime();
 367         }
 368
 369         if (period_abstime > 0 && period_abstime < min_period) {
 370                 period_abstime = min_period;
 371         }
 372
 373         kperf_timerv[timerid].period = period_abstime;
 374
 375         /* FIXME: re-program running timers? */
 376
 377         return 0;
 378 }
 379
 380 int
 381 kperf_timer_get_action(unsigned int timerid, uint32_t *action)
 382 {
 383         if (timerid >= kperf_timerc) {
 384                 return EINVAL;
 385         }
 386
 387         *action = kperf_timerv[timerid].actionid;
 388         return 0;
 389 }
 390
 391 int
 392 kperf_timer_set_action(unsigned int timerid, uint32_t action)
 393 {
 394         if (timerid >= kperf_timerc) {
 395                 return EINVAL;
 396         }
 397
 398         kperf_timerv[timerid].actionid = action;
 399         return 0;
 400 }
 401
 402 unsigned int
 403 kperf_timer_get_count(void)
 404 {
 405         return kperf_timerc;
 406 }
 407
 408 void
 409 kperf_timer_reset(void)
 410 {
 411         kperf_timer_set_petid(999);
 412         kperf_set_pet_idle_rate(KPERF_PET_DEFAULT_IDLE_RATE);
 413         kperf_set_lightweight_pet(0);
 414         for (unsigned int i = 0; i < kperf_timerc; i++) {
 415                 kperf_timerv[i].period = 0;
 416                 kperf_timerv[i].actionid = 0;
 417                 kperf_timerv[i].pending_cpus = 0;
 418         }
 419 }
 420
 421 extern int
 422 kperf_timer_set_count(unsigned int count)
 423 {
 424         struct kperf_timer *new_timerv = NULL, *old_timerv = NULL;
 425         unsigned int old_count;
 426
 427         if (min_period_abstime == 0) {
 428                 nanoseconds_to_absolutetime(KP_MIN_PERIOD_NS, &min_period_abstime);
 429                 nanoseconds_to_absolutetime(KP_MIN_PERIOD_BG_NS, &min_period_bg_abstime);
 430                 nanoseconds_to_absolutetime(KP_MIN_PERIOD_PET_NS, &min_period_pet_abstime);
 431                 nanoseconds_to_absolutetime(KP_MIN_PERIOD_PET_BG_NS,
 432                         &min_period_pet_bg_abstime);
 433                 assert(min_period_abstime > 0);
 434         }
 435
 436         if (count == kperf_timerc) {
 437                 return 0;
 438         }
 439         if (count > TIMER_MAX) {
 440                 return EINVAL;
 441         }
 442
 443         /* TODO: allow shrinking? */
 444         if (count < kperf_timerc) {
 445                 return EINVAL;
 446         }
 447
 448         /*
 449          * Make sure kperf is initialized when creating the array for the first
 450          * time.
 451          */
 452         if (kperf_timerc == 0) {
 453                 int r;
 454
 455                 /* main kperf */
 456                 if ((r = kperf_init())) {
 457                         return r;
 458                 }
 459         }
 460
 461         /*
 462          * Shut down any running timers since we will be messing with the timer
 463          * call structures.
 464          */
 465         kperf_timer_stop();
 466
 467         /* create a new array */
 468         new_timerv = kalloc_tag(count * sizeof(struct kperf_timer),
 469                 VM_KERN_MEMORY_DIAG);
 470         if (new_timerv == NULL) {
 471                 return ENOMEM;
 472         }
 473         old_timerv = kperf_timerv;
 474         old_count = kperf_timerc;
 475
 476         if (old_timerv != NULL) {
 477                 bcopy(kperf_timerv, new_timerv,
 478                         kperf_timerc * sizeof(struct kperf_timer));
 479         }
 480
 481         /* zero the new entries */
 482         bzero(&(new_timerv[kperf_timerc]),
 483                 (count - old_count) * sizeof(struct kperf_timer));
 484
 485         /* (re-)setup the timer call info for all entries */
 486         for (unsigned int i = 0; i < count; i++) {
 487                 timer_call_setup(&new_timerv[i].tcall, kperf_timer_handler, &new_timerv[i]);
 488         }
 489
 490         kperf_timerv = new_timerv;
 491         kperf_timerc = count;
 492
 493         if (old_timerv != NULL) {
 494                 kfree(old_timerv, old_count * sizeof(struct kperf_timer));
 495         }
 496
 497         return 0;
 498 }