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         struct kperf_context ctx;
 109
 110         assert(timer != NULL);
 111
 112         /* Always cut a tracepoint to show a sample event occurred */
 113         BUF_DATA(PERF_TM_HNDLR | DBG_FUNC_START, 0);
 114
 115         int ncpu = cpu_number();
 116
 117         struct kperf_sample *intbuf = kperf_intr_sample_buffer();
 118
 119         /* On a timer, we can see the "real" current thread */
 120         ctx.cur_thread = current_thread();
 121         ctx.cur_pid = task_pid(get_threadtask(ctx.cur_thread));
 122
 123         /* who fired */
 124         ctx.trigger_type = TRIGGER_TYPE_TIMER;
 125         ctx.trigger_id = (unsigned int)(timer - kperf_timerv);
 126
 127         if (ctx.trigger_id == pet_timer_id && ncpu < machine_info.logical_cpu_max) {
 128                 kperf_tid_on_cpus[ncpu] = thread_tid(ctx.cur_thread);
 129         }
 130
 131         /* make sure sampling is on */
 132         unsigned int status = kperf_sampling_status();
 133         if (status == KPERF_SAMPLING_OFF) {
 134                 BUF_INFO(PERF_TM_HNDLR | DBG_FUNC_END, SAMPLE_OFF);
 135                 return;
 136         } else if (status == KPERF_SAMPLING_SHUTDOWN) {
 137                 BUF_INFO(PERF_TM_HNDLR | DBG_FUNC_END, SAMPLE_SHUTDOWN);
 138                 return;
 139         }
 140
 141         /* call the action -- kernel-only from interrupt, pend user */
 142         int r = kperf_sample(intbuf, &ctx, timer->actionid,
 143                         SAMPLE_FLAG_PEND_USER | (system_sample ? SAMPLE_FLAG_SYSTEM : 0) |
 144                         (only_system ? SAMPLE_FLAG_ONLY_SYSTEM : 0));
 145
 146         /* end tracepoint is informational */
 147         BUF_INFO(PERF_TM_HNDLR | DBG_FUNC_END, r);
 148
 149         (void)atomic_fetch_and_explicit(&timer->pending_cpus,
 150                         ~(UINT64_C(1) << ncpu), memory_order_relaxed);
 151 }
 152
 153 void
 154 kperf_ipi_handler(void *param)
 155 {
 156         kperf_sample_cpu((struct kperf_timer *)param, false, false);
 157 }
 158
 159 static void
 160 kperf_timer_handler(void *param0, __unused void *param1)
 161 {
 162         struct kperf_timer *timer = param0;
 163         unsigned int ntimer = (unsigned int)(timer - kperf_timerv);
 164         unsigned int ncpus  = machine_info.logical_cpu_max;
 165         bool system_only_self = true;
 166
 167         if (timer->actionid == 0) {
 168                 return;
 169         }
 170
 171         timer->active = 1;
 172
 173         /* along the lines of do not ipi if we are all shutting down */
 174         if (kperf_sampling_status() == KPERF_SAMPLING_SHUTDOWN) {
 175                 goto deactivate;
 176         }
 177
 178         BUF_DATA(PERF_TM_FIRE, ntimer, ntimer == pet_timer_id, timer->period,
 179                                timer->actionid);
 180
 181         if (ntimer == pet_timer_id) {
 182                 kperf_pet_fire_before();
 183
 184                 /* clean-up the thread-on-CPUs cache */
 185                 bzero(kperf_tid_on_cpus, ncpus * sizeof(*kperf_tid_on_cpus));
 186         }
 187
 188         /*
 189          * IPI other cores only if the action has non-system samplers.
 190          */
 191         if (kperf_sample_has_non_system(timer->actionid)) {
 192                 /*
 193                  * If the core that's handling the timer is not scheduling
 194                  * threads, only run system samplers.
 195                  */
 196                 system_only_self = kperf_mp_broadcast_other_running(timer);
 197         }
 198         kperf_sample_cpu(timer, true, system_only_self);
 199
 200         /* release the pet thread? */
 201         if (ntimer == pet_timer_id) {
 202                 /* PET mode is responsible for rearming the timer */
 203                 kperf_pet_fire_after();
 204         } else {
 205                 /*
 206                   * FIXME: Get the current time from elsewhere.  The next
 207                   * timer's period now includes the time taken to reach this
 208                   * point.  This causes a bias towards longer sampling periods
 209                   * than requested.
 210                   */
 211                 kperf_timer_schedule(timer, mach_absolute_time());
 212         }
 213
 214 deactivate:
 215         timer->active = 0;
 216 }
 217
 218 /* program the timer from the PET thread */
 219 void
 220 kperf_timer_pet_rearm(uint64_t elapsed_ticks)
 221 {
 222         struct kperf_timer *timer = NULL;
 223         uint64_t period = 0;
 224         uint64_t deadline;
 225
 226         /*
 227          * If the pet_timer_id is invalid, it has been disabled, so this should
 228          * do nothing.
 229          */
 230         if (pet_timer_id >= kperf_timerc) {
 231                 return;
 232         }
 233
 234         unsigned int status = kperf_sampling_status();
 235         /* do not reprogram the timer if it has been shutdown or sampling is off */
 236         if (status == KPERF_SAMPLING_OFF) {
 237                 BUF_INFO(PERF_PET_END, SAMPLE_OFF);
 238                 return;
 239         } else if (status == KPERF_SAMPLING_SHUTDOWN) {
 240                 BUF_INFO(PERF_PET_END, SAMPLE_SHUTDOWN);
 241                 return;
 242         }
 243
 244         timer = &(kperf_timerv[pet_timer_id]);
 245
 246         /* if we re-programmed the timer to zero, just drop it */
 247         if (!timer->period) {
 248                 return;
 249         }
 250
 251         /* subtract the time the pet sample took being careful not to underflow */
 252         if (timer->period > elapsed_ticks) {
 253                 period = timer->period - elapsed_ticks;
 254         }
 255
 256         /* make sure we don't set the next PET sample to happen too soon */
 257         if (period < min_period_pet_abstime) {
 258                 period = min_period_pet_abstime;
 259         }
 260
 261         /* we probably took so long in the PET thread, it makes sense to take
 262          * the time again.
 263          */
 264         deadline = mach_absolute_time() + period;
 265
 266         BUF_INFO(PERF_PET_SCHED, timer->period, period, elapsed_ticks, deadline);
 267
 268         /* re-schedule the timer, making sure we don't apply slop */
 269         timer_call_enter(&timer->tcall, deadline, TIMER_CALL_SYS_CRITICAL);
 270
 271         return;
 272 }
 273
 274 /* turn on all the timers */
 275 void
 276 kperf_timer_go(void)
 277 {
 278         /* get the PET thread going */
 279         if (pet_timer_id < kperf_timerc) {
 280                 kperf_pet_config(kperf_timerv[pet_timer_id].actionid);
 281         }
 282
 283         uint64_t now = mach_absolute_time();
 284
 285         for (unsigned int i = 0; i < kperf_timerc; i++) {
 286                 if (kperf_timerv[i].period == 0) {
 287                         continue;
 288                 }
 289
 290                 kperf_timer_schedule(&(kperf_timerv[i]), now);
 291         }
 292 }
 293
 294 void
 295 kperf_timer_stop(void)
 296 {
 297         for (unsigned int i = 0; i < kperf_timerc; i++) {
 298                 if (kperf_timerv[i].period == 0) {
 299                         continue;
 300                 }
 301
 302                 /* wait for the timer to stop */
 303                 while (kperf_timerv[i].active);
 304
 305                 timer_call_cancel(&kperf_timerv[i].tcall);
 306         }
 307
 308         /* wait for PET to stop, too */
 309         kperf_pet_config(0);
 310 }
 311
 312 unsigned int
 313 kperf_timer_get_petid(void)
 314 {
 315         return pet_timer_id;
 316 }
 317
 318 int
 319 kperf_timer_set_petid(unsigned int timerid)
 320 {
 321         if (timerid < kperf_timerc) {
 322                 uint64_t min_period;
 323
 324                 min_period = kperf_timer_min_pet_period_abstime();
 325                 if (kperf_timerv[timerid].period < min_period) {
 326                         kperf_timerv[timerid].period = min_period;
 327                 }
 328                 kperf_pet_config(kperf_timerv[timerid].actionid);
 329         } else {
 330                 /* clear the PET trigger if it's a bogus ID */
 331                 kperf_pet_config(0);
 332         }
 333
 334         pet_timer_id = timerid;
 335
 336         return 0;
 337 }
 338
 339 int
 340 kperf_timer_get_period(unsigned int timerid, uint64_t *period_abstime)
 341 {
 342         if (timerid >= kperf_timerc) {
 343                 return EINVAL;
 344         }
 345
 346         *period_abstime = kperf_timerv[timerid].period;
 347         return 0;
 348 }
 349
 350 int
 351 kperf_timer_set_period(unsigned int timerid, uint64_t period_abstime)
 352 {
 353         uint64_t min_period;
 354
 355         if (timerid >= kperf_timerc) {
 356                 return EINVAL;
 357         }
 358
 359         if (pet_timer_id == timerid) {
 360                 min_period = kperf_timer_min_pet_period_abstime();
 361         } else {
 362                 min_period = kperf_timer_min_period_abstime();
 363         }
 364
 365         if (period_abstime > 0 && period_abstime < min_period) {
 366                 period_abstime = min_period;
 367         }
 368
 369         kperf_timerv[timerid].period = period_abstime;
 370
 371         /* FIXME: re-program running timers? */
 372
 373         return 0;
 374 }
 375
 376 int
 377 kperf_timer_get_action(unsigned int timerid, uint32_t *action)
 378 {
 379         if (timerid >= kperf_timerc) {
 380                 return EINVAL;
 381         }
 382
 383         *action = kperf_timerv[timerid].actionid;
 384         return 0;
 385 }
 386
 387 int
 388 kperf_timer_set_action(unsigned int timerid, uint32_t action)
 389 {
 390         if (timerid >= kperf_timerc) {
 391                 return EINVAL;
 392         }
 393
 394         kperf_timerv[timerid].actionid = action;
 395         return 0;
 396 }
 397
 398 unsigned int
 399 kperf_timer_get_count(void)
 400 {
 401         return kperf_timerc;
 402 }
 403
 404 void
 405 kperf_timer_reset(void)
 406 {
 407         kperf_timer_set_petid(999);
 408         kperf_set_pet_idle_rate(KPERF_PET_DEFAULT_IDLE_RATE);
 409         kperf_set_lightweight_pet(0);
 410         for (unsigned int i = 0; i < kperf_timerc; i++) {
 411                 kperf_timerv[i].period = 0;
 412                 kperf_timerv[i].actionid = 0;
 413                 kperf_timerv[i].pending_cpus = 0;
 414         }
 415 }
 416
 417 extern int
 418 kperf_timer_set_count(unsigned int count)
 419 {
 420         struct kperf_timer *new_timerv = NULL, *old_timerv = NULL;
 421         unsigned int old_count;
 422
 423         if (min_period_abstime == 0) {
 424                 nanoseconds_to_absolutetime(KP_MIN_PERIOD_NS, &min_period_abstime);
 425                 nanoseconds_to_absolutetime(KP_MIN_PERIOD_BG_NS, &min_period_bg_abstime);
 426                 nanoseconds_to_absolutetime(KP_MIN_PERIOD_PET_NS, &min_period_pet_abstime);
 427                 nanoseconds_to_absolutetime(KP_MIN_PERIOD_PET_BG_NS,
 428                         &min_period_pet_bg_abstime);
 429                 assert(min_period_abstime > 0);
 430         }
 431
 432         if (count == kperf_timerc) {
 433                 return 0;
 434         }
 435         if (count > TIMER_MAX) {
 436                 return EINVAL;
 437         }
 438
 439         /* TODO: allow shrinking? */
 440         if (count < kperf_timerc) {
 441                 return EINVAL;
 442         }
 443
 444         /*
 445          * Make sure kperf is initialized when creating the array for the first
 446          * time.
 447          */
 448         if (kperf_timerc == 0) {
 449                 int r;
 450
 451                 /* main kperf */
 452                 if ((r = kperf_init())) {
 453                         return r;
 454                 }
 455         }
 456
 457         /*
 458          * Shut down any running timers since we will be messing with the timer
 459          * call structures.
 460          */
 461         kperf_timer_stop();
 462
 463         /* create a new array */
 464         new_timerv = kalloc_tag(count * sizeof(struct kperf_timer),
 465                 VM_KERN_MEMORY_DIAG);
 466         if (new_timerv == NULL) {
 467                 return ENOMEM;
 468         }
 469         old_timerv = kperf_timerv;
 470         old_count = kperf_timerc;
 471
 472         if (old_timerv != NULL) {
 473                 bcopy(kperf_timerv, new_timerv,
 474                         kperf_timerc * sizeof(struct kperf_timer));
 475         }
 476
 477         /* zero the new entries */
 478         bzero(&(new_timerv[kperf_timerc]),
 479                 (count - old_count) * sizeof(struct kperf_timer));
 480
 481         /* (re-)setup the timer call info for all entries */
 482         for (unsigned int i = 0; i < count; i++) {
 483                 timer_call_setup(&new_timerv[i].tcall, kperf_timer_handler, &new_timerv[i]);
 484         }
 485
 486         kperf_timerv = new_timerv;
 487         kperf_timerc = count;
 488
 489         if (old_timerv != NULL) {
 490                 kfree(old_timerv, old_count * sizeof(struct kperf_timer));
 491         }
 492
 493         return 0;
 494 }