FastMixer.cpp revision bc6ae9bd4ee93e91e212ad35be038f21797ef31e
1/* 2 * Copyright (C) 2012 The Android Open Source Project 3 * 4 * Licensed under the Apache License, Version 2.0 (the "License"); 5 * you may not use this file except in compliance with the License. 6 * You may obtain a copy of the License at 7 * 8 * http://www.apache.org/licenses/LICENSE-2.0 9 * 10 * Unless required by applicable law or agreed to in writing, software 11 * distributed under the License is distributed on an "AS IS" BASIS, 12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 13 * See the License for the specific language governing permissions and 14 * limitations under the License. 15 */ 16 17// <IMPORTANT_WARNING> 18// Design rules for threadLoop() are given in the comments at section "Fast mixer thread" of 19// StateQueue.h. In particular, avoid library and system calls except at well-known points. 20// The design rules are only for threadLoop(), and don't apply to FastMixerDumpState methods. 21// </IMPORTANT_WARNING> 22 23#define LOG_TAG "FastMixer" 24//#define LOG_NDEBUG 0 25 26#define ATRACE_TAG ATRACE_TAG_AUDIO 27 28#include "Configuration.h" 29#include <sys/atomics.h> 30#include <time.h> 31#include <utils/Log.h> 32#include <utils/Trace.h> 33#include <system/audio.h> 34#ifdef FAST_MIXER_STATISTICS 35#include <cpustats/CentralTendencyStatistics.h> 36#ifdef CPU_FREQUENCY_STATISTICS 37#include <cpustats/ThreadCpuUsage.h> 38#endif 39#endif 40#include "AudioMixer.h" 41#include "FastMixer.h" 42 43#define FAST_HOT_IDLE_NS 1000000L // 1 ms: time to sleep while hot idling 44#define FAST_DEFAULT_NS 999999999L // ~1 sec: default time to sleep 45#define MIN_WARMUP_CYCLES 2 // minimum number of loop cycles to wait for warmup 46#define MAX_WARMUP_CYCLES 10 // maximum number of loop cycles to wait for warmup 47 48#define FCC_2 2 // fixed channel count assumption 49 50namespace android { 51 52// Fast mixer thread 53bool FastMixer::threadLoop() 54{ 55 static const FastMixerState initial; 56 const FastMixerState *previous = &initial, *current = &initial; 57 FastMixerState preIdle; // copy of state before we went into idle 58 struct timespec oldTs = {0, 0}; 59 bool oldTsValid = false; 60 long slopNs = 0; // accumulated time we've woken up too early (> 0) or too late (< 0) 61 long sleepNs = -1; // -1: busy wait, 0: sched_yield, > 0: nanosleep 62 int fastTrackNames[FastMixerState::kMaxFastTracks]; // handles used by mixer to identify tracks 63 int generations[FastMixerState::kMaxFastTracks]; // last observed mFastTracks[i].mGeneration 64 unsigned i; 65 for (i = 0; i < FastMixerState::kMaxFastTracks; ++i) { 66 fastTrackNames[i] = -1; 67 generations[i] = 0; 68 } 69 NBAIO_Sink *outputSink = NULL; 70 int outputSinkGen = 0; 71 AudioMixer* mixer = NULL; 72 short *mixBuffer = NULL; 73 enum {UNDEFINED, MIXED, ZEROED} mixBufferState = UNDEFINED; 74 NBAIO_Format format = Format_Invalid; 75 unsigned sampleRate = 0; 76 int fastTracksGen = 0; 77 long periodNs = 0; // expected period; the time required to render one mix buffer 78 long underrunNs = 0; // underrun likely when write cycle is greater than this value 79 long overrunNs = 0; // overrun likely when write cycle is less than this value 80 long forceNs = 0; // if overrun detected, force the write cycle to take this much time 81 long warmupNs = 0; // warmup complete when write cycle is greater than to this value 82 FastMixerDumpState dummyDumpState, *dumpState = &dummyDumpState; 83 bool ignoreNextOverrun = true; // used to ignore initial overrun and first after an underrun 84#ifdef FAST_MIXER_STATISTICS 85 struct timespec oldLoad = {0, 0}; // previous value of clock_gettime(CLOCK_THREAD_CPUTIME_ID) 86 bool oldLoadValid = false; // whether oldLoad is valid 87 uint32_t bounds = 0; 88 bool full = false; // whether we have collected at least mSamplingN samples 89#ifdef CPU_FREQUENCY_STATISTICS 90 ThreadCpuUsage tcu; // for reading the current CPU clock frequency in kHz 91#endif 92#endif 93 unsigned coldGen = 0; // last observed mColdGen 94 bool isWarm = false; // true means ready to mix, false means wait for warmup before mixing 95 struct timespec measuredWarmupTs = {0, 0}; // how long did it take for warmup to complete 96 uint32_t warmupCycles = 0; // counter of number of loop cycles required to warmup 97 NBAIO_Sink* teeSink = NULL; // if non-NULL, then duplicate write() to this non-blocking sink 98 NBLog::Writer dummyLogWriter, *logWriter = &dummyLogWriter; 99 uint32_t totalNativeFramesWritten = 0; // copied to dumpState->mFramesWritten 100 101 // next 2 fields are valid only when timestampStatus == NO_ERROR 102 AudioTimestamp timestamp; 103 uint32_t nativeFramesWrittenButNotPresented = 0; // the = 0 is to silence the compiler 104 status_t timestampStatus = INVALID_OPERATION; 105 106 for (;;) { 107 108 // either nanosleep, sched_yield, or busy wait 109 if (sleepNs >= 0) { 110 if (sleepNs > 0) { 111 ALOG_ASSERT(sleepNs < 1000000000); 112 const struct timespec req = {0, sleepNs}; 113 nanosleep(&req, NULL); 114 } else { 115 sched_yield(); 116 } 117 } 118 // default to long sleep for next cycle 119 sleepNs = FAST_DEFAULT_NS; 120 121 // poll for state change 122 const FastMixerState *next = mSQ.poll(); 123 if (next == NULL) { 124 // continue to use the default initial state until a real state is available 125 ALOG_ASSERT(current == &initial && previous == &initial); 126 next = current; 127 } 128 129 FastMixerState::Command command = next->mCommand; 130 if (next != current) { 131 132 // As soon as possible of learning of a new dump area, start using it 133 dumpState = next->mDumpState != NULL ? next->mDumpState : &dummyDumpState; 134 teeSink = next->mTeeSink; 135 logWriter = next->mNBLogWriter != NULL ? next->mNBLogWriter : &dummyLogWriter; 136 if (mixer != NULL) { 137 mixer->setLog(logWriter); 138 } 139 140 // We want to always have a valid reference to the previous (non-idle) state. 141 // However, the state queue only guarantees access to current and previous states. 142 // So when there is a transition from a non-idle state into an idle state, we make a 143 // copy of the last known non-idle state so it is still available on return from idle. 144 // The possible transitions are: 145 // non-idle -> non-idle update previous from current in-place 146 // non-idle -> idle update previous from copy of current 147 // idle -> idle don't update previous 148 // idle -> non-idle don't update previous 149 if (!(current->mCommand & FastMixerState::IDLE)) { 150 if (command & FastMixerState::IDLE) { 151 preIdle = *current; 152 current = &preIdle; 153 oldTsValid = false; 154#ifdef FAST_MIXER_STATISTICS 155 oldLoadValid = false; 156#endif 157 ignoreNextOverrun = true; 158 } 159 previous = current; 160 } 161 current = next; 162 } 163#if !LOG_NDEBUG 164 next = NULL; // not referenced again 165#endif 166 167 dumpState->mCommand = command; 168 169 switch (command) { 170 case FastMixerState::INITIAL: 171 case FastMixerState::HOT_IDLE: 172 sleepNs = FAST_HOT_IDLE_NS; 173 continue; 174 case FastMixerState::COLD_IDLE: 175 // only perform a cold idle command once 176 // FIXME consider checking previous state and only perform if previous != COLD_IDLE 177 if (current->mColdGen != coldGen) { 178 int32_t *coldFutexAddr = current->mColdFutexAddr; 179 ALOG_ASSERT(coldFutexAddr != NULL); 180 int32_t old = android_atomic_dec(coldFutexAddr); 181 if (old <= 0) { 182 __futex_syscall4(coldFutexAddr, FUTEX_WAIT_PRIVATE, old - 1, NULL); 183 } 184 int policy = sched_getscheduler(0); 185 if (!(policy == SCHED_FIFO || policy == SCHED_RR)) { 186 ALOGE("did not receive expected priority boost"); 187 } 188 // This may be overly conservative; there could be times that the normal mixer 189 // requests such a brief cold idle that it doesn't require resetting this flag. 190 isWarm = false; 191 measuredWarmupTs.tv_sec = 0; 192 measuredWarmupTs.tv_nsec = 0; 193 warmupCycles = 0; 194 sleepNs = -1; 195 coldGen = current->mColdGen; 196#ifdef FAST_MIXER_STATISTICS 197 bounds = 0; 198 full = false; 199#endif 200 oldTsValid = !clock_gettime(CLOCK_MONOTONIC, &oldTs); 201 timestampStatus = INVALID_OPERATION; 202 } else { 203 sleepNs = FAST_HOT_IDLE_NS; 204 } 205 continue; 206 case FastMixerState::EXIT: 207 delete mixer; 208 delete[] mixBuffer; 209 return false; 210 case FastMixerState::MIX: 211 case FastMixerState::WRITE: 212 case FastMixerState::MIX_WRITE: 213 break; 214 default: 215 LOG_ALWAYS_FATAL("bad command %d", command); 216 } 217 218 // there is a non-idle state available to us; did the state change? 219 size_t frameCount = current->mFrameCount; 220 if (current != previous) { 221 222 // handle state change here, but since we want to diff the state, 223 // we're prepared for previous == &initial the first time through 224 unsigned previousTrackMask; 225 226 // check for change in output HAL configuration 227 NBAIO_Format previousFormat = format; 228 if (current->mOutputSinkGen != outputSinkGen) { 229 outputSink = current->mOutputSink; 230 outputSinkGen = current->mOutputSinkGen; 231 if (outputSink == NULL) { 232 format = Format_Invalid; 233 sampleRate = 0; 234 } else { 235 format = outputSink->format(); 236 sampleRate = Format_sampleRate(format); 237 ALOG_ASSERT(Format_channelCount(format) == FCC_2); 238 } 239 dumpState->mSampleRate = sampleRate; 240 } 241 242 if ((!Format_isEqual(format, previousFormat)) || (frameCount != previous->mFrameCount)) { 243 // FIXME to avoid priority inversion, don't delete here 244 delete mixer; 245 mixer = NULL; 246 delete[] mixBuffer; 247 mixBuffer = NULL; 248 if (frameCount > 0 && sampleRate > 0) { 249 // FIXME new may block for unbounded time at internal mutex of the heap 250 // implementation; it would be better to have normal mixer allocate for us 251 // to avoid blocking here and to prevent possible priority inversion 252 mixer = new AudioMixer(frameCount, sampleRate, FastMixerState::kMaxFastTracks); 253 mixBuffer = new short[frameCount * FCC_2]; 254 periodNs = (frameCount * 1000000000LL) / sampleRate; // 1.00 255 underrunNs = (frameCount * 1750000000LL) / sampleRate; // 1.75 256 overrunNs = (frameCount * 500000000LL) / sampleRate; // 0.50 257 forceNs = (frameCount * 950000000LL) / sampleRate; // 0.95 258 warmupNs = (frameCount * 500000000LL) / sampleRate; // 0.50 259 } else { 260 periodNs = 0; 261 underrunNs = 0; 262 overrunNs = 0; 263 forceNs = 0; 264 warmupNs = 0; 265 } 266 mixBufferState = UNDEFINED; 267#if !LOG_NDEBUG 268 for (i = 0; i < FastMixerState::kMaxFastTracks; ++i) { 269 fastTrackNames[i] = -1; 270 } 271#endif 272 // we need to reconfigure all active tracks 273 previousTrackMask = 0; 274 fastTracksGen = current->mFastTracksGen - 1; 275 dumpState->mFrameCount = frameCount; 276 } else { 277 previousTrackMask = previous->mTrackMask; 278 } 279 280 // check for change in active track set 281 unsigned currentTrackMask = current->mTrackMask; 282 dumpState->mTrackMask = currentTrackMask; 283 if (current->mFastTracksGen != fastTracksGen) { 284 ALOG_ASSERT(mixBuffer != NULL); 285 int name; 286 287 // process removed tracks first to avoid running out of track names 288 unsigned removedTracks = previousTrackMask & ~currentTrackMask; 289 while (removedTracks != 0) { 290 i = __builtin_ctz(removedTracks); 291 removedTracks &= ~(1 << i); 292 const FastTrack* fastTrack = ¤t->mFastTracks[i]; 293 ALOG_ASSERT(fastTrack->mBufferProvider == NULL); 294 if (mixer != NULL) { 295 name = fastTrackNames[i]; 296 ALOG_ASSERT(name >= 0); 297 mixer->deleteTrackName(name); 298 } 299#if !LOG_NDEBUG 300 fastTrackNames[i] = -1; 301#endif 302 // don't reset track dump state, since other side is ignoring it 303 generations[i] = fastTrack->mGeneration; 304 } 305 306 // now process added tracks 307 unsigned addedTracks = currentTrackMask & ~previousTrackMask; 308 while (addedTracks != 0) { 309 i = __builtin_ctz(addedTracks); 310 addedTracks &= ~(1 << i); 311 const FastTrack* fastTrack = ¤t->mFastTracks[i]; 312 AudioBufferProvider *bufferProvider = fastTrack->mBufferProvider; 313 ALOG_ASSERT(bufferProvider != NULL && fastTrackNames[i] == -1); 314 if (mixer != NULL) { 315 // calling getTrackName with default channel mask and a random invalid 316 // sessionId (no effects here) 317 name = mixer->getTrackName(AUDIO_CHANNEL_OUT_STEREO, -555); 318 ALOG_ASSERT(name >= 0); 319 fastTrackNames[i] = name; 320 mixer->setBufferProvider(name, bufferProvider); 321 mixer->setParameter(name, AudioMixer::TRACK, AudioMixer::MAIN_BUFFER, 322 (void *) mixBuffer); 323 // newly allocated track names default to full scale volume 324 mixer->setParameter(name, AudioMixer::TRACK, AudioMixer::CHANNEL_MASK, 325 (void *)(uintptr_t)fastTrack->mChannelMask); 326 mixer->enable(name); 327 } 328 generations[i] = fastTrack->mGeneration; 329 } 330 331 // finally process (potentially) modified tracks; these use the same slot 332 // but may have a different buffer provider or volume provider 333 unsigned modifiedTracks = currentTrackMask & previousTrackMask; 334 while (modifiedTracks != 0) { 335 i = __builtin_ctz(modifiedTracks); 336 modifiedTracks &= ~(1 << i); 337 const FastTrack* fastTrack = ¤t->mFastTracks[i]; 338 if (fastTrack->mGeneration != generations[i]) { 339 // this track was actually modified 340 AudioBufferProvider *bufferProvider = fastTrack->mBufferProvider; 341 ALOG_ASSERT(bufferProvider != NULL); 342 if (mixer != NULL) { 343 name = fastTrackNames[i]; 344 ALOG_ASSERT(name >= 0); 345 mixer->setBufferProvider(name, bufferProvider); 346 if (fastTrack->mVolumeProvider == NULL) { 347 mixer->setParameter(name, AudioMixer::VOLUME, AudioMixer::VOLUME0, 348 (void *)0x1000); 349 mixer->setParameter(name, AudioMixer::VOLUME, AudioMixer::VOLUME1, 350 (void *)0x1000); 351 } 352 mixer->setParameter(name, AudioMixer::RESAMPLE, 353 AudioMixer::REMOVE, NULL); 354 mixer->setParameter(name, AudioMixer::TRACK, AudioMixer::CHANNEL_MASK, 355 (void *)(uintptr_t) fastTrack->mChannelMask); 356 // already enabled 357 } 358 generations[i] = fastTrack->mGeneration; 359 } 360 } 361 362 fastTracksGen = current->mFastTracksGen; 363 364 dumpState->mNumTracks = popcount(currentTrackMask); 365 } 366 367#if 1 // FIXME shouldn't need this 368 // only process state change once 369 previous = current; 370#endif 371 } 372 373 // do work using current state here 374 if ((command & FastMixerState::MIX) && (mixer != NULL) && isWarm) { 375 ALOG_ASSERT(mixBuffer != NULL); 376 // for each track, update volume and check for underrun 377 unsigned currentTrackMask = current->mTrackMask; 378 while (currentTrackMask != 0) { 379 i = __builtin_ctz(currentTrackMask); 380 currentTrackMask &= ~(1 << i); 381 const FastTrack* fastTrack = ¤t->mFastTracks[i]; 382 383 // Refresh the per-track timestamp 384 if (timestampStatus == NO_ERROR) { 385 uint32_t trackFramesWrittenButNotPresented = 386 nativeFramesWrittenButNotPresented; 387 uint32_t trackFramesWritten = fastTrack->mBufferProvider->framesReleased(); 388 // Can't provide an AudioTimestamp before first frame presented, 389 // or during the brief 32-bit wraparound window 390 if (trackFramesWritten >= trackFramesWrittenButNotPresented) { 391 AudioTimestamp perTrackTimestamp; 392 perTrackTimestamp.mPosition = 393 trackFramesWritten - trackFramesWrittenButNotPresented; 394 perTrackTimestamp.mTime = timestamp.mTime; 395 fastTrack->mBufferProvider->onTimestamp(perTrackTimestamp); 396 } 397 } 398 399 int name = fastTrackNames[i]; 400 ALOG_ASSERT(name >= 0); 401 if (fastTrack->mVolumeProvider != NULL) { 402 uint32_t vlr = fastTrack->mVolumeProvider->getVolumeLR(); 403 mixer->setParameter(name, AudioMixer::VOLUME, AudioMixer::VOLUME0, 404 (void *)(uintptr_t)(vlr & 0xFFFF)); 405 mixer->setParameter(name, AudioMixer::VOLUME, AudioMixer::VOLUME1, 406 (void *)(uintptr_t)(vlr >> 16)); 407 } 408 // FIXME The current implementation of framesReady() for fast tracks 409 // takes a tryLock, which can block 410 // up to 1 ms. If enough active tracks all blocked in sequence, this would result 411 // in the overall fast mix cycle being delayed. Should use a non-blocking FIFO. 412 size_t framesReady = fastTrack->mBufferProvider->framesReady(); 413 if (ATRACE_ENABLED()) { 414 // I wish we had formatted trace names 415 char traceName[16]; 416 strcpy(traceName, "fRdy"); 417 traceName[4] = i + (i < 10 ? '0' : 'A' - 10); 418 traceName[5] = '\0'; 419 ATRACE_INT(traceName, framesReady); 420 } 421 FastTrackDump *ftDump = &dumpState->mTracks[i]; 422 FastTrackUnderruns underruns = ftDump->mUnderruns; 423 if (framesReady < frameCount) { 424 if (framesReady == 0) { 425 underruns.mBitFields.mEmpty++; 426 underruns.mBitFields.mMostRecent = UNDERRUN_EMPTY; 427 mixer->disable(name); 428 } else { 429 // allow mixing partial buffer 430 underruns.mBitFields.mPartial++; 431 underruns.mBitFields.mMostRecent = UNDERRUN_PARTIAL; 432 mixer->enable(name); 433 } 434 } else { 435 underruns.mBitFields.mFull++; 436 underruns.mBitFields.mMostRecent = UNDERRUN_FULL; 437 mixer->enable(name); 438 } 439 ftDump->mUnderruns = underruns; 440 ftDump->mFramesReady = framesReady; 441 } 442 443 int64_t pts; 444 if (outputSink == NULL || (OK != outputSink->getNextWriteTimestamp(&pts))) { 445 pts = AudioBufferProvider::kInvalidPTS; 446 } 447 448 // process() is CPU-bound 449 mixer->process(pts); 450 mixBufferState = MIXED; 451 } else if (mixBufferState == MIXED) { 452 mixBufferState = UNDEFINED; 453 } 454 bool attemptedWrite = false; 455 //bool didFullWrite = false; // dumpsys could display a count of partial writes 456 if ((command & FastMixerState::WRITE) && (outputSink != NULL) && (mixBuffer != NULL)) { 457 if (mixBufferState == UNDEFINED) { 458 memset(mixBuffer, 0, frameCount * FCC_2 * sizeof(short)); 459 mixBufferState = ZEROED; 460 } 461 if (teeSink != NULL) { 462 (void) teeSink->write(mixBuffer, frameCount); 463 } 464 // FIXME write() is non-blocking and lock-free for a properly implemented NBAIO sink, 465 // but this code should be modified to handle both non-blocking and blocking sinks 466 dumpState->mWriteSequence++; 467 ATRACE_BEGIN("write"); 468 ssize_t framesWritten = outputSink->write(mixBuffer, frameCount); 469 ATRACE_END(); 470 dumpState->mWriteSequence++; 471 if (framesWritten >= 0) { 472 ALOG_ASSERT((size_t) framesWritten <= frameCount); 473 totalNativeFramesWritten += framesWritten; 474 dumpState->mFramesWritten = totalNativeFramesWritten; 475 //if ((size_t) framesWritten == frameCount) { 476 // didFullWrite = true; 477 //} 478 } else { 479 dumpState->mWriteErrors++; 480 } 481 attemptedWrite = true; 482 // FIXME count # of writes blocked excessively, CPU usage, etc. for dump 483 484 timestampStatus = outputSink->getTimestamp(timestamp); 485 if (timestampStatus == NO_ERROR) { 486 uint32_t totalNativeFramesPresented = timestamp.mPosition; 487 if (totalNativeFramesPresented <= totalNativeFramesWritten) { 488 nativeFramesWrittenButNotPresented = 489 totalNativeFramesWritten - totalNativeFramesPresented; 490 } else { 491 // HAL reported that more frames were presented than were written 492 timestampStatus = INVALID_OPERATION; 493 } 494 } 495 } 496 497 // To be exactly periodic, compute the next sleep time based on current time. 498 // This code doesn't have long-term stability when the sink is non-blocking. 499 // FIXME To avoid drift, use the local audio clock or watch the sink's fill status. 500 struct timespec newTs; 501 int rc = clock_gettime(CLOCK_MONOTONIC, &newTs); 502 if (rc == 0) { 503 //logWriter->logTimestamp(newTs); 504 if (oldTsValid) { 505 time_t sec = newTs.tv_sec - oldTs.tv_sec; 506 long nsec = newTs.tv_nsec - oldTs.tv_nsec; 507 ALOGE_IF(sec < 0 || (sec == 0 && nsec < 0), 508 "clock_gettime(CLOCK_MONOTONIC) failed: was %ld.%09ld but now %ld.%09ld", 509 oldTs.tv_sec, oldTs.tv_nsec, newTs.tv_sec, newTs.tv_nsec); 510 if (nsec < 0) { 511 --sec; 512 nsec += 1000000000; 513 } 514 // To avoid an initial underrun on fast tracks after exiting standby, 515 // do not start pulling data from tracks and mixing until warmup is complete. 516 // Warmup is considered complete after the earlier of: 517 // MIN_WARMUP_CYCLES write() attempts and last one blocks for at least warmupNs 518 // MAX_WARMUP_CYCLES write() attempts. 519 // This is overly conservative, but to get better accuracy requires a new HAL API. 520 if (!isWarm && attemptedWrite) { 521 measuredWarmupTs.tv_sec += sec; 522 measuredWarmupTs.tv_nsec += nsec; 523 if (measuredWarmupTs.tv_nsec >= 1000000000) { 524 measuredWarmupTs.tv_sec++; 525 measuredWarmupTs.tv_nsec -= 1000000000; 526 } 527 ++warmupCycles; 528 if ((nsec > warmupNs && warmupCycles >= MIN_WARMUP_CYCLES) || 529 (warmupCycles >= MAX_WARMUP_CYCLES)) { 530 isWarm = true; 531 dumpState->mMeasuredWarmupTs = measuredWarmupTs; 532 dumpState->mWarmupCycles = warmupCycles; 533 } 534 } 535 sleepNs = -1; 536 if (isWarm) { 537 if (sec > 0 || nsec > underrunNs) { 538 ATRACE_NAME("underrun"); 539 // FIXME only log occasionally 540 ALOGV("underrun: time since last cycle %d.%03ld sec", 541 (int) sec, nsec / 1000000L); 542 dumpState->mUnderruns++; 543 ignoreNextOverrun = true; 544 } else if (nsec < overrunNs) { 545 if (ignoreNextOverrun) { 546 ignoreNextOverrun = false; 547 } else { 548 // FIXME only log occasionally 549 ALOGV("overrun: time since last cycle %d.%03ld sec", 550 (int) sec, nsec / 1000000L); 551 dumpState->mOverruns++; 552 } 553 // This forces a minimum cycle time. It: 554 // - compensates for an audio HAL with jitter due to sample rate conversion 555 // - works with a variable buffer depth audio HAL that never pulls at a 556 // rate < than overrunNs per buffer. 557 // - recovers from overrun immediately after underrun 558 // It doesn't work with a non-blocking audio HAL. 559 sleepNs = forceNs - nsec; 560 } else { 561 ignoreNextOverrun = false; 562 } 563 } 564#ifdef FAST_MIXER_STATISTICS 565 if (isWarm) { 566 // advance the FIFO queue bounds 567 size_t i = bounds & (dumpState->mSamplingN - 1); 568 bounds = (bounds & 0xFFFF0000) | ((bounds + 1) & 0xFFFF); 569 if (full) { 570 bounds += 0x10000; 571 } else if (!(bounds & (dumpState->mSamplingN - 1))) { 572 full = true; 573 } 574 // compute the delta value of clock_gettime(CLOCK_MONOTONIC) 575 uint32_t monotonicNs = nsec; 576 if (sec > 0 && sec < 4) { 577 monotonicNs += sec * 1000000000; 578 } 579 // compute raw CPU load = delta value of clock_gettime(CLOCK_THREAD_CPUTIME_ID) 580 uint32_t loadNs = 0; 581 struct timespec newLoad; 582 rc = clock_gettime(CLOCK_THREAD_CPUTIME_ID, &newLoad); 583 if (rc == 0) { 584 if (oldLoadValid) { 585 sec = newLoad.tv_sec - oldLoad.tv_sec; 586 nsec = newLoad.tv_nsec - oldLoad.tv_nsec; 587 if (nsec < 0) { 588 --sec; 589 nsec += 1000000000; 590 } 591 loadNs = nsec; 592 if (sec > 0 && sec < 4) { 593 loadNs += sec * 1000000000; 594 } 595 } else { 596 // first time through the loop 597 oldLoadValid = true; 598 } 599 oldLoad = newLoad; 600 } 601#ifdef CPU_FREQUENCY_STATISTICS 602 // get the absolute value of CPU clock frequency in kHz 603 int cpuNum = sched_getcpu(); 604 uint32_t kHz = tcu.getCpukHz(cpuNum); 605 kHz = (kHz << 4) | (cpuNum & 0xF); 606#endif 607 // save values in FIFO queues for dumpsys 608 // these stores #1, #2, #3 are not atomic with respect to each other, 609 // or with respect to store #4 below 610 dumpState->mMonotonicNs[i] = monotonicNs; 611 dumpState->mLoadNs[i] = loadNs; 612#ifdef CPU_FREQUENCY_STATISTICS 613 dumpState->mCpukHz[i] = kHz; 614#endif 615 // this store #4 is not atomic with respect to stores #1, #2, #3 above, but 616 // the newest open & oldest closed halves are atomic with respect to each other 617 dumpState->mBounds = bounds; 618 ATRACE_INT("cycle_ms", monotonicNs / 1000000); 619 ATRACE_INT("load_us", loadNs / 1000); 620 } 621#endif 622 } else { 623 // first time through the loop 624 oldTsValid = true; 625 sleepNs = periodNs; 626 ignoreNextOverrun = true; 627 } 628 oldTs = newTs; 629 } else { 630 // monotonic clock is broken 631 oldTsValid = false; 632 sleepNs = periodNs; 633 } 634 635 636 } // for (;;) 637 638 // never return 'true'; Thread::_threadLoop() locks mutex which can result in priority inversion 639} 640 641FastMixerDumpState::FastMixerDumpState( 642#ifdef FAST_MIXER_STATISTICS 643 uint32_t samplingN 644#endif 645 ) : 646 mCommand(FastMixerState::INITIAL), mWriteSequence(0), mFramesWritten(0), 647 mNumTracks(0), mWriteErrors(0), mUnderruns(0), mOverruns(0), 648 mSampleRate(0), mFrameCount(0), /* mMeasuredWarmupTs({0, 0}), */ mWarmupCycles(0), 649 mTrackMask(0) 650#ifdef FAST_MIXER_STATISTICS 651 , mSamplingN(0), mBounds(0) 652#endif 653{ 654 mMeasuredWarmupTs.tv_sec = 0; 655 mMeasuredWarmupTs.tv_nsec = 0; 656#ifdef FAST_MIXER_STATISTICS 657 increaseSamplingN(samplingN); 658#endif 659} 660 661#ifdef FAST_MIXER_STATISTICS 662void FastMixerDumpState::increaseSamplingN(uint32_t samplingN) 663{ 664 if (samplingN <= mSamplingN || samplingN > kSamplingN || roundup(samplingN) != samplingN) { 665 return; 666 } 667 uint32_t additional = samplingN - mSamplingN; 668 // sample arrays aren't accessed atomically with respect to the bounds, 669 // so clearing reduces chance for dumpsys to read random uninitialized samples 670 memset(&mMonotonicNs[mSamplingN], 0, sizeof(mMonotonicNs[0]) * additional); 671 memset(&mLoadNs[mSamplingN], 0, sizeof(mLoadNs[0]) * additional); 672#ifdef CPU_FREQUENCY_STATISTICS 673 memset(&mCpukHz[mSamplingN], 0, sizeof(mCpukHz[0]) * additional); 674#endif 675 mSamplingN = samplingN; 676} 677#endif 678 679FastMixerDumpState::~FastMixerDumpState() 680{ 681} 682 683// helper function called by qsort() 684static int compare_uint32_t(const void *pa, const void *pb) 685{ 686 uint32_t a = *(const uint32_t *)pa; 687 uint32_t b = *(const uint32_t *)pb; 688 if (a < b) { 689 return -1; 690 } else if (a > b) { 691 return 1; 692 } else { 693 return 0; 694 } 695} 696 697void FastMixerDumpState::dump(int fd) const 698{ 699 if (mCommand == FastMixerState::INITIAL) { 700 fdprintf(fd, " FastMixer not initialized\n"); 701 return; 702 } 703#define COMMAND_MAX 32 704 char string[COMMAND_MAX]; 705 switch (mCommand) { 706 case FastMixerState::INITIAL: 707 strcpy(string, "INITIAL"); 708 break; 709 case FastMixerState::HOT_IDLE: 710 strcpy(string, "HOT_IDLE"); 711 break; 712 case FastMixerState::COLD_IDLE: 713 strcpy(string, "COLD_IDLE"); 714 break; 715 case FastMixerState::EXIT: 716 strcpy(string, "EXIT"); 717 break; 718 case FastMixerState::MIX: 719 strcpy(string, "MIX"); 720 break; 721 case FastMixerState::WRITE: 722 strcpy(string, "WRITE"); 723 break; 724 case FastMixerState::MIX_WRITE: 725 strcpy(string, "MIX_WRITE"); 726 break; 727 default: 728 snprintf(string, COMMAND_MAX, "%d", mCommand); 729 break; 730 } 731 double measuredWarmupMs = (mMeasuredWarmupTs.tv_sec * 1000.0) + 732 (mMeasuredWarmupTs.tv_nsec / 1000000.0); 733 double mixPeriodSec = (double) mFrameCount / (double) mSampleRate; 734 fdprintf(fd, " FastMixer command=%s writeSequence=%u framesWritten=%u\n" 735 " numTracks=%u writeErrors=%u underruns=%u overruns=%u\n" 736 " sampleRate=%u frameCount=%zu measuredWarmup=%.3g ms, warmupCycles=%u\n" 737 " mixPeriod=%.2f ms\n", 738 string, mWriteSequence, mFramesWritten, 739 mNumTracks, mWriteErrors, mUnderruns, mOverruns, 740 mSampleRate, mFrameCount, measuredWarmupMs, mWarmupCycles, 741 mixPeriodSec * 1e3); 742#ifdef FAST_MIXER_STATISTICS 743 // find the interval of valid samples 744 uint32_t bounds = mBounds; 745 uint32_t newestOpen = bounds & 0xFFFF; 746 uint32_t oldestClosed = bounds >> 16; 747 uint32_t n = (newestOpen - oldestClosed) & 0xFFFF; 748 if (n > mSamplingN) { 749 ALOGE("too many samples %u", n); 750 n = mSamplingN; 751 } 752 // statistics for monotonic (wall clock) time, thread raw CPU load in time, CPU clock frequency, 753 // and adjusted CPU load in MHz normalized for CPU clock frequency 754 CentralTendencyStatistics wall, loadNs; 755#ifdef CPU_FREQUENCY_STATISTICS 756 CentralTendencyStatistics kHz, loadMHz; 757 uint32_t previousCpukHz = 0; 758#endif 759 // Assuming a normal distribution for cycle times, three standard deviations on either side of 760 // the mean account for 99.73% of the population. So if we take each tail to be 1/1000 of the 761 // sample set, we get 99.8% combined, or close to three standard deviations. 762 static const uint32_t kTailDenominator = 1000; 763 uint32_t *tail = n >= kTailDenominator ? new uint32_t[n] : NULL; 764 // loop over all the samples 765 for (uint32_t j = 0; j < n; ++j) { 766 size_t i = oldestClosed++ & (mSamplingN - 1); 767 uint32_t wallNs = mMonotonicNs[i]; 768 if (tail != NULL) { 769 tail[j] = wallNs; 770 } 771 wall.sample(wallNs); 772 uint32_t sampleLoadNs = mLoadNs[i]; 773 loadNs.sample(sampleLoadNs); 774#ifdef CPU_FREQUENCY_STATISTICS 775 uint32_t sampleCpukHz = mCpukHz[i]; 776 // skip bad kHz samples 777 if ((sampleCpukHz & ~0xF) != 0) { 778 kHz.sample(sampleCpukHz >> 4); 779 if (sampleCpukHz == previousCpukHz) { 780 double megacycles = (double) sampleLoadNs * (double) (sampleCpukHz >> 4) * 1e-12; 781 double adjMHz = megacycles / mixPeriodSec; // _not_ wallNs * 1e9 782 loadMHz.sample(adjMHz); 783 } 784 } 785 previousCpukHz = sampleCpukHz; 786#endif 787 } 788 if (n) { 789 fdprintf(fd, " Simple moving statistics over last %.1f seconds:\n", 790 wall.n() * mixPeriodSec); 791 fdprintf(fd, " wall clock time in ms per mix cycle:\n" 792 " mean=%.2f min=%.2f max=%.2f stddev=%.2f\n", 793 wall.mean()*1e-6, wall.minimum()*1e-6, wall.maximum()*1e-6, 794 wall.stddev()*1e-6); 795 fdprintf(fd, " raw CPU load in us per mix cycle:\n" 796 " mean=%.0f min=%.0f max=%.0f stddev=%.0f\n", 797 loadNs.mean()*1e-3, loadNs.minimum()*1e-3, loadNs.maximum()*1e-3, 798 loadNs.stddev()*1e-3); 799 } else { 800 fdprintf(fd, " No FastMixer statistics available currently\n"); 801 } 802#ifdef CPU_FREQUENCY_STATISTICS 803 fdprintf(fd, " CPU clock frequency in MHz:\n" 804 " mean=%.0f min=%.0f max=%.0f stddev=%.0f\n", 805 kHz.mean()*1e-3, kHz.minimum()*1e-3, kHz.maximum()*1e-3, kHz.stddev()*1e-3); 806 fdprintf(fd, " adjusted CPU load in MHz (i.e. normalized for CPU clock frequency):\n" 807 " mean=%.1f min=%.1f max=%.1f stddev=%.1f\n", 808 loadMHz.mean(), loadMHz.minimum(), loadMHz.maximum(), loadMHz.stddev()); 809#endif 810 if (tail != NULL) { 811 qsort(tail, n, sizeof(uint32_t), compare_uint32_t); 812 // assume same number of tail samples on each side, left and right 813 uint32_t count = n / kTailDenominator; 814 CentralTendencyStatistics left, right; 815 for (uint32_t i = 0; i < count; ++i) { 816 left.sample(tail[i]); 817 right.sample(tail[n - (i + 1)]); 818 } 819 fdprintf(fd, " Distribution of mix cycle times in ms for the tails (> ~3 stddev outliers):\n" 820 " left tail: mean=%.2f min=%.2f max=%.2f stddev=%.2f\n" 821 " right tail: mean=%.2f min=%.2f max=%.2f stddev=%.2f\n", 822 left.mean()*1e-6, left.minimum()*1e-6, left.maximum()*1e-6, left.stddev()*1e-6, 823 right.mean()*1e-6, right.minimum()*1e-6, right.maximum()*1e-6, 824 right.stddev()*1e-6); 825 delete[] tail; 826 } 827#endif 828 // The active track mask and track states are updated non-atomically. 829 // So if we relied on isActive to decide whether to display, 830 // then we might display an obsolete track or omit an active track. 831 // Instead we always display all tracks, with an indication 832 // of whether we think the track is active. 833 uint32_t trackMask = mTrackMask; 834 fdprintf(fd, " Fast tracks: kMaxFastTracks=%u activeMask=%#x\n", 835 FastMixerState::kMaxFastTracks, trackMask); 836 fdprintf(fd, " Index Active Full Partial Empty Recent Ready\n"); 837 for (uint32_t i = 0; i < FastMixerState::kMaxFastTracks; ++i, trackMask >>= 1) { 838 bool isActive = trackMask & 1; 839 const FastTrackDump *ftDump = &mTracks[i]; 840 const FastTrackUnderruns& underruns = ftDump->mUnderruns; 841 const char *mostRecent; 842 switch (underruns.mBitFields.mMostRecent) { 843 case UNDERRUN_FULL: 844 mostRecent = "full"; 845 break; 846 case UNDERRUN_PARTIAL: 847 mostRecent = "partial"; 848 break; 849 case UNDERRUN_EMPTY: 850 mostRecent = "empty"; 851 break; 852 default: 853 mostRecent = "?"; 854 break; 855 } 856 fdprintf(fd, " %5u %6s %4u %7u %5u %7s %5zu\n", i, isActive ? "yes" : "no", 857 (underruns.mBitFields.mFull) & UNDERRUN_MASK, 858 (underruns.mBitFields.mPartial) & UNDERRUN_MASK, 859 (underruns.mBitFields.mEmpty) & UNDERRUN_MASK, 860 mostRecent, ftDump->mFramesReady); 861 } 862} 863 864} // namespace android 865