FastMixer.cpp revision b220884bf3129253cc5bc8d030bc475411ea4911
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_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 if (fastTrack->mSampleRate != 0 && fastTrack->mSampleRate != sampleRate) { 325 mixer->setParameter(name, AudioMixer::RESAMPLE, 326 AudioMixer::SAMPLE_RATE, (void*) fastTrack->mSampleRate); 327 } 328 mixer->setParameter(name, AudioMixer::TRACK, AudioMixer::CHANNEL_MASK, 329 (void *) fastTrack->mChannelMask); 330 mixer->enable(name); 331 } 332 generations[i] = fastTrack->mGeneration; 333 } 334 335 // finally process (potentially) modified tracks; these use the same slot 336 // but may have a different buffer provider or volume provider 337 unsigned modifiedTracks = currentTrackMask & previousTrackMask; 338 while (modifiedTracks != 0) { 339 i = __builtin_ctz(modifiedTracks); 340 modifiedTracks &= ~(1 << i); 341 const FastTrack* fastTrack = ¤t->mFastTracks[i]; 342 if (fastTrack->mGeneration != generations[i]) { 343 // this track was actually modified 344 AudioBufferProvider *bufferProvider = fastTrack->mBufferProvider; 345 ALOG_ASSERT(bufferProvider != NULL); 346 if (mixer != NULL) { 347 name = fastTrackNames[i]; 348 ALOG_ASSERT(name >= 0); 349 mixer->setBufferProvider(name, bufferProvider); 350 if (fastTrack->mVolumeProvider == NULL) { 351 mixer->setParameter(name, AudioMixer::VOLUME, AudioMixer::VOLUME0, 352 (void *)0x1000); 353 mixer->setParameter(name, AudioMixer::VOLUME, AudioMixer::VOLUME1, 354 (void *)0x1000); 355 } 356 if (fastTrack->mSampleRate != 0 && 357 fastTrack->mSampleRate != sampleRate) { 358 mixer->setParameter(name, AudioMixer::RESAMPLE, 359 AudioMixer::SAMPLE_RATE, (void*) fastTrack->mSampleRate); 360 } else { 361 mixer->setParameter(name, AudioMixer::RESAMPLE, 362 AudioMixer::REMOVE, NULL); 363 } 364 mixer->setParameter(name, AudioMixer::TRACK, AudioMixer::CHANNEL_MASK, 365 (void *) fastTrack->mChannelMask); 366 // already enabled 367 } 368 generations[i] = fastTrack->mGeneration; 369 } 370 } 371 372 fastTracksGen = current->mFastTracksGen; 373 374 dumpState->mNumTracks = popcount(currentTrackMask); 375 } 376 377#if 1 // FIXME shouldn't need this 378 // only process state change once 379 previous = current; 380#endif 381 } 382 383 // do work using current state here 384 if ((command & FastMixerState::MIX) && (mixer != NULL) && isWarm) { 385 ALOG_ASSERT(mixBuffer != NULL); 386 // for each track, update volume and check for underrun 387 unsigned currentTrackMask = current->mTrackMask; 388 while (currentTrackMask != 0) { 389 i = __builtin_ctz(currentTrackMask); 390 currentTrackMask &= ~(1 << i); 391 const FastTrack* fastTrack = ¤t->mFastTracks[i]; 392 393 // Refresh the per-track timestamp 394 if (timestampStatus == NO_ERROR) { 395 uint32_t trackFramesWrittenButNotPresented; 396 uint32_t trackSampleRate = fastTrack->mSampleRate; 397 // There is currently no sample rate conversion for fast tracks currently 398 if (trackSampleRate != 0 && trackSampleRate != sampleRate) { 399 trackFramesWrittenButNotPresented = 400 ((int64_t) nativeFramesWrittenButNotPresented * trackSampleRate) / 401 sampleRate; 402 } else { 403 trackFramesWrittenButNotPresented = nativeFramesWrittenButNotPresented; 404 } 405 uint32_t trackFramesWritten = fastTrack->mBufferProvider->framesReleased(); 406 // Can't provide an AudioTimestamp before first frame presented, 407 // or during the brief 32-bit wraparound window 408 if (trackFramesWritten >= trackFramesWrittenButNotPresented) { 409 AudioTimestamp perTrackTimestamp; 410 perTrackTimestamp.mPosition = 411 trackFramesWritten - trackFramesWrittenButNotPresented; 412 perTrackTimestamp.mTime = timestamp.mTime; 413 fastTrack->mBufferProvider->onTimestamp(perTrackTimestamp); 414 } 415 } 416 417 int name = fastTrackNames[i]; 418 ALOG_ASSERT(name >= 0); 419 if (fastTrack->mVolumeProvider != NULL) { 420 uint32_t vlr = fastTrack->mVolumeProvider->getVolumeLR(); 421 mixer->setParameter(name, AudioMixer::VOLUME, AudioMixer::VOLUME0, 422 (void *)(vlr & 0xFFFF)); 423 mixer->setParameter(name, AudioMixer::VOLUME, AudioMixer::VOLUME1, 424 (void *)(vlr >> 16)); 425 } 426 // FIXME The current implementation of framesReady() for fast tracks 427 // takes a tryLock, which can block 428 // up to 1 ms. If enough active tracks all blocked in sequence, this would result 429 // in the overall fast mix cycle being delayed. Should use a non-blocking FIFO. 430 size_t framesReady = fastTrack->mBufferProvider->framesReady(); 431 if (ATRACE_ENABLED()) { 432 // I wish we had formatted trace names 433 char traceName[16]; 434 strcpy(traceName, "fRdy"); 435 traceName[4] = i + (i < 10 ? '0' : 'A' - 10); 436 traceName[5] = '\0'; 437 ATRACE_INT(traceName, framesReady); 438 } 439 FastTrackDump *ftDump = &dumpState->mTracks[i]; 440 FastTrackUnderruns underruns = ftDump->mUnderruns; 441 if (framesReady < frameCount) { 442 if (framesReady == 0) { 443 underruns.mBitFields.mEmpty++; 444 underruns.mBitFields.mMostRecent = UNDERRUN_EMPTY; 445 mixer->disable(name); 446 } else { 447 // allow mixing partial buffer 448 underruns.mBitFields.mPartial++; 449 underruns.mBitFields.mMostRecent = UNDERRUN_PARTIAL; 450 mixer->enable(name); 451 } 452 } else { 453 underruns.mBitFields.mFull++; 454 underruns.mBitFields.mMostRecent = UNDERRUN_FULL; 455 mixer->enable(name); 456 } 457 ftDump->mUnderruns = underruns; 458 ftDump->mFramesReady = framesReady; 459 } 460 461 int64_t pts; 462 if (outputSink == NULL || (OK != outputSink->getNextWriteTimestamp(&pts))) { 463 pts = AudioBufferProvider::kInvalidPTS; 464 } 465 466 // process() is CPU-bound 467 mixer->process(pts); 468 mixBufferState = MIXED; 469 } else if (mixBufferState == MIXED) { 470 mixBufferState = UNDEFINED; 471 } 472 bool attemptedWrite = false; 473 //bool didFullWrite = false; // dumpsys could display a count of partial writes 474 if ((command & FastMixerState::WRITE) && (outputSink != NULL) && (mixBuffer != NULL)) { 475 if (mixBufferState == UNDEFINED) { 476 memset(mixBuffer, 0, frameCount * FCC_2 * sizeof(short)); 477 mixBufferState = ZEROED; 478 } 479 if (teeSink != NULL) { 480 (void) teeSink->write(mixBuffer, frameCount); 481 } 482 // FIXME write() is non-blocking and lock-free for a properly implemented NBAIO sink, 483 // but this code should be modified to handle both non-blocking and blocking sinks 484 dumpState->mWriteSequence++; 485 ATRACE_BEGIN("write"); 486 ssize_t framesWritten = outputSink->write(mixBuffer, frameCount); 487 ATRACE_END(); 488 dumpState->mWriteSequence++; 489 if (framesWritten >= 0) { 490 ALOG_ASSERT((size_t) framesWritten <= frameCount); 491 totalNativeFramesWritten += framesWritten; 492 dumpState->mFramesWritten = totalNativeFramesWritten; 493 //if ((size_t) framesWritten == frameCount) { 494 // didFullWrite = true; 495 //} 496 } else { 497 dumpState->mWriteErrors++; 498 } 499 attemptedWrite = true; 500 // FIXME count # of writes blocked excessively, CPU usage, etc. for dump 501 502 timestampStatus = outputSink->getTimestamp(timestamp); 503 if (timestampStatus == NO_ERROR) { 504 uint32_t totalNativeFramesPresented = timestamp.mPosition; 505 if (totalNativeFramesPresented <= totalNativeFramesWritten) { 506 nativeFramesWrittenButNotPresented = 507 totalNativeFramesWritten - totalNativeFramesPresented; 508 } else { 509 // HAL reported that more frames were presented than were written 510 timestampStatus = INVALID_OPERATION; 511 } 512 } 513 } 514 515 // To be exactly periodic, compute the next sleep time based on current time. 516 // This code doesn't have long-term stability when the sink is non-blocking. 517 // FIXME To avoid drift, use the local audio clock or watch the sink's fill status. 518 struct timespec newTs; 519 int rc = clock_gettime(CLOCK_MONOTONIC, &newTs); 520 if (rc == 0) { 521 //logWriter->logTimestamp(newTs); 522 if (oldTsValid) { 523 time_t sec = newTs.tv_sec - oldTs.tv_sec; 524 long nsec = newTs.tv_nsec - oldTs.tv_nsec; 525 ALOGE_IF(sec < 0 || (sec == 0 && nsec < 0), 526 "clock_gettime(CLOCK_MONOTONIC) failed: was %ld.%09ld but now %ld.%09ld", 527 oldTs.tv_sec, oldTs.tv_nsec, newTs.tv_sec, newTs.tv_nsec); 528 if (nsec < 0) { 529 --sec; 530 nsec += 1000000000; 531 } 532 // To avoid an initial underrun on fast tracks after exiting standby, 533 // do not start pulling data from tracks and mixing until warmup is complete. 534 // Warmup is considered complete after the earlier of: 535 // MIN_WARMUP_CYCLES write() attempts and last one blocks for at least warmupNs 536 // MAX_WARMUP_CYCLES write() attempts. 537 // This is overly conservative, but to get better accuracy requires a new HAL API. 538 if (!isWarm && attemptedWrite) { 539 measuredWarmupTs.tv_sec += sec; 540 measuredWarmupTs.tv_nsec += nsec; 541 if (measuredWarmupTs.tv_nsec >= 1000000000) { 542 measuredWarmupTs.tv_sec++; 543 measuredWarmupTs.tv_nsec -= 1000000000; 544 } 545 ++warmupCycles; 546 if ((nsec > warmupNs && warmupCycles >= MIN_WARMUP_CYCLES) || 547 (warmupCycles >= MAX_WARMUP_CYCLES)) { 548 isWarm = true; 549 dumpState->mMeasuredWarmupTs = measuredWarmupTs; 550 dumpState->mWarmupCycles = warmupCycles; 551 } 552 } 553 sleepNs = -1; 554 if (isWarm) { 555 if (sec > 0 || nsec > underrunNs) { 556 ATRACE_NAME("underrun"); 557 // FIXME only log occasionally 558 ALOGV("underrun: time since last cycle %d.%03ld sec", 559 (int) sec, nsec / 1000000L); 560 dumpState->mUnderruns++; 561 ignoreNextOverrun = true; 562 } else if (nsec < overrunNs) { 563 if (ignoreNextOverrun) { 564 ignoreNextOverrun = false; 565 } else { 566 // FIXME only log occasionally 567 ALOGV("overrun: time since last cycle %d.%03ld sec", 568 (int) sec, nsec / 1000000L); 569 dumpState->mOverruns++; 570 } 571 // This forces a minimum cycle time. It: 572 // - compensates for an audio HAL with jitter due to sample rate conversion 573 // - works with a variable buffer depth audio HAL that never pulls at a 574 // rate < than overrunNs per buffer. 575 // - recovers from overrun immediately after underrun 576 // It doesn't work with a non-blocking audio HAL. 577 sleepNs = forceNs - nsec; 578 } else { 579 ignoreNextOverrun = false; 580 } 581 } 582#ifdef FAST_MIXER_STATISTICS 583 if (isWarm) { 584 // advance the FIFO queue bounds 585 size_t i = bounds & (dumpState->mSamplingN - 1); 586 bounds = (bounds & 0xFFFF0000) | ((bounds + 1) & 0xFFFF); 587 if (full) { 588 bounds += 0x10000; 589 } else if (!(bounds & (dumpState->mSamplingN - 1))) { 590 full = true; 591 } 592 // compute the delta value of clock_gettime(CLOCK_MONOTONIC) 593 uint32_t monotonicNs = nsec; 594 if (sec > 0 && sec < 4) { 595 monotonicNs += sec * 1000000000; 596 } 597 // compute raw CPU load = delta value of clock_gettime(CLOCK_THREAD_CPUTIME_ID) 598 uint32_t loadNs = 0; 599 struct timespec newLoad; 600 rc = clock_gettime(CLOCK_THREAD_CPUTIME_ID, &newLoad); 601 if (rc == 0) { 602 if (oldLoadValid) { 603 sec = newLoad.tv_sec - oldLoad.tv_sec; 604 nsec = newLoad.tv_nsec - oldLoad.tv_nsec; 605 if (nsec < 0) { 606 --sec; 607 nsec += 1000000000; 608 } 609 loadNs = nsec; 610 if (sec > 0 && sec < 4) { 611 loadNs += sec * 1000000000; 612 } 613 } else { 614 // first time through the loop 615 oldLoadValid = true; 616 } 617 oldLoad = newLoad; 618 } 619#ifdef CPU_FREQUENCY_STATISTICS 620 // get the absolute value of CPU clock frequency in kHz 621 int cpuNum = sched_getcpu(); 622 uint32_t kHz = tcu.getCpukHz(cpuNum); 623 kHz = (kHz << 4) | (cpuNum & 0xF); 624#endif 625 // save values in FIFO queues for dumpsys 626 // these stores #1, #2, #3 are not atomic with respect to each other, 627 // or with respect to store #4 below 628 dumpState->mMonotonicNs[i] = monotonicNs; 629 dumpState->mLoadNs[i] = loadNs; 630#ifdef CPU_FREQUENCY_STATISTICS 631 dumpState->mCpukHz[i] = kHz; 632#endif 633 // this store #4 is not atomic with respect to stores #1, #2, #3 above, but 634 // the newest open & oldest closed halves are atomic with respect to each other 635 dumpState->mBounds = bounds; 636 ATRACE_INT("cycle_ms", monotonicNs / 1000000); 637 ATRACE_INT("load_us", loadNs / 1000); 638 } 639#endif 640 } else { 641 // first time through the loop 642 oldTsValid = true; 643 sleepNs = periodNs; 644 ignoreNextOverrun = true; 645 } 646 oldTs = newTs; 647 } else { 648 // monotonic clock is broken 649 oldTsValid = false; 650 sleepNs = periodNs; 651 } 652 653 654 } // for (;;) 655 656 // never return 'true'; Thread::_threadLoop() locks mutex which can result in priority inversion 657} 658 659FastMixerDumpState::FastMixerDumpState( 660#ifdef FAST_MIXER_STATISTICS 661 uint32_t samplingN 662#endif 663 ) : 664 mCommand(FastMixerState::INITIAL), mWriteSequence(0), mFramesWritten(0), 665 mNumTracks(0), mWriteErrors(0), mUnderruns(0), mOverruns(0), 666 mSampleRate(0), mFrameCount(0), /* mMeasuredWarmupTs({0, 0}), */ mWarmupCycles(0), 667 mTrackMask(0) 668#ifdef FAST_MIXER_STATISTICS 669 , mSamplingN(0), mBounds(0) 670#endif 671{ 672 mMeasuredWarmupTs.tv_sec = 0; 673 mMeasuredWarmupTs.tv_nsec = 0; 674#ifdef FAST_MIXER_STATISTICS 675 increaseSamplingN(samplingN); 676#endif 677} 678 679#ifdef FAST_MIXER_STATISTICS 680void FastMixerDumpState::increaseSamplingN(uint32_t samplingN) 681{ 682 if (samplingN <= mSamplingN || samplingN > kSamplingN || roundup(samplingN) != samplingN) { 683 return; 684 } 685 uint32_t additional = samplingN - mSamplingN; 686 // sample arrays aren't accessed atomically with respect to the bounds, 687 // so clearing reduces chance for dumpsys to read random uninitialized samples 688 memset(&mMonotonicNs[mSamplingN], 0, sizeof(mMonotonicNs[0]) * additional); 689 memset(&mLoadNs[mSamplingN], 0, sizeof(mLoadNs[0]) * additional); 690#ifdef CPU_FREQUENCY_STATISTICS 691 memset(&mCpukHz[mSamplingN], 0, sizeof(mCpukHz[0]) * additional); 692#endif 693 mSamplingN = samplingN; 694} 695#endif 696 697FastMixerDumpState::~FastMixerDumpState() 698{ 699} 700 701// helper function called by qsort() 702static int compare_uint32_t(const void *pa, const void *pb) 703{ 704 uint32_t a = *(const uint32_t *)pa; 705 uint32_t b = *(const uint32_t *)pb; 706 if (a < b) { 707 return -1; 708 } else if (a > b) { 709 return 1; 710 } else { 711 return 0; 712 } 713} 714 715void FastMixerDumpState::dump(int fd) const 716{ 717 if (mCommand == FastMixerState::INITIAL) { 718 fdprintf(fd, " FastMixer not initialized\n"); 719 return; 720 } 721#define COMMAND_MAX 32 722 char string[COMMAND_MAX]; 723 switch (mCommand) { 724 case FastMixerState::INITIAL: 725 strcpy(string, "INITIAL"); 726 break; 727 case FastMixerState::HOT_IDLE: 728 strcpy(string, "HOT_IDLE"); 729 break; 730 case FastMixerState::COLD_IDLE: 731 strcpy(string, "COLD_IDLE"); 732 break; 733 case FastMixerState::EXIT: 734 strcpy(string, "EXIT"); 735 break; 736 case FastMixerState::MIX: 737 strcpy(string, "MIX"); 738 break; 739 case FastMixerState::WRITE: 740 strcpy(string, "WRITE"); 741 break; 742 case FastMixerState::MIX_WRITE: 743 strcpy(string, "MIX_WRITE"); 744 break; 745 default: 746 snprintf(string, COMMAND_MAX, "%d", mCommand); 747 break; 748 } 749 double measuredWarmupMs = (mMeasuredWarmupTs.tv_sec * 1000.0) + 750 (mMeasuredWarmupTs.tv_nsec / 1000000.0); 751 double mixPeriodSec = (double) mFrameCount / (double) mSampleRate; 752 fdprintf(fd, " FastMixer command=%s writeSequence=%u framesWritten=%u\n" 753 " numTracks=%u writeErrors=%u underruns=%u overruns=%u\n" 754 " sampleRate=%u frameCount=%u measuredWarmup=%.3g ms, warmupCycles=%u\n" 755 " mixPeriod=%.2f ms\n", 756 string, mWriteSequence, mFramesWritten, 757 mNumTracks, mWriteErrors, mUnderruns, mOverruns, 758 mSampleRate, mFrameCount, measuredWarmupMs, mWarmupCycles, 759 mixPeriodSec * 1e3); 760#ifdef FAST_MIXER_STATISTICS 761 // find the interval of valid samples 762 uint32_t bounds = mBounds; 763 uint32_t newestOpen = bounds & 0xFFFF; 764 uint32_t oldestClosed = bounds >> 16; 765 uint32_t n = (newestOpen - oldestClosed) & 0xFFFF; 766 if (n > mSamplingN) { 767 ALOGE("too many samples %u", n); 768 n = mSamplingN; 769 } 770 // statistics for monotonic (wall clock) time, thread raw CPU load in time, CPU clock frequency, 771 // and adjusted CPU load in MHz normalized for CPU clock frequency 772 CentralTendencyStatistics wall, loadNs; 773#ifdef CPU_FREQUENCY_STATISTICS 774 CentralTendencyStatistics kHz, loadMHz; 775 uint32_t previousCpukHz = 0; 776#endif 777 // Assuming a normal distribution for cycle times, three standard deviations on either side of 778 // the mean account for 99.73% of the population. So if we take each tail to be 1/1000 of the 779 // sample set, we get 99.8% combined, or close to three standard deviations. 780 static const uint32_t kTailDenominator = 1000; 781 uint32_t *tail = n >= kTailDenominator ? new uint32_t[n] : NULL; 782 // loop over all the samples 783 for (uint32_t j = 0; j < n; ++j) { 784 size_t i = oldestClosed++ & (mSamplingN - 1); 785 uint32_t wallNs = mMonotonicNs[i]; 786 if (tail != NULL) { 787 tail[j] = wallNs; 788 } 789 wall.sample(wallNs); 790 uint32_t sampleLoadNs = mLoadNs[i]; 791 loadNs.sample(sampleLoadNs); 792#ifdef CPU_FREQUENCY_STATISTICS 793 uint32_t sampleCpukHz = mCpukHz[i]; 794 // skip bad kHz samples 795 if ((sampleCpukHz & ~0xF) != 0) { 796 kHz.sample(sampleCpukHz >> 4); 797 if (sampleCpukHz == previousCpukHz) { 798 double megacycles = (double) sampleLoadNs * (double) (sampleCpukHz >> 4) * 1e-12; 799 double adjMHz = megacycles / mixPeriodSec; // _not_ wallNs * 1e9 800 loadMHz.sample(adjMHz); 801 } 802 } 803 previousCpukHz = sampleCpukHz; 804#endif 805 } 806 if (n) { 807 fdprintf(fd, " Simple moving statistics over last %.1f seconds:\n", 808 wall.n() * mixPeriodSec); 809 fdprintf(fd, " wall clock time in ms per mix cycle:\n" 810 " mean=%.2f min=%.2f max=%.2f stddev=%.2f\n", 811 wall.mean()*1e-6, wall.minimum()*1e-6, wall.maximum()*1e-6, 812 wall.stddev()*1e-6); 813 fdprintf(fd, " raw CPU load in us per mix cycle:\n" 814 " mean=%.0f min=%.0f max=%.0f stddev=%.0f\n", 815 loadNs.mean()*1e-3, loadNs.minimum()*1e-3, loadNs.maximum()*1e-3, 816 loadNs.stddev()*1e-3); 817 } else { 818 fdprintf(fd, " No FastMixer statistics available currently\n"); 819 } 820#ifdef CPU_FREQUENCY_STATISTICS 821 fdprintf(fd, " CPU clock frequency in MHz:\n" 822 " mean=%.0f min=%.0f max=%.0f stddev=%.0f\n", 823 kHz.mean()*1e-3, kHz.minimum()*1e-3, kHz.maximum()*1e-3, kHz.stddev()*1e-3); 824 fdprintf(fd, " adjusted CPU load in MHz (i.e. normalized for CPU clock frequency):\n" 825 " mean=%.1f min=%.1f max=%.1f stddev=%.1f\n", 826 loadMHz.mean(), loadMHz.minimum(), loadMHz.maximum(), loadMHz.stddev()); 827#endif 828 if (tail != NULL) { 829 qsort(tail, n, sizeof(uint32_t), compare_uint32_t); 830 // assume same number of tail samples on each side, left and right 831 uint32_t count = n / kTailDenominator; 832 CentralTendencyStatistics left, right; 833 for (uint32_t i = 0; i < count; ++i) { 834 left.sample(tail[i]); 835 right.sample(tail[n - (i + 1)]); 836 } 837 fdprintf(fd, " Distribution of mix cycle times in ms for the tails (> ~3 stddev outliers):\n" 838 " left tail: mean=%.2f min=%.2f max=%.2f stddev=%.2f\n" 839 " right tail: mean=%.2f min=%.2f max=%.2f stddev=%.2f\n", 840 left.mean()*1e-6, left.minimum()*1e-6, left.maximum()*1e-6, left.stddev()*1e-6, 841 right.mean()*1e-6, right.minimum()*1e-6, right.maximum()*1e-6, 842 right.stddev()*1e-6); 843 delete[] tail; 844 } 845#endif 846 // The active track mask and track states are updated non-atomically. 847 // So if we relied on isActive to decide whether to display, 848 // then we might display an obsolete track or omit an active track. 849 // Instead we always display all tracks, with an indication 850 // of whether we think the track is active. 851 uint32_t trackMask = mTrackMask; 852 fdprintf(fd, " Fast tracks: kMaxFastTracks=%u activeMask=%#x\n", 853 FastMixerState::kMaxFastTracks, trackMask); 854 fdprintf(fd, " Index Active Full Partial Empty Recent Ready\n"); 855 for (uint32_t i = 0; i < FastMixerState::kMaxFastTracks; ++i, trackMask >>= 1) { 856 bool isActive = trackMask & 1; 857 const FastTrackDump *ftDump = &mTracks[i]; 858 const FastTrackUnderruns& underruns = ftDump->mUnderruns; 859 const char *mostRecent; 860 switch (underruns.mBitFields.mMostRecent) { 861 case UNDERRUN_FULL: 862 mostRecent = "full"; 863 break; 864 case UNDERRUN_PARTIAL: 865 mostRecent = "partial"; 866 break; 867 case UNDERRUN_EMPTY: 868 mostRecent = "empty"; 869 break; 870 default: 871 mostRecent = "?"; 872 break; 873 } 874 fdprintf(fd, " %5u %6s %4u %7u %5u %7s %5u\n", i, isActive ? "yes" : "no", 875 (underruns.mBitFields.mFull) & UNDERRUN_MASK, 876 (underruns.mBitFields.mPartial) & UNDERRUN_MASK, 877 (underruns.mBitFields.mEmpty) & UNDERRUN_MASK, 878 mostRecent, ftDump->mFramesReady); 879 } 880} 881 882} // namespace android 883