FastMixer.cpp revision 377b2ec9a2885f9b6405b07ba900a9e3f4349c38
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 } 240 241 if ((format != previousFormat) || (frameCount != previous->mFrameCount)) { 242 // FIXME to avoid priority inversion, don't delete here 243 delete mixer; 244 mixer = NULL; 245 delete[] mixBuffer; 246 mixBuffer = NULL; 247 if (frameCount > 0 && sampleRate > 0) { 248 // FIXME new may block for unbounded time at internal mutex of the heap 249 // implementation; it would be better to have normal mixer allocate for us 250 // to avoid blocking here and to prevent possible priority inversion 251 mixer = new AudioMixer(frameCount, sampleRate, FastMixerState::kMaxFastTracks); 252 mixBuffer = new short[frameCount * FCC_2]; 253 periodNs = (frameCount * 1000000000LL) / sampleRate; // 1.00 254 underrunNs = (frameCount * 1750000000LL) / sampleRate; // 1.75 255 overrunNs = (frameCount * 500000000LL) / sampleRate; // 0.50 256 forceNs = (frameCount * 950000000LL) / sampleRate; // 0.95 257 warmupNs = (frameCount * 500000000LL) / sampleRate; // 0.50 258 } else { 259 periodNs = 0; 260 underrunNs = 0; 261 overrunNs = 0; 262 forceNs = 0; 263 warmupNs = 0; 264 } 265 mixBufferState = UNDEFINED; 266#if !LOG_NDEBUG 267 for (i = 0; i < FastMixerState::kMaxFastTracks; ++i) { 268 fastTrackNames[i] = -1; 269 } 270#endif 271 // we need to reconfigure all active tracks 272 previousTrackMask = 0; 273 fastTracksGen = current->mFastTracksGen - 1; 274 dumpState->mFrameCount = frameCount; 275 } else { 276 previousTrackMask = previous->mTrackMask; 277 } 278 279 // check for change in active track set 280 unsigned currentTrackMask = current->mTrackMask; 281 dumpState->mTrackMask = currentTrackMask; 282 if (current->mFastTracksGen != fastTracksGen) { 283 ALOG_ASSERT(mixBuffer != NULL); 284 int name; 285 286 // process removed tracks first to avoid running out of track names 287 unsigned removedTracks = previousTrackMask & ~currentTrackMask; 288 while (removedTracks != 0) { 289 i = __builtin_ctz(removedTracks); 290 removedTracks &= ~(1 << i); 291 const FastTrack* fastTrack = ¤t->mFastTracks[i]; 292 ALOG_ASSERT(fastTrack->mBufferProvider == NULL); 293 if (mixer != NULL) { 294 name = fastTrackNames[i]; 295 ALOG_ASSERT(name >= 0); 296 mixer->deleteTrackName(name); 297 } 298#if !LOG_NDEBUG 299 fastTrackNames[i] = -1; 300#endif 301 // don't reset track dump state, since other side is ignoring it 302 generations[i] = fastTrack->mGeneration; 303 } 304 305 // now process added tracks 306 unsigned addedTracks = currentTrackMask & ~previousTrackMask; 307 while (addedTracks != 0) { 308 i = __builtin_ctz(addedTracks); 309 addedTracks &= ~(1 << i); 310 const FastTrack* fastTrack = ¤t->mFastTracks[i]; 311 AudioBufferProvider *bufferProvider = fastTrack->mBufferProvider; 312 ALOG_ASSERT(bufferProvider != NULL && fastTrackNames[i] == -1); 313 if (mixer != NULL) { 314 // calling getTrackName with default channel mask and a random invalid 315 // sessionId (no effects here) 316 name = mixer->getTrackName(AUDIO_CHANNEL_OUT_STEREO, -555); 317 ALOG_ASSERT(name >= 0); 318 fastTrackNames[i] = name; 319 mixer->setBufferProvider(name, bufferProvider); 320 mixer->setParameter(name, AudioMixer::TRACK, AudioMixer::MAIN_BUFFER, 321 (void *) mixBuffer); 322 // newly allocated track names default to full scale volume 323 mixer->setParameter(name, AudioMixer::TRACK, AudioMixer::CHANNEL_MASK, 324 (void *)(uintptr_t)fastTrack->mChannelMask); 325 mixer->enable(name); 326 } 327 generations[i] = fastTrack->mGeneration; 328 } 329 330 // finally process (potentially) modified tracks; these use the same slot 331 // but may have a different buffer provider or volume provider 332 unsigned modifiedTracks = currentTrackMask & previousTrackMask; 333 while (modifiedTracks != 0) { 334 i = __builtin_ctz(modifiedTracks); 335 modifiedTracks &= ~(1 << i); 336 const FastTrack* fastTrack = ¤t->mFastTracks[i]; 337 if (fastTrack->mGeneration != generations[i]) { 338 // this track was actually modified 339 AudioBufferProvider *bufferProvider = fastTrack->mBufferProvider; 340 ALOG_ASSERT(bufferProvider != NULL); 341 if (mixer != NULL) { 342 name = fastTrackNames[i]; 343 ALOG_ASSERT(name >= 0); 344 mixer->setBufferProvider(name, bufferProvider); 345 if (fastTrack->mVolumeProvider == NULL) { 346 mixer->setParameter(name, AudioMixer::VOLUME, AudioMixer::VOLUME0, 347 (void *)0x1000); 348 mixer->setParameter(name, AudioMixer::VOLUME, AudioMixer::VOLUME1, 349 (void *)0x1000); 350 } 351 mixer->setParameter(name, AudioMixer::RESAMPLE, 352 AudioMixer::REMOVE, NULL); 353 mixer->setParameter(name, AudioMixer::TRACK, AudioMixer::CHANNEL_MASK, 354 (void *)(uintptr_t) fastTrack->mChannelMask); 355 // already enabled 356 } 357 generations[i] = fastTrack->mGeneration; 358 } 359 } 360 361 fastTracksGen = current->mFastTracksGen; 362 363 dumpState->mNumTracks = popcount(currentTrackMask); 364 } 365 366#if 1 // FIXME shouldn't need this 367 // only process state change once 368 previous = current; 369#endif 370 } 371 372 // do work using current state here 373 if ((command & FastMixerState::MIX) && (mixer != NULL) && isWarm) { 374 ALOG_ASSERT(mixBuffer != NULL); 375 // for each track, update volume and check for underrun 376 unsigned currentTrackMask = current->mTrackMask; 377 while (currentTrackMask != 0) { 378 i = __builtin_ctz(currentTrackMask); 379 currentTrackMask &= ~(1 << i); 380 const FastTrack* fastTrack = ¤t->mFastTracks[i]; 381 382 // Refresh the per-track timestamp 383 if (timestampStatus == NO_ERROR) { 384 uint32_t trackFramesWrittenButNotPresented = 385 nativeFramesWrittenButNotPresented; 386 uint32_t trackFramesWritten = fastTrack->mBufferProvider->framesReleased(); 387 // Can't provide an AudioTimestamp before first frame presented, 388 // or during the brief 32-bit wraparound window 389 if (trackFramesWritten >= trackFramesWrittenButNotPresented) { 390 AudioTimestamp perTrackTimestamp; 391 perTrackTimestamp.mPosition = 392 trackFramesWritten - trackFramesWrittenButNotPresented; 393 perTrackTimestamp.mTime = timestamp.mTime; 394 fastTrack->mBufferProvider->onTimestamp(perTrackTimestamp); 395 } 396 } 397 398 int name = fastTrackNames[i]; 399 ALOG_ASSERT(name >= 0); 400 if (fastTrack->mVolumeProvider != NULL) { 401 uint32_t vlr = fastTrack->mVolumeProvider->getVolumeLR(); 402 mixer->setParameter(name, AudioMixer::VOLUME, AudioMixer::VOLUME0, 403 (void *)(uintptr_t)(vlr & 0xFFFF)); 404 mixer->setParameter(name, AudioMixer::VOLUME, AudioMixer::VOLUME1, 405 (void *)(uintptr_t)(vlr >> 16)); 406 } 407 // FIXME The current implementation of framesReady() for fast tracks 408 // takes a tryLock, which can block 409 // up to 1 ms. If enough active tracks all blocked in sequence, this would result 410 // in the overall fast mix cycle being delayed. Should use a non-blocking FIFO. 411 size_t framesReady = fastTrack->mBufferProvider->framesReady(); 412 if (ATRACE_ENABLED()) { 413 // I wish we had formatted trace names 414 char traceName[16]; 415 strcpy(traceName, "fRdy"); 416 traceName[4] = i + (i < 10 ? '0' : 'A' - 10); 417 traceName[5] = '\0'; 418 ATRACE_INT(traceName, framesReady); 419 } 420 FastTrackDump *ftDump = &dumpState->mTracks[i]; 421 FastTrackUnderruns underruns = ftDump->mUnderruns; 422 if (framesReady < frameCount) { 423 if (framesReady == 0) { 424 underruns.mBitFields.mEmpty++; 425 underruns.mBitFields.mMostRecent = UNDERRUN_EMPTY; 426 mixer->disable(name); 427 } else { 428 // allow mixing partial buffer 429 underruns.mBitFields.mPartial++; 430 underruns.mBitFields.mMostRecent = UNDERRUN_PARTIAL; 431 mixer->enable(name); 432 } 433 } else { 434 underruns.mBitFields.mFull++; 435 underruns.mBitFields.mMostRecent = UNDERRUN_FULL; 436 mixer->enable(name); 437 } 438 ftDump->mUnderruns = underruns; 439 ftDump->mFramesReady = framesReady; 440 } 441 442 int64_t pts; 443 if (outputSink == NULL || (OK != outputSink->getNextWriteTimestamp(&pts))) 444 pts = AudioBufferProvider::kInvalidPTS; 445 446 // process() is CPU-bound 447 mixer->process(pts); 448 mixBufferState = MIXED; 449 } else if (mixBufferState == MIXED) { 450 mixBufferState = UNDEFINED; 451 } 452 bool attemptedWrite = false; 453 //bool didFullWrite = false; // dumpsys could display a count of partial writes 454 if ((command & FastMixerState::WRITE) && (outputSink != NULL) && (mixBuffer != NULL)) { 455 if (mixBufferState == UNDEFINED) { 456 memset(mixBuffer, 0, frameCount * FCC_2 * sizeof(short)); 457 mixBufferState = ZEROED; 458 } 459 if (teeSink != NULL) { 460 (void) teeSink->write(mixBuffer, frameCount); 461 } 462 // FIXME write() is non-blocking and lock-free for a properly implemented NBAIO sink, 463 // but this code should be modified to handle both non-blocking and blocking sinks 464 dumpState->mWriteSequence++; 465 ATRACE_BEGIN("write"); 466 ssize_t framesWritten = outputSink->write(mixBuffer, frameCount); 467 ATRACE_END(); 468 dumpState->mWriteSequence++; 469 if (framesWritten >= 0) { 470 ALOG_ASSERT((size_t) framesWritten <= frameCount); 471 totalNativeFramesWritten += framesWritten; 472 dumpState->mFramesWritten = totalNativeFramesWritten; 473 //if ((size_t) framesWritten == frameCount) { 474 // didFullWrite = true; 475 //} 476 } else { 477 dumpState->mWriteErrors++; 478 } 479 attemptedWrite = true; 480 // FIXME count # of writes blocked excessively, CPU usage, etc. for dump 481 482 timestampStatus = outputSink->getTimestamp(timestamp); 483 if (timestampStatus == NO_ERROR) { 484 uint32_t totalNativeFramesPresented = timestamp.mPosition; 485 if (totalNativeFramesPresented <= totalNativeFramesWritten) { 486 nativeFramesWrittenButNotPresented = 487 totalNativeFramesWritten - totalNativeFramesPresented; 488 } else { 489 // HAL reported that more frames were presented than were written 490 timestampStatus = INVALID_OPERATION; 491 } 492 } 493 } 494 495 // To be exactly periodic, compute the next sleep time based on current time. 496 // This code doesn't have long-term stability when the sink is non-blocking. 497 // FIXME To avoid drift, use the local audio clock or watch the sink's fill status. 498 struct timespec newTs; 499 int rc = clock_gettime(CLOCK_MONOTONIC, &newTs); 500 if (rc == 0) { 501 //logWriter->logTimestamp(newTs); 502 if (oldTsValid) { 503 time_t sec = newTs.tv_sec - oldTs.tv_sec; 504 long nsec = newTs.tv_nsec - oldTs.tv_nsec; 505 ALOGE_IF(sec < 0 || (sec == 0 && nsec < 0), 506 "clock_gettime(CLOCK_MONOTONIC) failed: was %ld.%09ld but now %ld.%09ld", 507 oldTs.tv_sec, oldTs.tv_nsec, newTs.tv_sec, newTs.tv_nsec); 508 if (nsec < 0) { 509 --sec; 510 nsec += 1000000000; 511 } 512 // To avoid an initial underrun on fast tracks after exiting standby, 513 // do not start pulling data from tracks and mixing until warmup is complete. 514 // Warmup is considered complete after the earlier of: 515 // MIN_WARMUP_CYCLES write() attempts and last one blocks for at least warmupNs 516 // MAX_WARMUP_CYCLES write() attempts. 517 // This is overly conservative, but to get better accuracy requires a new HAL API. 518 if (!isWarm && attemptedWrite) { 519 measuredWarmupTs.tv_sec += sec; 520 measuredWarmupTs.tv_nsec += nsec; 521 if (measuredWarmupTs.tv_nsec >= 1000000000) { 522 measuredWarmupTs.tv_sec++; 523 measuredWarmupTs.tv_nsec -= 1000000000; 524 } 525 ++warmupCycles; 526 if ((nsec > warmupNs && warmupCycles >= MIN_WARMUP_CYCLES) || 527 (warmupCycles >= MAX_WARMUP_CYCLES)) { 528 isWarm = true; 529 dumpState->mMeasuredWarmupTs = measuredWarmupTs; 530 dumpState->mWarmupCycles = warmupCycles; 531 } 532 } 533 sleepNs = -1; 534 if (isWarm) { 535 if (sec > 0 || nsec > underrunNs) { 536 ATRACE_NAME("underrun"); 537 // FIXME only log occasionally 538 ALOGV("underrun: time since last cycle %d.%03ld sec", 539 (int) sec, nsec / 1000000L); 540 dumpState->mUnderruns++; 541 ignoreNextOverrun = true; 542 } else if (nsec < overrunNs) { 543 if (ignoreNextOverrun) { 544 ignoreNextOverrun = false; 545 } else { 546 // FIXME only log occasionally 547 ALOGV("overrun: time since last cycle %d.%03ld sec", 548 (int) sec, nsec / 1000000L); 549 dumpState->mOverruns++; 550 } 551 // This forces a minimum cycle time. It: 552 // - compensates for an audio HAL with jitter due to sample rate conversion 553 // - works with a variable buffer depth audio HAL that never pulls at a 554 // rate < than overrunNs per buffer. 555 // - recovers from overrun immediately after underrun 556 // It doesn't work with a non-blocking audio HAL. 557 sleepNs = forceNs - nsec; 558 } else { 559 ignoreNextOverrun = false; 560 } 561 } 562#ifdef FAST_MIXER_STATISTICS 563 if (isWarm) { 564 // advance the FIFO queue bounds 565 size_t i = bounds & (dumpState->mSamplingN - 1); 566 bounds = (bounds & 0xFFFF0000) | ((bounds + 1) & 0xFFFF); 567 if (full) { 568 bounds += 0x10000; 569 } else if (!(bounds & (dumpState->mSamplingN - 1))) { 570 full = true; 571 } 572 // compute the delta value of clock_gettime(CLOCK_MONOTONIC) 573 uint32_t monotonicNs = nsec; 574 if (sec > 0 && sec < 4) { 575 monotonicNs += sec * 1000000000; 576 } 577 // compute raw CPU load = delta value of clock_gettime(CLOCK_THREAD_CPUTIME_ID) 578 uint32_t loadNs = 0; 579 struct timespec newLoad; 580 rc = clock_gettime(CLOCK_THREAD_CPUTIME_ID, &newLoad); 581 if (rc == 0) { 582 if (oldLoadValid) { 583 sec = newLoad.tv_sec - oldLoad.tv_sec; 584 nsec = newLoad.tv_nsec - oldLoad.tv_nsec; 585 if (nsec < 0) { 586 --sec; 587 nsec += 1000000000; 588 } 589 loadNs = nsec; 590 if (sec > 0 && sec < 4) { 591 loadNs += sec * 1000000000; 592 } 593 } else { 594 // first time through the loop 595 oldLoadValid = true; 596 } 597 oldLoad = newLoad; 598 } 599#ifdef CPU_FREQUENCY_STATISTICS 600 // get the absolute value of CPU clock frequency in kHz 601 int cpuNum = sched_getcpu(); 602 uint32_t kHz = tcu.getCpukHz(cpuNum); 603 kHz = (kHz << 4) | (cpuNum & 0xF); 604#endif 605 // save values in FIFO queues for dumpsys 606 // these stores #1, #2, #3 are not atomic with respect to each other, 607 // or with respect to store #4 below 608 dumpState->mMonotonicNs[i] = monotonicNs; 609 dumpState->mLoadNs[i] = loadNs; 610#ifdef CPU_FREQUENCY_STATISTICS 611 dumpState->mCpukHz[i] = kHz; 612#endif 613 // this store #4 is not atomic with respect to stores #1, #2, #3 above, but 614 // the newest open & oldest closed halves are atomic with respect to each other 615 dumpState->mBounds = bounds; 616 ATRACE_INT("cycle_ms", monotonicNs / 1000000); 617 ATRACE_INT("load_us", loadNs / 1000); 618 } 619#endif 620 } else { 621 // first time through the loop 622 oldTsValid = true; 623 sleepNs = periodNs; 624 ignoreNextOverrun = true; 625 } 626 oldTs = newTs; 627 } else { 628 // monotonic clock is broken 629 oldTsValid = false; 630 sleepNs = periodNs; 631 } 632 633 634 } // for (;;) 635 636 // never return 'true'; Thread::_threadLoop() locks mutex which can result in priority inversion 637} 638 639FastMixerDumpState::FastMixerDumpState( 640#ifdef FAST_MIXER_STATISTICS 641 uint32_t samplingN 642#endif 643 ) : 644 mCommand(FastMixerState::INITIAL), mWriteSequence(0), mFramesWritten(0), 645 mNumTracks(0), mWriteErrors(0), mUnderruns(0), mOverruns(0), 646 mSampleRate(0), mFrameCount(0), /* mMeasuredWarmupTs({0, 0}), */ mWarmupCycles(0), 647 mTrackMask(0) 648#ifdef FAST_MIXER_STATISTICS 649 , mSamplingN(0), mBounds(0) 650#endif 651{ 652 mMeasuredWarmupTs.tv_sec = 0; 653 mMeasuredWarmupTs.tv_nsec = 0; 654#ifdef FAST_MIXER_STATISTICS 655 increaseSamplingN(samplingN); 656#endif 657} 658 659#ifdef FAST_MIXER_STATISTICS 660void FastMixerDumpState::increaseSamplingN(uint32_t samplingN) 661{ 662 if (samplingN <= mSamplingN || samplingN > kSamplingN || roundup(samplingN) != samplingN) { 663 return; 664 } 665 uint32_t additional = samplingN - mSamplingN; 666 // sample arrays aren't accessed atomically with respect to the bounds, 667 // so clearing reduces chance for dumpsys to read random uninitialized samples 668 memset(&mMonotonicNs[mSamplingN], 0, sizeof(mMonotonicNs[0]) * additional); 669 memset(&mLoadNs[mSamplingN], 0, sizeof(mLoadNs[0]) * additional); 670#ifdef CPU_FREQUENCY_STATISTICS 671 memset(&mCpukHz[mSamplingN], 0, sizeof(mCpukHz[0]) * additional); 672#endif 673 mSamplingN = samplingN; 674} 675#endif 676 677FastMixerDumpState::~FastMixerDumpState() 678{ 679} 680 681// helper function called by qsort() 682static int compare_uint32_t(const void *pa, const void *pb) 683{ 684 uint32_t a = *(const uint32_t *)pa; 685 uint32_t b = *(const uint32_t *)pb; 686 if (a < b) { 687 return -1; 688 } else if (a > b) { 689 return 1; 690 } else { 691 return 0; 692 } 693} 694 695void FastMixerDumpState::dump(int fd) const 696{ 697 if (mCommand == FastMixerState::INITIAL) { 698 fdprintf(fd, "FastMixer not initialized\n"); 699 return; 700 } 701#define COMMAND_MAX 32 702 char string[COMMAND_MAX]; 703 switch (mCommand) { 704 case FastMixerState::INITIAL: 705 strcpy(string, "INITIAL"); 706 break; 707 case FastMixerState::HOT_IDLE: 708 strcpy(string, "HOT_IDLE"); 709 break; 710 case FastMixerState::COLD_IDLE: 711 strcpy(string, "COLD_IDLE"); 712 break; 713 case FastMixerState::EXIT: 714 strcpy(string, "EXIT"); 715 break; 716 case FastMixerState::MIX: 717 strcpy(string, "MIX"); 718 break; 719 case FastMixerState::WRITE: 720 strcpy(string, "WRITE"); 721 break; 722 case FastMixerState::MIX_WRITE: 723 strcpy(string, "MIX_WRITE"); 724 break; 725 default: 726 snprintf(string, COMMAND_MAX, "%d", mCommand); 727 break; 728 } 729 double measuredWarmupMs = (mMeasuredWarmupTs.tv_sec * 1000.0) + 730 (mMeasuredWarmupTs.tv_nsec / 1000000.0); 731 double mixPeriodSec = (double) mFrameCount / (double) mSampleRate; 732 fdprintf(fd, "FastMixer command=%s writeSequence=%u framesWritten=%u\n" 733 " numTracks=%u writeErrors=%u underruns=%u overruns=%u\n" 734 " sampleRate=%u frameCount=%zu measuredWarmup=%.3g ms, warmupCycles=%u\n" 735 " mixPeriod=%.2f ms\n", 736 string, mWriteSequence, mFramesWritten, 737 mNumTracks, mWriteErrors, mUnderruns, mOverruns, 738 mSampleRate, mFrameCount, measuredWarmupMs, mWarmupCycles, 739 mixPeriodSec * 1e3); 740#ifdef FAST_MIXER_STATISTICS 741 // find the interval of valid samples 742 uint32_t bounds = mBounds; 743 uint32_t newestOpen = bounds & 0xFFFF; 744 uint32_t oldestClosed = bounds >> 16; 745 uint32_t n = (newestOpen - oldestClosed) & 0xFFFF; 746 if (n > mSamplingN) { 747 ALOGE("too many samples %u", n); 748 n = mSamplingN; 749 } 750 // statistics for monotonic (wall clock) time, thread raw CPU load in time, CPU clock frequency, 751 // and adjusted CPU load in MHz normalized for CPU clock frequency 752 CentralTendencyStatistics wall, loadNs; 753#ifdef CPU_FREQUENCY_STATISTICS 754 CentralTendencyStatistics kHz, loadMHz; 755 uint32_t previousCpukHz = 0; 756#endif 757 // Assuming a normal distribution for cycle times, three standard deviations on either side of 758 // the mean account for 99.73% of the population. So if we take each tail to be 1/1000 of the 759 // sample set, we get 99.8% combined, or close to three standard deviations. 760 static const uint32_t kTailDenominator = 1000; 761 uint32_t *tail = n >= kTailDenominator ? new uint32_t[n] : NULL; 762 // loop over all the samples 763 for (uint32_t j = 0; j < n; ++j) { 764 size_t i = oldestClosed++ & (mSamplingN - 1); 765 uint32_t wallNs = mMonotonicNs[i]; 766 if (tail != NULL) { 767 tail[j] = wallNs; 768 } 769 wall.sample(wallNs); 770 uint32_t sampleLoadNs = mLoadNs[i]; 771 loadNs.sample(sampleLoadNs); 772#ifdef CPU_FREQUENCY_STATISTICS 773 uint32_t sampleCpukHz = mCpukHz[i]; 774 // skip bad kHz samples 775 if ((sampleCpukHz & ~0xF) != 0) { 776 kHz.sample(sampleCpukHz >> 4); 777 if (sampleCpukHz == previousCpukHz) { 778 double megacycles = (double) sampleLoadNs * (double) (sampleCpukHz >> 4) * 1e-12; 779 double adjMHz = megacycles / mixPeriodSec; // _not_ wallNs * 1e9 780 loadMHz.sample(adjMHz); 781 } 782 } 783 previousCpukHz = sampleCpukHz; 784#endif 785 } 786 fdprintf(fd, "Simple moving statistics over last %.1f seconds:\n", wall.n() * mixPeriodSec); 787 fdprintf(fd, " wall clock time in ms per mix cycle:\n" 788 " mean=%.2f min=%.2f max=%.2f stddev=%.2f\n", 789 wall.mean()*1e-6, wall.minimum()*1e-6, wall.maximum()*1e-6, wall.stddev()*1e-6); 790 fdprintf(fd, " raw CPU load in us per mix cycle:\n" 791 " mean=%.0f min=%.0f max=%.0f stddev=%.0f\n", 792 loadNs.mean()*1e-3, loadNs.minimum()*1e-3, loadNs.maximum()*1e-3, 793 loadNs.stddev()*1e-3); 794#ifdef CPU_FREQUENCY_STATISTICS 795 fdprintf(fd, " CPU clock frequency in MHz:\n" 796 " mean=%.0f min=%.0f max=%.0f stddev=%.0f\n", 797 kHz.mean()*1e-3, kHz.minimum()*1e-3, kHz.maximum()*1e-3, kHz.stddev()*1e-3); 798 fdprintf(fd, " adjusted CPU load in MHz (i.e. normalized for CPU clock frequency):\n" 799 " mean=%.1f min=%.1f max=%.1f stddev=%.1f\n", 800 loadMHz.mean(), loadMHz.minimum(), loadMHz.maximum(), loadMHz.stddev()); 801#endif 802 if (tail != NULL) { 803 qsort(tail, n, sizeof(uint32_t), compare_uint32_t); 804 // assume same number of tail samples on each side, left and right 805 uint32_t count = n / kTailDenominator; 806 CentralTendencyStatistics left, right; 807 for (uint32_t i = 0; i < count; ++i) { 808 left.sample(tail[i]); 809 right.sample(tail[n - (i + 1)]); 810 } 811 fdprintf(fd, "Distribution of mix cycle times in ms for the tails (> ~3 stddev outliers):\n" 812 " left tail: mean=%.2f min=%.2f max=%.2f stddev=%.2f\n" 813 " right tail: mean=%.2f min=%.2f max=%.2f stddev=%.2f\n", 814 left.mean()*1e-6, left.minimum()*1e-6, left.maximum()*1e-6, left.stddev()*1e-6, 815 right.mean()*1e-6, right.minimum()*1e-6, right.maximum()*1e-6, 816 right.stddev()*1e-6); 817 delete[] tail; 818 } 819#endif 820 // The active track mask and track states are updated non-atomically. 821 // So if we relied on isActive to decide whether to display, 822 // then we might display an obsolete track or omit an active track. 823 // Instead we always display all tracks, with an indication 824 // of whether we think the track is active. 825 uint32_t trackMask = mTrackMask; 826 fdprintf(fd, "Fast tracks: kMaxFastTracks=%u activeMask=%#x\n", 827 FastMixerState::kMaxFastTracks, trackMask); 828 fdprintf(fd, "Index Active Full Partial Empty Recent Ready\n"); 829 for (uint32_t i = 0; i < FastMixerState::kMaxFastTracks; ++i, trackMask >>= 1) { 830 bool isActive = trackMask & 1; 831 const FastTrackDump *ftDump = &mTracks[i]; 832 const FastTrackUnderruns& underruns = ftDump->mUnderruns; 833 const char *mostRecent; 834 switch (underruns.mBitFields.mMostRecent) { 835 case UNDERRUN_FULL: 836 mostRecent = "full"; 837 break; 838 case UNDERRUN_PARTIAL: 839 mostRecent = "partial"; 840 break; 841 case UNDERRUN_EMPTY: 842 mostRecent = "empty"; 843 break; 844 default: 845 mostRecent = "?"; 846 break; 847 } 848 fdprintf(fd, "%5u %6s %4u %7u %5u %7s %5zu\n", i, isActive ? "yes" : "no", 849 (underruns.mBitFields.mFull) & UNDERRUN_MASK, 850 (underruns.mBitFields.mPartial) & UNDERRUN_MASK, 851 (underruns.mBitFields.mEmpty) & UNDERRUN_MASK, 852 mostRecent, ftDump->mFramesReady); 853 } 854} 855 856} // namespace android 857