FastMixer.cpp revision 0ddd56316262ac74a95e9edb595697c163136d6d
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 <sys/atomics.h> 29#include <time.h> 30#include <utils/Log.h> 31#include <utils/Trace.h> 32#include <system/audio.h> 33#ifdef FAST_MIXER_STATISTICS 34#include <cpustats/CentralTendencyStatistics.h> 35#ifdef CPU_FREQUENCY_STATISTICS 36#include <cpustats/ThreadCpuUsage.h> 37#endif 38#endif 39#include "AudioMixer.h" 40#include "FastMixer.h" 41 42#define FAST_HOT_IDLE_NS 1000000L // 1 ms: time to sleep while hot idling 43#define FAST_DEFAULT_NS 999999999L // ~1 sec: default time to sleep 44#define MIN_WARMUP_CYCLES 2 // minimum number of loop cycles to wait for warmup 45#define MAX_WARMUP_CYCLES 10 // maximum number of loop cycles to wait for warmup 46 47namespace android { 48 49// Fast mixer thread 50bool FastMixer::threadLoop() 51{ 52 static const FastMixerState initial; 53 const FastMixerState *previous = &initial, *current = &initial; 54 FastMixerState preIdle; // copy of state before we went into idle 55 struct timespec oldTs = {0, 0}; 56 bool oldTsValid = false; 57 long slopNs = 0; // accumulated time we've woken up too early (> 0) or too late (< 0) 58 long sleepNs = -1; // -1: busy wait, 0: sched_yield, > 0: nanosleep 59 int fastTrackNames[FastMixerState::kMaxFastTracks]; // handles used by mixer to identify tracks 60 int generations[FastMixerState::kMaxFastTracks]; // last observed mFastTracks[i].mGeneration 61 unsigned i; 62 for (i = 0; i < FastMixerState::kMaxFastTracks; ++i) { 63 fastTrackNames[i] = -1; 64 generations[i] = 0; 65 } 66 NBAIO_Sink *outputSink = NULL; 67 int outputSinkGen = 0; 68 AudioMixer* mixer = NULL; 69 short *mixBuffer = NULL; 70 enum {UNDEFINED, MIXED, ZEROED} mixBufferState = UNDEFINED; 71 NBAIO_Format format = Format_Invalid; 72 unsigned sampleRate = 0; 73 int fastTracksGen = 0; 74 long periodNs = 0; // expected period; the time required to render one mix buffer 75 long underrunNs = 0; // underrun likely when write cycle is greater than this value 76 long overrunNs = 0; // overrun likely when write cycle is less than this value 77 long forceNs = 0; // if overrun detected, force the write cycle to take this much time 78 long warmupNs = 0; // warmup complete when write cycle is greater than to this value 79 FastMixerDumpState dummyDumpState, *dumpState = &dummyDumpState; 80 bool ignoreNextOverrun = true; // used to ignore initial overrun and first after an underrun 81#ifdef FAST_MIXER_STATISTICS 82 struct timespec oldLoad = {0, 0}; // previous value of clock_gettime(CLOCK_THREAD_CPUTIME_ID) 83 bool oldLoadValid = false; // whether oldLoad is valid 84 uint32_t bounds = 0; 85 bool full = false; // whether we have collected at least kSamplingN samples 86#ifdef CPU_FREQUENCY_STATISTICS 87 ThreadCpuUsage tcu; // for reading the current CPU clock frequency in kHz 88#endif 89#endif 90 unsigned coldGen = 0; // last observed mColdGen 91 bool isWarm = false; // true means ready to mix, false means wait for warmup before mixing 92 struct timespec measuredWarmupTs = {0, 0}; // how long did it take for warmup to complete 93 uint32_t warmupCycles = 0; // counter of number of loop cycles required to warmup 94 NBAIO_Sink* teeSink = NULL; // if non-NULL, then duplicate write() to this non-blocking sink 95 NBLog::Writer dummyLogWriter, *logWriter = &dummyLogWriter; 96 97 for (;;) { 98 99 // either nanosleep, sched_yield, or busy wait 100 if (sleepNs >= 0) { 101 if (sleepNs > 0) { 102 ALOG_ASSERT(sleepNs < 1000000000); 103 const struct timespec req = {0, sleepNs}; 104 nanosleep(&req, NULL); 105 } else { 106 sched_yield(); 107 } 108 } 109 // default to long sleep for next cycle 110 sleepNs = FAST_DEFAULT_NS; 111 112 // poll for state change 113 const FastMixerState *next = mSQ.poll(); 114 if (next == NULL) { 115 // continue to use the default initial state until a real state is available 116 ALOG_ASSERT(current == &initial && previous == &initial); 117 next = current; 118 } 119 120 FastMixerState::Command command = next->mCommand; 121 if (next != current) { 122 123 logWriter->logTimestamp(); 124 logWriter->log("next != current"); 125 126 // As soon as possible of learning of a new dump area, start using it 127 dumpState = next->mDumpState != NULL ? next->mDumpState : &dummyDumpState; 128 teeSink = next->mTeeSink; 129 logWriter = next->mNBLogWriter != NULL ? next->mNBLogWriter : &dummyLogWriter; 130 131 // We want to always have a valid reference to the previous (non-idle) state. 132 // However, the state queue only guarantees access to current and previous states. 133 // So when there is a transition from a non-idle state into an idle state, we make a 134 // copy of the last known non-idle state so it is still available on return from idle. 135 // The possible transitions are: 136 // non-idle -> non-idle update previous from current in-place 137 // non-idle -> idle update previous from copy of current 138 // idle -> idle don't update previous 139 // idle -> non-idle don't update previous 140 if (!(current->mCommand & FastMixerState::IDLE)) { 141 if (command & FastMixerState::IDLE) { 142 preIdle = *current; 143 current = &preIdle; 144 oldTsValid = false; 145 oldLoadValid = false; 146 ignoreNextOverrun = true; 147 } 148 previous = current; 149 } 150 current = next; 151 } 152#if !LOG_NDEBUG 153 next = NULL; // not referenced again 154#endif 155 156 dumpState->mCommand = command; 157 158 switch (command) { 159 case FastMixerState::INITIAL: 160 case FastMixerState::HOT_IDLE: 161 sleepNs = FAST_HOT_IDLE_NS; 162 continue; 163 case FastMixerState::COLD_IDLE: 164 // only perform a cold idle command once 165 // FIXME consider checking previous state and only perform if previous != COLD_IDLE 166 if (current->mColdGen != coldGen) { 167 int32_t *coldFutexAddr = current->mColdFutexAddr; 168 ALOG_ASSERT(coldFutexAddr != NULL); 169 int32_t old = android_atomic_dec(coldFutexAddr); 170 if (old <= 0) { 171 logWriter->log("wait"); 172 __futex_syscall4(coldFutexAddr, FUTEX_WAIT_PRIVATE, old - 1, NULL); 173 } 174 // This may be overly conservative; there could be times that the normal mixer 175 // requests such a brief cold idle that it doesn't require resetting this flag. 176 isWarm = false; 177 measuredWarmupTs.tv_sec = 0; 178 measuredWarmupTs.tv_nsec = 0; 179 warmupCycles = 0; 180 sleepNs = -1; 181 coldGen = current->mColdGen; 182 bounds = 0; 183 full = false; 184 oldTsValid = !clock_gettime(CLOCK_MONOTONIC, &oldTs); 185 } else { 186 sleepNs = FAST_HOT_IDLE_NS; 187 } 188 continue; 189 case FastMixerState::EXIT: 190 logWriter->log("exit"); 191 delete mixer; 192 delete[] mixBuffer; 193 return false; 194 case FastMixerState::MIX: 195 case FastMixerState::WRITE: 196 case FastMixerState::MIX_WRITE: 197 break; 198 default: 199 LOG_FATAL("bad command %d", command); 200 } 201 202 // there is a non-idle state available to us; did the state change? 203 size_t frameCount = current->mFrameCount; 204 if (current != previous) { 205 206 // handle state change here, but since we want to diff the state, 207 // we're prepared for previous == &initial the first time through 208 unsigned previousTrackMask; 209 210 // check for change in output HAL configuration 211 NBAIO_Format previousFormat = format; 212 if (current->mOutputSinkGen != outputSinkGen) { 213 outputSink = current->mOutputSink; 214 outputSinkGen = current->mOutputSinkGen; 215 if (outputSink == NULL) { 216 format = Format_Invalid; 217 sampleRate = 0; 218 } else { 219 format = outputSink->format(); 220 sampleRate = Format_sampleRate(format); 221 ALOG_ASSERT(Format_channelCount(format) == 2); 222 } 223 dumpState->mSampleRate = sampleRate; 224 } 225 226 if ((format != previousFormat) || (frameCount != previous->mFrameCount)) { 227 // FIXME to avoid priority inversion, don't delete here 228 delete mixer; 229 mixer = NULL; 230 delete[] mixBuffer; 231 mixBuffer = NULL; 232 if (frameCount > 0 && sampleRate > 0) { 233 // FIXME new may block for unbounded time at internal mutex of the heap 234 // implementation; it would be better to have normal mixer allocate for us 235 // to avoid blocking here and to prevent possible priority inversion 236 mixer = new AudioMixer(frameCount, sampleRate, FastMixerState::kMaxFastTracks); 237 mixBuffer = new short[frameCount * 2]; 238 periodNs = (frameCount * 1000000000LL) / sampleRate; // 1.00 239 underrunNs = (frameCount * 1750000000LL) / sampleRate; // 1.75 240 overrunNs = (frameCount * 500000000LL) / sampleRate; // 0.50 241 forceNs = (frameCount * 950000000LL) / sampleRate; // 0.95 242 warmupNs = (frameCount * 500000000LL) / sampleRate; // 0.50 243 } else { 244 periodNs = 0; 245 underrunNs = 0; 246 overrunNs = 0; 247 forceNs = 0; 248 warmupNs = 0; 249 } 250 mixBufferState = UNDEFINED; 251#if !LOG_NDEBUG 252 for (i = 0; i < FastMixerState::kMaxFastTracks; ++i) { 253 fastTrackNames[i] = -1; 254 } 255#endif 256 // we need to reconfigure all active tracks 257 previousTrackMask = 0; 258 fastTracksGen = current->mFastTracksGen - 1; 259 dumpState->mFrameCount = frameCount; 260 } else { 261 previousTrackMask = previous->mTrackMask; 262 } 263 264 // check for change in active track set 265 unsigned currentTrackMask = current->mTrackMask; 266 dumpState->mTrackMask = currentTrackMask; 267 if (current->mFastTracksGen != fastTracksGen) { 268 logWriter->logf("gen %d", current->mFastTracksGen); 269 ALOG_ASSERT(mixBuffer != NULL); 270 int name; 271 272 // process removed tracks first to avoid running out of track names 273 unsigned removedTracks = previousTrackMask & ~currentTrackMask; 274 if (removedTracks) { 275 logWriter->logf("removed %#x", removedTracks); 276 } 277 while (removedTracks != 0) { 278 i = __builtin_ctz(removedTracks); 279 removedTracks &= ~(1 << i); 280 const FastTrack* fastTrack = ¤t->mFastTracks[i]; 281 ALOG_ASSERT(fastTrack->mBufferProvider == NULL); 282 if (mixer != NULL) { 283 name = fastTrackNames[i]; 284 ALOG_ASSERT(name >= 0); 285 mixer->deleteTrackName(name); 286 } 287#if !LOG_NDEBUG 288 fastTrackNames[i] = -1; 289#endif 290 // don't reset track dump state, since other side is ignoring it 291 generations[i] = fastTrack->mGeneration; 292 } 293 294 // now process added tracks 295 unsigned addedTracks = currentTrackMask & ~previousTrackMask; 296 if (addedTracks) { 297 logWriter->logf("added %#x", addedTracks); 298 } 299 while (addedTracks != 0) { 300 i = __builtin_ctz(addedTracks); 301 addedTracks &= ~(1 << i); 302 const FastTrack* fastTrack = ¤t->mFastTracks[i]; 303 AudioBufferProvider *bufferProvider = fastTrack->mBufferProvider; 304 ALOG_ASSERT(bufferProvider != NULL && fastTrackNames[i] == -1); 305 if (bufferProvider == NULL || 306 bufferProvider->getValid() != AudioBufferProvider::kValid) { 307 logWriter->logTimestamp(); 308 logWriter->logf("added invalid %#x", i); 309 } 310 if (mixer != NULL) { 311 // calling getTrackName with default channel mask and a random invalid 312 // sessionId (no effects here) 313 name = mixer->getTrackName(AUDIO_CHANNEL_OUT_STEREO, -555); 314 ALOG_ASSERT(name >= 0); 315 fastTrackNames[i] = name; 316 mixer->setParameter(name, AudioMixer::TRACK, AudioMixer::FAST_INDEX, 317 (void *) i); 318 mixer->setBufferProvider(name, bufferProvider); 319 mixer->setParameter(name, AudioMixer::TRACK, AudioMixer::MAIN_BUFFER, 320 (void *) mixBuffer); 321 // newly allocated track names default to full scale volume 322 if (fastTrack->mSampleRate != 0 && fastTrack->mSampleRate != sampleRate) { 323 mixer->setParameter(name, AudioMixer::RESAMPLE, 324 AudioMixer::SAMPLE_RATE, (void*) fastTrack->mSampleRate); 325 } 326 mixer->setParameter(name, AudioMixer::TRACK, AudioMixer::CHANNEL_MASK, 327 (void *) fastTrack->mChannelMask); 328 mixer->enable(name); 329 } 330 generations[i] = fastTrack->mGeneration; 331 } 332 333 // finally process (potentially) modified tracks; these use the same slot 334 // but may have a different buffer provider or volume provider 335 unsigned modifiedTracks = currentTrackMask & previousTrackMask; 336 if (modifiedTracks) { 337 logWriter->logf("pot. mod. %#x", modifiedTracks); 338 } 339 unsigned actuallyModifiedTracks = 0; 340 while (modifiedTracks != 0) { 341 i = __builtin_ctz(modifiedTracks); 342 modifiedTracks &= ~(1 << i); 343 const FastTrack* fastTrack = ¤t->mFastTracks[i]; 344 if (fastTrack->mGeneration != generations[i]) { 345 actuallyModifiedTracks |= 1 << i; 346 AudioBufferProvider *bufferProvider = fastTrack->mBufferProvider; 347 ALOG_ASSERT(bufferProvider != NULL); 348 if (bufferProvider == NULL || 349 bufferProvider->getValid() != AudioBufferProvider::kValid) { 350 logWriter->logTimestamp(); 351 logWriter->logf("modified invalid %#x", i); 352 } 353 if (mixer != NULL) { 354 name = fastTrackNames[i]; 355 ALOG_ASSERT(name >= 0); 356 mixer->setParameter(name, AudioMixer::TRACK, AudioMixer::FAST_INDEX, 357 (void *) i); 358 mixer->setBufferProvider(name, bufferProvider); 359 if (fastTrack->mVolumeProvider == NULL) { 360 mixer->setParameter(name, AudioMixer::VOLUME, AudioMixer::VOLUME0, 361 (void *)0x1000); 362 mixer->setParameter(name, AudioMixer::VOLUME, AudioMixer::VOLUME1, 363 (void *)0x1000); 364 } 365 if (fastTrack->mSampleRate != 0 && 366 fastTrack->mSampleRate != sampleRate) { 367 mixer->setParameter(name, AudioMixer::RESAMPLE, 368 AudioMixer::SAMPLE_RATE, (void*) fastTrack->mSampleRate); 369 } else { 370 mixer->setParameter(name, AudioMixer::RESAMPLE, 371 AudioMixer::REMOVE, NULL); 372 } 373 mixer->setParameter(name, AudioMixer::TRACK, AudioMixer::CHANNEL_MASK, 374 (void *) fastTrack->mChannelMask); 375 // already enabled 376 } 377 generations[i] = fastTrack->mGeneration; 378 } 379 } 380 if (actuallyModifiedTracks) { 381 logWriter->logf("act. mod. %#x", actuallyModifiedTracks); 382 } 383 384 fastTracksGen = current->mFastTracksGen; 385 386 dumpState->mNumTracks = popcount(currentTrackMask); 387 } 388 389#if 1 // FIXME shouldn't need this 390 // only process state change once 391 previous = current; 392#endif 393 } 394 395 // do work using current state here 396 if ((command & FastMixerState::MIX) && (mixer != NULL) && isWarm) { 397 ALOG_ASSERT(mixBuffer != NULL); 398 // for each track, update volume and check for underrun 399 unsigned currentTrackMask = current->mTrackMask; 400 while (currentTrackMask != 0) { 401 i = __builtin_ctz(currentTrackMask); 402 currentTrackMask &= ~(1 << i); 403 const FastTrack* fastTrack = ¤t->mFastTracks[i]; 404 int name = fastTrackNames[i]; 405 ALOG_ASSERT(name >= 0); 406 if (fastTrack->mVolumeProvider != NULL) { 407 uint32_t vlr = fastTrack->mVolumeProvider->getVolumeLR(); 408 mixer->setParameter(name, AudioMixer::VOLUME, AudioMixer::VOLUME0, 409 (void *)(vlr & 0xFFFF)); 410 mixer->setParameter(name, AudioMixer::VOLUME, AudioMixer::VOLUME1, 411 (void *)(vlr >> 16)); 412 } 413 // FIXME The current implementation of framesReady() for fast tracks 414 // takes a tryLock, which can block 415 // up to 1 ms. If enough active tracks all blocked in sequence, this would result 416 // in the overall fast mix cycle being delayed. Should use a non-blocking FIFO. 417 size_t framesReady = fastTrack->mBufferProvider->framesReady(); 418 if (ATRACE_ENABLED()) { 419 // I wish we had formatted trace names 420 char traceName[16]; 421 strcpy(traceName, "framesReady"); 422 traceName[11] = i + (i < 10 ? '0' : 'A' - 10); 423 traceName[12] = '\0'; 424 ATRACE_INT(traceName, framesReady); 425 } 426 FastTrackDump *ftDump = &dumpState->mTracks[i]; 427 FastTrackUnderruns underruns = ftDump->mUnderruns; 428 if (framesReady < frameCount) { 429 if (framesReady == 0) { 430 underruns.mBitFields.mEmpty++; 431 underruns.mBitFields.mMostRecent = UNDERRUN_EMPTY; 432 mixer->disable(name); 433 } else { 434 // allow mixing partial buffer 435 underruns.mBitFields.mPartial++; 436 underruns.mBitFields.mMostRecent = UNDERRUN_PARTIAL; 437 mixer->enable(name); 438 } 439 } else { 440 underruns.mBitFields.mFull++; 441 underruns.mBitFields.mMostRecent = UNDERRUN_FULL; 442 mixer->enable(name); 443 } 444 ftDump->mUnderruns = underruns; 445 ftDump->mFramesReady = framesReady; 446 AudioBufferProvider *bufferProvider = fastTrack->mBufferProvider; 447 if (bufferProvider == NULL || 448 bufferProvider->getValid() != AudioBufferProvider::kValid) { 449 logWriter->logTimestamp(); 450 logWriter->logf("mixing invalid %#x", i); 451 } 452 } 453 454 int64_t pts; 455 if (outputSink == NULL || (OK != outputSink->getNextWriteTimestamp(&pts))) 456 pts = AudioBufferProvider::kInvalidPTS; 457 458 // process() is CPU-bound 459 mixer->process(pts); 460 mixBufferState = MIXED; 461 } else if (mixBufferState == MIXED) { 462 mixBufferState = UNDEFINED; 463 } 464 bool attemptedWrite = false; 465 //bool didFullWrite = false; // dumpsys could display a count of partial writes 466 if ((command & FastMixerState::WRITE) && (outputSink != NULL) && (mixBuffer != NULL)) { 467 if (mixBufferState == UNDEFINED) { 468 memset(mixBuffer, 0, frameCount * 2 * sizeof(short)); 469 mixBufferState = ZEROED; 470 } 471 if (teeSink != NULL) { 472 (void) teeSink->write(mixBuffer, frameCount); 473 } 474 // FIXME write() is non-blocking and lock-free for a properly implemented NBAIO sink, 475 // but this code should be modified to handle both non-blocking and blocking sinks 476 dumpState->mWriteSequence++; 477 ATRACE_BEGIN("write"); 478 ssize_t framesWritten = outputSink->write(mixBuffer, frameCount); 479 ATRACE_END(); 480 dumpState->mWriteSequence++; 481 if (framesWritten >= 0) { 482 ALOG_ASSERT(framesWritten <= frameCount); 483 dumpState->mFramesWritten += framesWritten; 484 //if ((size_t) framesWritten == frameCount) { 485 // didFullWrite = true; 486 //} 487 } else { 488 dumpState->mWriteErrors++; 489 } 490 attemptedWrite = true; 491 // FIXME count # of writes blocked excessively, CPU usage, etc. for dump 492 } 493 494 // To be exactly periodic, compute the next sleep time based on current time. 495 // This code doesn't have long-term stability when the sink is non-blocking. 496 // FIXME To avoid drift, use the local audio clock or watch the sink's fill status. 497 struct timespec newTs; 498 int rc = clock_gettime(CLOCK_MONOTONIC, &newTs); 499 if (rc == 0) { 500 logWriter->logTimestamp(newTs); 501 if (oldTsValid) { 502 time_t sec = newTs.tv_sec - oldTs.tv_sec; 503 long nsec = newTs.tv_nsec - oldTs.tv_nsec; 504 ALOGE_IF(sec < 0 || (sec == 0 && nsec < 0), 505 "clock_gettime(CLOCK_MONOTONIC) failed: was %ld.%09ld but now %ld.%09ld", 506 oldTs.tv_sec, oldTs.tv_nsec, newTs.tv_sec, newTs.tv_nsec); 507 if (nsec < 0) { 508 --sec; 509 nsec += 1000000000; 510 } 511 // To avoid an initial underrun on fast tracks after exiting standby, 512 // do not start pulling data from tracks and mixing until warmup is complete. 513 // Warmup is considered complete after the earlier of: 514 // MIN_WARMUP_CYCLES write() attempts and last one blocks for at least warmupNs 515 // MAX_WARMUP_CYCLES write() attempts. 516 // This is overly conservative, but to get better accuracy requires a new HAL API. 517 if (!isWarm && attemptedWrite) { 518 measuredWarmupTs.tv_sec += sec; 519 measuredWarmupTs.tv_nsec += nsec; 520 if (measuredWarmupTs.tv_nsec >= 1000000000) { 521 measuredWarmupTs.tv_sec++; 522 measuredWarmupTs.tv_nsec -= 1000000000; 523 } 524 ++warmupCycles; 525 if ((nsec > warmupNs && warmupCycles >= MIN_WARMUP_CYCLES) || 526 (warmupCycles >= MAX_WARMUP_CYCLES)) { 527 isWarm = true; 528 dumpState->mMeasuredWarmupTs = measuredWarmupTs; 529 dumpState->mWarmupCycles = warmupCycles; 530 } 531 } 532 sleepNs = -1; 533 if (isWarm) { 534 if (sec > 0 || nsec > underrunNs) { 535 ATRACE_NAME("underrun"); 536 // FIXME only log occasionally 537 ALOGV("underrun: time since last cycle %d.%03ld sec", 538 (int) sec, nsec / 1000000L); 539 dumpState->mUnderruns++; 540 ignoreNextOverrun = true; 541 } else if (nsec < overrunNs) { 542 if (ignoreNextOverrun) { 543 ignoreNextOverrun = false; 544 } else { 545 // FIXME only log occasionally 546 ALOGV("overrun: time since last cycle %d.%03ld sec", 547 (int) sec, nsec / 1000000L); 548 dumpState->mOverruns++; 549 } 550 // This forces a minimum cycle time. It: 551 // - compensates for an audio HAL with jitter due to sample rate conversion 552 // - works with a variable buffer depth audio HAL that never pulls at a rate 553 // < than overrunNs per buffer. 554 // - recovers from overrun immediately after underrun 555 // It doesn't work with a non-blocking audio HAL. 556 sleepNs = forceNs - nsec; 557 } else { 558 ignoreNextOverrun = false; 559 } 560 } 561#ifdef FAST_MIXER_STATISTICS 562 if (isWarm) { 563 // advance the FIFO queue bounds 564 size_t i = bounds & (FastMixerDumpState::kSamplingN - 1); 565 bounds = (bounds & 0xFFFF0000) | ((bounds + 1) & 0xFFFF); 566 if (full) { 567 bounds += 0x10000; 568 } else if (!(bounds & (FastMixerDumpState::kSamplingN - 1))) { 569 full = true; 570 } 571 // compute the delta value of clock_gettime(CLOCK_MONOTONIC) 572 uint32_t monotonicNs = nsec; 573 if (sec > 0 && sec < 4) { 574 monotonicNs += sec * 1000000000; 575 } 576 // compute the raw CPU load = delta value of clock_gettime(CLOCK_THREAD_CPUTIME_ID) 577 uint32_t loadNs = 0; 578 struct timespec newLoad; 579 rc = clock_gettime(CLOCK_THREAD_CPUTIME_ID, &newLoad); 580 if (rc == 0) { 581 if (oldLoadValid) { 582 sec = newLoad.tv_sec - oldLoad.tv_sec; 583 nsec = newLoad.tv_nsec - oldLoad.tv_nsec; 584 if (nsec < 0) { 585 --sec; 586 nsec += 1000000000; 587 } 588 loadNs = nsec; 589 if (sec > 0 && sec < 4) { 590 loadNs += sec * 1000000000; 591 } 592 } else { 593 // first time through the loop 594 oldLoadValid = true; 595 } 596 oldLoad = newLoad; 597 } 598#ifdef CPU_FREQUENCY_STATISTICS 599 // get the absolute value of CPU clock frequency in kHz 600 int cpuNum = sched_getcpu(); 601 uint32_t kHz = tcu.getCpukHz(cpuNum); 602 kHz = (kHz << 4) | (cpuNum & 0xF); 603#endif 604 // save values in FIFO queues for dumpsys 605 // these stores #1, #2, #3 are not atomic with respect to each other, 606 // or with respect to store #4 below 607 dumpState->mMonotonicNs[i] = monotonicNs; 608 dumpState->mLoadNs[i] = loadNs; 609#ifdef CPU_FREQUENCY_STATISTICS 610 dumpState->mCpukHz[i] = kHz; 611#endif 612 // this store #4 is not atomic with respect to stores #1, #2, #3 above, but 613 // the newest open and oldest closed halves are atomic with respect to each other 614 dumpState->mBounds = bounds; 615 ATRACE_INT("cycle_ms", monotonicNs / 1000000); 616 ATRACE_INT("load_us", loadNs / 1000); 617 } 618#endif 619 } else { 620 // first time through the loop 621 oldTsValid = true; 622 sleepNs = periodNs; 623 ignoreNextOverrun = true; 624 } 625 oldTs = newTs; 626 } else { 627 // monotonic clock is broken 628 oldTsValid = false; 629 sleepNs = periodNs; 630 } 631 632 633 } // for (;;) 634 635 // never return 'true'; Thread::_threadLoop() locks mutex which can result in priority inversion 636} 637 638FastMixerDumpState::FastMixerDumpState() : 639 mCommand(FastMixerState::INITIAL), mWriteSequence(0), mFramesWritten(0), 640 mNumTracks(0), mWriteErrors(0), mUnderruns(0), mOverruns(0), 641 mSampleRate(0), mFrameCount(0), /* mMeasuredWarmupTs({0, 0}), */ mWarmupCycles(0), 642 mTrackMask(0) 643#ifdef FAST_MIXER_STATISTICS 644 , mBounds(0) 645#endif 646{ 647 mMeasuredWarmupTs.tv_sec = 0; 648 mMeasuredWarmupTs.tv_nsec = 0; 649 // sample arrays aren't accessed atomically with respect to the bounds, 650 // so clearing reduces chance for dumpsys to read random uninitialized samples 651 memset(&mMonotonicNs, 0, sizeof(mMonotonicNs)); 652 memset(&mLoadNs, 0, sizeof(mLoadNs)); 653#ifdef CPU_FREQUENCY_STATISTICS 654 memset(&mCpukHz, 0, sizeof(mCpukHz)); 655#endif 656} 657 658FastMixerDumpState::~FastMixerDumpState() 659{ 660} 661 662// helper function called by qsort() 663static int compare_uint32_t(const void *pa, const void *pb) 664{ 665 uint32_t a = *(const uint32_t *)pa; 666 uint32_t b = *(const uint32_t *)pb; 667 if (a < b) { 668 return -1; 669 } else if (a > b) { 670 return 1; 671 } else { 672 return 0; 673 } 674} 675 676void FastMixerDumpState::dump(int fd) 677{ 678 if (mCommand == FastMixerState::INITIAL) { 679 fdprintf(fd, "FastMixer not initialized\n"); 680 return; 681 } 682#define COMMAND_MAX 32 683 char string[COMMAND_MAX]; 684 switch (mCommand) { 685 case FastMixerState::INITIAL: 686 strcpy(string, "INITIAL"); 687 break; 688 case FastMixerState::HOT_IDLE: 689 strcpy(string, "HOT_IDLE"); 690 break; 691 case FastMixerState::COLD_IDLE: 692 strcpy(string, "COLD_IDLE"); 693 break; 694 case FastMixerState::EXIT: 695 strcpy(string, "EXIT"); 696 break; 697 case FastMixerState::MIX: 698 strcpy(string, "MIX"); 699 break; 700 case FastMixerState::WRITE: 701 strcpy(string, "WRITE"); 702 break; 703 case FastMixerState::MIX_WRITE: 704 strcpy(string, "MIX_WRITE"); 705 break; 706 default: 707 snprintf(string, COMMAND_MAX, "%d", mCommand); 708 break; 709 } 710 double measuredWarmupMs = (mMeasuredWarmupTs.tv_sec * 1000.0) + 711 (mMeasuredWarmupTs.tv_nsec / 1000000.0); 712 double mixPeriodSec = (double) mFrameCount / (double) mSampleRate; 713 fdprintf(fd, "FastMixer command=%s writeSequence=%u framesWritten=%u\n" 714 " numTracks=%u writeErrors=%u underruns=%u overruns=%u\n" 715 " sampleRate=%u frameCount=%u measuredWarmup=%.3g ms, warmupCycles=%u\n" 716 " mixPeriod=%.2f ms\n", 717 string, mWriteSequence, mFramesWritten, 718 mNumTracks, mWriteErrors, mUnderruns, mOverruns, 719 mSampleRate, mFrameCount, measuredWarmupMs, mWarmupCycles, 720 mixPeriodSec * 1e3); 721#ifdef FAST_MIXER_STATISTICS 722 // find the interval of valid samples 723 uint32_t bounds = mBounds; 724 uint32_t newestOpen = bounds & 0xFFFF; 725 uint32_t oldestClosed = bounds >> 16; 726 uint32_t n = (newestOpen - oldestClosed) & 0xFFFF; 727 if (n > kSamplingN) { 728 ALOGE("too many samples %u", n); 729 n = kSamplingN; 730 } 731 // statistics for monotonic (wall clock) time, thread raw CPU load in time, CPU clock frequency, 732 // and adjusted CPU load in MHz normalized for CPU clock frequency 733 CentralTendencyStatistics wall, loadNs; 734#ifdef CPU_FREQUENCY_STATISTICS 735 CentralTendencyStatistics kHz, loadMHz; 736 uint32_t previousCpukHz = 0; 737#endif 738 // Assuming a normal distribution for cycle times, three standard deviations on either side of 739 // the mean account for 99.73% of the population. So if we take each tail to be 1/1000 of the 740 // sample set, we get 99.8% combined, or close to three standard deviations. 741 static const uint32_t kTailDenominator = 1000; 742 uint32_t *tail = n >= kTailDenominator ? new uint32_t[n] : NULL; 743 // loop over all the samples 744 for (uint32_t j = 0; j < n; ++j) { 745 size_t i = oldestClosed++ & (kSamplingN - 1); 746 uint32_t wallNs = mMonotonicNs[i]; 747 if (tail != NULL) { 748 tail[j] = wallNs; 749 } 750 wall.sample(wallNs); 751 uint32_t sampleLoadNs = mLoadNs[i]; 752 loadNs.sample(sampleLoadNs); 753#ifdef CPU_FREQUENCY_STATISTICS 754 uint32_t sampleCpukHz = mCpukHz[i]; 755 // skip bad kHz samples 756 if ((sampleCpukHz & ~0xF) != 0) { 757 kHz.sample(sampleCpukHz >> 4); 758 if (sampleCpukHz == previousCpukHz) { 759 double megacycles = (double) sampleLoadNs * (double) (sampleCpukHz >> 4) * 1e-12; 760 double adjMHz = megacycles / mixPeriodSec; // _not_ wallNs * 1e9 761 loadMHz.sample(adjMHz); 762 } 763 } 764 previousCpukHz = sampleCpukHz; 765#endif 766 } 767 fdprintf(fd, "Simple moving statistics over last %.1f seconds:\n", wall.n() * mixPeriodSec); 768 fdprintf(fd, " wall clock time in ms per mix cycle:\n" 769 " mean=%.2f min=%.2f max=%.2f stddev=%.2f\n", 770 wall.mean()*1e-6, wall.minimum()*1e-6, wall.maximum()*1e-6, wall.stddev()*1e-6); 771 fdprintf(fd, " raw CPU load in us per mix cycle:\n" 772 " mean=%.0f min=%.0f max=%.0f stddev=%.0f\n", 773 loadNs.mean()*1e-3, loadNs.minimum()*1e-3, loadNs.maximum()*1e-3, 774 loadNs.stddev()*1e-3); 775#ifdef CPU_FREQUENCY_STATISTICS 776 fdprintf(fd, " CPU clock frequency in MHz:\n" 777 " mean=%.0f min=%.0f max=%.0f stddev=%.0f\n", 778 kHz.mean()*1e-3, kHz.minimum()*1e-3, kHz.maximum()*1e-3, kHz.stddev()*1e-3); 779 fdprintf(fd, " adjusted CPU load in MHz (i.e. normalized for CPU clock frequency):\n" 780 " mean=%.1f min=%.1f max=%.1f stddev=%.1f\n", 781 loadMHz.mean(), loadMHz.minimum(), loadMHz.maximum(), loadMHz.stddev()); 782#endif 783 if (tail != NULL) { 784 qsort(tail, n, sizeof(uint32_t), compare_uint32_t); 785 // assume same number of tail samples on each side, left and right 786 uint32_t count = n / kTailDenominator; 787 CentralTendencyStatistics left, right; 788 for (uint32_t i = 0; i < count; ++i) { 789 left.sample(tail[i]); 790 right.sample(tail[n - (i + 1)]); 791 } 792 fdprintf(fd, "Distribution of mix cycle times in ms for the tails (> ~3 stddev outliers):\n" 793 " left tail: mean=%.2f min=%.2f max=%.2f stddev=%.2f\n" 794 " right tail: mean=%.2f min=%.2f max=%.2f stddev=%.2f\n", 795 left.mean()*1e-6, left.minimum()*1e-6, left.maximum()*1e-6, left.stddev()*1e-6, 796 right.mean()*1e-6, right.minimum()*1e-6, right.maximum()*1e-6, 797 right.stddev()*1e-6); 798 delete[] tail; 799 } 800#endif 801 // The active track mask and track states are updated non-atomically. 802 // So if we relied on isActive to decide whether to display, 803 // then we might display an obsolete track or omit an active track. 804 // Instead we always display all tracks, with an indication 805 // of whether we think the track is active. 806 uint32_t trackMask = mTrackMask; 807 fdprintf(fd, "Fast tracks: kMaxFastTracks=%u activeMask=%#x\n", 808 FastMixerState::kMaxFastTracks, trackMask); 809 fdprintf(fd, "Index Active Full Partial Empty Recent Ready\n"); 810 for (uint32_t i = 0; i < FastMixerState::kMaxFastTracks; ++i, trackMask >>= 1) { 811 bool isActive = trackMask & 1; 812 const FastTrackDump *ftDump = &mTracks[i]; 813 const FastTrackUnderruns& underruns = ftDump->mUnderruns; 814 const char *mostRecent; 815 switch (underruns.mBitFields.mMostRecent) { 816 case UNDERRUN_FULL: 817 mostRecent = "full"; 818 break; 819 case UNDERRUN_PARTIAL: 820 mostRecent = "partial"; 821 break; 822 case UNDERRUN_EMPTY: 823 mostRecent = "empty"; 824 break; 825 default: 826 mostRecent = "?"; 827 break; 828 } 829 fdprintf(fd, "%5u %6s %4u %7u %5u %7s %5u\n", i, isActive ? "yes" : "no", 830 (underruns.mBitFields.mFull) & UNDERRUN_MASK, 831 (underruns.mBitFields.mPartial) & UNDERRUN_MASK, 832 (underruns.mBitFields.mEmpty) & UNDERRUN_MASK, 833 mostRecent, ftDump->mFramesReady); 834 } 835} 836 837} // namespace android 838