FastMixer.cpp revision 4b76d27d6c4751b31a1cb8ac5e6da1d4b7724a7b
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_isEqual(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 *) 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 *) 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 *)(vlr & 0xFFFF)); 404 mixer->setParameter(name, AudioMixer::VOLUME, AudioMixer::VOLUME1, 405 (void *)(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 } 446 447 // process() is CPU-bound 448 mixer->process(pts); 449 mixBufferState = MIXED; 450 } else if (mixBufferState == MIXED) { 451 mixBufferState = UNDEFINED; 452 } 453 bool attemptedWrite = false; 454 //bool didFullWrite = false; // dumpsys could display a count of partial writes 455 if ((command & FastMixerState::WRITE) && (outputSink != NULL) && (mixBuffer != NULL)) { 456 if (mixBufferState == UNDEFINED) { 457 memset(mixBuffer, 0, frameCount * FCC_2 * sizeof(short)); 458 mixBufferState = ZEROED; 459 } 460 if (teeSink != NULL) { 461 (void) teeSink->write(mixBuffer, frameCount); 462 } 463 // FIXME write() is non-blocking and lock-free for a properly implemented NBAIO sink, 464 // but this code should be modified to handle both non-blocking and blocking sinks 465 dumpState->mWriteSequence++; 466 ATRACE_BEGIN("write"); 467 ssize_t framesWritten = outputSink->write(mixBuffer, frameCount); 468 ATRACE_END(); 469 dumpState->mWriteSequence++; 470 if (framesWritten >= 0) { 471 ALOG_ASSERT((size_t) framesWritten <= frameCount); 472 totalNativeFramesWritten += framesWritten; 473 dumpState->mFramesWritten = totalNativeFramesWritten; 474 //if ((size_t) framesWritten == frameCount) { 475 // didFullWrite = true; 476 //} 477 } else { 478 dumpState->mWriteErrors++; 479 } 480 attemptedWrite = true; 481 // FIXME count # of writes blocked excessively, CPU usage, etc. for dump 482 483 timestampStatus = outputSink->getTimestamp(timestamp); 484 if (timestampStatus == NO_ERROR) { 485 uint32_t totalNativeFramesPresented = timestamp.mPosition; 486 if (totalNativeFramesPresented <= totalNativeFramesWritten) { 487 nativeFramesWrittenButNotPresented = 488 totalNativeFramesWritten - totalNativeFramesPresented; 489 } else { 490 // HAL reported that more frames were presented than were written 491 timestampStatus = INVALID_OPERATION; 492 } 493 } 494 } 495 496 // To be exactly periodic, compute the next sleep time based on current time. 497 // This code doesn't have long-term stability when the sink is non-blocking. 498 // FIXME To avoid drift, use the local audio clock or watch the sink's fill status. 499 struct timespec newTs; 500 int rc = clock_gettime(CLOCK_MONOTONIC, &newTs); 501 if (rc == 0) { 502 //logWriter->logTimestamp(newTs); 503 if (oldTsValid) { 504 time_t sec = newTs.tv_sec - oldTs.tv_sec; 505 long nsec = newTs.tv_nsec - oldTs.tv_nsec; 506 ALOGE_IF(sec < 0 || (sec == 0 && nsec < 0), 507 "clock_gettime(CLOCK_MONOTONIC) failed: was %ld.%09ld but now %ld.%09ld", 508 oldTs.tv_sec, oldTs.tv_nsec, newTs.tv_sec, newTs.tv_nsec); 509 if (nsec < 0) { 510 --sec; 511 nsec += 1000000000; 512 } 513 // To avoid an initial underrun on fast tracks after exiting standby, 514 // do not start pulling data from tracks and mixing until warmup is complete. 515 // Warmup is considered complete after the earlier of: 516 // MIN_WARMUP_CYCLES write() attempts and last one blocks for at least warmupNs 517 // MAX_WARMUP_CYCLES write() attempts. 518 // This is overly conservative, but to get better accuracy requires a new HAL API. 519 if (!isWarm && attemptedWrite) { 520 measuredWarmupTs.tv_sec += sec; 521 measuredWarmupTs.tv_nsec += nsec; 522 if (measuredWarmupTs.tv_nsec >= 1000000000) { 523 measuredWarmupTs.tv_sec++; 524 measuredWarmupTs.tv_nsec -= 1000000000; 525 } 526 ++warmupCycles; 527 if ((nsec > warmupNs && warmupCycles >= MIN_WARMUP_CYCLES) || 528 (warmupCycles >= MAX_WARMUP_CYCLES)) { 529 isWarm = true; 530 dumpState->mMeasuredWarmupTs = measuredWarmupTs; 531 dumpState->mWarmupCycles = warmupCycles; 532 } 533 } 534 sleepNs = -1; 535 if (isWarm) { 536 if (sec > 0 || nsec > underrunNs) { 537 ATRACE_NAME("underrun"); 538 // FIXME only log occasionally 539 ALOGV("underrun: time since last cycle %d.%03ld sec", 540 (int) sec, nsec / 1000000L); 541 dumpState->mUnderruns++; 542 ignoreNextOverrun = true; 543 } else if (nsec < overrunNs) { 544 if (ignoreNextOverrun) { 545 ignoreNextOverrun = false; 546 } else { 547 // FIXME only log occasionally 548 ALOGV("overrun: time since last cycle %d.%03ld sec", 549 (int) sec, nsec / 1000000L); 550 dumpState->mOverruns++; 551 } 552 // This forces a minimum cycle time. It: 553 // - compensates for an audio HAL with jitter due to sample rate conversion 554 // - works with a variable buffer depth audio HAL that never pulls at a 555 // rate < than overrunNs per buffer. 556 // - recovers from overrun immediately after underrun 557 // It doesn't work with a non-blocking audio HAL. 558 sleepNs = forceNs - nsec; 559 } else { 560 ignoreNextOverrun = false; 561 } 562 } 563#ifdef FAST_MIXER_STATISTICS 564 if (isWarm) { 565 // advance the FIFO queue bounds 566 size_t i = bounds & (dumpState->mSamplingN - 1); 567 bounds = (bounds & 0xFFFF0000) | ((bounds + 1) & 0xFFFF); 568 if (full) { 569 bounds += 0x10000; 570 } else if (!(bounds & (dumpState->mSamplingN - 1))) { 571 full = true; 572 } 573 // compute the delta value of clock_gettime(CLOCK_MONOTONIC) 574 uint32_t monotonicNs = nsec; 575 if (sec > 0 && sec < 4) { 576 monotonicNs += sec * 1000000000; 577 } 578 // compute raw CPU load = delta value of clock_gettime(CLOCK_THREAD_CPUTIME_ID) 579 uint32_t loadNs = 0; 580 struct timespec newLoad; 581 rc = clock_gettime(CLOCK_THREAD_CPUTIME_ID, &newLoad); 582 if (rc == 0) { 583 if (oldLoadValid) { 584 sec = newLoad.tv_sec - oldLoad.tv_sec; 585 nsec = newLoad.tv_nsec - oldLoad.tv_nsec; 586 if (nsec < 0) { 587 --sec; 588 nsec += 1000000000; 589 } 590 loadNs = nsec; 591 if (sec > 0 && sec < 4) { 592 loadNs += sec * 1000000000; 593 } 594 } else { 595 // first time through the loop 596 oldLoadValid = true; 597 } 598 oldLoad = newLoad; 599 } 600#ifdef CPU_FREQUENCY_STATISTICS 601 // get the absolute value of CPU clock frequency in kHz 602 int cpuNum = sched_getcpu(); 603 uint32_t kHz = tcu.getCpukHz(cpuNum); 604 kHz = (kHz << 4) | (cpuNum & 0xF); 605#endif 606 // save values in FIFO queues for dumpsys 607 // these stores #1, #2, #3 are not atomic with respect to each other, 608 // or with respect to store #4 below 609 dumpState->mMonotonicNs[i] = monotonicNs; 610 dumpState->mLoadNs[i] = loadNs; 611#ifdef CPU_FREQUENCY_STATISTICS 612 dumpState->mCpukHz[i] = kHz; 613#endif 614 // this store #4 is not atomic with respect to stores #1, #2, #3 above, but 615 // the newest open & oldest closed halves are atomic with respect to each other 616 dumpState->mBounds = bounds; 617 ATRACE_INT("cycle_ms", monotonicNs / 1000000); 618 ATRACE_INT("load_us", loadNs / 1000); 619 } 620#endif 621 } else { 622 // first time through the loop 623 oldTsValid = true; 624 sleepNs = periodNs; 625 ignoreNextOverrun = true; 626 } 627 oldTs = newTs; 628 } else { 629 // monotonic clock is broken 630 oldTsValid = false; 631 sleepNs = periodNs; 632 } 633 634 635 } // for (;;) 636 637 // never return 'true'; Thread::_threadLoop() locks mutex which can result in priority inversion 638} 639 640FastMixerDumpState::FastMixerDumpState( 641#ifdef FAST_MIXER_STATISTICS 642 uint32_t samplingN 643#endif 644 ) : 645 mCommand(FastMixerState::INITIAL), mWriteSequence(0), mFramesWritten(0), 646 mNumTracks(0), mWriteErrors(0), mUnderruns(0), mOverruns(0), 647 mSampleRate(0), mFrameCount(0), /* mMeasuredWarmupTs({0, 0}), */ mWarmupCycles(0), 648 mTrackMask(0) 649#ifdef FAST_MIXER_STATISTICS 650 , mSamplingN(0), mBounds(0) 651#endif 652{ 653 mMeasuredWarmupTs.tv_sec = 0; 654 mMeasuredWarmupTs.tv_nsec = 0; 655#ifdef FAST_MIXER_STATISTICS 656 increaseSamplingN(samplingN); 657#endif 658} 659 660#ifdef FAST_MIXER_STATISTICS 661void FastMixerDumpState::increaseSamplingN(uint32_t samplingN) 662{ 663 if (samplingN <= mSamplingN || samplingN > kSamplingN || roundup(samplingN) != samplingN) { 664 return; 665 } 666 uint32_t additional = samplingN - mSamplingN; 667 // sample arrays aren't accessed atomically with respect to the bounds, 668 // so clearing reduces chance for dumpsys to read random uninitialized samples 669 memset(&mMonotonicNs[mSamplingN], 0, sizeof(mMonotonicNs[0]) * additional); 670 memset(&mLoadNs[mSamplingN], 0, sizeof(mLoadNs[0]) * additional); 671#ifdef CPU_FREQUENCY_STATISTICS 672 memset(&mCpukHz[mSamplingN], 0, sizeof(mCpukHz[0]) * additional); 673#endif 674 mSamplingN = samplingN; 675} 676#endif 677 678FastMixerDumpState::~FastMixerDumpState() 679{ 680} 681 682// helper function called by qsort() 683static int compare_uint32_t(const void *pa, const void *pb) 684{ 685 uint32_t a = *(const uint32_t *)pa; 686 uint32_t b = *(const uint32_t *)pb; 687 if (a < b) { 688 return -1; 689 } else if (a > b) { 690 return 1; 691 } else { 692 return 0; 693 } 694} 695 696void FastMixerDumpState::dump(int fd) const 697{ 698 if (mCommand == FastMixerState::INITIAL) { 699 fdprintf(fd, " FastMixer not initialized\n"); 700 return; 701 } 702#define COMMAND_MAX 32 703 char string[COMMAND_MAX]; 704 switch (mCommand) { 705 case FastMixerState::INITIAL: 706 strcpy(string, "INITIAL"); 707 break; 708 case FastMixerState::HOT_IDLE: 709 strcpy(string, "HOT_IDLE"); 710 break; 711 case FastMixerState::COLD_IDLE: 712 strcpy(string, "COLD_IDLE"); 713 break; 714 case FastMixerState::EXIT: 715 strcpy(string, "EXIT"); 716 break; 717 case FastMixerState::MIX: 718 strcpy(string, "MIX"); 719 break; 720 case FastMixerState::WRITE: 721 strcpy(string, "WRITE"); 722 break; 723 case FastMixerState::MIX_WRITE: 724 strcpy(string, "MIX_WRITE"); 725 break; 726 default: 727 snprintf(string, COMMAND_MAX, "%d", mCommand); 728 break; 729 } 730 double measuredWarmupMs = (mMeasuredWarmupTs.tv_sec * 1000.0) + 731 (mMeasuredWarmupTs.tv_nsec / 1000000.0); 732 double mixPeriodSec = (double) mFrameCount / (double) mSampleRate; 733 fdprintf(fd, " FastMixer command=%s writeSequence=%u framesWritten=%u\n" 734 " numTracks=%u writeErrors=%u underruns=%u overruns=%u\n" 735 " sampleRate=%u frameCount=%u measuredWarmup=%.3g ms, warmupCycles=%u\n" 736 " mixPeriod=%.2f ms\n", 737 string, mWriteSequence, mFramesWritten, 738 mNumTracks, mWriteErrors, mUnderruns, mOverruns, 739 mSampleRate, mFrameCount, measuredWarmupMs, mWarmupCycles, 740 mixPeriodSec * 1e3); 741#ifdef FAST_MIXER_STATISTICS 742 // find the interval of valid samples 743 uint32_t bounds = mBounds; 744 uint32_t newestOpen = bounds & 0xFFFF; 745 uint32_t oldestClosed = bounds >> 16; 746 uint32_t n = (newestOpen - oldestClosed) & 0xFFFF; 747 if (n > mSamplingN) { 748 ALOGE("too many samples %u", n); 749 n = mSamplingN; 750 } 751 // statistics for monotonic (wall clock) time, thread raw CPU load in time, CPU clock frequency, 752 // and adjusted CPU load in MHz normalized for CPU clock frequency 753 CentralTendencyStatistics wall, loadNs; 754#ifdef CPU_FREQUENCY_STATISTICS 755 CentralTendencyStatistics kHz, loadMHz; 756 uint32_t previousCpukHz = 0; 757#endif 758 // Assuming a normal distribution for cycle times, three standard deviations on either side of 759 // the mean account for 99.73% of the population. So if we take each tail to be 1/1000 of the 760 // sample set, we get 99.8% combined, or close to three standard deviations. 761 static const uint32_t kTailDenominator = 1000; 762 uint32_t *tail = n >= kTailDenominator ? new uint32_t[n] : NULL; 763 // loop over all the samples 764 for (uint32_t j = 0; j < n; ++j) { 765 size_t i = oldestClosed++ & (mSamplingN - 1); 766 uint32_t wallNs = mMonotonicNs[i]; 767 if (tail != NULL) { 768 tail[j] = wallNs; 769 } 770 wall.sample(wallNs); 771 uint32_t sampleLoadNs = mLoadNs[i]; 772 loadNs.sample(sampleLoadNs); 773#ifdef CPU_FREQUENCY_STATISTICS 774 uint32_t sampleCpukHz = mCpukHz[i]; 775 // skip bad kHz samples 776 if ((sampleCpukHz & ~0xF) != 0) { 777 kHz.sample(sampleCpukHz >> 4); 778 if (sampleCpukHz == previousCpukHz) { 779 double megacycles = (double) sampleLoadNs * (double) (sampleCpukHz >> 4) * 1e-12; 780 double adjMHz = megacycles / mixPeriodSec; // _not_ wallNs * 1e9 781 loadMHz.sample(adjMHz); 782 } 783 } 784 previousCpukHz = sampleCpukHz; 785#endif 786 } 787 if (n) { 788 fdprintf(fd, " Simple moving statistics over last %.1f seconds:\n", 789 wall.n() * mixPeriodSec); 790 fdprintf(fd, " wall clock time in ms per mix cycle:\n" 791 " mean=%.2f min=%.2f max=%.2f stddev=%.2f\n", 792 wall.mean()*1e-6, wall.minimum()*1e-6, wall.maximum()*1e-6, 793 wall.stddev()*1e-6); 794 fdprintf(fd, " raw CPU load in us per mix cycle:\n" 795 " mean=%.0f min=%.0f max=%.0f stddev=%.0f\n", 796 loadNs.mean()*1e-3, loadNs.minimum()*1e-3, loadNs.maximum()*1e-3, 797 loadNs.stddev()*1e-3); 798 } else { 799 fdprintf(fd, " No FastMixer statistics available currently\n"); 800 } 801#ifdef CPU_FREQUENCY_STATISTICS 802 fdprintf(fd, " CPU clock frequency in MHz:\n" 803 " mean=%.0f min=%.0f max=%.0f stddev=%.0f\n", 804 kHz.mean()*1e-3, kHz.minimum()*1e-3, kHz.maximum()*1e-3, kHz.stddev()*1e-3); 805 fdprintf(fd, " adjusted CPU load in MHz (i.e. normalized for CPU clock frequency):\n" 806 " mean=%.1f min=%.1f max=%.1f stddev=%.1f\n", 807 loadMHz.mean(), loadMHz.minimum(), loadMHz.maximum(), loadMHz.stddev()); 808#endif 809 if (tail != NULL) { 810 qsort(tail, n, sizeof(uint32_t), compare_uint32_t); 811 // assume same number of tail samples on each side, left and right 812 uint32_t count = n / kTailDenominator; 813 CentralTendencyStatistics left, right; 814 for (uint32_t i = 0; i < count; ++i) { 815 left.sample(tail[i]); 816 right.sample(tail[n - (i + 1)]); 817 } 818 fdprintf(fd, " Distribution of mix cycle times in ms for the tails (> ~3 stddev outliers):\n" 819 " left tail: mean=%.2f min=%.2f max=%.2f stddev=%.2f\n" 820 " right tail: mean=%.2f min=%.2f max=%.2f stddev=%.2f\n", 821 left.mean()*1e-6, left.minimum()*1e-6, left.maximum()*1e-6, left.stddev()*1e-6, 822 right.mean()*1e-6, right.minimum()*1e-6, right.maximum()*1e-6, 823 right.stddev()*1e-6); 824 delete[] tail; 825 } 826#endif 827 // The active track mask and track states are updated non-atomically. 828 // So if we relied on isActive to decide whether to display, 829 // then we might display an obsolete track or omit an active track. 830 // Instead we always display all tracks, with an indication 831 // of whether we think the track is active. 832 uint32_t trackMask = mTrackMask; 833 fdprintf(fd, " Fast tracks: kMaxFastTracks=%u activeMask=%#x\n", 834 FastMixerState::kMaxFastTracks, trackMask); 835 fdprintf(fd, " Index Active Full Partial Empty Recent Ready\n"); 836 for (uint32_t i = 0; i < FastMixerState::kMaxFastTracks; ++i, trackMask >>= 1) { 837 bool isActive = trackMask & 1; 838 const FastTrackDump *ftDump = &mTracks[i]; 839 const FastTrackUnderruns& underruns = ftDump->mUnderruns; 840 const char *mostRecent; 841 switch (underruns.mBitFields.mMostRecent) { 842 case UNDERRUN_FULL: 843 mostRecent = "full"; 844 break; 845 case UNDERRUN_PARTIAL: 846 mostRecent = "partial"; 847 break; 848 case UNDERRUN_EMPTY: 849 mostRecent = "empty"; 850 break; 851 default: 852 mostRecent = "?"; 853 break; 854 } 855 fdprintf(fd, " %5u %6s %4u %7u %5u %7s %5u\n", i, isActive ? "yes" : "no", 856 (underruns.mBitFields.mFull) & UNDERRUN_MASK, 857 (underruns.mBitFields.mPartial) & UNDERRUN_MASK, 858 (underruns.mBitFields.mEmpty) & UNDERRUN_MASK, 859 mostRecent, ftDump->mFramesReady); 860 } 861} 862 863} // namespace android 864