FastMixer.cpp revision 7f5d335f7b4caecd0dfb8f1085f352f1d2da5d2e
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->log("next != current"); 124 125 // As soon as possible of learning of a new dump area, start using it 126 dumpState = next->mDumpState != NULL ? next->mDumpState : &dummyDumpState; 127 teeSink = next->mTeeSink; 128 logWriter = next->mNBLogWriter != NULL ? next->mNBLogWriter : &dummyLogWriter; 129 130 // We want to always have a valid reference to the previous (non-idle) state. 131 // However, the state queue only guarantees access to current and previous states. 132 // So when there is a transition from a non-idle state into an idle state, we make a 133 // copy of the last known non-idle state so it is still available on return from idle. 134 // The possible transitions are: 135 // non-idle -> non-idle update previous from current in-place 136 // non-idle -> idle update previous from copy of current 137 // idle -> idle don't update previous 138 // idle -> non-idle don't update previous 139 if (!(current->mCommand & FastMixerState::IDLE)) { 140 if (command & FastMixerState::IDLE) { 141 preIdle = *current; 142 current = &preIdle; 143 oldTsValid = false; 144 oldLoadValid = false; 145 ignoreNextOverrun = true; 146 } 147 previous = current; 148 } 149 current = next; 150 } 151#if !LOG_NDEBUG 152 next = NULL; // not referenced again 153#endif 154 155 dumpState->mCommand = command; 156 157 switch (command) { 158 case FastMixerState::INITIAL: 159 case FastMixerState::HOT_IDLE: 160 sleepNs = FAST_HOT_IDLE_NS; 161 continue; 162 case FastMixerState::COLD_IDLE: 163 // only perform a cold idle command once 164 // FIXME consider checking previous state and only perform if previous != COLD_IDLE 165 if (current->mColdGen != coldGen) { 166 int32_t *coldFutexAddr = current->mColdFutexAddr; 167 ALOG_ASSERT(coldFutexAddr != NULL); 168 int32_t old = android_atomic_dec(coldFutexAddr); 169 if (old <= 0) { 170 logWriter->log("wait"); 171 __futex_syscall4(coldFutexAddr, FUTEX_WAIT_PRIVATE, old - 1, NULL); 172 } 173 // This may be overly conservative; there could be times that the normal mixer 174 // requests such a brief cold idle that it doesn't require resetting this flag. 175 isWarm = false; 176 measuredWarmupTs.tv_sec = 0; 177 measuredWarmupTs.tv_nsec = 0; 178 warmupCycles = 0; 179 sleepNs = -1; 180 coldGen = current->mColdGen; 181 bounds = 0; 182 full = false; 183 oldTsValid = !clock_gettime(CLOCK_MONOTONIC, &oldTs); 184 } else { 185 sleepNs = FAST_HOT_IDLE_NS; 186 } 187 continue; 188 case FastMixerState::EXIT: 189 logWriter->log("exit"); 190 delete mixer; 191 delete[] mixBuffer; 192 return false; 193 case FastMixerState::MIX: 194 case FastMixerState::WRITE: 195 case FastMixerState::MIX_WRITE: 196 break; 197 default: 198 LOG_FATAL("bad command %d", command); 199 } 200 201 // there is a non-idle state available to us; did the state change? 202 size_t frameCount = current->mFrameCount; 203 if (current != previous) { 204 205 // handle state change here, but since we want to diff the state, 206 // we're prepared for previous == &initial the first time through 207 unsigned previousTrackMask; 208 209 // check for change in output HAL configuration 210 NBAIO_Format previousFormat = format; 211 if (current->mOutputSinkGen != outputSinkGen) { 212 outputSink = current->mOutputSink; 213 outputSinkGen = current->mOutputSinkGen; 214 if (outputSink == NULL) { 215 format = Format_Invalid; 216 sampleRate = 0; 217 } else { 218 format = outputSink->format(); 219 sampleRate = Format_sampleRate(format); 220 ALOG_ASSERT(Format_channelCount(format) == 2); 221 } 222 dumpState->mSampleRate = sampleRate; 223 } 224 225 if ((format != previousFormat) || (frameCount != previous->mFrameCount)) { 226 // FIXME to avoid priority inversion, don't delete here 227 delete mixer; 228 mixer = NULL; 229 delete[] mixBuffer; 230 mixBuffer = NULL; 231 if (frameCount > 0 && sampleRate > 0) { 232 // FIXME new may block for unbounded time at internal mutex of the heap 233 // implementation; it would be better to have normal mixer allocate for us 234 // to avoid blocking here and to prevent possible priority inversion 235 mixer = new AudioMixer(frameCount, sampleRate, FastMixerState::kMaxFastTracks); 236 mixBuffer = new short[frameCount * 2]; 237 periodNs = (frameCount * 1000000000LL) / sampleRate; // 1.00 238 underrunNs = (frameCount * 1750000000LL) / sampleRate; // 1.75 239 overrunNs = (frameCount * 500000000LL) / sampleRate; // 0.50 240 forceNs = (frameCount * 950000000LL) / sampleRate; // 0.95 241 warmupNs = (frameCount * 500000000LL) / sampleRate; // 0.50 242 } else { 243 periodNs = 0; 244 underrunNs = 0; 245 overrunNs = 0; 246 forceNs = 0; 247 warmupNs = 0; 248 } 249 mixBufferState = UNDEFINED; 250#if !LOG_NDEBUG 251 for (i = 0; i < FastMixerState::kMaxFastTracks; ++i) { 252 fastTrackNames[i] = -1; 253 } 254#endif 255 // we need to reconfigure all active tracks 256 previousTrackMask = 0; 257 fastTracksGen = current->mFastTracksGen - 1; 258 dumpState->mFrameCount = frameCount; 259 } else { 260 previousTrackMask = previous->mTrackMask; 261 } 262 263 // check for change in active track set 264 unsigned currentTrackMask = current->mTrackMask; 265 dumpState->mTrackMask = currentTrackMask; 266 if (current->mFastTracksGen != fastTracksGen) { 267 logWriter->logf("gen %d", current->mFastTracksGen); 268 ALOG_ASSERT(mixBuffer != NULL); 269 int name; 270 271 // process removed tracks first to avoid running out of track names 272 unsigned removedTracks = previousTrackMask & ~currentTrackMask; 273 if (removedTracks) { 274 logWriter->logf("removed %#x", removedTracks); 275 } 276 while (removedTracks != 0) { 277 i = __builtin_ctz(removedTracks); 278 removedTracks &= ~(1 << i); 279 const FastTrack* fastTrack = ¤t->mFastTracks[i]; 280 ALOG_ASSERT(fastTrack->mBufferProvider == NULL); 281 if (mixer != NULL) { 282 name = fastTrackNames[i]; 283 ALOG_ASSERT(name >= 0); 284 mixer->deleteTrackName(name); 285 } 286#if !LOG_NDEBUG 287 fastTrackNames[i] = -1; 288#endif 289 // don't reset track dump state, since other side is ignoring it 290 generations[i] = fastTrack->mGeneration; 291 } 292 293 // now process added tracks 294 unsigned addedTracks = currentTrackMask & ~previousTrackMask; 295 if (addedTracks) { 296 logWriter->logf("added %#x", addedTracks); 297 } 298 while (addedTracks != 0) { 299 i = __builtin_ctz(addedTracks); 300 addedTracks &= ~(1 << i); 301 const FastTrack* fastTrack = ¤t->mFastTracks[i]; 302 AudioBufferProvider *bufferProvider = fastTrack->mBufferProvider; 303 ALOG_ASSERT(bufferProvider != NULL && fastTrackNames[i] == -1); 304 if (mixer != NULL) { 305 // calling getTrackName with default channel mask and a random invalid 306 // sessionId (no effects here) 307 name = mixer->getTrackName(AUDIO_CHANNEL_OUT_STEREO, -555); 308 ALOG_ASSERT(name >= 0); 309 fastTrackNames[i] = name; 310 mixer->setBufferProvider(name, bufferProvider); 311 mixer->setParameter(name, AudioMixer::TRACK, AudioMixer::MAIN_BUFFER, 312 (void *) mixBuffer); 313 // newly allocated track names default to full scale volume 314 if (fastTrack->mSampleRate != 0 && fastTrack->mSampleRate != sampleRate) { 315 mixer->setParameter(name, AudioMixer::RESAMPLE, 316 AudioMixer::SAMPLE_RATE, (void*) fastTrack->mSampleRate); 317 } 318 mixer->setParameter(name, AudioMixer::TRACK, AudioMixer::CHANNEL_MASK, 319 (void *) fastTrack->mChannelMask); 320 mixer->enable(name); 321 } 322 generations[i] = fastTrack->mGeneration; 323 } 324 325 // finally process modified tracks; these use the same slot 326 // but may have a different buffer provider or volume provider 327 unsigned modifiedTracks = currentTrackMask & previousTrackMask; 328 if (modifiedTracks) { 329 logWriter->logf("modified %#x", modifiedTracks); 330 } 331 while (modifiedTracks != 0) { 332 i = __builtin_ctz(modifiedTracks); 333 modifiedTracks &= ~(1 << i); 334 const FastTrack* fastTrack = ¤t->mFastTracks[i]; 335 if (fastTrack->mGeneration != generations[i]) { 336 AudioBufferProvider *bufferProvider = fastTrack->mBufferProvider; 337 ALOG_ASSERT(bufferProvider != NULL); 338 if (mixer != NULL) { 339 name = fastTrackNames[i]; 340 ALOG_ASSERT(name >= 0); 341 mixer->setBufferProvider(name, bufferProvider); 342 if (fastTrack->mVolumeProvider == NULL) { 343 mixer->setParameter(name, AudioMixer::VOLUME, AudioMixer::VOLUME0, 344 (void *)0x1000); 345 mixer->setParameter(name, AudioMixer::VOLUME, AudioMixer::VOLUME1, 346 (void *)0x1000); 347 } 348 if (fastTrack->mSampleRate != 0 && 349 fastTrack->mSampleRate != sampleRate) { 350 mixer->setParameter(name, AudioMixer::RESAMPLE, 351 AudioMixer::SAMPLE_RATE, (void*) fastTrack->mSampleRate); 352 } else { 353 mixer->setParameter(name, AudioMixer::RESAMPLE, 354 AudioMixer::REMOVE, NULL); 355 } 356 mixer->setParameter(name, AudioMixer::TRACK, AudioMixer::CHANNEL_MASK, 357 (void *) fastTrack->mChannelMask); 358 // already enabled 359 } 360 generations[i] = fastTrack->mGeneration; 361 } 362 } 363 364 fastTracksGen = current->mFastTracksGen; 365 366 dumpState->mNumTracks = popcount(currentTrackMask); 367 } 368 369#if 1 // FIXME shouldn't need this 370 // only process state change once 371 previous = current; 372#endif 373 } 374 375 // do work using current state here 376 if ((command & FastMixerState::MIX) && (mixer != NULL) && isWarm) { 377 ALOG_ASSERT(mixBuffer != NULL); 378 // for each track, update volume and check for underrun 379 unsigned currentTrackMask = current->mTrackMask; 380 while (currentTrackMask != 0) { 381 i = __builtin_ctz(currentTrackMask); 382 currentTrackMask &= ~(1 << i); 383 const FastTrack* fastTrack = ¤t->mFastTracks[i]; 384 int name = fastTrackNames[i]; 385 ALOG_ASSERT(name >= 0); 386 if (fastTrack->mVolumeProvider != NULL) { 387 uint32_t vlr = fastTrack->mVolumeProvider->getVolumeLR(); 388 mixer->setParameter(name, AudioMixer::VOLUME, AudioMixer::VOLUME0, 389 (void *)(vlr & 0xFFFF)); 390 mixer->setParameter(name, AudioMixer::VOLUME, AudioMixer::VOLUME1, 391 (void *)(vlr >> 16)); 392 } 393 // FIXME The current implementation of framesReady() for fast tracks 394 // takes a tryLock, which can block 395 // up to 1 ms. If enough active tracks all blocked in sequence, this would result 396 // in the overall fast mix cycle being delayed. Should use a non-blocking FIFO. 397 size_t framesReady = fastTrack->mBufferProvider->framesReady(); 398 if (ATRACE_ENABLED()) { 399 // I wish we had formatted trace names 400 char traceName[16]; 401 strcpy(traceName, "framesReady"); 402 traceName[11] = i + (i < 10 ? '0' : 'A' - 10); 403 traceName[12] = '\0'; 404 ATRACE_INT(traceName, framesReady); 405 } 406 FastTrackDump *ftDump = &dumpState->mTracks[i]; 407 FastTrackUnderruns underruns = ftDump->mUnderruns; 408 if (framesReady < frameCount) { 409 if (framesReady == 0) { 410 underruns.mBitFields.mEmpty++; 411 underruns.mBitFields.mMostRecent = UNDERRUN_EMPTY; 412 mixer->disable(name); 413 } else { 414 // allow mixing partial buffer 415 underruns.mBitFields.mPartial++; 416 underruns.mBitFields.mMostRecent = UNDERRUN_PARTIAL; 417 mixer->enable(name); 418 } 419 } else { 420 underruns.mBitFields.mFull++; 421 underruns.mBitFields.mMostRecent = UNDERRUN_FULL; 422 mixer->enable(name); 423 } 424 ftDump->mUnderruns = underruns; 425 ftDump->mFramesReady = framesReady; 426 } 427 428 int64_t pts; 429 if (outputSink == NULL || (OK != outputSink->getNextWriteTimestamp(&pts))) 430 pts = AudioBufferProvider::kInvalidPTS; 431 432 // process() is CPU-bound 433 mixer->process(pts); 434 mixBufferState = MIXED; 435 } else if (mixBufferState == MIXED) { 436 mixBufferState = UNDEFINED; 437 } 438 bool attemptedWrite = false; 439 //bool didFullWrite = false; // dumpsys could display a count of partial writes 440 if ((command & FastMixerState::WRITE) && (outputSink != NULL) && (mixBuffer != NULL)) { 441 if (mixBufferState == UNDEFINED) { 442 memset(mixBuffer, 0, frameCount * 2 * sizeof(short)); 443 mixBufferState = ZEROED; 444 } 445 if (teeSink != NULL) { 446 (void) teeSink->write(mixBuffer, frameCount); 447 } 448 // FIXME write() is non-blocking and lock-free for a properly implemented NBAIO sink, 449 // but this code should be modified to handle both non-blocking and blocking sinks 450 dumpState->mWriteSequence++; 451 ATRACE_BEGIN("write"); 452 ssize_t framesWritten = outputSink->write(mixBuffer, frameCount); 453 ATRACE_END(); 454 dumpState->mWriteSequence++; 455 if (framesWritten >= 0) { 456 ALOG_ASSERT(framesWritten <= frameCount); 457 dumpState->mFramesWritten += framesWritten; 458 //if ((size_t) framesWritten == frameCount) { 459 // didFullWrite = true; 460 //} 461 } else { 462 dumpState->mWriteErrors++; 463 } 464 attemptedWrite = true; 465 // FIXME count # of writes blocked excessively, CPU usage, etc. for dump 466 } 467 468 // To be exactly periodic, compute the next sleep time based on current time. 469 // This code doesn't have long-term stability when the sink is non-blocking. 470 // FIXME To avoid drift, use the local audio clock or watch the sink's fill status. 471 struct timespec newTs; 472 int rc = clock_gettime(CLOCK_MONOTONIC, &newTs); 473 if (rc == 0) { 474 logWriter->logTimestamp(newTs); 475 if (oldTsValid) { 476 time_t sec = newTs.tv_sec - oldTs.tv_sec; 477 long nsec = newTs.tv_nsec - oldTs.tv_nsec; 478 ALOGE_IF(sec < 0 || (sec == 0 && nsec < 0), 479 "clock_gettime(CLOCK_MONOTONIC) failed: was %ld.%09ld but now %ld.%09ld", 480 oldTs.tv_sec, oldTs.tv_nsec, newTs.tv_sec, newTs.tv_nsec); 481 if (nsec < 0) { 482 --sec; 483 nsec += 1000000000; 484 } 485 // To avoid an initial underrun on fast tracks after exiting standby, 486 // do not start pulling data from tracks and mixing until warmup is complete. 487 // Warmup is considered complete after the earlier of: 488 // MIN_WARMUP_CYCLES write() attempts and last one blocks for at least warmupNs 489 // MAX_WARMUP_CYCLES write() attempts. 490 // This is overly conservative, but to get better accuracy requires a new HAL API. 491 if (!isWarm && attemptedWrite) { 492 measuredWarmupTs.tv_sec += sec; 493 measuredWarmupTs.tv_nsec += nsec; 494 if (measuredWarmupTs.tv_nsec >= 1000000000) { 495 measuredWarmupTs.tv_sec++; 496 measuredWarmupTs.tv_nsec -= 1000000000; 497 } 498 ++warmupCycles; 499 if ((nsec > warmupNs && warmupCycles >= MIN_WARMUP_CYCLES) || 500 (warmupCycles >= MAX_WARMUP_CYCLES)) { 501 isWarm = true; 502 dumpState->mMeasuredWarmupTs = measuredWarmupTs; 503 dumpState->mWarmupCycles = warmupCycles; 504 } 505 } 506 sleepNs = -1; 507 if (isWarm) { 508 if (sec > 0 || nsec > underrunNs) { 509 ATRACE_NAME("underrun"); 510 // FIXME only log occasionally 511 ALOGV("underrun: time since last cycle %d.%03ld sec", 512 (int) sec, nsec / 1000000L); 513 dumpState->mUnderruns++; 514 ignoreNextOverrun = true; 515 } else if (nsec < overrunNs) { 516 if (ignoreNextOverrun) { 517 ignoreNextOverrun = false; 518 } else { 519 // FIXME only log occasionally 520 ALOGV("overrun: time since last cycle %d.%03ld sec", 521 (int) sec, nsec / 1000000L); 522 dumpState->mOverruns++; 523 } 524 // This forces a minimum cycle time. It: 525 // - compensates for an audio HAL with jitter due to sample rate conversion 526 // - works with a variable buffer depth audio HAL that never pulls at a rate 527 // < than overrunNs per buffer. 528 // - recovers from overrun immediately after underrun 529 // It doesn't work with a non-blocking audio HAL. 530 sleepNs = forceNs - nsec; 531 } else { 532 ignoreNextOverrun = false; 533 } 534 } 535#ifdef FAST_MIXER_STATISTICS 536 if (isWarm) { 537 // advance the FIFO queue bounds 538 size_t i = bounds & (FastMixerDumpState::kSamplingN - 1); 539 bounds = (bounds & 0xFFFF0000) | ((bounds + 1) & 0xFFFF); 540 if (full) { 541 bounds += 0x10000; 542 } else if (!(bounds & (FastMixerDumpState::kSamplingN - 1))) { 543 full = true; 544 } 545 // compute the delta value of clock_gettime(CLOCK_MONOTONIC) 546 uint32_t monotonicNs = nsec; 547 if (sec > 0 && sec < 4) { 548 monotonicNs += sec * 1000000000; 549 } 550 // compute the raw CPU load = delta value of clock_gettime(CLOCK_THREAD_CPUTIME_ID) 551 uint32_t loadNs = 0; 552 struct timespec newLoad; 553 rc = clock_gettime(CLOCK_THREAD_CPUTIME_ID, &newLoad); 554 if (rc == 0) { 555 if (oldLoadValid) { 556 sec = newLoad.tv_sec - oldLoad.tv_sec; 557 nsec = newLoad.tv_nsec - oldLoad.tv_nsec; 558 if (nsec < 0) { 559 --sec; 560 nsec += 1000000000; 561 } 562 loadNs = nsec; 563 if (sec > 0 && sec < 4) { 564 loadNs += sec * 1000000000; 565 } 566 } else { 567 // first time through the loop 568 oldLoadValid = true; 569 } 570 oldLoad = newLoad; 571 } 572#ifdef CPU_FREQUENCY_STATISTICS 573 // get the absolute value of CPU clock frequency in kHz 574 int cpuNum = sched_getcpu(); 575 uint32_t kHz = tcu.getCpukHz(cpuNum); 576 kHz = (kHz << 4) | (cpuNum & 0xF); 577#endif 578 // save values in FIFO queues for dumpsys 579 // these stores #1, #2, #3 are not atomic with respect to each other, 580 // or with respect to store #4 below 581 dumpState->mMonotonicNs[i] = monotonicNs; 582 dumpState->mLoadNs[i] = loadNs; 583#ifdef CPU_FREQUENCY_STATISTICS 584 dumpState->mCpukHz[i] = kHz; 585#endif 586 // this store #4 is not atomic with respect to stores #1, #2, #3 above, but 587 // the newest open and oldest closed halves are atomic with respect to each other 588 dumpState->mBounds = bounds; 589 ATRACE_INT("cycle_ms", monotonicNs / 1000000); 590 ATRACE_INT("load_us", loadNs / 1000); 591 } 592#endif 593 } else { 594 // first time through the loop 595 oldTsValid = true; 596 sleepNs = periodNs; 597 ignoreNextOverrun = true; 598 } 599 oldTs = newTs; 600 } else { 601 // monotonic clock is broken 602 oldTsValid = false; 603 sleepNs = periodNs; 604 } 605 606 607 } // for (;;) 608 609 // never return 'true'; Thread::_threadLoop() locks mutex which can result in priority inversion 610} 611 612FastMixerDumpState::FastMixerDumpState() : 613 mCommand(FastMixerState::INITIAL), mWriteSequence(0), mFramesWritten(0), 614 mNumTracks(0), mWriteErrors(0), mUnderruns(0), mOverruns(0), 615 mSampleRate(0), mFrameCount(0), /* mMeasuredWarmupTs({0, 0}), */ mWarmupCycles(0), 616 mTrackMask(0) 617#ifdef FAST_MIXER_STATISTICS 618 , mBounds(0) 619#endif 620{ 621 mMeasuredWarmupTs.tv_sec = 0; 622 mMeasuredWarmupTs.tv_nsec = 0; 623 // sample arrays aren't accessed atomically with respect to the bounds, 624 // so clearing reduces chance for dumpsys to read random uninitialized samples 625 memset(&mMonotonicNs, 0, sizeof(mMonotonicNs)); 626 memset(&mLoadNs, 0, sizeof(mLoadNs)); 627#ifdef CPU_FREQUENCY_STATISTICS 628 memset(&mCpukHz, 0, sizeof(mCpukHz)); 629#endif 630} 631 632FastMixerDumpState::~FastMixerDumpState() 633{ 634} 635 636// helper function called by qsort() 637static int compare_uint32_t(const void *pa, const void *pb) 638{ 639 uint32_t a = *(const uint32_t *)pa; 640 uint32_t b = *(const uint32_t *)pb; 641 if (a < b) { 642 return -1; 643 } else if (a > b) { 644 return 1; 645 } else { 646 return 0; 647 } 648} 649 650void FastMixerDumpState::dump(int fd) 651{ 652 if (mCommand == FastMixerState::INITIAL) { 653 fdprintf(fd, "FastMixer not initialized\n"); 654 return; 655 } 656#define COMMAND_MAX 32 657 char string[COMMAND_MAX]; 658 switch (mCommand) { 659 case FastMixerState::INITIAL: 660 strcpy(string, "INITIAL"); 661 break; 662 case FastMixerState::HOT_IDLE: 663 strcpy(string, "HOT_IDLE"); 664 break; 665 case FastMixerState::COLD_IDLE: 666 strcpy(string, "COLD_IDLE"); 667 break; 668 case FastMixerState::EXIT: 669 strcpy(string, "EXIT"); 670 break; 671 case FastMixerState::MIX: 672 strcpy(string, "MIX"); 673 break; 674 case FastMixerState::WRITE: 675 strcpy(string, "WRITE"); 676 break; 677 case FastMixerState::MIX_WRITE: 678 strcpy(string, "MIX_WRITE"); 679 break; 680 default: 681 snprintf(string, COMMAND_MAX, "%d", mCommand); 682 break; 683 } 684 double measuredWarmupMs = (mMeasuredWarmupTs.tv_sec * 1000.0) + 685 (mMeasuredWarmupTs.tv_nsec / 1000000.0); 686 double mixPeriodSec = (double) mFrameCount / (double) mSampleRate; 687 fdprintf(fd, "FastMixer command=%s writeSequence=%u framesWritten=%u\n" 688 " numTracks=%u writeErrors=%u underruns=%u overruns=%u\n" 689 " sampleRate=%u frameCount=%u measuredWarmup=%.3g ms, warmupCycles=%u\n" 690 " mixPeriod=%.2f ms\n", 691 string, mWriteSequence, mFramesWritten, 692 mNumTracks, mWriteErrors, mUnderruns, mOverruns, 693 mSampleRate, mFrameCount, measuredWarmupMs, mWarmupCycles, 694 mixPeriodSec * 1e3); 695#ifdef FAST_MIXER_STATISTICS 696 // find the interval of valid samples 697 uint32_t bounds = mBounds; 698 uint32_t newestOpen = bounds & 0xFFFF; 699 uint32_t oldestClosed = bounds >> 16; 700 uint32_t n = (newestOpen - oldestClosed) & 0xFFFF; 701 if (n > kSamplingN) { 702 ALOGE("too many samples %u", n); 703 n = kSamplingN; 704 } 705 // statistics for monotonic (wall clock) time, thread raw CPU load in time, CPU clock frequency, 706 // and adjusted CPU load in MHz normalized for CPU clock frequency 707 CentralTendencyStatistics wall, loadNs; 708#ifdef CPU_FREQUENCY_STATISTICS 709 CentralTendencyStatistics kHz, loadMHz; 710 uint32_t previousCpukHz = 0; 711#endif 712 // Assuming a normal distribution for cycle times, three standard deviations on either side of 713 // the mean account for 99.73% of the population. So if we take each tail to be 1/1000 of the 714 // sample set, we get 99.8% combined, or close to three standard deviations. 715 static const uint32_t kTailDenominator = 1000; 716 uint32_t *tail = n >= kTailDenominator ? new uint32_t[n] : NULL; 717 // loop over all the samples 718 for (uint32_t j = 0; j < n; ++j) { 719 size_t i = oldestClosed++ & (kSamplingN - 1); 720 uint32_t wallNs = mMonotonicNs[i]; 721 if (tail != NULL) { 722 tail[j] = wallNs; 723 } 724 wall.sample(wallNs); 725 uint32_t sampleLoadNs = mLoadNs[i]; 726 loadNs.sample(sampleLoadNs); 727#ifdef CPU_FREQUENCY_STATISTICS 728 uint32_t sampleCpukHz = mCpukHz[i]; 729 // skip bad kHz samples 730 if ((sampleCpukHz & ~0xF) != 0) { 731 kHz.sample(sampleCpukHz >> 4); 732 if (sampleCpukHz == previousCpukHz) { 733 double megacycles = (double) sampleLoadNs * (double) (sampleCpukHz >> 4) * 1e-12; 734 double adjMHz = megacycles / mixPeriodSec; // _not_ wallNs * 1e9 735 loadMHz.sample(adjMHz); 736 } 737 } 738 previousCpukHz = sampleCpukHz; 739#endif 740 } 741 fdprintf(fd, "Simple moving statistics over last %.1f seconds:\n", wall.n() * mixPeriodSec); 742 fdprintf(fd, " wall clock time in ms per mix cycle:\n" 743 " mean=%.2f min=%.2f max=%.2f stddev=%.2f\n", 744 wall.mean()*1e-6, wall.minimum()*1e-6, wall.maximum()*1e-6, wall.stddev()*1e-6); 745 fdprintf(fd, " raw CPU load in us per mix cycle:\n" 746 " mean=%.0f min=%.0f max=%.0f stddev=%.0f\n", 747 loadNs.mean()*1e-3, loadNs.minimum()*1e-3, loadNs.maximum()*1e-3, 748 loadNs.stddev()*1e-3); 749#ifdef CPU_FREQUENCY_STATISTICS 750 fdprintf(fd, " CPU clock frequency in MHz:\n" 751 " mean=%.0f min=%.0f max=%.0f stddev=%.0f\n", 752 kHz.mean()*1e-3, kHz.minimum()*1e-3, kHz.maximum()*1e-3, kHz.stddev()*1e-3); 753 fdprintf(fd, " adjusted CPU load in MHz (i.e. normalized for CPU clock frequency):\n" 754 " mean=%.1f min=%.1f max=%.1f stddev=%.1f\n", 755 loadMHz.mean(), loadMHz.minimum(), loadMHz.maximum(), loadMHz.stddev()); 756#endif 757 if (tail != NULL) { 758 qsort(tail, n, sizeof(uint32_t), compare_uint32_t); 759 // assume same number of tail samples on each side, left and right 760 uint32_t count = n / kTailDenominator; 761 CentralTendencyStatistics left, right; 762 for (uint32_t i = 0; i < count; ++i) { 763 left.sample(tail[i]); 764 right.sample(tail[n - (i + 1)]); 765 } 766 fdprintf(fd, "Distribution of mix cycle times in ms for the tails (> ~3 stddev outliers):\n" 767 " left tail: mean=%.2f min=%.2f max=%.2f stddev=%.2f\n" 768 " right tail: mean=%.2f min=%.2f max=%.2f stddev=%.2f\n", 769 left.mean()*1e-6, left.minimum()*1e-6, left.maximum()*1e-6, left.stddev()*1e-6, 770 right.mean()*1e-6, right.minimum()*1e-6, right.maximum()*1e-6, 771 right.stddev()*1e-6); 772 delete[] tail; 773 } 774#endif 775 // The active track mask and track states are updated non-atomically. 776 // So if we relied on isActive to decide whether to display, 777 // then we might display an obsolete track or omit an active track. 778 // Instead we always display all tracks, with an indication 779 // of whether we think the track is active. 780 uint32_t trackMask = mTrackMask; 781 fdprintf(fd, "Fast tracks: kMaxFastTracks=%u activeMask=%#x\n", 782 FastMixerState::kMaxFastTracks, trackMask); 783 fdprintf(fd, "Index Active Full Partial Empty Recent Ready\n"); 784 for (uint32_t i = 0; i < FastMixerState::kMaxFastTracks; ++i, trackMask >>= 1) { 785 bool isActive = trackMask & 1; 786 const FastTrackDump *ftDump = &mTracks[i]; 787 const FastTrackUnderruns& underruns = ftDump->mUnderruns; 788 const char *mostRecent; 789 switch (underruns.mBitFields.mMostRecent) { 790 case UNDERRUN_FULL: 791 mostRecent = "full"; 792 break; 793 case UNDERRUN_PARTIAL: 794 mostRecent = "partial"; 795 break; 796 case UNDERRUN_EMPTY: 797 mostRecent = "empty"; 798 break; 799 default: 800 mostRecent = "?"; 801 break; 802 } 803 fdprintf(fd, "%5u %6s %4u %7u %5u %7s %5u\n", i, isActive ? "yes" : "no", 804 (underruns.mBitFields.mFull) & UNDERRUN_MASK, 805 (underruns.mBitFields.mPartial) & UNDERRUN_MASK, 806 (underruns.mBitFields.mEmpty) & UNDERRUN_MASK, 807 mostRecent, ftDump->mFramesReady); 808 } 809} 810 811} // namespace android 812