FastThread.cpp revision e4a7ce250cb94a00aa2f76e5edca1c4479dc5401
1/*
2 * Copyright (C) 2014 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#define LOG_TAG "FastThread"
18//#define LOG_NDEBUG 0
19
20#define ATRACE_TAG ATRACE_TAG_AUDIO
21
22#include "Configuration.h"
23#include <linux/futex.h>
24#include <sys/syscall.h>
25#include <utils/Log.h>
26#include <utils/Trace.h>
27#include "FastThread.h"
28#include "FastThreadDumpState.h"
29
30#define FAST_DEFAULT_NS    999999999L   // ~1 sec: default time to sleep
31#define FAST_HOT_IDLE_NS     1000000L   // 1 ms: time to sleep while hot idling
32#define MIN_WARMUP_CYCLES          2    // minimum number of consecutive in-range loop cycles
33                                        // to wait for warmup
34#define MAX_WARMUP_CYCLES         10    // maximum number of loop cycles to wait for warmup
35
36namespace android {
37
38FastThread::FastThread() : Thread(false /*canCallJava*/),
39    // re-initialized to &sInitial by subclass constructor
40    mPrevious(NULL), mCurrent(NULL),
41    /* mOldTs({0, 0}), */
42    mOldTsValid(false),
43    mSleepNs(-1),
44    mPeriodNs(0),
45    mUnderrunNs(0),
46    mOverrunNs(0),
47    mForceNs(0),
48    mWarmupNsMin(0),
49    mWarmupNsMax(LONG_MAX),
50    // re-initialized to &mDummySubclassDumpState by subclass constructor
51    mDummyDumpState(NULL),
52    mDumpState(NULL),
53    mIgnoreNextOverrun(true),
54#ifdef FAST_THREAD_STATISTICS
55    // mOldLoad
56    mOldLoadValid(false),
57    mBounds(0),
58    mFull(false),
59    // mTcu
60#endif
61    mColdGen(0),
62    mIsWarm(false),
63    /* mMeasuredWarmupTs({0, 0}), */
64    mWarmupCycles(0),
65    mWarmupConsecutiveInRangeCycles(0),
66    // mDummyLogWriter
67    mLogWriter(&mDummyLogWriter),
68    mTimestampStatus(INVALID_OPERATION),
69
70    mCommand(FastThreadState::INITIAL),
71#if 0
72    frameCount(0),
73#endif
74    mAttemptedWrite(false)
75{
76    mOldTs.tv_sec = 0;
77    mOldTs.tv_nsec = 0;
78    mMeasuredWarmupTs.tv_sec = 0;
79    mMeasuredWarmupTs.tv_nsec = 0;
80}
81
82FastThread::~FastThread()
83{
84}
85
86bool FastThread::threadLoop()
87{
88    for (;;) {
89
90        // either nanosleep, sched_yield, or busy wait
91        if (mSleepNs >= 0) {
92            if (mSleepNs > 0) {
93                ALOG_ASSERT(mSleepNs < 1000000000);
94                const struct timespec req = {0, mSleepNs};
95                nanosleep(&req, NULL);
96            } else {
97                sched_yield();
98            }
99        }
100        // default to long sleep for next cycle
101        mSleepNs = FAST_DEFAULT_NS;
102
103        // poll for state change
104        const FastThreadState *next = poll();
105        if (next == NULL) {
106            // continue to use the default initial state until a real state is available
107            // FIXME &sInitial not available, should save address earlier
108            //ALOG_ASSERT(mCurrent == &sInitial && previous == &sInitial);
109            next = mCurrent;
110        }
111
112        mCommand = next->mCommand;
113        if (next != mCurrent) {
114
115            // As soon as possible of learning of a new dump area, start using it
116            mDumpState = next->mDumpState != NULL ? next->mDumpState : mDummyDumpState;
117            mLogWriter = next->mNBLogWriter != NULL ? next->mNBLogWriter : &mDummyLogWriter;
118            setLog(mLogWriter);
119
120            // We want to always have a valid reference to the previous (non-idle) state.
121            // However, the state queue only guarantees access to current and previous states.
122            // So when there is a transition from a non-idle state into an idle state, we make a
123            // copy of the last known non-idle state so it is still available on return from idle.
124            // The possible transitions are:
125            //  non-idle -> non-idle    update previous from current in-place
126            //  non-idle -> idle        update previous from copy of current
127            //  idle     -> idle        don't update previous
128            //  idle     -> non-idle    don't update previous
129            if (!(mCurrent->mCommand & FastThreadState::IDLE)) {
130                if (mCommand & FastThreadState::IDLE) {
131                    onIdle();
132                    mOldTsValid = false;
133#ifdef FAST_THREAD_STATISTICS
134                    mOldLoadValid = false;
135#endif
136                    mIgnoreNextOverrun = true;
137                }
138                mPrevious = mCurrent;
139            }
140            mCurrent = next;
141        }
142#if !LOG_NDEBUG
143        next = NULL;    // not referenced again
144#endif
145
146        mDumpState->mCommand = mCommand;
147
148        // FIXME what does this comment mean?
149        // << current, previous, command, dumpState >>
150
151        switch (mCommand) {
152        case FastThreadState::INITIAL:
153        case FastThreadState::HOT_IDLE:
154            mSleepNs = FAST_HOT_IDLE_NS;
155            continue;
156        case FastThreadState::COLD_IDLE:
157            // only perform a cold idle command once
158            // FIXME consider checking previous state and only perform if previous != COLD_IDLE
159            if (mCurrent->mColdGen != mColdGen) {
160                int32_t *coldFutexAddr = mCurrent->mColdFutexAddr;
161                ALOG_ASSERT(coldFutexAddr != NULL);
162                int32_t old = android_atomic_dec(coldFutexAddr);
163                if (old <= 0) {
164                    syscall(__NR_futex, coldFutexAddr, FUTEX_WAIT_PRIVATE, old - 1, NULL);
165                }
166                int policy = sched_getscheduler(0);
167                if (!(policy == SCHED_FIFO || policy == SCHED_RR)) {
168                    ALOGE("did not receive expected priority boost");
169                }
170                // This may be overly conservative; there could be times that the normal mixer
171                // requests such a brief cold idle that it doesn't require resetting this flag.
172                mIsWarm = false;
173                mMeasuredWarmupTs.tv_sec = 0;
174                mMeasuredWarmupTs.tv_nsec = 0;
175                mWarmupCycles = 0;
176                mWarmupConsecutiveInRangeCycles = 0;
177                mSleepNs = -1;
178                mColdGen = mCurrent->mColdGen;
179#ifdef FAST_THREAD_STATISTICS
180                mBounds = 0;
181                mFull = false;
182#endif
183                mOldTsValid = !clock_gettime(CLOCK_MONOTONIC, &mOldTs);
184                mTimestampStatus = INVALID_OPERATION;
185            } else {
186                mSleepNs = FAST_HOT_IDLE_NS;
187            }
188            continue;
189        case FastThreadState::EXIT:
190            onExit();
191            return false;
192        default:
193            LOG_ALWAYS_FATAL_IF(!isSubClassCommand(mCommand));
194            break;
195        }
196
197        // there is a non-idle state available to us; did the state change?
198        if (mCurrent != mPrevious) {
199            onStateChange();
200#if 1   // FIXME shouldn't need this
201            // only process state change once
202            mPrevious = mCurrent;
203#endif
204        }
205
206        // do work using current state here
207        mAttemptedWrite = false;
208        onWork();
209
210        // To be exactly periodic, compute the next sleep time based on current time.
211        // This code doesn't have long-term stability when the sink is non-blocking.
212        // FIXME To avoid drift, use the local audio clock or watch the sink's fill status.
213        struct timespec newTs;
214        int rc = clock_gettime(CLOCK_MONOTONIC, &newTs);
215        if (rc == 0) {
216            //mLogWriter->logTimestamp(newTs);
217            if (mOldTsValid) {
218                time_t sec = newTs.tv_sec - mOldTs.tv_sec;
219                long nsec = newTs.tv_nsec - mOldTs.tv_nsec;
220                ALOGE_IF(sec < 0 || (sec == 0 && nsec < 0),
221                        "clock_gettime(CLOCK_MONOTONIC) failed: was %ld.%09ld but now %ld.%09ld",
222                        mOldTs.tv_sec, mOldTs.tv_nsec, newTs.tv_sec, newTs.tv_nsec);
223                if (nsec < 0) {
224                    --sec;
225                    nsec += 1000000000;
226                }
227                // To avoid an initial underrun on fast tracks after exiting standby,
228                // do not start pulling data from tracks and mixing until warmup is complete.
229                // Warmup is considered complete after the earlier of:
230                //      MIN_WARMUP_CYCLES consecutive in-range write() attempts,
231                //          where "in-range" means mWarmupNsMin <= cycle time <= mWarmupNsMax
232                //      MAX_WARMUP_CYCLES write() attempts.
233                // This is overly conservative, but to get better accuracy requires a new HAL API.
234                if (!mIsWarm && mAttemptedWrite) {
235                    mMeasuredWarmupTs.tv_sec += sec;
236                    mMeasuredWarmupTs.tv_nsec += nsec;
237                    if (mMeasuredWarmupTs.tv_nsec >= 1000000000) {
238                        mMeasuredWarmupTs.tv_sec++;
239                        mMeasuredWarmupTs.tv_nsec -= 1000000000;
240                    }
241                    ++mWarmupCycles;
242                    if (mWarmupNsMin <= nsec && nsec <= mWarmupNsMax) {
243                        ALOGV("warmup cycle %d in range: %.03f ms", mWarmupCycles, nsec * 1e-9);
244                        ++mWarmupConsecutiveInRangeCycles;
245                    } else {
246                        ALOGV("warmup cycle %d out of range: %.03f ms", mWarmupCycles, nsec * 1e-9);
247                        mWarmupConsecutiveInRangeCycles = 0;
248                    }
249                    if ((mWarmupConsecutiveInRangeCycles >= MIN_WARMUP_CYCLES) ||
250                            (mWarmupCycles >= MAX_WARMUP_CYCLES)) {
251                        mIsWarm = true;
252                        mDumpState->mMeasuredWarmupTs = mMeasuredWarmupTs;
253                        mDumpState->mWarmupCycles = mWarmupCycles;
254                    }
255                }
256                mSleepNs = -1;
257                if (mIsWarm) {
258                    if (sec > 0 || nsec > mUnderrunNs) {
259                        ATRACE_NAME("underrun");
260                        // FIXME only log occasionally
261                        ALOGV("underrun: time since last cycle %d.%03ld sec",
262                                (int) sec, nsec / 1000000L);
263                        mDumpState->mUnderruns++;
264                        mIgnoreNextOverrun = true;
265                    } else if (nsec < mOverrunNs) {
266                        if (mIgnoreNextOverrun) {
267                            mIgnoreNextOverrun = false;
268                        } else {
269                            // FIXME only log occasionally
270                            ALOGV("overrun: time since last cycle %d.%03ld sec",
271                                    (int) sec, nsec / 1000000L);
272                            mDumpState->mOverruns++;
273                        }
274                        // This forces a minimum cycle time. It:
275                        //  - compensates for an audio HAL with jitter due to sample rate conversion
276                        //  - works with a variable buffer depth audio HAL that never pulls at a
277                        //    rate < than mOverrunNs per buffer.
278                        //  - recovers from overrun immediately after underrun
279                        // It doesn't work with a non-blocking audio HAL.
280                        mSleepNs = mForceNs - nsec;
281                    } else {
282                        mIgnoreNextOverrun = false;
283                    }
284                }
285#ifdef FAST_THREAD_STATISTICS
286                if (mIsWarm) {
287                    // advance the FIFO queue bounds
288                    size_t i = mBounds & (mDumpState->mSamplingN - 1);
289                    mBounds = (mBounds & 0xFFFF0000) | ((mBounds + 1) & 0xFFFF);
290                    if (mFull) {
291                        mBounds += 0x10000;
292                    } else if (!(mBounds & (mDumpState->mSamplingN - 1))) {
293                        mFull = true;
294                    }
295                    // compute the delta value of clock_gettime(CLOCK_MONOTONIC)
296                    uint32_t monotonicNs = nsec;
297                    if (sec > 0 && sec < 4) {
298                        monotonicNs += sec * 1000000000;
299                    }
300                    // compute raw CPU load = delta value of clock_gettime(CLOCK_THREAD_CPUTIME_ID)
301                    uint32_t loadNs = 0;
302                    struct timespec newLoad;
303                    rc = clock_gettime(CLOCK_THREAD_CPUTIME_ID, &newLoad);
304                    if (rc == 0) {
305                        if (mOldLoadValid) {
306                            sec = newLoad.tv_sec - mOldLoad.tv_sec;
307                            nsec = newLoad.tv_nsec - mOldLoad.tv_nsec;
308                            if (nsec < 0) {
309                                --sec;
310                                nsec += 1000000000;
311                            }
312                            loadNs = nsec;
313                            if (sec > 0 && sec < 4) {
314                                loadNs += sec * 1000000000;
315                            }
316                        } else {
317                            // first time through the loop
318                            mOldLoadValid = true;
319                        }
320                        mOldLoad = newLoad;
321                    }
322#ifdef CPU_FREQUENCY_STATISTICS
323                    // get the absolute value of CPU clock frequency in kHz
324                    int cpuNum = sched_getcpu();
325                    uint32_t kHz = mTcu.getCpukHz(cpuNum);
326                    kHz = (kHz << 4) | (cpuNum & 0xF);
327#endif
328                    // save values in FIFO queues for dumpsys
329                    // these stores #1, #2, #3 are not atomic with respect to each other,
330                    // or with respect to store #4 below
331                    mDumpState->mMonotonicNs[i] = monotonicNs;
332                    mDumpState->mLoadNs[i] = loadNs;
333#ifdef CPU_FREQUENCY_STATISTICS
334                    mDumpState->mCpukHz[i] = kHz;
335#endif
336                    // this store #4 is not atomic with respect to stores #1, #2, #3 above, but
337                    // the newest open & oldest closed halves are atomic with respect to each other
338                    mDumpState->mBounds = mBounds;
339                    ATRACE_INT("cycle_ms", monotonicNs / 1000000);
340                    ATRACE_INT("load_us", loadNs / 1000);
341                }
342#endif
343            } else {
344                // first time through the loop
345                mOldTsValid = true;
346                mSleepNs = mPeriodNs;
347                mIgnoreNextOverrun = true;
348            }
349            mOldTs = newTs;
350        } else {
351            // monotonic clock is broken
352            mOldTsValid = false;
353            mSleepNs = mPeriodNs;
354        }
355
356    }   // for (;;)
357
358    // never return 'true'; Thread::_threadLoop() locks mutex which can result in priority inversion
359}
360
361}   // namespace android
362