You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

491 lines
20 KiB

/*
* Copyright (C) 2019 The Android Open Source Project
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
#pragma once
#include "Stream.h"
#include <condition_variable>
#include <future>
#include <list>
#include <map>
#include <memory>
#include <mutex>
#include <string>
#include <unordered_set>
#include <vector>
#include <utils/AndroidThreads.h>
namespace android::soundpool {
// TODO: Move helper classes to a utility file, with separate test.
/**
* JavaThread is used like std::thread but for threads that may call the JVM.
*
* std::thread does not easily attach to the JVM. We need JVM capable threads
* from createThreadEtc() since android binder call optimization may attempt to
* call back into Java if the SoundPool runs in system server.
*
*
* No locking is required - the member variables are inherently thread-safe.
*/
class JavaThread {
public:
JavaThread(std::function<void()> f, const char *name)
: mF{std::move(f)} {
createThreadEtc(staticFunction, this, name);
}
JavaThread(JavaThread &&) = delete; // uses "this" ptr, not moveable.
~JavaThread() {
join(); // manually block until the future is ready as std::future
// destructor doesn't block unless it comes from std::async
// and it is the last reference to shared state.
}
void join() const {
mFuture.wait();
}
bool isClosed() const {
return mIsClosed;
}
private:
static int staticFunction(void *data) {
JavaThread *jt = static_cast<JavaThread *>(data);
jt->mF();
jt->mIsClosed = true; // set the flag that we are closed
// now before we allow the destructor to execute;
// otherwise there may be a use after free.
jt->mPromise.set_value();
return 0;
}
// No locking is provided as these variables are initialized in the constructor
// and the members referenced are thread-safe objects.
// (mFuture.wait() can block multiple threads.)
// Note the order of member variables is reversed for destructor.
const std::function<void()> mF;
// Used in join() to block until the thread completes.
// See https://en.cppreference.com/w/cpp/thread/promise for the void specialization of
// promise.
std::promise<void> mPromise;
std::future<void> mFuture{mPromise.get_future()};
std::atomic_bool mIsClosed = false;
};
/**
* The ThreadPool manages thread lifetimes of SoundPool worker threads.
*
* TODO: the (eventual) goal of ThreadPool is to transparently and cooperatively
* maximize CPU utilization while avoiding starvation of other applications.
* Some possibilities:
*
* We should create worker threads when we have SoundPool work and the system is idle.
* CPU cycles are "use-it-or-lose-it" when the system is idle.
*
* We should adjust the priority of worker threads so that the second (and subsequent) worker
* threads have lower priority (should we try to promote priority also?).
*
* We should throttle the spawning of new worker threads, spacing over time, to avoid
* creating too many new threads all at once, on initialization.
*/
class ThreadPool {
public:
ThreadPool(size_t maxThreadCount, std::string name)
: mMaxThreadCount(maxThreadCount)
, mName{std::move(name)} { }
~ThreadPool() { quit(); }
size_t getActiveThreadCount() const { return mActiveThreadCount; }
size_t getMaxThreadCount() const { return mMaxThreadCount; }
void quit() {
std::list<std::unique_ptr<JavaThread>> threads;
{
std::lock_guard lock(mThreadLock);
if (mQuit) return; // already joined.
mQuit = true;
threads = std::move(mThreads);
mThreads.clear();
}
// mQuit set under lock, no more threads will be created.
for (auto &thread : threads) {
thread->join();
thread.reset();
}
LOG_ALWAYS_FATAL_IF(mActiveThreadCount != 0,
"Invalid Active Threads: %zu", (size_t)mActiveThreadCount);
}
// returns a non-zero id if successful, the id is to help logging messages.
int32_t launch(std::function<void(int32_t /* id */)> f) {
std::list<std::unique_ptr<JavaThread>> threadsToRelease; // release outside of lock.
std::lock_guard lock(mThreadLock);
if (mQuit) return 0; // ignore if we have quit
// clean up threads.
for (auto it = mThreads.begin(); it != mThreads.end(); ) {
if ((*it)->isClosed()) {
threadsToRelease.emplace_back(std::move(*it));
it = mThreads.erase(it);
} else {
++it;
}
}
const size_t threadCount = mThreads.size();
if (threadCount < mMaxThreadCount) {
// if the id wraps, we don't care about collisions. it's just for logging.
mNextThreadId = mNextThreadId == INT32_MAX ? 1 : ++mNextThreadId;
const int32_t id = mNextThreadId;
mThreads.emplace_back(std::make_unique<JavaThread>(
[this, id, mf = std::move(f)] { mf(id); --mActiveThreadCount; },
(mName + std::to_string(id)).c_str()));
++mActiveThreadCount;
return id;
}
return 0;
}
// TODO: launch only if load average is low.
// This gets the load average
// See also std::thread::hardware_concurrency() for the concurrent capability.
static double getLoadAvg() {
double loadAvg[1];
if (getloadavg(loadAvg, std::size(loadAvg)) > 0) {
return loadAvg[0];
}
return -1.;
}
private:
const size_t mMaxThreadCount;
const std::string mName;
std::atomic_size_t mActiveThreadCount = 0;
std::mutex mThreadLock;
bool mQuit GUARDED_BY(mThreadLock) = false;
int32_t mNextThreadId GUARDED_BY(mThreadLock) = 0;
std::list<std::unique_ptr<JavaThread>> mThreads GUARDED_BY(mThreadLock);
};
/**
* A Perfect HashTable for IDs (key) to pointers (value).
*
* There are no collisions. Why? because we generate the IDs for you to look up :-).
*
* The goal of this hash table is to map an integer ID handle > 0 to a pointer.
* We give these IDs in monotonic order (though we may skip if it were to cause a collision).
*
* The size of the hashtable must be large enough to accommodate the max number of keys.
* We suggest 2x.
*
* Readers are lockless
* Single writer could be lockless, but we allow multiple writers through an internal lock.
*
* For the Key type K, valid keys generated are > 0 (signed or unsigned)
* For the Value type V, values are pointers - nullptr means empty.
*/
template <typename K, typename V>
class PerfectHash {
public:
PerfectHash(size_t hashCapacity)
: mHashCapacity(hashCapacity)
, mK2V{new std::atomic<V>[hashCapacity]()} {
}
// Generate a key for a value V.
// There is a testing function getKforV() which checks what the value reports as its key.
//
// Calls back into getKforV under lock.
//
// We expect that the hashCapacity is 2x the number of stored keys in order
// to have one or two tries to find an empty slot
K generateKey(V value, std::function<K(V)> getKforV, K oldKey = 0) {
std::lock_guard lock(mHashLock);
// try to remove the old key.
if (oldKey > 0) { // key valid
const V v = getValue(oldKey);
if (v != nullptr) { // value still valid
const K atPosition = getKforV(v);
if (atPosition < 0 || // invalid value
atPosition == oldKey || // value's key still valid and matches old key
((atPosition ^ oldKey) & (mHashCapacity - 1)) != 0) { // stale key entry
getValue(oldKey) = nullptr; // invalidate
}
} // else if value is invalid, no need to invalidate.
}
// check if we are invalidating only.
if (value == nullptr) return 0;
// now insert the new value and return the key.
size_t tries = 0;
for (; tries < mHashCapacity; ++tries) {
mNextKey = mNextKey == std::numeric_limits<K>::max() ? 1 : mNextKey + 1;
const V v = getValue(mNextKey);
//ALOGD("tries: %zu, key:%d value:%p", tries, (int)mNextKey, v);
if (v == nullptr) break; // empty
const K atPosition = getKforV(v);
//ALOGD("tries: %zu key atPosition:%d", tries, (int)atPosition);
if (atPosition < 0 || // invalid value
((atPosition ^ mNextKey) & (mHashCapacity - 1)) != 0) { // stale key entry
break;
}
}
LOG_ALWAYS_FATAL_IF(tries == mHashCapacity, "hash table overflow!");
//ALOGD("%s: found after %zu tries", __func__, tries);
getValue(mNextKey) = value;
return mNextKey;
}
std::atomic<V> &getValue(K key) { return mK2V[key & (mHashCapacity - 1)]; }
const std::atomic_int32_t &getValue(K key) const { return mK2V[key & (mHashCapacity - 1)]; }
private:
mutable std::mutex mHashLock;
const size_t mHashCapacity; // size of mK2V no lock needed.
std::unique_ptr<std::atomic<V>[]> mK2V; // no lock needed for read access.
K mNextKey GUARDED_BY(mHashLock) {};
};
/**
* StreamMap contains the all the valid streams available to SoundPool.
*
* There is no Lock required for this class because the streams are
* allocated in the constructor, the lookup is lockless, and the Streams
* returned are locked internally.
*
* The lookup uses a perfect hash.
* It is possible to use a lockless hash table or to use a stripe-locked concurrent
* hashmap for essentially lock-free lookup.
*
* This follows Map-Reduce parallelism model.
* https://en.wikipedia.org/wiki/MapReduce
*
* Conceivably the forEach could be parallelized using std::for_each with a
* std::execution::par policy.
*
* https://en.cppreference.com/w/cpp/algorithm/for_each
*/
class StreamMap {
public:
explicit StreamMap(int32_t streams);
// Returns the stream associated with streamID or nullptr if not found.
// This need not be locked.
// The stream ID will never migrate to another Stream, but it may change
// underneath you. The Stream operations that take a streamID will confirm
// that the streamID matches under the Stream lock before executing otherwise
// it ignores the command as stale.
Stream* findStream(int32_t streamID) const;
// Iterates through the stream pool applying the function f.
// Since this enumerates over every single stream, it is unlocked.
//
// See related: https://en.cppreference.com/w/cpp/algorithm/for_each
void forEach(std::function<void(const Stream *)>f) const {
for (size_t i = 0; i < mStreamPoolSize; ++i) {
f(&mStreamPool[i]);
}
}
void forEach(std::function<void(Stream *)>f) {
for (size_t i = 0; i < mStreamPoolSize; ++i) {
f(&mStreamPool[i]);
}
}
// Returns the pair stream for a given Stream.
// This need not be locked as it is a property of the pointer address.
Stream* getPairStream(const Stream* stream) const {
const size_t index = streamPosition(stream);
return &mStreamPool[index ^ 1];
}
// find the position of the stream in mStreamPool array.
size_t streamPosition(const Stream* stream) const; // no lock needed
size_t getStreamMapSize() const {
return mStreamPoolSize;
}
// find the next valid ID for a stream and store in hash table.
int32_t getNextIdForStream(Stream* stream) const;
private:
// use the hash table to attempt to find the stream.
// nullptr is returned if the lookup fails.
Stream* lookupStreamFromId(int32_t streamID) const;
// The stream pool is initialized in the constructor, effectively const.
// no locking required for access.
//
// The constructor parameter "streams" results in streams pairs of streams.
// We have twice as many streams because we wish to return a streamID "handle"
// back to the app immediately, while we may be stopping the other stream in the
// pair to get its AudioTrack :-).
//
// Of the stream pair, only one of the streams may have an AudioTrack.
// The fixed association of a stream pair allows callbacks from the AudioTrack
// to be associated properly to either one or the other of the stream pair.
//
// TODO: The stream pair arrangement can be removed if we have better AudioTrack
// callback handling (being able to remove and change the callback after construction).
//
// Streams may be accessed anytime off of the stream pool
// as there is internal locking on each stream.
std::unique_ptr<Stream[]> mStreamPool; // no lock needed for access.
size_t mStreamPoolSize; // no lock needed for access.
// In order to find the Stream from a StreamID, we could do a linear lookup in mStreamPool.
// As an alternative, one could use stripe-locked or lock-free concurrent hashtables.
//
// When considering linear search vs hashmap, verify the typical use-case size.
// Linear search is faster than std::unordered_map (circa 2018) for less than 40 elements.
// [ Skarupke, M. (2018), "You Can Do Better than std::unordered_map: New and Recent
// Improvements to Hash Table Performance." C++Now 2018. cppnow.org, see
// https://www.youtube.com/watch?v=M2fKMP47slQ ]
//
// Here, we use a PerfectHash of Id to Stream *, since we can control the
// StreamID returned to the user. This allows O(1) read access to mStreamPool lock-free.
//
// We prefer that the next stream ID is monotonic for aesthetic reasons
// (if we didn't care about monotonicity, a simple method is to apply a generation count
// to each stream in the unused upper bits of its index in mStreamPool for the id).
//
std::unique_ptr<PerfectHash<int32_t, Stream *>> mPerfectHash;
};
/**
* StreamManager is used to manage the streams (accessed by StreamID from Java).
*
* Locking order (proceeds from application to component).
* SoundPool mApiLock (if needed) -> StreamManager mStreamManagerLock
* -> pair Stream mLock -> queued Stream mLock
*/
class StreamManager : public StreamMap {
public:
// Note: the SoundPool pointer is only used for stream initialization.
// It is not stored in StreamManager.
StreamManager(int32_t streams, size_t threads, const audio_attributes_t* attributes,
std::string opPackageName);
~StreamManager();
// Returns positive streamID on success, 0 on failure. This is locked.
int32_t queueForPlay(const std::shared_ptr<Sound> &sound,
int32_t soundID, float leftVolume, float rightVolume,
int32_t priority, int32_t loop, float rate)
NO_THREAD_SAFETY_ANALYSIS; // uses unique_lock
///////////////////////////////////////////////////////////////////////
// Called from soundpool::Stream
const audio_attributes_t* getAttributes() const { return &mAttributes; }
const std::string& getOpPackageName() const { return mOpPackageName; }
// Moves the stream to the restart queue (called upon BUFFER_END of the static track)
// this is locked internally.
// If activeStreamIDToMatch is nonzero, it will only move to the restart queue
// if the streamIDToMatch is found on the active queue.
void moveToRestartQueue(Stream* stream, int32_t activeStreamIDToMatch = 0);
private:
void run(int32_t id) NO_THREAD_SAFETY_ANALYSIS; // worker thread, takes unique_lock.
void dump() const; // no lock needed
// returns true if more worker threads are needed.
bool needMoreThreads_l() REQUIRES(mStreamManagerLock) {
return mRestartStreams.size() > 0 &&
(mThreadPool->getActiveThreadCount() == 0
|| std::distance(mRestartStreams.begin(),
mRestartStreams.upper_bound(systemTime()))
> (ptrdiff_t)mThreadPool->getActiveThreadCount());
}
// returns true if the stream was added.
bool moveToRestartQueue_l(
Stream* stream, int32_t activeStreamIDToMatch = 0) REQUIRES(mStreamManagerLock);
// returns number of queues the stream was removed from (should be 0 or 1);
// a special code of -1 is returned if activeStreamIDToMatch is > 0 and
// the stream wasn't found on the active queue.
ssize_t removeFromQueues_l(
Stream* stream, int32_t activeStreamIDToMatch = 0) REQUIRES(mStreamManagerLock);
void addToRestartQueue_l(Stream *stream) REQUIRES(mStreamManagerLock);
void addToActiveQueue_l(Stream *stream) REQUIRES(mStreamManagerLock);
void sanityCheckQueue_l() const REQUIRES(mStreamManagerLock);
const audio_attributes_t mAttributes;
const std::string mOpPackageName;
// For legacy compatibility, we lock the stream manager on stop when
// there is only one stream. This allows a play to be called immediately
// after stopping, otherwise it is possible that the play might be discarded
// (returns 0) because that stream may be in the worker thread call to stop.
const bool mLockStreamManagerStop;
std::unique_ptr<ThreadPool> mThreadPool; // locked internally
// mStreamManagerLock is used to lock access for transitions between the
// 4 stream queues by the Manager Thread or by the user initiated play().
// A stream pair has exactly one stream on exactly one of the queues.
std::mutex mStreamManagerLock;
std::condition_variable mStreamManagerCondition GUARDED_BY(mStreamManagerLock);
bool mQuit GUARDED_BY(mStreamManagerLock) = false;
// There are constructor arg "streams" pairs of streams, only one of each
// pair on the 4 stream queues below. The other stream in the pair serves as
// placeholder to accumulate user changes, pending actual availability of the
// AudioTrack, as it may be in use, requiring stop-then-restart.
//
// The 4 queues are implemented in the appropriate STL container based on perceived
// optimality.
// 1) mRestartStreams: Streams awaiting stop.
// The paired stream may be active (but with no AudioTrack), and will be restarted
// with an active AudioTrack when the current stream is stopped.
std::multimap<int64_t /* stopTimeNs */, Stream*>
mRestartStreams GUARDED_BY(mStreamManagerLock);
// 2) mActiveStreams: Streams that are active.
// The paired stream will be inactive.
// This is in order of specified by kStealActiveStream_OldestFirst
std::list<Stream*> mActiveStreams GUARDED_BY(mStreamManagerLock);
// 3) mAvailableStreams: Streams that are inactive.
// The paired stream will also be inactive.
// No particular order.
std::unordered_set<Stream*> mAvailableStreams GUARDED_BY(mStreamManagerLock);
// 4) mProcessingStreams: Streams that are being processed by the ManagerThreads
// When on this queue, the stream and its pair are not available for stealing.
// Each ManagerThread will have at most one stream on the mProcessingStreams queue.
// The paired stream may be active or restarting.
// No particular order.
std::unordered_set<Stream*> mProcessingStreams GUARDED_BY(mStreamManagerLock);
};
} // namespace android::soundpool