Even if a developer synchronizes every method, thread safety is still not guaranteed
because the internal callback methods can't be synced from outside.
Issue:#3773
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=184122072
1. Add string for STATE_CANCELED. Lint doesn't like that the
switch statement on the state IntDef doesn't have a case
for STATE_CANCELED. May as well add one, even if we're not
planning on our demo app showing notifications for this
state.
2. Replace non-human-readable error message with one provided
by ErrorMessageProvider.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=184120892
In the case of the components we deliberately access via
reflection, it's normal that they might not be resolved
due to proguarding (i.e. if the app isn't being built to
include them). Don't note their omission.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=184120611
I think (?) they're harmless, but lint doesn't like them.
Using them within the class body means the TargetApi
annotation applies, which makes lint happy.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=184117951
In MediaCodecRenderer, we currently uses codec.getInput/OutputBuffers event
though these APIs are deprecated and are not recommended from API 21+. This
change makes sure that:
- On API 20 and below, we will keep using codec.getInput/OutputBuffers.
- On API 21+, we will use getInput/OutputBuffer(index) APIs instead.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=184112329
*** Reason for rollback ***
Broke everything
*** Original change description ***
Clean up message naming in EPII
***
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=184061352
This make sure all media sources can be reprepared after being released.
Issue:#3498
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=183990416
When the dynamic media source contains multiple empty timelines in a row and some
of them dynamically change to a non-empty timeline, the window and period indices
are not updated correctly because the index of the changed child source is wrong.
To fix this bug, the child index is added to the media period holder to have direct
access on the current child index to prevent ambiguity.
Furthermore, the uid is changed to be the hash code of the MediaSourceHolder not the
MediaSource itself to allow adding the same MediaSource twice without violating the
unique uid policy.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=183973462
Also fixes a bug where deferred media periods were kept in the list for
an unprepared media source although the media period was already released.
Issue:#3498
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=183800889
This enables caching manifest files for DASH, HLS and SmoothStreaming.
To disable caching a non cache DataSource should be provided for reading
manifest to the used MediaSource.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=183794252
In startReadWrite*() methods a new CachedContent is created if the there
isn't one already for the given key. If the span is release without
writing any content, this fix removes the added CachedContent.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=183666821
Before this change, the ad playback state stored the number of played ads in
each ad group. There was no way to represent that an ad had failed to load (and
it wouldn't be possible just to increment the played ad count to signal a load
error because there might be an unplayed ad before the ad that failed to load).
Represent the state of each ad (unavailable, available, skipped, played, error)
in each ad group. In a later change the player will use this information to
update its loaded MediaPeriods in response to future ads failing to load.
Also make the AdPlaybackState immutable and remove copying/duplication of its
fields in the ad timeline and period.
Issue: #3584
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=183655308
1. When we try and load something via reflection and find the
class, always throw rather than failing silently if we
subsequently fail to instantiate an instance. This is
indicative of a broken proguard setup, and failing silently
makes it hard to spot.
2. Add library/core proguard configuration to ensure extension
renderer constructors that we access via reflection are kept.
3. Add demos/main proguard configuration to ensure ImaAdsLoader
constructor that we access via reflection is kept.
4. Added IMA proguard file to hopefully fix#3723, although I
wasn't actually able to reproduce the issue.
Issue: #3723
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=183648187
- Renderers becoming ready is asynchronous, so the change wasn't
well thought through :(.
- This will bring back the possibility of getting stuck in the
buffering-but-not-loading anything state. This will need to be
addressed in a future CL.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=183646837
100ms is unrealistically short and, for example, causes the player to buffer
many periods ahead when looping.
Previously this was not feasible, because ExoPlayerTest as instrumentation test
actually needed to wait for the realtime playback duration.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=183646772
MediaPeriodInfoSequence has functionality for determining what MediaPeriod
should be loaded next. Move this into the queue as an initial step towards
moving logic concerning updating the queue of media periods out of
ExoPlayerImplInternal.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=183391114
This removes some boiler-plate code for compostite sources and will also
simplify resuing media source in the future (because this class can keep track
of child listeners).
Issue:#3498
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=183387123
- This gives LoadControl enough information in shouldContinueLoading
to know whether returning false will result in a terminal non-playback
state.
- DefaultLoadControl will always return true when returning false will
result in a terminal non-playback state, unless the target buffer size
is exceeded. This can help to avoid getting stuck in the case that a
MediaPeriod is providing samples from an unexpected starting time.
- Make the terminal state actually terminal. Previously the player would
end up in an indefinite buffering state. We now fail with an error.
- Also remove the opportunity for LoadControl implementations to livelock
playback. No sane LoadControl should simultaneously tell the player that
it doesn't want to load anything and that it doesn't want to start
playback. So this change removes the opportunity and starts playback in
EPII instead in this case.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=183114797
Try to delay failure for as long as possible. That is, propagate
DRM session failures only after an encrypted buffer arrives or clear
sample playback without session is not allowed.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=183076348
setBytesRemaining() doesn't work when called after closeCurrentSource()
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=182999254
So far this wasn't possible because Robolectric's Looper and MessageQueue
implementations have multiple shortcomings:
1. The message loop of new HandlerThreads is an not an actual loop and
scheduled messages are executed on the thread the message is enqueued
(not the handler thread).
2. The scheduler used to replace the message queue is synchronizing all its
methods. Thus, when a test attempts to add messages to a Handler from
two different threads, it may easily run into a deadlock.
3. The scheduler doesn't correctly emulate the order of messages as they
would be in an actual MessageQueue:
a. If the message is enqueued on the handler thread, it gets executed
immediately (and not after all other messages at the same time).
b. The list of messages is always re-sorted by time, meaning that the
order of execution for messages at the same time is indeterminate.
4. Robolectric's SystemClock implementation returns the current scheduler
time of the main UI thread. So, unless this scheduler is used to add
messages in the future, the SystemClock time never advances.
This CL adds two helper classes which extend and replace Robolectric's
ShadowLooper and ShadowMessageQueue.
1. We intercept messages being enqueued or deleted in the message queue.
Thus Robolectric's faulty scheduler gets never used. Instead, we keep
a blocking priority queue of messages, sorted first by execution time
and then by FIFO order to correctly emulate the real MessageQueue.
2. We also keep a list of deleted messages to know which messages to ignore
when they come up in the looper.
3. When a new Looper is started, we override the dummy loop to an actual
eternal while loop which waits for new messages, checks if they haven't
been deleted, and runs the messages (similar to what Robolectric's
MessageQueue would have done at this point).
Because we don't actually use the main UI thread in our tests, we can't rely
on the SystemClock to progress in any sensible manner. To overcome this issue,
we can use the auto-advancing FakeClock also used for the simulation tests.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=182912510
This achieves two things:
1. All our tests use the same type of assertions.
2. The tests currently run as instrumentation test can be moved to
Robolectric without changing the assertions.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=182910542
These were caused by two issues:
1. The FakeMediaSource can be updated with a new timeline. The setNewSourceInfo
is called from a different thread than prepareSource and both access local
variables without synchronization.
2. For multi-window playback, the FakeRenderer claims that isReady and isEnded
are both set to false if it read the end of the stream. However isReady should be
true because it is able to "render" its data until the end of the stream.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=182785169
Also disable use of dummy surface for devices that require the
workaround. It's only useful in the case that we can use
setOutputSurfaceWorkaround, so if it's disabled the dummy surface
has no purpose (it actually makes things worse by consuming past
the key-frame prior to the current position, which doesn't happen
if you have no surface at all).
Issue: #3724
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=182750068
This fixes a very specific case where the data read has non-cached gaps
and a read-only CDS switches to read from upstream in a gap then the
cached data is deleted. When the CDS reaches the end of the gap, it
tries to open the next source. As there is no cached data, it tries to
continue with the already opened upstream data source but as it reached
end of the gap range, the code starts looping.
Also fixes infinite lock which occurs when in the previous case CDS isn't
readonly. It locks the content while filling the gap in the cache. At the
end of the gap, as the following data is deleted it tries to lock the
content for writing but the content is already locked by itself.
The last fix is preventing removal of CachedContent entry from
CachedContentIndex while associated key is locked.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=182595426
Some tests in ExoPlayerTest issue commands to the player from the test thread
while the player is actively playing media (playWhenReady=true). Due to the
indeterminate time taken to enqueue the commands on the playback thread, they
may arrive when the player already proceeded to another window or finished
playback.
To ensure the tests are always deterministic, this change pauses playback in
the tests where this may happen before issuing the commands.
Also, for tests where we need to wait for a new window before issuing the
next command, a new action is added which allows to play until a specified
position.
-------------
Created by MOE: https://github.com/google/moe
MOE_MIGRATED_REVID=182535096
