This cleans up a ton of race conditions, and starts moving towards something
we can use with Clang's -Wthread-safety (but that has a ways to go still).
This can happen if you close the stream's underlying device directly, which
removes the binding but doesn't destroy the object.
In this case, the stream remains valid until destroyed, but still should not
be able to be bound to a new device.
This does a ton of work that can deadlock, because several crucial WASAPI
things that we want to do in response to this will block until the
notification callback has returned, so we can't call them from the handler
directly, or we'll be waiting until the thing that called us returns.
Almost nothing checks these return values, and there's no reason a valid
lock should fail to operate. The cases where a lock isn't valid (it's a
bogus pointer, it was previously destroyed, a thread is unlocking a lock it
doesn't own, etc) are undefined behavior and always were, and should be
treated as an application bug.
Reference Issue #8096.
ALSA expects handles to be of type ALSA_Device, and passing the handle for the default device as a plain string causes a crash as it attempts to deference the string contents itself as a pointer to a string.
Create immutable static ALSA_Device structs for the default devices and pass those as the handles. They are not placed in the hotplug list, and the audio layer doesn't attempt to free ALSA handles, so there is no need to worry about them being erroneously freed.
Both strings are _right there_ for comparing, so we can just set a flag to
note the device definitely changed.
Also simplified string management further; hotplug thread now makes a copy
of the string before releasing the lock if there was a change event, so when
the lock releases further events don't see a NULL and assume it's a new
device, causing a lot of work to ripple out and decide nothing has changed,
until the system stabilizes again. Now, it just does the right thing once.
We don't match dev->name by string, since we might use the same string for both capture and output devices. Instead use the device pointer itself as the handle.
@icculus, are we guaranteed the device pointer is valid in ALSA_OpenDevice()?
This fixes problems where Pulse callbacks don't fire in the order we expect,
or fail to fire at all, and avoids extra round trips to the Pulse server to
lookup information we could have trivially obtained already.
The end result is we would occasionally miss default device changes, etc, and
this resolves that better.
This patch reverts the previous reversion, and then adds code to queue up
events to be sent the next time SDL pumps the event queue. This guarantees
that the event watcher/filter _never_ runs from an SDL audio device thread
or some other backend-specific internal thread.
This reverts commit 76f81797b7.
This worked in the normal cases, but:
A device thread that calls SDL_DisconnectAudioDevice due to failure will fire
the disconnect event from the device thread...and if there's an event watcher
that uses that moment to close the device, we still end up in the same
situation, where the device thread tries to join on itself.
Better solutions are still pending.
Otherwise, they risk the device thread joining on itself.
Now we make sure the reference is held at the logical device level until
the physical device is closed, so it can't destroy the device in normal
usage until the thread is joined, etc.
Fully committing to it...!
This left SDL_wave.* alone for now, since there's a ton of comments in there
and this code hasn't changed much from SDL2 so far. But as SDL2 ages out a
little more, I'll likely switch this over, too.
First stage happens before we destroy objects, and is generally used to
shut down hotplug. The second stage is the usual deinit, which cleans up
the lowlevel API, unloads shared libraries, etc.
- No more tapdance to either join the audio device thread or have it detach
itself. Significant simplication of and fixes to the locking code to prevent
deadlocks.
- Physical devices now keep a refcount. Each logical device increments it,
as does the existence of a device thread, etc. Last unref destroys the
device and takes it out of the device_hash. Since there's a lot of moving
parts that might be holding a reference to a physical device, this seemed
like a safer way to protect the object.
- Disconnected devices now continue to function as zombie devices. Playback
devices will still consume data (and just throw it away), and capture devices
will continue to produce data (which always be silence). This helps apps
that don't handle disconnect events; the device still stops playing/capturing,
but bound audio streams will still consume data so they don't allocate more
data infinitely, and apps that depend on an audio callback firing regularly
to make progress won't hang.
Please note that disconnected audio devices must now be explicitly closed!
They always _should_ have been, but before this commit, SDL3 would destroy the
disconnected device for you (and manually closing afterwards was a safe no-op).
Reference Issue #8331.
Fixes#8386.
(and probably others).
