The scheduler is awaken when aborting a task loop, but not after
a triggering event is pushed. This can cause throttling to induce
long state transitions for pipelines with many streams.
Observed for Unprepare with:
GST_DEBUG=ts-benchmark:4 ../../target/debug/examples/benchmark 2000 ts-udpsrc 2 20 5000
During MR !793, the socket configuration mechanism was changed to
use commands passed to the Task via a channel. This worked properly
for user changes via settings and signals, however the default
clients setting was not used.
A simple solution could have been to send a command at initialization
to add the default clients, but it was considered a better solution
to just wait for the Task preparation to configure the sockets based
on the value of settings.clients at that time, thus avoiding
unnecessary successive removals and additions of clients which could
have happened before preparation.
Of course, users can still add or remove clients as before, before
and after Task preparation.
See also https://gitlab.freedesktop.org/gstreamer/gst-plugins-rs/-/merge_requests/793
The way the runtime::Task is implemented, UdpSinkTask is available
as a mutable ref in all the TaskImpl functions, which offers the
opportunity to avoid using Mutexes.
Main higlights:
- Removed the back and forth calls between UdpSinkPadHandler
and UdpSinkTask.
- Udp sockets are now part of UdpSinkTask, which is also in
charge of preparing them instead of leaving this to UdpSink.
This removed the need for Context::enter since
TaskImpl::prepare already operates under the target Context.
- In order for the clients list to be visible from the UdpSink,
this list was maintained by UdpSinkPadHandler which was also
in charge of (un)configuring the Udp sockets. The sockets are
now part of UdpSinkTask, which is also in charge of the
(un)configuration. Add/remove/replace requests are passed as
commands to the UdpSinkTask via a channel.
- The clients list visible to the UdpSink is now part of the
Settings (it is also a read/write property). Since the actual
socket (un)configuration is asynchronously handled by the Task,
the clients list is updated by the add/remove/replace signals
and set_property("clients", ..). Should a problem occur during
the async (un)configuration, and only in this case, the
UdpSinkTask would update the clients lists in Settings
accordingly so that it stays consistent with the internal state.
- The function clear_clients was renamed as replace_with_clients.
- clients is now based on a BTreeSet instead of a Vec. All the
managing functions perform some sort of lookup prior to updating
the collection. It also ease implementation.
- Removed the UdpSinkPadHandler RwLock. Using flume channels, we
are able to clone the Receiver so it can be stored in UdpSink
and reused when preparing the UdpSinkTask.
The I/O handle was dropped prior to removing it from the reactor,
which caused `Poller::delete` to fail due to an invalid file
descriptor. This used to happen silently unless the same fd was
added again, e.g. by changing states in the pipeline as follow:
Null -> Playing -> Null -> Playing.
In which case `Poller::add` failed due to an already existing file.
This commit makes sure the fd is removed from the reactor prior to
dropping the handle. In order to achieve this, a new task is spawned
on the `Context` on which the I/O was originally registered, allowing
it to access the proper `Reactor`. The I/O can then safely be dropped.
Because the I/O handle is moved to the spawned future, this solution
requires adding the `Send + 'static` bounds to the I/O handle used
within the `Async` wrapper. This appears not too restrictive for
existing implementations though. Other attempts were considered,
but they would cause deadlocks.
This new approach also solves a potential race condition where a
fd could be re-registered in a `Reactor` before it was removed.
The internal (C) jitterbuffer needs to know about the configured
latency when calculating a PTS, as it otherwise may consider that
the packet is too late, trigger a resync and cause the element to
discard the packet altogether.
I could not identify when this was broken, but the net effect was
that in the current state, ts-jitterbuffer was discarding up to
half of all the incoming packets.
Previous version used the Context::block_on_or_add_sub_task which
spawns a full-fledged executor with timer and io Reactor for no
reason when we just need to wait for a Receiver or JoinHandle.
When the iteration loop is throttling, the call to `abort` on the
`loop_abort_handle` returns immediately, but the actual `Future`
for the iteration loop is aborted only when the scheduler throttling
completes. State transitions which requires the loop to be aborted &
which are serialized at the pipeline level can incur long delays.
This commit makes sure the Task Context's scheduler is awaken as soon
as the task loop is aborted.
Previous version relied on a plain loop / match / break because
I experimented different strategies. The while variant is better
for the final solution.
The function `enter` is executed in a blocking way from the caller's
point of view. This means that we can guaranty that the provided
function and its output will outlive the underlying Scheduler Task
execution. This requires an unsafe call to
`async_task::spawn_unchecked`. See:
https://docs.rs/async-task/latest/async_task/fn.spawn_unchecked.html
The threadshare executor was based on a modified version of tokio
which implemented the throttling strategy in the BasicScheduler.
Upstream tokio codebase has significantly diverged from what it
was when the throttling strategy was implemented making it hard
to follow. This means that we can hardly get updates from the
upstream project and when we cherry pick fixes, we can't reflect
the state of the project on our fork's version. As a consequence,
tools such as cargo-deny can't check for RUSTSEC fixes in our fork.
The smol ecosystem makes it quite easy to implement and maintain
a custom async executor. This MR imports the smol parts that
need modifications to comply with the threadshare model and implements
a throttling executor in place of the tokio fork.
Networking tokio specific types are replaced with Async wrappers
in the spirit of [smol-rs/async-io]. Note however that the Async
wrappers needed modifications in order to use the per thread
Reactor model. This means that higher level upstream networking
crates such as [async-net] can not be used with our Async
implementation.
Based on the example benchmark with ts-udpsrc, performances seem on par
with what we achieved using the tokio fork.
Fixes https://gitlab.freedesktop.org/gstreamer/gst-plugins-rs/-/issues/118
Related to https://gitlab.freedesktop.org/gstreamer/gst-plugins-rs/-/merge_requests/604
The time driver for the threadshare runtime assigns the timer
entries to the nearest throttling time frame so that the timer
fires as close as possible to the expected instant. This means
that the timer might fire before or after the expected instant
(at most `wait / 2` away).
In some cases, we don't want the timer to fire early. The new
function `delay_for_at_least` ensures that the timer is assigned
to the time frame after the expected instant.
cargo-c will produce a pkg-config file making it easier to statically
link plugins.
Also add 'static' features for plugins depending on < 1.14 as this is the
minimal required version to use static linking because of ABI changes in
core.
There is no way to dynamically ask Cargo to build static or dynamic lib
so we have to build both and pick the one we care when doing the meson
processing.
Fix#88
error[E0412]: cannot find type `Error` in this scope
--> generic\threadshare\src\socket.rs:237:56
|
237 | pub fn set_tos(&self, qos_dscp: i32) -> Result<(), Error> {
| ^^^^^ not found in this scope
Otherwise we can deadlock because of a lock order issue:
- render() is called with the sink pad handler lock and takes the
settings lock
- clearing clients takes the sink pad handler lock
Only two uses of unsafely setting the pad functions is left:
- fallbacksrc for overriding the chain function of the proxy pad of a
ghost pad
- threadshare for overriding the pad functions after creationg, which
probably needs some fixing at some point
Since those are using the clock for sync, they need to also
provide a clock for good measure. The reason is that even if
downstream elements provide a clock, we don't want to have
that clock selected because it might not be running yet.
Pad{Src,Sink}[Ref] delegate some functions to their respective
Pad{Src,Sink}Inner. Since they act as smart pointers, we can
safely implement the Deref trait to simplify the implementations.
StateMachines are spawned on a runtime::Context which uses a tokio
runtime. The StateMachine doesn't need all the features from tokio
such as the IO and timers drivers.
This commit makes use of a light-weight futures executor to spawn
the StateMachines.
When initializing Pad functions in `Pad{Src,Sink}`, we downgrade the
`Pad{Src,Sink}` and upgrade it when necessary. This was implemented
to avoid reference cycles:
`gst::Pad` -> pad function -> `Pad{Src,Sink}` -> `gst::Pad`.
Since `Pad{Src,Sink}` reset the pad functions when dropping, there is
no cycles, so we can use an `Arc<Pad{Src,Sink}>` in the pad functions,
thus saving an `upgrade`.
This is similar to what the standard jitterbuffer does, and
is necessary to avoid deadlocks with the rtpbin session lock,
if the user calls onto any API that requires us to take the
state lock at the wrong time (eg setting the latency property,
clearing the pt map)
They're actually used as a hackish intrusive list around GQueue so using
a different allocator is rather dangerous. Or rather, this is generally
dangerous and shouldn't be done but ...
Also make sure to free items manually in finalize() so that any
contained buffers can also be unreffed.
We don't want to run it every time a strong reference is dropped but
only at the very end. Otherwise dropping the socket stream will cause a
panic because the socket itself is still running.
This can only happen if the receiver is dropped, which only happens when
the task is stopped. As such, Flushing should be returned instead of
panicking.
