## Motivation
In many cases, it is desirable to spawn a set of tasks associated with
keys, with the ability to cancel them by key. As an example use case for
this sort of thing, see Tower's [`ReadyCache` type][1].
Now that PR #4530 adds a way of cancelling tasks in a
`tokio::task::JoinSet`, we can implement a map-like API based on the
same `IdleNotifiedSet` primitive.
## Solution
This PR adds an implementation of a `JoinMap` type to
`tokio_util::task`, using the `JoinSet` type from `tokio::task`, the
`AbortHandle` type added in #4530, and the new task IDs added in #4630.
Individual tasks can be aborted by key using the `JoinMap::abort`
method, and a set of tasks whose key match a given predicate can be
aborted using `JoinMap::abort_matching`.
When tasks complete, `JoinMap::join_one` returns their associated key
alongside the output from the spawned future, or the key and the
`JoinError` if the task did not complete successfully.
Overall, I think the way this works is pretty straightforward; much of
this PR is just API boilerplate to implement the union of applicable
APIs from `JoinSet` and `HashMap`. Unlike previous iterations on the
`JoinMap` API (e.g. #4538), this version is implemented entirely in
`tokio_util`, using only public APIs from the `tokio` crate. Currently,
the required `tokio` APIs are unstable, but implementing `JoinMap` in
`tokio-util` means we will never have to make stability commitments for
the `JoinMap` API itself.
[1]: https://github.com/tower-rs/tower/blob/master/tower/src/ready_cache/cache.rs
Signed-off-by: Eliza Weisman <[email protected]>
## Motivation
PR #4538 adds a prototype implementation of a `JoinMap` API in
`tokio::task`. In [this comment][1] on that PR, @carllerche pointed out
that a much simpler `JoinMap` type could be implemented outside of
`tokio` (either in `tokio-util` or in user code) if we just modified
`JoinSet` to return a task ID type when spawning new tasks, and when
tasks complete. This seems like a better approach for the following
reasons:
* A `JoinMap`-like type need not become a permanent part of `tokio`'s
stable API
* Task IDs seem like something that could be generally useful outside of
a `JoinMap` implementation
## Solution
This branch adds a `tokio::task::Id` type that uniquely identifies a
task relative to all other spawned tasks. Task IDs are assigned
sequentially based on an atomic `usize` counter of spawned tasks.
In addition, I modified `JoinSet` to add a `join_with_id` method that
behaves identically to `join_one` but also returns an ID. This can be
used to implement a `JoinMap` type.
Note that because `join_with_id` must return a task ID regardless of
whether the task completes successfully or returns a `JoinError`, I've
also changed `JoinError` to carry the ID of the task that errored, and
added a `JoinError::id` method for accessing it. Alternatively, we could
have done one of the following:
* have `join_with_id` return `Option<(Id, Result<T, JoinError>)>`, which
would be inconsistent with the return type of `join_one` (which we've
[already bikeshedded over once][2]...)
* have `join_with_id` return `Result<Option<(Id, T)>, (Id, JoinError)>>`,
which just feels gross.
I thought adding the task ID to `JoinError` was the nicest option, and
is potentially useful for other stuff as well, so it's probably a good API to
have anyway.
[1]: https://github.com/tokio-rs/tokio/pull/4538#issuecomment-1065614755
[2]: https://github.com/tokio-rs/tokio/pull/4335#discussion_r773377901Closes#4538
Signed-off-by: Eliza Weisman <[email protected]>
Previously, `runtime::Handle` was a single struct composed of the
internal handles for each runtime component. This patch splits the
`Handle` struct into a `HandleInner` which contains everything
**except** the task scheduler handle. Now, `HandleInner` is passed to
the task scheduler during creation and the task scheduler is responsible
for storing it. `Handle` only needs to hold the scheduler handle and
can access the rest of the component handles by querying the task
scheduler.
The motivation for this change is it now enables the multi-threaded
scheduler to have direct access to the blocking spawner handle.
Previously, when spawning a new thread, the multi-threaded scheduler had
to access the blocking spawner by accessing a thread-local variable.
Now, in theory, the multi-threaded scheduler can use `HandleInner`
directly. However, this change hasn't been done in this PR yet.
Also, now the `Handle` struct is much smaller.
This change is intended to make it easier for the multi-threaded
scheduler to shutdown idle threads and respawn them on demand.
Handle::current docs say it's not possible to call it on any non-runtime thread, but you can call it from a runtime context created by an EnterGuard. This updates the docs to mention EnterGuard as a way to avoid this panic.
It turns out that the CI job for testing `tokio_unstable` features isn't
actually running doctests for `tokio_unstable`, just lib and integration
tests. This is because RustDoc is responsible for running doctests, and
it needs the unstable cfg passed to it separately from `RUSTFLAGS`.
This means that if the examples for unstable APIs are broken, CI won't
catch this, which is not great!
This commit changes the `test-unstable` CI job to pass `--cfg
tokio_unstable` in `RUSTDOCFLAGS` as well as `RUSTFLAGS`. This way,
doctests for unstable APIs should actually run.
I also fixed a typo in one of the runtime metrics doctests that was
causing a compilation error, which was caught as a result of actually
testing the unstable API docs on CI. :)
Signed-off-by: Eliza Weisman <[email protected]>
## Motivation
Recent Clippy releases have added some new lints that trigger on some
code in Tokio. These aren't a big deal, but seeing them in my editor is
mildly annoying.
## Solution
This branch fixes the following issues flagged by Clippy:
* manual `Option::map` implementation
* use of `.map(...).flatten(...)` that could be replaced with
`.and_then(...)`
* manual implementation of saturating arithmetic on `Duration`s
* simplify some boolean expressions in assertions (`!res.is_ok()` can be
`res.is_err()`)
* fix redundant field names in initializers
* replace an unnecessary cast to `usize` with an explicitly typed
integer literal
Signed-off-by: Eliza Weisman <[email protected]>
## Motivation
Before we stabilize the `JoinSet` API, we intend to add a method for
individual tasks in the `JoinSet` to be aborted. Because the
`JoinHandle`s for the tasks spawned on a `JoinSet` are owned by the
`JoinSet`, the user can no longer use them to abort tasks on the
`JoinSet`. Therefore, we need another way to cause a remote abort of a
task on a `JoinSet` without holding its `JoinHandle`.
## Solution
This branch adds a new `AbortHandle` type in `tokio::task`, which
represents the owned permission to remotely cancel a task, but _not_ to
await its output. The `AbortHandle` type holds an additional reference
to the task cell.
A crate-private method is added to `JoinHandle` that returns an
`AbortHandle` for the same task, incrementing its ref count.
`AbortHandle` provides a single method, `AbortHandle::abort(self)`, that
remotely cancels the task. Dropping an `AbortHandle` decrements the
task's ref count but does not cancel it. The `AbortHandle` type is
currently marked as unstable.
The spawning methods on `JoinSet` are modified to return an
`AbortHandle` that can be used to cancel the spawned task.
## Future Work
- Currently, the `AbortHandle` type is _only_ available in the public
API through a `JoinSet`. We could also make the
`JoinHandle::abort_handle` method public, to allow users to use the
`AbortHandle` type in other contexts. I didn't do that in this PR,
because I wanted to make the API addition as minimal as possible, but we
could make this method public later.
- Currently, `AbortHandle` is not `Clone`. We could easily make it
`Clone` by incrementing the task's ref count. Since this adds more trait
impls to the API, we may want to be cautious about this, but I see no
obvious reason we would need to remove a `Clone` implementation if one
was added...
- There's been some discussion of adding a `JoinMap` type that allows
aborting tasks by key, and manages a hash map of keys to `AbortHandle`s,
and removes the tasks from the map when they complete. This would make
aborting by key much easier, since the user wouldn't have to worry about
keeping the state of the map of abort handles and the tasks actually
active on the `JoinSet` in sync. After thinking about it a bit, I
thought this is probably best as a `tokio-util` API --- it can currently
be implemented in `tokio-util` with the APIs added in `tokio` in this
PR.
- I noticed while working on this that `JoinSet::join_one` and
`JoinSet::poll_join_one` return a cancelled `JoinError` when a task is
cancelled. I'm not sure if I love this behavior --- it seems like it
would be nicer to just skip cancelled tasks and continue polling. But,
there are currently tests that expect a cancelled `JoinError` to be
returned for each cancelled task, so I didn't want to change it in
_this_ PR. I think this is worth revisiting before stabilizing the API,
though?
Signed-off-by: Eliza Weisman <[email protected]>
## Motivation
Rust error messages seem to have changed a bit in 1.59.0
## Solution
Update the `trybuild` stderr output.
This should unbreak the tests on the latest Rust (and fix CI).
## Motivation
The `JoinSet` type is currently missing the `tokio_unstable` and
`feature = "rt"` `doc(cfg(...))` attributes, making it erroneously
appear to be available without the required feature and without unstable
features enabled. This is incorrect.
I believe this is because `doc(cfg(...))` on a re-export doesn't
actually add the required cfgs to the type itself, and the
`cfg_unstable!` is currently only guarding a re-export and module.
## Solution
This PR fixes the missing attributes.
Alice Ryhl