This patch includes an initial implementation of a new multi-threaded
runtime. The new runtime aims to increase the scheduler throughput by
speeding up how it dispatches work to peer worker threads. This
implementation improves most benchmarks by about ~10% when the number of
threads is below 16. As threads increase, mutex contention deteriorates
performance.
Because the new scheduler is not yet ready to replace the old one, the
patch introduces it as an unstable runtime flavor with a warning that it
isn't production ready. Work to improve the scalability of the runtime
will most likely require more intrusive changes across Tokio, so I am
opting to merge with master to avoid larger conflicts.
Previously, the current_thread scheduler used its own injection queue
instead of sharing the same one as the multi-threaded scheduler. This
patch updates the current_thread scheduler to use the same injection
queue as the multi-threaded one (`task::Inject`).
`task::Inject` includes an optimization where it does not need to
acquire the mutex if the queue is empty.
Previously, calling `task::yield_now().await` would yield the current
task to the scheduler, but the scheduler would poll it again before
polling the resource drivers. This behavior can result in starving the
resource drivers.
This patch creates a queue tracking yielded tasks. The scheduler
notifies those tasks **after** polling the resource drivers.
Refs: #5209
This patch updates CI to use `cross` to run Tokio tests on virtualized
ARM and i686 VMs. Because ipv6 doesn't work on Github action running in
a docker instance, those tests are disabled
An earlier change updated `enter_runtime` to also set the current
handle. However, the change did not store the `SetCurrentGuard`, so the
"current handle" was immediately unset. This patch stores the
`SetCurrentGuard` in the `EnterRuntimeGuard`.
No existing test exposed this bug because all tests went via `Runtime`
instead of `Handle`. Currently, `Runtime` is still explicitly setting
the handle before entering runtime, so all tests still passed. A new
test is added that covers the case of calling `Handle::block_on` and
accessing the current handle.
For historical reasons, the current-thread runtime's scheduler was named
BasicScheduler and the multi-thread runtime's scheduler were named
ThreadPool. This patch renames the schedulers to mirror the runtime
names to increase consistency. This patch also moves the scheduler
implementations into a new `runtime::scheduler` module.
This adds initial, unstable, support for the wasm32-wasi target. Not all of Tokio's
features are supported yet as WASI's non-blocking APIs are still limited.
Refs: tokio-rs/tokio#4827
* task: add consume_budget for cooperative scheduling
For cpu-only computations that do not use any Tokio resources,
budgeting does not really kick in in order to yield and prevent
other tasks from starvation. The new mechanism - consume_budget,
performs a budget check, consumes a unit of it, and yields only
if the task exceeded the budget. That allows cpu-intenstive
computations to define points in the program which indicate that
some significant work was performed. It will yield only if the budget
is gone, which is a much better alternative to unconditional yielding,
which is a potentially heavy operation.
* tests: add a test case for task::consume_budget
The test case ensures that the task::consume_budget utility
actually burns budget and makes the task yield once the whole
budget is gone.
Uses the infrastructure added by #2828 to enable switching
`TcpListener::accept` to use `&self`.
This also switches `poll_accept` to use `&self`. While doing introduces
a hazard, `poll_*` style functions are considered low-level. Most users
will use the `async fn` variants which are more misuse-resistant.
TcpListener::incoming() is temporarily removed as it has the same
problem as `TcpSocket::by_ref()` and will be implemented later.
This refactors I/O registration in a few ways:
- Cleans up the cached readiness in `PollEvented`. This cache used to
be helpful when readiness was a linked list of `*mut Node`s in
`Registration`. Previous refactors have turned `Registration` into just
an `AtomicUsize` holding the current readiness, so the cache is just
extra work and complexity. Gone.
- Polling the `Registration` for readiness now gives a `ReadyEvent`,
which includes the driver tick. This event must be passed back into
`clear_readiness`, so that the readiness is only cleared from `Registration`
if the tick hasn't changed. Previously, it was possible to clear the
readiness even though another thread had *just* polled the driver and
found the socket ready again.
- Registration now also contains an `async fn readiness`, which stores
wakers in an instrusive linked list. This allows an unbounded number
of tasks to register for readiness (previously, only 1 per direction (read
and write)). By using the intrusive linked list, there is no concern of
leaking the storage of the wakers, since they are stored inside the `async fn`
and released when the future is dropped.
- Registration retains a `poll_readiness(Direction)` method, to support
`AsyncRead` and `AsyncWrite`. They aren't able to use `async fn`s, and
so there are 2 reserved slots for those methods.
- IO types where it makes sense to have multiple tasks waiting on them
now take advantage of this new `async fn readiness`, such as `UdpSocket`
and `UnixDatagram`.
Additionally, this makes the `io-driver` "feature" internal-only (no longer
documented, not part of public API), and adds a second internal-only
feature, `io-readiness`, to group together linked list part of registration
that is only used by some of the IO types.
After a bit of discussion, changing stream-based transports (like
`TcpStream`) to have `async fn read(&self)` is punted, since that
is likely too easy of a footgun to activate.
Refs: #2779, #2728
JoinHandle of threads created by the pool are now tracked and properly joined at
shutdown. If the thread does not return within the timeout, then it's not joined and
left to the OS for cleanup.
Also, break a cycle between wakers held by the timer and the runtime.
Fixes#2641, #2535
Previously, dropping the Write handle would issue a `shutdown(Both)`. However,
shutting down the read half is not portable and not the correct action to take.
This changes the behavior of OwnedWriteHalf to only perform a `shutdown(Write)`
on drop.
The new queue uses `u8` to track offsets. Cursors are expected to wrap.
An operation was performed with `+` instead of `wrapping_add`. This was
not _obviously_ issue before as it is difficult to wrap a `usize` on
64bit platforms, but wrapping a `u8` is trivial.
The fix is to use `wrapping_add` instead of `+`. A new test is added
that catches the issue.
Fixes#2361
The work-stealing scheduler includes an optimization where each worker
includes a single slot to store the **last** scheduled task. Tasks in
scheduler's LIFO slot are executed next. This speeds up and reduces
latency with message passing patterns.
Previously, this optimization was susceptible to starving other tasks in
certain cases. If two tasks ping-ping between each other without ever
yielding, the worker would never execute other tasks.
An early PR (#2160) introduced a form of pre-emption. Each task is
allocated a per-poll operation budget. Tokio resources will return ready
until the budget is depleted, at which point, Tokio resources will
always return `Pending`.
This patch leverages the operation budget to limit the LIFO scheduler
optimization. When executing tasks from the LIFO slot, the budget is
**not** reset. Once the budget goes to zero, the task in the LIFO slot
is pushed to the back of the queue.
A refactor of the scheduler internals focusing on simplifying and
reducing unsafety. There are no fundamental logic changes.
* The state transitions of the core task component are refined and
reduced.
* `basic_scheduler` has most unsafety removed.
* `local_set` has most unsafety removed.
* `threaded_scheduler` limits most unsafety to its queue implementation.
Previously, if an IO event was received during the runtime shutdown
process, it was possible to enter a deadlock. This was due to the
scheduler shutdown logic not expecting tasks to get scheduled once the
worker was in the shutdown process.
This patch fixes the deadlock by checking the queues for new tasks after
each call to park. If a new task is received, it is forcefully shutdown.
Fixes#2061
Storing a `Runtime` value in a thread-local resulted in a panic due to
the inability to access the parker.
This fixes the bug by skipping parking if it fails. In general, there
isn't much that we can do besides not parking.
Fixes#593
Nested spawn_blocking calls would result in a panic due to the necessary
context not being setup. This patch sets the blocking pool context from
within a blocking pool.
Fixes#1982
The "global executor" thread-local is to track where to spawn new tasks,
**not** which scheduler is active on the current thread. This fixes a
bug with scheduling tasks on the basic_scheduler by tracking the
currently active basic_scheduler with a dedicated thread-local variable.
Fixes: #1851
## Motivation
In earlier versions of `tokio`, the `current_thread::Runtime` type could
be used to run `!Send` futures. However, PR #1716 merged the
current-thread and threadpool runtimes into a single type, which can no
longer run `!Send` futures. There is still a need in some cases to
support futures that don't implement `Send`, and the `tokio-compat`
crate requires this in order to provide APIs that existed in `tokio`
0.1.
## Solution
This branch implements the API described by @carllerche in
https://github.com/tokio-rs/tokio/pull/1716#issuecomment-549496309. It
adds a new `LocalSet` type and `spawn_local` function to `tokio::task`.
The `LocalSet` type is used to group together a set of tasks which must
run on the same thread and don't implement `Send`. These are available
when a new "rt-util" feature flag is enabled.
Currently, the local task set is run by passing it a reference to a
`Runtime` and a future to `block_on`. In the future, we may also want
to investigate allowing spawned futures to construct their own local
task sets, which would be executed on the worker that the future is
executing on.
In order to implement the new API, I've made some internal changes to
the `task` module and `Schedule` trait to support scheduling both `Send`
and `!Send` futures.
Signed-off-by: Eliza Weisman <eliza@buoyant.io>
