This patch reduces the number of times worker threads wake up without having
work to do in the multi-threaded scheduler. Unnecessary wake-ups are expensive
and slow down the scheduler. I have observed this change reduce no-op wakes
by up to 50%.
The multi-threaded scheduler is work-stealing. When a worker has tasks to process,
and other workers are idle (parked), these idle workers must be unparked so that
they can steal work from the busy worker. However, unparking threads is expensive,
so there is an optimization that avoids unparking a worker if there already exists
workers in a "searching" state (the worker is unparked and looking for work). This
works pretty well, but transitioning from 1 "searching" worker to 0 searching workers
introduces a race condition where a thread unpark can be lost:
* thread 1: last searching worker about to exit searching state
* thread 2: needs to unpark a thread, but skip because there is a searching worker.
* thread 1: exits searching state w/o seeing thread 2's work.
Because this should be a rare condition, Tokio solves this by always unparking a
new worker when the current worker:
* is the last searching worker
* is transitioning out of searching
* has work to process.
When the newly unparked worker wakes, if the race condition described above
happened, "thread 2"'s work will be found. Otherwise, it will just go back to sleep.
Now we come to the issue at hand. A bug incorrectly set a worker to "searching"
when the I/O driver unparked the thread. In a situation where the scheduler was
only partially under load and is able to operate with 1 active worker, the I/O driver
would unpark the thread when new I/O events are received, incorrectly transition
it to "searching", find new work generated by inbound I/O events, incorrectly
transition itself from the last searcher -> no searchers, and unpark a new thread.
This new thread would wake, find no work and go back to sleep.
Note that, when the scheduler is fully saturated, this change will make no impact
as most workers are always unparked and the optimization to avoid unparking
threads described at the top apply.
Re-applies #4377 and fixes the bug resulting in Hyper's double panic.
Revert: #4394
Original PR:
This PR does some refactoring to the current-thread scheduler bringing it closer to the structure of the
multi-threaded scheduler. More specifically, the core scheduler data is stored in a Core struct and that
struct is passed around as a "token" indicating permission to do work. The Core structure is also stored
in the thread-local context.
This refactor is intended to support #4373, making it easier to track counters in more locations in the
current-thread scheduler.
I tried to keep commits small, but the "set Core in thread-local context" is both the biggest commit and
the key one.
This patch does some refactoring to the current-thread scheduler bringing it closer to the
structure of the multi-threaded scheduler. More specifically, the core scheduler data is stored
in a Core struct and that struct is passed around as a "token" indicating permission to do
work. The Core structure is also stored in the thread-local context.
This refactor is intended to support #4373, making it easier to track counters in more locations
in the current-thread scheduler.
Tokio 1.7.0 introduced a change intended to eagerly shutdown newly
spawned tasks if the runtime is in the process of shutting down.
However, it introduced a bug where already spawned tasks could be
shutdown too early, resulting in the potential introduction of deadlocks
if tasks acquired mutexes in drop handlers.
Fixes#3869
Instead of using sleep in time::advance, this fixes the root of the issue. When futures passed
to `Runtime::block_on` are woken, it bypassed all the machinery around advancing time. By
intercepting wakes in the time driver, we know when the block_on task is woken and skip
advancing time in that case.
Fixes#3837
Previously, `time::advance` would set the mocked clock forward the
requested amount, then yield. However, if there was no work ready to
perform immediately, this would result in advancing to the next expiring
sleep.
Now, `time::advance(...)` will unblock at the requested time. The
difference between `time::advance(...)` and `time::sleep(...)` is a bit
fuzzy. The main difference is `time::sleep(...)` operates on the current
task and `time::advance(...)` operates at the runtime level.
Fixes#3710
In some cases, a cycle is created between I/O driver wakers and the I/O
driver resource slab. This patch clears stored wakers when an I/O
resource is dropped, breaking the cycle.
Fixes#3228
Removes the `stream` feature flag from `Cargo.toml` and removes the
`futures-core` dependency. Once `Stream` lands in `std`, a feature flag
is most likely not needed.
Pausing time is a capability added to assist with testing Tokio code
dependent on time. Currently, the capability implicitly requires the
current_thread runtime.
This change enforces the requirement by panicking if called from a
multi-threaded runtime.
Enabling `tracing` integration now requires compiling with `--cfg
tokio_unstable`. Once `tracing-core` guarantees the same level of
stability as Tokio 1.0, unstable can be removed.
Closes#3258
Instead of using OS specific extension traits, OS specific methods are
moved onto the structs themselves and guarded with `cfg`. The API
documentation should highlight the function is platform specific.
Closes#2925
The mpsc `try_recv()` functions have an issue where a sent message
happens-before a call to `try_recv()` but `try_recv()` returns `None`.
Fixing this is non-trivial, so the function is removed for 1.0. When the
bug is fixed, the function can be added back.
Closes#2020
Adds `ready()`, `readable()`, and `writable()` async methods for waiting
for socket readiness. Adds `try_send`, `try_send_to`, `try_recv`, and
`try_recv_from` for performing non-blocking operations on the socket.
This is the UDP equivalent of #3130.
Adds function to await for readiness on the TcpStream and non-blocking read/write functions.
`async fn TcpStream::ready(Interest)` waits for socket readiness satisfying **any** of the specified
interest. There are also two shorthand functions, `readable()` and `writable()`.
Once the stream is in a ready state, the caller may perform non-blocking operations on it using
`try_read()` and `try_write()`. These function return `WouldBlock` if the stream is not, in fact, ready.
The await readiness function are similar to `AsyncFd`, but do not require a guard. The guard in
`AsyncFd` protect against a potential race between receiving the readiness notification and clearing
it. The guard is needed as Tokio does not control the operations. With `TcpStream`, the `try_read()`
and `try_write()` function handle clearing stream readiness as needed.
This also exposes `Interest` and `Ready`, both defined in Tokio as wrappers for Mio types. These
types will also be useful for fixing #3072 .
Other I/O types, such as `TcpListener`, `UdpSocket`, `Unix*` should get similar functions, but this
is left for later PRs.
Refs: #3130
* Removes duplicated code by moving it to `Registration`.
* impl `Deref` for `PollEvented` to avoid `get_ref()`.
* Avoid extra waker clones in I/O driver.
* Add `Interest` wrapper around `mio::Interest`.
This reverts commit fe2b997.
We are avoiding adding poll_read_buf to tokio itself for now. The patch is
reverted now in order to not block the v0.3.2 release (#3059).
The `Receiver` handle maintains a position in the broadcast channel for
itself. Cloning implies copying the state of the value. Intuitively,
cloning a `broadcast::Receiver` would return a new receiver with an
identical position. However, the current implementation returns a new
`Receiver` positioned at the tail of the channel.
This behavior subtlety is why `new_subscriber()` is used to create
`Receiver` handles. An alternate API should consider the position issue.
Refs: #2933
This combines the `dns` and `net` feature flags. Previously, `dns` was
included as part of `net`. Given that is is rare that one would want
`dns` without `net`, DNS is now entirely gated w/ `net`.
The `parking_lot` feature is included as part of `full`.
Some misc docs are tweaked to reflect feature flag changes.
Changes inherent methods to take `&self` instead of `&mut self`. This
brings the API in line with `std`.
This patch is implemented by using a `tokio::sync::Mutex` to guard the
internal `File` state. This is not an ideal implementation strategy
doesn't make a big impact compared to having to dispatch operations to a
background thread followed by a blocking syscall.
In the future, the implementation can be improved as we explore async
file-system APIs provided by the operating-system (iocp / io_uring).
Closes#2927
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 enables the caller to configure the socket and to explicitly bind
the socket before converting it to a `TcpStream` or `TcpListener`.
Closes: #2902
Updates the mpsc channel to use the intrusive waker based sempahore.
This enables using `Sender` with `&self`.
Instead of using `Sender::poll_ready` to ensure capacity and updating
the `Sender` state, `async fn Sender::reserve()` is added. This function
returns a `Permit` value representing the reserved capacity.
Fixes: #2637
Refs: #2718 (intrusive waiters)
These functions have object safety issues. It also has been decided to
avoid vectored operations on the I/O traits. A later PR will bring back
vectored operations on specific types that support them.
Refs: #2879, #2716
When the mpsc channel receiver closes the channel, receiving should
return `None` once all in-progress sends have completed. When a sender
reserves capacity, this prevents the receiver from fully shutting down.
Previously, when the sender, after reserving capacity, dropped without
sending a message, the receiver was not notified. This results in
blocking the shutdown process until all sender handles drop.
This patch adds a receiver notification when the channel is both closed
and all outstanding sends have completed.
When the mpsc channel receiver closes the channel, receiving should
return `None` once all in-progress sends have completed. When a sender
reserves capacity, this prevents the receiver from fully shutting down.
Previously, when the sender, after reserving capacity, dropped without
sending a message, the receiver was not notified. This results in
blocking the shutdown process until all sender handles drop.
This patch adds a receiver notification when the channel is both closed
and all outstanding sends have completed.
Decouples getting the latest `watch` value from receiving the change
notification. The `Receiver` async method becomes
`Receiver::changed()`. The latest value is obtained from
`Receiver::borrow()`.
The implementation is updated to use `Notify`. This requires adding
`Notify::notify_waiters`. This method is generally useful but is kept
private for now.
* sync: move CancellationToken to tokio-util
The `CancellationToken` utility is only available with the
`tokio_unstable` flag. This was done as the API is not final, but it
adds friction for users.
This patch moves `CancellationToken` to tokio-util where it is generally
available. The tokio-util crate does not have any constraints on
breaking change releases.
* fix clippy
* clippy again
The I/O driver uses a slab to store per-resource state. Doing this
provides two benefits. First, allocating state is streamlined. Second,
resources may be safely indexed using a `usize` type. The `usize` is
used passed to the OS's selector when registering for receiving events.
The original slab implementation used a `Vec` backed by `RwLock`. This
primarily caused contention when reading state. This implementation also
only **grew** the slab capacity but never shrank. In #1625, the slab was
rewritten to use a lock-free strategy. The lock contention was removed
but this implementation was still grow-only.
This change adds the ability to release memory. Similar to the previous
implementation, it structures the slab to use a vector of pages. This
enables growing the slab without having to move any previous entries. It
also adds the ability to release pages. This is done by introducing a
lock when allocating/releasing slab entries. This does not impact
benchmarks, primarily due to the existing implementation not being
"done" and also having a lock around allocating and releasing.
A `Slab::compact()` function is added. Pages are iterated. When a page
is found with no slots in use, the page is freed. The `compact()`
function is called occasionally by the I/O driver.
Fixes#2505
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.
Previously, in the broadcast channel, receiver wakers were passed to the
sender via an atomic stack with allocated nodes. When a message was
sent, the stack was drained. This caused a problem when many receivers
pushed a waiter node then dropped. The waiter node remained indefinitely
in cases where no values were sent.
This patch switches broadcast to use the intrusive linked-list waiter
strategy used by `Notify` and `Semaphore.
In some cases, when a call to `block_in_place` completes, the runtime is
reinstated on the thread. In this case, the task budget must also be set
in order to avoid starving other tasks on the worker.
Do not export the `scoped_thread_local` macro outside of the Tokio
crate. This is not considered a breaking change as the macro never
worked if used from outside of the crate due to the generated code
referencing crate-private types.
Previously, the function picking the default number of threads for the
threaded runtime did not factor in `max_threads`. Instead, it only used
the value returned by `num_cpus`. However, if `num_cpus` returns a value
greater than `max_threads`, then the function would panic.
This patch fixes the function by limiting the default number of threads
by `max_threads`.
Fixes#2452
Fixes a couple bugs in the work-stealing queue introduced as
part of #2315. First, the cursor needs to be able to represent more
values than the size of the buffer. This is to be able to track if
`tail` is ahead of `head` or if they are identical. This bug resulted in
the "overflow" path being taken before the buffer was full.
The second bug can happen when a queue is being stolen from concurrently
with stealing into. In this case, it is possible for buffer slots to be
overwritten before they are released by the stealer. This is harder to
happen in practice due to the first bug preventing the queue from
filling up 100%, but could still happen. It triggered an assertion in
`steal_into`. This bug slipped through due to a bug in loom not
correctly catching the case. The loom bug is fixed as part of
tokio-rs/loom#119.
Fixes: #2382
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
