* rt: fix `Runtime::reactor()` as used by tokio-core
Up until Tokio v0.1.11, the handle returned by `Runtime::reactor()`
pointed to a reactor instance running in a background thread. The thread
was eagerly spawned.
As of v0.1.12, a reactor instance is created per runtime worker thread.
`Runtime::reactor()` was deprecated and updated to point to the reactor
for one of the worker threads.
A problem occurs when attempting to use the reactor before spawning a
task. Worker threads are spawned lazily, which means that the reactor
referenced by `Runtime::reactor()` is not yet running.
This patch changes `Runtime::reactor` back to a dedicated reactor
running on a background thread. However, the background thread is now
spawned lazily when the deprecated function is first called.
Fixes#720
* Fix comment
Co-Authored-By: carllerche <me@carllerche.com>
- `tokio::run` checks Enter before creating a new threadpool and
spawning the main future.
- `Runtime::block_on` now checks Enter
- `Runtime::block_on_all` now checks Enter
Fixes: #681
## Motivation
Currently, a potential panic exists in `LengthDelimitedCodec::encode`.
Writing the length field to the `dst` buffer can exceed the buffer
capacity, as `BufMut::put_uint_{le,be}` doesn't reserve more capacity.
## Solution
This branch adds a call to `dst.reserve` to ensure that there's
sufficient remaining buffer capacity to hold the length field and
the frame, prior to writing the length field. Previously, capacity
was only reserved later in the function, when writing the frame
to the buffer, and we never reserved capacity for the length field.
I've also added a test that reproduces the issue. The test panics on
master, but passes after making this change.
Signed-off-by: Eliza Weisman <eliza@buoyant.io>
This patch refactors `length_delimited` to be implemented as a `Codec` and
use the default `Framed` wrapper types.
The original implementation did not do this in order to support vectored writes in the
write half. However, this implementation would be more efficient with small frames anyway.
If vectored writes are to be explored in the future, then it should be done holistically.
Signed-off-by: Eliza Weisman <eliza@buoyant.io>
The futures 0.2 crate is not intended for widespread usage. Also, the
futures team is exploring the compat shim route.
If futures 0.3 support is added to Tokio 0.1, then a different
integration route will be explored, making the current code unhelpful.
Extract `tokio::executor::current_thread` to a tokio-current-thread
crate. Deprecated fns stay in the old location. The new crate only
contains thee most recent API.
Currently, the timer uses a `Now` trait to abstract the source of time.
This allows time to be mocked out. However, the current implementation
has a number of limitations as represented by #288 and #296.
The main issues are that `Now` requires `&mut self` which prevents a
value from being easily used in a concurrent environment. Also, when
wanting to write code that is abstract over the source of time, generics
get out of hand.
This patch provides an alternate solution. A new type, `Clock` is
provided which defaults to `Instant::now` as the source of time, but
allows configuring the actual source using a new iteration of the `Now`
trait. This time, `Now` is `Send + Sync + 'static`. Internally, `Clock`
stores the now value in an `Arc<Now>` value, which introduces dynamism
and allows `Clock` values to be cloned and be `Sync`.
Also, the current clock can be set for the current execution context
using the `with_default` pattern.
Because using the `Instant::now` will be the most common case by far, it
is special cased in order to avoid the need to allocate an `Arc` and use
dynamic dispatch.
This patch introduces a version of `Runtime` that runs all components on
the current thread. This allows users to spawn futures that do not implement
`Send`.
This ensures that all fd-based futures are put into the queue for the
current tick, if the CurrentThread is parking via the Reactor.
Otherwise, if there are queued up futures already, only those would be
polled in the turn. These futures could then notify others/themselves to
have the queue still non-empty on the next turn. Which then potentially
allows the reactor to never be polled, and thus fd-based futures are
never queued up and polled.
Also return in the Turn return value whether any futures were polled at
all, which allows the caller to know if any work was done at all in this
turn and based on that adjust behavior.
This patch renames `Sleep` from tokio-timer and the tokio facade to
`Delay`. Given that the future does not actually put anything to sleep,
the `Delay` name feels more appropriate.
Fixes#263
This patch integrate the new timer implementation with the runtime by
initializing a timer per worker thread. This allows minimizing the
amount of synchronization needed for using timers.
* Fix races.
This mostly pulls in changes from rust-lang-nursery/futures-rs#881, but
also updates Registration to be a bit more obvious as to what is going
on.
* Reduce spurious wakeups caused by Reactor
This patch adds an ABA guard on token values before registering them
with Mio. This allows catching token reuse and avoid the notification.
This is needed for OS X as the notification is used to determine that a
TCP connect has completed. A spurious notification can potentially cause
write failures.
This patch updates `poll_read_ready` to take a `mask` argument, enabling
the caller to specify the desired readiness. `need_read` is renamed to
`clear_read_ready` and also takes a mask.
This enables a caller to listen for HUP events without requiring reading
from the I/O resource.
Tokio is moving away from using `WouldBlock`, instead favoring
`Async::NotReady`.
This patch updates the TCP and UDP types, deprecating any function that
returns `WouldBlock` and adding a poll_ prefixed equivalent.
Currently, `tokio::spawn` matched the `spawn` function from futures 0.2.
However, this adds additional ergonomic overhead and removes the ability
to spawn from a drop fn. See rust-lang-nursery/futures-rs#830.
This patch switches the behavior to access the thread-local variable
referencing the default executor directly in the `spawn` function.
This patch makes a significant change to how I/O resources bind to a
reactor. Currently, an I/O resource (TCP, UDP, PollEvented) will bind
itself with a reactor upon creation.
First, some history.
Originally, tokio-core required that I/O resources be explicitly
associated with a reactor upon creation by passing in a `&Handle`. Tokio
reform introduced a default reactor. If I/O resources do not specify a
reactor upon creation, then the default reactor is used.
However, futures tend to favor being lazy. Creating a future should do
no work, instead it is defining a computation to be performed once the
future is executed. Binding an I/O resource with a reactor on creation
goes against this pattern.
This patch fixes this by allowing I/O resources to lazily bind to a
reactor. An explicit `&Handle` can still be used on creation, but if no
reactor is specified, then the default reactor is used. However, this
binding happens during execution time (read / write) and not creation.
Currently, the thread-local tracking the current thread executor is not
set when a task is dropped. This means that one cannot spawn a new
future from within the drop implementation of another future.
This patch adds support for this by setting the thread-local before
releasing a task.
This implementation is a bit messy. It probably could be cleaned up, but
this is being put off in favor of trying a more comprehensive
reorganization once the current thread executor is feature complete.
The logic that enables `CurrentThread::turn` to avoid unbounded
iteration was incorrect. It was possible for unfortunate timing to
result in a dead lock.
This patch provides a fix as well as a test.
CurrentThread::turn uses a turn count strategy to allow `turn` to not
run infinitely. Currently, there is a bug where spawned tasks will not
get executed in calls to `turn`.
This patch fixes the bug by correctly setting the turn count for newly
spawned tasks.
This patch is an intial implementation of the Tokio runtime. The Tokio
runtime provides an out of the box configuration for running I/O heavy
asynchronous applications.
As of now, the Tokio runtime is a combination of a work-stealing thread
pool as well as a background reactor to drive I/O resources.
This patch also includes tokio-executor, a hopefully short lived crate
that is based on the futures 0.2 executor RFC.
* Implement `Park` for `Reactor`
This enables the reactor to be used as the thread parker for executors.
This also adds an `Error` component to `Park`. With this change, a
`Reactor` and a `CurrentThread` can be combined to achieve the
capabilities of tokio-core.
The `Handle` type is intended to be used by end users of Tokio. The
wakeup functionlity is needed by executor implementations. It doesn't
make sense to put executor specific functionality on a type that is
intended for end users.
Generally speaking, it is unsafe to access to perform asynchronous
operations using `&self`. Taking `&self` allows usage from a `Sync`
context, which has unexpected results.
Taking `&mut self` to perform these operations prevents using these
asynchronous values from across tasks (unless they are wrapped in
`RefCell` or `Mutex`.
