Today the Unix and Windows implementations have similar yet differing
implementations of hooking into OS events and propagating them to any
listening futures. Rather than re-implement the same behavior two
different ways, we should factor out any commonality into a shared
module and keep the Unix/Windows modules focused solely on OS
integrations.
Reusing the same implementation across OS versions also allows for more
consistent behavior between platforms, which also makes squashing bugs
much easier.
This change introduces the `registry` module which handles creating and
initializing a global map of signals/events and their registered
listeners. Each OS specific module is expected to implement the OS hooks
which delegate to invoking the registry module's methods for
distributing the event notifications.
# Use registry module for Windows implementation
Note this still uses the same architecture as previously: a driver task
is spawned by the first registered event, and that task is responsible
for delivering any events to registered futures. (If that first event
loop goes away, all events will deadlock). A solution to this issue will
be explored at a later time.
- Use `Handle::default` over `Handle::current` for consistent semantics
- Make all `windows::Event` constructors lazily invoke `global_init`
so they can be safely constructed off-task
- Don't assume the reactor is alive and event registration will be done
when calling `global_init`
Add windows regression tests. Unfortunately, Windows doesn't have a
reliable way of programmatically sending CTRL_C or CTRL_BREAK events
to a progress, so the tests can only exercise our internal machinery by
invoking the handler that we register with the OS
Fixes#999
* Minimize allocation needed for channels
* Use a newtype for signal ids
* We can just cast the raw pointer to a `usize` and still perform a
simple identity check, without incurring any implications of storing a
raw pointer (e.g. previously Signal was !Sync and had an unsafe impl of
Send, and now it is naturally Sync+Send)
* Broadcast with `try_send` instead of `start_send`
The `Stream::start_send` method uses backpressure and schedules the
current task to be notified whenever the channel has additional room,
which means we'll generate a lot of unnecessary wakeups whenever a
channel gets full
By changing to `try_send` and handling any errors, we ensure the
Driver's task won't get woken up when a Signal finally consumes its
notification, since we're coalescing things anyway
Since the CI runs all tests for all tokio crates, it is possible that a
sporadic failure in one crate can mask failures/successes of other
crates' tests.
Using the `--no-fail-fast` flag instructs cargo to run *all* tests
before failing the build. This will allow checking to see if any
relevant test cases still pass even if an unrelated test has failed.
* Originally reported in alexcrichton/tokio-process#42
* The root cause appears to be due to two different PollEvented
instances trying to consume readiness events from the same file
descriptor.
* Previously we would simply swallow any `AlreadyExists` errors when
attempting to register the pipe receiver with the event loop. I'm not
sure if this means the PollEvented wrapper wasn't fully registered to
receive events, or maybe there is a potential race condition with how
PollEvented consumes mio readiness events. Using a fresh/duplicate file
descriptor appears to mitigate the issue, however.
* I was also not able to reproduce the issue as an isolated test case so
there is no regression test available within this crate (but we can add
one in tokio-process)
* Added timeouts to all tests that were missing them
- any issue we have will likely result in deadlocks/starvation so its
best if all tests quickly timeout rather than require getting killed
or have the CI timeout itself
* Added a `support` module and put a bunch of helpers there to DRY the
tests
* As observed in alexcrichton/tokio-signal#38, Signal instances can starve based on the order
they are created in, and this ordering appears to be platform/OS
specific
* The crux of the issue is that we woud only *attempt* to broadcast any
pending signals if we successfully read out at least one byte from the
global pipe.
* For reasons unclear to me, the affected Signal instance would get
woken up after the signal handler writes to the global pipe, but it
would immediately hit a WouldBlock error and give up, bypassing the
broadcast attempt (even though the pending flag was correctly set).
- Maybe this has to do with OS specifics with how the bytes are
delivered (or not), or with some complex interaction with tokio and
the pipe registration. It seems fishy since strace logs didn't show
the signal handler pipe write fail either, but I'm all out of ideas
* The fix appears simple: unconditionally attempt to broadcast any
pending signals *any* time a Driver instance is woken up.
* Since we perform an atomic check for each pending signal, we know that
each (coalesced) signal broadcast will happen at most once. If we were
supuriously woken up and no signals were pending, then nothing will be
yielded to any pollers of Signal
* The down side is that since each Signal instance polls a Driver
instance, each poll to Signal will essentially perform N atomic
operations (N = number of signals we support) in an attempt to broadcast
any pending signals.
- However, we can revisit optimizing this better in the future
Fixesalexcrichton/tokio-signal#38
* We introduce a new global structure which keeps track of how many
signal streams have been registered with a given event loop (the event
loop is identified by its OS file descriptor)
* We only attempt to deregister our global evented pipe from any event
loop if and only if we are the last signal that was registered with it
* Currently, whenever a new signal stream is created we attempt to
register a global pipe with the event loop to drive events.
* We also (correctly) swallow any descriptor-already-registered errors
since the same pipe is always used
* However, we currently *deregister* the same global pipe *any time* a
Signal stream is dropped.
* This means that if 2 or more of Signal instances exist simultaneously
(even if listening for different signals) and one of them is dropped,
the remainder will starve (until any new signal is created again).
* Cargo runs each integration-style-test in its own process. Since the
tests use global data structures specific to the process, we should run
them in an isolated manner to avoid having cross-test interactions
* Fixesalexcrichton/tokio-signal#39
Ivan Petkov