thomasdezeeuw / gaea Goto Github PK

Also a good test of the flexibility and usability of the event::Source trait.

The current tests aren't enough.

Write benchmarks for the following:

• Duplicated kqueue file descriptor vs. none-duplicated, see pr #45.
• Timer space event structure, see issue #2.
• User space events, see issue #25.

Currently it still refers to the Timer struct, but that was removed. Now Poller only has add and remove deadline methods, this name should reflect that.

Same treatment as #12.

Kqueue supports signals with the same API and Linux has signalfd.

Most example use a try_main function, but Result can now be returned by main directly.

From the paper: "Hashed and Hierarchical Timing Wheels: Efficient Data Structures for Implementing a Timer Facility" by George Varghese and Anthony Lauck.

Tracking issue: rust-lang/rust#58452 and stabilize pr: rust-lang/rust#60334.

See the error log here: https://cirrus-ci.com/task/5941804527517696.

One that simply describes what the examples do.

When documentation the crate from another I get the following warnings:

warning: `[Capacity::Growable]` cannot be resolved, ignoring it...
   --> /Users/thomas/.cargo/git/checkouts/mio-st-36dfa78dd2dc66e1/518d5ec/src/event.rs:208:26
    |
208 |     /// If this returns [`Capacity::Growable`] and the capacity left is
    |                          ^^^^^^^^^^^^^^^^^^^^ cannot be resolved, ignoring
    |
    = help: to escape `[` and `]` characters, just add '\' before them like `\[` or `\]`

warning: `[Source::blocking_poll]` cannot be resolved, ignoring it...
  --> /Users/thomas/.cargo/git/checkouts/mio-st-36dfa78dd2dc66e1/518d5ec/src/event.rs:91:26
   |
91 |     /// [blocking call]: Source::blocking_poll
   |                          ^^^^^^^^^^^^^^^^^^^^^ cannot be resolved, ignoring
   |
   = help: to escape `[` and `]` characters, just add '\' before them like `\[` or `\]`

warning: `[Events]` cannot be resolved, ignoring it...
  --> /Users/thomas/.cargo/git/checkouts/mio-st-36dfa78dd2dc66e1/518d5ec/src/event.rs:99:33
   |
99 |     /// Some implementation of [`Events`] have a limited available capacity.
   |                                 ^^^^^^^^ cannot be resolved, ignoring
   |
   = help: to escape `[` and `]` characters, just add '\' before them like `\[` or `\]`

warning: `[Events::capacity_left]` cannot be resolved, ignoring it...
   --> /Users/thomas/.cargo/git/checkouts/mio-st-36dfa78dd2dc66e1/518d5ec/src/event.rs:100:51
    |
100 |     /// This method may not add more events then [`Events::capacity_left`]
    |                                                   ^^^^^^^^^^^^^^^^^^^^^^^ cannot be resolved, ignoring
    |
    = help: to escape `[` and `]` characters, just add '\' before them like `\[` or `\]`

warning: `[Source::poll]` cannot be resolved, ignoring it...
   --> /Users/thomas/.cargo/git/checkouts/mio-st-36dfa78dd2dc66e1/518d5ec/src/event.rs:109:30
    |
109 |     /// This is the same as [`Source::poll`] and all requirements of that method
    |                              ^^^^^^^^^^^^^^ cannot be resolved, ignoring
    |
    = help: to escape `[` and `]` characters, just add '\' before them like `\[` or `\]`

warning: `[crate::os::OsQueue::register]` cannot be resolved, ignoring it...
  --> /Users/thomas/.cargo/git/checkouts/mio-st-36dfa78dd2dc66e1/518d5ec/src/os/evented.rs:76:31
   |
76 |     /// [`OsQueue.register`]: crate::os::OsQueue::register
   |                               ^^^^^^^^^^^^^^^^^^^^^^^^^^^^ cannot be resolved, ignoring
   |
   = help: to escape `[` and `]` characters, just add '\' before them like `\[` or `\]`

warning: `[crate::os::OsQueue::reregister]` cannot be resolved, ignoring it...
  --> /Users/thomas/.cargo/git/checkouts/mio-st-36dfa78dd2dc66e1/518d5ec/src/os/evented.rs:84:33
   |
84 |     /// [`OsQueue.reregister`]: crate::os::OsQueue::reregister
   |                                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ cannot be resolved, ignoring
   |
   = help: to escape `[` and `]` characters, just add '\' before them like `\[` or `\]`

warning: `[crate::os::OsQueue::deregister]` cannot be resolved, ignoring it...
  --> /Users/thomas/.cargo/git/checkouts/mio-st-36dfa78dd2dc66e1/518d5ec/src/os/evented.rs:92:33
   |
92 |     /// [`OsQueue.deregister`]: crate::os::OsQueue::deregister
   |                                 ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ cannot be resolved, ignoring
   |
   = help: to escape `[` and `]` characters, just add '\' before them like `\[` or `\]`

    Finished dev [unoptimized + debuginfo] target(s) in 8.90s

This would require a number of tests, following tokio-rs/mio#933, tokio-rs/mio#939 and the discussion in mio: tokio-rs/mio#941.

Some of them actually use UdpSocket.

• Update changelog.
• Bump version in Cargo.toml.
• Create git tab.
• Publish to crates.io.

Commit 64ce7af disabled it.

NetBSD doesn't have EVFILT_USER which the Awakener uses.

• Add an expect_no_events.
• Deduplicate expect_events functions in tests.

Also some other various small things after the rewrite.

If we drop bitflags and ensure that the interests provided to register and reregister are never we can drop the not_empty check. This also means we can drop PollCalled.

https://github.com/Thomasdezeeuw/mio-st/blob/090c62073c3f8cf3a9fbfd969e02b437e7aa3c2d/src/sys/unix/awakener.rs#L30

I think this may be wrong. It should be edge-triggered instead. Here is why:

https://github.com/Thomasdezeeuw/mio-st/blob/090c62073c3f8cf3a9fbfd969e02b437e7aa3c2d/tests/awakener.rs#L48-L62

If you look at the test above, it seems to be checking that two different events are triggered subsequently. But this is not true. This simply works because the eventfd has been set level-triggered, meaning that each time epoll_wait is called, it will produce an event that data is available, because of the fact that we are not reading from the eventfd at-all!

This observation can be reproduced by simply commenting-out the lines 51-54. You will see that the test passes, because you are actually testing some other semantics and not the actual one here.

Solution: The eventfd should be registered edge-triggered.

Looking at EV_RECEIPT in https://www.freebsd.org/cgi/man.cgi?query=kqueue&manpath=FreeBSD+11.2-RELEASE+and+Ports (which we use here) I don't think the call to check_errors is needed here: https://github.com/Thomasdezeeuw/mio-st/blob/master/src/sys/unix/kqueue.rs#L255.

From the manual for setsockopt:

SO_KEEPALIVE    enables keep connections alive
SO_LINGER       linger on close if data present
SO_OOBINLINE    enables reception of out-of-band data in band
SO_SNDBUF       set buffer size for output
SO_RCVBUF       set buffer size for input
SO_SNDLOWAT     set minimum count for output
SO_RCVLOWAT     set minimum count for input
SO_SNDTIMEO     set timeout value for output
SO_RCVTIMEO     set timeout value for input

It still mentions ConnectedUdpSocket, but that is removed.

Expand the testing of RegisterOption::EDGE/LEVEL | RegisterOption::ONESHOT for all types that implement Evented.

#17 (d902405) disabled testing with a stable compiler on Travis.

Currently use after fork is somewhat vague. For exampe kqueue won't work after fork, but epoll might. We need to make it consistent.

Currently poll has two generic parameters of which the error is almost never inferred, e.g. from the example of the crate docs.

poll::<_, io::Error>(&mut [&mut os_queue], &mut events, None)?;
//    ^^^^^^^^^^^^^^ Don't like this.

Currently the tests for the deadline module fail:

• deadline::multiple_deadline_same_id
• deadline::multiple_timers
• deadline::poll_should_set_timeout

All with roughly the same error; the time ran too long.

Maybe dion from https://en.wikipedia.org/wiki/Dionysus.

For example a TcpListener what Ready value should it have when registering to get all events? Try to document that for each type.

EINTR (io::ErrorKind::Interrupted), EWOULDBLOCK and EAGAIN (io::ErrorKind::WouldBlock) are three errors that aren't actually errors and can usually be safely ignored.

Since the change to a trait Events might not be the best name. Something like event::Sink might be better.

I don't like the name. The function indicates when the next event might be available and is used to reduce any provided timeout.

Even though we can't run the tests for more OSes on Travis we can at least build those targets. Locally I use the following script.

cargo build --target x86_64-apple-darwin
cargo build --target x86_64-unknown-freebsd
# Doesn't have EVFILT_USER, used in Awakener.
#cargo build --target x86_64-unknown-netbsd
cargo build --target x86_64-unknown-linux-gnu
cargo build --target x86_64-unknown-linux-musl

See Q6 of the Q&A in http://man7.org/linux/man-pages/man7/epoll.7.html and EV_DELETE on https://www.freebsd.org/cgi/man.cgi?query=kqueue&manpath=FreeBSD+11.2-RELEASE+and+Ports.

This could as simple as adding the PollOpt to RegistrationInner and checking for it to be oneshot, if so the first call to notify will also call deregister.

For TcpListener, TcpStream, UdpSocket and ConnectedUdpSocket. [::1] should do the trick, but I'm not sure if Travis supports ipv6.

Some of the use of unsafe code is incorrect I think. For example the use of mem::uninitialized should be replaced with mem::MaybeUninit (currently unstable).