/ior

IO Uring bindings for Crystal.

Primary LanguageCrystalMIT LicenseMIT

ior

Bindings for liburing. These bindings are not intended to be safe, or to replace the built-in ways of doing IO in Crystal. Rather this is intended to be the base other libraries and applications use so that they can accomplish that.

Installation

  1. Install liburing ( https://github.com/axboe/liburing ) and a recent Linux kernel.

  2. Add the dependency to your shard.yml:

    dependencies:
  ior:
    github: yxhuvud/ior

  3. Run shards install

Usage

The basic abstraction is the uring. A uring consists of a submission queue and a completion queue. A user writes to the submission queue and reads from the completion queue. An instance can be created by calling IOR::IOUring.new (there is also a block version available).

Operations on a ring include:

  • ring.submit: Submits the currently unsubmitted events in the submission queue. Multiple events can be submitted at once.

  • ring.wait, ring.wait(n) : Wait until there is at least one (n) entry in the completion queue. Then it returns the first of those.

Both variants supports a timeout parameter.

  • ring.submit_and_wait(n = 1): Do both submit and wait at once

  • ring.peek, ring.peek(buf): If there is an unprocessed entry in the submission queue, return it. Otherwise return nil.

If a block is supplied, then the entry will be yielded and then marked as seen.

If a block and buffer is supplied, then peek will fill as much of the buffer with completed entries as possible and then yield them and mark them as seen.

  • ring.seen(cqe): Marks the given completion queue event as seen.

  • ring.sqe, ring.sqe!: Fetches a submission queue event. Note that it can return nil if submission queue is full.

    The following operations are then supported to configure the sqe. See sqe.cr for detailed parameter options.

    • sqe.nop: Submit an event and get it back in the completion queue, without actually doing anything.
    • sqe.read, sqe.write: Reads and writes. Same as read and write syscalls. If no offset is provided it will read from current position.
    • sqe.readv, sqe.writev: Vectored writes. Same as readv and writev syscalls.
    • sqe.fsync: File sync. Same as fsync syscall, but note that the queue doesn't promise to handle events in order, by default.
    • sqe.poll_add: Similar to poll and epoll.
    • sqe.send, sqe.recv: Similar to send and recv syscalls.
    • sqe.sendmsg, sqe.recvmsg: Network read/write. Same as sendmsg and recvmsg syscalls.
    • sqe.splice: Similar to splice syscall.
    • sqe.timeout: Wait until the specified amount of events have completed, or until the given time has elapsed.
    • sqe.link_timeout: Timeout the previous op. Requires the io_link flag to be set in previous message.
    • sqe.timeout_remove: Remove a previously specified timeout.
    • sqe.async_cancel: Cancel a previously submitted operation (identified by userdata).
    • sqe.accept, sqe.connect: Similar to accept and connect syscalls.
    • sqe.fallocate: Similar to fallocate syscall.
    • sqe.openat: Similar to openat syscall.
    • sqe.close, sqe.shutdown: Similar to close and shutdown syscalls.
    • sqe.renameat, sqe.unlinkat: Similar to renameat and unlinkat syscalls.
    • sqe.files_update: Update the list of registred files. Similar to #register_files on the ring, but async.

    All SQEs supports the following options:

    • fixed_file : Use one of the previously registered files. See IOR::IOUring#register_files.
    • io_drain: Process all other entries in the ring before processing this.
    • io_link: If this is set, then the next event will not be started before the processing of this even is done.
    • io_hardlink: Same as above but different. See man pages for description.

    See https://github.com/yxhuvud/ior/blob/master/src/ior/sqe.cr and man io_uring_enter for details.

  • ring.close: Tear down the ring.

  • ring.register_files, ring.unregister_files: In addition to working with normal file descriptors, IOUring can avoid the cost of setting up internal structures for each and every call by registering a list of files in advance. To make use of these files, the submission queue event (sqe) must have the fixed_file flag set.

  • ring.sq_ready: Shows how many unsubmitted events are present in the submission queue.

  • ring.space_left: Shows how many more events can be sent before submit is necessary.

  • ring.cq_ready: Shows how many events are waiting in the completion queue.

  • ring.unsubmitted?: Returns true if there are any unsubmitted events.

  • ring.full_submission_queue?: Returns true if the submission queue needs to be submitted.

Example:

require "ior"

fh = File.open(".test/readv")
buf = Slice(UInt8).new(32) { 0u8 }
vec = LibC::IOVec.new
vec.base = buf.to_unsafe
vec.len = buf.size
iovecs = Slice(LibC::IOVec).new(1, vec)

# First create a ring:
ring = IOR::IOUring.new

# Set up a submission queue event. Multiple submissions can be sent
# out at once. They will be executed in undefined order by default,
# but it is possible to define the order if need be.

# The user_data field that is set is also set on the corresponding
# completion event, and can for example contain a pointer to something
# that instructs the program on how to handle the result.
ring.sqe.readv(fh, buf, 0, user_data: 4711)

# Then tell the kernel about it. This is a syscall that doesn't block.
ring.submit

# Then wait for the result. This will return immediately without
# having to do any syscall if there is already events ready to
# consume, and if not it will block until there is at least one event
# ready.

# wait returns an event from the completion queue.
cqe = ring.wait

# These contain the user_data and and result from the call. In this
# case the content of the file.
string = String.new(buf[0, cqe.value.res])
cqe.value.user_data # => 4711

# When processed a completion event should be marked as consumed:
ring.seen(cqe)

# Cleanup:
fh.close
ring.close

Read more

Additional resources can be found in the man pages, and at https://kernel.dk/io_uring.pdf or https://github.com/axboe/liburing .

Development

See installation instructions. Additionally some specs need the -Dpreview_mt flag set or they will block forever.

Contributing

  1. Fork it (https://github.com/yxhuvud/ior/fork)
  2. Create your feature branch (git checkout -b my-new-feature)
  3. Commit your changes (git commit -am 'Add some feature')
  4. Push to the branch (git push origin my-new-feature)
  5. Create a new Pull Request

Contributors