Skip to content

Latest commit

 

History

History
212 lines (147 loc) · 6.37 KB

Readme.md

File metadata and controls

212 lines (147 loc) · 6.37 KB
Raw

nimlang io_uring (nimuring)

This is a pure implementation of io_uring in nim.

The library was based on several implementations of io_uring in other languages at once, such as:

Mostly on liburing

Documentation

I inserted lines from man for convenience into the binding to the io_uring kernel, but the documentation generator could not display it in normal form, so use man and read the code.

The library consists of 3 modules:

  1. actual ABI to kernel io_uring src/io_uring.nim
  2. interface for easy usage src/queue.nim
  3. and builder pattern for making SQE src/ops.nim

also there is utility module to make atomic reads/writes but I concidering it belongs to queue.

Creating Submition and Completion queues

To start submitting commands to the io_uring subsystem and then recieve completion, a convenient wrapper was made that creates and manages both rings.

import nimuring

let queue = newQueue(2, { SETUP_SQPOLL })

we have created sqe/cqe queue with ability to submiting 2 entries and revieving 4 complitions

by default completion ring is 2x size of submition ring to survive overflows (remark if overflow is actually happened you cannot submit more sqe's but it's advanced usage) There is examples/sqOverflow.nim to see an example of how to bypass this

second parameter is io_uring flags. List of available flags you can see in src/io_uring.nim

There are others parameters you can use but it's also is advanced theme.

Submitting OPS

after creating queue you can submit operations to it. The library provides the following API:

  1. you asks queue to allocate SQE in submission ring
  2. and after getting pointer to it filling it with parameters you need
  3. to not remembering all ops parameters theris utility templates to fill it.
import nimuring

let queue = newQueue(2, { SETUP_SQPOLL })
let sqe = queue.getSqe()
sqe.write(stdout.getFileHandle, "Hello world\n")
# proc write*(sqe: SqePointer; fd: FileHandle; str: string; offset: int = 0): SqePointer

queue.submit()
# all allocated previosly sqes will be submited

echo q.copyCqes(1)
# waits until completing

as you can see every op returning SqePointer back and by this builder pattern can be made

queue.getSqe().write(stdout.getFileHandle, "Hello world\n").setUserData(0)

every op utility is just set of commands which filling sqe with right converted values. so setUserData is setting sqe.userData field with given value.

Recieving Completions

io_uring completion is fully parallel. But if you wish to await completion you can wait for them blocking current thread.

  1. you can await completion when submiting
queue.getSqe().write(stdout.getFileHandle, "Hello world\n")
queue.submit(1)
## proc submit*(queue: var Queue; waitNr: uint = 0): int
# submit will wait until waitNr count of completions
  1. waiting for completions then copying them
queue.getSqe().write(stdout.getFileHandle, "Hello world\n")
queue.submit()
echo queue.copyCqes(1)
## proc copyCqes*(queue: var Queue; waitNr: uint = 0): seq[Cqe]
# waiting waitNr the same as submit(waitNr=1)
  1. dont wait just check for completions (convenient when implementing your own event loop)
queue.getSqe().write(stdout.getFileHandle, "Hello world\n")
queue.submit()
echo queue.copyCqes()
# maybe empty @[]

# also can copy to prealocated array

let cqes = newSeq[Cqe](16)
queue.copyCqes(cqes)
## will not allocate new seq but copy to specified one

How to pass user data

queue.getSqe().userData(41)
# default is NOOP
# userData can be any pointer of number
queue.submit()
echo queue.copyCqes(1)[0].userData
# must be 41!

with that there is 2 ways to pass complex user data

  1. unref data and then free it yourself
type
  myObj = object
    a: int
    b: int
let myobj = myObj(13, 37)
echo myobj
queue.getSqe().userData(myobj.addr)
queue.submit()

let cqes = queue.copyCqes(1)
let myobj2 = cast[myObj](cqes[0].userData)
echo myobj2

it works now because myobj is not freed by GC but for example if it was in a function call it will be a problem

proc example() =
  # let myobj = myObj(13, 37)
  # firstly it cannot be stack allocated at all
  
  let myobj = new myObj(13, 37)
  # ok not we allocated on heap but GC will clear it when example function is returned
  GC_ref(myobj)
  # then we say GC do not delete this object

  queue.getSqe().userData(myobj.addr)
  queue.submit()

example()

let cqes = queue.copyCqes(1)
let myobj2 = cast[myObj](cqes[0].userData)
echo myobj2

# now we can track myobj again
GC_unref(myobj2)
  1. persist it in global stash so GC know it allways here (for me it works faster)
let myData = newSqe[ref myObj](4)

proc example() =
  myData.add(new myObj(13, 37))
  let dataIndex = myData.len

  queue.getSqe().userData(dataIndex)
  queue.submit()

example()

let cqes = queue.copyCqes(1)

let myobj = myData[cqes[0].userData]
myData.delete(cqes[0].userData)

echo myobj

That's all the api

There are of course more but it all documented in man files like registering fd/buffers/files and many flags and ops

async/await

There is a way to trigger event fd when result is added into CQring. And if it is registered in asyncdispatch loop it will be dispatched in callback, which should read CQring and do something upon results

example in c: https://unixism.net/loti/tutorial/register_eventfd.html

But since the nature of io_uring is different from the nature of selectors, i don't think it is possible to combine the io_uring queue and at the same time maintain acceptable performance.

Because of this, I tried to implement another event loop, which preserve same async/await api from the standard library. https://github.com/blackmius/uasync

Examples

examples can be found in examples folder see examples of udp and tcp server in benchmark folder and also check the usage of OPS in ./tests/ops