more asyncio foo

runtime/asyncio/Cargo.toml

@@ -13,4 +13,13 @@ nix = "0.23"
 bitflags = "1.3"
 ptr_meta = "0.1"
 
-crossbeam-queue = "0.3"
+# SegQueue for task waiting on CQE or SQE being available again.
+crossbeam-queue = "0.3"
+
+# AsyncRead, AsyncWrite, AsyncSeek and related traits
+futures-io = "0.3"
+
+[dev-dependencies]
+# As Mr. Torgue would put it: THE MOST EXTREME F*CKING ASYNC FUNCTION RUNNNER! EXPLOSIONS!
+extreme = "666.666.666666"
+futures-lite = "1.12"

runtime/asyncio/examples/future.rs

@@ -0,0 +1,63 @@
+use std::fs::File;
+use std::future::Future;
+use std::io;
+use std::os::unix::prelude::AsRawFd;
+use std::sync::Arc;
+use std::task::{Context, Poll, RawWaker, RawWakerVTable, Waker};
+use asyncio::ctypes::IORING_OP;
+use asyncio::io_uring::IoUring;
+
+use futures_lite::io::AsyncReadExt;
+
+pub fn drive<T>(iouring: &IoUring, mut f: impl Future<Output = io::Result<T>>) -> io::Result<T> {
+    static VTABLE: RawWakerVTable = RawWakerVTable::new(
+        |clone_me| unsafe {
+            let arc = Arc::from_raw(clone_me);
+            std::mem::forget(arc.clone());
+            RawWaker::new(Arc::into_raw(arc) as *const (), &VTABLE)
+        },
+        |wake_me| unsafe { Arc::from_raw(wake_me); },
+        |wake_by_ref_me| unsafe {},
+        |drop_me| unsafe { drop(Arc::from_raw(drop_me)) },
+    );
+
+    let mut f = unsafe { std::pin::Pin::new_unchecked(&mut f) };
+    let park = Arc::new(());
+    let sender = Arc::into_raw(park.clone());
+    let raw_waker = RawWaker::new(sender as *const _, &VTABLE);
+    let waker = unsafe { Waker::from_raw(raw_waker) };
+    let mut cx = Context::from_waker(&waker);
+
+    loop {
+        match f.as_mut().poll(&mut cx) {
+            Poll::Ready(t) => return t,
+            Poll::Pending => {
+                iouring.handle_completions();
+                match iouring.submit_wait() {
+                    Ok(_) => {}
+                    Err(e) => return Err(e),
+                }
+            }
+        }
+    }
+}
+
+fn main() {
+    let file = File::open("/tmp/poem").unwrap();
+    let fd = file.as_raw_fd();
+
+    let mut ring: &'static IoUring = Box::leak(Box::new(IoUring::setup(4).unwrap()));
+
+    let mut async_file = asyncio::fs::File::new(fd, ring);
+
+    let mut buf = Box::new([0u8; 4096]);
+
+    let f = async move {
+        let len = async_file.read(&mut buf[..]).await?;
+        println!("Read {} bytes:", len);
+        let str = unsafe { std::str::from_utf8_unchecked(&buf[..len]) };
+        println!("{}", str);
+        Ok(())
+    };
+    drive(ring, f);
+}

runtime/asyncio/examples/raw.rs

@@ -1,13 +1,54 @@
 use std::fs::File;
+use std::os::unix::prelude::AsRawFd;
 use asyncio::ctypes::IORING_OP;
 use asyncio::io_uring::IoUring;
 
+
 fn main() {
-    let file = File::open("")
+    let file = File::open("/tmp/poem").unwrap();
-    let ring = IoUring::setup(64).unwrap();
+    let fd = file.as_raw_fd();
-    let cqes = ring.cqes();
+
-    ring.try_prepare(1, |sqes| {
+    let ring = IoUring::setup(4).unwrap();
-        let sqe = sqes.next().unwrap();
+    let mut cqes = ring.cqes();
-        sqe.set_opcode(IORING_OP::READ)
+
+    let buf = Box::new([0u8; 4096]);
+    ring.try_prepare(3, |mut sqes| {
+        let mut sqe = sqes.next().unwrap();
+        sqe.set_opcode(IORING_OP::READ);
+        sqe.set_address(buf.as_ptr() as u64);
+        sqe.set_fd(fd);
+        sqe.set_len(4096);
+
+        let mut sqe = sqes.next().unwrap();
+        sqe.set_opcode(IORING_OP::NOP);
+        sqe.set_userdata(0xCAFEBABE);
+
+        let mut sqe = sqes.next().unwrap();
+        sqe.set_opcode(IORING_OP::NOP);
+        sqe.set_userdata(0xDEADBEEF);
     }).unwrap();
+    let mut amt  = 0;
+    while amt == 0 {
+        amt = ring.submit().unwrap();
+    }
+    println!("{}", amt);
+
+    for _ in 0..3 {
+        let mut cqe = None;
+        while cqe.is_none() {
+            cqe = cqes.next();
+        }
+        let cqe = cqe.unwrap();
+        println!("{:?}", cqe);
+        if cqe.user_data == 0xCAFEBABE {
+            println!("cafebabe");
+        } else if cqe.user_data == 0xDEADBEEF {
+            println!("deadbeef");
+        }
+
+        if let Ok(len) = cqe.result() {
+            let out = unsafe { std::str::from_utf8_unchecked(&buf[0..len as usize]) };
+            println!("{}", out);
+        }
+    }
 }

runtime/asyncio/poem

@@ -1,19 +0,0 @@
-Dunkel war’s, der Mond schien helle,
-schneebedeckt die grüne Flur,
-als ein Wagen blitzesschnelle,
-langsam um die Ecke fuhr.
-
-Drinnen saßen stehend Leute,
-schweigend ins Gespräch vertieft,
-als ein totgeschoss’ner Hase
-auf der Sandbank Schlittschuh lief.
-
-Und ein blondgelockter Jüngling
-mit kohlrabenschwarzem Haar
-saß auf einer grünen Kiste,
-die rot angestrichen war.
-
-Neben ihm ’ne alte Schrulle,
-zählte kaum erst sechzehn Jahr,
-in der Hand ’ne Butterstulle,
-die mit Schmalz bestrichen war.

runtime/asyncio/src/completion.rs

@@ -5,23 +5,32 @@ use std::mem::ManuallyDrop;
 use std::task::Waker;
 use crate::cancellation::Cancellation;
 
-pub struct Completion<'cx> {
+// TODO: Completions for linked requests? How would you handle having multiple results? In one
+//       Completion struct or using multiple? If the latter, prepare needs to set user_data
+//       for all intermediary SQE explicitly.
+pub struct Completion {
     state: ManuallyDrop<Box<Cell<State>>>,
-    marker: PhantomData<fn(&'cx ()) -> &'cx ()>,
 }
 
 enum State {
     Submitted(Waker),
-    Completed(io::Result<u32>),
+    Completed(io::Result<i32>),
     Cancelled(Cancellation),
     Empty,
 }
 
-impl<'cx> Completion<'cx> {
+impl Completion {
     pub fn new(waker: Waker) -> Self {
         Self {
             state: ManuallyDrop::new(Box::new(Cell::new(State::Submitted(waker)))),
-            marker: PhantomData,
+        }
+    }
+
+    pub(crate) unsafe fn from_raw(ptr: u64) -> Self {
+        let ptr = ptr as usize as *mut Cell<State>;
+        let state = ManuallyDrop::new(Box::from_raw(ptr));
+        Self {
+            state,
         }
     }
 
@@ -29,7 +38,7 @@ impl<'cx> Completion<'cx> {
         self.state.as_ptr() as *const _ as usize as u64
     }
 
-    pub fn check(self, waker: &Waker) -> Result<io::Result<u32>, Self> {
+    pub fn check(self, waker: &Waker) -> Result<io::Result<i32>, Self> {
         match self.state.replace(State::Empty) {
             State::Submitted(old_waker) => {
                 // If the given waker wakes a different task than the one we were constructed
@@ -59,7 +68,7 @@ impl<'cx> Completion<'cx> {
         }
     }
 
-    pub fn complete(self, result: io::Result<u32>) {
+    pub fn complete(self, result: io::Result<i32>) {
         match self.state.replace(State::Completed(result)) {
             State::Submitted(waker) => {
                 waker.wake();

runtime/asyncio/src/cq.rs

@@ -1,8 +1,12 @@
 use std::cell::UnsafeCell;
 use std::os::unix::prelude::RawFd;
+use std::pin::Pin;
 use std::ptr::NonNull;
-use std::sync::atomic::{AtomicU32, Ordering};
+use std::sync::atomic::{AtomicU32, compiler_fence, Ordering};
+use std::task::{Context, Poll, Waker};
+use crossbeam_queue::SegQueue;
 use nix::sys::mman::munmap;
+use crate::completion::Completion;
 use crate::cqe::CQE;
 use crate::ctypes::{CQOffsets, IORING_CQ};
 
@@ -15,18 +19,24 @@ pub struct CQ {
     /// main problem that can happen otherwise is that the kernel assumes it lost completions
     /// which we already successfully pulled from the queue.
     head: &'static AtomicU32,
-    cached_head: UnsafeCell<u32>,
 
     /// Tail of the completion queue. Moved by the kernel when new completions are stored.
     ///
     /// Since this is modified by the kernel we should use atomic operations to read it, making
     /// sure both the kernel and any program have a consistent view of its contents.
     tail: &'static AtomicU32,
+
+    /// A cached version of `tail` which additionally counts reserved slots for future
+    /// completions, i.e. slots that the kernel will fill in the future.
+    predicted_tail: UnsafeCell<u32>,
+
     ring_mask: u32,
     num_entries: u32,
     flags: &'static AtomicU32,
     entries: &'static [CQE],
 
+    waiters: SegQueue<Waker>,
+
     // cq_ptr is set to `None` if we used a single mmap for both SQ and CQ.
     cq_ptr: *mut libc::c_void,
     cq_map_size: usize,
@@ -58,8 +68,8 @@ impl CQ {
         let num_entries = *(ptr.offset(offs.ring_entries as isize).cast());
 
         let head: &AtomicU32 = &*(ptr.offset(offs.head as isize).cast());
-        let cached_head = UnsafeCell::new(head.load(Ordering::Acquire));
         let tail: &AtomicU32 = &*(ptr.offset(offs.tail as isize).cast());
+        let predicted_tail = UnsafeCell::new(head.load(Ordering::Acquire));
         let flags: &AtomicU32 = &*(ptr.offset(offs.flags as isize).cast());
         let entries = std::slice::from_raw_parts(
             ptr.offset(offs.cqes as isize).cast(),
@@ -68,7 +78,7 @@ impl CQ {
 
         Self {
             head,
-            cached_head,
+            predicted_tail,
             tail,
             ring_mask,
             num_entries,
@@ -76,18 +86,61 @@ impl CQ {
 
             entries,
 
+            waiters: SegQueue::new(),
+
             // Only store a pointer if we used a separate mmap() syscall for the CQ
             cq_ptr: if split_mmap { ptr } else { std::ptr::null_mut() },
             cq_map_size,
         }
     }
 
+    #[inline(always)]
+    fn predicted_tail(&self) -> &mut u32 {
+        unsafe { &mut *self.predicted_tail.get() }
+    }
+
+    #[inline(always)]
+    /// Currently used + reserved slots
+    pub fn used(&self) -> u32 {
+        let tail = *self.predicted_tail();
+        let head = self.head.load(Ordering::Relaxed);
+        compiler_fence(Ordering::Acquire);
+        tail.wrapping_sub(head)
+    }
+
+    #[inline(always)]
+    /// Amount of available slots taking reservations into account.
+    pub fn available(&self) -> u32 {
+        self.num_entries - self.used()
+    }
+
+    /// Try to reserve a number of CQ slots to make sure that
+    pub fn try_reserve(&self, count: u32) -> bool {
+        if self.available() >= count {
+            let tail = self.predicted_tail();
+            *tail = (*tail).wrapping_add(count);
+            true
+        } else {
+            false
+        }
+    }
+
+    pub fn poll_reserve(self: Pin<&mut Self>, ctx: &mut Context<'_>, count: u32) -> Poll<()> {
+        if self.available() >= count {
+            Poll::Ready(())
+        } else {
+            self.waiters.push(ctx.waker().clone());
+            Poll::Pending
+        }
+    }
+
     pub fn get_next(&self) -> Option<&CQE> {
-        let tail = self.tail.load(Ordering::Acquire);
+        let tail = self.tail.load(Ordering::Relaxed);
-        let head = self.head.load(Ordering::Acquire);
+        let head = self.head.load(Ordering::Relaxed);
         if tail == head {
             None
         } else {
+            compiler_fence(Ordering::Acquire);
             self.head.fetch_add(1, Ordering::Release);
             let index = (head & self.ring_mask) as usize;
             Some(&self.entries[index])
@@ -95,8 +148,153 @@ impl CQ {
     }
 
     pub fn ready(&self) -> u32 {
-        let tail = self.tail.load(Ordering::Acquire);
+        let tail = self.tail.load(Ordering::Relaxed);
-        let head = self.head.load(Ordering::Acquire);
+        let head = self.head.load(Ordering::Relaxed);
+        compiler_fence(Ordering::Acquire);
         tail.wrapping_sub(head)
     }
+
+    pub fn handle(&self, handler: impl Fn(&CQE)) {
+        let tail = self.tail.load(Ordering::Relaxed);
+        let head = self.head.load(Ordering::Relaxed);
+
+        for i in head..tail {
+            let index = (i & self.ring_mask) as usize;
+            let cqe = &self.entries[index];
+            handler(cqe);
+        }
+
+        compiler_fence(Ordering::Acquire);
+        self.head.store(tail, Ordering::Release);
+    }
+
+    #[cfg(test)]
+    fn test_insert_cqe(&self, cqe: impl Iterator<Item=CQE>) {
+        let head = self.head.load(Ordering::Relaxed);
+        let mut tail = self.tail.load(Ordering::Acquire);
+        unsafe {
+            for entry in cqe {
+                let index = (tail & self.ring_mask) as usize;
+                // Yes, this is absolutely not safe or defined behaviour in the first place. This
+                // function must *never* be used outside simple testing setups.
+                let ptr = &self.entries[index] as *const _ as *mut CQE;
+                ptr.write(entry);
+                tail += 1;
+
+                // If we would overflow, crash instead
+                assert!((tail - head) <= self.num_entries, "test_insert_cqe overflowed the buffer");
+            }
+        }
+        self.tail.store(tail, Ordering::Release);
+    }
+}
+
+mod tests {
+    use std::sync::atomic::AtomicU64;
+    use super::*;
+
+    fn gen_cq(num_entries: u32) -> CQ {
+        let head = Box::leak(Box::new(AtomicU32::new(0)));
+        let tail = Box::leak(Box::new(AtomicU32::new(0)));
+        let flags = Box::leak(Box::new(AtomicU32::new(0)));
+        let entries = Box::leak((0..num_entries).map(|_| CQE::default()).collect());
+
+        CQ {
+            head,
+            tail,
+            predicted_tail: UnsafeCell::new(0),
+            ring_mask: num_entries - 1,
+            num_entries,
+            flags,
+            entries,
+            cq_ptr: std::ptr::null_mut(),
+            cq_map_size: 0,
+            waiters: SegQueue::new(),
+        }
+    }
+
+    #[test]
+    fn test_test_insert_cqe() {
+        let cq = gen_cq(4);
+        cq.test_insert_cqe([
+            CQE {
+                user_data: 1,
+                .. Default::default()
+            },
+            CQE {
+                user_data: 2,
+                .. Default::default()
+            },
+            CQE {
+                user_data: 3,
+                .. Default::default()
+            },
+            CQE {
+                user_data: 4,
+                .. Default::default()
+            },
+        ].into_iter());
+        println!("{:?}", cq.entries);
+        for i in 0..4 {
+            assert_eq!(cq.entries[i].user_data, (i+1) as u64);
+        }
+    }
+
+    #[test]
+    #[should_panic]
+    fn test_test_insert_cqe_overflow() {
+        let cq = gen_cq(2);
+        cq.test_insert_cqe([
+            CQE {
+                user_data: 1,
+                .. Default::default()
+            },
+            CQE {
+                user_data: 2,
+                .. Default::default()
+            },
+            CQE {
+                user_data: 3,
+                .. Default::default()
+            },
+            CQE {
+                user_data: 4,
+                .. Default::default()
+            },
+        ].into_iter());
+        println!("{:?}", cq.entries);
+    }
+
+    #[test]
+    fn test_cq_reserve_insert() {
+        let cq = gen_cq(4);
+        assert_eq!(cq.tail.load(Ordering::Relaxed), 0);
+        assert_eq!(cq.head.load(Ordering::Relaxed), 0);
+        assert_eq!(*cq.predicted_tail(), 0);
+
+        cq.try_reserve(2);
+        assert_eq!(cq.tail.load(Ordering::Relaxed), 0);
+        assert_eq!(*cq.predicted_tail(), 2);
+
+        cq.test_insert_cqe([
+            CQE {
+                user_data: 1,
+                .. Default::default()
+            },
+            CQE {
+                user_data: 2,
+                .. Default::default()
+            },
+        ].into_iter());
+
+        assert_eq!(cq.head.load(Ordering::Relaxed), 0);
+        assert_eq!(cq.tail.load(Ordering::Relaxed), 2);
+        assert_eq!(*cq.predicted_tail(), 2);
+
+        let mut o = AtomicU64::new(1);
+        cq.handle(|cqe| {
+            assert_eq!(cqe.user_data, o.fetch_add(1, Ordering::Relaxed))
+        });
+        assert_eq!(o.load(Ordering::Relaxed), 3);
+    }
 }

runtime/asyncio/src/cqe.rs

@@ -9,7 +9,7 @@ use crate::io_uring::{IoUring};
 /// Completion Queue Event
 pub struct CQE {
     pub user_data: u64,
-    res: i32,
+    pub(crate) res: i32,
     pub flags: IOCQE,
 }
 

runtime/asyncio/src/fs.rs

@@ -0,0 +1,47 @@
+use std::cell::Cell;
+use std::io::IoSliceMut;
+use std::os::unix::prelude::RawFd;
+use std::pin::Pin;
+use std::task::{Context, Poll};
+use futures_io::AsyncRead;
+use crate::completion::Completion;
+use crate::ctypes::IORING_OP;
+use crate::io_uring::IoUring;
+use crate::sqe::{SQE, SQEs};
+use crate::submission::Submission;
+
+pub struct File {
+    fd: RawFd,
+    submission: Submission,
+}
+
+impl File {
+    pub fn new(fd: RawFd, io_uring: &'static IoUring) -> Self {
+        Self { fd, submission: Submission::new(io_uring) }
+    }
+
+    fn prepare_read<'sq>(
+        fd: RawFd,
+        buf: &mut [u8],
+        sqes: &mut SQEs<'sq>,
+    ) -> SQE<'sq>
+    {
+        let mut sqe = sqes.next().expect("prepare_read requires at least one SQE");
+        sqe.set_opcode(IORING_OP::READ);
+        sqe.set_address(buf.as_ptr() as u64);
+        sqe.set_fd(fd);
+        sqe.set_len(buf.len() as i32);
+        sqe
+    }
+}
+
+impl AsyncRead for File {
+    fn poll_read(mut self: Pin<&mut Self>, ctx: &mut Context<'_>, buf: &mut [u8])
+        -> Poll<std::io::Result<usize>>
+    {
+        let fd = self.fd;
+        Pin::new(&mut self.submission).poll(ctx, 1, |sqes| {
+            Self::prepare_read(fd, buf, sqes)
+        }).map(|res| res.map(|val| val as usize))
+    }
+}

runtime/asyncio/src/io_uring.rs

@@ -1,20 +1,22 @@
 use std::fmt::{Debug, Formatter};
+use std::future::Future;
 use std::io;
 use std::marker::PhantomData;
 use std::mem::{size_of, align_of};
+use std::ops::Deref;
 use std::sync::atomic::{AtomicU32, Ordering};
 use std::os::unix::prelude::RawFd;
 use std::pin::Pin;
 use std::ptr::NonNull;
-use std::task::{Context, Poll, Waker};
+use std::sync::Arc;
-use crossbeam_queue::SegQueue;
+use std::task::{Context, Poll, RawWaker, RawWakerVTable, Waker};
 use nix::sys::{mman, mman::{MapFlags, ProtFlags}};
 use crate::completion::Completion;
 use crate::cq::CQ;
 use crate::cqe::{CQE, CQEs};
 use crate::ctypes::{CQOffsets, IORING_ENTER, SQOffsets};
 use crate::sq::SQ;
-use crate::sqe::SQEs;
+use crate::sqe::{SQE, SQEs};
 use super::ctypes::{Params, io_uring_sqe, IORING_CQ, IORING_FEAT,
                     IORING_OFF_CQ_RING, IORING_OFF_SQ_RING, IORING_OFF_SQES, IORING_SQ};
 use super::syscall;
@@ -25,8 +27,6 @@ pub struct IoUring {
     params: Params,
     sq: SQ,
     cq: CQ,
-
-    waiting: SegQueue<(u32, Waker)>,
 }
 
 unsafe fn mmap(map_size: usize, fd: RawFd, offset: i64) -> nix::Result<*mut libc::c_void> {
@@ -49,7 +49,6 @@ impl IoUring {
             (params.sq_entries as usize) * size_of::<u32>();
         let mut cq_map_size = (params.cq_off.cqes as usize) +
             (params.cq_entries as usize) * size_of::<CQE>();
-        println!("{:?} {}", params.sq_off, sq_map_size);
 
         // If we can use a single mmap() syscall to map sq, cq and cqe the size of the total map
         // is the largest of `sq_map_size` and `cq_map_size`.
@@ -58,7 +57,6 @@ impl IoUring {
             cq_map_size = sq_map_size;
         }
 
-        println!("{:?}", params.cq_off);
         let sq_ptr = unsafe {
             mmap(sq_map_size as usize, fd, IORING_OFF_SQ_RING as i64)?
         };
@@ -99,29 +97,72 @@ impl IoUring {
             params,
             sq,
             cq,
-            waiting: SegQueue::new(),
         })
     }
 
-    pub fn try_prepare<'cx>(
+    pub fn try_prepare(
         &self,
         count: u32,
         prepare: impl FnOnce(SQEs<'_>)
     ) -> Option<()> {
-        // TODO: Lock the required amount of slots on both submission and completion queue, then
+        self.handle_completions();
-        //       construct the sqes.
+        if !self.cq.try_reserve(count) {
+            return None;
+        }
+
         if let Some(sqes) = self.sq.try_reserve(count) {
-            Some(prepare(sqes))
+            let start = sqes.start();
+            prepare(sqes);
+            self.sq.prepare(start, count);
+            Some(())
         } else {
             None
         }
     }
 
-    pub fn submit(&self) -> io::Result<u32> {
+    pub fn poll_prepare<'cx>(
-        self.sq.submit(self.fd)
+        mut self: Pin<&mut Self>,
+        ctx: &mut Context<'cx>,
+        count: u32,
+        prepare: impl for<'sq> FnOnce(SQEs<'sq>, &mut Context<'cx>) -> Completion
+    ) -> Poll<Completion> {
+        Pin::new(&mut self.sq).poll_prepare(ctx, count, prepare)
     }
 
-    pub fn cqes(&self) -> CQEs {
+    pub fn poll_submit<'cx>(
-        CQEs::new(&self.cq)
+        mut self: Pin<&mut Self>,
+        ctx: &mut Context<'cx>,
+        head: u32,
+    ) -> Poll<()> {
+        let fd = self.fd;
+        Pin::new(&mut self.sq).poll_submit(ctx, fd, head)
+    }
+
+    pub fn submit_wait(&self) -> io::Result<u32> {
+        self.sq.submit_wait(self.fd)
+    }
+
+    pub fn handle_completions(&self) {
+        self.cq.handle(|cqe| {
+            let udata = cqe.user_data;
+            if udata != 0 {
+                let completion = unsafe {
+                    Completion::from_raw(udata)
+                };
+                completion.complete(cqe.result())
+            }
+        });
+    }
+}
+
+impl Future for &IoUring {
+    type Output = io::Result<()>;
+
+    fn poll(self: Pin<&mut Self>, cx: &mut Context<'_>) -> Poll<Self::Output> {
+        self.handle_completions();
+        match self.sq.submit(self.fd, Some(cx.waker())) {
+            Ok(_) => Poll::Pending,
+            Err(e) => Poll::Ready(Err(e)),
+        }
     }
 }

runtime/asyncio/src/lib.rs

@@ -9,9 +9,11 @@ mod sqe;
 mod cq;
 mod cqe;
 
+mod submission;
 mod completion;
 mod cancellation;
 
+pub mod fs;
 
 #[macro_export]
 macro_rules! ready {

runtime/asyncio/src/sq.rs

@@ -1,13 +1,17 @@
-use std::cell::UnsafeCell;
+use std::cell::{Cell, UnsafeCell};
 use std::fmt::{Debug, Formatter};
 use std::io;
 use std::mem::ManuallyDrop;
 use std::os::unix::prelude::RawFd;
+use std::pin::Pin;
 use std::ptr::NonNull;
 use std::sync::atomic::{AtomicU32, compiler_fence, fence, Ordering};
+use std::task::{Context, Poll, Waker};
+use crossbeam_queue::SegQueue;
 use nix::sys::mman::munmap;
+use crate::completion::Completion;
 use crate::ctypes::{IORING_ENTER, IORING_SQ, io_uring_sqe, SQOffsets};
-use crate::sqe::SQEs;
+use crate::sqe::{SQE, SQEs};
 use crate::syscall;
 
 pub struct SQ {
@@ -46,6 +50,10 @@ pub struct SQ {
     sq_ptr: NonNull<()>,
     sq_map_size: usize,
     sqes_map_size: usize,
+
+    /// Queue of tasks waiting for a submission, either because they need free slots or because
+    waiters: SegQueue<Waker>,
+    submitter: Cell<Option<Waker>>,
 }
 
 static_assertions::assert_not_impl_any!(SQ: Send, Sync);
@@ -125,6 +133,8 @@ impl SQ {
             sq_ptr,
             sq_map_size,
             sqes_map_size,
+            waiters: SegQueue::new(),
+            submitter: Cell::new(None),
         }
     }
 
@@ -169,9 +179,50 @@ impl SQ {
         self.num_entries - self.used()
     }
 
+    #[inline(always)]
+    fn to_submit(&self) -> u32 {
+        let shared_tail = self.array_tail.load(Ordering::Relaxed);
+        let cached_tail = *self.cached_tail();
+        cached_tail.wrapping_sub(shared_tail)
+    }
+
+    pub fn submit_wait(&self, fd: RawFd) -> io::Result<u32> {
+        // Ensure that the writes into the array are not moved after the write of the tail.
+        // Otherwise kernelside may read completely wrong indices from array.
+        compiler_fence(Ordering::Release);
+        self.array_tail.store(*self.cached_tail(), Ordering::Release);
+
+        let retval = syscall::enter(
+            fd,
+            self.num_entries,
+            1,
+            IORING_ENTER::GETEVENTS,
+            std::ptr::null(),
+            0,
+        )? as u32;
+        // Return SQE into circulation that we successfully submitted to the kernel.
+        self.increment_head(retval);
+        self.notify();
+        Ok(retval)
+    }
+
     /// Submit all prepared entries to the kernel. This function will return the number of
     /// entries successfully submitted to the kernel.
-    pub fn submit(&self, fd: RawFd) -> io::Result<u32> {
+    pub fn submit(&self, fd: RawFd, waker: Option<&Waker>) -> io::Result<u32> {
+        if let Some(waker) = waker {
+            let new = if let Some(old) = self.submitter.take() {
+                if old.will_wake(waker) { old } else { waker.clone() }
+            } else {
+                waker.clone()
+            };
+            self.submitter.set(Some(new));
+        }
+
+        // Ensure that the writes into the array are not moved after the write of the tail.
+        // Otherwise kernelside may read completely wrong indices from array.
+        compiler_fence(Ordering::Release);
+        self.array_tail.store(*self.cached_tail(), Ordering::Release);
+
         let retval = syscall::enter(
             fd,
             self.num_entries,
@@ -182,6 +233,7 @@ impl SQ {
         )? as u32;
         // Return SQE into circulation that we successfully submitted to the kernel.
         self.increment_head(retval);
+        self.notify();
         Ok(retval)
     }
 
@@ -204,7 +256,7 @@ impl SQ {
     /// point to `count` entries in `self.sqes` starting at `start`. This allows actions to be
     /// submitted to the kernel even when there are still reserved SQE in between that weren't yet
     /// filled.
-    fn prepare(&self, start: u32, count: u32) {
+    pub fn prepare(&self, start: u32, count: u32) {
         // Load the tail of the array (i.e. where we will start filling)
         let tail = self.cached_tail();
         let mut head = start;
@@ -221,10 +273,59 @@ impl SQ {
             head = head.wrapping_add(1);
             *tail = (*tail).wrapping_add(1);
         }
-        // Ensure that the writes into the array are not moved after the write of the tail.
+
-        // Otherwise kernelside may read completely wrong indices from array.
+        // FIXME: This should really be done by epoll
-        compiler_fence(Ordering::Release);
+        if let Some(waker) = self.submitter.take() {
-        self.array_tail.store(*tail, Ordering::Release);
+            waker.wake_by_ref();
+            self.submitter.set(Some(waker));
+        }
+    }
+
+    pub fn poll_prepare<'cx>(
+        self: Pin<&mut Self>,
+        ctx: &mut Context<'cx>,
+        count: u32,
+        prepare: impl for<'sq> FnOnce(SQEs<'sq>, &mut Context<'cx>) -> Completion
+    ) -> Poll<Completion> {
+        if let Some(sqes) = self.try_reserve(count) {
+            let start = sqes.start();
+            let completion = prepare(sqes, ctx);
+            self.prepare(start, count);
+            Poll::Ready(completion)
+        } else {
+            self.waiters.push(ctx.waker().clone());
+            Poll::Pending
+        }
+    }
+
+    /// Suggest to submit pending events to the kernel. Returns `Ready` when the relevant event
+    /// was submitted to the kernel, i.e. when kernelside `head` >= the given `head`.
+    pub fn poll_submit(self: Pin<&mut Self>, ctx: &mut Context<'_>, fd: RawFd, head: u32)
+        -> Poll<()>
+    {
+        let shared_tail = self.array_tail.load(Ordering::Relaxed);
+        let cached_tail = *self.cached_tail();
+        let to_submit = cached_tail.wrapping_sub(shared_tail);
+
+        // TODO: Do some smart cookie thinking here and batch submissions in a sensible way
+        if to_submit > 4 {
+            self.submit(fd, None);
+        }
+
+        if *self.sqes_head() < head {
+            self.waiters.push(ctx.waker().clone());
+            Poll::Pending
+        } else {
+            Poll::Ready(())
+        }
+    }
+
+    pub fn notify(&self) {
+        if self.waiters.len() > 0 && self.available() > 0 {
+            while let Some(waker) = self.waiters.pop() {
+                waker.wake()
+            }
+        }
     }
 
     pub fn try_reserve(&self, count: u32) -> Option<SQEs<'_>> {
@@ -273,7 +374,9 @@ mod tests {
                 sqes,
                 sq_ptr: NonNull::dangling(),
                 sq_map_size: 0,
-                sqes_map_size: 0
+                sqes_map_size: 0,
+                waiters: SegQueue::new(),
+                submitter: Cell::new(None),
             })
         }
     }

runtime/asyncio/src/sqe.rs

@@ -1,7 +1,8 @@
 use std::cell::UnsafeCell;
 use std::ops::{Deref, DerefMut};
+use std::os::unix::prelude::RawFd;
 use std::slice::IterMut;
-use crate::ctypes::{IORING_OP, IOSQE, io_uring_sqe};
+use crate::ctypes::{IORING_OP, IOSQE, io_uring_sqe, SQEOpFlags};
 
 #[derive(Debug)]
 pub struct SQE<'iou> {
@@ -37,6 +38,26 @@ impl<'iou> SQE<'iou> {
     pub fn set_len(&mut self, len: i32) {
         self.sqe.len = len;
     }
+
+    #[inline(always)]
+    pub fn set_fd(&mut self, fd: RawFd) {
+        self.sqe.fd = fd;
+    }
+
+    #[inline(always)]
+    pub fn set_offset(&mut self, offset: u64) {
+        self.sqe.offset = offset;
+    }
+
+    #[inline(always)]
+    pub fn set_op_flags(&mut self, op_flags: SQEOpFlags) {
+        self.sqe.op_flags = op_flags;
+    }
+
+    pub fn prepare_cancel(&mut self, user_data: u64) {
+        self.set_opcode(IORING_OP::ASYNC_CANCEL);
+        self.set_address(user_data);
+    }
 }
 
 pub struct SQEs<'iou> {
@@ -90,6 +111,10 @@ impl<'iou> SQEs<'iou> {
         self.capacity
     }
 
+    pub fn used(&self) -> u32 {
+        self.count
+    }
+
     fn consume(&mut self) -> Option<SQE<'iou>> {
         if self.count >= self.capacity {
             None

runtime/asyncio/src/submission.rs

@@ -0,0 +1,136 @@
+use std::io;
+use std::pin::Pin;
+use std::task::{Context, Poll};
+use crate::cancellation::Cancellation;
+use crate::completion::Completion;
+use crate::io_uring::IoUring;
+use crate::sq::SQ;
+use crate::sqe::{SQE, SQEs};
+
+pub struct Submission {
+    iouring: &'static IoUring,
+    state: State,
+}
+
+enum State {
+    Inert,
+    Prepared(u32, Completion),
+    Submitted(Completion),
+    Cancelled(u64),
+    Lost,
+}
+
+impl Submission {
+    pub fn new(iouring: &'static IoUring) -> Self {
+        Self { iouring, state: State::Inert }
+    }
+
+    fn split_pinned(self: Pin<&mut Self>) -> (Pin<&mut IoUring>, &mut State) {
+        unsafe {
+            let this = Pin::get_unchecked_mut(self);
+            let iouring = &mut *(this.iouring as *const _ as *mut _);
+            (Pin::new_unchecked(iouring), &mut this.state)
+        }
+    }
+
+    pub fn poll(
+        mut self: Pin<&mut Self>,
+        ctx: &mut Context<'_>,
+        count: u32,
+        prepare: impl for<'sq> FnOnce(&mut SQEs<'sq>) -> SQE<'sq>
+    ) -> Poll<io::Result<i32>> {
+        match self.state {
+            State::Inert | State::Cancelled(_) => {
+                let head = crate::ready!(self.as_mut().poll_prepare(ctx, count, prepare));
+                crate::ready!(self.as_mut().poll_submit(ctx, head));
+                self.poll_complete(ctx)
+            },
+            State::Prepared(head, _) => {
+                crate::ready!(self.as_mut().poll_submit(ctx, head));
+                self.poll_complete(ctx)
+            },
+            State::Submitted(_) => self.poll_complete(ctx),
+            State::Lost => {
+                panic!("Ring in invalid state")
+            },
+        }
+    }
+
+    pub fn poll_prepare(
+        self: Pin<&mut Self>,
+        ctx: &mut Context<'_>,
+        count: u32,
+        prepare: impl for<'sq> FnOnce(&mut SQEs<'sq>) -> SQE<'sq>
+    ) -> Poll<u32> {
+        let (sq, state) = self.split_pinned();
+        let mut head = 0u32;
+        let completion = match *state {
+            State::Inert => {
+                crate::ready!(sq.poll_prepare(ctx, count, |mut sqes, ctx| {
+                    *state = State::Lost;
+
+                    let mut sqe = prepare(&mut sqes);
+                    let completion = Completion::new(ctx.waker().clone());
+                    sqe.set_userdata(completion.addr());
+
+                    head = sqes.used();
+                    completion
+                }))
+            },
+            State::Cancelled(prev) => {
+                crate::ready!(sq.poll_prepare(ctx, count + 1, |mut sqes, ctx| {
+                    *state = State::Lost;
+
+                    sqes.soft_linked().next().unwrap().prepare_cancel(prev);
+
+                    let mut sqe = prepare(&mut sqes);
+                    let completion = Completion::new(ctx.waker().clone());
+                    sqe.set_userdata(completion.addr());
+
+                    head = sqes.used();
+                    completion
+                }))
+            },
+            _ => unreachable!(),
+        };
+        *state = State::Prepared(head, completion);
+        Poll::Ready(head)
+    }
+
+    pub fn poll_submit(
+        self: Pin<&mut Self>,
+        ctx: &mut Context<'_>,
+        head: u32,
+    ) -> Poll<()> {
+        let (iouring, state) = self.split_pinned();
+        match iouring.poll_submit(ctx, head) {
+            Poll::Ready(()) => {
+                match std::mem::replace(state, State::Lost) {
+                    State::Prepared(_, completion) => {
+                        *state = State::Submitted(completion);
+                    },
+                    _ => unreachable!(),
+                }
+                Poll::Ready(())
+            },
+            Poll::Pending => Poll::Pending,
+        }
+    }
+
+    pub fn poll_complete(
+        self: Pin<&mut Self>,
+        ctx: &mut Context<'_>,
+    ) -> Poll<io::Result<i32>> {
+        let (_, state) = self.split_pinned();
+        if let State::Submitted(completion) = std::mem::replace(state, State::Inert) {
+            match completion.check(ctx.waker()) {
+                Ok(result) => return Poll::Ready(result),
+                Err(completion) => {
+                    *state = State::Submitted(completion)
+                }
+            }
+        }
+
+        Poll::Pending
+    }
+}