Files
cloud-hypervisor/block/src/raw_async.rs
Wei Liu 57e331db0e block: drop the ASYNC flag from io_uring entries
The ASYNC flag forces requests to go to worker threads. Worker threads
are expensive. Let the kernel decide what to do.

With this change, I no longer see an excessive amount of io_uring worker
threads.

Quote from the manual for io_uring_sqe_set_flags(3):

```
   IOSQE_ASYNC
          Normal operation for io_uring is to try and issue an sqe
          as non-blocking first, and if that fails, execute it in an
          async manner. To support more efficient overlapped
          operation of requests that the application knows/assumes
          will always (or most of the time) block, the application
          can ask for an sqe to be issued async from the start. Note
          that this flag immediately causes the SQE to be offloaded
          to an async helper thread with no initial non-blocking
          attempt.  This may be less efficient and should not be
          used liberally or without understanding the performance
          and efficiency tradeoffs.
```

Signed-off-by: Wei Liu <liuwe@microsoft.com>
2024-12-21 10:47:29 +00:00

168 lines
4.8 KiB
Rust

// Copyright © 2021 Intel Corporation
//
// SPDX-License-Identifier: Apache-2.0 AND BSD-3-Clause
use std::fs::File;
use std::io::{Error, Seek, SeekFrom};
use std::os::unix::io::{AsRawFd, RawFd};
use io_uring::{opcode, types, IoUring};
use vmm_sys_util::eventfd::EventFd;
use crate::async_io::{
AsyncIo, AsyncIoError, AsyncIoResult, DiskFile, DiskFileError, DiskFileResult,
};
use crate::DiskTopology;
pub struct RawFileDisk {
file: File,
}
impl RawFileDisk {
pub fn new(file: File) -> Self {
RawFileDisk { file }
}
}
impl DiskFile for RawFileDisk {
fn size(&mut self) -> DiskFileResult<u64> {
self.file
.seek(SeekFrom::End(0))
.map_err(DiskFileError::Size)
}
fn new_async_io(&self, ring_depth: u32) -> DiskFileResult<Box<dyn AsyncIo>> {
Ok(Box::new(
RawFileAsync::new(self.file.as_raw_fd(), ring_depth)
.map_err(DiskFileError::NewAsyncIo)?,
) as Box<dyn AsyncIo>)
}
fn topology(&mut self) -> DiskTopology {
if let Ok(topology) = DiskTopology::probe(&self.file) {
topology
} else {
warn!("Unable to get device topology. Using default topology");
DiskTopology::default()
}
}
}
pub struct RawFileAsync {
fd: RawFd,
io_uring: IoUring,
eventfd: EventFd,
}
impl RawFileAsync {
pub fn new(fd: RawFd, ring_depth: u32) -> std::io::Result<Self> {
let io_uring = IoUring::new(ring_depth)?;
let eventfd = EventFd::new(libc::EFD_NONBLOCK)?;
// Register the io_uring eventfd that will notify when something in
// the completion queue is ready.
io_uring.submitter().register_eventfd(eventfd.as_raw_fd())?;
Ok(RawFileAsync {
fd,
io_uring,
eventfd,
})
}
}
impl AsyncIo for RawFileAsync {
fn notifier(&self) -> &EventFd {
&self.eventfd
}
fn read_vectored(
&mut self,
offset: libc::off_t,
iovecs: &[libc::iovec],
user_data: u64,
) -> AsyncIoResult<()> {
let (submitter, mut sq, _) = self.io_uring.split();
// SAFETY: we know the file descriptor is valid and we
// relied on vm-memory to provide the buffer address.
unsafe {
sq.push(
&opcode::Readv::new(types::Fd(self.fd), iovecs.as_ptr(), iovecs.len() as u32)
.offset(offset.try_into().unwrap())
.build()
.user_data(user_data),
)
.map_err(|_| AsyncIoError::ReadVectored(Error::other("Submission queue is full")))?
};
// Update the submission queue and submit new operations to the
// io_uring instance.
sq.sync();
submitter.submit().map_err(AsyncIoError::ReadVectored)?;
Ok(())
}
fn write_vectored(
&mut self,
offset: libc::off_t,
iovecs: &[libc::iovec],
user_data: u64,
) -> AsyncIoResult<()> {
let (submitter, mut sq, _) = self.io_uring.split();
// SAFETY: we know the file descriptor is valid and we
// relied on vm-memory to provide the buffer address.
unsafe {
sq.push(
&opcode::Writev::new(types::Fd(self.fd), iovecs.as_ptr(), iovecs.len() as u32)
.offset(offset.try_into().unwrap())
.build()
.user_data(user_data),
)
.map_err(|_| AsyncIoError::WriteVectored(Error::other("Submission queue is full")))?
};
// Update the submission queue and submit new operations to the
// io_uring instance.
sq.sync();
submitter.submit().map_err(AsyncIoError::WriteVectored)?;
Ok(())
}
fn fsync(&mut self, user_data: Option<u64>) -> AsyncIoResult<()> {
if let Some(user_data) = user_data {
let (submitter, mut sq, _) = self.io_uring.split();
// SAFETY: we know the file descriptor is valid.
unsafe {
sq.push(
&opcode::Fsync::new(types::Fd(self.fd))
.build()
.user_data(user_data),
)
.map_err(|_| AsyncIoError::Fsync(Error::other("Submission queue is full")))?
};
// Update the submission queue and submit new operations to the
// io_uring instance.
sq.sync();
submitter.submit().map_err(AsyncIoError::Fsync)?;
} else {
// SAFETY: FFI call with a valid fd
unsafe { libc::fsync(self.fd) };
}
Ok(())
}
fn next_completed_request(&mut self) -> Option<(u64, i32)> {
self.io_uring
.completion()
.next()
.map(|entry| (entry.user_data(), entry.result()))
}
}