Re: [Qemu-devel] [Qemu-block] [PATCH v5 00/12] Add support for io_uring

From: Stefan Hajnoczi <stefanha@gmail.com>
To: Stefan Hajnoczi <stefanha@redhat.com>
Cc: Fam Zheng <fam@euphon.net>, Kevin Wolf <kwolf@redhat.com>,
	qemu block <qemu-block@nongnu.org>,
	Markus Armbruster <armbru@redhat.com>,
	qemu-devel <qemu-devel@nongnu.org>, Max Reitz <mreitz@redhat.com>,
	saket.sinha89@gmail.com, Paolo Bonzini <pbonzini@redhat.com>,
	Julia Suvorova <jusual@mail.ru>,
	Aarushi Mehta <mehta.aaru20@gmail.com>
Subject: Re: [Qemu-devel] [Qemu-block] [PATCH v5 00/12] Add support for io_uring
Date: Sun, 23 Jun 2019 15:03:58 +0100	[thread overview]
Message-ID: <CAJSP0QUwPoEi8-23-Qvfos5UtWi_7NaTdq7BW4T0SeZ47ET2UQ@mail.gmail.com> (raw)
In-Reply-To: <CAJSP0QWZKJ3KiAM2ZU8spYd13UWxW9xVhMAg_s-PxWcjCp9AdQ@mail.gmail.com>

On Sat, Jun 22, 2019 at 4:13 PM Stefan Hajnoczi <stefanha@gmail.com> wrote:
> On Tue, Jun 11, 2019 at 10:57 AM Stefan Hajnoczi <stefanha@redhat.com> wrote:
> > On Mon, Jun 10, 2019 at 07:18:53PM +0530, Aarushi Mehta wrote:
> > > This patch series adds support for the newly developed io_uring Linux AIO
> > > interface. Linux io_uring is faster than Linux's AIO asynchronous I/O code,
> > > offers efficient buffered asynchronous I/O support, the ability to do I/O
> > > without performing a system call via polled I/O, and other efficiency enhancements.
> > >
> > > Testing it requires a host kernel (5.1+) and the liburing library.
> > > Use the option -drive aio=io_uring to enable it.
> > >
> > > v5:
> > > - Adds completion polling
> > > - Extends qemu-io
> > > - Adds qemu-iotest
> >
> > Flush is not hooked up.  Please use the io_uring IOURING_OP_FSYNC that
> > you've already written and connect it to file-posix.c.
>
> IOURING_OP_FSYNC is in fact synchronous.  This means io_uring_enter()
> blocks until this operation completes.  This is not desirable since
> the AIO engine should not block the QEMU thread it's running from for
> a long time (e.g. network file system that is not responding).
>
> I think it's best *not* to use io_uring for fsync.  Instead we can
> continue to use the thread pool, just like Linux AIO.

Looking more closely, this is wrong.  Although fsync is synchronous,
io_uring takes care to bounce it to the workqueue when submitted via
io_uring_enter().  Therefore it appears asynchronous to userspace and
we can and should use io_uring for fsync.

Stefan