Re: [PATCH v4] Use io_uring_register_ring_fd() to skip fd operations

[Stefan Hajnoczi]

On Fri, Apr 22, 2022 at 11:08:39AM +0100, Daniel P. Berrangé wrote:
> On Fri, Apr 22, 2022 at 11:00:47AM +0100, Fam Zheng wrote:
> > On 2022-04-22 09:52, Daniel P. Berrangé wrote:
> > > On Fri, Apr 22, 2022 at 09:34:28AM +0100, Fam Zheng wrote:
> > > > On 2022-04-22 00:36, Sam Li wrote:
> > > > > Linux recently added a new io_uring(7) optimization API that QEMU
> > > > > doesn't take advantage of yet. The liburing library that QEMU uses
> > > > > has added a corresponding new API calling io_uring_register_ring_fd().
> > > > > When this API is called after creating the ring, the io_uring_submit()
> > > > > library function passes a flag to the io_uring_enter(2) syscall
> > > > > allowing it to skip the ring file descriptor fdget()/fdput()
> > > > > operations. This saves some CPU cycles.
> > > > > 
> > > > > Signed-off-by: Sam Li <faithilikerun@gmail.com>
> > > > > ---
> > > > >  block/io_uring.c | 10 +++++++++-
> > > > >  1 file changed, 9 insertions(+), 1 deletion(-)
> > > > > 
> > > > > diff --git a/block/io_uring.c b/block/io_uring.c
> > > > > index 782afdb433..5247fb79e2 100644
> > > > > --- a/block/io_uring.c
> > > > > +++ b/block/io_uring.c
> > > > > @@ -435,8 +435,16 @@ LuringState *luring_init(Error **errp)
> > > > >      }
> > > > >  
> > > > >      ioq_init(&s->io_q);
> > > > > -    return s;
> > > > > +    if (io_uring_register_ring_fd(&s->ring) < 0) {
> > > > > +        /*
> > > > > +         * Only warn about this error: we will fallback to the 
> > > > > non-optimized
> > > > > +         * io_uring operations.
> > > > > +         */
> > > > > +        error_reportf_err(*errp,
> > > > > +                         "failed to register linux io_uring ring 
> > > > > file descriptor");
> > > > 
> > > > IIUC errp can be NULL, so let's not dereference it without checking. 
> > > > So, just
> > > > use error_report?
> > > 
> > > Plenty of people will be running kernels that lack the new feature,
> > > so this "failure" will be an expected scenario. We shouldn't be
> > > spamming the logs with any error or warning message. Assuming  QEMU
> > > remains fully functional, merely not as optimized, we should be
> > > totally silent.
> > 
> > Functionally, that's a very valid point. But performance wise, is it good to
> > have some visibility of this? Since people use io_uring instead of other
> > options almost certainly for performance, and here the issue does matter 
> > quite
> > a bit.
> 
> IMHO what you describe is largely a documentation issue, and/or something
> for OS vendors to worry about if they want to maximise their users'
> performance. As long as io_uring is fully functional we shouldn't print
> errors on every QEMU startup, as it leads to pointless bug reports/support
> escalations about something that is operating normally, wasting users and
> vendors' time.

Also, this is a minor optimization. It's nice to save a CPU cycles when
possible, but it's probably not significant enough that users would
bother to upgrade their kernel.

I think no warning is necessary.

Stefan

signature.asc
Description: PGP signature