Re: [pve-devel] [PATCH qemu] add patch fixing io_uring short read slow path

[Fabian Ebner <f.ebner@proxmox.com>]

Am 01.07.22 um 13:53 schrieb Thomas Lamprecht:
> How likely is this to trigger? Did you came across such events or is this a backport
> for the sake of being cautious?
> 
> Mostly asking because 7.0 is already on staging, so it'd be good to know if we should
> add another 6.0 release between (a bit of an hassle but possible due it not having yet
> seen any public repo), or if there isn't as much time pressure and we can just roll it
> out with the slower rollout of QEMU 7.0.

I didn't run into this recently. I do remember that I wondered about a
luring_resubmit_short_read trace for one VM while looking at the
LVM+io_uring issues a year ago, but IIRC that one behaved a bit
differently - got a weird GRUB error rather than the hang reported in
the forum, which I was able to reproduce with another VM then. But maybe
this was actually the root cause for these issues? Would have to find a
reproducer again to see if the patch fixes it.

v1 of the patch mentions:
> Normally recent versions of linux will not issue short reads,
> but apparently btrfs with O_DIRECT (cache=none) does.

but might not be the only situation. I tried a bit with a VM with an
image on BTRFS, but didn't manage to have the call show up in a trace.

So I can't really tell how likely. It can happen with BTRFS, it could
happen with LVM in the past (we don't allow io_uring+LVM currently). I'd
say since there is a big unknown, it might be worth a backport.