Re: [PATCH] ovl: disable IOCB_DIO_CALLER_COMP

[Jens Axboe <axboe@kernel.dk>]

On 9/25/23 3:18 AM, Amir Goldstein wrote:
> On Mon, Sep 25, 2023 at 9:21?AM Jens Axboe <axboe@kernel.dk> wrote:
>>
>> overlayfs copies the kiocb flags when it sets up a new kiocb to handle
>> a write, but it doesn't properly support dealing with the deferred
>> caller completions of the kiocb. This means it doesn't get the final
>> write completion value, and hence will complete the write with '0' as
>> the result.
>>
>> We could support the caller completions in overlayfs, but for now let's
>> just disable them in the generated write kiocb.
>>
>> Reported-by: Zorro Lang <zlang@redhat.com>
>> Link: https://lore.kernel.org/io-uring/20230924142754.ejwsjen5pvyc32l4@dell-per750-06-vm-08.rhts.eng.pek2.redhat.com/
>> Fixes: 8c052fb3002e ("iomap: support IOCB_DIO_CALLER_COMP")
>> Signed-off-by: Jens Axboe <axboe@kernel.dk>
>>
> 
> Thanks for fixing this Jens!
> If you or Christian want to send this fix to Linus, you have my ACK.

No problem - and thanks, maybe Christian can pick this one up? I
tentatively queued it up here just so I don't forget it:

https://git.kernel.dk/cgit/linux/log/?h=ovl-kiocb

> On the bright side, I am glad that you are aware of the overlayfs
> "kiocb_clone" use case, which delegates/forwards the io request to
> another file in another fs.
> 
> I have already posted an RFC [1] for moving this functionality to
> common vfs code. My main goal was to expose it to other filesystem
> (fuse), but a very desired side effect is that this functionality gets
> more vfs reviewer eyes and then the chances of catching a regression
> like this one during review of vfs changes hopefully increases.

Ah that's great! Yeah it's a bit hidden in there if you don't know about
it, and I did grep today when writing this patch to ensure we didn't
have any others like it. So I think we're good for now, at least.

> As for test coverage, I need to check why my tests did not catch
> this - I suspect fsx may not have been rebuilt with io_uring support,
> but not sure (not near workstation atm).

I'm guessing it's because you don't have liburing installed on the test
box, then fsx etc don't get built with io_uring support in xfstests.

> If you would like to add overlayfs to your test coverage, as Zorro
> explained, it is as simple as running ./check -overlay with your
> existing fstests config.
> ./check -overlay is a relatively faster test run because many of the
> tests do _notrun on overlayfs.
> I don't have to tell you that io_uring code will end up running on
> overlayfs in many container workloads, so it is not a niche setup.

Will add it to the mix! Thanks for the details.

-- 
Jens Axboe