Re: [PATCH v2] io_uring: fix short read slow path

[Jens Axboe]

On 7/5/22 7:28 AM, Stefan Hajnoczi wrote:
> On Fri, Jul 01, 2022 at 07:52:31AM +0900, Dominique Martinet wrote:
>> Stefano Garzarella wrote on Thu, Jun 30, 2022 at 05:49:21PM +0200:
>>>> so when we ask for more we issue an extra short reads, making sure we go
>>>> through the two short reads path.
>>>> (Unfortunately I wasn't quite sure what to fiddle with to issue short
>>>> reads in the first place, I tried cutting one of the iovs short in
>>>> luring_do_submit() but I must not have been doing it properly as I ended
>>>> up with 0 return values which are handled by filling in with 0 (reads
>>>> after eof) and that didn't work well)
>>>
>>> Do you remember the kernel version where you first saw these problems?
>>
>> Since you're quoting my paragraph about testing two short reads, I've
>> never seen any that I know of; but there's also no reason these couldn't
>> happen.
>>
>> Single short reads have been happening for me with O_DIRECT (cache=none)
>> on btrfs for a while, but unfortunately I cannot remember which was the
>> first kernel I've seen this on -- I think rather than a kernel update it
>> was due to file manipulations that made the file eligible for short
>> reads in the first place (I started running deduplication on the backing
>> file)
>>
>> The older kernel I have installed right now is 5.16 and that can
>> reproduce it --  I'll give my laptop some work over the weekend to test
>> still maintained stable branches if that's useful.
> 
> Hi Dominique,
> Linux 5.16 contains commit 9d93a3f5a0c ("io_uring: punt short reads to
> async context"). The comment above QEMU's luring_resubmit_short_read()
> claims that short reads are a bug that was fixed by Linux commit
> 9d93a3f5a0c.
> 
> If the comment is inaccurate it needs to be fixed. Maybe short writes
> need to be handled too.
> 
> I have CCed Jens and the io_uring mailing list to clarify:
> 1. Are short IORING_OP_READV reads possible on files/block devices?
> 2. Are short IORING_OP_WRITEV writes possible on files/block devices?

In general we try very hard to avoid them, but if eg we get a short read
or write from blocking context (eg io-wq), then io_uring does return
that. There's really not much we can do here, it seems futile to retry
IO which was issued just like it would've been from a normal blocking
syscall yet it is still short.

-- 
Jens Axboe