Re: io_uring_prep_openat_direct() and link/drain

[Jens Axboe <axboe@kernel.dk>]

On 3/30/22 6:43 AM, Miklos Szeredi wrote:
> On Wed, 30 Mar 2022 at 14:35, Jens Axboe <axboe@kernel.dk> wrote:
>>
>> On 3/30/22 2:18 AM, Miklos Szeredi wrote:
>>> On Tue, 29 Mar 2022 at 22:03, Jens Axboe <axboe@kernel.dk> wrote:
>>>
>>>> Can you try and repull the branch? I rebased the top two, so reset to
>>>> v5.17 first and then pull it.
>>>
>>> This one works in all cases, except if there's a link flag on the
>>> read.  In that case the read itself will succeed but following
>>> requests on the link chain will fail with -ECANCELED.
>>
>> That sounds like you're asking for more data than what is being read,
>> in which case the link should correctly break and cancel the rest.
>> That's expected behavior, a short read is not a successful read in that
>> regard.
> 
> Sorry, I don't get it.  Does the link flag has other meaning than
> "wait until this request completes before starting the next"?

IOSQE_IO_LINK means "wait until this request SUCCESSFULLY completes
before starting the next one". You can't reliably use a link from an eg
read, if you don't know how much data it read. If you don't care about
success, then use IOSQE_IO_HARDLINK. That is semantically like a link,
but it doesn't break on an unexpected result.

If you're using LINK from the read _and_ the read returns less than you
asked for, IOSQE_IO_LINK will break the chain as that is an unexpected
result.

-- 
Jens Axboe