Re: [Bug] io_uring_register_files_update broken

[Pavel Begunkov <asml.silence@gmail.com>]

On 7/1/21 3:01 PM, Daniele Salvatore Albano wrote:
> On Thu, 1 Jul 2021 at 00:28, Pavel Begunkov <asml.silence@gmail.com> wrote:
>>
>> On 6/30/21 10:14 PM, Victor Stewart wrote:
>>> i'm fairly confident there is something broken with
>>> io_uring_register_files_update,
>>> especially the offset parameter.
>>>
>>> when trying to update a single fd, and getting a successful result of
>>> 1, proceeding
>>> operations with IOSQE_FIXED_FILE fail with -9. but if i update all of
>>> the fds with
>>> then my recv operations succeed, but close still fails with -9.
>>>
>>> on Clear LInux 5.12.13-1050.native
>>
>> Thanks for letting know, I'll take a look
>>

[...]

> I sent an email a while ago to raise a question about a potential bug
> related to close.
> 
> Looks like the close doesn't support registered files (although I saw
> some code within a patch from Jens to fix it while I was
> investigating).

It have never been supported, don't remember it being discussed to
be implemented at any point, and the links below are about
unrelated things.

If you want to close a fixed file, you can use
IORING_OP_FILES_UPDATE with fd=-1 instead. Easy enough to
add it to close, but not sure I see the reason for that considering
existence of IORING_OP_FILES_UPDATE.

fwiw, it doesn't sound related to Victor's report, but still need
to look at it.

> Attached below
> 
> On Fri, 21 May 2021 at 21:28, Daniele Salvatore Albano
> <d.albano@gmail.com> wrote:
>>
>> Hi,
>>
>> Is there any specific reason for which io_close_prep returns EBADF if
>> using REQ_F_FIXED_FILE?
>>
>> I discovered my software was failing to close sockets when using fixed
>> files a while ago but I put it to the side, initially thinking it was
>> a bug I introduced in my code.
>> In recent days I picked it up again and after investigating it, it
>> looks like that, instead, that's the expected behaviour.
>>
>> From what I see, although the behaviour was slightly changed with a
>> couple of commits (ie. with
>> https://github.com/torvalds/linux/commit/cf3040ca55f2085b0a372a620ee2cb93ae19b686
>> ) the io_close_prep have had this behaviour from the very beginning
>> https://github.com/torvalds/linux/commit/b5dba59e0cf7e2cc4d3b3b1ac5fe81ddf21959eb
>> .
>>
>> @Jens during my researches I have also found
>> https://lkml.org/lkml/2020/5/7/1575 where there is a patch that
>> allows, at least from what it looks like at a first glance, fixed
>> files with io_close_prep but seems that the email thread died there.
>>
>> Shouldn't the close op match the behaviour of the other I/O related
>> ops when it comes to fds?
>>
>> If there aren't specific reasons, happy to look into it and write a patch.
>>
>>
>> Thanks,
>> Daniele

-- 
Pavel Begunkov