Re: [PATCHSET v2 0/6] Support selectable file descriptors

[Jens Axboe <axboe@kernel.dk>]

On 3/4/20 1:09 PM, Josh Triplett wrote:
> On Wed, Mar 04, 2020 at 01:00:05PM -0700, Jens Axboe wrote:
>> On 3/4/20 12:56 PM, Josh Triplett wrote:
>>> On Wed, Mar 04, 2020 at 12:10:08PM -0700, Jens Axboe wrote:
>>>> On 3/4/20 12:03 PM, Josh Triplett wrote:
>>>>> On Wed, Mar 04, 2020 at 11:00:10AM -0700, Jens Axboe wrote:
>>>>>> One of the fabled features with chains has long been the desire to
>>>>>> support things like:
>>>>>>
>>>>>> <open fileX><read from fileX><close fileX>
>>>>>>
>>>>>> in a single chain. This currently doesn't work, since the read/close
>>>>>> depends on what file descriptor we get on open.
>>>>>>
>>>>>> The original attempt at solving this provided a means to pass
>>>>>> descriptors between chains in a link, this version takes a different
>>>>>> route. Based on Josh's support for O_SPECIFIC_FD, we can instead control
>>>>>> what fd value we're going to get out of open (or accept). With that in
>>>>>> place, we don't need to do any magic to make this work. The above chain
>>>>>> then becomes:
>>>>>>
>>>>>> <open fileX with fd Y><read from fd Y><close fd Y>
>>>>>>
>>>>>> which is a lot more useful, and allows any sort of weird chains without
>>>>>> needing to nest "last open" file descriptors.
>>>>>>
>>>>>> Updated the test program to use this approach:
>>>>>>
>>>>>> https://git.kernel.dk/cgit/liburing/plain/test/orc.c?h=fd-select
>>>>>>
>>>>>> which forces the use of fd==89 for the open, and then uses that for the
>>>>>> read and close.
>>>>>>
>>>>>> Outside of this adaptation, fixed a few bugs and cleaned things up.
>>>>>
>>>>> I posted one comment about an issue in patch 6.
>>>>>
>>>>> Patches 2-5 look great; for those:
>>>>> Reviewed-by: Josh Triplett <josh@joshtriplett.org>
>>>>>
>>>>> Thanks for picking this up and running with it!
>>>>
>>>> Thanks for doing the prep work! I think it turned out that much better
>>>> for it.
>>>>
>>>> Are you going to post your series for general review? I just stole
>>>> your 1 patch that was needed for me.
>>>
>>> Since your patch series depends on mine, please feel free to run with
>>> the series. Would you mind adding my patch 1 and 3 at the end of your
>>> series? You need patch 1 to make this more usable for userspace, and
>>> patch 3 would allow for an OP_PIPE which I'd love to have.
>>
>> Let me add patch 1 to the top of the stack, for the pipe part that
>> probably should be taken in separately. But not a huge deal to me,
>> as long as we can get it reviewed.
> 
> That works for me; I don't mind if the pipe support goes in a bit later.
> And there are many other fd-producing syscalls that need support for
> userspace-selected FDs, including signalfd4, eventfd2, timerfd_create,
> epoll_create1, memfd_create, userfaultfd, and the pidfd family.

Right, at least on the io_uring front, adding IORING_OP_SOCKET and
providing support for SOCK_SPECIFIC_FD through that would be trivial and
a few lines of change. In general, we can more easily do all that
through io_uring, as we have room to shove in that extra 'fd'.

I've queued up patch 1 as well.

>>> Do you plan to submit this during the next merge window?
>>
>> Maybe? In terms of timing, I think we're well within the opportunity
>> to do so, at least.
> 
> I look forward to seeing it go in.

Me too :-)

-- 
Jens Axboe