Re: [PATCH 5/5] io_uring: add IORING_OP_WAITID support

[Jens Axboe <axboe@kernel.dk>]

On 7/15/23 1:12 AM, Arnd Bergmann wrote:
> On Fri, Jul 14, 2023, at 22:14, Jens Axboe wrote:
>> On 7/14/23 12:33?PM, Arnd Bergmann wrote:
>>> On Fri, Jul 14, 2023, at 17:47, Christian Brauner wrote:
>>>> On Tue, Jul 11, 2023 at 04:18:13PM -0600, Jens Axboe wrote:
>>>>>>> Does this require argument conversion for compat tasks?
>>>>>>>
>>>>>>> Even without the rusage argument, I think the siginfo
>>>>>>> remains incompatible with 32-bit tasks, unfortunately.
>>>>>>
>>>>>> Hmm yes good point, if compat_siginfo and siginfo are different, then it
>>>>>> does need handling for that. Would be a trivial addition, I'll make that
>>>>>> change. Thanks Arnd!
>>>>>
>>>>> Should be fixed in the current version:
>>>>>
>>>>> https://git.kernel.dk/cgit/linux/commit/?h=io_uring-waitid&id=08f3dc9b7cedbd20c0f215f25c9a7814c6c601cc
>>>>
>>>> In kernel/signal.c in pidfd_send_signal() we have
>>>> copy_siginfo_from_user_any() it seems that a similar version
>>>> copy_siginfo_to_user_any() might be something to consider. We do have
>>>> copy_siginfo_to_user32() and copy_siginfo_to_user(). But I may lack
>>>> context why this wouldn't work here.
>>>
>>> We could add a copy_siginfo_to_user_any(), but I think open-coding
>>> it is easier here, since the in_compat_syscall() check does not
>>> work inside of the io_uring kernel thread, it has to be
>>> "if (req->ctx->compat)" in order to match the wordsize of the task
>>> that started the request.
>>
>> Yeah, unifying this stuff did cross my mind when adding another one.
>> Which I think could still be done, you'd just need to pass in a 'compat'
>> parameter similar to how it's done for iovec importing.
>>
>> But if it's ok with everybody I'd rather do that as a cleanup post this.
> 
> Sure, keeping that separate seem best.
> 
> Looking at what copy_siginfo_from_user_any() actually does, I don't
> even think it's worth adapting copy_siginfo_to_user_any() for io_uring,
> since it's already just a trivial wrapper, and adding another
> argument would add more complexity overall than it saves.

Yeah, took a look too this morning, and not sure there's much to reduce
here that would make it cleaner. I'm going to send out a v2 with this
unchanged, holler if people disagree.

-- 
Jens Axboe