Re: [PATCH] io_uring: use __kernel_timespec in timeout ABI

[Jens Axboe]

On 10/1/19 9:49 AM, Arnd Bergmann wrote:
> On Tue, Oct 1, 2019 at 5:38 PM Jens Axboe <axboe@xxxxxxxxx> wrote:
>>
>> On 10/1/19 8:09 AM, Jens Axboe wrote:
>>> On 9/30/19 2:20 PM, Arnd Bergmann wrote:
>>>> All system calls use struct __kernel_timespec instead of the old struct
>>>> timespec, but this one was just added with the old-style ABI. Change it
>>>> now to enforce the use of __kernel_timespec, avoiding ABI confusion and
>>>> the need for compat handlers on 32-bit architectures.
>>>>
>>>> Any user space caller will have to use __kernel_timespec now, but this
>>>> is unambiguous and works for any C library regardless of the time_t
>>>> definition. A nicer way to specify the timeout would have been a less
>>>> ambiguous 64-bit nanosecond value, but I suppose it's too late now to
>>>> change that as this would impact both 32-bit and 64-bit users.
>>>
>>> Thanks for catching that, Arnd. Applied.
>>
>> On second thought - since there appears to be no good 64-bit timespec
>> available to userspace, the alternative here is including on in liburing.
> 
> What's wrong with using __kernel_timespec? Just the name?
> I suppose liburing could add a macro to give it a different name
> for its users.

Just that it seems I need to make it available through liburing on
systems that don't have it yet. Not a big deal, though.

>> That seems kinda crappy in terms of API, so why not just use a 64-bit nsec
>> value as you suggest? There's on released kernel with this feature yet, so
>> there's nothing stopping us from just changing the API to be based on
>> a single 64-bit nanosecond timeout.
> 
> Certainly fine with me.
> 
>> +       timeout = READ_ONCE(sqe->addr);
>>          hrtimer_init(&req->timeout.timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
>>          req->timeout.timer.function = io_timeout_fn;
>> -       hrtimer_start(&req->timeout.timer, timespec_to_ktime(ts),
>> +       hrtimer_start(&req->timeout.timer, ns_to_ktime(timeout),
> 
> It seems a little odd to use the 'addr' field as something that's not
> an address,
> and I'm not sure I understand the logic behind when you use a READ_ONCE()
> as opposed to simply accessing the sqe the way it is done a few lines
> earlier.
> 
> The time handling definitely looks good to me.

One thing that struck me about this approach - we then lose the ability to
differentiate between "don't want a timed timeout" with ts == NULL, vs
tv_sec and tv_nsec both being 0.

I think I'll stuck with that you had and just use __kernel_timespec in
liburing.

-- 
Jens Axboe