On 2/17/20 9:12 AM, Jens Axboe wrote:
> On 2/17/20 5:09 AM, Peter Zijlstra wrote:
>> On Fri, Feb 14, 2020 at 01:44:32PM -0700, Jens Axboe wrote:
>>
>> I've not looked at git trees yet, but the below doesn't apply to
>> anything I have at hand.
>>
>> Anyway, I think I can still make sense of it -- just a rename or two
>> seems to be missing.
>>
>> A few notes on the below...
>
> Thanks for continuing to look at it, while we both try and make sense of
> it :-)
>
>>> diff --git a/include/linux/sched.h b/include/linux/sched.h
>>> index 04278493bf15..447b06c6bed0 100644
>>> --- a/include/linux/sched.h
>>> +++ b/include/linux/sched.h
>>> @@ -685,6 +685,11 @@ struct task_struct {
>>> #endif
>>> struct sched_dl_entity dl;
>>>
>>> +#ifdef CONFIG_IO_URING
>>> + struct list_head uring_work;
>>> + raw_spinlock_t uring_lock;
>>> +#endif
>>> +
>>
>> Could we pretty please use struct callback_head for this, just like
>> task_work() and RCU ? Look at task_work_add() for inspiration.
>
> Sure, so add a new one, sched_work, and have it get this sched-in or
> sched-out behavior.
>
> Only potential hitch I see there is related to ordering, which is more
> of a fairness thab correctness issue. I'm going to ignore that for now,
> and we can always revisit later.
OK, did the conversion, and it turned out pretty trivial, and reduces my
lines as well since I don't have to manage the list side. See here:
https://git.kernel.dk/cgit/linux-block/log/?h=io_uring-task-poll
Three small prep patches:
sched: move io-wq/workqueue worker sched in/out into helpers
kernel: abstract out task work helpers
sched: add a sched_work list
and then the two main patches before on top of that, one which uses this
just for the poll command, then the last one which enables any request
to use this path if it's pollable.
Let me know what you think of the direction, and I'll send out a
"proper" series for perusing.
--
Jens Axboe
