Re: [PATCH 3/3] io_uring: move struct io_kiocb from task_struct to io_uring_task

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

On 11/4/24 16:16, Jens Axboe wrote:
> On 11/4/24 8:41 AM, Pavel Begunkov wrote:
>> On 11/3/24 17:49, Jens Axboe wrote:
>>> Rather than store the task_struct itself in struct io_kiocb, store
>>> the io_uring specific task_struct. The life times are the same in terms
>>> of io_uring, and this avoids doing some dereferences through the
>>> task_struct. For the hot path of putting local task references, we can
>>
>> Makes me wonder, is __io_submit_flush_completions() the only hot
>> place it tries to improve? It doesn't have to look into the task
>> there but on the other hand we need to do it that init.
>> If that's costly, for DEFER_TASKRUN we can get rid of per task
>> counting, the task is pinned together with the ctx, and the task
>> exit path can count the number of requests allocated.
>>
>> if (!(ctx->flags & DEFER_TASKRUN))
>>      io_task_get_ref();
>>
>> if (!(ctx->flags & DEFER_TASKRUN))
>>      io_task_put_ref();
>>
>> But can be further improved
> 
> Avoid task refs would surely be useful. For SINGLE_ISSUER, no?

Perhaps, but it doesn't imply single waiter / completer task.
IOPOLL would need to be checked and possibly there might be
races with io-wq. In general, all optimisations just got
shifted into DEFER_TASKRUN and SINGLE_ISSUER is not that useful
apart from carrying the semantics.

-- 
Pavel Begunkov