Re: [PATCH] io_uring: run dependent links inline if possible

[Jens Axboe <axboe@kernel.dk>]

On 9/29/19 3:47 AM, Jackie Liu wrote:
> 
> 
>> 在 2019年9月29日，07:23，Jens Axboe <axboe@kernel.dk> 写道：
>>
>> Currently any dependent link is executed from a new workqueue context,
>> which means that we'll be doing a context switch per link in the chain.
>> If we are running the completion of the current request from our async
>> workqueue and find that the next request is a link, then run it directly
>> from the workqueue context instead of forcing another switch.
>>
>> This improves the performance of linked SQEs, and reduces the CPU
>> overhead.
>>
>> Signed-off-by: Jens Axboe <axboe@kernel.dk>
>>
>> ---
>>
>> 2-3x speedup doing read-write links, where the read often ends up
>> blocking. Tested with examples/link-cp.c
>>
>> diff --git a/fs/io_uring.c b/fs/io_uring.c
>> index aa8ac557493c..742d95563a54 100644
>> --- a/fs/io_uring.c
>> +++ b/fs/io_uring.c
>> @@ -667,7 +667,7 @@ static void __io_free_req(struct io_kiocb *req)
>> 	kmem_cache_free(req_cachep, req);
>> }
>>
>> -static void io_req_link_next(struct io_kiocb *req)
>> +struct io_kiocb *io_req_link_next(struct io_kiocb *req)
>> {
>> 	struct io_kiocb *nxt;
>>
>> @@ -686,9 +686,19 @@ static void io_req_link_next(struct io_kiocb *req)
>> 		}
>>
>> 		nxt->flags |= REQ_F_LINK_DONE;
>> +		/*
>> +		 * If we're in async work, we can continue processing this,
>> +		 * we can continue processing the chain in this context instead
>> +		 * of having to queue up new async work.
>> +		 */
>> +		if (current_work())
>> +			return nxt;
>> 		INIT_WORK(&nxt->work, io_sq_wq_submit_work);
>> 		io_queue_async_work(req->ctx, nxt);
>> +		nxt = NULL;
>> 	}
>> +
>> +	return nxt;
>> }
>>
>> /*
>> @@ -707,8 +717,10 @@ static void io_fail_links(struct io_kiocb *req)
>> 	}
>> }
>>
>> -static void io_free_req(struct io_kiocb *req)
>> +static struct io_kiocb *io_free_req(struct io_kiocb *req)
>> {
>> +	struct io_kiocb *nxt = NULL;
>> +
>> 	/*
>> 	 * If LINK is set, we have dependent requests in this chain. If we
>> 	 * didn't fail this request, queue the first one up, moving any other
>> @@ -719,16 +731,30 @@ static void io_free_req(struct io_kiocb *req)
>> 		if (req->flags & REQ_F_FAIL_LINK)
>> 			io_fail_links(req);
>> 		else
>> -			io_req_link_next(req);
>> +			nxt = io_req_link_next(req);
>> 	}
>>
>> 	__io_free_req(req);
>> +	return nxt;
>> }
>>
> 
> LGTM, Reviewed-by: Jackie Liu <liuyun01@kylinos.cn>
> 
> The function io_free_req has been used not only for free req, but also for the task
> of finding the next link entry. I think it is possible to change a name to avoid
> confusion, of course, only personal opinion.

That's a good point. I also changed how we handle the return of that, so there's
no confusion as to a caller getting a nxt request returned and not handling it.
See here:

http://git.kernel.dk/cgit/linux-block/commit/?h=for-5.5/io_uring&id=778fd7a24868b329ff8da2784fd8ced5e35af78c

I'll send out a v2 that's the above, and your naming suggestion.

-- 
Jens Axboe