Re: [PATCH 3/3] io_uring: move *queue_link_head() from common path

[Jens Axboe]

On 12/17/19 11:05 AM, Pavel Begunkov wrote:
> On 17/12/2019 21:01, Jens Axboe wrote:
>> On 12/17/19 10:52 AM, Pavel Begunkov wrote:
>>> On 17/12/2019 20:37, Jens Axboe wrote:
>>>> On 12/17/19 9:45 AM, Jens Axboe wrote:
>>>>> On 12/16/19 4:38 PM, Pavel Begunkov wrote:
>>>>>> On 17/12/2019 02:22, Pavel Begunkov wrote:
>>>>>>> Move io_queue_link_head() to links handling code in io_submit_sqe(),
>>>>>>> so it wouldn't need extra checks and would have better data locality.
>>>>>>>
>>>>>>> Signed-off-by: Pavel Begunkov <asml.silence@xxxxxxxxx>
>>>>>>> ---
>>>>>>>  fs/io_uring.c | 32 ++++++++++++++------------------
>>>>>>>  1 file changed, 14 insertions(+), 18 deletions(-)
>>>>>>>
>>>>>>> diff --git a/fs/io_uring.c b/fs/io_uring.c
>>>>>>> index bac9e711e38d..a880ed1409cb 100644
>>>>>>> --- a/fs/io_uring.c
>>>>>>> +++ b/fs/io_uring.c
>>>>>>> @@ -3373,13 +3373,15 @@ static bool io_submit_sqe(struct io_kiocb *req, struct io_submit_state *state,
>>>>>>>  			  struct io_kiocb **link)
>>>>>>>  {
>>>>>>>  	struct io_ring_ctx *ctx = req->ctx;
>>>>>>> +	unsigned int sqe_flags;
>>>>>>>  	int ret;
>>>>>>>  
>>>>>>> +	sqe_flags = READ_ONCE(req->sqe->flags);
>>>>>>>  	req->user_data = READ_ONCE(req->sqe->user_data);
>>>>>>>  	trace_io_uring_submit_sqe(ctx, req->user_data, true, req->in_async);
>>>>>>>  
>>>>>>>  	/* enforce forwards compatibility on users */
>>>>>>> -	if (unlikely(req->sqe->flags & ~SQE_VALID_FLAGS)) {
>>>>>>> +	if (unlikely(sqe_flags & ~SQE_VALID_FLAGS)) {
>>>>>>>  		ret = -EINVAL;
>>>>>>>  		goto err_req;
>>>>>>>  	}
>>>>>>> @@ -3402,10 +3404,10 @@ static bool io_submit_sqe(struct io_kiocb *req, struct io_submit_state *state,
>>>>>>>  	if (*link) {
>>>>>>>  		struct io_kiocb *head = *link;
>>>>>>>  
>>>>>>> -		if (req->sqe->flags & IOSQE_IO_DRAIN)
>>>>>>> +		if (sqe_flags & IOSQE_IO_DRAIN)
>>>>>>>  			head->flags |= REQ_F_DRAIN_LINK | REQ_F_IO_DRAIN;
>>>>>>>  
>>>>>>> -		if (req->sqe->flags & IOSQE_IO_HARDLINK)
>>>>>>> +		if (sqe_flags & IOSQE_IO_HARDLINK)
>>>>>>>  			req->flags |= REQ_F_HARDLINK;
>>>>>>>  
>>>>>>>  		if (io_alloc_async_ctx(req)) {
>>>>>>> @@ -3421,9 +3423,15 @@ static bool io_submit_sqe(struct io_kiocb *req, struct io_submit_state *state,
>>>>>>>  		}
>>>>>>>  		trace_io_uring_link(ctx, req, head);
>>>>>>>  		list_add_tail(&req->link_list, &head->link_list);
>>>>>>> -	} else if (req->sqe->flags & (IOSQE_IO_LINK|IOSQE_IO_HARDLINK)) {
>>>>>>> +
>>>>>>> +		/* last request of a link, enqueue the link */
>>>>>>> +		if (!(sqe_flags & IOSQE_IO_LINK)) {
>>>>>>
>>>>>> This looks suspicious (as well as in the current revision). Returning back
>>>>>> to my questions a few days ago can sqe->flags have IOSQE_IO_HARDLINK, but not
>>>>>> IOSQE_IO_LINK? I don't find any check.
>>>>>>
>>>>>> In other words, should it be as follows?
>>>>>> !(sqe_flags & (IOSQE_IO_LINK|IOSQE_IO_HARDLINK))
>>>>>
>>>>> Yeah, I think that should check for both. I'm fine with either approach
>>>>> in general:
>>>>>
>>>>> - IOSQE_IO_HARDLINK must have IOSQE_IO_LINK set
>>>>>
>>>>> or
>>>>>
>>>>> - IOSQE_IO_HARDLINK implies IOSQE_IO_LINK
>>>>>
>>>>> Seems like the former is easier to verify in terms of functionality,
>>>>> since we can rest easy if we check this early and -EINVAL if that isn't
>>>>> the case.
>>>>>
>>>>> What do you think?
>>>>
>>>> If you agree, want to send in a patch for that for 5.5? Then I can respin
>>>> for-5.6/io_uring on top of that, and we can apply your cleanups there.
>>>>
>>> Yes, that's the idea. Already got a patch, if you haven't done it yet.
>>
>> I haven't.
>>
>>> Just was thinking, whether to add a check for not setting both flags
>>> at the same moment in the "imply" case. Would give us 1 state in 2 bits
>>> for future use.
>>
>> Not sure I follow what you're saying here, can you elaborate?
>>
> 
> Sure
> 
> #define IOSQE_IO_LINK		(1U << 2)	/* links next sqe */
> #define IOSQE_IO_HARDLINK	(1U << 3)	/* like LINK, but stronger */
> 
> That's 2 consequent bits, so 4 states:
> 0,0 -> not a link
> 1,0 -> common link
> 0,1 -> hard link
> 1,1 -> reserved, space for another link-quirk type
> 
> But that would require additional check, i.e.
> 
> if (flags&(LINK|HARDLINK) == (LINK|HARDLINK)) ...

Ah, I see. In terms of usability, I think it makes more sense to have

IOSQE_LINK | IOSQE_HARDLINK

be the same as just IOSQE_LINK. It would be nice to retain that for
something else, but I think it'll be more confusing to users.

-- 
Jens Axboe