Re: [PATCH 4/4] io_uring: mark opcodes that always need io-wq punt

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

On 4/25/23 16:07, Ming Lei wrote:
> On Tue, Apr 25, 2023 at 08:50:33AM -0600, Jens Axboe wrote:
>> On 4/25/23 8:42?AM, Ming Lei wrote:
>>> On Tue, Apr 25, 2023 at 07:31:10AM -0600, Jens Axboe wrote:
>>>> On 4/24/23 8:50?PM, Ming Lei wrote:
>>>>> On Mon, Apr 24, 2023 at 08:18:02PM -0600, Jens Axboe wrote:
>>>>>> On 4/24/23 8:13?PM, Ming Lei wrote:
>>>>>>> On Mon, Apr 24, 2023 at 08:08:09PM -0600, Jens Axboe wrote:
>>>>>>>> On 4/24/23 6:57?PM, Ming Lei wrote:
>>>>>>>>> On Mon, Apr 24, 2023 at 09:24:33AM -0600, Jens Axboe wrote:
>>>>>>>>>> On 4/24/23 1:30?AM, Ming Lei wrote:
>>>>>>>>>>> On Thu, Apr 20, 2023 at 12:31:35PM -0600, Jens Axboe wrote:
>>>>>>>>>>>> Add an opdef bit for them, and set it for the opcodes where we always
>>>>>>>>>>>> need io-wq punt. With that done, exclude them from the file_can_poll()
>>>>>>>>>>>> check in terms of whether or not we need to punt them if any of the
>>>>>>>>>>>> NO_OFFLOAD flags are set.
>>>>>>>>>>>>
>>>>>>>>>>>> Signed-off-by: Jens Axboe <axboe@kernel.dk>
>>>>>>>>>>>> ---
>>>>>>>>>>>>   io_uring/io_uring.c |  2 +-
>>>>>>>>>>>>   io_uring/opdef.c    | 22 ++++++++++++++++++++--
>>>>>>>>>>>>   io_uring/opdef.h    |  2 ++
>>>>>>>>>>>>   3 files changed, 23 insertions(+), 3 deletions(-)
>>>>>>>>>>>>
>>>>>>>>>>>> diff --git a/io_uring/io_uring.c b/io_uring/io_uring.c
>>>>>>>>>>>> index fee3e461e149..420cfd35ebc6 100644
>>>>>>>>>>>> --- a/io_uring/io_uring.c
>>>>>>>>>>>> +++ b/io_uring/io_uring.c
>>>>>>>>>>>> @@ -1948,7 +1948,7 @@ static int io_issue_sqe(struct io_kiocb *req, unsigned int issue_flags)
>>>>>>>>>>>>   		return -EBADF;
>>>>>>>>>>>>   
>>>>>>>>>>>>   	if (issue_flags & IO_URING_F_NO_OFFLOAD &&
>>>>>>>>>>>> -	    (!req->file || !file_can_poll(req->file)))
>>>>>>>>>>>> +	    (!req->file || !file_can_poll(req->file) || def->always_iowq))
>>>>>>>>>>>>   		issue_flags &= ~IO_URING_F_NONBLOCK;
>>>>>>>>>>>
>>>>>>>>>>> I guess the check should be !def->always_iowq?
>>>>>>>>>>
>>>>>>>>>> How so? Nobody that takes pollable files should/is setting
>>>>>>>>>> ->always_iowq. If we can poll the file, we should not force inline
>>>>>>>>>> submission. Basically the ones setting ->always_iowq always do -EAGAIN
>>>>>>>>>> returns if nonblock == true.
>>>>>>>>>
>>>>>>>>> I meant IO_URING_F_NONBLOCK is cleared here for  ->always_iowq, and
>>>>>>>>> these OPs won't return -EAGAIN, then run in the current task context
>>>>>>>>> directly.
>>>>>>>>
>>>>>>>> Right, of IO_URING_F_NO_OFFLOAD is set, which is entirely the point of
>>>>>>>> it :-)
>>>>>>>
>>>>>>> But ->always_iowq isn't actually _always_ since fallocate/fsync/... are
>>>>>>> not punted to iowq in case of IO_URING_F_NO_OFFLOAD, looks the naming of
>>>>>>> ->always_iowq is a bit confusing?
>>>>>>
>>>>>> Yeah naming isn't that great, I can see how that's bit confusing. I'll
>>>>>> be happy to take suggestions on what would make it clearer.
>>>>>
>>>>> Except for the naming, I am also wondering why these ->always_iowq OPs
>>>>> aren't punted to iowq in case of IO_URING_F_NO_OFFLOAD, given it
>>>>> shouldn't improve performance by doing so because these OPs are supposed
>>>>> to be slow and always slept, not like others(buffered writes, ...),
>>>>> can you provide one hint about not offloading these OPs? Or is it just that
>>>>> NO_OFFLOAD needs to not offload every OPs?
>>>>
>>>> The whole point of NO_OFFLOAD is that items that would normally be
>>>> passed to io-wq are just run inline. This provides a way to reap the
>>>> benefits of batched submissions and syscall reductions. Some opcodes
>>>> will just never be async, and io-wq offloads are not very fast. Some of
>>>
>>> Yeah, seems io-wq is much slower than inline issue, maybe it needs
>>> to be looked into, and it is easy to run into io-wq for IOSQE_IO_LINK.
>>
>> Indeed, depending on what is being linked, you may see io-wq activity
>> which is not ideal.
> 
> That is why I prefer to fused command for ublk zero copy, because the
> registering buffer approach suggested by Pavel and Ziyang has to link
> register buffer OP with the actual IO OP, and it is observed that
> IOPS drops to 1/2 in 4k random io test with registered buffer approach.

What's good about it is that you can use linked requests with it
but you _don't have to_.

Curiously, I just recently compared submitting 8 two-request links
(16 reqs in total) vs submit(8)+submit(8), all that in a loop.
The latter was faster. It wasn't a clean experiment, but shows
that links are not super fast and would be nice to get them better.

For the register buf approach, I tried it out, looked good to me.
It outperforms splice requests (with a hack that removes force
iowq execution) by 5-10% with synthetic benchmark. Works better than
splice(2) for QD>=2. Let me send it out, perhaps today, so we can
figure out how it compares against ublk/fused and see the margin is.

-- 
Pavel Begunkov