Re: [RFC v2 09/13] io_uring: separate wq for ring polling

[Jens Axboe <axboe@kernel.dk>]

On 1/4/23 1:34?PM, Jens Axboe wrote:
> On 1/4/23 1:28?PM, Pavel Begunkov wrote:
>> On 1/4/23 18:08, Jens Axboe wrote:
>>> On 1/2/23 8:04?PM, Pavel Begunkov wrote:
>>>> Don't use ->cq_wait for ring polling but add a separate wait queue for
>>>> it. We need it for following patches.
>>>>
>>>> Signed-off-by: Pavel Begunkov <asml.silence@gmail.com>
>>>> ---
>>>>   include/linux/io_uring_types.h | 1 +
>>>>   io_uring/io_uring.c            | 3 ++-
>>>>   io_uring/io_uring.h            | 9 +++++++++
>>>>   3 files changed, 12 insertions(+), 1 deletion(-)
>>>>
>>>> diff --git a/include/linux/io_uring_types.h b/include/linux/io_uring_types.h
>>>> index dcd8a563ab52..cbcd3aaddd9d 100644
>>>> --- a/include/linux/io_uring_types.h
>>>> +++ b/include/linux/io_uring_types.h
>>>> @@ -286,6 +286,7 @@ struct io_ring_ctx {
>>>>           unsigned        cq_entries;
>>>>           struct io_ev_fd    __rcu    *io_ev_fd;
>>>>           struct wait_queue_head    cq_wait;
>>>> +        struct wait_queue_head    poll_wq;
>>>>           unsigned        cq_extra;
>>>>       } ____cacheline_aligned_in_smp;
>>>>   
>>>
>>> Should we move poll_wq somewhere else, more out of the way?
>>
>> If we care about polling perf and cache collisions with
>> cq_wait, yeah we can. In any case it's a good idea to at
>> least move it after cq_extra.
>>
>>> Would need to gate the check a flag or something.
>>
>> Not sure I follow
> 
> I guess I could've been a bit more verbose... If we consider poll on the
> io_uring rather uncommon, then moving the poll_wq outside of the hotter
> cq_wait cacheline(s) would make sense. Each wait_queue_head is more than
> a cacheline. Then we could have a flag in a spot that's hot anyway
> whether to check it or not, eg in that same section as cq_wait.
> 
> Looking at the layout right now, we're at 116 bytes for that section, or
> two cachelines with 12 bytes to spare. If we add poll_wq, then we'll be
> at 196 bytes, which is 4 bytes over the next cacheline. So it'd
> essentially double the size of that section. If we moved it outside of
> the aligned sections, then it'd pack better.

Just after writing this, I noticed that a spinlock took 64 bytes... In
other words, I have LOCKDEP enabled. The correct number is 24 bytes for
wait_queue_head which is obviously a lot more reasonable. It'd still
make that section one more cacheline since it's now at 60 bytes and
would grow to 84 bytes. But it's obviously not as big of a problem as I
had originally assumed.

-- 
Jens Axboe