On 12/18/2019 2:31 AM, Jens Axboe wrote:
> On 12/17/19 4:21 PM, Jens Axboe wrote:
>> On 12/17/19 3:28 PM, Pavel Begunkov wrote:
>>> percpu_ref_tryget() has its own overhead. Instead getting a reference
>>> for each request, grab a bunch once per io_submit_sqes().
>>>
>>> basic benchmark with submit and wait 128 non-linked nops showed ~5%
>>> performance gain. (7044 KIOPS vs 7423)
>>>
>>> Signed-off-by: Pavel Begunkov <asml.silence@xxxxxxxxx>
>>> ---
>>>
>>> For notice: it could be done without @extra_refs variable,
>>> but looked too tangled because of gotos.
>>>
>>>
>>> fs/io_uring.c | 11 ++++++++---
>>> 1 file changed, 8 insertions(+), 3 deletions(-)
>>>
>>> diff --git a/fs/io_uring.c b/fs/io_uring.c
>>> index cf4138f0e504..6c85dfc62224 100644
>>> --- a/fs/io_uring.c
>>> +++ b/fs/io_uring.c
>>> @@ -845,9 +845,6 @@ static struct io_kiocb *io_get_req(struct io_ring_ctx *ctx,
>>> gfp_t gfp = GFP_KERNEL | __GFP_NOWARN;
>>> struct io_kiocb *req;
>>>
>>> - if (!percpu_ref_tryget(&ctx->refs))
>>> - return NULL;
>>> -
>>> if (!state) {
>>> req = kmem_cache_alloc(req_cachep, gfp);
>>> if (unlikely(!req))
>>> @@ -3929,6 +3926,7 @@ static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
>>> struct io_submit_state state, *statep = NULL;
>>> struct io_kiocb *link = NULL;
>>> int i, submitted = 0;
>>> + unsigned int extra_refs;
>>> bool mm_fault = false;
>>>
>>> /* if we have a backlog and couldn't flush it all, return BUSY */
>>> @@ -3941,6 +3939,10 @@ static int io_submit_sqes(struct io_ring_ctx *ctx, unsigned int nr,
>>> statep = &state;
>>> }
>>>
>>> + if (!percpu_ref_tryget_many(&ctx->refs, nr))
>>> + return -EAGAIN;
>>> + extra_refs = nr;
>>> +
>>> for (i = 0; i < nr; i++) {
>>> struct io_kiocb *req = io_get_req(ctx, statep);
>>>
>
> This also needs to come before the submit_state_start().
>
Good catch, forgot to handle this. Thanks
> I forgot to mention that I really like the idea, no point in NOT batching
> the refs when we know exactly how many refs we need.
>
--
Pavel Begunkov
