Re: [PATCH v3] io_uring: switch cancel_hash to use per entry spinlock

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

On 6/11/22 05:34, Hao Xu wrote:
> On 6/10/22 18:21, Pavel Begunkov wrote:
>> On 6/8/22 12:12, Hao Xu wrote:
>>> From: Hao Xu <howeyxu@tencent.com>
>>>
>>> Add a new io_hash_bucket structure so that each bucket in cancel_hash
>>> has separate spinlock. Use per entry lock for cancel_hash, this removes
>>> some completion lock invocation and remove contension between different
>>> cancel_hash entries.
>>>
>>> Signed-off-by: Hao Xu <howeyxu@tencent.com>
>>> ---
>>>
>>> v1->v2:
>>>   - Add per entry lock for poll/apoll task work code which was missed
>>>     in v1
>>>   - add an member in io_kiocb to track req's indice in cancel_hash
>>>
>>> v2->v3:
>>>   - make struct io_hash_bucket align with cacheline to avoid cacheline
>>>     false sharing.
>>>   - re-calculate hash value when deleting an entry from cancel_hash.
>>>     (cannot leverage struct io_poll to store the indice since it's
>>>      already 64 Bytes)
>>>
>>>   io_uring/cancel.c         | 14 +++++++--
>>>   io_uring/cancel.h         |  6 ++++
>>>   io_uring/fdinfo.c         |  9 ++++--
>>>   io_uring/io_uring.c       |  8 +++--
>>>   io_uring/io_uring_types.h |  2 +-
>>>   io_uring/poll.c           | 64 +++++++++++++++++++++------------------
>>>   6 files changed, 65 insertions(+), 38 deletions(-)
>>>
>>> diff --git a/io_uring/cancel.c b/io_uring/cancel.c
>>> index 83cceb52d82d..bced5d6b9294 100644
>>> --- a/io_uring/cancel.c
>>> +++ b/io_uring/cancel.c
>>> @@ -93,14 +93,14 @@ int io_try_cancel(struct io_kiocb *req, struct io_cancel_data *cd)
>>>       if (!ret)
>>>           return 0;
>>> -    spin_lock(&ctx->completion_lock);
>>>       ret = io_poll_cancel(ctx, cd);
>>>       if (ret != -ENOENT)
>>>           goto out;
>>> +    spin_lock(&ctx->completion_lock);
>>>       if (!(cd->flags & IORING_ASYNC_CANCEL_FD))
>>>           ret = io_timeout_cancel(ctx, cd);
>>> -out:
>>>       spin_unlock(&ctx->completion_lock);
>>> +out:
>>>       return ret;
>>>   }
>>> @@ -192,3 +192,13 @@ int io_async_cancel(struct io_kiocb *req, unsigned int issue_flags)
>>>       io_req_set_res(req, ret, 0);
>>>       return IOU_OK;
>>>   }
>>> +
>>> +inline void init_hash_table(struct io_hash_bucket *hash_table, unsigned size)
>>
>> Not inline, it can break builds
>>
> 
> What do you mean? It's compiled well.

I might be wrong, but IIRC there could be linking issues. Anyway,
it's used only from another obj and the function is cold, so no reason
to have it inline.

-- 
Pavel Begunkov