Re: [PATCH 2/2] io_uring: switch cancel_hash to use per list spinlock

[Hao Xu <haoxu.linux@icloud.com>]

On 5/30/22 08:18, Pavel Begunkov wrote:
> On 5/30/22 00:34, Jens Axboe wrote:
>> On 5/29/22 4:50 PM, Pavel Begunkov wrote:
>>> On 5/29/22 19:40, Jens Axboe wrote:
>>>> On 5/29/22 12:07 PM, Hao Xu wrote:
>>>>> On 5/30/22 00:25, Jens Axboe wrote:
>>>>>> On 5/29/22 10:20 AM, Hao Xu wrote:
>>>>>>> From: Hao Xu <howeyxu@tencent.com>
>>>>>>>
>>>>>>> From: Hao Xu <howeyxu@tencent.com>
>>>>>>>
>>>>>>> Use per list lock for cancel_hash, this removes some completion lock
>>>>>>> invocation and remove contension between different cancel_hash 
>>>>>>> entries
>>>>>>
>>>>>> Interesting, do you have any numbers on this?
>>>>>
>>>>> Just Theoretically for now, I'll do some tests tomorrow. This is
>>>>> actually RFC, forgot to change the subject.
>>>>>
>>>>>>
>>>>>> Also, I'd make a hash bucket struct:
>>>>>>
>>>>>> struct io_hash_bucket {
>>>>>>       spinlock_t lock;
>>>>>>       struct hlist_head list;
>>>>>> };
>>>>>>
>>>>>> rather than two separate structs, that'll have nicer memory 
>>>>>> locality too
>>>>>> and should further improve it. Could be done as a prep patch with the
>>>>>> old locking in place, making the end patch doing the per-bucket lock
>>>>>> simpler as well.
>>>>>
>>>>> Sure, if the test number make sense, I'll send v2. I'll test the
>>>>> hlist_bl list as well(the comment of it says it is much slower than
>>>>> normal spin_lock, but we may not care the efficiency of poll
>>>>> cancellation very much?).
>>>>
>>>> I don't think the bit spinlocks are going to be useful, we should
>>>> stick with a spinlock for this. They are indeed slower and generally 
>>>> not
>>>> used for that reason. For a use case where you need a ton of locks and
>>>> saving the 4 bytes for a spinlock would make sense (or maybe not
>>>> changing some struct?), maybe they have a purpose. But not for this.
>>>
>>> We can put the cancel hashes under uring_lock and completely kill
>>> the hash spinlocking (2 lock/unlock pairs per single-shot). The code
>>> below won't even compile and missing cancellation bits, I'll pick it
>>> up in a week.
>>>
>>> Even better would be to have two hash tables, and auto-magically apply
>>> the feature to SINGLE_SUBMITTER, SQPOLL (both will have uring_lock held)
>>> and apoll (need uring_lock after anyway).
>>
>> My hope was that it'd take us closer to being able to use more granular
>> locking for hashing in general. I don't care too much about the
>> cancelation, but the normal hash locking would be useful to do.
>>
>> However, for cancelations, under uring_lock would indeed be preferable
>> to doing per-bucket locks there. Guess I'll wait and see what your final
>> patch looks like, not sure why it'd be a ctx conditional?
> 
> It replaces 2 spin lock/unlock with one io_tw_lock() in the completion
> path, which is done once per tw batch and grabbed anyway if
> there is no contention (see handle_tw_list()).
> 
> It could be unconditional, but I'd say for those 3 cases we have
> non-existing chance to regress perf/latency, but I can think of
> some cases where it might screw latencies, all share io_uring
> b/w threads.
> 
> Should benefit the cancellation path as well, but I don't care
> about it as well.
> 
>> What about io_poll_remove_all()?
> 
> As mentioned, it's not handled in the diff, but easily doable,
> it should just traverse both hash tables.
> 

Two hash tables looks good to me. If I don't get you wrong, one table
under uring_lock, the other one for normal handling(like per bucket
locking)?