Re: [PATCH v2] io-wq: handle hashed writes in chains

[Jens Axboe <axboe@kernel.dk>]

On 3/22/20 2:25 PM, Pavel Begunkov wrote:
> On 22/03/2020 22:51, Jens Axboe wrote:
>> commit f1d96a8fcbbbb22d4fbc1d69eaaa678bbb0ff6e2
>> Author: Pavel Begunkov <asml.silence@gmail.com>
>> Date:   Fri Mar 13 22:29:14 2020 +0300
>>
>>     io_uring: NULL-deref for IOSQE_{ASYNC,DRAIN}
>>
>> which is what I ran into as well last week...
> 
> I picked it before testing
> 
>> The extra memory isn't a bit deal, it's very minor. My main concern
>> would be fairness, since we'd then be grabbing non-contig hashed chunks,
>> before we did not. May not be a concern as long as we ensure the
>> non-hasned (and differently hashed) work can proceed in parallel. For my
>> end, I deliberately added:
> 
> Don't think it's really a problem, all ordering/scheduling is up to
> users (i.e.  io_uring), and it can't infinitely postpone a work,
> because it's processing spliced requests without taking more, even if
> new ones hash to the same bit.

I don't disagree with you, just wanted to bring it up!

>> +	/* already have hashed work, let new worker get this */
>> +	if (ret) {
>> +		struct io_wqe_acct *acct;
>> +
>> +		/* get new worker for unhashed, if none now */
>> +		acct = io_work_get_acct(wqe, work);
>> +		if (!atomic_read(&acct->nr_running))
>> +			io_wqe_wake_worker(wqe, acct);
>> +		break;
>> +	}
>>
>> to try and improve that.
> 
> Is there performance problems with your patch without this chunk? I
> may see another problem with yours, I need to think it through.

No, and in fact it probably should be a separate thing, but I kind of
like your approach so not moving forward with mine. I do think it's
worth looking into separately, as there's no reason why we can't wake a
non-hashed worker if we're just doing hashed work from the existing
thread. If that thread is just doing copies and not blocking, the
unhashed (or next hashed) work is just sitting idle while it could be
running instead.

Hence I added that hunk, to kick a new worker to proceed in parallel.

-- 
Jens Axboe