Re: [RFC 0/2] 3 cacheline io_kiocb

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

 On 25/07/2020 22:40, Jens Axboe wrote:
> On 7/25/20 12:24 PM, Pavel Begunkov wrote:
>> On 25/07/2020 18:45, Jens Axboe wrote:
>>> On 7/25/20 2:31 AM, Pavel Begunkov wrote:
>>>> That's not final for a several reasons, but good enough for discussion.
>>>> That brings io_kiocb down to 192B. I didn't try to benchmark it
>>>> properly, but quick nop test gave +5% throughput increase.
>>>> 7531 vs 7910 KIOPS with fio/t/io_uring
>>>>
>>>> The whole situation is obviously a bunch of tradeoffs. For instance,
>>>> instead of shrinking it, we can inline apoll to speed apoll path.
>>>>
>>>> [2/2] just for a reference, I'm thinking about other ways to shrink it.
>>>> e.g. ->link_list can be a single-linked list with linked tiemouts
>>>> storing a back-reference. This can turn out to be better, because
>>>> that would move ->fixed_file_refs to the 2nd cacheline, so we won't
>>>> ever touch 3rd cacheline in the submission path.
>>>> Any other ideas?
>>>
>>> Nothing noticeable for me, still about the same performance. But
>>> generally speaking, I don't necessarily think we need to go all in on
>>> making this as tiny as possible. It's much more important to chase the
>>> items where we only use 2 cachelines for the hot path, and then we have
>>> the extra space in there already for the semi hot paths like poll driven
>>> retry. Yes, we're still allocating from a pool that has slightly larger
>>> objects, but that doesn't really matter _that_ much. Avoiding an extra
>>> kmalloc+kfree for the semi hot paths are a bigger deal than making
>>> io_kiocb smaller and smaller.
>>>
>>> That said, for no-brainer changes, we absolutely should make it smaller.
>>> I just don't want to jump through convoluted hoops to get there.
>>
>> Agree, but that's not the end goal. The first point is to kill the union,
>> but it already has enough space for that.
> 
> Right
> 
>> The second is to see, whether we can use the space in a better way. From
>> the high level perspective ->apoll and ->work are alike and both serve to
>> provide asynchronous paths, hence the idea to swap them naturally comes to
>> mind.
> 
> Totally agree, which is why the union of those kind of makes sense.
> We're definitely NOT using them at the same time, but the fact that we
> had various mm/creds/whatnot in the work_struct made that a bit iffy.

Thinking of it, if combined with work de-init as you proposed before, it's
probably possible to make a layout similar to the one below

struct io_kiocb {
	...
	struct hlist_node	hash_node;
	struct callback_head	task_work;	
	union {
		struct io_wq_work	work;
		struct async_poll	apoll;
	};
};

Saves ->apoll kmalloc(), and the actual work de-init would be negligibly
rare. Worth to try


> 
>> TBH, I don't think it'd do much, because init of ->io would probably
>> hide any benefit.
> 
> There should be no ->io init/alloc for this test case.

I mean, before getting into io_arm_poll_handler(), it should get -EAGAIN
in io_{read,write}() and initialise ->io in io_setup_async_rw(), at least
for READV, WRITEV.

-- 
Pavel Begunkov