Re: [PATCH] aio: Convert ioctx_table to XArray

[Jens Axboe <axboe@kernel.dk>]

On 12/11/18 11:32 AM, Jens Axboe wrote:
> On 12/11/18 11:05 AM, Jens Axboe wrote:
>> On 12/11/18 11:02 AM, Jeff Moyer wrote:
>>> Matthew Wilcox <willy@infradead.org> writes:
>>>
>>>> On Tue, Dec 11, 2018 at 12:21:52PM -0500, Jeff Moyer wrote:
>>>>> I'm going to submit this version formally.  If you're interested in
>>>>> converting the ioctx_table to xarray, you can do that separately from a
>>>>> security fix.  I would include a performance analysis with that patch,
>>>>> though.  The idea of using a radix tree for the ioctx table was
>>>>> discarded due to performance reasons--see commit db446a08c23d5 ("aio:
>>>>> convert the ioctx list to table lookup v3").  I suspect using the xarray
>>>>> will perform similarly.
>>>>
>>>> There's a big difference between Octavian's patch and mine.  That patch
>>>> indexed into the radix tree by 'ctx_id' directly, which was pretty
>>>> much guaranteed to exhibit some close-to-worst-case behaviour from the
>>>> radix tree due to IDs being sparsely assigned.  My patch uses the ring
>>>> ID which _we_ assigned, and so is nicely behaved, being usually a very
>>>> small integer.
>>>
>>> OK, good to know.  I obviously didn't look too closely at the two.
>>>
>>>> What performance analysis would you find compelling?  Octavian's original
>>>> fio script:
>>>>
>>>>> rw=randrw; size=256k ;directory=/mnt/fio; ioengine=libaio; iodepth=1
>>>>> blocksize=1024; numjobs=512; thread; loops=100
>>>>>
>>>>> on an EXT2 filesystem mounted on top of a ramdisk
>>>>
>>>> or something else?
>>>
>>> I think the most common use case is a small number of ioctx-s, so I'd
>>> like to see that use case not regress (that should be easy, right?).
>>> Kent, what were the tests you were using when doing this work?  Jens,
>>> since you're doing performance work in this area now, are there any
>>> particular test cases you care about?
>>
>> I can give it a spin, ioctx lookup is in the fast path, and for "classic"
>> aio we do it twice for each IO...
> 
> Don't see any regressions. But if we're fiddling with it anyway, can't
> we do something smarter? Make the fast path just index a table, and put
> all the big hammers in setup/destroy. We're spending a non-substantial
> amount of time doing lookups, that's really no different before and
> after the patch.

Looks like it's the percpu ref get, in terms of "lookup" we already
look pretty good.

-- 
Jens Axboe