Re: [RFC 0/8] Replacing the readpages a_op

From: Chris Mason <clm@fb.com>
To: Matthew Wilcox <willy@infradead.org>
Cc: Jens Axboe <axboe@kernel.dk>,
	"linux-xfs@vger.kernel.org" <linux-xfs@vger.kernel.org>,
	"linux-fsdevel@vger.kernel.org" <linux-fsdevel@vger.kernel.org>,
	"linux-mm@kvack.org" <linux-mm@kvack.org>,
	"jlayton@kernel.org" <jlayton@kernel.org>,
	"hch@infradead.org" <hch@infradead.org>
Subject: Re: [RFC 0/8] Replacing the readpages a_op
Date: Mon, 13 Jan 2020 22:34:16 +0000	[thread overview]
Message-ID: <F1FD3E8B-AC7E-48AB-99CD-E5D8E71851EE@fb.com> (raw)
In-Reply-To: <20200113222704.GC18216@bombadil.infradead.org>

On 13 Jan 2020, at 17:27, Matthew Wilcox wrote:

> On Mon, Jan 13, 2020 at 03:14:26PM -0700, Jens Axboe wrote:
>> On 1/13/20 3:10 PM, Matthew Wilcox wrote:
>>> On Mon, Jan 13, 2020 at 03:00:40PM -0700, Jens Axboe wrote:
>>>> On 1/13/20 2:58 PM, Matthew Wilcox wrote:
>>>>> On Mon, Jan 13, 2020 at 06:00:52PM +0000, Chris Mason wrote:
>>>>>> This is true, I didn't explain that part well ;)  Depending on
>>>>>> compression etc we might end up poking the xarray inside the 
>>>>>> actual IO
>>>>>> functions, but the main difference is that btrfs is building a 
>>>>>> single
>>>>>> bio.  You're moving the plug so you'll merge into single bio, but 
>>>>>> I'd
>>>>>> rather build 2MB bios than merge them.
>>>>>
>>>>> Why don't we store a bio pointer inside the plug?  You're 
>>>>> opencoding that,
>>>>> iomap is opencoding that, and I bet there's a dozen other places 
>>>>> where
>>>>> we pass a bio around.  Then blk_finish_plug can submit the bio.
>>>>
>>>> Plugs aren't necessarily a bio, they can be callbacks too.
>>>
>>> I'm thinking something as simple as this:
>>
>> It's a little odd imho, the plugging generally collect requests. 
>> Sounds
>> what you're looking for is some plug owner private data, which just
>> happens to be a bio in this case?
>>
>> Is this over repeated calls to some IO generating helper? Would it be
>> more efficient if that helper could generate the full bio in one go,
>> instead of piecemeal?
>
> The issue is around ->readpages.  Take a look at how iomap_readpages
> works, for example.  We're under a plug (taken in mm/readahead.c),
> but we still go through the rigamarole of keeping a pointer to the bio
> in ctx.bio and passing ctx around so that we don't end up with many
> fragments which have to be recombined into a single bio at the end.
>
> I think what I want is a bio I can reach from current, somehow.  And 
> the
> plug feels like a natural place to keep it because it's basically 
> saying
> "I want to do lots of little IOs and have them combined".  The fact 
> that
> the iomap code has a bio that it precombines fragments into suggests 
> to
> me that the existing antifragmentation code in the plugging mechanism
> isn't good enough.  So let's make it better by storing a bio in the 
> plug
> and then we can get rid of the bio in the iomap code.

Both btrfs and xfs do this, we have a bio that we pass around and build 
and submit.  We both also do some gymnastics in writepages to avoid 
waiting for the bios we've been building to finish while we're building 
them.

I love the idea of the plug api having a way to hold that for us, but 
sometimes we really are building the bios, and we don't want the plug to 
let it go if we happen to schedule.

-chris