[LSF/MM/BPF TOPIC] Memory folios

[Matthew Wilcox]

On Mon, May 10, 2021 at 06:56:17PM +0100, Matthew Wilcox wrote:
> I don't know exactly how much will be left to discuss about supporting
> larger memory allocation units in the page cache by December.  In my
> ideal world, all the patches I've submitted so far are accepted, I
> persuade every filesystem maintainer to convert their own filesystem
> and struct page is nothing but a bad memory by December.  In reality,
> I'm just not that persuasive.
> 
> So, probably some kind of discussion will be worthwhile about
> converting the remaining filesystems to use folios, when it's worth
> having filesystems opt-in to multi-page folios, what we can do about
> buffer-head based filesystems, and so on.
> 
> Hopefully we aren't still discussing whether folios are a good idea
> or not by then.

I got an email from Hannes today asking about memory folios as they
pertain to the block layer, and I thought this would be a good chance
to talk about them.  If you're not familiar with the term "folio",
https://lore.kernel.org/lkml/20210505150628.111735-10-willy@infradead.org/
is not a bad introduction.

Thanks to the work done by Ming Lei in 2017, the block layer already
supports multipage bvecs, so to a first order of approximation, I don't
need anything from the block layer on down through the various storage
layers.  Which is why I haven't been talking to anyone in storage!

It might change (slightly) the contents of bios.  For example,
bvec[n]->bv_offset might now be larger than PAGE_SIZE.  Drivers should
handle this OK, but probably haven't been audited to make sure they do.
Mostly, it's simply that drivers will now see fewer, larger, segments
in their bios.  Once a filesystem supports multipage folios, we will
allocate order-N pages as part of readahead (and sufficiently large
writes).  Dirtiness is tracked on a per-folio basis (not per page),
so folios take trips around the LRU as a single unit and finally make
it to being written back as a single unit.

Drivers still need to cope with sub-folio-sized reads and writes.
O_DIRECT still exists and (eg) doing a sub-page, block-aligned write
will not necessarily cause readaround to happen.  Filesystems may read
and write their own metadata at whatever granularity and alignment they
see fit.  But the vast majority of pagecache I/O will be folio-sized
and folio-aligned.

I do have two small patches which make it easier for the one
filesystem that I've converted so far (iomap/xfs) to add folios to bios
and get folios back out of bios:

https://lore.kernel.org/lkml/20210505150628.111735-72-willy@infradead.org/
https://lore.kernel.org/lkml/20210505150628.111735-73-willy@infradead.org/

as well as a third patch that estimates how large a bio to allocate,
given the current folio that it's working on:
https://git.infradead.org/users/willy/pagecache.git/commitdiff/89541b126a59dc7319ad618767e2d880fcadd6c2

It would be possible to make other changes in future.  For example, if
we decide it'd be better, we could change bvecs from being (page, offset,
length) to (folio, offset, length).  I don't know that it's worth doing;
it would need to be evaluated on its merits.  Personally, I'd rather
see us move to a (phys_addr, length) pair, but I'm a little busy at the
moment.

Hannes has some fun ideas about using the folio work to support larger
sector sizes, and I think they're doable.