Re: Discontiguous folios/pagesets

[Matthew Wilcox <willy@infradead.org>]

On Mon, Aug 30, 2021 at 11:28:18AM -0700, Darrick J. Wong wrote:
> On Sat, Aug 28, 2021 at 01:27:29PM -0600, Andreas Dilger wrote:
> > On Aug 28, 2021, at 1:04 PM, Matthew Wilcox <willy@infradead.org> wrote:
> > > 
> > > The current folio work is focused on permitting the VM to use
> > > physically contiguous chunks of memory.  Both Darrick and Johannes
> > > have pointed out the advantages of supporting logically-contiguous,
> > > physically-discontiguous chunks of memory.  Johannes wants to be able to
> > > use order-0 allocations to allocate larger folios, getting the benefit
> > > of managing the memory in larger chunks without requiring the memory
> > > allocator to be able to find contiguous chunks.  Darrick wants to support
> > > non-power-of-two block sizes.
> > 
> > What is the use case for non-power-of-two block sizes?  The main question
> > is whether that use case is important enough to add the complexity and
> > overhead in order to support it?
> 
> For copy-on-write to a XFS realtime volume where the allocation extent
> size (we support bigalloc too! :P) is not a power of two (e.g. you set
> up a 4 disk raid5 with 64k stripes, now the extent size is 192k).
> 
> Granted, I don't think folios handling 192k chunks is absolutely
> *required* for folios; the only hard requirement is that if any page in
> a 192k extent becomes dirty, the rest have to get written out all the
> same time, and the cow remap can only happen after the last page
> finishes writeback.

I /think/ "all pages get written out at the same time" is basically the
same thing as "support a non-power-of-two block size".

If we only have page A in the cache at the time it's going to be written
back, we have to read in pages B and C in order to calculate the parity P.
That will annoy writeback-because-we're-low-on-memory; I know we allow
a certain amount of allocation to happen in the writeback path, but
requiring 128kB to be allocated is a bit much.

So we have to allow page A being dirty to pin pages B and C in the cache.
I suppose that's possible; we could make (clean) pages B and C follow
page A on the LRU, so they're going to still be in RAM at the time that
page A is written back.  I don't fully understand how the LRU works,
but I assume it'd be a nightmare to ensure that A, B and C all move
around the system in the same way.  Much easier to ensure that ABC stay
linked together and all get written back at once.