Re: [PATCH 0/3] btrfs: Replace kmap() with kmap_local_page()

[Matthew Wilcox]

On Thu, Jun 02, 2022 at 09:22:15AM -0700, Ira Weiny wrote:
> On Wed, Jun 01, 2022 at 03:25:45PM +0200, David Sterba wrote:
> > On Tue, May 31, 2022 at 04:53:32PM +0200, Fabio M. De Francesco wrote:
> > > This is the first series of patches aimed towards the conversion of Btrfs
> > > filesystem from the use of kmap() to kmap_local_page().
> > 
> > We've already had patches converting kmaps and you're changing the last
> > ones, so this is could be the last series, with two exceptions.
> > 
> > 1) You've changed lzo.c and zlib. but the same kmap/kunmap pattern is
> >    used in zstd.c.
> 
> I checked out zstd.c and one of the issues there is the way that the input
> workspace is mapped page by page while iterating those pages.
> 
> This got me thinking about what Willy said at LSFmm concerning something like
> kmap_local_range().  Mapping more than 1 page at a time could save some
> unmap/remap of output pages required for kmap_local_page() ordering.

Umm ... Not entirely sure what I said, but it'd be really hard to kmap
multiple pages with the current PAE implementation.  I've steered away
from doing that for now, and kmap_local_folio() just guarantees the
page that the offset lands in is mapped.

I don't think the right answer is having a kmap_folio() that will map
the entire folio.  I'd be more tempted to add vmap_folio() for that.
My understanding is that PAE systems have more address space available
for vmap than they do for kmap.

> Unfortunately, I think the length of the input is probably to long in many
> cases.  And some remapping may still be required.
> 
> Cc: Willy
> 
> As an aside, Willy, I'm thinking that a kmap_local_range() should return a
> folio in some way.  Would you agree?

I imagine it taking a folio to describe the range that's being accessed.
But maybe it should be a phys_addr_t?  I tend to prefer phys_addr_t over
pfn + offset as it's more compact on 64-bit systems and the same on
32-bit systems.