On Fri, Jan 27, 2023 at 08:21:46AM -0800, Eric Biggers wrote:
> On Fri, Jan 27, 2023 at 04:13:37PM +0000, Matthew Wilcox wrote:
> > On Thu, Jan 26, 2023 at 07:02:14PM -0800, Eric Biggers wrote:
> > > On Thu, Jan 26, 2023 at 08:23:46PM +0000, Matthew Wilcox (Oracle) wrote:
> > > > fscrypt_is_bounce_folio() is the equivalent of fscrypt_is_bounce_page()
> > > > and fscrypt_pagecache_folio() is the equivalent of fscrypt_pagecache_page().
> > > >
> > > > Signed-off-by: Matthew Wilcox (Oracle) <willy@xxxxxxxxxxxxx>
> > > > ---
> > > > include/linux/fscrypt.h | 21 +++++++++++++++++++++
> > > > 1 file changed, 21 insertions(+)
> > > >
> > > > diff --git a/include/linux/fscrypt.h b/include/linux/fscrypt.h
> > > > index 4f5f8a651213..c2c07d36fb3a 100644
> > > > --- a/include/linux/fscrypt.h
> > > > +++ b/include/linux/fscrypt.h
> > > > @@ -273,6 +273,16 @@ static inline struct page *fscrypt_pagecache_page(struct page *bounce_page)
> > > > return (struct page *)page_private(bounce_page);
> > > > }
> > > >
> > > > +static inline bool fscrypt_is_bounce_folio(struct folio *folio)
> > > > +{
> > > > + return folio->mapping == NULL;
> > > > +}
> > > > +
> > > > +static inline struct folio *fscrypt_pagecache_folio(struct folio *bounce_folio)
> > > > +{
> > > > + return bounce_folio->private;
> > > > +}
> > >
> > > ext4_bio_write_folio() is still doing:
> > >
> > > bounce_page = fscrypt_encrypt_pagecache_blocks(&folio->page, ...);
> > >
> > > Should it be creating a "bounce folio" instead, or is that not in the scope of
> > > this patchset?
> >
> > It's out of scope for _this_ patchset. I think it's a patchset that
> > could come either before or after, and is needed to support large folios
> > with ext4. The biggest problem with doing that conversion is that
> > bounce pages are allocated from a mempool which obviously only allocates
> > order-0 folios. I don't know what to do about that. Have a mempool
> > for each order of folio that the filesystem supports? Try to allocate
> > folios without a mempool and then split the folio if allocation fails?
> > Have a mempool containing PMD-order pages and split them ourselves if
> > we need to allocate from the mempool?
> >
> > Nothing's really standing out to me as the perfect answer. There are
> > probably other alternatives.
>
> Would it be possible to keep using bounce *pages* all the time, even when the
> pagecache contains large folios?
I _think_ so. Probably the best solution is to attempt to allocate an
order-N folio (with GFP_NOWAIT?) and then fall back to allocating 2^N
pages from the mempool. It'll require some surgery to ext4_finish_bio()
as well as ext4_bio_write_folio(), fscrypt_encrypt_pagecache_blocks()
and fscrypt_free_bounce_page(), but I think it's doable. I'll try
to whip something up.
