On 03/26, Matthew Wilcox wrote:
> On Fri, Mar 14, 2025 at 09:38:15PM +0000, Jaegeuk Kim wrote:
> > On 03/14, Matthew Wilcox wrote:
> > > Unfortunately, I thnk I have to abandon this effort. It's only going
> > > to make supporting large folios harder (ie there would then need to be
> > > an equivalently disruptive series adding support for large folios).
> > >
> > > The fundamental problem is that f2fs has no concept of block size !=
> > > PAGE_SIZE. So if you create a filesystem on a 4kB PAGE_SIZE kernel,
> > > you can't mount it on a 16kB PAGE_SIZE kernel. An example:
> > >
> > > int f2fs_recover_inline_xattr(struct inode *inode, struct page *page)
> > > {
> > > struct f2fs_inode *ri;
> > > ipage = f2fs_get_node_page(F2FS_I_SB(inode), inode->i_ino);
> > > ri = F2FS_INODE(page);
> > >
> > > so an inode number is an index into the filesystem in PAGE_SIZE units,
> > > not in filesystem block size units. Fixing this is a major effort, and
> > > I lack the confidence in my abilities to do it without breaking anything.
> > >
> > > As an outline of what needs to happen, I think that rather than passing
> > > around so many struct page pointers, we should be passing around either
> > > folio + offset, or we should be passing around struct f2fs_inode pointers.
> > > My preference is for the latter. We can always convert back to the
> > > folio containing the inode if we need to (eg to mark it dirty) and it
> > > adds some typesafety by ensuring that we're passing around pointers that
> > > we believe belong to an inode and not, say, a struct page which happens
> > > to contain a directory entry.
> > >
> > > This is a monster task, I think. I'm going to have to disable f2fs
> > > from testing with split page/folio. This is going to be a big problem
> > > for Android.
> >
> > I see. fyi; in Android, I'm thinking to run 16KB page kernel with 16KB format
> > natively to keep block_size = PAGE_SIZE. Wasn't large folio to support a set
> > of pages while keeping block_size = PAGE_SIZE?
>
> Oh, I think I do see a possible argument for continuing this work.
>
> If we have an f2fs filesystem with a 16kB block size, we can use order-0
> folios with a 16kB PAGE_SIZE kernel, and if we want to mount it on a
> kernel with a 4kB PAGE_SIZE kernel, then we can use order-2 folios to
> do that.
>
> Is that a useful improvement to f2fs? It's not really the intent of
> large folios; it's supposed to be used to support arbitrary order folios.
> But we have all the pieces in place such that we could tell the page
> cache min-order = max-order = 2.
It may be helpful in case where someone wants to try 4KB page kernel back,
after Android ships 16KB page/block products. Does it require a big surgery?
