Re: [PATCH v10 10/10] xfs: enable block size larger than page size support

["Darrick J. Wong" <djwong@kernel.org>]

On Tue, Jul 16, 2024 at 06:46:40PM +0100, Matthew Wilcox wrote:
> On Tue, Jul 16, 2024 at 10:40:16AM -0700, Darrick J. Wong wrote:
> > On Tue, Jul 16, 2024 at 04:29:05PM +0100, Matthew Wilcox wrote:
> > > On Mon, Jul 15, 2024 at 11:44:57AM +0200, Pankaj Raghav (Samsung) wrote:
> > > > +++ b/fs/xfs/xfs_super.c
> > > > @@ -1638,16 +1638,30 @@ xfs_fs_fill_super(
> > > >  		goto out_free_sb;
> > > >  	}
> > > >  
> > > > -	/*
> > > > -	 * Until this is fixed only page-sized or smaller data blocks work.
> > > > -	 */
> > > >  	if (mp->m_sb.sb_blocksize > PAGE_SIZE) {
> > > > -		xfs_warn(mp,
> > > > -		"File system with blocksize %d bytes. "
> > > > -		"Only pagesize (%ld) or less will currently work.",
> > > > +		size_t max_folio_size = mapping_max_folio_size_supported();
> > > > +
> > > > +		if (!xfs_has_crc(mp)) {
> > > > +			xfs_warn(mp,
> > > > +"V4 Filesystem with blocksize %d bytes. Only pagesize (%ld) or less is supported.",
> > > >  				mp->m_sb.sb_blocksize, PAGE_SIZE);
> > > > -		error = -ENOSYS;
> > > > -		goto out_free_sb;
> > > > +			error = -ENOSYS;
> > > > +			goto out_free_sb;
> > > > +		}
> > > > +
> > > > +		if (mp->m_sb.sb_blocksize > max_folio_size) {
> > > > +			xfs_warn(mp,
> > > > +"block size (%u bytes) not supported; maximum folio size supported in "\
> > > > +"the page cache is (%ld bytes). Check MAX_PAGECACHE_ORDER (%d)",
> > > > +			mp->m_sb.sb_blocksize, max_folio_size,
> > > > +			MAX_PAGECACHE_ORDER);
> > > 
> > > Again, too much message.  Way too much.  We shouldn't even allow block
> > > devices to be created if their block size is larger than the max supported
> > > by the page cache.
> > 
> > Filesystem blocksize != block device blocksize.  xfs still needs this
> > check because one can xfs_copy a 64k-fsblock xfs to a hdd with 512b
> > sectors and try to mount that on x86.
> > 
> > Assuming there /is/ some fs that allows 1G blocksize, you'd then really
> > want a mount check that would prevent you from mounting that.
> 
> Absolutely, we need to have an fs blocksize check in the fs (if only
> because fs fuzzers will put random values in fields and expect the system
> to not crash).  But that should have nothing to do with page cache size.

I don't understand your objection -- we're setting the minimum folio
order on a file's pagecache to match the fs-wide blocksize.  If the
pagecache can't possibly fulfill our fs-wide requirement, then why would
we continue the mount?

Let's pretend that MAX_PAGECACHE_ORDER is 1.  The filesystem has 16k
blocks, the CPU has 4k base pages.  xfs will try to set the min folio
order to 2 via mapping_set_folio_order_range.  That function clamps it
to 1, so we try to cache a 16k fsblock with 8k pages.  Does that
actually work?

If not, then doesn't it make more more sense to fail the mount?

--D