Re: [PATCH 2/2] xfs: Extend xattr extent counter to 32-bits

["Darrick J. Wong" <darrick.wong@oracle.com>]

On Tue, Apr 07, 2020 at 09:30:02AM +1000, Dave Chinner wrote:
> On Mon, Apr 06, 2020 at 10:06:03AM -0700, Darrick J. Wong wrote:
> > On Sat, Apr 04, 2020 at 02:22:03PM +0530, Chandan Rajendra wrote:
> > > XFS has a per-inode xattr extent counter which is 16 bits wide. A workload
> > > which
> > > 1. Creates 1,000,000 255-byte sized xattrs,
> > > 2. Deletes 50% of these xattrs in an alternating manner,
> > > 3. Tries to create 400,000 new 255-byte sized xattrs
> > > causes the following message to be printed on the console,
> > > 
> > > XFS (loop0): xfs_iflush_int: detected corrupt incore inode 131, total extents = -19916, nblocks = 102937, ptr ffff9ce33b098c00
> > > XFS (loop0): xfs_do_force_shutdown(0x8) called from line 3739 of file fs/xfs/xfs_inode.c. Return address = ffffffffa4a94173
> > > 
> > > This indicates that we overflowed the 16-bits wide xattr extent counter.
> > > 
> > > I have been informed that there are instances where a single file has
> > >  > 100 million hardlinks. With parent pointers being stored in xattr,
> > > we will overflow the 16-bits wide xattr extent counter when large
> > > number of hardlinks are created.
> > > 
> > > Hence this commit extends xattr extent counter to 32-bits. It also introduces
> > > an incompat flag to prevent older kernels from mounting newer filesystems with
> > > 32-bit wide xattr extent counter.
> > > 
> > > Signed-off-by: Chandan Rajendra <chandanrlinux@gmail.com>
> > > ---
> > >  fs/xfs/libxfs/xfs_format.h     | 28 +++++++++++++++++++++-------
> > >  fs/xfs/libxfs/xfs_inode_buf.c  | 27 +++++++++++++++++++--------
> > >  fs/xfs/libxfs/xfs_inode_fork.c |  3 ++-
> > >  fs/xfs/libxfs/xfs_log_format.h |  5 +++--
> > >  fs/xfs/libxfs/xfs_types.h      |  4 ++--
> > >  fs/xfs/scrub/inode.c           |  7 ++++---
> > >  fs/xfs/xfs_inode_item.c        |  3 ++-
> > >  fs/xfs/xfs_log_recover.c       | 13 ++++++++++---
> > >  8 files changed, 63 insertions(+), 27 deletions(-)
> > > 
> > > diff --git a/fs/xfs/libxfs/xfs_format.h b/fs/xfs/libxfs/xfs_format.h
> > > index 045556e78ee2c..0a4266b0d46e1 100644
> > > --- a/fs/xfs/libxfs/xfs_format.h
> > > +++ b/fs/xfs/libxfs/xfs_format.h
> > > @@ -465,10 +465,12 @@ xfs_sb_has_ro_compat_feature(
> > >  #define XFS_SB_FEAT_INCOMPAT_FTYPE	(1 << 0)	/* filetype in dirent */
> > >  #define XFS_SB_FEAT_INCOMPAT_SPINODES	(1 << 1)	/* sparse inode chunks */
> > >  #define XFS_SB_FEAT_INCOMPAT_META_UUID	(1 << 2)	/* metadata UUID */
> > > +#define XFS_SB_FEAT_INCOMPAT_32BIT_AEXT_CNTR (1 << 3)
> > 
> > If you're going to introduce an INCOMPAT feature, please also use the
> > opportunity to convert xattrs to something resembling the dir v3 format,
> > where we index free space within each block so that we can speed up attr
> > setting with 100 million attrs.
> 
> Not necessary. Chandan has already spent a lot of time investigating
> that - I suggested doing the investigation probably a year ago when
> he was looking for stuff to do knowing that this could be a problem
> parent pointers hit.

Oh, I didn't realize that analysis work has already been done.

Chandan, could you please mention that somewhere in the cover letter?
It does mention that you tried creating 1M xattrs, but I guess it needed
to be more explicit about not uncovering any gigantic performance holes.

> Long story short - there's no degradation in
> performance in the dabtree out to tens of millions of records with
> different fixed size or random sized attributes, nor does various
> combinations of insert/lookup/remove/replace operations seem to
> impact the tree performance at scale. IOWs, we hit the 16 bit extent
> limits of the attribute trees without finding any degradation in
> performance.

Ok.  I'll take "attr v3 upgrade" off my list of things to look out for.

> Hence we concluded that the dabtree structure does not require
> significant modification or optimisation to work well with typical
> parent pointer attribute demands...
> 
> As for free space indexes....
> 
> The issue with the directory structure that requires external free
> space is that the directory data is not part of the dabtree itself.
> The attribute fork stores all the attributes at the leaves of the
> dabtree, while the directory structure stores the directory data in
> external blocks and the dabtree only contains the name hash index
> that points to the external data.
> 
> i.e. When we add an attribute to the dabtree, we split/merge leaves
> of the tree based on where the name hash index tells us it needs to
> be inserted/removed from. i.e. we make space available or collapse
> sparse leaves of the dabtree as a side effect of inserting or
> removing objects.
> 
> The directory structure is very different. The dirents cannot change
> location as their logical offset into the dir data segment is used
> as the readdir/seekdir/telldir cookie. Therefore that location is
> not allowed to change for the life of the dirent and so we can't
> store them in the leaves of a dabtree indexed in hash order because
> the offset into the tree would change as other entries are inserted
> and removed.  Hence when we remove dirents, we must leave holes in
> the data segment so the rest of the dirent data does not change
> logical offset.
> 
> The directory name hash index - the dabtree bit - is in a separate
> segment (the 2nd one). Because it only stores pointers to dirents in
> the data segment, it doesn't need to leave holes - the dabtree just
> merge/splits as required as pointers to the dir data segment are
> added/removed - and has no free space tracking.
> 
> Hence when we go to add a dirent, we need to find the best free
> space in the dir data segment to add that dirent. This requires a
> dir data segment free space index, and that is held in the 3rd dir
> segment.  Once we've found the best free space via lookup in the
> free space index, we go modify the dir data block it points to, then
> update the dabtree to point the name hash at that new dirent.
> 
> IOWs, the requirement for a free space map in the directory
> structure results from storing the dirent data externally to the
> dabtree. Attributes are stored directly in the leaves of the
> dabtree - except for remote attributes which can be anywhere in the
> BMBT address space - and hence do no need external free space
> tracking to determine where to best insert them...

<nod> Got it.  I've suspected this property about the xattr structures
for a long time, so I'm glad to hear someone else echo that. :)

Dave: May I try to rework the above into something suitable for the
ondisk format documentation?

--D

> Cheers,
> 
> Dave.
> -- 
> Dave Chinner
> david@fromorbit.com