Re: [PATCH 2/2] xfs: remove metafile inodes from the active inode stat

[Christoph Hellwig <hch@lst.de>]

On Tue, Feb 03, 2026 at 12:41:23PM +0530, Nirjhar Roy (IBM) wrote:
> On Mon, 2026-02-02 at 15:14 +0100, Christoph Hellwig wrote:
> > The active inode (or active vnode until recently) stat can get much larger
> > than expected on file systems with a lot of metafile inodes like zoned
> > file systems on SMR hard disks with 10.000s of rtg rmap inodes.
> And this was causing (or could have caused) some sort of counter overflows or something?  

Not really an overflow.  But if you have a lot of metadir inodes it
messes up the stats with extra counts that are not user visible.

> > This fixes xfs/177 on SMR hard drives.
> I can see that xfs/177 has a couple of sub-test cases (like Round 1,2, ...) - do you remember if 1
> particular round was causing problems or were there issues with all/most of them?

Comparing the cached values to the expected ones.

> So is it like then there is a state(or at some function) where
> xs_inodes_active counter was bumped up even though "ip" was a metadir
> inode and here in the above line it is corrected (i.e, decremented
> by 1) and xs_inodes_meta is incremented - shouldn't the appropriate
> counter have been directly bumped up whenever it was created?

xfs_inode_alloc doesn't know if the inode is going to be a meta inode,
as it hasn't been read from disk yet for the common case.  For the
less common inode allocation case we'd know it, but passing it down
would be a bit annoying.

> > +/* Metafile counters */
> > +	uint32_t		xs_inodes_meta;
> uint64_t would be an overkill, isn't it?

Yes.  Sticking to the same type as the active inodes here.