Re: [PATCH for 6.19-rc1] fs: preserve file type in make_bad_inode() unless invalid

[Al Viro <viro@zeniv.linux.org.uk>]

On Wed, Dec 10, 2025 at 08:43:38PM +0000, Al Viro wrote:
> On Wed, Dec 10, 2025 at 03:35:31PM +0000, Al Viro wrote:
> > On Wed, Dec 10, 2025 at 11:09:24AM +0100, Mateusz Guzik wrote:
> > > On Wed, Dec 10, 2025 at 10:45 AM Tetsuo Handa
> > > <penguin-kernel@i-love.sakura.ne.jp> wrote:
> > > >
> > > > syzbot is hitting VFS_BUG_ON_INODE(!S_ISDIR(inode->i_mode)) check
> > > > introduced by commit e631df89cd5d ("fs: speed up path lookup with cheaper
> > > > handling of MAY_EXEC"), for make_bad_inode() is blindly changing file type
> > > > to S_IFREG. Since make_bad_inode() might be called after an inode is fully
> > > > constructed, make_bad_inode() should not needlessly change file type.
> > > >
> > > 
> > > ouch
> > > 
> > > So let's say calls to make_bad_inode *after* d_instantiate are unavoidable.
> > 
> > ... and each one is a bug.
> 
> FWIW, I'm very tempted to fold make_bad_inode() into iget_failed().  Other
> callers tend to be either pointless (e.g. ext2_new_inode() after reaching
> fail: label - we only get there if inode has never reached inode hash
> table; make_bad_inode() in there should've been gone for a long time)
> or outright broken.
> 
> There's not a lot of callers, thankfully; I'm going through those at the
> moment, but so far the impression is that we should be able to simply bury
> the damn thing.

While we are at it, 73861970938a "minixfs: Verify inode mode when loading from
disk" that introduced one of those is seriously misguided - sanity check belongs
in V1_minix_iget/V2_minix_iget, and should be handled there the same way we
deal with zero i_nlink.

We really ought to take that function out - as it is, it's an attractive
nuisance...