Re: [PATCH] fuse: prevent exchange/revalidate races

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

On Wed, Sep 17, 2025 at 06:43:50PM +0100, Al Viro wrote:
> On Wed, Sep 17, 2025 at 07:27:25PM +0200, Miklos Szeredi wrote:
> > On Wed, 17 Sept 2025 at 18:41, Al Viro <viro@zeniv.linux.org.uk> wrote:
> > >
> > > On Wed, Sep 17, 2025 at 05:42:18PM +0200, Miklos Szeredi wrote:
> > > > > ... and if the call of ->d_revalidate() had been with parent locked, you've
> > > > > just got a deadlock in that case.
> > > >
> > > > Why?
> > >
> > > Because the locking order on directories is "ancestors first"; you are trying
> > > to grab an inode that is also a directory and might be anywhere in the tree
> > > by that point - that's precisely what you are trying to check, isn't it?
> > 
> > But if parent is locked, it must not have been renamed, hence
> > parent-child relationship holds.
> 
> Not if parent is held shared.

Note that lookup will reparent a subdirectory just fine - exclusive lock on
that directory's parent would make it fail, but shared one is not a problem.