Re: processes hung after sys_renameat, and 'missing' processes

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

On Thu, Jun 07, 2012 at 07:08:04PM -0700, Eric W. Biederman wrote:
> > +	dentry->d_fsdata = sysfs_get(sd);
> > +	ret = d_materialise_unique(dentry, inode);
> 
> I have a small problem with d_materialise_unique.  For renames of files
> d_materialise_unique calls __d_instantiate_unique. __d_instantiate_unique
> does not detect renames of files.  Which at least misses the rename
> of sysfs symlinks.

Er... yes, but why do we care?  It's not as if you had a hardwired
reference to dentry from your objects, after all (can't, with multiple
superblocks).  So you get old stale dentry at the old location and
a new one where we'd moved that sucker.  They have the same inode
and each holds a reference to the same sd; ->d_revalidate() at the
old location must invalidate the old instance anyway, since you are
not guaranteed that lookup at the new one will happen before repeated
lookup at the old one.

Directories *are* special in that respect, but symlinks are trivial...
VFS doesn't care if you have extra dentries for those and neither does
sysfs, AFAICS.

It's not that we couldn't teach d_materialise_unique() about those (e.g.
introduce a new dentry flag and treat dentries with it as directories
for d_materialise_unique() purposes); I would like to understand the
reasons for doing that, though.