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

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

On Thu, Jun 07, 2012 at 04:12:45PM -0700, Eric W. Biederman wrote:
> We take the approprate dentry locks in the approparite order so d_move
> and the dcache should not care in the slightest about the inode
> mutecies.
> 
> If we need the inode mutecies to make the dcache bits safe then
> something really is insane.  There may be subtle insanities in the
> vfs that require the inode muticies of the parents in d_move but I am
> certainly not seeing them.  At least as I read it the code in __d_move
> only touches and modifies dentry fields.

Yes.  Now, go take a look at e.g. the locking order on ->d_lock.  No,
I'm not saying that I like it.  Not at all.  But we do rely on the
non-local protections for tree topology, just to make sure that the
damn thing has the locking order consistent - not changing between
the moments you take locks you've ordered, for starters.

I realize that "serialize all operations on a single per-machine mutex" is a
solution.  It's just not something feasible when we are talking about all
directory tree modifications on a general-purpose filesystem.  So no,
sysfs approach to that kind of problems is not usable here.

I doubt that we have something sysfs-related in the deadlocks davej is seeing,
but I seriously suspect that I can cook one based on sysfs_rename() setting
the things up for silent topology changes on ->lookup().  I would suggest
using d_materialise_unique() there - that one *does* take care to take
locks needed.

BTW, looking at the code in sysfs_lookup()... why bother with d_set_d_op()
instead of just sb->s_d_op = &sysfs_dentry_ops; once during sysfs_fill_super()?
In the worst case you need to do that after you've allocated the root
dentry, depending on whether you are willing or not to make ->d_revalidate()
return 1 whenever it's called on the root dentry...