Re: [RFC] [PATCH 1/5] cgroups: revamp subsys array

[Ben Blum <bblum@andrew.cmu.edu>]

On Wed, Dec 09, 2009 at 02:07:53PM +0800, Li Zefan wrote:
> Ben Blum wrote:
> > On Tue, Dec 08, 2009 at 03:38:43PM +0800, Li Zefan wrote:
> >>> @@ -1291,6 +1324,7 @@ static int cgroup_get_sb(struct file_system_type *fs_type,
> >>>  	struct cgroupfs_root *new_root;
> >>>  
> >>>  	/* First find the desired set of subsystems */
> >>> +	down_read(&subsys_mutex);
> >> Hmm.. this can lead to deadlock. sget() returns success with sb->s_umount
> >> held, so here we have:
> >>
> >> 	down_read(&subsys_mutex);
> >>
> >> 	down_write(&sb->s_umount);
> >>
> >> On the other hand, sb->s_umount is held before calling kill_sb(),
> >> so when umounting we have:
> >>
> >> 	down_write(&sb->s_umount);
> >>
> >> 	down_read(&subsys_mutex);
> > 
> > Unless I'm gravely mistaken, you can't have deadlock on an rwsem when
> > it's being taken for reading in both cases? You would have to have at
> > least one of the cases being down_write.
> > 
> 
> lockdep will warn on this..

Hm. Why did I not see this warning...?

> And it can really lead to deadlock, though not so obivously:
> 
>   thread 1       thread 2        thread 3
> -------------------------------------------
> | read(A)        write(B)
> |
> |                                write(A)
> |
> |                read(A)
> |
> | write(B)
> |
> 
> t3 is waiting for t1 to release the lock, then t2 tries to
> acquire A lock to read, but it has to wait because of t3,
> and t1 has to wait t2.
> 
> Note: a read lock has to wait if a write lock is already
> waiting for the lock.

Okay, clever, the deadlock happens because of a behavioural optimization
of the rwsems. Good catch on the whole issue.

How does this sound as a possible solution, in cgroup_get_sb:

1) Take subsys_mutex
2) Call parse_cgroupfs_options()
3) Drop subsys_mutex
4) Call sget(), which gets sb->s_umount without subsys_mutex held
5) Take subsys_mutex
6) Call verify_cgroupfs_options()
7) Proceed as normal

In which verify_cgroupfs_options will be a new function that ensures the
invariants that rebind_subsystems expects are still there; if not, bail
out by jumping to drop_new_super just as if parse_cgroupfs_options had
failed in the first place.

Another question: What's the justification for having an interface of
seemingly symmetrical "initialize" and "destroy" functions, one of which
has to take a lock and the other gets called with the lock already held?
Seems like it's asking for trouble.

-- bblum