[PATCH 3/6] fs: distinguish between user initiated freeze and kernel initiated freeze

[Darrick J. Wong]

On Thu, Jun 08, 2023 at 11:11:30AM +0200, Jan Kara wrote:
> On Wed 07-06-23 22:29:04, Christoph Hellwig wrote:
> > On Thu, May 25, 2023 at 04:14:30PM +0200, Jan Kara wrote:
> > > Yes, this is exactly how I'd imagine it. Thanks for writing the patch!
> > > 
> > > I'd just note that this would need rebasing on top of Luis' patches 1 and
> > > 2. Also:
> > 
> > I'd not do that for now.  1 needs a lot more work, and 2 seems rather
> > questionable.
> 
> OK, I agree the wrappers could be confusing (they didn't confuse me but
> when you spelled it out, I agree).
> 
> > > Now the only remaining issue with the code is that the two different
> > > holders can be attempting to freeze the filesystem at once and in that case
> > > one of them has to wait for the other one instead of returning -EBUSY as
> > > would happen currently. This can happen because we temporarily drop
> > > s_umount in freeze_super() due to lock ordering issues. I think we could
> > > do something like:
> > > 
> > > 	if (!sb_unfrozen(sb)) {
> > > 		up_write(&sb->s_umount);
> > > 		wait_var_event(&sb->s_writers.frozen,
> > > 			       sb_unfrozen(sb) || sb_frozen(sb));
> > > 		down_write(&sb->s_umount);
> > > 		goto retry;
> > > 	}
> > > 
> > > and then sprinkle wake_up_var(&sb->s_writers.frozen) at appropriate places
> > > in freeze_super().
> > 
> > Let's do that separately as a follow on..
> 
> Well, we need to somehow settle on how to deal with a race when both kernel
> & userspace race to freeze the filesystem and make the result consistent
> with the situation when the fs is already frozen by someone.

<nod> I'll see what I can do about that.

> > > BTW, when reading this code, I've spotted attached cleanup opportunity but
> > > I'll queue that separately so that is JFYI.
> > > 
> > > > +#define FREEZE_HOLDER_USERSPACE	(1U << 1)	/* userspace froze fs */
> > > > +#define FREEZE_HOLDER_KERNEL	(1U << 2)	/* kernel froze fs */
> > > 
> > > Why not start from 1U << 0? And bonus points for using BIT() macro :).
> > 
> > BIT() is a nasty thing and actually makes code harder to read. And it
> > doesn't interact very well with the __bitwise annotation that might
> > actually be useful here.
> 
> OK. I'm not too hung up on BIT() macro.
> 
> 								Honza
> -- 
> Jan Kara <jack at suse.com>
> SUSE Labs, CR