In message <20080424124245.GC15214@ZenIV.linux.org.uk>, Al Viro writes:
quoted text
> On Thu, Apr 24, 2008 at 01:39:50PM +0200, Miklos Szeredi wrote:
> > Then I did this series, which basically guarantees, that that cannot
> > happen.  Al rejected it, and rather fixed some of the remaining
> > places.  He still missed several, which sort of proves my point.
> 
> Which ones have I missed?
>  
> > I think it's totally pointless to continue trying to fix the symptoms
> > instead of getting at the root of the problem.
> > 
> > I know that VFS interfaces are a sensitive question, but it would be
> > nice it we could have some sanity back in this discussion.
> 
> Yes, it would.  How about that, for starters:
> 
> path_create() et.al. are *wrong* for nfsd; if nothing else, I'm not at
> all convinced that even apparmour wants export path + relative there
> _and_ r/o vs. r/w is decision that doesn't clearly map to ex_mnt flags.
> 
> Moreover, it's not at all obvious that we want to drop write access as
> soon as vfs_...() is over in case of nfsd.  Some of the stuff done
> immeidately afterwards might very well qualify for inclusion into
> protected area; some of the stuff done immediately _prior_ very likely
> needs that as well - look at fh_verify() and tell me why we don't want
> that "I'll hold write access to vfsmount" to span the area including
> that sucker.  If we want the r/o vs r/w policy directly vfsmount-based
> for nfsd, that is.
> 
> For ecryptfs it's also bogus - at the very least we need to decide what
> should happen when underlying vfsmount is remounted.  Again, I'm less
> than convinced that we want the same way to express r/o vs. r/w policy.

I tend to agree that the mnt_want* bracketing may need to encompass more
than one vfs_* method.  When copyup is involved, the problem is only
exacerbated.  A copyup may want to perform several actions "atomically", for
example:

1. vfs_mkdir /a
2. vfs_mkdir /a/b
3. vfs_mkdir /a/b/c
4. vfs_create /a/b/c/f
5. vfs_read(source file) -> vfs_write(/a/b/c/f)

These actions are often done in close proximity in unionfs, and in such a
case, it'd be nice if I could want_write only once around this sequence.

However, I'm not sure that for copyup (or ecryptfs), that an vfsmount-level
want/drop write is the right interface.  As you pointed out before, Al, some
other mechanism might be needed, perhaps at the superblock level, to declare
one's desire to write the f/s "atomically" (to avoid concurrent topology
changes).

Thanks,
Erez.
--
unsubscribe notice
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/