On Thu, 2011-12-22 at 08:11 -0800, Sage Weil wrote:
> On Thu, 22 Dec 2011, Alex Elder wrote:
> > I'll send all this to ceph-devel later, I just wanted
> [added cc]
>
> > to document it while it's fresh and send it to *someone*
> > and you made this change so I'm sending it to you. The
> > change in question is:
> >
> > be655596b3de5873f994ddbe205751a5ffb4de39
> > ceph: use i_ceph_lock instead of i_lock
> >
> > First, after updating my ceph tree and rebuilding I found
> > that every time I completed a mount in my UML environment
> > it would immediately crash the UML.
> >
> > This led to me learning how to attach a debugger to
> > the UML (good) and this morning I was able to capture
> > this situation in gdb.
> >
> > The problem is in ceph_dir_set_complete(), which is
> > blindly dereferencing the return value of
> > __d_find_any_alias(inode), and that function is
> > returning NULL because the inode's i_dentry list
> > is empty.
> >
> > So that's problem 1.
>
> Yeah, looks like I broke that with
> 9f6f8f50663e6cf42cca64f1f5ee0ca438c33c46. Just moving the dentry->d_sb
> inside the if should do it! Strange that this didn't trigger on sepia...
> :/
I have a patch that fixes this one and I have confirmed I
can now get my UML filesystem mounted with it applied.
> > In looking at this I find that __d_find_any_alias()
> > is duplicated in ceph, to (almost) match the Linux
> > implementation which is not exported. But unlike
> > the Linux version, the ceph version does NOT grab
> > a reference to the found alias. I consider this a
> > bug, because a function with an identical name should
> > perform identically to the original, otherwise it
> > should be named something different.
> >
> > That's problem 2.
> > Furthermore, I looked at the locking that's done for
> > the Linux version of __d_find_any_alias, and that
> > truly does require taking the Linux inode lock--the
> > (new) ceph inode i_ceph_lock is not sufficient. Note
> > that __d_find_any_alias() really only operates on the
> > Linux inode and does nothing with the ceph inode, so
> > it may be that simply changing the locking used *in
> > this case* back to using the Linux inode lock is
> > enough to fix the problem. But I'm not familiar
> > enough at this point to know the locking context
> > where these calls are made, so can't be sure doing
> > this would cause problems (like lock inversions).
> >
> > There are three places this __d_find_any_alias()
> > is called: ceph_dir_{set,clear,test}_complete().
> > Those are straightforward, but the calls to *those*
> > functions are mostly embedded in long strings of
> > logical conditions so it's going to take a little
> > work to unwind this.
> >
> > So that's problem 3, which is the biggest one.
>
> I think the __d_find_any_alias() needs to take the i_lock and a reference
> to make it fully safe. i_lock is ordered inside i_ceph_lock, now, so that
> should be safe in all of the call chains. Better yet, we should just
> export the VFS version and use that.
I'll put together a patch that both adds the reference,
and fixes callers so they release it properly.
> What do you think? A simple fix is best so we can get it into 3.2 (which
> is about a week away?).
What I described is simplest. I don't want to suggest
a VFS change at this point.
> > Potential problem 4 is that this suggests the
> > mechanical switch to using i_ceph_lock is a little
> > suspect and might need a closer look.
>
> Yeah. Well, this is one of the few cases where i_lock where we muddle
> with the dcache internal structure. The others should pretty much all be
> in dir.c, and maybe inode.c's splice_dentry() and friends. The dcache
> readdir is by far the most suspect of all of those, but also probably the
> lowest priority at the moment. Maybe your visit in January would be a
> good time to go over it in detail?
Sure. Or I could review the committed patch. For now
I'll get something togehter for the 3.2 release.
-Alex
--
To unsubscribe from this list: send the line "unsubscribe ceph-devel" in
the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
