Hi Aneesh,
I'm going to rely on your memory here... A long, long time ago (2011!),
you've authored this commit:
-------
commit ed80fcfac2565fa866d93ba14f0e75de17a8223e
Author: Aneesh Kumar K.V <aneesh.kumar@xxxxxxxxxxxxxxxxxx>
Date: Wed Jul 6 16:32:31 2011 +0530
fs/9p: Always ask new inode in create
This make sure we don't end up reusing the unlinked inode object.
The ideal way is to use inode i_generation. But i_generation is
not available in userspace always.
Signed-off-by: Aneesh Kumar K.V <aneesh.kumar@xxxxxxxxxxxxxxxxxx>
Signed-off-by: Eric Van Hensbergen <ericvh@xxxxxxxxx>
-------
Do you happen to remember or know *why* you wanted to make sure we don't
reuse the unlinked inode object?
I'm asking because that's causing problems with (at least) fscache
cookie, iget5_locked() gets us a new inode in v9fs_qid_iget_dotl()
and tries to get a new cookie before the evict has happened and
relinquished the former inode's.
There's also problems with coherency in sight -- evict is also in charge
of flushing all dirty pages, so the new inode can in theory issue IO and
read from server data which has been written by another process on the
same client and while 9p isn't known for coherency with multiple clients
it's a different story with a single one! (didn't reproduce that one)
Anyway, it'd be great to know why you did that so we can try another
workaround.
In theory I'd love to have qemu and others export fsid + a fhandle from
name_to_handle_at that includes i_generation and full inode number in
the qid path, but we're limited by the 64bits of the protocol so it's a
tough one... In practice I don't see generation being used all that much
by filesystems to reuse inode numbers though, so wondering which is the
most problematic?
You can find the rest of the thread here if you're not subscribed to
v9fs-developer or linux-fsdevel:
https://lkml.kernel.org/r/87czu45gcs.fsf@xxxxxxx
Luis Henriques wrote on Fri, May 14, 2021 at 05:10:56PM +0100:
> So, from our last chat on IRC, we have the following happening:
>
> v9fs_vfs_atomic_open_dotl
> v9fs_vfs_lookup
> v9fs_get_new_inode_from_fid
> v9fs_inode_from_fid_dotl
> v9fs_qid_iget_dotl
>
> At this point, iget5_locked() gets called with the test function set to
> v9fs_test_new_inode_dotl(), which *always* returns 0. It's still not
> clear to me why commit ed80fcfac256 ("fs/9p: Always ask new inode in
> create") has introduced this behavior but even if that's not correct, we
> still have a race regarding cookies handling, right?
>
> I'm still seeing:
>
> CPU0 CPU1
> v9fs_drop_inode ...
> v9fs_evict_inode /* atomic_open */
> v9fs_cache_inode_get_cookie <= COLLISION
> fscache_relinquish
Do you mean you still have that problem after ed80fcfac256 has been
reverted?
> So, the question remains: would it be possible to do the relinquish
> earlier (->drop_inode)? Or is 9p really shooting itself in the foot by
> forcing iget5_locked() to always create a new inode here?
Ugh that is the kind of things I don't want to experiment with...
->drop_inode() seems to be called with i_lock taken and meant to be just
a test, not something that can wait, but from what I'm reading it might
be possible to set I_WILL_FREE, drop the lock, do our stuff and reaquire
the lock at the end.. perhaps? It looks like inode lookup will just loop
around on ilookup5_nowait while I_WILL_FREE is set so new inodes can't
be taken at this point, it's a de-facto spin lock with iget5_locked and
friends.
I have no idea what will break though, I'd really rather leave it to the
vfs and have 9p do the right thing with inode recycling.
--
Dominique
