nfs: fix NULL deference in nfs4_get_valid_delegation
authorJ. Bruce Fields <bfields@redhat.com>
Mon, 11 May 2020 14:02:48 +0000 (10:02 -0400)
committerTrond Myklebust <trond.myklebust@hammerspace.com>
Mon, 11 May 2020 18:05:58 +0000 (14:05 -0400)
We add the new state to the nfsi->open_states list, making it
potentially visible to other threads, before we've finished initializing
it.

That wasn't a problem when all the readers were also taking the i_lock
(as we do here), but since we switched to RCU, there's now a possibility
that a reader could see the partially initialized state.

Symptoms observed were a crash when another thread called
nfs4_get_valid_delegation() on a NULL inode, resulting in an oops like:

BUG: unable to handle page fault for address: ffffffffffffffb0 ...
RIP: 0010:nfs4_get_valid_delegation+0x6/0x30 [nfsv4] ...
Call Trace:
 nfs4_open_prepare+0x80/0x1c0 [nfsv4]
 __rpc_execute+0x75/0x390 [sunrpc]
 ? finish_task_switch+0x75/0x260
 rpc_async_schedule+0x29/0x40 [sunrpc]
 process_one_work+0x1ad/0x370
 worker_thread+0x30/0x390
 ? create_worker+0x1a0/0x1a0
 kthread+0x10c/0x130
 ? kthread_park+0x80/0x80
 ret_from_fork+0x22/0x30

Fixes: 9ae075fdd190 "NFSv4: Convert open state lookup to use RCU"
Reviewed-by: Seiichi Ikarashi <s.ikarashi@fujitsu.com>
Tested-by: Daisuke Matsuda <matsuda-daisuke@fujitsu.com>
Tested-by: Masayoshi Mizuma <m.mizuma@jp.fujitsu.com>
Signed-off-by: J. Bruce Fields <bfields@redhat.com>
Cc: stable@vger.kernel.org # v4.20+
Signed-off-by: Trond Myklebust <trond.myklebust@hammerspace.com>
fs/nfs/nfs4state.c

index ac93715..a8dc25c 100644 (file)
@@ -734,9 +734,9 @@ nfs4_get_open_state(struct inode *inode, struct nfs4_state_owner *owner)
                state = new;
                state->owner = owner;
                atomic_inc(&owner->so_count);
-               list_add_rcu(&state->inode_states, &nfsi->open_states);
                ihold(inode);
                state->inode = inode;
+               list_add_rcu(&state->inode_states, &nfsi->open_states);
                spin_unlock(&inode->i_lock);
                /* Note: The reclaim code dictates that we add stateless
                 * and read-only stateids to the end of the list */