Commit 9a99af49 authored Feb 04, 2013 by Trond Myklebust

NFSv4.1: Prevent deadlocks between state recovery and file locking



We currently have a deadlock in which the state recovery thread
ends up blocking due to one of the locks which it is trying to
recover holding the nfs_inode->rwsem.
The situation is as follows: the state recovery thread is
scheduled in order to recover from a reboot. It immediately
drains the session, forcing all ordinary NFSv4.1 calls to
nfs41_setup_sequence() to be put to sleep.  This includes the
file locking process that holds the nfs_inode->rwsem.
When the thread gets to nfs4_reclaim_locks(), it tries to
grab a write lock on nfs_inode->rwsem, and boom...

Fix is to have the lock drop the nfs_inode->rwsem while it is
doing RPC calls. We use a sequence lock in order to signal to
the locking process whether or not a state recovery thread has
run on that inode, in which case it should retry the lock.

Reported-by: Andy Adamson <andros@netapp.com>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>

parent c137afab

fs/nfs/nfs4proc.c

+9 −0

Original line number	Diff line number	Diff line
		@@ -4813,8 +4813,10 @@ static int nfs41_lock_expired(struct nfs4_state state, struct file_lock reques

		static int _nfs4_proc_setlk(struct nfs4_state state, int cmd, struct file_lock request)
		{
		struct nfs4_state_owner *sp = state->owner;
		struct nfs_inode *nfsi = NFS_I(state->inode);
		unsigned char fl_flags = request->fl_flags;
		unsigned int seq;
		int status = -ENOLCK;

		if ((fl_flags & FL_POSIX) &&
		@@ -4836,9 +4838,16 @@ static int _nfs4_proc_setlk(struct nfs4_state *state, int cmd, struct file_lock
		status = do_vfs_lock(request->fl_file, request);
		goto out_unlock;
		}
		seq = raw_seqcount_begin(&sp->so_reclaim_seqcount);
		up_read(&nfsi->rwsem);
		status = _nfs4_do_setlk(state, cmd, request, NFS_LOCK_NEW);
		if (status != 0)
		goto out;
		down_read(&nfsi->rwsem);
		if (read_seqcount_retry(&sp->so_reclaim_seqcount, seq)) {
		status = -NFS4ERR_DELAY;
		goto out_unlock;
		}
		/* Note: we always want to sleep here! */
		request->fl_flags = fl_flags \| FL_SLEEP;
		if (do_vfs_lock(request->fl_file, request) < 0)