NFSv4: Fix a potential CLOSE race

 Once the state_owner and lock_owner semaphores get removed, it will be
 possible for other OPEN requests to reopen the same file if they have
 lower sequence ids than our CLOSE call.
 This patch ensures that we recheck the file state once
 nfs_wait_on_sequence() has completed waiting.

 Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>

1 files changed, 77 insertions, 33 deletions

diff --git a/fs/nfs/nfs4proc.c b/fs/nfs/nfs4proc.c
index 9ba89e7cdd28..5154ddf6d9a5 100644
--- a/fs/nfs/nfs4proc.c
+++ b/fs/nfs/nfs4proc.c
@@ -189,6 +189,21 @@ static void update_changeattr(struct inode *inode, struct nfs4_change_info *cinf
 		nfsi->change_attr = cinfo->after;
 }
 
+/* Helper for asynchronous RPC calls */
+static int nfs4_call_async(struct rpc_clnt *clnt, rpc_action tk_begin,
+		rpc_action tk_exit, void *calldata)
+{
+	struct rpc_task *task;
+
+	if (!(task = rpc_new_task(clnt, tk_exit, RPC_TASK_ASYNC)))
+		return -ENOMEM;
+
+	task->tk_calldata = calldata;
+	task->tk_action = tk_begin;
+	rpc_execute(task);
+	return 0;
+}
+
 static void update_open_stateid(struct nfs4_state *state, nfs4_stateid *stateid, int open_flags)
 {
 	struct inode *inode = state->inode;
@@ -810,11 +825,24 @@ struct nfs4_closedata {
 	struct nfs_closeres res;
 };
 
+static void nfs4_free_closedata(struct nfs4_closedata *calldata)
+{
+	struct nfs4_state *state = calldata->state;
+	struct nfs4_state_owner *sp = state->owner;
+	struct nfs_server *server = NFS_SERVER(calldata->inode);
+
+	nfs4_put_open_state(calldata->state);
+	nfs_free_seqid(calldata->arg.seqid);
+	up(&sp->so_sema);
+	nfs4_put_state_owner(sp);
+	up_read(&server->nfs4_state->cl_sem);
+	kfree(calldata);
+}
+
 static void nfs4_close_done(struct rpc_task *task)
 {
 	struct nfs4_closedata *calldata = (struct nfs4_closedata *)task->tk_calldata;
 	struct nfs4_state *state = calldata->state;
-	struct nfs4_state_owner *sp = state->owner;
 	struct nfs_server *server = NFS_SERVER(calldata->inode);
 
         /* hmm. we are done with the inode, and in the process of freeing
@@ -838,25 +866,46 @@ static void nfs4_close_done(struct rpc_task *task)
 			}
 	}
 	state->state = calldata->arg.open_flags;
-	nfs4_put_open_state(state);
-	nfs_free_seqid(calldata->arg.seqid);
-	up(&sp->so_sema);
-	nfs4_put_state_owner(sp);
-	up_read(&server->nfs4_state->cl_sem);
-	kfree(calldata);
+	nfs4_free_closedata(calldata);
 }
 
-static inline int nfs4_close_call(struct rpc_clnt *clnt, struct nfs4_closedata *calldata)
+static void nfs4_close_begin(struct rpc_task *task)
 {
+	struct nfs4_closedata *calldata = (struct nfs4_closedata *)task->tk_calldata;
+	struct nfs4_state *state = calldata->state;
 	struct rpc_message msg = {
 		.rpc_proc = &nfs4_procedures[NFSPROC4_CLNT_CLOSE],
 		.rpc_argp = &calldata->arg,
 		.rpc_resp = &calldata->res,
-		.rpc_cred = calldata->state->owner->so_cred,
+		.rpc_cred = state->owner->so_cred,
 	};
-	if (calldata->arg.open_flags != 0)
+	int mode = 0;
+	int status;
+
+	status = nfs_wait_on_sequence(calldata->arg.seqid, task);
+	if (status != 0)
+		return;
+	/* Don't reorder reads */
+	smp_rmb();
+	/* Recalculate the new open mode in case someone reopened the file
+	 * while we were waiting in line to be scheduled.
+	 */
+	if (state->nreaders != 0)
+		mode |= FMODE_READ;
+	if (state->nwriters != 0)
+		mode |= FMODE_WRITE;
+	if (test_bit(NFS_DELEGATED_STATE, &state->flags))
+		state->state = mode;
+	if (mode == state->state) {
+		nfs4_free_closedata(calldata);
+		task->tk_exit = NULL;
+		rpc_exit(task, 0);
+		return;
+	}
+	if (mode != 0)
 		msg.rpc_proc = &nfs4_procedures[NFSPROC4_CLNT_OPEN_DOWNGRADE];
-	return rpc_call_async(clnt, &msg, 0, nfs4_close_done, calldata);
+	calldata->arg.open_flags = mode;
+	rpc_call_setup(task, &msg, 0);
 }
 
 /* 
@@ -873,35 +922,30 @@ static inline int nfs4_close_call(struct rpc_clnt *clnt, struct nfs4_closedata *
 int nfs4_do_close(struct inode *inode, struct nfs4_state *state, mode_t mode) 
 {
 	struct nfs4_closedata *calldata;
-	int status;
+	int status = -ENOMEM;
 
-	/* Tell caller we're done */
-	if (test_bit(NFS_DELEGATED_STATE, &state->flags)) {
-		state->state = mode;
-		return 0;
-	}
-	calldata = (struct nfs4_closedata *)kmalloc(sizeof(*calldata), GFP_KERNEL);
+	calldata = kmalloc(sizeof(*calldata), GFP_KERNEL);
 	if (calldata == NULL)
-		return -ENOMEM;
+		goto out;
 	calldata->inode = inode;
 	calldata->state = state;
 	calldata->arg.fh = NFS_FH(inode);
+	calldata->arg.stateid = &state->stateid;
 	/* Serialization for the sequence id */
 	calldata->arg.seqid = nfs_alloc_seqid(&state->owner->so_seqid);
-	if (calldata->arg.seqid == NULL) {
-		kfree(calldata);
-		return -ENOMEM;
-	}
-	calldata->arg.open_flags = mode;
-	memcpy(&calldata->arg.stateid, &state->stateid,
-			sizeof(calldata->arg.stateid));
-	status = nfs4_close_call(NFS_SERVER(inode)->client, calldata);
-	/*
-	 * Return -EINPROGRESS on success in order to indicate to the
-	 * caller that an asynchronous RPC call has been launched, and
-	 * that it will release the semaphores on completion.
-	 */
-	return (status == 0) ? -EINPROGRESS : status;
+	if (calldata->arg.seqid == NULL)
+		goto out_free_calldata;
+
+	status = nfs4_call_async(NFS_SERVER(inode)->client, nfs4_close_begin,
+			nfs4_close_done, calldata);
+	if (status == 0)
+		goto out;
+
+	nfs_free_seqid(calldata->arg.seqid);
+out_free_calldata:
+	kfree(calldata);
+out:
+	return status;
 }
 
 struct inode *