Skip to contentSkip to conversation

NFSv4.1 Fix a pNFS session draining deadlock

PATCH In review [email protected] 3 participants 3 messages May 20, 2013

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 NFSv4.1 Fix a pNFS session draining deadlock

From: Andy Adamson <[email protected]>

On a CB_RECALL the callback service thread flushes the inode using filemap_flush prior to scheduling the state manager thread to return the delegation. When pNFS is used and I/O has not yet gone to the data server servicing the inode, a LAYOUTGET can preceed the I/O. Unlike the async filemap_flush call, the LAYOUTGET must proceed to completion.

If the state manager starts to recover data while the inode flush is sending the LAYOUTGET, a deadlock occurs as the callback service thread holds the single callback session slot until the flushing is done which blocks the state manager thread, and the state manager thread has set the session draining bit which puts the inode flush LAYOUTGET RPC to sleep on the forechannel slot table waitq.

Separate the draining of the back channel from the draining of the fore channel by moving the NFS4_SESSION_DRAINING bit from session scope into the fore and back slot tables. Drain the back channel first allowing the LAYOUTGET call to proceed (and fail) so the callback service thread frees the callback slot. Then proceed with draining the forechannel.

---
 fs/nfs/callback_proc.c |  2 +-
 fs/nfs/callback_xdr.c  |  2 +-
 fs/nfs/nfs4proc.c      |  2 +-
 fs/nfs/nfs4session.c   |  6 ++++--
 fs/nfs/nfs4session.h   | 13 ++++++++-----
 fs/nfs/nfs4state.c     | 15 +++++++--------
 6 files changed, 22 insertions(+), 18 deletions(-)
fs/nfs/callback_proc.c +1 -1
index a13d26e..0bc2768 100644
@@ -414,7 +414,7 @@ __be32 nfs4_callback_sequence(struct cb_sequenceargs *args,
 
 	spin_lock(&tbl->slot_tbl_lock);
 	/* state manager is resetting the session */
-	if (test_bit(NFS4_SESSION_DRAINING, &clp->cl_session->session_state)) {
+	if (test_bit(NFS4_SLOT_TBL_DRAINING, &tbl->slot_tbl_state)) {
 		spin_unlock(&tbl->slot_tbl_lock);
 		status = htonl(NFS4ERR_DELAY);
 		/* Return NFS4ERR_BADSESSION if we're draining the session
fs/nfs/callback_xdr.c +1 -1
index 59461c9..a35582c 100644
@@ -763,7 +763,7 @@ static void nfs4_callback_free_slot(struct nfs4_session *session)
 	 * A single slot, so highest used slotid is either 0 or -1
 	 */
 	tbl->highest_used_slotid = NFS4_NO_SLOT;
-	nfs4_session_drain_complete(session, tbl);
+	nfs4_slot_tbl_drain_complete(tbl);
 	spin_unlock(&tbl->slot_tbl_lock);
 }
 
fs/nfs/nfs4proc.c +1 -1
index 8fbc100..4e2fe71 100644
@@ -572,7 +572,7 @@ int nfs41_setup_sequence(struct nfs4_session *session,
 	task->tk_timeout = 0;
 
 	spin_lock(&tbl->slot_tbl_lock);
-	if (test_bit(NFS4_SESSION_DRAINING, &session->session_state) &&
+	if (test_bit(NFS4_SLOT_TBL_DRAINING, &tbl->slot_tbl_state) &&
 	    !args->sa_privileged) {
 		/* The state manager will wait until the slot table is empty */
 		dprintk("%s session is draining\n", __func__);
fs/nfs/nfs4session.c +4 -2
index ebda5f4..7ecfb75 100644
@@ -73,7 +73,7 @@ void nfs4_free_slot(struct nfs4_slot_table *tbl, struct nfs4_slot *slot)
 			tbl->highest_used_slotid = new_max;
 		else {
 			tbl->highest_used_slotid = NFS4_NO_SLOT;
-			nfs4_session_drain_complete(tbl->session, tbl);
+			nfs4_slot_tbl_drain_complete(tbl);
 		}
 	}
 	dprintk("%s: slotid %u highest_used_slotid %d\n", __func__,
@@ -226,7 +226,7 @@ static bool nfs41_assign_slot(struct rpc_task *task, void *pslot)
 	struct nfs4_slot *slot = pslot;
 	struct nfs4_slot_table *tbl = slot->table;
 
-	if (nfs4_session_draining(tbl->session) && !args->sa_privileged)
+	if (nfs4_slot_tbl_draining(tbl) && !args->sa_privileged)
 		return false;
 	slot->generation = tbl->generation;
 	args->sa_slot = slot;
@@ -395,12 +395,14 @@ int nfs4_setup_session_slot_tables(struct nfs4_session *ses)
 	/* Fore channel */
 	tbl = &ses->fc_slot_table;
 	tbl->session = ses;
+	tbl->slot_tbl_state = 0;
 	status = nfs4_realloc_slot_table(tbl, ses->fc_attrs.max_reqs, 1);
 	if (status) /* -ENOMEM */
 		return status;
 	/* Back channel */
 	tbl = &ses->bc_slot_table;
 	tbl->session = ses;
+	tbl->slot_tbl_state = 0;
 	status = nfs4_realloc_slot_table(tbl, ses->bc_attrs.max_reqs, 0);
 	if (status && tbl->slots == NULL)
 		/* Fore and back channel share a connection so get
fs/nfs/nfs4session.h +8 -5
index 6f3cb39..ff7d9f0 100644
@@ -25,6 +25,10 @@ struct nfs4_slot {
 };
 
 /* Sessions */
+enum nfs4_slot_tbl_state {
+	NFS4_SLOT_TBL_DRAINING,
+};
+
 #define SLOT_TABLE_SZ DIV_ROUND_UP(NFS4_MAX_SLOT_TABLE, 8*sizeof(long))
 struct nfs4_slot_table {
 	struct nfs4_session *session;		/* Parent session */
@@ -43,6 +47,7 @@ struct nfs4_slot_table {
 	unsigned long	generation;		/* Generation counter for
 						   target_highest_slotid */
 	struct completion complete;
+	unsigned long	slot_tbl_state;
 };
 
 /*
@@ -68,7 +73,6 @@ struct nfs4_session {
 
 enum nfs4_session_state {
 	NFS4_SESSION_INITING,
-	NFS4_SESSION_DRAINING,
 };
 
 #if defined(CONFIG_NFS_V4_1)
@@ -88,12 +92,11 @@ extern void nfs4_destroy_session(struct nfs4_session *session);
 extern int nfs4_init_session(struct nfs_server *server);
 extern int nfs4_init_ds_session(struct nfs_client *, unsigned long);
 
-extern void nfs4_session_drain_complete(struct nfs4_session *session,
-		struct nfs4_slot_table *tbl);
+extern void nfs4_slot_tbl_drain_complete(struct nfs4_slot_table *tbl);
 
-static inline bool nfs4_session_draining(struct nfs4_session *session)
+static inline bool nfs4_slot_tbl_draining(struct nfs4_slot_table *tbl)
 {
-	return !!test_bit(NFS4_SESSION_DRAINING, &session->session_state);
+	return !!test_bit(NFS4_SLOT_TBL_DRAINING, &tbl->slot_tbl_state);
 }
 
 bool nfs41_wake_and_assign_slot(struct nfs4_slot_table *tbl,
fs/nfs/nfs4state.c +7 -8
index 300d17d..1fab140 100644
@@ -241,7 +241,7 @@ static void nfs4_end_drain_session(struct nfs_client *clp)
 	if (ses == NULL)
 		return;
 	tbl = &ses->fc_slot_table;
-	if (test_and_clear_bit(NFS4_SESSION_DRAINING, &ses->session_state)) {
+	if (test_and_clear_bit(NFS4_SLOT_TBL_DRAINING, &tbl->slot_tbl_state)) {
 		spin_lock(&tbl->slot_tbl_lock);
 		nfs41_wake_slot_table(tbl);
 		spin_unlock(&tbl->slot_tbl_lock);
@@ -251,15 +251,15 @@ static void nfs4_end_drain_session(struct nfs_client *clp)
 /*
  * Signal state manager thread if session fore channel is drained
  */
-void nfs4_session_drain_complete(struct nfs4_session *session,
-		struct nfs4_slot_table *tbl)
+void nfs4_slot_tbl_drain_complete(struct nfs4_slot_table *tbl)
 {
-	if (nfs4_session_draining(session))
+	if (nfs4_slot_tbl_draining(tbl))
 		complete(&tbl->complete);
 }
 
-static int nfs4_wait_on_slot_tbl(struct nfs4_slot_table *tbl)
+static int nfs4_drain_slot_tbl(struct nfs4_slot_table *tbl)
 {
+	set_bit(NFS4_SLOT_TBL_DRAINING, &tbl->slot_tbl_state);
 	spin_lock(&tbl->slot_tbl_lock);
 	if (tbl->highest_used_slotid != NFS4_NO_SLOT) {
 		INIT_COMPLETION(tbl->complete);
@@ -275,13 +275,12 @@ static int nfs4_begin_drain_session(struct nfs_client *clp)
 	struct nfs4_session *ses = clp->cl_session;
 	int ret = 0;
 
-	set_bit(NFS4_SESSION_DRAINING, &ses->session_state);
 	/* back channel */
-	ret = nfs4_wait_on_slot_tbl(&ses->bc_slot_table);
+	ret = nfs4_drain_slot_tbl(&ses->bc_slot_table);
 	if (ret)
 		return ret;
 	/* fore channel */
-	return nfs4_wait_on_slot_tbl(&ses->fc_slot_table);
+	return nfs4_drain_slot_tbl(&ses->fc_slot_table);
 }
 
 static void nfs41_finish_session_reset(struct nfs_client *clp)
Signature · 1 line
1.7.11.7
#1 lore raw
Quoted 37 lines · On Mon, 2013-05-20 at 14:13 -0400, [email protected] wrote:

From: Andy Adamson <[email protected]>

On a CB_RECALL the callback service thread flushes the inode using filemap_flush prior to scheduling the state manager thread to return the delegation. When pNFS is used and I/O has not yet gone to the data server servicing the inode, a LAYOUTGET can preceed the I/O. Unlike the async filemap_flush call, the LAYOUTGET must proceed to completion.

If the state manager starts to recover data while the inode flush is sending the LAYOUTGET, a deadlock occurs as the callback service thread holds the single callback session slot until the flushing is done which blocks the state manager thread, and the state manager thread has set the session draining bit which puts the inode flush LAYOUTGET RPC to sleep on the forechannel slot table waitq.

Separate the draining of the back channel from the draining of the fore channel by moving the NFS4_SESSION_DRAINING bit from session scope into the fore and back slot tables. Drain the back channel first allowing the LAYOUTGET call to proceed (and fail) so the callback service thread frees the callback slot. Then proceed with draining the forechannel.

---
 fs/nfs/callback_proc.c |  2 +-
 fs/nfs/callback_xdr.c  |  2 +-
 fs/nfs/nfs4proc.c      |  2 +-
 fs/nfs/nfs4session.c   |  6 ++++--
 fs/nfs/nfs4session.h   | 13 ++++++++-----
 fs/nfs/nfs4state.c     | 15 +++++++--------
 6 files changed, 22 insertions(+), 18 deletions(-)
fs/nfs/nfs4session.c +1 -0
@@ -395,12 +395,14 @@ int nfs4_setup_session_slot_tables(struct nfs4_session *ses)
 	/* Fore channel */
 	tbl = &ses->fc_slot_table;
 	tbl->session = ses;
+	tbl->slot_tbl_state = 0;

fs/nfs/nfs4session.c int nfs4_setup_session_slot_tables(struct nfs4_session *ses) · line 408

Quoted 10 lines
fs/nfs/nfs4session.c +1 -0
 	status = nfs4_realloc_slot_table(tbl, ses->fc_attrs.max_reqs, 1);
 	if (status) /* -ENOMEM */
 		return status;
 	/* Back channel */
 	tbl = &ses->bc_slot_table;
 	tbl->session = ses;
+	tbl->slot_tbl_state = 0;
 	status = nfs4_realloc_slot_table(tbl, ses->bc_attrs.max_reqs, 0);
 	if (status && tbl->slots == NULL)
 		/* Fore and back channel share a connection so get

Hi Andy,

The above hunk is buggy. nfs4_setup_session_slot_tables() is called from nfs4_proc_create_session(), so it must not reset the NFS4_SLOT_TBL_DRAINING flag.

I'm therefore removing the above 2 lines. I've applied the rest of the patch.

Thanks!
Trond

Signature · 6 lines
Trond Myklebust
Linux NFS client maintainer

NetApp
[email protected]
www.netapp.com
#2 lore raw
Quoted 42 lines · Trond
Quoted 41 lines · On Mon, 2013-05-20 at 14:13 -0400, [email protected] wrote:

From: Andy Adamson <[email protected]>

On a CB_RECALL the callback service thread flushes the inode using filemap_flush prior to scheduling the state manager thread to return the delegation. When pNFS is used and I/O has not yet gone to the data server servicing the inode, a LAYOUTGET can preceed the I/O. Unlike the async filemap_flush call, the LAYOUTGET must proceed to completion.

If the state manager starts to recover data while the inode flush is sending the LAYOUTGET, a deadlock occurs as the callback service thread holds the single callback session slot until the flushing is done which blocks the state manager thread, and the state manager thread has set the session draining bit which puts the inode flush LAYOUTGET RPC to sleep on the forechannel slot table waitq.

Separate the draining of the back channel from the draining of the fore channel by moving the NFS4_SESSION_DRAINING bit from session scope into the fore and back slot tables. Drain the back channel first allowing the LAYOUTGET call to proceed (and fail) so the callback service thread frees the callback slot. Then proceed with draining the forechannel.

---

fs/nfs/callback_proc.c |  2 +-
fs/nfs/callback_xdr.c  |  2 +-
fs/nfs/nfs4proc.c      |  2 +-
fs/nfs/nfs4session.c   |  6 ++++--
fs/nfs/nfs4session.h   | 13 ++++++++-----
fs/nfs/nfs4state.c     | 15 +++++++--------
6 files changed, 22 insertions(+), 18 deletions(-)
fs/nfs/nfs4session.c +1 -0
@@ -395,12 +395,14 @@ int nfs4_setup_session_slot_tables(struct nfs4_session *ses)
 	/* Fore channel */
 	tbl = &ses->fc_slot_table;
 	tbl->session = ses;
+	tbl->slot_tbl_state = 0;
 	status = nfs4_realloc_slot_table(tbl, ses->fc_attrs.max_reqs, 1);
 	if (status) /* -ENOMEM */
 		return status;
 	/* Back channel */
Quoted 9 lines

fs/nfs/nfs4session.c int nfs4_setup_session_slot_tables(struct nfs4_session *ses) · line 408

Quoted 6 lines
fs/nfs/nfs4session.c +1 -0
 	tbl = &ses->bc_slot_table;
 	tbl->session = ses;
+	tbl->slot_tbl_state = 0;
 	status = nfs4_realloc_slot_table(tbl, ses->bc_attrs.max_reqs, 0);
 	if (status && tbl->slots == NULL)
 		/* Fore and back channel share a connection so get

Hi Andy,

Hi Trond

Quoted 3 lines

The above hunk is buggy. nfs4_setup_session_slot_tables() is called from nfs4_proc_create_session(), so it must not reset the NFS4_SLOT_TBL_DRAINING flag.

Yes - I see. Good catch

Quoted 4 lines

I'm therefore removing the above 2 lines. I've applied the rest of the patch.

Thanks!

-->Andy

Quoted 10 lines

Thanks!
Trond

Signature · 6 lines
Trond Myklebust
Linux NFS client maintainer

NetApp
[email protected]
www.netapp.com