NFSv4.1 Fix a pNFS session draining deadlock
authorAndy Adamson <andros@netapp.com>
Mon, 20 May 2013 18:13:50 +0000 (14:13 -0400)
committerTrond Myklebust <Trond.Myklebust@netapp.com>
Mon, 20 May 2013 18:20:14 +0000 (14:20 -0400)
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.

Signed-off-by: Andy Adamson <andros@netapp.com>
Signed-off-by: Trond Myklebust <Trond.Myklebust@netapp.com>
fs/nfs/callback_proc.c
fs/nfs/callback_xdr.c
fs/nfs/nfs4proc.c
fs/nfs/nfs4session.c
fs/nfs/nfs4session.h
fs/nfs/nfs4state.c

index a13d26ede254357d46bc1f24c7c404555048987f..0bc27684ebfa338d0b77c295427c31472ec66049 100644 (file)
@@ -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
index 59461c957d9d7ed7204e09f9fad1aa0b154d8486..a35582c9d4440f8fe907192427b5c4f9a9a3b061 100644 (file)
@@ -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);
 }
 
index 8fbc100541154cbd31253dc261eb5eaafbe71e29..4e2fe714d5c292ae7776cb070c0c21fbfffdf7d1 100644 (file)
@@ -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__);
index ebda5f4a031b74d2c890d4c1e6cf932d869e8f2b..c4e225e4a9afc382b1eb689314ec2e7875cc1a45 100644 (file)
@@ -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;
index 6f3cb39386d4e400d509e6526924fa0b5233ed74..ff7d9f0f8a65179fbf9bc795c4b1cdbf1ee56bf1 100644 (file)
@@ -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,
index 300d17d85c0e03397d352746fc6fc775b39941e1..1fab140764c42756867f064a81b61903b8858a91 100644 (file)
@@ -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)