]> err.no Git - linux-2.6/commitdiff
ocfs2: make dlm recovery finalization 2 stage
authorKurt Hackel <kurt.hackel@oracle.com>
Mon, 1 May 2006 18:11:13 +0000 (11:11 -0700)
committerMark Fasheh <mark.fasheh@oracle.com>
Mon, 26 Jun 2006 21:42:58 +0000 (14:42 -0700)
Makes it easier for the recovery process to deal with node death.

Signed-off-by: Kurt Hackel <kurt.hackel@oracle.com>
Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>
fs/ocfs2/dlm/dlmcommon.h
fs/ocfs2/dlm/dlmrecovery.c

index 9e052445b0b5d6abc228d5e8731f5d3a2ad3c431..78eccd0951e4ea847c4ffd9062eb8f9537b8a77e 100644 (file)
@@ -71,7 +71,8 @@ static inline int dlm_is_recovery_lock(const char *lock_name, int name_len)
        return 0;
 }
 
-#define DLM_RECO_STATE_ACTIVE  0x0001
+#define DLM_RECO_STATE_ACTIVE    0x0001
+#define DLM_RECO_STATE_FINALIZE  0x0002
 
 struct dlm_recovery_ctxt
 {
@@ -633,7 +634,8 @@ struct dlm_finalize_reco
 {
        u8 node_idx;
        u8 dead_node;
-       __be16 pad1;
+       u8 flags;
+       u8 pad1;
        __be32 pad2;
 };
 
index 6ee8b324712949be9ad93dbdfb373aab4ff7a8be..19123ce8b3064a7e9aa45b231d7653107354933b 100644 (file)
@@ -134,12 +134,18 @@ static inline void dlm_set_reco_master(struct dlm_ctxt *dlm,
        dlm->reco.new_master = master;
 }
 
-static inline void dlm_reset_recovery(struct dlm_ctxt *dlm)
+static inline void __dlm_reset_recovery(struct dlm_ctxt *dlm)
 {
-       spin_lock(&dlm->spinlock);
+       assert_spin_locked(&dlm->spinlock);
        clear_bit(dlm->reco.dead_node, dlm->recovery_map);
        dlm_set_reco_dead_node(dlm, O2NM_INVALID_NODE_NUM);
        dlm_set_reco_master(dlm, O2NM_INVALID_NODE_NUM);
+}
+
+static inline void dlm_reset_recovery(struct dlm_ctxt *dlm)
+{
+       spin_lock(&dlm->spinlock);
+       __dlm_reset_recovery(dlm);
        spin_unlock(&dlm->spinlock);
 }
 
@@ -2074,6 +2080,20 @@ static void __dlm_hb_node_down(struct dlm_ctxt *dlm, int idx)
 {
        assert_spin_locked(&dlm->spinlock);
 
+       if (dlm->reco.new_master == idx) {
+               mlog(0, "%s: recovery master %d just died\n",
+                    dlm->name, idx);
+               if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) {
+                       /* finalize1 was reached, so it is safe to clear
+                        * the new_master and dead_node.  that recovery
+                        * is complete. */
+                       mlog(0, "%s: dead master %d had reached "
+                            "finalize1 state, clearing\n", dlm->name, idx);
+                       dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE;
+                       __dlm_reset_recovery(dlm);
+               }
+       }
+
        /* check to see if the node is already considered dead */
        if (!test_bit(idx, dlm->live_nodes_map)) {
                mlog(0, "for domain %s, node %d is already dead. "
@@ -2364,6 +2384,14 @@ retry:
                         * another ENOMEM */
                        msleep(100);
                        goto retry;
+               } else if (ret == EAGAIN) {
+                       mlog(0, "%s: trying to start recovery of node "
+                            "%u, but node %u is waiting for last recovery "
+                            "to complete, backoff for a bit\n", dlm->name,
+                            dead_node, nodenum);
+                       /* TODO Look into replacing msleep with cond_resched() */
+                       msleep(100);
+                       goto retry;
                }
        }
 
@@ -2379,6 +2407,17 @@ int dlm_begin_reco_handler(struct o2net_msg *msg, u32 len, void *data)
        if (!dlm_grab(dlm))
                return 0;
 
+       spin_lock(&dlm->spinlock);
+       if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) {
+               mlog(0, "%s: node %u wants to recover node %u (%u:%u) "
+                    "but this node is in finalize state, waiting on finalize2\n",
+                    dlm->name, br->node_idx, br->dead_node,
+                    dlm->reco.dead_node, dlm->reco.new_master);
+               spin_unlock(&dlm->spinlock);
+               return EAGAIN;
+       }
+       spin_unlock(&dlm->spinlock);
+
        mlog(0, "%s: node %u wants to recover node %u (%u:%u)\n",
             dlm->name, br->node_idx, br->dead_node,
             dlm->reco.dead_node, dlm->reco.new_master);
@@ -2432,6 +2471,7 @@ int dlm_begin_reco_handler(struct o2net_msg *msg, u32 len, void *data)
        return 0;
 }
 
+#define DLM_FINALIZE_STAGE2  0x01
 static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm)
 {
        int ret = 0;
@@ -2439,25 +2479,31 @@ static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm)
        struct dlm_node_iter iter;
        int nodenum;
        int status;
+       int stage = 1;
 
-       mlog(0, "finishing recovery for node %s:%u\n",
-            dlm->name, dlm->reco.dead_node);
+       mlog(0, "finishing recovery for node %s:%u, "
+            "stage %d\n", dlm->name, dlm->reco.dead_node, stage);
 
        spin_lock(&dlm->spinlock);
        dlm_node_iter_init(dlm->domain_map, &iter);
        spin_unlock(&dlm->spinlock);
 
+stage2:
        memset(&fr, 0, sizeof(fr));
        fr.node_idx = dlm->node_num;
        fr.dead_node = dlm->reco.dead_node;
+       if (stage == 2)
+               fr.flags |= DLM_FINALIZE_STAGE2;
 
        while ((nodenum = dlm_node_iter_next(&iter)) >= 0) {
                if (nodenum == dlm->node_num)
                        continue;
                ret = o2net_send_message(DLM_FINALIZE_RECO_MSG, dlm->key,
                                         &fr, sizeof(fr), nodenum, &status);
-               if (ret >= 0) {
+               if (ret >= 0)
                        ret = status;
+               if (ret < 0) {
+                       mlog_errno(ret);
                        if (dlm_is_host_down(ret)) {
                                /* this has no effect on this recovery 
                                 * session, so set the status to zero to 
@@ -2466,12 +2512,15 @@ static int dlm_send_finalize_reco_message(struct dlm_ctxt *dlm)
                                     "node finished recovery.\n", nodenum);
                                ret = 0;
                        }
-               }
-               if (ret < 0) {
-                       mlog_errno(ret);
                        break;
                }
        }
+       if (stage == 1) {
+               /* reset the node_iter back to the top and send finalize2 */
+               iter.curnode = -1;
+               stage = 2;
+               goto stage2;
+       }
 
        return ret;
 }
@@ -2480,15 +2529,19 @@ int dlm_finalize_reco_handler(struct o2net_msg *msg, u32 len, void *data)
 {
        struct dlm_ctxt *dlm = data;
        struct dlm_finalize_reco *fr = (struct dlm_finalize_reco *)msg->buf;
+       int stage = 1;
 
        /* ok to return 0, domain has gone away */
        if (!dlm_grab(dlm))
                return 0;
 
-       mlog(0, "%s: node %u finalizing recovery of node %u (%u:%u)\n",
-            dlm->name, fr->node_idx, fr->dead_node,
-            dlm->reco.dead_node, dlm->reco.new_master);
+       if (fr->flags & DLM_FINALIZE_STAGE2)
+               stage = 2;
 
+       mlog(0, "%s: node %u finalizing recovery stage%d of "
+            "node %u (%u:%u)\n", dlm->name, fr->node_idx, stage,
+            fr->dead_node, dlm->reco.dead_node, dlm->reco.new_master);
        spin_lock(&dlm->spinlock);
 
        if (dlm->reco.new_master != fr->node_idx) {
@@ -2504,13 +2557,38 @@ int dlm_finalize_reco_handler(struct o2net_msg *msg, u32 len, void *data)
                BUG();
        }
 
-       dlm_finish_local_lockres_recovery(dlm, fr->dead_node, fr->node_idx);
-
-       spin_unlock(&dlm->spinlock);
-
-       dlm_reset_recovery(dlm);
+       switch (stage) {
+               case 1:
+                       dlm_finish_local_lockres_recovery(dlm, fr->dead_node, fr->node_idx);
+                       if (dlm->reco.state & DLM_RECO_STATE_FINALIZE) {
+                               mlog(ML_ERROR, "%s: received finalize1 from "
+                                    "new master %u for dead node %u, but "
+                                    "this node has already received it!\n",
+                                    dlm->name, fr->node_idx, fr->dead_node);
+                               dlm_print_reco_node_status(dlm);
+                               BUG();
+                       }
+                       dlm->reco.state |= DLM_RECO_STATE_FINALIZE;
+                       spin_unlock(&dlm->spinlock);
+                       break;
+               case 2:
+                       if (!(dlm->reco.state & DLM_RECO_STATE_FINALIZE)) {
+                               mlog(ML_ERROR, "%s: received finalize2 from "
+                                    "new master %u for dead node %u, but "
+                                    "this node did not have finalize1!\n",
+                                    dlm->name, fr->node_idx, fr->dead_node);
+                               dlm_print_reco_node_status(dlm);
+                               BUG();
+                       }
+                       dlm->reco.state &= ~DLM_RECO_STATE_FINALIZE;
+                       spin_unlock(&dlm->spinlock);
+                       dlm_reset_recovery(dlm);
+                       dlm_kick_recovery_thread(dlm);
+                       break;
+               default:
+                       BUG();
+       }
 
-       dlm_kick_recovery_thread(dlm);
        mlog(0, "%s: recovery done, reco master was %u, dead now %u, master now %u\n",
             dlm->name, fr->node_idx, dlm->reco.dead_node, dlm->reco.new_master);