2 commits - src/paxos_lease.c src/resource.c

[David Teigland]

 src/paxos_lease.c |   22 ++++++++++++++++++++++
 src/resource.c    |    3 +++
 2 files changed, 25 insertions(+)

New commits:
commit fe03cafde177757ffe3c9e24d269c4ad5328f538
Author: David Teigland <teigland(a)redhat.com&gt;
Date:   Fri Feb 24 16:25:51 2012 -0600

    sanlock: fix problem in paxos
    
    The special case in the comment should be recognized
    and handled properly.
    
    Signed-off-by: David Teigland <teigland(a)redhat.com&gt;

diff --git a/src/paxos_lease.c b/src/paxos_lease.c
index 9fa62b1..d1243cb 100644
--- a/src/paxos_lease.c
+++ b/src/paxos_lease.c
@@ -1220,6 +1220,11 @@ int paxos_lease_acquire(struct task *task,
 	 * current host_id and generation?
 	 */
 
+	/* This next_lver assignment is based on the original cur_leader, not a
+	   re-reading of the leader here, i.e. we cannot just re-read the leader
+	   here, and make next_lver one more than that.  This is because another
+	   node may have made us the owner of next_lver as it is now. */
+
 	next_lver = cur_leader.lver + 1;
 
 	if (!our_dblock.mbal)
@@ -1271,6 +1276,23 @@ int paxos_lease_acquire(struct task *task,
 		goto out;
 	}
 
+	if (tmp_leader.lver > next_lver) {
+		/*
+		 * A case where this was observed: for next_lver 65 we abort1, and delay.
+		 * While sleeping, the lease v65 (which was acquired during our abort1) is
+		 * released and then reacquired as v66.  When we goto retry_ballot, our
+		 * next_lver is 65, but the current lver on disk is 66, causing us to
+		 * we fail in the larger1 check.)
+		 */
+		log_token(token, "paxos_acquire stale next_lver %llu now %llu owner %llu %llu
%llu",
+			  (unsigned long long)next_lver,
+			  (unsigned long long)tmp_leader.lver,
+			  (unsigned long long)tmp_leader.owner_id,
+			  (unsigned long long)tmp_leader.owner_generation,
+			  (unsigned long long)tmp_leader.timestamp);
+		goto restart;
+	}
+
 	error = run_ballot(task, token, cur_leader.num_hosts, next_lver, our_mbal,
 			   &dblock);
 


commit 74383a19e5a1254fc661630b2c6bc1d9b3b098d6
Author: David Teigland <teigland(a)redhat.com&gt;
Date:   Fri Feb 24 15:25:13 2012 -0600

    sanlock: fix missing close_disks
    
    which was leaking fd's
    
    Signed-off-by: David Teigland <teigland(a)redhat.com&gt;

diff --git a/src/resource.c b/src/resource.c
index 0df17a4..a76b88d 100644
--- a/src/resource.c
+++ b/src/resource.c
@@ -349,17 +349,20 @@ static int _release_token(struct task *task, struct token *token,
int opened,
 	if ((r->flags & R_SHARED) && !last_token) {
 		/* will release when final sh token is released */
 		log_token(token, "release_token more shared");
+		close_disks(token->disks, token->r.num_disks);
 		return SANLK_OK;
 	}
 
 	if (!last_token) {
 		/* should never happen */
 		log_errot(token, "release_token exclusive not last");
+		close_disks(token->disks, token->r.num_disks);
 		return SANLK_ERROR;
 	}
 
 	if (!lver) {
 		/* never acquired on disk so no need to release on disk */
+		close_disks(token->disks, token->r.num_disks);
 		rv = SANLK_OK;
 		goto out;
 	}