ceph: handle race between cap reconnect and cap release

[karo-tx-linux.git] / fs / ceph / mds_client.c

diff --git a/fs/ceph/mds_client.c b/fs/ceph/mds_client.c
index 4a93d69abc8bcc43d0ec2d8473ff8dc65f9f46ee..6d953ab0ac06c08463c6134639d658a7ce90f304 100644 (file)
--- a/fs/ceph/mds_client.c
+++ b/fs/ceph/mds_client.c
@@ -444,6 +444,7 @@ static struct ceph_mds_session *register_session(struct ceph_mds_client *mdsc,
        INIT_LIST_HEAD(&s->s_waiting);
        INIT_LIST_HEAD(&s->s_unsafe);
        s->s_num_cap_releases = 0;
+       s->s_cap_reconnect = 0;
        s->s_cap_iterator = NULL;
        INIT_LIST_HEAD(&s->s_cap_releases);
        INIT_LIST_HEAD(&s->s_cap_releases_done);
@@ -1415,7 +1416,6 @@ static void discard_cap_releases(struct ceph_mds_client *mdsc,
        unsigned num;
 
        dout("discard_cap_releases mds%d\n", session->s_mds);
-       spin_lock(&session->s_cap_lock);
 
        /* zero out the in-progress message */
        msg = list_first_entry(&session->s_cap_releases,
@@ -1442,8 +1442,6 @@ static void discard_cap_releases(struct ceph_mds_client *mdsc,
                msg->front.iov_len = sizeof(*head);
                list_add(&msg->list_head, &session->s_cap_releases);
        }
-
-       spin_unlock(&session->s_cap_lock);
 }
 
 /*
@@ -2488,6 +2486,7 @@ static int encode_caps_cb(struct inode *inode, struct ceph_cap *cap,
        cap->seq = 0;        /* reset cap seq */
        cap->issue_seq = 0;  /* and issue_seq */
        cap->mseq = 0;       /* and migrate_seq */
+       cap->cap_gen = cap->session->s_cap_gen;
 
        if (recon_state->flock) {
                rec.v2.cap_id = cpu_to_le64(cap->cap_id);
@@ -2611,8 +2610,20 @@ static void send_mds_reconnect(struct ceph_mds_client *mdsc,
        dout("session %p state %s\n", session,
             session_state_name(session->s_state));
 
+       spin_lock(&session->s_gen_ttl_lock);
+       session->s_cap_gen++;
+       spin_unlock(&session->s_gen_ttl_lock);
+
+       spin_lock(&session->s_cap_lock);
+       /*
+        * notify __ceph_remove_cap() that we are composing cap reconnect.
+        * If a cap get released before being added to the cap reconnect,
+        * __ceph_remove_cap() should skip queuing cap release.
+        */
+       session->s_cap_reconnect = 1;
        /* drop old cap expires; we're about to reestablish that state */
        discard_cap_releases(mdsc, session);
+       spin_unlock(&session->s_cap_lock);
 
        /* traverse this session's caps */
        s_nr_caps = session->s_nr_caps;
@@ -2627,6 +2638,10 @@ static void send_mds_reconnect(struct ceph_mds_client *mdsc,
        if (err < 0)
                goto fail;
 
+       spin_lock(&session->s_cap_lock);
+       session->s_cap_reconnect = 0;
+       spin_unlock(&session->s_cap_lock);
+
        /*
         * snaprealms.  we provide mds with the ino, seq (version), and
         * parent for all of our realms.  If the mds has any newer info,