ceph: only send one flushsnap per cap_snap per mds session
authorSage Weil <sage@newdream.net>
Fri, 17 Sep 2010 15:03:08 +0000 (08:03 -0700)
committerSage Weil <sage@newdream.net>
Fri, 17 Sep 2010 15:03:08 +0000 (08:03 -0700)
Sending multiple flushsnap messages is problematic because we ignore
the response if the tid doesn't match, and the server may only respond to
each one once.  It's also a waste.

So, skip cap_snaps that are already on the flushing list, unless the caller
tells us to resend (because we are reconnecting).

Signed-off-by: Sage Weil <sage@newdream.net>
fs/ceph/caps.c
fs/ceph/snap.c
fs/ceph/super.h

index 9fbe9019155c196e53dbc82732cfe620b5f39596..b01c316a8148700693ce4a4212166a288a162a3d 100644 (file)
@@ -1195,10 +1195,14 @@ static int __send_cap(struct ceph_mds_client *mdsc, struct ceph_cap *cap,
  * asynchronously back to the MDS once sync writes complete and dirty
  * data is written out.
  *
+ * Unless @again is true, skip cap_snaps that were already sent to
+ * the MDS (i.e., during this session).
+ *
  * Called under i_lock.  Takes s_mutex as needed.
  */
 void __ceph_flush_snaps(struct ceph_inode_info *ci,
-                       struct ceph_mds_session **psession)
+                       struct ceph_mds_session **psession,
+                       int again)
                __releases(ci->vfs_inode->i_lock)
                __acquires(ci->vfs_inode->i_lock)
 {
@@ -1240,6 +1244,13 @@ retry:
                        dout("no auth cap (migrating?), doing nothing\n");
                        goto out;
                }
+
+               /* only flush each capsnap once */
+               if (!again && !list_empty(&capsnap->flushing_item)) {
+                       dout("already flushed %p, skipping\n", capsnap);
+                       continue;
+               }
+
                mds = ci->i_auth_cap->session->s_mds;
                mseq = ci->i_auth_cap->mseq;
 
@@ -1314,7 +1325,7 @@ static void ceph_flush_snaps(struct ceph_inode_info *ci)
        struct inode *inode = &ci->vfs_inode;
 
        spin_lock(&inode->i_lock);
-       __ceph_flush_snaps(ci, NULL);
+       __ceph_flush_snaps(ci, NULL, 0);
        spin_unlock(&inode->i_lock);
 }
 
@@ -1477,7 +1488,7 @@ void ceph_check_caps(struct ceph_inode_info *ci, int flags,
 
        /* flush snaps first time around only */
        if (!list_empty(&ci->i_cap_snaps))
-               __ceph_flush_snaps(ci, &session);
+               __ceph_flush_snaps(ci, &session, 0);
        goto retry_locked;
 retry:
        spin_lock(&inode->i_lock);
@@ -1894,7 +1905,7 @@ static void kick_flushing_capsnaps(struct ceph_mds_client *mdsc,
                if (cap && cap->session == session) {
                        dout("kick_flushing_caps %p cap %p capsnap %p\n", inode,
                             cap, capsnap);
-                       __ceph_flush_snaps(ci, &session);
+                       __ceph_flush_snaps(ci, &session, 1);
                } else {
                        pr_err("%p auth cap %p not mds%d ???\n", inode,
                               cap, session->s_mds);
index 9e6eef14b7df4fef762ca95d5f5fbda9a6b5816d..190b6c4a6f2b91aace658f8de7664f75cc0bdcc4 100644 (file)
@@ -717,7 +717,7 @@ static void flush_snaps(struct ceph_mds_client *mdsc)
                igrab(inode);
                spin_unlock(&mdsc->snap_flush_lock);
                spin_lock(&inode->i_lock);
-               __ceph_flush_snaps(ci, &session);
+               __ceph_flush_snaps(ci, &session, 0);
                spin_unlock(&inode->i_lock);
                iput(inode);
                spin_lock(&mdsc->snap_flush_lock);
index c80bfbe27b05248f236e2aebd17a2bd21e0744e1..b87638e84c4bc266e9b325f12e7fb98fc78e8986 100644 (file)
@@ -828,7 +828,8 @@ extern void ceph_put_cap_refs(struct ceph_inode_info *ci, int had);
 extern void ceph_put_wrbuffer_cap_refs(struct ceph_inode_info *ci, int nr,
                                       struct ceph_snap_context *snapc);
 extern void __ceph_flush_snaps(struct ceph_inode_info *ci,
-                              struct ceph_mds_session **psession);
+                              struct ceph_mds_session **psession,
+                              int again);
 extern void ceph_check_caps(struct ceph_inode_info *ci, int flags,
                            struct ceph_mds_session *session);
 extern void ceph_check_delayed_caps(struct ceph_mds_client *mdsc);