ceph: fix cap flush race reentrancy
authorSage Weil <sage@newdream.net>
Tue, 24 May 2011 18:46:31 +0000 (11:46 -0700)
committerSage Weil <sage@newdream.net>
Tue, 24 May 2011 18:52:12 +0000 (11:52 -0700)
In e9964c10 we change cap flushing to do a delicate dance because some
inodes on the cap_dirty list could be in a migrating state (got EXPORT but
not IMPORT) in which we couldn't actually flush and move from
dirty->flushing, breaking the while (!empty) { process first } loop
structure.  It worked for a single sync thread, but was not reentrant and
triggered infinite loops when multiple syncers came along.

Instead, move inodes with dirty to a separate cap_dirty_migrating list
when in the limbo export-but-no-import state, allowing us to go back to
the simple loop structure (which was reentrant).  This is cleaner and more
robust.

Audited the cap_dirty users and this looks fine:
list_empty(&ci->i_dirty_item) is still a reliable indicator of whether we
have dirty caps (which list we're on is irrelevant) and list_del_init()
calls still do the right thing.

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

index 591202bc96684647ffc8522985f58750971abe2d..1f72b00447c40e6383330496421a177b79150405 100644 (file)
@@ -2635,6 +2635,7 @@ static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
                              struct ceph_mds_session *session,
                              int *open_target_sessions)
 {
+       struct ceph_mds_client *mdsc = ceph_inode_to_client(inode)->mdsc;
        struct ceph_inode_info *ci = ceph_inode(inode);
        int mds = session->s_mds;
        unsigned mseq = le32_to_cpu(ex->migrate_seq);
@@ -2671,6 +2672,19 @@ static void handle_cap_export(struct inode *inode, struct ceph_mds_caps *ex,
                         * export targets, so that we get the matching IMPORT
                         */
                        *open_target_sessions = 1;
+
+                       /*
+                        * we can't flush dirty caps that we've seen the
+                        * EXPORT but no IMPORT for
+                        */
+                       spin_lock(&mdsc->cap_dirty_lock);
+                       if (!list_empty(&ci->i_dirty_item)) {
+                               dout(" moving %p to cap_dirty_migrating\n",
+                                    inode);
+                               list_move(&ci->i_dirty_item,
+                                         &mdsc->cap_dirty_migrating);
+                       }
+                       spin_unlock(&mdsc->cap_dirty_lock);
                }
                __ceph_remove_cap(cap);
        }
@@ -2708,6 +2722,13 @@ static void handle_cap_import(struct ceph_mds_client *mdsc,
                ci->i_cap_exporting_issued = 0;
                ci->i_cap_exporting_mseq = 0;
                ci->i_cap_exporting_mds = -1;
+
+               spin_lock(&mdsc->cap_dirty_lock);
+               if (!list_empty(&ci->i_dirty_item)) {
+                       dout(" moving %p back to cap_dirty\n", inode);
+                       list_move(&ci->i_dirty_item, &mdsc->cap_dirty);
+               }
+               spin_unlock(&mdsc->cap_dirty_lock);
        } else {
                dout("handle_cap_import inode %p ci %p mds%d mseq %d\n",
                     inode, ci, mds, mseq);
@@ -2911,38 +2932,16 @@ void ceph_check_delayed_caps(struct ceph_mds_client *mdsc)
  */
 void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
 {
-       struct ceph_inode_info *ci, *nci = NULL;
-       struct inode *inode, *ninode = NULL;
-       struct list_head *p, *n;
+       struct ceph_inode_info *ci;
+       struct inode *inode;
 
        dout("flush_dirty_caps\n");
        spin_lock(&mdsc->cap_dirty_lock);
-       list_for_each_safe(p, n, &mdsc->cap_dirty) {
-               if (nci) {
-                       ci = nci;
-                       inode = ninode;
-                       ci->i_ceph_flags &= ~CEPH_I_NOFLUSH;
-                       dout("flush_dirty_caps inode %p (was next inode)\n",
-                            inode);
-               } else {
-                       ci = list_entry(p, struct ceph_inode_info,
-                                       i_dirty_item);
-                       inode = igrab(&ci->vfs_inode);
-                       BUG_ON(!inode);
-                       dout("flush_dirty_caps inode %p\n", inode);
-               }
-               if (n != &mdsc->cap_dirty) {
-                       nci = list_entry(n, struct ceph_inode_info,
-                                        i_dirty_item);
-                       ninode = igrab(&nci->vfs_inode);
-                       BUG_ON(!ninode);
-                       nci->i_ceph_flags |= CEPH_I_NOFLUSH;
-                       dout("flush_dirty_caps next inode %p, noflush\n",
-                            ninode);
-               } else {
-                       nci = NULL;
-                       ninode = NULL;
-               }
+       while (!list_empty(&mdsc->cap_dirty)) {
+               ci = list_first_entry(&mdsc->cap_dirty, struct ceph_inode_info,
+                                     i_dirty_item);
+               inode = igrab(&ci->vfs_inode);
+               dout("flush_dirty_caps %p\n", inode);
                spin_unlock(&mdsc->cap_dirty_lock);
                if (inode) {
                        ceph_check_caps(ci, CHECK_CAPS_NODELAY|CHECK_CAPS_FLUSH,
@@ -2952,6 +2951,7 @@ void ceph_flush_dirty_caps(struct ceph_mds_client *mdsc)
                spin_lock(&mdsc->cap_dirty_lock);
        }
        spin_unlock(&mdsc->cap_dirty_lock);
+       dout("flush_dirty_caps done\n");
 }
 
 /*
index c12d2e9a0ec67a276068b8c236d4d4c73a39f4d5..79743d146be69ec8ce3f279a032e1a5459eb462c 100644 (file)
@@ -3004,6 +3004,7 @@ int ceph_mdsc_init(struct ceph_fs_client *fsc)
        spin_lock_init(&mdsc->snap_flush_lock);
        mdsc->cap_flush_seq = 0;
        INIT_LIST_HEAD(&mdsc->cap_dirty);
+       INIT_LIST_HEAD(&mdsc->cap_dirty_migrating);
        mdsc->num_cap_flushing = 0;
        spin_lock_init(&mdsc->cap_dirty_lock);
        init_waitqueue_head(&mdsc->cap_flushing_wq);
index 4e3a9cc0bba6f1f89a77e6355a8f72a52614d43f..7d8a0d662d56cfa73876d302f8de892a046a9972 100644 (file)
@@ -278,6 +278,7 @@ struct ceph_mds_client {
 
        u64               cap_flush_seq;
        struct list_head  cap_dirty;        /* inodes with dirty caps */
+       struct list_head  cap_dirty_migrating; /* ...that are migration... */
        int               num_cap_flushing; /* # caps we are flushing */
        spinlock_t        cap_dirty_lock;   /* protects above items */
        wait_queue_head_t cap_flushing_wq;