ceph: fix cap_snap and realm split
authorSage Weil <sage@newdream.net>
Thu, 16 Sep 2010 23:26:51 +0000 (16:26 -0700)
committerSage Weil <sage@newdream.net>
Thu, 16 Sep 2010 23:26:51 +0000 (16:26 -0700)
The cap_snap creation/queueing relies on both the current i_head_snapc
_and_ the i_snap_realm pointers being correct, so that the new cap_snap
can properly reference the old context and the new i_head_snapc can be
updated to reference the new snaprealm's context.  To fix this, we:

 - move inodes completely to the new (split) realm so that i_snap_realm
   is correct, and
 - generate the new snapc's _before_ queueing the cap_snaps in
   ceph_update_snap_trace().

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

index 50461b8c23a4e5db84719c67350cb7663a2fe0fc..efbc604001c8bbfa3b96eb107529f0f03256b1a1 100644 (file)
@@ -411,8 +411,8 @@ static int writepage_nounlock(struct page *page, struct writeback_control *wbc)
        if (i_size < page_off + len)
                len = i_size - page_off;
 
-       dout("writepage %p page %p index %lu on %llu~%u\n",
-            inode, page, page->index, page_off, len);
+       dout("writepage %p page %p index %lu on %llu~%u snapc %p\n",
+            inode, page, page->index, page_off, len, snapc);
 
        writeback_stat = atomic_long_inc_return(&client->writeback_count);
        if (writeback_stat >
index 9e836afba34115383408705ee3f899891cde935c..9e6eef14b7df4fef762ca95d5f5fbda9a6b5816d 100644 (file)
@@ -119,6 +119,7 @@ static struct ceph_snap_realm *ceph_create_snap_realm(
        INIT_LIST_HEAD(&realm->children);
        INIT_LIST_HEAD(&realm->child_item);
        INIT_LIST_HEAD(&realm->empty_item);
+       INIT_LIST_HEAD(&realm->dirty_item);
        INIT_LIST_HEAD(&realm->inodes_with_caps);
        spin_lock_init(&realm->inodes_with_caps_lock);
        __insert_snap_realm(&mdsc->snap_realms, realm);
@@ -604,6 +605,7 @@ int ceph_update_snap_trace(struct ceph_mds_client *mdsc,
        struct ceph_snap_realm *realm;
        int invalidate = 0;
        int err = -ENOMEM;
+       LIST_HEAD(dirty_realms);
 
        dout("update_snap_trace deletion=%d\n", deletion);
 more:
@@ -626,24 +628,6 @@ more:
                }
        }
 
-       if (le64_to_cpu(ri->seq) > realm->seq) {
-               dout("update_snap_trace updating %llx %p %lld -> %lld\n",
-                    realm->ino, realm, realm->seq, le64_to_cpu(ri->seq));
-               /*
-                * if the realm seq has changed, queue a cap_snap for every
-                * inode with open caps.  we do this _before_ we update
-                * the realm info so that we prepare for writeback under the
-                * _previous_ snap context.
-                *
-                * ...unless it's a snap deletion!
-                */
-               if (!deletion)
-                       queue_realm_cap_snaps(realm);
-       } else {
-               dout("update_snap_trace %llx %p seq %lld unchanged\n",
-                    realm->ino, realm, realm->seq);
-       }
-
        /* ensure the parent is correct */
        err = adjust_snap_realm_parent(mdsc, realm, le64_to_cpu(ri->parent));
        if (err < 0)
@@ -651,6 +635,8 @@ more:
        invalidate += err;
 
        if (le64_to_cpu(ri->seq) > realm->seq) {
+               dout("update_snap_trace updating %llx %p %lld -> %lld\n",
+                    realm->ino, realm, realm->seq, le64_to_cpu(ri->seq));
                /* update realm parameters, snap lists */
                realm->seq = le64_to_cpu(ri->seq);
                realm->created = le64_to_cpu(ri->created);
@@ -668,9 +654,17 @@ more:
                if (err < 0)
                        goto fail;
 
+               /* queue realm for cap_snap creation */
+               list_add(&realm->dirty_item, &dirty_realms);
+
                invalidate = 1;
        } else if (!realm->cached_context) {
+               dout("update_snap_trace %llx %p seq %lld new\n",
+                    realm->ino, realm, realm->seq);
                invalidate = 1;
+       } else {
+               dout("update_snap_trace %llx %p seq %lld unchanged\n",
+                    realm->ino, realm, realm->seq);
        }
 
        dout("done with %llx %p, invalidated=%d, %p %p\n", realm->ino,
@@ -683,6 +677,14 @@ more:
        if (invalidate)
                rebuild_snap_realms(realm);
 
+       /*
+        * queue cap snaps _after_ we've built the new snap contexts,
+        * so that i_head_snapc can be set appropriately.
+        */
+       list_for_each_entry(realm, &dirty_realms, dirty_item) {
+               queue_realm_cap_snaps(realm);
+       }
+
        __cleanup_empty_realms(mdsc);
        return 0;
 
@@ -816,6 +818,7 @@ void ceph_handle_snap(struct ceph_mds_client *mdsc,
                        };
                        struct inode *inode = ceph_find_inode(sb, vino);
                        struct ceph_inode_info *ci;
+                       struct ceph_snap_realm *oldrealm;
 
                        if (!inode)
                                continue;
@@ -841,18 +844,19 @@ void ceph_handle_snap(struct ceph_mds_client *mdsc,
                        dout(" will move %p to split realm %llx %p\n",
                             inode, realm->ino, realm);
                        /*
-                        * Remove the inode from the realm's inode
-                        * list, but don't add it to the new realm
-                        * yet.  We don't want the cap_snap to be
-                        * queued (again) by ceph_update_snap_trace()
-                        * below.  Queue it _now_, under the old context.
+                        * Move the inode to the new realm
                         */
                        spin_lock(&realm->inodes_with_caps_lock);
                        list_del_init(&ci->i_snap_realm_item);
+                       list_add(&ci->i_snap_realm_item,
+                                &realm->inodes_with_caps);
+                       oldrealm = ci->i_snap_realm;
+                       ci->i_snap_realm = realm;
                        spin_unlock(&realm->inodes_with_caps_lock);
                        spin_unlock(&inode->i_lock);
 
-                       ceph_queue_cap_snap(ci);
+                       ceph_get_snap_realm(mdsc, realm);
+                       ceph_put_snap_realm(mdsc, oldrealm);
 
                        iput(inode);
                        continue;
@@ -880,43 +884,9 @@ skip_inode:
        ceph_update_snap_trace(mdsc, p, e,
                               op == CEPH_SNAP_OP_DESTROY);
 
-       if (op == CEPH_SNAP_OP_SPLIT) {
-               /*
-                * ok, _now_ add the inodes into the new realm.
-                */
-               for (i = 0; i < num_split_inos; i++) {
-                       struct ceph_vino vino = {
-                               .ino = le64_to_cpu(split_inos[i]),
-                               .snap = CEPH_NOSNAP,
-                       };
-                       struct inode *inode = ceph_find_inode(sb, vino);
-                       struct ceph_inode_info *ci;
-
-                       if (!inode)
-                               continue;
-                       ci = ceph_inode(inode);
-                       spin_lock(&inode->i_lock);
-                       if (list_empty(&ci->i_snap_realm_item)) {
-                               struct ceph_snap_realm *oldrealm =
-                                       ci->i_snap_realm;
-
-                               dout(" moving %p to split realm %llx %p\n",
-                                    inode, realm->ino, realm);
-                               spin_lock(&realm->inodes_with_caps_lock);
-                               list_add(&ci->i_snap_realm_item,
-                                        &realm->inodes_with_caps);
-                               ci->i_snap_realm = realm;
-                               spin_unlock(&realm->inodes_with_caps_lock);
-                               ceph_get_snap_realm(mdsc, realm);
-                               ceph_put_snap_realm(mdsc, oldrealm);
-                       }
-                       spin_unlock(&inode->i_lock);
-                       iput(inode);
-               }
-
+       if (op == CEPH_SNAP_OP_SPLIT)
                /* we took a reference when we created the realm, above */
                ceph_put_snap_realm(mdsc, realm);
-       }
 
        __cleanup_empty_realms(mdsc);
 
index c33897ae5725e82ca269606b78d54214d8abf7af..c80bfbe27b05248f236e2aebd17a2bd21e0744e1 100644 (file)
@@ -690,6 +690,8 @@ struct ceph_snap_realm {
 
        struct list_head empty_item;     /* if i have ref==0 */
 
+       struct list_head dirty_item;     /* if realm needs new context */
+
        /* the current set of snaps for this realm */
        struct ceph_snap_context *cached_context;