From 8bef9239ee1a42eb37d3f83bacf6a75f019c028d Mon Sep 17 00:00:00 2001 From: Sage Weil Date: Tue, 14 Sep 2010 15:45:44 -0700 Subject: ceph: correctly set 'follows' in flushsnap messages The 'follows' should match the seq for the snap context for the given snap cap, which is the context under which we have been dirtying and writing data and metadata. The snapshot that _contains_ those updates thus _follows_ that context's seq #. Signed-off-by: Sage Weil --- fs/ceph/snap.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'fs/ceph/snap.c') diff --git a/fs/ceph/snap.c b/fs/ceph/snap.c index 4868b9dcac5a..9e836afba341 100644 --- a/fs/ceph/snap.c +++ b/fs/ceph/snap.c @@ -467,7 +467,7 @@ void ceph_queue_cap_snap(struct ceph_inode_info *ci) INIT_LIST_HEAD(&capsnap->ci_item); INIT_LIST_HEAD(&capsnap->flushing_item); - capsnap->follows = snapc->seq - 1; + capsnap->follows = snapc->seq; capsnap->issued = __ceph_caps_issued(ci, NULL); capsnap->dirty = dirty; -- cgit v1.2.3 From ae00d4f37f4df56821331deb1028748110dd6dc9 Mon Sep 17 00:00:00 2001 From: Sage Weil Date: Thu, 16 Sep 2010 16:26:51 -0700 Subject: ceph: fix cap_snap and realm split 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 --- fs/ceph/snap.c | 88 +++++++++++++++++++--------------------------------------- 1 file changed, 29 insertions(+), 59 deletions(-) (limited to 'fs/ceph/snap.c') diff --git a/fs/ceph/snap.c b/fs/ceph/snap.c index 9e836afba341..9e6eef14b7df 100644 --- a/fs/ceph/snap.c +++ b/fs/ceph/snap.c @@ -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); -- cgit v1.2.3 From e835124c2be289515b918f2688ced4249e2de566 Mon Sep 17 00:00:00 2001 From: Sage Weil Date: Fri, 17 Sep 2010 08:03:08 -0700 Subject: ceph: only send one flushsnap per cap_snap per mds session 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 --- fs/ceph/snap.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'fs/ceph/snap.c') diff --git a/fs/ceph/snap.c b/fs/ceph/snap.c index 9e6eef14b7df..190b6c4a6f2b 100644 --- a/fs/ceph/snap.c +++ b/fs/ceph/snap.c @@ -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); -- cgit v1.2.3