mds: update CInode snapids (first & last) during replaywip-4248-snapid-journaling

A user reported on the mailing list that they were hitting the assert
assert(in->first == p->dnfirst ||
	       (in->is_multiversion() && in->first > p->dnfirst));
in EMetablob::replay. It turned out that it was on an unlink operation
(which moved the inode into a straydir), and the stray dentry had a
first of 4 while the inode had a first of 2. Investigation revealed
that we weren't updating "first" or "last" of the CInode for any CInodes that
existed in cache before the replay! (Non-existent inodes had the correct
values set from the p->dnfirst and p->dnlast on construction).

To fix this, we modify EMetaBlob::fullbit::update_inode() so that,
*if* old_inodes is non-empty, we set in->first as 1 past the last
entry's "last" value. This is good because old_inodes will have been
modified any time the inode was (assuming it was already multi-version).

Signed-off-by: Greg Farnum <greg@inktank.com>

1 files changed, 5 insertions, 0 deletions

diff --git a/src/mds/journal.cc b/src/mds/journal.cc
index 09072cef0d8..e19b5c5a179 100644
--- a/src/mds/journal.cc
+++ b/src/mds/journal.cc
@@ -516,6 +516,11 @@ void EMetaBlob::fullbit::update_inode(MDS *mds, CInode *in)
     in->symlink = symlink;
   }
   in->old_inodes = old_inodes;
+  if (!in->old_inodes.empty()) {
+    map<snapid_t, old_inode_t>::iterator p = in->old_inodes.end();
+    --p; // get the actual last entry
+    in->first = p->first + 1;
+  }
 }
 
 // EMetaBlob::remotebit