Ceph » CephFS

I should note: the error itself is very concerning because the only way for dump_cache to fail is if it's operating on a file, which it isn't in this case.

This could be memory corruption on the stack but I don't immediately see the reason why. I've seen this appear in two jobs for the last test run. I'm seeing if it's still reproducible in the next run.

Sorry, the bug was introduced by my commit:

commit 7b9eae62c8c654ff82684451c222257d2c93be64 (HEAD)
Author: Yan, Zheng <zyan@redhat.com>
Date:   Wed Aug 2 17:26:56 2017 +0800

    mds: track snap inodes through sorted map

    Current mds track both head inodes and snap inodes through unsorted
    map. The unsorted map makes finding snap inode that follows a given
    snapid difficult. Currnt MDCache::pick_inode_snap() use snap set to
    guess snap inode's last. The method isn't reliable because snap set
    may change after creating the snap inode. For example:

    MDS cows inode[2,head] with snap set[5,6], which results inode[2,6]
    and inode[7,head].

    Later mds wants to find snap inode that follows snapid 2. But the
    snap set become [5], mds can't find snap inode [2,5].

    Signed-off-by: "Yan, Zheng" <zyan@redhat.com>

below incremental patch can fix it

diff --git a/src/mds/MDCache.cc b/src/mds/MDCache.cc
index 2fbccccd6e..0bb1cda9b9 100644
--- a/src/mds/MDCache.cc
+++ b/src/mds/MDCache.cc
@@ -11979,6 +11979,7 @@ int MDCache::dump_cache(const char *fn, Formatter *f,
     if (r < 0)
       goto out;
   }
+  r = 0;

  out:
   if (f) {

https://github.com/ceph/ceph/pull/16778/commits/f519fca9dd958121a289676edf5175fb8be9894f