]> git.apps.os.sepia.ceph.com Git - ceph.git/commit
osd/OSDMap: CRUSH_TUNABLES5 added in jewel, not kraken 23227/head
authorSage Weil <sage@redhat.com>
Wed, 25 Jul 2018 01:30:54 +0000 (20:30 -0500)
committerSage Weil <sage@redhat.com>
Wed, 25 Jul 2018 01:32:01 +0000 (20:32 -0500)
commit82e80366d842ada257b2562ab84c25ba35c1aa72
treeb6b6e0ad9cd509a20f35af6f6c51a2eb7b442de7
parent27728c38c9b9d5cb6f93e06bac8b2e1b723aac7c
osd/OSDMap: CRUSH_TUNABLES5 added in jewel, not kraken

In 0ceb5c0f3694725c7043ea99b077816da2ca44a7 I screwed up the version that
CRUSH_TUANBLES5 (chooseleaf_stable) was added: it comes from
043a7378ec4e828f2568ea3906d9076848161ccd, v10.0.3, which leads up to
jewel.

This, in turn, means that when jewel tunables are set we do not actually
encode the jewel tunables.  At least, luminous doesn't; jewel uses a
different code path to decode how to encode the OSDMap.

This, in turn, leads to crc errors:

- luminous mons running
- require_osd_release is jewel
- jewel tunables are set.  mon in-memory crushmap gets chooseleaf_stable=1
- osds are luminous, but don't see the new field in the crushmap
- upgrade finishes, and require_osd_Release becomes luminous
- the full osdmap encodes the chooseleaf_stable field, finally, and it
  is 1
- the osds try to encode the same thing, but their crushmap has 0.
-> crc mismatch

This is a regression from 0ceb5c0f3694725c7043ea99b077816da2ca44a7,
backported to luminous in 686b054f48c50859bd5e8bad93007bf9d7d7899c in
12.2.6.

Fixes: 686b054f48c50859bd5e8bad93007bf9d7d7899c
Fixes: http://tracker.ceph.com/issues/25057
Signed-off-by: Sage Weil <sage@redhat.com>
(cherry picked from commit 7e9d24d0602abefa4cca245f5f86dcd4b412b97a)
src/osd/OSDMap.cc