From: Tao Ning <ningtao@sangfor.com.cn>
Date: Tue, 24 Sep 2019 04:45:24 +0000 (+0800)
Subject: osd/PGLog: reset log.complete_to when recover obect failed
X-Git-Tag: v15.1.0~1405^2
X-Git-Url: http://git-server-git.apps.pok.os.sepia.ceph.com/?a=commitdiff_plain;h=refs%2Fpull%2F30533%2Fhead;p=ceph.git

osd/PGLog: reset log.complete_to when recover obect failed

1. Injection read fault, recover failed object is added to the missing set, log.complete_to == log.end()
2. Before the first object recovery, the new object writes, update info.last_update, because num_missing() != 0, so last_complete will not be updated
3. After the first object recovery, PeeringState::recover_got generates the error: ceph_assert(info.last_complete == info.last_update);

Fixed: https://tracker.ceph.com/issues/42019

Signed-off-by: Tao Ning <ningtao@sangfor.com.cn>
---

diff --git a/src/osd/PeeringState.cc b/src/osd/PeeringState.cc
index df942f3f9b4e..008fa52ba98c 100644
--- a/src/osd/PeeringState.cc
+++ b/src/osd/PeeringState.cc
@@ -3735,6 +3735,7 @@ void PeeringState::force_object_missing(
       peer_missing[peer].add(soid, version, eversion_t(), false);
     } else {
       pg_log.missing_add(soid, version, eversion_t());
+      pg_log.reset_complete_to(&info);
       pg_log.set_last_requested(0);
     }
   }