osd/PeeringState: do not trim pg log past last_update_ondisk

Trimming past last_update_ondisk would be really bad, e.g.,
a new interval change would cancel&redo a previous op, and if
we trim past last_update_ondisk, there could be potential
object inconsistencies as log merging won't necessarily be able
to find all divergent entries later (we lost track of the unfinished
op that should really be reverted).

Fixes: https://tracker.ceph.com/issues/44532
Signed-off-by: xie xingguo <xie.xingguo@zte.com.cn>
(cherry picked from commit eef1928e0153f09dc2522bb885b0ad9f5c65f4bd)

diff --git a/src/osd/PeeringState.cc b/src/osd/PeeringState.cc
index 10ffec25291013069699401b9355b261f93ea96f..5a49aaf4514e94c02f786dfd7949740f7f577bfa 100644 (file)
--- a/src/osd/PeeringState.cc
+++ b/src/osd/PeeringState.cc
@@ -4093,9 +4093,10 @@ void PeeringState::calc_trim_to_aggressive()
   size_t target = pl->get_target_pg_log_entries();
 
   // limit pg log trimming up to the can_rollback_to value
-  eversion_t limit = std::min(
+  eversion_t limit = std::min({
     pg_log.get_head(),
-    pg_log.get_can_rollback_to());
+    pg_log.get_can_rollback_to(),
+    last_update_ondisk});
   psdout(10) << __func__ << " limit = " << limit << dendl;
 
   if (limit != eversion_t() &&