osd/PeeringState: recover_got - add special handler for empty log

If we use a small osd_min[max]_pg_log_entries setting, say, 10, and
do splitting, a child pg may only be able to keep some divergent log
entries in the list. Any follow-up peering cycle on that child pg
would then rewind all the divergent log entries and lead to an empty
log list, which as a result could reliably fire the

  ceph_assert(info.last_complete == info.last_update)

assert if we start to recover any of the missing divergent objects
later.

Fixes: https://tracker.ceph.com/issues/41816
Signed-off-by: xie xingguo <xie.xingguo@zte.com.cn>

diff --git a/src/osd/PeeringState.cc b/src/osd/PeeringState.cc
index 616cdd72b2ff83c7d638b7b0cf512ed326577eda..a180db783159fccb48fb2ef23c1cf99c27db2d75 100644 (file)
--- a/src/osd/PeeringState.cc
+++ b/src/osd/PeeringState.cc
@@ -3659,7 +3659,10 @@ void PeeringState::recover_got(
 
   psdout(10) << "got missing " << oid << " v " << v << dendl;
   pg_log.recover_got(oid, v, info);
-  if (pg_log.get_log().complete_to != pg_log.get_log().log.end()) {
+  if (pg_log.get_log().log.empty()) {
+    psdout(10) << "last_complete now " << info.last_complete
+               << " while log is empty" << dendl;
+  } else if (pg_log.get_log().complete_to != pg_log.get_log().log.end()) {
     psdout(10) << "last_complete now " << info.last_complete
               << " log.complete_to " << pg_log.get_log().complete_to->version
               << dendl;