There are multiple problems which may cause this symptom as follows:
#476 - incomplete replica is left, if memory shortage happens in remove_replica_entity() (already merged into gfarm-2.5.6 in r6568)
#478 - incomplete replica is left, if a back channel is disconnected just before a reply from gfmd to gfsd (r6664 for 2.5, r7341 for 2.5.7)
#479 - incomplete replica is left due to a chicken-or-egg problem to maintain a reference count of struct peer(r6694 and r6695 for 2.5, r7343 for 2.5.7)
#480 - incomplete replica is left due to missing decrement of a reference count of struct peer (r6672 for 2.5, r7342 for 2.5.7)
#481 - incomplete replica is left at disconnection of a back channel, if the disconnection happens after successful GFS_PROTO_REPLICATION_REQUEST but before GFM_PROTO_REPLICATION_RESULT (r6799 for 2.5, r7345 for 2.5.7)
#482 - incomplete replica is left due to missing reply about GFS_PROTO_REPLICATION_REQUEST from gfsd to gfmd (r6809 for 2.5, r7346 for 2.5.7)
#490 - incomplete replica is left, when virtual memory space for mmap(2) is exhausted (not a bug, just memory shortage)
#500 - incomplete replica is left, if malloc(3) fails (r6909 for 2.5, r7363 for 2.5.7)
These are superset of #486 ([pullup-2.5.7.3] #483 - descriptor leak about gfsd back channel due to problem of replication handling)
NOTE
This is same issue with #408 - incomplete replica is left,
made to keep track changes on the 2.5.7 release branch.
Problem caused by this ticket
fix to #479 -> #602 - [pullup-2.5.7.4] #601 - gfmd crashes when gfmd-initiated replication is invoked against a filesystem node which is down
There are multiple problems which may cause this symptom as follows:
These are superset of #486 ([pullup-2.5.7.3] #483 - descriptor leak about gfsd back channel due to problem of replication handling)
NOTE
This is same issue with #408 - incomplete replica is left,
made to keep track changes on the 2.5.7 release branch.
Problem caused by this ticket
Reported by: n-soda
Original Ticket: gfarm/tickets/485