-
Bug
-
Resolution: Fixed
-
Medium
-
None
-
None
-
3
-
9223372036854775807
Description
A peer with both `lp_ping_mdh` and `lp_push_mdh` outstanding gets both unlinked by `lnet_peer_cancel_discovery()` during queue cleanup 1. `LNetMDUnlink()` only delivers the unlink event inline when `md_refcount == 0`; a busy MD defers it to the LND completion path. The inline event moves the peer off `ln_dc_expired`, so cleanup 2 stops waiting and cleanup 3 dequeues it via `lnet_peer_discovery_complete(lp, -ESHUTDOWN)` — which sets `LNET_PEER_REDISCOVER`, so `lnet_peer_is_uptodate()` stays false. When the deferred event finally arrives, `lnet_peer_queue_for_discovery()` finds `lp_dc_list` empty, takes a fresh ref, and puts the peer back on `ln_dc_request`, ultimately causing a reference leak.
Trigger conditions & timing
1. A peer has both `lp_ping_mdh` and `lp_push_mdh` outstanding;
2. `lnet_peer_cancel_discovery()` triggers unlink of both MDs during queue cleanup 1;
3. The unlink includes one inline if idle and one deferred event if busy;
4. The inline unlink completes cleanup 2 and cleanup 3, and drains the queue;
5. If the deferred unlink arrives afterward, it re-inserts the peer into `ln_dc_request`, and this reference will never be released, causing a leak.
Impact & errors
1. May trigger `LASSERT(list_empty(&the_lnet.ln_dc_request))`;
2. May report "Memory leaks detected" at module unload;
3. May trigger `LASSERT(md->md_handler != handler)` in `lnet_assert_handler_unused()`.
Once the erroneous timing window is hit, a reference leak occurs; the final error point depends on the concurrency timing of the peer re-inserting and the detection point.