r=johann,shadow
- fixes ptlrpcd blocking on very long reply unlink waiting. To do so new rpc phase introduced
RQ_PHASE_UNREGISTERING in which request stay until we have reply_in_callback() called by lnet
signaling that reply is unlinked. All requests in this state are skipped in processing by prlrcd
instead of waiting n * 300s on each of them. This allows ptlrpcd to process other rpcs in the set;
- make sure that inflight count is coherent with being present on sending or delay list. That is,
if we see inflight != 0, rpc must be on one of these lists. This is very helpful in
ptlrpc_invalidate_import() to show all rpcs still waiting after invalidating import;
- in ptlrpc_invalidate_import() wait maximal rq_deadline - now from all inflight rpcs instead of
obd_timeout which may be much longer. If calculated timeout is 0, obd_timeout is used. This fixes
the issue that rq_deadline - now > obd_timeout (very easy to see in logs) which led to inflight !=
0 assert because inflight rpcs timed out later than our wait period is finished;
- in ptlrpc_invalidate_import() wait forever for rpcs in UNREGISTERING phase. Check in assert for
inflight == 0 for wait timed out case if no rpcs in UNREGISTERING phase. Only those in
UNREGISTERING phase are allowed to stay longer than obd_timeout;
- added ptlrpc_move_rqphase() function. All phase changes go through it. Add debug_req() there to
track down all phase changes;
- conf_sanity.sh test_45 added to emulate very long reply unlink and also situation when
rq_deadline - now > obd_timeout;
- fixed using rq_timedout in debug_req();
- do not wait forever in ptlrpc_unregister_reply() for async case (using it from sets). Sync case
left unchanged.
#define OBD_FAIL_PTLRPC_IMP_DEACTIVE 0x50d
#define OBD_FAIL_PTLRPC_DUMP_LOG 0x50e
+#define OBD_FAIL_PTLRPC_LONG_UNLINK 0x50f
#define OBD_FAIL_OBD_PING_NET 0x600
#define OBD_FAIL_OBD_LOG_CANCEL_NET 0x601
git://git.whamcloud.com / fs / lustre-release.git / commitdiff