diff --git a/lustre/ptlrpc/client.c b/lustre/ptlrpc/client.c
index 7e1e6cd0106ea2864dfd2a1bf32aa6f52d560433..affe53fc6012d24b2c4d2678fadcbd5fea412636 100644
--- a/lustre/ptlrpc/client.c
+++ b/lustre/ptlrpc/client.c
@@ -1063,7 +1063,7 @@ int ptlrpc_check_set(struct ptlrpc_request_set *set)
if (ptlrpc_client_recv_or_unlink(req) ||
ptlrpc_client_bulk_active(req))
continue;
-
+
/* Turn repl fail_loc off to prevent it from looping
* forever. */
if (OBD_FAIL_CHECK(OBD_FAIL_PTLRPC_LONG_REPL_UNLINK)) {
@@ -1091,7 +1091,7 @@ int ptlrpc_check_set(struct ptlrpc_request_set *set)
/* Note that this also will start async reply unlink */
if (req->rq_net_err && !req->rq_timedout) {
ptlrpc_expire_one_request(req, 1);
-
+
/* Check if we still need to wait for unlink. */
if (ptlrpc_client_recv_or_unlink(req) ||
ptlrpc_client_bulk_active(req))
@@ -1124,8 +1124,11 @@ int ptlrpc_check_set(struct ptlrpc_request_set *set)
continue;
spin_lock(&imp->imp_lock);
-
if (ptlrpc_import_delay_req(imp, req, &status)){
+ /* put on delay list - only if we wait
+ * recovery finished - before send */
+ list_del_init(&req->rq_list);
+ list_add_tail(&req->rq_list, &imp->imp_delayed_list);
spin_unlock(&imp->imp_lock);
continue;
}
@@ -1157,8 +1160,6 @@ int ptlrpc_check_set(struct ptlrpc_request_set *set)
/* This is re-sending anyways,
* let's mark req as resend. */
req->rq_resend = 1;
- lustre_msg_add_flags(req->rq_reqmsg,
- MSG_RESENT);
if (req->rq_bulk) {
__u64 old_xid;
@@ -1209,17 +1210,8 @@ int ptlrpc_check_set(struct ptlrpc_request_set *set)
spin_unlock(&req->rq_lock);
req->rq_status = after_reply(req);
- if (req->rq_resend) {
- /* Add this req to the delayed list so
- it can be errored if the import is
- evicted after recovery. */
- spin_lock(&imp->imp_lock);
- list_del_init(&req->rq_list);
- list_add_tail(&req->rq_list,
- &imp->imp_delayed_list);
- spin_unlock(&imp->imp_lock);
+ if (req->rq_resend)
continue;
- }
/* If there is no bulk associated with this request,
* then we're done and should let the interpreter
@@ -1967,8 +1959,6 @@ restart:
}
if (req->rq_resend) {
- lustre_msg_add_flags(req->rq_reqmsg, MSG_RESENT);
-
if (req->rq_bulk != NULL) {
ptlrpc_unregister_bulk(req, 0);
@@ -2046,18 +2036,18 @@ restart:
}
/* Resend if we need to */
- if (req->rq_resend) {
+ if (req->rq_resend||req->rq_timedout) {
/* ...unless we were specifically told otherwise. */
if (req->rq_no_resend)
GOTO(out, rc = -ETIMEDOUT);
spin_lock(&imp->imp_lock);
+ /* we can have rq_timeout on dlm fake import which not support
+ * recovery - but me need resend request on this import instead
+ * of return error */
+ req->rq_resend = 1;
goto restart;
}
- if (req->rq_timedout) { /* non-recoverable timeout */
- GOTO(out, rc = -ETIMEDOUT);
- }
-
if (!ptlrpc_client_replied(req)) {
/* How can this be? -eeb */
DEBUG_REQ(D_ERROR, req, "!rq_replied: ");
diff --git a/lustre/ptlrpc/niobuf.c b/lustre/ptlrpc/niobuf.c
index 613ef6f8060f945f80449c276dc80545073db254..7bb6103c42ddf7f0aa2120dc5dc7f2663612755d 100644
--- a/lustre/ptlrpc/niobuf.c
+++ b/lustre/ptlrpc/niobuf.c
@@ -523,6 +523,9 @@ int ptl_send_rpc(struct ptlrpc_request *request, int noreply)
lustre_msghdr_set_flags(request->rq_reqmsg,
request->rq_import->imp_msghdr_flags);
+ if (request->rq_resend)
+ lustre_msg_add_flags(request->rq_reqmsg, MSG_RESENT);
+
if (!noreply) {
LASSERT (request->rq_replen != 0);
if (request->rq_repbuf == NULL)
diff --git a/lustre/ptlrpc/recov_thread.c b/lustre/ptlrpc/recov_thread.c
index 20ad67402ef0a613028632c3c5c03166bb9c30b7..5c12b63be81b3c3bd32af2fef0d361d669888249 100644
--- a/lustre/ptlrpc/recov_thread.c
+++ b/lustre/ptlrpc/recov_thread.c
@@ -223,7 +223,7 @@ static int llcd_send(struct llog_canceld_ctxt *llcd)
GOTO(exit, rc = 0);
lcm = llcd->llcd_lcm;
-
+
/*
* Check if we're in exit stage. Do not send llcd in
* this case.
@@ -271,6 +271,11 @@ static int llcd_send(struct llog_canceld_ctxt *llcd)
ptlrpc_at_set_req_timeout(req);
req->rq_interpret_reply = llcd_interpret;
req->rq_async_args.pointer_arg[0] = llcd;
+
+ /* llog cancels will be replayed after reconnect so this will do twice
+ * first from replay llog, second for resended rpc */
+ req->rq_no_delay = req->rq_no_resend = 1;
+
rc = ptlrpc_set_add_new_req(&lcm->lcm_pc, req);
if (rc) {
ptlrpc_req_finished(req);
diff --git a/lustre/tests/replay-single.sh b/lustre/tests/replay-single.sh
index d5b0273da69815bbc33d1b2893e7fd2f12bb54f1..15bcc98b87b69eca50a4299dd78997d3d6c88241 100755
--- a/lustre/tests/replay-single.sh
+++ b/lustre/tests/replay-single.sh
@@ -1332,33 +1332,6 @@ test_59() {
}
run_test 59 "test log_commit_thread vs filter_destroy race"
-# bug 17323
-test_59b() {
- do_facet mds "lctl set_param debug=+rpctrace"
- mkdir -p $DIR/$tdir
- createmany -o $DIR/$tdir/$tfile-%d 2000
- sync
-#define OBD_FAIL_OBD_LOG_CANCEL_REP 0x606
- do_facet mds "lctl set_param fail_loc=0x606"
- unlinkmany $DIR/$tdir/$tfile-%d 2000
-
- # make sure that all llcds left ost and nothing left cached
- sync
- sleep 10
- do_facet mds "lctl set_param fail_loc=0x0"
-
- # sleep 2 obd_timeouts from ost to make sure that we get resents.
- local timeout=$(do_facet ost1 lctl get_param -n timeout)
- timeout=$((timeout * 2))
- log "Sleep $timeout"
- sleep $timeout
- do_facet mds $LCTL dk | grep -q "RESENT cancel req"
- local res=$?
- rmdir $DIR/$tdir
- return $res
-}
-run_test 59b "resent handle in llog_origin_handle_cancel"
-
# race between add unlink llog vs cat log init in post_recovery (only for b1_6)
# bug 12086: should no oops and No ctxt error for this test
test_60() {