[Sheepdog] [PATCH 11/14] return SD_RES_NEW_NODE_VER during node membership change

[MORITA Kazutaka]

Sheepdog cannot process any I/O requests during the node membership
change to keep strong consistency, so we must return
SD_RES_NEW_NODE_VER instead of adding them to a waitlist queue.

Signed-off-by: MORITA Kazutaka <morita.kazutaka at lab.ntt.co.jp>
---
 sheep/group.c |   24 ++++++++++++++++++++++--
 1 files changed, 22 insertions(+), 2 deletions(-)

diff --git a/sheep/group.c b/sheep/group.c
index 28d211f..5d9409d 100644
--- a/sheep/group.c
+++ b/sheep/group.c
@@ -1227,8 +1227,13 @@ out:
 	cpg_event_free(cevent);
 	cpg_event_clear_running();
 
-	if (!list_empty(&sys->cpg_event_siblings) && !cpg_event_suspended())
-		start_cpg_event_work();
+	if (!list_empty(&sys->cpg_event_siblings)) {
+		if (cpg_event_joining())
+			/* io requests need to return SD_RES_NEW_NODE_VER */
+			start_cpg_event_work();
+		else if (!cpg_event_suspended())
+			start_cpg_event_work();
+	}
 }
 
 static int check_epoch(struct request *req)
@@ -1318,6 +1323,21 @@ void start_cpg_event_work(void)
 	if (cpg_event_joining()) {
 		if (!cpg_event_suspended())
 			panic("should not happen\n");
+
+		if (cevent->ctype == CPG_EVENT_REQUEST) {
+			struct request *req = container_of(cevent, struct request, cev);
+			if (is_io_request(req->rq.opcode) && req->rq.flags & SD_FLAG_CMD_DIRECT) {
+				list_del(&cevent->cpg_event_list);
+
+				req->rp.result = SD_RES_NEW_NODE_VER;
+
+				/* TODO: cleanup */
+				list_add_tail(&req->r_wlist, &sys->outstanding_req_list);
+				sys->nr_outstanding_io++;
+
+				req->work.done(&req->work, 0);
+			}
+		}
 		return;
 	}
 
-- 
1.5.6.5