migration/rdma: Fix race on source

Fix a race where the destination might try and send the source a
WRID_READY before the source has done a post-recv for it.

rdma_post_recv has to happen after the qp exists, and we're
OK since we've already called qemu_rdma_source_init that calls
qemu_alloc_qp.

This corresponds to:
https://bugzilla.redhat.com/show_bug.cgi?id=1285044

The race can be triggered by adding a few ms wait before this
post_recv_control (which was originally due to me turning on loads of
debug).

Signed-off-by: Dr. David Alan Gilbert <dgilbert@redhat.com>
Reviewed-by: Peter Xu <peterx@redhat.com>
Message-Id: <20170717110936.23314-2-dgilbert@redhat.com>
Signed-off-by: Juan Quintela <quintela@redhat.com>

diff --git a/migration/rdma.c b/migration/rdma.c
index c6bc607a03b2d9ed38b3943d7c43299f9909a1b4..6111e10c70b6178550dfc242ec08d6d05c6ac40f 100644 (file)
--- a/migration/rdma.c
+++ b/migration/rdma.c
@@ -2365,6 +2365,12 @@ static int qemu_rdma_connect(RDMAContext *rdma, Error **errp)
 
     caps_to_network(&cap);
 
+    ret = qemu_rdma_post_recv_control(rdma, RDMA_WRID_READY);
+    if (ret) {
+        ERROR(errp, "posting second control recv");
+        goto err_rdma_source_connect;
+    }
+
     ret = rdma_connect(rdma->cm_id, &conn_param);
     if (ret) {
         perror("rdma_connect");
@@ -2405,12 +2411,6 @@ static int qemu_rdma_connect(RDMAContext *rdma, Error **errp)
 
     rdma_ack_cm_event(cm_event);
 
-    ret = qemu_rdma_post_recv_control(rdma, RDMA_WRID_READY);
-    if (ret) {
-        ERROR(errp, "posting second control recv!");
-        goto err_rdma_source_connect;
-    }
-
     rdma->control_ready_expected = 1;
     rdma->nb_sent = 0;
     return 0;