Fix starvation of pending writes in CTDB queues

CTDB uses a queue to receive requests and send answers. It works
asynchronously using the tevent framework. However there was an issue
that gave priority to the receiving side so, when a request was
processed and the answer posted to the queue, if another incoming
request arrived, it was served before sending the previous answer.

This scenario could repeat for long periods of time if the frequency of
incoming requests was high enough.

Eventually, a small time gap between incoming request gave a chance to
process the pending output queue, sending many answers in a burst.

This patch makes sure that both queues (input and output) are processed
if the event contains the appropriate flag.

Signed-off-by: Xavi Hernandez <xhernandez@redhat.com>
Reviewed-by: Volker Lendecke <vl@samba.org>
Reviewed-by: Martin Schwenke <martin@meltin.net>
Autobuild-User(master): Volker Lendecke <vl@samba.org>
Autobuild-Date(master): Mon Jul  1 09:17:43 UTC 2024 on atb-devel-224

diff --git a/ctdb/common/ctdb_io.c b/ctdb/common/ctdb_io.c
index bf8bc73b77dc81575be1346412ebb454adb72df4..9ac9b84a7fb2dd5930874f0ceb0892509db08fd7 100644 (file)
--- a/ctdb/common/ctdb_io.c
+++ b/ctdb/common/ctdb_io.c
@@ -272,7 +272,7 @@ static void queue_dead(struct tevent_context *ev, struct tevent_immediate *im,
 /*
   called when an incoming connection is writeable
 */
-static void queue_io_write(struct ctdb_queue *queue)
+static bool queue_io_write(struct ctdb_queue *queue)
 {
        while (queue->out_queue) {
                struct ctdb_queue_pkt *pkt = queue->out_queue;
@@ -294,14 +294,14 @@ static void queue_io_write(struct ctdb_queue *queue)
                        queue->fd = -1;
                        tevent_schedule_immediate(queue->im, queue->ctdb->ev,
                                                  queue_dead, queue);
-                       return;
+                       return false;
                }
-               if (n <= 0) return;
+               if (n <= 0) return true;
                
                if (n != pkt->length) {
                        pkt->length -= n;
                        pkt->data += n;
-                       return;
+                       return true;
                }
 
                DLIST_REMOVE(queue->out_queue, pkt);
@@ -310,6 +310,8 @@ static void queue_io_write(struct ctdb_queue *queue)
        }
 
        TEVENT_FD_NOT_WRITEABLE(queue->fde);
+
+       return true;
 }
 
 /*
@@ -320,10 +322,13 @@ static void queue_io_handler(struct tevent_context *ev, struct tevent_fd *fde,
 {
        struct ctdb_queue *queue = talloc_get_type(private_data, struct ctdb_queue);
 
+       if (flags & TEVENT_FD_WRITE) {
+               if (!queue_io_write(queue)) {
+                       return;
+               }
+       }
        if (flags & TEVENT_FD_READ) {
                queue_io_read(queue);
-       } else {
-               queue_io_write(queue);
        }
 }