Improve implementation of the Nagle algorithm for SCTP:

Don't delay the final fragment of a fragmented user message.

Approved by: re
MFC after: 4 weeks

1 files changed, 12 insertions, 6 deletions

diff --git a/sys/netinet/sctp_output.c b/sys/netinet/sctp_output.c
index 81c9e70..d91f898 100644
--- a/sys/netinet/sctp_output.c
+++ b/sys/netinet/sctp_output.c
@@ -9821,19 +9821,22 @@ sctp_chunk_output(struct sctp_inpcb *inp,
 	unsigned int burst_cnt = 0;
 	struct timeval now;
 	int now_filled = 0;
-	int nagle_on = 0;
+	int nagle_on;
 	int frag_point = sctp_get_frag_point(stcb, &stcb->asoc);
 	int un_sent = 0;
 	int fr_done;
 	unsigned int tot_frs = 0;
 
 	asoc = &stcb->asoc;
+	/* The Nagle algorithm is only applied when handling a send call. */
 	if (from_where == SCTP_OUTPUT_FROM_USR_SEND) {
 		if (sctp_is_feature_on(inp, SCTP_PCB_FLAGS_NODELAY)) {
 			nagle_on = 0;
 		} else {
 			nagle_on = 1;
 		}
+	} else {
+		nagle_on = 0;
 	}
 	SCTP_TCB_LOCK_ASSERT(stcb);
 
@@ -10007,15 +10010,18 @@ sctp_chunk_output(struct sctp_inpcb *inp,
 			}
 		}
 		if (nagle_on) {
-			/*-
-			 * When nagle is on, we look at how much is un_sent, then
-			 * if its smaller than an MTU and we have data in
-			 * flight we stop.
+			/*
+			 * When the Nagle algorithm is used, look at how
+			 * much is unsent, then if its smaller than an MTU
+			 * and we have data in flight we stop, except if we
+			 * are handling a fragmented user message.
 			 */
 			un_sent = ((stcb->asoc.total_output_queue_size - stcb->asoc.total_flight) +
 			    (stcb->asoc.stream_queue_cnt * sizeof(struct sctp_data_chunk)));
 			if ((un_sent < (int)(stcb->asoc.smallest_mtu - SCTP_MIN_OVERHEAD)) &&
-			    (stcb->asoc.total_flight > 0)) {
+			    (stcb->asoc.total_flight > 0) &&
+			    ((stcb->asoc.locked_on_sending == NULL) ||
+			    sctp_is_feature_on(inp, SCTP_PCB_FLAGS_EXPLICIT_EOR))) {
 				break;
 			}
 		}