xen-netback: batch copies for multiple to-guest rx packets

Instead of flushing the copy ops when an packet is complete, complete
packets when their copy ops are done.  This improves performance by
reducing the number of grant copy hypercalls.

Latency is still limited by the relatively small size of the copy
batch.

Signed-off-by: David Vrabel <david.vrabel@citrix.com>
[re-based]
Signed-off-by: Paul Durrant <paul.durrant@citrix.com>

diff --git a/drivers/net/xen-netback/common.h b/drivers/net/xen-netback/common.h
index adef4822ed6c3a933b1771ddc6a5f0d29715cf13..5d406033a09dc6121478f15c7a2c96bdff92850e 100644 (file)
--- a/drivers/net/xen-netback/common.h
+++ b/drivers/net/xen-netback/common.h
@@ -132,6 +132,7 @@ struct xenvif_copy_state {
        struct gnttab_copy op[COPY_BATCH_SIZE];
        RING_IDX idx[COPY_BATCH_SIZE];
        unsigned int num;
+       struct sk_buff_head *completed;
 };
 
 struct xenvif_queue { /* Per-queue data for xenvif */

diff --git a/drivers/net/xen-netback/rx.c b/drivers/net/xen-netback/rx.c
index 03a7c0eb0b1861d1629891fad580b0ba46188683..3e6333e0abd96160d5dce5455446a0b4aa919f32 100644 (file)
--- a/drivers/net/xen-netback/rx.c
+++ b/drivers/net/xen-netback/rx.c
@@ -132,6 +132,7 @@ static void xenvif_rx_queue_drop_expired(struct xenvif_queue *queue)
 static void xenvif_rx_copy_flush(struct xenvif_queue *queue)
 {
        unsigned int i;
+       int notify;
 
        gnttab_batch_copy(queue->rx_copy.op, queue->rx_copy.num);
 
@@ -153,6 +154,13 @@ static void xenvif_rx_copy_flush(struct xenvif_queue *queue)
        }
 
        queue->rx_copy.num = 0;
+
+       /* Push responses for all completed packets. */
+       RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->rx, notify);
+       if (notify)
+               notify_remote_via_irq(queue->rx_irq);
+
+       __skb_queue_purge(queue->rx_copy.completed);
 }
 
 static void xenvif_rx_copy_add(struct xenvif_queue *queue,
@@ -278,18 +286,10 @@ static void xenvif_rx_next_skb(struct xenvif_queue *queue,
 static void xenvif_rx_complete(struct xenvif_queue *queue,
                               struct xenvif_pkt_state *pkt)
 {
-       int notify;
-
-       /* Complete any outstanding copy ops for this skb. */
-       xenvif_rx_copy_flush(queue);
-
-       /* Push responses and notify. */
+       /* All responses are ready to be pushed. */
        queue->rx.rsp_prod_pvt = queue->rx.req_cons;
-       RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&queue->rx, notify);
-       if (notify)
-               notify_remote_via_irq(queue->rx_irq);
 
-       dev_kfree_skb(pkt->skb);
+       __skb_queue_tail(queue->rx_copy.completed, pkt->skb);
 }
 
 static void xenvif_rx_next_chunk(struct xenvif_queue *queue,
@@ -428,13 +428,20 @@ void xenvif_rx_skb(struct xenvif_queue *queue)
 
 void xenvif_rx_action(struct xenvif_queue *queue)
 {
+       struct sk_buff_head completed_skbs;
        unsigned int work_done = 0;
 
+       __skb_queue_head_init(&completed_skbs);
+       queue->rx_copy.completed = &completed_skbs;
+
        while (xenvif_rx_ring_slots_available(queue) &&
               work_done < RX_BATCH_SIZE) {
                xenvif_rx_skb(queue);
                work_done++;
        }
+
+       /* Flush any pending copies and complete all skbs. */
+       xenvif_rx_copy_flush(queue);
 }
 
 static bool xenvif_rx_queue_stalled(struct xenvif_queue *queue)