net/ena: reuse zero length Rx descriptor
authorMichal Krawczyk <mk@semihalf.com>
Wed, 8 Apr 2020 08:29:19 +0000 (10:29 +0200)
committerFerruh Yigit <ferruh.yigit@intel.com>
Tue, 21 Apr 2020 11:57:07 +0000 (13:57 +0200)
Some ENA devices can pass to the driver descriptor with length 0. To
avoid extra allocation, the descriptor can be reused by simply putting
it back to the device.

Signed-off-by: Michal Krawczyk <mk@semihalf.com>
Reviewed-by: Igor Chauskin <igorch@amazon.com>
Reviewed-by: Guy Tzalik <gtzalik@amazon.com>
doc/guides/rel_notes/release_20_05.rst
drivers/net/ena/ena_ethdev.c

index 82ecb1e..fce4fb4 100644 (file)
@@ -89,6 +89,7 @@ New Features
   * Added support for large LLQ (Low-latency queue) headers.
   * Added Tx drops as new extended driver statistic.
   * Added support for accelerated LLQ mode.
+  * Handling of the 0 length descriptors on the Rx path.
 
 * **Updated Intel i40e driver.**
 
index 1a7cc68..156a3e4 100644 (file)
@@ -195,6 +195,8 @@ static struct rte_mbuf *ena_rx_mbuf(struct ena_ring *rx_ring,
                                    uint8_t offset);
 static uint16_t eth_ena_recv_pkts(void *rx_queue,
                                  struct rte_mbuf **rx_pkts, uint16_t nb_pkts);
+static int ena_add_single_rx_desc(struct ena_com_io_sq *io_sq,
+                                 struct rte_mbuf *mbuf, uint16_t id);
 static int ena_populate_rx_queue(struct ena_ring *rxq, unsigned int count);
 static void ena_init_rings(struct ena_adapter *adapter,
                           bool disable_meta_caching);
@@ -1414,6 +1416,24 @@ static int ena_rx_queue_setup(struct rte_eth_dev *dev,
        return 0;
 }
 
+static int ena_add_single_rx_desc(struct ena_com_io_sq *io_sq,
+                                 struct rte_mbuf *mbuf, uint16_t id)
+{
+       struct ena_com_buf ebuf;
+       int rc;
+
+       /* prepare physical address for DMA transaction */
+       ebuf.paddr = mbuf->buf_iova + RTE_PKTMBUF_HEADROOM;
+       ebuf.len = mbuf->buf_len - RTE_PKTMBUF_HEADROOM;
+
+       /* pass resource to device */
+       rc = ena_com_add_single_rx_desc(io_sq, &ebuf, id);
+       if (unlikely(rc != 0))
+               PMD_DRV_LOG(WARNING, "failed adding rx desc\n");
+
+       return rc;
+}
+
 static int ena_populate_rx_queue(struct ena_ring *rxq, unsigned int count)
 {
        unsigned int i;
@@ -1441,7 +1461,6 @@ static int ena_populate_rx_queue(struct ena_ring *rxq, unsigned int count)
 
        for (i = 0; i < count; i++) {
                struct rte_mbuf *mbuf = mbufs[i];
-               struct ena_com_buf ebuf;
                struct ena_rx_buffer *rx_info;
 
                if (likely((i + 4) < count))
@@ -1454,16 +1473,10 @@ static int ena_populate_rx_queue(struct ena_ring *rxq, unsigned int count)
 
                rx_info = &rxq->rx_buffer_info[req_id];
 
-               /* prepare physical address for DMA transaction */
-               ebuf.paddr = mbuf->buf_iova + RTE_PKTMBUF_HEADROOM;
-               ebuf.len = mbuf->buf_len - RTE_PKTMBUF_HEADROOM;
-               /* pass resource to device */
-               rc = ena_com_add_single_rx_desc(rxq->ena_com_io_sq,
-                                               &ebuf, req_id);
-               if (unlikely(rc)) {
-                       PMD_DRV_LOG(WARNING, "failed adding rx desc\n");
+               rc = ena_add_single_rx_desc(rxq->ena_com_io_sq, mbuf, req_id);
+               if (unlikely(rc != 0))
                        break;
-               }
+
                rx_info->mbuf = mbuf;
                next_to_use = ENA_IDX_NEXT_MASKED(next_to_use, rxq->size_mask);
        }
@@ -2079,6 +2092,7 @@ static struct rte_mbuf *ena_rx_mbuf(struct ena_ring *rx_ring,
        struct rte_mbuf *mbuf;
        struct rte_mbuf *mbuf_head;
        struct ena_rx_buffer *rx_info;
+       int rc;
        uint16_t ntc, len, req_id, buf = 0;
 
        if (unlikely(descs == 0))
@@ -2121,13 +2135,44 @@ static struct rte_mbuf *ena_rx_mbuf(struct ena_ring *rx_ring,
                rx_info = &rx_ring->rx_buffer_info[req_id];
                RTE_ASSERT(rx_info->mbuf != NULL);
 
-               /* Create an mbuf chain. */
-               mbuf->next = rx_info->mbuf;
-               mbuf = mbuf->next;
+               if (unlikely(len == 0)) {
+                       /*
+                        * Some devices can pass descriptor with the length 0.
+                        * To avoid confusion, the PMD is simply putting the
+                        * descriptor back, as it was never used. We'll avoid
+                        * mbuf allocation that way.
+                        */
+                       rc = ena_add_single_rx_desc(rx_ring->ena_com_io_sq,
+                               rx_info->mbuf, req_id);
+                       if (unlikely(rc != 0)) {
+                               /* Free the mbuf in case of an error. */
+                               rte_mbuf_raw_free(rx_info->mbuf);
+                       } else {
+                               /*
+                                * If there was no error, just exit the loop as
+                                * 0 length descriptor is always the last one.
+                                */
+                               break;
+                       }
+               } else {
+                       /* Create an mbuf chain. */
+                       mbuf->next = rx_info->mbuf;
+                       mbuf = mbuf->next;
 
-               ena_init_rx_mbuf(mbuf, len);
-               mbuf_head->pkt_len += len;
+                       ena_init_rx_mbuf(mbuf, len);
+                       mbuf_head->pkt_len += len;
+               }
 
+               /*
+                * Mark the descriptor as depleted and perform necessary
+                * cleanup.
+                * This code will execute in two cases:
+                *  1. Descriptor len was greater than 0 - normal situation.
+                *  2. Descriptor len was 0 and we failed to add the descriptor
+                *     to the device. In that situation, we should try to add
+                *     the mbuf again in the populate routine and mark the
+                *     descriptor as used up by the device.
+                */
                rx_info->mbuf = NULL;
                rx_ring->empty_rx_reqs[ntc] = req_id;
                ntc = ENA_IDX_NEXT_MASKED(ntc, rx_ring->size_mask);