ixgbe: remove simple scalar scattered Rx method
authorVlad Zolotarov <vladz@cloudius-systems.com>
Wed, 29 Apr 2015 08:38:16 +0000 (11:38 +0300)
committerThomas Monjalon <thomas.monjalon@6wind.com>
Thu, 7 May 2015 17:07:28 +0000 (19:07 +0200)
Kill ixgbe_recv_scattered_pkts() - use ixgbe_recv_pkts_lro_single_alloc()
instead.

Work against HW queues in LRO and scattered Rx cases is exactly the same.
Therefore we may drop the inferior callback.

This patch also changes the sw_rsc_ring allocation in the
ixgbe_dev_rx_queue_setup() to always allocate sw_rsc_ring instead of
explicitly allocating it in all possible cases when it may be needed:
LRO and/or scattered Rx.

This will only impose sizeof(void*) * IXGBE_MAX_RING_DESC = 32KB overhead
per Rx queue as a price for a much simpler code, which seems reasonable.

Signed-off-by: Vlad Zolotarov <vladz@cloudius-systems.com>
Acked-by: Konstantin Ananyev <konstantin.ananyev@intel.com>
lib/librte_pmd_ixgbe/ixgbe_ethdev.c
lib/librte_pmd_ixgbe/ixgbe_ethdev.h
lib/librte_pmd_ixgbe/ixgbe_rxtx.c

index aec1de9..5f9a1cf 100644 (file)
@@ -986,7 +986,7 @@ eth_ixgbevf_dev_init(struct rte_eth_dev *eth_dev)
         * RX function */
        if (rte_eal_process_type() != RTE_PROC_PRIMARY){
                if (eth_dev->data->scattered_rx)
-                       eth_dev->rx_pkt_burst = ixgbe_recv_scattered_pkts;
+                       eth_dev->rx_pkt_burst = ixgbe_recv_pkts_lro_single_alloc;
                return 0;
        }
 
index 5b90115..419ea5d 100644 (file)
@@ -352,9 +352,6 @@ void ixgbevf_dev_rxtx_start(struct rte_eth_dev *dev);
 uint16_t ixgbe_recv_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
                uint16_t nb_pkts);
 
-uint16_t ixgbe_recv_scattered_pkts(void *rx_queue,
-               struct rte_mbuf **rx_pkts, uint16_t nb_pkts);
-
 uint16_t ixgbe_recv_pkts_lro_single_alloc(void *rx_queue,
                struct rte_mbuf **rx_pkts, uint16_t nb_pkts);
 uint16_t ixgbe_recv_pkts_lro_bulk_alloc(void *rx_queue,
index 5ce00ab..e3cf4fc 100644 (file)
@@ -1722,239 +1722,6 @@ ixgbe_recv_pkts_lro_bulk_alloc(void *rx_queue, struct rte_mbuf **rx_pkts,
        return ixgbe_recv_pkts_lro(rx_queue, rx_pkts, nb_pkts, true);
 }
 
-uint16_t
-ixgbe_recv_scattered_pkts(void *rx_queue, struct rte_mbuf **rx_pkts,
-                         uint16_t nb_pkts)
-{
-       struct ixgbe_rx_queue *rxq;
-       volatile union ixgbe_adv_rx_desc *rx_ring;
-       volatile union ixgbe_adv_rx_desc *rxdp;
-       struct ixgbe_rx_entry *sw_ring;
-       struct ixgbe_rx_entry *rxe;
-       struct rte_mbuf *first_seg;
-       struct rte_mbuf *last_seg;
-       struct rte_mbuf *rxm;
-       struct rte_mbuf *nmb;
-       union ixgbe_adv_rx_desc rxd;
-       uint64_t dma; /* Physical address of mbuf data buffer */
-       uint32_t staterr;
-       uint16_t rx_id;
-       uint16_t nb_rx;
-       uint16_t nb_hold;
-       uint16_t data_len;
-
-       nb_rx = 0;
-       nb_hold = 0;
-       rxq = rx_queue;
-       rx_id = rxq->rx_tail;
-       rx_ring = rxq->rx_ring;
-       sw_ring = rxq->sw_ring;
-
-       /*
-        * Retrieve RX context of current packet, if any.
-        */
-       first_seg = rxq->pkt_first_seg;
-       last_seg = rxq->pkt_last_seg;
-
-       while (nb_rx < nb_pkts) {
-       next_desc:
-               /*
-                * The order of operations here is important as the DD status
-                * bit must not be read after any other descriptor fields.
-                * rx_ring and rxdp are pointing to volatile data so the order
-                * of accesses cannot be reordered by the compiler. If they were
-                * not volatile, they could be reordered which could lead to
-                * using invalid descriptor fields when read from rxd.
-                */
-               rxdp = &rx_ring[rx_id];
-               staterr = rxdp->wb.upper.status_error;
-               if (! (staterr & rte_cpu_to_le_32(IXGBE_RXDADV_STAT_DD)))
-                       break;
-               rxd = *rxdp;
-
-               /*
-                * Descriptor done.
-                *
-                * Allocate a new mbuf to replenish the RX ring descriptor.
-                * If the allocation fails:
-                *    - arrange for that RX descriptor to be the first one
-                *      being parsed the next time the receive function is
-                *      invoked [on the same queue].
-                *
-                *    - Stop parsing the RX ring and return immediately.
-                *
-                * This policy does not drop the packet received in the RX
-                * descriptor for which the allocation of a new mbuf failed.
-                * Thus, it allows that packet to be later retrieved if
-                * mbuf have been freed in the mean time.
-                * As a side effect, holding RX descriptors instead of
-                * systematically giving them back to the NIC may lead to
-                * RX ring exhaustion situations.
-                * However, the NIC can gracefully prevent such situations
-                * to happen by sending specific "back-pressure" flow control
-                * frames to its peer(s).
-                */
-               PMD_RX_LOG(DEBUG, "port_id=%u queue_id=%u rx_id=%u "
-                          "staterr=0x%x data_len=%u",
-                          (unsigned) rxq->port_id, (unsigned) rxq->queue_id,
-                          (unsigned) rx_id, (unsigned) staterr,
-                          (unsigned) rte_le_to_cpu_16(rxd.wb.upper.length));
-
-               nmb = rte_rxmbuf_alloc(rxq->mb_pool);
-               if (nmb == NULL) {
-                       PMD_RX_LOG(DEBUG, "RX mbuf alloc failed port_id=%u "
-                                  "queue_id=%u", (unsigned) rxq->port_id,
-                                  (unsigned) rxq->queue_id);
-                       rte_eth_devices[rxq->port_id].data->rx_mbuf_alloc_failed++;
-                       break;
-               }
-
-               nb_hold++;
-               rxe = &sw_ring[rx_id];
-               rx_id++;
-               if (rx_id == rxq->nb_rx_desc)
-                       rx_id = 0;
-
-               /* Prefetch next mbuf while processing current one. */
-               rte_ixgbe_prefetch(sw_ring[rx_id].mbuf);
-
-               /*
-                * When next RX descriptor is on a cache-line boundary,
-                * prefetch the next 4 RX descriptors and the next 8 pointers
-                * to mbufs.
-                */
-               if ((rx_id & 0x3) == 0) {
-                       rte_ixgbe_prefetch(&rx_ring[rx_id]);
-                       rte_ixgbe_prefetch(&sw_ring[rx_id]);
-               }
-
-               /*
-                * Update RX descriptor with the physical address of the new
-                * data buffer of the new allocated mbuf.
-                */
-               rxm = rxe->mbuf;
-               rxe->mbuf = nmb;
-               dma = rte_cpu_to_le_64(RTE_MBUF_DATA_DMA_ADDR_DEFAULT(nmb));
-               rxdp->read.hdr_addr = dma;
-               rxdp->read.pkt_addr = dma;
-
-               /*
-                * Set data length & data buffer address of mbuf.
-                */
-               data_len = rte_le_to_cpu_16(rxd.wb.upper.length);
-               rxm->data_len = data_len;
-               rxm->data_off = RTE_PKTMBUF_HEADROOM;
-
-               /*
-                * If this is the first buffer of the received packet,
-                * set the pointer to the first mbuf of the packet and
-                * initialize its context.
-                * Otherwise, update the total length and the number of segments
-                * of the current scattered packet, and update the pointer to
-                * the last mbuf of the current packet.
-                */
-               if (first_seg == NULL) {
-                       first_seg = rxm;
-                       first_seg->pkt_len = data_len;
-                       first_seg->nb_segs = 1;
-               } else {
-                       first_seg->pkt_len = (uint16_t)(first_seg->pkt_len
-                                       + data_len);
-                       first_seg->nb_segs++;
-                       last_seg->next = rxm;
-               }
-
-               /*
-                * If this is not the last buffer of the received packet,
-                * update the pointer to the last mbuf of the current scattered
-                * packet and continue to parse the RX ring.
-                */
-               if (! (staterr & IXGBE_RXDADV_STAT_EOP)) {
-                       last_seg = rxm;
-                       goto next_desc;
-               }
-
-               /*
-                * This is the last buffer of the received packet.
-                * If the CRC is not stripped by the hardware:
-                *   - Subtract the CRC length from the total packet length.
-                *   - If the last buffer only contains the whole CRC or a part
-                *     of it, free the mbuf associated to the last buffer.
-                *     If part of the CRC is also contained in the previous
-                *     mbuf, subtract the length of that CRC part from the
-                *     data length of the previous mbuf.
-                */
-               rxm->next = NULL;
-               if (unlikely(rxq->crc_len > 0)) {
-                       first_seg->pkt_len -= ETHER_CRC_LEN;
-                       if (data_len <= ETHER_CRC_LEN) {
-                               rte_pktmbuf_free_seg(rxm);
-                               first_seg->nb_segs--;
-                               last_seg->data_len = (uint16_t)
-                                       (last_seg->data_len -
-                                        (ETHER_CRC_LEN - data_len));
-                               last_seg->next = NULL;
-                       } else
-                               rxm->data_len =
-                                       (uint16_t) (data_len - ETHER_CRC_LEN);
-               }
-
-               /* Initialize the first mbuf of the returned packet */
-               ixgbe_fill_cluster_head_buf(first_seg, &rxd, rxq->port_id,
-                                           staterr);
-
-               /* Prefetch data of first segment, if configured to do so. */
-               rte_packet_prefetch((char *)first_seg->buf_addr +
-                       first_seg->data_off);
-
-               /*
-                * Store the mbuf address into the next entry of the array
-                * of returned packets.
-                */
-               rx_pkts[nb_rx++] = first_seg;
-
-               /*
-                * Setup receipt context for a new packet.
-                */
-               first_seg = NULL;
-       }
-
-       /*
-        * Record index of the next RX descriptor to probe.
-        */
-       rxq->rx_tail = rx_id;
-
-       /*
-        * Save receive context.
-        */
-       rxq->pkt_first_seg = first_seg;
-       rxq->pkt_last_seg = last_seg;
-
-       /*
-        * If the number of free RX descriptors is greater than the RX free
-        * threshold of the queue, advance the Receive Descriptor Tail (RDT)
-        * register.
-        * Update the RDT with the value of the last processed RX descriptor
-        * minus 1, to guarantee that the RDT register is never equal to the
-        * RDH register, which creates a "full" ring situtation from the
-        * hardware point of view...
-        */
-       nb_hold = (uint16_t) (nb_hold + rxq->nb_rx_hold);
-       if (nb_hold > rxq->rx_free_thresh) {
-               PMD_RX_LOG(DEBUG, "port_id=%u queue_id=%u rx_tail=%u "
-                          "nb_hold=%u nb_rx=%u",
-                          (unsigned) rxq->port_id, (unsigned) rxq->queue_id,
-                          (unsigned) rx_id, (unsigned) nb_hold,
-                          (unsigned) nb_rx);
-               rx_id = (uint16_t) ((rx_id == 0) ?
-                                    (rxq->nb_rx_desc - 1) : (rx_id - 1));
-               IXGBE_PCI_REG_WRITE(rxq->rdt_reg_addr, rx_id);
-               nb_hold = 0;
-       }
-       rxq->nb_rx_hold = nb_hold;
-       return (nb_rx);
-}
-
 /*********************************************************************
  *
  *  Queue management functions
@@ -2505,14 +2272,6 @@ ixgbe_dev_rx_queue_setup(struct rte_eth_dev *dev,
        uint16_t len;
        struct ixgbe_adapter *adapter =
                (struct ixgbe_adapter *)dev->data->dev_private;
-       struct rte_eth_dev_info dev_info = { 0 };
-       struct rte_eth_rxmode *dev_rx_mode = &dev->data->dev_conf.rxmode;
-       bool rsc_requested = false;
-
-       dev->dev_ops->dev_infos_get(dev, &dev_info);
-       if ((dev_info.rx_offload_capa & DEV_RX_OFFLOAD_TCP_LRO) &&
-           dev_rx_mode->enable_lro)
-               rsc_requested = true;
 
        PMD_INIT_FUNC_TRACE();
        hw = IXGBE_DEV_PRIVATE_TO_HW(dev->data->dev_private);
@@ -2623,17 +2382,22 @@ ixgbe_dev_rx_queue_setup(struct rte_eth_dev *dev,
                return (-ENOMEM);
        }
 
-       if (rsc_requested) {
-               rxq->sw_sc_ring =
-                       rte_zmalloc_socket("rxq->sw_sc_ring",
-                                          sizeof(struct ixgbe_scattered_rx_entry) * len,
-                                          RTE_CACHE_LINE_SIZE, socket_id);
-               if (!rxq->sw_sc_ring) {
-                       ixgbe_rx_queue_release(rxq);
-                       return (-ENOMEM);
-               }
-       } else
-               rxq->sw_sc_ring = NULL;
+       /*
+        * Always allocate even if it's not going to be needed in order to
+        * simplify the code.
+        *
+        * This ring is used in LRO and Scattered Rx cases and Scattered Rx may
+        * be requested in ixgbe_dev_rx_init(), which is called later from
+        * dev_start() flow.
+        */
+       rxq->sw_sc_ring =
+               rte_zmalloc_socket("rxq->sw_sc_ring",
+                                  sizeof(struct ixgbe_scattered_rx_entry) * len,
+                                  RTE_CACHE_LINE_SIZE, socket_id);
+       if (!rxq->sw_sc_ring) {
+               ixgbe_rx_queue_release(rxq);
+               return (-ENOMEM);
+       }
 
        PMD_INIT_LOG(DEBUG, "sw_ring=%p sw_sc_ring=%p hw_ring=%p "
                            "dma_addr=0x%"PRIx64,
@@ -4017,12 +3781,13 @@ void ixgbe_set_rx_function(struct rte_eth_dev *dev)
 
                        dev->rx_pkt_burst = ixgbe_recv_scattered_pkts_vec;
                } else {
-                       PMD_INIT_LOG(DEBUG, "Using Regualr (non-vector) "
+                       PMD_INIT_LOG(DEBUG, "Using Regualr (non-vector, "
+                                           "single allocation) "
                                            "Scattered Rx callback "
                                            "(port=%d).",
                                     dev->data->port_id);
 
-                       dev->rx_pkt_burst = ixgbe_recv_scattered_pkts;
+                       dev->rx_pkt_burst = ixgbe_recv_pkts_lro_single_alloc;
                }
        /*
         * Below we set "simple" callbacks according to port/queues parameters.
@@ -4851,7 +4616,8 @@ ixgbevf_dev_rx_init(struct rte_eth_dev *dev)
                                        ixgbe_recv_scattered_pkts_vec;
                        else
 #endif
-                               dev->rx_pkt_burst = ixgbe_recv_scattered_pkts;
+                               dev->rx_pkt_burst =
+                                       ixgbe_recv_pkts_lro_single_alloc;
                }
        }