net/sfc: support Rx scatter in EF10 Rx datapath
authorAndrew Rybchenko <arybchenko@solarflare.com>
Wed, 3 Oct 2018 09:03:55 +0000 (10:03 +0100)
committerFerruh Yigit <ferruh.yigit@intel.com>
Thu, 11 Oct 2018 16:53:49 +0000 (18:53 +0200)
Signed-off-by: Andrew Rybchenko <arybchenko@solarflare.com>
Reviewed-by: Ivan Malov <ivan.malov@oktetlabs.ru>
doc/guides/nics/sfc_efx.rst
doc/guides/rel_notes/release_18_11.rst
drivers/net/sfc/sfc_ef10_rx.c

index 425e669..a241f00 100644 (file)
@@ -322,7 +322,7 @@ boolean parameters value.
   **efx** chooses libefx-based datapath which supports Rx scatter.
   **ef10** chooses EF10 (SFN7xxx, SFN8xxx, X2xxx) native datapath which is
   more efficient than libefx-based and provides richer packet type
-  classification, but lacks Rx scatter support.
+  classification.
   **ef10_esps** chooses SFNX2xxx equal stride packed stream datapath
   which may be used on DPDK firmware variant only
   (see notes about its limitations above).
index b92c204..3d19191 100644 (file)
@@ -88,6 +88,12 @@ New Features
   Added the new enetc driver for NXP enetc platform. See the
   "ENETC Poll Mode Driver" document for more details on this new driver.
 
+* **Updated Solarflare network PMD.**
+
+  Updated the sfc_efx driver including the following changes:
+
+  * Added support for Rx scatter in EF10 datapath implementation.
+
 * **Updated failsafe driver.**
 
   Updated the failsafe driver including the following changes:
index 78602bc..ff52a02 100644 (file)
@@ -63,6 +63,7 @@ struct sfc_ef10_rxq {
        efx_qword_t                     *evq_hw_ring;
        struct sfc_ef10_rx_sw_desc      *sw_ring;
        uint64_t                        rearm_data;
+       struct rte_mbuf                 *scatter_pkt;
        uint16_t                        prefix_size;
 
        /* Used on refill */
@@ -192,6 +193,8 @@ sfc_ef10_rx_pending(struct sfc_ef10_rxq *rxq, struct rte_mbuf **rx_pkts,
 {
        uint16_t n_rx_pkts = RTE_MIN(nb_pkts, rxq->pending - rxq->completed);
 
+       SFC_ASSERT(rxq->pending == rxq->completed || rxq->scatter_pkt == NULL);
+
        if (n_rx_pkts != 0) {
                unsigned int completed = rxq->completed;
 
@@ -234,7 +237,13 @@ sfc_ef10_rx_process_event(struct sfc_ef10_rxq *rxq, efx_qword_t rx_ev,
 
        ready = (EFX_QWORD_FIELD(rx_ev, ESF_DZ_RX_DSC_PTR_LBITS) - pending) &
                EFX_MASK32(ESF_DZ_RX_DSC_PTR_LBITS);
-       SFC_ASSERT(ready > 0);
+
+       if (ready == 0) {
+               /* Rx abort - it was no enough descriptors for Rx packet */
+               rte_pktmbuf_free(rxq->scatter_pkt);
+               rxq->scatter_pkt = NULL;
+               return rx_pkts;
+       }
 
        rxq->pending = pending + ready;
 
@@ -250,15 +259,43 @@ sfc_ef10_rx_process_event(struct sfc_ef10_rxq *rxq, efx_qword_t rx_ev,
                return rx_pkts;
        }
 
+       /* If scattered packet is in progress */
+       if (rxq->scatter_pkt != NULL) {
+               /* Events for scattered packet frags are not merged */
+               SFC_ASSERT(ready == 1);
+               SFC_ASSERT(rxq->completed == pending);
+
+               /* There is no pseudo-header in scatter segments. */
+               seg_len = EFX_QWORD_FIELD(rx_ev, ESF_DZ_RX_BYTES);
+
+               rxd = &rxq->sw_ring[pending++ & ptr_mask];
+               m = rxd->mbuf;
+
+               MBUF_RAW_ALLOC_CHECK(m);
+
+               m->data_off = RTE_PKTMBUF_HEADROOM;
+               rte_pktmbuf_data_len(m) = seg_len;
+               rte_pktmbuf_pkt_len(m) = seg_len;
+
+               rxq->scatter_pkt->nb_segs++;
+               rte_pktmbuf_pkt_len(rxq->scatter_pkt) += seg_len;
+               rte_pktmbuf_lastseg(rxq->scatter_pkt)->next = m;
+
+               if (~rx_ev.eq_u64[0] &
+                   rte_cpu_to_le_64(1ull << ESF_DZ_RX_CONT_LBN)) {
+                       *rx_pkts++ = rxq->scatter_pkt;
+                       rxq->scatter_pkt = NULL;
+               }
+               rxq->completed = pending;
+               return rx_pkts;
+       }
+
        rxd = &rxq->sw_ring[pending++ & ptr_mask];
 
        sfc_ef10_rx_prefetch_next(rxq, pending & ptr_mask);
 
        m = rxd->mbuf;
 
-       *rx_pkts++ = m;
-       rxq->completed = pending;
-
        RTE_BUILD_BUG_ON(sizeof(m->rearm_data[0]) != sizeof(rxq->rearm_data));
        m->rearm_data[0] = rxq->rearm_data;
 
@@ -289,6 +326,17 @@ sfc_ef10_rx_process_event(struct sfc_ef10_rxq *rxq, efx_qword_t rx_ev,
 
        SFC_ASSERT(m->next == NULL);
 
+       if (~rx_ev.eq_u64[0] & rte_cpu_to_le_64(1ull << ESF_DZ_RX_CONT_LBN)) {
+               *rx_pkts++ = m;
+               rxq->completed = pending;
+       } else {
+               /* Events with CONT bit are not merged */
+               SFC_ASSERT(ready == 1);
+               rxq->scatter_pkt = m;
+               rxq->completed = pending;
+               return rx_pkts;
+       }
+
        /* Remember mbuf to copy offload flags and packet type from */
        m0 = m;
        while (pending != rxq->pending) {
@@ -649,6 +697,9 @@ sfc_ef10_rx_qpurge(struct sfc_dp_rxq *dp_rxq)
        unsigned int i;
        struct sfc_ef10_rx_sw_desc *rxd;
 
+       rte_pktmbuf_free(rxq->scatter_pkt);
+       rxq->scatter_pkt = NULL;
+
        for (i = rxq->completed; i != rxq->added; ++i) {
                rxd = &rxq->sw_ring[i & rxq->ptr_mask];
                rte_mbuf_raw_free(rxd->mbuf);
@@ -666,7 +717,8 @@ struct sfc_dp_rx sfc_ef10_rx = {
                .type           = SFC_DP_RX,
                .hw_fw_caps     = SFC_DP_HW_FW_CAP_EF10,
        },
-       .features               = SFC_DP_RX_FEAT_MULTI_PROCESS |
+       .features               = SFC_DP_RX_FEAT_SCATTER |
+                                 SFC_DP_RX_FEAT_MULTI_PROCESS |
                                  SFC_DP_RX_FEAT_TUNNELS |
                                  SFC_DP_RX_FEAT_CHECKSUM,
        .get_dev_info           = sfc_ef10_rx_get_dev_info,