git.droids-corp.org / dpdk.git / blobdiff
? search:
summary | shortlog | log | commit | commitdiff | tree
raw | inline | side by side
net/ngbe: support MAC filters
[dpdk.git] / drivers / net / iavf / iavf_rxtx_vec_avx2.c
diff --git a/drivers/net/iavf/iavf_rxtx_vec_avx2.c b/drivers/net/iavf/iavf_rxtx_vec_avx2.c
index f5646d6..b6ef1ae 100644 (file)
--- a/drivers/net/iavf/iavf_rxtx_vec_avx2.c
+++ b/drivers/net/iavf/iavf_rxtx_vec_avx2.c
@@ -4,7 +4,7 @@
 
 #include "iavf_rxtx_vec_common.h"
 
-#include <x86intrin.h>
+#include <rte_vect.h>
 
 #ifndef __INTEL_COMPILER
 #pragma GCC diagnostic ignored "-Wcast-qual"
@@ -127,8 +127,8 @@ _iavf_recv_raw_pkts_vec_avx2(struct iavf_rx_queue *rxq,
         * destination
         */
        const __m256i vlan_flags_shuf =
-               _mm256_set_epi32(0, 0, PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED, 0,
-                                0, 0, PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED, 0);
+               _mm256_set_epi32(0, 0, RTE_MBUF_F_RX_VLAN | RTE_MBUF_F_RX_VLAN_STRIPPED, 0,
+                                0, 0, RTE_MBUF_F_RX_VLAN | RTE_MBUF_F_RX_VLAN_STRIPPED, 0);
        /**
         * data to be shuffled by result of flag mask, shifted down 11.
         * If RSS/FDIR bits are set, shuffle moves appropriate flags in
@@ -136,11 +136,11 @@ _iavf_recv_raw_pkts_vec_avx2(struct iavf_rx_queue *rxq,
         */
        const __m256i rss_flags_shuf =
                _mm256_set_epi8(0, 0, 0, 0, 0, 0, 0, 0,
-                               PKT_RX_RSS_HASH | PKT_RX_FDIR, PKT_RX_RSS_HASH,
-                               0, 0, 0, 0, PKT_RX_FDIR, 0,/* end up 128-bits */
+                               RTE_MBUF_F_RX_RSS_HASH | RTE_MBUF_F_RX_FDIR, RTE_MBUF_F_RX_RSS_HASH,
+                               0, 0, 0, 0, RTE_MBUF_F_RX_FDIR, 0,/* end up 128-bits */
                                0, 0, 0, 0, 0, 0, 0, 0,
-                               PKT_RX_RSS_HASH | PKT_RX_FDIR, PKT_RX_RSS_HASH,
-                               0, 0, 0, 0, PKT_RX_FDIR, 0);
+                               RTE_MBUF_F_RX_RSS_HASH | RTE_MBUF_F_RX_FDIR, RTE_MBUF_F_RX_RSS_HASH,
+                               0, 0, 0, 0, RTE_MBUF_F_RX_FDIR, 0);
 
        /**
         * data to be shuffled by the result of the flags mask shifted by 22
@@ -148,33 +148,33 @@ _iavf_recv_raw_pkts_vec_avx2(struct iavf_rx_queue *rxq,
         */
        const __m256i l3_l4_flags_shuf = _mm256_set_epi8(0, 0, 0, 0, 0, 0, 0, 0,
                        /* shift right 1 bit to make sure it not exceed 255 */
-                       (PKT_RX_OUTER_IP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD |
-                        PKT_RX_IP_CKSUM_BAD) >> 1,
-                       (PKT_RX_IP_CKSUM_GOOD | PKT_RX_OUTER_IP_CKSUM_BAD |
-                        PKT_RX_L4_CKSUM_BAD) >> 1,
-                       (PKT_RX_OUTER_IP_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
-                       (PKT_RX_IP_CKSUM_GOOD | PKT_RX_OUTER_IP_CKSUM_BAD) >> 1,
-                       (PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
-                       (PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_BAD) >> 1,
-                       PKT_RX_IP_CKSUM_BAD >> 1,
-                       (PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD) >> 1,
+                       (RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD | RTE_MBUF_F_RX_L4_CKSUM_BAD |
+                        RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
+                       (RTE_MBUF_F_RX_IP_CKSUM_GOOD | RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD |
+                        RTE_MBUF_F_RX_L4_CKSUM_BAD) >> 1,
+                       (RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD | RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
+                       (RTE_MBUF_F_RX_IP_CKSUM_GOOD | RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD) >> 1,
+                       (RTE_MBUF_F_RX_L4_CKSUM_BAD | RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
+                       (RTE_MBUF_F_RX_IP_CKSUM_GOOD | RTE_MBUF_F_RX_L4_CKSUM_BAD) >> 1,
+                       RTE_MBUF_F_RX_IP_CKSUM_BAD >> 1,
+                       (RTE_MBUF_F_RX_IP_CKSUM_GOOD | RTE_MBUF_F_RX_L4_CKSUM_GOOD) >> 1,
                        /* second 128-bits */
                        0, 0, 0, 0, 0, 0, 0, 0,
-                       (PKT_RX_OUTER_IP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD |
-                        PKT_RX_IP_CKSUM_BAD) >> 1,
-                       (PKT_RX_IP_CKSUM_GOOD | PKT_RX_OUTER_IP_CKSUM_BAD |
-                        PKT_RX_L4_CKSUM_BAD) >> 1,
-                       (PKT_RX_OUTER_IP_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
-                       (PKT_RX_IP_CKSUM_GOOD | PKT_RX_OUTER_IP_CKSUM_BAD) >> 1,
-                       (PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
-                       (PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_BAD) >> 1,
-                       PKT_RX_IP_CKSUM_BAD >> 1,
-                       (PKT_RX_IP_CKSUM_GOOD | PKT_RX_L4_CKSUM_GOOD) >> 1);
+                       (RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD | RTE_MBUF_F_RX_L4_CKSUM_BAD |
+                        RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
+                       (RTE_MBUF_F_RX_IP_CKSUM_GOOD | RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD |
+                        RTE_MBUF_F_RX_L4_CKSUM_BAD) >> 1,
+                       (RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD | RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
+                       (RTE_MBUF_F_RX_IP_CKSUM_GOOD | RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD) >> 1,
+                       (RTE_MBUF_F_RX_L4_CKSUM_BAD | RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
+                       (RTE_MBUF_F_RX_IP_CKSUM_GOOD | RTE_MBUF_F_RX_L4_CKSUM_BAD) >> 1,
+                       RTE_MBUF_F_RX_IP_CKSUM_BAD >> 1,
+                       (RTE_MBUF_F_RX_IP_CKSUM_GOOD | RTE_MBUF_F_RX_L4_CKSUM_GOOD) >> 1);
 
        const __m256i cksum_mask =
-                _mm256_set1_epi32(PKT_RX_IP_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD |
-                                  PKT_RX_L4_CKSUM_GOOD | PKT_RX_L4_CKSUM_BAD |
-                                  PKT_RX_OUTER_IP_CKSUM_BAD);
+                _mm256_set1_epi32(RTE_MBUF_F_RX_IP_CKSUM_GOOD | RTE_MBUF_F_RX_IP_CKSUM_BAD |
+                                  RTE_MBUF_F_RX_L4_CKSUM_GOOD | RTE_MBUF_F_RX_L4_CKSUM_BAD |
+                                  RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD);
 
        RTE_SET_USED(avx_aligned); /* for 32B descriptors we don't use this */
 
@@ -502,10 +502,10 @@ static inline __m256i
 flex_rxd_to_fdir_flags_vec_avx2(const __m256i fdir_id0_7)
 {
 #define FDID_MIS_MAGIC 0xFFFFFFFF
-       RTE_BUILD_BUG_ON(PKT_RX_FDIR != (1 << 2));
-       RTE_BUILD_BUG_ON(PKT_RX_FDIR_ID != (1 << 13));
-       const __m256i pkt_fdir_bit = _mm256_set1_epi32(PKT_RX_FDIR |
-                       PKT_RX_FDIR_ID);
+       RTE_BUILD_BUG_ON(RTE_MBUF_F_RX_FDIR != (1 << 2));
+       RTE_BUILD_BUG_ON(RTE_MBUF_F_RX_FDIR_ID != (1 << 13));
+       const __m256i pkt_fdir_bit = _mm256_set1_epi32(RTE_MBUF_F_RX_FDIR |
+                       RTE_MBUF_F_RX_FDIR_ID);
        /* desc->flow_id field == 0xFFFFFFFF means fdir mismatch */
        const __m256i fdir_mis_mask = _mm256_set1_epi32(FDID_MIS_MAGIC);
        __m256i fdir_mask = _mm256_cmpeq_epi32(fdir_id0_7,
@@ -524,7 +524,10 @@ _iavf_recv_raw_pkts_vec_avx2_flex_rxd(struct iavf_rx_queue *rxq,
 {
 #define IAVF_DESCS_PER_LOOP_AVX 8
 
-       const uint32_t *type_table = rxq->vsi->adapter->ptype_tbl;
+       struct iavf_adapter *adapter = rxq->vsi->adapter;
+
+       uint64_t offloads = adapter->dev_data->dev_conf.rxmode.offloads;
+       const uint32_t *type_table = adapter->ptype_tbl;
 
        const __m256i mbuf_init = _mm256_set_epi64x(0, 0,
                        0, rxq->mbuf_initializer);
@@ -626,36 +629,36 @@ _iavf_recv_raw_pkts_vec_avx2_flex_rxd(struct iavf_rx_queue *rxq,
         */
        const __m256i l3_l4_flags_shuf = _mm256_set_epi8(0, 0, 0, 0, 0, 0, 0, 0,
                        /* shift right 1 bit to make sure it not exceed 255 */
-                       (PKT_RX_OUTER_IP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD |
-                        PKT_RX_IP_CKSUM_BAD) >> 1,
-                       (PKT_RX_OUTER_IP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD |
-                        PKT_RX_IP_CKSUM_GOOD) >> 1,
-                       (PKT_RX_OUTER_IP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD |
-                        PKT_RX_IP_CKSUM_BAD) >> 1,
-                       (PKT_RX_OUTER_IP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD |
-                        PKT_RX_IP_CKSUM_GOOD) >> 1,
-                       (PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
-                       (PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_GOOD) >> 1,
-                       (PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD) >> 1,
-                       (PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_CKSUM_GOOD) >> 1,
+                       (RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD | RTE_MBUF_F_RX_L4_CKSUM_BAD |
+                        RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
+                       (RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD | RTE_MBUF_F_RX_L4_CKSUM_BAD |
+                        RTE_MBUF_F_RX_IP_CKSUM_GOOD) >> 1,
+                       (RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD | RTE_MBUF_F_RX_L4_CKSUM_GOOD |
+                        RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
+                       (RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD | RTE_MBUF_F_RX_L4_CKSUM_GOOD |
+                        RTE_MBUF_F_RX_IP_CKSUM_GOOD) >> 1,
+                       (RTE_MBUF_F_RX_L4_CKSUM_BAD | RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
+                       (RTE_MBUF_F_RX_L4_CKSUM_BAD | RTE_MBUF_F_RX_IP_CKSUM_GOOD) >> 1,
+                       (RTE_MBUF_F_RX_L4_CKSUM_GOOD | RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
+                       (RTE_MBUF_F_RX_L4_CKSUM_GOOD | RTE_MBUF_F_RX_IP_CKSUM_GOOD) >> 1,
                        /* second 128-bits */
                        0, 0, 0, 0, 0, 0, 0, 0,
-                       (PKT_RX_OUTER_IP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD |
-                        PKT_RX_IP_CKSUM_BAD) >> 1,
-                       (PKT_RX_OUTER_IP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD |
-                        PKT_RX_IP_CKSUM_GOOD) >> 1,
-                       (PKT_RX_OUTER_IP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD |
-                        PKT_RX_IP_CKSUM_BAD) >> 1,
-                       (PKT_RX_OUTER_IP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD |
-                        PKT_RX_IP_CKSUM_GOOD) >> 1,
-                       (PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1,
-                       (PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_GOOD) >> 1,
-                       (PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD) >> 1,
-                       (PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_CKSUM_GOOD) >> 1);
+                       (RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD | RTE_MBUF_F_RX_L4_CKSUM_BAD |
+                        RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
+                       (RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD | RTE_MBUF_F_RX_L4_CKSUM_BAD |
+                        RTE_MBUF_F_RX_IP_CKSUM_GOOD) >> 1,
+                       (RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD | RTE_MBUF_F_RX_L4_CKSUM_GOOD |
+                        RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
+                       (RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD | RTE_MBUF_F_RX_L4_CKSUM_GOOD |
+                        RTE_MBUF_F_RX_IP_CKSUM_GOOD) >> 1,
+                       (RTE_MBUF_F_RX_L4_CKSUM_BAD | RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
+                       (RTE_MBUF_F_RX_L4_CKSUM_BAD | RTE_MBUF_F_RX_IP_CKSUM_GOOD) >> 1,
+                       (RTE_MBUF_F_RX_L4_CKSUM_GOOD | RTE_MBUF_F_RX_IP_CKSUM_BAD) >> 1,
+                       (RTE_MBUF_F_RX_L4_CKSUM_GOOD | RTE_MBUF_F_RX_IP_CKSUM_GOOD) >> 1);
        const __m256i cksum_mask =
-                _mm256_set1_epi32(PKT_RX_IP_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD |
-                                  PKT_RX_L4_CKSUM_GOOD | PKT_RX_L4_CKSUM_BAD |
-                                  PKT_RX_OUTER_IP_CKSUM_BAD);
+                _mm256_set1_epi32(RTE_MBUF_F_RX_IP_CKSUM_GOOD | RTE_MBUF_F_RX_IP_CKSUM_BAD |
+                                  RTE_MBUF_F_RX_L4_CKSUM_GOOD | RTE_MBUF_F_RX_L4_CKSUM_BAD |
+                                  RTE_MBUF_F_RX_OUTER_IP_CKSUM_BAD);
        /**
         * data to be shuffled by result of flag mask, shifted down 12.
         * If RSS(bit12)/VLAN(bit13) are set,
@@ -664,27 +667,27 @@ _iavf_recv_raw_pkts_vec_avx2_flex_rxd(struct iavf_rx_queue *rxq,
        const __m256i rss_flags_shuf = _mm256_set_epi8(0, 0, 0, 0,
                        0, 0, 0, 0,
                        0, 0, 0, 0,
-                       PKT_RX_RSS_HASH, 0,
-                       PKT_RX_RSS_HASH, 0,
+                       RTE_MBUF_F_RX_RSS_HASH, 0,
+                       RTE_MBUF_F_RX_RSS_HASH, 0,
                        /* end up 128-bits */
                        0, 0, 0, 0,
                        0, 0, 0, 0,
                        0, 0, 0, 0,
-                       PKT_RX_RSS_HASH, 0,
-                       PKT_RX_RSS_HASH, 0);
+                       RTE_MBUF_F_RX_RSS_HASH, 0,
+                       RTE_MBUF_F_RX_RSS_HASH, 0);
 
        const __m256i vlan_flags_shuf = _mm256_set_epi8(0, 0, 0, 0,
                        0, 0, 0, 0,
                        0, 0, 0, 0,
-                       PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
-                       PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
+                       RTE_MBUF_F_RX_VLAN | RTE_MBUF_F_RX_VLAN_STRIPPED,
+                       RTE_MBUF_F_RX_VLAN | RTE_MBUF_F_RX_VLAN_STRIPPED,
                        0, 0,
                        /* end up 128-bits */
                        0, 0, 0, 0,
                        0, 0, 0, 0,
                        0, 0, 0, 0,
-                       PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
-                       PKT_RX_VLAN | PKT_RX_VLAN_STRIPPED,
+                       RTE_MBUF_F_RX_VLAN | RTE_MBUF_F_RX_VLAN_STRIPPED,
+                       RTE_MBUF_F_RX_VLAN | RTE_MBUF_F_RX_VLAN_STRIPPED,
                        0, 0);
 
        uint16_t i, received;
@@ -903,9 +906,8 @@ _iavf_recv_raw_pkts_vec_avx2_flex_rxd(struct iavf_rx_queue *rxq,
                 * needs to load 2nd 16B of each desc for RSS hash parsing,
                 * will cause performance drop to get into this context.
                 */
-               if (rxq->vsi->adapter->eth_dev->data->dev_conf.rxmode.offloads &
-                               DEV_RX_OFFLOAD_RSS_HASH ||
-                               rxq->rx_flags & IAVF_RX_FLAGS_VLAN_TAG_LOC_L2TAG2_2) {
+               if (offloads & RTE_ETH_RX_OFFLOAD_RSS_HASH ||
+                   rxq->rx_flags & IAVF_RX_FLAGS_VLAN_TAG_LOC_L2TAG2_2) {
                        /* load bottom half of every 32B desc */
                        const __m128i raw_desc_bh7 =
                                _mm_load_si128
@@ -956,8 +958,7 @@ _iavf_recv_raw_pkts_vec_avx2_flex_rxd(struct iavf_rx_queue *rxq,
                                        (_mm256_castsi128_si256(raw_desc_bh0),
                                        raw_desc_bh1, 1);
 
-                       if (rxq->vsi->adapter->eth_dev->data->dev_conf.rxmode.offloads &
-                                       DEV_RX_OFFLOAD_RSS_HASH) {
+                       if (offloads & RTE_ETH_RX_OFFLOAD_RSS_HASH) {
                                /**
                                 * to shift the 32b RSS hash value to the
                                 * highest 32b of each 128b before mask
@@ -1025,8 +1026,8 @@ _iavf_recv_raw_pkts_vec_avx2_flex_rxd(struct iavf_rx_queue *rxq,
                                                        0, 0, 0, 0,
                                                        0, 0, 0, 0,
                                                        0, 0,
-                                                       PKT_RX_VLAN |
-                                                       PKT_RX_VLAN_STRIPPED,
+                                                       RTE_MBUF_F_RX_VLAN |
+                                                       RTE_MBUF_F_RX_VLAN_STRIPPED,
                                                        0);
 
                                vlan_flags =
@@ -1500,7 +1501,7 @@ iavf_xmit_fixed_burst_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts,
 
        txq->tx_tail = tx_id;
 
-       IAVF_PCI_REG_WRITE(txq->qtx_tail, txq->tx_tail);
+       IAVF_PCI_REG_WC_WRITE(txq->qtx_tail, txq->tx_tail);
 
        return nb_pkts;
 }
DPDK repo used for reviews