X-Git-Url: http://git.droids-corp.org/?a=blobdiff_plain;f=drivers%2Fnet%2Fice%2Fice_rxtx_vec_avx2.c;h=9725ac01804357c4ea4c68eb306681e77d81f63a;hb=05b405d581486651305551a9f7295f40388d95db;hp=83dcdf15d424754b23a973dca0e6e7cd1eed0806;hpb=20daa1c978b7d34b447453b3f25483c700745456;p=dpdk.git diff --git a/drivers/net/ice/ice_rxtx_vec_avx2.c b/drivers/net/ice/ice_rxtx_vec_avx2.c index 83dcdf15d4..9725ac0180 100644 --- a/drivers/net/ice/ice_rxtx_vec_avx2.c +++ b/drivers/net/ice/ice_rxtx_vec_avx2.c @@ -16,7 +16,7 @@ ice_rxq_rearm(struct ice_rx_queue *rxq) return ice_rxq_rearm_common(rxq, false); } -static inline __m256i +static __rte_always_inline __m256i ice_flex_rxd_to_fdir_flags_vec_avx2(const __m256i fdir_id0_7) { #define FDID_MIS_MAGIC 0xFFFFFFFF @@ -35,9 +35,10 @@ ice_flex_rxd_to_fdir_flags_vec_avx2(const __m256i fdir_id0_7) return fdir_flags; } -static inline uint16_t +static __rte_always_inline uint16_t _ice_recv_raw_pkts_vec_avx2(struct ice_rx_queue *rxq, struct rte_mbuf **rx_pkts, - uint16_t nb_pkts, uint8_t *split_packet) + uint16_t nb_pkts, uint8_t *split_packet, + bool offload) { #define ICE_DESCS_PER_LOOP_AVX 8 @@ -385,39 +386,43 @@ _ice_recv_raw_pkts_vec_avx2(struct ice_rx_queue *rxq, struct rte_mbuf **rx_pkts, */ __m256i status0_7 = _mm256_unpacklo_epi64(status4_7, status0_3); + __m256i mbuf_flags = _mm256_set1_epi32(0); - /* now do flag manipulation */ + if (offload) { + /* now do flag manipulation */ - /* get only flag/error bits we want */ - const __m256i flag_bits = - _mm256_and_si256(status0_7, flags_mask); - /** - * l3_l4_error flags, shuffle, then shift to correct adjustment - * of flags in flags_shuf, and finally mask out extra bits - */ - __m256i l3_l4_flags = _mm256_shuffle_epi8(l3_l4_flags_shuf, - _mm256_srli_epi32(flag_bits, 4)); - l3_l4_flags = _mm256_slli_epi32(l3_l4_flags, 1); - - __m256i l4_outer_mask = _mm256_set1_epi32(0x6); - __m256i l4_outer_flags = - _mm256_and_si256(l3_l4_flags, l4_outer_mask); - l4_outer_flags = _mm256_slli_epi32(l4_outer_flags, 20); - - __m256i l3_l4_mask = _mm256_set1_epi32(~0x6); - l3_l4_flags = _mm256_and_si256(l3_l4_flags, l3_l4_mask); - l3_l4_flags = _mm256_or_si256(l3_l4_flags, l4_outer_flags); - l3_l4_flags = _mm256_and_si256(l3_l4_flags, cksum_mask); - /* set rss and vlan flags */ - const __m256i rss_vlan_flag_bits = - _mm256_srli_epi32(flag_bits, 12); - const __m256i rss_vlan_flags = - _mm256_shuffle_epi8(rss_vlan_flags_shuf, - rss_vlan_flag_bits); - - /* merge flags */ - __m256i mbuf_flags = _mm256_or_si256(l3_l4_flags, - rss_vlan_flags); + /* get only flag/error bits we want */ + const __m256i flag_bits = + _mm256_and_si256(status0_7, flags_mask); + /** + * l3_l4_error flags, shuffle, then shift to correct adjustment + * of flags in flags_shuf, and finally mask out extra bits + */ + __m256i l3_l4_flags = _mm256_shuffle_epi8(l3_l4_flags_shuf, + _mm256_srli_epi32(flag_bits, 4)); + l3_l4_flags = _mm256_slli_epi32(l3_l4_flags, 1); + + __m256i l4_outer_mask = _mm256_set1_epi32(0x6); + __m256i l4_outer_flags = + _mm256_and_si256(l3_l4_flags, l4_outer_mask); + l4_outer_flags = _mm256_slli_epi32(l4_outer_flags, 20); + + __m256i l3_l4_mask = _mm256_set1_epi32(~0x6); + + l3_l4_flags = _mm256_and_si256(l3_l4_flags, l3_l4_mask); + l3_l4_flags = _mm256_or_si256(l3_l4_flags, l4_outer_flags); + l3_l4_flags = _mm256_and_si256(l3_l4_flags, cksum_mask); + /* set rss and vlan flags */ + const __m256i rss_vlan_flag_bits = + _mm256_srli_epi32(flag_bits, 12); + const __m256i rss_vlan_flags = + _mm256_shuffle_epi8(rss_vlan_flags_shuf, + rss_vlan_flag_bits); + + /* merge flags */ + mbuf_flags = _mm256_or_si256(l3_l4_flags, + rss_vlan_flags); + } if (rxq->fdir_enabled) { const __m256i fdir_id4_7 = @@ -461,95 +466,97 @@ _ice_recv_raw_pkts_vec_avx2(struct ice_rx_queue *rxq, struct rte_mbuf **rx_pkts, _mm256_extract_epi32(fdir_id0_7, 4); } /* if() on fdir_enabled */ + if (offload) { #ifndef RTE_LIBRTE_ICE_16BYTE_RX_DESC - /** - * 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) { - /* load bottom half of every 32B desc */ - const __m128i raw_desc_bh7 = - _mm_load_si128 - ((void *)(&rxdp[7].wb.status_error1)); - rte_compiler_barrier(); - const __m128i raw_desc_bh6 = - _mm_load_si128 - ((void *)(&rxdp[6].wb.status_error1)); - rte_compiler_barrier(); - const __m128i raw_desc_bh5 = - _mm_load_si128 - ((void *)(&rxdp[5].wb.status_error1)); - rte_compiler_barrier(); - const __m128i raw_desc_bh4 = - _mm_load_si128 - ((void *)(&rxdp[4].wb.status_error1)); - rte_compiler_barrier(); - const __m128i raw_desc_bh3 = - _mm_load_si128 - ((void *)(&rxdp[3].wb.status_error1)); - rte_compiler_barrier(); - const __m128i raw_desc_bh2 = - _mm_load_si128 - ((void *)(&rxdp[2].wb.status_error1)); - rte_compiler_barrier(); - const __m128i raw_desc_bh1 = - _mm_load_si128 - ((void *)(&rxdp[1].wb.status_error1)); - rte_compiler_barrier(); - const __m128i raw_desc_bh0 = - _mm_load_si128 - ((void *)(&rxdp[0].wb.status_error1)); - - __m256i raw_desc_bh6_7 = - _mm256_inserti128_si256 - (_mm256_castsi128_si256(raw_desc_bh6), - raw_desc_bh7, 1); - __m256i raw_desc_bh4_5 = - _mm256_inserti128_si256 - (_mm256_castsi128_si256(raw_desc_bh4), - raw_desc_bh5, 1); - __m256i raw_desc_bh2_3 = - _mm256_inserti128_si256 - (_mm256_castsi128_si256(raw_desc_bh2), - raw_desc_bh3, 1); - __m256i raw_desc_bh0_1 = - _mm256_inserti128_si256 - (_mm256_castsi128_si256(raw_desc_bh0), - raw_desc_bh1, 1); - /** - * to shift the 32b RSS hash value to the - * highest 32b of each 128b before mask + * needs to load 2nd 16B of each desc for RSS hash parsing, + * will cause performance drop to get into this context. */ - __m256i rss_hash6_7 = - _mm256_slli_epi64(raw_desc_bh6_7, 32); - __m256i rss_hash4_5 = - _mm256_slli_epi64(raw_desc_bh4_5, 32); - __m256i rss_hash2_3 = - _mm256_slli_epi64(raw_desc_bh2_3, 32); - __m256i rss_hash0_1 = - _mm256_slli_epi64(raw_desc_bh0_1, 32); - - __m256i rss_hash_msk = - _mm256_set_epi32(0xFFFFFFFF, 0, 0, 0, - 0xFFFFFFFF, 0, 0, 0); - - rss_hash6_7 = _mm256_and_si256 - (rss_hash6_7, rss_hash_msk); - rss_hash4_5 = _mm256_and_si256 - (rss_hash4_5, rss_hash_msk); - rss_hash2_3 = _mm256_and_si256 - (rss_hash2_3, rss_hash_msk); - rss_hash0_1 = _mm256_and_si256 - (rss_hash0_1, rss_hash_msk); - - mb6_7 = _mm256_or_si256(mb6_7, rss_hash6_7); - mb4_5 = _mm256_or_si256(mb4_5, rss_hash4_5); - mb2_3 = _mm256_or_si256(mb2_3, rss_hash2_3); - mb0_1 = _mm256_or_si256(mb0_1, rss_hash0_1); - } /* if() on RSS hash parsing */ + if (rxq->vsi->adapter->pf.dev_data->dev_conf.rxmode.offloads & + DEV_RX_OFFLOAD_RSS_HASH) { + /* load bottom half of every 32B desc */ + const __m128i raw_desc_bh7 = + _mm_load_si128 + ((void *)(&rxdp[7].wb.status_error1)); + rte_compiler_barrier(); + const __m128i raw_desc_bh6 = + _mm_load_si128 + ((void *)(&rxdp[6].wb.status_error1)); + rte_compiler_barrier(); + const __m128i raw_desc_bh5 = + _mm_load_si128 + ((void *)(&rxdp[5].wb.status_error1)); + rte_compiler_barrier(); + const __m128i raw_desc_bh4 = + _mm_load_si128 + ((void *)(&rxdp[4].wb.status_error1)); + rte_compiler_barrier(); + const __m128i raw_desc_bh3 = + _mm_load_si128 + ((void *)(&rxdp[3].wb.status_error1)); + rte_compiler_barrier(); + const __m128i raw_desc_bh2 = + _mm_load_si128 + ((void *)(&rxdp[2].wb.status_error1)); + rte_compiler_barrier(); + const __m128i raw_desc_bh1 = + _mm_load_si128 + ((void *)(&rxdp[1].wb.status_error1)); + rte_compiler_barrier(); + const __m128i raw_desc_bh0 = + _mm_load_si128 + ((void *)(&rxdp[0].wb.status_error1)); + + __m256i raw_desc_bh6_7 = + _mm256_inserti128_si256 + (_mm256_castsi128_si256(raw_desc_bh6), + raw_desc_bh7, 1); + __m256i raw_desc_bh4_5 = + _mm256_inserti128_si256 + (_mm256_castsi128_si256(raw_desc_bh4), + raw_desc_bh5, 1); + __m256i raw_desc_bh2_3 = + _mm256_inserti128_si256 + (_mm256_castsi128_si256(raw_desc_bh2), + raw_desc_bh3, 1); + __m256i raw_desc_bh0_1 = + _mm256_inserti128_si256 + (_mm256_castsi128_si256(raw_desc_bh0), + raw_desc_bh1, 1); + + /** + * to shift the 32b RSS hash value to the + * highest 32b of each 128b before mask + */ + __m256i rss_hash6_7 = + _mm256_slli_epi64(raw_desc_bh6_7, 32); + __m256i rss_hash4_5 = + _mm256_slli_epi64(raw_desc_bh4_5, 32); + __m256i rss_hash2_3 = + _mm256_slli_epi64(raw_desc_bh2_3, 32); + __m256i rss_hash0_1 = + _mm256_slli_epi64(raw_desc_bh0_1, 32); + + __m256i rss_hash_msk = + _mm256_set_epi32(0xFFFFFFFF, 0, 0, 0, + 0xFFFFFFFF, 0, 0, 0); + + rss_hash6_7 = _mm256_and_si256 + (rss_hash6_7, rss_hash_msk); + rss_hash4_5 = _mm256_and_si256 + (rss_hash4_5, rss_hash_msk); + rss_hash2_3 = _mm256_and_si256 + (rss_hash2_3, rss_hash_msk); + rss_hash0_1 = _mm256_and_si256 + (rss_hash0_1, rss_hash_msk); + + mb6_7 = _mm256_or_si256(mb6_7, rss_hash6_7); + mb4_5 = _mm256_or_si256(mb4_5, rss_hash4_5); + mb2_3 = _mm256_or_si256(mb2_3, rss_hash2_3); + mb0_1 = _mm256_or_si256(mb0_1, rss_hash0_1); + } /* if() on RSS hash parsing */ #endif + } /** * At this point, we have the 8 sets of flags in the low 16-bits @@ -701,7 +708,16 @@ uint16_t ice_recv_pkts_vec_avx2(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts) { - return _ice_recv_raw_pkts_vec_avx2(rx_queue, rx_pkts, nb_pkts, NULL); + return _ice_recv_raw_pkts_vec_avx2(rx_queue, rx_pkts, + nb_pkts, NULL, false); +} + +uint16_t +ice_recv_pkts_vec_avx2_offload(void *rx_queue, struct rte_mbuf **rx_pkts, + uint16_t nb_pkts) +{ + return _ice_recv_raw_pkts_vec_avx2(rx_queue, rx_pkts, + nb_pkts, NULL, true); } /** @@ -709,16 +725,16 @@ ice_recv_pkts_vec_avx2(void *rx_queue, struct rte_mbuf **rx_pkts, * Notice: * - nb_pkts < ICE_DESCS_PER_LOOP, just return no packet */ -static uint16_t +static __rte_always_inline uint16_t ice_recv_scattered_burst_vec_avx2(void *rx_queue, struct rte_mbuf **rx_pkts, - uint16_t nb_pkts) + uint16_t nb_pkts, bool offload) { struct ice_rx_queue *rxq = rx_queue; uint8_t split_flags[ICE_VPMD_RX_BURST] = {0}; /* get some new buffers */ uint16_t nb_bufs = _ice_recv_raw_pkts_vec_avx2(rxq, rx_pkts, nb_pkts, - split_flags); + split_flags, offload); if (nb_bufs == 0) return 0; @@ -751,48 +767,74 @@ ice_recv_scattered_burst_vec_avx2(void *rx_queue, struct rte_mbuf **rx_pkts, * Notice: * - nb_pkts < ICE_DESCS_PER_LOOP, just return no packet */ -uint16_t -ice_recv_scattered_pkts_vec_avx2(void *rx_queue, struct rte_mbuf **rx_pkts, - uint16_t nb_pkts) +static __rte_always_inline uint16_t +ice_recv_scattered_pkts_vec_avx2_common(void *rx_queue, + struct rte_mbuf **rx_pkts, + uint16_t nb_pkts, + bool offload) { uint16_t retval = 0; while (nb_pkts > ICE_VPMD_RX_BURST) { uint16_t burst = ice_recv_scattered_burst_vec_avx2(rx_queue, - rx_pkts + retval, ICE_VPMD_RX_BURST); + rx_pkts + retval, ICE_VPMD_RX_BURST, offload); retval += burst; nb_pkts -= burst; if (burst < ICE_VPMD_RX_BURST) return retval; } return retval + ice_recv_scattered_burst_vec_avx2(rx_queue, - rx_pkts + retval, nb_pkts); + rx_pkts + retval, nb_pkts, offload); } -static inline void +uint16_t +ice_recv_scattered_pkts_vec_avx2(void *rx_queue, + struct rte_mbuf **rx_pkts, + uint16_t nb_pkts) +{ + return ice_recv_scattered_pkts_vec_avx2_common(rx_queue, + rx_pkts, + nb_pkts, + false); +} + +uint16_t +ice_recv_scattered_pkts_vec_avx2_offload(void *rx_queue, + struct rte_mbuf **rx_pkts, + uint16_t nb_pkts) +{ + return ice_recv_scattered_pkts_vec_avx2_common(rx_queue, + rx_pkts, + nb_pkts, + true); +} + +static __rte_always_inline void ice_vtx1(volatile struct ice_tx_desc *txdp, - struct rte_mbuf *pkt, uint64_t flags) + struct rte_mbuf *pkt, uint64_t flags, bool offload) { uint64_t high_qw = (ICE_TX_DESC_DTYPE_DATA | ((uint64_t)flags << ICE_TXD_QW1_CMD_S) | ((uint64_t)pkt->data_len << ICE_TXD_QW1_TX_BUF_SZ_S)); + if (offload) + ice_txd_enable_offload(pkt, &high_qw); __m128i descriptor = _mm_set_epi64x(high_qw, pkt->buf_iova + pkt->data_off); _mm_store_si128((__m128i *)txdp, descriptor); } -static inline void +static __rte_always_inline void ice_vtx(volatile struct ice_tx_desc *txdp, - struct rte_mbuf **pkt, uint16_t nb_pkts, uint64_t flags) + struct rte_mbuf **pkt, uint16_t nb_pkts, uint64_t flags, bool offload) { const uint64_t hi_qw_tmpl = (ICE_TX_DESC_DTYPE_DATA | ((uint64_t)flags << ICE_TXD_QW1_CMD_S)); /* if unaligned on 32-bit boundary, do one to align */ if (((uintptr_t)txdp & 0x1F) != 0 && nb_pkts != 0) { - ice_vtx1(txdp, *pkt, flags); + ice_vtx1(txdp, *pkt, flags, offload); nb_pkts--, txdp++, pkt++; } @@ -802,18 +844,26 @@ ice_vtx(volatile struct ice_tx_desc *txdp, hi_qw_tmpl | ((uint64_t)pkt[3]->data_len << ICE_TXD_QW1_TX_BUF_SZ_S); + if (offload) + ice_txd_enable_offload(pkt[3], &hi_qw3); uint64_t hi_qw2 = hi_qw_tmpl | ((uint64_t)pkt[2]->data_len << ICE_TXD_QW1_TX_BUF_SZ_S); + if (offload) + ice_txd_enable_offload(pkt[2], &hi_qw2); uint64_t hi_qw1 = hi_qw_tmpl | ((uint64_t)pkt[1]->data_len << ICE_TXD_QW1_TX_BUF_SZ_S); + if (offload) + ice_txd_enable_offload(pkt[1], &hi_qw1); uint64_t hi_qw0 = hi_qw_tmpl | ((uint64_t)pkt[0]->data_len << ICE_TXD_QW1_TX_BUF_SZ_S); + if (offload) + ice_txd_enable_offload(pkt[0], &hi_qw0); __m256i desc2_3 = _mm256_set_epi64x @@ -833,14 +883,14 @@ ice_vtx(volatile struct ice_tx_desc *txdp, /* do any last ones */ while (nb_pkts) { - ice_vtx1(txdp, *pkt, flags); + ice_vtx1(txdp, *pkt, flags, offload); txdp++, pkt++, nb_pkts--; } } -static inline uint16_t +static __rte_always_inline uint16_t ice_xmit_fixed_burst_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts, - uint16_t nb_pkts) + uint16_t nb_pkts, bool offload) { struct ice_tx_queue *txq = (struct ice_tx_queue *)tx_queue; volatile struct ice_tx_desc *txdp; @@ -853,7 +903,7 @@ ice_xmit_fixed_burst_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts, nb_pkts = RTE_MIN(nb_pkts, txq->tx_rs_thresh); if (txq->nb_tx_free < txq->tx_free_thresh) - ice_tx_free_bufs(txq); + ice_tx_free_bufs_vec(txq); nb_commit = nb_pkts = (uint16_t)RTE_MIN(txq->nb_tx_free, nb_pkts); if (unlikely(nb_pkts == 0)) @@ -869,11 +919,11 @@ ice_xmit_fixed_burst_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts, if (nb_commit >= n) { ice_tx_backlog_entry(txep, tx_pkts, n); - ice_vtx(txdp, tx_pkts, n - 1, flags); + ice_vtx(txdp, tx_pkts, n - 1, flags, offload); tx_pkts += (n - 1); txdp += (n - 1); - ice_vtx1(txdp, *tx_pkts++, rs); + ice_vtx1(txdp, *tx_pkts++, rs, offload); nb_commit = (uint16_t)(nb_commit - n); @@ -887,7 +937,7 @@ ice_xmit_fixed_burst_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts, ice_tx_backlog_entry(txep, tx_pkts, nb_commit); - ice_vtx(txdp, tx_pkts, nb_commit, flags); + ice_vtx(txdp, tx_pkts, nb_commit, flags, offload); tx_id = (uint16_t)(tx_id + nb_commit); if (tx_id > txq->tx_next_rs) { @@ -905,9 +955,9 @@ ice_xmit_fixed_burst_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts, return nb_pkts; } -uint16_t -ice_xmit_pkts_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts, - uint16_t nb_pkts) +static __rte_always_inline uint16_t +ice_xmit_pkts_vec_avx2_common(void *tx_queue, struct rte_mbuf **tx_pkts, + uint16_t nb_pkts, bool offload) { uint16_t nb_tx = 0; struct ice_tx_queue *txq = (struct ice_tx_queue *)tx_queue; @@ -917,7 +967,7 @@ ice_xmit_pkts_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts, num = (uint16_t)RTE_MIN(nb_pkts, txq->tx_rs_thresh); ret = ice_xmit_fixed_burst_vec_avx2(tx_queue, &tx_pkts[nb_tx], - num); + num, offload); nb_tx += ret; nb_pkts -= ret; if (ret < num) @@ -926,3 +976,17 @@ ice_xmit_pkts_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts, return nb_tx; } + +uint16_t +ice_xmit_pkts_vec_avx2(void *tx_queue, struct rte_mbuf **tx_pkts, + uint16_t nb_pkts) +{ + return ice_xmit_pkts_vec_avx2_common(tx_queue, tx_pkts, nb_pkts, false); +} + +uint16_t +ice_xmit_pkts_vec_avx2_offload(void *tx_queue, struct rte_mbuf **tx_pkts, + uint16_t nb_pkts) +{ + return ice_xmit_pkts_vec_avx2_common(tx_queue, tx_pkts, nb_pkts, true); +}