X-Git-Url: http://git.droids-corp.org/?a=blobdiff_plain;f=drivers%2Fnet%2Fice%2Fice_rxtx_vec_avx512.c;h=5bba9887d296b8908a91f9e6b948192175dbc3b2;hb=a972972fe00a9f06f6fa8632981c4ee6ef8cdcd1;hp=e5e7cc148226397d303fe7bb8b38043b82611f5b;hpb=a4e480de268eaabbb2251b2f1782ae0f05f005a9;p=dpdk.git diff --git a/drivers/net/ice/ice_rxtx_vec_avx512.c b/drivers/net/ice/ice_rxtx_vec_avx512.c index e5e7cc1482..5bba9887d2 100644 --- a/drivers/net/ice/ice_rxtx_vec_avx512.c +++ b/drivers/net/ice/ice_rxtx_vec_avx512.c @@ -4,7 +4,7 @@ #include "ice_rxtx_vec_common.h" -#include +#include #ifndef __INTEL_COMPILER #pragma GCC diagnostic ignored "-Wcast-qual" @@ -12,7 +12,7 @@ #define ICE_DESCS_PER_LOOP_AVX 8 -static inline void +static __rte_always_inline void ice_rxq_rearm(struct ice_rx_queue *rxq) { int i; @@ -24,6 +24,9 @@ ice_rxq_rearm(struct ice_rx_queue *rxq) rxdp = rxq->rx_ring + rxq->rxrearm_start; + if (unlikely(!cache)) + return ice_rxq_rearm_common(rxq, true); + /* We need to pull 'n' more MBUFs into the software ring */ if (cache->len < ICE_RXQ_REARM_THRESH) { uint32_t req = ICE_RXQ_REARM_THRESH + (cache->size - @@ -125,13 +128,34 @@ ice_rxq_rearm(struct ice_rx_queue *rxq) (rxq->nb_rx_desc - 1) : (rxq->rxrearm_start - 1)); /* Update the tail pointer on the NIC */ - ICE_PCI_REG_WRITE(rxq->qrx_tail, rx_id); + ICE_PCI_REG_WC_WRITE(rxq->qrx_tail, rx_id); +} + +static inline __m256i +ice_flex_rxd_to_fdir_flags_vec_avx512(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); + /* 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, + fdir_mis_mask); + /* this XOR op results to bit-reverse the fdir_mask */ + fdir_mask = _mm256_xor_si256(fdir_mask, fdir_mis_mask); + const __m256i fdir_flags = _mm256_and_si256(fdir_mask, pkt_fdir_bit); + + return fdir_flags; } -static inline uint16_t +static __rte_always_inline uint16_t _ice_recv_raw_pkts_vec_avx512(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 do_offload) { const uint32_t *ptype_tbl = rxq->vsi->adapter->ptype_tbl; const __m256i mbuf_init = _mm256_set_epi64x(0, 0, @@ -202,6 +226,7 @@ _ice_recv_raw_pkts_vec_avx512(struct ice_rx_queue *rxq, RTE_BUILD_BUG_ON(offsetof(struct rte_mbuf, hash) != offsetof(struct rte_mbuf, rx_descriptor_fields1) + 12); + /* following code block is for Rx Checksum Offload */ /* Status/Error flag masks */ /** * mask everything except Checksum Reports, RSS indication @@ -211,43 +236,88 @@ _ice_recv_raw_pkts_vec_avx512(struct ice_rx_queue *rxq, * bit13 is for VLAN indication. */ const __m256i flags_mask = - _mm256_set1_epi32((7 << 4) | (1 << 12) | (1 << 13)); + _mm256_set1_epi32((0xF << 4) | (1 << 12) | (1 << 13)); /** * data to be shuffled by the result of the flags mask shifted by 4 * bits. This gives use the l3_l4 flags. */ - 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_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD | - PKT_RX_IP_CKSUM_BAD) >> 1, - (PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD | - PKT_RX_IP_CKSUM_GOOD) >> 1, - (PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD | - PKT_RX_IP_CKSUM_BAD) >> 1, - (PKT_RX_EIP_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, - /* 2nd 128-bits */ - 0, 0, 0, 0, 0, 0, 0, 0, - (PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD | - PKT_RX_IP_CKSUM_BAD) >> 1, - (PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD | - PKT_RX_IP_CKSUM_GOOD) >> 1, - (PKT_RX_EIP_CKSUM_BAD | PKT_RX_L4_CKSUM_GOOD | - PKT_RX_IP_CKSUM_BAD) >> 1, - (PKT_RX_EIP_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); + const __m256i l3_l4_flags_shuf = + _mm256_set_epi8((PKT_RX_OUTER_L4_CKSUM_BAD >> 20 | + PKT_RX_OUTER_IP_CKSUM_BAD | PKT_RX_L4_CKSUM_BAD | + PKT_RX_IP_CKSUM_BAD) >> 1, + (PKT_RX_OUTER_L4_CKSUM_BAD >> 20 | PKT_RX_OUTER_IP_CKSUM_BAD | + PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_GOOD) >> 1, + (PKT_RX_OUTER_L4_CKSUM_BAD >> 20 | PKT_RX_OUTER_IP_CKSUM_BAD | + PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD) >> 1, + (PKT_RX_OUTER_L4_CKSUM_BAD >> 20 | PKT_RX_OUTER_IP_CKSUM_BAD | + PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_CKSUM_GOOD) >> 1, + (PKT_RX_OUTER_L4_CKSUM_BAD >> 20 | PKT_RX_L4_CKSUM_BAD | + PKT_RX_IP_CKSUM_BAD) >> 1, + (PKT_RX_OUTER_L4_CKSUM_BAD >> 20 | PKT_RX_L4_CKSUM_BAD | + PKT_RX_IP_CKSUM_GOOD) >> 1, + (PKT_RX_OUTER_L4_CKSUM_BAD >> 20 | PKT_RX_L4_CKSUM_GOOD | + PKT_RX_IP_CKSUM_BAD) >> 1, + (PKT_RX_OUTER_L4_CKSUM_BAD >> 20 | PKT_RX_L4_CKSUM_GOOD | + PKT_RX_IP_CKSUM_GOOD) >> 1, + (PKT_RX_OUTER_L4_CKSUM_GOOD >> 20 | PKT_RX_OUTER_IP_CKSUM_BAD | + PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1, + (PKT_RX_OUTER_L4_CKSUM_GOOD >> 20 | PKT_RX_OUTER_IP_CKSUM_BAD | + PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_GOOD) >> 1, + (PKT_RX_OUTER_L4_CKSUM_GOOD >> 20 | PKT_RX_OUTER_IP_CKSUM_BAD | + PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD) >> 1, + (PKT_RX_OUTER_L4_CKSUM_GOOD >> 20 | PKT_RX_OUTER_IP_CKSUM_BAD | + PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_CKSUM_GOOD) >> 1, + (PKT_RX_OUTER_L4_CKSUM_GOOD >> 20 | PKT_RX_L4_CKSUM_BAD | + PKT_RX_IP_CKSUM_BAD) >> 1, + (PKT_RX_OUTER_L4_CKSUM_GOOD >> 20 | PKT_RX_L4_CKSUM_BAD | + PKT_RX_IP_CKSUM_GOOD) >> 1, + (PKT_RX_OUTER_L4_CKSUM_GOOD >> 20 | PKT_RX_L4_CKSUM_GOOD | + PKT_RX_IP_CKSUM_BAD) >> 1, + (PKT_RX_OUTER_L4_CKSUM_GOOD >> 20 | PKT_RX_L4_CKSUM_GOOD | + PKT_RX_IP_CKSUM_GOOD) >> 1, + /** + * second 128-bits + * shift right 20 bits to use the low two bits to indicate + * outer checksum status + * shift right 1 bit to make sure it not exceed 255 + */ + (PKT_RX_OUTER_L4_CKSUM_BAD >> 20 | PKT_RX_OUTER_IP_CKSUM_BAD | + PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1, + (PKT_RX_OUTER_L4_CKSUM_BAD >> 20 | PKT_RX_OUTER_IP_CKSUM_BAD | + PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_GOOD) >> 1, + (PKT_RX_OUTER_L4_CKSUM_BAD >> 20 | PKT_RX_OUTER_IP_CKSUM_BAD | + PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD) >> 1, + (PKT_RX_OUTER_L4_CKSUM_BAD >> 20 | PKT_RX_OUTER_IP_CKSUM_BAD | + PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_CKSUM_GOOD) >> 1, + (PKT_RX_OUTER_L4_CKSUM_BAD >> 20 | PKT_RX_L4_CKSUM_BAD | + PKT_RX_IP_CKSUM_BAD) >> 1, + (PKT_RX_OUTER_L4_CKSUM_BAD >> 20 | PKT_RX_L4_CKSUM_BAD | + PKT_RX_IP_CKSUM_GOOD) >> 1, + (PKT_RX_OUTER_L4_CKSUM_BAD >> 20 | PKT_RX_L4_CKSUM_GOOD | + PKT_RX_IP_CKSUM_BAD) >> 1, + (PKT_RX_OUTER_L4_CKSUM_BAD >> 20 | PKT_RX_L4_CKSUM_GOOD | + PKT_RX_IP_CKSUM_GOOD) >> 1, + (PKT_RX_OUTER_L4_CKSUM_GOOD >> 20 | PKT_RX_OUTER_IP_CKSUM_BAD | + PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_BAD) >> 1, + (PKT_RX_OUTER_L4_CKSUM_GOOD >> 20 | PKT_RX_OUTER_IP_CKSUM_BAD | + PKT_RX_L4_CKSUM_BAD | PKT_RX_IP_CKSUM_GOOD) >> 1, + (PKT_RX_OUTER_L4_CKSUM_GOOD >> 20 | PKT_RX_OUTER_IP_CKSUM_BAD | + PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_CKSUM_BAD) >> 1, + (PKT_RX_OUTER_L4_CKSUM_GOOD >> 20 | PKT_RX_OUTER_IP_CKSUM_BAD | + PKT_RX_L4_CKSUM_GOOD | PKT_RX_IP_CKSUM_GOOD) >> 1, + (PKT_RX_OUTER_L4_CKSUM_GOOD >> 20 | PKT_RX_L4_CKSUM_BAD | + PKT_RX_IP_CKSUM_BAD) >> 1, + (PKT_RX_OUTER_L4_CKSUM_GOOD >> 20 | PKT_RX_L4_CKSUM_BAD | + PKT_RX_IP_CKSUM_GOOD) >> 1, + (PKT_RX_OUTER_L4_CKSUM_GOOD >> 20 | PKT_RX_L4_CKSUM_GOOD | + PKT_RX_IP_CKSUM_BAD) >> 1, + (PKT_RX_OUTER_L4_CKSUM_GOOD >> 20 | PKT_RX_L4_CKSUM_GOOD | + PKT_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_EIP_CKSUM_BAD); + _mm256_set1_epi32(PKT_RX_IP_CKSUM_MASK | + PKT_RX_L4_CKSUM_MASK | + PKT_RX_OUTER_IP_CKSUM_BAD | + PKT_RX_OUTER_L4_CKSUM_MASK); /** * data to be shuffled by result of flag mask, shifted down 12. * If RSS(bit12)/VLAN(bit13) are set, @@ -420,119 +490,184 @@ _ice_recv_raw_pkts_vec_avx512(struct ice_rx_queue *rxq, __m256i status0_7 = _mm512_extracti64x4_epi64 (raw_status0_7, 0); - /* now do flag manipulation */ + __m256i mbuf_flags = _mm256_set1_epi32(0); - /* 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); - 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 */ - const __m256i mbuf_flags = _mm256_or_si256(l3_l4_flags, - rss_vlan_flags); + if (do_offload) { + /* now do flag manipulation */ -#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); + /* 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 = + _mm256_unpackhi_epi32(raw_desc6_7, raw_desc4_5); + + const __m256i fdir_id0_3 = + _mm256_unpackhi_epi32(raw_desc2_3, raw_desc0_1); + + const __m256i fdir_id0_7 = + _mm256_unpackhi_epi64(fdir_id4_7, fdir_id0_3); + + if (do_offload) { + const __m256i fdir_flags = + ice_flex_rxd_to_fdir_flags_vec_avx512 + (fdir_id0_7); + + /* merge with fdir_flags */ + mbuf_flags = _mm256_or_si256 + (mbuf_flags, fdir_flags); + } else { + mbuf_flags = + ice_flex_rxd_to_fdir_flags_vec_avx512 + (fdir_id0_7); + } + + /* write to mbuf: have to use scalar store here */ + rx_pkts[i + 0]->hash.fdir.hi = + _mm256_extract_epi32(fdir_id0_7, 3); + + rx_pkts[i + 1]->hash.fdir.hi = + _mm256_extract_epi32(fdir_id0_7, 7); + + rx_pkts[i + 2]->hash.fdir.hi = + _mm256_extract_epi32(fdir_id0_7, 2); + + rx_pkts[i + 3]->hash.fdir.hi = + _mm256_extract_epi32(fdir_id0_7, 6); + + rx_pkts[i + 4]->hash.fdir.hi = + _mm256_extract_epi32(fdir_id0_7, 1); + + rx_pkts[i + 5]->hash.fdir.hi = + _mm256_extract_epi32(fdir_id0_7, 5); + + rx_pkts[i + 6]->hash.fdir.hi = + _mm256_extract_epi32(fdir_id0_7, 0); + + rx_pkts[i + 7]->hash.fdir.hi = + _mm256_extract_epi32(fdir_id0_7, 4); + } /* if() on fdir_enabled */ + + if (do_offload) { +#ifndef RTE_LIBRTE_ICE_16BYTE_RX_DESC /** - * 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 @@ -688,7 +823,19 @@ uint16_t ice_recv_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts, uint16_t nb_pkts) { - return _ice_recv_raw_pkts_vec_avx512(rx_queue, rx_pkts, nb_pkts, NULL); + return _ice_recv_raw_pkts_vec_avx512(rx_queue, rx_pkts, nb_pkts, NULL, false); +} + +/** + * Notice: + * - nb_pkts < ICE_DESCS_PER_LOOP, just return no packet + */ +uint16_t +ice_recv_pkts_vec_avx512_offload(void *rx_queue, struct rte_mbuf **rx_pkts, + uint16_t nb_pkts) +{ + return _ice_recv_raw_pkts_vec_avx512(rx_queue, rx_pkts, + nb_pkts, NULL, true); } /** @@ -705,7 +852,49 @@ ice_recv_scattered_burst_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts, /* get some new buffers */ uint16_t nb_bufs = _ice_recv_raw_pkts_vec_avx512(rxq, rx_pkts, nb_pkts, - split_flags); + split_flags, false); + if (nb_bufs == 0) + return 0; + + /* happy day case, full burst + no packets to be joined */ + const uint64_t *split_fl64 = (uint64_t *)split_flags; + + if (!rxq->pkt_first_seg && + split_fl64[0] == 0 && split_fl64[1] == 0 && + split_fl64[2] == 0 && split_fl64[3] == 0) + return nb_bufs; + + /* reassemble any packets that need reassembly */ + unsigned int i = 0; + + if (!rxq->pkt_first_seg) { + /* find the first split flag, and only reassemble then */ + while (i < nb_bufs && !split_flags[i]) + i++; + if (i == nb_bufs) + return nb_bufs; + rxq->pkt_first_seg = rx_pkts[i]; + } + return i + ice_rx_reassemble_packets(rxq, &rx_pkts[i], nb_bufs - i, + &split_flags[i]); +} + +/** + * vPMD receive routine that reassembles single burst of 32 scattered packets + * Notice: + * - nb_pkts < ICE_DESCS_PER_LOOP, just return no packet + */ +static uint16_t +ice_recv_scattered_burst_vec_avx512_offload(void *rx_queue, + struct rte_mbuf **rx_pkts, + uint16_t nb_pkts) +{ + 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_avx512(rxq, + rx_pkts, nb_pkts, split_flags, true); if (nb_bufs == 0) return 0; @@ -756,6 +945,32 @@ ice_recv_scattered_pkts_vec_avx512(void *rx_queue, struct rte_mbuf **rx_pkts, rx_pkts + retval, nb_pkts); } +/** + * vPMD receive routine that reassembles scattered packets. + * Main receive routine that can handle arbitrary burst sizes + * Notice: + * - nb_pkts < ICE_DESCS_PER_LOOP, just return no packet + */ +uint16_t +ice_recv_scattered_pkts_vec_avx512_offload(void *rx_queue, + struct rte_mbuf **rx_pkts, + uint16_t nb_pkts) +{ + uint16_t retval = 0; + + while (nb_pkts > ICE_VPMD_RX_BURST) { + uint16_t burst = + ice_recv_scattered_burst_vec_avx512_offload(rx_queue, + rx_pkts + retval, ICE_VPMD_RX_BURST); + retval += burst; + nb_pkts -= burst; + if (burst < ICE_VPMD_RX_BURST) + return retval; + } + return retval + ice_recv_scattered_burst_vec_avx512_offload(rx_queue, + rx_pkts + retval, nb_pkts); +} + static __rte_always_inline int ice_tx_free_bufs_avx512(struct ice_tx_queue *txq) { @@ -781,9 +996,14 @@ ice_tx_free_bufs_avx512(struct ice_tx_queue *txq) if (txq->offloads & DEV_TX_OFFLOAD_MBUF_FAST_FREE && (n & 31) == 0) { struct rte_mempool *mp = txep[0].mbuf->pool; + void **cache_objs; struct rte_mempool_cache *cache = rte_mempool_default_cache(mp, rte_lcore_id()); - void **cache_objs = &cache->objs[cache->len]; + + if (!cache || cache->len == 0) + goto normal; + + cache_objs = &cache->objs[cache->len]; if (n > RTE_MEMPOOL_CACHE_MAX_SIZE) { rte_mempool_ops_enqueue_bulk(mp, (void *)txep, n); @@ -821,6 +1041,7 @@ ice_tx_free_bufs_avx512(struct ice_tx_queue *txq) goto done; } +normal: m = rte_pktmbuf_prefree_seg(txep[0].mbuf); if (likely(m)) { free[0] = m; @@ -858,23 +1079,26 @@ done: return txq->tx_rs_thresh; } -static inline void +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 do_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 (do_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 -ice_vtx(volatile struct ice_tx_desc *txdp, - struct rte_mbuf **pkt, uint16_t nb_pkts, uint64_t flags) +static __rte_always_inline void +ice_vtx(volatile struct ice_tx_desc *txdp, struct rte_mbuf **pkt, + uint16_t nb_pkts, uint64_t flags, bool do_offload) { const uint64_t hi_qw_tmpl = (ICE_TX_DESC_DTYPE_DATA | ((uint64_t)flags << ICE_TXD_QW1_CMD_S)); @@ -884,18 +1108,26 @@ 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 (do_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 (do_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 (do_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 (do_offload) + ice_txd_enable_offload(pkt[0], &hi_qw0); __m512i desc0_3 = _mm512_set_epi64 @@ -912,7 +1144,7 @@ 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, do_offload); txdp++, pkt++, nb_pkts--; } } @@ -927,9 +1159,9 @@ ice_tx_backlog_entry_avx512(struct ice_vec_tx_entry *txep, txep[i].mbuf = tx_pkts[i]; } -static inline uint16_t +static __rte_always_inline uint16_t ice_xmit_fixed_burst_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts, - uint16_t nb_pkts) + uint16_t nb_pkts, bool do_offload) { struct ice_tx_queue *txq = (struct ice_tx_queue *)tx_queue; volatile struct ice_tx_desc *txdp; @@ -959,11 +1191,11 @@ ice_xmit_fixed_burst_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts, if (nb_commit >= n) { ice_tx_backlog_entry_avx512(txep, tx_pkts, n); - ice_vtx(txdp, tx_pkts, n - 1, flags); + ice_vtx(txdp, tx_pkts, n - 1, flags, do_offload); tx_pkts += (n - 1); txdp += (n - 1); - ice_vtx1(txdp, *tx_pkts++, rs); + ice_vtx1(txdp, *tx_pkts++, rs, do_offload); nb_commit = (uint16_t)(nb_commit - n); @@ -977,7 +1209,7 @@ ice_xmit_fixed_burst_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts, ice_tx_backlog_entry_avx512(txep, tx_pkts, nb_commit); - ice_vtx(txdp, tx_pkts, nb_commit, flags); + ice_vtx(txdp, tx_pkts, nb_commit, flags, do_offload); tx_id = (uint16_t)(tx_id + nb_commit); if (tx_id > txq->tx_next_rs) { @@ -990,7 +1222,7 @@ ice_xmit_fixed_burst_vec_avx512(void *tx_queue, struct rte_mbuf **tx_pkts, txq->tx_tail = tx_id; - ICE_PCI_REG_WRITE(txq->qtx_tail, txq->tx_tail); + ICE_PCI_REG_WC_WRITE(txq->qtx_tail, txq->tx_tail); return nb_pkts; } @@ -1007,7 +1239,30 @@ ice_xmit_pkts_vec_avx512(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_avx512(tx_queue, - &tx_pkts[nb_tx], num); + &tx_pkts[nb_tx], num, false); + nb_tx += ret; + nb_pkts -= ret; + if (ret < num) + break; + } + + return nb_tx; +} + +uint16_t +ice_xmit_pkts_vec_avx512_offload(void *tx_queue, struct rte_mbuf **tx_pkts, + uint16_t nb_pkts) +{ + uint16_t nb_tx = 0; + struct ice_tx_queue *txq = (struct ice_tx_queue *)tx_queue; + + while (nb_pkts) { + uint16_t ret, num; + + num = (uint16_t)RTE_MIN(nb_pkts, txq->tx_rs_thresh); + ret = ice_xmit_fixed_burst_vec_avx512(tx_queue, + &tx_pkts[nb_tx], num, true); + nb_tx += ret; nb_pkts -= ret; if (ret < num)