lib/node/ip4_lookup_sse.h

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(C) 2020 Marvell International Ltd.
   3  */
   4
   5 #ifndef __INCLUDE_IP4_LOOKUP_SSE_H__
   6 #define __INCLUDE_IP4_LOOKUP_SSE_H__
   7
   8 /* X86 SSE */
   9 static uint16_t
  10 ip4_lookup_node_process_vec(struct rte_graph *graph, struct rte_node *node,
  11                         void **objs, uint16_t nb_objs)
  12 {
  13         struct rte_mbuf *mbuf0, *mbuf1, *mbuf2, *mbuf3, **pkts;
  14         struct rte_lpm *lpm = IP4_LOOKUP_NODE_LPM(node->ctx);
  15         const int dyn = IP4_LOOKUP_NODE_PRIV1_OFF(node->ctx);
  16         rte_edge_t next0, next1, next2, next3, next_index;
  17         struct rte_ipv4_hdr *ipv4_hdr;
  18         uint32_t ip0, ip1, ip2, ip3;
  19         void **to_next, **from;
  20         uint16_t last_spec = 0;
  21         uint16_t n_left_from;
  22         uint16_t held = 0;
  23         uint32_t drop_nh;
  24         rte_xmm_t dst;
  25         __m128i dip; /* SSE register */
  26         int rc, i;
  27
  28         /* Speculative next */
  29         next_index = RTE_NODE_IP4_LOOKUP_NEXT_REWRITE;
  30         /* Drop node */
  31         drop_nh = ((uint32_t)RTE_NODE_IP4_LOOKUP_NEXT_PKT_DROP) << 16;
  32
  33         pkts = (struct rte_mbuf **)objs;
  34         from = objs;
  35         n_left_from = nb_objs;
  36
  37         if (n_left_from >= 4) {
  38                 for (i = 0; i < 4; i++)
  39                         rte_prefetch0(rte_pktmbuf_mtod_offset(pkts[i], void *,
  40                                                 sizeof(struct rte_ether_hdr)));
  41         }
  42
  43         /* Get stream for the speculated next node */
  44         to_next = rte_node_next_stream_get(graph, node, next_index, nb_objs);
  45         while (n_left_from >= 4) {
  46                 /* Prefetch next-next mbufs */
  47                 if (likely(n_left_from > 11)) {
  48                         rte_prefetch0(pkts[8]);
  49                         rte_prefetch0(pkts[9]);
  50                         rte_prefetch0(pkts[10]);
  51                         rte_prefetch0(pkts[11]);
  52                 }
  53
  54                 /* Prefetch next mbuf data */
  55                 if (likely(n_left_from > 7)) {
  56                         rte_prefetch0(rte_pktmbuf_mtod_offset(pkts[4], void *,
  57                                                 sizeof(struct rte_ether_hdr)));
  58                         rte_prefetch0(rte_pktmbuf_mtod_offset(pkts[5], void *,
  59                                                 sizeof(struct rte_ether_hdr)));
  60                         rte_prefetch0(rte_pktmbuf_mtod_offset(pkts[6], void *,
  61                                                 sizeof(struct rte_ether_hdr)));
  62                         rte_prefetch0(rte_pktmbuf_mtod_offset(pkts[7], void *,
  63                                                 sizeof(struct rte_ether_hdr)));
  64                 }
  65
  66                 mbuf0 = pkts[0];
  67                 mbuf1 = pkts[1];
  68                 mbuf2 = pkts[2];
  69                 mbuf3 = pkts[3];
  70
  71                 pkts += 4;
  72                 n_left_from -= 4;
  73
  74                 /* Extract DIP of mbuf0 */
  75                 ipv4_hdr = rte_pktmbuf_mtod_offset(mbuf0, struct rte_ipv4_hdr *,
  76                                                 sizeof(struct rte_ether_hdr));
  77                 ip0 = ipv4_hdr->dst_addr;
  78                 /* Extract cksum, ttl as ipv4 hdr is in cache */
  79                 node_mbuf_priv1(mbuf0, dyn)->cksum = ipv4_hdr->hdr_checksum;
  80                 node_mbuf_priv1(mbuf0, dyn)->ttl = ipv4_hdr->time_to_live;
  81
  82                 /* Extract DIP of mbuf1 */
  83                 ipv4_hdr = rte_pktmbuf_mtod_offset(mbuf1, struct rte_ipv4_hdr *,
  84                                                 sizeof(struct rte_ether_hdr));
  85                 ip1 = ipv4_hdr->dst_addr;
  86                 /* Extract cksum, ttl as ipv4 hdr is in cache */
  87                 node_mbuf_priv1(mbuf1, dyn)->cksum = ipv4_hdr->hdr_checksum;
  88                 node_mbuf_priv1(mbuf1, dyn)->ttl = ipv4_hdr->time_to_live;
  89
  90                 /* Extract DIP of mbuf2 */
  91                 ipv4_hdr = rte_pktmbuf_mtod_offset(mbuf2, struct rte_ipv4_hdr *,
  92                                                 sizeof(struct rte_ether_hdr));
  93                 ip2 = ipv4_hdr->dst_addr;
  94                 /* Extract cksum, ttl as ipv4 hdr is in cache */
  95                 node_mbuf_priv1(mbuf2, dyn)->cksum = ipv4_hdr->hdr_checksum;
  96                 node_mbuf_priv1(mbuf2, dyn)->ttl = ipv4_hdr->time_to_live;
  97
  98                 /* Extract DIP of mbuf3 */
  99                 ipv4_hdr = rte_pktmbuf_mtod_offset(mbuf3, struct rte_ipv4_hdr *,
 100                                                 sizeof(struct rte_ether_hdr));
 101                 ip3 = ipv4_hdr->dst_addr;
 102
 103                 /* Prepare for lookup x4 */
 104                 dip = _mm_set_epi32(ip3, ip2, ip1, ip0);
 105
 106                 /* Byte swap 4 IPV4 addresses. */
 107                 const __m128i bswap_mask = _mm_set_epi8(
 108                         12, 13, 14, 15, 8, 9, 10, 11, 4, 5, 6, 7, 0, 1, 2, 3);
 109                 dip = _mm_shuffle_epi8(dip, bswap_mask);
 110
 111                 /* Extract cksum, ttl as ipv4 hdr is in cache */
 112                 node_mbuf_priv1(mbuf3, dyn)->cksum = ipv4_hdr->hdr_checksum;
 113                 node_mbuf_priv1(mbuf3, dyn)->ttl = ipv4_hdr->time_to_live;
 114
 115                 /* Perform LPM lookup to get NH and next node */
 116                 rte_lpm_lookupx4(lpm, dip, dst.u32, drop_nh);
 117
 118                 /* Extract next node id and NH */
 119                 node_mbuf_priv1(mbuf0, dyn)->nh = dst.u32[0] & 0xFFFF;
 120                 next0 = (dst.u32[0] >> 16);
 121
 122                 node_mbuf_priv1(mbuf1, dyn)->nh = dst.u32[1] & 0xFFFF;
 123                 next1 = (dst.u32[1] >> 16);
 124
 125                 node_mbuf_priv1(mbuf2, dyn)->nh = dst.u32[2] & 0xFFFF;
 126                 next2 = (dst.u32[2] >> 16);
 127
 128                 node_mbuf_priv1(mbuf3, dyn)->nh = dst.u32[3] & 0xFFFF;
 129                 next3 = (dst.u32[3] >> 16);
 130
 131                 /* Enqueue four to next node */
 132                 rte_edge_t fix_spec =
 133                         (next_index ^ next0) | (next_index ^ next1) |
 134                         (next_index ^ next2) | (next_index ^ next3);
 135
 136                 if (unlikely(fix_spec)) {
 137                         /* Copy things successfully speculated till now */
 138                         rte_memcpy(to_next, from, last_spec * sizeof(from[0]));
 139                         from += last_spec;
 140                         to_next += last_spec;
 141                         held += last_spec;
 142                         last_spec = 0;
 143
 144                         /* Next0 */
 145                         if (next_index == next0) {
 146                                 to_next[0] = from[0];
 147                                 to_next++;
 148                                 held++;
 149                         } else {
 150                                 rte_node_enqueue_x1(graph, node, next0,
 151                                                     from[0]);
 152                         }
 153
 154                         /* Next1 */
 155                         if (next_index == next1) {
 156                                 to_next[0] = from[1];
 157                                 to_next++;
 158                                 held++;
 159                         } else {
 160                                 rte_node_enqueue_x1(graph, node, next1,
 161                                                     from[1]);
 162                         }
 163
 164                         /* Next2 */
 165                         if (next_index == next2) {
 166                                 to_next[0] = from[2];
 167                                 to_next++;
 168                                 held++;
 169                         } else {
 170                                 rte_node_enqueue_x1(graph, node, next2,
 171                                                     from[2]);
 172                         }
 173
 174                         /* Next3 */
 175                         if (next_index == next3) {
 176                                 to_next[0] = from[3];
 177                                 to_next++;
 178                                 held++;
 179                         } else {
 180                                 rte_node_enqueue_x1(graph, node, next3,
 181                                                     from[3]);
 182                         }
 183
 184                         from += 4;
 185
 186                 } else {
 187                         last_spec += 4;
 188                 }
 189         }
 190
 191         while (n_left_from > 0) {
 192                 uint32_t next_hop;
 193
 194                 mbuf0 = pkts[0];
 195
 196                 pkts += 1;
 197                 n_left_from -= 1;
 198
 199                 /* Extract DIP of mbuf0 */
 200                 ipv4_hdr = rte_pktmbuf_mtod_offset(mbuf0, struct rte_ipv4_hdr *,
 201                                                 sizeof(struct rte_ether_hdr));
 202                 /* Extract cksum, ttl as ipv4 hdr is in cache */
 203                 node_mbuf_priv1(mbuf0, dyn)->cksum = ipv4_hdr->hdr_checksum;
 204                 node_mbuf_priv1(mbuf0, dyn)->ttl = ipv4_hdr->time_to_live;
 205
 206                 rc = rte_lpm_lookup(lpm, rte_be_to_cpu_32(ipv4_hdr->dst_addr),
 207                                     &next_hop);
 208                 next_hop = (rc == 0) ? next_hop : drop_nh;
 209
 210                 node_mbuf_priv1(mbuf0, dyn)->nh = next_hop & 0xFFFF;
 211                 next0 = (next_hop >> 16);
 212
 213                 if (unlikely(next_index ^ next0)) {
 214                         /* Copy things successfully speculated till now */
 215                         rte_memcpy(to_next, from, last_spec * sizeof(from[0]));
 216                         from += last_spec;
 217                         to_next += last_spec;
 218                         held += last_spec;
 219                         last_spec = 0;
 220
 221                         rte_node_enqueue_x1(graph, node, next0, from[0]);
 222                         from += 1;
 223                 } else {
 224                         last_spec += 1;
 225                 }
 226         }
 227
 228         /* !!! Home run !!! */
 229         if (likely(last_spec == nb_objs)) {
 230                 rte_node_next_stream_move(graph, node, next_index);
 231                 return nb_objs;
 232         }
 233
 234         held += last_spec;
 235         /* Copy things successfully speculated till now */
 236         rte_memcpy(to_next, from, last_spec * sizeof(from[0]));
 237         rte_node_next_stream_put(graph, node, next_index, held);
 238
 239         return nb_objs;
 240 }
 241
 242 #endif /* __INCLUDE_IP4_LOOKUP_SSE_H__ */