node: add IPv4 lookup for arm64
authorNithin Dabilpuram <ndabilpuram@marvell.com>
Sat, 11 Apr 2020 14:14:19 +0000 (19:44 +0530)
committerThomas Monjalon <thomas@monjalon.net>
Tue, 5 May 2020 21:40:36 +0000 (23:40 +0200)
Add arm64 specific IPv4 lookup process function
for ip4_lookup node. This node performs LPM lookup
on every packet received and forwards it to a next
node that is identified by lookup result.

Signed-off-by: Nithin Dabilpuram <ndabilpuram@marvell.com>
Signed-off-by: Kiran Kumar K <kirankumark@marvell.com>
Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
lib/librte_node/ip4_lookup.c
lib/librte_node/ip4_lookup_neon.h [new file with mode: 0644]

index 988bc5e..8ec4f06 100644 (file)
@@ -28,6 +28,10 @@ struct ip4_lookup_node_main {
        struct rte_lpm *lpm_tbl[RTE_MAX_NUMA_NODES];
 };
 
+#if defined(RTE_MACHINE_CPUFLAG_NEON)
+#include "ip4_lookup_neon.h"
+#else
+
 static uint16_t
 ip4_lookup_node_process(struct rte_graph *graph, struct rte_node *node,
                        void **objs, uint16_t nb_objs)
@@ -101,6 +105,8 @@ ip4_lookup_node_process(struct rte_graph *graph, struct rte_node *node,
        return nb_objs;
 }
 
+#endif
+
 static int
 ip4_lookup_node_init(const struct rte_graph *graph, struct rte_node *node)
 {
diff --git a/lib/librte_node/ip4_lookup_neon.h b/lib/librte_node/ip4_lookup_neon.h
new file mode 100644 (file)
index 0000000..bb3150f
--- /dev/null
@@ -0,0 +1,238 @@
+/* SPDX-License-Identifier: BSD-3-Clause
+ * Copyright(C) 2020 Marvell International Ltd.
+ */
+
+#ifndef __INCLUDE_IP4_LOOKUP_NEON_H__
+#define __INCLUDE_IP4_LOOKUP_NEON_H__
+
+/* ARM64 NEON */
+static uint16_t
+ip4_lookup_node_process(struct rte_graph *graph, struct rte_node *node,
+                       void **objs, uint16_t nb_objs)
+{
+       struct rte_mbuf *mbuf0, *mbuf1, *mbuf2, *mbuf3, **pkts;
+       struct rte_ipv4_hdr *ipv4_hdr;
+       void **to_next, **from;
+       uint16_t last_spec = 0;
+       rte_edge_t next_index;
+       uint16_t n_left_from;
+       struct rte_lpm *lpm;
+       uint16_t held = 0;
+       uint32_t drop_nh;
+       rte_xmm_t result;
+       rte_xmm_t priv01;
+       rte_xmm_t priv23;
+       int32x4_t dip;
+       int rc, i;
+
+       /* Speculative next */
+       next_index = RTE_NODE_IP4_LOOKUP_NEXT_REWRITE;
+       /* Drop node */
+       drop_nh = ((uint32_t)RTE_NODE_IP4_LOOKUP_NEXT_PKT_DROP) << 16;
+
+       /* Get socket specific LPM from ctx */
+       lpm = *((struct rte_lpm **)node->ctx);
+
+       pkts = (struct rte_mbuf **)objs;
+       from = objs;
+       n_left_from = nb_objs;
+
+#define OBJS_PER_CLINE (RTE_CACHE_LINE_SIZE / sizeof(void *))
+       for (i = OBJS_PER_CLINE; i < RTE_GRAPH_BURST_SIZE; i += OBJS_PER_CLINE)
+               rte_prefetch0(&objs[i]);
+
+       for (i = 0; i < 4 && i < n_left_from; i++)
+               rte_prefetch0(rte_pktmbuf_mtod_offset(pkts[i], void *,
+                                               sizeof(struct rte_ether_hdr)));
+
+       /* Get stream for the speculated next node */
+       to_next = rte_node_next_stream_get(graph, node, next_index, nb_objs);
+       while (n_left_from >= 4) {
+#if RTE_GRAPH_BURST_SIZE > 64
+               /* Prefetch next-next mbufs */
+               if (likely(n_left_from > 11)) {
+                       rte_prefetch0(pkts[8]);
+                       rte_prefetch0(pkts[9]);
+                       rte_prefetch0(pkts[10]);
+                       rte_prefetch0(pkts[11]);
+               }
+#endif
+               /* Prefetch next mbuf data */
+               if (likely(n_left_from > 7)) {
+                       rte_prefetch0(rte_pktmbuf_mtod_offset(pkts[4], void *,
+                                               sizeof(struct rte_ether_hdr)));
+                       rte_prefetch0(rte_pktmbuf_mtod_offset(pkts[5], void *,
+                                               sizeof(struct rte_ether_hdr)));
+                       rte_prefetch0(rte_pktmbuf_mtod_offset(pkts[6], void *,
+                                               sizeof(struct rte_ether_hdr)));
+                       rte_prefetch0(rte_pktmbuf_mtod_offset(pkts[7], void *,
+                                               sizeof(struct rte_ether_hdr)));
+               }
+
+               mbuf0 = pkts[0];
+               mbuf1 = pkts[1];
+               mbuf2 = pkts[2];
+               mbuf3 = pkts[3];
+
+               pkts += 4;
+               n_left_from -= 4;
+
+               /* Extract DIP of mbuf0 */
+               ipv4_hdr = rte_pktmbuf_mtod_offset(mbuf0, struct rte_ipv4_hdr *,
+                                               sizeof(struct rte_ether_hdr));
+               dip = vsetq_lane_s32(ipv4_hdr->dst_addr, dip, 0);
+               /* Extract cksum, ttl as ipv4 hdr is in cache */
+               priv01.u16[1] = ipv4_hdr->time_to_live;
+               priv01.u32[1] = ipv4_hdr->hdr_checksum;
+
+               /* Extract DIP of mbuf1 */
+               ipv4_hdr = rte_pktmbuf_mtod_offset(mbuf1, struct rte_ipv4_hdr *,
+                                               sizeof(struct rte_ether_hdr));
+               dip = vsetq_lane_s32(ipv4_hdr->dst_addr, dip, 1);
+               /* Extract cksum, ttl as ipv4 hdr is in cache */
+               priv01.u16[5] = ipv4_hdr->time_to_live;
+               priv01.u32[3] = ipv4_hdr->hdr_checksum;
+
+               /* Extract DIP of mbuf2 */
+               ipv4_hdr = rte_pktmbuf_mtod_offset(mbuf2, struct rte_ipv4_hdr *,
+                                               sizeof(struct rte_ether_hdr));
+               dip = vsetq_lane_s32(ipv4_hdr->dst_addr, dip, 2);
+               /* Extract cksum, ttl as ipv4 hdr is in cache */
+               priv23.u16[1] = ipv4_hdr->time_to_live;
+               priv23.u32[1] = ipv4_hdr->hdr_checksum;
+
+               /* Extract DIP of mbuf3 */
+               ipv4_hdr = rte_pktmbuf_mtod_offset(mbuf3, struct rte_ipv4_hdr *,
+                                               sizeof(struct rte_ether_hdr));
+               dip = vsetq_lane_s32(ipv4_hdr->dst_addr, dip, 3);
+
+               dip = vreinterpretq_s32_u8(
+                       vrev32q_u8(vreinterpretq_u8_s32(dip)));
+               /* Extract cksum, ttl as ipv4 hdr is in cache */
+               priv23.u16[5] = ipv4_hdr->time_to_live;
+               priv23.u32[3] = ipv4_hdr->hdr_checksum;
+
+               /* Perform LPM lookup to get NH and next node */
+               rte_lpm_lookupx4(lpm, dip, result.u32, drop_nh);
+               priv01.u16[0] = result.u16[0];
+               priv01.u16[4] = result.u16[2];
+               priv23.u16[0] = result.u16[4];
+               priv23.u16[4] = result.u16[6];
+
+               node_mbuf_priv1(mbuf0)->u = priv01.u64[0];
+               node_mbuf_priv1(mbuf1)->u = priv01.u64[1];
+               node_mbuf_priv1(mbuf2)->u = priv23.u64[0];
+               node_mbuf_priv1(mbuf3)->u = priv23.u64[1];
+
+               /* Enqueue four to next node */
+               rte_edge_t fix_spec = ((next_index == result.u16[1]) &&
+                                      (result.u16[1] == result.u16[3]) &&
+                                      (result.u16[3] == result.u16[5]) &&
+                                      (result.u16[5] == result.u16[7]));
+
+               if (unlikely(fix_spec == 0)) {
+                       /* Copy things successfully speculated till now */
+                       rte_memcpy(to_next, from, last_spec * sizeof(from[0]));
+                       from += last_spec;
+                       to_next += last_spec;
+                       held += last_spec;
+                       last_spec = 0;
+
+                       /* Next0 */
+                       if (next_index == result.u16[1]) {
+                               to_next[0] = from[0];
+                               to_next++;
+                               held++;
+                       } else {
+                               rte_node_enqueue_x1(graph, node, result.u16[1],
+                                                   from[0]);
+                       }
+
+                       /* Next1 */
+                       if (next_index == result.u16[3]) {
+                               to_next[0] = from[1];
+                               to_next++;
+                               held++;
+                       } else {
+                               rte_node_enqueue_x1(graph, node, result.u16[3],
+                                                   from[1]);
+                       }
+
+                       /* Next2 */
+                       if (next_index == result.u16[5]) {
+                               to_next[0] = from[2];
+                               to_next++;
+                               held++;
+                       } else {
+                               rte_node_enqueue_x1(graph, node, result.u16[5],
+                                                   from[2]);
+                       }
+
+                       /* Next3 */
+                       if (next_index == result.u16[7]) {
+                               to_next[0] = from[3];
+                               to_next++;
+                               held++;
+                       } else {
+                               rte_node_enqueue_x1(graph, node, result.u16[7],
+                                                   from[3]);
+                       }
+
+                       from += 4;
+               } else {
+                       last_spec += 4;
+               }
+       }
+
+       while (n_left_from > 0) {
+               uint32_t next_hop;
+               uint16_t next0;
+
+               mbuf0 = pkts[0];
+
+               pkts += 1;
+               n_left_from -= 1;
+
+               /* Extract DIP of mbuf0 */
+               ipv4_hdr = rte_pktmbuf_mtod_offset(mbuf0, struct rte_ipv4_hdr *,
+                                               sizeof(struct rte_ether_hdr));
+               /* Extract cksum, ttl as ipv4 hdr is in cache */
+               node_mbuf_priv1(mbuf0)->cksum = ipv4_hdr->hdr_checksum;
+               node_mbuf_priv1(mbuf0)->ttl = ipv4_hdr->time_to_live;
+
+               rc = rte_lpm_lookup(lpm, rte_be_to_cpu_32(ipv4_hdr->dst_addr),
+                                   &next_hop);
+               next_hop = (rc == 0) ? next_hop : drop_nh;
+
+               node_mbuf_priv1(mbuf0)->nh = (uint16_t)next_hop;
+               next_hop = next_hop >> 16;
+               next0 = (uint16_t)next_hop;
+
+               if (unlikely(next_index ^ next0)) {
+                       /* Copy things successfully speculated till now */
+                       rte_memcpy(to_next, from, last_spec * sizeof(from[0]));
+                       from += last_spec;
+                       to_next += last_spec;
+                       held += last_spec;
+                       last_spec = 0;
+
+                       rte_node_enqueue_x1(graph, node, next0, from[0]);
+                       from += 1;
+               } else {
+                       last_spec += 1;
+               }
+       }
+
+       /* !!! Home run !!! */
+       if (likely(last_spec == nb_objs)) {
+               rte_node_next_stream_move(graph, node, next_index);
+               return nb_objs;
+       }
+       held += last_spec;
+       rte_memcpy(to_next, from, last_spec * sizeof(from[0]));
+       rte_node_next_stream_put(graph, node, next_index, held);
+
+       return nb_objs;
+}
+
+#endif /* __INCLUDE_IP4_LOOKUP_NEON_H__ */