drivers/net/tap/tap_bpf_program.c

   1 /* SPDX-License-Identifier: BSD-3-Clause OR GPL-2.0
   2  * Copyright 2017 Mellanox Technologies, Ltd.
   3  */
   4
   5 #include <stdint.h>
   6 #include <stdbool.h>
   7 #include <sys/types.h>
   8 #include <sys/socket.h>
   9 #include <asm/types.h>
  10 #include <linux/in.h>
  11 #include <linux/if.h>
  12 #include <linux/if_ether.h>
  13 #include <linux/ip.h>
  14 #include <linux/ipv6.h>
  15 #include <linux/if_tunnel.h>
  16 #include <linux/filter.h>
  17 #include <linux/bpf.h>
  18
  19 #include "tap_rss.h"
  20
  21 /** Create IPv4 address */
  22 #define IPv4(a, b, c, d) ((__u32)(((a) & 0xff) << 24) | \
  23                 (((b) & 0xff) << 16) | \
  24                 (((c) & 0xff) << 8)  | \
  25                 ((d) & 0xff))
  26
  27 #define PORT(a, b) ((__u16)(((a) & 0xff) << 8) | \
  28                 ((b) & 0xff))
  29
  30 /*
  31  * The queue number is offset by 1, to distinguish packets that have
  32  * gone through this rule (skb->cb[1] != 0) from others.
  33  */
  34 #define QUEUE_OFFSET            1
  35 #define PIN_GLOBAL_NS           2
  36
  37 #define KEY_IDX                 0
  38 #define BPF_MAP_ID_KEY  1
  39
  40 struct vlan_hdr {
  41         __be16 proto;
  42         __be16 tci;
  43 };
  44
  45 struct bpf_elf_map __attribute__((section("maps"), used))
  46 map_keys = {
  47         .type           =       BPF_MAP_TYPE_HASH,
  48         .id             =       BPF_MAP_ID_KEY,
  49         .size_key       =       sizeof(__u32),
  50         .size_value     =       sizeof(struct rss_key),
  51         .max_elem       =       256,
  52         .pinning        =       PIN_GLOBAL_NS,
  53 };
  54
  55 __section("cls_q") int
  56 match_q(struct __sk_buff *skb)
  57 {
  58         __u32 queue = skb->cb[1];
  59         volatile __u32 q = 0xdeadbeef;
  60         __u32 match_queue = QUEUE_OFFSET + q;
  61
  62         /* printt("match_q$i() queue = %d\n", queue); */
  63
  64         if (queue != match_queue)
  65                 return TC_ACT_OK;
  66         return TC_ACT_UNSPEC;
  67 }
  68
  69
  70 struct ipv4_l3_l4_tuple {
  71         __u32    src_addr;
  72         __u32    dst_addr;
  73         __u16    dport;
  74         __u16    sport;
  75 } __attribute__((packed));
  76
  77 struct ipv6_l3_l4_tuple {
  78         __u8        src_addr[16];
  79         __u8        dst_addr[16];
  80         __u16       dport;
  81         __u16       sport;
  82 } __attribute__((packed));
  83
  84 static const __u8 def_rss_key[] = {
  85         0xd1, 0x81, 0xc6, 0x2c,
  86         0xf7, 0xf4, 0xdb, 0x5b,
  87         0x19, 0x83, 0xa2, 0xfc,
  88         0x94, 0x3e, 0x1a, 0xdb,
  89         0xd9, 0x38, 0x9e, 0x6b,
  90         0xd1, 0x03, 0x9c, 0x2c,
  91         0xa7, 0x44, 0x99, 0xad,
  92         0x59, 0x3d, 0x56, 0xd9,
  93         0xf3, 0x25, 0x3c, 0x06,
  94         0x2a, 0xdc, 0x1f, 0xfc,
  95 };
  96
  97 static __u32  __attribute__((always_inline))
  98 rte_softrss_be(const __u32 *input_tuple, const uint8_t *rss_key,
  99                 __u8 input_len)
 100 {
 101         __u32 i, j, hash = 0;
 102 #pragma unroll
 103         for (j = 0; j < input_len; j++) {
 104 #pragma unroll
 105                 for (i = 0; i < 32; i++) {
 106                         if (input_tuple[j] & (1 << (31 - i))) {
 107                                 hash ^= ((const __u32 *)def_rss_key)[j] << i |
 108                                 (__u32)((uint64_t)
 109                                 (((const __u32 *)def_rss_key)[j + 1])
 110                                         >> (32 - i));
 111                         }
 112                 }
 113         }
 114         return hash;
 115 }
 116
 117 static int __attribute__((always_inline))
 118 rss_l3_l4(struct __sk_buff *skb)
 119 {
 120         void *data_end = (void *)(long)skb->data_end;
 121         void *data = (void *)(long)skb->data;
 122         __u16 proto = (__u16)skb->protocol;
 123         __u32 key_idx = 0xdeadbeef;
 124         __u32 hash;
 125         struct rss_key *rsskey;
 126         __u64 off = ETH_HLEN;
 127         int j;
 128         __u8 *key = 0;
 129         __u32 len;
 130         __u32 queue = 0;
 131
 132         rsskey = map_lookup_elem(&map_keys, &key_idx);
 133         if (!rsskey) {
 134                 printt("hash(): rss key is not configured\n");
 135                 return TC_ACT_OK;
 136         }
 137         key = (__u8 *)rsskey->key;
 138
 139         /* Get correct proto for 802.1ad */
 140         if (skb->vlan_present && skb->vlan_proto == htons(ETH_P_8021AD)) {
 141                 if (data + ETH_ALEN * 2 + sizeof(struct vlan_hdr) +
 142                     sizeof(proto) > data_end)
 143                         return TC_ACT_OK;
 144                 proto = *(__u16 *)(data + ETH_ALEN * 2 +
 145                                    sizeof(struct vlan_hdr));
 146                 off += sizeof(struct vlan_hdr);
 147         }
 148
 149         if (proto == htons(ETH_P_IP)) {
 150                 if (data + off + sizeof(struct iphdr) + sizeof(__u32)
 151                         > data_end)
 152                         return TC_ACT_OK;
 153
 154                 __u8 *src_dst_addr = data + off + offsetof(struct iphdr, saddr);
 155                 __u8 *src_dst_port = data + off + sizeof(struct iphdr);
 156                 struct ipv4_l3_l4_tuple v4_tuple = {
 157                         .src_addr = IPv4(*(src_dst_addr + 0),
 158                                         *(src_dst_addr + 1),
 159                                         *(src_dst_addr + 2),
 160                                         *(src_dst_addr + 3)),
 161                         .dst_addr = IPv4(*(src_dst_addr + 4),
 162                                         *(src_dst_addr + 5),
 163                                         *(src_dst_addr + 6),
 164                                         *(src_dst_addr + 7)),
 165                         .sport = PORT(*(src_dst_port + 0),
 166                                         *(src_dst_port + 1)),
 167                         .dport = PORT(*(src_dst_port + 2),
 168                                         *(src_dst_port + 3)),
 169                 };
 170                 __u8 input_len = sizeof(v4_tuple) / sizeof(__u32);
 171                 if (rsskey->hash_fields & (1 << HASH_FIELD_IPV4_L3))
 172                         input_len--;
 173                 hash = rte_softrss_be((__u32 *)&v4_tuple, key, 3);
 174         } else if (proto == htons(ETH_P_IPV6)) {
 175                 if (data + off + sizeof(struct ipv6hdr) +
 176                                         sizeof(__u32) > data_end)
 177                         return TC_ACT_OK;
 178                 __u8 *src_dst_addr = data + off +
 179                                         offsetof(struct ipv6hdr, saddr);
 180                 __u8 *src_dst_port = data + off +
 181                                         sizeof(struct ipv6hdr);
 182                 struct ipv6_l3_l4_tuple v6_tuple;
 183                 for (j = 0; j < 4; j++)
 184                         *((uint32_t *)&v6_tuple.src_addr + j) =
 185                                 __builtin_bswap32(*((uint32_t *)
 186                                                 src_dst_addr + j));
 187                 for (j = 0; j < 4; j++)
 188                         *((uint32_t *)&v6_tuple.dst_addr + j) =
 189                                 __builtin_bswap32(*((uint32_t *)
 190                                                 src_dst_addr + 4 + j));
 191                 v6_tuple.sport = PORT(*(src_dst_port + 0),
 192                               *(src_dst_port + 1));
 193                 v6_tuple.dport = PORT(*(src_dst_port + 2),
 194                               *(src_dst_port + 3));
 195
 196                 __u8 input_len = sizeof(v6_tuple) / sizeof(__u32);
 197                 if (rsskey->hash_fields & (1 << HASH_FIELD_IPV6_L3))
 198                         input_len--;
 199                 hash = rte_softrss_be((__u32 *)&v6_tuple, key, 9);
 200         } else {
 201                 return TC_ACT_PIPE;
 202         }
 203
 204         queue = rsskey->queues[(hash % rsskey->nb_queues) &
 205                                        (TAP_MAX_QUEUES - 1)];
 206         skb->cb[1] = QUEUE_OFFSET + queue;
 207         /* printt(">>>>> rss_l3_l4 hash=0x%x queue=%u\n", hash, queue); */
 208
 209         return TC_ACT_RECLASSIFY;
 210 }
 211
 212 #define RSS(L)                                          \
 213         __section(#L) int                               \
 214                 L ## _hash(struct __sk_buff *skb)       \
 215         {                                               \
 216                 return rss_ ## L (skb);                 \
 217         }
 218
 219 RSS(l3_l4)
 220
 221 BPF_LICENSE("Dual BSD/GPL");