X-Git-Url: http://git.bitcoin.ninja/index.cgi?a=blobdiff_plain;f=xdp.c;h=a32f605d3c4219f8713dda28b83452f4a7ab8f03;hb=36cf3e162d8ca62a573f7b7989a9ec5b50ebba8c;hp=26a12313097539af3477af65dece5d8945016d98;hpb=e6d17f2612e4274d79de87b5d6c0cb992c1aa888;p=flowspec-xdp diff --git a/xdp.c b/xdp.c index 26a1231..a32f605 100644 --- a/xdp.c +++ b/xdp.c @@ -77,9 +77,12 @@ struct tcphdr { // Note that all operations on uint128s *stay* in Network byte order! #if defined(__LITTLE_ENDIAN) -#define BIGEND32(v) ((v >> 3*8) | ((v >> 8) & 0xff00) | ((v << 8) & 0xff0000) | (v << 3*8) & 0xff000000) +#define BIGEND32(v) (((((uint32_t)(v)) >> 3*8) & 0xff) | \ + ((((uint32_t)(v)) >> 1*8) & 0xff00) | \ + ((((uint32_t)(v)) << 1*8) & 0xff0000) | \ + ((((uint32_t)(v)) << 3*8) & 0xff000000)) #elif defined(__BIG_ENDIAN) -#define BIGEND32(v) (v) +#define BIGEND32(v) ((uint32_t)(v)) #else #error "Need endian info" #endif @@ -92,10 +95,11 @@ struct tcphdr { (((uint128_t)BIGEND32(a)) << 0*32)) #define HTON128(a) BIGEND128(a >> 3*32, a >> 2*32, a >> 1*32, a>> 0*32) // Yes, somehow macro'ing this changes LLVM's view of htons... -#define BE16(a) ((((uint16_t)(a & 0xff00)) >> 8) | (((uint16_t)(a & 0xff)) << 8)) +#define BE16(a) (((((uint16_t)a) & 0xff00) >> 8) | ((((uint16_t)a) & 0xff) << 8)) #elif defined(__BIG_ENDIAN) -#define BIGEND128(a, b, c, d) ((((uint128_t)a) << 3*32) | (((uint128_t)b) << 2*32) | (((uint128_t)c) << 1*32) | (((uint128_t)d) << 0*32)) -#define HTON128(a) (a) +#define BIGEND128(a, b, c, d) ((((uint128_t)(a)) << 3*32) | (((uint128_t)(b)) << 2*32) | (((uint128_t)(c)) << 1*32) | (((uint128_t)(d)) << 0*32)) +#define HTON128(a) ((uint128_t)(a)) +#define BE16(a) ((uint16_t)(a)) #else #error "Need endian info" #endif @@ -104,13 +108,29 @@ struct tcphdr { #define MASK6(pfxlen) HTON128(~((((uint128_t)1) << (128 - pfxlen)) - 1)) #define MASK6_OFFS(offs, pfxlen) HTON128((~((((uint128_t)1) << (128 - pfxlen)) - 1)) & ((((uint128_t)1) << (128 - offs)) - 1)) -// Note rules.h may also define PARSE_8021Q and REQ_8021Q -// Note rules.h may also define PARSE_IHL +// PARSE is used as a preprocessor flag to indicate parsing fields +#define PARSE 42 #include "rules.h" #define unlikely(a) __builtin_expect(a, 0) #define likely(a) __builtin_expect(a, 1) +static const uint32_t PKT_LEN_DROP = 0; +static const uint32_t VLAN_DROP = 1; +static const uint32_t IHL_DROP = 2; +static const uint32_t V6FRAG_DROP = 3; +#define STATIC_RULE_CNT 4 + +#define DO_RETURN(reason, ret) {\ + if (ret == XDP_DROP) { INCREMENT_MATCH(reason); } \ + return ret; \ + } + +// It seems (based on drop counts) that data_end points to the last byte, not one-past-the-end. +// This feels strange, but some documentation suggests > here as well, so we stick with that. +#define CHECK_LEN(start, struc) \ + if (unlikely((void*)(start) + sizeof(struct struc) > data_end)) DO_RETURN(PKT_LEN_DROP, XDP_DROP); + #ifdef TEST // 64 bit version of xdp_md for testing struct xdp_md { @@ -125,12 +145,84 @@ struct xdp_md { }; static const int XDP_PASS = 0; static const int XDP_DROP = 1; -#else + +static long drop_cnt_map[RULECNT + STATIC_RULE_CNT]; +#define INCREMENT_MATCH(reason) { drop_cnt_map[reason] += 1; drop_cnt_map[reason] += data_end - pktdata; } + +#else /* TEST */ #include #include +struct match_counter { + uint64_t bytes; + uint64_t packets; +}; +struct { + __uint(type, BPF_MAP_TYPE_PERCPU_ARRAY); + __uint(max_entries, RULECNT + STATIC_RULE_CNT); + __u32 *key; + struct match_counter *value; +} drop_cnt_map SEC(".maps"); + +#define INCREMENT_MATCH(reason) { \ + struct match_counter *value = bpf_map_lookup_elem(&drop_cnt_map, &reason); \ + if (value) { \ + value->bytes += data_end - pktdata; \ + value->packets += 1; \ + } \ +} + +#ifdef RATE_CNT +struct ratelimit { + struct bpf_spin_lock lock; + union { + int64_t sent_bytes; + int64_t sent_packets; + } rate; + int64_t sent_time; +}; +struct { + __uint(type, BPF_MAP_TYPE_ARRAY); + __uint(max_entries, RATE_CNT); + __u32 *key; + struct ratelimit *value; +} rate_map SEC(".maps"); +#endif /* RATE_CNT */ + +// For per-source rate limiting, we have to use per-CPU hash maps as Linux +// doesn't support spinlocks inside of a LRU_HASH (see if block in +// map_check_btf as of Linux 5.10). +// This isn't exactly accurate, but at least its faster. +struct percpu_ratelimit { + union { + int64_t sent_bytes; + int64_t sent_packets; + } rate; + int64_t sent_time; +}; + +#define V6_SRC_RATE_DEFINE(n, limit) \ +struct { \ + __uint(type, BPF_MAP_TYPE_LRU_PERCPU_HASH); \ + __uint(max_entries, limit); \ + uint128_t *key; \ + struct percpu_ratelimit *value; \ +} v6_src_rate_##n SEC(".maps"); + +#define V4_SRC_RATE_DEFINE(n, limit) \ +struct { \ + __uint(type, BPF_MAP_TYPE_LRU_PERCPU_HASH); \ + __uint(max_entries, limit); \ + __u32 *key; \ + struct percpu_ratelimit *value; \ +} v4_src_rate_##n SEC(".maps"); + +#include "maps.h" + +#ifndef HAVE_WRAPPER // Set this to call xdp_drop externally SEC("xdp_drop") -#endif +#endif /* HAVE_WRAPPER */ +#endif /* not TEST */ int xdp_drop_prog(struct xdp_md *ctx) { const void *const data_end = (void *)(size_t)ctx->data_end; @@ -139,116 +231,125 @@ int xdp_drop_prog(struct xdp_md *ctx) unsigned short eth_proto; { - if (unlikely((void*)(size_t)ctx->data + sizeof(struct ethhdr) > data_end)) - return XDP_DROP; + // DO_RETURN in CHECK_LEN relies on pktdata being set to calculate packet length. + // That said, we don't want to overflow, so just set packet length to 0 here. + pktdata = data_end; + CHECK_LEN((size_t)ctx->data, ethhdr); const struct ethhdr *const eth = (void*)(size_t)ctx->data; + pktdata = (const void *)(long)ctx->data + sizeof(struct ethhdr); -#ifdef PARSE_8021Q +#if PARSE_8021Q == PARSE if (likely(eth->h_proto == BE16(ETH_P_8021Q))) { - if (unlikely((void*)(size_t)ctx->data + sizeof(struct ethhdr_vlan) > data_end)) - return XDP_DROP; + CHECK_LEN((size_t)ctx->data, ethhdr_vlan); const struct ethhdr_vlan *const eth_vlan = (void*)(size_t)ctx->data; - + pktdata = (const void *)(long)ctx->data + sizeof(struct ethhdr_vlan); #ifdef REQ_8021Q if (unlikely((eth_vlan->tci & BE16(0xfff)) != BE16(REQ_8021Q))) - return XDP_DROP; + DO_RETURN(VLAN_DROP, XDP_DROP); #endif - eth_proto = eth_vlan->h_proto; +#else + if (unlikely(eth->h_proto == BE16(ETH_P_8021Q))) { pktdata = (const void *)(long)ctx->data + sizeof(struct ethhdr_vlan); + DO_RETURN(VLAN_DROP, PARSE_8021Q); +#endif } else { #ifdef REQ_8021Q - return XDP_DROP; + DO_RETURN(VLAN_DROP, XDP_DROP); #else - pktdata = (const void *)(long)ctx->data + sizeof(struct ethhdr); eth_proto = eth->h_proto; #endif } -#else - pktdata = (const void *)(long)ctx->data + sizeof(struct ethhdr); - eth_proto = eth->h_proto; -#endif } - const struct tcphdr *tcp = NULL; - const struct udphdr *udp = NULL; - const struct icmphdr *icmp = NULL; - const struct icmp6hdr *icmpv6 = NULL; - const struct ip6_fraghdr *frag6 = NULL; - const struct iphdr *ip = NULL; - const struct ip6hdr *ip6 = NULL; const void *l4hdr = NULL; + const struct tcphdr *tcp = NULL; + uint8_t ports_valid = 0; + uint16_t sport = 0, dport = 0; // Host Endian! Only valid with tcp || udp + +#ifdef NEED_V4_PARSE if (eth_proto == BE16(ETH_P_IP)) { - if (unlikely(pktdata + sizeof(struct iphdr) > data_end)) - return XDP_DROP; - ip = (struct iphdr*) pktdata; + CHECK_LEN(pktdata, iphdr); + struct iphdr *ip = (struct iphdr*) pktdata; -#ifdef PARSE_IHL - if (unlikely(ip->ihl < 5)) return XDP_DROP; +#if PARSE_IHL == PARSE + if (unlikely(ip->ihl < 5)) DO_RETURN(IHL_DROP, XDP_DROP); l4hdr = pktdata + ip->ihl * 4; #else - if (ip->ihl != 5) return XDP_DROP; + if (ip->ihl != 5) DO_RETURN(IHL_DROP, PARSE_IHL); l4hdr = pktdata + 5*4; #endif - if (ip->protocol == IP_PROTO_TCP) { - if (unlikely(l4hdr + sizeof(struct tcphdr) > data_end)) - return XDP_DROP; - tcp = (struct tcphdr*) l4hdr; - } else if (ip->protocol == IP_PROTO_UDP) { - if (unlikely(l4hdr + sizeof(struct udphdr) > data_end)) - return XDP_DROP; - udp = (struct udphdr*) l4hdr; - } else if (ip->protocol == IP_PROTO_ICMP) { - if (unlikely(l4hdr + sizeof(struct icmphdr) > data_end)) - return XDP_DROP; - icmp = (struct icmphdr*) l4hdr; + + const struct icmphdr *icmp = NULL; + if ((ip->frag_off & BE16(IP_OFFSET)) == 0) { + if (ip->protocol == IP_PROTO_TCP) { + CHECK_LEN(l4hdr, tcphdr); + tcp = (struct tcphdr*) l4hdr; + sport = BE16(tcp->source); + dport = BE16(tcp->dest); + ports_valid = 1; + } else if (ip->protocol == IP_PROTO_UDP) { + CHECK_LEN(l4hdr, udphdr); + const struct udphdr *udp = (struct udphdr*) l4hdr; + sport = BE16(udp->source); + dport = BE16(udp->dest); + ports_valid = 1; + } else if (ip->protocol == IP_PROTO_ICMP) { + CHECK_LEN(l4hdr, icmphdr); + icmp = (struct icmphdr*) l4hdr; + } } - } else if (eth_proto == BE16(ETH_P_IPV6)) { - if (unlikely(pktdata + sizeof(struct ip6hdr) > data_end)) - return XDP_DROP; - ip6 = (struct ip6hdr*) pktdata; + + RULES4 + } +#endif +#ifdef NEED_V6_PARSE + if (eth_proto == BE16(ETH_P_IPV6)) { + CHECK_LEN(pktdata, ip6hdr); + struct ip6hdr *ip6 = (struct ip6hdr*) pktdata; l4hdr = pktdata + 40; - uint8_t v6nexthdr; + uint8_t v6nexthdr = ip6->nexthdr; + const struct ip6_fraghdr *frag6 = NULL; +#ifdef PARSE_V6_FRAG +#if PARSE_V6_FRAG == PARSE if (ip6->nexthdr == IP6_PROTO_FRAG) { - if (unlikely(l4hdr + sizeof(struct ip6_fraghdr) > data_end)) - return XDP_DROP; + CHECK_LEN(l4hdr, ip6_fraghdr); frag6 = (struct ip6_fraghdr*) l4hdr; l4hdr = l4hdr + sizeof(struct ip6_fraghdr); v6nexthdr = frag6->nexthdr; - } else { - v6nexthdr = ip6->nexthdr; +#else + if (unlikely(ip6->nexthdr == IP6_PROTO_FRAG)) { + DO_RETURN(V6FRAG_DROP, PARSE_V6_FRAG); +#endif } - - if (v6nexthdr == IP_PROTO_TCP) { - if (unlikely(l4hdr + sizeof(struct tcphdr) > data_end)) - return XDP_DROP; - tcp = (struct tcphdr*) l4hdr; - } else if (v6nexthdr == IP_PROTO_UDP) { - if (unlikely(l4hdr + sizeof(struct udphdr) > data_end)) - return XDP_DROP; - udp = (struct udphdr*) l4hdr; - } else if (v6nexthdr == IP6_PROTO_ICMPV6) { - if (unlikely(l4hdr + sizeof(struct icmp6hdr) > data_end)) - return XDP_DROP; - icmpv6 = (struct icmp6hdr*) l4hdr; +#endif + // TODO: Handle more options? + + const struct icmp6hdr *icmpv6 = NULL; + if (frag6 == NULL || (frag6->frag_off & BE16(IP6_FRAGOFF)) == 0) { + if (v6nexthdr == IP_PROTO_TCP) { + CHECK_LEN(l4hdr, tcphdr); + tcp = (struct tcphdr*) l4hdr; + sport = BE16(tcp->source); + dport = BE16(tcp->dest); + ports_valid = 1; + } else if (v6nexthdr == IP_PROTO_UDP) { + CHECK_LEN(l4hdr, udphdr); + const struct udphdr *udp = (struct udphdr*) l4hdr; + sport = BE16(udp->source); + dport = BE16(udp->dest); + ports_valid = 1; + } else if (v6nexthdr == IP6_PROTO_ICMPV6) { + CHECK_LEN(l4hdr, icmp6hdr); + icmpv6 = (struct icmp6hdr*) l4hdr; + } } - // TODO: Handle some options? - } else { - return XDP_PASS; - } - uint16_t sport, dport; // Host Endian! Only valid with tcp || udp - if (tcp != NULL) { - sport = BE16(tcp->source); - dport = BE16(tcp->dest); - } else if (udp != NULL) { - sport = BE16(udp->source); - dport = BE16(udp->dest); + RULES6 } - - RULES +#endif return XDP_PASS; } @@ -257,7 +358,7 @@ int xdp_drop_prog(struct xdp_md *ctx) #include #include -const char d[] = TEST; +char d[] = TEST; int main() { struct xdp_md test = { .data = (uint64_t)d,