X-Git-Url: http://git.bitcoin.ninja/index.cgi?a=blobdiff_plain;f=xdp.c;h=29e2cf3e3f716e8f595d5905b00dee62ebdc66d5;hb=0ac7dea853a16579072fe9d85cfc791226167491;hp=74e8e8cb0e850c8556b68877c0faa6de194f51b4;hpb=b2597a12f3a4c500b43b9888b5174385b0227bc3;p=flowspec-xdp diff --git a/xdp.c b/xdp.c index 74e8e8c..29e2cf3 100644 --- a/xdp.c +++ b/xdp.c @@ -7,7 +7,7 @@ #include #include -#define NULL (void*)0 +#include "siphash.h" /* IP flags. */ #define IP_CE 0x8000 /* Flag: "Congestion" */ @@ -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 @@ -117,6 +121,16 @@ 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 { @@ -133,32 +147,135 @@ static const int XDP_PASS = 0; static const int XDP_DROP = 1; static long drop_cnt_map[RULECNT + STATIC_RULE_CNT]; -#define DO_RETURN(reason, ret) { \ - if (ret == XDP_DROP) drop_cnt_map[reason] += 1; \ - return ret; \ - } +#define INCREMENT_MATCH(reason) { drop_cnt_map[reason] += 1; drop_cnt_map[reason] += data_end - pktdata; } -#else +#else /* TEST */ #include #include -struct bpf_map_def SEC("maps") drop_cnt_map = { - .type = BPF_MAP_TYPE_PERCPU_ARRAY, - .key_size = sizeof(uint32_t), - .value_size = sizeof(long), - .max_entries = RULECNT + STATIC_RULE_CNT, +struct match_counter { + uint64_t bytes; + uint64_t packets; }; -#define DO_RETURN(reason, ret) {\ - if (ret == XDP_DROP) { \ - long *value = bpf_map_lookup_elem(&drop_cnt_map, &reason); \ - if (value) \ - *value += 1; \ +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; + int64_t sent_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 */ + +// We implement a rather naive hashtable here instead of using a BPF map because +// (a) the BPF map hashtables are similarly naive (no rehashing, etc), +// (b) the BPF map LRU hashtables don't support locking. +// +// We first separate into a few top-level buckets with per-bucket locks, limiting +// us to 2^SRC_HASH_MAX_PARALLELISM parallel accessors. +// +// Then we build an array of MAX_ENTRIES/2**SRC_HASH_MAX_PARALLELISM_POW entries, +// which are split into buckets of size SRC_HASH_BUCKET_COUNT. An entry can appear +// in any of the SRC_HASH_BUCKET_COUNT buckets at it's hash value. +#define SRC_HASH_MAX_PARALLELISM_POW 9 +#define SRC_HASH_MAX_PARALLELISM (1 << SRC_HASH_MAX_PARALLELISM_POW) +#define SRC_HASH_BUCKET_COUNT_POW 3 +#define SRC_HASH_BUCKET_COUNT (1 << SRC_HASH_BUCKET_COUNT_POW) + +#include "rand.h" + +#define CREATE_PERSRC_LOOKUP(IPV, IP_TYPE) \ +struct persrc_rate##IPV##_entry { \ + int64_t sent_rate; \ + int64_t sent_time; \ + IP_TYPE srcip; \ +}; \ + \ +struct persrc_rate##IPV##_bucket { \ + struct bpf_spin_lock lock; \ + struct persrc_rate##IPV##_entry entries[]; \ +}; \ + \ +struct persrc_rate##IPV##_ptr { \ + struct persrc_rate##IPV##_entry *rate; \ + struct bpf_spin_lock *lock; \ +}; \ + \ +__attribute__((always_inline)) \ +static inline struct persrc_rate##IPV##_ptr get_v##IPV##_persrc_ratelimit(IP_TYPE key, void *map, size_t map_limit) { \ + struct persrc_rate##IPV##_ptr res = { .rate = NULL, .lock = NULL }; \ + uint64_t hash = siphash(&key, sizeof(key), COMPILE_TIME_RAND); \ + \ + const uint32_t map_key = hash % SRC_HASH_MAX_PARALLELISM; \ + struct persrc_rate##IPV##_bucket *buckets = bpf_map_lookup_elem(map, &map_key); \ + if (!buckets) return res; \ + \ + hash >>= SRC_HASH_MAX_PARALLELISM_POW; \ + map_limit >>= SRC_HASH_MAX_PARALLELISM_POW; \ + \ + struct persrc_rate##IPV##_entry *first_bucket = &buckets->entries[(hash % map_limit) & (~(SRC_HASH_BUCKET_COUNT - 1))]; \ + bpf_spin_lock(&buckets->lock); \ + \ + int min_sent_idx = 0; \ + int64_t min_sent_time = INT64_MAX; \ + for (int i = 0; i < SRC_HASH_BUCKET_COUNT; i++) { \ + if (first_bucket[i].srcip == key) { \ + res.rate = &first_bucket[i]; \ + res.lock = &buckets->lock; \ + return res; \ + } else if (min_sent_time > first_bucket[i].sent_time) { \ + min_sent_time = first_bucket[i].sent_time; \ + min_sent_idx = i; \ } \ - return XDP_DROP; \ - } + } \ + res.rate = &first_bucket[min_sent_idx]; \ + res.rate->srcip = key; \ + res.rate->sent_rate = 0; \ + res.rate->sent_time = 0; \ + res.lock = &buckets->lock; \ + return res; \ +} + +CREATE_PERSRC_LOOKUP(6, uint128_t) +CREATE_PERSRC_LOOKUP(4, uint32_t) + +#define SRC_RATE_DEFINE(IPV, n, limit) \ +struct persrc_rate##IPV##_bucket_##n { \ + struct bpf_spin_lock lock; \ + struct persrc_rate##IPV##_entry entries[limit / SRC_HASH_MAX_PARALLELISM]; \ +}; \ +struct { \ + __uint(type, BPF_MAP_TYPE_ARRAY); \ + __uint(max_entries, SRC_HASH_MAX_PARALLELISM); \ + uint32_t *key; \ + struct persrc_rate##IPV##_bucket_##n *value; \ +} v##IPV##_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; @@ -167,56 +284,45 @@ int xdp_drop_prog(struct xdp_md *ctx) unsigned short eth_proto; { - if (unlikely((void*)(size_t)ctx->data + sizeof(struct ethhdr) > data_end)) - DO_RETURN(PKT_LEN_DROP, 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); #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)) - DO_RETURN(PKT_LEN_DROP, 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))) DO_RETURN(VLAN_DROP, XDP_DROP); #endif - eth_proto = eth_vlan->h_proto; - pktdata = (const void *)(long)ctx->data + sizeof(struct ethhdr_vlan); #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 DO_RETURN(VLAN_DROP, XDP_DROP); #else - pktdata = (const void *)(long)ctx->data + sizeof(struct ethhdr); eth_proto = eth->h_proto; #endif } } -#ifdef NEED_V4_PARSE - const struct iphdr *ip = NULL; - const struct icmphdr *icmp = NULL; -#endif -#ifdef NEED_V6_PARSE - const struct ip6hdr *ip6 = NULL; - const struct icmp6hdr *icmpv6 = NULL; - const struct ip6_fraghdr *frag6 = NULL; -#endif - const void *l4hdr = NULL; const struct tcphdr *tcp = NULL; - const struct udphdr *udp = NULL; + int32_t sport = -1, dport = -1; // 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)) - DO_RETURN(PKT_LEN_DROP, XDP_DROP); - ip = (struct iphdr*) pktdata; + CHECK_LEN(pktdata, iphdr); + struct iphdr *ip = (struct iphdr*) pktdata; #if PARSE_IHL == PARSE if (unlikely(ip->ihl < 5)) DO_RETURN(IHL_DROP, XDP_DROP); @@ -226,37 +332,40 @@ int xdp_drop_prog(struct xdp_md *ctx) l4hdr = pktdata + 5*4; #endif + const struct icmphdr *icmp = NULL; if ((ip->frag_off & BE16(IP_OFFSET)) == 0) { if (ip->protocol == IP_PROTO_TCP) { - if (unlikely(l4hdr + sizeof(struct tcphdr) > data_end)) - DO_RETURN(PKT_LEN_DROP, XDP_DROP); + CHECK_LEN(l4hdr, tcphdr); tcp = (struct tcphdr*) l4hdr; + sport = BE16(tcp->source); + dport = BE16(tcp->dest); } else if (ip->protocol == IP_PROTO_UDP) { - if (unlikely(l4hdr + sizeof(struct udphdr) > data_end)) - DO_RETURN(PKT_LEN_DROP, XDP_DROP); - udp = (struct udphdr*) l4hdr; + CHECK_LEN(l4hdr, udphdr); + const struct udphdr *udp = (struct udphdr*) l4hdr; + sport = BE16(udp->source); + dport = BE16(udp->dest); } else if (ip->protocol == IP_PROTO_ICMP) { - if (unlikely(l4hdr + sizeof(struct icmphdr) > data_end)) - DO_RETURN(PKT_LEN_DROP, XDP_DROP); + CHECK_LEN(l4hdr, icmphdr); icmp = (struct icmphdr*) l4hdr; } } + + RULES4 } #endif #ifdef NEED_V6_PARSE if (eth_proto == BE16(ETH_P_IPV6)) { - if (unlikely(pktdata + sizeof(struct ip6hdr) > data_end)) - DO_RETURN(PKT_LEN_DROP, XDP_DROP); - ip6 = (struct ip6hdr*) pktdata; + CHECK_LEN(pktdata, ip6hdr); + struct ip6hdr *ip6 = (struct ip6hdr*) pktdata; l4hdr = pktdata + 40; 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)) - DO_RETURN(PKT_LEN_DROP, XDP_DROP); + CHECK_LEN(l4hdr, ip6_fraghdr); frag6 = (struct ip6_fraghdr*) l4hdr; l4hdr = l4hdr + sizeof(struct ip6_fraghdr); v6nexthdr = frag6->nexthdr; @@ -268,34 +377,27 @@ int xdp_drop_prog(struct xdp_md *ctx) #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) { - if (unlikely(l4hdr + sizeof(struct tcphdr) > data_end)) - DO_RETURN(PKT_LEN_DROP, XDP_DROP); + CHECK_LEN(l4hdr, tcphdr); tcp = (struct tcphdr*) l4hdr; + sport = BE16(tcp->source); + dport = BE16(tcp->dest); } else if (v6nexthdr == IP_PROTO_UDP) { - if (unlikely(l4hdr + sizeof(struct udphdr) > data_end)) - DO_RETURN(PKT_LEN_DROP, XDP_DROP); - udp = (struct udphdr*) l4hdr; + CHECK_LEN(l4hdr, udphdr); + const struct udphdr *udp = (struct udphdr*) l4hdr; + sport = BE16(udp->source); + dport = BE16(udp->dest); } else if (v6nexthdr == IP6_PROTO_ICMPV6) { - if (unlikely(l4hdr + sizeof(struct icmp6hdr) > data_end)) - DO_RETURN(PKT_LEN_DROP, XDP_DROP); + CHECK_LEN(l4hdr, icmp6hdr); icmpv6 = (struct icmp6hdr*) l4hdr; } } - } -#endif - 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; } @@ -304,7 +406,7 @@ 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,