#include <linux/icmpv6.h>
#include <arpa/inet.h>
-#define NULL (void*)0
+#include "siphash.h"
/* IP flags. */
#define IP_CE 0x8000 /* Flag: "Congestion" */
// 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
(((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))
+#define BE128BEHIGH64(val) ((uint64_t)((uint128_t)(val)))
+
#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 BE16(a) ((uint16_t)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))
+#define BE128BEHIGH64(val) ((uint64_t)(((uint128_t)(val)) >> 64))
+
#else
#error "Need endian info"
#endif
static const int XDP_PASS = 0;
static const int XDP_DROP = 1;
-static long drop_cnt_map[RULECNT + STATIC_RULE_CNT];
-#define INCREMENT_MATCH(reason) drop_cnt_map[reason] += 1;
+static long drop_cnt_map[STATS_RULECNT + STATIC_RULE_CNT];
+#define INCREMENT_MATCH(reason) { drop_cnt_map[reason] += 1; drop_cnt_map[reason] += data_end - pktdata; }
-#else
+#else /* TEST */
#include <linux/bpf.h>
#include <bpf/bpf_helpers.h>
-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;
};
+struct {
+ __uint(type, BPF_MAP_TYPE_PERCPU_ARRAY);
+ __uint(max_entries, STATS_RULECNT + STATIC_RULE_CNT);
+ __u32 *key;
+ struct match_counter *value;
+} drop_cnt_map SEC(".maps");
+
#define INCREMENT_MATCH(reason) { \
- long *value = bpf_map_lookup_elem(&drop_cnt_map, &reason); \
- if (value) \
- *value += 1; \
+ struct match_counter *value = bpf_map_lookup_elem(&drop_cnt_map, &reason); \
+ if (value) { \
+ value->bytes += data_end - pktdata; \
+ value->packets += 1; \
+ } \
}
+// Rate limits are done in a static-sized leaky bucket with a decimal counter
+// Bucket size is always exactly (1 << RATE_BUCKET_INTEGER_BITS)
+#define RATE_BUCKET_DECIMAL_BITS 8
+#define RATE_BUCKET_INTEGER_BITS 4
+
+#define RATE_BUCKET_BITS (RATE_BUCKET_DECIMAL_BITS + RATE_BUCKET_INTEGER_BITS)
+#define RATE_TIME_MASK ((1ULL << (64 - RATE_BUCKET_BITS)) - 1)
+
+// Time going backwards 10ms+ or forward 32sec+ implies we should consider it
+// an overflow, or at least stale enough that we should reset the entry.
+#define RATE_MIN_TIME_OFFSET -10000000LL
+#define RATE_MAX_TIME_OFFSET 32000000000LL
+
#ifdef RATE_CNT
struct ratelimit {
struct bpf_spin_lock lock;
- int64_t sent_bytes;
- int64_t sent_time;
+ uint64_t sent_time;
};
struct {
__uint(type, BPF_MAP_TYPE_ARRAY);
__u32 *key;
struct ratelimit *value;
} rate_map SEC(".maps");
-#endif
+#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.
+//
+// Because we use buckets of size 16, see collision_prob.py, the number of
+// elements we can hold with only a 1% probability of overflowing a bucket is:
+//
+// 128K-entry hash table (2MiB): ~33K sources
+// 256K-entry hash table (4MiB): ~63K sources
+// 512K-entry hash table (8MiB): ~119K sources
+// 1M-entry hash table (16MiB): ~227K sources
+#define SRC_HASH_MAX_PARALLELISM_POW 8
+#define SRC_HASH_MAX_PARALLELISM (1 << SRC_HASH_MAX_PARALLELISM_POW)
+#define SRC_HASH_BUCKET_COUNT_POW 4
+#define SRC_HASH_BUCKET_COUNT (1 << SRC_HASH_BUCKET_COUNT_POW)
+
+#define DO_RATE_LIMIT(do_lock, rate, time_masked, amt_in_pkt, limit_ns_per_pkt, matchbool) do { \
+if (rate) { \
+ do_lock; \
+ int64_t bucket_pkts = (rate->sent_time & (~RATE_TIME_MASK)) >> (64 - RATE_BUCKET_BITS); \
+ /* We mask the top 12 bits, so date overflows every 52 days, handled below */ \
+ int64_t time_diff = time_masked - ((int64_t)(rate->sent_time & RATE_TIME_MASK)); \
+ if (unlikely(time_diff < -1000000000 || time_diff > 16000000000)) { \
+ bucket_pkts = 0; \
+ } else { \
+ if (unlikely(time_diff < 0)) { time_diff = 0; } \
+ int64_t pkts_since_last = (time_diff << RATE_BUCKET_BITS) * ((uint64_t)amt_in_pkt) / ((uint64_t)limit_ns_per_pkt); \
+ bucket_pkts -= pkts_since_last; \
+ } \
+ if (bucket_pkts < (((1 << RATE_BUCKET_INTEGER_BITS) - 1) << RATE_BUCKET_DECIMAL_BITS)) { \
+ if (unlikely(bucket_pkts < 0)) bucket_pkts = 0; \
+ rate->sent_time = time_masked | ((bucket_pkts + (1 << RATE_BUCKET_DECIMAL_BITS)) << (64 - RATE_BUCKET_BITS)); \
+ matchbool = 0; \
+ } else { \
+ matchbool = 1; \
+ } \
+} \
+} while(0);
+
+#define CREATE_PERSRC_LOOKUP(IPV, IP_TYPE) \
+struct persrc_rate##IPV##_entry { \
+ uint64_t sent_time; \
+ IP_TYPE srcip; \
+}; \
+ \
+struct persrc_rate##IPV##_bucket { \
+ struct bpf_spin_lock lock; \
+ struct persrc_rate##IPV##_entry entries[]; \
+}; \
+ \
+static int check_v##IPV##_persrc_ratelimit(IP_TYPE key, void *map, size_t map_limit, int64_t cur_time_masked, uint64_t amt, uint64_t limit_ns_per_pkt) { \
+ uint64_t hash = siphash_##IP_TYPE(key); \
+ \
+ 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 0; \
+ \
+ 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); \
+ \
+ uint64_t min_sent_idx = 0; /* Must be uint64_t or BPF verifier gets lost and thinks it can be any value */ \
+ uint64_t min_sent_time = UINT64_MAX; \
+ for (uint64_t i = 0; i < SRC_HASH_BUCKET_COUNT; i++) { \
+ if (first_bucket[i].srcip == key) { \
+ min_sent_idx = i; \
+ break; \
+ } \
+ int64_t time_offset = ((int64_t)cur_time_masked) - (first_bucket[i].sent_time & RATE_TIME_MASK); \
+ if (time_offset < RATE_MIN_TIME_OFFSET || time_offset > RATE_MAX_TIME_OFFSET) { \
+ min_sent_idx = i; \
+ break; \
+ } \
+ if ((first_bucket[i].sent_time & RATE_TIME_MASK) < min_sent_time) { \
+ min_sent_time = first_bucket[i].sent_time & RATE_TIME_MASK; \
+ min_sent_idx = i; \
+ } \
+ } \
+ struct persrc_rate##IPV##_entry *entry = &first_bucket[min_sent_idx]; \
+ if (entry->srcip != key) { \
+ entry->srcip = key; \
+ entry->sent_time = 0; \
+ } \
+ int matched = 0; \
+ DO_RATE_LIMIT(, entry, cur_time_masked, amt, limit_ns_per_pkt, matched); \
+ bpf_spin_unlock(&buckets->lock); \
+ return matched; \
+}
+CREATE_PERSRC_LOOKUP(6, uint128_t)
+CREATE_PERSRC_LOOKUP(5, uint64_t) // IPv6 matching no more than a /64
+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;
unsigned short eth_proto;
{
+ // 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))) {
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
}
const void *l4hdr = NULL;
const struct tcphdr *tcp = NULL;
- uint8_t ports_valid = 0;
- uint16_t sport, dport; // Host Endian! Only valid with tcp || udp
+ int32_t sport = -1, dport = -1; // Host Endian! Only valid with tcp || udp
#ifdef NEED_V4_PARSE
if (eth_proto == BE16(ETH_P_IP)) {
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;
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;
projects / flowspec-xdp / blobdiff