X-Git-Url: http://git.bitcoin.ninja/index.cgi?a=blobdiff_plain;f=xdp.c;h=29e2cf3e3f716e8f595d5905b00dee62ebdc66d5;hb=0ac7dea853a16579072fe9d85cfc791226167491;hp=113c4a3306d2ed7fc65b3161cfabfbf348b36690;hpb=e63afbe176e13e0b1e3092837807cd5e06886da6;p=flowspec-xdp

diff --git a/xdp.c b/xdp.c
index 113c4a3..29e2cf3 100644
--- a/xdp.c
+++ b/xdp.c
@@ -7,7 +7,7 @@
 #include <linux/icmpv6.h>
 #include <arpa/inet.h>
 
-#define NULL (void*)0
+#include "siphash.h"
 
 /* IP flags. */
 #define IP_CE		0x8000		/* Flag: "Congestion"		*/
@@ -149,7 +149,7 @@ static const int XDP_DROP = 1;
 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
+#else /* TEST */
 #include <linux/bpf.h>
 #include <bpf/bpf_helpers.h>
 
@@ -175,10 +175,7 @@ struct {
 #ifdef RATE_CNT
 struct ratelimit {
 	struct bpf_spin_lock lock;
-	union {
-		int64_t sent_bytes;
-		int64_t sent_packets;
-	} rate;
+	int64_t sent_rate;
 	int64_t sent_time;
 };
 struct {
@@ -187,12 +184,98 @@ struct {
 	__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.
+#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; \
+		} \
+	} \
+	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
+#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;
@@ -234,8 +317,7 @@ int xdp_drop_prog(struct xdp_md *ctx)
 
 	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)) {
@@ -257,13 +339,11 @@ int xdp_drop_prog(struct xdp_md *ctx)
 				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;
@@ -304,13 +384,11 @@ int xdp_drop_prog(struct xdp_md *ctx)
 				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;