X-Git-Url: http://git.bitcoin.ninja/index.cgi?p=flowspec-xdp;a=blobdiff_plain;f=xdp.c;h=187c220b0261fdaad0acc5ec7d48c6d36a4b86e5;hp=1344acd1aac2558494ad47b58b0ed89caca74f07;hb=HEAD;hpb=e7e5b2c8192b715dcfb574051c8d098c8360d746

diff --git a/xdp.c b/xdp.c
index 1344acd..187c220 100644
--- a/xdp.c
+++ b/xdp.c
@@ -96,10 +96,15 @@ struct tcphdr {
 #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 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) ((uint128_t)(a))
 #define BE16(a) ((uint16_t)(a))
+#define BE128BEHIGH64(val) ((uint64_t)(((uint128_t)(val)) >> 64))
+
 #else
 #error "Need endian info"
 #endif
@@ -146,7 +151,7 @@ struct xdp_md {
 static const int XDP_PASS = 0;
 static const int XDP_DROP = 1;
 
-static long drop_cnt_map[RULECNT + STATIC_RULE_CNT];
+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 /* TEST */
@@ -159,7 +164,7 @@ struct match_counter {
 };
 struct {
 	__uint(type, BPF_MAP_TYPE_PERCPU_ARRAY);
-	__uint(max_entries, RULECNT + STATIC_RULE_CNT);
+	__uint(max_entries, STATS_RULECNT + STATIC_RULE_CNT);
 	__u32 *key;
 	struct match_counter *value;
 } drop_cnt_map SEC(".maps");
@@ -172,11 +177,23 @@ struct {
 	} \
 }
 
+// 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_rate;
-	int64_t sent_time;
+	uint64_t sent_time;
 };
 struct {
 	__uint(type, BPF_MAP_TYPE_ARRAY);
@@ -186,32 +203,119 @@ struct {
 } 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 {
-	int64_t sent_rate;
-	int64_t sent_time;
-};
+// 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; \
+}
 
-#define V6_SRC_RATE_DEFINE(n, limit) \
-struct { \
-	__uint(type, BPF_MAP_TYPE_LRU_PERCPU_HASH); \
-	__uint(map_flags, BPF_F_NO_COMMON_LRU); \
-	__uint(max_entries, limit); \
-	uint128_t *key; \
-	struct percpu_ratelimit *value; \
-} v6_src_rate_##n SEC(".maps");
-
-#define V4_SRC_RATE_DEFINE(n, limit) \
+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_LRU_PERCPU_HASH); \
-	__uint(map_flags, BPF_F_NO_COMMON_LRU); \
-	__uint(max_entries, limit); \
-	__u32 *key; \
-	struct percpu_ratelimit *value; \
-} v4_src_rate_##n SEC(".maps");
+	__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"