#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
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 */
};
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");
} \
}
+// 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);
// 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
+//
+// 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 3
+#define SRC_HASH_BUCKET_COUNT_POW 4
#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; \
+ uint64_t sent_time; \
IP_TYPE srcip; \
}; \
\
}; \
\
__attribute__((always_inline)) \
-static inline struct persrc_rate##IPV##_ptr get_v##IPV##_persrc_ratelimit(IP_TYPE key, void *map, size_t map_limit) { \
+static inline struct persrc_rate##IPV##_ptr get_v##IPV##_persrc_ratelimit(IP_TYPE key, void *map, size_t map_limit, int64_t cur_time_masked) { \
struct persrc_rate##IPV##_ptr res = { .rate = NULL, .lock = NULL }; \
uint64_t hash = siphash(&key, sizeof(key), COMPILE_TIME_RAND); \
\
bpf_spin_lock(&buckets->lock); \
\
int min_sent_idx = 0; \
- int64_t min_sent_time = INT64_MAX; \
+ uint64_t min_sent_time = UINT64_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; \
+ } \
+ 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; \
} \
} \
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(5, uint64_t) // IPv6 matching no more than a /64
CREATE_PERSRC_LOOKUP(4, uint32_t)
#define SRC_RATE_DEFINE(IPV, n, limit) \