Fix/better handling of no-stats-tracking rules

[flowspec-xdp] / xdp.c

diff --git a/xdp.c b/xdp.c
index 95ab7df9743b85e16d701f79463d989c69b0832c..e63bec379ee143672211c59c9bb4aaab19af9c2a 100644 (file)
--- a/xdp.c
+++ b/xdp.c
@@ -151,7 +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 */
@@ -164,7 +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");
@@ -177,11 +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);
@@ -201,17 +213,24 @@ struct {
 // 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; \
 }; \
  \
@@ -226,7 +245,7 @@ struct persrc_rate##IPV##_ptr { \
 }; \
  \
 __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); \
  \
@@ -241,20 +260,25 @@ static inline struct persrc_rate##IPV##_ptr get_v##IPV##_persrc_ratelimit(IP_TYP
        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; \