Reduce bash CPU time for high-core-count machines in dropcount.sh

[flowspec-xdp] / xdp.c

diff --git a/xdp.c b/xdp.c
index 6a8c19e4f852bad1e108093332043d0033f8e55b..62120d2ef52107511c5f4a4f245c666ef5d4dcb2 100644 (file)
--- a/xdp.c
+++ b/xdp.c
@@ -185,6 +185,11 @@ struct {
 #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;
@@ -208,9 +213,17 @@ 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"
@@ -232,7 +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); \
  \
@@ -253,7 +266,13 @@ static inline struct persrc_rate##IPV##_ptr get_v##IPV##_persrc_ratelimit(IP_TYP
                        res.rate = &first_bucket[i]; \
                        res.lock = &buckets->lock; \
                        return res; \
-               } else if (min_sent_time > (first_bucket[i].sent_time & RATE_TIME_MASK)) { \
+               } \
+               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; \
                } \