Optimize per-src v6 matching on <= /64s to avoid always storing 4 0-bytes

diff --git a/genrules.py b/genrules.py
index 1ab90b05a80bca0b9d1615cb5276489f9cce85d8..b2bf796af9a86e43218f70e2a7c8f21eda034f01 100755 (executable)
--- a/genrules.py
+++ b/genrules.py
@@ -297,6 +297,7 @@ with open("rules.h", "w") as out:
     rulecnt = 0
     ratelimitcnt = 0
     v4persrcratelimits = []
+    v5persrcratelimits = []
     v6persrcratelimits = []
 
     lastrule = None
@@ -422,6 +423,12 @@ with open("rules.h", "w") as out:
                                 first_action += f"struct persrc_rate4_ptr rate_ptr = get_v4_persrc_ratelimit(srcip, rate_map, {(high_byte + 1) * 4096});\n"
                                 first_action += f"struct persrc_rate4_entry *rate = rate_ptr.rate;\n"
                                 v4persrcratelimits.append((high_byte + 1) * 4096)
+                            elif mid_byte <= 64:
+                                first_action += f"const uint64_t srcip = BE128BEHIGH64(ip6->saddr & MASK6({mid_byte}));\n"
+                                first_action += f"void *rate_map = &v5_src_rate_{len(v5persrcratelimits)};\n"
+                                first_action += f"struct persrc_rate5_ptr rate_ptr = get_v5_persrc_ratelimit(srcip, rate_map, {(high_byte + 1) * 4096});\n"
+                                first_action += f"struct persrc_rate5_entry *rate = rate_ptr.rate;\n"
+                                v5persrcratelimits.append((high_byte + 1) * 4096)
                             else:
                                 if mid_byte > 128:
                                     continue
@@ -504,5 +511,7 @@ with open("rules.h", "w") as out:
     with open("maps.h", "w") as out:
         for idx, limit in enumerate(v4persrcratelimits):
             out.write(f"SRC_RATE_DEFINE(4, {idx}, {limit})\n")
+        for idx, limit in enumerate(v5persrcratelimits):
+            out.write(f"SRC_RATE_DEFINE(5, {idx}, {limit})\n")
         for idx, limit in enumerate(v6persrcratelimits):
             out.write(f"SRC_RATE_DEFINE(6, {idx}, {limit})\n")

diff --git a/xdp.c b/xdp.c
index 29e2cf3e3f716e8f595d5905b00dee62ebdc66d5..95ab7df9743b85e16d701f79463d989c69b0832c 100644 (file)
--- 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
@@ -256,6 +261,7 @@ static inline struct persrc_rate##IPV##_ptr get_v##IPV##_persrc_ratelimit(IP_TYP
 }
 
 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) \