X-Git-Url: http://git.bitcoin.ninja/index.cgi?a=blobdiff_plain;f=xdp.c;h=a32f605d3c4219f8713dda28b83452f4a7ab8f03;hb=36cf3e162d8ca62a573f7b7989a9ec5b50ebba8c;hp=506d4224030168ff32f98032199bf144e6630ec0;hpb=2400b2559ac2947300b90173a3fe25d632dba523;p=flowspec-xdp

diff --git a/xdp.c b/xdp.c
index 506d422..a32f605 100644
--- a/xdp.c
+++ b/xdp.c
@@ -77,9 +77,12 @@ struct tcphdr {
 // Note that all operations on uint128s *stay* in Network byte order!
 
 #if defined(__LITTLE_ENDIAN)
-#define BIGEND32(v) ((v >> 3*8) | ((v >> 8) & 0xff00) | ((v << 8) & 0xff0000) | (v << 3*8) & 0xff000000)
+#define BIGEND32(v) (((((uint32_t)(v)) >> 3*8) & 0xff) | \
+                     ((((uint32_t)(v)) >> 1*8) & 0xff00) | \
+                     ((((uint32_t)(v)) << 1*8) & 0xff0000) | \
+                     ((((uint32_t)(v)) << 3*8) & 0xff000000))
 #elif defined(__BIG_ENDIAN)
-#define BIGEND32(v) (v)
+#define BIGEND32(v) ((uint32_t)(v))
 #else
 #error "Need endian info"
 #endif
@@ -92,11 +95,11 @@ struct tcphdr {
 		(((uint128_t)BIGEND32(a)) << 0*32))
 #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 BE16(a) (((((uint16_t)a) & 0xff00) >> 8) | ((((uint16_t)a) & 0xff) << 8))
 #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) (a)
-#define BE16(a) ((uint16_t)a)
+#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))
 #else
 #error "Need endian info"
 #endif
@@ -146,7 +149,7 @@ static const int XDP_DROP = 1;
 static long drop_cnt_map[RULECNT + STATIC_RULE_CNT];
 #define INCREMENT_MATCH(reason) { drop_cnt_map[reason] += 1; drop_cnt_map[reason] += data_end - pktdata; }
 
-#else
+#else /* TEST */
 #include <linux/bpf.h>
 #include <bpf/bpf_helpers.h>
 
@@ -172,7 +175,10 @@ struct {
 #ifdef RATE_CNT
 struct ratelimit {
 	struct bpf_spin_lock lock;
-	int64_t sent_bytes;
+	union {
+		int64_t sent_bytes;
+		int64_t sent_packets;
+	} rate;
 	int64_t sent_time;
 };
 struct {
@@ -181,10 +187,42 @@ struct {
 	__u32 *key;
 	struct ratelimit *value;
 } rate_map SEC(".maps");
-#endif
+#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 {
+	union {
+		int64_t sent_bytes;
+		int64_t sent_packets;
+	} rate;
+	int64_t sent_time;
+};
 
+#define V6_SRC_RATE_DEFINE(n, limit) \
+struct { \
+	__uint(type, BPF_MAP_TYPE_LRU_PERCPU_HASH); \
+	__uint(max_entries, limit); \
+	uint128_t *key; \
+	struct percpu_ratelimit *value; \
+} v6_src_rate_##n SEC(".maps");
+
+#define V4_SRC_RATE_DEFINE(n, limit) \
+struct { \
+	__uint(type, BPF_MAP_TYPE_LRU_PERCPU_HASH); \
+	__uint(max_entries, limit); \
+	__u32 *key; \
+	struct percpu_ratelimit *value; \
+} v4_src_rate_##n SEC(".maps");
+
+#include "maps.h"
+
+#ifndef HAVE_WRAPPER // Set this to call xdp_drop externally
 SEC("xdp_drop")
-#endif
+#endif /* HAVE_WRAPPER */
+#endif /* not TEST */
 int xdp_drop_prog(struct xdp_md *ctx)
 {
 	const void *const data_end = (void *)(size_t)ctx->data_end;
@@ -227,7 +265,7 @@ int xdp_drop_prog(struct xdp_md *ctx)
 	const void *l4hdr = NULL;
 	const struct tcphdr *tcp = NULL;
 	uint8_t ports_valid = 0;
-	uint16_t sport, dport; // Host Endian! Only valid with tcp || udp
+	uint16_t sport = 0, dport = 0; // Host Endian! Only valid with tcp || udp
 
 #ifdef NEED_V4_PARSE
 	if (eth_proto == BE16(ETH_P_IP)) {