Fix/better handling of no-stats-tracking rules
[flowspec-xdp] / genrules.py
index db138a8724e7ec4169d66db37bc95cbfbe3eb63d..7e71d755fbd43e1187e76651b3e4c3ede8716eeb 100755 (executable)
@@ -45,20 +45,26 @@ class ASTNode:
         if self.action == ASTAction.EXPR:
             return self.left.write(expr_param, expr_param2)
 
-def parse_ast(expr, parse_expr):
+def parse_ast(expr, parse_expr, comma_is_or):
     expr = expr.strip()
 
     comma_split = expr.split(",", 1)
     or_split = expr.split("||", 1)
     if len(comma_split) > 1 and not "||" in comma_split[0]:
-        return ASTNode(ASTAction.OR, parse_ast(comma_split[0], parse_expr), parse_ast(comma_split[1], parse_expr))
+        # Confusingly, BIRD uses `,` as either || or &&, depending on the type
+        # of expression being parsed. Specifically, a `numbers-match` uses `,`
+        # as OR, whereas a `bitmask-match` uses `,` as AND.
+        if comma_is_or:
+            return ASTNode(ASTAction.OR, parse_ast(comma_split[0], parse_expr, comma_is_or), parse_ast(comma_split[1], parse_expr, comma_is_or))
+        else:
+            return ASTNode(ASTAction.AND, parse_ast(comma_split[0], parse_expr, comma_is_or), parse_ast(comma_split[1], parse_expr, comma_is_or))
     if len(or_split) > 1:
         assert not "," in or_split[0]
-        return ASTNode(ASTAction.OR, parse_ast(or_split[0], parse_expr), parse_ast(or_split[1], parse_expr))
+        return ASTNode(ASTAction.OR, parse_ast(or_split[0], parse_expr, comma_is_or), parse_ast(or_split[1], parse_expr, comma_is_or))
 
     and_split = expr.split("&&", 1)
     if len(and_split) > 1:
-        return ASTNode(ASTAction.AND, parse_ast(and_split[0], parse_expr), parse_ast(and_split[1], parse_expr))
+        return ASTNode(ASTAction.AND, parse_ast(and_split[0], parse_expr, comma_is_or), parse_ast(and_split[1], parse_expr, comma_is_or))
 
     if expr.strip() == "true":
         return ASTNode(ASTAction.TRUE, None)
@@ -66,7 +72,7 @@ def parse_ast(expr, parse_expr):
         return ASTNode(ASTAction.FALSE, None)
 
     if expr.startswith("!"):
-        return ASTNode(ASTAction.NOT, parse_ast(expr[1:], parse_expr))
+        return ASTNode(ASTAction.NOT, parse_ast(expr[1:], parse_expr, comma_is_or))
 
     return parse_expr(expr)
 
@@ -193,15 +199,15 @@ def ip_to_rule(proto, inip, ty, offset):
        break;"""
 
 def fragment_to_rule(ipproto, rules):
-    ast = parse_ast(rules, parse_frag_expr)
+    ast = parse_ast(rules, parse_frag_expr, False)
     return "if (!( " + ast.write(ipproto) + " )) break;"
 
 def len_to_rule(rules):
-    ast = parse_ast(rules, parse_numbers_expr)
+    ast = parse_ast(rules, parse_numbers_expr, True)
     return "if (!( " + ast.write("(data_end - pktdata)") + " )) break;"
  
 def proto_to_rule(ipproto, proto):
-    ast = parse_ast(proto, parse_numbers_expr)
+    ast = parse_ast(proto, parse_numbers_expr, True)
 
     if ipproto == 4:
         return "if (!( " + ast.write("ip->protocol") + " )) break;"
@@ -209,21 +215,21 @@ def proto_to_rule(ipproto, proto):
         return "if (!( " + ast.write("ip6->nexthdr") + " )) break;"
 
 def icmp_type_to_rule(proto, ty):
-    ast = parse_ast(ty, parse_numbers_expr)
+    ast = parse_ast(ty, parse_numbers_expr, True)
     if proto == 4:
         return "if (icmp == NULL) break;\nif (!( " + ast.write("icmp->type") + " )) break;"
     else:
         return "if (icmpv6 == NULL) break;\nif (!( " + ast.write("icmpv6->icmp6_type") + " )) break;"
 
 def icmp_code_to_rule(proto, code):
-    ast = parse_ast(code, parse_numbers_expr)
+    ast = parse_ast(code, parse_numbers_expr, True)
     if proto == 4:
         return "if (icmp == NULL) break;\nif (!( " + ast.write("icmp->code") + " )) break;"
     else:
         return "if (icmpv6 == NULL) break;\nif (!( " + ast.write("icmpv6->icmp6_code") + " )) break;"
 
 def dscp_to_rule(proto, rules):
-    ast = parse_ast(rules, parse_numbers_expr)
+    ast = parse_ast(rules, parse_numbers_expr, True)
 
     if proto == 4:
         return "if (!( " + ast.write("((ip->tos & 0xfc) >> 2)") + " )) break;"
@@ -232,20 +238,20 @@ def dscp_to_rule(proto, rules):
 
 def port_to_rule(ty, rules):
     if ty == "port" :
-        ast = parse_ast(rules, parse_numbers_expr)
-        return "if (!ports_valid) break;\nif (!( " + ast.write("sport", "dport") + " )) break;"
+        ast = parse_ast(rules, parse_numbers_expr, True)
+        return "if (sport == -1 || dport == -1) break;\nif (!( " + ast.write("sport", "dport") + " )) break;"
 
-    ast = parse_ast(rules, parse_numbers_expr)
-    return "if (!ports_valid) break;\nif (!( " + ast.write(ty) + " )) break;"
+    ast = parse_ast(rules, parse_numbers_expr, True)
+    return "if (" + ty + " == -1) break;\nif (!( " + ast.write(ty) + " )) break;"
 
 def tcp_flags_to_rule(rules):
-    ast = parse_ast(rules, parse_bit_expr)
+    ast = parse_ast(rules, parse_bit_expr, False)
 
     return f"""if (tcp == NULL) break;
 if (!( {ast.write("(ntohs(tcp->flags) & 0xfff)")} )) break;"""
 
 def flow_label_to_rule(rules):
-    ast = parse_ast(rules, parse_bit_expr)
+    ast = parse_ast(rules, parse_bit_expr, False)
 
     return f"""if (ip6 == NULL) break;
 if (!( {ast.write("((((uint32_t)(ip6->flow_lbl[0] & 0xf)) << 2*8) | (((uint32_t)ip6->flow_lbl[1]) << 1*8) | (uint32_t)ip6->flow_lbl[0])")} )) break;"""
@@ -288,8 +294,11 @@ with open("rules.h", "w") as out:
     rules6 = ""
     rules4 = ""
     use_v6_frags = False
-    rulecnt = 0
+    stats_rulecnt = 0
     ratelimitcnt = 0
+    v4persrcratelimits = []
+    v5persrcratelimits = []
+    v6persrcratelimits = []
 
     lastrule = None
     for line in sys.stdin.readlines():
@@ -360,6 +369,7 @@ with open("rules.h", "w") as out:
 
             # Now write the match handling!
             first_action = None
+            stats_action = ""
             last_action = None
             for community in line.split("("):
                 if not community.startswith("generic, "):
@@ -369,66 +379,96 @@ with open("rules.h", "w") as out:
                 if len(blocks[1].strip()) != 10: # Should be 0x12345678
                     continue
                 ty = blocks[1].strip()[:6]
+                high_byte = int(blocks[1].strip()[6:8], 16)
+                mid_byte = int(blocks[1].strip()[8:], 16)
                 low_bytes = int(blocks[2].strip(") \n"), 16)
-                if ty == "0x8006" or ty == "0x800c":
+                if ty == "0x8006" or ty == "0x800c" or ty == "0x8306" or ty == "0x830c":
                     if first_action is not None:
                         # Two ratelimit actions, just drop the old one. RFC 8955 says we can.
                         first_action = None
                     exp = (low_bytes & (0xff << 23)) >> 23
                     if low_bytes == 0:
-                        first_action = "return XDP_DROP;"
+                        first_action = "{stats_replace}\nreturn XDP_DROP;"
                     elif low_bytes & (1 <<  31) != 0:
                         # Negative limit, just drop
-                        first_action = "return XDP_DROP;"
+                        first_action = "{stats_replace}\nreturn XDP_DROP;"
                     elif exp == 0xff:
                         # NaN/INF. Just treat as INF and accept
                         first_action = None
-                    elif exp <= 127: # < 1
-                        first_action = "return XDP_DROP;"
-                    elif exp >= 127 + 63: # The count won't even fit in 64-bits, just accept
+                    elif exp < 127: # < 1
+                        first_action = "{stats_replace}\nreturn XDP_DROP;"
+                    elif exp >= 127 + 29: # We can't handle the precision required with ns this high
                         first_action = None
                     else:
                         mantissa = low_bytes & ((1 << 23) - 1)
                         value = 1.0 + mantissa / (2**23)
                         value *= 2**(exp-127)
-                        if ty == "0x8006":
-                            accessor = "rate->rate.sent_bytes"
+
+                        first_action =   "int64_t time_masked = bpf_ktime_get_ns() & RATE_TIME_MASK;\n"
+                        first_action += f"int64_t per_pkt_ns = (1000000000LL << RATE_BUCKET_INTEGER_BITS) / {math.floor(value)};\n"
+                        if ty == "0x8006" or ty == "0x800c":
+                            spin_lock = "bpf_spin_lock(&rate->lock);"
+                            spin_unlock = "bpf_spin_unlock(&rate->lock);"
+                            first_action += f"const uint32_t ratelimitidx = {ratelimitcnt};\n"
+                            first_action += "struct ratelimit *rate = bpf_map_lookup_elem(&rate_map, &ratelimitidx);\n"
+                            ratelimitcnt += 1
                         else:
-                            accessor = "rate->rate.sent_packets"
-                        # Note that int64_t will overflow after 292 years of uptime
-                        first_action = "int64_t time = bpf_ktime_get_ns();\n"
-                        first_action += f"const uint32_t ratelimitidx = {ratelimitcnt};\n"
-                        first_action +=  "uint64_t allowed_since_last = 0;\n"
-                        first_action +=  "struct ratelimit *rate = bpf_map_lookup_elem(&rate_map, &ratelimitidx);\n"
-                        first_action +=  "if (rate) {\n"
-                        first_action +=  "\tbpf_spin_lock(&rate->lock);\n"
-                        first_action += f"\tif (likely({accessor} > 0))" + " {\n"
-                        first_action +=  "\t\tint64_t diff = time - rate->sent_time;\n"
-                        # Unlikely or not, if the flow is slow, take a perf hit (though with the else if branch it doesn't matter)
-                        first_action +=  "\t\tif (unlikely(diff > 1000000000))\n"
-                        first_action += f"\t\t\t{accessor} = 0;\n"
-                        first_action +=  "\t\telse if (likely(diff > 0))\n"
-                        first_action += f"\t\t\tallowed_since_last = ((uint64_t)diff) * {math.floor(value)} / 1000000000;\n"
-                        first_action +=  "\t}\n"
-                        first_action += f"\tif ({accessor} - ((int64_t)allowed_since_last) <= 0)" + " {\n"
-                        if ty == "0x8006":
-                            first_action += f"\t\t{accessor} = data_end - pktdata;\n"
+                            spin_lock = "/* No locking as we're locked in get_v*_persrc_ratelimit */"
+                            spin_unlock = "bpf_spin_unlock(rate_ptr.lock);"
+                            if proto == 4:
+                                if mid_byte > 32:
+                                    continue
+                                first_action += f"const uint32_t srcip = ip->saddr & MASK4({mid_byte});\n"
+                                first_action += f"void *rate_map = &v4_src_rate_{len(v4persrcratelimits)};\n"
+                                first_action += f"struct persrc_rate4_ptr rate_ptr = get_v4_persrc_ratelimit(srcip, rate_map, {(high_byte + 1) * 4096}, time_masked);\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}, time_masked);\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
+                                first_action += f"const uint128_t srcip = ip6->saddr & MASK6({mid_byte});\n"
+                                first_action += f"void *rate_map = &v6_src_rate_{len(v6persrcratelimits)};\n"
+                                first_action += f"struct persrc_rate6_ptr rate_ptr = get_v6_persrc_ratelimit(srcip, rate_map, {(high_byte + 1) * 4096}, time_masked);\n"
+                                first_action += f"struct persrc_rate6_entry *rate = rate_ptr.rate;\n"
+                                v6persrcratelimits.append((high_byte + 1) * 4096)
+                        if ty == "0x8006" or ty == "0x8306":
+                            first_action += "uint64_t amt = data_end - pktdata;\n"
                         else:
-                            first_action += f"\t\t{accessor} = 1;\n"
-                        first_action +=  "\t\trate->sent_time = time;\n"
-                        first_action +=  "\t\tbpf_spin_unlock(&rate->lock);\n"
+                            first_action += "uint64_t amt = 1;\n"
+                        first_action +=  "if (rate) {\n"
+                        first_action += f"\t{spin_lock}\n"
+                        first_action +=  "\tint64_t bucket_pkts = (rate->sent_time & (~RATE_TIME_MASK)) >> (64 - RATE_BUCKET_BITS);\n"
+                        # We mask the top 12 bits, so date overflows every 52 days, handled below
+                        first_action +=  "\tint64_t time_diff = time_masked - ((int64_t)(rate->sent_time & RATE_TIME_MASK));\n"
+                        first_action +=  "\tif (unlikely(time_diff < -1000000000 || time_diff > 16000000000)) {\n"
+                        first_action +=  "\t\tbucket_pkts = 0;\n"
                         first_action +=  "\t} else {\n"
-                        first_action +=  "\t\tbpf_spin_unlock(&rate->lock);\n"
+                        first_action +=  "\t\tif (unlikely(time_diff < 0)) { time_diff = 0; }\n"
+                        first_action += f"\t\tint64_t pkts_since_last = (time_diff << RATE_BUCKET_BITS) * amt / per_pkt_ns;\n"
+                        first_action +=  "\t\tbucket_pkts -= pkts_since_last;\n"
+                        first_action +=  "\t}\n"
+                        first_action +=  "\tif (bucket_pkts >= (((1 << RATE_BUCKET_INTEGER_BITS) - 1) << RATE_BUCKET_DECIMAL_BITS)) {\n"
+                        first_action += f"\t\t{spin_unlock}\n"
+                        first_action +=  "\t\t{stats_replace}\n"
                         first_action +=  "\t\treturn XDP_DROP;\n"
+                        first_action +=  "\t} else {\n"
+                        first_action +=  "\t\tif (unlikely(bucket_pkts < 0)) bucket_pkts = 0;\n"
+                        first_action += f"\t\trate->sent_time = time_masked | ((bucket_pkts + (1 << RATE_BUCKET_DECIMAL_BITS)) << (64 - RATE_BUCKET_BITS));\n"
+                        first_action += f"\t\t{spin_unlock}\n"
                         first_action +=  "\t}\n"
                         first_action +=  "}\n"
-                        ratelimitcnt += 1
                 elif ty == "0x8007":
                     if low_bytes & 1 == 0:
                         last_action = "return XDP_PASS;"
                     if low_bytes & 2 == 2:
-                        write_rule(f"const uint32_t ruleidx = STATIC_RULE_CNT + {rulecnt};")
-                        write_rule("INCREMENT_MATCH(ruleidx);")
+                        stats_action = f"const uint32_t ruleidx = STATIC_RULE_CNT + {stats_rulecnt};\n"
+                        stats_action += "INCREMENT_MATCH(ruleidx);"
                 elif ty == "0x8008":
                     assert False # We do not implement the redirect action
                 elif ty == "0x8009":
@@ -439,24 +479,29 @@ with open("rules.h", "w") as out:
                         write_rule("uint8_t orig_tos = ip->tos;")
                         write_rule("ip->tos = (ip->tos & 3) | " + str(low_bytes << 2) + ";")
                         write_rule("chk = (chk - orig_tos + ip->tos);")
-                        write_rule("if (unlikely(chk < 0)) { chk += 65534; }")
+                        write_rule("if (unlikely(chk > 0xffff)) { chk -= 65535; }")
+                        write_rule("else if (unlikely(chk < 0)) { chk += 65535; }")
                         write_rule("ip->check = ~BE16(chk);")
                     else:
                         write_rule("ip6->priority = " + str(low_bytes >> 2) + ";")
                         write_rule("ip6->flow_lbl[0] = (ip6->flow_lbl[0] & 0x3f) | " + str((low_bytes & 3) << 6) + ";")
             if first_action is not None:
-                write_rule(first_action)
+                write_rule(first_action.replace("{stats_replace}", stats_action))
+            if stats_action != "" and (first_action is None or "{stats_replace}" not in first_action):
+                write_rule(stats_action)
             if last_action is not None:
                 write_rule(last_action)
             if proto == 6:
                 rules6 += "\t} while(0);\\\n"
             else:
                 rules4 += "\t} while(0);\\\n"
-            rulecnt += 1
+            if stats_action != "":
+                print(rule)
+                stats_rulecnt += 1
             lastrule = None
 
     out.write("\n")
-    out.write(f"#define RULECNT {rulecnt}\n")
+    out.write(f"#define STATS_RULECNT {stats_rulecnt}\n")
     if ratelimitcnt != 0:
         out.write(f"#define RATE_CNT {ratelimitcnt}\n")
     if rules4 != "":
@@ -468,3 +513,10 @@ with open("rules.h", "w") as out:
     if args.v6frag == "ignore-parse-if-rule":
         if use_v6_frags:
             out.write("#define PARSE_V6_FRAG PARSE\n")
+    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")