X-Git-Url: http://git.bitcoin.ninja/index.cgi?a=blobdiff_plain;f=genrules.py;h=10f138f277cb2657931488743f0f3f3aa81f2390;hb=9257387510c6760e67d28d24fe520e6ef9cd59d4;hp=51e547d2d28a12f346aac036cd63f052637b757a;hpb=348583843f9499b0f1c335edc684a077bb2c2710;p=flowspec-xdp diff --git a/genrules.py b/genrules.py index 51e547d..10f138f 100755 --- a/genrules.py +++ b/genrules.py @@ -4,6 +4,7 @@ import sys import ipaddress from enum import Enum import argparse +import math IP_PROTO_ICMP = 1 @@ -44,20 +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) @@ -65,7 +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) @@ -192,15 +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;" @@ -208,21 +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;" @@ -231,20 +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;""" @@ -287,7 +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(): @@ -309,9 +320,19 @@ with open("rules.h", "w") as out: else: continue - def write_rule(r): - global rules4, rules6 - if proto == 6: + # LLVM can be pretty bad at optimizing out common subexpressions. Ideally we'd optimize + # by pulling common subexpressions in back-to-back rules out into a single check, but + # that's a bunch of work that LLVM really should do for us. Instead, we blindly guess + # that source-address is the least likely to be a common subexpression and rely on LLVM + # managing to pull out common subexpressions as long as they're the first check(s). By + # placing source-address checks last, LLVM should do at least some work for us. + # See https://bugs.llvm.org/show_bug.cgi?id=52455 + last_checks = "" + def write_rule(r, place_at_end=False): + global rules4, rules6, last_checks + if place_at_end: + last_checks += "\t\t" + r.replace("\n", " \\\n\t\t") + " \\\n" + elif proto == 6: rules6 += "\t\t" + r.replace("\n", " \\\n\t\t") + " \\\n" else: rules4 += "\t\t" + r.replace("\n", " \\\n\t\t") + " \\\n" @@ -326,7 +347,7 @@ with open("rules.h", "w") as out: else: offset = None if step.strip().startswith("src"): - write_rule(ip_to_rule(proto, nets[0], "saddr", offset)) + write_rule(ip_to_rule(proto, nets[0], "saddr", offset), True) else: write_rule(ip_to_rule(proto, nets[0], "daddr", offset)) elif step.strip().startswith("proto") and proto == 4: @@ -356,8 +377,14 @@ with open("rules.h", "w") as out: else: assert False + if proto == 6: + rules6 += last_checks + else: + rules4 += last_checks + # Now write the match handling! first_action = None + stats_action = "" last_action = None for community in line.split("("): if not community.startswith("generic, "): @@ -367,18 +394,74 @@ 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": + 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 = "{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 = "{stats_replace}\nreturn XDP_DROP;" + elif exp >= 127 + 29: # We can't handle the precision required with ns this high + first_action = None else: - assert False # Not yet supported + mantissa = low_bytes & ((1 << 23) - 1) + value = 1.0 + mantissa / (2**23) + value *= 2**(exp-127) + + 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 == "0x8306": + first_action += "uint64_t amt = data_end - pktdata;\n" + else: + first_action += "uint64_t amt = 1;\n" + if ty == "0x8006" or ty == "0x800c": + first_action += f"const uint32_t ratelimitidx = {ratelimitcnt};\n" + first_action += "struct ratelimit *rate = bpf_map_lookup_elem(&rate_map, &ratelimitidx);\n" + ratelimitcnt += 1 + first_action += "int matched = 0;\n" + first_action += "DO_RATE_LIMIT(bpf_spin_lock(&rate->lock), rate, time_masked, amt, per_pkt_ns, matched);\n" + first_action += "if (rate) { bpf_spin_unlock(&rate->lock); }\n" + else: + 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"int matched = check_v4_persrc_ratelimit(srcip, rate_map, {(high_byte + 1) * 4096}, time_masked, amt, per_pkt_ns);\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"int matched = check_v5_persrc_ratelimit(srcip, rate_map, {(high_byte + 1) * 4096}, time_masked, amt, per_pkt_ns);\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"int matched = check_v6_persrc_ratelimit(srcip, rate_map, {(high_byte + 1) * 4096}, time_masked, amt, per_pkt_ns);\n" + v6persrcratelimits.append((high_byte + 1) * 4096) + first_action += "if (matched) {\n" + first_action += "\t{stats_replace}\n" + first_action += "\treturn XDP_DROP;\n" + first_action += "}\n" 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": @@ -389,24 +472,31 @@ 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 != "": out.write("#define NEED_V4_PARSE\n") out.write("#define RULES4 {\\\n" + rules4 + "}\n") @@ -416,3 +506,10 @@ 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")