X-Git-Url: http://git.bitcoin.ninja/index.cgi?a=blobdiff_plain;f=genrules.py;h=da42e01732795f85a6c30b3e53913d7ac9a304f9;hb=9779a53938e32d7d81f3c45346731cb06401ae94;hp=4468c1e00e52d58fb12c9d4769ae12c95c7d2704;hpb=a8be3b609743627f58a464d7b75217a778eb4fa9;p=flowspec-xdp diff --git a/genrules.py b/genrules.py index 4468c1e..da42e01 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 @@ -227,7 +228,7 @@ def dscp_to_rule(proto, rules): if proto == 4: return "if (!( " + ast.write("((ip->tos & 0xfc) >> 2)") + " )) break;" else: - return "if (!( " + ast.write("((ip6->priority << 4) | ((ip6->flow_lbl[0] & 0xc0) >> 4) >> 2)") + " )) break;" + return "if (!( " + ast.write("((ip6->priority << 2) | ((ip6->flow_lbl[0] & 0xc0) >> 6))") + " )) break;" def port_to_rule(ty, rules): if ty == "port" : @@ -249,6 +250,7 @@ def flow_label_to_rule(rules): 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;""" + with open("rules.h", "w") as out: parse = argparse.ArgumentParser() parse.add_argument("--ihl", dest="ihl", required=True, choices=["drop-options","accept-options","parse-options"]) @@ -287,76 +289,177 @@ with open("rules.h", "w") as out: rules4 = "" use_v6_frags = False rulecnt = 0 + ratelimitcnt = 0 + lastrule = None for line in sys.stdin.readlines(): - t = line.split("{") - if len(t) != 2: - continue - if t[0].strip() == "flow4": - proto = 4 - rules4 += "\tdo {\\\n" - elif t[0].strip() == "flow6": - proto = 6 - rules6 += "\tdo {\\\n" - else: + if "{" in line: + if lastrule is not None: + print("Skipped rule due to lack of understood community tag: " + lastrule) + lastrule = line continue - - def write_rule(r): - global rules4, rules6 - if proto == 6: - rules6 += "\t\t" + r.replace("\n", " \\\n\t\t") + " \\\n" + if "BGP.ext_community: " in line: + assert lastrule is not None + + t = lastrule.split("{") + if t[0].strip() == "flow4": + proto = 4 + rules4 += "\tdo {\\\n" + elif t[0].strip() == "flow6": + proto = 6 + rules6 += "\tdo {\\\n" else: - rules4 += "\t\t" + r.replace("\n", " \\\n\t\t") + " \\\n" - - rule = t[1].split("}")[0].strip() - for step in rule.split(";"): - if step.strip().startswith("src") or step.strip().startswith("dst"): - nets = step.strip()[3:].strip().split(" ") - if len(nets) > 1: - assert nets[1] == "offset" - offset = nets[2] + continue + + def write_rule(r): + global rules4, rules6 + if proto == 6: + rules6 += "\t\t" + r.replace("\n", " \\\n\t\t") + " \\\n" else: - offset = None - if step.strip().startswith("src"): - write_rule(ip_to_rule(proto, nets[0], "saddr", offset)) + rules4 += "\t\t" + r.replace("\n", " \\\n\t\t") + " \\\n" + + rule = t[1].split("}")[0].strip() + for step in rule.split(";"): + if step.strip().startswith("src") or step.strip().startswith("dst"): + nets = step.strip()[3:].strip().split(" ") + if len(nets) > 1: + assert nets[1] == "offset" + offset = nets[2] + else: + offset = None + if step.strip().startswith("src"): + write_rule(ip_to_rule(proto, nets[0], "saddr", offset)) + else: + write_rule(ip_to_rule(proto, nets[0], "daddr", offset)) + elif step.strip().startswith("proto") and proto == 4: + write_rule(proto_to_rule(4, step.strip()[6:])) + elif step.strip().startswith("next header") and proto == 6: + write_rule(proto_to_rule(6, step.strip()[12:])) + elif step.strip().startswith("icmp type"): + write_rule(icmp_type_to_rule(proto, step.strip()[10:])) + elif step.strip().startswith("icmp code"): + write_rule(icmp_code_to_rule(proto, step.strip()[10:])) + elif step.strip().startswith("sport") or step.strip().startswith("dport") or step.strip().startswith("port"): + write_rule(port_to_rule(step.strip().split(" ")[0], step.strip().split(" ", 1)[1])) + elif step.strip().startswith("length"): + write_rule(len_to_rule(step.strip()[7:])) + elif step.strip().startswith("dscp"): + write_rule(dscp_to_rule(proto, step.strip()[5:])) + elif step.strip().startswith("tcp flags"): + write_rule(tcp_flags_to_rule(step.strip()[10:])) + elif step.strip().startswith("label"): + write_rule(flow_label_to_rule(step.strip()[6:])) + elif step.strip().startswith("fragment"): + if proto == 6: + use_v6_frags = True + write_rule(fragment_to_rule(proto, step.strip()[9:])) + elif step.strip() == "": + pass else: - write_rule(ip_to_rule(proto, nets[0], "daddr", offset)) - elif step.strip().startswith("proto") and proto == 4: - write_rule(proto_to_rule(4, step.strip()[6:])) - elif step.strip().startswith("next header") and proto == 6: - write_rule(proto_to_rule(6, step.strip()[12:])) - elif step.strip().startswith("icmp type"): - write_rule(icmp_type_to_rule(proto, step.strip()[10:])) - elif step.strip().startswith("icmp code"): - write_rule(icmp_code_to_rule(proto, step.strip()[10:])) - elif step.strip().startswith("sport") or step.strip().startswith("dport") or step.strip().startswith("port"): - write_rule(port_to_rule(step.strip().split(" ")[0], step.strip().split(" ", 1)[1])) - elif step.strip().startswith("length"): - write_rule(len_to_rule(step.strip()[7:])) - elif step.strip().startswith("dscp"): - write_rule(dscp_to_rule(proto, step.strip()[5:])) - elif step.strip().startswith("tcp flags"): - write_rule(tcp_flags_to_rule(step.strip()[10:])) - elif step.strip().startswith("label"): - write_rule(flow_label_to_rule(step.strip()[6:])) - elif step.strip().startswith("fragment"): - if proto == 6: - use_v6_frags = True - write_rule(fragment_to_rule(proto, step.strip()[9:])) - elif step.strip() == "": - pass + assert False + + # Now write the match handling! + first_action = None + last_action = None + for community in line.split("("): + if not community.startswith("generic, "): + continue + blocks = community.split(",") + assert len(blocks) == 3 + if len(blocks[1].strip()) != 10: # Should be 0x12345678 + continue + ty = blocks[1].strip()[:6] + low_bytes = int(blocks[2].strip(") \n"), 16) + if ty == "0x8006" or ty == "0x800c": + 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;" + elif low_bytes & (1 << 31) != 0: + # Negative limit, just drop + first_action = "return 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 + 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" + 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" + 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 += "\t} else {\n" + first_action += "\t\tbpf_spin_unlock(&rate->lock);\n" + first_action += "\t\treturn XDP_DROP;\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);") + elif ty == "0x8008": + assert False # We do not implement the redirect action + elif ty == "0x8009": + if low_bytes & ~0b111111 != 0: + assert False # Invalid DSCP value + if proto == 4: + write_rule("int32_t chk = ~BE16(ip->check) & 0xffff;") + 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 > 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) + if last_action is not None: + write_rule(last_action) + if proto == 6: + rules6 += "\t} while(0);\\\n" else: - assert False - write_rule(f"const uint32_t ruleidx = STATIC_RULE_CNT + {rulecnt};") - write_rule("DO_RETURN(ruleidx, XDP_DROP);") - if proto == 6: - rules6 += "\t} while(0);\\\n" - else: - rules4 += "\t} while(0);\\\n" - rulecnt += 1 + rules4 += "\t} while(0);\\\n" + rulecnt += 1 + lastrule = None out.write("\n") out.write(f"#define RULECNT {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")