Now that mem is more compact, bump max tracked src IPs to 1M
[flowspec-xdp] / genrules.py
index edef6175e264f03f6edec910069c55ea92ae1c41..1ab90b05a80bca0b9d1615cb5276489f9cce85d8 100755 (executable)
@@ -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;"
@@ -208,47 +215,48 @@ 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;"
     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" :
-        ast = parse_ast(rules, parse_numbers_expr)
-        return "if (tcp == NULL && udp == NULL) 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 (tcp == NULL && udp == NULL) 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;"""
 
+
 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 +295,203 @@ with open("rules.h", "w") as out:
     rules4 = ""
     use_v6_frags = False
     rulecnt = 0
+    ratelimitcnt = 0
+    v4persrcratelimits = []
+    v6persrcratelimits = []
 
+    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
+            stats_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]
+                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" 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 = "{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 + 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)
+                        # Note that int64_t will overflow after 292 years of uptime
+                        first_action = "int64_t time = bpf_ktime_get_ns();\n"
+                        first_action +=  "uint64_t allowed_since_last = 0;\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:
+                            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});\n"
+                                first_action += f"struct persrc_rate4_entry *rate = rate_ptr.rate;\n"
+                                v4persrcratelimits.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});\n"
+                                first_action += f"struct persrc_rate6_entry *rate = rate_ptr.rate;\n"
+                                v6persrcratelimits.append((high_byte + 1) * 4096)
+                        first_action +=  "if (rate) {\n"
+                        first_action += f"\t{spin_lock}\n"
+                        first_action +=  "\tif (likely(rate->sent_rate > 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 +=  "\t\t\trate->sent_rate = 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 +=  "\tif (rate->sent_rate - ((int64_t)allowed_since_last) <= 0)" + " {\n"
+                        if ty == "0x8006" or ty == "0x8306":
+                            first_action += "\t\trate->sent_rate = data_end - pktdata;\n"
+                        else:
+                            first_action += "\t\trate->sent_rate = 1;\n"
+                        first_action +=  "\t\trate->sent_time = time;\n"
+                        first_action += f"\t\t{spin_unlock}\n"
+                        first_action +=  "\t} else {\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}\n"
+                        first_action +=  "}\n"
+                elif ty == "0x8007":
+                    if low_bytes & 1 == 0:
+                        last_action = "return XDP_PASS;"
+                    if low_bytes & 2 == 2:
+                        stats_action = f"const uint32_t ruleidx = STATIC_RULE_CNT + {rulecnt};\n"
+                        stats_action += "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.replace("{stats_replace}", stats_action))
+            if stats_action is not None 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:
-                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")
@@ -366,3 +501,8 @@ 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(v6persrcratelimits):
+            out.write(f"SRC_RATE_DEFINE(6, {idx}, {limit})\n")