Combine redundant rule conditions to work around LLVM bug 52455 fully
[flowspec-xdp] / genrules.py
index 93c22f5c9059dfb7f7cccefa7707dda1fb849871..3865b6b69f42edf1cab46786213bc380d8196df8 100755 (executable)
@@ -256,6 +256,69 @@ 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;"""
 
+class RuleAction(Enum):
+    CONDITIONS = 1
+    ACTION = 2
+    LIST = 3
+class RuleNode:
+    def __init__(self, ty, action, inner):
+        self.ty = ty
+        self.action = action
+        self.inner = inner
+        if ty == RuleAction.ACTION:
+            assert inner is None
+            assert type(action) == str
+        elif ty == RuleAction.LIST:
+            assert type(inner) == list
+            assert action is None
+            for item in inner:
+                assert type(item) == RuleNode
+        else:
+            assert ty == RuleAction.CONDITIONS
+            assert type(action) == list
+            assert type(inner) == RuleNode
+
+    def maybe_join(self, neighbor):
+        if self.ty == RuleAction.CONDITIONS and neighbor.ty == RuleAction.CONDITIONS:
+            overlapping_conditions = [x for x in self.action if x in neighbor.action]
+            if len(overlapping_conditions) != 0:
+                us = RuleNode(RuleAction.CONDITIONS, [x for x in self.action if x not in overlapping_conditions], self.inner)
+                them = RuleNode(RuleAction.CONDITIONS, [x for x in neighbor.action if x not in overlapping_conditions], neighbor.inner)
+                self.action = overlapping_conditions
+                if self.inner.ty == RuleAction.LIST and us.action == []:
+                    self.inner.inner.append(them)
+                else:
+                    self.inner = RuleNode(RuleAction.LIST, None, [us, them])
+                self.inner.flatten()
+                return True
+        return False
+
+    def flatten(self):
+        # LLVM can be pretty bad at optimizing out common subexpressions. Thus, we have to do a
+        # pass here to toptimize out common subexpressions in back-to-back rules.
+        # See https://bugs.llvm.org/show_bug.cgi?id=52455
+        assert self.ty == RuleAction.LIST
+        did_update = True
+        while did_update:
+            did_update = False
+            for i in range(0, len(self.inner) - 1):
+                if self.inner[i].maybe_join(self.inner[i + 1]):
+                    del self.inner[i + 1]
+                    did_update = True
+                    break
+
+    def write(self, out, pfx="\t"):
+        if self.ty == RuleAction.CONDITIONS:
+            out.write(pfx + "do {\\\n")
+            for cond in self.action:
+                out.write("\t" + pfx + cond.strip().replace("\n", " \\\n\t" + pfx) + " \\\n")
+            self.inner.write(out, pfx)
+            out.write(pfx + "} while(0);\\\n")
+        elif self.ty == RuleAction.LIST:
+            for item in self.inner:
+                item.write(out, pfx + "\t")
+        else:
+            out.write("\t" + pfx + self.action.strip().replace("\n", " \\\n\t" + pfx) + " \\\n")
 
 with open("rules.h", "w") as out:
     parse = argparse.ArgumentParser()
@@ -291,10 +354,10 @@ with open("rules.h", "w") as out:
             assert False
         out.write("#define REQ_8021Q " + args.vlan_tag + "\n")
 
-    rules6 = ""
-    rules4 = ""
+    rules6 = []
+    rules4 = []
     use_v6_frags = False
-    rulecnt = 0
+    stats_rulecnt = 0
     ratelimitcnt = 0
     v4persrcratelimits = []
     v5persrcratelimits = []
@@ -313,19 +376,15 @@ with open("rules.h", "w") as out:
             t = lastrule.split("{")
             if t[0].strip() == "flow4":
                 proto = 4
-                rules4 += "\tdo {\\\n"
             elif t[0].strip() == "flow6":
                 proto = 6
-                rules6 += "\tdo {\\\n"
             else:
                 continue
 
+            conditions = []
             def write_rule(r):
-                global rules4, rules6
-                if proto == 6:
-                    rules6 += "\t\t" + r.replace("\n", " \\\n\t\t") + " \\\n"
-                else:
-                    rules4 += "\t\t" + r.replace("\n", " \\\n\t\t") + " \\\n"
+                global conditions
+                conditions.append(r + "\n")
 
             rule = t[1].split("}")[0].strip()
             for step in rule.split(";"):
@@ -367,9 +426,14 @@ with open("rules.h", "w") as out:
                 else:
                     assert False
 
+            actions = ""
+            def write_rule(r):
+                global actions
+                actions += r + "\n"
+
             # Now write the match handling!
             first_action = None
-            stats_action = None
+            stats_action = ""
             last_action = None
             for community in line.split("("):
                 if not community.startswith("generic, "):
@@ -404,70 +468,48 @@ with open("rules.h", "w") as out:
                         value = 1.0 + mantissa / (2**23)
                         value *= 2**(exp-127)
 
-                        first_action =   "int64_t time = bpf_ktime_get_ns() & RATE_TIME_MASK;\n"
+                        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":
-                            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
+                            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:
-                            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"
+                                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"struct persrc_rate5_ptr rate_ptr = get_v5_persrc_ratelimit(srcip, rate_map, {(high_byte + 1) * 4096});\n"
-                                first_action += f"struct persrc_rate5_entry *rate = rate_ptr.rate;\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"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"
+                                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)
-                        if ty == "0x8006" or ty == "0x8306":
-                            first_action += "uint64_t amt = data_end - pktdata;\n"
-                        else:
-                            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 - ((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\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 | ((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 +=  "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:
-                        stats_action = f"const uint32_t ruleidx = STATIC_RULE_CNT + {rulecnt};\n"
+                        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
@@ -487,27 +529,43 @@ with open("rules.h", "w") as out:
                         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):
+            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"
+                rules6.append(RuleNode(RuleAction.CONDITIONS, conditions, RuleNode(RuleAction.ACTION, actions, None)))
             else:
-                rules4 += "\t} while(0);\\\n"
-            rulecnt += 1
+                rules4.append(RuleNode(RuleAction.CONDITIONS, conditions, RuleNode(RuleAction.ACTION, actions, None)))
+            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 != "":
+
+    # Here we should probably sort the rules according to flowspec's sorting rules. We don't bother
+    # however, because its annoying.
+
+    if len(rules4) != 0:
         out.write("#define NEED_V4_PARSE\n")
-        out.write("#define RULES4 {\\\n" + rules4 + "}\n")
-    if rules6:
+        out.write("#define RULES4 {\\\n")
+        rules4 = RuleNode(RuleAction.LIST, None, rules4)
+        rules4.flatten()
+        rules4.write(out)
+        out.write("}\n")
+
+    if len(rules6) != 0:
         out.write("#define NEED_V6_PARSE\n")
-        out.write("#define RULES6 {\\\n" + rules6 + "}\n")
+        out.write("#define RULES6 {\\\n")
+        rules6 = RuleNode(RuleAction.LIST, None, rules6)
+        rules6.flatten()
+        rules6.write(out)
+        out.write("}\n")
+
     if args.v6frag == "ignore-parse-if-rule":
         if use_v6_frags:
             out.write("#define PARSE_V6_FRAG PARSE\n")