rulecnt = 0
ratelimitcnt = 0
v4persrcratelimits = []
+ v5persrcratelimits = []
v6persrcratelimits = []
lastrule = None
elif exp == 0xff:
# NaN/INF. Just treat as INF and accept
first_action = None
- elif exp <= 127: # < 1
+ 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
+ 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" or ty == "0x8306":
- 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 += "uint64_t allowed_since_last = 0;\n"
+
+ first_action = "int64_t time = 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 += "struct ratelimit *rate = bpf_map_lookup_elem(&rate_map, &ratelimitidx);\n"
ratelimitcnt += 1
else:
- spin_lock = "/* No locking as we're per-CPU */"
- spin_unlock = "/* No locking as we're per-CPU */"
+ 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"
- v4persrcratelimits.append((high_byte + 1) * 1024)
+ 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)
+ 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"
+ 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"
- v6persrcratelimits.append((high_byte + 1) * 1024)
- first_action += f"struct percpu_ratelimit *rate = bpf_map_lookup_elem(rate_map, &srcip);\n"
- first_action += "if (rate) {\n"
- first_action += f"\t{spin_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"
+ 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)
if ty == "0x8006" or ty == "0x8306":
- first_action += f"\t\t{accessor} = data_end - pktdata;\n"
+ 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 += f"\t\t{spin_unlock}\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 - ((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"
- if ty == "0x8306" or ty == "0x830c":
- first_action += "} else {\n"
- first_action += "\tstruct percpu_ratelimit new_rate = { .sent_time = time, };\n"
- first_action += "\trate = &new_rate;\n"
- if ty == "0x8006" or ty == "0x8306":
- first_action += f"\t\t{accessor} = data_end - pktdata;\n"
- else:
- first_action += f"\t\t{accessor} = 1;\n"
- first_action += "\tbpf_map_update_elem(rate_map, &srcip, rate, BPF_ANY);\n"
first_action += "}\n"
elif ty == "0x8007":
if low_bytes & 1 == 0:
out.write("#define PARSE_V6_FRAG PARSE\n")
with open("maps.h", "w") as out:
for idx, limit in enumerate(v4persrcratelimits):
- out.write(f"V4_SRC_RATE_DEFINE({idx}, {limit})\n")
+ 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"V6_SRC_RATE_DEFINE({idx}, {limit})\n")
+ out.write(f"SRC_RATE_DEFINE(6, {idx}, {limit})\n")