1 /*
2 * LLCStat Show LLC hit ratio for each process on each CPU core.
3 * For Linux, uses BCC, eBPF. Embedded C.
4 *
5 * Basic example of BCC timed sampling perf event.
6 *
7 * USAGE: LLCStat [duration]
8 *
9 * Copyright (c) Facebook, Inc.
10 * Licensed under the Apache License, Version 2.0 (the "License")
11 */
12
13 #include <linux/perf_event.h>
14 #include <unistd.h>
15 #include <iomanip>
16 #include <iostream>
17 #include <string>
18
19 #include "BPF.h"
20
21 const std::string BPF_PROGRAM = R"(
22 #include <linux/ptrace.h>
23 #include <uapi/linux/bpf_perf_event.h>
24
25 struct event_t {
26 int cpu;
27 int pid;
28 char name[16];
29 };
30
31 BPF_HASH(ref_count, struct event_t);
32 BPF_HASH(miss_count, struct event_t);
33
34 static inline __attribute__((always_inline)) void get_key(struct event_t* key) {
35 key->cpu = bpf_get_smp_processor_id();
36 key->pid = bpf_get_current_pid_tgid();
37 bpf_get_current_comm(&(key->name), sizeof(key->name));
38 }
39
40 int on_cache_miss(struct bpf_perf_event_data *ctx) {
41 struct event_t key = {};
42 get_key(&key);
43
44 u64 zero = 0, *val;
45 val = miss_count.lookup_or_init(&key, &zero);
46 (*val) += ctx->sample_period;
47
48 return 0;
49 }
50
51 int on_cache_ref(struct bpf_perf_event_data *ctx) {
52 struct event_t key = {};
53 get_key(&key);
54
55 u64 zero = 0, *val;
56 val = ref_count.lookup_or_init(&key, &zero);
57 (*val) += ctx->sample_period;
58
59 return 0;
60 }
61 )";
62
63 struct event_t {
64 int cpu;
65 int pid;
66 char name[16];
67 };
68
main(int argc,char ** argv)69 int main(int argc, char** argv) {
70 ebpf::BPF bpf;
71 auto init_res = bpf.init(BPF_PROGRAM);
72 if (init_res.code() != 0) {
73 std::cerr << init_res.msg() << std::endl;
74 return 1;
75 }
76
77 auto attach_ref_res =
78 bpf.attach_perf_event(PERF_TYPE_HARDWARE, PERF_COUNT_HW_CACHE_REFERENCES,
79 "on_cache_ref", 100, 0);
80 if (attach_ref_res.code() != 0) {
81 std::cerr << attach_ref_res.msg() << std::endl;
82 return 1;
83 }
84 auto attach_miss_res = bpf.attach_perf_event(
85 PERF_TYPE_HARDWARE, PERF_COUNT_HW_CACHE_MISSES, "on_cache_miss", 100, 0);
86 if (attach_miss_res.code() != 0) {
87 std::cerr << attach_miss_res.msg() << std::endl;
88 return 1;
89 }
90
91 int probe_time = 10;
92 if (argc == 2) {
93 probe_time = atoi(argv[1]);
94 }
95 std::cout << "Probing for " << probe_time << " seconds" << std::endl;
96 sleep(probe_time);
97 bpf.detach_perf_event(PERF_TYPE_HARDWARE, PERF_COUNT_HW_CACHE_REFERENCES);
98 bpf.detach_perf_event(PERF_TYPE_HARDWARE, PERF_COUNT_HW_CACHE_MISSES);
99
100 auto refs = bpf.get_hash_table<event_t, uint64_t>("ref_count");
101 auto misses = bpf.get_hash_table<event_t, uint64_t>("miss_count");
102 for (auto it : refs.get_table_offline()) {
103 uint64_t hit;
104 try {
105 auto miss = misses[it.first];
106 hit = miss <= it.second ? it.second - miss : 0;
107 } catch (...) {
108 hit = it.second;
109 }
110 double ratio = (double(hit) / double(it.second)) * 100.0;
111 std::cout << "PID " << std::setw(8) << std::setfill(' ') << it.first.pid;
112 std::cout << std::setw(20) << std::setfill(' ') << std::left
113 << " (" + std::string(it.first.name) + ") " << std::right;
114 std::cout << "on CPU " << std::setw(2) << std::setfill(' ') << it.first.cpu;
115 std::cout << " Hit Rate " << std::setprecision(4) << ratio << "% ";
116 std::cout << "(" << hit << "/" << it.second << ")" << std::endl;
117 }
118 return 0;
119 }
120