samples/bpf/trace_event_user.c

   1 /* Copyright (c) 2016 Facebook
   2  *
   3  * This program is free software; you can redistribute it and/or
   4  * modify it under the terms of version 2 of the GNU General Public
   5  * License as published by the Free Software Foundation.
   6  */
   7 #include <stdio.h>
   8 #include <unistd.h>
   9 #include <stdlib.h>
  10 #include <stdbool.h>
  11 #include <string.h>
  12 #include <fcntl.h>
  13 #include <poll.h>
  14 #include <sys/ioctl.h>
  15 #include <linux/perf_event.h>
  16 #include <linux/bpf.h>
  17 #include <signal.h>
  18 #include <assert.h>
  19 #include <errno.h>
  20 #include <sys/resource.h>
  21 #include "libbpf.h"
  22 #include "bpf_load.h"
  23 #include "perf-sys.h"
  24 #include "trace_helpers.h"
  25
  26 #define SAMPLE_FREQ 50
  27
  28 static bool sys_read_seen, sys_write_seen;
  29
  30 static void print_ksym(__u64 addr)
  31 {
  32         struct ksym *sym;
  33
  34         if (!addr)
  35                 return;
  36         sym = ksym_search(addr);
  37         printf("%s;", sym->name);
  38         if (!strcmp(sym->name, "sys_read"))
  39                 sys_read_seen = true;
  40         else if (!strcmp(sym->name, "sys_write"))
  41                 sys_write_seen = true;
  42 }
  43
  44 static void print_addr(__u64 addr)
  45 {
  46         if (!addr)
  47                 return;
  48         printf("%llx;", addr);
  49 }
  50
  51 #define TASK_COMM_LEN 16
  52
  53 struct key_t {
  54         char comm[TASK_COMM_LEN];
  55         __u32 kernstack;
  56         __u32 userstack;
  57 };
  58
  59 static void print_stack(struct key_t *key, __u64 count)
  60 {
  61         __u64 ip[PERF_MAX_STACK_DEPTH] = {};
  62         static bool warned;
  63         int i;
  64
  65         printf("%3lld %s;", count, key->comm);
  66         if (bpf_map_lookup_elem(map_fd[1], &key->kernstack, ip) != 0) {
  67                 printf("---;");
  68         } else {
  69                 for (i = PERF_MAX_STACK_DEPTH - 1; i >= 0; i--)
  70                         print_ksym(ip[i]);
  71         }
  72         printf("-;");
  73         if (bpf_map_lookup_elem(map_fd[1], &key->userstack, ip) != 0) {
  74                 printf("---;");
  75         } else {
  76                 for (i = PERF_MAX_STACK_DEPTH - 1; i >= 0; i--)
  77                         print_addr(ip[i]);
  78         }
  79         if (count < 6)
  80                 printf("\r");
  81         else
  82                 printf("\n");
  83
  84         if (key->kernstack == -EEXIST && !warned) {
  85                 printf("stackmap collisions seen. Consider increasing size\n");
  86                 warned = true;
  87         } else if ((int)key->kernstack < 0 && (int)key->userstack < 0) {
  88                 printf("err stackid %d %d\n", key->kernstack, key->userstack);
  89         }
  90 }
  91
  92 static void int_exit(int sig)
  93 {
  94         kill(0, SIGKILL);
  95         exit(0);
  96 }
  97
  98 static void print_stacks(void)
  99 {
 100         struct key_t key = {}, next_key;
 101         __u64 value;
 102         __u32 stackid = 0, next_id;
 103         int fd = map_fd[0], stack_map = map_fd[1];
 104
 105         sys_read_seen = sys_write_seen = false;
 106         while (bpf_map_get_next_key(fd, &key, &next_key) == 0) {
 107                 bpf_map_lookup_elem(fd, &next_key, &value);
 108                 print_stack(&next_key, value);
 109                 bpf_map_delete_elem(fd, &next_key);
 110                 key = next_key;
 111         }
 112         printf("\n");
 113         if (!sys_read_seen || !sys_write_seen) {
 114                 printf("BUG kernel stack doesn't contain sys_read() and sys_write()\n");
 115                 int_exit(0);
 116         }
 117
 118         /* clear stack map */
 119         while (bpf_map_get_next_key(stack_map, &stackid, &next_id) == 0) {
 120                 bpf_map_delete_elem(stack_map, &next_id);
 121                 stackid = next_id;
 122         }
 123 }
 124
 125 static inline int generate_load(void)
 126 {
 127         if (system("dd if=/dev/zero of=/dev/null count=5000k status=none") < 0) {
 128                 printf("failed to generate some load with dd: %s\n", strerror(errno));
 129                 return -1;
 130         }
 131
 132         return 0;
 133 }
 134
 135 static void test_perf_event_all_cpu(struct perf_event_attr *attr)
 136 {
 137         int nr_cpus = sysconf(_SC_NPROCESSORS_CONF);
 138         int *pmu_fd = malloc(nr_cpus * sizeof(int));
 139         int i, error = 0;
 140
 141         /* system wide perf event, no need to inherit */
 142         attr->inherit = 0;
 143
 144         /* open perf_event on all cpus */
 145         for (i = 0; i < nr_cpus; i++) {
 146                 pmu_fd[i] = sys_perf_event_open(attr, -1, i, -1, 0);
 147                 if (pmu_fd[i] < 0) {
 148                         printf("sys_perf_event_open failed\n");
 149                         error = 1;
 150                         goto all_cpu_err;
 151                 }
 152                 assert(ioctl(pmu_fd[i], PERF_EVENT_IOC_SET_BPF, prog_fd[0]) == 0);
 153                 assert(ioctl(pmu_fd[i], PERF_EVENT_IOC_ENABLE) == 0);
 154         }
 155
 156         if (generate_load() < 0) {
 157                 error = 1;
 158                 goto all_cpu_err;
 159         }
 160         print_stacks();
 161 all_cpu_err:
 162         for (i--; i >= 0; i--) {
 163                 ioctl(pmu_fd[i], PERF_EVENT_IOC_DISABLE);
 164                 close(pmu_fd[i]);
 165         }
 166         free(pmu_fd);
 167         if (error)
 168                 int_exit(0);
 169 }
 170
 171 static void test_perf_event_task(struct perf_event_attr *attr)
 172 {
 173         int pmu_fd, error = 0;
 174
 175         /* per task perf event, enable inherit so the "dd ..." command can be traced properly.
 176          * Enabling inherit will cause bpf_perf_prog_read_time helper failure.
 177          */
 178         attr->inherit = 1;
 179
 180         /* open task bound event */
 181         pmu_fd = sys_perf_event_open(attr, 0, -1, -1, 0);
 182         if (pmu_fd < 0) {
 183                 printf("sys_perf_event_open failed\n");
 184                 int_exit(0);
 185         }
 186         assert(ioctl(pmu_fd, PERF_EVENT_IOC_SET_BPF, prog_fd[0]) == 0);
 187         assert(ioctl(pmu_fd, PERF_EVENT_IOC_ENABLE) == 0);
 188
 189         if (generate_load() < 0) {
 190                 error = 1;
 191                 goto err;
 192         }
 193         print_stacks();
 194 err:
 195         ioctl(pmu_fd, PERF_EVENT_IOC_DISABLE);
 196         close(pmu_fd);
 197         if (error)
 198                 int_exit(0);
 199 }
 200
 201 static void test_bpf_perf_event(void)
 202 {
 203         struct perf_event_attr attr_type_hw = {
 204                 .sample_freq = SAMPLE_FREQ,
 205                 .freq = 1,
 206                 .type = PERF_TYPE_HARDWARE,
 207                 .config = PERF_COUNT_HW_CPU_CYCLES,
 208         };
 209         struct perf_event_attr attr_type_sw = {
 210                 .sample_freq = SAMPLE_FREQ,
 211                 .freq = 1,
 212                 .type = PERF_TYPE_SOFTWARE,
 213                 .config = PERF_COUNT_SW_CPU_CLOCK,
 214         };
 215         struct perf_event_attr attr_hw_cache_l1d = {
 216                 .sample_freq = SAMPLE_FREQ,
 217                 .freq = 1,
 218                 .type = PERF_TYPE_HW_CACHE,
 219                 .config =
 220                         PERF_COUNT_HW_CACHE_L1D |
 221                         (PERF_COUNT_HW_CACHE_OP_READ << 8) |
 222                         (PERF_COUNT_HW_CACHE_RESULT_ACCESS << 16),
 223         };
 224         struct perf_event_attr attr_hw_cache_branch_miss = {
 225                 .sample_freq = SAMPLE_FREQ,
 226                 .freq = 1,
 227                 .type = PERF_TYPE_HW_CACHE,
 228                 .config =
 229                         PERF_COUNT_HW_CACHE_BPU |
 230                         (PERF_COUNT_HW_CACHE_OP_READ << 8) |
 231                         (PERF_COUNT_HW_CACHE_RESULT_MISS << 16),
 232         };
 233         struct perf_event_attr attr_type_raw = {
 234                 .sample_freq = SAMPLE_FREQ,
 235                 .freq = 1,
 236                 .type = PERF_TYPE_RAW,
 237                 /* Intel Instruction Retired */
 238                 .config = 0xc0,
 239         };
 240         struct perf_event_attr attr_type_raw_lock_load = {
 241                 .sample_freq = SAMPLE_FREQ,
 242                 .freq = 1,
 243                 .type = PERF_TYPE_RAW,
 244                 /* Intel MEM_UOPS_RETIRED.LOCK_LOADS */
 245                 .config = 0x21d0,
 246                 /* Request to record lock address from PEBS */
 247                 .sample_type = PERF_SAMPLE_ADDR,
 248                 /* Record address value requires precise event */
 249                 .precise_ip = 2,
 250         };
 251
 252         printf("Test HW_CPU_CYCLES\n");
 253         test_perf_event_all_cpu(&attr_type_hw);
 254         test_perf_event_task(&attr_type_hw);
 255
 256         printf("Test SW_CPU_CLOCK\n");
 257         test_perf_event_all_cpu(&attr_type_sw);
 258         test_perf_event_task(&attr_type_sw);
 259
 260         printf("Test HW_CACHE_L1D\n");
 261         test_perf_event_all_cpu(&attr_hw_cache_l1d);
 262         test_perf_event_task(&attr_hw_cache_l1d);
 263
 264         printf("Test HW_CACHE_BPU\n");
 265         test_perf_event_all_cpu(&attr_hw_cache_branch_miss);
 266         test_perf_event_task(&attr_hw_cache_branch_miss);
 267
 268         printf("Test Instruction Retired\n");
 269         test_perf_event_all_cpu(&attr_type_raw);
 270         test_perf_event_task(&attr_type_raw);
 271
 272         printf("Test Lock Load\n");
 273         test_perf_event_all_cpu(&attr_type_raw_lock_load);
 274         test_perf_event_task(&attr_type_raw_lock_load);
 275
 276         printf("*** PASS ***\n");
 277 }
 278
 279
 280 int main(int argc, char **argv)
 281 {
 282         struct rlimit r = {RLIM_INFINITY, RLIM_INFINITY};
 283         char filename[256];
 284
 285         snprintf(filename, sizeof(filename), "%s_kern.o", argv[0]);
 286         setrlimit(RLIMIT_MEMLOCK, &r);
 287
 288         signal(SIGINT, int_exit);
 289         signal(SIGTERM, int_exit);
 290
 291         if (load_kallsyms()) {
 292                 printf("failed to process /proc/kallsyms\n");
 293                 return 1;
 294         }
 295
 296         if (load_bpf_file(filename)) {
 297                 printf("%s", bpf_log_buf);
 298                 return 2;
 299         }
 300
 301         if (fork() == 0) {
 302                 read_trace_pipe();
 303                 return 0;
 304         }
 305         test_bpf_perf_event();
 306         int_exit(0);
 307         return 0;
 308 }