sampleip_user.c 4.2 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198
  1. /*
  2. * sampleip: sample instruction pointer and frequency count in a BPF map.
  3. *
  4. * Copyright 2016 Netflix, Inc.
  5. *
  6. * This program is free software; you can redistribute it and/or
  7. * modify it under the terms of version 2 of the GNU General Public
  8. * License as published by the Free Software Foundation.
  9. */
  10. #include <stdio.h>
  11. #include <stdlib.h>
  12. #include <stdio.h>
  13. #include <unistd.h>
  14. #include <errno.h>
  15. #include <signal.h>
  16. #include <string.h>
  17. #include <assert.h>
  18. #include <linux/perf_event.h>
  19. #include <linux/ptrace.h>
  20. #include <linux/bpf.h>
  21. #include <sys/ioctl.h>
  22. #include "libbpf.h"
  23. #include "bpf_load.h"
  24. #include "perf-sys.h"
  25. #define DEFAULT_FREQ 99
  26. #define DEFAULT_SECS 5
  27. #define MAX_IPS 8192
  28. #define PAGE_OFFSET 0xffff880000000000
  29. static int nr_cpus;
  30. static void usage(void)
  31. {
  32. printf("USAGE: sampleip [-F freq] [duration]\n");
  33. printf(" -F freq # sample frequency (Hertz), default 99\n");
  34. printf(" duration # sampling duration (seconds), default 5\n");
  35. }
  36. static int sampling_start(int *pmu_fd, int freq)
  37. {
  38. int i;
  39. struct perf_event_attr pe_sample_attr = {
  40. .type = PERF_TYPE_SOFTWARE,
  41. .freq = 1,
  42. .sample_period = freq,
  43. .config = PERF_COUNT_SW_CPU_CLOCK,
  44. .inherit = 1,
  45. };
  46. for (i = 0; i < nr_cpus; i++) {
  47. pmu_fd[i] = sys_perf_event_open(&pe_sample_attr, -1 /* pid */, i,
  48. -1 /* group_fd */, 0 /* flags */);
  49. if (pmu_fd[i] < 0) {
  50. fprintf(stderr, "ERROR: Initializing perf sampling\n");
  51. return 1;
  52. }
  53. assert(ioctl(pmu_fd[i], PERF_EVENT_IOC_SET_BPF,
  54. prog_fd[0]) == 0);
  55. assert(ioctl(pmu_fd[i], PERF_EVENT_IOC_ENABLE, 0) == 0);
  56. }
  57. return 0;
  58. }
  59. static void sampling_end(int *pmu_fd)
  60. {
  61. int i;
  62. for (i = 0; i < nr_cpus; i++)
  63. close(pmu_fd[i]);
  64. }
  65. struct ipcount {
  66. __u64 ip;
  67. __u32 count;
  68. };
  69. /* used for sorting */
  70. struct ipcount counts[MAX_IPS];
  71. static int count_cmp(const void *p1, const void *p2)
  72. {
  73. return ((struct ipcount *)p1)->count - ((struct ipcount *)p2)->count;
  74. }
  75. static void print_ip_map(int fd)
  76. {
  77. struct ksym *sym;
  78. __u64 key, next_key;
  79. __u32 value;
  80. int i, max;
  81. printf("%-19s %-32s %s\n", "ADDR", "KSYM", "COUNT");
  82. /* fetch IPs and counts */
  83. key = 0, i = 0;
  84. while (bpf_map_get_next_key(fd, &key, &next_key) == 0) {
  85. bpf_map_lookup_elem(fd, &next_key, &value);
  86. counts[i].ip = next_key;
  87. counts[i++].count = value;
  88. key = next_key;
  89. }
  90. max = i;
  91. /* sort and print */
  92. qsort(counts, max, sizeof(struct ipcount), count_cmp);
  93. for (i = 0; i < max; i++) {
  94. if (counts[i].ip > PAGE_OFFSET) {
  95. sym = ksym_search(counts[i].ip);
  96. printf("0x%-17llx %-32s %u\n", counts[i].ip, sym->name,
  97. counts[i].count);
  98. } else {
  99. printf("0x%-17llx %-32s %u\n", counts[i].ip, "(user)",
  100. counts[i].count);
  101. }
  102. }
  103. if (max == MAX_IPS) {
  104. printf("WARNING: IP hash was full (max %d entries); ", max);
  105. printf("may have dropped samples\n");
  106. }
  107. }
  108. static void int_exit(int sig)
  109. {
  110. printf("\n");
  111. print_ip_map(map_fd[0]);
  112. exit(0);
  113. }
  114. int main(int argc, char **argv)
  115. {
  116. char filename[256];
  117. int *pmu_fd, opt, freq = DEFAULT_FREQ, secs = DEFAULT_SECS;
  118. /* process arguments */
  119. while ((opt = getopt(argc, argv, "F:h")) != -1) {
  120. switch (opt) {
  121. case 'F':
  122. freq = atoi(optarg);
  123. break;
  124. case 'h':
  125. default:
  126. usage();
  127. return 0;
  128. }
  129. }
  130. if (argc - optind == 1)
  131. secs = atoi(argv[optind]);
  132. if (freq == 0 || secs == 0) {
  133. usage();
  134. return 1;
  135. }
  136. /* initialize kernel symbol translation */
  137. if (load_kallsyms()) {
  138. fprintf(stderr, "ERROR: loading /proc/kallsyms\n");
  139. return 2;
  140. }
  141. /* create perf FDs for each CPU */
  142. nr_cpus = sysconf(_SC_NPROCESSORS_CONF);
  143. pmu_fd = malloc(nr_cpus * sizeof(int));
  144. if (pmu_fd == NULL) {
  145. fprintf(stderr, "ERROR: malloc of pmu_fd\n");
  146. return 1;
  147. }
  148. /* load BPF program */
  149. snprintf(filename, sizeof(filename), "%s_kern.o", argv[0]);
  150. if (load_bpf_file(filename)) {
  151. fprintf(stderr, "ERROR: loading BPF program (errno %d):\n",
  152. errno);
  153. if (strcmp(bpf_log_buf, "") == 0)
  154. fprintf(stderr, "Try: ulimit -l unlimited\n");
  155. else
  156. fprintf(stderr, "%s", bpf_log_buf);
  157. return 1;
  158. }
  159. signal(SIGINT, int_exit);
  160. /* do sampling */
  161. printf("Sampling at %d Hertz for %d seconds. Ctrl-C also ends.\n",
  162. freq, secs);
  163. if (sampling_start(pmu_fd, freq) != 0)
  164. return 1;
  165. sleep(secs);
  166. sampling_end(pmu_fd);
  167. free(pmu_fd);
  168. /* output sample counts */
  169. print_ip_map(map_fd[0]);
  170. return 0;
  171. }