trace_syscalls.c 10 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456
  1. #include <trace/syscall.h>
  2. #include <linux/kernel.h>
  3. #include <linux/ftrace.h>
  4. #include <linux/perf_counter.h>
  5. #include <asm/syscall.h>
  6. #include "trace_output.h"
  7. #include "trace.h"
  8. static DEFINE_MUTEX(syscall_trace_lock);
  9. static int sys_refcount_enter;
  10. static int sys_refcount_exit;
  11. static DECLARE_BITMAP(enabled_enter_syscalls, FTRACE_SYSCALL_MAX);
  12. static DECLARE_BITMAP(enabled_exit_syscalls, FTRACE_SYSCALL_MAX);
  13. /* Option to display the parameters types */
  14. enum {
  15. TRACE_SYSCALLS_OPT_TYPES = 0x1,
  16. };
  17. static struct tracer_opt syscalls_opts[] = {
  18. { TRACER_OPT(syscall_arg_type, TRACE_SYSCALLS_OPT_TYPES) },
  19. { }
  20. };
  21. static struct tracer_flags syscalls_flags = {
  22. .val = 0, /* By default: no parameters types */
  23. .opts = syscalls_opts
  24. };
  25. enum print_line_t
  26. print_syscall_enter(struct trace_iterator *iter, int flags)
  27. {
  28. struct trace_seq *s = &iter->seq;
  29. struct trace_entry *ent = iter->ent;
  30. struct syscall_trace_enter *trace;
  31. struct syscall_metadata *entry;
  32. int i, ret, syscall;
  33. trace = (typeof(trace))ent;
  34. syscall = trace->nr;
  35. entry = syscall_nr_to_meta(syscall);
  36. if (!entry)
  37. goto end;
  38. if (entry->enter_id != ent->type) {
  39. WARN_ON_ONCE(1);
  40. goto end;
  41. }
  42. ret = trace_seq_printf(s, "%s(", entry->name);
  43. if (!ret)
  44. return TRACE_TYPE_PARTIAL_LINE;
  45. for (i = 0; i < entry->nb_args; i++) {
  46. /* parameter types */
  47. if (syscalls_flags.val & TRACE_SYSCALLS_OPT_TYPES) {
  48. ret = trace_seq_printf(s, "%s ", entry->types[i]);
  49. if (!ret)
  50. return TRACE_TYPE_PARTIAL_LINE;
  51. }
  52. /* parameter values */
  53. ret = trace_seq_printf(s, "%s: %lx%s ", entry->args[i],
  54. trace->args[i],
  55. i == entry->nb_args - 1 ? ")" : ",");
  56. if (!ret)
  57. return TRACE_TYPE_PARTIAL_LINE;
  58. }
  59. end:
  60. trace_seq_printf(s, "\n");
  61. return TRACE_TYPE_HANDLED;
  62. }
  63. enum print_line_t
  64. print_syscall_exit(struct trace_iterator *iter, int flags)
  65. {
  66. struct trace_seq *s = &iter->seq;
  67. struct trace_entry *ent = iter->ent;
  68. struct syscall_trace_exit *trace;
  69. int syscall;
  70. struct syscall_metadata *entry;
  71. int ret;
  72. trace = (typeof(trace))ent;
  73. syscall = trace->nr;
  74. entry = syscall_nr_to_meta(syscall);
  75. if (!entry) {
  76. trace_seq_printf(s, "\n");
  77. return TRACE_TYPE_HANDLED;
  78. }
  79. if (entry->exit_id != ent->type) {
  80. WARN_ON_ONCE(1);
  81. return TRACE_TYPE_UNHANDLED;
  82. }
  83. ret = trace_seq_printf(s, "%s -> 0x%lx\n", entry->name,
  84. trace->ret);
  85. if (!ret)
  86. return TRACE_TYPE_PARTIAL_LINE;
  87. return TRACE_TYPE_HANDLED;
  88. }
  89. int ftrace_format_syscall(struct ftrace_event_call *call, struct trace_seq *s)
  90. {
  91. int i;
  92. int nr;
  93. int ret = 0;
  94. struct syscall_metadata *entry;
  95. int offset = sizeof(struct trace_entry);
  96. nr = syscall_name_to_nr((char *)call->data);
  97. entry = syscall_nr_to_meta(nr);
  98. if (!entry)
  99. return ret;
  100. for (i = 0; i < entry->nb_args; i++) {
  101. ret = trace_seq_printf(s, "\tfield:%s %s;", entry->types[i],
  102. entry->args[i]);
  103. if (!ret)
  104. return 0;
  105. ret = trace_seq_printf(s, "\toffset:%d;\tsize:%lu;\n", offset,
  106. sizeof(unsigned long));
  107. if (!ret)
  108. return 0;
  109. offset += sizeof(unsigned long);
  110. }
  111. trace_seq_printf(s, "\nprint fmt: \"");
  112. for (i = 0; i < entry->nb_args; i++) {
  113. ret = trace_seq_printf(s, "%s: 0x%%0%lulx%s", entry->args[i],
  114. sizeof(unsigned long),
  115. i == entry->nb_args - 1 ? "\", " : ", ");
  116. if (!ret)
  117. return 0;
  118. }
  119. for (i = 0; i < entry->nb_args; i++) {
  120. ret = trace_seq_printf(s, "((unsigned long)(REC->%s))%s",
  121. entry->args[i],
  122. i == entry->nb_args - 1 ? "\n" : ", ");
  123. if (!ret)
  124. return 0;
  125. }
  126. return ret;
  127. }
  128. void ftrace_syscall_enter(struct pt_regs *regs, long id)
  129. {
  130. struct syscall_trace_enter *entry;
  131. struct syscall_metadata *sys_data;
  132. struct ring_buffer_event *event;
  133. int size;
  134. int syscall_nr;
  135. syscall_nr = syscall_get_nr(current, regs);
  136. if (!test_bit(syscall_nr, enabled_enter_syscalls))
  137. return;
  138. sys_data = syscall_nr_to_meta(syscall_nr);
  139. if (!sys_data)
  140. return;
  141. size = sizeof(*entry) + sizeof(unsigned long) * sys_data->nb_args;
  142. event = trace_current_buffer_lock_reserve(sys_data->enter_id, size,
  143. 0, 0);
  144. if (!event)
  145. return;
  146. entry = ring_buffer_event_data(event);
  147. entry->nr = syscall_nr;
  148. syscall_get_arguments(current, regs, 0, sys_data->nb_args, entry->args);
  149. trace_current_buffer_unlock_commit(event, 0, 0);
  150. trace_wake_up();
  151. }
  152. void ftrace_syscall_exit(struct pt_regs *regs, long ret)
  153. {
  154. struct syscall_trace_exit *entry;
  155. struct syscall_metadata *sys_data;
  156. struct ring_buffer_event *event;
  157. int syscall_nr;
  158. syscall_nr = syscall_get_nr(current, regs);
  159. if (!test_bit(syscall_nr, enabled_exit_syscalls))
  160. return;
  161. sys_data = syscall_nr_to_meta(syscall_nr);
  162. if (!sys_data)
  163. return;
  164. event = trace_current_buffer_lock_reserve(sys_data->exit_id,
  165. sizeof(*entry), 0, 0);
  166. if (!event)
  167. return;
  168. entry = ring_buffer_event_data(event);
  169. entry->nr = syscall_nr;
  170. entry->ret = syscall_get_return_value(current, regs);
  171. trace_current_buffer_unlock_commit(event, 0, 0);
  172. trace_wake_up();
  173. }
  174. int reg_event_syscall_enter(void *ptr)
  175. {
  176. int ret = 0;
  177. int num;
  178. char *name;
  179. name = (char *)ptr;
  180. num = syscall_name_to_nr(name);
  181. if (num < 0 || num >= FTRACE_SYSCALL_MAX)
  182. return -ENOSYS;
  183. mutex_lock(&syscall_trace_lock);
  184. if (!sys_refcount_enter)
  185. ret = register_trace_syscall_enter(ftrace_syscall_enter);
  186. if (ret) {
  187. pr_info("event trace: Could not activate"
  188. "syscall entry trace point");
  189. } else {
  190. set_bit(num, enabled_enter_syscalls);
  191. sys_refcount_enter++;
  192. }
  193. mutex_unlock(&syscall_trace_lock);
  194. return ret;
  195. }
  196. void unreg_event_syscall_enter(void *ptr)
  197. {
  198. int num;
  199. char *name;
  200. name = (char *)ptr;
  201. num = syscall_name_to_nr(name);
  202. if (num < 0 || num >= FTRACE_SYSCALL_MAX)
  203. return;
  204. mutex_lock(&syscall_trace_lock);
  205. sys_refcount_enter--;
  206. clear_bit(num, enabled_enter_syscalls);
  207. if (!sys_refcount_enter)
  208. unregister_trace_syscall_enter(ftrace_syscall_enter);
  209. mutex_unlock(&syscall_trace_lock);
  210. }
  211. int reg_event_syscall_exit(void *ptr)
  212. {
  213. int ret = 0;
  214. int num;
  215. char *name;
  216. name = (char *)ptr;
  217. num = syscall_name_to_nr(name);
  218. if (num < 0 || num >= FTRACE_SYSCALL_MAX)
  219. return -ENOSYS;
  220. mutex_lock(&syscall_trace_lock);
  221. if (!sys_refcount_exit)
  222. ret = register_trace_syscall_exit(ftrace_syscall_exit);
  223. if (ret) {
  224. pr_info("event trace: Could not activate"
  225. "syscall exit trace point");
  226. } else {
  227. set_bit(num, enabled_exit_syscalls);
  228. sys_refcount_exit++;
  229. }
  230. mutex_unlock(&syscall_trace_lock);
  231. return ret;
  232. }
  233. void unreg_event_syscall_exit(void *ptr)
  234. {
  235. int num;
  236. char *name;
  237. name = (char *)ptr;
  238. num = syscall_name_to_nr(name);
  239. if (num < 0 || num >= FTRACE_SYSCALL_MAX)
  240. return;
  241. mutex_lock(&syscall_trace_lock);
  242. sys_refcount_exit--;
  243. clear_bit(num, enabled_exit_syscalls);
  244. if (!sys_refcount_exit)
  245. unregister_trace_syscall_exit(ftrace_syscall_exit);
  246. mutex_unlock(&syscall_trace_lock);
  247. }
  248. struct trace_event event_syscall_enter = {
  249. .trace = print_syscall_enter,
  250. };
  251. struct trace_event event_syscall_exit = {
  252. .trace = print_syscall_exit,
  253. };
  254. #ifdef CONFIG_EVENT_PROFILE
  255. struct syscall_enter_record {
  256. struct trace_entry entry;
  257. unsigned long args[0];
  258. };
  259. struct syscall_exit_record {
  260. struct trace_entry entry;
  261. unsigned long ret;
  262. };
  263. static DECLARE_BITMAP(enabled_prof_enter_syscalls, FTRACE_SYSCALL_MAX);
  264. static DECLARE_BITMAP(enabled_prof_exit_syscalls, FTRACE_SYSCALL_MAX);
  265. static int sys_prof_refcount_enter;
  266. static int sys_prof_refcount_exit;
  267. static void prof_syscall_enter(struct pt_regs *regs, long id)
  268. {
  269. struct syscall_enter_record *rec;
  270. struct syscall_metadata *sys_data;
  271. int syscall_nr;
  272. int size;
  273. syscall_nr = syscall_get_nr(current, regs);
  274. if (!test_bit(syscall_nr, enabled_prof_enter_syscalls))
  275. return;
  276. sys_data = syscall_nr_to_meta(syscall_nr);
  277. if (!sys_data)
  278. return;
  279. /* get the size after alignment with the u32 buffer size field */
  280. size = sizeof(unsigned long) * sys_data->nb_args + sizeof(*rec);
  281. size = ALIGN(size + sizeof(u32), sizeof(u64));
  282. size -= sizeof(u32);
  283. do {
  284. char raw_data[size];
  285. /* zero the dead bytes from align to not leak stack to user */
  286. *(u64 *)(&raw_data[size - sizeof(u64)]) = 0ULL;
  287. rec = (struct syscall_enter_record *) raw_data;
  288. tracing_generic_entry_update(&rec->entry, 0, 0);
  289. rec->entry.type = sys_data->enter_id;
  290. syscall_get_arguments(current, regs, 0, sys_data->nb_args,
  291. (unsigned long *)&rec->args);
  292. perf_tpcounter_event(sys_data->enter_id, 0, 1, rec, size);
  293. } while(0);
  294. }
  295. int reg_prof_syscall_enter(char *name)
  296. {
  297. int ret = 0;
  298. int num;
  299. num = syscall_name_to_nr(name);
  300. if (num < 0 || num >= FTRACE_SYSCALL_MAX)
  301. return -ENOSYS;
  302. mutex_lock(&syscall_trace_lock);
  303. if (!sys_prof_refcount_enter)
  304. ret = register_trace_syscall_enter(prof_syscall_enter);
  305. if (ret) {
  306. pr_info("event trace: Could not activate"
  307. "syscall entry trace point");
  308. } else {
  309. set_bit(num, enabled_prof_enter_syscalls);
  310. sys_prof_refcount_enter++;
  311. }
  312. mutex_unlock(&syscall_trace_lock);
  313. return ret;
  314. }
  315. void unreg_prof_syscall_enter(char *name)
  316. {
  317. int num;
  318. num = syscall_name_to_nr(name);
  319. if (num < 0 || num >= FTRACE_SYSCALL_MAX)
  320. return;
  321. mutex_lock(&syscall_trace_lock);
  322. sys_prof_refcount_enter--;
  323. clear_bit(num, enabled_prof_enter_syscalls);
  324. if (!sys_prof_refcount_enter)
  325. unregister_trace_syscall_enter(prof_syscall_enter);
  326. mutex_unlock(&syscall_trace_lock);
  327. }
  328. static void prof_syscall_exit(struct pt_regs *regs, long ret)
  329. {
  330. struct syscall_metadata *sys_data;
  331. struct syscall_exit_record rec;
  332. int syscall_nr;
  333. syscall_nr = syscall_get_nr(current, regs);
  334. if (!test_bit(syscall_nr, enabled_prof_exit_syscalls))
  335. return;
  336. sys_data = syscall_nr_to_meta(syscall_nr);
  337. if (!sys_data)
  338. return;
  339. tracing_generic_entry_update(&rec.entry, 0, 0);
  340. rec.entry.type = sys_data->exit_id;
  341. rec.ret = syscall_get_return_value(current, regs);
  342. perf_tpcounter_event(sys_data->exit_id, 0, 1, &rec, sizeof(rec));
  343. }
  344. int reg_prof_syscall_exit(char *name)
  345. {
  346. int ret = 0;
  347. int num;
  348. num = syscall_name_to_nr(name);
  349. if (num < 0 || num >= FTRACE_SYSCALL_MAX)
  350. return -ENOSYS;
  351. mutex_lock(&syscall_trace_lock);
  352. if (!sys_prof_refcount_exit)
  353. ret = register_trace_syscall_exit(prof_syscall_exit);
  354. if (ret) {
  355. pr_info("event trace: Could not activate"
  356. "syscall entry trace point");
  357. } else {
  358. set_bit(num, enabled_prof_exit_syscalls);
  359. sys_prof_refcount_exit++;
  360. }
  361. mutex_unlock(&syscall_trace_lock);
  362. return ret;
  363. }
  364. void unreg_prof_syscall_exit(char *name)
  365. {
  366. int num;
  367. num = syscall_name_to_nr(name);
  368. if (num < 0 || num >= FTRACE_SYSCALL_MAX)
  369. return;
  370. mutex_lock(&syscall_trace_lock);
  371. sys_prof_refcount_exit--;
  372. clear_bit(num, enabled_prof_exit_syscalls);
  373. if (!sys_prof_refcount_exit)
  374. unregister_trace_syscall_exit(prof_syscall_exit);
  375. mutex_unlock(&syscall_trace_lock);
  376. }
  377. #endif