trace_syscalls.c 12 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462463464465466467468469470471472473474475476477478479480481482483484485486487488489490491492493494495496497498499500501502503504505506507508509510511512513514515
  1. #include <trace/syscall.h>
  2. #include <trace/events/syscalls.h>
  3. #include <linux/kernel.h>
  4. #include <linux/ftrace.h>
  5. #include <linux/perf_counter.h>
  6. #include <asm/syscall.h>
  7. #include "trace_output.h"
  8. #include "trace.h"
  9. static DEFINE_MUTEX(syscall_trace_lock);
  10. static int sys_refcount_enter;
  11. static int sys_refcount_exit;
  12. static DECLARE_BITMAP(enabled_enter_syscalls, FTRACE_SYSCALL_MAX);
  13. static DECLARE_BITMAP(enabled_exit_syscalls, FTRACE_SYSCALL_MAX);
  14. enum print_line_t
  15. print_syscall_enter(struct trace_iterator *iter, int flags)
  16. {
  17. struct trace_seq *s = &iter->seq;
  18. struct trace_entry *ent = iter->ent;
  19. struct syscall_trace_enter *trace;
  20. struct syscall_metadata *entry;
  21. int i, ret, syscall;
  22. trace = (typeof(trace))ent;
  23. syscall = trace->nr;
  24. entry = syscall_nr_to_meta(syscall);
  25. if (!entry)
  26. goto end;
  27. if (entry->enter_id != ent->type) {
  28. WARN_ON_ONCE(1);
  29. goto end;
  30. }
  31. ret = trace_seq_printf(s, "%s(", entry->name);
  32. if (!ret)
  33. return TRACE_TYPE_PARTIAL_LINE;
  34. for (i = 0; i < entry->nb_args; i++) {
  35. /* parameter types */
  36. if (trace_flags & TRACE_ITER_VERBOSE) {
  37. ret = trace_seq_printf(s, "%s ", entry->types[i]);
  38. if (!ret)
  39. return TRACE_TYPE_PARTIAL_LINE;
  40. }
  41. /* parameter values */
  42. ret = trace_seq_printf(s, "%s: %lx%s", entry->args[i],
  43. trace->args[i],
  44. i == entry->nb_args - 1 ? "" : ", ");
  45. if (!ret)
  46. return TRACE_TYPE_PARTIAL_LINE;
  47. }
  48. ret = trace_seq_putc(s, ')');
  49. if (!ret)
  50. return TRACE_TYPE_PARTIAL_LINE;
  51. end:
  52. ret = trace_seq_putc(s, '\n');
  53. if (!ret)
  54. return TRACE_TYPE_PARTIAL_LINE;
  55. return TRACE_TYPE_HANDLED;
  56. }
  57. enum print_line_t
  58. print_syscall_exit(struct trace_iterator *iter, int flags)
  59. {
  60. struct trace_seq *s = &iter->seq;
  61. struct trace_entry *ent = iter->ent;
  62. struct syscall_trace_exit *trace;
  63. int syscall;
  64. struct syscall_metadata *entry;
  65. int ret;
  66. trace = (typeof(trace))ent;
  67. syscall = trace->nr;
  68. entry = syscall_nr_to_meta(syscall);
  69. if (!entry) {
  70. trace_seq_printf(s, "\n");
  71. return TRACE_TYPE_HANDLED;
  72. }
  73. if (entry->exit_id != ent->type) {
  74. WARN_ON_ONCE(1);
  75. return TRACE_TYPE_UNHANDLED;
  76. }
  77. ret = trace_seq_printf(s, "%s -> 0x%lx\n", entry->name,
  78. trace->ret);
  79. if (!ret)
  80. return TRACE_TYPE_PARTIAL_LINE;
  81. return TRACE_TYPE_HANDLED;
  82. }
  83. extern char *__bad_type_size(void);
  84. #define SYSCALL_FIELD(type, name) \
  85. sizeof(type) != sizeof(trace.name) ? \
  86. __bad_type_size() : \
  87. #type, #name, offsetof(typeof(trace), name), sizeof(trace.name)
  88. int syscall_enter_format(struct ftrace_event_call *call, struct trace_seq *s)
  89. {
  90. int i;
  91. int nr;
  92. int ret;
  93. struct syscall_metadata *entry;
  94. struct syscall_trace_enter trace;
  95. int offset = offsetof(struct syscall_trace_enter, args);
  96. nr = syscall_name_to_nr(call->data);
  97. entry = syscall_nr_to_meta(nr);
  98. if (!entry)
  99. return 0;
  100. ret = trace_seq_printf(s, "\tfield:%s %s;\toffset:%zu;\tsize:%zu;\n",
  101. SYSCALL_FIELD(int, nr));
  102. if (!ret)
  103. return 0;
  104. for (i = 0; i < entry->nb_args; i++) {
  105. ret = trace_seq_printf(s, "\tfield:%s %s;", entry->types[i],
  106. entry->args[i]);
  107. if (!ret)
  108. return 0;
  109. ret = trace_seq_printf(s, "\toffset:%d;\tsize:%zu;\n", offset,
  110. sizeof(unsigned long));
  111. if (!ret)
  112. return 0;
  113. offset += sizeof(unsigned long);
  114. }
  115. trace_seq_puts(s, "\nprint fmt: \"");
  116. for (i = 0; i < entry->nb_args; i++) {
  117. ret = trace_seq_printf(s, "%s: 0x%%0%zulx%s", entry->args[i],
  118. sizeof(unsigned long),
  119. i == entry->nb_args - 1 ? "" : ", ");
  120. if (!ret)
  121. return 0;
  122. }
  123. trace_seq_putc(s, '"');
  124. for (i = 0; i < entry->nb_args; i++) {
  125. ret = trace_seq_printf(s, ", ((unsigned long)(REC->%s))",
  126. entry->args[i]);
  127. if (!ret)
  128. return 0;
  129. }
  130. return trace_seq_putc(s, '\n');
  131. }
  132. int syscall_exit_format(struct ftrace_event_call *call, struct trace_seq *s)
  133. {
  134. int ret;
  135. struct syscall_trace_exit trace;
  136. ret = trace_seq_printf(s,
  137. "\tfield:%s %s;\toffset:%zu;\tsize:%zu;\n"
  138. "\tfield:%s %s;\toffset:%zu;\tsize:%zu;\n",
  139. SYSCALL_FIELD(int, nr),
  140. SYSCALL_FIELD(unsigned long, ret));
  141. if (!ret)
  142. return 0;
  143. return trace_seq_printf(s, "\nprint fmt: \"0x%%lx\", REC->ret\n");
  144. }
  145. int syscall_enter_define_fields(struct ftrace_event_call *call)
  146. {
  147. struct syscall_trace_enter trace;
  148. struct syscall_metadata *meta;
  149. int ret;
  150. int nr;
  151. int i;
  152. int offset = offsetof(typeof(trace), args);
  153. nr = syscall_name_to_nr(call->data);
  154. meta = syscall_nr_to_meta(nr);
  155. if (!meta)
  156. return 0;
  157. ret = trace_define_common_fields(call);
  158. if (ret)
  159. return ret;
  160. for (i = 0; i < meta->nb_args; i++) {
  161. ret = trace_define_field(call, meta->types[i],
  162. meta->args[i], offset,
  163. sizeof(unsigned long), 0,
  164. FILTER_OTHER);
  165. offset += sizeof(unsigned long);
  166. }
  167. return ret;
  168. }
  169. int syscall_exit_define_fields(struct ftrace_event_call *call)
  170. {
  171. struct syscall_trace_exit trace;
  172. int ret;
  173. ret = trace_define_common_fields(call);
  174. if (ret)
  175. return ret;
  176. ret = trace_define_field(call, SYSCALL_FIELD(unsigned long, ret), 0,
  177. FILTER_OTHER);
  178. return ret;
  179. }
  180. void ftrace_syscall_enter(struct pt_regs *regs, long id)
  181. {
  182. struct syscall_trace_enter *entry;
  183. struct syscall_metadata *sys_data;
  184. struct ring_buffer_event *event;
  185. int size;
  186. int syscall_nr;
  187. syscall_nr = syscall_get_nr(current, regs);
  188. if (!test_bit(syscall_nr, enabled_enter_syscalls))
  189. return;
  190. sys_data = syscall_nr_to_meta(syscall_nr);
  191. if (!sys_data)
  192. return;
  193. size = sizeof(*entry) + sizeof(unsigned long) * sys_data->nb_args;
  194. event = trace_current_buffer_lock_reserve(sys_data->enter_id, size,
  195. 0, 0);
  196. if (!event)
  197. return;
  198. entry = ring_buffer_event_data(event);
  199. entry->nr = syscall_nr;
  200. syscall_get_arguments(current, regs, 0, sys_data->nb_args, entry->args);
  201. if (!filter_current_check_discard(sys_data->enter_event, entry, event))
  202. trace_current_buffer_unlock_commit(event, 0, 0);
  203. }
  204. void ftrace_syscall_exit(struct pt_regs *regs, long ret)
  205. {
  206. struct syscall_trace_exit *entry;
  207. struct syscall_metadata *sys_data;
  208. struct ring_buffer_event *event;
  209. int syscall_nr;
  210. syscall_nr = syscall_get_nr(current, regs);
  211. if (!test_bit(syscall_nr, enabled_exit_syscalls))
  212. return;
  213. sys_data = syscall_nr_to_meta(syscall_nr);
  214. if (!sys_data)
  215. return;
  216. event = trace_current_buffer_lock_reserve(sys_data->exit_id,
  217. sizeof(*entry), 0, 0);
  218. if (!event)
  219. return;
  220. entry = ring_buffer_event_data(event);
  221. entry->nr = syscall_nr;
  222. entry->ret = syscall_get_return_value(current, regs);
  223. if (!filter_current_check_discard(sys_data->exit_event, entry, event))
  224. trace_current_buffer_unlock_commit(event, 0, 0);
  225. }
  226. int reg_event_syscall_enter(struct ftrace_event_call *call)
  227. {
  228. int ret = 0;
  229. int num;
  230. char *name;
  231. name = (char *)call->data;
  232. num = syscall_name_to_nr(name);
  233. if (num < 0 || num >= FTRACE_SYSCALL_MAX)
  234. return -ENOSYS;
  235. mutex_lock(&syscall_trace_lock);
  236. if (!sys_refcount_enter)
  237. ret = register_trace_sys_enter(ftrace_syscall_enter);
  238. if (ret) {
  239. pr_info("event trace: Could not activate"
  240. "syscall entry trace point");
  241. } else {
  242. set_bit(num, enabled_enter_syscalls);
  243. sys_refcount_enter++;
  244. }
  245. mutex_unlock(&syscall_trace_lock);
  246. return ret;
  247. }
  248. void unreg_event_syscall_enter(struct ftrace_event_call *call)
  249. {
  250. int num;
  251. char *name;
  252. name = (char *)call->data;
  253. num = syscall_name_to_nr(name);
  254. if (num < 0 || num >= FTRACE_SYSCALL_MAX)
  255. return;
  256. mutex_lock(&syscall_trace_lock);
  257. sys_refcount_enter--;
  258. clear_bit(num, enabled_enter_syscalls);
  259. if (!sys_refcount_enter)
  260. unregister_trace_sys_enter(ftrace_syscall_enter);
  261. mutex_unlock(&syscall_trace_lock);
  262. }
  263. int reg_event_syscall_exit(struct ftrace_event_call *call)
  264. {
  265. int ret = 0;
  266. int num;
  267. char *name;
  268. name = call->data;
  269. num = syscall_name_to_nr(name);
  270. if (num < 0 || num >= FTRACE_SYSCALL_MAX)
  271. return -ENOSYS;
  272. mutex_lock(&syscall_trace_lock);
  273. if (!sys_refcount_exit)
  274. ret = register_trace_sys_exit(ftrace_syscall_exit);
  275. if (ret) {
  276. pr_info("event trace: Could not activate"
  277. "syscall exit trace point");
  278. } else {
  279. set_bit(num, enabled_exit_syscalls);
  280. sys_refcount_exit++;
  281. }
  282. mutex_unlock(&syscall_trace_lock);
  283. return ret;
  284. }
  285. void unreg_event_syscall_exit(struct ftrace_event_call *call)
  286. {
  287. int num;
  288. char *name;
  289. name = call->data;
  290. num = syscall_name_to_nr(name);
  291. if (num < 0 || num >= FTRACE_SYSCALL_MAX)
  292. return;
  293. mutex_lock(&syscall_trace_lock);
  294. sys_refcount_exit--;
  295. clear_bit(num, enabled_exit_syscalls);
  296. if (!sys_refcount_exit)
  297. unregister_trace_sys_exit(ftrace_syscall_exit);
  298. mutex_unlock(&syscall_trace_lock);
  299. }
  300. struct trace_event event_syscall_enter = {
  301. .trace = print_syscall_enter,
  302. };
  303. struct trace_event event_syscall_exit = {
  304. .trace = print_syscall_exit,
  305. };
  306. #ifdef CONFIG_EVENT_PROFILE
  307. static DECLARE_BITMAP(enabled_prof_enter_syscalls, FTRACE_SYSCALL_MAX);
  308. static DECLARE_BITMAP(enabled_prof_exit_syscalls, FTRACE_SYSCALL_MAX);
  309. static int sys_prof_refcount_enter;
  310. static int sys_prof_refcount_exit;
  311. static void prof_syscall_enter(struct pt_regs *regs, long id)
  312. {
  313. struct syscall_trace_enter *rec;
  314. struct syscall_metadata *sys_data;
  315. int syscall_nr;
  316. int size;
  317. syscall_nr = syscall_get_nr(current, regs);
  318. if (!test_bit(syscall_nr, enabled_prof_enter_syscalls))
  319. return;
  320. sys_data = syscall_nr_to_meta(syscall_nr);
  321. if (!sys_data)
  322. return;
  323. /* get the size after alignment with the u32 buffer size field */
  324. size = sizeof(unsigned long) * sys_data->nb_args + sizeof(*rec);
  325. size = ALIGN(size + sizeof(u32), sizeof(u64));
  326. size -= sizeof(u32);
  327. do {
  328. char raw_data[size];
  329. /* zero the dead bytes from align to not leak stack to user */
  330. *(u64 *)(&raw_data[size - sizeof(u64)]) = 0ULL;
  331. rec = (struct syscall_trace_enter *) raw_data;
  332. tracing_generic_entry_update(&rec->ent, 0, 0);
  333. rec->ent.type = sys_data->enter_id;
  334. rec->nr = syscall_nr;
  335. syscall_get_arguments(current, regs, 0, sys_data->nb_args,
  336. (unsigned long *)&rec->args);
  337. perf_tpcounter_event(sys_data->enter_id, 0, 1, rec, size);
  338. } while(0);
  339. }
  340. int reg_prof_syscall_enter(char *name)
  341. {
  342. int ret = 0;
  343. int num;
  344. num = syscall_name_to_nr(name);
  345. if (num < 0 || num >= FTRACE_SYSCALL_MAX)
  346. return -ENOSYS;
  347. mutex_lock(&syscall_trace_lock);
  348. if (!sys_prof_refcount_enter)
  349. ret = register_trace_sys_enter(prof_syscall_enter);
  350. if (ret) {
  351. pr_info("event trace: Could not activate"
  352. "syscall entry trace point");
  353. } else {
  354. set_bit(num, enabled_prof_enter_syscalls);
  355. sys_prof_refcount_enter++;
  356. }
  357. mutex_unlock(&syscall_trace_lock);
  358. return ret;
  359. }
  360. void unreg_prof_syscall_enter(char *name)
  361. {
  362. int num;
  363. num = syscall_name_to_nr(name);
  364. if (num < 0 || num >= FTRACE_SYSCALL_MAX)
  365. return;
  366. mutex_lock(&syscall_trace_lock);
  367. sys_prof_refcount_enter--;
  368. clear_bit(num, enabled_prof_enter_syscalls);
  369. if (!sys_prof_refcount_enter)
  370. unregister_trace_sys_enter(prof_syscall_enter);
  371. mutex_unlock(&syscall_trace_lock);
  372. }
  373. static void prof_syscall_exit(struct pt_regs *regs, long ret)
  374. {
  375. struct syscall_metadata *sys_data;
  376. struct syscall_trace_exit rec;
  377. int syscall_nr;
  378. syscall_nr = syscall_get_nr(current, regs);
  379. if (!test_bit(syscall_nr, enabled_prof_exit_syscalls))
  380. return;
  381. sys_data = syscall_nr_to_meta(syscall_nr);
  382. if (!sys_data)
  383. return;
  384. tracing_generic_entry_update(&rec.ent, 0, 0);
  385. rec.ent.type = sys_data->exit_id;
  386. rec.nr = syscall_nr;
  387. rec.ret = syscall_get_return_value(current, regs);
  388. perf_tpcounter_event(sys_data->exit_id, 0, 1, &rec, sizeof(rec));
  389. }
  390. int reg_prof_syscall_exit(char *name)
  391. {
  392. int ret = 0;
  393. int num;
  394. num = syscall_name_to_nr(name);
  395. if (num < 0 || num >= FTRACE_SYSCALL_MAX)
  396. return -ENOSYS;
  397. mutex_lock(&syscall_trace_lock);
  398. if (!sys_prof_refcount_exit)
  399. ret = register_trace_sys_exit(prof_syscall_exit);
  400. if (ret) {
  401. pr_info("event trace: Could not activate"
  402. "syscall entry trace point");
  403. } else {
  404. set_bit(num, enabled_prof_exit_syscalls);
  405. sys_prof_refcount_exit++;
  406. }
  407. mutex_unlock(&syscall_trace_lock);
  408. return ret;
  409. }
  410. void unreg_prof_syscall_exit(char *name)
  411. {
  412. int num;
  413. num = syscall_name_to_nr(name);
  414. if (num < 0 || num >= FTRACE_SYSCALL_MAX)
  415. return;
  416. mutex_lock(&syscall_trace_lock);
  417. sys_prof_refcount_exit--;
  418. clear_bit(num, enabled_prof_exit_syscalls);
  419. if (!sys_prof_refcount_exit)
  420. unregister_trace_sys_exit(prof_syscall_exit);
  421. mutex_unlock(&syscall_trace_lock);
  422. }
  423. #endif