trace_syscalls.c 11 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398399400401402403404405406407408409410411412413414415416417418419420421422423424425426427428429430431432433434435436437438439440441442443444445446447448449450451452453454455456457458459460461462
  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. enum print_line_t
  14. print_syscall_enter(struct trace_iterator *iter, int flags)
  15. {
  16. struct trace_seq *s = &iter->seq;
  17. struct trace_entry *ent = iter->ent;
  18. struct syscall_trace_enter *trace;
  19. struct syscall_metadata *entry;
  20. int i, ret, syscall;
  21. trace = (typeof(trace))ent;
  22. syscall = trace->nr;
  23. entry = syscall_nr_to_meta(syscall);
  24. if (!entry)
  25. goto end;
  26. if (entry->enter_id != ent->type) {
  27. WARN_ON_ONCE(1);
  28. goto end;
  29. }
  30. ret = trace_seq_printf(s, "%s(", entry->name);
  31. if (!ret)
  32. return TRACE_TYPE_PARTIAL_LINE;
  33. for (i = 0; i < entry->nb_args; i++) {
  34. /* parameter types */
  35. if (trace_flags & TRACE_ITER_VERBOSE) {
  36. ret = trace_seq_printf(s, "%s ", entry->types[i]);
  37. if (!ret)
  38. return TRACE_TYPE_PARTIAL_LINE;
  39. }
  40. /* parameter values */
  41. ret = trace_seq_printf(s, "%s: %lx%s ", entry->args[i],
  42. trace->args[i],
  43. i == entry->nb_args - 1 ? ")" : ",");
  44. if (!ret)
  45. return TRACE_TYPE_PARTIAL_LINE;
  46. }
  47. end:
  48. trace_seq_printf(s, "\n");
  49. return TRACE_TYPE_HANDLED;
  50. }
  51. enum print_line_t
  52. print_syscall_exit(struct trace_iterator *iter, int flags)
  53. {
  54. struct trace_seq *s = &iter->seq;
  55. struct trace_entry *ent = iter->ent;
  56. struct syscall_trace_exit *trace;
  57. int syscall;
  58. struct syscall_metadata *entry;
  59. int ret;
  60. trace = (typeof(trace))ent;
  61. syscall = trace->nr;
  62. entry = syscall_nr_to_meta(syscall);
  63. if (!entry) {
  64. trace_seq_printf(s, "\n");
  65. return TRACE_TYPE_HANDLED;
  66. }
  67. if (entry->exit_id != ent->type) {
  68. WARN_ON_ONCE(1);
  69. return TRACE_TYPE_UNHANDLED;
  70. }
  71. ret = trace_seq_printf(s, "%s -> 0x%lx\n", entry->name,
  72. trace->ret);
  73. if (!ret)
  74. return TRACE_TYPE_PARTIAL_LINE;
  75. return TRACE_TYPE_HANDLED;
  76. }
  77. extern char *__bad_type_size(void);
  78. #define SYSCALL_FIELD(type, name) \
  79. sizeof(type) != sizeof(trace.name) ? \
  80. __bad_type_size() : \
  81. #type, #name, offsetof(typeof(trace), name), sizeof(trace.name)
  82. int syscall_enter_format(struct ftrace_event_call *call, struct trace_seq *s)
  83. {
  84. int i;
  85. int nr;
  86. int ret;
  87. struct syscall_metadata *entry;
  88. struct syscall_trace_enter trace;
  89. int offset = offsetof(struct syscall_trace_enter, args);
  90. nr = syscall_name_to_nr(call->data);
  91. entry = syscall_nr_to_meta(nr);
  92. if (!entry)
  93. return 0;
  94. ret = trace_seq_printf(s, "\tfield:%s %s;\toffset:%zu;\tsize:%zu;\n",
  95. SYSCALL_FIELD(int, nr));
  96. if (!ret)
  97. return 0;
  98. for (i = 0; i < entry->nb_args; i++) {
  99. ret = trace_seq_printf(s, "\tfield:%s %s;", entry->types[i],
  100. entry->args[i]);
  101. if (!ret)
  102. return 0;
  103. ret = trace_seq_printf(s, "\toffset:%d;\tsize:%zu;\n", offset,
  104. sizeof(unsigned long));
  105. if (!ret)
  106. return 0;
  107. offset += sizeof(unsigned long);
  108. }
  109. trace_seq_printf(s, "\nprint fmt: \"");
  110. for (i = 0; i < entry->nb_args; i++) {
  111. ret = trace_seq_printf(s, "%s: 0x%%0%zulx%s", entry->args[i],
  112. sizeof(unsigned long),
  113. i == entry->nb_args - 1 ? "\", " : ", ");
  114. if (!ret)
  115. return 0;
  116. }
  117. for (i = 0; i < entry->nb_args; i++) {
  118. ret = trace_seq_printf(s, "((unsigned long)(REC->%s))%s",
  119. entry->args[i],
  120. i == entry->nb_args - 1 ? "\n" : ", ");
  121. if (!ret)
  122. return 0;
  123. }
  124. return ret;
  125. }
  126. int syscall_exit_format(struct ftrace_event_call *call, struct trace_seq *s)
  127. {
  128. int ret;
  129. struct syscall_trace_exit trace;
  130. ret = trace_seq_printf(s,
  131. "\tfield:%s %s;\toffset:%zu;\tsize:%zu;\n"
  132. "\tfield:%s %s;\toffset:%zu;\tsize:%zu;\n",
  133. SYSCALL_FIELD(int, nr),
  134. SYSCALL_FIELD(unsigned long, ret));
  135. if (!ret)
  136. return 0;
  137. return trace_seq_printf(s, "\nprint fmt: \"0x%%lx\", REC->ret\n");
  138. }
  139. void ftrace_syscall_enter(struct pt_regs *regs, long id)
  140. {
  141. struct syscall_trace_enter *entry;
  142. struct syscall_metadata *sys_data;
  143. struct ring_buffer_event *event;
  144. int size;
  145. int syscall_nr;
  146. syscall_nr = syscall_get_nr(current, regs);
  147. if (!test_bit(syscall_nr, enabled_enter_syscalls))
  148. return;
  149. sys_data = syscall_nr_to_meta(syscall_nr);
  150. if (!sys_data)
  151. return;
  152. size = sizeof(*entry) + sizeof(unsigned long) * sys_data->nb_args;
  153. event = trace_current_buffer_lock_reserve(sys_data->enter_id, size,
  154. 0, 0);
  155. if (!event)
  156. return;
  157. entry = ring_buffer_event_data(event);
  158. entry->nr = syscall_nr;
  159. syscall_get_arguments(current, regs, 0, sys_data->nb_args, entry->args);
  160. trace_current_buffer_unlock_commit(event, 0, 0);
  161. trace_wake_up();
  162. }
  163. void ftrace_syscall_exit(struct pt_regs *regs, long ret)
  164. {
  165. struct syscall_trace_exit *entry;
  166. struct syscall_metadata *sys_data;
  167. struct ring_buffer_event *event;
  168. int syscall_nr;
  169. syscall_nr = syscall_get_nr(current, regs);
  170. if (!test_bit(syscall_nr, enabled_exit_syscalls))
  171. return;
  172. sys_data = syscall_nr_to_meta(syscall_nr);
  173. if (!sys_data)
  174. return;
  175. event = trace_current_buffer_lock_reserve(sys_data->exit_id,
  176. sizeof(*entry), 0, 0);
  177. if (!event)
  178. return;
  179. entry = ring_buffer_event_data(event);
  180. entry->nr = syscall_nr;
  181. entry->ret = syscall_get_return_value(current, regs);
  182. trace_current_buffer_unlock_commit(event, 0, 0);
  183. trace_wake_up();
  184. }
  185. int reg_event_syscall_enter(void *ptr)
  186. {
  187. int ret = 0;
  188. int num;
  189. char *name;
  190. name = (char *)ptr;
  191. num = syscall_name_to_nr(name);
  192. if (num < 0 || num >= FTRACE_SYSCALL_MAX)
  193. return -ENOSYS;
  194. mutex_lock(&syscall_trace_lock);
  195. if (!sys_refcount_enter)
  196. ret = register_trace_syscall_enter(ftrace_syscall_enter);
  197. if (ret) {
  198. pr_info("event trace: Could not activate"
  199. "syscall entry trace point");
  200. } else {
  201. set_bit(num, enabled_enter_syscalls);
  202. sys_refcount_enter++;
  203. }
  204. mutex_unlock(&syscall_trace_lock);
  205. return ret;
  206. }
  207. void unreg_event_syscall_enter(void *ptr)
  208. {
  209. int num;
  210. char *name;
  211. name = (char *)ptr;
  212. num = syscall_name_to_nr(name);
  213. if (num < 0 || num >= FTRACE_SYSCALL_MAX)
  214. return;
  215. mutex_lock(&syscall_trace_lock);
  216. sys_refcount_enter--;
  217. clear_bit(num, enabled_enter_syscalls);
  218. if (!sys_refcount_enter)
  219. unregister_trace_syscall_enter(ftrace_syscall_enter);
  220. mutex_unlock(&syscall_trace_lock);
  221. }
  222. int reg_event_syscall_exit(void *ptr)
  223. {
  224. int ret = 0;
  225. int num;
  226. char *name;
  227. name = (char *)ptr;
  228. num = syscall_name_to_nr(name);
  229. if (num < 0 || num >= FTRACE_SYSCALL_MAX)
  230. return -ENOSYS;
  231. mutex_lock(&syscall_trace_lock);
  232. if (!sys_refcount_exit)
  233. ret = register_trace_syscall_exit(ftrace_syscall_exit);
  234. if (ret) {
  235. pr_info("event trace: Could not activate"
  236. "syscall exit trace point");
  237. } else {
  238. set_bit(num, enabled_exit_syscalls);
  239. sys_refcount_exit++;
  240. }
  241. mutex_unlock(&syscall_trace_lock);
  242. return ret;
  243. }
  244. void unreg_event_syscall_exit(void *ptr)
  245. {
  246. int num;
  247. char *name;
  248. name = (char *)ptr;
  249. num = syscall_name_to_nr(name);
  250. if (num < 0 || num >= FTRACE_SYSCALL_MAX)
  251. return;
  252. mutex_lock(&syscall_trace_lock);
  253. sys_refcount_exit--;
  254. clear_bit(num, enabled_exit_syscalls);
  255. if (!sys_refcount_exit)
  256. unregister_trace_syscall_exit(ftrace_syscall_exit);
  257. mutex_unlock(&syscall_trace_lock);
  258. }
  259. struct trace_event event_syscall_enter = {
  260. .trace = print_syscall_enter,
  261. };
  262. struct trace_event event_syscall_exit = {
  263. .trace = print_syscall_exit,
  264. };
  265. #ifdef CONFIG_EVENT_PROFILE
  266. static DECLARE_BITMAP(enabled_prof_enter_syscalls, FTRACE_SYSCALL_MAX);
  267. static DECLARE_BITMAP(enabled_prof_exit_syscalls, FTRACE_SYSCALL_MAX);
  268. static int sys_prof_refcount_enter;
  269. static int sys_prof_refcount_exit;
  270. static void prof_syscall_enter(struct pt_regs *regs, long id)
  271. {
  272. struct syscall_trace_enter *rec;
  273. struct syscall_metadata *sys_data;
  274. int syscall_nr;
  275. int size;
  276. syscall_nr = syscall_get_nr(current, regs);
  277. if (!test_bit(syscall_nr, enabled_prof_enter_syscalls))
  278. return;
  279. sys_data = syscall_nr_to_meta(syscall_nr);
  280. if (!sys_data)
  281. return;
  282. /* get the size after alignment with the u32 buffer size field */
  283. size = sizeof(unsigned long) * sys_data->nb_args + sizeof(*rec);
  284. size = ALIGN(size + sizeof(u32), sizeof(u64));
  285. size -= sizeof(u32);
  286. do {
  287. char raw_data[size];
  288. /* zero the dead bytes from align to not leak stack to user */
  289. *(u64 *)(&raw_data[size - sizeof(u64)]) = 0ULL;
  290. rec = (struct syscall_trace_enter *) raw_data;
  291. tracing_generic_entry_update(&rec->ent, 0, 0);
  292. rec->ent.type = sys_data->enter_id;
  293. rec->nr = syscall_nr;
  294. syscall_get_arguments(current, regs, 0, sys_data->nb_args,
  295. (unsigned long *)&rec->args);
  296. perf_tpcounter_event(sys_data->enter_id, 0, 1, rec, size);
  297. } while(0);
  298. }
  299. int reg_prof_syscall_enter(char *name)
  300. {
  301. int ret = 0;
  302. int num;
  303. num = syscall_name_to_nr(name);
  304. if (num < 0 || num >= FTRACE_SYSCALL_MAX)
  305. return -ENOSYS;
  306. mutex_lock(&syscall_trace_lock);
  307. if (!sys_prof_refcount_enter)
  308. ret = register_trace_syscall_enter(prof_syscall_enter);
  309. if (ret) {
  310. pr_info("event trace: Could not activate"
  311. "syscall entry trace point");
  312. } else {
  313. set_bit(num, enabled_prof_enter_syscalls);
  314. sys_prof_refcount_enter++;
  315. }
  316. mutex_unlock(&syscall_trace_lock);
  317. return ret;
  318. }
  319. void unreg_prof_syscall_enter(char *name)
  320. {
  321. int num;
  322. num = syscall_name_to_nr(name);
  323. if (num < 0 || num >= FTRACE_SYSCALL_MAX)
  324. return;
  325. mutex_lock(&syscall_trace_lock);
  326. sys_prof_refcount_enter--;
  327. clear_bit(num, enabled_prof_enter_syscalls);
  328. if (!sys_prof_refcount_enter)
  329. unregister_trace_syscall_enter(prof_syscall_enter);
  330. mutex_unlock(&syscall_trace_lock);
  331. }
  332. static void prof_syscall_exit(struct pt_regs *regs, long ret)
  333. {
  334. struct syscall_metadata *sys_data;
  335. struct syscall_trace_exit rec;
  336. int syscall_nr;
  337. syscall_nr = syscall_get_nr(current, regs);
  338. if (!test_bit(syscall_nr, enabled_prof_exit_syscalls))
  339. return;
  340. sys_data = syscall_nr_to_meta(syscall_nr);
  341. if (!sys_data)
  342. return;
  343. tracing_generic_entry_update(&rec.ent, 0, 0);
  344. rec.ent.type = sys_data->exit_id;
  345. rec.nr = syscall_nr;
  346. rec.ret = syscall_get_return_value(current, regs);
  347. perf_tpcounter_event(sys_data->exit_id, 0, 1, &rec, sizeof(rec));
  348. }
  349. int reg_prof_syscall_exit(char *name)
  350. {
  351. int ret = 0;
  352. int num;
  353. num = syscall_name_to_nr(name);
  354. if (num < 0 || num >= FTRACE_SYSCALL_MAX)
  355. return -ENOSYS;
  356. mutex_lock(&syscall_trace_lock);
  357. if (!sys_prof_refcount_exit)
  358. ret = register_trace_syscall_exit(prof_syscall_exit);
  359. if (ret) {
  360. pr_info("event trace: Could not activate"
  361. "syscall entry trace point");
  362. } else {
  363. set_bit(num, enabled_prof_exit_syscalls);
  364. sys_prof_refcount_exit++;
  365. }
  366. mutex_unlock(&syscall_trace_lock);
  367. return ret;
  368. }
  369. void unreg_prof_syscall_exit(char *name)
  370. {
  371. int num;
  372. num = syscall_name_to_nr(name);
  373. if (num < 0 || num >= FTRACE_SYSCALL_MAX)
  374. return;
  375. mutex_lock(&syscall_trace_lock);
  376. sys_prof_refcount_exit--;
  377. clear_bit(num, enabled_prof_exit_syscalls);
  378. if (!sys_prof_refcount_exit)
  379. unregister_trace_syscall_exit(prof_syscall_exit);
  380. mutex_unlock(&syscall_trace_lock);
  381. }
  382. #endif