1 #include <linux/types.h>
2 #include <stdlib.h>
3 #include <unistd.h>
4 #include <stdio.h>
5 #include <ctype.h>
6 #include <string.h>
7
8 #include "parse-events.h"
9 #include "evlist.h"
10 #include "evsel.h"
11 #include "thread_map.h"
12 #include "cpumap.h"
13 #include "machine.h"
14 #include "event.h"
15 #include "thread.h"
16
17 #include "tests.h"
18
19 #define BUFSZ 1024
20 #define READLEN 128
21
22 struct state {
23 u64 done[1024];
24 size_t done_cnt;
25 };
26
hex(char c)27 static unsigned int hex(char c)
28 {
29 if (c >= '0' && c <= '9')
30 return c - '0';
31 if (c >= 'a' && c <= 'f')
32 return c - 'a' + 10;
33 return c - 'A' + 10;
34 }
35
read_objdump_chunk(const char ** line,unsigned char ** buf,size_t * buf_len)36 static size_t read_objdump_chunk(const char **line, unsigned char **buf,
37 size_t *buf_len)
38 {
39 size_t bytes_read = 0;
40 unsigned char *chunk_start = *buf;
41
42 /* Read bytes */
43 while (*buf_len > 0) {
44 char c1, c2;
45
46 /* Get 2 hex digits */
47 c1 = *(*line)++;
48 if (!isxdigit(c1))
49 break;
50 c2 = *(*line)++;
51 if (!isxdigit(c2))
52 break;
53
54 /* Store byte and advance buf */
55 **buf = (hex(c1) << 4) | hex(c2);
56 (*buf)++;
57 (*buf_len)--;
58 bytes_read++;
59
60 /* End of chunk? */
61 if (isspace(**line))
62 break;
63 }
64
65 /*
66 * objdump will display raw insn as LE if code endian
67 * is LE and bytes_per_chunk > 1. In that case reverse
68 * the chunk we just read.
69 *
70 * see disassemble_bytes() at binutils/objdump.c for details
71 * how objdump chooses display endian)
72 */
73 if (bytes_read > 1 && !bigendian()) {
74 unsigned char *chunk_end = chunk_start + bytes_read - 1;
75 unsigned char tmp;
76
77 while (chunk_start < chunk_end) {
78 tmp = *chunk_start;
79 *chunk_start = *chunk_end;
80 *chunk_end = tmp;
81 chunk_start++;
82 chunk_end--;
83 }
84 }
85
86 return bytes_read;
87 }
88
read_objdump_line(const char * line,unsigned char * buf,size_t buf_len)89 static size_t read_objdump_line(const char *line, unsigned char *buf,
90 size_t buf_len)
91 {
92 const char *p;
93 size_t ret, bytes_read = 0;
94
95 /* Skip to a colon */
96 p = strchr(line, ':');
97 if (!p)
98 return 0;
99 p++;
100
101 /* Skip initial spaces */
102 while (*p) {
103 if (!isspace(*p))
104 break;
105 p++;
106 }
107
108 do {
109 ret = read_objdump_chunk(&p, &buf, &buf_len);
110 bytes_read += ret;
111 p++;
112 } while (ret > 0);
113
114 /* return number of successfully read bytes */
115 return bytes_read;
116 }
117
read_objdump_output(FILE * f,void * buf,size_t * len,u64 start_addr)118 static int read_objdump_output(FILE *f, void *buf, size_t *len, u64 start_addr)
119 {
120 char *line = NULL;
121 size_t line_len, off_last = 0;
122 ssize_t ret;
123 int err = 0;
124 u64 addr, last_addr = start_addr;
125
126 while (off_last < *len) {
127 size_t off, read_bytes, written_bytes;
128 unsigned char tmp[BUFSZ];
129
130 ret = getline(&line, &line_len, f);
131 if (feof(f))
132 break;
133 if (ret < 0) {
134 pr_debug("getline failed\n");
135 err = -1;
136 break;
137 }
138
139 /* read objdump data into temporary buffer */
140 read_bytes = read_objdump_line(line, tmp, sizeof(tmp));
141 if (!read_bytes)
142 continue;
143
144 if (sscanf(line, "%"PRIx64, &addr) != 1)
145 continue;
146 if (addr < last_addr) {
147 pr_debug("addr going backwards, read beyond section?\n");
148 break;
149 }
150 last_addr = addr;
151
152 /* copy it from temporary buffer to 'buf' according
153 * to address on current objdump line */
154 off = addr - start_addr;
155 if (off >= *len)
156 break;
157 written_bytes = MIN(read_bytes, *len - off);
158 memcpy(buf + off, tmp, written_bytes);
159 off_last = off + written_bytes;
160 }
161
162 /* len returns number of bytes that could not be read */
163 *len -= off_last;
164
165 free(line);
166
167 return err;
168 }
169
read_via_objdump(const char * filename,u64 addr,void * buf,size_t len)170 static int read_via_objdump(const char *filename, u64 addr, void *buf,
171 size_t len)
172 {
173 char cmd[PATH_MAX * 2];
174 const char *fmt;
175 FILE *f;
176 int ret;
177
178 fmt = "%s -z -d --start-address=0x%"PRIx64" --stop-address=0x%"PRIx64" %s";
179 ret = snprintf(cmd, sizeof(cmd), fmt, "objdump", addr, addr + len,
180 filename);
181 if (ret <= 0 || (size_t)ret >= sizeof(cmd))
182 return -1;
183
184 pr_debug("Objdump command is: %s\n", cmd);
185
186 /* Ignore objdump errors */
187 strcat(cmd, " 2>/dev/null");
188
189 f = popen(cmd, "r");
190 if (!f) {
191 pr_debug("popen failed\n");
192 return -1;
193 }
194
195 ret = read_objdump_output(f, buf, &len, addr);
196 if (len) {
197 pr_debug("objdump read too few bytes: %zd\n", len);
198 if (!ret)
199 ret = len;
200 }
201
202 pclose(f);
203
204 return ret;
205 }
206
dump_buf(unsigned char * buf,size_t len)207 static void dump_buf(unsigned char *buf, size_t len)
208 {
209 size_t i;
210
211 for (i = 0; i < len; i++) {
212 pr_debug("0x%02x ", buf[i]);
213 if (i % 16 == 15)
214 pr_debug("\n");
215 }
216 pr_debug("\n");
217 }
218
read_object_code(u64 addr,size_t len,u8 cpumode,struct thread * thread,struct state * state)219 static int read_object_code(u64 addr, size_t len, u8 cpumode,
220 struct thread *thread, struct state *state)
221 {
222 struct addr_location al;
223 unsigned char buf1[BUFSZ];
224 unsigned char buf2[BUFSZ];
225 size_t ret_len;
226 u64 objdump_addr;
227 int ret;
228
229 pr_debug("Reading object code for memory address: %#"PRIx64"\n", addr);
230
231 thread__find_addr_map(thread, cpumode, MAP__FUNCTION, addr, &al);
232 if (!al.map || !al.map->dso) {
233 pr_debug("thread__find_addr_map failed\n");
234 return -1;
235 }
236
237 pr_debug("File is: %s\n", al.map->dso->long_name);
238
239 if (al.map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
240 !dso__is_kcore(al.map->dso)) {
241 pr_debug("Unexpected kernel address - skipping\n");
242 return 0;
243 }
244
245 pr_debug("On file address is: %#"PRIx64"\n", al.addr);
246
247 if (len > BUFSZ)
248 len = BUFSZ;
249
250 /* Do not go off the map */
251 if (addr + len > al.map->end)
252 len = al.map->end - addr;
253
254 /* Read the object code using perf */
255 ret_len = dso__data_read_offset(al.map->dso, thread->mg->machine,
256 al.addr, buf1, len);
257 if (ret_len != len) {
258 pr_debug("dso__data_read_offset failed\n");
259 return -1;
260 }
261
262 /*
263 * Converting addresses for use by objdump requires more information.
264 * map__load() does that. See map__rip_2objdump() for details.
265 */
266 if (map__load(al.map))
267 return -1;
268
269 /* objdump struggles with kcore - try each map only once */
270 if (dso__is_kcore(al.map->dso)) {
271 size_t d;
272
273 for (d = 0; d < state->done_cnt; d++) {
274 if (state->done[d] == al.map->start) {
275 pr_debug("kcore map tested already");
276 pr_debug(" - skipping\n");
277 return 0;
278 }
279 }
280 if (state->done_cnt >= ARRAY_SIZE(state->done)) {
281 pr_debug("Too many kcore maps - skipping\n");
282 return 0;
283 }
284 state->done[state->done_cnt++] = al.map->start;
285 }
286
287 /* Read the object code using objdump */
288 objdump_addr = map__rip_2objdump(al.map, al.addr);
289 ret = read_via_objdump(al.map->dso->long_name, objdump_addr, buf2, len);
290 if (ret > 0) {
291 /*
292 * The kernel maps are inaccurate - assume objdump is right in
293 * that case.
294 */
295 if (cpumode == PERF_RECORD_MISC_KERNEL ||
296 cpumode == PERF_RECORD_MISC_GUEST_KERNEL) {
297 len -= ret;
298 if (len) {
299 pr_debug("Reducing len to %zu\n", len);
300 } else if (dso__is_kcore(al.map->dso)) {
301 /*
302 * objdump cannot handle very large segments
303 * that may be found in kcore.
304 */
305 pr_debug("objdump failed for kcore");
306 pr_debug(" - skipping\n");
307 return 0;
308 } else {
309 return -1;
310 }
311 }
312 }
313 if (ret < 0) {
314 pr_debug("read_via_objdump failed\n");
315 return -1;
316 }
317
318 /* The results should be identical */
319 if (memcmp(buf1, buf2, len)) {
320 pr_debug("Bytes read differ from those read by objdump\n");
321 pr_debug("buf1 (dso):\n");
322 dump_buf(buf1, len);
323 pr_debug("buf2 (objdump):\n");
324 dump_buf(buf2, len);
325 return -1;
326 }
327 pr_debug("Bytes read match those read by objdump\n");
328
329 return 0;
330 }
331
process_sample_event(struct machine * machine,struct perf_evlist * evlist,union perf_event * event,struct state * state)332 static int process_sample_event(struct machine *machine,
333 struct perf_evlist *evlist,
334 union perf_event *event, struct state *state)
335 {
336 struct perf_sample sample;
337 struct thread *thread;
338 int ret;
339
340 if (perf_evlist__parse_sample(evlist, event, &sample)) {
341 pr_debug("perf_evlist__parse_sample failed\n");
342 return -1;
343 }
344
345 thread = machine__findnew_thread(machine, sample.pid, sample.tid);
346 if (!thread) {
347 pr_debug("machine__findnew_thread failed\n");
348 return -1;
349 }
350
351 ret = read_object_code(sample.ip, READLEN, sample.cpumode, thread, state);
352 thread__put(thread);
353 return ret;
354 }
355
process_event(struct machine * machine,struct perf_evlist * evlist,union perf_event * event,struct state * state)356 static int process_event(struct machine *machine, struct perf_evlist *evlist,
357 union perf_event *event, struct state *state)
358 {
359 if (event->header.type == PERF_RECORD_SAMPLE)
360 return process_sample_event(machine, evlist, event, state);
361
362 if (event->header.type == PERF_RECORD_THROTTLE ||
363 event->header.type == PERF_RECORD_UNTHROTTLE)
364 return 0;
365
366 if (event->header.type < PERF_RECORD_MAX) {
367 int ret;
368
369 ret = machine__process_event(machine, event, NULL);
370 if (ret < 0)
371 pr_debug("machine__process_event failed, event type %u\n",
372 event->header.type);
373 return ret;
374 }
375
376 return 0;
377 }
378
process_events(struct machine * machine,struct perf_evlist * evlist,struct state * state)379 static int process_events(struct machine *machine, struct perf_evlist *evlist,
380 struct state *state)
381 {
382 union perf_event *event;
383 int i, ret;
384
385 for (i = 0; i < evlist->nr_mmaps; i++) {
386 while ((event = perf_evlist__mmap_read(evlist, i)) != NULL) {
387 ret = process_event(machine, evlist, event, state);
388 perf_evlist__mmap_consume(evlist, i);
389 if (ret < 0)
390 return ret;
391 }
392 }
393 return 0;
394 }
395
comp(const void * a,const void * b)396 static int comp(const void *a, const void *b)
397 {
398 return *(int *)a - *(int *)b;
399 }
400
do_sort_something(void)401 static void do_sort_something(void)
402 {
403 int buf[40960], i;
404
405 for (i = 0; i < (int)ARRAY_SIZE(buf); i++)
406 buf[i] = ARRAY_SIZE(buf) - i - 1;
407
408 qsort(buf, ARRAY_SIZE(buf), sizeof(int), comp);
409
410 for (i = 0; i < (int)ARRAY_SIZE(buf); i++) {
411 if (buf[i] != i) {
412 pr_debug("qsort failed\n");
413 break;
414 }
415 }
416 }
417
sort_something(void)418 static void sort_something(void)
419 {
420 int i;
421
422 for (i = 0; i < 10; i++)
423 do_sort_something();
424 }
425
syscall_something(void)426 static void syscall_something(void)
427 {
428 int pipefd[2];
429 int i;
430
431 for (i = 0; i < 1000; i++) {
432 if (pipe(pipefd) < 0) {
433 pr_debug("pipe failed\n");
434 break;
435 }
436 close(pipefd[1]);
437 close(pipefd[0]);
438 }
439 }
440
fs_something(void)441 static void fs_something(void)
442 {
443 const char *test_file_name = "temp-perf-code-reading-test-file--";
444 FILE *f;
445 int i;
446
447 for (i = 0; i < 1000; i++) {
448 f = fopen(test_file_name, "w+");
449 if (f) {
450 fclose(f);
451 unlink(test_file_name);
452 }
453 }
454 }
455
do_something(void)456 static void do_something(void)
457 {
458 fs_something();
459
460 sort_something();
461
462 syscall_something();
463 }
464
465 enum {
466 TEST_CODE_READING_OK,
467 TEST_CODE_READING_NO_VMLINUX,
468 TEST_CODE_READING_NO_KCORE,
469 TEST_CODE_READING_NO_ACCESS,
470 TEST_CODE_READING_NO_KERNEL_OBJ,
471 };
472
do_test_code_reading(bool try_kcore)473 static int do_test_code_reading(bool try_kcore)
474 {
475 struct machine *machine;
476 struct thread *thread;
477 struct record_opts opts = {
478 .mmap_pages = UINT_MAX,
479 .user_freq = UINT_MAX,
480 .user_interval = ULLONG_MAX,
481 .freq = 500,
482 .target = {
483 .uses_mmap = true,
484 },
485 };
486 struct state state = {
487 .done_cnt = 0,
488 };
489 struct thread_map *threads = NULL;
490 struct cpu_map *cpus = NULL;
491 struct perf_evlist *evlist = NULL;
492 struct perf_evsel *evsel = NULL;
493 int err = -1, ret;
494 pid_t pid;
495 struct map *map;
496 bool have_vmlinux, have_kcore, excl_kernel = false;
497
498 pid = getpid();
499
500 machine = machine__new_host();
501
502 ret = machine__create_kernel_maps(machine);
503 if (ret < 0) {
504 pr_debug("machine__create_kernel_maps failed\n");
505 goto out_err;
506 }
507
508 /* Force the use of kallsyms instead of vmlinux to try kcore */
509 if (try_kcore)
510 symbol_conf.kallsyms_name = "/proc/kallsyms";
511
512 /* Load kernel map */
513 map = machine__kernel_map(machine);
514 ret = map__load(map);
515 if (ret < 0) {
516 pr_debug("map__load failed\n");
517 goto out_err;
518 }
519 have_vmlinux = dso__is_vmlinux(map->dso);
520 have_kcore = dso__is_kcore(map->dso);
521
522 /* 2nd time through we just try kcore */
523 if (try_kcore && !have_kcore)
524 return TEST_CODE_READING_NO_KCORE;
525
526 /* No point getting kernel events if there is no kernel object */
527 if (!have_vmlinux && !have_kcore)
528 excl_kernel = true;
529
530 threads = thread_map__new_by_tid(pid);
531 if (!threads) {
532 pr_debug("thread_map__new_by_tid failed\n");
533 goto out_err;
534 }
535
536 ret = perf_event__synthesize_thread_map(NULL, threads,
537 perf_event__process, machine, false, 500);
538 if (ret < 0) {
539 pr_debug("perf_event__synthesize_thread_map failed\n");
540 goto out_err;
541 }
542
543 thread = machine__findnew_thread(machine, pid, pid);
544 if (!thread) {
545 pr_debug("machine__findnew_thread failed\n");
546 goto out_put;
547 }
548
549 cpus = cpu_map__new(NULL);
550 if (!cpus) {
551 pr_debug("cpu_map__new failed\n");
552 goto out_put;
553 }
554
555 while (1) {
556 const char *str;
557
558 evlist = perf_evlist__new();
559 if (!evlist) {
560 pr_debug("perf_evlist__new failed\n");
561 goto out_put;
562 }
563
564 perf_evlist__set_maps(evlist, cpus, threads);
565
566 if (excl_kernel)
567 str = "cycles:u";
568 else
569 str = "cycles";
570 pr_debug("Parsing event '%s'\n", str);
571 ret = parse_events(evlist, str, NULL);
572 if (ret < 0) {
573 pr_debug("parse_events failed\n");
574 goto out_put;
575 }
576
577 perf_evlist__config(evlist, &opts, NULL);
578
579 evsel = perf_evlist__first(evlist);
580
581 evsel->attr.comm = 1;
582 evsel->attr.disabled = 1;
583 evsel->attr.enable_on_exec = 0;
584
585 ret = perf_evlist__open(evlist);
586 if (ret < 0) {
587 if (!excl_kernel) {
588 excl_kernel = true;
589 /*
590 * Both cpus and threads are now owned by evlist
591 * and will be freed by following perf_evlist__set_maps
592 * call. Getting refference to keep them alive.
593 */
594 cpu_map__get(cpus);
595 thread_map__get(threads);
596 perf_evlist__set_maps(evlist, NULL, NULL);
597 perf_evlist__delete(evlist);
598 evlist = NULL;
599 continue;
600 }
601
602 if (verbose) {
603 char errbuf[512];
604 perf_evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
605 pr_debug("perf_evlist__open() failed!\n%s\n", errbuf);
606 }
607
608 goto out_put;
609 }
610 break;
611 }
612
613 ret = perf_evlist__mmap(evlist, UINT_MAX, false);
614 if (ret < 0) {
615 pr_debug("perf_evlist__mmap failed\n");
616 goto out_put;
617 }
618
619 perf_evlist__enable(evlist);
620
621 do_something();
622
623 perf_evlist__disable(evlist);
624
625 ret = process_events(machine, evlist, &state);
626 if (ret < 0)
627 goto out_put;
628
629 if (!have_vmlinux && !have_kcore && !try_kcore)
630 err = TEST_CODE_READING_NO_KERNEL_OBJ;
631 else if (!have_vmlinux && !try_kcore)
632 err = TEST_CODE_READING_NO_VMLINUX;
633 else if (excl_kernel)
634 err = TEST_CODE_READING_NO_ACCESS;
635 else
636 err = TEST_CODE_READING_OK;
637 out_put:
638 thread__put(thread);
639 out_err:
640
641 if (evlist) {
642 perf_evlist__delete(evlist);
643 } else {
644 cpu_map__put(cpus);
645 thread_map__put(threads);
646 }
647 machine__delete_threads(machine);
648 machine__delete(machine);
649
650 return err;
651 }
652
test__code_reading(int subtest __maybe_unused)653 int test__code_reading(int subtest __maybe_unused)
654 {
655 int ret;
656
657 ret = do_test_code_reading(false);
658 if (!ret)
659 ret = do_test_code_reading(true);
660
661 switch (ret) {
662 case TEST_CODE_READING_OK:
663 return 0;
664 case TEST_CODE_READING_NO_VMLINUX:
665 pr_debug("no vmlinux\n");
666 return 0;
667 case TEST_CODE_READING_NO_KCORE:
668 pr_debug("no kcore\n");
669 return 0;
670 case TEST_CODE_READING_NO_ACCESS:
671 pr_debug("no access\n");
672 return 0;
673 case TEST_CODE_READING_NO_KERNEL_OBJ:
674 pr_debug("no kernel obj\n");
675 return 0;
676 default:
677 return -1;
678 };
679 }
680