xref: /openbmc/linux/tools/perf/tests/code-reading.c (revision 8cb5d748)
1 #include <errno.h>
2 #include <linux/kernel.h>
3 #include <linux/types.h>
4 #include <inttypes.h>
5 #include <stdlib.h>
6 #include <unistd.h>
7 #include <stdio.h>
8 #include <string.h>
9 #include <sys/param.h>
10 
11 #include "parse-events.h"
12 #include "evlist.h"
13 #include "evsel.h"
14 #include "thread_map.h"
15 #include "cpumap.h"
16 #include "machine.h"
17 #include "event.h"
18 #include "thread.h"
19 
20 #include "tests.h"
21 
22 #include "sane_ctype.h"
23 
24 #define BUFSZ	1024
25 #define READLEN	128
26 
27 struct state {
28 	u64 done[1024];
29 	size_t done_cnt;
30 };
31 
32 static unsigned int hex(char c)
33 {
34 	if (c >= '0' && c <= '9')
35 		return c - '0';
36 	if (c >= 'a' && c <= 'f')
37 		return c - 'a' + 10;
38 	return c - 'A' + 10;
39 }
40 
41 static size_t read_objdump_chunk(const char **line, unsigned char **buf,
42 				 size_t *buf_len)
43 {
44 	size_t bytes_read = 0;
45 	unsigned char *chunk_start = *buf;
46 
47 	/* Read bytes */
48 	while (*buf_len > 0) {
49 		char c1, c2;
50 
51 		/* Get 2 hex digits */
52 		c1 = *(*line)++;
53 		if (!isxdigit(c1))
54 			break;
55 		c2 = *(*line)++;
56 		if (!isxdigit(c2))
57 			break;
58 
59 		/* Store byte and advance buf */
60 		**buf = (hex(c1) << 4) | hex(c2);
61 		(*buf)++;
62 		(*buf_len)--;
63 		bytes_read++;
64 
65 		/* End of chunk? */
66 		if (isspace(**line))
67 			break;
68 	}
69 
70 	/*
71 	 * objdump will display raw insn as LE if code endian
72 	 * is LE and bytes_per_chunk > 1. In that case reverse
73 	 * the chunk we just read.
74 	 *
75 	 * see disassemble_bytes() at binutils/objdump.c for details
76 	 * how objdump chooses display endian)
77 	 */
78 	if (bytes_read > 1 && !bigendian()) {
79 		unsigned char *chunk_end = chunk_start + bytes_read - 1;
80 		unsigned char tmp;
81 
82 		while (chunk_start < chunk_end) {
83 			tmp = *chunk_start;
84 			*chunk_start = *chunk_end;
85 			*chunk_end = tmp;
86 			chunk_start++;
87 			chunk_end--;
88 		}
89 	}
90 
91 	return bytes_read;
92 }
93 
94 static size_t read_objdump_line(const char *line, unsigned char *buf,
95 				size_t buf_len)
96 {
97 	const char *p;
98 	size_t ret, bytes_read = 0;
99 
100 	/* Skip to a colon */
101 	p = strchr(line, ':');
102 	if (!p)
103 		return 0;
104 	p++;
105 
106 	/* Skip initial spaces */
107 	while (*p) {
108 		if (!isspace(*p))
109 			break;
110 		p++;
111 	}
112 
113 	do {
114 		ret = read_objdump_chunk(&p, &buf, &buf_len);
115 		bytes_read += ret;
116 		p++;
117 	} while (ret > 0);
118 
119 	/* return number of successfully read bytes */
120 	return bytes_read;
121 }
122 
123 static int read_objdump_output(FILE *f, void *buf, size_t *len, u64 start_addr)
124 {
125 	char *line = NULL;
126 	size_t line_len, off_last = 0;
127 	ssize_t ret;
128 	int err = 0;
129 	u64 addr, last_addr = start_addr;
130 
131 	while (off_last < *len) {
132 		size_t off, read_bytes, written_bytes;
133 		unsigned char tmp[BUFSZ];
134 
135 		ret = getline(&line, &line_len, f);
136 		if (feof(f))
137 			break;
138 		if (ret < 0) {
139 			pr_debug("getline failed\n");
140 			err = -1;
141 			break;
142 		}
143 
144 		/* read objdump data into temporary buffer */
145 		read_bytes = read_objdump_line(line, tmp, sizeof(tmp));
146 		if (!read_bytes)
147 			continue;
148 
149 		if (sscanf(line, "%"PRIx64, &addr) != 1)
150 			continue;
151 		if (addr < last_addr) {
152 			pr_debug("addr going backwards, read beyond section?\n");
153 			break;
154 		}
155 		last_addr = addr;
156 
157 		/* copy it from temporary buffer to 'buf' according
158 		 * to address on current objdump line */
159 		off = addr - start_addr;
160 		if (off >= *len)
161 			break;
162 		written_bytes = MIN(read_bytes, *len - off);
163 		memcpy(buf + off, tmp, written_bytes);
164 		off_last = off + written_bytes;
165 	}
166 
167 	/* len returns number of bytes that could not be read */
168 	*len -= off_last;
169 
170 	free(line);
171 
172 	return err;
173 }
174 
175 static int read_via_objdump(const char *filename, u64 addr, void *buf,
176 			    size_t len)
177 {
178 	char cmd[PATH_MAX * 2];
179 	const char *fmt;
180 	FILE *f;
181 	int ret;
182 
183 	fmt = "%s -z -d --start-address=0x%"PRIx64" --stop-address=0x%"PRIx64" %s";
184 	ret = snprintf(cmd, sizeof(cmd), fmt, "objdump", addr, addr + len,
185 		       filename);
186 	if (ret <= 0 || (size_t)ret >= sizeof(cmd))
187 		return -1;
188 
189 	pr_debug("Objdump command is: %s\n", cmd);
190 
191 	/* Ignore objdump errors */
192 	strcat(cmd, " 2>/dev/null");
193 
194 	f = popen(cmd, "r");
195 	if (!f) {
196 		pr_debug("popen failed\n");
197 		return -1;
198 	}
199 
200 	ret = read_objdump_output(f, buf, &len, addr);
201 	if (len) {
202 		pr_debug("objdump read too few bytes: %zd\n", len);
203 		if (!ret)
204 			ret = len;
205 	}
206 
207 	pclose(f);
208 
209 	return ret;
210 }
211 
212 static void dump_buf(unsigned char *buf, size_t len)
213 {
214 	size_t i;
215 
216 	for (i = 0; i < len; i++) {
217 		pr_debug("0x%02x ", buf[i]);
218 		if (i % 16 == 15)
219 			pr_debug("\n");
220 	}
221 	pr_debug("\n");
222 }
223 
224 static int read_object_code(u64 addr, size_t len, u8 cpumode,
225 			    struct thread *thread, struct state *state)
226 {
227 	struct addr_location al;
228 	unsigned char buf1[BUFSZ];
229 	unsigned char buf2[BUFSZ];
230 	size_t ret_len;
231 	u64 objdump_addr;
232 	const char *objdump_name;
233 	char decomp_name[KMOD_DECOMP_LEN];
234 	int ret;
235 
236 	pr_debug("Reading object code for memory address: %#"PRIx64"\n", addr);
237 
238 	thread__find_addr_map(thread, cpumode, MAP__FUNCTION, addr, &al);
239 	if (!al.map || !al.map->dso) {
240 		if (cpumode == PERF_RECORD_MISC_HYPERVISOR) {
241 			pr_debug("Hypervisor address can not be resolved - skipping\n");
242 			return 0;
243 		}
244 
245 		pr_debug("thread__find_addr_map failed\n");
246 		return -1;
247 	}
248 
249 	pr_debug("File is: %s\n", al.map->dso->long_name);
250 
251 	if (al.map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
252 	    !dso__is_kcore(al.map->dso)) {
253 		pr_debug("Unexpected kernel address - skipping\n");
254 		return 0;
255 	}
256 
257 	pr_debug("On file address is: %#"PRIx64"\n", al.addr);
258 
259 	if (len > BUFSZ)
260 		len = BUFSZ;
261 
262 	/* Do not go off the map */
263 	if (addr + len > al.map->end)
264 		len = al.map->end - addr;
265 
266 	/* Read the object code using perf */
267 	ret_len = dso__data_read_offset(al.map->dso, thread->mg->machine,
268 					al.addr, buf1, len);
269 	if (ret_len != len) {
270 		pr_debug("dso__data_read_offset failed\n");
271 		return -1;
272 	}
273 
274 	/*
275 	 * Converting addresses for use by objdump requires more information.
276 	 * map__load() does that.  See map__rip_2objdump() for details.
277 	 */
278 	if (map__load(al.map))
279 		return -1;
280 
281 	/* objdump struggles with kcore - try each map only once */
282 	if (dso__is_kcore(al.map->dso)) {
283 		size_t d;
284 
285 		for (d = 0; d < state->done_cnt; d++) {
286 			if (state->done[d] == al.map->start) {
287 				pr_debug("kcore map tested already");
288 				pr_debug(" - skipping\n");
289 				return 0;
290 			}
291 		}
292 		if (state->done_cnt >= ARRAY_SIZE(state->done)) {
293 			pr_debug("Too many kcore maps - skipping\n");
294 			return 0;
295 		}
296 		state->done[state->done_cnt++] = al.map->start;
297 	}
298 
299 	objdump_name = al.map->dso->long_name;
300 	if (dso__needs_decompress(al.map->dso)) {
301 		if (dso__decompress_kmodule_path(al.map->dso, objdump_name,
302 						 decomp_name,
303 						 sizeof(decomp_name)) < 0) {
304 			pr_debug("decompression failed\n");
305 			return -1;
306 		}
307 
308 		objdump_name = decomp_name;
309 	}
310 
311 	/* Read the object code using objdump */
312 	objdump_addr = map__rip_2objdump(al.map, al.addr);
313 	ret = read_via_objdump(objdump_name, objdump_addr, buf2, len);
314 
315 	if (dso__needs_decompress(al.map->dso))
316 		unlink(objdump_name);
317 
318 	if (ret > 0) {
319 		/*
320 		 * The kernel maps are inaccurate - assume objdump is right in
321 		 * that case.
322 		 */
323 		if (cpumode == PERF_RECORD_MISC_KERNEL ||
324 		    cpumode == PERF_RECORD_MISC_GUEST_KERNEL) {
325 			len -= ret;
326 			if (len) {
327 				pr_debug("Reducing len to %zu\n", len);
328 			} else if (dso__is_kcore(al.map->dso)) {
329 				/*
330 				 * objdump cannot handle very large segments
331 				 * that may be found in kcore.
332 				 */
333 				pr_debug("objdump failed for kcore");
334 				pr_debug(" - skipping\n");
335 				return 0;
336 			} else {
337 				return -1;
338 			}
339 		}
340 	}
341 	if (ret < 0) {
342 		pr_debug("read_via_objdump failed\n");
343 		return -1;
344 	}
345 
346 	/* The results should be identical */
347 	if (memcmp(buf1, buf2, len)) {
348 		pr_debug("Bytes read differ from those read by objdump\n");
349 		pr_debug("buf1 (dso):\n");
350 		dump_buf(buf1, len);
351 		pr_debug("buf2 (objdump):\n");
352 		dump_buf(buf2, len);
353 		return -1;
354 	}
355 	pr_debug("Bytes read match those read by objdump\n");
356 
357 	return 0;
358 }
359 
360 static int process_sample_event(struct machine *machine,
361 				struct perf_evlist *evlist,
362 				union perf_event *event, struct state *state)
363 {
364 	struct perf_sample sample;
365 	struct thread *thread;
366 	int ret;
367 
368 	if (perf_evlist__parse_sample(evlist, event, &sample)) {
369 		pr_debug("perf_evlist__parse_sample failed\n");
370 		return -1;
371 	}
372 
373 	thread = machine__findnew_thread(machine, sample.pid, sample.tid);
374 	if (!thread) {
375 		pr_debug("machine__findnew_thread failed\n");
376 		return -1;
377 	}
378 
379 	ret = read_object_code(sample.ip, READLEN, sample.cpumode, thread, state);
380 	thread__put(thread);
381 	return ret;
382 }
383 
384 static int process_event(struct machine *machine, struct perf_evlist *evlist,
385 			 union perf_event *event, struct state *state)
386 {
387 	if (event->header.type == PERF_RECORD_SAMPLE)
388 		return process_sample_event(machine, evlist, event, state);
389 
390 	if (event->header.type == PERF_RECORD_THROTTLE ||
391 	    event->header.type == PERF_RECORD_UNTHROTTLE)
392 		return 0;
393 
394 	if (event->header.type < PERF_RECORD_MAX) {
395 		int ret;
396 
397 		ret = machine__process_event(machine, event, NULL);
398 		if (ret < 0)
399 			pr_debug("machine__process_event failed, event type %u\n",
400 				 event->header.type);
401 		return ret;
402 	}
403 
404 	return 0;
405 }
406 
407 static int process_events(struct machine *machine, struct perf_evlist *evlist,
408 			  struct state *state)
409 {
410 	union perf_event *event;
411 	int i, ret;
412 
413 	for (i = 0; i < evlist->nr_mmaps; i++) {
414 		while ((event = perf_evlist__mmap_read(evlist, i)) != NULL) {
415 			ret = process_event(machine, evlist, event, state);
416 			perf_evlist__mmap_consume(evlist, i);
417 			if (ret < 0)
418 				return ret;
419 		}
420 	}
421 	return 0;
422 }
423 
424 static int comp(const void *a, const void *b)
425 {
426 	return *(int *)a - *(int *)b;
427 }
428 
429 static void do_sort_something(void)
430 {
431 	int buf[40960], i;
432 
433 	for (i = 0; i < (int)ARRAY_SIZE(buf); i++)
434 		buf[i] = ARRAY_SIZE(buf) - i - 1;
435 
436 	qsort(buf, ARRAY_SIZE(buf), sizeof(int), comp);
437 
438 	for (i = 0; i < (int)ARRAY_SIZE(buf); i++) {
439 		if (buf[i] != i) {
440 			pr_debug("qsort failed\n");
441 			break;
442 		}
443 	}
444 }
445 
446 static void sort_something(void)
447 {
448 	int i;
449 
450 	for (i = 0; i < 10; i++)
451 		do_sort_something();
452 }
453 
454 static void syscall_something(void)
455 {
456 	int pipefd[2];
457 	int i;
458 
459 	for (i = 0; i < 1000; i++) {
460 		if (pipe(pipefd) < 0) {
461 			pr_debug("pipe failed\n");
462 			break;
463 		}
464 		close(pipefd[1]);
465 		close(pipefd[0]);
466 	}
467 }
468 
469 static void fs_something(void)
470 {
471 	const char *test_file_name = "temp-perf-code-reading-test-file--";
472 	FILE *f;
473 	int i;
474 
475 	for (i = 0; i < 1000; i++) {
476 		f = fopen(test_file_name, "w+");
477 		if (f) {
478 			fclose(f);
479 			unlink(test_file_name);
480 		}
481 	}
482 }
483 
484 static void do_something(void)
485 {
486 	fs_something();
487 
488 	sort_something();
489 
490 	syscall_something();
491 }
492 
493 enum {
494 	TEST_CODE_READING_OK,
495 	TEST_CODE_READING_NO_VMLINUX,
496 	TEST_CODE_READING_NO_KCORE,
497 	TEST_CODE_READING_NO_ACCESS,
498 	TEST_CODE_READING_NO_KERNEL_OBJ,
499 };
500 
501 static int do_test_code_reading(bool try_kcore)
502 {
503 	struct machine *machine;
504 	struct thread *thread;
505 	struct record_opts opts = {
506 		.mmap_pages	     = UINT_MAX,
507 		.user_freq	     = UINT_MAX,
508 		.user_interval	     = ULLONG_MAX,
509 		.freq		     = 500,
510 		.target		     = {
511 			.uses_mmap   = true,
512 		},
513 	};
514 	struct state state = {
515 		.done_cnt = 0,
516 	};
517 	struct thread_map *threads = NULL;
518 	struct cpu_map *cpus = NULL;
519 	struct perf_evlist *evlist = NULL;
520 	struct perf_evsel *evsel = NULL;
521 	int err = -1, ret;
522 	pid_t pid;
523 	struct map *map;
524 	bool have_vmlinux, have_kcore, excl_kernel = false;
525 
526 	pid = getpid();
527 
528 	machine = machine__new_host();
529 
530 	ret = machine__create_kernel_maps(machine);
531 	if (ret < 0) {
532 		pr_debug("machine__create_kernel_maps failed\n");
533 		goto out_err;
534 	}
535 
536 	/* Force the use of kallsyms instead of vmlinux to try kcore */
537 	if (try_kcore)
538 		symbol_conf.kallsyms_name = "/proc/kallsyms";
539 
540 	/* Load kernel map */
541 	map = machine__kernel_map(machine);
542 	ret = map__load(map);
543 	if (ret < 0) {
544 		pr_debug("map__load failed\n");
545 		goto out_err;
546 	}
547 	have_vmlinux = dso__is_vmlinux(map->dso);
548 	have_kcore = dso__is_kcore(map->dso);
549 
550 	/* 2nd time through we just try kcore */
551 	if (try_kcore && !have_kcore)
552 		return TEST_CODE_READING_NO_KCORE;
553 
554 	/* No point getting kernel events if there is no kernel object */
555 	if (!have_vmlinux && !have_kcore)
556 		excl_kernel = true;
557 
558 	threads = thread_map__new_by_tid(pid);
559 	if (!threads) {
560 		pr_debug("thread_map__new_by_tid failed\n");
561 		goto out_err;
562 	}
563 
564 	ret = perf_event__synthesize_thread_map(NULL, threads,
565 						perf_event__process, machine, false, 500);
566 	if (ret < 0) {
567 		pr_debug("perf_event__synthesize_thread_map failed\n");
568 		goto out_err;
569 	}
570 
571 	thread = machine__findnew_thread(machine, pid, pid);
572 	if (!thread) {
573 		pr_debug("machine__findnew_thread failed\n");
574 		goto out_put;
575 	}
576 
577 	cpus = cpu_map__new(NULL);
578 	if (!cpus) {
579 		pr_debug("cpu_map__new failed\n");
580 		goto out_put;
581 	}
582 
583 	while (1) {
584 		const char *str;
585 
586 		evlist = perf_evlist__new();
587 		if (!evlist) {
588 			pr_debug("perf_evlist__new failed\n");
589 			goto out_put;
590 		}
591 
592 		perf_evlist__set_maps(evlist, cpus, threads);
593 
594 		if (excl_kernel)
595 			str = "cycles:u";
596 		else
597 			str = "cycles";
598 		pr_debug("Parsing event '%s'\n", str);
599 		ret = parse_events(evlist, str, NULL);
600 		if (ret < 0) {
601 			pr_debug("parse_events failed\n");
602 			goto out_put;
603 		}
604 
605 		perf_evlist__config(evlist, &opts, NULL);
606 
607 		evsel = perf_evlist__first(evlist);
608 
609 		evsel->attr.comm = 1;
610 		evsel->attr.disabled = 1;
611 		evsel->attr.enable_on_exec = 0;
612 
613 		ret = perf_evlist__open(evlist);
614 		if (ret < 0) {
615 			if (!excl_kernel) {
616 				excl_kernel = true;
617 				/*
618 				 * Both cpus and threads are now owned by evlist
619 				 * and will be freed by following perf_evlist__set_maps
620 				 * call. Getting refference to keep them alive.
621 				 */
622 				cpu_map__get(cpus);
623 				thread_map__get(threads);
624 				perf_evlist__set_maps(evlist, NULL, NULL);
625 				perf_evlist__delete(evlist);
626 				evlist = NULL;
627 				continue;
628 			}
629 
630 			if (verbose > 0) {
631 				char errbuf[512];
632 				perf_evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
633 				pr_debug("perf_evlist__open() failed!\n%s\n", errbuf);
634 			}
635 
636 			goto out_put;
637 		}
638 		break;
639 	}
640 
641 	ret = perf_evlist__mmap(evlist, UINT_MAX, false);
642 	if (ret < 0) {
643 		pr_debug("perf_evlist__mmap failed\n");
644 		goto out_put;
645 	}
646 
647 	perf_evlist__enable(evlist);
648 
649 	do_something();
650 
651 	perf_evlist__disable(evlist);
652 
653 	ret = process_events(machine, evlist, &state);
654 	if (ret < 0)
655 		goto out_put;
656 
657 	if (!have_vmlinux && !have_kcore && !try_kcore)
658 		err = TEST_CODE_READING_NO_KERNEL_OBJ;
659 	else if (!have_vmlinux && !try_kcore)
660 		err = TEST_CODE_READING_NO_VMLINUX;
661 	else if (excl_kernel)
662 		err = TEST_CODE_READING_NO_ACCESS;
663 	else
664 		err = TEST_CODE_READING_OK;
665 out_put:
666 	thread__put(thread);
667 out_err:
668 
669 	if (evlist) {
670 		perf_evlist__delete(evlist);
671 	} else {
672 		cpu_map__put(cpus);
673 		thread_map__put(threads);
674 	}
675 	machine__delete_threads(machine);
676 	machine__delete(machine);
677 
678 	return err;
679 }
680 
681 int test__code_reading(struct test *test __maybe_unused, int subtest __maybe_unused)
682 {
683 	int ret;
684 
685 	ret = do_test_code_reading(false);
686 	if (!ret)
687 		ret = do_test_code_reading(true);
688 
689 	switch (ret) {
690 	case TEST_CODE_READING_OK:
691 		return 0;
692 	case TEST_CODE_READING_NO_VMLINUX:
693 		pr_debug("no vmlinux\n");
694 		return 0;
695 	case TEST_CODE_READING_NO_KCORE:
696 		pr_debug("no kcore\n");
697 		return 0;
698 	case TEST_CODE_READING_NO_ACCESS:
699 		pr_debug("no access\n");
700 		return 0;
701 	case TEST_CODE_READING_NO_KERNEL_OBJ:
702 		pr_debug("no kernel obj\n");
703 		return 0;
704 	default:
705 		return -1;
706 	};
707 }
708