xref: /openbmc/linux/tools/perf/tests/code-reading.c (revision 2359ccdd)
1 // SPDX-License-Identifier: GPL-2.0
2 #include <errno.h>
3 #include <linux/kernel.h>
4 #include <linux/types.h>
5 #include <inttypes.h>
6 #include <stdlib.h>
7 #include <unistd.h>
8 #include <stdio.h>
9 #include <string.h>
10 #include <sys/param.h>
11 
12 #include "parse-events.h"
13 #include "evlist.h"
14 #include "evsel.h"
15 #include "thread_map.h"
16 #include "cpumap.h"
17 #include "machine.h"
18 #include "event.h"
19 #include "thread.h"
20 
21 #include "tests.h"
22 
23 #include "sane_ctype.h"
24 
25 #define BUFSZ	1024
26 #define READLEN	128
27 
28 struct state {
29 	u64 done[1024];
30 	size_t done_cnt;
31 };
32 
33 static unsigned int hex(char c)
34 {
35 	if (c >= '0' && c <= '9')
36 		return c - '0';
37 	if (c >= 'a' && c <= 'f')
38 		return c - 'a' + 10;
39 	return c - 'A' + 10;
40 }
41 
42 static size_t read_objdump_chunk(const char **line, unsigned char **buf,
43 				 size_t *buf_len)
44 {
45 	size_t bytes_read = 0;
46 	unsigned char *chunk_start = *buf;
47 
48 	/* Read bytes */
49 	while (*buf_len > 0) {
50 		char c1, c2;
51 
52 		/* Get 2 hex digits */
53 		c1 = *(*line)++;
54 		if (!isxdigit(c1))
55 			break;
56 		c2 = *(*line)++;
57 		if (!isxdigit(c2))
58 			break;
59 
60 		/* Store byte and advance buf */
61 		**buf = (hex(c1) << 4) | hex(c2);
62 		(*buf)++;
63 		(*buf_len)--;
64 		bytes_read++;
65 
66 		/* End of chunk? */
67 		if (isspace(**line))
68 			break;
69 	}
70 
71 	/*
72 	 * objdump will display raw insn as LE if code endian
73 	 * is LE and bytes_per_chunk > 1. In that case reverse
74 	 * the chunk we just read.
75 	 *
76 	 * see disassemble_bytes() at binutils/objdump.c for details
77 	 * how objdump chooses display endian)
78 	 */
79 	if (bytes_read > 1 && !bigendian()) {
80 		unsigned char *chunk_end = chunk_start + bytes_read - 1;
81 		unsigned char tmp;
82 
83 		while (chunk_start < chunk_end) {
84 			tmp = *chunk_start;
85 			*chunk_start = *chunk_end;
86 			*chunk_end = tmp;
87 			chunk_start++;
88 			chunk_end--;
89 		}
90 	}
91 
92 	return bytes_read;
93 }
94 
95 static size_t read_objdump_line(const char *line, unsigned char *buf,
96 				size_t buf_len)
97 {
98 	const char *p;
99 	size_t ret, bytes_read = 0;
100 
101 	/* Skip to a colon */
102 	p = strchr(line, ':');
103 	if (!p)
104 		return 0;
105 	p++;
106 
107 	/* Skip initial spaces */
108 	while (*p) {
109 		if (!isspace(*p))
110 			break;
111 		p++;
112 	}
113 
114 	do {
115 		ret = read_objdump_chunk(&p, &buf, &buf_len);
116 		bytes_read += ret;
117 		p++;
118 	} while (ret > 0);
119 
120 	/* return number of successfully read bytes */
121 	return bytes_read;
122 }
123 
124 static int read_objdump_output(FILE *f, void *buf, size_t *len, u64 start_addr)
125 {
126 	char *line = NULL;
127 	size_t line_len, off_last = 0;
128 	ssize_t ret;
129 	int err = 0;
130 	u64 addr, last_addr = start_addr;
131 
132 	while (off_last < *len) {
133 		size_t off, read_bytes, written_bytes;
134 		unsigned char tmp[BUFSZ];
135 
136 		ret = getline(&line, &line_len, f);
137 		if (feof(f))
138 			break;
139 		if (ret < 0) {
140 			pr_debug("getline failed\n");
141 			err = -1;
142 			break;
143 		}
144 
145 		/* read objdump data into temporary buffer */
146 		read_bytes = read_objdump_line(line, tmp, sizeof(tmp));
147 		if (!read_bytes)
148 			continue;
149 
150 		if (sscanf(line, "%"PRIx64, &addr) != 1)
151 			continue;
152 		if (addr < last_addr) {
153 			pr_debug("addr going backwards, read beyond section?\n");
154 			break;
155 		}
156 		last_addr = addr;
157 
158 		/* copy it from temporary buffer to 'buf' according
159 		 * to address on current objdump line */
160 		off = addr - start_addr;
161 		if (off >= *len)
162 			break;
163 		written_bytes = MIN(read_bytes, *len - off);
164 		memcpy(buf + off, tmp, written_bytes);
165 		off_last = off + written_bytes;
166 	}
167 
168 	/* len returns number of bytes that could not be read */
169 	*len -= off_last;
170 
171 	free(line);
172 
173 	return err;
174 }
175 
176 static int read_via_objdump(const char *filename, u64 addr, void *buf,
177 			    size_t len)
178 {
179 	char cmd[PATH_MAX * 2];
180 	const char *fmt;
181 	FILE *f;
182 	int ret;
183 
184 	fmt = "%s -z -d --start-address=0x%"PRIx64" --stop-address=0x%"PRIx64" %s";
185 	ret = snprintf(cmd, sizeof(cmd), fmt, "objdump", addr, addr + len,
186 		       filename);
187 	if (ret <= 0 || (size_t)ret >= sizeof(cmd))
188 		return -1;
189 
190 	pr_debug("Objdump command is: %s\n", cmd);
191 
192 	/* Ignore objdump errors */
193 	strcat(cmd, " 2>/dev/null");
194 
195 	f = popen(cmd, "r");
196 	if (!f) {
197 		pr_debug("popen failed\n");
198 		return -1;
199 	}
200 
201 	ret = read_objdump_output(f, buf, &len, addr);
202 	if (len) {
203 		pr_debug("objdump read too few bytes: %zd\n", len);
204 		if (!ret)
205 			ret = len;
206 	}
207 
208 	pclose(f);
209 
210 	return ret;
211 }
212 
213 static void dump_buf(unsigned char *buf, size_t len)
214 {
215 	size_t i;
216 
217 	for (i = 0; i < len; i++) {
218 		pr_debug("0x%02x ", buf[i]);
219 		if (i % 16 == 15)
220 			pr_debug("\n");
221 	}
222 	pr_debug("\n");
223 }
224 
225 static int read_object_code(u64 addr, size_t len, u8 cpumode,
226 			    struct thread *thread, struct state *state)
227 {
228 	struct addr_location al;
229 	unsigned char buf1[BUFSZ];
230 	unsigned char buf2[BUFSZ];
231 	size_t ret_len;
232 	u64 objdump_addr;
233 	const char *objdump_name;
234 	char decomp_name[KMOD_DECOMP_LEN];
235 	int ret;
236 
237 	pr_debug("Reading object code for memory address: %#"PRIx64"\n", addr);
238 
239 	thread__find_addr_map(thread, cpumode, MAP__FUNCTION, addr, &al);
240 	if (!al.map || !al.map->dso) {
241 		if (cpumode == PERF_RECORD_MISC_HYPERVISOR) {
242 			pr_debug("Hypervisor address can not be resolved - skipping\n");
243 			return 0;
244 		}
245 
246 		pr_debug("thread__find_addr_map failed\n");
247 		return -1;
248 	}
249 
250 	pr_debug("File is: %s\n", al.map->dso->long_name);
251 
252 	if (al.map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
253 	    !dso__is_kcore(al.map->dso)) {
254 		pr_debug("Unexpected kernel address - skipping\n");
255 		return 0;
256 	}
257 
258 	pr_debug("On file address is: %#"PRIx64"\n", al.addr);
259 
260 	if (len > BUFSZ)
261 		len = BUFSZ;
262 
263 	/* Do not go off the map */
264 	if (addr + len > al.map->end)
265 		len = al.map->end - addr;
266 
267 	/* Read the object code using perf */
268 	ret_len = dso__data_read_offset(al.map->dso, thread->mg->machine,
269 					al.addr, buf1, len);
270 	if (ret_len != len) {
271 		pr_debug("dso__data_read_offset failed\n");
272 		return -1;
273 	}
274 
275 	/*
276 	 * Converting addresses for use by objdump requires more information.
277 	 * map__load() does that.  See map__rip_2objdump() for details.
278 	 */
279 	if (map__load(al.map))
280 		return -1;
281 
282 	/* objdump struggles with kcore - try each map only once */
283 	if (dso__is_kcore(al.map->dso)) {
284 		size_t d;
285 
286 		for (d = 0; d < state->done_cnt; d++) {
287 			if (state->done[d] == al.map->start) {
288 				pr_debug("kcore map tested already");
289 				pr_debug(" - skipping\n");
290 				return 0;
291 			}
292 		}
293 		if (state->done_cnt >= ARRAY_SIZE(state->done)) {
294 			pr_debug("Too many kcore maps - skipping\n");
295 			return 0;
296 		}
297 		state->done[state->done_cnt++] = al.map->start;
298 	}
299 
300 	objdump_name = al.map->dso->long_name;
301 	if (dso__needs_decompress(al.map->dso)) {
302 		if (dso__decompress_kmodule_path(al.map->dso, objdump_name,
303 						 decomp_name,
304 						 sizeof(decomp_name)) < 0) {
305 			pr_debug("decompression failed\n");
306 			return -1;
307 		}
308 
309 		objdump_name = decomp_name;
310 	}
311 
312 	/* Read the object code using objdump */
313 	objdump_addr = map__rip_2objdump(al.map, al.addr);
314 	ret = read_via_objdump(objdump_name, objdump_addr, buf2, len);
315 
316 	if (dso__needs_decompress(al.map->dso))
317 		unlink(objdump_name);
318 
319 	if (ret > 0) {
320 		/*
321 		 * The kernel maps are inaccurate - assume objdump is right in
322 		 * that case.
323 		 */
324 		if (cpumode == PERF_RECORD_MISC_KERNEL ||
325 		    cpumode == PERF_RECORD_MISC_GUEST_KERNEL) {
326 			len -= ret;
327 			if (len) {
328 				pr_debug("Reducing len to %zu\n", len);
329 			} else if (dso__is_kcore(al.map->dso)) {
330 				/*
331 				 * objdump cannot handle very large segments
332 				 * that may be found in kcore.
333 				 */
334 				pr_debug("objdump failed for kcore");
335 				pr_debug(" - skipping\n");
336 				return 0;
337 			} else {
338 				return -1;
339 			}
340 		}
341 	}
342 	if (ret < 0) {
343 		pr_debug("read_via_objdump failed\n");
344 		return -1;
345 	}
346 
347 	/* The results should be identical */
348 	if (memcmp(buf1, buf2, len)) {
349 		pr_debug("Bytes read differ from those read by objdump\n");
350 		pr_debug("buf1 (dso):\n");
351 		dump_buf(buf1, len);
352 		pr_debug("buf2 (objdump):\n");
353 		dump_buf(buf2, len);
354 		return -1;
355 	}
356 	pr_debug("Bytes read match those read by objdump\n");
357 
358 	return 0;
359 }
360 
361 static int process_sample_event(struct machine *machine,
362 				struct perf_evlist *evlist,
363 				union perf_event *event, struct state *state)
364 {
365 	struct perf_sample sample;
366 	struct thread *thread;
367 	int ret;
368 
369 	if (perf_evlist__parse_sample(evlist, event, &sample)) {
370 		pr_debug("perf_evlist__parse_sample failed\n");
371 		return -1;
372 	}
373 
374 	thread = machine__findnew_thread(machine, sample.pid, sample.tid);
375 	if (!thread) {
376 		pr_debug("machine__findnew_thread failed\n");
377 		return -1;
378 	}
379 
380 	ret = read_object_code(sample.ip, READLEN, sample.cpumode, thread, state);
381 	thread__put(thread);
382 	return ret;
383 }
384 
385 static int process_event(struct machine *machine, struct perf_evlist *evlist,
386 			 union perf_event *event, struct state *state)
387 {
388 	if (event->header.type == PERF_RECORD_SAMPLE)
389 		return process_sample_event(machine, evlist, event, state);
390 
391 	if (event->header.type == PERF_RECORD_THROTTLE ||
392 	    event->header.type == PERF_RECORD_UNTHROTTLE)
393 		return 0;
394 
395 	if (event->header.type < PERF_RECORD_MAX) {
396 		int ret;
397 
398 		ret = machine__process_event(machine, event, NULL);
399 		if (ret < 0)
400 			pr_debug("machine__process_event failed, event type %u\n",
401 				 event->header.type);
402 		return ret;
403 	}
404 
405 	return 0;
406 }
407 
408 static int process_events(struct machine *machine, struct perf_evlist *evlist,
409 			  struct state *state)
410 {
411 	union perf_event *event;
412 	struct perf_mmap *md;
413 	int i, ret;
414 
415 	for (i = 0; i < evlist->nr_mmaps; i++) {
416 		md = &evlist->mmap[i];
417 		if (perf_mmap__read_init(md) < 0)
418 			continue;
419 
420 		while ((event = perf_mmap__read_event(md)) != NULL) {
421 			ret = process_event(machine, evlist, event, state);
422 			perf_mmap__consume(md);
423 			if (ret < 0)
424 				return ret;
425 		}
426 		perf_mmap__read_done(md);
427 	}
428 	return 0;
429 }
430 
431 static int comp(const void *a, const void *b)
432 {
433 	return *(int *)a - *(int *)b;
434 }
435 
436 static void do_sort_something(void)
437 {
438 	int buf[40960], i;
439 
440 	for (i = 0; i < (int)ARRAY_SIZE(buf); i++)
441 		buf[i] = ARRAY_SIZE(buf) - i - 1;
442 
443 	qsort(buf, ARRAY_SIZE(buf), sizeof(int), comp);
444 
445 	for (i = 0; i < (int)ARRAY_SIZE(buf); i++) {
446 		if (buf[i] != i) {
447 			pr_debug("qsort failed\n");
448 			break;
449 		}
450 	}
451 }
452 
453 static void sort_something(void)
454 {
455 	int i;
456 
457 	for (i = 0; i < 10; i++)
458 		do_sort_something();
459 }
460 
461 static void syscall_something(void)
462 {
463 	int pipefd[2];
464 	int i;
465 
466 	for (i = 0; i < 1000; i++) {
467 		if (pipe(pipefd) < 0) {
468 			pr_debug("pipe failed\n");
469 			break;
470 		}
471 		close(pipefd[1]);
472 		close(pipefd[0]);
473 	}
474 }
475 
476 static void fs_something(void)
477 {
478 	const char *test_file_name = "temp-perf-code-reading-test-file--";
479 	FILE *f;
480 	int i;
481 
482 	for (i = 0; i < 1000; i++) {
483 		f = fopen(test_file_name, "w+");
484 		if (f) {
485 			fclose(f);
486 			unlink(test_file_name);
487 		}
488 	}
489 }
490 
491 static const char *do_determine_event(bool excl_kernel)
492 {
493 	const char *event = excl_kernel ? "cycles:u" : "cycles";
494 
495 #ifdef __s390x__
496 	char cpuid[128], model[16], model_c[16], cpum_cf_v[16];
497 	unsigned int family;
498 	int ret, cpum_cf_a;
499 
500 	if (get_cpuid(cpuid, sizeof(cpuid)))
501 		goto out_clocks;
502 	ret = sscanf(cpuid, "%*[^,],%u,%[^,],%[^,],%[^,],%x", &family, model_c,
503 		     model, cpum_cf_v, &cpum_cf_a);
504 	if (ret != 5)		 /* Not available */
505 		goto out_clocks;
506 	if (excl_kernel && (cpum_cf_a & 4))
507 		return event;
508 	if (!excl_kernel && (cpum_cf_a & 2))
509 		return event;
510 
511 	/* Fall through: missing authorization */
512 out_clocks:
513 	event = excl_kernel ? "cpu-clock:u" : "cpu-clock";
514 
515 #endif
516 	return event;
517 }
518 
519 static void do_something(void)
520 {
521 	fs_something();
522 
523 	sort_something();
524 
525 	syscall_something();
526 }
527 
528 enum {
529 	TEST_CODE_READING_OK,
530 	TEST_CODE_READING_NO_VMLINUX,
531 	TEST_CODE_READING_NO_KCORE,
532 	TEST_CODE_READING_NO_ACCESS,
533 	TEST_CODE_READING_NO_KERNEL_OBJ,
534 };
535 
536 static int do_test_code_reading(bool try_kcore)
537 {
538 	struct machine *machine;
539 	struct thread *thread;
540 	struct record_opts opts = {
541 		.mmap_pages	     = UINT_MAX,
542 		.user_freq	     = UINT_MAX,
543 		.user_interval	     = ULLONG_MAX,
544 		.freq		     = 500,
545 		.target		     = {
546 			.uses_mmap   = true,
547 		},
548 	};
549 	struct state state = {
550 		.done_cnt = 0,
551 	};
552 	struct thread_map *threads = NULL;
553 	struct cpu_map *cpus = NULL;
554 	struct perf_evlist *evlist = NULL;
555 	struct perf_evsel *evsel = NULL;
556 	int err = -1, ret;
557 	pid_t pid;
558 	struct map *map;
559 	bool have_vmlinux, have_kcore, excl_kernel = false;
560 
561 	pid = getpid();
562 
563 	machine = machine__new_host();
564 
565 	ret = machine__create_kernel_maps(machine);
566 	if (ret < 0) {
567 		pr_debug("machine__create_kernel_maps failed\n");
568 		goto out_err;
569 	}
570 
571 	/* Force the use of kallsyms instead of vmlinux to try kcore */
572 	if (try_kcore)
573 		symbol_conf.kallsyms_name = "/proc/kallsyms";
574 
575 	/* Load kernel map */
576 	map = machine__kernel_map(machine);
577 	ret = map__load(map);
578 	if (ret < 0) {
579 		pr_debug("map__load failed\n");
580 		goto out_err;
581 	}
582 	have_vmlinux = dso__is_vmlinux(map->dso);
583 	have_kcore = dso__is_kcore(map->dso);
584 
585 	/* 2nd time through we just try kcore */
586 	if (try_kcore && !have_kcore)
587 		return TEST_CODE_READING_NO_KCORE;
588 
589 	/* No point getting kernel events if there is no kernel object */
590 	if (!have_vmlinux && !have_kcore)
591 		excl_kernel = true;
592 
593 	threads = thread_map__new_by_tid(pid);
594 	if (!threads) {
595 		pr_debug("thread_map__new_by_tid failed\n");
596 		goto out_err;
597 	}
598 
599 	ret = perf_event__synthesize_thread_map(NULL, threads,
600 						perf_event__process, machine, false, 500);
601 	if (ret < 0) {
602 		pr_debug("perf_event__synthesize_thread_map failed\n");
603 		goto out_err;
604 	}
605 
606 	thread = machine__findnew_thread(machine, pid, pid);
607 	if (!thread) {
608 		pr_debug("machine__findnew_thread failed\n");
609 		goto out_put;
610 	}
611 
612 	cpus = cpu_map__new(NULL);
613 	if (!cpus) {
614 		pr_debug("cpu_map__new failed\n");
615 		goto out_put;
616 	}
617 
618 	while (1) {
619 		const char *str;
620 
621 		evlist = perf_evlist__new();
622 		if (!evlist) {
623 			pr_debug("perf_evlist__new failed\n");
624 			goto out_put;
625 		}
626 
627 		perf_evlist__set_maps(evlist, cpus, threads);
628 
629 		str = do_determine_event(excl_kernel);
630 		pr_debug("Parsing event '%s'\n", str);
631 		ret = parse_events(evlist, str, NULL);
632 		if (ret < 0) {
633 			pr_debug("parse_events failed\n");
634 			goto out_put;
635 		}
636 
637 		perf_evlist__config(evlist, &opts, NULL);
638 
639 		evsel = perf_evlist__first(evlist);
640 
641 		evsel->attr.comm = 1;
642 		evsel->attr.disabled = 1;
643 		evsel->attr.enable_on_exec = 0;
644 
645 		ret = perf_evlist__open(evlist);
646 		if (ret < 0) {
647 			if (!excl_kernel) {
648 				excl_kernel = true;
649 				/*
650 				 * Both cpus and threads are now owned by evlist
651 				 * and will be freed by following perf_evlist__set_maps
652 				 * call. Getting refference to keep them alive.
653 				 */
654 				cpu_map__get(cpus);
655 				thread_map__get(threads);
656 				perf_evlist__set_maps(evlist, NULL, NULL);
657 				perf_evlist__delete(evlist);
658 				evlist = NULL;
659 				continue;
660 			}
661 
662 			if (verbose > 0) {
663 				char errbuf[512];
664 				perf_evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
665 				pr_debug("perf_evlist__open() failed!\n%s\n", errbuf);
666 			}
667 
668 			goto out_put;
669 		}
670 		break;
671 	}
672 
673 	ret = perf_evlist__mmap(evlist, UINT_MAX);
674 	if (ret < 0) {
675 		pr_debug("perf_evlist__mmap failed\n");
676 		goto out_put;
677 	}
678 
679 	perf_evlist__enable(evlist);
680 
681 	do_something();
682 
683 	perf_evlist__disable(evlist);
684 
685 	ret = process_events(machine, evlist, &state);
686 	if (ret < 0)
687 		goto out_put;
688 
689 	if (!have_vmlinux && !have_kcore && !try_kcore)
690 		err = TEST_CODE_READING_NO_KERNEL_OBJ;
691 	else if (!have_vmlinux && !try_kcore)
692 		err = TEST_CODE_READING_NO_VMLINUX;
693 	else if (excl_kernel)
694 		err = TEST_CODE_READING_NO_ACCESS;
695 	else
696 		err = TEST_CODE_READING_OK;
697 out_put:
698 	thread__put(thread);
699 out_err:
700 
701 	if (evlist) {
702 		perf_evlist__delete(evlist);
703 	} else {
704 		cpu_map__put(cpus);
705 		thread_map__put(threads);
706 	}
707 	machine__delete_threads(machine);
708 	machine__delete(machine);
709 
710 	return err;
711 }
712 
713 int test__code_reading(struct test *test __maybe_unused, int subtest __maybe_unused)
714 {
715 	int ret;
716 
717 	ret = do_test_code_reading(false);
718 	if (!ret)
719 		ret = do_test_code_reading(true);
720 
721 	switch (ret) {
722 	case TEST_CODE_READING_OK:
723 		return 0;
724 	case TEST_CODE_READING_NO_VMLINUX:
725 		pr_debug("no vmlinux\n");
726 		return 0;
727 	case TEST_CODE_READING_NO_KCORE:
728 		pr_debug("no kcore\n");
729 		return 0;
730 	case TEST_CODE_READING_NO_ACCESS:
731 		pr_debug("no access\n");
732 		return 0;
733 	case TEST_CODE_READING_NO_KERNEL_OBJ:
734 		pr_debug("no kernel obj\n");
735 		return 0;
736 	default:
737 		return -1;
738 	};
739 }
740