xref: /openbmc/linux/tools/perf/tests/code-reading.c (revision 52cdded0)
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 #include <perf/cpumap.h>
12 #include <perf/evlist.h>
13 #include <perf/mmap.h>
14 
15 #include "debug.h"
16 #include "dso.h"
17 #include "env.h"
18 #include "parse-events.h"
19 #include "trace-event.h"
20 #include "evlist.h"
21 #include "evsel.h"
22 #include "thread_map.h"
23 #include "machine.h"
24 #include "map.h"
25 #include "symbol.h"
26 #include "event.h"
27 #include "record.h"
28 #include "util/mmap.h"
29 #include "util/synthetic-events.h"
30 #include "thread.h"
31 
32 #include "tests.h"
33 
34 #include <linux/ctype.h>
35 
36 #define BUFSZ	1024
37 #define READLEN	128
38 
39 struct state {
40 	u64 done[1024];
41 	size_t done_cnt;
42 };
43 
44 static unsigned int hex(char c)
45 {
46 	if (c >= '0' && c <= '9')
47 		return c - '0';
48 	if (c >= 'a' && c <= 'f')
49 		return c - 'a' + 10;
50 	return c - 'A' + 10;
51 }
52 
53 static size_t read_objdump_chunk(const char **line, unsigned char **buf,
54 				 size_t *buf_len)
55 {
56 	size_t bytes_read = 0;
57 	unsigned char *chunk_start = *buf;
58 
59 	/* Read bytes */
60 	while (*buf_len > 0) {
61 		char c1, c2;
62 
63 		/* Get 2 hex digits */
64 		c1 = *(*line)++;
65 		if (!isxdigit(c1))
66 			break;
67 		c2 = *(*line)++;
68 		if (!isxdigit(c2))
69 			break;
70 
71 		/* Store byte and advance buf */
72 		**buf = (hex(c1) << 4) | hex(c2);
73 		(*buf)++;
74 		(*buf_len)--;
75 		bytes_read++;
76 
77 		/* End of chunk? */
78 		if (isspace(**line))
79 			break;
80 	}
81 
82 	/*
83 	 * objdump will display raw insn as LE if code endian
84 	 * is LE and bytes_per_chunk > 1. In that case reverse
85 	 * the chunk we just read.
86 	 *
87 	 * see disassemble_bytes() at binutils/objdump.c for details
88 	 * how objdump chooses display endian)
89 	 */
90 	if (bytes_read > 1 && !bigendian()) {
91 		unsigned char *chunk_end = chunk_start + bytes_read - 1;
92 		unsigned char tmp;
93 
94 		while (chunk_start < chunk_end) {
95 			tmp = *chunk_start;
96 			*chunk_start = *chunk_end;
97 			*chunk_end = tmp;
98 			chunk_start++;
99 			chunk_end--;
100 		}
101 	}
102 
103 	return bytes_read;
104 }
105 
106 static size_t read_objdump_line(const char *line, unsigned char *buf,
107 				size_t buf_len)
108 {
109 	const char *p;
110 	size_t ret, bytes_read = 0;
111 
112 	/* Skip to a colon */
113 	p = strchr(line, ':');
114 	if (!p)
115 		return 0;
116 	p++;
117 
118 	/* Skip initial spaces */
119 	while (*p) {
120 		if (!isspace(*p))
121 			break;
122 		p++;
123 	}
124 
125 	do {
126 		ret = read_objdump_chunk(&p, &buf, &buf_len);
127 		bytes_read += ret;
128 		p++;
129 	} while (ret > 0);
130 
131 	/* return number of successfully read bytes */
132 	return bytes_read;
133 }
134 
135 static int read_objdump_output(FILE *f, void *buf, size_t *len, u64 start_addr)
136 {
137 	char *line = NULL;
138 	size_t line_len, off_last = 0;
139 	ssize_t ret;
140 	int err = 0;
141 	u64 addr, last_addr = start_addr;
142 
143 	while (off_last < *len) {
144 		size_t off, read_bytes, written_bytes;
145 		unsigned char tmp[BUFSZ];
146 
147 		ret = getline(&line, &line_len, f);
148 		if (feof(f))
149 			break;
150 		if (ret < 0) {
151 			pr_debug("getline failed\n");
152 			err = -1;
153 			break;
154 		}
155 
156 		/* read objdump data into temporary buffer */
157 		read_bytes = read_objdump_line(line, tmp, sizeof(tmp));
158 		if (!read_bytes)
159 			continue;
160 
161 		if (sscanf(line, "%"PRIx64, &addr) != 1)
162 			continue;
163 		if (addr < last_addr) {
164 			pr_debug("addr going backwards, read beyond section?\n");
165 			break;
166 		}
167 		last_addr = addr;
168 
169 		/* copy it from temporary buffer to 'buf' according
170 		 * to address on current objdump line */
171 		off = addr - start_addr;
172 		if (off >= *len)
173 			break;
174 		written_bytes = MIN(read_bytes, *len - off);
175 		memcpy(buf + off, tmp, written_bytes);
176 		off_last = off + written_bytes;
177 	}
178 
179 	/* len returns number of bytes that could not be read */
180 	*len -= off_last;
181 
182 	free(line);
183 
184 	return err;
185 }
186 
187 static int read_via_objdump(const char *filename, u64 addr, void *buf,
188 			    size_t len)
189 {
190 	char cmd[PATH_MAX * 2];
191 	const char *fmt;
192 	FILE *f;
193 	int ret;
194 
195 	fmt = "%s -z -d --start-address=0x%"PRIx64" --stop-address=0x%"PRIx64" %s";
196 	ret = snprintf(cmd, sizeof(cmd), fmt, "objdump", addr, addr + len,
197 		       filename);
198 	if (ret <= 0 || (size_t)ret >= sizeof(cmd))
199 		return -1;
200 
201 	pr_debug("Objdump command is: %s\n", cmd);
202 
203 	/* Ignore objdump errors */
204 	strcat(cmd, " 2>/dev/null");
205 
206 	f = popen(cmd, "r");
207 	if (!f) {
208 		pr_debug("popen failed\n");
209 		return -1;
210 	}
211 
212 	ret = read_objdump_output(f, buf, &len, addr);
213 	if (len) {
214 		pr_debug("objdump read too few bytes: %zd\n", len);
215 		if (!ret)
216 			ret = len;
217 	}
218 
219 	pclose(f);
220 
221 	return ret;
222 }
223 
224 static void dump_buf(unsigned char *buf, size_t len)
225 {
226 	size_t i;
227 
228 	for (i = 0; i < len; i++) {
229 		pr_debug("0x%02x ", buf[i]);
230 		if (i % 16 == 15)
231 			pr_debug("\n");
232 	}
233 	pr_debug("\n");
234 }
235 
236 static int read_object_code(u64 addr, size_t len, u8 cpumode,
237 			    struct thread *thread, struct state *state)
238 {
239 	struct addr_location al;
240 	unsigned char buf1[BUFSZ];
241 	unsigned char buf2[BUFSZ];
242 	size_t ret_len;
243 	u64 objdump_addr;
244 	const char *objdump_name;
245 	char decomp_name[KMOD_DECOMP_LEN];
246 	bool decomp = false;
247 	int ret;
248 
249 	pr_debug("Reading object code for memory address: %#"PRIx64"\n", addr);
250 
251 	if (!thread__find_map(thread, cpumode, addr, &al) || !al.map->dso) {
252 		if (cpumode == PERF_RECORD_MISC_HYPERVISOR) {
253 			pr_debug("Hypervisor address can not be resolved - skipping\n");
254 			return 0;
255 		}
256 
257 		pr_debug("thread__find_map failed\n");
258 		return -1;
259 	}
260 
261 	pr_debug("File is: %s\n", al.map->dso->long_name);
262 
263 	if (al.map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
264 	    !dso__is_kcore(al.map->dso)) {
265 		pr_debug("Unexpected kernel address - skipping\n");
266 		return 0;
267 	}
268 
269 	pr_debug("On file address is: %#"PRIx64"\n", al.addr);
270 
271 	if (len > BUFSZ)
272 		len = BUFSZ;
273 
274 	/* Do not go off the map */
275 	if (addr + len > al.map->end)
276 		len = al.map->end - addr;
277 
278 	/* Read the object code using perf */
279 	ret_len = dso__data_read_offset(al.map->dso, thread->maps->machine,
280 					al.addr, buf1, len);
281 	if (ret_len != len) {
282 		pr_debug("dso__data_read_offset failed\n");
283 		return -1;
284 	}
285 
286 	/*
287 	 * Converting addresses for use by objdump requires more information.
288 	 * map__load() does that.  See map__rip_2objdump() for details.
289 	 */
290 	if (map__load(al.map))
291 		return -1;
292 
293 	/* objdump struggles with kcore - try each map only once */
294 	if (dso__is_kcore(al.map->dso)) {
295 		size_t d;
296 
297 		for (d = 0; d < state->done_cnt; d++) {
298 			if (state->done[d] == al.map->start) {
299 				pr_debug("kcore map tested already");
300 				pr_debug(" - skipping\n");
301 				return 0;
302 			}
303 		}
304 		if (state->done_cnt >= ARRAY_SIZE(state->done)) {
305 			pr_debug("Too many kcore maps - skipping\n");
306 			return 0;
307 		}
308 		state->done[state->done_cnt++] = al.map->start;
309 	}
310 
311 	objdump_name = al.map->dso->long_name;
312 	if (dso__needs_decompress(al.map->dso)) {
313 		if (dso__decompress_kmodule_path(al.map->dso, objdump_name,
314 						 decomp_name,
315 						 sizeof(decomp_name)) < 0) {
316 			pr_debug("decompression failed\n");
317 			return -1;
318 		}
319 
320 		decomp = true;
321 		objdump_name = decomp_name;
322 	}
323 
324 	/* Read the object code using objdump */
325 	objdump_addr = map__rip_2objdump(al.map, al.addr);
326 	ret = read_via_objdump(objdump_name, objdump_addr, buf2, len);
327 
328 	if (decomp)
329 		unlink(objdump_name);
330 
331 	if (ret > 0) {
332 		/*
333 		 * The kernel maps are inaccurate - assume objdump is right in
334 		 * that case.
335 		 */
336 		if (cpumode == PERF_RECORD_MISC_KERNEL ||
337 		    cpumode == PERF_RECORD_MISC_GUEST_KERNEL) {
338 			len -= ret;
339 			if (len) {
340 				pr_debug("Reducing len to %zu\n", len);
341 			} else if (dso__is_kcore(al.map->dso)) {
342 				/*
343 				 * objdump cannot handle very large segments
344 				 * that may be found in kcore.
345 				 */
346 				pr_debug("objdump failed for kcore");
347 				pr_debug(" - skipping\n");
348 				return 0;
349 			} else {
350 				return -1;
351 			}
352 		}
353 	}
354 	if (ret < 0) {
355 		pr_debug("read_via_objdump failed\n");
356 		return -1;
357 	}
358 
359 	/* The results should be identical */
360 	if (memcmp(buf1, buf2, len)) {
361 		pr_debug("Bytes read differ from those read by objdump\n");
362 		pr_debug("buf1 (dso):\n");
363 		dump_buf(buf1, len);
364 		pr_debug("buf2 (objdump):\n");
365 		dump_buf(buf2, len);
366 		return -1;
367 	}
368 	pr_debug("Bytes read match those read by objdump\n");
369 
370 	return 0;
371 }
372 
373 static int process_sample_event(struct machine *machine,
374 				struct evlist *evlist,
375 				union perf_event *event, struct state *state)
376 {
377 	struct perf_sample sample;
378 	struct thread *thread;
379 	int ret;
380 
381 	if (perf_evlist__parse_sample(evlist, event, &sample)) {
382 		pr_debug("perf_evlist__parse_sample failed\n");
383 		return -1;
384 	}
385 
386 	thread = machine__findnew_thread(machine, sample.pid, sample.tid);
387 	if (!thread) {
388 		pr_debug("machine__findnew_thread failed\n");
389 		return -1;
390 	}
391 
392 	ret = read_object_code(sample.ip, READLEN, sample.cpumode, thread, state);
393 	thread__put(thread);
394 	return ret;
395 }
396 
397 static int process_event(struct machine *machine, struct evlist *evlist,
398 			 union perf_event *event, struct state *state)
399 {
400 	if (event->header.type == PERF_RECORD_SAMPLE)
401 		return process_sample_event(machine, evlist, event, state);
402 
403 	if (event->header.type == PERF_RECORD_THROTTLE ||
404 	    event->header.type == PERF_RECORD_UNTHROTTLE)
405 		return 0;
406 
407 	if (event->header.type < PERF_RECORD_MAX) {
408 		int ret;
409 
410 		ret = machine__process_event(machine, event, NULL);
411 		if (ret < 0)
412 			pr_debug("machine__process_event failed, event type %u\n",
413 				 event->header.type);
414 		return ret;
415 	}
416 
417 	return 0;
418 }
419 
420 static int process_events(struct machine *machine, struct evlist *evlist,
421 			  struct state *state)
422 {
423 	union perf_event *event;
424 	struct mmap *md;
425 	int i, ret;
426 
427 	for (i = 0; i < evlist->core.nr_mmaps; i++) {
428 		md = &evlist->mmap[i];
429 		if (perf_mmap__read_init(&md->core) < 0)
430 			continue;
431 
432 		while ((event = perf_mmap__read_event(&md->core)) != NULL) {
433 			ret = process_event(machine, evlist, event, state);
434 			perf_mmap__consume(&md->core);
435 			if (ret < 0)
436 				return ret;
437 		}
438 		perf_mmap__read_done(&md->core);
439 	}
440 	return 0;
441 }
442 
443 static int comp(const void *a, const void *b)
444 {
445 	return *(int *)a - *(int *)b;
446 }
447 
448 static void do_sort_something(void)
449 {
450 	int buf[40960], i;
451 
452 	for (i = 0; i < (int)ARRAY_SIZE(buf); i++)
453 		buf[i] = ARRAY_SIZE(buf) - i - 1;
454 
455 	qsort(buf, ARRAY_SIZE(buf), sizeof(int), comp);
456 
457 	for (i = 0; i < (int)ARRAY_SIZE(buf); i++) {
458 		if (buf[i] != i) {
459 			pr_debug("qsort failed\n");
460 			break;
461 		}
462 	}
463 }
464 
465 static void sort_something(void)
466 {
467 	int i;
468 
469 	for (i = 0; i < 10; i++)
470 		do_sort_something();
471 }
472 
473 static void syscall_something(void)
474 {
475 	int pipefd[2];
476 	int i;
477 
478 	for (i = 0; i < 1000; i++) {
479 		if (pipe(pipefd) < 0) {
480 			pr_debug("pipe failed\n");
481 			break;
482 		}
483 		close(pipefd[1]);
484 		close(pipefd[0]);
485 	}
486 }
487 
488 static void fs_something(void)
489 {
490 	const char *test_file_name = "temp-perf-code-reading-test-file--";
491 	FILE *f;
492 	int i;
493 
494 	for (i = 0; i < 1000; i++) {
495 		f = fopen(test_file_name, "w+");
496 		if (f) {
497 			fclose(f);
498 			unlink(test_file_name);
499 		}
500 	}
501 }
502 
503 #ifdef __s390x__
504 #include "header.h" // for get_cpuid()
505 #endif
506 
507 static const char *do_determine_event(bool excl_kernel)
508 {
509 	const char *event = excl_kernel ? "cycles:u" : "cycles";
510 
511 #ifdef __s390x__
512 	char cpuid[128], model[16], model_c[16], cpum_cf_v[16];
513 	unsigned int family;
514 	int ret, cpum_cf_a;
515 
516 	if (get_cpuid(cpuid, sizeof(cpuid)))
517 		goto out_clocks;
518 	ret = sscanf(cpuid, "%*[^,],%u,%[^,],%[^,],%[^,],%x", &family, model_c,
519 		     model, cpum_cf_v, &cpum_cf_a);
520 	if (ret != 5)		 /* Not available */
521 		goto out_clocks;
522 	if (excl_kernel && (cpum_cf_a & 4))
523 		return event;
524 	if (!excl_kernel && (cpum_cf_a & 2))
525 		return event;
526 
527 	/* Fall through: missing authorization */
528 out_clocks:
529 	event = excl_kernel ? "cpu-clock:u" : "cpu-clock";
530 
531 #endif
532 	return event;
533 }
534 
535 static void do_something(void)
536 {
537 	fs_something();
538 
539 	sort_something();
540 
541 	syscall_something();
542 }
543 
544 enum {
545 	TEST_CODE_READING_OK,
546 	TEST_CODE_READING_NO_VMLINUX,
547 	TEST_CODE_READING_NO_KCORE,
548 	TEST_CODE_READING_NO_ACCESS,
549 	TEST_CODE_READING_NO_KERNEL_OBJ,
550 };
551 
552 static int do_test_code_reading(bool try_kcore)
553 {
554 	struct machine *machine;
555 	struct thread *thread;
556 	struct record_opts opts = {
557 		.mmap_pages	     = UINT_MAX,
558 		.user_freq	     = UINT_MAX,
559 		.user_interval	     = ULLONG_MAX,
560 		.freq		     = 500,
561 		.target		     = {
562 			.uses_mmap   = true,
563 		},
564 	};
565 	struct state state = {
566 		.done_cnt = 0,
567 	};
568 	struct perf_thread_map *threads = NULL;
569 	struct perf_cpu_map *cpus = NULL;
570 	struct evlist *evlist = NULL;
571 	struct evsel *evsel = NULL;
572 	int err = -1, ret;
573 	pid_t pid;
574 	struct map *map;
575 	bool have_vmlinux, have_kcore, excl_kernel = false;
576 
577 	pid = getpid();
578 
579 	machine = machine__new_host();
580 	machine->env = &perf_env;
581 
582 	ret = machine__create_kernel_maps(machine);
583 	if (ret < 0) {
584 		pr_debug("machine__create_kernel_maps failed\n");
585 		goto out_err;
586 	}
587 
588 	/* Force the use of kallsyms instead of vmlinux to try kcore */
589 	if (try_kcore)
590 		symbol_conf.kallsyms_name = "/proc/kallsyms";
591 
592 	/* Load kernel map */
593 	map = machine__kernel_map(machine);
594 	ret = map__load(map);
595 	if (ret < 0) {
596 		pr_debug("map__load failed\n");
597 		goto out_err;
598 	}
599 	have_vmlinux = dso__is_vmlinux(map->dso);
600 	have_kcore = dso__is_kcore(map->dso);
601 
602 	/* 2nd time through we just try kcore */
603 	if (try_kcore && !have_kcore)
604 		return TEST_CODE_READING_NO_KCORE;
605 
606 	/* No point getting kernel events if there is no kernel object */
607 	if (!have_vmlinux && !have_kcore)
608 		excl_kernel = true;
609 
610 	threads = thread_map__new_by_tid(pid);
611 	if (!threads) {
612 		pr_debug("thread_map__new_by_tid failed\n");
613 		goto out_err;
614 	}
615 
616 	ret = perf_event__synthesize_thread_map(NULL, threads,
617 						perf_event__process, machine, false);
618 	if (ret < 0) {
619 		pr_debug("perf_event__synthesize_thread_map failed\n");
620 		goto out_err;
621 	}
622 
623 	thread = machine__findnew_thread(machine, pid, pid);
624 	if (!thread) {
625 		pr_debug("machine__findnew_thread failed\n");
626 		goto out_put;
627 	}
628 
629 	cpus = perf_cpu_map__new(NULL);
630 	if (!cpus) {
631 		pr_debug("perf_cpu_map__new failed\n");
632 		goto out_put;
633 	}
634 
635 	while (1) {
636 		const char *str;
637 
638 		evlist = evlist__new();
639 		if (!evlist) {
640 			pr_debug("perf_evlist__new failed\n");
641 			goto out_put;
642 		}
643 
644 		perf_evlist__set_maps(&evlist->core, cpus, threads);
645 
646 		str = do_determine_event(excl_kernel);
647 		pr_debug("Parsing event '%s'\n", str);
648 		ret = parse_events(evlist, str, NULL);
649 		if (ret < 0) {
650 			pr_debug("parse_events failed\n");
651 			goto out_put;
652 		}
653 
654 		perf_evlist__config(evlist, &opts, NULL);
655 
656 		evsel = evlist__first(evlist);
657 
658 		evsel->core.attr.comm = 1;
659 		evsel->core.attr.disabled = 1;
660 		evsel->core.attr.enable_on_exec = 0;
661 
662 		ret = evlist__open(evlist);
663 		if (ret < 0) {
664 			if (!excl_kernel) {
665 				excl_kernel = true;
666 				/*
667 				 * Both cpus and threads are now owned by evlist
668 				 * and will be freed by following perf_evlist__set_maps
669 				 * call. Getting refference to keep them alive.
670 				 */
671 				perf_cpu_map__get(cpus);
672 				perf_thread_map__get(threads);
673 				perf_evlist__set_maps(&evlist->core, NULL, NULL);
674 				evlist__delete(evlist);
675 				evlist = NULL;
676 				continue;
677 			}
678 
679 			if (verbose > 0) {
680 				char errbuf[512];
681 				evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
682 				pr_debug("perf_evlist__open() failed!\n%s\n", errbuf);
683 			}
684 
685 			goto out_put;
686 		}
687 		break;
688 	}
689 
690 	ret = evlist__mmap(evlist, UINT_MAX);
691 	if (ret < 0) {
692 		pr_debug("evlist__mmap failed\n");
693 		goto out_put;
694 	}
695 
696 	evlist__enable(evlist);
697 
698 	do_something();
699 
700 	evlist__disable(evlist);
701 
702 	ret = process_events(machine, evlist, &state);
703 	if (ret < 0)
704 		goto out_put;
705 
706 	if (!have_vmlinux && !have_kcore && !try_kcore)
707 		err = TEST_CODE_READING_NO_KERNEL_OBJ;
708 	else if (!have_vmlinux && !try_kcore)
709 		err = TEST_CODE_READING_NO_VMLINUX;
710 	else if (excl_kernel)
711 		err = TEST_CODE_READING_NO_ACCESS;
712 	else
713 		err = TEST_CODE_READING_OK;
714 out_put:
715 	thread__put(thread);
716 out_err:
717 
718 	if (evlist) {
719 		evlist__delete(evlist);
720 	} else {
721 		perf_cpu_map__put(cpus);
722 		perf_thread_map__put(threads);
723 	}
724 	machine__delete_threads(machine);
725 	machine__delete(machine);
726 
727 	return err;
728 }
729 
730 int test__code_reading(struct test *test __maybe_unused, int subtest __maybe_unused)
731 {
732 	int ret;
733 
734 	ret = do_test_code_reading(false);
735 	if (!ret)
736 		ret = do_test_code_reading(true);
737 
738 	switch (ret) {
739 	case TEST_CODE_READING_OK:
740 		return 0;
741 	case TEST_CODE_READING_NO_VMLINUX:
742 		pr_debug("no vmlinux\n");
743 		return 0;
744 	case TEST_CODE_READING_NO_KCORE:
745 		pr_debug("no kcore\n");
746 		return 0;
747 	case TEST_CODE_READING_NO_ACCESS:
748 		pr_debug("no access\n");
749 		return 0;
750 	case TEST_CODE_READING_NO_KERNEL_OBJ:
751 		pr_debug("no kernel obj\n");
752 		return 0;
753 	default:
754 		return -1;
755 	};
756 }
757