xref: /openbmc/linux/tools/perf/tests/code-reading.c (revision 9e3bd0f6)
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 
14 #include "debug.h"
15 #include "dso.h"
16 #include "env.h"
17 #include "parse-events.h"
18 #include "trace-event.h"
19 #include "evlist.h"
20 #include "evsel.h"
21 #include "thread_map.h"
22 #include "machine.h"
23 #include "map.h"
24 #include "symbol.h"
25 #include "event.h"
26 #include "record.h"
27 #include "util/mmap.h"
28 #include "util/synthetic-events.h"
29 #include "thread.h"
30 
31 #include "tests.h"
32 
33 #include <linux/ctype.h>
34 
35 #define BUFSZ	1024
36 #define READLEN	128
37 
38 struct state {
39 	u64 done[1024];
40 	size_t done_cnt;
41 };
42 
43 static unsigned int hex(char c)
44 {
45 	if (c >= '0' && c <= '9')
46 		return c - '0';
47 	if (c >= 'a' && c <= 'f')
48 		return c - 'a' + 10;
49 	return c - 'A' + 10;
50 }
51 
52 static size_t read_objdump_chunk(const char **line, unsigned char **buf,
53 				 size_t *buf_len)
54 {
55 	size_t bytes_read = 0;
56 	unsigned char *chunk_start = *buf;
57 
58 	/* Read bytes */
59 	while (*buf_len > 0) {
60 		char c1, c2;
61 
62 		/* Get 2 hex digits */
63 		c1 = *(*line)++;
64 		if (!isxdigit(c1))
65 			break;
66 		c2 = *(*line)++;
67 		if (!isxdigit(c2))
68 			break;
69 
70 		/* Store byte and advance buf */
71 		**buf = (hex(c1) << 4) | hex(c2);
72 		(*buf)++;
73 		(*buf_len)--;
74 		bytes_read++;
75 
76 		/* End of chunk? */
77 		if (isspace(**line))
78 			break;
79 	}
80 
81 	/*
82 	 * objdump will display raw insn as LE if code endian
83 	 * is LE and bytes_per_chunk > 1. In that case reverse
84 	 * the chunk we just read.
85 	 *
86 	 * see disassemble_bytes() at binutils/objdump.c for details
87 	 * how objdump chooses display endian)
88 	 */
89 	if (bytes_read > 1 && !bigendian()) {
90 		unsigned char *chunk_end = chunk_start + bytes_read - 1;
91 		unsigned char tmp;
92 
93 		while (chunk_start < chunk_end) {
94 			tmp = *chunk_start;
95 			*chunk_start = *chunk_end;
96 			*chunk_end = tmp;
97 			chunk_start++;
98 			chunk_end--;
99 		}
100 	}
101 
102 	return bytes_read;
103 }
104 
105 static size_t read_objdump_line(const char *line, unsigned char *buf,
106 				size_t buf_len)
107 {
108 	const char *p;
109 	size_t ret, bytes_read = 0;
110 
111 	/* Skip to a colon */
112 	p = strchr(line, ':');
113 	if (!p)
114 		return 0;
115 	p++;
116 
117 	/* Skip initial spaces */
118 	while (*p) {
119 		if (!isspace(*p))
120 			break;
121 		p++;
122 	}
123 
124 	do {
125 		ret = read_objdump_chunk(&p, &buf, &buf_len);
126 		bytes_read += ret;
127 		p++;
128 	} while (ret > 0);
129 
130 	/* return number of successfully read bytes */
131 	return bytes_read;
132 }
133 
134 static int read_objdump_output(FILE *f, void *buf, size_t *len, u64 start_addr)
135 {
136 	char *line = NULL;
137 	size_t line_len, off_last = 0;
138 	ssize_t ret;
139 	int err = 0;
140 	u64 addr, last_addr = start_addr;
141 
142 	while (off_last < *len) {
143 		size_t off, read_bytes, written_bytes;
144 		unsigned char tmp[BUFSZ];
145 
146 		ret = getline(&line, &line_len, f);
147 		if (feof(f))
148 			break;
149 		if (ret < 0) {
150 			pr_debug("getline failed\n");
151 			err = -1;
152 			break;
153 		}
154 
155 		/* read objdump data into temporary buffer */
156 		read_bytes = read_objdump_line(line, tmp, sizeof(tmp));
157 		if (!read_bytes)
158 			continue;
159 
160 		if (sscanf(line, "%"PRIx64, &addr) != 1)
161 			continue;
162 		if (addr < last_addr) {
163 			pr_debug("addr going backwards, read beyond section?\n");
164 			break;
165 		}
166 		last_addr = addr;
167 
168 		/* copy it from temporary buffer to 'buf' according
169 		 * to address on current objdump line */
170 		off = addr - start_addr;
171 		if (off >= *len)
172 			break;
173 		written_bytes = MIN(read_bytes, *len - off);
174 		memcpy(buf + off, tmp, written_bytes);
175 		off_last = off + written_bytes;
176 	}
177 
178 	/* len returns number of bytes that could not be read */
179 	*len -= off_last;
180 
181 	free(line);
182 
183 	return err;
184 }
185 
186 static int read_via_objdump(const char *filename, u64 addr, void *buf,
187 			    size_t len)
188 {
189 	char cmd[PATH_MAX * 2];
190 	const char *fmt;
191 	FILE *f;
192 	int ret;
193 
194 	fmt = "%s -z -d --start-address=0x%"PRIx64" --stop-address=0x%"PRIx64" %s";
195 	ret = snprintf(cmd, sizeof(cmd), fmt, "objdump", addr, addr + len,
196 		       filename);
197 	if (ret <= 0 || (size_t)ret >= sizeof(cmd))
198 		return -1;
199 
200 	pr_debug("Objdump command is: %s\n", cmd);
201 
202 	/* Ignore objdump errors */
203 	strcat(cmd, " 2>/dev/null");
204 
205 	f = popen(cmd, "r");
206 	if (!f) {
207 		pr_debug("popen failed\n");
208 		return -1;
209 	}
210 
211 	ret = read_objdump_output(f, buf, &len, addr);
212 	if (len) {
213 		pr_debug("objdump read too few bytes: %zd\n", len);
214 		if (!ret)
215 			ret = len;
216 	}
217 
218 	pclose(f);
219 
220 	return ret;
221 }
222 
223 static void dump_buf(unsigned char *buf, size_t len)
224 {
225 	size_t i;
226 
227 	for (i = 0; i < len; i++) {
228 		pr_debug("0x%02x ", buf[i]);
229 		if (i % 16 == 15)
230 			pr_debug("\n");
231 	}
232 	pr_debug("\n");
233 }
234 
235 static int read_object_code(u64 addr, size_t len, u8 cpumode,
236 			    struct thread *thread, struct state *state)
237 {
238 	struct addr_location al;
239 	unsigned char buf1[BUFSZ];
240 	unsigned char buf2[BUFSZ];
241 	size_t ret_len;
242 	u64 objdump_addr;
243 	const char *objdump_name;
244 	char decomp_name[KMOD_DECOMP_LEN];
245 	bool decomp = false;
246 	int ret;
247 
248 	pr_debug("Reading object code for memory address: %#"PRIx64"\n", addr);
249 
250 	if (!thread__find_map(thread, cpumode, addr, &al) || !al.map->dso) {
251 		if (cpumode == PERF_RECORD_MISC_HYPERVISOR) {
252 			pr_debug("Hypervisor address can not be resolved - skipping\n");
253 			return 0;
254 		}
255 
256 		pr_debug("thread__find_map failed\n");
257 		return -1;
258 	}
259 
260 	pr_debug("File is: %s\n", al.map->dso->long_name);
261 
262 	if (al.map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
263 	    !dso__is_kcore(al.map->dso)) {
264 		pr_debug("Unexpected kernel address - skipping\n");
265 		return 0;
266 	}
267 
268 	pr_debug("On file address is: %#"PRIx64"\n", al.addr);
269 
270 	if (len > BUFSZ)
271 		len = BUFSZ;
272 
273 	/* Do not go off the map */
274 	if (addr + len > al.map->end)
275 		len = al.map->end - addr;
276 
277 	/* Read the object code using perf */
278 	ret_len = dso__data_read_offset(al.map->dso, thread->mg->machine,
279 					al.addr, buf1, len);
280 	if (ret_len != len) {
281 		pr_debug("dso__data_read_offset failed\n");
282 		return -1;
283 	}
284 
285 	/*
286 	 * Converting addresses for use by objdump requires more information.
287 	 * map__load() does that.  See map__rip_2objdump() for details.
288 	 */
289 	if (map__load(al.map))
290 		return -1;
291 
292 	/* objdump struggles with kcore - try each map only once */
293 	if (dso__is_kcore(al.map->dso)) {
294 		size_t d;
295 
296 		for (d = 0; d < state->done_cnt; d++) {
297 			if (state->done[d] == al.map->start) {
298 				pr_debug("kcore map tested already");
299 				pr_debug(" - skipping\n");
300 				return 0;
301 			}
302 		}
303 		if (state->done_cnt >= ARRAY_SIZE(state->done)) {
304 			pr_debug("Too many kcore maps - skipping\n");
305 			return 0;
306 		}
307 		state->done[state->done_cnt++] = al.map->start;
308 	}
309 
310 	objdump_name = al.map->dso->long_name;
311 	if (dso__needs_decompress(al.map->dso)) {
312 		if (dso__decompress_kmodule_path(al.map->dso, objdump_name,
313 						 decomp_name,
314 						 sizeof(decomp_name)) < 0) {
315 			pr_debug("decompression failed\n");
316 			return -1;
317 		}
318 
319 		decomp = true;
320 		objdump_name = decomp_name;
321 	}
322 
323 	/* Read the object code using objdump */
324 	objdump_addr = map__rip_2objdump(al.map, al.addr);
325 	ret = read_via_objdump(objdump_name, objdump_addr, buf2, len);
326 
327 	if (decomp)
328 		unlink(objdump_name);
329 
330 	if (ret > 0) {
331 		/*
332 		 * The kernel maps are inaccurate - assume objdump is right in
333 		 * that case.
334 		 */
335 		if (cpumode == PERF_RECORD_MISC_KERNEL ||
336 		    cpumode == PERF_RECORD_MISC_GUEST_KERNEL) {
337 			len -= ret;
338 			if (len) {
339 				pr_debug("Reducing len to %zu\n", len);
340 			} else if (dso__is_kcore(al.map->dso)) {
341 				/*
342 				 * objdump cannot handle very large segments
343 				 * that may be found in kcore.
344 				 */
345 				pr_debug("objdump failed for kcore");
346 				pr_debug(" - skipping\n");
347 				return 0;
348 			} else {
349 				return -1;
350 			}
351 		}
352 	}
353 	if (ret < 0) {
354 		pr_debug("read_via_objdump failed\n");
355 		return -1;
356 	}
357 
358 	/* The results should be identical */
359 	if (memcmp(buf1, buf2, len)) {
360 		pr_debug("Bytes read differ from those read by objdump\n");
361 		pr_debug("buf1 (dso):\n");
362 		dump_buf(buf1, len);
363 		pr_debug("buf2 (objdump):\n");
364 		dump_buf(buf2, len);
365 		return -1;
366 	}
367 	pr_debug("Bytes read match those read by objdump\n");
368 
369 	return 0;
370 }
371 
372 static int process_sample_event(struct machine *machine,
373 				struct evlist *evlist,
374 				union perf_event *event, struct state *state)
375 {
376 	struct perf_sample sample;
377 	struct thread *thread;
378 	int ret;
379 
380 	if (perf_evlist__parse_sample(evlist, event, &sample)) {
381 		pr_debug("perf_evlist__parse_sample failed\n");
382 		return -1;
383 	}
384 
385 	thread = machine__findnew_thread(machine, sample.pid, sample.tid);
386 	if (!thread) {
387 		pr_debug("machine__findnew_thread failed\n");
388 		return -1;
389 	}
390 
391 	ret = read_object_code(sample.ip, READLEN, sample.cpumode, thread, state);
392 	thread__put(thread);
393 	return ret;
394 }
395 
396 static int process_event(struct machine *machine, struct evlist *evlist,
397 			 union perf_event *event, struct state *state)
398 {
399 	if (event->header.type == PERF_RECORD_SAMPLE)
400 		return process_sample_event(machine, evlist, event, state);
401 
402 	if (event->header.type == PERF_RECORD_THROTTLE ||
403 	    event->header.type == PERF_RECORD_UNTHROTTLE)
404 		return 0;
405 
406 	if (event->header.type < PERF_RECORD_MAX) {
407 		int ret;
408 
409 		ret = machine__process_event(machine, event, NULL);
410 		if (ret < 0)
411 			pr_debug("machine__process_event failed, event type %u\n",
412 				 event->header.type);
413 		return ret;
414 	}
415 
416 	return 0;
417 }
418 
419 static int process_events(struct machine *machine, struct evlist *evlist,
420 			  struct state *state)
421 {
422 	union perf_event *event;
423 	struct mmap *md;
424 	int i, ret;
425 
426 	for (i = 0; i < evlist->core.nr_mmaps; i++) {
427 		md = &evlist->mmap[i];
428 		if (perf_mmap__read_init(md) < 0)
429 			continue;
430 
431 		while ((event = perf_mmap__read_event(md)) != NULL) {
432 			ret = process_event(machine, evlist, event, state);
433 			perf_mmap__consume(md);
434 			if (ret < 0)
435 				return ret;
436 		}
437 		perf_mmap__read_done(md);
438 	}
439 	return 0;
440 }
441 
442 static int comp(const void *a, const void *b)
443 {
444 	return *(int *)a - *(int *)b;
445 }
446 
447 static void do_sort_something(void)
448 {
449 	int buf[40960], i;
450 
451 	for (i = 0; i < (int)ARRAY_SIZE(buf); i++)
452 		buf[i] = ARRAY_SIZE(buf) - i - 1;
453 
454 	qsort(buf, ARRAY_SIZE(buf), sizeof(int), comp);
455 
456 	for (i = 0; i < (int)ARRAY_SIZE(buf); i++) {
457 		if (buf[i] != i) {
458 			pr_debug("qsort failed\n");
459 			break;
460 		}
461 	}
462 }
463 
464 static void sort_something(void)
465 {
466 	int i;
467 
468 	for (i = 0; i < 10; i++)
469 		do_sort_something();
470 }
471 
472 static void syscall_something(void)
473 {
474 	int pipefd[2];
475 	int i;
476 
477 	for (i = 0; i < 1000; i++) {
478 		if (pipe(pipefd) < 0) {
479 			pr_debug("pipe failed\n");
480 			break;
481 		}
482 		close(pipefd[1]);
483 		close(pipefd[0]);
484 	}
485 }
486 
487 static void fs_something(void)
488 {
489 	const char *test_file_name = "temp-perf-code-reading-test-file--";
490 	FILE *f;
491 	int i;
492 
493 	for (i = 0; i < 1000; i++) {
494 		f = fopen(test_file_name, "w+");
495 		if (f) {
496 			fclose(f);
497 			unlink(test_file_name);
498 		}
499 	}
500 }
501 
502 #ifdef __s390x__
503 #include "header.h" // for get_cpuid()
504 #endif
505 
506 static const char *do_determine_event(bool excl_kernel)
507 {
508 	const char *event = excl_kernel ? "cycles:u" : "cycles";
509 
510 #ifdef __s390x__
511 	char cpuid[128], model[16], model_c[16], cpum_cf_v[16];
512 	unsigned int family;
513 	int ret, cpum_cf_a;
514 
515 	if (get_cpuid(cpuid, sizeof(cpuid)))
516 		goto out_clocks;
517 	ret = sscanf(cpuid, "%*[^,],%u,%[^,],%[^,],%[^,],%x", &family, model_c,
518 		     model, cpum_cf_v, &cpum_cf_a);
519 	if (ret != 5)		 /* Not available */
520 		goto out_clocks;
521 	if (excl_kernel && (cpum_cf_a & 4))
522 		return event;
523 	if (!excl_kernel && (cpum_cf_a & 2))
524 		return event;
525 
526 	/* Fall through: missing authorization */
527 out_clocks:
528 	event = excl_kernel ? "cpu-clock:u" : "cpu-clock";
529 
530 #endif
531 	return event;
532 }
533 
534 static void do_something(void)
535 {
536 	fs_something();
537 
538 	sort_something();
539 
540 	syscall_something();
541 }
542 
543 enum {
544 	TEST_CODE_READING_OK,
545 	TEST_CODE_READING_NO_VMLINUX,
546 	TEST_CODE_READING_NO_KCORE,
547 	TEST_CODE_READING_NO_ACCESS,
548 	TEST_CODE_READING_NO_KERNEL_OBJ,
549 };
550 
551 static int do_test_code_reading(bool try_kcore)
552 {
553 	struct machine *machine;
554 	struct thread *thread;
555 	struct record_opts opts = {
556 		.mmap_pages	     = UINT_MAX,
557 		.user_freq	     = UINT_MAX,
558 		.user_interval	     = ULLONG_MAX,
559 		.freq		     = 500,
560 		.target		     = {
561 			.uses_mmap   = true,
562 		},
563 	};
564 	struct state state = {
565 		.done_cnt = 0,
566 	};
567 	struct perf_thread_map *threads = NULL;
568 	struct perf_cpu_map *cpus = NULL;
569 	struct evlist *evlist = NULL;
570 	struct evsel *evsel = NULL;
571 	int err = -1, ret;
572 	pid_t pid;
573 	struct map *map;
574 	bool have_vmlinux, have_kcore, excl_kernel = false;
575 
576 	pid = getpid();
577 
578 	machine = machine__new_host();
579 	machine->env = &perf_env;
580 
581 	ret = machine__create_kernel_maps(machine);
582 	if (ret < 0) {
583 		pr_debug("machine__create_kernel_maps failed\n");
584 		goto out_err;
585 	}
586 
587 	/* Force the use of kallsyms instead of vmlinux to try kcore */
588 	if (try_kcore)
589 		symbol_conf.kallsyms_name = "/proc/kallsyms";
590 
591 	/* Load kernel map */
592 	map = machine__kernel_map(machine);
593 	ret = map__load(map);
594 	if (ret < 0) {
595 		pr_debug("map__load failed\n");
596 		goto out_err;
597 	}
598 	have_vmlinux = dso__is_vmlinux(map->dso);
599 	have_kcore = dso__is_kcore(map->dso);
600 
601 	/* 2nd time through we just try kcore */
602 	if (try_kcore && !have_kcore)
603 		return TEST_CODE_READING_NO_KCORE;
604 
605 	/* No point getting kernel events if there is no kernel object */
606 	if (!have_vmlinux && !have_kcore)
607 		excl_kernel = true;
608 
609 	threads = thread_map__new_by_tid(pid);
610 	if (!threads) {
611 		pr_debug("thread_map__new_by_tid failed\n");
612 		goto out_err;
613 	}
614 
615 	ret = perf_event__synthesize_thread_map(NULL, threads,
616 						perf_event__process, machine, false);
617 	if (ret < 0) {
618 		pr_debug("perf_event__synthesize_thread_map failed\n");
619 		goto out_err;
620 	}
621 
622 	thread = machine__findnew_thread(machine, pid, pid);
623 	if (!thread) {
624 		pr_debug("machine__findnew_thread failed\n");
625 		goto out_put;
626 	}
627 
628 	cpus = perf_cpu_map__new(NULL);
629 	if (!cpus) {
630 		pr_debug("perf_cpu_map__new failed\n");
631 		goto out_put;
632 	}
633 
634 	while (1) {
635 		const char *str;
636 
637 		evlist = evlist__new();
638 		if (!evlist) {
639 			pr_debug("perf_evlist__new failed\n");
640 			goto out_put;
641 		}
642 
643 		perf_evlist__set_maps(&evlist->core, cpus, threads);
644 
645 		str = do_determine_event(excl_kernel);
646 		pr_debug("Parsing event '%s'\n", str);
647 		ret = parse_events(evlist, str, NULL);
648 		if (ret < 0) {
649 			pr_debug("parse_events failed\n");
650 			goto out_put;
651 		}
652 
653 		perf_evlist__config(evlist, &opts, NULL);
654 
655 		evsel = evlist__first(evlist);
656 
657 		evsel->core.attr.comm = 1;
658 		evsel->core.attr.disabled = 1;
659 		evsel->core.attr.enable_on_exec = 0;
660 
661 		ret = evlist__open(evlist);
662 		if (ret < 0) {
663 			if (!excl_kernel) {
664 				excl_kernel = true;
665 				/*
666 				 * Both cpus and threads are now owned by evlist
667 				 * and will be freed by following perf_evlist__set_maps
668 				 * call. Getting refference to keep them alive.
669 				 */
670 				perf_cpu_map__get(cpus);
671 				perf_thread_map__get(threads);
672 				perf_evlist__set_maps(&evlist->core, NULL, NULL);
673 				evlist__delete(evlist);
674 				evlist = NULL;
675 				continue;
676 			}
677 
678 			if (verbose > 0) {
679 				char errbuf[512];
680 				perf_evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
681 				pr_debug("perf_evlist__open() failed!\n%s\n", errbuf);
682 			}
683 
684 			goto out_put;
685 		}
686 		break;
687 	}
688 
689 	ret = evlist__mmap(evlist, UINT_MAX);
690 	if (ret < 0) {
691 		pr_debug("evlist__mmap failed\n");
692 		goto out_put;
693 	}
694 
695 	evlist__enable(evlist);
696 
697 	do_something();
698 
699 	evlist__disable(evlist);
700 
701 	ret = process_events(machine, evlist, &state);
702 	if (ret < 0)
703 		goto out_put;
704 
705 	if (!have_vmlinux && !have_kcore && !try_kcore)
706 		err = TEST_CODE_READING_NO_KERNEL_OBJ;
707 	else if (!have_vmlinux && !try_kcore)
708 		err = TEST_CODE_READING_NO_VMLINUX;
709 	else if (excl_kernel)
710 		err = TEST_CODE_READING_NO_ACCESS;
711 	else
712 		err = TEST_CODE_READING_OK;
713 out_put:
714 	thread__put(thread);
715 out_err:
716 
717 	if (evlist) {
718 		evlist__delete(evlist);
719 	} else {
720 		perf_cpu_map__put(cpus);
721 		perf_thread_map__put(threads);
722 	}
723 	machine__delete_threads(machine);
724 	machine__delete(machine);
725 
726 	return err;
727 }
728 
729 int test__code_reading(struct test *test __maybe_unused, int subtest __maybe_unused)
730 {
731 	int ret;
732 
733 	ret = do_test_code_reading(false);
734 	if (!ret)
735 		ret = do_test_code_reading(true);
736 
737 	switch (ret) {
738 	case TEST_CODE_READING_OK:
739 		return 0;
740 	case TEST_CODE_READING_NO_VMLINUX:
741 		pr_debug("no vmlinux\n");
742 		return 0;
743 	case TEST_CODE_READING_NO_KCORE:
744 		pr_debug("no kcore\n");
745 		return 0;
746 	case TEST_CODE_READING_NO_ACCESS:
747 		pr_debug("no access\n");
748 		return 0;
749 	case TEST_CODE_READING_NO_KERNEL_OBJ:
750 		pr_debug("no kernel obj\n");
751 		return 0;
752 	default:
753 		return -1;
754 	};
755 }
756