xref: /openbmc/linux/tools/perf/tests/code-reading.c (revision 4ce94eab)
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/string2.h"
30 #include "util/synthetic-events.h"
31 #include "thread.h"
32 
33 #include "tests.h"
34 
35 #include <linux/ctype.h>
36 
37 #define BUFSZ	1024
38 #define READLEN	128
39 
40 struct state {
41 	u64 done[1024];
42 	size_t done_cnt;
43 };
44 
45 static size_t read_objdump_chunk(const char **line, unsigned char **buf,
46 				 size_t *buf_len)
47 {
48 	size_t bytes_read = 0;
49 	unsigned char *chunk_start = *buf;
50 
51 	/* Read bytes */
52 	while (*buf_len > 0) {
53 		char c1, c2;
54 
55 		/* Get 2 hex digits */
56 		c1 = *(*line)++;
57 		if (!isxdigit(c1))
58 			break;
59 		c2 = *(*line)++;
60 		if (!isxdigit(c2))
61 			break;
62 
63 		/* Store byte and advance buf */
64 		**buf = (hex(c1) << 4) | hex(c2);
65 		(*buf)++;
66 		(*buf_len)--;
67 		bytes_read++;
68 
69 		/* End of chunk? */
70 		if (isspace(**line))
71 			break;
72 	}
73 
74 	/*
75 	 * objdump will display raw insn as LE if code endian
76 	 * is LE and bytes_per_chunk > 1. In that case reverse
77 	 * the chunk we just read.
78 	 *
79 	 * see disassemble_bytes() at binutils/objdump.c for details
80 	 * how objdump chooses display endian)
81 	 */
82 	if (bytes_read > 1 && !bigendian()) {
83 		unsigned char *chunk_end = chunk_start + bytes_read - 1;
84 		unsigned char tmp;
85 
86 		while (chunk_start < chunk_end) {
87 			tmp = *chunk_start;
88 			*chunk_start = *chunk_end;
89 			*chunk_end = tmp;
90 			chunk_start++;
91 			chunk_end--;
92 		}
93 	}
94 
95 	return bytes_read;
96 }
97 
98 static size_t read_objdump_line(const char *line, unsigned char *buf,
99 				size_t buf_len)
100 {
101 	const char *p;
102 	size_t ret, bytes_read = 0;
103 
104 	/* Skip to a colon */
105 	p = strchr(line, ':');
106 	if (!p)
107 		return 0;
108 	p++;
109 
110 	/* Skip initial spaces */
111 	while (*p) {
112 		if (!isspace(*p))
113 			break;
114 		p++;
115 	}
116 
117 	do {
118 		ret = read_objdump_chunk(&p, &buf, &buf_len);
119 		bytes_read += ret;
120 		p++;
121 	} while (ret > 0);
122 
123 	/* return number of successfully read bytes */
124 	return bytes_read;
125 }
126 
127 static int read_objdump_output(FILE *f, void *buf, size_t *len, u64 start_addr)
128 {
129 	char *line = NULL;
130 	size_t line_len, off_last = 0;
131 	ssize_t ret;
132 	int err = 0;
133 	u64 addr, last_addr = start_addr;
134 
135 	while (off_last < *len) {
136 		size_t off, read_bytes, written_bytes;
137 		unsigned char tmp[BUFSZ];
138 
139 		ret = getline(&line, &line_len, f);
140 		if (feof(f))
141 			break;
142 		if (ret < 0) {
143 			pr_debug("getline failed\n");
144 			err = -1;
145 			break;
146 		}
147 
148 		/* read objdump data into temporary buffer */
149 		read_bytes = read_objdump_line(line, tmp, sizeof(tmp));
150 		if (!read_bytes)
151 			continue;
152 
153 		if (sscanf(line, "%"PRIx64, &addr) != 1)
154 			continue;
155 		if (addr < last_addr) {
156 			pr_debug("addr going backwards, read beyond section?\n");
157 			break;
158 		}
159 		last_addr = addr;
160 
161 		/* copy it from temporary buffer to 'buf' according
162 		 * to address on current objdump line */
163 		off = addr - start_addr;
164 		if (off >= *len)
165 			break;
166 		written_bytes = MIN(read_bytes, *len - off);
167 		memcpy(buf + off, tmp, written_bytes);
168 		off_last = off + written_bytes;
169 	}
170 
171 	/* len returns number of bytes that could not be read */
172 	*len -= off_last;
173 
174 	free(line);
175 
176 	return err;
177 }
178 
179 static int read_via_objdump(const char *filename, u64 addr, void *buf,
180 			    size_t len)
181 {
182 	char cmd[PATH_MAX * 2];
183 	const char *fmt;
184 	FILE *f;
185 	int ret;
186 
187 	fmt = "%s -z -d --start-address=0x%"PRIx64" --stop-address=0x%"PRIx64" %s";
188 	ret = snprintf(cmd, sizeof(cmd), fmt, "objdump", addr, addr + len,
189 		       filename);
190 	if (ret <= 0 || (size_t)ret >= sizeof(cmd))
191 		return -1;
192 
193 	pr_debug("Objdump command is: %s\n", cmd);
194 
195 	/* Ignore objdump errors */
196 	strcat(cmd, " 2>/dev/null");
197 
198 	f = popen(cmd, "r");
199 	if (!f) {
200 		pr_debug("popen failed\n");
201 		return -1;
202 	}
203 
204 	ret = read_objdump_output(f, buf, &len, addr);
205 	if (len) {
206 		pr_debug("objdump read too few bytes: %zd\n", len);
207 		if (!ret)
208 			ret = len;
209 	}
210 
211 	pclose(f);
212 
213 	return ret;
214 }
215 
216 static void dump_buf(unsigned char *buf, size_t len)
217 {
218 	size_t i;
219 
220 	for (i = 0; i < len; i++) {
221 		pr_debug("0x%02x ", buf[i]);
222 		if (i % 16 == 15)
223 			pr_debug("\n");
224 	}
225 	pr_debug("\n");
226 }
227 
228 static int read_object_code(u64 addr, size_t len, u8 cpumode,
229 			    struct thread *thread, struct state *state)
230 {
231 	struct addr_location al;
232 	unsigned char buf1[BUFSZ];
233 	unsigned char buf2[BUFSZ];
234 	size_t ret_len;
235 	u64 objdump_addr;
236 	const char *objdump_name;
237 	char decomp_name[KMOD_DECOMP_LEN];
238 	bool decomp = false;
239 	int ret;
240 
241 	pr_debug("Reading object code for memory address: %#"PRIx64"\n", addr);
242 
243 	if (!thread__find_map(thread, cpumode, addr, &al) || !al.map->dso) {
244 		if (cpumode == PERF_RECORD_MISC_HYPERVISOR) {
245 			pr_debug("Hypervisor address can not be resolved - skipping\n");
246 			return 0;
247 		}
248 
249 		pr_debug("thread__find_map failed\n");
250 		return -1;
251 	}
252 
253 	pr_debug("File is: %s\n", al.map->dso->long_name);
254 
255 	if (al.map->dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS &&
256 	    !dso__is_kcore(al.map->dso)) {
257 		pr_debug("Unexpected kernel address - skipping\n");
258 		return 0;
259 	}
260 
261 	pr_debug("On file address is: %#"PRIx64"\n", al.addr);
262 
263 	if (len > BUFSZ)
264 		len = BUFSZ;
265 
266 	/* Do not go off the map */
267 	if (addr + len > al.map->end)
268 		len = al.map->end - addr;
269 
270 	/* Read the object code using perf */
271 	ret_len = dso__data_read_offset(al.map->dso, thread->maps->machine,
272 					al.addr, buf1, len);
273 	if (ret_len != len) {
274 		pr_debug("dso__data_read_offset failed\n");
275 		return -1;
276 	}
277 
278 	/*
279 	 * Converting addresses for use by objdump requires more information.
280 	 * map__load() does that.  See map__rip_2objdump() for details.
281 	 */
282 	if (map__load(al.map))
283 		return -1;
284 
285 	/* objdump struggles with kcore - try each map only once */
286 	if (dso__is_kcore(al.map->dso)) {
287 		size_t d;
288 
289 		for (d = 0; d < state->done_cnt; d++) {
290 			if (state->done[d] == al.map->start) {
291 				pr_debug("kcore map tested already");
292 				pr_debug(" - skipping\n");
293 				return 0;
294 			}
295 		}
296 		if (state->done_cnt >= ARRAY_SIZE(state->done)) {
297 			pr_debug("Too many kcore maps - skipping\n");
298 			return 0;
299 		}
300 		state->done[state->done_cnt++] = al.map->start;
301 	}
302 
303 	objdump_name = al.map->dso->long_name;
304 	if (dso__needs_decompress(al.map->dso)) {
305 		if (dso__decompress_kmodule_path(al.map->dso, objdump_name,
306 						 decomp_name,
307 						 sizeof(decomp_name)) < 0) {
308 			pr_debug("decompression failed\n");
309 			return -1;
310 		}
311 
312 		decomp = true;
313 		objdump_name = decomp_name;
314 	}
315 
316 	/* Read the object code using objdump */
317 	objdump_addr = map__rip_2objdump(al.map, al.addr);
318 	ret = read_via_objdump(objdump_name, objdump_addr, buf2, len);
319 
320 	if (decomp)
321 		unlink(objdump_name);
322 
323 	if (ret > 0) {
324 		/*
325 		 * The kernel maps are inaccurate - assume objdump is right in
326 		 * that case.
327 		 */
328 		if (cpumode == PERF_RECORD_MISC_KERNEL ||
329 		    cpumode == PERF_RECORD_MISC_GUEST_KERNEL) {
330 			len -= ret;
331 			if (len) {
332 				pr_debug("Reducing len to %zu\n", len);
333 			} else if (dso__is_kcore(al.map->dso)) {
334 				/*
335 				 * objdump cannot handle very large segments
336 				 * that may be found in kcore.
337 				 */
338 				pr_debug("objdump failed for kcore");
339 				pr_debug(" - skipping\n");
340 				return 0;
341 			} else {
342 				return -1;
343 			}
344 		}
345 	}
346 	if (ret < 0) {
347 		pr_debug("read_via_objdump failed\n");
348 		return -1;
349 	}
350 
351 	/* The results should be identical */
352 	if (memcmp(buf1, buf2, len)) {
353 		pr_debug("Bytes read differ from those read by objdump\n");
354 		pr_debug("buf1 (dso):\n");
355 		dump_buf(buf1, len);
356 		pr_debug("buf2 (objdump):\n");
357 		dump_buf(buf2, len);
358 		return -1;
359 	}
360 	pr_debug("Bytes read match those read by objdump\n");
361 
362 	return 0;
363 }
364 
365 static int process_sample_event(struct machine *machine,
366 				struct evlist *evlist,
367 				union perf_event *event, struct state *state)
368 {
369 	struct perf_sample sample;
370 	struct thread *thread;
371 	int ret;
372 
373 	if (evlist__parse_sample(evlist, event, &sample)) {
374 		pr_debug("evlist__parse_sample failed\n");
375 		return -1;
376 	}
377 
378 	thread = machine__findnew_thread(machine, sample.pid, sample.tid);
379 	if (!thread) {
380 		pr_debug("machine__findnew_thread failed\n");
381 		return -1;
382 	}
383 
384 	ret = read_object_code(sample.ip, READLEN, sample.cpumode, thread, state);
385 	thread__put(thread);
386 	return ret;
387 }
388 
389 static int process_event(struct machine *machine, struct evlist *evlist,
390 			 union perf_event *event, struct state *state)
391 {
392 	if (event->header.type == PERF_RECORD_SAMPLE)
393 		return process_sample_event(machine, evlist, event, state);
394 
395 	if (event->header.type == PERF_RECORD_THROTTLE ||
396 	    event->header.type == PERF_RECORD_UNTHROTTLE)
397 		return 0;
398 
399 	if (event->header.type < PERF_RECORD_MAX) {
400 		int ret;
401 
402 		ret = machine__process_event(machine, event, NULL);
403 		if (ret < 0)
404 			pr_debug("machine__process_event failed, event type %u\n",
405 				 event->header.type);
406 		return ret;
407 	}
408 
409 	return 0;
410 }
411 
412 static int process_events(struct machine *machine, struct evlist *evlist,
413 			  struct state *state)
414 {
415 	union perf_event *event;
416 	struct mmap *md;
417 	int i, ret;
418 
419 	for (i = 0; i < evlist->core.nr_mmaps; i++) {
420 		md = &evlist->mmap[i];
421 		if (perf_mmap__read_init(&md->core) < 0)
422 			continue;
423 
424 		while ((event = perf_mmap__read_event(&md->core)) != NULL) {
425 			ret = process_event(machine, evlist, event, state);
426 			perf_mmap__consume(&md->core);
427 			if (ret < 0)
428 				return ret;
429 		}
430 		perf_mmap__read_done(&md->core);
431 	}
432 	return 0;
433 }
434 
435 static int comp(const void *a, const void *b)
436 {
437 	return *(int *)a - *(int *)b;
438 }
439 
440 static void do_sort_something(void)
441 {
442 	int buf[40960], i;
443 
444 	for (i = 0; i < (int)ARRAY_SIZE(buf); i++)
445 		buf[i] = ARRAY_SIZE(buf) - i - 1;
446 
447 	qsort(buf, ARRAY_SIZE(buf), sizeof(int), comp);
448 
449 	for (i = 0; i < (int)ARRAY_SIZE(buf); i++) {
450 		if (buf[i] != i) {
451 			pr_debug("qsort failed\n");
452 			break;
453 		}
454 	}
455 }
456 
457 static void sort_something(void)
458 {
459 	int i;
460 
461 	for (i = 0; i < 10; i++)
462 		do_sort_something();
463 }
464 
465 static void syscall_something(void)
466 {
467 	int pipefd[2];
468 	int i;
469 
470 	for (i = 0; i < 1000; i++) {
471 		if (pipe(pipefd) < 0) {
472 			pr_debug("pipe failed\n");
473 			break;
474 		}
475 		close(pipefd[1]);
476 		close(pipefd[0]);
477 	}
478 }
479 
480 static void fs_something(void)
481 {
482 	const char *test_file_name = "temp-perf-code-reading-test-file--";
483 	FILE *f;
484 	int i;
485 
486 	for (i = 0; i < 1000; i++) {
487 		f = fopen(test_file_name, "w+");
488 		if (f) {
489 			fclose(f);
490 			unlink(test_file_name);
491 		}
492 	}
493 }
494 
495 #ifdef __s390x__
496 #include "header.h" // for get_cpuid()
497 #endif
498 
499 static const char *do_determine_event(bool excl_kernel)
500 {
501 	const char *event = excl_kernel ? "cycles:u" : "cycles";
502 
503 #ifdef __s390x__
504 	char cpuid[128], model[16], model_c[16], cpum_cf_v[16];
505 	unsigned int family;
506 	int ret, cpum_cf_a;
507 
508 	if (get_cpuid(cpuid, sizeof(cpuid)))
509 		goto out_clocks;
510 	ret = sscanf(cpuid, "%*[^,],%u,%[^,],%[^,],%[^,],%x", &family, model_c,
511 		     model, cpum_cf_v, &cpum_cf_a);
512 	if (ret != 5)		 /* Not available */
513 		goto out_clocks;
514 	if (excl_kernel && (cpum_cf_a & 4))
515 		return event;
516 	if (!excl_kernel && (cpum_cf_a & 2))
517 		return event;
518 
519 	/* Fall through: missing authorization */
520 out_clocks:
521 	event = excl_kernel ? "cpu-clock:u" : "cpu-clock";
522 
523 #endif
524 	return event;
525 }
526 
527 static void do_something(void)
528 {
529 	fs_something();
530 
531 	sort_something();
532 
533 	syscall_something();
534 }
535 
536 enum {
537 	TEST_CODE_READING_OK,
538 	TEST_CODE_READING_NO_VMLINUX,
539 	TEST_CODE_READING_NO_KCORE,
540 	TEST_CODE_READING_NO_ACCESS,
541 	TEST_CODE_READING_NO_KERNEL_OBJ,
542 };
543 
544 static int do_test_code_reading(bool try_kcore)
545 {
546 	struct machine *machine;
547 	struct thread *thread;
548 	struct record_opts opts = {
549 		.mmap_pages	     = UINT_MAX,
550 		.user_freq	     = UINT_MAX,
551 		.user_interval	     = ULLONG_MAX,
552 		.freq		     = 500,
553 		.target		     = {
554 			.uses_mmap   = true,
555 		},
556 	};
557 	struct state state = {
558 		.done_cnt = 0,
559 	};
560 	struct perf_thread_map *threads = NULL;
561 	struct perf_cpu_map *cpus = NULL;
562 	struct evlist *evlist = NULL;
563 	struct evsel *evsel = NULL;
564 	int err = -1, ret;
565 	pid_t pid;
566 	struct map *map;
567 	bool have_vmlinux, have_kcore, excl_kernel = false;
568 
569 	pid = getpid();
570 
571 	machine = machine__new_host();
572 	machine->env = &perf_env;
573 
574 	ret = machine__create_kernel_maps(machine);
575 	if (ret < 0) {
576 		pr_debug("machine__create_kernel_maps failed\n");
577 		goto out_err;
578 	}
579 
580 	/* Force the use of kallsyms instead of vmlinux to try kcore */
581 	if (try_kcore)
582 		symbol_conf.kallsyms_name = "/proc/kallsyms";
583 
584 	/* Load kernel map */
585 	map = machine__kernel_map(machine);
586 	ret = map__load(map);
587 	if (ret < 0) {
588 		pr_debug("map__load failed\n");
589 		goto out_err;
590 	}
591 	have_vmlinux = dso__is_vmlinux(map->dso);
592 	have_kcore = dso__is_kcore(map->dso);
593 
594 	/* 2nd time through we just try kcore */
595 	if (try_kcore && !have_kcore)
596 		return TEST_CODE_READING_NO_KCORE;
597 
598 	/* No point getting kernel events if there is no kernel object */
599 	if (!have_vmlinux && !have_kcore)
600 		excl_kernel = true;
601 
602 	threads = thread_map__new_by_tid(pid);
603 	if (!threads) {
604 		pr_debug("thread_map__new_by_tid failed\n");
605 		goto out_err;
606 	}
607 
608 	ret = perf_event__synthesize_thread_map(NULL, threads,
609 						perf_event__process, machine, false);
610 	if (ret < 0) {
611 		pr_debug("perf_event__synthesize_thread_map failed\n");
612 		goto out_err;
613 	}
614 
615 	thread = machine__findnew_thread(machine, pid, pid);
616 	if (!thread) {
617 		pr_debug("machine__findnew_thread failed\n");
618 		goto out_put;
619 	}
620 
621 	cpus = perf_cpu_map__new(NULL);
622 	if (!cpus) {
623 		pr_debug("perf_cpu_map__new failed\n");
624 		goto out_put;
625 	}
626 
627 	while (1) {
628 		const char *str;
629 
630 		evlist = evlist__new();
631 		if (!evlist) {
632 			pr_debug("evlist__new failed\n");
633 			goto out_put;
634 		}
635 
636 		perf_evlist__set_maps(&evlist->core, cpus, threads);
637 
638 		str = do_determine_event(excl_kernel);
639 		pr_debug("Parsing event '%s'\n", str);
640 		ret = parse_events(evlist, str, NULL);
641 		if (ret < 0) {
642 			pr_debug("parse_events failed\n");
643 			goto out_put;
644 		}
645 
646 		evlist__config(evlist, &opts, NULL);
647 
648 		evsel = evlist__first(evlist);
649 
650 		evsel->core.attr.comm = 1;
651 		evsel->core.attr.disabled = 1;
652 		evsel->core.attr.enable_on_exec = 0;
653 
654 		ret = evlist__open(evlist);
655 		if (ret < 0) {
656 			if (!excl_kernel) {
657 				excl_kernel = true;
658 				/*
659 				 * Both cpus and threads are now owned by evlist
660 				 * and will be freed by following perf_evlist__set_maps
661 				 * call. Getting refference to keep them alive.
662 				 */
663 				perf_cpu_map__get(cpus);
664 				perf_thread_map__get(threads);
665 				perf_evlist__set_maps(&evlist->core, NULL, NULL);
666 				evlist__delete(evlist);
667 				evlist = NULL;
668 				continue;
669 			}
670 
671 			if (verbose > 0) {
672 				char errbuf[512];
673 				evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
674 				pr_debug("perf_evlist__open() failed!\n%s\n", errbuf);
675 			}
676 
677 			goto out_put;
678 		}
679 		break;
680 	}
681 
682 	ret = evlist__mmap(evlist, UINT_MAX);
683 	if (ret < 0) {
684 		pr_debug("evlist__mmap failed\n");
685 		goto out_put;
686 	}
687 
688 	evlist__enable(evlist);
689 
690 	do_something();
691 
692 	evlist__disable(evlist);
693 
694 	ret = process_events(machine, evlist, &state);
695 	if (ret < 0)
696 		goto out_put;
697 
698 	if (!have_vmlinux && !have_kcore && !try_kcore)
699 		err = TEST_CODE_READING_NO_KERNEL_OBJ;
700 	else if (!have_vmlinux && !try_kcore)
701 		err = TEST_CODE_READING_NO_VMLINUX;
702 	else if (excl_kernel)
703 		err = TEST_CODE_READING_NO_ACCESS;
704 	else
705 		err = TEST_CODE_READING_OK;
706 out_put:
707 	thread__put(thread);
708 out_err:
709 
710 	if (evlist) {
711 		evlist__delete(evlist);
712 	} else {
713 		perf_cpu_map__put(cpus);
714 		perf_thread_map__put(threads);
715 	}
716 	machine__delete_threads(machine);
717 	machine__delete(machine);
718 
719 	return err;
720 }
721 
722 int test__code_reading(struct test *test __maybe_unused, int subtest __maybe_unused)
723 {
724 	int ret;
725 
726 	ret = do_test_code_reading(false);
727 	if (!ret)
728 		ret = do_test_code_reading(true);
729 
730 	switch (ret) {
731 	case TEST_CODE_READING_OK:
732 		return 0;
733 	case TEST_CODE_READING_NO_VMLINUX:
734 		pr_debug("no vmlinux\n");
735 		return 0;
736 	case TEST_CODE_READING_NO_KCORE:
737 		pr_debug("no kcore\n");
738 		return 0;
739 	case TEST_CODE_READING_NO_ACCESS:
740 		pr_debug("no access\n");
741 		return 0;
742 	case TEST_CODE_READING_NO_KERNEL_OBJ:
743 		pr_debug("no kernel obj\n");
744 		return 0;
745 	default:
746 		return -1;
747 	};
748 }
749