xref: /openbmc/linux/tools/perf/tests/code-reading.c (revision da097dcc)
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 "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/string2.h"
29 #include "util/synthetic-events.h"
30 #include "util/util.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 && !host_is_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] = {0};
233 	unsigned char buf2[BUFSZ] = {0};
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, err = 0;
240 	struct dso *dso;
241 
242 	pr_debug("Reading object code for memory address: %#"PRIx64"\n", addr);
243 
244 	addr_location__init(&al);
245 	if (!thread__find_map(thread, cpumode, addr, &al) || !map__dso(al.map)) {
246 		if (cpumode == PERF_RECORD_MISC_HYPERVISOR) {
247 			pr_debug("Hypervisor address can not be resolved - skipping\n");
248 			goto out;
249 		}
250 
251 		pr_debug("thread__find_map failed\n");
252 		err = -1;
253 		goto out;
254 	}
255 	dso = map__dso(al.map);
256 	pr_debug("File is: %s\n", dso->long_name);
257 
258 	if (dso->symtab_type == DSO_BINARY_TYPE__KALLSYMS && !dso__is_kcore(dso)) {
259 		pr_debug("Unexpected kernel address - skipping\n");
260 		goto out;
261 	}
262 
263 	pr_debug("On file address is: %#"PRIx64"\n", al.addr);
264 
265 	if (len > BUFSZ)
266 		len = BUFSZ;
267 
268 	/* Do not go off the map */
269 	if (addr + len > map__end(al.map))
270 		len = map__end(al.map) - addr;
271 
272 	/* Read the object code using perf */
273 	ret_len = dso__data_read_offset(dso, maps__machine(thread__maps(thread)),
274 					al.addr, buf1, len);
275 	if (ret_len != len) {
276 		pr_debug("dso__data_read_offset failed\n");
277 		err = -1;
278 		goto out;
279 	}
280 
281 	/*
282 	 * Converting addresses for use by objdump requires more information.
283 	 * map__load() does that.  See map__rip_2objdump() for details.
284 	 */
285 	if (map__load(al.map)) {
286 		err = -1;
287 		goto out;
288 	}
289 
290 	/* objdump struggles with kcore - try each map only once */
291 	if (dso__is_kcore(dso)) {
292 		size_t d;
293 
294 		for (d = 0; d < state->done_cnt; d++) {
295 			if (state->done[d] == map__start(al.map)) {
296 				pr_debug("kcore map tested already");
297 				pr_debug(" - skipping\n");
298 				goto out;
299 			}
300 		}
301 		if (state->done_cnt >= ARRAY_SIZE(state->done)) {
302 			pr_debug("Too many kcore maps - skipping\n");
303 			goto out;
304 		}
305 		state->done[state->done_cnt++] = map__start(al.map);
306 	}
307 
308 	objdump_name = dso->long_name;
309 	if (dso__needs_decompress(dso)) {
310 		if (dso__decompress_kmodule_path(dso, objdump_name,
311 						 decomp_name,
312 						 sizeof(decomp_name)) < 0) {
313 			pr_debug("decompression failed\n");
314 			err = -1;
315 			goto out;
316 		}
317 
318 		decomp = true;
319 		objdump_name = decomp_name;
320 	}
321 
322 	/* Read the object code using objdump */
323 	objdump_addr = map__rip_2objdump(al.map, al.addr);
324 	ret = read_via_objdump(objdump_name, objdump_addr, buf2, len);
325 
326 	if (decomp)
327 		unlink(objdump_name);
328 
329 	if (ret > 0) {
330 		/*
331 		 * The kernel maps are inaccurate - assume objdump is right in
332 		 * that case.
333 		 */
334 		if (cpumode == PERF_RECORD_MISC_KERNEL ||
335 		    cpumode == PERF_RECORD_MISC_GUEST_KERNEL) {
336 			len -= ret;
337 			if (len) {
338 				pr_debug("Reducing len to %zu\n", len);
339 			} else if (dso__is_kcore(dso)) {
340 				/*
341 				 * objdump cannot handle very large segments
342 				 * that may be found in kcore.
343 				 */
344 				pr_debug("objdump failed for kcore");
345 				pr_debug(" - skipping\n");
346 			} else {
347 				err = -1;
348 			}
349 			goto out;
350 		}
351 	}
352 	if (ret < 0) {
353 		pr_debug("read_via_objdump failed\n");
354 		err = -1;
355 		goto out;
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 		err = -1;
366 		goto out;
367 	}
368 	pr_debug("Bytes read match those read by objdump\n");
369 out:
370 	addr_location__exit(&al);
371 	return err;
372 }
373 
374 static int process_sample_event(struct machine *machine,
375 				struct evlist *evlist,
376 				union perf_event *event, struct state *state)
377 {
378 	struct perf_sample sample;
379 	struct thread *thread;
380 	int ret;
381 
382 	if (evlist__parse_sample(evlist, event, &sample)) {
383 		pr_debug("evlist__parse_sample failed\n");
384 		return -1;
385 	}
386 
387 	thread = machine__findnew_thread(machine, sample.pid, sample.tid);
388 	if (!thread) {
389 		pr_debug("machine__findnew_thread failed\n");
390 		return -1;
391 	}
392 
393 	ret = read_object_code(sample.ip, READLEN, sample.cpumode, thread, state);
394 	thread__put(thread);
395 	return ret;
396 }
397 
398 static int process_event(struct machine *machine, struct evlist *evlist,
399 			 union perf_event *event, struct state *state)
400 {
401 	if (event->header.type == PERF_RECORD_SAMPLE)
402 		return process_sample_event(machine, evlist, event, state);
403 
404 	if (event->header.type == PERF_RECORD_THROTTLE ||
405 	    event->header.type == PERF_RECORD_UNTHROTTLE)
406 		return 0;
407 
408 	if (event->header.type < PERF_RECORD_MAX) {
409 		int ret;
410 
411 		ret = machine__process_event(machine, event, NULL);
412 		if (ret < 0)
413 			pr_debug("machine__process_event failed, event type %u\n",
414 				 event->header.type);
415 		return ret;
416 	}
417 
418 	return 0;
419 }
420 
421 static int process_events(struct machine *machine, struct evlist *evlist,
422 			  struct state *state)
423 {
424 	union perf_event *event;
425 	struct mmap *md;
426 	int i, ret;
427 
428 	for (i = 0; i < evlist->core.nr_mmaps; i++) {
429 		md = &evlist->mmap[i];
430 		if (perf_mmap__read_init(&md->core) < 0)
431 			continue;
432 
433 		while ((event = perf_mmap__read_event(&md->core)) != NULL) {
434 			ret = process_event(machine, evlist, event, state);
435 			perf_mmap__consume(&md->core);
436 			if (ret < 0)
437 				return ret;
438 		}
439 		perf_mmap__read_done(&md->core);
440 	}
441 	return 0;
442 }
443 
444 static int comp(const void *a, const void *b)
445 {
446 	return *(int *)a - *(int *)b;
447 }
448 
449 static void do_sort_something(void)
450 {
451 	int buf[40960], i;
452 
453 	for (i = 0; i < (int)ARRAY_SIZE(buf); i++)
454 		buf[i] = ARRAY_SIZE(buf) - i - 1;
455 
456 	qsort(buf, ARRAY_SIZE(buf), sizeof(int), comp);
457 
458 	for (i = 0; i < (int)ARRAY_SIZE(buf); i++) {
459 		if (buf[i] != i) {
460 			pr_debug("qsort failed\n");
461 			break;
462 		}
463 	}
464 }
465 
466 static void sort_something(void)
467 {
468 	int i;
469 
470 	for (i = 0; i < 10; i++)
471 		do_sort_something();
472 }
473 
474 static void syscall_something(void)
475 {
476 	int pipefd[2];
477 	int i;
478 
479 	for (i = 0; i < 1000; i++) {
480 		if (pipe(pipefd) < 0) {
481 			pr_debug("pipe failed\n");
482 			break;
483 		}
484 		close(pipefd[1]);
485 		close(pipefd[0]);
486 	}
487 }
488 
489 static void fs_something(void)
490 {
491 	const char *test_file_name = "temp-perf-code-reading-test-file--";
492 	FILE *f;
493 	int i;
494 
495 	for (i = 0; i < 1000; i++) {
496 		f = fopen(test_file_name, "w+");
497 		if (f) {
498 			fclose(f);
499 			unlink(test_file_name);
500 		}
501 	}
502 }
503 
504 #ifdef __s390x__
505 #include "header.h" // for get_cpuid()
506 #endif
507 
508 static const char *do_determine_event(bool excl_kernel)
509 {
510 	const char *event = excl_kernel ? "cycles:u" : "cycles";
511 
512 #ifdef __s390x__
513 	char cpuid[128], model[16], model_c[16], cpum_cf_v[16];
514 	unsigned int family;
515 	int ret, cpum_cf_a;
516 
517 	if (get_cpuid(cpuid, sizeof(cpuid)))
518 		goto out_clocks;
519 	ret = sscanf(cpuid, "%*[^,],%u,%[^,],%[^,],%[^,],%x", &family, model_c,
520 		     model, cpum_cf_v, &cpum_cf_a);
521 	if (ret != 5)		 /* Not available */
522 		goto out_clocks;
523 	if (excl_kernel && (cpum_cf_a & 4))
524 		return event;
525 	if (!excl_kernel && (cpum_cf_a & 2))
526 		return event;
527 
528 	/* Fall through: missing authorization */
529 out_clocks:
530 	event = excl_kernel ? "cpu-clock:u" : "cpu-clock";
531 
532 #endif
533 	return event;
534 }
535 
536 static void do_something(void)
537 {
538 	fs_something();
539 
540 	sort_something();
541 
542 	syscall_something();
543 }
544 
545 enum {
546 	TEST_CODE_READING_OK,
547 	TEST_CODE_READING_NO_VMLINUX,
548 	TEST_CODE_READING_NO_KCORE,
549 	TEST_CODE_READING_NO_ACCESS,
550 	TEST_CODE_READING_NO_KERNEL_OBJ,
551 };
552 
553 static int do_test_code_reading(bool try_kcore)
554 {
555 	struct machine *machine;
556 	struct thread *thread;
557 	struct record_opts opts = {
558 		.mmap_pages	     = UINT_MAX,
559 		.user_freq	     = UINT_MAX,
560 		.user_interval	     = ULLONG_MAX,
561 		.freq		     = 500,
562 		.target		     = {
563 			.uses_mmap   = true,
564 		},
565 	};
566 	struct state state = {
567 		.done_cnt = 0,
568 	};
569 	struct perf_thread_map *threads = NULL;
570 	struct perf_cpu_map *cpus = NULL;
571 	struct evlist *evlist = NULL;
572 	struct evsel *evsel = NULL;
573 	int err = -1, ret;
574 	pid_t pid;
575 	struct map *map;
576 	bool have_vmlinux, have_kcore, excl_kernel = false;
577 	struct dso *dso;
578 
579 	pid = getpid();
580 
581 	machine = machine__new_host();
582 	machine->env = &perf_env;
583 
584 	ret = machine__create_kernel_maps(machine);
585 	if (ret < 0) {
586 		pr_debug("machine__create_kernel_maps failed\n");
587 		goto out_err;
588 	}
589 
590 	/* Force the use of kallsyms instead of vmlinux to try kcore */
591 	if (try_kcore)
592 		symbol_conf.kallsyms_name = "/proc/kallsyms";
593 
594 	/* Load kernel map */
595 	map = machine__kernel_map(machine);
596 	ret = map__load(map);
597 	if (ret < 0) {
598 		pr_debug("map__load failed\n");
599 		goto out_err;
600 	}
601 	dso = map__dso(map);
602 	have_vmlinux = dso__is_vmlinux(dso);
603 	have_kcore = dso__is_kcore(dso);
604 
605 	/* 2nd time through we just try kcore */
606 	if (try_kcore && !have_kcore)
607 		return TEST_CODE_READING_NO_KCORE;
608 
609 	/* No point getting kernel events if there is no kernel object */
610 	if (!have_vmlinux && !have_kcore)
611 		excl_kernel = true;
612 
613 	threads = thread_map__new_by_tid(pid);
614 	if (!threads) {
615 		pr_debug("thread_map__new_by_tid failed\n");
616 		goto out_err;
617 	}
618 
619 	ret = perf_event__synthesize_thread_map(NULL, threads,
620 						perf_event__process, machine,
621 						true, false);
622 	if (ret < 0) {
623 		pr_debug("perf_event__synthesize_thread_map failed\n");
624 		goto out_err;
625 	}
626 
627 	thread = machine__findnew_thread(machine, pid, pid);
628 	if (!thread) {
629 		pr_debug("machine__findnew_thread failed\n");
630 		goto out_put;
631 	}
632 
633 	cpus = perf_cpu_map__new(NULL);
634 	if (!cpus) {
635 		pr_debug("perf_cpu_map__new failed\n");
636 		goto out_put;
637 	}
638 
639 	while (1) {
640 		const char *str;
641 
642 		evlist = evlist__new();
643 		if (!evlist) {
644 			pr_debug("evlist__new failed\n");
645 			goto out_put;
646 		}
647 
648 		perf_evlist__set_maps(&evlist->core, cpus, threads);
649 
650 		str = do_determine_event(excl_kernel);
651 		pr_debug("Parsing event '%s'\n", str);
652 		ret = parse_event(evlist, str);
653 		if (ret < 0) {
654 			pr_debug("parse_events failed\n");
655 			goto out_put;
656 		}
657 
658 		evlist__config(evlist, &opts, NULL);
659 
660 		evlist__for_each_entry(evlist, evsel) {
661 			evsel->core.attr.comm = 1;
662 			evsel->core.attr.disabled = 1;
663 			evsel->core.attr.enable_on_exec = 0;
664 		}
665 
666 		ret = evlist__open(evlist);
667 		if (ret < 0) {
668 			if (!excl_kernel) {
669 				excl_kernel = true;
670 				/*
671 				 * Both cpus and threads are now owned by evlist
672 				 * and will be freed by following perf_evlist__set_maps
673 				 * call. Getting reference to keep them alive.
674 				 */
675 				perf_cpu_map__get(cpus);
676 				perf_thread_map__get(threads);
677 				perf_evlist__set_maps(&evlist->core, NULL, NULL);
678 				evlist__delete(evlist);
679 				evlist = NULL;
680 				continue;
681 			}
682 
683 			if (verbose > 0) {
684 				char errbuf[512];
685 				evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
686 				pr_debug("perf_evlist__open() failed!\n%s\n", errbuf);
687 			}
688 
689 			goto out_put;
690 		}
691 		break;
692 	}
693 
694 	ret = evlist__mmap(evlist, UINT_MAX);
695 	if (ret < 0) {
696 		pr_debug("evlist__mmap failed\n");
697 		goto out_put;
698 	}
699 
700 	evlist__enable(evlist);
701 
702 	do_something();
703 
704 	evlist__disable(evlist);
705 
706 	ret = process_events(machine, evlist, &state);
707 	if (ret < 0)
708 		goto out_put;
709 
710 	if (!have_vmlinux && !have_kcore && !try_kcore)
711 		err = TEST_CODE_READING_NO_KERNEL_OBJ;
712 	else if (!have_vmlinux && !try_kcore)
713 		err = TEST_CODE_READING_NO_VMLINUX;
714 	else if (excl_kernel)
715 		err = TEST_CODE_READING_NO_ACCESS;
716 	else
717 		err = TEST_CODE_READING_OK;
718 out_put:
719 	thread__put(thread);
720 out_err:
721 	evlist__delete(evlist);
722 	perf_cpu_map__put(cpus);
723 	perf_thread_map__put(threads);
724 	machine__delete(machine);
725 
726 	return err;
727 }
728 
729 static int test__code_reading(struct test_suite *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 
757 DEFINE_SUITE("Object code reading", code_reading);
758