xref: /linux/tools/perf/tests/code-reading.c (revision 5027ec19f1049a07df5b0a37b1f462514cf2724b)
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 	/*
273 	 * Some architectures (ex: powerpc) have stubs (trampolines) in kernel
274 	 * modules to manage long jumps. Check if the ip offset falls in stubs
275 	 * sections for kernel modules. And skip module address after text end
276 	 */
277 	if (dso->is_kmod && al.addr > dso->text_end) {
278 		pr_debug("skipping the module address %#"PRIx64" after text end\n", al.addr);
279 		goto out;
280 	}
281 
282 	/* Read the object code using perf */
283 	ret_len = dso__data_read_offset(dso, maps__machine(thread__maps(thread)),
284 					al.addr, buf1, len);
285 	if (ret_len != len) {
286 		pr_debug("dso__data_read_offset failed\n");
287 		err = -1;
288 		goto out;
289 	}
290 
291 	/*
292 	 * Converting addresses for use by objdump requires more information.
293 	 * map__load() does that.  See map__rip_2objdump() for details.
294 	 */
295 	if (map__load(al.map)) {
296 		err = -1;
297 		goto out;
298 	}
299 
300 	/* objdump struggles with kcore - try each map only once */
301 	if (dso__is_kcore(dso)) {
302 		size_t d;
303 
304 		for (d = 0; d < state->done_cnt; d++) {
305 			if (state->done[d] == map__start(al.map)) {
306 				pr_debug("kcore map tested already");
307 				pr_debug(" - skipping\n");
308 				goto out;
309 			}
310 		}
311 		if (state->done_cnt >= ARRAY_SIZE(state->done)) {
312 			pr_debug("Too many kcore maps - skipping\n");
313 			goto out;
314 		}
315 		state->done[state->done_cnt++] = map__start(al.map);
316 	}
317 
318 	objdump_name = dso->long_name;
319 	if (dso__needs_decompress(dso)) {
320 		if (dso__decompress_kmodule_path(dso, objdump_name,
321 						 decomp_name,
322 						 sizeof(decomp_name)) < 0) {
323 			pr_debug("decompression failed\n");
324 			err = -1;
325 			goto out;
326 		}
327 
328 		decomp = true;
329 		objdump_name = decomp_name;
330 	}
331 
332 	/* Read the object code using objdump */
333 	objdump_addr = map__rip_2objdump(al.map, al.addr);
334 	ret = read_via_objdump(objdump_name, objdump_addr, buf2, len);
335 
336 	if (decomp)
337 		unlink(objdump_name);
338 
339 	if (ret > 0) {
340 		/*
341 		 * The kernel maps are inaccurate - assume objdump is right in
342 		 * that case.
343 		 */
344 		if (cpumode == PERF_RECORD_MISC_KERNEL ||
345 		    cpumode == PERF_RECORD_MISC_GUEST_KERNEL) {
346 			len -= ret;
347 			if (len) {
348 				pr_debug("Reducing len to %zu\n", len);
349 			} else if (dso__is_kcore(dso)) {
350 				/*
351 				 * objdump cannot handle very large segments
352 				 * that may be found in kcore.
353 				 */
354 				pr_debug("objdump failed for kcore");
355 				pr_debug(" - skipping\n");
356 			} else {
357 				err = -1;
358 			}
359 			goto out;
360 		}
361 	}
362 	if (ret < 0) {
363 		pr_debug("read_via_objdump failed\n");
364 		err = -1;
365 		goto out;
366 	}
367 
368 	/* The results should be identical */
369 	if (memcmp(buf1, buf2, len)) {
370 		pr_debug("Bytes read differ from those read by objdump\n");
371 		pr_debug("buf1 (dso):\n");
372 		dump_buf(buf1, len);
373 		pr_debug("buf2 (objdump):\n");
374 		dump_buf(buf2, len);
375 		err = -1;
376 		goto out;
377 	}
378 	pr_debug("Bytes read match those read by objdump\n");
379 out:
380 	addr_location__exit(&al);
381 	return err;
382 }
383 
384 static int process_sample_event(struct machine *machine,
385 				struct evlist *evlist,
386 				union perf_event *event, struct state *state)
387 {
388 	struct perf_sample sample;
389 	struct thread *thread;
390 	int ret;
391 
392 	if (evlist__parse_sample(evlist, event, &sample)) {
393 		pr_debug("evlist__parse_sample failed\n");
394 		return -1;
395 	}
396 
397 	thread = machine__findnew_thread(machine, sample.pid, sample.tid);
398 	if (!thread) {
399 		pr_debug("machine__findnew_thread failed\n");
400 		return -1;
401 	}
402 
403 	ret = read_object_code(sample.ip, READLEN, sample.cpumode, thread, state);
404 	thread__put(thread);
405 	return ret;
406 }
407 
408 static int process_event(struct machine *machine, struct evlist *evlist,
409 			 union perf_event *event, struct state *state)
410 {
411 	if (event->header.type == PERF_RECORD_SAMPLE)
412 		return process_sample_event(machine, evlist, event, state);
413 
414 	if (event->header.type == PERF_RECORD_THROTTLE ||
415 	    event->header.type == PERF_RECORD_UNTHROTTLE)
416 		return 0;
417 
418 	if (event->header.type < PERF_RECORD_MAX) {
419 		int ret;
420 
421 		ret = machine__process_event(machine, event, NULL);
422 		if (ret < 0)
423 			pr_debug("machine__process_event failed, event type %u\n",
424 				 event->header.type);
425 		return ret;
426 	}
427 
428 	return 0;
429 }
430 
431 static int process_events(struct machine *machine, struct evlist *evlist,
432 			  struct state *state)
433 {
434 	union perf_event *event;
435 	struct mmap *md;
436 	int i, ret;
437 
438 	for (i = 0; i < evlist->core.nr_mmaps; i++) {
439 		md = &evlist->mmap[i];
440 		if (perf_mmap__read_init(&md->core) < 0)
441 			continue;
442 
443 		while ((event = perf_mmap__read_event(&md->core)) != NULL) {
444 			ret = process_event(machine, evlist, event, state);
445 			perf_mmap__consume(&md->core);
446 			if (ret < 0)
447 				return ret;
448 		}
449 		perf_mmap__read_done(&md->core);
450 	}
451 	return 0;
452 }
453 
454 static int comp(const void *a, const void *b)
455 {
456 	return *(int *)a - *(int *)b;
457 }
458 
459 static void do_sort_something(void)
460 {
461 	int buf[40960], i;
462 
463 	for (i = 0; i < (int)ARRAY_SIZE(buf); i++)
464 		buf[i] = ARRAY_SIZE(buf) - i - 1;
465 
466 	qsort(buf, ARRAY_SIZE(buf), sizeof(int), comp);
467 
468 	for (i = 0; i < (int)ARRAY_SIZE(buf); i++) {
469 		if (buf[i] != i) {
470 			pr_debug("qsort failed\n");
471 			break;
472 		}
473 	}
474 }
475 
476 static void sort_something(void)
477 {
478 	int i;
479 
480 	for (i = 0; i < 10; i++)
481 		do_sort_something();
482 }
483 
484 static void syscall_something(void)
485 {
486 	int pipefd[2];
487 	int i;
488 
489 	for (i = 0; i < 1000; i++) {
490 		if (pipe(pipefd) < 0) {
491 			pr_debug("pipe failed\n");
492 			break;
493 		}
494 		close(pipefd[1]);
495 		close(pipefd[0]);
496 	}
497 }
498 
499 static void fs_something(void)
500 {
501 	const char *test_file_name = "temp-perf-code-reading-test-file--";
502 	FILE *f;
503 	int i;
504 
505 	for (i = 0; i < 1000; i++) {
506 		f = fopen(test_file_name, "w+");
507 		if (f) {
508 			fclose(f);
509 			unlink(test_file_name);
510 		}
511 	}
512 }
513 
514 #ifdef __s390x__
515 #include "header.h" // for get_cpuid()
516 #endif
517 
518 static const char *do_determine_event(bool excl_kernel)
519 {
520 	const char *event = excl_kernel ? "cycles:u" : "cycles";
521 
522 #ifdef __s390x__
523 	char cpuid[128], model[16], model_c[16], cpum_cf_v[16];
524 	unsigned int family;
525 	int ret, cpum_cf_a;
526 
527 	if (get_cpuid(cpuid, sizeof(cpuid)))
528 		goto out_clocks;
529 	ret = sscanf(cpuid, "%*[^,],%u,%[^,],%[^,],%[^,],%x", &family, model_c,
530 		     model, cpum_cf_v, &cpum_cf_a);
531 	if (ret != 5)		 /* Not available */
532 		goto out_clocks;
533 	if (excl_kernel && (cpum_cf_a & 4))
534 		return event;
535 	if (!excl_kernel && (cpum_cf_a & 2))
536 		return event;
537 
538 	/* Fall through: missing authorization */
539 out_clocks:
540 	event = excl_kernel ? "cpu-clock:u" : "cpu-clock";
541 
542 #endif
543 	return event;
544 }
545 
546 static void do_something(void)
547 {
548 	fs_something();
549 
550 	sort_something();
551 
552 	syscall_something();
553 }
554 
555 enum {
556 	TEST_CODE_READING_OK,
557 	TEST_CODE_READING_NO_VMLINUX,
558 	TEST_CODE_READING_NO_KCORE,
559 	TEST_CODE_READING_NO_ACCESS,
560 	TEST_CODE_READING_NO_KERNEL_OBJ,
561 };
562 
563 static int do_test_code_reading(bool try_kcore)
564 {
565 	struct machine *machine;
566 	struct thread *thread;
567 	struct record_opts opts = {
568 		.mmap_pages	     = UINT_MAX,
569 		.user_freq	     = UINT_MAX,
570 		.user_interval	     = ULLONG_MAX,
571 		.freq		     = 500,
572 		.target		     = {
573 			.uses_mmap   = true,
574 		},
575 	};
576 	struct state state = {
577 		.done_cnt = 0,
578 	};
579 	struct perf_thread_map *threads = NULL;
580 	struct perf_cpu_map *cpus = NULL;
581 	struct evlist *evlist = NULL;
582 	struct evsel *evsel = NULL;
583 	int err = -1, ret;
584 	pid_t pid;
585 	struct map *map;
586 	bool have_vmlinux, have_kcore, excl_kernel = false;
587 	struct dso *dso;
588 
589 	pid = getpid();
590 
591 	machine = machine__new_host();
592 	machine->env = &perf_env;
593 
594 	ret = machine__create_kernel_maps(machine);
595 	if (ret < 0) {
596 		pr_debug("machine__create_kernel_maps failed\n");
597 		goto out_err;
598 	}
599 
600 	/* Force the use of kallsyms instead of vmlinux to try kcore */
601 	if (try_kcore)
602 		symbol_conf.kallsyms_name = "/proc/kallsyms";
603 
604 	/* Load kernel map */
605 	map = machine__kernel_map(machine);
606 	ret = map__load(map);
607 	if (ret < 0) {
608 		pr_debug("map__load failed\n");
609 		goto out_err;
610 	}
611 	dso = map__dso(map);
612 	have_vmlinux = dso__is_vmlinux(dso);
613 	have_kcore = dso__is_kcore(dso);
614 
615 	/* 2nd time through we just try kcore */
616 	if (try_kcore && !have_kcore)
617 		return TEST_CODE_READING_NO_KCORE;
618 
619 	/* No point getting kernel events if there is no kernel object */
620 	if (!have_vmlinux && !have_kcore)
621 		excl_kernel = true;
622 
623 	threads = thread_map__new_by_tid(pid);
624 	if (!threads) {
625 		pr_debug("thread_map__new_by_tid failed\n");
626 		goto out_err;
627 	}
628 
629 	ret = perf_event__synthesize_thread_map(NULL, threads,
630 						perf_event__process, machine,
631 						true, false);
632 	if (ret < 0) {
633 		pr_debug("perf_event__synthesize_thread_map failed\n");
634 		goto out_err;
635 	}
636 
637 	thread = machine__findnew_thread(machine, pid, pid);
638 	if (!thread) {
639 		pr_debug("machine__findnew_thread failed\n");
640 		goto out_put;
641 	}
642 
643 	cpus = perf_cpu_map__new(NULL);
644 	if (!cpus) {
645 		pr_debug("perf_cpu_map__new failed\n");
646 		goto out_put;
647 	}
648 
649 	while (1) {
650 		const char *str;
651 
652 		evlist = evlist__new();
653 		if (!evlist) {
654 			pr_debug("evlist__new failed\n");
655 			goto out_put;
656 		}
657 
658 		perf_evlist__set_maps(&evlist->core, cpus, threads);
659 
660 		str = do_determine_event(excl_kernel);
661 		pr_debug("Parsing event '%s'\n", str);
662 		ret = parse_event(evlist, str);
663 		if (ret < 0) {
664 			pr_debug("parse_events failed\n");
665 			goto out_put;
666 		}
667 
668 		evlist__config(evlist, &opts, NULL);
669 
670 		evsel = evlist__first(evlist);
671 
672 		evsel->core.attr.comm = 1;
673 		evsel->core.attr.disabled = 1;
674 		evsel->core.attr.enable_on_exec = 0;
675 
676 		ret = evlist__open(evlist);
677 		if (ret < 0) {
678 			if (!excl_kernel) {
679 				excl_kernel = true;
680 				/*
681 				 * Both cpus and threads are now owned by evlist
682 				 * and will be freed by following perf_evlist__set_maps
683 				 * call. Getting reference to keep them alive.
684 				 */
685 				perf_cpu_map__get(cpus);
686 				perf_thread_map__get(threads);
687 				perf_evlist__set_maps(&evlist->core, NULL, NULL);
688 				evlist__delete(evlist);
689 				evlist = NULL;
690 				continue;
691 			}
692 
693 			if (verbose > 0) {
694 				char errbuf[512];
695 				evlist__strerror_open(evlist, errno, errbuf, sizeof(errbuf));
696 				pr_debug("perf_evlist__open() failed!\n%s\n", errbuf);
697 			}
698 
699 			goto out_put;
700 		}
701 		break;
702 	}
703 
704 	ret = evlist__mmap(evlist, UINT_MAX);
705 	if (ret < 0) {
706 		pr_debug("evlist__mmap failed\n");
707 		goto out_put;
708 	}
709 
710 	evlist__enable(evlist);
711 
712 	do_something();
713 
714 	evlist__disable(evlist);
715 
716 	ret = process_events(machine, evlist, &state);
717 	if (ret < 0)
718 		goto out_put;
719 
720 	if (!have_vmlinux && !have_kcore && !try_kcore)
721 		err = TEST_CODE_READING_NO_KERNEL_OBJ;
722 	else if (!have_vmlinux && !try_kcore)
723 		err = TEST_CODE_READING_NO_VMLINUX;
724 	else if (excl_kernel)
725 		err = TEST_CODE_READING_NO_ACCESS;
726 	else
727 		err = TEST_CODE_READING_OK;
728 out_put:
729 	thread__put(thread);
730 out_err:
731 	evlist__delete(evlist);
732 	perf_cpu_map__put(cpus);
733 	perf_thread_map__put(threads);
734 	machine__delete(machine);
735 
736 	return err;
737 }
738 
739 static int test__code_reading(struct test_suite *test __maybe_unused, int subtest __maybe_unused)
740 {
741 	int ret;
742 
743 	ret = do_test_code_reading(false);
744 	if (!ret)
745 		ret = do_test_code_reading(true);
746 
747 	switch (ret) {
748 	case TEST_CODE_READING_OK:
749 		return 0;
750 	case TEST_CODE_READING_NO_VMLINUX:
751 		pr_debug("no vmlinux\n");
752 		return 0;
753 	case TEST_CODE_READING_NO_KCORE:
754 		pr_debug("no kcore\n");
755 		return 0;
756 	case TEST_CODE_READING_NO_ACCESS:
757 		pr_debug("no access\n");
758 		return 0;
759 	case TEST_CODE_READING_NO_KERNEL_OBJ:
760 		pr_debug("no kernel obj\n");
761 		return 0;
762 	default:
763 		return -1;
764 	};
765 }
766 
767 DEFINE_SUITE("Object code reading", code_reading);
768