1 /*
2  * Post mortem Dwarf CFI based unwinding on top of regs and stack dumps.
3  *
4  * Lots of this code have been borrowed or heavily inspired from parts of
5  * the libunwind 0.99 code which are (amongst other contributors I may have
6  * forgotten):
7  *
8  * Copyright (C) 2002-2007 Hewlett-Packard Co
9  *	Contributed by David Mosberger-Tang <davidm@hpl.hp.com>
10  *
11  * And the bugs have been added by:
12  *
13  * Copyright (C) 2010, Frederic Weisbecker <fweisbec@gmail.com>
14  * Copyright (C) 2012, Jiri Olsa <jolsa@redhat.com>
15  *
16  */
17 
18 #include <elf.h>
19 #include <gelf.h>
20 #include <fcntl.h>
21 #include <string.h>
22 #include <unistd.h>
23 #include <sys/mman.h>
24 #include <linux/list.h>
25 #include <libunwind.h>
26 #include <libunwind-ptrace.h>
27 #include "callchain.h"
28 #include "thread.h"
29 #include "session.h"
30 #include "perf_regs.h"
31 #include "unwind.h"
32 #include "symbol.h"
33 #include "util.h"
34 #include "debug.h"
35 
36 extern int
37 UNW_OBJ(dwarf_search_unwind_table) (unw_addr_space_t as,
38 				    unw_word_t ip,
39 				    unw_dyn_info_t *di,
40 				    unw_proc_info_t *pi,
41 				    int need_unwind_info, void *arg);
42 
43 #define dwarf_search_unwind_table UNW_OBJ(dwarf_search_unwind_table)
44 
45 extern int
46 UNW_OBJ(dwarf_find_debug_frame) (int found, unw_dyn_info_t *di_debug,
47 				 unw_word_t ip,
48 				 unw_word_t segbase,
49 				 const char *obj_name, unw_word_t start,
50 				 unw_word_t end);
51 
52 #define dwarf_find_debug_frame UNW_OBJ(dwarf_find_debug_frame)
53 
54 #define DW_EH_PE_FORMAT_MASK	0x0f	/* format of the encoded value */
55 #define DW_EH_PE_APPL_MASK	0x70	/* how the value is to be applied */
56 
57 /* Pointer-encoding formats: */
58 #define DW_EH_PE_omit		0xff
59 #define DW_EH_PE_ptr		0x00	/* pointer-sized unsigned value */
60 #define DW_EH_PE_udata4		0x03	/* unsigned 32-bit value */
61 #define DW_EH_PE_udata8		0x04	/* unsigned 64-bit value */
62 #define DW_EH_PE_sdata4		0x0b	/* signed 32-bit value */
63 #define DW_EH_PE_sdata8		0x0c	/* signed 64-bit value */
64 
65 /* Pointer-encoding application: */
66 #define DW_EH_PE_absptr		0x00	/* absolute value */
67 #define DW_EH_PE_pcrel		0x10	/* rel. to addr. of encoded value */
68 
69 /*
70  * The following are not documented by LSB v1.3, yet they are used by
71  * GCC, presumably they aren't documented by LSB since they aren't
72  * used on Linux:
73  */
74 #define DW_EH_PE_funcrel	0x40	/* start-of-procedure-relative */
75 #define DW_EH_PE_aligned	0x50	/* aligned pointer */
76 
77 /* Flags intentionaly not handled, since they're not needed:
78  * #define DW_EH_PE_indirect      0x80
79  * #define DW_EH_PE_uleb128       0x01
80  * #define DW_EH_PE_udata2        0x02
81  * #define DW_EH_PE_sleb128       0x09
82  * #define DW_EH_PE_sdata2        0x0a
83  * #define DW_EH_PE_textrel       0x20
84  * #define DW_EH_PE_datarel       0x30
85  */
86 
87 struct unwind_info {
88 	struct perf_sample	*sample;
89 	struct machine		*machine;
90 	struct thread		*thread;
91 };
92 
93 #define dw_read(ptr, type, end) ({	\
94 	type *__p = (type *) ptr;	\
95 	type  __v;			\
96 	if ((__p + 1) > (type *) end)	\
97 		return -EINVAL;		\
98 	__v = *__p++;			\
99 	ptr = (typeof(ptr)) __p;	\
100 	__v;				\
101 	})
102 
103 static int __dw_read_encoded_value(u8 **p, u8 *end, u64 *val,
104 				   u8 encoding)
105 {
106 	u8 *cur = *p;
107 	*val = 0;
108 
109 	switch (encoding) {
110 	case DW_EH_PE_omit:
111 		*val = 0;
112 		goto out;
113 	case DW_EH_PE_ptr:
114 		*val = dw_read(cur, unsigned long, end);
115 		goto out;
116 	default:
117 		break;
118 	}
119 
120 	switch (encoding & DW_EH_PE_APPL_MASK) {
121 	case DW_EH_PE_absptr:
122 		break;
123 	case DW_EH_PE_pcrel:
124 		*val = (unsigned long) cur;
125 		break;
126 	default:
127 		return -EINVAL;
128 	}
129 
130 	if ((encoding & 0x07) == 0x00)
131 		encoding |= DW_EH_PE_udata4;
132 
133 	switch (encoding & DW_EH_PE_FORMAT_MASK) {
134 	case DW_EH_PE_sdata4:
135 		*val += dw_read(cur, s32, end);
136 		break;
137 	case DW_EH_PE_udata4:
138 		*val += dw_read(cur, u32, end);
139 		break;
140 	case DW_EH_PE_sdata8:
141 		*val += dw_read(cur, s64, end);
142 		break;
143 	case DW_EH_PE_udata8:
144 		*val += dw_read(cur, u64, end);
145 		break;
146 	default:
147 		return -EINVAL;
148 	}
149 
150  out:
151 	*p = cur;
152 	return 0;
153 }
154 
155 #define dw_read_encoded_value(ptr, end, enc) ({			\
156 	u64 __v;						\
157 	if (__dw_read_encoded_value(&ptr, end, &__v, enc)) {	\
158 		return -EINVAL;                                 \
159 	}                                                       \
160 	__v;                                                    \
161 	})
162 
163 static u64 elf_section_offset(int fd, const char *name)
164 {
165 	Elf *elf;
166 	GElf_Ehdr ehdr;
167 	GElf_Shdr shdr;
168 	u64 offset = 0;
169 
170 	elf = elf_begin(fd, PERF_ELF_C_READ_MMAP, NULL);
171 	if (elf == NULL)
172 		return 0;
173 
174 	do {
175 		if (gelf_getehdr(elf, &ehdr) == NULL)
176 			break;
177 
178 		if (!elf_section_by_name(elf, &ehdr, &shdr, name, NULL))
179 			break;
180 
181 		offset = shdr.sh_offset;
182 	} while (0);
183 
184 	elf_end(elf);
185 	return offset;
186 }
187 
188 #ifndef NO_LIBUNWIND_DEBUG_FRAME
189 static int elf_is_exec(int fd, const char *name)
190 {
191 	Elf *elf;
192 	GElf_Ehdr ehdr;
193 	int retval = 0;
194 
195 	elf = elf_begin(fd, PERF_ELF_C_READ_MMAP, NULL);
196 	if (elf == NULL)
197 		return 0;
198 	if (gelf_getehdr(elf, &ehdr) == NULL)
199 		goto out;
200 
201 	retval = (ehdr.e_type == ET_EXEC);
202 
203 out:
204 	elf_end(elf);
205 	pr_debug("unwind: elf_is_exec(%s): %d\n", name, retval);
206 	return retval;
207 }
208 #endif
209 
210 struct table_entry {
211 	u32 start_ip_offset;
212 	u32 fde_offset;
213 };
214 
215 struct eh_frame_hdr {
216 	unsigned char version;
217 	unsigned char eh_frame_ptr_enc;
218 	unsigned char fde_count_enc;
219 	unsigned char table_enc;
220 
221 	/*
222 	 * The rest of the header is variable-length and consists of the
223 	 * following members:
224 	 *
225 	 *	encoded_t eh_frame_ptr;
226 	 *	encoded_t fde_count;
227 	 */
228 
229 	/* A single encoded pointer should not be more than 8 bytes. */
230 	u64 enc[2];
231 
232 	/*
233 	 * struct {
234 	 *    encoded_t start_ip;
235 	 *    encoded_t fde_addr;
236 	 * } binary_search_table[fde_count];
237 	 */
238 	char data[0];
239 } __packed;
240 
241 static int unwind_spec_ehframe(struct dso *dso, struct machine *machine,
242 			       u64 offset, u64 *table_data, u64 *segbase,
243 			       u64 *fde_count)
244 {
245 	struct eh_frame_hdr hdr;
246 	u8 *enc = (u8 *) &hdr.enc;
247 	u8 *end = (u8 *) &hdr.data;
248 	ssize_t r;
249 
250 	r = dso__data_read_offset(dso, machine, offset,
251 				  (u8 *) &hdr, sizeof(hdr));
252 	if (r != sizeof(hdr))
253 		return -EINVAL;
254 
255 	/* We dont need eh_frame_ptr, just skip it. */
256 	dw_read_encoded_value(enc, end, hdr.eh_frame_ptr_enc);
257 
258 	*fde_count  = dw_read_encoded_value(enc, end, hdr.fde_count_enc);
259 	*segbase    = offset;
260 	*table_data = (enc - (u8 *) &hdr) + offset;
261 	return 0;
262 }
263 
264 static int read_unwind_spec_eh_frame(struct dso *dso, struct machine *machine,
265 				     u64 *table_data, u64 *segbase,
266 				     u64 *fde_count)
267 {
268 	int ret = -EINVAL, fd;
269 	u64 offset = dso->data.eh_frame_hdr_offset;
270 
271 	if (offset == 0) {
272 		fd = dso__data_get_fd(dso, machine);
273 		if (fd < 0)
274 			return -EINVAL;
275 
276 		/* Check the .eh_frame section for unwinding info */
277 		offset = elf_section_offset(fd, ".eh_frame_hdr");
278 		dso->data.eh_frame_hdr_offset = offset;
279 		dso__data_put_fd(dso);
280 	}
281 
282 	if (offset)
283 		ret = unwind_spec_ehframe(dso, machine, offset,
284 					  table_data, segbase,
285 					  fde_count);
286 
287 	return ret;
288 }
289 
290 #ifndef NO_LIBUNWIND_DEBUG_FRAME
291 static int read_unwind_spec_debug_frame(struct dso *dso,
292 					struct machine *machine, u64 *offset)
293 {
294 	int fd;
295 	u64 ofs = dso->data.debug_frame_offset;
296 
297 	if (ofs == 0) {
298 		fd = dso__data_get_fd(dso, machine);
299 		if (fd < 0)
300 			return -EINVAL;
301 
302 		/* Check the .debug_frame section for unwinding info */
303 		ofs = elf_section_offset(fd, ".debug_frame");
304 		dso->data.debug_frame_offset = ofs;
305 		dso__data_put_fd(dso);
306 	}
307 
308 	*offset = ofs;
309 	if (*offset)
310 		return 0;
311 
312 	return -EINVAL;
313 }
314 #endif
315 
316 static struct map *find_map(unw_word_t ip, struct unwind_info *ui)
317 {
318 	struct addr_location al;
319 
320 	thread__find_addr_map(ui->thread, PERF_RECORD_MISC_USER,
321 			      MAP__FUNCTION, ip, &al);
322 	return al.map;
323 }
324 
325 static int
326 find_proc_info(unw_addr_space_t as, unw_word_t ip, unw_proc_info_t *pi,
327 	       int need_unwind_info, void *arg)
328 {
329 	struct unwind_info *ui = arg;
330 	struct map *map;
331 	unw_dyn_info_t di;
332 	u64 table_data, segbase, fde_count;
333 	int ret = -EINVAL;
334 
335 	map = find_map(ip, ui);
336 	if (!map || !map->dso)
337 		return -EINVAL;
338 
339 	pr_debug("unwind: find_proc_info dso %s\n", map->dso->name);
340 
341 	/* Check the .eh_frame section for unwinding info */
342 	if (!read_unwind_spec_eh_frame(map->dso, ui->machine,
343 				       &table_data, &segbase, &fde_count)) {
344 		memset(&di, 0, sizeof(di));
345 		di.format   = UNW_INFO_FORMAT_REMOTE_TABLE;
346 		di.start_ip = map->start;
347 		di.end_ip   = map->end;
348 		di.u.rti.segbase    = map->start + segbase;
349 		di.u.rti.table_data = map->start + table_data;
350 		di.u.rti.table_len  = fde_count * sizeof(struct table_entry)
351 				      / sizeof(unw_word_t);
352 		ret = dwarf_search_unwind_table(as, ip, &di, pi,
353 						need_unwind_info, arg);
354 	}
355 
356 #ifndef NO_LIBUNWIND_DEBUG_FRAME
357 	/* Check the .debug_frame section for unwinding info */
358 	if (ret < 0 &&
359 	    !read_unwind_spec_debug_frame(map->dso, ui->machine, &segbase)) {
360 		int fd = dso__data_get_fd(map->dso, ui->machine);
361 		int is_exec = elf_is_exec(fd, map->dso->name);
362 		unw_word_t base = is_exec ? 0 : map->start;
363 		const char *symfile;
364 
365 		if (fd >= 0)
366 			dso__data_put_fd(map->dso);
367 
368 		symfile = map->dso->symsrc_filename ?: map->dso->name;
369 
370 		memset(&di, 0, sizeof(di));
371 		if (dwarf_find_debug_frame(0, &di, ip, base, symfile,
372 					   map->start, map->end))
373 			return dwarf_search_unwind_table(as, ip, &di, pi,
374 							 need_unwind_info, arg);
375 	}
376 #endif
377 
378 	return ret;
379 }
380 
381 static int access_fpreg(unw_addr_space_t __maybe_unused as,
382 			unw_regnum_t __maybe_unused num,
383 			unw_fpreg_t __maybe_unused *val,
384 			int __maybe_unused __write,
385 			void __maybe_unused *arg)
386 {
387 	pr_err("unwind: access_fpreg unsupported\n");
388 	return -UNW_EINVAL;
389 }
390 
391 static int get_dyn_info_list_addr(unw_addr_space_t __maybe_unused as,
392 				  unw_word_t __maybe_unused *dil_addr,
393 				  void __maybe_unused *arg)
394 {
395 	return -UNW_ENOINFO;
396 }
397 
398 static int resume(unw_addr_space_t __maybe_unused as,
399 		  unw_cursor_t __maybe_unused *cu,
400 		  void __maybe_unused *arg)
401 {
402 	pr_err("unwind: resume unsupported\n");
403 	return -UNW_EINVAL;
404 }
405 
406 static int
407 get_proc_name(unw_addr_space_t __maybe_unused as,
408 	      unw_word_t __maybe_unused addr,
409 		char __maybe_unused *bufp, size_t __maybe_unused buf_len,
410 		unw_word_t __maybe_unused *offp, void __maybe_unused *arg)
411 {
412 	pr_err("unwind: get_proc_name unsupported\n");
413 	return -UNW_EINVAL;
414 }
415 
416 static int access_dso_mem(struct unwind_info *ui, unw_word_t addr,
417 			  unw_word_t *data)
418 {
419 	struct addr_location al;
420 	ssize_t size;
421 
422 	thread__find_addr_map(ui->thread, PERF_RECORD_MISC_USER,
423 			      MAP__FUNCTION, addr, &al);
424 	if (!al.map) {
425 		pr_debug("unwind: no map for %lx\n", (unsigned long)addr);
426 		return -1;
427 	}
428 
429 	if (!al.map->dso)
430 		return -1;
431 
432 	size = dso__data_read_addr(al.map->dso, al.map, ui->machine,
433 				   addr, (u8 *) data, sizeof(*data));
434 
435 	return !(size == sizeof(*data));
436 }
437 
438 static int access_mem(unw_addr_space_t __maybe_unused as,
439 		      unw_word_t addr, unw_word_t *valp,
440 		      int __write, void *arg)
441 {
442 	struct unwind_info *ui = arg;
443 	struct stack_dump *stack = &ui->sample->user_stack;
444 	u64 start, end;
445 	int offset;
446 	int ret;
447 
448 	/* Don't support write, probably not needed. */
449 	if (__write || !stack || !ui->sample->user_regs.regs) {
450 		*valp = 0;
451 		return 0;
452 	}
453 
454 	ret = perf_reg_value(&start, &ui->sample->user_regs, PERF_REG_SP);
455 	if (ret)
456 		return ret;
457 
458 	end = start + stack->size;
459 
460 	/* Check overflow. */
461 	if (addr + sizeof(unw_word_t) < addr)
462 		return -EINVAL;
463 
464 	if (addr < start || addr + sizeof(unw_word_t) >= end) {
465 		ret = access_dso_mem(ui, addr, valp);
466 		if (ret) {
467 			pr_debug("unwind: access_mem %p not inside range"
468 				 " 0x%" PRIx64 "-0x%" PRIx64 "\n",
469 				 (void *) (uintptr_t) addr, start, end);
470 			*valp = 0;
471 			return ret;
472 		}
473 		return 0;
474 	}
475 
476 	offset = addr - start;
477 	*valp  = *(unw_word_t *)&stack->data[offset];
478 	pr_debug("unwind: access_mem addr %p val %lx, offset %d\n",
479 		 (void *) (uintptr_t) addr, (unsigned long)*valp, offset);
480 	return 0;
481 }
482 
483 static int access_reg(unw_addr_space_t __maybe_unused as,
484 		      unw_regnum_t regnum, unw_word_t *valp,
485 		      int __write, void *arg)
486 {
487 	struct unwind_info *ui = arg;
488 	int id, ret;
489 	u64 val;
490 
491 	/* Don't support write, I suspect we don't need it. */
492 	if (__write) {
493 		pr_err("unwind: access_reg w %d\n", regnum);
494 		return 0;
495 	}
496 
497 	if (!ui->sample->user_regs.regs) {
498 		*valp = 0;
499 		return 0;
500 	}
501 
502 	id = libunwind__arch_reg_id(regnum);
503 	if (id < 0)
504 		return -EINVAL;
505 
506 	ret = perf_reg_value(&val, &ui->sample->user_regs, id);
507 	if (ret) {
508 		pr_err("unwind: can't read reg %d\n", regnum);
509 		return ret;
510 	}
511 
512 	*valp = (unw_word_t) val;
513 	pr_debug("unwind: reg %d, val %lx\n", regnum, (unsigned long)*valp);
514 	return 0;
515 }
516 
517 static void put_unwind_info(unw_addr_space_t __maybe_unused as,
518 			    unw_proc_info_t *pi __maybe_unused,
519 			    void *arg __maybe_unused)
520 {
521 	pr_debug("unwind: put_unwind_info called\n");
522 }
523 
524 static int entry(u64 ip, struct thread *thread,
525 		 unwind_entry_cb_t cb, void *arg)
526 {
527 	struct unwind_entry e;
528 	struct addr_location al;
529 
530 	thread__find_addr_location(thread, PERF_RECORD_MISC_USER,
531 				   MAP__FUNCTION, ip, &al);
532 
533 	e.ip = ip;
534 	e.map = al.map;
535 	e.sym = al.sym;
536 
537 	pr_debug("unwind: %s:ip = 0x%" PRIx64 " (0x%" PRIx64 ")\n",
538 		 al.sym ? al.sym->name : "''",
539 		 ip,
540 		 al.map ? al.map->map_ip(al.map, ip) : (u64) 0);
541 
542 	return cb(&e, arg);
543 }
544 
545 static void display_error(int err)
546 {
547 	switch (err) {
548 	case UNW_EINVAL:
549 		pr_err("unwind: Only supports local.\n");
550 		break;
551 	case UNW_EUNSPEC:
552 		pr_err("unwind: Unspecified error.\n");
553 		break;
554 	case UNW_EBADREG:
555 		pr_err("unwind: Register unavailable.\n");
556 		break;
557 	default:
558 		break;
559 	}
560 }
561 
562 static unw_accessors_t accessors = {
563 	.find_proc_info		= find_proc_info,
564 	.put_unwind_info	= put_unwind_info,
565 	.get_dyn_info_list_addr	= get_dyn_info_list_addr,
566 	.access_mem		= access_mem,
567 	.access_reg		= access_reg,
568 	.access_fpreg		= access_fpreg,
569 	.resume			= resume,
570 	.get_proc_name		= get_proc_name,
571 };
572 
573 int unwind__prepare_access(struct thread *thread)
574 {
575 	unw_addr_space_t addr_space;
576 
577 	if (callchain_param.record_mode != CALLCHAIN_DWARF)
578 		return 0;
579 
580 	addr_space = unw_create_addr_space(&accessors, 0);
581 	if (!addr_space) {
582 		pr_err("unwind: Can't create unwind address space.\n");
583 		return -ENOMEM;
584 	}
585 
586 	unw_set_caching_policy(addr_space, UNW_CACHE_GLOBAL);
587 	thread__set_priv(thread, addr_space);
588 
589 	return 0;
590 }
591 
592 void unwind__flush_access(struct thread *thread)
593 {
594 	unw_addr_space_t addr_space;
595 
596 	if (callchain_param.record_mode != CALLCHAIN_DWARF)
597 		return;
598 
599 	addr_space = thread__priv(thread);
600 	unw_flush_cache(addr_space, 0, 0);
601 }
602 
603 void unwind__finish_access(struct thread *thread)
604 {
605 	unw_addr_space_t addr_space;
606 
607 	if (callchain_param.record_mode != CALLCHAIN_DWARF)
608 		return;
609 
610 	addr_space = thread__priv(thread);
611 	unw_destroy_addr_space(addr_space);
612 }
613 
614 static int get_entries(struct unwind_info *ui, unwind_entry_cb_t cb,
615 		       void *arg, int max_stack)
616 {
617 	u64 val;
618 	unw_word_t ips[max_stack];
619 	unw_addr_space_t addr_space;
620 	unw_cursor_t c;
621 	int ret, i = 0;
622 
623 	ret = perf_reg_value(&val, &ui->sample->user_regs, PERF_REG_IP);
624 	if (ret)
625 		return ret;
626 
627 	ips[i++] = (unw_word_t) val;
628 
629 	/*
630 	 * If we need more than one entry, do the DWARF
631 	 * unwind itself.
632 	 */
633 	if (max_stack - 1 > 0) {
634 		addr_space = thread__priv(ui->thread);
635 		if (addr_space == NULL)
636 			return -1;
637 
638 		ret = unw_init_remote(&c, addr_space, ui);
639 		if (ret)
640 			display_error(ret);
641 
642 		while (!ret && (unw_step(&c) > 0) && i < max_stack) {
643 			unw_get_reg(&c, UNW_REG_IP, &ips[i]);
644 			++i;
645 		}
646 
647 		max_stack = i;
648 	}
649 
650 	/*
651 	 * Display what we got based on the order setup.
652 	 */
653 	for (i = 0; i < max_stack && !ret; i++) {
654 		int j = i;
655 
656 		if (callchain_param.order == ORDER_CALLER)
657 			j = max_stack - i - 1;
658 		ret = ips[j] ? entry(ips[j], ui->thread, cb, arg) : 0;
659 	}
660 
661 	return ret;
662 }
663 
664 int unwind__get_entries(unwind_entry_cb_t cb, void *arg,
665 			struct thread *thread,
666 			struct perf_sample *data, int max_stack)
667 {
668 	struct unwind_info ui = {
669 		.sample       = data,
670 		.thread       = thread,
671 		.machine      = thread->mg->machine,
672 	};
673 
674 	if (!data->user_regs.regs)
675 		return -EINVAL;
676 
677 	if (max_stack <= 0)
678 		return -EINVAL;
679 
680 	return get_entries(&ui, cb, arg, max_stack);
681 }
682