1 /*
2 * Copyright (C) 2009 Matt Fleming <matt@console-pimps.org>
3 *
4 * This file is subject to the terms and conditions of the GNU General Public
5 * License. See the file "COPYING" in the main directory of this archive
6 * for more details.
7 *
8 * This is an implementation of a DWARF unwinder. Its main purpose is
9 * for generating stacktrace information. Based on the DWARF 3
10 * specification from http://www.dwarfstd.org.
11 *
12 * TODO:
13 * - DWARF64 doesn't work.
14 * - Registers with DWARF_VAL_OFFSET rules aren't handled properly.
15 */
16
17 /* #define DEBUG */
18 #include <linux/kernel.h>
19 #include <linux/io.h>
20 #include <linux/list.h>
21 #include <linux/mempool.h>
22 #include <linux/mm.h>
23 #include <linux/elf.h>
24 #include <linux/ftrace.h>
25 #include <linux/module.h>
26 #include <linux/slab.h>
27 #include <asm/dwarf.h>
28 #include <asm/unwinder.h>
29 #include <asm/sections.h>
30 #include <asm/unaligned.h>
31 #include <asm/stacktrace.h>
32
33 /* Reserve enough memory for two stack frames */
34 #define DWARF_FRAME_MIN_REQ 2
35 /* ... with 4 registers per frame. */
36 #define DWARF_REG_MIN_REQ (DWARF_FRAME_MIN_REQ * 4)
37
38 static struct kmem_cache *dwarf_frame_cachep;
39 static mempool_t *dwarf_frame_pool;
40
41 static struct kmem_cache *dwarf_reg_cachep;
42 static mempool_t *dwarf_reg_pool;
43
44 static struct rb_root cie_root;
45 static DEFINE_SPINLOCK(dwarf_cie_lock);
46
47 static struct rb_root fde_root;
48 static DEFINE_SPINLOCK(dwarf_fde_lock);
49
50 static struct dwarf_cie *cached_cie;
51
52 static unsigned int dwarf_unwinder_ready;
53
54 /**
55 * dwarf_frame_alloc_reg - allocate memory for a DWARF register
56 * @frame: the DWARF frame whose list of registers we insert on
57 * @reg_num: the register number
58 *
59 * Allocate space for, and initialise, a dwarf reg from
60 * dwarf_reg_pool and insert it onto the (unsorted) linked-list of
61 * dwarf registers for @frame.
62 *
63 * Return the initialised DWARF reg.
64 */
dwarf_frame_alloc_reg(struct dwarf_frame * frame,unsigned int reg_num)65 static struct dwarf_reg *dwarf_frame_alloc_reg(struct dwarf_frame *frame,
66 unsigned int reg_num)
67 {
68 struct dwarf_reg *reg;
69
70 reg = mempool_alloc(dwarf_reg_pool, GFP_ATOMIC);
71 if (!reg) {
72 printk(KERN_WARNING "Unable to allocate a DWARF register\n");
73 /*
74 * Let's just bomb hard here, we have no way to
75 * gracefully recover.
76 */
77 UNWINDER_BUG();
78 }
79
80 reg->number = reg_num;
81 reg->addr = 0;
82 reg->flags = 0;
83
84 list_add(®->link, &frame->reg_list);
85
86 return reg;
87 }
88
dwarf_frame_free_regs(struct dwarf_frame * frame)89 static void dwarf_frame_free_regs(struct dwarf_frame *frame)
90 {
91 struct dwarf_reg *reg, *n;
92
93 list_for_each_entry_safe(reg, n, &frame->reg_list, link) {
94 list_del(®->link);
95 mempool_free(reg, dwarf_reg_pool);
96 }
97 }
98
99 /**
100 * dwarf_frame_reg - return a DWARF register
101 * @frame: the DWARF frame to search in for @reg_num
102 * @reg_num: the register number to search for
103 *
104 * Lookup and return the dwarf reg @reg_num for this frame. Return
105 * NULL if @reg_num is an register invalid number.
106 */
dwarf_frame_reg(struct dwarf_frame * frame,unsigned int reg_num)107 static struct dwarf_reg *dwarf_frame_reg(struct dwarf_frame *frame,
108 unsigned int reg_num)
109 {
110 struct dwarf_reg *reg;
111
112 list_for_each_entry(reg, &frame->reg_list, link) {
113 if (reg->number == reg_num)
114 return reg;
115 }
116
117 return NULL;
118 }
119
120 /**
121 * dwarf_read_addr - read dwarf data
122 * @src: source address of data
123 * @dst: destination address to store the data to
124 *
125 * Read 'n' bytes from @src, where 'n' is the size of an address on
126 * the native machine. We return the number of bytes read, which
127 * should always be 'n'. We also have to be careful when reading
128 * from @src and writing to @dst, because they can be arbitrarily
129 * aligned. Return 'n' - the number of bytes read.
130 */
dwarf_read_addr(unsigned long * src,unsigned long * dst)131 static inline int dwarf_read_addr(unsigned long *src, unsigned long *dst)
132 {
133 u32 val = get_unaligned(src);
134 put_unaligned(val, dst);
135 return sizeof(unsigned long *);
136 }
137
138 /**
139 * dwarf_read_uleb128 - read unsigned LEB128 data
140 * @addr: the address where the ULEB128 data is stored
141 * @ret: address to store the result
142 *
143 * Decode an unsigned LEB128 encoded datum. The algorithm is taken
144 * from Appendix C of the DWARF 3 spec. For information on the
145 * encodings refer to section "7.6 - Variable Length Data". Return
146 * the number of bytes read.
147 */
dwarf_read_uleb128(char * addr,unsigned int * ret)148 static inline unsigned long dwarf_read_uleb128(char *addr, unsigned int *ret)
149 {
150 unsigned int result;
151 unsigned char byte;
152 int shift, count;
153
154 result = 0;
155 shift = 0;
156 count = 0;
157
158 while (1) {
159 byte = __raw_readb(addr);
160 addr++;
161 count++;
162
163 result |= (byte & 0x7f) << shift;
164 shift += 7;
165
166 if (!(byte & 0x80))
167 break;
168 }
169
170 *ret = result;
171
172 return count;
173 }
174
175 /**
176 * dwarf_read_leb128 - read signed LEB128 data
177 * @addr: the address of the LEB128 encoded data
178 * @ret: address to store the result
179 *
180 * Decode signed LEB128 data. The algorithm is taken from Appendix
181 * C of the DWARF 3 spec. Return the number of bytes read.
182 */
dwarf_read_leb128(char * addr,int * ret)183 static inline unsigned long dwarf_read_leb128(char *addr, int *ret)
184 {
185 unsigned char byte;
186 int result, shift;
187 int num_bits;
188 int count;
189
190 result = 0;
191 shift = 0;
192 count = 0;
193
194 while (1) {
195 byte = __raw_readb(addr);
196 addr++;
197 result |= (byte & 0x7f) << shift;
198 shift += 7;
199 count++;
200
201 if (!(byte & 0x80))
202 break;
203 }
204
205 /* The number of bits in a signed integer. */
206 num_bits = 8 * sizeof(result);
207
208 if ((shift < num_bits) && (byte & 0x40))
209 result |= (-1 << shift);
210
211 *ret = result;
212
213 return count;
214 }
215
216 /**
217 * dwarf_read_encoded_value - return the decoded value at @addr
218 * @addr: the address of the encoded value
219 * @val: where to write the decoded value
220 * @encoding: the encoding with which we can decode @addr
221 *
222 * GCC emits encoded address in the .eh_frame FDE entries. Decode
223 * the value at @addr using @encoding. The decoded value is written
224 * to @val and the number of bytes read is returned.
225 */
dwarf_read_encoded_value(char * addr,unsigned long * val,char encoding)226 static int dwarf_read_encoded_value(char *addr, unsigned long *val,
227 char encoding)
228 {
229 unsigned long decoded_addr = 0;
230 int count = 0;
231
232 switch (encoding & 0x70) {
233 case DW_EH_PE_absptr:
234 break;
235 case DW_EH_PE_pcrel:
236 decoded_addr = (unsigned long)addr;
237 break;
238 default:
239 pr_debug("encoding=0x%x\n", (encoding & 0x70));
240 UNWINDER_BUG();
241 }
242
243 if ((encoding & 0x07) == 0x00)
244 encoding |= DW_EH_PE_udata4;
245
246 switch (encoding & 0x0f) {
247 case DW_EH_PE_sdata4:
248 case DW_EH_PE_udata4:
249 count += 4;
250 decoded_addr += get_unaligned((u32 *)addr);
251 __raw_writel(decoded_addr, val);
252 break;
253 default:
254 pr_debug("encoding=0x%x\n", encoding);
255 UNWINDER_BUG();
256 }
257
258 return count;
259 }
260
261 /**
262 * dwarf_entry_len - return the length of an FDE or CIE
263 * @addr: the address of the entry
264 * @len: the length of the entry
265 *
266 * Read the initial_length field of the entry and store the size of
267 * the entry in @len. We return the number of bytes read. Return a
268 * count of 0 on error.
269 */
dwarf_entry_len(char * addr,unsigned long * len)270 static inline int dwarf_entry_len(char *addr, unsigned long *len)
271 {
272 u32 initial_len;
273 int count;
274
275 initial_len = get_unaligned((u32 *)addr);
276 count = 4;
277
278 /*
279 * An initial length field value in the range DW_LEN_EXT_LO -
280 * DW_LEN_EXT_HI indicates an extension, and should not be
281 * interpreted as a length. The only extension that we currently
282 * understand is the use of DWARF64 addresses.
283 */
284 if (initial_len >= DW_EXT_LO && initial_len <= DW_EXT_HI) {
285 /*
286 * The 64-bit length field immediately follows the
287 * compulsory 32-bit length field.
288 */
289 if (initial_len == DW_EXT_DWARF64) {
290 *len = get_unaligned((u64 *)addr + 4);
291 count = 12;
292 } else {
293 printk(KERN_WARNING "Unknown DWARF extension\n");
294 count = 0;
295 }
296 } else
297 *len = initial_len;
298
299 return count;
300 }
301
302 /**
303 * dwarf_lookup_cie - locate the cie
304 * @cie_ptr: pointer to help with lookup
305 */
dwarf_lookup_cie(unsigned long cie_ptr)306 static struct dwarf_cie *dwarf_lookup_cie(unsigned long cie_ptr)
307 {
308 struct rb_node **rb_node = &cie_root.rb_node;
309 struct dwarf_cie *cie = NULL;
310 unsigned long flags;
311
312 spin_lock_irqsave(&dwarf_cie_lock, flags);
313
314 /*
315 * We've cached the last CIE we looked up because chances are
316 * that the FDE wants this CIE.
317 */
318 if (cached_cie && cached_cie->cie_pointer == cie_ptr) {
319 cie = cached_cie;
320 goto out;
321 }
322
323 while (*rb_node) {
324 struct dwarf_cie *cie_tmp;
325
326 cie_tmp = rb_entry(*rb_node, struct dwarf_cie, node);
327 BUG_ON(!cie_tmp);
328
329 if (cie_ptr == cie_tmp->cie_pointer) {
330 cie = cie_tmp;
331 cached_cie = cie_tmp;
332 goto out;
333 } else {
334 if (cie_ptr < cie_tmp->cie_pointer)
335 rb_node = &(*rb_node)->rb_left;
336 else
337 rb_node = &(*rb_node)->rb_right;
338 }
339 }
340
341 out:
342 spin_unlock_irqrestore(&dwarf_cie_lock, flags);
343 return cie;
344 }
345
346 /**
347 * dwarf_lookup_fde - locate the FDE that covers pc
348 * @pc: the program counter
349 */
dwarf_lookup_fde(unsigned long pc)350 struct dwarf_fde *dwarf_lookup_fde(unsigned long pc)
351 {
352 struct rb_node **rb_node = &fde_root.rb_node;
353 struct dwarf_fde *fde = NULL;
354 unsigned long flags;
355
356 spin_lock_irqsave(&dwarf_fde_lock, flags);
357
358 while (*rb_node) {
359 struct dwarf_fde *fde_tmp;
360 unsigned long tmp_start, tmp_end;
361
362 fde_tmp = rb_entry(*rb_node, struct dwarf_fde, node);
363 BUG_ON(!fde_tmp);
364
365 tmp_start = fde_tmp->initial_location;
366 tmp_end = fde_tmp->initial_location + fde_tmp->address_range;
367
368 if (pc < tmp_start) {
369 rb_node = &(*rb_node)->rb_left;
370 } else {
371 if (pc < tmp_end) {
372 fde = fde_tmp;
373 goto out;
374 } else
375 rb_node = &(*rb_node)->rb_right;
376 }
377 }
378
379 out:
380 spin_unlock_irqrestore(&dwarf_fde_lock, flags);
381
382 return fde;
383 }
384
385 /**
386 * dwarf_cfa_execute_insns - execute instructions to calculate a CFA
387 * @insn_start: address of the first instruction
388 * @insn_end: address of the last instruction
389 * @cie: the CIE for this function
390 * @fde: the FDE for this function
391 * @frame: the instructions calculate the CFA for this frame
392 * @pc: the program counter of the address we're interested in
393 *
394 * Execute the Call Frame instruction sequence starting at
395 * @insn_start and ending at @insn_end. The instructions describe
396 * how to calculate the Canonical Frame Address of a stackframe.
397 * Store the results in @frame.
398 */
dwarf_cfa_execute_insns(unsigned char * insn_start,unsigned char * insn_end,struct dwarf_cie * cie,struct dwarf_fde * fde,struct dwarf_frame * frame,unsigned long pc)399 static int dwarf_cfa_execute_insns(unsigned char *insn_start,
400 unsigned char *insn_end,
401 struct dwarf_cie *cie,
402 struct dwarf_fde *fde,
403 struct dwarf_frame *frame,
404 unsigned long pc)
405 {
406 unsigned char insn;
407 unsigned char *current_insn;
408 unsigned int count, delta, reg, expr_len, offset;
409 struct dwarf_reg *regp;
410
411 current_insn = insn_start;
412
413 while (current_insn < insn_end && frame->pc <= pc) {
414 insn = __raw_readb(current_insn++);
415
416 /*
417 * Firstly, handle the opcodes that embed their operands
418 * in the instructions.
419 */
420 switch (DW_CFA_opcode(insn)) {
421 case DW_CFA_advance_loc:
422 delta = DW_CFA_operand(insn);
423 delta *= cie->code_alignment_factor;
424 frame->pc += delta;
425 continue;
426 /* NOTREACHED */
427 case DW_CFA_offset:
428 reg = DW_CFA_operand(insn);
429 count = dwarf_read_uleb128(current_insn, &offset);
430 current_insn += count;
431 offset *= cie->data_alignment_factor;
432 regp = dwarf_frame_alloc_reg(frame, reg);
433 regp->addr = offset;
434 regp->flags |= DWARF_REG_OFFSET;
435 continue;
436 /* NOTREACHED */
437 case DW_CFA_restore:
438 reg = DW_CFA_operand(insn);
439 continue;
440 /* NOTREACHED */
441 }
442
443 /*
444 * Secondly, handle the opcodes that don't embed their
445 * operands in the instruction.
446 */
447 switch (insn) {
448 case DW_CFA_nop:
449 continue;
450 case DW_CFA_advance_loc1:
451 delta = *current_insn++;
452 frame->pc += delta * cie->code_alignment_factor;
453 break;
454 case DW_CFA_advance_loc2:
455 delta = get_unaligned((u16 *)current_insn);
456 current_insn += 2;
457 frame->pc += delta * cie->code_alignment_factor;
458 break;
459 case DW_CFA_advance_loc4:
460 delta = get_unaligned((u32 *)current_insn);
461 current_insn += 4;
462 frame->pc += delta * cie->code_alignment_factor;
463 break;
464 case DW_CFA_offset_extended:
465 count = dwarf_read_uleb128(current_insn, ®);
466 current_insn += count;
467 count = dwarf_read_uleb128(current_insn, &offset);
468 current_insn += count;
469 offset *= cie->data_alignment_factor;
470 break;
471 case DW_CFA_restore_extended:
472 count = dwarf_read_uleb128(current_insn, ®);
473 current_insn += count;
474 break;
475 case DW_CFA_undefined:
476 count = dwarf_read_uleb128(current_insn, ®);
477 current_insn += count;
478 regp = dwarf_frame_alloc_reg(frame, reg);
479 regp->flags |= DWARF_UNDEFINED;
480 break;
481 case DW_CFA_def_cfa:
482 count = dwarf_read_uleb128(current_insn,
483 &frame->cfa_register);
484 current_insn += count;
485 count = dwarf_read_uleb128(current_insn,
486 &frame->cfa_offset);
487 current_insn += count;
488
489 frame->flags |= DWARF_FRAME_CFA_REG_OFFSET;
490 break;
491 case DW_CFA_def_cfa_register:
492 count = dwarf_read_uleb128(current_insn,
493 &frame->cfa_register);
494 current_insn += count;
495 frame->flags |= DWARF_FRAME_CFA_REG_OFFSET;
496 break;
497 case DW_CFA_def_cfa_offset:
498 count = dwarf_read_uleb128(current_insn, &offset);
499 current_insn += count;
500 frame->cfa_offset = offset;
501 break;
502 case DW_CFA_def_cfa_expression:
503 count = dwarf_read_uleb128(current_insn, &expr_len);
504 current_insn += count;
505
506 frame->cfa_expr = current_insn;
507 frame->cfa_expr_len = expr_len;
508 current_insn += expr_len;
509
510 frame->flags |= DWARF_FRAME_CFA_REG_EXP;
511 break;
512 case DW_CFA_offset_extended_sf:
513 count = dwarf_read_uleb128(current_insn, ®);
514 current_insn += count;
515 count = dwarf_read_leb128(current_insn, &offset);
516 current_insn += count;
517 offset *= cie->data_alignment_factor;
518 regp = dwarf_frame_alloc_reg(frame, reg);
519 regp->flags |= DWARF_REG_OFFSET;
520 regp->addr = offset;
521 break;
522 case DW_CFA_val_offset:
523 count = dwarf_read_uleb128(current_insn, ®);
524 current_insn += count;
525 count = dwarf_read_leb128(current_insn, &offset);
526 offset *= cie->data_alignment_factor;
527 regp = dwarf_frame_alloc_reg(frame, reg);
528 regp->flags |= DWARF_VAL_OFFSET;
529 regp->addr = offset;
530 break;
531 case DW_CFA_GNU_args_size:
532 count = dwarf_read_uleb128(current_insn, &offset);
533 current_insn += count;
534 break;
535 case DW_CFA_GNU_negative_offset_extended:
536 count = dwarf_read_uleb128(current_insn, ®);
537 current_insn += count;
538 count = dwarf_read_uleb128(current_insn, &offset);
539 offset *= cie->data_alignment_factor;
540
541 regp = dwarf_frame_alloc_reg(frame, reg);
542 regp->flags |= DWARF_REG_OFFSET;
543 regp->addr = -offset;
544 break;
545 default:
546 pr_debug("unhandled DWARF instruction 0x%x\n", insn);
547 UNWINDER_BUG();
548 break;
549 }
550 }
551
552 return 0;
553 }
554
555 /**
556 * dwarf_free_frame - free the memory allocated for @frame
557 * @frame: the frame to free
558 */
dwarf_free_frame(struct dwarf_frame * frame)559 void dwarf_free_frame(struct dwarf_frame *frame)
560 {
561 dwarf_frame_free_regs(frame);
562 mempool_free(frame, dwarf_frame_pool);
563 }
564
565 extern void ret_from_irq(void);
566
567 /**
568 * dwarf_unwind_stack - unwind the stack
569 *
570 * @pc: address of the function to unwind
571 * @prev: struct dwarf_frame of the previous stackframe on the callstack
572 *
573 * Return a struct dwarf_frame representing the most recent frame
574 * on the callstack. Each of the lower (older) stack frames are
575 * linked via the "prev" member.
576 */
dwarf_unwind_stack(unsigned long pc,struct dwarf_frame * prev)577 struct dwarf_frame *dwarf_unwind_stack(unsigned long pc,
578 struct dwarf_frame *prev)
579 {
580 struct dwarf_frame *frame;
581 struct dwarf_cie *cie;
582 struct dwarf_fde *fde;
583 struct dwarf_reg *reg;
584 unsigned long addr;
585
586 /*
587 * If we've been called in to before initialization has
588 * completed, bail out immediately.
589 */
590 if (!dwarf_unwinder_ready)
591 return NULL;
592
593 /*
594 * If we're starting at the top of the stack we need get the
595 * contents of a physical register to get the CFA in order to
596 * begin the virtual unwinding of the stack.
597 *
598 * NOTE: the return address is guaranteed to be setup by the
599 * time this function makes its first function call.
600 */
601 if (!pc || !prev)
602 pc = (unsigned long)current_text_addr();
603
604 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
605 /*
606 * If our stack has been patched by the function graph tracer
607 * then we might see the address of return_to_handler() where we
608 * expected to find the real return address.
609 */
610 if (pc == (unsigned long)&return_to_handler) {
611 int index = current->curr_ret_stack;
612
613 /*
614 * We currently have no way of tracking how many
615 * return_to_handler()'s we've seen. If there is more
616 * than one patched return address on our stack,
617 * complain loudly.
618 */
619 WARN_ON(index > 0);
620
621 pc = current->ret_stack[index].ret;
622 }
623 #endif
624
625 frame = mempool_alloc(dwarf_frame_pool, GFP_ATOMIC);
626 if (!frame) {
627 printk(KERN_ERR "Unable to allocate a dwarf frame\n");
628 UNWINDER_BUG();
629 }
630
631 INIT_LIST_HEAD(&frame->reg_list);
632 frame->flags = 0;
633 frame->prev = prev;
634 frame->return_addr = 0;
635
636 fde = dwarf_lookup_fde(pc);
637 if (!fde) {
638 /*
639 * This is our normal exit path. There are two reasons
640 * why we might exit here,
641 *
642 * a) pc has no asscociated DWARF frame info and so
643 * we don't know how to unwind this frame. This is
644 * usually the case when we're trying to unwind a
645 * frame that was called from some assembly code
646 * that has no DWARF info, e.g. syscalls.
647 *
648 * b) the DEBUG info for pc is bogus. There's
649 * really no way to distinguish this case from the
650 * case above, which sucks because we could print a
651 * warning here.
652 */
653 goto bail;
654 }
655
656 cie = dwarf_lookup_cie(fde->cie_pointer);
657
658 frame->pc = fde->initial_location;
659
660 /* CIE initial instructions */
661 dwarf_cfa_execute_insns(cie->initial_instructions,
662 cie->instructions_end, cie, fde,
663 frame, pc);
664
665 /* FDE instructions */
666 dwarf_cfa_execute_insns(fde->instructions, fde->end, cie,
667 fde, frame, pc);
668
669 /* Calculate the CFA */
670 switch (frame->flags) {
671 case DWARF_FRAME_CFA_REG_OFFSET:
672 if (prev) {
673 reg = dwarf_frame_reg(prev, frame->cfa_register);
674 UNWINDER_BUG_ON(!reg);
675 UNWINDER_BUG_ON(reg->flags != DWARF_REG_OFFSET);
676
677 addr = prev->cfa + reg->addr;
678 frame->cfa = __raw_readl(addr);
679
680 } else {
681 /*
682 * Again, we're starting from the top of the
683 * stack. We need to physically read
684 * the contents of a register in order to get
685 * the Canonical Frame Address for this
686 * function.
687 */
688 frame->cfa = dwarf_read_arch_reg(frame->cfa_register);
689 }
690
691 frame->cfa += frame->cfa_offset;
692 break;
693 default:
694 UNWINDER_BUG();
695 }
696
697 reg = dwarf_frame_reg(frame, DWARF_ARCH_RA_REG);
698
699 /*
700 * If we haven't seen the return address register or the return
701 * address column is undefined then we must assume that this is
702 * the end of the callstack.
703 */
704 if (!reg || reg->flags == DWARF_UNDEFINED)
705 goto bail;
706
707 UNWINDER_BUG_ON(reg->flags != DWARF_REG_OFFSET);
708
709 addr = frame->cfa + reg->addr;
710 frame->return_addr = __raw_readl(addr);
711
712 /*
713 * Ah, the joys of unwinding through interrupts.
714 *
715 * Interrupts are tricky - the DWARF info needs to be _really_
716 * accurate and unfortunately I'm seeing a lot of bogus DWARF
717 * info. For example, I've seen interrupts occur in epilogues
718 * just after the frame pointer (r14) had been restored. The
719 * problem was that the DWARF info claimed that the CFA could be
720 * reached by using the value of the frame pointer before it was
721 * restored.
722 *
723 * So until the compiler can be trusted to produce reliable
724 * DWARF info when it really matters, let's stop unwinding once
725 * we've calculated the function that was interrupted.
726 */
727 if (prev && prev->pc == (unsigned long)ret_from_irq)
728 frame->return_addr = 0;
729
730 return frame;
731
732 bail:
733 dwarf_free_frame(frame);
734 return NULL;
735 }
736
dwarf_parse_cie(void * entry,void * p,unsigned long len,unsigned char * end,struct module * mod)737 static int dwarf_parse_cie(void *entry, void *p, unsigned long len,
738 unsigned char *end, struct module *mod)
739 {
740 struct rb_node **rb_node = &cie_root.rb_node;
741 struct rb_node *parent = *rb_node;
742 struct dwarf_cie *cie;
743 unsigned long flags;
744 int count;
745
746 cie = kzalloc(sizeof(*cie), GFP_KERNEL);
747 if (!cie)
748 return -ENOMEM;
749
750 cie->length = len;
751
752 /*
753 * Record the offset into the .eh_frame section
754 * for this CIE. It allows this CIE to be
755 * quickly and easily looked up from the
756 * corresponding FDE.
757 */
758 cie->cie_pointer = (unsigned long)entry;
759
760 cie->version = *(char *)p++;
761 UNWINDER_BUG_ON(cie->version != 1);
762
763 cie->augmentation = p;
764 p += strlen(cie->augmentation) + 1;
765
766 count = dwarf_read_uleb128(p, &cie->code_alignment_factor);
767 p += count;
768
769 count = dwarf_read_leb128(p, &cie->data_alignment_factor);
770 p += count;
771
772 /*
773 * Which column in the rule table contains the
774 * return address?
775 */
776 if (cie->version == 1) {
777 cie->return_address_reg = __raw_readb(p);
778 p++;
779 } else {
780 count = dwarf_read_uleb128(p, &cie->return_address_reg);
781 p += count;
782 }
783
784 if (cie->augmentation[0] == 'z') {
785 unsigned int length, count;
786 cie->flags |= DWARF_CIE_Z_AUGMENTATION;
787
788 count = dwarf_read_uleb128(p, &length);
789 p += count;
790
791 UNWINDER_BUG_ON((unsigned char *)p > end);
792
793 cie->initial_instructions = p + length;
794 cie->augmentation++;
795 }
796
797 while (*cie->augmentation) {
798 /*
799 * "L" indicates a byte showing how the
800 * LSDA pointer is encoded. Skip it.
801 */
802 if (*cie->augmentation == 'L') {
803 p++;
804 cie->augmentation++;
805 } else if (*cie->augmentation == 'R') {
806 /*
807 * "R" indicates a byte showing
808 * how FDE addresses are
809 * encoded.
810 */
811 cie->encoding = *(char *)p++;
812 cie->augmentation++;
813 } else if (*cie->augmentation == 'P') {
814 /*
815 * "R" indicates a personality
816 * routine in the CIE
817 * augmentation.
818 */
819 UNWINDER_BUG();
820 } else if (*cie->augmentation == 'S') {
821 UNWINDER_BUG();
822 } else {
823 /*
824 * Unknown augmentation. Assume
825 * 'z' augmentation.
826 */
827 p = cie->initial_instructions;
828 UNWINDER_BUG_ON(!p);
829 break;
830 }
831 }
832
833 cie->initial_instructions = p;
834 cie->instructions_end = end;
835
836 /* Add to list */
837 spin_lock_irqsave(&dwarf_cie_lock, flags);
838
839 while (*rb_node) {
840 struct dwarf_cie *cie_tmp;
841
842 cie_tmp = rb_entry(*rb_node, struct dwarf_cie, node);
843
844 parent = *rb_node;
845
846 if (cie->cie_pointer < cie_tmp->cie_pointer)
847 rb_node = &parent->rb_left;
848 else if (cie->cie_pointer >= cie_tmp->cie_pointer)
849 rb_node = &parent->rb_right;
850 else
851 WARN_ON(1);
852 }
853
854 rb_link_node(&cie->node, parent, rb_node);
855 rb_insert_color(&cie->node, &cie_root);
856
857 #ifdef CONFIG_MODULES
858 if (mod != NULL)
859 list_add_tail(&cie->link, &mod->arch.cie_list);
860 #endif
861
862 spin_unlock_irqrestore(&dwarf_cie_lock, flags);
863
864 return 0;
865 }
866
dwarf_parse_fde(void * entry,u32 entry_type,void * start,unsigned long len,unsigned char * end,struct module * mod)867 static int dwarf_parse_fde(void *entry, u32 entry_type,
868 void *start, unsigned long len,
869 unsigned char *end, struct module *mod)
870 {
871 struct rb_node **rb_node = &fde_root.rb_node;
872 struct rb_node *parent = *rb_node;
873 struct dwarf_fde *fde;
874 struct dwarf_cie *cie;
875 unsigned long flags;
876 int count;
877 void *p = start;
878
879 fde = kzalloc(sizeof(*fde), GFP_KERNEL);
880 if (!fde)
881 return -ENOMEM;
882
883 fde->length = len;
884
885 /*
886 * In a .eh_frame section the CIE pointer is the
887 * delta between the address within the FDE
888 */
889 fde->cie_pointer = (unsigned long)(p - entry_type - 4);
890
891 cie = dwarf_lookup_cie(fde->cie_pointer);
892 fde->cie = cie;
893
894 if (cie->encoding)
895 count = dwarf_read_encoded_value(p, &fde->initial_location,
896 cie->encoding);
897 else
898 count = dwarf_read_addr(p, &fde->initial_location);
899
900 p += count;
901
902 if (cie->encoding)
903 count = dwarf_read_encoded_value(p, &fde->address_range,
904 cie->encoding & 0x0f);
905 else
906 count = dwarf_read_addr(p, &fde->address_range);
907
908 p += count;
909
910 if (fde->cie->flags & DWARF_CIE_Z_AUGMENTATION) {
911 unsigned int length;
912 count = dwarf_read_uleb128(p, &length);
913 p += count + length;
914 }
915
916 /* Call frame instructions. */
917 fde->instructions = p;
918 fde->end = end;
919
920 /* Add to list. */
921 spin_lock_irqsave(&dwarf_fde_lock, flags);
922
923 while (*rb_node) {
924 struct dwarf_fde *fde_tmp;
925 unsigned long tmp_start, tmp_end;
926 unsigned long start, end;
927
928 fde_tmp = rb_entry(*rb_node, struct dwarf_fde, node);
929
930 start = fde->initial_location;
931 end = fde->initial_location + fde->address_range;
932
933 tmp_start = fde_tmp->initial_location;
934 tmp_end = fde_tmp->initial_location + fde_tmp->address_range;
935
936 parent = *rb_node;
937
938 if (start < tmp_start)
939 rb_node = &parent->rb_left;
940 else if (start >= tmp_end)
941 rb_node = &parent->rb_right;
942 else
943 WARN_ON(1);
944 }
945
946 rb_link_node(&fde->node, parent, rb_node);
947 rb_insert_color(&fde->node, &fde_root);
948
949 #ifdef CONFIG_MODULES
950 if (mod != NULL)
951 list_add_tail(&fde->link, &mod->arch.fde_list);
952 #endif
953
954 spin_unlock_irqrestore(&dwarf_fde_lock, flags);
955
956 return 0;
957 }
958
dwarf_unwinder_dump(struct task_struct * task,struct pt_regs * regs,unsigned long * sp,const struct stacktrace_ops * ops,void * data)959 static void dwarf_unwinder_dump(struct task_struct *task,
960 struct pt_regs *regs,
961 unsigned long *sp,
962 const struct stacktrace_ops *ops,
963 void *data)
964 {
965 struct dwarf_frame *frame, *_frame;
966 unsigned long return_addr;
967
968 _frame = NULL;
969 return_addr = 0;
970
971 while (1) {
972 frame = dwarf_unwind_stack(return_addr, _frame);
973
974 if (_frame)
975 dwarf_free_frame(_frame);
976
977 _frame = frame;
978
979 if (!frame || !frame->return_addr)
980 break;
981
982 return_addr = frame->return_addr;
983 ops->address(data, return_addr, 1);
984 }
985
986 if (frame)
987 dwarf_free_frame(frame);
988 }
989
990 static struct unwinder dwarf_unwinder = {
991 .name = "dwarf-unwinder",
992 .dump = dwarf_unwinder_dump,
993 .rating = 150,
994 };
995
dwarf_unwinder_cleanup(void)996 static void dwarf_unwinder_cleanup(void)
997 {
998 struct rb_node **fde_rb_node = &fde_root.rb_node;
999 struct rb_node **cie_rb_node = &cie_root.rb_node;
1000
1001 /*
1002 * Deallocate all the memory allocated for the DWARF unwinder.
1003 * Traverse all the FDE/CIE lists and remove and free all the
1004 * memory associated with those data structures.
1005 */
1006 while (*fde_rb_node) {
1007 struct dwarf_fde *fde;
1008
1009 fde = rb_entry(*fde_rb_node, struct dwarf_fde, node);
1010 rb_erase(*fde_rb_node, &fde_root);
1011 kfree(fde);
1012 }
1013
1014 while (*cie_rb_node) {
1015 struct dwarf_cie *cie;
1016
1017 cie = rb_entry(*cie_rb_node, struct dwarf_cie, node);
1018 rb_erase(*cie_rb_node, &cie_root);
1019 kfree(cie);
1020 }
1021
1022 kmem_cache_destroy(dwarf_reg_cachep);
1023 kmem_cache_destroy(dwarf_frame_cachep);
1024 }
1025
1026 /**
1027 * dwarf_parse_section - parse DWARF section
1028 * @eh_frame_start: start address of the .eh_frame section
1029 * @eh_frame_end: end address of the .eh_frame section
1030 * @mod: the kernel module containing the .eh_frame section
1031 *
1032 * Parse the information in a .eh_frame section.
1033 */
dwarf_parse_section(char * eh_frame_start,char * eh_frame_end,struct module * mod)1034 static int dwarf_parse_section(char *eh_frame_start, char *eh_frame_end,
1035 struct module *mod)
1036 {
1037 u32 entry_type;
1038 void *p, *entry;
1039 int count, err = 0;
1040 unsigned long len = 0;
1041 unsigned int c_entries, f_entries;
1042 unsigned char *end;
1043
1044 c_entries = 0;
1045 f_entries = 0;
1046 entry = eh_frame_start;
1047
1048 while ((char *)entry < eh_frame_end) {
1049 p = entry;
1050
1051 count = dwarf_entry_len(p, &len);
1052 if (count == 0) {
1053 /*
1054 * We read a bogus length field value. There is
1055 * nothing we can do here apart from disabling
1056 * the DWARF unwinder. We can't even skip this
1057 * entry and move to the next one because 'len'
1058 * tells us where our next entry is.
1059 */
1060 err = -EINVAL;
1061 goto out;
1062 } else
1063 p += count;
1064
1065 /* initial length does not include itself */
1066 end = p + len;
1067
1068 entry_type = get_unaligned((u32 *)p);
1069 p += 4;
1070
1071 if (entry_type == DW_EH_FRAME_CIE) {
1072 err = dwarf_parse_cie(entry, p, len, end, mod);
1073 if (err < 0)
1074 goto out;
1075 else
1076 c_entries++;
1077 } else {
1078 err = dwarf_parse_fde(entry, entry_type, p, len,
1079 end, mod);
1080 if (err < 0)
1081 goto out;
1082 else
1083 f_entries++;
1084 }
1085
1086 entry = (char *)entry + len + 4;
1087 }
1088
1089 printk(KERN_INFO "DWARF unwinder initialised: read %u CIEs, %u FDEs\n",
1090 c_entries, f_entries);
1091
1092 return 0;
1093
1094 out:
1095 return err;
1096 }
1097
1098 #ifdef CONFIG_MODULES
module_dwarf_finalize(const Elf_Ehdr * hdr,const Elf_Shdr * sechdrs,struct module * me)1099 int module_dwarf_finalize(const Elf_Ehdr *hdr, const Elf_Shdr *sechdrs,
1100 struct module *me)
1101 {
1102 unsigned int i, err;
1103 unsigned long start, end;
1104 char *secstrings = (void *)hdr + sechdrs[hdr->e_shstrndx].sh_offset;
1105
1106 start = end = 0;
1107
1108 for (i = 1; i < hdr->e_shnum; i++) {
1109 /* Alloc bit cleared means "ignore it." */
1110 if ((sechdrs[i].sh_flags & SHF_ALLOC)
1111 && !strcmp(secstrings+sechdrs[i].sh_name, ".eh_frame")) {
1112 start = sechdrs[i].sh_addr;
1113 end = start + sechdrs[i].sh_size;
1114 break;
1115 }
1116 }
1117
1118 /* Did we find the .eh_frame section? */
1119 if (i != hdr->e_shnum) {
1120 INIT_LIST_HEAD(&me->arch.cie_list);
1121 INIT_LIST_HEAD(&me->arch.fde_list);
1122 err = dwarf_parse_section((char *)start, (char *)end, me);
1123 if (err) {
1124 printk(KERN_WARNING "%s: failed to parse DWARF info\n",
1125 me->name);
1126 return err;
1127 }
1128 }
1129
1130 return 0;
1131 }
1132
1133 /**
1134 * module_dwarf_cleanup - remove FDE/CIEs associated with @mod
1135 * @mod: the module that is being unloaded
1136 *
1137 * Remove any FDEs and CIEs from the global lists that came from
1138 * @mod's .eh_frame section because @mod is being unloaded.
1139 */
module_dwarf_cleanup(struct module * mod)1140 void module_dwarf_cleanup(struct module *mod)
1141 {
1142 struct dwarf_fde *fde, *ftmp;
1143 struct dwarf_cie *cie, *ctmp;
1144 unsigned long flags;
1145
1146 spin_lock_irqsave(&dwarf_cie_lock, flags);
1147
1148 list_for_each_entry_safe(cie, ctmp, &mod->arch.cie_list, link) {
1149 list_del(&cie->link);
1150 rb_erase(&cie->node, &cie_root);
1151 kfree(cie);
1152 }
1153
1154 spin_unlock_irqrestore(&dwarf_cie_lock, flags);
1155
1156 spin_lock_irqsave(&dwarf_fde_lock, flags);
1157
1158 list_for_each_entry_safe(fde, ftmp, &mod->arch.fde_list, link) {
1159 list_del(&fde->link);
1160 rb_erase(&fde->node, &fde_root);
1161 kfree(fde);
1162 }
1163
1164 spin_unlock_irqrestore(&dwarf_fde_lock, flags);
1165 }
1166 #endif /* CONFIG_MODULES */
1167
1168 /**
1169 * dwarf_unwinder_init - initialise the dwarf unwinder
1170 *
1171 * Build the data structures describing the .dwarf_frame section to
1172 * make it easier to lookup CIE and FDE entries. Because the
1173 * .eh_frame section is packed as tightly as possible it is not
1174 * easy to lookup the FDE for a given PC, so we build a list of FDE
1175 * and CIE entries that make it easier.
1176 */
dwarf_unwinder_init(void)1177 static int __init dwarf_unwinder_init(void)
1178 {
1179 int err = -ENOMEM;
1180
1181 dwarf_frame_cachep = kmem_cache_create("dwarf_frames",
1182 sizeof(struct dwarf_frame), 0,
1183 SLAB_PANIC | SLAB_HWCACHE_ALIGN | SLAB_NOTRACK, NULL);
1184
1185 dwarf_reg_cachep = kmem_cache_create("dwarf_regs",
1186 sizeof(struct dwarf_reg), 0,
1187 SLAB_PANIC | SLAB_HWCACHE_ALIGN | SLAB_NOTRACK, NULL);
1188
1189 dwarf_frame_pool = mempool_create(DWARF_FRAME_MIN_REQ,
1190 mempool_alloc_slab,
1191 mempool_free_slab,
1192 dwarf_frame_cachep);
1193 if (!dwarf_frame_pool)
1194 goto out;
1195
1196 dwarf_reg_pool = mempool_create(DWARF_REG_MIN_REQ,
1197 mempool_alloc_slab,
1198 mempool_free_slab,
1199 dwarf_reg_cachep);
1200 if (!dwarf_reg_pool)
1201 goto out;
1202
1203 err = dwarf_parse_section(__start_eh_frame, __stop_eh_frame, NULL);
1204 if (err)
1205 goto out;
1206
1207 err = unwinder_register(&dwarf_unwinder);
1208 if (err)
1209 goto out;
1210
1211 dwarf_unwinder_ready = 1;
1212
1213 return 0;
1214
1215 out:
1216 printk(KERN_ERR "Failed to initialise DWARF unwinder: %d\n", err);
1217 dwarf_unwinder_cleanup();
1218 return err;
1219 }
1220 early_initcall(dwarf_unwinder_init);
1221