bd1c497280a66b927b1b07e8eaa4817fd278920a
[linux-2.6.git] / arch / sh / kernel / dwarf.c
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 <asm/dwarf.h>
26 #include <asm/unwinder.h>
27 #include <asm/sections.h>
28 #include <asm/unaligned.h>
29 #include <asm/stacktrace.h>
30
31 /* Reserve enough memory for two stack frames */
32 #define DWARF_FRAME_MIN_REQ     2
33 /* ... with 4 registers per frame. */
34 #define DWARF_REG_MIN_REQ       (DWARF_FRAME_MIN_REQ * 4)
35
36 static struct kmem_cache *dwarf_frame_cachep;
37 static mempool_t *dwarf_frame_pool;
38
39 static struct kmem_cache *dwarf_reg_cachep;
40 static mempool_t *dwarf_reg_pool;
41
42 static struct rb_root cie_root;
43 static DEFINE_SPINLOCK(dwarf_cie_lock);
44
45 static struct rb_root fde_root;
46 static DEFINE_SPINLOCK(dwarf_fde_lock);
47
48 static struct dwarf_cie *cached_cie;
49
50 /**
51  *      dwarf_frame_alloc_reg - allocate memory for a DWARF register
52  *      @frame: the DWARF frame whose list of registers we insert on
53  *      @reg_num: the register number
54  *
55  *      Allocate space for, and initialise, a dwarf reg from
56  *      dwarf_reg_pool and insert it onto the (unsorted) linked-list of
57  *      dwarf registers for @frame.
58  *
59  *      Return the initialised DWARF reg.
60  */
61 static struct dwarf_reg *dwarf_frame_alloc_reg(struct dwarf_frame *frame,
62                                                unsigned int reg_num)
63 {
64         struct dwarf_reg *reg;
65
66         reg = mempool_alloc(dwarf_reg_pool, GFP_ATOMIC);
67         if (!reg) {
68                 printk(KERN_WARNING "Unable to allocate a DWARF register\n");
69                 /*
70                  * Let's just bomb hard here, we have no way to
71                  * gracefully recover.
72                  */
73                 UNWINDER_BUG();
74         }
75
76         reg->number = reg_num;
77         reg->addr = 0;
78         reg->flags = 0;
79
80         list_add(&reg->link, &frame->reg_list);
81
82         return reg;
83 }
84
85 static void dwarf_frame_free_regs(struct dwarf_frame *frame)
86 {
87         struct dwarf_reg *reg, *n;
88
89         list_for_each_entry_safe(reg, n, &frame->reg_list, link) {
90                 list_del(&reg->link);
91                 mempool_free(reg, dwarf_reg_pool);
92         }
93 }
94
95 /**
96  *      dwarf_frame_reg - return a DWARF register
97  *      @frame: the DWARF frame to search in for @reg_num
98  *      @reg_num: the register number to search for
99  *
100  *      Lookup and return the dwarf reg @reg_num for this frame. Return
101  *      NULL if @reg_num is an register invalid number.
102  */
103 static struct dwarf_reg *dwarf_frame_reg(struct dwarf_frame *frame,
104                                          unsigned int reg_num)
105 {
106         struct dwarf_reg *reg;
107
108         list_for_each_entry(reg, &frame->reg_list, link) {
109                 if (reg->number == reg_num)
110                         return reg;
111         }
112
113         return NULL;
114 }
115
116 /**
117  *      dwarf_read_addr - read dwarf data
118  *      @src: source address of data
119  *      @dst: destination address to store the data to
120  *
121  *      Read 'n' bytes from @src, where 'n' is the size of an address on
122  *      the native machine. We return the number of bytes read, which
123  *      should always be 'n'. We also have to be careful when reading
124  *      from @src and writing to @dst, because they can be arbitrarily
125  *      aligned. Return 'n' - the number of bytes read.
126  */
127 static inline int dwarf_read_addr(unsigned long *src, unsigned long *dst)
128 {
129         u32 val = get_unaligned(src);
130         put_unaligned(val, dst);
131         return sizeof(unsigned long *);
132 }
133
134 /**
135  *      dwarf_read_uleb128 - read unsigned LEB128 data
136  *      @addr: the address where the ULEB128 data is stored
137  *      @ret: address to store the result
138  *
139  *      Decode an unsigned LEB128 encoded datum. The algorithm is taken
140  *      from Appendix C of the DWARF 3 spec. For information on the
141  *      encodings refer to section "7.6 - Variable Length Data". Return
142  *      the number of bytes read.
143  */
144 static inline unsigned long dwarf_read_uleb128(char *addr, unsigned int *ret)
145 {
146         unsigned int result;
147         unsigned char byte;
148         int shift, count;
149
150         result = 0;
151         shift = 0;
152         count = 0;
153
154         while (1) {
155                 byte = __raw_readb(addr);
156                 addr++;
157                 count++;
158
159                 result |= (byte & 0x7f) << shift;
160                 shift += 7;
161
162                 if (!(byte & 0x80))
163                         break;
164         }
165
166         *ret = result;
167
168         return count;
169 }
170
171 /**
172  *      dwarf_read_leb128 - read signed LEB128 data
173  *      @addr: the address of the LEB128 encoded data
174  *      @ret: address to store the result
175  *
176  *      Decode signed LEB128 data. The algorithm is taken from Appendix
177  *      C of the DWARF 3 spec. Return the number of bytes read.
178  */
179 static inline unsigned long dwarf_read_leb128(char *addr, int *ret)
180 {
181         unsigned char byte;
182         int result, shift;
183         int num_bits;
184         int count;
185
186         result = 0;
187         shift = 0;
188         count = 0;
189
190         while (1) {
191                 byte = __raw_readb(addr);
192                 addr++;
193                 result |= (byte & 0x7f) << shift;
194                 shift += 7;
195                 count++;
196
197                 if (!(byte & 0x80))
198                         break;
199         }
200
201         /* The number of bits in a signed integer. */
202         num_bits = 8 * sizeof(result);
203
204         if ((shift < num_bits) && (byte & 0x40))
205                 result |= (-1 << shift);
206
207         *ret = result;
208
209         return count;
210 }
211
212 /**
213  *      dwarf_read_encoded_value - return the decoded value at @addr
214  *      @addr: the address of the encoded value
215  *      @val: where to write the decoded value
216  *      @encoding: the encoding with which we can decode @addr
217  *
218  *      GCC emits encoded address in the .eh_frame FDE entries. Decode
219  *      the value at @addr using @encoding. The decoded value is written
220  *      to @val and the number of bytes read is returned.
221  */
222 static int dwarf_read_encoded_value(char *addr, unsigned long *val,
223                                     char encoding)
224 {
225         unsigned long decoded_addr = 0;
226         int count = 0;
227
228         switch (encoding & 0x70) {
229         case DW_EH_PE_absptr:
230                 break;
231         case DW_EH_PE_pcrel:
232                 decoded_addr = (unsigned long)addr;
233                 break;
234         default:
235                 pr_debug("encoding=0x%x\n", (encoding & 0x70));
236                 UNWINDER_BUG();
237         }
238
239         if ((encoding & 0x07) == 0x00)
240                 encoding |= DW_EH_PE_udata4;
241
242         switch (encoding & 0x0f) {
243         case DW_EH_PE_sdata4:
244         case DW_EH_PE_udata4:
245                 count += 4;
246                 decoded_addr += get_unaligned((u32 *)addr);
247                 __raw_writel(decoded_addr, val);
248                 break;
249         default:
250                 pr_debug("encoding=0x%x\n", encoding);
251                 UNWINDER_BUG();
252         }
253
254         return count;
255 }
256
257 /**
258  *      dwarf_entry_len - return the length of an FDE or CIE
259  *      @addr: the address of the entry
260  *      @len: the length of the entry
261  *
262  *      Read the initial_length field of the entry and store the size of
263  *      the entry in @len. We return the number of bytes read. Return a
264  *      count of 0 on error.
265  */
266 static inline int dwarf_entry_len(char *addr, unsigned long *len)
267 {
268         u32 initial_len;
269         int count;
270
271         initial_len = get_unaligned((u32 *)addr);
272         count = 4;
273
274         /*
275          * An initial length field value in the range DW_LEN_EXT_LO -
276          * DW_LEN_EXT_HI indicates an extension, and should not be
277          * interpreted as a length. The only extension that we currently
278          * understand is the use of DWARF64 addresses.
279          */
280         if (initial_len >= DW_EXT_LO && initial_len <= DW_EXT_HI) {
281                 /*
282                  * The 64-bit length field immediately follows the
283                  * compulsory 32-bit length field.
284                  */
285                 if (initial_len == DW_EXT_DWARF64) {
286                         *len = get_unaligned((u64 *)addr + 4);
287                         count = 12;
288                 } else {
289                         printk(KERN_WARNING "Unknown DWARF extension\n");
290                         count = 0;
291                 }
292         } else
293                 *len = initial_len;
294
295         return count;
296 }
297
298 /**
299  *      dwarf_lookup_cie - locate the cie
300  *      @cie_ptr: pointer to help with lookup
301  */
302 static struct dwarf_cie *dwarf_lookup_cie(unsigned long cie_ptr)
303 {
304         struct rb_node **rb_node = &cie_root.rb_node;
305         struct dwarf_cie *cie = NULL;
306         unsigned long flags;
307
308         spin_lock_irqsave(&dwarf_cie_lock, flags);
309
310         /*
311          * We've cached the last CIE we looked up because chances are
312          * that the FDE wants this CIE.
313          */
314         if (cached_cie && cached_cie->cie_pointer == cie_ptr) {
315                 cie = cached_cie;
316                 goto out;
317         }
318
319         while (*rb_node) {
320                 struct dwarf_cie *cie_tmp;
321
322                 cie_tmp = rb_entry(*rb_node, struct dwarf_cie, node);
323                 BUG_ON(!cie_tmp);
324
325                 if (cie_ptr == cie_tmp->cie_pointer) {
326                         cie = cie_tmp;
327                         cached_cie = cie_tmp;
328                         goto out;
329                 } else {
330                         if (cie_ptr < cie_tmp->cie_pointer)
331                                 rb_node = &(*rb_node)->rb_left;
332                         else
333                                 rb_node = &(*rb_node)->rb_right;
334                 }
335         }
336
337 out:
338         spin_unlock_irqrestore(&dwarf_cie_lock, flags);
339         return cie;
340 }
341
342 /**
343  *      dwarf_lookup_fde - locate the FDE that covers pc
344  *      @pc: the program counter
345  */
346 struct dwarf_fde *dwarf_lookup_fde(unsigned long pc)
347 {
348         struct rb_node **rb_node = &fde_root.rb_node;
349         struct dwarf_fde *fde = NULL;
350         unsigned long flags;
351
352         spin_lock_irqsave(&dwarf_fde_lock, flags);
353
354         while (*rb_node) {
355                 struct dwarf_fde *fde_tmp;
356                 unsigned long tmp_start, tmp_end;
357
358                 fde_tmp = rb_entry(*rb_node, struct dwarf_fde, node);
359                 BUG_ON(!fde_tmp);
360
361                 tmp_start = fde_tmp->initial_location;
362                 tmp_end = fde_tmp->initial_location + fde_tmp->address_range;
363
364                 if (pc < tmp_start) {
365                         rb_node = &(*rb_node)->rb_left;
366                 } else {
367                         if (pc < tmp_end) {
368                                 fde = fde_tmp;
369                                 goto out;
370                         } else
371                                 rb_node = &(*rb_node)->rb_right;
372                 }
373         }
374
375 out:
376         spin_unlock_irqrestore(&dwarf_fde_lock, flags);
377
378         return fde;
379 }
380
381 /**
382  *      dwarf_cfa_execute_insns - execute instructions to calculate a CFA
383  *      @insn_start: address of the first instruction
384  *      @insn_end: address of the last instruction
385  *      @cie: the CIE for this function
386  *      @fde: the FDE for this function
387  *      @frame: the instructions calculate the CFA for this frame
388  *      @pc: the program counter of the address we're interested in
389  *
390  *      Execute the Call Frame instruction sequence starting at
391  *      @insn_start and ending at @insn_end. The instructions describe
392  *      how to calculate the Canonical Frame Address of a stackframe.
393  *      Store the results in @frame.
394  */
395 static int dwarf_cfa_execute_insns(unsigned char *insn_start,
396                                    unsigned char *insn_end,
397                                    struct dwarf_cie *cie,
398                                    struct dwarf_fde *fde,
399                                    struct dwarf_frame *frame,
400                                    unsigned long pc)
401 {
402         unsigned char insn;
403         unsigned char *current_insn;
404         unsigned int count, delta, reg, expr_len, offset;
405         struct dwarf_reg *regp;
406
407         current_insn = insn_start;
408
409         while (current_insn < insn_end && frame->pc <= pc) {
410                 insn = __raw_readb(current_insn++);
411
412                 /*
413                  * Firstly, handle the opcodes that embed their operands
414                  * in the instructions.
415                  */
416                 switch (DW_CFA_opcode(insn)) {
417                 case DW_CFA_advance_loc:
418                         delta = DW_CFA_operand(insn);
419                         delta *= cie->code_alignment_factor;
420                         frame->pc += delta;
421                         continue;
422                         /* NOTREACHED */
423                 case DW_CFA_offset:
424                         reg = DW_CFA_operand(insn);
425                         count = dwarf_read_uleb128(current_insn, &offset);
426                         current_insn += count;
427                         offset *= cie->data_alignment_factor;
428                         regp = dwarf_frame_alloc_reg(frame, reg);
429                         regp->addr = offset;
430                         regp->flags |= DWARF_REG_OFFSET;
431                         continue;
432                         /* NOTREACHED */
433                 case DW_CFA_restore:
434                         reg = DW_CFA_operand(insn);
435                         continue;
436                         /* NOTREACHED */
437                 }
438
439                 /*
440                  * Secondly, handle the opcodes that don't embed their
441                  * operands in the instruction.
442                  */
443                 switch (insn) {
444                 case DW_CFA_nop:
445                         continue;
446                 case DW_CFA_advance_loc1:
447                         delta = *current_insn++;
448                         frame->pc += delta * cie->code_alignment_factor;
449                         break;
450                 case DW_CFA_advance_loc2:
451                         delta = get_unaligned((u16 *)current_insn);
452                         current_insn += 2;
453                         frame->pc += delta * cie->code_alignment_factor;
454                         break;
455                 case DW_CFA_advance_loc4:
456                         delta = get_unaligned((u32 *)current_insn);
457                         current_insn += 4;
458                         frame->pc += delta * cie->code_alignment_factor;
459                         break;
460                 case DW_CFA_offset_extended:
461                         count = dwarf_read_uleb128(current_insn, &reg);
462                         current_insn += count;
463                         count = dwarf_read_uleb128(current_insn, &offset);
464                         current_insn += count;
465                         offset *= cie->data_alignment_factor;
466                         break;
467                 case DW_CFA_restore_extended:
468                         count = dwarf_read_uleb128(current_insn, &reg);
469                         current_insn += count;
470                         break;
471                 case DW_CFA_undefined:
472                         count = dwarf_read_uleb128(current_insn, &reg);
473                         current_insn += count;
474                         regp = dwarf_frame_alloc_reg(frame, reg);
475                         regp->flags |= DWARF_UNDEFINED;
476                         break;
477                 case DW_CFA_def_cfa:
478                         count = dwarf_read_uleb128(current_insn,
479                                                    &frame->cfa_register);
480                         current_insn += count;
481                         count = dwarf_read_uleb128(current_insn,
482                                                    &frame->cfa_offset);
483                         current_insn += count;
484
485                         frame->flags |= DWARF_FRAME_CFA_REG_OFFSET;
486                         break;
487                 case DW_CFA_def_cfa_register:
488                         count = dwarf_read_uleb128(current_insn,
489                                                    &frame->cfa_register);
490                         current_insn += count;
491                         frame->flags |= DWARF_FRAME_CFA_REG_OFFSET;
492                         break;
493                 case DW_CFA_def_cfa_offset:
494                         count = dwarf_read_uleb128(current_insn, &offset);
495                         current_insn += count;
496                         frame->cfa_offset = offset;
497                         break;
498                 case DW_CFA_def_cfa_expression:
499                         count = dwarf_read_uleb128(current_insn, &expr_len);
500                         current_insn += count;
501
502                         frame->cfa_expr = current_insn;
503                         frame->cfa_expr_len = expr_len;
504                         current_insn += expr_len;
505
506                         frame->flags |= DWARF_FRAME_CFA_REG_EXP;
507                         break;
508                 case DW_CFA_offset_extended_sf:
509                         count = dwarf_read_uleb128(current_insn, &reg);
510                         current_insn += count;
511                         count = dwarf_read_leb128(current_insn, &offset);
512                         current_insn += count;
513                         offset *= cie->data_alignment_factor;
514                         regp = dwarf_frame_alloc_reg(frame, reg);
515                         regp->flags |= DWARF_REG_OFFSET;
516                         regp->addr = offset;
517                         break;
518                 case DW_CFA_val_offset:
519                         count = dwarf_read_uleb128(current_insn, &reg);
520                         current_insn += count;
521                         count = dwarf_read_leb128(current_insn, &offset);
522                         offset *= cie->data_alignment_factor;
523                         regp = dwarf_frame_alloc_reg(frame, reg);
524                         regp->flags |= DWARF_VAL_OFFSET;
525                         regp->addr = offset;
526                         break;
527                 case DW_CFA_GNU_args_size:
528                         count = dwarf_read_uleb128(current_insn, &offset);
529                         current_insn += count;
530                         break;
531                 case DW_CFA_GNU_negative_offset_extended:
532                         count = dwarf_read_uleb128(current_insn, &reg);
533                         current_insn += count;
534                         count = dwarf_read_uleb128(current_insn, &offset);
535                         offset *= cie->data_alignment_factor;
536
537                         regp = dwarf_frame_alloc_reg(frame, reg);
538                         regp->flags |= DWARF_REG_OFFSET;
539                         regp->addr = -offset;
540                         break;
541                 default:
542                         pr_debug("unhandled DWARF instruction 0x%x\n", insn);
543                         UNWINDER_BUG();
544                         break;
545                 }
546         }
547
548         return 0;
549 }
550
551 /**
552  *      dwarf_free_frame - free the memory allocated for @frame
553  *      @frame: the frame to free
554  */
555 void dwarf_free_frame(struct dwarf_frame *frame)
556 {
557         dwarf_frame_free_regs(frame);
558         mempool_free(frame, dwarf_frame_pool);
559 }
560
561 extern void ret_from_irq(void);
562
563 /**
564  *      dwarf_unwind_stack - unwind the stack
565  *
566  *      @pc: address of the function to unwind
567  *      @prev: struct dwarf_frame of the previous stackframe on the callstack
568  *
569  *      Return a struct dwarf_frame representing the most recent frame
570  *      on the callstack. Each of the lower (older) stack frames are
571  *      linked via the "prev" member.
572  */
573 struct dwarf_frame *dwarf_unwind_stack(unsigned long pc,
574                                        struct dwarf_frame *prev)
575 {
576         struct dwarf_frame *frame;
577         struct dwarf_cie *cie;
578         struct dwarf_fde *fde;
579         struct dwarf_reg *reg;
580         unsigned long addr;
581
582         /*
583          * If we're starting at the top of the stack we need get the
584          * contents of a physical register to get the CFA in order to
585          * begin the virtual unwinding of the stack.
586          *
587          * NOTE: the return address is guaranteed to be setup by the
588          * time this function makes its first function call.
589          */
590         if (!pc || !prev)
591                 pc = (unsigned long)current_text_addr();
592
593 #ifdef CONFIG_FUNCTION_GRAPH_TRACER
594         /*
595          * If our stack has been patched by the function graph tracer
596          * then we might see the address of return_to_handler() where we
597          * expected to find the real return address.
598          */
599         if (pc == (unsigned long)&return_to_handler) {
600                 int index = current->curr_ret_stack;
601
602                 /*
603                  * We currently have no way of tracking how many
604                  * return_to_handler()'s we've seen. If there is more
605                  * than one patched return address on our stack,
606                  * complain loudly.
607                  */
608                 WARN_ON(index > 0);
609
610                 pc = current->ret_stack[index].ret;
611         }
612 #endif
613
614         frame = mempool_alloc(dwarf_frame_pool, GFP_ATOMIC);
615         if (!frame) {
616                 printk(KERN_ERR "Unable to allocate a dwarf frame\n");
617                 UNWINDER_BUG();
618         }
619
620         INIT_LIST_HEAD(&frame->reg_list);
621         frame->flags = 0;
622         frame->prev = prev;
623         frame->return_addr = 0;
624
625         fde = dwarf_lookup_fde(pc);
626         if (!fde) {
627                 /*
628                  * This is our normal exit path. There are two reasons
629                  * why we might exit here,
630                  *
631                  *      a) pc has no asscociated DWARF frame info and so
632                  *      we don't know how to unwind this frame. This is
633                  *      usually the case when we're trying to unwind a
634                  *      frame that was called from some assembly code
635                  *      that has no DWARF info, e.g. syscalls.
636                  *
637                  *      b) the DEBUG info for pc is bogus. There's
638                  *      really no way to distinguish this case from the
639                  *      case above, which sucks because we could print a
640                  *      warning here.
641                  */
642                 goto bail;
643         }
644
645         cie = dwarf_lookup_cie(fde->cie_pointer);
646
647         frame->pc = fde->initial_location;
648
649         /* CIE initial instructions */
650         dwarf_cfa_execute_insns(cie->initial_instructions,
651                                 cie->instructions_end, cie, fde,
652                                 frame, pc);
653
654         /* FDE instructions */
655         dwarf_cfa_execute_insns(fde->instructions, fde->end, cie,
656                                 fde, frame, pc);
657
658         /* Calculate the CFA */
659         switch (frame->flags) {
660         case DWARF_FRAME_CFA_REG_OFFSET:
661                 if (prev) {
662                         reg = dwarf_frame_reg(prev, frame->cfa_register);
663                         UNWINDER_BUG_ON(!reg);
664                         UNWINDER_BUG_ON(reg->flags != DWARF_REG_OFFSET);
665
666                         addr = prev->cfa + reg->addr;
667                         frame->cfa = __raw_readl(addr);
668
669                 } else {
670                         /*
671                          * Again, we're starting from the top of the
672                          * stack. We need to physically read
673                          * the contents of a register in order to get
674                          * the Canonical Frame Address for this
675                          * function.
676                          */
677                         frame->cfa = dwarf_read_arch_reg(frame->cfa_register);
678                 }
679
680                 frame->cfa += frame->cfa_offset;
681                 break;
682         default:
683                 UNWINDER_BUG();
684         }
685
686         reg = dwarf_frame_reg(frame, DWARF_ARCH_RA_REG);
687
688         /*
689          * If we haven't seen the return address register or the return
690          * address column is undefined then we must assume that this is
691          * the end of the callstack.
692          */
693         if (!reg || reg->flags == DWARF_UNDEFINED)
694                 goto bail;
695
696         UNWINDER_BUG_ON(reg->flags != DWARF_REG_OFFSET);
697
698         addr = frame->cfa + reg->addr;
699         frame->return_addr = __raw_readl(addr);
700
701         /*
702          * Ah, the joys of unwinding through interrupts.
703          *
704          * Interrupts are tricky - the DWARF info needs to be _really_
705          * accurate and unfortunately I'm seeing a lot of bogus DWARF
706          * info. For example, I've seen interrupts occur in epilogues
707          * just after the frame pointer (r14) had been restored. The
708          * problem was that the DWARF info claimed that the CFA could be
709          * reached by using the value of the frame pointer before it was
710          * restored.
711          *
712          * So until the compiler can be trusted to produce reliable
713          * DWARF info when it really matters, let's stop unwinding once
714          * we've calculated the function that was interrupted.
715          */
716         if (prev && prev->pc == (unsigned long)ret_from_irq)
717                 frame->return_addr = 0;
718
719         return frame;
720
721 bail:
722         dwarf_free_frame(frame);
723         return NULL;
724 }
725
726 static int dwarf_parse_cie(void *entry, void *p, unsigned long len,
727                            unsigned char *end, struct module *mod)
728 {
729         struct rb_node **rb_node = &cie_root.rb_node;
730         struct rb_node *parent;
731         struct dwarf_cie *cie;
732         unsigned long flags;
733         int count;
734
735         cie = kzalloc(sizeof(*cie), GFP_KERNEL);
736         if (!cie)
737                 return -ENOMEM;
738
739         cie->length = len;
740
741         /*
742          * Record the offset into the .eh_frame section
743          * for this CIE. It allows this CIE to be
744          * quickly and easily looked up from the
745          * corresponding FDE.
746          */
747         cie->cie_pointer = (unsigned long)entry;
748
749         cie->version = *(char *)p++;
750         UNWINDER_BUG_ON(cie->version != 1);
751
752         cie->augmentation = p;
753         p += strlen(cie->augmentation) + 1;
754
755         count = dwarf_read_uleb128(p, &cie->code_alignment_factor);
756         p += count;
757
758         count = dwarf_read_leb128(p, &cie->data_alignment_factor);
759         p += count;
760
761         /*
762          * Which column in the rule table contains the
763          * return address?
764          */
765         if (cie->version == 1) {
766                 cie->return_address_reg = __raw_readb(p);
767                 p++;
768         } else {
769                 count = dwarf_read_uleb128(p, &cie->return_address_reg);
770                 p += count;
771         }
772
773         if (cie->augmentation[0] == 'z') {
774                 unsigned int length, count;
775                 cie->flags |= DWARF_CIE_Z_AUGMENTATION;
776
777                 count = dwarf_read_uleb128(p, &length);
778                 p += count;
779
780                 UNWINDER_BUG_ON((unsigned char *)p > end);
781
782                 cie->initial_instructions = p + length;
783                 cie->augmentation++;
784         }
785
786         while (*cie->augmentation) {
787                 /*
788                  * "L" indicates a byte showing how the
789                  * LSDA pointer is encoded. Skip it.
790                  */
791                 if (*cie->augmentation == 'L') {
792                         p++;
793                         cie->augmentation++;
794                 } else if (*cie->augmentation == 'R') {
795                         /*
796                          * "R" indicates a byte showing
797                          * how FDE addresses are
798                          * encoded.
799                          */
800                         cie->encoding = *(char *)p++;
801                         cie->augmentation++;
802                 } else if (*cie->augmentation == 'P') {
803                         /*
804                          * "R" indicates a personality
805                          * routine in the CIE
806                          * augmentation.
807                          */
808                         UNWINDER_BUG();
809                 } else if (*cie->augmentation == 'S') {
810                         UNWINDER_BUG();
811                 } else {
812                         /*
813                          * Unknown augmentation. Assume
814                          * 'z' augmentation.
815                          */
816                         p = cie->initial_instructions;
817                         UNWINDER_BUG_ON(!p);
818                         break;
819                 }
820         }
821
822         cie->initial_instructions = p;
823         cie->instructions_end = end;
824
825         /* Add to list */
826         spin_lock_irqsave(&dwarf_cie_lock, flags);
827
828         while (*rb_node) {
829                 struct dwarf_cie *cie_tmp;
830
831                 cie_tmp = rb_entry(*rb_node, struct dwarf_cie, node);
832
833                 parent = *rb_node;
834
835                 if (cie->cie_pointer < cie_tmp->cie_pointer)
836                         rb_node = &parent->rb_left;
837                 else if (cie->cie_pointer >= cie_tmp->cie_pointer)
838                         rb_node = &parent->rb_right;
839                 else
840                         WARN_ON(1);
841         }
842
843         rb_link_node(&cie->node, parent, rb_node);
844         rb_insert_color(&cie->node, &cie_root);
845
846         if (mod != NULL)
847                 list_add_tail(&cie->link, &mod->arch.cie_list);
848
849         spin_unlock_irqrestore(&dwarf_cie_lock, flags);
850
851         return 0;
852 }
853
854 static int dwarf_parse_fde(void *entry, u32 entry_type,
855                            void *start, unsigned long len,
856                            unsigned char *end, struct module *mod)
857 {
858         struct rb_node **rb_node = &fde_root.rb_node;
859         struct rb_node *parent;
860         struct dwarf_fde *fde;
861         struct dwarf_cie *cie;
862         unsigned long flags;
863         int count;
864         void *p = start;
865
866         fde = kzalloc(sizeof(*fde), GFP_KERNEL);
867         if (!fde)
868                 return -ENOMEM;
869
870         fde->length = len;
871
872         /*
873          * In a .eh_frame section the CIE pointer is the
874          * delta between the address within the FDE
875          */
876         fde->cie_pointer = (unsigned long)(p - entry_type - 4);
877
878         cie = dwarf_lookup_cie(fde->cie_pointer);
879         fde->cie = cie;
880
881         if (cie->encoding)
882                 count = dwarf_read_encoded_value(p, &fde->initial_location,
883                                                  cie->encoding);
884         else
885                 count = dwarf_read_addr(p, &fde->initial_location);
886
887         p += count;
888
889         if (cie->encoding)
890                 count = dwarf_read_encoded_value(p, &fde->address_range,
891                                                  cie->encoding & 0x0f);
892         else
893                 count = dwarf_read_addr(p, &fde->address_range);
894
895         p += count;
896
897         if (fde->cie->flags & DWARF_CIE_Z_AUGMENTATION) {
898                 unsigned int length;
899                 count = dwarf_read_uleb128(p, &length);
900                 p += count + length;
901         }
902
903         /* Call frame instructions. */
904         fde->instructions = p;
905         fde->end = end;
906
907         /* Add to list. */
908         spin_lock_irqsave(&dwarf_fde_lock, flags);
909
910         while (*rb_node) {
911                 struct dwarf_fde *fde_tmp;
912                 unsigned long tmp_start, tmp_end;
913                 unsigned long start, end;
914
915                 fde_tmp = rb_entry(*rb_node, struct dwarf_fde, node);
916
917                 start = fde->initial_location;
918                 end = fde->initial_location + fde->address_range;
919
920                 tmp_start = fde_tmp->initial_location;
921                 tmp_end = fde_tmp->initial_location + fde_tmp->address_range;
922
923                 parent = *rb_node;
924
925                 if (start < tmp_start)
926                         rb_node = &parent->rb_left;
927                 else if (start >= tmp_end)
928                         rb_node = &parent->rb_right;
929                 else
930                         WARN_ON(1);
931         }
932
933         rb_link_node(&fde->node, parent, rb_node);
934         rb_insert_color(&fde->node, &fde_root);
935
936         if (mod != NULL)
937                 list_add_tail(&fde->link, &mod->arch.fde_list);
938
939         spin_unlock_irqrestore(&dwarf_fde_lock, flags);
940
941         return 0;
942 }
943
944 static void dwarf_unwinder_dump(struct task_struct *task,
945                                 struct pt_regs *regs,
946                                 unsigned long *sp,
947                                 const struct stacktrace_ops *ops,
948                                 void *data)
949 {
950         struct dwarf_frame *frame, *_frame;
951         unsigned long return_addr;
952
953         _frame = NULL;
954         return_addr = 0;
955
956         while (1) {
957                 frame = dwarf_unwind_stack(return_addr, _frame);
958
959                 if (_frame)
960                         dwarf_free_frame(_frame);
961
962                 _frame = frame;
963
964                 if (!frame || !frame->return_addr)
965                         break;
966
967                 return_addr = frame->return_addr;
968                 ops->address(data, return_addr, 1);
969         }
970
971         if (frame)
972                 dwarf_free_frame(frame);
973 }
974
975 static struct unwinder dwarf_unwinder = {
976         .name = "dwarf-unwinder",
977         .dump = dwarf_unwinder_dump,
978         .rating = 150,
979 };
980
981 static void dwarf_unwinder_cleanup(void)
982 {
983         struct rb_node **fde_rb_node = &fde_root.rb_node;
984         struct rb_node **cie_rb_node = &cie_root.rb_node;
985
986         /*
987          * Deallocate all the memory allocated for the DWARF unwinder.
988          * Traverse all the FDE/CIE lists and remove and free all the
989          * memory associated with those data structures.
990          */
991         while (*fde_rb_node) {
992                 struct dwarf_fde *fde;
993
994                 fde = rb_entry(*fde_rb_node, struct dwarf_fde, node);
995                 rb_erase(*fde_rb_node, &fde_root);
996                 kfree(fde);
997         }
998
999         while (*cie_rb_node) {
1000                 struct dwarf_cie *cie;
1001
1002                 cie = rb_entry(*cie_rb_node, struct dwarf_cie, node);
1003                 rb_erase(*cie_rb_node, &cie_root);
1004                 kfree(cie);
1005         }
1006
1007         kmem_cache_destroy(dwarf_reg_cachep);
1008         kmem_cache_destroy(dwarf_frame_cachep);
1009 }
1010
1011 /**
1012  *      dwarf_parse_section - parse DWARF section
1013  *      @eh_frame_start: start address of the .eh_frame section
1014  *      @eh_frame_end: end address of the .eh_frame section
1015  *      @mod: the kernel module containing the .eh_frame section
1016  *
1017  *      Parse the information in a .eh_frame section.
1018  */
1019 static int dwarf_parse_section(char *eh_frame_start, char *eh_frame_end,
1020                                struct module *mod)
1021 {
1022         u32 entry_type;
1023         void *p, *entry;
1024         int count, err = 0;
1025         unsigned long len = 0;
1026         unsigned int c_entries, f_entries;
1027         unsigned char *end;
1028
1029         c_entries = 0;
1030         f_entries = 0;
1031         entry = eh_frame_start;
1032
1033         while ((char *)entry < eh_frame_end) {
1034                 p = entry;
1035
1036                 count = dwarf_entry_len(p, &len);
1037                 if (count == 0) {
1038                         /*
1039                          * We read a bogus length field value. There is
1040                          * nothing we can do here apart from disabling
1041                          * the DWARF unwinder. We can't even skip this
1042                          * entry and move to the next one because 'len'
1043                          * tells us where our next entry is.
1044                          */
1045                         err = -EINVAL;
1046                         goto out;
1047                 } else
1048                         p += count;
1049
1050                 /* initial length does not include itself */
1051                 end = p + len;
1052
1053                 entry_type = get_unaligned((u32 *)p);
1054                 p += 4;
1055
1056                 if (entry_type == DW_EH_FRAME_CIE) {
1057                         err = dwarf_parse_cie(entry, p, len, end, mod);
1058                         if (err < 0)
1059                                 goto out;
1060                         else
1061                                 c_entries++;
1062                 } else {
1063                         err = dwarf_parse_fde(entry, entry_type, p, len,
1064                                               end, mod);
1065                         if (err < 0)
1066                                 goto out;
1067                         else
1068                                 f_entries++;
1069                 }
1070
1071                 entry = (char *)entry + len + 4;
1072         }
1073
1074         printk(KERN_INFO "DWARF unwinder initialised: read %u CIEs, %u FDEs\n",
1075                c_entries, f_entries);
1076
1077         return 0;
1078
1079 out:
1080         return err;
1081 }
1082
1083 #ifdef CONFIG_MODULES
1084 int module_dwarf_finalize(const Elf_Ehdr *hdr, const Elf_Shdr *sechdrs,
1085                           struct module *me)
1086 {
1087         unsigned int i, err;
1088         unsigned long start, end;
1089         char *secstrings = (void *)hdr + sechdrs[hdr->e_shstrndx].sh_offset;
1090
1091         start = end = 0;
1092
1093         for (i = 1; i < hdr->e_shnum; i++) {
1094                 /* Alloc bit cleared means "ignore it." */
1095                 if ((sechdrs[i].sh_flags & SHF_ALLOC)
1096                     && !strcmp(secstrings+sechdrs[i].sh_name, ".eh_frame")) {
1097                         start = sechdrs[i].sh_addr;
1098                         end = start + sechdrs[i].sh_size;
1099                         break;
1100                 }
1101         }
1102
1103         /* Did we find the .eh_frame section? */
1104         if (i != hdr->e_shnum) {
1105                 INIT_LIST_HEAD(&me->arch.cie_list);
1106                 INIT_LIST_HEAD(&me->arch.fde_list);
1107                 err = dwarf_parse_section((char *)start, (char *)end, me);
1108                 if (err) {
1109                         printk(KERN_WARNING "%s: failed to parse DWARF info\n",
1110                                me->name);
1111                         return err;
1112                 }
1113         }
1114
1115         return 0;
1116 }
1117
1118 /**
1119  *      module_dwarf_cleanup - remove FDE/CIEs associated with @mod
1120  *      @mod: the module that is being unloaded
1121  *
1122  *      Remove any FDEs and CIEs from the global lists that came from
1123  *      @mod's .eh_frame section because @mod is being unloaded.
1124  */
1125 void module_dwarf_cleanup(struct module *mod)
1126 {
1127         struct dwarf_fde *fde, *ftmp;
1128         struct dwarf_cie *cie, *ctmp;
1129         unsigned long flags;
1130
1131         spin_lock_irqsave(&dwarf_cie_lock, flags);
1132
1133         list_for_each_entry_safe(cie, ctmp, &mod->arch.cie_list, link) {
1134                 list_del(&cie->link);
1135                 rb_erase(&cie->node, &cie_root);
1136                 kfree(cie);
1137         }
1138
1139         spin_unlock_irqrestore(&dwarf_cie_lock, flags);
1140
1141         spin_lock_irqsave(&dwarf_fde_lock, flags);
1142
1143         list_for_each_entry_safe(fde, ftmp, &mod->arch.fde_list, link) {
1144                 list_del(&fde->link);
1145                 rb_erase(&fde->node, &fde_root);
1146                 kfree(fde);
1147         }
1148
1149         spin_unlock_irqrestore(&dwarf_fde_lock, flags);
1150 }
1151 #endif /* CONFIG_MODULES */
1152
1153 /**
1154  *      dwarf_unwinder_init - initialise the dwarf unwinder
1155  *
1156  *      Build the data structures describing the .dwarf_frame section to
1157  *      make it easier to lookup CIE and FDE entries. Because the
1158  *      .eh_frame section is packed as tightly as possible it is not
1159  *      easy to lookup the FDE for a given PC, so we build a list of FDE
1160  *      and CIE entries that make it easier.
1161  */
1162 static int __init dwarf_unwinder_init(void)
1163 {
1164         int err;
1165
1166         dwarf_frame_cachep = kmem_cache_create("dwarf_frames",
1167                         sizeof(struct dwarf_frame), 0,
1168                         SLAB_PANIC | SLAB_HWCACHE_ALIGN | SLAB_NOTRACK, NULL);
1169
1170         dwarf_reg_cachep = kmem_cache_create("dwarf_regs",
1171                         sizeof(struct dwarf_reg), 0,
1172                         SLAB_PANIC | SLAB_HWCACHE_ALIGN | SLAB_NOTRACK, NULL);
1173
1174         dwarf_frame_pool = mempool_create(DWARF_FRAME_MIN_REQ,
1175                                           mempool_alloc_slab,
1176                                           mempool_free_slab,
1177                                           dwarf_frame_cachep);
1178
1179         dwarf_reg_pool = mempool_create(DWARF_REG_MIN_REQ,
1180                                          mempool_alloc_slab,
1181                                          mempool_free_slab,
1182                                          dwarf_reg_cachep);
1183
1184         err = dwarf_parse_section(__start_eh_frame, __stop_eh_frame, NULL);
1185         if (err)
1186                 goto out;
1187
1188         err = unwinder_register(&dwarf_unwinder);
1189         if (err)
1190                 goto out;
1191
1192         return 0;
1193
1194 out:
1195         printk(KERN_ERR "Failed to initialise DWARF unwinder: %d\n", err);
1196         dwarf_unwinder_cleanup();
1197         return -EINVAL;
1198 }
1199 early_initcall(dwarf_unwinder_init);