2 * Copyright (C) 2009 Matt Fleming <matt@console-pimps.org>
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
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.
13 * - DWARF64 doesn't work.
17 #include <linux/kernel.h>
19 #include <linux/list.h>
21 #include <asm/dwarf.h>
22 #include <asm/unwinder.h>
23 #include <asm/sections.h>
24 #include <asm/unaligned.h>
25 #include <asm/dwarf.h>
26 #include <asm/stacktrace.h>
28 static LIST_HEAD(dwarf_cie_list
);
29 DEFINE_SPINLOCK(dwarf_cie_lock
);
31 static LIST_HEAD(dwarf_fde_list
);
32 DEFINE_SPINLOCK(dwarf_fde_lock
);
34 static struct dwarf_cie
*cached_cie
;
37 * Figure out whether we need to allocate some dwarf registers. If dwarf
38 * registers have already been allocated then we may need to realloc
39 * them. "reg" is a register number that we need to be able to access
42 * Register numbers start at zero, therefore we need to allocate space
43 * for "reg" + 1 registers.
45 static void dwarf_frame_alloc_regs(struct dwarf_frame
*frame
,
48 struct dwarf_reg
*regs
;
49 unsigned int num_regs
= reg
+ 1;
53 new_size
= num_regs
* sizeof(*regs
);
54 old_size
= frame
->num_regs
* sizeof(*regs
);
56 /* Fast path: don't allocate any regs if we've already got enough. */
57 if (frame
->num_regs
>= num_regs
)
60 regs
= kzalloc(new_size
, GFP_ATOMIC
);
62 printk(KERN_WARNING
"Unable to allocate DWARF registers\n");
64 * Let's just bomb hard here, we have no way to
71 memcpy(regs
, frame
->regs
, old_size
);
76 frame
->num_regs
= num_regs
;
80 * dwarf_read_addr - read dwarf data
81 * @src: source address of data
82 * @dst: destination address to store the data to
84 * Read 'n' bytes from @src, where 'n' is the size of an address on
85 * the native machine. We return the number of bytes read, which
86 * should always be 'n'. We also have to be careful when reading
87 * from @src and writing to @dst, because they can be arbitrarily
88 * aligned. Return 'n' - the number of bytes read.
90 static inline int dwarf_read_addr(unsigned long *src
, unsigned long *dst
)
92 u32 val
= get_unaligned(src
);
93 put_unaligned(val
, dst
);
94 return sizeof(unsigned long *);
98 * dwarf_read_uleb128 - read unsigned LEB128 data
99 * @addr: the address where the ULEB128 data is stored
100 * @ret: address to store the result
102 * Decode an unsigned LEB128 encoded datum. The algorithm is taken
103 * from Appendix C of the DWARF 3 spec. For information on the
104 * encodings refer to section "7.6 - Variable Length Data". Return
105 * the number of bytes read.
107 static inline unsigned long dwarf_read_uleb128(char *addr
, unsigned int *ret
)
118 byte
= __raw_readb(addr
);
122 result
|= (byte
& 0x7f) << shift
;
135 * dwarf_read_leb128 - read signed LEB128 data
136 * @addr: the address of the LEB128 encoded data
137 * @ret: address to store the result
139 * Decode signed LEB128 data. The algorithm is taken from Appendix
140 * C of the DWARF 3 spec. Return the number of bytes read.
142 static inline unsigned long dwarf_read_leb128(char *addr
, int *ret
)
154 byte
= __raw_readb(addr
);
156 result
|= (byte
& 0x7f) << shift
;
164 /* The number of bits in a signed integer. */
165 num_bits
= 8 * sizeof(result
);
167 if ((shift
< num_bits
) && (byte
& 0x40))
168 result
|= (-1 << shift
);
176 * dwarf_read_encoded_value - return the decoded value at @addr
177 * @addr: the address of the encoded value
178 * @val: where to write the decoded value
179 * @encoding: the encoding with which we can decode @addr
181 * GCC emits encoded address in the .eh_frame FDE entries. Decode
182 * the value at @addr using @encoding. The decoded value is written
183 * to @val and the number of bytes read is returned.
185 static int dwarf_read_encoded_value(char *addr
, unsigned long *val
,
188 unsigned long decoded_addr
= 0;
191 switch (encoding
& 0x70) {
192 case DW_EH_PE_absptr
:
195 decoded_addr
= (unsigned long)addr
;
198 pr_debug("encoding=0x%x\n", (encoding
& 0x70));
202 if ((encoding
& 0x07) == 0x00)
203 encoding
|= DW_EH_PE_udata4
;
205 switch (encoding
& 0x0f) {
206 case DW_EH_PE_sdata4
:
207 case DW_EH_PE_udata4
:
209 decoded_addr
+= get_unaligned((u32
*)addr
);
210 __raw_writel(decoded_addr
, val
);
213 pr_debug("encoding=0x%x\n", encoding
);
221 * dwarf_entry_len - return the length of an FDE or CIE
222 * @addr: the address of the entry
223 * @len: the length of the entry
225 * Read the initial_length field of the entry and store the size of
226 * the entry in @len. We return the number of bytes read. Return a
227 * count of 0 on error.
229 static inline int dwarf_entry_len(char *addr
, unsigned long *len
)
234 initial_len
= get_unaligned((u32
*)addr
);
238 * An initial length field value in the range DW_LEN_EXT_LO -
239 * DW_LEN_EXT_HI indicates an extension, and should not be
240 * interpreted as a length. The only extension that we currently
241 * understand is the use of DWARF64 addresses.
243 if (initial_len
>= DW_EXT_LO
&& initial_len
<= DW_EXT_HI
) {
245 * The 64-bit length field immediately follows the
246 * compulsory 32-bit length field.
248 if (initial_len
== DW_EXT_DWARF64
) {
249 *len
= get_unaligned((u64
*)addr
+ 4);
252 printk(KERN_WARNING
"Unknown DWARF extension\n");
262 * dwarf_lookup_cie - locate the cie
263 * @cie_ptr: pointer to help with lookup
265 static struct dwarf_cie
*dwarf_lookup_cie(unsigned long cie_ptr
)
267 struct dwarf_cie
*cie
, *n
;
270 spin_lock_irqsave(&dwarf_cie_lock
, flags
);
273 * We've cached the last CIE we looked up because chances are
274 * that the FDE wants this CIE.
276 if (cached_cie
&& cached_cie
->cie_pointer
== cie_ptr
) {
281 list_for_each_entry_safe(cie
, n
, &dwarf_cie_list
, link
) {
282 if (cie
->cie_pointer
== cie_ptr
) {
288 /* Couldn't find the entry in the list. */
289 if (&cie
->link
== &dwarf_cie_list
)
292 spin_unlock_irqrestore(&dwarf_cie_lock
, flags
);
297 * dwarf_lookup_fde - locate the FDE that covers pc
298 * @pc: the program counter
300 struct dwarf_fde
*dwarf_lookup_fde(unsigned long pc
)
303 struct dwarf_fde
*fde
, *n
;
305 spin_lock_irqsave(&dwarf_fde_lock
, flags
);
306 list_for_each_entry_safe(fde
, n
, &dwarf_fde_list
, link
) {
307 unsigned long start
, end
;
309 start
= fde
->initial_location
;
310 end
= fde
->initial_location
+ fde
->address_range
;
312 if (pc
>= start
&& pc
< end
)
316 /* Couldn't find the entry in the list. */
317 if (&fde
->link
== &dwarf_fde_list
)
320 spin_unlock_irqrestore(&dwarf_fde_lock
, flags
);
326 * dwarf_cfa_execute_insns - execute instructions to calculate a CFA
327 * @insn_start: address of the first instruction
328 * @insn_end: address of the last instruction
329 * @cie: the CIE for this function
330 * @fde: the FDE for this function
331 * @frame: the instructions calculate the CFA for this frame
332 * @pc: the program counter of the address we're interested in
333 * @define_ra: keep executing insns until the return addr reg is defined?
335 * Execute the Call Frame instruction sequence starting at
336 * @insn_start and ending at @insn_end. The instructions describe
337 * how to calculate the Canonical Frame Address of a stackframe.
338 * Store the results in @frame.
340 static int dwarf_cfa_execute_insns(unsigned char *insn_start
,
341 unsigned char *insn_end
,
342 struct dwarf_cie
*cie
,
343 struct dwarf_fde
*fde
,
344 struct dwarf_frame
*frame
,
349 unsigned char *current_insn
;
350 unsigned int count
, delta
, reg
, expr_len
, offset
;
353 current_insn
= insn_start
;
356 * If we're executing instructions for the dwarf_unwind_stack()
357 * FDE we need to keep executing instructions until the value of
358 * DWARF_ARCH_RA_REG is defined. See the comment in
359 * dwarf_unwind_stack() for more details.
366 while (current_insn
< insn_end
&& (frame
->pc
<= pc
|| !seen_ra_reg
) ) {
367 insn
= __raw_readb(current_insn
++);
370 if (frame
->num_regs
>= DWARF_ARCH_RA_REG
&&
371 frame
->regs
[DWARF_ARCH_RA_REG
].flags
)
376 * Firstly, handle the opcodes that embed their operands
377 * in the instructions.
379 switch (DW_CFA_opcode(insn
)) {
380 case DW_CFA_advance_loc
:
381 delta
= DW_CFA_operand(insn
);
382 delta
*= cie
->code_alignment_factor
;
387 reg
= DW_CFA_operand(insn
);
388 count
= dwarf_read_uleb128(current_insn
, &offset
);
389 current_insn
+= count
;
390 offset
*= cie
->data_alignment_factor
;
391 dwarf_frame_alloc_regs(frame
, reg
);
392 frame
->regs
[reg
].addr
= offset
;
393 frame
->regs
[reg
].flags
|= DWARF_REG_OFFSET
;
397 reg
= DW_CFA_operand(insn
);
403 * Secondly, handle the opcodes that don't embed their
404 * operands in the instruction.
409 case DW_CFA_advance_loc1
:
410 delta
= *current_insn
++;
411 frame
->pc
+= delta
* cie
->code_alignment_factor
;
413 case DW_CFA_advance_loc2
:
414 delta
= get_unaligned((u16
*)current_insn
);
416 frame
->pc
+= delta
* cie
->code_alignment_factor
;
418 case DW_CFA_advance_loc4
:
419 delta
= get_unaligned((u32
*)current_insn
);
421 frame
->pc
+= delta
* cie
->code_alignment_factor
;
423 case DW_CFA_offset_extended
:
424 count
= dwarf_read_uleb128(current_insn
, ®
);
425 current_insn
+= count
;
426 count
= dwarf_read_uleb128(current_insn
, &offset
);
427 current_insn
+= count
;
428 offset
*= cie
->data_alignment_factor
;
430 case DW_CFA_restore_extended
:
431 count
= dwarf_read_uleb128(current_insn
, ®
);
432 current_insn
+= count
;
434 case DW_CFA_undefined
:
435 count
= dwarf_read_uleb128(current_insn
, ®
);
436 current_insn
+= count
;
439 count
= dwarf_read_uleb128(current_insn
,
440 &frame
->cfa_register
);
441 current_insn
+= count
;
442 count
= dwarf_read_uleb128(current_insn
,
444 current_insn
+= count
;
446 frame
->flags
|= DWARF_FRAME_CFA_REG_OFFSET
;
448 case DW_CFA_def_cfa_register
:
449 count
= dwarf_read_uleb128(current_insn
,
450 &frame
->cfa_register
);
451 current_insn
+= count
;
452 frame
->cfa_offset
= 0;
453 frame
->flags
|= DWARF_FRAME_CFA_REG_OFFSET
;
455 case DW_CFA_def_cfa_offset
:
456 count
= dwarf_read_uleb128(current_insn
, &offset
);
457 current_insn
+= count
;
458 frame
->cfa_offset
= offset
;
460 case DW_CFA_def_cfa_expression
:
461 count
= dwarf_read_uleb128(current_insn
, &expr_len
);
462 current_insn
+= count
;
464 frame
->cfa_expr
= current_insn
;
465 frame
->cfa_expr_len
= expr_len
;
466 current_insn
+= expr_len
;
468 frame
->flags
|= DWARF_FRAME_CFA_REG_EXP
;
470 case DW_CFA_offset_extended_sf
:
471 count
= dwarf_read_uleb128(current_insn
, ®
);
472 current_insn
+= count
;
473 count
= dwarf_read_leb128(current_insn
, &offset
);
474 current_insn
+= count
;
475 offset
*= cie
->data_alignment_factor
;
476 dwarf_frame_alloc_regs(frame
, reg
);
477 frame
->regs
[reg
].flags
|= DWARF_REG_OFFSET
;
478 frame
->regs
[reg
].addr
= offset
;
480 case DW_CFA_val_offset
:
481 count
= dwarf_read_uleb128(current_insn
, ®
);
482 current_insn
+= count
;
483 count
= dwarf_read_leb128(current_insn
, &offset
);
484 offset
*= cie
->data_alignment_factor
;
485 frame
->regs
[reg
].flags
|= DWARF_REG_OFFSET
;
486 frame
->regs
[reg
].addr
= offset
;
489 pr_debug("unhandled DWARF instruction 0x%x\n", insn
);
498 * dwarf_unwind_stack - recursively unwind the stack
499 * @pc: address of the function to unwind
500 * @prev: struct dwarf_frame of the previous stackframe on the callstack
502 * Return a struct dwarf_frame representing the most recent frame
503 * on the callstack. Each of the lower (older) stack frames are
504 * linked via the "prev" member.
506 struct dwarf_frame
*dwarf_unwind_stack(unsigned long pc
,
507 struct dwarf_frame
*prev
)
509 struct dwarf_frame
*frame
;
510 struct dwarf_cie
*cie
;
511 struct dwarf_fde
*fde
;
514 bool define_ra
= false;
517 * If this is the first invocation of this recursive function we
518 * need get the contents of a physical register to get the CFA
519 * in order to begin the virtual unwinding of the stack.
521 * Setting "define_ra" to true indictates that we want
522 * dwarf_cfa_execute_insns() to continue executing instructions
523 * until we know how to calculate the value of DWARF_ARCH_RA_REG
524 * (which we need in order to kick off the whole unwinding
527 * NOTE: the return address is guaranteed to be setup by the
528 * time this function makes its first function call.
531 pc
= (unsigned long)&dwarf_unwind_stack
;
535 frame
= kzalloc(sizeof(*frame
), GFP_ATOMIC
);
541 fde
= dwarf_lookup_fde(pc
);
544 * This is our normal exit path - the one that stops the
545 * recursion. There's two reasons why we might exit
548 * a) pc has no asscociated DWARF frame info and so
549 * we don't know how to unwind this frame. This is
550 * usually the case when we're trying to unwind a
551 * frame that was called from some assembly code
552 * that has no DWARF info, e.g. syscalls.
554 * b) the DEBUG info for pc is bogus. There's
555 * really no way to distinguish this case from the
556 * case above, which sucks because we could print a
562 cie
= dwarf_lookup_cie(fde
->cie_pointer
);
564 frame
->pc
= fde
->initial_location
;
566 /* CIE initial instructions */
567 dwarf_cfa_execute_insns(cie
->initial_instructions
,
568 cie
->instructions_end
, cie
, fde
,
571 /* FDE instructions */
572 dwarf_cfa_execute_insns(fde
->instructions
, fde
->end
, cie
,
573 fde
, frame
, pc
, define_ra
);
575 /* Calculate the CFA */
576 switch (frame
->flags
) {
577 case DWARF_FRAME_CFA_REG_OFFSET
:
579 BUG_ON(!prev
->regs
[frame
->cfa_register
].flags
);
582 addr
+= prev
->regs
[frame
->cfa_register
].addr
;
583 frame
->cfa
= __raw_readl(addr
);
587 * Again, this is the first invocation of this
588 * recurisve function. We need to physically
589 * read the contents of a register in order to
590 * get the Canonical Frame Address for this
593 frame
->cfa
= dwarf_read_arch_reg(frame
->cfa_register
);
596 frame
->cfa
+= frame
->cfa_offset
;
602 /* If we haven't seen the return address reg, we're screwed. */
603 BUG_ON(!frame
->regs
[DWARF_ARCH_RA_REG
].flags
);
605 for (i
= 0; i
<= frame
->num_regs
; i
++) {
606 struct dwarf_reg
*reg
= &frame
->regs
[i
];
612 offset
+= frame
->cfa
;
615 addr
= frame
->cfa
+ frame
->regs
[DWARF_ARCH_RA_REG
].addr
;
616 frame
->return_addr
= __raw_readl(addr
);
618 frame
->next
= dwarf_unwind_stack(frame
->return_addr
, frame
);
622 static int dwarf_parse_cie(void *entry
, void *p
, unsigned long len
,
625 struct dwarf_cie
*cie
;
629 cie
= kzalloc(sizeof(*cie
), GFP_KERNEL
);
636 * Record the offset into the .eh_frame section
637 * for this CIE. It allows this CIE to be
638 * quickly and easily looked up from the
641 cie
->cie_pointer
= (unsigned long)entry
;
643 cie
->version
= *(char *)p
++;
644 BUG_ON(cie
->version
!= 1);
646 cie
->augmentation
= p
;
647 p
+= strlen(cie
->augmentation
) + 1;
649 count
= dwarf_read_uleb128(p
, &cie
->code_alignment_factor
);
652 count
= dwarf_read_leb128(p
, &cie
->data_alignment_factor
);
656 * Which column in the rule table contains the
659 if (cie
->version
== 1) {
660 cie
->return_address_reg
= __raw_readb(p
);
663 count
= dwarf_read_uleb128(p
, &cie
->return_address_reg
);
667 if (cie
->augmentation
[0] == 'z') {
668 unsigned int length
, count
;
669 cie
->flags
|= DWARF_CIE_Z_AUGMENTATION
;
671 count
= dwarf_read_uleb128(p
, &length
);
674 BUG_ON((unsigned char *)p
> end
);
676 cie
->initial_instructions
= p
+ length
;
680 while (*cie
->augmentation
) {
682 * "L" indicates a byte showing how the
683 * LSDA pointer is encoded. Skip it.
685 if (*cie
->augmentation
== 'L') {
688 } else if (*cie
->augmentation
== 'R') {
690 * "R" indicates a byte showing
691 * how FDE addresses are
694 cie
->encoding
= *(char *)p
++;
696 } else if (*cie
->augmentation
== 'P') {
698 * "R" indicates a personality
703 } else if (*cie
->augmentation
== 'S') {
707 * Unknown augmentation. Assume
710 p
= cie
->initial_instructions
;
716 cie
->initial_instructions
= p
;
717 cie
->instructions_end
= end
;
720 spin_lock_irqsave(&dwarf_cie_lock
, flags
);
721 list_add_tail(&cie
->link
, &dwarf_cie_list
);
722 spin_unlock_irqrestore(&dwarf_cie_lock
, flags
);
727 static int dwarf_parse_fde(void *entry
, u32 entry_type
,
728 void *start
, unsigned long len
)
730 struct dwarf_fde
*fde
;
731 struct dwarf_cie
*cie
;
736 fde
= kzalloc(sizeof(*fde
), GFP_KERNEL
);
743 * In a .eh_frame section the CIE pointer is the
744 * delta between the address within the FDE
746 fde
->cie_pointer
= (unsigned long)(p
- entry_type
- 4);
748 cie
= dwarf_lookup_cie(fde
->cie_pointer
);
752 count
= dwarf_read_encoded_value(p
, &fde
->initial_location
,
755 count
= dwarf_read_addr(p
, &fde
->initial_location
);
760 count
= dwarf_read_encoded_value(p
, &fde
->address_range
,
761 cie
->encoding
& 0x0f);
763 count
= dwarf_read_addr(p
, &fde
->address_range
);
767 if (fde
->cie
->flags
& DWARF_CIE_Z_AUGMENTATION
) {
769 count
= dwarf_read_uleb128(p
, &length
);
773 /* Call frame instructions. */
774 fde
->instructions
= p
;
775 fde
->end
= start
+ len
;
778 spin_lock_irqsave(&dwarf_fde_lock
, flags
);
779 list_add_tail(&fde
->link
, &dwarf_fde_list
);
780 spin_unlock_irqrestore(&dwarf_fde_lock
, flags
);
785 static void dwarf_unwinder_dump(struct task_struct
*task
, struct pt_regs
*regs
,
787 const struct stacktrace_ops
*ops
, void *data
)
789 struct dwarf_frame
*frame
;
791 frame
= dwarf_unwind_stack(0, NULL
);
793 while (frame
&& frame
->return_addr
) {
794 ops
->address(data
, frame
->return_addr
, 1);
799 static struct unwinder dwarf_unwinder
= {
800 .name
= "dwarf-unwinder",
801 .dump
= dwarf_unwinder_dump
,
805 static void dwarf_unwinder_cleanup(void)
807 struct dwarf_cie
*cie
, *m
;
808 struct dwarf_fde
*fde
, *n
;
812 * Deallocate all the memory allocated for the DWARF unwinder.
813 * Traverse all the FDE/CIE lists and remove and free all the
814 * memory associated with those data structures.
816 spin_lock_irqsave(&dwarf_cie_lock
, flags
);
817 list_for_each_entry_safe(cie
, m
, &dwarf_cie_list
, link
)
819 spin_unlock_irqrestore(&dwarf_cie_lock
, flags
);
821 spin_lock_irqsave(&dwarf_fde_lock
, flags
);
822 list_for_each_entry_safe(fde
, n
, &dwarf_fde_list
, link
)
824 spin_unlock_irqrestore(&dwarf_fde_lock
, flags
);
828 * dwarf_unwinder_init - initialise the dwarf unwinder
830 * Build the data structures describing the .dwarf_frame section to
831 * make it easier to lookup CIE and FDE entries. Because the
832 * .eh_frame section is packed as tightly as possible it is not
833 * easy to lookup the FDE for a given PC, so we build a list of FDE
834 * and CIE entries that make it easier.
836 void dwarf_unwinder_init(void)
842 unsigned int c_entries
, f_entries
;
844 INIT_LIST_HEAD(&dwarf_cie_list
);
845 INIT_LIST_HEAD(&dwarf_fde_list
);
849 entry
= &__start_eh_frame
;
851 while ((char *)entry
< __stop_eh_frame
) {
854 count
= dwarf_entry_len(p
, &len
);
857 * We read a bogus length field value. There is
858 * nothing we can do here apart from disabling
859 * the DWARF unwinder. We can't even skip this
860 * entry and move to the next one because 'len'
861 * tells us where our next entry is.
867 /* initial length does not include itself */
870 entry_type
= get_unaligned((u32
*)p
);
873 if (entry_type
== DW_EH_FRAME_CIE
) {
874 err
= dwarf_parse_cie(entry
, p
, len
, end
);
880 err
= dwarf_parse_fde(entry
, entry_type
, p
, len
);
887 entry
= (char *)entry
+ len
+ 4;
890 printk(KERN_INFO
"DWARF unwinder initialised: read %u CIEs, %u FDEs\n",
891 c_entries
, f_entries
);
893 err
= unwinder_register(&dwarf_unwinder
);
900 printk(KERN_ERR
"Failed to initialise DWARF unwinder: %d\n", err
);
901 dwarf_unwinder_cleanup();