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
334 * Execute the Call Frame instruction sequence starting at
335 * @insn_start and ending at @insn_end. The instructions describe
336 * how to calculate the Canonical Frame Address of a stackframe.
337 * Store the results in @frame.
339 static int dwarf_cfa_execute_insns(unsigned char *insn_start,
340 unsigned char *insn_end,
341 struct dwarf_cie *cie,
342 struct dwarf_fde *fde,
343 struct dwarf_frame *frame,
347 unsigned char *current_insn;
348 unsigned int count, delta, reg, expr_len, offset;
350 current_insn = insn_start;
352 while (current_insn < insn_end && frame->pc <= pc) {
353 insn = __raw_readb(current_insn++);
356 * Firstly, handle the opcodes that embed their operands
357 * in the instructions.
359 switch (DW_CFA_opcode(insn)) {
360 case DW_CFA_advance_loc:
361 delta = DW_CFA_operand(insn);
362 delta *= cie->code_alignment_factor;
367 reg = DW_CFA_operand(insn);
368 count = dwarf_read_uleb128(current_insn, &offset);
369 current_insn += count;
370 offset *= cie->data_alignment_factor;
371 dwarf_frame_alloc_regs(frame, reg);
372 frame->regs[reg].addr = offset;
373 frame->regs[reg].flags |= DWARF_REG_OFFSET;
377 reg = DW_CFA_operand(insn);
383 * Secondly, handle the opcodes that don't embed their
384 * operands in the instruction.
389 case DW_CFA_advance_loc1:
390 delta = *current_insn++;
391 frame->pc += delta * cie->code_alignment_factor;
393 case DW_CFA_advance_loc2:
394 delta = get_unaligned((u16 *)current_insn);
396 frame->pc += delta * cie->code_alignment_factor;
398 case DW_CFA_advance_loc4:
399 delta = get_unaligned((u32 *)current_insn);
401 frame->pc += delta * cie->code_alignment_factor;
403 case DW_CFA_offset_extended:
404 count = dwarf_read_uleb128(current_insn, ®);
405 current_insn += count;
406 count = dwarf_read_uleb128(current_insn, &offset);
407 current_insn += count;
408 offset *= cie->data_alignment_factor;
410 case DW_CFA_restore_extended:
411 count = dwarf_read_uleb128(current_insn, ®);
412 current_insn += count;
414 case DW_CFA_undefined:
415 count = dwarf_read_uleb128(current_insn, ®);
416 current_insn += count;
419 count = dwarf_read_uleb128(current_insn,
420 &frame->cfa_register);
421 current_insn += count;
422 count = dwarf_read_uleb128(current_insn,
424 current_insn += count;
426 frame->flags |= DWARF_FRAME_CFA_REG_OFFSET;
428 case DW_CFA_def_cfa_register:
429 count = dwarf_read_uleb128(current_insn,
430 &frame->cfa_register);
431 current_insn += count;
432 frame->cfa_offset = 0;
433 frame->flags |= DWARF_FRAME_CFA_REG_OFFSET;
435 case DW_CFA_def_cfa_offset:
436 count = dwarf_read_uleb128(current_insn, &offset);
437 current_insn += count;
438 frame->cfa_offset = offset;
440 case DW_CFA_def_cfa_expression:
441 count = dwarf_read_uleb128(current_insn, &expr_len);
442 current_insn += count;
444 frame->cfa_expr = current_insn;
445 frame->cfa_expr_len = expr_len;
446 current_insn += expr_len;
448 frame->flags |= DWARF_FRAME_CFA_REG_EXP;
450 case DW_CFA_offset_extended_sf:
451 count = dwarf_read_uleb128(current_insn, ®);
452 current_insn += count;
453 count = dwarf_read_leb128(current_insn, &offset);
454 current_insn += count;
455 offset *= cie->data_alignment_factor;
456 dwarf_frame_alloc_regs(frame, reg);
457 frame->regs[reg].flags |= DWARF_REG_OFFSET;
458 frame->regs[reg].addr = offset;
460 case DW_CFA_val_offset:
461 count = dwarf_read_uleb128(current_insn, ®);
462 current_insn += count;
463 count = dwarf_read_leb128(current_insn, &offset);
464 offset *= cie->data_alignment_factor;
465 frame->regs[reg].flags |= DWARF_REG_OFFSET;
466 frame->regs[reg].addr = offset;
469 pr_debug("unhandled DWARF instruction 0x%x\n", insn);
478 * dwarf_unwind_stack - recursively unwind the stack
479 * @pc: address of the function to unwind
480 * @prev: struct dwarf_frame of the previous stackframe on the callstack
482 * Return a struct dwarf_frame representing the most recent frame
483 * on the callstack. Each of the lower (older) stack frames are
484 * linked via the "prev" member.
486 struct dwarf_frame *dwarf_unwind_stack(unsigned long pc,
487 struct dwarf_frame *prev)
489 struct dwarf_frame *frame;
490 struct dwarf_cie *cie;
491 struct dwarf_fde *fde;
496 * If this is the first invocation of this recursive function we
497 * need get the contents of a physical register to get the CFA
498 * in order to begin the virtual unwinding of the stack.
500 * NOTE: the return address is guaranteed to be setup by the
501 * time this function makes its first function call.
504 pc = (unsigned long)current_text_addr();
506 frame = kzalloc(sizeof(*frame), GFP_ATOMIC);
512 fde = dwarf_lookup_fde(pc);
515 * This is our normal exit path - the one that stops the
516 * recursion. There's two reasons why we might exit
519 * a) pc has no asscociated DWARF frame info and so
520 * we don't know how to unwind this frame. This is
521 * usually the case when we're trying to unwind a
522 * frame that was called from some assembly code
523 * that has no DWARF info, e.g. syscalls.
525 * b) the DEBUG info for pc is bogus. There's
526 * really no way to distinguish this case from the
527 * case above, which sucks because we could print a
533 cie = dwarf_lookup_cie(fde->cie_pointer);
535 frame->pc = fde->initial_location;
537 /* CIE initial instructions */
538 dwarf_cfa_execute_insns(cie->initial_instructions,
539 cie->instructions_end, cie, fde,
542 /* FDE instructions */
543 dwarf_cfa_execute_insns(fde->instructions, fde->end, cie,
546 /* Calculate the CFA */
547 switch (frame->flags) {
548 case DWARF_FRAME_CFA_REG_OFFSET:
550 BUG_ON(!prev->regs[frame->cfa_register].flags);
553 addr += prev->regs[frame->cfa_register].addr;
554 frame->cfa = __raw_readl(addr);
558 * Again, this is the first invocation of this
559 * recurisve function. We need to physically
560 * read the contents of a register in order to
561 * get the Canonical Frame Address for this
564 frame->cfa = dwarf_read_arch_reg(frame->cfa_register);
567 frame->cfa += frame->cfa_offset;
573 /* If we haven't seen the return address reg, we're screwed. */
574 BUG_ON(!frame->regs[DWARF_ARCH_RA_REG].flags);
576 for (i = 0; i <= frame->num_regs; i++) {
577 struct dwarf_reg *reg = &frame->regs[i];
583 offset += frame->cfa;
586 addr = frame->cfa + frame->regs[DWARF_ARCH_RA_REG].addr;
587 frame->return_addr = __raw_readl(addr);
589 frame->next = dwarf_unwind_stack(frame->return_addr, frame);
593 static int dwarf_parse_cie(void *entry, void *p, unsigned long len,
596 struct dwarf_cie *cie;
600 cie = kzalloc(sizeof(*cie), GFP_KERNEL);
607 * Record the offset into the .eh_frame section
608 * for this CIE. It allows this CIE to be
609 * quickly and easily looked up from the
612 cie->cie_pointer = (unsigned long)entry;
614 cie->version = *(char *)p++;
615 BUG_ON(cie->version != 1);
617 cie->augmentation = p;
618 p += strlen(cie->augmentation) + 1;
620 count = dwarf_read_uleb128(p, &cie->code_alignment_factor);
623 count = dwarf_read_leb128(p, &cie->data_alignment_factor);
627 * Which column in the rule table contains the
630 if (cie->version == 1) {
631 cie->return_address_reg = __raw_readb(p);
634 count = dwarf_read_uleb128(p, &cie->return_address_reg);
638 if (cie->augmentation[0] == 'z') {
639 unsigned int length, count;
640 cie->flags |= DWARF_CIE_Z_AUGMENTATION;
642 count = dwarf_read_uleb128(p, &length);
645 BUG_ON((unsigned char *)p > end);
647 cie->initial_instructions = p + length;
651 while (*cie->augmentation) {
653 * "L" indicates a byte showing how the
654 * LSDA pointer is encoded. Skip it.
656 if (*cie->augmentation == 'L') {
659 } else if (*cie->augmentation == 'R') {
661 * "R" indicates a byte showing
662 * how FDE addresses are
665 cie->encoding = *(char *)p++;
667 } else if (*cie->augmentation == 'P') {
669 * "R" indicates a personality
674 } else if (*cie->augmentation == 'S') {
678 * Unknown augmentation. Assume
681 p = cie->initial_instructions;
687 cie->initial_instructions = p;
688 cie->instructions_end = end;
691 spin_lock_irqsave(&dwarf_cie_lock, flags);
692 list_add_tail(&cie->link, &dwarf_cie_list);
693 spin_unlock_irqrestore(&dwarf_cie_lock, flags);
698 static int dwarf_parse_fde(void *entry, u32 entry_type,
699 void *start, unsigned long len)
701 struct dwarf_fde *fde;
702 struct dwarf_cie *cie;
707 fde = kzalloc(sizeof(*fde), GFP_KERNEL);
714 * In a .eh_frame section the CIE pointer is the
715 * delta between the address within the FDE
717 fde->cie_pointer = (unsigned long)(p - entry_type - 4);
719 cie = dwarf_lookup_cie(fde->cie_pointer);
723 count = dwarf_read_encoded_value(p, &fde->initial_location,
726 count = dwarf_read_addr(p, &fde->initial_location);
731 count = dwarf_read_encoded_value(p, &fde->address_range,
732 cie->encoding & 0x0f);
734 count = dwarf_read_addr(p, &fde->address_range);
738 if (fde->cie->flags & DWARF_CIE_Z_AUGMENTATION) {
740 count = dwarf_read_uleb128(p, &length);
744 /* Call frame instructions. */
745 fde->instructions = p;
746 fde->end = start + len;
749 spin_lock_irqsave(&dwarf_fde_lock, flags);
750 list_add_tail(&fde->link, &dwarf_fde_list);
751 spin_unlock_irqrestore(&dwarf_fde_lock, flags);
756 static void dwarf_unwinder_dump(struct task_struct *task, struct pt_regs *regs,
758 const struct stacktrace_ops *ops, void *data)
760 struct dwarf_frame *frame;
762 frame = dwarf_unwind_stack(0, NULL);
764 while (frame && frame->return_addr) {
765 ops->address(data, frame->return_addr, 1);
770 static struct unwinder dwarf_unwinder = {
771 .name = "dwarf-unwinder",
772 .dump = dwarf_unwinder_dump,
776 static void dwarf_unwinder_cleanup(void)
778 struct dwarf_cie *cie, *m;
779 struct dwarf_fde *fde, *n;
783 * Deallocate all the memory allocated for the DWARF unwinder.
784 * Traverse all the FDE/CIE lists and remove and free all the
785 * memory associated with those data structures.
787 spin_lock_irqsave(&dwarf_cie_lock, flags);
788 list_for_each_entry_safe(cie, m, &dwarf_cie_list, link)
790 spin_unlock_irqrestore(&dwarf_cie_lock, flags);
792 spin_lock_irqsave(&dwarf_fde_lock, flags);
793 list_for_each_entry_safe(fde, n, &dwarf_fde_list, link)
795 spin_unlock_irqrestore(&dwarf_fde_lock, flags);
799 * dwarf_unwinder_init - initialise the dwarf unwinder
801 * Build the data structures describing the .dwarf_frame section to
802 * make it easier to lookup CIE and FDE entries. Because the
803 * .eh_frame section is packed as tightly as possible it is not
804 * easy to lookup the FDE for a given PC, so we build a list of FDE
805 * and CIE entries that make it easier.
807 void dwarf_unwinder_init(void)
813 unsigned int c_entries, f_entries;
815 INIT_LIST_HEAD(&dwarf_cie_list);
816 INIT_LIST_HEAD(&dwarf_fde_list);
820 entry = &__start_eh_frame;
822 while ((char *)entry < __stop_eh_frame) {
825 count = dwarf_entry_len(p, &len);
828 * We read a bogus length field value. There is
829 * nothing we can do here apart from disabling
830 * the DWARF unwinder. We can't even skip this
831 * entry and move to the next one because 'len'
832 * tells us where our next entry is.
838 /* initial length does not include itself */
841 entry_type = get_unaligned((u32 *)p);
844 if (entry_type == DW_EH_FRAME_CIE) {
845 err = dwarf_parse_cie(entry, p, len, end);
851 err = dwarf_parse_fde(entry, entry_type, p, len);
858 entry = (char *)entry + len + 4;
861 printk(KERN_INFO "DWARF unwinder initialised: read %u CIEs, %u FDEs\n",
862 c_entries, f_entries);
864 err = unwinder_register(&dwarf_unwinder);
871 printk(KERN_ERR "Failed to initialise DWARF unwinder: %d\n", err);
872 dwarf_unwinder_cleanup();