sh: dwarf unwinder support.

This is a first cut at a generic DWARF unwinder for the kernel. It's
still lacking DWARF64 support and the DWARF expression support hasn't
been tested very well but it is generating proper stacktraces on SH for
WARN_ON() and NULL dereferences.

Signed-off-by: Matt Fleming <matt@console-pimps.org>
Signed-off-by: Paul Mundt <lethal@linux-sh.org>

arch/sh/Kconfig.debug

@@ -110,6 +110,14 @@ config DUMP_CODE
 
 	  Those looking for more verbose debugging output should say Y.
 
+config DWARF_UNWINDER
+	bool "Enable the DWARF unwinder for stacktraces"
+	select FRAME_POINTER
+	default n
+	help
+	  Enabling this option will make stacktraces more accurate, at
+	  the cost of an increase in overall kernel size.
+
 config SH_NO_BSS_INIT
 	bool "Avoid zeroing BSS (to speed-up startup on suitable platforms)"
 	depends on DEBUG_KERNEL

arch/sh/Makefile

@@ -191,6 +191,10 @@ ifeq ($(CONFIG_MCOUNT),y)
   KBUILD_CFLAGS += -pg
 endif
 
+ifeq ($(CONFIG_DWARF_UNWINDER),y)
+  KBUILD_CFLAGS += -fasynchronous-unwind-tables
+endif
+
 libs-$(CONFIG_SUPERH32)		:= arch/sh/lib/	$(libs-y)
 libs-$(CONFIG_SUPERH64)		:= arch/sh/lib64/ $(libs-y)
 

arch/sh/include/asm/dwarf.h

@@ -0,0 +1,407 @@
+/*
+ * Copyright (C) 2009 Matt Fleming <matt@console-pimps.org>
+ *
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ */
+#ifndef __ASM_SH_DWARF_H
+#define __ASM_SH_DWARF_H
+
+#ifdef CONFIG_DWARF_UNWINDER
+
+/*
+ * DWARF expression operations
+ */
+#define DW_OP_addr	0x03
+#define DW_OP_deref	0x06
+#define DW_OP_const1u	0x08
+#define DW_OP_const1s	0x09
+#define DW_OP_const2u	0x0a
+#define DW_OP_const2s	0x0b
+#define DW_OP_const4u	0x0c
+#define DW_OP_const4s	0x0d
+#define DW_OP_const8u	0x0e
+#define DW_OP_const8s	0x0f
+#define DW_OP_constu	0x10
+#define DW_OP_consts	0x11
+#define DW_OP_dup	0x12
+#define DW_OP_drop	0x13
+#define DW_OP_over	0x14
+#define DW_OP_pick	0x15
+#define DW_OP_swap	0x16
+#define DW_OP_rot	0x17
+#define DW_OP_xderef	0x18
+#define DW_OP_abs	0x19
+#define DW_OP_and	0x1a
+#define DW_OP_div	0x1b
+#define DW_OP_minus	0x1c
+#define DW_OP_mod	0x1d
+#define DW_OP_mul	0x1e
+#define DW_OP_neg	0x1f
+#define DW_OP_not	0x20
+#define DW_OP_or	0x21
+#define DW_OP_plus	0x22
+#define DW_OP_plus_uconst	0x23
+#define DW_OP_shl	0x24
+#define DW_OP_shr	0x25
+#define DW_OP_shra	0x26
+#define DW_OP_xor	0x27
+#define DW_OP_skip	0x2f
+#define DW_OP_bra	0x28
+#define DW_OP_eq	0x29
+#define DW_OP_ge	0x2a
+#define DW_OP_gt	0x2b
+#define DW_OP_le	0x2c
+#define DW_OP_lt	0x2d
+#define DW_OP_ne	0x2e
+#define DW_OP_lit0	0x30
+#define DW_OP_lit1	0x31
+#define DW_OP_lit2	0x32
+#define DW_OP_lit3	0x33
+#define DW_OP_lit4	0x34
+#define DW_OP_lit5	0x35
+#define DW_OP_lit6	0x36
+#define DW_OP_lit7	0x37
+#define DW_OP_lit8	0x38
+#define DW_OP_lit9	0x39
+#define DW_OP_lit10	0x3a
+#define DW_OP_lit11	0x3b
+#define DW_OP_lit12	0x3c
+#define DW_OP_lit13	0x3d
+#define DW_OP_lit14	0x3e
+#define DW_OP_lit15	0x3f
+#define DW_OP_lit16	0x40
+#define DW_OP_lit17	0x41
+#define DW_OP_lit18	0x42
+#define DW_OP_lit19	0x43
+#define DW_OP_lit20	0x44
+#define DW_OP_lit21	0x45
+#define DW_OP_lit22	0x46
+#define DW_OP_lit23	0x47
+#define DW_OP_lit24	0x48
+#define DW_OP_lit25	0x49
+#define DW_OP_lit26	0x4a
+#define DW_OP_lit27	0x4b
+#define DW_OP_lit28	0x4c
+#define DW_OP_lit29	0x4d
+#define DW_OP_lit30	0x4e
+#define DW_OP_lit31	0x4f
+#define DW_OP_reg0	0x50
+#define DW_OP_reg1	0x51
+#define DW_OP_reg2	0x52
+#define DW_OP_reg3	0x53
+#define DW_OP_reg4	0x54
+#define DW_OP_reg5	0x55
+#define DW_OP_reg6	0x56
+#define DW_OP_reg7	0x57
+#define DW_OP_reg8	0x58
+#define DW_OP_reg9	0x59
+#define DW_OP_reg10	0x5a
+#define DW_OP_reg11	0x5b
+#define DW_OP_reg12	0x5c
+#define DW_OP_reg13	0x5d
+#define DW_OP_reg14	0x5e
+#define DW_OP_reg15	0x5f
+#define DW_OP_reg16	0x60
+#define DW_OP_reg17	0x61
+#define DW_OP_reg18	0x62
+#define DW_OP_reg19	0x63
+#define DW_OP_reg20	0x64
+#define DW_OP_reg21	0x65
+#define DW_OP_reg22	0x66
+#define DW_OP_reg23	0x67
+#define DW_OP_reg24	0x68
+#define DW_OP_reg25	0x69
+#define DW_OP_reg26	0x6a
+#define DW_OP_reg27	0x6b
+#define DW_OP_reg28	0x6c
+#define DW_OP_reg29	0x6d
+#define DW_OP_reg30	0x6e
+#define DW_OP_reg31	0x6f
+#define DW_OP_breg0	0x70
+#define DW_OP_breg1	0x71
+#define DW_OP_breg2	0x72
+#define DW_OP_breg3	0x73
+#define DW_OP_breg4	0x74
+#define DW_OP_breg5	0x75
+#define DW_OP_breg6	0x76
+#define DW_OP_breg7	0x77
+#define DW_OP_breg8	0x78
+#define DW_OP_breg9	0x79
+#define DW_OP_breg10	0x7a
+#define DW_OP_breg11	0x7b
+#define DW_OP_breg12	0x7c
+#define DW_OP_breg13	0x7d
+#define DW_OP_breg14	0x7e
+#define DW_OP_breg15	0x7f
+#define DW_OP_breg16	0x80
+#define DW_OP_breg17	0x81
+#define DW_OP_breg18	0x82
+#define DW_OP_breg19	0x83
+#define DW_OP_breg20	0x84
+#define DW_OP_breg21	0x85
+#define DW_OP_breg22	0x86
+#define DW_OP_breg23	0x87
+#define DW_OP_breg24	0x88
+#define DW_OP_breg25	0x89
+#define DW_OP_breg26	0x8a
+#define DW_OP_breg27	0x8b
+#define DW_OP_breg28	0x8c
+#define DW_OP_breg29	0x8d
+#define DW_OP_breg30	0x8e
+#define DW_OP_breg31	0x8f
+#define DW_OP_regx	0x90
+#define DW_OP_fbreg	0x91
+#define DW_OP_bregx	0x92
+#define DW_OP_piece	0x93
+#define DW_OP_deref_size	0x94
+#define DW_OP_xderef_size	0x95
+#define DW_OP_nop	0x96
+#define DW_OP_push_object_address	0x97
+#define DW_OP_call2	0x98
+#define DW_OP_call4	0x99
+#define DW_OP_call_ref	0x9a
+#define DW_OP_form_tls_address	0x9b
+#define DW_OP_call_frame_cfa	0x9c
+#define DW_OP_bit_piece	0x9d
+#define DW_OP_lo_user	0xe0
+#define DW_OP_hi_user	0xff
+
+/*
+ * Addresses used in FDE entries in the .eh_frame section may be encoded
+ * using one of the following encodings.
+ */
+#define DW_EH_PE_absptr	0x00
+#define DW_EH_PE_omit	0xff
+#define DW_EH_PE_uleb128	0x01
+#define DW_EH_PE_udata2	0x02
+#define DW_EH_PE_udata4	0x03
+#define DW_EH_PE_udata8	0x04
+#define DW_EH_PE_sleb128	0x09
+#define DW_EH_PE_sdata2	0x0a
+#define DW_EH_PE_sdata4	0x0b
+#define DW_EH_PE_sdata8	0x0c
+#define DW_EH_PE_signed	0x09
+
+#define DW_EH_PE_pcrel	0x10
+
+/*
+ * The architecture-specific register number that contains the return
+ * address in the .debug_frame table.
+ */
+#define DWARF_ARCH_RA_REG	17
+
+/*
+ * At what offset into dwarf_unwind_stack() is DWARF_ARCH_RA_REG setup?
+ */
+#define DWARF_ARCH_UNWIND_OFFSET	0x20
+
+#ifndef __ASSEMBLY__
+/*
+ * Read either the frame pointer (r14) or the stack pointer (r15).
+ * NOTE: this MUST be inlined.
+ */
+static __always_inline unsigned long dwarf_read_arch_reg(unsigned int reg)
+{
+	unsigned long value;
+
+	switch (reg) {
+	case 14:
+		__asm__ __volatile__("mov r14, %0\n" : "=r" (value));
+		break;
+	case 15:
+		__asm__ __volatile__("mov r15, %0\n" : "=r" (value));
+		break;
+	default:
+		BUG();
+	}
+
+	return value;
+}
+
+/**
+ *	dwarf_cie - Common Information Entry
+ */
+struct dwarf_cie {
+	unsigned long length;
+	unsigned long cie_id;
+	unsigned char version;
+	const char *augmentation;
+	unsigned int code_alignment_factor;
+	int data_alignment_factor;
+
+	/* Which column in the rule table represents return addr of func. */
+	unsigned int return_address_reg;
+
+	unsigned char *initial_instructions;
+	unsigned char *instructions_end;
+
+	unsigned char encoding;
+
+	unsigned long cie_pointer;
+
+	struct list_head link;
+
+	unsigned long flags;
+#define DWARF_CIE_Z_AUGMENTATION	(1 << 0)
+};
+
+/**
+ *	dwarf_fde - Frame Description Entry
+ */
+struct dwarf_fde {
+	unsigned long length;
+	unsigned long cie_pointer;
+	struct dwarf_cie *cie;
+	unsigned long initial_location;
+	unsigned long address_range;
+	unsigned char *instructions;
+	unsigned char *end;
+	struct list_head link;
+};
+
+/**
+ *	dwarf_frame - DWARF information for a frame in the call stack
+ */
+struct dwarf_frame {
+	struct dwarf_frame *prev, *next;
+
+	unsigned long pc;
+
+	struct dwarf_reg *regs;
+	unsigned int num_regs;	/* how many regs are allocated? */
+
+	unsigned int depth;	/* what level are we in the callstack? */
+
+	unsigned long cfa;
+
+	/* Valid when DW_FRAME_CFA_REG_OFFSET is set in flags */
+	unsigned int cfa_register;
+	unsigned int cfa_offset;
+
+	/* Valid when DW_FRAME_CFA_REG_EXP is set in flags */
+	unsigned char *cfa_expr;
+	unsigned int cfa_expr_len;
+
+	unsigned long flags;
+#define DWARF_FRAME_CFA_REG_OFFSET	(1 << 0)
+#define DWARF_FRAME_CFA_REG_EXP		(1 << 1)
+
+	unsigned long return_addr;
+};
+
+/**
+ *	dwarf_reg - DWARF register
+ *	@flags: Describes how to calculate the value of this register
+ */
+struct dwarf_reg {
+	unsigned long addr;
+	unsigned long flags;
+#define DWARF_REG_OFFSET	(1 << 0)
+};
+
+/**
+ *	dwarf_stack - a DWARF stack contains a collection of DWARF frames
+ *	@depth: the number of frames in the stack
+ *	@level: an array of DWARF frames, indexed by stack level
+ *
+ */
+struct dwarf_stack {
+	unsigned int depth;
+	struct dwarf_frame **level;
+};
+
+/*
+ * Call Frame instruction opcodes.
+ */
+#define DW_CFA_advance_loc	0x40
+#define DW_CFA_offset		0x80
+#define DW_CFA_restore		0xc0
+#define DW_CFA_nop		0x00
+#define DW_CFA_set_loc		0x01
+#define DW_CFA_advance_loc1	0x02
+#define DW_CFA_advance_loc2	0x03
+#define DW_CFA_advance_loc4	0x04
+#define DW_CFA_offset_extended	0x05
+#define DW_CFA_restore_extended	0x06
+#define DW_CFA_undefined	0x07
+#define DW_CFA_same_value	0x08
+#define DW_CFA_register		0x09
+#define DW_CFA_remember_state	0x0a
+#define DW_CFA_restore_state	0x0b
+#define DW_CFA_def_cfa		0x0c
+#define DW_CFA_def_cfa_register	0x0d
+#define DW_CFA_def_cfa_offset	0x0e
+#define DW_CFA_def_cfa_expression	0x0f
+#define DW_CFA_expression	0x10
+#define DW_CFA_offset_extended_sf	0x11
+#define DW_CFA_def_cfa_sf	0x12
+#define DW_CFA_def_cfa_offset_sf	0x13
+#define DW_CFA_val_offset	0x14
+#define DW_CFA_val_offset_sf	0x15
+#define DW_CFA_val_expression	0x16
+#define DW_CFA_lo_user		0x1c
+#define DW_CFA_hi_user		0x3f
+
+/*
+ * Some call frame instructions encode their operands in the opcode. We
+ * need some helper functions to extract both the opcode and operands
+ * from an instruction.
+ */
+static inline unsigned int DW_CFA_opcode(unsigned long insn)
+{
+	return (insn & 0xc0);
+}
+
+static inline unsigned int DW_CFA_operand(unsigned long insn)
+{
+	return (insn & 0x3f);
+}
+
+#define DW_EH_FRAME_CIE	0		/* .eh_frame CIE IDs are 0 */
+#define DW_CIE_ID	0xffffffff
+#define DW64_CIE_ID	0xffffffffffffffffULL
+
+/*
+ * DWARF FDE/CIE length field values.
+ */
+#define DW_EXT_LO	0xfffffff0
+#define DW_EXT_HI	0xffffffff
+#define DW_EXT_DWARF64	DW_EXT_HI
+
+extern void dwarf_unwinder_init(void);
+
+extern struct dwarf_frame *dwarf_unwind_stack(unsigned long,
+					      struct dwarf_frame *);
+#endif /* __ASSEMBLY__ */
+
+#define CFI_STARTPROC	.cfi_startproc
+#define CFI_ENDPROC	.cfi_endproc
+#define CFI_DEF_CFA	.cfi_def_cfa
+#define CFI_REGISTER	.cfi_register
+#define CFI_REL_OFFSET	.cfi_rel_offset
+
+#else
+
+/*
+ * Use the asm comment character to ignore the rest of the line.
+ */
+#define CFI_IGNORE	!
+
+#define CFI_STARTPROC	CFI_IGNORE
+#define CFI_ENDPROC	CFI_IGNORE
+#define CFI_DEF_CFA	CFI_IGNORE
+#define CFI_REGISTER	CFI_IGNORE
+#define CFI_REL_OFFSET	CFI_IGNORE
+
+#ifndef __ASSEMBLY__
+static inline void dwarf_unwinder_init(void)
+{
+}
+#endif
+
+#endif /* CONFIG_DWARF_UNWINDER */
+
+#endif /* __ASM_SH_DWARF_H */

arch/sh/include/asm/sections.h

@@ -7,6 +7,7 @@ extern void __nosave_begin, __nosave_end;
 extern long __machvec_start, __machvec_end;
 extern char __uncached_start, __uncached_end;
 extern char _ebss[];
+extern char __start_eh_frame[], __stop_eh_frame[];
 
 #endif /* __ASM_SH_SECTIONS_H */
 

arch/sh/include/asm/vmlinux.lds.h

@@ -0,0 +1,17 @@
+#ifndef __ASM_SH_VMLINUX_LDS_H
+#define __ASM_SH_VMLINUX_LDS_H
+
+#include <asm-generic/vmlinux.lds.h>
+
+#ifdef CONFIG_DWARF_UNWINDER
+#define DWARF_EH_FRAME							\
+	.eh_frame : AT(ADDR(.eh_frame) - LOAD_OFFSET) {			\
+		  VMLINUX_SYMBOL(__start_eh_frame) = .;			\
+		  *(.eh_frame)						\
+		  VMLINUX_SYMBOL(__stop_eh_frame) = .;			\
+	}
+#else
+#define DWARF_EH_FRAME
+#endif
+
+#endif /* __ASM_SH_VMLINUX_LDS_H */

arch/sh/kernel/Makefile_32

@@ -33,6 +33,7 @@ obj-$(CONFIG_FTRACE_SYSCALLS)	+= ftrace.o
 obj-$(CONFIG_FUNCTION_GRAPH_TRACER) += ftrace.o
 obj-$(CONFIG_DUMP_CODE)		+= disassemble.o
 obj-$(CONFIG_HIBERNATION)	+= swsusp.o
+obj-$(CONFIG_DWARF_UNWINDER)	+= dwarf.o
 
 obj-$(CONFIG_GENERIC_CLOCKEVENTS_BROADCAST)	+= localtimer.o
 

arch/sh/kernel/Makefile_64

@@ -13,6 +13,7 @@ obj-$(CONFIG_CRASH_DUMP)	+= crash_dump.o
 obj-$(CONFIG_STACKTRACE)	+= stacktrace.o
 obj-$(CONFIG_IO_TRAPPED)	+= io_trapped.o
 obj-$(CONFIG_GENERIC_GPIO)	+= gpio.o
+obj-$(CONFIG_DWARF_UNWINDER)	+= dwarf.o
 
 obj-$(CONFIG_GENERIC_CLOCKEVENTS_BROADCAST)	+= localtimer.o
 

arch/sh/kernel/dwarf.c

@@ -0,0 +1,876 @@
+/*
+ * Copyright (C) 2009 Matt Fleming <matt@console-pimps.org>
+ *
+ * This file is subject to the terms and conditions of the GNU General Public
+ * License.  See the file "COPYING" in the main directory of this archive
+ * for more details.
+ *
+ * This is an implementation of a DWARF unwinder. Its main purpose is
+ * for generating stacktrace information. Based on the DWARF 3
+ * specification from http://www.dwarfstd.org.
+ *
+ * TODO:
+ *	- DWARF64 doesn't work.
+ */
+
+/* #define DEBUG */
+#include <linux/kernel.h>
+#include <linux/io.h>
+#include <linux/list.h>
+#include <linux/mm.h>
+#include <asm/dwarf.h>
+#include <asm/unwinder.h>
+#include <asm/sections.h>
+#include <asm-generic/unaligned.h>
+#include <asm/dwarf.h>
+#include <asm/stacktrace.h>
+
+static LIST_HEAD(dwarf_cie_list);
+DEFINE_SPINLOCK(dwarf_cie_lock);
+
+static LIST_HEAD(dwarf_fde_list);
+DEFINE_SPINLOCK(dwarf_fde_lock);
+
+static struct dwarf_cie *cached_cie;
+
+/*
+ * Figure out whether we need to allocate some dwarf registers. If dwarf
+ * registers have already been allocated then we may need to realloc
+ * them. "reg" is a register number that we need to be able to access
+ * after this call.
+ *
+ * Register numbers start at zero, therefore we need to allocate space
+ * for "reg" + 1 registers.
+ */
+static void dwarf_frame_alloc_regs(struct dwarf_frame *frame,
+				   unsigned int reg)
+{
+	struct dwarf_reg *regs;
+	unsigned int num_regs = reg + 1;
+	size_t new_size;
+	size_t old_size;
+
+	new_size = num_regs * sizeof(*regs);
+	old_size = frame->num_regs * sizeof(*regs);
+
+	/* Fast path: don't allocate any regs if we've already got enough. */
+	if (frame->num_regs >= num_regs)
+		return;
+
+	regs = kzalloc(new_size, GFP_KERNEL);
+	if (!regs) {
+		printk(KERN_WARNING "Unable to allocate DWARF registers\n");
+		/*
+		 * Let's just bomb hard here, we have no way to
+		 * gracefully recover.
+		 */
+		BUG();
+	}
+
+	if (frame->regs) {
+		memcpy(regs, frame->regs, old_size);
+		kfree(frame->regs);
+	}
+
+	frame->regs = regs;
+	frame->num_regs = num_regs;
+}
+
+/**
+ *	dwarf_read_addr - read dwarf data
+ *	@src: source address of data
+ *	@dst: destination address to store the data to
+ *
+ *	Read 'n' bytes from @src, where 'n' is the size of an address on
+ *	the native machine. We return the number of bytes read, which
+ *	should always be 'n'. We also have to be careful when reading
+ *	from @src and writing to @dst, because they can be arbitrarily
+ *	aligned. Return 'n' - the number of bytes read.
+ */
+static inline int dwarf_read_addr(void *src, void *dst)
+{
+	u32 val = __get_unaligned_cpu32(src);
+	__put_unaligned_cpu32(val, dst);
+
+	return sizeof(unsigned long *);
+}
+
+/**
+ *	dwarf_read_uleb128 - read unsigned LEB128 data
+ *	@addr: the address where the ULEB128 data is stored
+ *	@ret: address to store the result
+ *
+ *	Decode an unsigned LEB128 encoded datum. The algorithm is taken
+ *	from Appendix C of the DWARF 3 spec. For information on the
+ *	encodings refer to section "7.6 - Variable Length Data". Return
+ *	the number of bytes read.
+ */
+static inline unsigned long dwarf_read_uleb128(char *addr, unsigned int *ret)
+{
+	unsigned int result;
+	unsigned char byte;
+	int shift, count;
+
+	result = 0;
+	shift = 0;
+	count = 0;
+
+	while (1) {
+		byte = __raw_readb(addr);
+		addr++;
+		count++;
+
+		result |= (byte & 0x7f) << shift;
+		shift += 7;
+
+		if (!(byte & 0x80))
+			break;
+	}
+
+	*ret = result;
+
+	return count;
+}
+
+/**
+ *	dwarf_read_leb128 - read signed LEB128 data
+ *	@addr: the address of the LEB128 encoded data
+ *	@ret: address to store the result
+ *
+ *	Decode signed LEB128 data. The algorithm is taken from Appendix
+ *	C of the DWARF 3 spec. Return the number of bytes read.
+ */
+static inline unsigned long dwarf_read_leb128(char *addr, int *ret)
+{
+	unsigned char byte;
+	int result, shift;
+	int num_bits;
+	int count;
+
+	result = 0;
+	shift = 0;
+	count = 0;
+
+	while (1) {
+		byte = __raw_readb(addr);
+		addr++;
+		result |= (byte & 0x7f) << shift;
+		shift += 7;
+		count++;
+
+		if (!(byte & 0x80))
+			break;
+	}
+
+	/* The number of bits in a signed integer. */
+	num_bits = 8 * sizeof(result);
+
+	if ((shift < num_bits) && (byte & 0x40))
+		result |= (-1 << shift);
+
+	*ret = result;
+
+	return count;
+}
+
+/**
+ *	dwarf_read_encoded_value - return the decoded value at @addr
+ *	@addr: the address of the encoded value
+ *	@val: where to write the decoded value
+ *	@encoding: the encoding with which we can decode @addr
+ *
+ *	GCC emits encoded address in the .eh_frame FDE entries. Decode
+ *	the value at @addr using @encoding. The decoded value is written
+ *	to @val and the number of bytes read is returned.
+ */
+static int dwarf_read_encoded_value(char *addr, unsigned long *val,
+				    char encoding)
+{
+	unsigned long decoded_addr = 0;
+	int count = 0;
+
+	switch (encoding & 0x70) {
+	case DW_EH_PE_absptr:
+		break;
+	case DW_EH_PE_pcrel:
+		decoded_addr = (unsigned long)addr;
+		break;
+	default:
+		pr_debug("encoding=0x%x\n", (encoding & 0x70));
+		BUG();
+	}
+
+	if ((encoding & 0x07) == 0x00)
+		encoding |= DW_EH_PE_udata4;
+
+	switch (encoding & 0x0f) {
+	case DW_EH_PE_sdata4:
+	case DW_EH_PE_udata4:
+		count += 4;
+		decoded_addr += __get_unaligned_cpu32(addr);
+		__raw_writel(decoded_addr, val);
+		break;
+	default:
+		pr_debug("encoding=0x%x\n", encoding);
+		BUG();
+	}
+
+	return count;
+}
+
+/**
+ *	dwarf_entry_len - return the length of an FDE or CIE
+ *	@addr: the address of the entry
+ *	@len: the length of the entry
+ *
+ *	Read the initial_length field of the entry and store the size of
+ *	the entry in @len. We return the number of bytes read. Return a
+ *	count of 0 on error.
+ */
+static inline int dwarf_entry_len(char *addr, unsigned long *len)
+{
+	u32 initial_len;
+	int count;
+
+	initial_len = __get_unaligned_cpu32(addr);
+	count = 4;
+
+	/*
+	 * An initial length field value in the range DW_LEN_EXT_LO -
+	 * DW_LEN_EXT_HI indicates an extension, and should not be
+	 * interpreted as a length. The only extension that we currently
+	 * understand is the use of DWARF64 addresses.
+	 */
+	if (initial_len >= DW_EXT_LO && initial_len <= DW_EXT_HI) {
+		/*
+		 * The 64-bit length field immediately follows the
+		 * compulsory 32-bit length field.
+		 */
+		if (initial_len == DW_EXT_DWARF64) {
+			*len = __get_unaligned_cpu64(addr + 4);
+			count = 12;
+		} else {
+			printk(KERN_WARNING "Unknown DWARF extension\n");
+			count = 0;
+		}
+	} else
+		*len = initial_len;
+
+	return count;
+}
+
+/**
+ *	dwarf_lookup_cie - locate the cie
+ *	@cie_ptr: pointer to help with lookup
+ */
+static struct dwarf_cie *dwarf_lookup_cie(unsigned long cie_ptr)
+{
+	struct dwarf_cie *cie, *n;
+	unsigned long flags;
+
+	spin_lock_irqsave(&dwarf_cie_lock, flags);
+
+	/*
+	 * We've cached the last CIE we looked up because chances are
+	 * that the FDE wants this CIE.
+	 */
+	if (cached_cie && cached_cie->cie_pointer == cie_ptr) {
+		cie = cached_cie;
+		goto out;
+	}
+
+	list_for_each_entry_safe(cie, n, &dwarf_cie_list, link) {
+		if (cie->cie_pointer == cie_ptr) {
+			cached_cie = cie;
+			break;
+		}
+	}
+
+	/* Couldn't find the entry in the list. */
+	if (&cie->link == &dwarf_cie_list)
+		cie = NULL;
+out:
+	spin_unlock_irqrestore(&dwarf_cie_lock, flags);
+	return cie;
+}
+
+/**
+ *	dwarf_lookup_fde - locate the FDE that covers pc
+ *	@pc: the program counter
+ */
+struct dwarf_fde *dwarf_lookup_fde(unsigned long pc)
+{
+	unsigned long flags;
+	struct dwarf_fde *fde, *n;
+
+	spin_lock_irqsave(&dwarf_fde_lock, flags);
+	list_for_each_entry_safe(fde, n, &dwarf_fde_list, link) {
+		unsigned long start, end;
+
+		start = fde->initial_location;
+		end = fde->initial_location + fde->address_range;
+
+		if (pc >= start && pc < end)
+			break;
+	}
+
+	/* Couldn't find the entry in the list. */
+	if (&fde->link == &dwarf_fde_list)
+		fde = NULL;
+
+	spin_unlock_irqrestore(&dwarf_fde_lock, flags);
+
+	return fde;
+}
+
+/**
+ *	dwarf_cfa_execute_insns - execute instructions to calculate a CFA
+ *	@insn_start: address of the first instruction
+ *	@insn_end: address of the last instruction
+ *	@cie: the CIE for this function
+ *	@fde: the FDE for this function
+ *	@frame: the instructions calculate the CFA for this frame
+ *	@pc: the program counter of the address we're interested in
+ *
+ *	Execute the Call Frame instruction sequence starting at
+ *	@insn_start and ending at @insn_end. The instructions describe
+ *	how to calculate the Canonical Frame Address of a stackframe.
+ *	Store the results in @frame.
+ */
+static int 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)
+{
+	unsigned char insn;
+	unsigned char *current_insn;
+	unsigned int count, delta, reg, expr_len, offset;
+
+	current_insn = insn_start;
+
+	while (current_insn < insn_end && frame->pc <= pc) {
+		insn = __raw_readb(current_insn++);
+
+		/*
+		 * Firstly, handle the opcodes that embed their operands
+		 * in the instructions.
+		 */
+		switch (DW_CFA_opcode(insn)) {
+		case DW_CFA_advance_loc:
+			delta = DW_CFA_operand(insn);
+			delta *= cie->code_alignment_factor;
+			frame->pc += delta;
+			continue;
+			/* NOTREACHED */
+		case DW_CFA_offset:
+			reg = DW_CFA_operand(insn);
+			count = dwarf_read_uleb128(current_insn, &offset);
+			current_insn += count;
+			offset *= cie->data_alignment_factor;
+			dwarf_frame_alloc_regs(frame, reg);
+			frame->regs[reg].addr = offset;
+			frame->regs[reg].flags |= DWARF_REG_OFFSET;
+			continue;
+			/* NOTREACHED */
+		case DW_CFA_restore:
+			reg = DW_CFA_operand(insn);
+			continue;
+			/* NOTREACHED */
+		}
+
+		/*
+		 * Secondly, handle the opcodes that don't embed their
+		 * operands in the instruction.
+		 */
+		switch (insn) {
+		case DW_CFA_nop:
+			continue;
+		case DW_CFA_advance_loc1:
+			delta = *current_insn++;
+			frame->pc += delta * cie->code_alignment_factor;
+			break;
+		case DW_CFA_advance_loc2:
+			delta = __get_unaligned_cpu16(current_insn);
+			current_insn += 2;
+			frame->pc += delta * cie->code_alignment_factor;
+			break;
+		case DW_CFA_advance_loc4:
+			delta = __get_unaligned_cpu32(current_insn);
+			current_insn += 4;
+			frame->pc += delta * cie->code_alignment_factor;
+			break;
+		case DW_CFA_offset_extended:
+			count = dwarf_read_uleb128(current_insn, &reg);
+			current_insn += count;
+			count = dwarf_read_uleb128(current_insn, &offset);
+			current_insn += count;
+			offset *= cie->data_alignment_factor;
+			break;
+		case DW_CFA_restore_extended:
+			count = dwarf_read_uleb128(current_insn, &reg);
+			current_insn += count;
+			break;
+		case DW_CFA_undefined:
+			count = dwarf_read_uleb128(current_insn, &reg);
+			current_insn += count;
+			break;
+		case DW_CFA_def_cfa:
+			count = dwarf_read_uleb128(current_insn,
+						   &frame->cfa_register);
+			current_insn += count;
+			count = dwarf_read_uleb128(current_insn,
+						   &frame->cfa_offset);
+			current_insn += count;
+
+			frame->flags |= DWARF_FRAME_CFA_REG_OFFSET;
+			break;
+		case DW_CFA_def_cfa_register:
+			count = dwarf_read_uleb128(current_insn,
+						   &frame->cfa_register);
+			current_insn += count;
+			frame->flags |= DWARF_FRAME_CFA_REG_OFFSET;
+			break;
+		case DW_CFA_def_cfa_offset:
+			count = dwarf_read_uleb128(current_insn, &offset);
+			current_insn += count;
+			frame->cfa_offset = offset;
+			break;
+		case DW_CFA_def_cfa_expression:
+			count = dwarf_read_uleb128(current_insn, &expr_len);
+			current_insn += count;
+
+			frame->cfa_expr = current_insn;
+			frame->cfa_expr_len = expr_len;
+			current_insn += expr_len;
+
+			frame->flags |= DWARF_FRAME_CFA_REG_EXP;
+			break;
+		case DW_CFA_offset_extended_sf:
+			count = dwarf_read_uleb128(current_insn, &reg);
+			current_insn += count;
+			count = dwarf_read_leb128(current_insn, &offset);
+			current_insn += count;
+			offset *= cie->data_alignment_factor;
+			dwarf_frame_alloc_regs(frame, reg);
+			frame->regs[reg].flags |= DWARF_REG_OFFSET;
+			frame->regs[reg].addr = offset;
+			break;
+		case DW_CFA_val_offset:
+			count = dwarf_read_uleb128(current_insn, &reg);
+			current_insn += count;
+			count = dwarf_read_leb128(current_insn, &offset);
+			offset *= cie->data_alignment_factor;
+			frame->regs[reg].flags |= DWARF_REG_OFFSET;
+			frame->regs[reg].addr = offset;
+			break;
+		default:
+			pr_debug("unhandled DWARF instruction 0x%x\n", insn);
+			break;
+		}
+	}
+
+	return 0;
+}
+
+/**
+ *	dwarf_unwind_stack - recursively unwind the stack
+ *	@pc: address of the function to unwind
+ *	@prev: struct dwarf_frame of the previous stackframe on the callstack
+ *
+ *	Return a struct dwarf_frame representing the most recent frame
+ *	on the callstack. Each of the lower (older) stack frames are
+ *	linked via the "prev" member.
+ */
+struct dwarf_frame *dwarf_unwind_stack(unsigned long pc,
+				       struct dwarf_frame *prev)
+{
+	struct dwarf_frame *frame;
+	struct dwarf_cie *cie;
+	struct dwarf_fde *fde;
+	unsigned long addr;
+	int i, offset;
+
+	/*
+	 * If this is the first invocation of this recursive function we
+	 * need get the contents of a physical register to get the CFA
+	 * in order to begin the virtual unwinding of the stack.
+	 *
+	 * The constant DWARF_ARCH_UNWIND_OFFSET is added to the address of
+	 * this function because the return address register
+	 * (DWARF_ARCH_RA_REG) will probably not be initialised until a
+	 * few instructions into the prologue.
+	 */
+	if (!pc && !prev) {
+		pc = (unsigned long)&dwarf_unwind_stack;
+		pc += DWARF_ARCH_UNWIND_OFFSET;
+	}
+
+	frame = kzalloc(sizeof(*frame), GFP_KERNEL);
+	if (!frame)
+		return NULL;
+
+	frame->prev = prev;
+
+	fde = dwarf_lookup_fde(pc);
+	if (!fde) {
+		/*
+		 * This is our normal exit path - the one that stops the
+		 * recursion. There's two reasons why we might exit
+		 * here,
+		 *
+		 *	a) pc has no asscociated DWARF frame info and so
+		 *	we don't know how to unwind this frame. This is
+		 *	usually the case when we're trying to unwind a
+		 *	frame that was called from some assembly code
+		 *	that has no DWARF info, e.g. syscalls.
+		 *
+		 *	b) the DEBUG info for pc is bogus. There's
+		 *	really no way to distinguish this case from the
+		 *	case above, which sucks because we could print a
+		 *	warning here.
+		 */
+		return NULL;
+	}
+
+	cie = dwarf_lookup_cie(fde->cie_pointer);
+
+	frame->pc = fde->initial_location;
+
+	/* CIE initial instructions */
+	dwarf_cfa_execute_insns(cie->initial_instructions,
+				cie->instructions_end, cie, fde, frame, pc);
+
+	/* FDE instructions */
+	dwarf_cfa_execute_insns(fde->instructions, fde->end, cie,
+				fde, frame, pc);
+
+	/* Calculate the CFA */
+	switch (frame->flags) {
+	case DWARF_FRAME_CFA_REG_OFFSET:
+		if (prev) {
+			BUG_ON(!prev->regs[frame->cfa_register].flags);
+
+			addr = prev->cfa;
+			addr += prev->regs[frame->cfa_register].addr;
+			frame->cfa = __raw_readl(addr);
+
+		} else {
+			/*
+			 * Again, this is the first invocation of this
+			 * recurisve function. We need to physically
+			 * read the contents of a register in order to
+			 * get the Canonical Frame Address for this
+			 * function.
+			 */
+			frame->cfa = dwarf_read_arch_reg(frame->cfa_register);
+		}
+
+		frame->cfa += frame->cfa_offset;
+		break;
+	default:
+		BUG();
+	}
+
+	/* If we haven't seen the return address reg, we're screwed. */
+	BUG_ON(!frame->regs[DWARF_ARCH_RA_REG].flags);
+
+	for (i = 0; i <= frame->num_regs; i++) {
+		struct dwarf_reg *reg = &frame->regs[i];
+
+		if (!reg->flags)
+			continue;
+
+		offset = reg->addr;
+		offset += frame->cfa;
+	}
+
+	addr = frame->cfa + frame->regs[DWARF_ARCH_RA_REG].addr;
+	frame->return_addr = __raw_readl(addr);
+
+	frame->next = dwarf_unwind_stack(frame->return_addr, frame);
+	return frame;
+}
+
+static int dwarf_parse_cie(void *entry, void *p, unsigned long len,
+			   unsigned char *end)
+{
+	struct dwarf_cie *cie;
+	unsigned long flags;
+	int count;
+
+	cie = kzalloc(sizeof(*cie), GFP_KERNEL);
+	if (!cie)
+		return -ENOMEM;
+
+	cie->length = len;
+
+	/*
+	 * Record the offset into the .eh_frame section
+	 * for this CIE. It allows this CIE to be
+	 * quickly and easily looked up from the
+	 * corresponding FDE.
+	 */
+	cie->cie_pointer = (unsigned long)entry;
+
+	cie->version = *(char *)p++;
+	BUG_ON(cie->version != 1);
+
+	cie->augmentation = p;
+	p += strlen(cie->augmentation) + 1;
+
+	count = dwarf_read_uleb128(p, &cie->code_alignment_factor);
+	p += count;
+
+	count = dwarf_read_leb128(p, &cie->data_alignment_factor);
+	p += count;
+
+	/*
+	 * Which column in the rule table contains the
+	 * return address?
+	 */
+	if (cie->version == 1) {
+		cie->return_address_reg = __raw_readb(p);
+		p++;
+	} else {
+		count = dwarf_read_uleb128(p, &cie->return_address_reg);
+		p += count;
+	}
+
+	if (cie->augmentation[0] == 'z') {
+		unsigned int length, count;
+		cie->flags |= DWARF_CIE_Z_AUGMENTATION;
+
+		count = dwarf_read_uleb128(p, &length);
+		p += count;
+
+		BUG_ON((unsigned char *)p > end);
+
+		cie->initial_instructions = p + length;
+		cie->augmentation++;
+	}
+
+	while (*cie->augmentation) {
+		/*
+		 * "L" indicates a byte showing how the
+		 * LSDA pointer is encoded. Skip it.
+		 */
+		if (*cie->augmentation == 'L') {
+			p++;
+			cie->augmentation++;
+		} else if (*cie->augmentation == 'R') {
+			/*
+			 * "R" indicates a byte showing
+			 * how FDE addresses are
+			 * encoded.
+			 */
+			cie->encoding = *(char *)p++;
+			cie->augmentation++;
+		} else if (*cie->augmentation == 'P') {
+			/*
+			 * "R" indicates a personality
+			 * routine in the CIE
+			 * augmentation.
+			 */
+			BUG();
+		} else if (*cie->augmentation == 'S') {
+			BUG();
+		} else {
+			/*
+			 * Unknown augmentation. Assume
+			 * 'z' augmentation.
+			 */
+			p = cie->initial_instructions;
+			BUG_ON(!p);
+			break;
+		}
+	}
+
+	cie->initial_instructions = p;
+	cie->instructions_end = end;
+
+	/* Add to list */
+	spin_lock_irqsave(&dwarf_cie_lock, flags);
+	list_add_tail(&cie->link, &dwarf_cie_list);
+	spin_unlock_irqrestore(&dwarf_cie_lock, flags);
+
+	return 0;
+}
+
+static int dwarf_parse_fde(void *entry, u32 entry_type,
+			   void *start, unsigned long len)
+{
+	struct dwarf_fde *fde;
+	struct dwarf_cie *cie;
+	unsigned long flags;
+	int count;
+	void *p = start;
+
+	fde = kzalloc(sizeof(*fde), GFP_KERNEL);
+	if (!fde)
+		return -ENOMEM;
+
+	fde->length = len;
+
+	/*
+	 * In a .eh_frame section the CIE pointer is the
+	 * delta between the address within the FDE
+	 */
+	fde->cie_pointer = (unsigned long)(p - entry_type - 4);
+
+	cie = dwarf_lookup_cie(fde->cie_pointer);
+	fde->cie = cie;
+
+	if (cie->encoding)
+		count = dwarf_read_encoded_value(p, &fde->initial_location,
+						 cie->encoding);
+	else
+		count = dwarf_read_addr(p, &fde->initial_location);
+
+	p += count;
+
+	if (cie->encoding)
+		count = dwarf_read_encoded_value(p, &fde->address_range,
+						 cie->encoding & 0x0f);
+	else
+		count = dwarf_read_addr(p, &fde->address_range);
+
+	p += count;
+
+	if (fde->cie->flags & DWARF_CIE_Z_AUGMENTATION) {
+		unsigned int length;
+		count = dwarf_read_uleb128(p, &length);
+		p += count + length;
+	}
+
+	/* Call frame instructions. */
+	fde->instructions = p;
+	fde->end = start + len;
+
+	/* Add to list. */
+	spin_lock_irqsave(&dwarf_fde_lock, flags);
+	list_add_tail(&fde->link, &dwarf_fde_list);
+	spin_unlock_irqrestore(&dwarf_fde_lock, flags);
+
+	return 0;
+}
+
+static void dwarf_unwinder_dump(struct task_struct *task, struct pt_regs *regs,
+				unsigned long *sp,
+				const struct stacktrace_ops *ops, void *data)
+{
+	struct dwarf_frame *frame;
+
+	frame = dwarf_unwind_stack(0, NULL);
+
+	while (frame && frame->return_addr) {
+		ops->address(data, frame->return_addr, 1);
+		frame = frame->next;
+	}
+}
+
+static struct unwinder dwarf_unwinder = {
+	.name = "dwarf-unwinder",
+	.dump = dwarf_unwinder_dump,
+	.rating = 150,
+};
+
+static void dwarf_unwinder_cleanup(void)
+{
+	struct dwarf_cie *cie, *m;
+	struct dwarf_fde *fde, *n;
+	unsigned long flags;
+
+	/*
+	 * Deallocate all the memory allocated for the DWARF unwinder.
+	 * Traverse all the FDE/CIE lists and remove and free all the
+	 * memory associated with those data structures.
+	 */
+	spin_lock_irqsave(&dwarf_cie_lock, flags);
+	list_for_each_entry_safe(cie, m, &dwarf_cie_list, link)
+		kfree(cie);
+	spin_unlock_irqrestore(&dwarf_cie_lock, flags);
+
+	spin_lock_irqsave(&dwarf_fde_lock, flags);
+	list_for_each_entry_safe(fde, n, &dwarf_fde_list, link)
+		kfree(fde);
+	spin_unlock_irqrestore(&dwarf_fde_lock, flags);
+}
+
+/**
+ *	dwarf_unwinder_init - initialise the dwarf unwinder
+ *
+ *	Build the data structures describing the .dwarf_frame section to
+ *	make it easier to lookup CIE and FDE entries. Because the
+ *	.eh_frame section is packed as tightly as possible it is not
+ *	easy to lookup the FDE for a given PC, so we build a list of FDE
+ *	and CIE entries that make it easier.
+ */
+void dwarf_unwinder_init(void)
+{
+	u32 entry_type;
+	void *p, *entry;
+	int count, err;
+	unsigned long len;
+	unsigned int c_entries, f_entries;
+	unsigned char *end;
+	INIT_LIST_HEAD(&dwarf_cie_list);
+	INIT_LIST_HEAD(&dwarf_fde_list);
+
+	c_entries = 0;
+	f_entries = 0;
+	entry = &__start_eh_frame;
+
+	while ((char *)entry < __stop_eh_frame) {
+		p = entry;
+
+		count = dwarf_entry_len(p, &len);
+		if (count == 0) {
+			/*
+			 * We read a bogus length field value. There is
+			 * nothing we can do here apart from disabling
+			 * the DWARF unwinder. We can't even skip this
+			 * entry and move to the next one because 'len'
+			 * tells us where our next entry is.
+			 */
+			goto out;
+		} else
+			p += count;
+
+		/* initial length does not include itself */
+		end = p + len;
+
+		entry_type = __get_unaligned_cpu32(p);
+		p += 4;
+
+		if (entry_type == DW_EH_FRAME_CIE) {
+			err = dwarf_parse_cie(entry, p, len, end);
+			if (err < 0)
+				goto out;
+			else
+				c_entries++;
+		} else {
+			err = dwarf_parse_fde(entry, entry_type, p, len);
+			if (err < 0)
+				goto out;
+			else
+				f_entries++;
+		}
+
+		entry = (char *)entry + len + 4;
+	}
+
+	printk(KERN_INFO "DWARF unwinder initialised: read %u CIEs, %u FDEs\n",
+	       c_entries, f_entries);
+
+	err = unwinder_register(&dwarf_unwinder);
+	if (err)
+		goto out;
+
+	return;
+
+out:
+	printk(KERN_ERR "Failed to initialise DWARF unwinder: %d\n", err);
+	dwarf_unwinder_cleanup();
+}

arch/sh/kernel/irq.c

@@ -14,6 +14,7 @@
 #include <asm/processor.h>
 #include <asm/machvec.h>
 #include <asm/uaccess.h>
+#include <asm/dwarf.h>
 #include <asm/thread_info.h>
 #include <cpu/mmu_context.h>
 
@@ -261,6 +262,9 @@ void __init init_IRQ(void)
 		sh_mv.mv_init_irq();
 
 	irq_ctx_init(smp_processor_id());
+
+	/* This needs to be early, but not too early.. */
+	dwarf_unwinder_init();
 }
 
 #ifdef CONFIG_SPARSE_IRQ

arch/sh/kernel/vmlinux.lds.S

@@ -12,7 +12,7 @@ OUTPUT_ARCH(sh)
 
 #include <asm/thread_info.h>
 #include <asm/cache.h>
-#include <asm-generic/vmlinux.lds.h>
+#include <asm/vmlinux.lds.h>
 
 ENTRY(_start)
 SECTIONS
@@ -70,6 +70,8 @@ SECTIONS
 
 	_edata = .;			/* End of data section */
 
+	DWARF_EH_FRAME
+
 	. = ALIGN(PAGE_SIZE);		/* Init code and data */
 	__init_begin = .;
 	INIT_TEXT_SECTION(PAGE_SIZE)