Skip to contentSkip to conversation

x86-64: Fix CFI data for common_interrupt

PATCH NAK Mark Wielaard 2 participants 9 messages Feb 21, 2012 – Feb 27, 2012 x86

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 x86-64: Fix CFI data for common_interrupt

Commit eab9e6 "x86-64: Fix CFI data for interrupt frames" introduced a DW_CFA_def_cfa_expression in the SAVE_ARGS_IRQ macro. To later define the CFA using a simple register+offset rule both register and offset need to be supplied. Just using CFI_DEF_CFA_REGISTER leaves the offset undefined. So use CFI_DEF_CFA with reg+off explicitly at the end of common_interrupt.

---
 arch/x86/kernel/entry_64.S |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)
arch/x86/kernel/entry_64.S +1 -1
index 3fe8239..e00ef55 100644
@@ -813,7 +813,7 @@ ret_from_intr:
 
 	/* Restore saved previous stack */
 	popq %rsi
-	CFI_DEF_CFA_REGISTER	rsi
+	CFI_DEF_CFA rsi,0	/* needed after def_cfa_expression */
 	leaq ARGOFFSET-RBP(%rsi), %rsp
 	CFI_DEF_CFA_REGISTER	rsp
 	CFI_ADJUST_CFA_OFFSET	RBP-ARGOFFSET
Signature · 1 line
1.7.7.6
#1 lore raw
Quoted 7 lines
Quoted 1 line
Quoted 1 line · Mark Wielaard

Commit eab9e6 "x86-64: Fix CFI data for interrupt frames" introduced a DW_CFA_def_cfa_expression in the SAVE_ARGS_IRQ macro. To later define the CFA using a simple register+offset rule both register and offset need to be supplied. Just using CFI_DEF_CFA_REGISTER leaves the offset undefined. So use CFI_DEF_CFA with reg+off explicitly at the end of common_interrupt.

NAK, unless you can prove a path via which the offset will remain unset until hitting a CFI_DEF_CFA_REGISTER. And if you indeed found such a path, the entry point of the path is where the problem ought to be fixed.

Are you perhaps thinking that .cfi_def_cfa_register invalidates the offset in any way? That, to my knowledge, isn't the case, it just replaces the CFA register with the one specified, leaving the offset unchanged.

Jan

Quoted 20 lines
---
 arch/x86/kernel/entry_64.S |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)
arch/x86/kernel/entry_64.S +1 -1
index 3fe8239..e00ef55 100644
@@ -813,7 +813,7 @@ ret_from_intr:
 
 	/* Restore saved previous stack */
 	popq %rsi
-	CFI_DEF_CFA_REGISTER	rsi
+	CFI_DEF_CFA rsi,0	/* needed after def_cfa_expression */
 	leaq ARGOFFSET-RBP(%rsi), %rsp
 	CFI_DEF_CFA_REGISTER	rsp
 	CFI_ADJUST_CFA_OFFSET	RBP-ARGOFFSET
Signature · 1 line
1.7.7.6
#2 lore raw
Quoted 13 lines · Jan Beulich
Quoted 7 lines
Quoted 1 line
Quoted 1 line · Mark Wielaard

Commit eab9e6 "x86-64: Fix CFI data for interrupt frames" introduced a DW_CFA_def_cfa_expression in the SAVE_ARGS_IRQ macro. To later define the CFA using a simple register+offset rule both register and offset need to be supplied. Just using CFI_DEF_CFA_REGISTER leaves the offset undefined. So use CFI_DEF_CFA with reg+off explicitly at the end of common_interrupt.

NAK, unless you can prove a path via which the offset will remain unset until hitting a CFI_DEF_CFA_REGISTER. And if you indeed found such a path, the entry point of the path is where the problem ought to be fixed.

Quoted 4 lines

Are you perhaps thinking that .cfi_def_cfa_register invalidates the offset in any way? That, to my knowledge, isn't the case, it just replaces the CFA register with the one specified, leaving the offset unchanged.

DW_CFA_def_cfa_expression invalidates the offset (and register). Used through the interrupt macro for do_IRQ which uses the SAVE_ARGS_IRQ to define common_interrupt. So after using DW_CFA_def_cfa_expression we get a CFI_DEF_REGISTER and the CFI for common_interrupt looks like:

[ 6e30] FDE length=148 cie=[ 6e18]

   CIE_pointer:              28184
   initial_location:         0xffffffff815e8d00 <common_interrupt>
   address_range:            0x1ba

Program:
[...]

     advance_loc 1 to 0x69
     def_cfa_expression 6
          [   0] breg7 0
          [   2] deref
          [   3] const1u 136
          [   5] plus
     advance_loc 22 to 0x7f
     def_cfa_register r4 (rsi)

[...]

For DW_CFA_def_register DWARF4 explicitly says so: "This operation is valid only if the current CFA rule is defined to use a register and offset." So one needs to use CFI_DEF_CFA with both a register and an offset here after the def_cfa_expression.

Thanks,

Mark

#3 lore raw
Quoted 40 lines
Quoted 1 line
Quoted 1 line · Mark Wielaard
Quoted 13 lines · Jan Beulich
Quoted 7 lines
Quoted 1 line
Quoted 1 line · Mark Wielaard

Commit eab9e6 "x86-64: Fix CFI data for interrupt frames" introduced a DW_CFA_def_cfa_expression in the SAVE_ARGS_IRQ macro. To later define the CFA using a simple register+offset rule both register and offset need to be supplied. Just using CFI_DEF_CFA_REGISTER leaves the offset undefined. So use CFI_DEF_CFA with reg+off explicitly at the end of common_interrupt.

NAK, unless you can prove a path via which the offset will remain unset until hitting a CFI_DEF_CFA_REGISTER. And if you indeed found such a path, the entry point of the path is where the problem ought to be fixed.

Quoted 4 lines

Are you perhaps thinking that .cfi_def_cfa_register invalidates the offset in any way? That, to my knowledge, isn't the case, it just replaces the CFA register with the one specified, leaving the offset unchanged.

DW_CFA_def_cfa_expression invalidates the offset (and register). Used through the interrupt macro for do_IRQ which uses the SAVE_ARGS_IRQ to define common_interrupt. So after using DW_CFA_def_cfa_expression we get a CFI_DEF_REGISTER and the CFI for common_interrupt looks like:

[ 6e30] FDE length=148 cie=[ 6e18]

   CIE_pointer:              28184
   initial_location:         0xffffffff815e8d00 <common_interrupt>
   address_range:            0x1ba

Program:
[...]

     advance_loc 1 to 0x69
     def_cfa_expression 6
          [   0] breg7 0
          [   2] deref
          [   3] const1u 136
          [   5] plus
     advance_loc 22 to 0x7f
     def_cfa_register r4 (rsi)
Quoted 6 lines

[...]

For DW_CFA_def_register DWARF4 explicitly says so: "This operation is valid only if the current CFA rule is defined to use a register and offset." So one needs to use CFI_DEF_CFA with both a register and an offset here after the def_cfa_expression.

Hmm, that's in contrast to the gas implementation (but I'd certainly give the specification preference if it explicitly states so, so gas should at least emit a warning here rather than considering this valid).

But provided the specification mandates this, I'm okay with the change in principle. Just that specifying an offset of 0 doesn't look right then.

Jan

#4 lore raw
Quoted 10 lines · Jan Beulich
Quoted 5 lines
Quoted 1 line
Quoted 1 line · Mark Wielaard

For DW_CFA_def_register DWARF4 explicitly says so: "This operation is valid only if the current CFA rule is defined to use a register and offset." So one needs to use CFI_DEF_CFA with both a register and an offset here after the def_cfa_expression.

Hmm, that's in contrast to the gas implementation (but I'd certainly give the specification preference if it explicitly states so, so gas should at least emit a warning here rather than considering this valid).

I am afraid gas cannot help us here. Since like you pointed out in your patch:

This requires the use of .cfi_escape (allowing arbitrary byte streams to be emitted into .eh_frame), as there is no .cfi_def_cfa_expression (which also cannot reasonably be expected, as it would require a full expression parser).

So we are on our own here.

Quoted 2 lines

But provided the specification mandates this, I'm okay with the change in principle. Just that specifying an offset of 0 doesn't look right then.

Yeah, I dunno what I was thinking. The offset should be set to the offset that was there before when rsi was pushed. The attached patch does that by using the same value as was used at the start of common_interrupt. Does that look OK?

Thanks,

Mark

Quoted 1 line

From 5b48b64319463a4ccab560304ce3cf72c71d081d Mon Sep 17 00:00:00 2001

From: Mark Wielaard <[email protected]>
Date: Tue, 21 Feb 2012 14:45:52 +0100
Subject: [PATCH] x86-64: Fix CFI data for common_interrupt

Commit eab9e6 "x86-64: Fix CFI data for interrupt frames" introduced a DW_CFA_def_cfa_expression in the SAVE_ARGS_IRQ macro. To later define the CFA using a simple register+offset rule both register and offset need to be supplied. Just using CFI_DEF_CFA_REGISTER leaves the offset undefined. So use CFI_DEF_CFA with reg+off explicitly at the end of common_interrupt.

---
 arch/x86/kernel/entry_64.S |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)
arch/x86/kernel/entry_64.S +1 -1
index 3fe8239..308b604 100644
@@ -813,7 +813,7 @@ ret_from_intr:
 
 	/* Restore saved previous stack */
 	popq %rsi
-	CFI_DEF_CFA_REGISTER	rsi
+	CFI_DEF_CFA rsi,16+SS-RBP	/* reg/off reset after def_cfa_expr */
 	leaq ARGOFFSET-RBP(%rsi), %rsp
 	CFI_DEF_CFA_REGISTER	rsp
 	CFI_ADJUST_CFA_OFFSET	RBP-ARGOFFSET
Signature · 1 line
1.7.7.6
#5 lore raw
Quoted 16 lines
Quoted 1 line
Quoted 1 line

Mark Wielaard <[email protected]> 02/21/12 11:08 PM >>>

Quoted 10 lines · Jan Beulich
Quoted 5 lines
Quoted 1 line
Quoted 1 line · Mark Wielaard

For DW_CFA_def_register DWARF4 explicitly says so: "This operation is valid only if the current CFA rule is defined to use a register and offset." So one needs to use CFI_DEF_CFA with both a register and an offset here after the def_cfa_expression.

Hmm, that's in contrast to the gas implementation (but I'd certainly give the specification preference if it explicitly states so, so gas should at least emit a warning here rather than considering this valid).

I am afraid gas cannot help us here. Since like you pointed out in your patch:

Quoted 6 lines

This requires the use of .cfi_escape (allowing arbitrary byte streams to be emitted into .eh_frame), as there is no .cfi_def_cfa_expression (which also cannot reasonably be expected, as it would require a full expression parser).

So we are on our own here.

Hmm, yes, probably it wouldn't be nice if gas reset all its state when .cfi_escape is used.

Quoted 7 lines
Quoted 2 lines

But provided the specification mandates this, I'm okay with the change in principle. Just that specifying an offset of 0 doesn't look right then.

Yeah, I dunno what I was thinking. The offset should be set to the offset that was there before when rsi was pushed. The attached patch does that by using the same value as was used at the start of common_interrupt. Does that look OK?

I would have thought that it should be SS+8-RBP (as %rbp is at the top of the stack at that point). I can't verify this immediately, though, as I'm not in the office today.

Jan

Acked-by
#6 lore raw
In reply to Mark Wielaard
Quoted 9 lines
Quoted 1 line
Quoted 1 line · Mark Wielaard
Quoted 2 lines · Jan Beulich

But provided the specification mandates this, I'm okay with the change in principle. Just that specifying an offset of 0 doesn't look right then.

Yeah, I dunno what I was thinking. The offset should be set to the offset that was there before when rsi was pushed. The attached patch does that by using the same value as was used at the start of common_interrupt. Does that look OK?

As written before, it ought to be

CFI_DEF_CFA rsi,SS+8-RBP /* reg/off reset after def_cfa_expr */

With that, feel free to add

when you re-submit.

Jan

#7 lore raw

PATCH 1/1 x86-64: Fix CFI data for common_interrupt

From: Mark Wielaard <[email protected]>

Commit eab9e6 "x86-64: Fix CFI data for interrupt frames" introduced a DW_CFA_def_cfa_expression in the SAVE_ARGS_IRQ macro. To later define the CFA using a simple register+offset rule both register and offset need to be supplied. Just using CFI_DEF_CFA_REGISTER leaves the offset undefined. So use CFI_DEF_CFA with reg+off explicitly at the end of common_interrupt.

---
 arch/x86/kernel/entry_64.S |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)
arch/x86/kernel/entry_64.S +1 -1
index 3fe8239..54be36b 100644
@@ -813,7 +813,7 @@ ret_from_intr:
 
 	/* Restore saved previous stack */
 	popq %rsi
-	CFI_DEF_CFA_REGISTER	rsi
+	CFI_DEF_CFA rsi,SS+8-RBP	/* reg/off reset after def_cfa_expr */
 	leaq ARGOFFSET-RBP(%rsi), %rsp
 	CFI_DEF_CFA_REGISTER	rsp
 	CFI_ADJUST_CFA_OFFSET	RBP-ARGOFFSET
Signature · 1 line
1.7.7.6
#8 lore raw

[tip:x86/debug] x86-64: Fix CFI data for common_interrupt()

Commit-ID:  928282e432ee584129a39da831ffa72c38e189b7
Gitweb:     http://git.kernel.org/tip/928282e432ee584129a39da831ffa72c38e189b7
Author:     Mark Wielaard <[email protected]>
AuthorDate: Fri, 24 Feb 2012 11:32:05 +0100
Committer:  Ingo Molnar <[email protected]>
CommitDate: Mon, 27 Feb 2012 10:46:14 +0100

x86-64: Fix CFI data for common_interrupt()

Commit eab9e6137f23 ("x86-64: Fix CFI data for interrupt frames") introduced a DW_CFA_def_cfa_expression in the SAVE_ARGS_IRQ macro. To later define the CFA using a simple register+offset rule both register and offset need to be supplied. Just using CFI_DEF_CFA_REGISTER leaves the offset undefined. So use CFI_DEF_CFA with reg+off explicitly at the end of common_interrupt.

---
 arch/x86/kernel/entry_64.S |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)
arch/x86/kernel/entry_64.S +1 -1
index 3fe8239..54be36b 100644
@@ -813,7 +813,7 @@ ret_from_intr:
 
 	/* Restore saved previous stack */
 	popq %rsi
-	CFI_DEF_CFA_REGISTER	rsi
+	CFI_DEF_CFA rsi,SS+8-RBP	/* reg/off reset after def_cfa_expr */
 	leaq ARGOFFSET-RBP(%rsi), %rsp
 	CFI_DEF_CFA_REGISTER	rsp
 	CFI_ADJUST_CFA_OFFSET	RBP-ARGOFFSET