Skip to contentSkip to conversation

x86/debug: define BUG() againfor !CONFIG_BUG

PATCH In review Arnd Bergmann 4 participants 9 messages Mar 29, 2017 – Mar 30, 2017 x86

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 x86/debug: define BUG() againfor !CONFIG_BUG

The latest change to the BUG() macro inadvertently reverted the earlier commit b06dd879f5db ("x86: always define BUG() and HAVE_ARCH_BUG, even with !CONFIG_BUG") that sanitized the behavior with CONFIG_BUG=n.

I noticed this as some warnings have appeared again that were previously fixed as a side effect of that patch:

kernel/seccomp.c: In function '__seccomp_filter':
kernel/seccomp.c:670:1: error: no return statement in function returning non-void [-Werror=return-type]

drivers/gpu/drm/i915/intel_sprite.c: In function 'intel_check_sprite_plane':
drivers/gpu/drm/i915/intel_sprite.c:936:20: error: 'src_h' may be used uninitialized in this function [-Werror=maybe-uninitialized]

   src->y2 = (src_y + src_h) << 16;
             ~~~~~~~^~~~~~~~
drivers/gpu/drm/i915/intel_sprite.c:934:20: error: 'src_w' may be used uninitialized in this function [-Werror=maybe-uninitialized]
   src->x2 = (src_x + src_w) << 16;
             ~~~~~~~^~~~~~~~
drivers/gpu/drm/i915/intel_sprite.c:936:20: error: 'src_y' may be used uninitialized in this function [-Werror=maybe-uninitialized]
   src->y2 = (src_y + src_h) << 16;
             ~~~~~~~^~~~~~~~
drivers/gpu/drm/i915/intel_sprite.c:934:20: error: 'src_x' may be used uninitialized in this function [-Werror=maybe-uninitialized]
   src->x2 = (src_x + src_w) << 16;
             ~~~~~~~^~~~~~~~

This combines the two patches and uses the ud2 macro to define BUG() in case of CONFIG_BUG=n.

Fixes 9a93848fe787 ("x86/debug: Implement __WARN() using UD0")
Cc Josh Triplett <[email protected]>
---
 arch/x86/include/asm/bug.h | 10 +++++++---
 1 file changed, 7 insertions(+), 3 deletions(-)
arch/x86/include/asm/bug.h +7 -3
index 4fde330c44b7..cecf559d0012 100644
@@ -23,7 +23,6 @@
 #define LEN_UD0		2
 
 #ifdef CONFIG_GENERIC_BUG
-#define HAVE_ARCH_BUG
 
 #ifdef CONFIG_X86_32
 # define __BUG_REL(val)	".long " __stringify(val)
@@ -64,6 +63,13 @@ do {									\
 
 #endif /* CONFIG_DEBUG_BUGVERBOSE */
 
+#else
+
+#define _BUG_FLAGS(ins, flags)  asm volatile(ins)
+
+#endif /* CONFIG_GENERIC_BUG */
+
+#define HAVE_ARCH_BUG
 #define BUG()							\
 do {								\
 	_BUG_FLAGS(ASM_UD2, 0);					\
@@ -72,8 +78,6 @@ do {								\
 
 #define __WARN_TAINT(taint)	_BUG_FLAGS(ASM_UD0, BUGFLAG_TAINT(taint))
 
-#endif /* CONFIG_GENERIC_BUG */
-
 #include <asm-generic/bug.h>
 
 #endif /* _ASM_X86_BUG_H */
Signature · 1 line
2.9.0
#1 lore raw
Quoted 24 lines · Arnd Bergmann

The latest change to the BUG() macro inadvertently reverted the earlier commit b06dd879f5db ("x86: always define BUG() and HAVE_ARCH_BUG, even with !CONFIG_BUG") that sanitized the behavior with CONFIG_BUG=n.

I noticed this as some warnings have appeared again that were previously fixed as a side effect of that patch:

kernel/seccomp.c: In function '__seccomp_filter':
kernel/seccomp.c:670:1: error: no return statement in function returning non-void [-Werror=return-type]

drivers/gpu/drm/i915/intel_sprite.c: In function 'intel_check_sprite_plane':
drivers/gpu/drm/i915/intel_sprite.c:936:20: error: 'src_h' may be used uninitialized in this function [-Werror=maybe-uninitialized]

   src->y2 = (src_y + src_h) << 16;
             ~~~~~~~^~~~~~~~
drivers/gpu/drm/i915/intel_sprite.c:934:20: error: 'src_w' may be used uninitialized in this function [-Werror=maybe-uninitialized]
   src->x2 = (src_x + src_w) << 16;
             ~~~~~~~^~~~~~~~
drivers/gpu/drm/i915/intel_sprite.c:936:20: error: 'src_y' may be used uninitialized in this function [-Werror=maybe-uninitialized]
   src->y2 = (src_y + src_h) << 16;
             ~~~~~~~^~~~~~~~
drivers/gpu/drm/i915/intel_sprite.c:934:20: error: 'src_x' may be used uninitialized in this function [-Werror=maybe-uninitialized]
   src->x2 = (src_x + src_w) << 16;
             ~~~~~~~^~~~~~~~
Quoted 2 lines

This combines the two patches and uses the ud2 macro to define BUG() in case of CONFIG_BUG=n.

OK, fair enough I suppose. However, I cribbed this from arm64. What does that do for BUG=n ?

#2 lore raw
Quoted 25 lines · Peter Zijlstra
Quoted 24 lines · Arnd Bergmann

The latest change to the BUG() macro inadvertently reverted the earlier commit b06dd879f5db ("x86: always define BUG() and HAVE_ARCH_BUG, even with !CONFIG_BUG") that sanitized the behavior with CONFIG_BUG=n.

I noticed this as some warnings have appeared again that were previously fixed as a side effect of that patch:

kernel/seccomp.c: In function '__seccomp_filter':
kernel/seccomp.c:670:1: error: no return statement in function returning non-void [-Werror=return-type]

drivers/gpu/drm/i915/intel_sprite.c: In function 'intel_check_sprite_plane':
drivers/gpu/drm/i915/intel_sprite.c:936:20: error: 'src_h' may be used uninitialized in this function [-Werror=maybe-uninitialized]

   src->y2 = (src_y + src_h) << 16;
             ~~~~~~~^~~~~~~~
drivers/gpu/drm/i915/intel_sprite.c:934:20: error: 'src_w' may be used uninitialized in this function [-Werror=maybe-uninitialized]
   src->x2 = (src_x + src_w) << 16;
             ~~~~~~~^~~~~~~~
drivers/gpu/drm/i915/intel_sprite.c:936:20: error: 'src_y' may be used uninitialized in this function [-Werror=maybe-uninitialized]
   src->y2 = (src_y + src_h) << 16;
             ~~~~~~~^~~~~~~~
drivers/gpu/drm/i915/intel_sprite.c:934:20: error: 'src_x' may be used uninitialized in this function [-Werror=maybe-uninitialized]
   src->x2 = (src_x + src_w) << 16;
             ~~~~~~~^~~~~~~~
Quoted 5 lines
Quoted 2 lines

This combines the two patches and uses the ud2 macro to define BUG() in case of CONFIG_BUG=n.

OK, fair enough I suppose. However, I cribbed this from arm64. What does that do for BUG=n ?

I think we'll get a U2D crash in this case, without any bug information.

I.e. only marginally debuggable, but it's a deterministic outcome - instead of the crazy GCC code generation variant of the day when the warning triggers, or the similarly crazy infinite loop hang.

I'm not entirely sure though, I don't think many people actually _use_ CONFIG_BUG=n, it's essentially a crazy thing to do even on constrainted hardware. Debugging and maintenance costs almost always trump marginal hardware costs of a bit more debugging code.

Thanks,

Ingo

#3 lore raw
Quoted 36 lines · Ingo Molnar
Quoted 25 lines · Peter Zijlstra
Quoted 24 lines · Arnd Bergmann

The latest change to the BUG() macro inadvertently reverted the earlier commit b06dd879f5db ("x86: always define BUG() and HAVE_ARCH_BUG, even with !CONFIG_BUG") that sanitized the behavior with CONFIG_BUG=n.

I noticed this as some warnings have appeared again that were previously fixed as a side effect of that patch:

kernel/seccomp.c: In function '__seccomp_filter':
kernel/seccomp.c:670:1: error: no return statement in function returning non-void [-Werror=return-type]

drivers/gpu/drm/i915/intel_sprite.c: In function 'intel_check_sprite_plane':
drivers/gpu/drm/i915/intel_sprite.c:936:20: error: 'src_h' may be used uninitialized in this function [-Werror=maybe-uninitialized]

   src->y2 = (src_y + src_h) << 16;
             ~~~~~~~^~~~~~~~
drivers/gpu/drm/i915/intel_sprite.c:934:20: error: 'src_w' may be used uninitialized in this function [-Werror=maybe-uninitialized]
   src->x2 = (src_x + src_w) << 16;
             ~~~~~~~^~~~~~~~
drivers/gpu/drm/i915/intel_sprite.c:936:20: error: 'src_y' may be used uninitialized in this function [-Werror=maybe-uninitialized]
   src->y2 = (src_y + src_h) << 16;
             ~~~~~~~^~~~~~~~
drivers/gpu/drm/i915/intel_sprite.c:934:20: error: 'src_x' may be used uninitialized in this function [-Werror=maybe-uninitialized]
   src->x2 = (src_x + src_w) << 16;
             ~~~~~~~^~~~~~~~
Quoted 5 lines
Quoted 2 lines

This combines the two patches and uses the ud2 macro to define BUG() in case of CONFIG_BUG=n.

OK, fair enough I suppose. However, I cribbed this from arm64. What does that do for BUG=n ?

I think we'll get a U2D crash in this case, without any bug information.

Quoted 8 lines

I.e. only marginally debuggable, but it's a deterministic outcome - instead of the crazy GCC code generation variant of the day when the warning triggers, or the similarly crazy infinite loop hang.

I'm not entirely sure though, I don't think many people actually _use_ CONFIG_BUG=n, it's essentially a crazy thing to do even on constrainted hardware. Debugging and maintenance costs almost always trump marginal hardware costs of a bit more debugging code.

So I've applied Arnd's patch to restore the previous behavior - but I agree with you that the situation isn't entirely logical at the moment. At minimum we need a comment explaining it or so.

Thanks,

Ingo

#4 lore raw
In reply to Ingo Molnar
Quoted 36 lines · Ingo Molnar
Quoted 25 lines · Peter Zijlstra
Quoted 24 lines · Arnd Bergmann

The latest change to the BUG() macro inadvertently reverted the earlier commit b06dd879f5db ("x86: always define BUG() and HAVE_ARCH_BUG, even with !CONFIG_BUG") that sanitized the behavior with CONFIG_BUG=n.

I noticed this as some warnings have appeared again that were previously fixed as a side effect of that patch:

kernel/seccomp.c: In function '__seccomp_filter':
kernel/seccomp.c:670:1: error: no return statement in function returning non-void [-Werror=return-type]

drivers/gpu/drm/i915/intel_sprite.c: In function 'intel_check_sprite_plane':
drivers/gpu/drm/i915/intel_sprite.c:936:20: error: 'src_h' may be used uninitialized in this function [-Werror=maybe-uninitialized]

   src->y2 = (src_y + src_h) << 16;
             ~~~~~~~^~~~~~~~
drivers/gpu/drm/i915/intel_sprite.c:934:20: error: 'src_w' may be used uninitialized in this function [-Werror=maybe-uninitialized]
   src->x2 = (src_x + src_w) << 16;
             ~~~~~~~^~~~~~~~
drivers/gpu/drm/i915/intel_sprite.c:936:20: error: 'src_y' may be used uninitialized in this function [-Werror=maybe-uninitialized]
   src->y2 = (src_y + src_h) << 16;
             ~~~~~~~^~~~~~~~
drivers/gpu/drm/i915/intel_sprite.c:934:20: error: 'src_x' may be used uninitialized in this function [-Werror=maybe-uninitialized]
   src->x2 = (src_x + src_w) << 16;
             ~~~~~~~^~~~~~~~
Quoted 5 lines
Quoted 2 lines

This combines the two patches and uses the ud2 macro to define BUG() in case of CONFIG_BUG=n.

OK, fair enough I suppose. However, I cribbed this from arm64. What does that do for BUG=n ?

I think we'll get a U2D crash in this case, without any bug information.

Quoted 8 lines

I.e. only marginally debuggable, but it's a deterministic outcome - instead of the crazy GCC code generation variant of the day when the warning triggers, or the similarly crazy infinite loop hang.

I'm not entirely sure though, I don't think many people actually _use_ CONFIG_BUG=n, it's essentially a crazy thing to do even on constrainted hardware. Debugging and maintenance costs almost always trump marginal hardware costs of a bit more debugging code.

So should we then, for x86, disable BUG=n instead?

#5 lore raw
Quoted 41 lines · Peter Zijlstra
Quoted 40 lines · Ingo Molnar
Quoted 25 lines · Peter Zijlstra
Quoted 24 lines · Arnd Bergmann

The latest change to the BUG() macro inadvertently reverted the earlier commit b06dd879f5db ("x86: always define BUG() and HAVE_ARCH_BUG, even with !CONFIG_BUG") that sanitized the behavior with CONFIG_BUG=n.

I noticed this as some warnings have appeared again that were previously fixed as a side effect of that patch:

kernel/seccomp.c: In function '__seccomp_filter':
kernel/seccomp.c:670:1: error: no return statement in function returning non-void [-Werror=return-type]

drivers/gpu/drm/i915/intel_sprite.c: In function 'intel_check_sprite_plane':
drivers/gpu/drm/i915/intel_sprite.c:936:20: error: 'src_h' may be used uninitialized in this function [-Werror=maybe-uninitialized]

   src->y2 = (src_y + src_h) << 16;
             ~~~~~~~^~~~~~~~
drivers/gpu/drm/i915/intel_sprite.c:934:20: error: 'src_w' may be used uninitialized in this function [-Werror=maybe-uninitialized]
   src->x2 = (src_x + src_w) << 16;
             ~~~~~~~^~~~~~~~
drivers/gpu/drm/i915/intel_sprite.c:936:20: error: 'src_y' may be used uninitialized in this function [-Werror=maybe-uninitialized]
   src->y2 = (src_y + src_h) << 16;
             ~~~~~~~^~~~~~~~
drivers/gpu/drm/i915/intel_sprite.c:934:20: error: 'src_x' may be used uninitialized in this function [-Werror=maybe-uninitialized]
   src->x2 = (src_x + src_w) << 16;
             ~~~~~~~^~~~~~~~
Quoted 5 lines
Quoted 2 lines

This combines the two patches and uses the ud2 macro to define BUG() in case of CONFIG_BUG=n.

OK, fair enough I suppose. However, I cribbed this from arm64. What does that do for BUG=n ?

I think we'll get a U2D crash in this case, without any bug information.

I.e. only marginally debuggable, but it's a deterministic outcome - instead of the crazy GCC code generation variant of the day when the warning triggers, or the similarly crazy infinite loop hang.

Quoted 6 lines
Quoted 4 lines

I'm not entirely sure though, I don't think many people actually _use_ CONFIG_BUG=n, it's essentially a crazy thing to do even on constrainted hardware. Debugging and maintenance costs almost always trump marginal hardware costs of a bit more debugging code.

So should we then, for x86, disable BUG=n instead?

Arnd, does CONFIG_BUG=n give any (marginal) text savings wrt. CONFIG_BUG=y && CONFIG_BUG_VERBOSE=n?

If not then we should indeed just disable CONFIG_BUG=n.

Thanks,

Ingo

#6 lore raw
Quoted 6 lines · Ingo Molnar
Quoted 1 line · Peter Zijlstra

So should we then, for x86, disable BUG=n instead?

Arnd, does CONFIG_BUG=n give any (marginal) text savings wrt. CONFIG_BUG=y && CONFIG_BUG_VERBOSE=n?

My patch makes it slightly worse, but the difference is still around 1% of the total vmlinux size:

# CONFIG_BUG=n without my patch
text data bss dec hex filename
9543165 4190288 843776 14577229 de6e4d obj-x86/vmlinux

# CONFIG_BUG=n with my patch
text data bss dec hex filename
9578700 4190480 843776 14612956 def9dc obj-x86/vmlinux

# CONFIG_BUG=y, CONFIG_BUGVERBOSE=n
text data bss dec hex filename
9698166 4212208 843776 14754150 e12166 obj-x86/vmlinux

Among the things we save are:

- WARN()/WARN_ON() gets left out entirely
- The bug_entry section (46192 bytes in defconfig)
- The code for decoding and printing the bug

Quoted 1 line

If not then we should indeed just disable CONFIG_BUG=n.

CONFIG_BUG is already guarded by CONFIG_EXPERT, along with this help text in Kconfig:

Disabling this option eliminates support for BUG and WARN, reducing the size of your kernel image and potentially quietly ignoring numerous fatal conditions. You should only consider disabling this option for embedded systems with no facilities for reporting errors. Just say Y.

I think this is reasonable to disable for systems that have no console at all.

Arnd

#7 lore raw
In reply to Peter Zijlstra
Quoted 25 lines · Peter Zijlstra
Quoted 24 lines · Arnd Bergmann

The latest change to the BUG() macro inadvertently reverted the earlier commit b06dd879f5db ("x86: always define BUG() and HAVE_ARCH_BUG, even with !CONFIG_BUG") that sanitized the behavior with CONFIG_BUG=n.

I noticed this as some warnings have appeared again that were previously fixed as a side effect of that patch:

kernel/seccomp.c: In function '__seccomp_filter':
kernel/seccomp.c:670:1: error: no return statement in function returning non-void [-Werror=return-type]

drivers/gpu/drm/i915/intel_sprite.c: In function 'intel_check_sprite_plane':
drivers/gpu/drm/i915/intel_sprite.c:936:20: error: 'src_h' may be used uninitialized in this function [-Werror=maybe-uninitialized]

   src->y2 = (src_y + src_h) << 16;
             ~~~~~~~^~~~~~~~
drivers/gpu/drm/i915/intel_sprite.c:934:20: error: 'src_w' may be used uninitialized in this function [-Werror=maybe-uninitialized]
   src->x2 = (src_x + src_w) << 16;
             ~~~~~~~^~~~~~~~
drivers/gpu/drm/i915/intel_sprite.c:936:20: error: 'src_y' may be used uninitialized in this function [-Werror=maybe-uninitialized]
   src->y2 = (src_y + src_h) << 16;
             ~~~~~~~^~~~~~~~
drivers/gpu/drm/i915/intel_sprite.c:934:20: error: 'src_x' may be used uninitialized in this function [-Werror=maybe-uninitialized]
   src->x2 = (src_x + src_w) << 16;
             ~~~~~~~^~~~~~~~
Quoted 4 lines
Quoted 2 lines

This combines the two patches and uses the ud2 macro to define BUG() in case of CONFIG_BUG=n.

OK, fair enough I suppose. However, I cribbed this from arm64.

Catalin just merged my patch to make arm64 work like x86 and arm32 ;-)

Quoted 1 line

What does that do for BUG=n ?

Without the patch, running into a BUG() often leads to undefined behavior, e.g. running off at the end of a function that is marked noreturn(), or simply continuing a thread after we know that this leads to data corruption (which is the reason to put the BUG() in there in the first place).

With my patch, we will insert a ud2 instruction and mark the context as unreachable(), so gcc can discard any assembler output after the unreachable() and the kernel will kill the task at that point, but will not have a bug_entry and just report an invalid instruction similar to what newer gcc generates for an unconditional division by zero or NULL pointer dereference.

Arnd

#8 lore raw

[tip:x86/asm] x86/debug: Define BUG() again for !CONFIG_BUG

In reply to Arnd Bergmann
x86/debug: Define BUG() again for !CONFIG_BUG
Commit-ID:  70579a86e3c8eb2ce57999e594a73b4dfe095959
Gitweb:     http://git.kernel.org/tip/70579a86e3c8eb2ce57999e594a73b4dfe095959
Author:     Arnd Bergmann <[email protected]>
AuthorDate: Wed, 29 Mar 2017 23:16:31 +0200
Committer:  Ingo Molnar <[email protected]>
CommitDate: Thu, 30 Mar 2017 09:12:10 +0200

x86/debug: Define BUG() again for !CONFIG_BUG

The latest change to the BUG() macro inadvertently reverted the earlier commit:

b06dd879f5db ("x86: always define BUG() and HAVE_ARCH_BUG, even with !CONFIG_BUG")

... that sanitized the behavior with CONFIG_BUG=n.

I noticed this as some warnings have appeared again that were previously fixed as a side effect of that patch:

kernel/seccomp.c: In function '__seccomp_filter':
kernel/seccomp.c:670:1: error: no return statement in function returning non-void [-Werror=return-type]
...

This combines the two patches and uses the ud2 macro to define BUG() in case of CONFIG_BUG=n.

Cc Andy Lutomirski <[email protected]>
Cc Borislav Petkov <[email protected]>
Cc Brian Gerst <[email protected]>
Cc Denys Vlasenko <[email protected]>
Cc H. Peter Anvin <[email protected]>
Cc Josh Poimboeuf <[email protected]>
Cc Josh Triplett <[email protected]>
Cc Linus Torvalds <[email protected]>
Cc Peter Zijlstra <[email protected]>
Cc Thomas Gleixner <[email protected]>
Fixes 9a93848fe787 ("x86/debug: Implement __WARN() using UD0")
---
 arch/x86/include/asm/bug.h | 10 +++++++---
 1 file changed, 7 insertions(+), 3 deletions(-)
arch/x86/include/asm/bug.h +7 -3
index 4fde330..cecf559 100644
@@ -23,7 +23,6 @@
 #define LEN_UD0		2
 
 #ifdef CONFIG_GENERIC_BUG
-#define HAVE_ARCH_BUG
 
 #ifdef CONFIG_X86_32
 # define __BUG_REL(val)	".long " __stringify(val)
@@ -64,6 +63,13 @@ do {									\
 
 #endif /* CONFIG_DEBUG_BUGVERBOSE */
 
+#else
+
+#define _BUG_FLAGS(ins, flags)  asm volatile(ins)
+
+#endif /* CONFIG_GENERIC_BUG */
+
+#define HAVE_ARCH_BUG
 #define BUG()							\
 do {								\
 	_BUG_FLAGS(ASM_UD2, 0);					\
@@ -72,8 +78,6 @@ do {								\
 
 #define __WARN_TAINT(taint)	_BUG_FLAGS(ASM_UD0, BUGFLAG_TAINT(taint))
 
-#endif /* CONFIG_GENERIC_BUG */
-
 #include <asm-generic/bug.h>
 
 #endif /* _ASM_X86_BUG_H */