Skip to contentSkip to conversation

Default setting of the ARM_UNWIND option

PATCH New Claudio Scordino 2 participants 9 messages Oct 19, 2009 – Oct 30, 2009 2 messages not in the archive Arm (32-bit)

lore
Share & tools
lore

A message that is not in the archive (sent off-list or not yet received): [email protected]

A message that is not in the archive (sent off-list or not yet received): [email protected]

#2 lore raw

PATCH 1/1 Default setting of the ARM_UNWIND option

[email cut-and-pasted from a private discussion...]

Quoted 9 lines
Quoted 9 lines

moving from release 2.6.27 to release 2.6.32-rc4, my ARM board
(derived from AT91SAM9623-EK) started hanging at the initial "Calibrating delay loop" message.

After some inspection, I found out the problem to be with commit adf8b37bafc1495393201a2ae4235846371870d0.

In practice, it enables stack unwinding support by default. However, if enabled, it hangs my machine.

The problem might be with my compiler (according to the help, stack unwinding only works with EABI compilers).

Quoted 8 lines
Quoted 2 lines

I'm using a CodeSourcery gcc 4.2.1 "arm-none-linux-gnueabi".

I think 4.2.1 generates really buggy code with -fno-frame-pointer, nothing to do with the stack unwinding implementation in the kernel. I would recommend you either use a newer compiler or disables ARM_UNWIND in your .config (which automatically re-enables FRAMEPOINTER).

Hi Catalin,

thank you for your answer.

In my opinion, we should change the default setting of this option.

Having "yes" as default policy for ARM_UNWIND means that, by default, the kernel does not boot if a user has a buggy compiler (like me) and he/she doesn't explicitly disable the ARM_UNWIND option.

Consider that people may not know that their problem is related to the ARM_UNWIND option. Therefore, they may waste time just bisecting to understand where their problem really is.

Therefore, my suggestion is to change the default policy of this option as shown in the attached patch, to avoid a potential waste of time of developers. Users who know that they can enable this option, will enable it.

Let me know what you think about this suggestion.

Many thanks and my best regards,

Claudio

From: Claudio Scordino <[email protected]>
Date: Mon, 19 Oct 2009 11:59:16 +0200
Subject: [PATCH 1/1] Disable stack unwinding support by default, since it does not work on some buggy or not-EABI compilers.

Disable stack unwinding support by default, since it does not work on some buggy or not-EABI compilers.

---
 arch/arm/Kconfig.debug |    4 ++--
 1 files changed, 2 insertions(+), 2 deletions(-)
arch/arm/Kconfig.debug +2 -2
index 1a6f70e..925e3fd 100644
@@ -19,13 +19,13 @@ config FRAME_POINTER
 config ARM_UNWIND
 	bool "Enable stack unwinding support"
 	depends on AEABI && EXPERIMENTAL
-	default y
+	default n
 	help
 	  This option enables stack unwinding support in the kernel
 	  using the information automatically generated by the
 	  compiler. The resulting kernel image is slightly bigger but
 	  the performance is not affected. Currently, this feature
-	  only works with EABI compilers. If unsure say Y.
+	  only works with EABI compilers. If unsure say N.
 
 config DEBUG_USER
 	bool "Verbose user fault messages"
Signature · 1 line
1.6.0.4
#3 lore raw

PATCH 1/1 Default setting of the ARM_UNWIND option

In reply to a message not in the archive

[email cut-and-pasted from a private discussion...]

Quoted 9 lines
Quoted 9 lines

moving from release 2.6.27 to release 2.6.32-rc4, my ARM board
(derived from AT91SAM9623-EK) started hanging at the initial "Calibrating delay loop" message.

After some inspection, I found out the problem to be with commit adf8b37bafc1495393201a2ae4235846371870d0.

In practice, it enables stack unwinding support by default. However, if enabled, it hangs my machine.

The problem might be with my compiler (according to the help, stack unwinding only works with EABI compilers).

Quoted 8 lines
Quoted 2 lines

I'm using a CodeSourcery gcc 4.2.1 "arm-none-linux-gnueabi".

I think 4.2.1 generates really buggy code with -fno-frame-pointer, nothing to do with the stack unwinding implementation in the kernel. I would recommend you either use a newer compiler or disables ARM_UNWIND in your .config (which automatically re-enables FRAMEPOINTER).

Hi Catalin,

thank you for your answer.

In my opinion, we should change the default setting of this option.

Having "yes" as default policy for ARM_UNWIND means that, by default, the kernel does not boot if a user has a buggy compiler (like me) and he/she doesn't explicitly disable the ARM_UNWIND option.

Consider that people may not know that their problem is related to the ARM_UNWIND option. Therefore, they may waste time just bisecting to understand where their problem really is.

Therefore, my suggestion is to change the default policy of this option as shown in the attached patch, to avoid a potential waste of time of developers. Users who know that they can enable this option, will enable it.

Let me know what you think about this suggestion.

Many thanks and my best regards,

Claudio

Quoted 1 line

From 3b688dbebb5fb3345e48cd491eb387434db165da Mon Sep 17 00:00:00 2001

From: Claudio Scordino <[email protected]>
Date: Mon, 19 Oct 2009 11:59:16 +0200
Subject: [PATCH 1/1] Disable stack unwinding support by default, since it does not work on some buggy or not-EABI compilers.

---
 arch/arm/Kconfig.debug |    4 ++--
 1 files changed, 2 insertions(+), 2 deletions(-)
arch/arm/Kconfig.debug +2 -2
index 1a6f70e..925e3fd 100644
@@ -19,13 +19,13 @@ config FRAME_POINTER
 config ARM_UNWIND
 	bool "Enable stack unwinding support"
 	depends on AEABI && EXPERIMENTAL
-	default y
+	default n
 	help
 	  This option enables stack unwinding support in the kernel
 	  using the information automatically generated by the
 	  compiler. The resulting kernel image is slightly bigger but
 	  the performance is not affected. Currently, this feature
-	  only works with EABI compilers. If unsure say Y.
+	  only works with EABI compilers. If unsure say N.
 
 config DEBUG_USER
 	bool "Verbose user fault messages"
Signature · 1 line
1.6.0.4
#4 lore raw

PATCH 1/1 Default setting of the ARM_UNWIND option

In reply to a message not in the archive

Hi all,

I didn't get any comment on this patch, so I try to re-submit.

My ARM board hanged at the initial "Calibrating delay loop" message.

After some inspection, I found out the problem to be with commit adf8b37bafc1495393201a2ae4235846371870d0. This commit introduces stack unwinding for ARM, and set it enabled by default. However, it seems to not work with buggy or not-EABI compilers.

My suggestion is to keep the feature (which is fine) but change the default setting of the option (see the attached patch).

Having this option enabled by default, in fact, means that the kernel does not boot if the user has a wrong compiler. If so, inspecting the reason of the hang may require too much time (especially for people not familiar with this option).

Consider that people may not know that their problem is related to the ARM_UNWIND option. Therefore, they may waste time bisecting just to understand where their problem really is.

If we disable this option by default, users who know the real meaning of the option will still be able of setting it to the proper value.

Any comment ?

Many thanks,

Claudio

From: Claudio Scordino <[email protected]>
Date: Mon, 19 Oct 2009 11:59:16 +0200
Subject: [PATCH 1/1] Disable stack unwinding support by default, since it does not work on some buggy or not-EABI compilers.

Disable stack unwinding support by default, since it does not work on some buggy or not-EABI compilers.

---
 arch/arm/Kconfig.debug |    4 ++--
 1 files changed, 2 insertions(+), 2 deletions(-)
arch/arm/Kconfig.debug +2 -2
index 1a6f70e..925e3fd 100644
@@ -19,13 +19,13 @@ config FRAME_POINTER
 config ARM_UNWIND
 	bool "Enable stack unwinding support"
 	depends on AEABI && EXPERIMENTAL
-	default y
+	default n
 	help
 	  This option enables stack unwinding support in the kernel
 	  using the information automatically generated by the
 	  compiler. The resulting kernel image is slightly bigger but
 	  the performance is not affected. Currently, this feature
-	  only works with EABI compilers. If unsure say Y.
+	  only works with EABI compilers. If unsure say N.
 
 config DEBUG_USER
 	bool "Verbose user fault messages"
Signature · 1 line
1.6.0.4
#5 lore raw
Quoted 9 lines · Claudio Scordino

My ARM board hanged at the initial "Calibrating delay loop" message.

After some inspection, I found out the problem to be with commit adf8b37bafc1495393201a2ae4235846371870d0. This commit introduces stack unwinding for ARM, and set it enabled by default. However, it seems to not work with buggy or not-EABI compilers.

My suggestion is to keep the feature (which is fine) but change the default setting of the option (see the attached patch).

The option still depends on EXPERIMENTAL, so you get ARM_UNWIND on when enabling that.

I'm more in favour of a #warning on #error in the unwind.c file based on the compiler version rather than not having it on by default. The reason is that people reported performance improvements when compiling the kernel without frame pointers.

Signature · 1 line
Catalin
#6 lore raw

Catalin Marinas ha scritto:

Quoted 13 lines · Claudio Scordino
Quoted 11 lines

My ARM board hanged at the initial "Calibrating delay loop" message.

After some inspection, I found out the problem to be with commit adf8b37bafc1495393201a2ae4235846371870d0. This commit introduces stack unwinding for ARM, and set it enabled by default. However, it seems to not work with buggy or not-EABI compilers.

My suggestion is to keep the feature (which is fine) but change the default setting of the option (see the attached patch).

Quoted 8 lines

The option still depends on EXPERIMENTAL, so you get ARM_UNWIND on when enabling that.

I'm more in favour of a #warning on #error in the unwind.c file based on the compiler version rather than not having it on by default. The reason is that people reported performance improvements when compiling the kernel without frame pointers.

This solution is fine too (even if I still think that changing the default setting is better).

Please, consider the patch in attachment: is it like you would have it ?

BTW, do we have any list of buggy or not-EABI versions of the gcc compiler ?

Many thanks,

Claudio

From: Claudio Scordino <[email protected]>
Date: Mon, 19 Oct 2009 11:59:16 +0200
Subject: [PATCH 1/1] Disable stack unwinding support by default, since it does not work on some buggy or not-EABI compilers.

Disable stack unwinding support by default, since it does not work on some buggy or not-EABI compilers.

---
 arch/arm/Kconfig.debug |    4 ++--
 1 files changed, 2 insertions(+), 2 deletions(-)
arch/arm/Kconfig.debug +2 -2
index 1a6f70e..925e3fd 100644
@@ -19,13 +19,13 @@ config FRAME_POINTER
 config ARM_UNWIND
 	bool "Enable stack unwinding support"
 	depends on AEABI && EXPERIMENTAL
-	default y
+	default n
 	help
 	  This option enables stack unwinding support in the kernel
 	  using the information automatically generated by the
 	  compiler. The resulting kernel image is slightly bigger but
 	  the performance is not affected. Currently, this feature
-	  only works with EABI compilers. If unsure say Y.
+	  only works with EABI compilers. If unsure say N.
 
 config DEBUG_USER
 	bool "Verbose user fault messages"
Signature · 1 line
1.6.0.4
#7 lore raw
In reply to Catalin Marinas

Catalin Marinas ha scritto:

Quoted 13 lines · Claudio Scordino
Quoted 11 lines

My ARM board hanged at the initial "Calibrating delay loop" message.

After some inspection, I found out the problem to be with commit adf8b37bafc1495393201a2ae4235846371870d0. This commit introduces stack unwinding for ARM, and set it enabled by default. However, it seems to not work with buggy or not-EABI compilers.

My suggestion is to keep the feature (which is fine) but change the default setting of the option (see the attached patch).

Quoted 8 lines

The option still depends on EXPERIMENTAL, so you get ARM_UNWIND on when enabling that.

I'm more in favour of a #warning on #error in the unwind.c file based on the compiler version rather than not having it on by default. The reason is that people reported performance improvements when compiling the kernel without frame pointers.

[Sorry, I sent the wrong patch... This is the new one.]

This solution is fine too (even if I still think that changing the default setting is better).

Please, consider the patch in attachment: is it like you would have it ?

BTW, do we have any list of buggy or not-EABI versions of the gcc compiler ?

Many thanks,

Claudio

From: Claudio Scordino <[email protected]>
Date: Mon, 26 Oct 2009 17:13:18 +0100
Subject: [PATCH 1/1] Some warnings when compiling ARM unwind support with a buggy or not-EABI compiler.

ARM unwind support is known to build only with EABI and not-buggy compilers. Now we check the compiler and raise a #warning in case of wrong compiler.

---
 arch/arm/kernel/unwind.c |    9 +++++++++
 1 files changed, 9 insertions(+), 0 deletions(-)
arch/arm/kernel/unwind.c +9 -0
index 39baf11..47345c1 100644
@@ -26,6 +26,15 @@
  * http://infocenter.arm.com/help/topic/com.arm.doc.subset.swdev.abi/index.html
  */
 
+#if !defined (__ARM_EABI__)
+#warning Your compiler does not have EABI support.
+#warning    ARM unwind support it is known to compile only on EABI compilers.
+#warning    Change compiler or disable ARM_UNWIND option.
+#elif (__GNUC__ == 4 && __GNUC_MINOR__ == 2)
+#warning Your compiler is too buggy; it is known to not compile ARM unwind support.
+#warning    Change compiler or disable ARM_UNWIND option.
+#endif
+
 #include <linux/kernel.h>
 #include <linux/init.h>
 #include <linux/module.h>
Signature · 1 line
1.6.0.4
#8 lore raw
Quoted 16 lines · Claudio Scordino

Catalin Marinas ha scritto:

Quoted 15 lines · Claudio Scordino
Quoted 10 lines

My ARM board hanged at the initial "Calibrating delay loop" message.

After some inspection, I found out the problem to be with commit adf8b37bafc1495393201a2ae4235846371870d0. This commit introduces stack unwinding for ARM, and set it enabled by default. However, it seems to not work with buggy or not-EABI compilers.

My suggestion is to keep the feature (which is fine) but change the default setting of the option (see the attached patch).

The option still depends on EXPERIMENTAL, so you get ARM_UNWIND on when enabling that.

Quoted 6 lines
Quoted 4 lines

I'm more in favour of a #warning on #error in the unwind.c file based on the compiler version rather than not having it on by default. The reason is that people reported performance improvements when compiling the kernel without frame pointers.

[Sorry, I sent the wrong patch... This is the new one.]

and the wrong ARM kernel list (I changed the Cc line).

Quoted 4 lines

This solution is fine too (even if I still think that changing the default setting is better).

Please, consider the patch in attachment: is it like you would have it ?

You can just send it to Russell's patch system once there are no more comments on it.

See below for my comments.

Quoted 2 lines

BTW, do we have any list of buggy or not-EABI versions of the gcc compiler ?

No, but for non-EABI compilers I think the __ARM_EABI__ check is enough.

Quoted 6 lines

From: Claudio Scordino <[email protected]>
Date: Mon, 26 Oct 2009 17:13:18 +0100
Subject: [PATCH 1/1] Some warnings when compiling ARM unwind support with a buggy or not-EABI compiler.

ARM unwind support is known to build only with EABI and not-buggy compilers. Now we check the compiler and raise a #warning in case of wrong compiler.

Maybe you can clarify a bit that the problem is not unwinding information but the -fno-frame-pointer option added as a result of !CONFIG_FRAME_POINTER.

Quoted 10 lines
---
 arch/arm/kernel/unwind.c |    9 +++++++++
 1 files changed, 9 insertions(+), 0 deletions(-)
arch/arm/kernel/unwind.c +0 -0
index 39baf11..47345c1 100644

arch/arm/kernel/unwind.c line 31

Quoted 6 lines
arch/arm/kernel/unwind.c +3 -0
@@ -26,6 +26,15 @@
  * http://infocenter.arm.com/help/topic/com.arm.doc.subset.swdev.abi/index.html
  */
 
+#if !defined (__ARM_EABI__)
+#warning Your compiler does not have EABI support.
+#warning    ARM unwind support it is known to compile only on EABI compilers.

Maybe "ARM unwind is known ... only with ..." (though English is not my first language).

arch/arm/kernel/unwind.c line 33

Quoted 2 lines
arch/arm/kernel/unwind.c +2 -0
+#warning    Change compiler or disable ARM_UNWIND option.
+#elif (__GNUC__ == 4 && __GNUC_MINOR__ == 2)

Could we assume this for (__GNUC_MINOR__ <= 2)?

Quoted 7 lines
arch/arm/kernel/unwind.c +4 -0
+#warning Your compiler is too buggy; it is known to not compile ARM unwind support.
+#warning    Change compiler or disable ARM_UNWIND option.
+#endif
+
 #include <linux/kernel.h>
 #include <linux/init.h>
 #include <linux/module.h>
Signature · 1 line
Catalin
#9 lore raw

Hi,

I modified the patch according to your suggestions.

Quoted 3 lines

You can just send it to Russell's patch system once there are no more comments on it.

Done. I'm attaching the latest version here too.

Many thanks,

Claudio
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 0001-ARM-unwind-is-known-to-compile-only-with-EABI-and-no.patch
Type: text/x-patch
Size: 1372 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/linux-arm-kernel/attachments/20091030/352f32a5/attachment.bin>

#10 lore raw
Quoted 5 lines · Claudio Scordino

From: Claudio Scordino <[email protected]>
Date: Fri, 30 Oct 2009 11:46:00 +0100
Subject: [PATCH 1/1] ARM unwind is known to compile only with EABI and not-buggy compilers.

Quoted 8 lines

ARM unwind is known to compile only with EABI and not-buggy compilers. The problem is not the unwinding information but the -fno-frame-pointer option added as a result of !CONFIG_FRAME_POINTER. Now we check the compiler and raise a #warning in case of wrong compiler.

Unless others have objections, here's my ack:

Signature · 1 line
Catalin