Skip to contentSkip to conversation

More ARM binutils fuckage

DISCUSSION Russell King 6 participants 9 messages Dec 5, 2006 – Dec 6, 2006

lore
Share & tools
lore
#0 lore raw

There's not much to say about this, other than scream and go hide in the corner. ARM toolchains are just basically fscked.

arm-linux-ld -EL -p --no-undefined -X -o .tmp_vmlinux1 -T
arch/arm/kernel/vmlinux.lds arch/arm/kernel/head.o arch/arm/kernel/init_task.o init/built-in.o --start-group
usr/built-in.o arch/arm/kernel/built-in.o arch/arm/mm/built-in.o arch/arm/common/built-in.o arch/arm/mach-versatile/built-in.o arch/arm/nwfpe/built-in.o arch/arm/vfp/built-in.o kernel/built-in.o mm/built-in.o fs/built-in.o ipc/built-in.o security/built-in.o crypto/built-in.o block/built-in.o arch/arm/lib/lib.a lib/lib.a arch/arm/lib/built-in.o lib/built-in.o drivers/built-in.o sound/built-in.o net/built-in.o --end-group

Produces no error, but:

$ arm-linux-nm ../build/versatile/.tmp_vmlinux1 |grep ' U '
         U __divdi3
         U __udivdi3
         U __umoddi3

Duh.

Result is you get successful kernel builds without any error, but a *bad* incomplete image. IOW, you do not know whether or not your kernel is _safe_ to execute.

That's binutils 2.17. binutils 2.16 had this problem and I tried to get a patch into the kernel to use 'nm' to detect this fsckage. We *need* this patch in the kernel.

Sam, can we resurrect my attempts to do this please?

Signature · 3 lines
Russell King
 Linux kernel    2.6 ARM Linux   - http://www.arm.linux.org.uk/
 maintainer of:
#1 lore raw
Quoted 2 lines · Russell King

There's not much to say about this, other than scream and go hide in the corner. ARM toolchains are just basically fscked.

And while we're on the topic of ARM linux toolchain fsckage, it would be nice to know what patches and incantations are currently recommended when configuring gcc for building various modern ARM kernel/ABI configurations (OABI + soft VFP, EABI, etc.).

There has been some discussion on the crossgcc mailing list, especially around recent compilers and NPTL/TLS, and crosstool has accumulated some of the relevant patches. If an expert (i. e., someone who has built their own toolchain and gotten an ARM EABI/NPTL system all the way up) were to reply to this message with some details, we might be able to coax chip vendors off of hard-float gcc-3.3/glibc-2.3.x/linuxthreads toolchains.

Cheers,
- Michael

#2 lore raw
Quoted 8 lines · Michael K. Edwards
Quoted 2 lines · Russell King

There's not much to say about this, other than scream and go hide in the corner. ARM toolchains are just basically fscked.

And while we're on the topic of ARM linux toolchain fsckage, it would be nice to know what patches and incantations are currently recommended when configuring gcc for building various modern ARM kernel/ABI configurations (OABI + soft VFP, EABI, etc.).

There is no such thing as soft VFP.

I can only talk from the requirements of the kernel. gcc 3.4.3 is the minimum for ARM, which with binutils 2.17 will allow you to build the kernel as OABI in *any* configuration. No patches required for either.

Enabling EABI needs a compiler which supports EABI. That's where I get fuzzy but recent gcc 4 should be suitable. I have had it suggested to me that EABI support in the toolchain isn't all that stable at the moment.

As for userspace... and NPTL, that's a different matter.

Signature · 3 lines
Russell King
 Linux kernel    2.6 ARM Linux   - http://www.arm.linux.org.uk/
 maintainer of:
#3 lore raw
Quoted 4 lines · Russell King

Enabling EABI needs a compiler which supports EABI. That's where I get fuzzy but recent gcc 4 should be suitable. I have had it suggested to me that EABI support in the toolchain isn't all that stable at the moment.

I use a bog-standard gcc 4.1.0 to cross-compile all my ARM kernels with, which allows me to build both old-ABI and EABI kernels. (These days I build all kernels in EABI mode with old-ABI compat.) I have not run into any code generation issues with this compiler yet.

On the ARM I am running an EABI userland with glibc 2.5, and build stuff natively with vanilla binutils 2.17.50.0.5 (2.17.5.0.3 is 'too old', as it doesn't understand the two argument form of the .movsp directive which one of the gcc ICE fix patches emits), and gcc 4.1.1 with some patches from gcc bugzilla to fix an ICE or two.

#4 lore raw

Hello Lennert,

Wednesday, December 6, 2006, 3:08:13 AM, you wrote:

[]

Quoted 3 lines

(These
days I build all kernels in EABI mode with old-ABI compat.) I have not run into any code generation issues with this compiler yet.

I wonder, if OABI-compat is known to actually work on OABI userspace,
I mean, on something real, like xserver-kdrive ;-). Because I'd really like to build single kernel for both old and new userspace too, but afraid to try that, fearing to be put down by another broken feature ;-).

Signature · 2 lines
Best regards,
 Paul                            mailto:[email protected]
#5 lore raw

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

Paul Sokolovsky schreef:

Quoted 9 lines

Hello Lennert,

Wednesday, December 6, 2006, 3:08:13 AM, you wrote:

[]

Quoted 3 lines

(These
days I build all kernels in EABI mode with old-ABI compat.) I have not run into any code generation issues with this compiler yet.

Quoted 5 lines

I wonder, if OABI-compat is known to actually work on OABI userspace,
I mean, on something real, like xserver-kdrive ;-). Because I'd really like to build single kernel for both old and new userspace too, but afraid to try that, fearing to be put down by another broken feature ;-).

It does work, I routinely switched between softfpa and eabi userspace without reflashing the kernel. You just need to make sure that both userspace images have the same modules.

regards,

Koen
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.5 (Darwin)

iD8DBQFFdnytMkyGM64RGpERAuvpAJ0ccwFvMWQrU1V3THSd1FlaDDwOTACfbj92 +PF+IsXCGxCJx1YThB9NHdQ=
=B7Ff
-----END PGP SIGNATURE-----

#6 lore raw
In reply to Paul Sokolovsky
Quoted 9 lines · Paul Sokolovsky

Hello Lennert,

Wednesday, December 6, 2006, 3:08:13 AM, you wrote:

[]

Quoted 3 lines

(These
days I build all kernels in EABI mode with old-ABI compat.) I have not run into any code generation issues with this compiler yet.

Quoted 5 lines

I wonder, if OABI-compat is known to actually work on OABI userspace,
I mean, on something real, like xserver-kdrive ;-). Because I'd really like to build single kernel for both old and new userspace too, but afraid to try that, fearing to be put down by another broken feature ;-).

I've used OABI-compat + EABI ARM kernels to routinely switch between OABI and EABI rootfs boots. There were some minor busybox issues on the OABI-compat side (w.r.t. syscalls), but nothing major. These were fairly complex applications, so "it should work" (and IIRC the patches were publicly available). But I'm sure I'm forgetting some significant detail and can't go check right now.

Quoted 12 lines

--
Best regards,
Paul mailto:[email protected]

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

Signature · 2 lines
"Thou shalt not follow the NULL pointer, for chaos and madness await
thee at its end"
#7 lore raw
In reply to Russell King
Quoted 1 line · Russell King

There is no such thing as soft VFP.

Patches have been floating around for quite some time that implement soft float with VFP parameter passing conventions (which notably implies native endianness, unlike FPA). They all seem to derive from Nicolas Pitre's patch at
http://lists.arm.linux.org.uk/pipermail/linux-arm/2003-October/006436.html
, which is said to have some bugs but to supply the majority of the needed functionality. (See, for instance,
http://www.busybox.net/cgi-bin/viewcvs.cgi/branches/buildroot.mjn3/toolchain/gcc/3.4.6/arm-softfloat.patch.conditional?rev=14854&view=auto.)
I assume that I'm not the first to fix it up for gcc 4.1.1.

Quoted 4 lines

I can only talk from the requirements of the kernel. gcc 3.4.3 is the minimum for ARM, which with binutils 2.17 will allow you to build the kernel as OABI in *any* configuration. No patches required for either.

It would be nice if this appeared prominently in Documentation/arm/whatever so that vendors who are stuck on gcc 3.3 (and binutils so old that they comment things out in vmlinux.lds) can be encouraged to move forward.

Cheers,
- Michael

#8 lore raw
In reply to Russell King
Quoted 14 lines · Russell King

There's not much to say about this, other than scream and go hide in the corner. ARM toolchains are just basically fscked.

arm-linux-ld -EL -p --no-undefined -X -o .tmp_vmlinux1 -T
arch/arm/kernel/vmlinux.lds arch/arm/kernel/head.o arch/arm/kernel/init_task.o init/built-in.o --start-group
usr/built-in.o arch/arm/kernel/built-in.o arch/arm/mm/built-in.o arch/arm/common/built-in.o arch/arm/mach-versatile/built-in.o arch/arm/nwfpe/built-in.o arch/arm/vfp/built-in.o kernel/built-in.o mm/built-in.o fs/built-in.o ipc/built-in.o security/built-in.o crypto/built-in.o block/built-in.o arch/arm/lib/lib.a lib/lib.a arch/arm/lib/built-in.o lib/built-in.o drivers/built-in.o sound/built-in.o net/built-in.o --end-group

Quoted 8 lines

Produces no error, but:

$ arm-linux-nm ../build/versatile/.tmp_vmlinux1 |grep ' U '
         U __divdi3
         U __udivdi3
         U __umoddi3

Duh.

I'm lead to believe that these are due to gcc issuing .globl directives for these symbols, but not actually referencing them. Hence the symbol is marked undefined in the symbol table, but no reloations actually exist.

Hence why the linker (correctly) doesn't fail.

Ergo, no problem. Please ignore the previous mail.

Signature · 3 lines
Russell King
 Linux kernel    2.6 ARM Linux   - http://www.arm.linux.org.uk/
 maintainer of: