possible bug in powerpc LE compat syscalls with 64-bit args

possible bug in powerpc LE compat syscalls with 64-bit args

[Chris Metcalf]

Adhemerval Zanella is cleaning up the preadv/pwritev implementation in glibc
and came across an issue with the tilepro/tilegx32 implementation of argument
passing for 64-bit arguments in split register pairs.

The glibc code base uses __LONG_LONG_PAIR to split 64-bit arguments,
passing "hi, lo" for BE systems and "lo, hi" for LE systems.  Tile is almost uniformly
LE, but we do support BE for specific customers.  The 32-bit compat mode on
tilegx was getting passed "hi, lo" arguments by glibc but looking for them to
be "lo, hi" arguments on the kernel side.  I fixed this here:

https://lkml.kernel.org/g/1466019219-10462-1-git-send-email-cmetcalf@mellanox.com

Reviewing what other platforms do, it seems like powerpc compat mode may
have the opposite problem in little-endian mode, since arguments are passed
in "hi, lo" order unconditionally in arch/powerpc/kernel/sys_ppc32.c.  You may
want to adopt the solution in the patch cited above, or similar solutions
elsewhere, eg regs_to_64 in arch/arm64/include/asm/assembler.h, or
merge_64 in arch/mips/kernel/linux32.c.

Here's the email thread from libc-alpha:

https://sourceware.org/ml/libc-alpha/2016-06/msg00638.html

-- 
Chris Metcalf, Mellanox Technologies
http://www.mellanox.com

Re: possible bug in powerpc LE compat syscalls with 64-bit args

[Adhemerval Zanella]

On 16/06/2016 12:36, Chris Metcalf wrote:
> Adhemerval Zanella is cleaning up the preadv/pwritev implementation in glibc
> and came across an issue with the tilepro/tilegx32 implementation of argument
> passing for 64-bit arguments in split register pairs.
> 
> The glibc code base uses __LONG_LONG_PAIR to split 64-bit arguments,
> passing "hi, lo" for BE systems and "lo, hi" for LE systems.  Tile is almost uniformly
> LE, but we do support BE for specific customers.  The 32-bit compat mode on
> tilegx was getting passed "hi, lo" arguments by glibc but looking for them to
> be "lo, hi" arguments on the kernel side.  I fixed this here:
> 
> https://lkml.kernel.org/g/1466019219-10462-1-git-send-email-cmetcalf@mellanox.com
> 
> Reviewing what other platforms do, it seems like powerpc compat mode may
> have the opposite problem in little-endian mode, since arguments are passed
> in "hi, lo" order unconditionally in arch/powerpc/kernel/sys_ppc32.c.  You may
> want to adopt the solution in the patch cited above, or similar solutions
> elsewhere, eg regs_to_64 in arch/arm64/include/asm/assembler.h, or
> merge_64 in arch/mips/kernel/linux32.c.
> 
> Here's the email thread from libc-alpha:
> 
> https://sourceware.org/ml/libc-alpha/2016-06/msg00638.html
> 

I am not fully familiar with kernel policy regarding ABI so is this considered
an ABI breakage or a real defect?

Re: possible bug in powerpc LE compat syscalls with 64-bit args

[Michael Ellerman]

On Thu, 2016-06-16 at 11:36 -0400, Chris Metcalf wrote:

> Adhemerval Zanella is cleaning up the preadv/pwritev implementation in glibc
> and came across an issue with the tilepro/tilegx32 implementation of argument
> passing for 64-bit arguments in split register pairs.
> 
> The glibc code base uses __LONG_LONG_PAIR to split 64-bit arguments,
> passing "hi, lo" for BE systems and "lo, hi" for LE systems.  Tile is almost uniformly
> LE, but we do support BE for specific customers.  The 32-bit compat mode on
> tilegx was getting passed "hi, lo" arguments by glibc but looking for them to
> be "lo, hi" arguments on the kernel side.  I fixed this here:
> 
> https://lkml.kernel.org/g/1466019219-10462-1-git-send-email-cmetcalf@mellanox.com
> 
> Reviewing what other platforms do, it seems like powerpc compat mode may
> have the opposite problem in little-endian mode, since arguments are passed
> in "hi, lo" order unconditionally in arch/powerpc/kernel/sys_ppc32.c.  You may
> want to adopt the solution in the patch cited above, or similar solutions
> elsewhere, eg regs_to_64 in arch/arm64/include/asm/assembler.h, or
> merge_64 in arch/mips/kernel/linux32.c.

We don't support compat mode on ppc64le. So I think we're OK?

Thanks for letting us know.

cheers

Re: possible bug in powerpc LE compat syscalls with 64-bit args

[Andreas Schwab]

Chris Metcalf <cmetcalf@mellanox.com> writes:

> Reviewing what other platforms do, it seems like powerpc compat mode may
> have the opposite problem in little-endian mode, since arguments are passed
> in "hi, lo" order unconditionally in arch/powerpc/kernel/sys_ppc32.c.

PPC32 is always big-endian.

Andreas.

-- 
Andreas Schwab, schwab@linux-m68k.org
GPG Key fingerprint = 58CA 54C7 6D53 942B 1756  01D3 44D5 214B 8276 4ED5
"And now for something completely different."

Re: possible bug in powerpc LE compat syscalls with 64-bit args

[Chris Metcalf]

On 6/16/2016 5:42 PM, Andreas Schwab wrote:
> Chris Metcalf <cmetcalf@mellanox.com> writes:
>
>> Reviewing what other platforms do, it seems like powerpc compat mode may
>> have the opposite problem in little-endian mode, since arguments are passed
>> in "hi, lo" order unconditionally in arch/powerpc/kernel/sys_ppc32.c.
> PPC32 is always big-endian.

Sounds good then.

Commit 422b9b9684db ("powerpc/compat: 32-bit little endian machine
name is ppcle, not ppc") made me think there was support for a
little-endian 32-bit mode.

-- 
Chris Metcalf, Mellanox Technologies
http://www.mellanox.com