Re: [PATCH] Fix siginfo._uid bug

[Maxim Kuvyrkov <maxim@codesourcery.com>]

Andreas Schwab wrote:
> Maxim Kuvyrkov <maxim@codesourcery.com> writes:
> 
>> The bug is rather elegant and has been present in sources for years.
> 
> For more than 9 years, to be precise, when 32 bit uids were introduced.
> 
>> The problem is that m68k uses a custom siginfo layout due to having
>> a 16-bit uid field for 'backward compatibility'.  I.e., siginfo._kill
>> fields are:
> 
> I've found this interesting thread from 2000:
> 
> http://marc.info/?l=linux-kernel&m=94829131029274&w=2
> 
> The change went in 2.3.41, and undid the change made in 2.3.39.  It was
> reintroduced (for asm-m68k/siginfo.h only and with __kernel_* types; no
> other siginfo header used them at that time) in 2.4.0-test12.
> 
> Now looking at the glibc side, m68k has always used the generic
> linux/bits/siginfo.h (with a single 32bit uid field), until Richard
> Sandiford introduced linux/m68k/siginfo.h in October 2006, copying the
> (broken) layout from the kernel.
> 
> Given that there is no way to keep backward compatibility it might be a
> good opportunity to do same cleanup here.  Like going back to the
> generic layout.

There is yet another bug in siginfo.  si_sigval is expected to be at the 
same offsets in _timer and _rt.  At the moment si_sigval is at offset 8 
in _timer and at offset 6 in _rt (the patch for fixing uids makes that 
8@_timer and 10@_rt).

This bug causes rt/tst-*timer* tests fail.

Moving on to the generic version of siginfo.h will certainly fix the 
problem, otherwise, one needs to pad extra 2 bytes in _timer.

--
Maxim K.
CodeSourcery