Re: [PATCH 2/2] coredump: add a new elf note with siginfo fields of the signal

[Pedro Alves <palves@redhat.com>]

On 09/13/2012 06:25 PM, Roland McGrath wrote:
> I agree with Oleg.  If there is an NT_SIGINFO note, it should contain
> exactly the siginfo_t layout and data that we otherwise expose to userland
> already.  That is, it must match what PTRACE_GETSIGINFO reports, which (I
> think) also matches exactly what appears on the stack for a signal
> delivery.  

Agreed.  Please.  One note though:

PTRACE_GETSIGINFO always returns the siginfo_t in the architecture/bitness
of the kernel.  IOW, PTRACE_GETSIGINFO on a 32-bit debuggee, running on a
64-bit kernel, returns the siginfo_t in 64-bit layout.  IOW, copy_siginfo_to_user32
doesn't apply).  Or maybe that's the architecture of the tracer, instead of the kernel.
Haven't really checked.  In any case, GDB has to basically duplicate the kernel's
copy_siginfo_to_user/from_user for each supported arch (that does biarch) because
of this, for "(gdb) print $_siginfo".
I wish we had a PTRACE_GETSIGINFO variant that returned the siginfo_t in
layout of the program, not the kernel's..

> Note also that compat_binfmt_elf must deliver the 32-bit version
> of siginfo_t, i.e. as copy_siginfo_to_user32 does.

-- 
Pedro Alves