[PATCH] display mask on enter to sigreturn, not exit

[Dmitry V. Levin]

On Wed, Jan 18, 2012 at 11:09:58AM +0100, Denys Vlasenko wrote:
> On 01/17/2012 10:46 PM, Dmitry V. Levin wrote:
> > On Tue, Jan 17, 2012 at 08:59:46PM +0100, Denys Vlasenko wrote:
> >> On 01/17/2012 08:43 PM, Dmitry V. Levin wrote:
> >>> On Tue, Jan 17, 2012 at 08:01:00PM +0100, Denys Vlasenko wrote:
> >>>> On 01/17/2012 07:53 PM, Dmitry V. Levin wrote:
> >>> [...]
> >>>>> [X86_64] (x86_64_regs): Remove.
> >>>>> * syscall.c [X86_64] (x86_64_regs): Make static.
> >>>>
> >>>> I think we will eventually need register structs to be accessible
> >>>> from all files. That's why I made x86_64_regs global too,
> >>>> even though it does not have to be global (so far).
> >>>>
> >>>> I would like to ask for it to remain global.
> >>>
> >>> I'd probably agree, but the most astonishing thing is that global
> >>> x86_64_regs doesn't work for me at all:
> >>>
> >>> $ ./strace /bin/true
> >>> Segmentation fault
> >>
> >> Mysterious...
> >
> > There is no mystic, just a commonplace data corruption.
> > On x86-64, sizeof(struct user_regs_struct) == 27 * sizeof(long),
> > but sizeof(struct pt_regs) == 21 * sizeof(long).
> >
> > strace passes x86_64_regs of type "struct pt_regs" to PTRACE_GETREGS,
> > which expects a pointer to "struct user_regs_struct", and happily
> > overwrites 6 extra long words.  It could crash much earlier than now.
> >
> > We probably had to use user_regs_struct instead of pt_regs.
> 
> Looks that way - I checked glibc and kernel headers and.
> 
> I just committed the following change:
> 
> diff --git a/syscall.c b/syscall.c
> index d356006..9c04cf9 100644
> --- a/syscall.c
> +++ b/syscall.c
> @@ -737,7 +737,13 @@ struct tcb *tcp_last = NULL;
>   # if defined(I386)
>   struct pt_regs i386_regs;
>   # elif defined(X86_64)
> -static struct pt_regs x86_64_regs;
> +/*
> + * On 32 bits, pt_regs and user_regs_struct are the same,
> + * but on 64 bits, user_regs_struct has six more fields:
> + * fs_base, gs_base, ds, es, fs, gs.
> + * PTRACE_GETREGS fills them too, so struct pt_regs would overflow.
> + */
> +static struct user_regs_struct x86_64_regs;

Thanks, this finally fixes regression introduced by v4.6-124-geb0e3e8.
btw, #include <asm/ptrace.h> in defs.h is no longer needed for X86_64.

I suggest to do the same change for i386_regs as well, just to keep it
simple.  Then the the whole change to defs.h made by v4.6-124-geb0e3e8
could be reverted.


-- 
ldv
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 198 bytes
Desc: not available
URL: <http://lists.strace.io/pipermail/strace-devel/attachments/20120118/10f6691c/attachment.bin>