Re: [PATCH] uprobes/core: handle breakpoint and signal step exception.

[Ingo Molnar <mingo@elte.hu>]

* Ingo Molnar <mingo@elte.hu> wrote:

> 
> * Srikar Dronamraju <srikar@linux.vnet.ibm.com> wrote:
> 
> > * Ingo Molnar <mingo@elte.hu> [2012-03-08 14:48:09]:
> > 
> > > 
> > > * Srikar Dronamraju <srikar@linux.vnet.ibm.com> wrote:
> > > 
> > > > @@ -233,9 +233,11 @@ static inline void __user *arch_compat_alloc_user_space(long len)
> > > >  
> > > >  	if (test_thread_flag(TIF_IA32)) {
> > > >  		sp = task_pt_regs(current)->sp;
> > > > +#ifdef CONFIG_X86_64
> > > >  	} else {
> > > >  		/* -128 for the x32 ABI redzone */
> > > >  		sp = __this_cpu_read(old_rsp) - 128;
> > > > +#endif
> > > >  	}
> > > >  
> > > >  	return (void __user *)round_down(sp - len, 16);
> > > 
> > > So 'sp' is undefined if that TIF check fails?
> > > 
> > > Also, on a 32-bit kernel the TIF check probably fails all the 
> > > time, because we don't set TIF_IA32 (and don't know that flag).
> > 
> > > 
> > > It would probably be better to make the whole helper inline 
> > > #ifdef 64-bit, it does not look very useful on 32-bit.
> > > 
> > 
> > arch_compat_alloc_user_space gets called from compat_alloc_user_space
> > which is arch agnostic and exported too.
> > 
> > So I will change this to
> > 
> > void __user *arch_compat_alloc_user_space(long len)
> > {
> >   	if (is_ia32_compat_task(current))
> >   		sp = task_pt_regs(current)->sp;
> > #ifdef CONFIG_X86_64
> >   	else
> >   		/* -128 for the x32 ABI redzone */
> >   		sp = __this_cpu_read(old_rsp) - 128;
> > #endif
> >   
> >   	return (void __user *)round_down(sp - len, 16);
> > }
> > 
> > where is_ia32_compat_task() is the new macro that you 
> > suggested we put in compat.h which would return true if the 
> > task is 32 bit emulated on x86_64 or running on i386 machine.
> > 
> > Hence we can avoid the case where sp is not set.
> 
> Ok - looks good at first glance.

It does not look good on a second glance though, once I checked 
your latest patches.

arch_compat_alloc_user_space() is arch agnostic on 
*CONFIG_COMPAT=y* kernels.

It's generally not available on 32-bit builds - CONFIG_COMPAT is 
a facility to provide 32-bit syscall compatibility on 64-bit 
kernels. Such a facility is not needed on 32-bit kernels.

So providing this:

> void __user *arch_compat_alloc_user_space(long len)
> {
>	if (is_ia32_compat_task(current))
>               sp = task_pt_regs(current)->sp;

on 32-bit systems makes little sense.

So ... instead of adding is_compat_task() to compat.h it would 
be better to add it to another x86 header (processor.h might be 
good but I have not checked very hard) and maybe name it 
is_32bit_task() or so, to make sure there's no confusion with 
CONFIG_COMPAT=y methods.

I.e. you could drop this patch altogether:

  x86/trivial: Fix 'old_rsp' undefined build failure when including asm/compat.h

And rework the is_ia32_compat_task() patch to use another header 
and to use the is_32bit_task() name. Also, you should double 
check whether the x32 execution model needs special 
consideration as well:

 #define TIF_IA32		17	/* IA32 compatibility process */
 #define TIF_X32		30	/* 32-bit native x86-64 binary */

otherwise uprobe will not work with x32 tasks properly.

Thanks,

	Ingo