Re: [patch 15/21] ptrace changes

[Martin Schwidefsky <schwidefsky@de.ibm.com>]

On Thu, 2008-11-06 at 12:24 -0600, David Smith wrote:
> Martin Schwidefsky wrote:
> > On Mon, 2008-11-03 at 11:14 -0600, David Smith wrote: 
> > 
> >> Is this correct, or should the 1st syscall argument be found in
> >> regs->orig_gpr2 by syscall_get_arguments()/syscall_set_arguments()?
> > 
> > The question is when do syscall_get_arguments and syscall_set_arguments
> > functions get called? If they are called on a call chain that started
> > from do_syscall_trace_enter then we'd have to use orig_gpr2 instead of
> > gprs[2] but if the functions are not called via do_syscall_trace_enter
> > the first argument is located in grprs[2]. As far as I can see the sole
> > user of syscall_get_arguments is collect_syscall which is used to get
> > the registers of a blocked process. In this case the kernel call chain
> > does not include do_syscall_trace_enter, therefore the first argument is
> > in gprs[2], no?
> 
> But, collect_syscall() also calls syscall_get_nr():
> 
> 	*callno = syscall_get_nr(target, regs);
> 	if (*callno != -1L && maxargs > 0)
> 		syscall_get_arguments(target, regs, 0, maxargs, args);
> 
> Both syscall_get_nr() *and* syscall_get_arguments() returning gprs[2]
> can't be right, can it?

Indeed, this cannot work. syscall_get_nr() requires that it is called on
a call chain that includes do_syscall_trace_enter(). And the fix for it
is not trivial. Probably the best would be to add another field to
pt_regs which contains the system call number. syscall_get_arguments()
could be improved to always use orig_gpr2 instead of grps[2] then it
doesn't matter when it is called. The problematic one is
syscall_set_arguments(), there it depends if grps[2] needs to be stored
of not. Hmm, this needs some thinking ..
Good spotting by the way :-)

-- 
blue skies,
  Martin.

"Reality continues to ruin my life." - Calvin.