Re: [1/3] kprobes-func-args-268-rc3.patch

[Suparna Bhattacharya <suparna@in.ibm.com>]

On Thu, Aug 05, 2004 at 03:51:12PM +0200, Andi Kleen wrote:
> On Thu, Aug 05, 2004 at 07:03:48PM +0530, Suparna Bhattacharya wrote:
> > On Thu, Aug 05, 2004 at 02:54:23PM +0200, Andi Kleen wrote:
> > > On Thu, Aug 05, 2004 at 05:54:31PM +0530, Suparna Bhattacharya wrote:
> > > > > I think you misunderstood Linus' suggestion.  The problem with
> > > > > modifying arguments on the stack frame is always there because the C
> > > > > ABI allows it. One suggested solution was to use a second function
> > > > 
> > > > I did realise that it is the ABI which allows this, but I thought
> > > > that the only situation in which we know gcc to actually clobber
> > > > arguments from the callee in practice is for tailcall optimization. 
> > > 
> > > It just breaks the most common workaround. 
> > 
> > Just curious, do you know if other cases/optimizations where the
> > callee clobbers arguments on stack ?
> 
> gcc can do it all the time. tail call is just a special case.
> 
> It happens relatively rarely because often it caches the data
> in registers. But when there is enough register pressure it can
> be written back to the original argument slot.
> 
> If you pass a structure by value it happens more often
> (at least in older gccs, 3.5 now loads this into registers too) 
> 
> On -O0 code I would also expect it to happen often.
> 
> > 
> > > 
> > > > I'm not sure if that can be guaranteed and yes saving bytes from
> > > > stack would avoid the problem totally (hence the comment) and make
> > > > it less tied to expected innards of the compiler. The only issue 
> > > > with that is deciding the maximum number of arguments so it is 
> > > > generic enough. 
> > > 
> > > 64bytes, aka 16 arguments seem far enough.
> > 
> > OK, is there is consensus on this ? 
> 
> I think there is a clear consensus that anybody who uses 16 arguments
> in a kernel function already did something very wrong.
> 
> Passing structures by value may be reasonable, but not supporting
> that for big structures is a reasonable restriction.
> 
> > We'd have to make the code check for stack boundary etc and probably 
> > compare and copy back only if there has been a change.
> 
> Why? And what stack boundary?
> 
> The probe is not supposed to modify the arguments, isn't it, so I don't
> see why you ever want to copy back.  
> 
> I would write the trampoline in assembly btw, doing such things in C is usually
> very fragile.

Oh, I guess you are suggesting actually copying args forward on stack 
before passing control to the probe routine. Hence the assembly 
trampoline.

What I had in mind originally was a little different - i.e. just 
make the kprobes breakpoint handler itself save the additional 64 KB 
from stack the way it does for pt_regs right now, before iret ing
into the jprobe handler function. And verify and restore that only 
if needed on the way back (i.e. second trap).

Regards
Suparna

> 
> > > > > >  
> > > > 
> > > > Even with CONFIG_REGPARM, if you have a large 
> > > > number of arguments for example, is spill over into stack 
> > > > a possibility ?
> > > 
> > > Yes. For more than three (Linux uses -mregparm=3) 
> > > Also varargs arguments will be always on the stack I think.
> > 
> > Right, so making the copy dependent on !CONFIG_REGPARM wouldn't
> > make sense would it ?
> 
> Yes, it wouldn't.
> 
> -Andi

-- 
Suparna Bhattacharya (suparna@in.ibm.com)
Linux Technology Center
IBM Software Lab, India