Re: [PATCH 12/X] uprobes: x86: introduce abort_xol()

[Oleg Nesterov <oleg@redhat.com>]

On 10/21, Ananth N Mavinakayanahalli wrote:
>
> On Fri, Oct 21, 2011 at 08:12:07PM +0530, Srikar Dronamraju wrote:
>
> > > +void abort_xol(struct pt_regs *regs)
> > > +{
> > > +	// !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
> > > +	// !!! Dear Srikar and Ananth, please implement me !!!
> > > +	// !!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
> > > +	struct uprobe_task *utask = current->utask;
> > > +	regs->ip = utask->vaddr;
> >
> > nit:
> > Shouldnt we be setting the ip to the next instruction after this
> > instruction?
>
> No, since we should re-execute the original instruction

Yes,

> after removing
> the breakpoint.

No? we should not remove this uprobe?

> Also, wrt ip being set to the next instruction on a breakpoint hit,
> that's arch specific.

Probably yes, I am not sure. But:

> For instance, on x86, it points to the next
> instruction,

No?

	/**
	 * get_uprobe_bkpt_addr - compute address of bkpt given post-bkpt regs
	 * @regs: Reflects the saved state of the task after it has hit a breakpoint
	 * instruction.
	 * Return the address of the breakpoint instruction.
	 */
	unsigned long __weak get_uprobe_bkpt_addr(struct pt_regs *regs)
	{
		return instruction_pointer(regs) - UPROBES_BKPT_INSN_SIZE;
	}

Yes, initially regs->ip points to the next insn after int3, but
utask->vaddr == get_uprobe_bkpt_addr() == addr of int3.

Right?

> while on powerpc, the nip points to the breakpoint vaddr
> at the time of exception.

I think get_uprobe_bkpt_addr() should be consistent on every arch.
That is why (I think) it is __weak.

Anyway, abort_xol() has to be arch-specific.

Oleg.