Re: [PATCHSET 00/13] tracing/uprobes: Add support for more fetch methods (v6)

[Oleg Nesterov <oleg@redhat.com>]

On 11/06, Namhyung Kim wrote:
>
> On Tue, 5 Nov 2013 18:45:35 +0100, Oleg Nesterov wrote:
> > On 11/05, Namhyung Kim wrote:
> >>
> >> This is what I have for now:
> >>
> >> static void __user *get_user_vaddr(struct pt_regs *regs, unsigned long addr,
> >> 				   struct trace_uprobe *tu)
> >> {
> >> 	unsigned long base_addr;
> >> 	unsigned long vaddr;
> >>
> >> 	base_addr = instruction_pointer(regs) - tu->offset;
> >> 	vaddr = base_addr + addr;
> >>
> >> 	return (void __force __user *) vaddr;
> >> }
> >>
> >> When I tested it, it was able to fetch global and bss data from both of
> >> executable and library properly.
> >
> > Heh ;) I didn't expect you will agree with this suggestion. But if you
> > think it can work - great!
>
> It seems to work for me well except the cross-fetch.

Yes, but cross-fetching needs something different anyway, so I think we
should discuss this separately.

> But I'm not sure it'll work for every cases.

I think "ip - tu->offset + vaddr" trick should always work, just we need
to calculate this "vaddr" passed as an argument correctly.

Except: user-space can create another executable mapping and call the
probed function via another address, but I think we can ignore this.
And I think we can do nothing in this case, because in this case we
can't even rely on tu->inode.

But,

> It would be great if some
> elf gurus come up and give some feedbacks.
>
> Masami?

Yes.

> > As for "-= tu->offset"... Can't we avoid it? User-space needs to calculate
> > the "@" argument anyway, why it can't also substruct this offset?
>
> Hmm.. it makes sense too. :)

I am no longer sure ;)

This way the "@" argument will look more confusing, it will depend on the
address/offset of the probed insn. But again, I do not know, this is up
to you.

> >> But it still doesn't work for uretprobes
> >> as you said before.
> >
> > This looks simple,
> >
> > 	+	if (is_ret_probe(tu)) {
> > 	+		saved_ip = instruction_pointer(regs);
> > 	+		instruction_pointer_set(func);
> > 	+	}
> > 		store_trace_args(...);
> > 	+	if (is_ret_probe(tu))
> > 	+		instruction_pointer_set(saved_ip);
> >
> > although not pretty.
>
> So for normal non-uretprobes, func == instruction_pointer(), right?

No, for normal non-uretprobes func == 0 (actually, undefined).

> If so, just passing func as you suggested looks better than this.

Not sure I understand... OK, we can change uprobe_trace_func() and
uprobe_perf_func()

		if (!is_ret_probe(tu))
	-		uprobe_trace_print(tu, 0, regs);
	+		uprobe_trace_print(tu, instruction_pointer(regs), regs);
		return 0;

but why?

We need the "saved_ip" ugly hack above only if is_ret_probe() == T and
thus instruction_pointer() doesn't match the address of the probed function.
And there is no way to pass some additional info to call_fetch/etc from
uprobe_*_print().

See also another email...

Oleg.