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

[Oleg Nesterov <oleg@redhat.com>]

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!

Let me clarify just in case. Yes, _personally_ I think we should try
to avoid the vma games, and it looks better to me this way. But I won't
argue if you change your mind, I understand this approach has its own
disadvantages.

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?

Or perhaps we can change parse_probe_arg("@") to update "param" ? Yes,
in this case it needs another argument, not sure...

> 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.

> This symbol offset calculation was done in the getsymoff which implemented
> like below (I'm sure there's a much simpler way to do this, but ...).

Perhaps I'll even try to read/understand it later, but this elf stuff is
the black magic to me ;)

Oleg.