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

[Oleg Nesterov <oleg@redhat.com>]

On 11/04, Oleg Nesterov wrote:
>
> On 11/04, Oleg Nesterov wrote:
> >
> > On 11/04, Oleg Nesterov wrote:
> > >
> > > But in any case, I strongly believe that it doesn't make any sense to
> > > rely on tu->inode in get_user_vaddr().
> >
> > Hmm. But I forgot about the case when you probe the function in libc
> > and want to dump the variable in libc...
> >
> > So probably I was wrong and this all needs more thinking. Damn.
> > Perhaps we really need to pass @file/offset, but it is not clear what
> > we can do with bss/anon-mapping.
>
> Or. Not that I really like this, but just for discussion...
>
> How about
>
> 	static void __user *get_user_vaddr(struct pt_regs *regs, unsigned long addr)
> 	{
> 		return (void __force __user *)addr + instruction_pointer(regs);
> 	}
>
> ?
>
> This should solve the problems with relocations/randomization/bss.
>
> The obvious disadvantage is that it is not easy to calculate the
> offset we need to pass as an argument, it depends on the probed
> function.

forgot to mention... and instruction_pointer() can't work in ret-probe,
we need to pass the "unsigned long func" arg somehow...

>
> And this still doesn't allow to, say, probe the executable but read
> the data from libc. Unless, again, we attach to the running process
> or randomize_va_space = 0, so we can know it in advance. But otherwise
> I do not think there is any solution.
>
> Oleg.