Re: [RFC PATCH] piped/ptraced coredump (was: Dump smaller VMAs first in ELF cores)

[Oleg Nesterov <oleg@redhat.com>]

On 08/04, Linus Torvalds wrote:
>
> On Sun, 4 Aug 2024 at 11:53, Oleg Nesterov <oleg@redhat.com> wrote:
> >
> > Apart from SIGKILL, the dumper already has the full control.
>
> What do you mean? It's a regular usermodehelper. It gets the dump data
> as input. That's all the control it has.

I meant, the dumping thread can't exit until the dumper reads the data
from stdin or closes the pipe. Until then the damper can read /proc/pid/mem
and do other things.

> > And note that the dumper can already use ptrace.
>
> .. with the normal ptrace() rules, yes.
>
> You realize that some setups literally disable ptrace() system calls,
> right? Which your patch now effectively sidesteps.

Well. If, say, selinux disables ptrace, then ptrace_attach() in this
patch should also fail.

But if some setups disable sys_ptrace() as a system call... then yes,
I didn't know that.

> THAT is why I don't like it. ptrace() is *dangerous*.

And horrible ;)

> Just adding some implicit tracing willy-nilly needs to be something
> people really worry about.

Ok, as I said I won't insist.

Oleg.