Re: ptrace(PTRACE_ATTACH) [no intervering wait] ptrace(PTRACE_DETACH) may leave tracee stuck

[Oleg Nesterov <oleg@redhat.com>]

On 07/23, Mike Galbraith wrote:
>
> I received a report that glibc:elf/pldd hangs occasionally, and indeed..
>
>   for i in `seq 1 1000`; do taskset -c 3 pldd $$ > /dev/null 2>&1; done
>
> ..will do so.  Rummage.....
>
> ptrace(PTRACE_DETACH) returns -ESRCH when the trap hasn't happened yet,
> which happens because pldd doesn't wait() before ptrace(PTRACE_DETACH).
>
> pldd source:
>
[...snip...]
>
> Seems this usually works only because cycles expended between attach and
> detach is usually enough to let trap happen so tracee can set its state
> to TASK_TRACED as PTRACE_DETACH expects it to be.
>
> Is this expected behavior?

Yes. PTRACE_ATTACH + PTRACE_DETACH is not correct without wait() in
between, this is expected.

PTRACE_DETACH like (almost) any other ptrace request needs the stopped
tracee. Otherwise, say, ptrace_disable() or flush_ptrace_hw_breakpoint()
are not safe.

We could probably add PTRACE_UNTRACE which only does __ptrace_unlink/etc
like the exiting tracer does. (In particular, it could help to detach a
zombie).

But note that even PTRACE_ATTACH + PTRACE_UNTRACE won't be really correct.
PTRACE_ATTACH sends SIGSTOP, so without sys_wait() in between the tracee
can stop in TASK_STOPPED.

Oleg.