Re: Kernel stack read with PTRACE_EVENT_EXIT and io_uring threads

From: Eric W. Biederman
Date: Wed Jun 23 2021 - 10:34:54 EST

Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx> writes:

> On Tue, Jun 22, 2021 at 1:53 PM Eric W. Biederman <ebiederm@xxxxxxxxxxxx> wrote:
>> Playing with it some more I think I have everything working working
>> except for PTRACE_EVENT_SECCOMP (which can stay ptrace_event) and
>> group_exit(2).
>> Basically in exit sending yourself a signal and then calling do_exit
>> from the signal handler is not unreasonable, as exit is an ordinary
>> system call.
> Ok, this is a bit odd, but I do like the concept of just making
> ptrace_event just post a signal, and have all ptrace things always be
> handled at signal time (or the special system call entry/exit, which
> is fine too).
>> For purposes of discussion this is my current draft implementation.
> I didn't check what is so different about exit_group() that you left
> that as an exercise for the reader, but if that ends up then removing
> the whole "wait synchromously for ptrace" cases for good I don't
> _hate_ this. It's a bit odd, but it would be really nice to limit
> where ptrace picks up data.

I am still figuring out exit_group. I am hoping for sometime today.
My intuition tells me I can do it, and I have a sense of what threads I
need to pull to get there. I just don't know what the code is going to
look like yet.

Basically solving exit_group means moving ptrace_event out of do_exit.

> We do end up doing that stuff in "get_signal()", and that means that
> we have the interaction with io_uring calling it directly, but it's at
> least not a new thing.

The ugliest bit is having to repeat the wait_for_vfork_done both in fork
and in get_signal.