Re: [PATCH 1/3] ptrace: ptrace_attach: check PF_KTHREAD + exit_state instead of ->mm

[Oleg Nesterov <oleg@redhat.com>]

On 05/05, Roland McGrath wrote:
>
> This changes the order of the already-traced and security checks.
> It would match the previous behavior to have the ->exit_state and ->ptrace
> checks before __ptrace_may_access().  This is a small nit, but it could
> affect whether some existing harmless usage pattern starts generating new
> access failure logging from security modules (e.g. SELinux avc denials).

Another subtle change I forgot to comment.

> I don't see any reason you can't just swap the order back as it was before.

The last patch in series, "do not use task_lock()", is the reason.

We need tasklist for writing to check (and set) ->ptrace, but we need
task_lock() to call __ptrace_may_access().

We can preserve the current behaviour, we can do get_task_mm() beforehand,
modify __ptrace_may_access() a bit, and call __ptrace_may_access() under
tasklist later (in fact, this was the very first version of this patch which
I didn't send).

But do we really care? If selinux denies to ptrace this task, can't we
return -EACESS regardless of ->ptrace?

Oleg.