Re: [PATCH v2] ptrace: fix ptrace vs tasklist_lock race on PREEMPT_RT.

[Peter Zijlstra <peterz@infradead.org>]

On Thu, Mar 31, 2022 at 04:25:42PM +0200, Sebastian Andrzej Siewior wrote:
> As explained by Alexander Fyodorov <halcy@yandex.ru>:
> 
> |read_lock(&tasklist_lock) in ptrace_stop() is converted to sleeping
> |lock on a PREEMPT_RT kernel, and it can remove __TASK_TRACED from
> |task->__state (by moving  it to task->saved_state). If parent does
> |wait() on child followed by a sys_ptrace call, the following race can
> |happen:
> |
> |- child sets __TASK_TRACED in ptrace_stop()
> |- parent does wait() which eventually calls wait_task_stopped() and returns
> |  child's pid
> |- child blocks on read_lock(&tasklist_lock) in ptrace_stop() and moves
> |  __TASK_TRACED flag to saved_state
> |- parent calls sys_ptrace, which calls ptrace_check_attach() and
> |  wait_task_inactive()
> 
> The patch is based on his initial patch where an additional check is
> added in case the __TASK_TRACED moved to ->saved_state. The pi_lock is
> acquired to have stable view on ->__state and ->saved_state.
> 
> wait_task_inactive() needs to check both task states while waiting for the
> expected task state. Should the expected task state be in ->saved_state then
> the task is blocked on a sleeping lock. In this case wait_task_inactive() needs
> to wait until the lock situtation has been resolved (the expected state is in
> ->__state). This ensures that the task is idle and does not wakeup as part of
> lock resolving and races for instance with __switch_to_xtra() while the
> debugger clears TIF_BLOCKSTEP() (noted by Oleg Nesterov).
> 
> [ Fix for ptrace_unfreeze_traced() by Oleg Nesterov ]
> 
> Signed-off-by: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
> ---
> v1…v2:
>   - Use also ->saved_state in task_state_match_and_set().
>   - Wait in wait_task_inactive() until the desired task state is in
>     ->__state so that the task won't wake up a as part of lock
>     resolving. Pointed out by Oleg Nesterov.
> 
>  include/linux/sched.h |  128 ++++++++++++++++++++++++++++++++++++++++++++++++--
>  kernel/ptrace.c       |   25 +++++----
>  kernel/sched/core.c   |   11 +++-
>  3 files changed, 146 insertions(+), 18 deletions(-)
> 
> --- a/include/linux/sched.h
> +++ b/include/linux/sched.h
> @@ -118,12 +118,8 @@ struct task_group;
>  
>  #define task_is_running(task)		(READ_ONCE((task)->__state) == TASK_RUNNING)
>  
> -#define task_is_traced(task)		((READ_ONCE(task->__state) & __TASK_TRACED) != 0)
> -
>  #define task_is_stopped(task)		((READ_ONCE(task->__state) & __TASK_STOPPED) != 0)
>  
> -#define task_is_stopped_or_traced(task)	((READ_ONCE(task->__state) & (__TASK_STOPPED | __TASK_TRACED)) != 0)
> -
>  /*
>   * Special states are those that do not use the normal wait-loop pattern. See
>   * the comment with set_special_state().

Urgh, so I have reworking all this somewhere on my todo list as well.
Except I mean to move it away from using p->__state entirely. We should
not be keeping canonical state in there.

As is, I think we can write task_is_stopped() like:

#define task_is_stopped(task)	((task)->jobctl & JOBCTL_STOP_PENDING)

Because jobctl is in fact the canonical state. I'm still not sure if we
can do the same with task_is_traced(), ideally that would be expressed
in terms of (task)->ptrace. But ptrace_stop() hurts my brain. All that
stuff is entirely to involved.

Anyway, let me see if I can page some of that back..