Re: strange interaction between fuse + pidns
From: Eric W. Biederman
Date: Mon Jul 11 2022 - 17:38:08 EST
Tycho Andersen <firstname.lastname@example.org> writes:
> Hi all,
> On Mon, Jul 11, 2022 at 03:59:15PM +0200, Miklos Szeredi wrote:
>> On Mon, 11 Jul 2022 at 12:35, Miklos Szeredi <miklos@xxxxxxxxxx> wrote:
>> > Can you try the attached untested patch?
>> Updated patch to avoid use after free on req->args.
>> Still mostly untested.
> Thanks, when I applied your patch, I still ended up with tasks stuck
> waiting with a SIGKILL pending. So I looked into that and came up with
> the patch below. With both your patch and mine, my testcase exits
> Eric (or Christian, or anyone), can you comment on the patch below? I
> have no idea what this will break. Maybe instead a better approach is
> some additional special case in __send_signal_locked()?
> From b7ea26adcf3546be5745063cc86658acb5ed37e9 Mon Sep 17 00:00:00 2001
> From: Tycho Andersen <email@example.com>
> Date: Mon, 11 Jul 2022 11:26:58 -0600
> Subject: [PATCH] sched: __fatal_signal_pending() should also check shared
> The wait_* code uses signal_pending_state() to test whether a thread has
> been interrupted, which ultimately uses __fatal_signal_pending() to detect
> if there is a fatal signal.
> When a pid ns dies, in zap_pid_ns_processes() it does:
> group_send_sig_info(SIGKILL, SEND_SIG_PRIV, task, PIDTYPE_MAX);
> for all the tasks in the pid ns. That calls through:
> group_send_sig_info() ->
> do_send_sig_info() ->
> send_signal_locked() ->
> which does:
> pending = (type != PIDTYPE_PID) ? &t->signal->shared_pending : &t->pending;
> which puts sigkill in the set of shared signals, but not the individual
> pending ones. If tasks are stuck in a killable wait (e.g. a fuse flush
> operation), they won't see this shared signal, and will hang forever, since
> TIF_SIGPENDING is set, but the fatal signal can't be detected.
That is perplexing.
__send_signal_locked calls complete_signal. Then if any of the tasks of
the process can receive the signal, complete_signal will loop through
all of the tasks of the process and set the per thread SIGKILL. Pretty
much by definition tasks can always receive SIGKILL.
Is complete_signal not being able to do that?
The patch below really should not be necessary, and I have pending work
that if I can push over the finish line won't even make sense.
As it is currently an abuse to use the per thread SIGKILL to indicate
that a fatal signal has been short circuit delivered. That abuse as
well as being unclean tends to confuse people reading the code.
> Signed-off-by: Tycho Andersen <firstname.lastname@example.org>
> include/linux/sched/signal.h | 3 ++-
> 1 file changed, 2 insertions(+), 1 deletion(-)
> diff --git a/include/linux/sched/signal.h b/include/linux/sched/signal.h
> index cafbe03eed01..a033ccb0a729 100644
> --- a/include/linux/sched/signal.h
> +++ b/include/linux/sched/signal.h
> @@ -402,7 +402,8 @@ static inline int signal_pending(struct task_struct *p)
> static inline int __fatal_signal_pending(struct task_struct *p)
> - return unlikely(sigismember(&p->pending.signal, SIGKILL));
> + return unlikely(sigismember(&p->pending.signal, SIGKILL) ||
> + sigismember(&p->signal->shared_pending.signal, SIGKILL));
> static inline int fatal_signal_pending(struct task_struct *p)
> base-commit: 32346491ddf24599decca06190ebca03ff9de7f8