Virtualization - Dec 2021 - [PATCH V6 05/10] signal: Perfom autoreap for PF_USER_WORKER

Userspace doesn't know about PF_USER_WORKER threads, so it can't do wait
to clean them up. For cases like where qemu will do dynamic/hot add/remove
of vhost devices, then we need to auto reap the thread like was done for
the kthread case, because qemu does not know what API the kernel/vhost
layer is using.

This has us do autoreaping for these threads similar to when the parent
ignores SIGCHLD and for kthreads.

Signed-off-by: Mike Christie <michael.christie at oracle.com>
---
 kernel/signal.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/kernel/signal.c b/kernel/signal.c
index a629b11bf3e0..4ce2cc195269 100644
--- a/kernel/signal.c
+++ b/kernel/signal.c
@@ -2071,9 +2071,9 @@ bool do_notify_parent(struct task_struct *tsk, int sig)
 
 	psig = tsk->parent->sighand;
 	spin_lock_irqsave(&psig->siglock, flags);
-	if (!tsk->ptrace && sig == SIGCHLD &&
+	if (!tsk->ptrace && (tsk->flags & PF_USER_WORKER || (sig ==
SIGCHLD &&
 	    (psig->action[SIGCHLD-1].sa.sa_handler == SIG_IGN ||
-	     (psig->action[SIGCHLD-1].sa.sa_flags & SA_NOCLDWAIT))) {
+	     (psig->action[SIGCHLD-1].sa.sa_flags & SA_NOCLDWAIT))))) {
 		/*
 		 * We are exiting and our parent doesn't care.  POSIX.1
 		 * defines special semantics for setting SIGCHLD to SIG_IGN
-- 
2.25.1

Mike Christie <michael.christie at oracle.com> writes:
> Userspace doesn't know about PF_USER_WORKER threads, so it can't do
wait
> to clean them up. For cases like where qemu will do dynamic/hot add/remove
> of vhost devices, then we need to auto reap the thread like was done for
> the kthread case, because qemu does not know what API the kernel/vhost
> layer is using.
>
> This has us do autoreaping for these threads similar to when the parent
> ignores SIGCHLD and for kthreads.
There is a lot wrong with this change.
1) you can just set "task->signal = SIGCHLD" to get this
   behavior so it is unnecessary.

2) This is not the autoreaping you want.  This autoreaping just kicks
   in when the parents signal handler is SIG_IGN.  Since I presume
   you are not controlling the parent this is just plain nonsense.

The autoreap you want is the autoreap in exit_notify, and you don't
want to call do_notify_parent at all.

Eric
> Signed-off-by: Mike Christie <michael.christie at oracle.com>
> ---
>  kernel/signal.c | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
>
> diff --git a/kernel/signal.c b/kernel/signal.c
> index a629b11bf3e0..4ce2cc195269 100644
> --- a/kernel/signal.c
> +++ b/kernel/signal.c
> @@ -2071,9 +2071,9 @@ bool do_notify_parent(struct task_struct *tsk, int
sig)
>  
>  	psig = tsk->parent->sighand;
>  	spin_lock_irqsave(&psig->siglock, flags);
> -	if (!tsk->ptrace && sig == SIGCHLD &&
> +	if (!tsk->ptrace && (tsk->flags & PF_USER_WORKER ||
(sig == SIGCHLD &&
>  	    (psig->action[SIGCHLD-1].sa.sa_handler == SIG_IGN ||
> -	     (psig->action[SIGCHLD-1].sa.sa_flags & SA_NOCLDWAIT))) {
> +	     (psig->action[SIGCHLD-1].sa.sa_flags & SA_NOCLDWAIT))))) {
>  		/*
>  		 * We are exiting and our parent doesn't care.  POSIX.1
>  		 * defines special semantics for setting SIGCHLD to SIG_IGN