Re: [patch] fix uml slowness caused by ptrace preemption bug on host

[Oleg Nesterov <oleg@redhat.com>]

On 03/19, Roland McGrath wrote:
>
> I'm no scheduler expert and I don't know whether the exact placement in
> your change is the optimal one.

Agreed, can't we do a bit more simple patch?

	--- kernel/signal.c
	+++ kernel/signal.c
	@@ -1572,8 +1572,10 @@ static void ptrace_stop(int exit_code, i
		spin_unlock_irq(&current->sighand->siglock);
		read_lock(&tasklist_lock);
		if (may_ptrace_stop()) {
	+		preempt_disable();
			do_notify_parent_cldstop(current, CLD_TRAPPED);
			read_unlock(&tasklist_lock);
	+		preempt_enable_no_resched();
			schedule();
		} else {
			/*

Yes, the task can be preempted right after spin_unlock(->siglock), but
this is unlikely. We need the "synchronous" wakeup, and this patch helps
as well.



Actually, I don't know which ptrace requests really need to make sure
the tracee was deactivated. Perhaps they can call wait_task_inactive()
themselves? I guess this is bad idea, but most of requests definitely
do not need wait_task_inactive().

Oleg.