Re: Making rcu_normal=1 in RT

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Wed, Nov 02, 2016 at 05:05:47PM +0100, Sebastian Andrzej Siewior wrote:
> On 2016-11-01 05:31:56 [+0200], Michael S. Tsirkin wrote:
> > > > > > > 
> > > > > > > > - optionally it has been requested to make synchronize_rcu() behave like
> > > > > > > >   synchronize_rcu_expedited() on shutdown and kexec().
> > 
> > I was trying to say that the fact net core calls expedited sync
> > implies that many people care about boot time.
> > 
> > Should rcu_normal_after_boot override rcu_expedited?
> > I'm not sure what happens if you specify both ATM.
> 
> I think what you asked for is something like this:
> 
> diff --git a/kernel/reboot.c b/kernel/reboot.c
> --- a/kernel/reboot.c
> +++ b/kernel/reboot.c
> @@ -283,6 +283,7 @@ SYSCALL_DEFINE4(reboot, int, magic1, int, magic2, unsigned int, cmd,
>  	struct pid_namespace *pid_ns = task_active_pid_ns(current);
>  	char buffer[256];
>  	int ret = 0;
> +	bool force_rcu_exp = false;
> 
>  	/* We only trust the superuser with rebooting the system. */
>  	if (!ns_capable(pid_ns->user_ns, CAP_SYS_BOOT))
> @@ -311,7 +312,19 @@ SYSCALL_DEFINE4(reboot, int, magic1, int, magic2, unsigned int, cmd,
>  	if ((cmd == LINUX_REBOOT_CMD_POWER_OFF) && !pm_power_off)
>  		cmd = LINUX_REBOOT_CMD_HALT;
> 
> +	switch (cmd) {
> +	case LINUX_REBOOT_CMD_RESTART:
> +	case LINUX_REBOOT_CMD_HALT:
> +	case LINUX_REBOOT_CMD_POWER_OFF:
> +	case LINUX_REBOOT_CMD_RESTART2:
> +	case LINUX_REBOOT_CMD_KEXEC:
> +		force_rcu_exp = true;
> +	}
> +
>  	mutex_lock(&reboot_mutex);
> +	if (force_rcu_exp)
> +		rcu_expedite_gp();
> +
>  	switch (cmd) {
>  	case LINUX_REBOOT_CMD_RESTART:
>  		kernel_restart(NULL);
> @@ -362,6 +375,8 @@ SYSCALL_DEFINE4(reboot, int, magic1, int, magic2, unsigned int, cmd,
>  		ret = -EINVAL;
>  		break;
>  	}
> +	if (force_rcu_exp)
> +		rcu_unexpedite_gp();
>  	mutex_unlock(&reboot_mutex);
>  	return ret;
>  }
> 
> and problem that this patch is pointless because it is too late. This
> RESTART, POWER_OFF, KEXEC hooks here called once the `shutdown' binary
> is done doing sync & cleanup and invokes the kernel for the final
> operation.
> So to achieve what you ask for, we would need a
> LINUX_REBOOT_CMD_RCU_FAST option which is invoked by the `shutdown'
> binary as the first thing - long before it gets to
> LINUX_REBOOT_CMD_RESTART (or KEXEC).

FWIW, this looks sane to me.

							Thanx, Paul