On Fri, Oct 24, 2008 at 11:33:42AM +0200, Oleg Nesterov wrote:
quoted text
> On 10/23, Cyrill Gorcunov wrote:
> >
> > [Gautham R Shenoy - Thu, Oct 23, 2008 at 11:51:19PM +0530]
> > | On Thu, Oct 23, 2008 at 07:02:12PM +0200, Oleg Nesterov wrote:
> > | > Hmm. arch/x86/kernel/smpboot.c:do_boot_cpu() can deadlock ?
> > | > 
> > | > It is called from _cpu_up() under cpu_hotplug_begin(), and it
> > | > waits for c_idle.work. Again, if we have the pending work which
> > | > needs get_online_cpus() we seem to have problems.
> > | 
> > | Good point. Though this code gets triggered mostly during boot time when
> > | the CPUs are being brought online for the first time. If we have some
> > | work-item pending at that time, which needs get_online_cpus(), we could
> > | possibly see this deadlock.
> > | 
> > | > 
> > | > Oleg.
> > | 
> > | -- 
> > | Thanks and Regards
> > | gautham
> > | 
> > 
> > May I ask? If I understand right we do use this part of do_boot_cpu
> > 
> > 	if (!keventd_up() || current_is_keventd())
> > 		c_idle.work.func(&c_idle.work);
> > 	else {
> > 		schedule_work(&c_idle.work);
> > 		wait_for_completion(&c_idle.done);
> > 	}
> > 
> > if only we've been called the first time after power on. And all
> > subsequent call of this do_boot_cpu would lead to
> > 
> > 	if (c_idle.idle) {
> > 		c_idle.idle->thread.sp = (unsigned long) (((struct pt_regs *)
> > 			(THREAD_SIZE +  task_stack_page(c_idle.idle))) - 1);
> > 		init_idle(c_idle.idle, cpu);
> > 		goto do_rest;
> > 	}
> > 
> > ie go to do_rest and no wait_for_completion/schedule_work at all.
> > Did I miss something? *Sorry* in advance if the question is quite
> > not related. This work-pending situation is in 'possible' scenario
> > only (ie we don't have such a callers for now... yet)?
> 
> There are no problems during boot time, afaics.
> 
> kernel_init() calls smp_init() before do_basic_setup()->init_workqueues().
> This means that do_boot_cpu() won't use workqueues due to !keventd_up().
> 
> But let's suppose we boot with maxcpus=1, and then bring up another CPU.
> Or we really add the new physical CPU (I don't really know if this is
> possible on x86).

Even I am not sure if physical hotplug is possible.

But think about the virtualization case when we want to add
additional CPU to a KVM guest at runtime? This
is not such a rare use-case. It could dead-lock that time, No?

quoted text
> 
> Oleg.

-- 
Thanks and Regards
gautham
--
unsubscribe notice
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/