Re: [PATCH 0/3] x86: fix hang when AP bringup is too slow

[Prarit Bhargava <prarit@redhat.com>]

On 03/19/2014 08:54 AM, Igor Mammedov wrote:
> On Wed, 19 Mar 2014 07:51:05 -0400
> Prarit Bhargava <prarit@redhat.com> wrote:
> 
>>
>>
>> On 03/18/2014 02:49 PM, Igor Mammedov wrote:
>>> On Tue, 18 Mar 2014 08:21:19 -0400
>>> Prarit Bhargava <prarit@redhat.com> wrote:
>>>
>>>>
>>>>
>>>> On 03/13/2014 10:25 AM, Igor Mammedov wrote:
>>>>> Hang is observed on virtual machines during CPU hotplug,
>>>>> especially in big guests with many CPUs. (It happens more
>>>>> often if host is over-committed).
>>>>>
>>>>
>>>> Hey Igor, I like this better than the previous version.  Thanks for taking into
>>>> account the possible races in this code.
>>>>
>>>> A quick question on system behaviour.  As you know I've been more concerned
>>>> lately with error handling, etc., through the cpu hotplug code as we've seen
>>>> several customer reports of silent failures or cascading failures in the cpu
>>>> hotplug code when users have been attempting to perform physical hotplug.
>>>>
>>>> After your patches have been applied, in theory the following can happen:
>>>>
>>>> The master CPU is completing the AP cpu's bring up.  The AP cpu is doing (sorry
>>>> for the cut-and-paste),
>>>>
>>>> void cpu_init(void)
>>>> {
>>>>         int cpu = smp_processor_id();
>>>>         struct task_struct *curr = current;
>>>>         struct tss_struct *t = &per_cpu(init_tss, cpu);
>>>>         struct thread_struct *thread = &curr->thread;
>>>>
>>>>         /*
>>>>          * wait till the master CPU completes it's STARTUP sequence,
>>>>          * and decides to wait till this AP boots
>>>>          */
>>>>         while (!cpumask_test_cpu(cpu, cpu_callout_mask)) {
>>>>                 cpu_relax();
>>>>                 if (per_cpu(x86_cpu_to_apicid, cpu) == BAD_APICID)
>>>>                         halt();
>>>>         }
>>>>
>>>> and is spinning on cpu_relax().  Suppose something goes wrong and the softlockup
>>>> watchdog fires on the AP cpu:
>>>>
>>>> 1.  Can it? :) ie) will the softlockup fire at this point of the AP init?  Okay,
>>>> I'm being really lazy and not looking at the code ;)
>>> It shouldn't, CPU is in pristine state and just came from boot trampoline at
>>> this point without interrupts configured yet.
>>
>> Okay, not a big problem.
>>
>>>
>>>>
>>>> 2.  Is there anything we can do in this code to notify the user of a problem?
>>>> Even a pr_crit() here I think would help to indicate what went wrong; it might
>>>> be useful for future debugging in this area to have some sort of output.  I
>>>> think a WARN() or BUG() is necessary here as there are several calls to cpu_init().
>>> Do you mean something like this:
>>>
>>> +		if (per_cpu(x86_cpu_to_apicid, cpu) == BAD_APICID) {
>>> +                       WARN(1);
>>> +			halt();
>>> +               }
>>
>> Yeah, maybe WARN_ON(1, "some comment") though.
> printk at so early stage might be cause issues, since it is quite complex.
> Its' disabling/enabling irqs, calls *_delay_*() functions and takes locks.
> The last is especially dangerous because if AP is shot down by another
> INIT/SIPI, system will hang on next printk if locks were acquired by AP
> at that time.

early_printk()?

> That case is possible if master CPU has got errors during wakeup_ap() and
> failed cpu_up() then it was unplugged + plugged via ACPI and attempted
> to be onlined again. 
> 
> It's much safer not to do anything complex at AP start-up so early.
> 
> BTW:
> when AP reaches halt() line, failure is not silent. the master CPU might
> print error message if debug level logging is active:
> see arch/x86/kernel/smpboot.c:native_cpu_up()
> ...
>         if (err) {
>                 pr_debug("do_boot_cpu failed %d\n", err);
>                 return -EIO;
>         }
> ...
> 
> perhaps we should change pr_debug to pr_crit here to make it more visible.
> something like:
> 
> @@ -858,7 +858,7 @@ int native_cpu_up(unsigned int cpu, struct task_struct *tidle)
>  
>         err = do_boot_cpu(apicid, cpu, tidle);
>         if (err) {
> -               pr_debug("do_boot_cpu failed %d\n", err);
> +               pr_crit("do_boot_cpu failed(%d) to wakeup CPU#%u\n", err, cpu);
>                 return -EIO;
>         }
> 

Yes, this is a good idea.

> 
>>
>>>
>>>>
>>>> 3.  Change this comment:
>>>>
>>>>          * wait till the master CPU completes it's STARTUP sequence,
>>>>          * and decides to wait till this AP boots
>>>>
>>>> to
>>>>
>>>> 	/* wait for the master CPU to complete this cpu's STARTUP. */ ?
>>> well, that is not quite the same as above, comment should underline that
>>> AP waits for ACK from master CPU before continuing with this AP initialization.
>>>
>>> How about:
>>> /* wait for ACK from master CPU before continuing with AP initialization */
>>
>> Awesome :)
>>
>> P.
>>
>>>
>>>>
>>>> Apologies for the late review,
>>>>
>>>> P.
>>>
>>>
>