Re: [PATCH RESEND] mm: fix tick_sched timer blocked by pgdat_resize_lock

[Kirill Tkhai <ktkhai@virtuozzo.com>]

On 13.01.2020 03:54, Shile Zhang wrote:
> 
> 
> On 2020/1/10 19:42, Kirill Tkhai wrote:
>> On 10.01.2020 12:30, Shile Zhang wrote:
>>> When 'CONFIG_DEFERRED_STRUCT_PAGE_INIT' is set, 'pgdat_resize_lock'
>>> will be called inside 'pgdatinit' kthread to initialise the deferred
>>> pages with local interrupts disabled. Which is introduced by
>>> commit 3a2d7fa8a3d5 ("mm: disable interrupts while initializing deferred
>>> pages").
>>>
>>> But 'pgdatinit' kthread is possible be pined on the boot CPU (CPU#0 by
>>> default), especially in small system with NRCPUS <= 2. In this case, the
>>> interrupts are disabled on boot CPU during memory initialising, which
>>> caused the tick_sched timer be blocked, leading to wall clock stuck.
>>>
>>> Fixes: commit 3a2d7fa8a3d5 ("mm: disable interrupts while initializing
>>> deferred pages")
>>>
>>> Signed-off-by: Shile Zhang <shile.zhang@linux.alibaba.com>
>>> ---
>>>   include/linux/memory_hotplug.h | 16 ++++++++++++++--
>>>   1 file changed, 14 insertions(+), 2 deletions(-)
>>>
>>> diff --git a/include/linux/memory_hotplug.h b/include/linux/memory_hotplug.h
>>> index ba0dca6aac6e..be69a6dc4fee 100644
>>> --- a/include/linux/memory_hotplug.h
>>> +++ b/include/linux/memory_hotplug.h
>>> @@ -6,6 +6,8 @@
>>>   #include <linux/spinlock.h>
>>>   #include <linux/notifier.h>
>>>   #include <linux/bug.h>
>>> +#include <linux/sched.h>
>>> +#include <linux/smp.h>
>>>     struct page;
>>>   struct zone;
>>> @@ -282,12 +284,22 @@ static inline bool movable_node_is_enabled(void)
>>>   static inline
>>>   void pgdat_resize_lock(struct pglist_data *pgdat, unsigned long *flags)
>>>   {
>>> -    spin_lock_irqsave(&pgdat->node_size_lock, *flags);
>>> +    /*
>>> +     * Disable local interrupts on boot CPU will stop the tick_sched
>>> +     * timer, which will block jiffies(wall clock) update.
>>> +     */
>>> +    if (current->cpu != get_boot_cpu_id())
>>> +        spin_lock_irqsave(&pgdat->node_size_lock, *flags);
>>> +    else
>>> +        spin_lock(&pgdat->node_size_lock);
>>>   }
>>>   static inline
>>>   void pgdat_resize_unlock(struct pglist_data *pgdat, unsigned long *flags)
>>>   {
>>> -    spin_unlock_irqrestore(&pgdat->node_size_lock, *flags);
>>> +    if (current->cpu != get_boot_cpu_id())
>>> +        spin_unlock_irqrestore(&pgdat->node_size_lock, *flags);
>>> +    else
>>> +        spin_unlock(&pgdat->node_size_lock);
>>>   }
>>>   static inline
>>>   void pgdat_resize_init(struct pglist_data *pgdat)
>> 1)Linux kernel is *preemptible*. Kernel with CONFIG_PREEMPT_RT option even may preempt
>> *kernel* code in the middle of function. When you are executing a code containing
>> pgdat_resize_lock() and pgdat_resize_unlock(), the process may migrate to another cpu
>> between them.
>>
>> bool cpu               another cpu
>> ----------------------------------
>> pgdat_resize_lock()
>>    spin_lock()
>>    --> migrate to another cpu
>>                        pgdat_resize_unlock()
>>                        spin_unlock_irqrestore(<uninitialized flags>)
>>
>> (Yes, in case of CONFIG_PREEMPT_RT, process is preemptible even after spin_lock() call).
>>
>> This looks like a bad helpers, and we should not introduce such the design.
> 
> Hi Kirill,
> 
> Thanks for your comments!
> Sorry for I'm not very clear about this lock/unlock, but I encountered this issue
> with "CONFIG_PREEMPT is not set".

The thing is we simply shouldn't introduce such the primitives since the thread
may migrate to another cpu, while you own the lock. This looks like a buggy design.

>> 2)I think there is no the problem this patch solves. Do we really this statistics?
>> Can't we simple remove print message from deferred_init_memmap() and solve this?
> 
> Sorry for I've not put this issue very clearly. It's *not* just one statistics log
> with wrong time calculate, but the wall clock is stuck.
> So the 'systemd-analyze' command also give a wrong time as I mentioned in the cover
> letter. I don't think is OK just remove the log, it cannot solve the wall clock latency.

Have you tried temporary enabling interrupts in the middle of cycle after a huge enough
memory block is initialized? Something like:

deferred_init_memmap()
{
	while (spfn < epfn) {
		nr_pages += deferred_init_maxorder(&i, zone, &spfn, &epfn);
		local_irq_enable();
		local_irq_disable();
	}
}

Or, maybe, enable/disable interrupts somewhere inside deferred_init_maxorder().

>> Also, you may try to check that sched_clock() gives better results with interrupts
>> disabled (on x86 it uses rdtsc, when it's possible. But it also may fallback to
>> jiffies-based clock in some hardware cases, and they also won't go with interrupts
>> disabled).