Re: [PATCH -rt] rt-slab: fix cpu inconsistency case

[Hiroshi Shimamoto <h-shimamoto@ct.jp.nec.com>]

Steven Rostedt wrote:
> 
> On Thu, 20 Mar 2008, Hiroshi Shimamoto wrote:
> 
>> mm/slab.c:cache_alloc_refill()
>> 	if (unlikely(!ac->avail)) {
>> 		int x;
>> 		x = cache_grow(cachep, flags | GFP_THISNODE, cpu_to_node(*this_cpu), NULL, this_cpu);
>>
>> 		/* cache_grow can reenable interrupts, then ac could change. */
>> 		ac = cpu_cache_get(cachep, *this_cpu);
>>
>> The comment says, "ac could change", but it never if *this_cpu is same.
>>
>> In cache_alloc_refill(), cpu_cache_get() should called with the valid cpu id
>> after cache_grow(). Because on !PREEMPT_RT, the array_cache is protected by
>> disabling irqs, so array_cache of other cpu shouldn't be accessed.
> 
> Ah, sorry I missed the !PREEMPT_RT part.

no problem. I also missed to say clearly this issue is on !PREEMPT_RT :-)

> 
>>>>  # define slab_irq_disable_rt(flags)	do { (void)(flags); } while (0)
>>>>  # define slab_irq_enable_rt(flags)	do { (void)(flags); } while (0)
>>>>  # define slab_spin_lock_irq(lock, cpu) \
>>>> @@ -160,8 +160,8 @@ DEFINE_PER_CPU_LOCKED(int, slab_irq_locks) = { 0, };
>>>>  	do { slab_irq_enable(cpu); (void) (flags); } while (0)
>>>>  # define slab_irq_disable_rt(cpu)	slab_irq_disable(cpu)
>>>>  # define slab_irq_enable_rt(cpu)	slab_irq_enable(cpu)
>>>> -# define slab_irq_disable_nort()	do { } while (0)
>>>> -# define slab_irq_enable_nort()		do { } while (0)
>>>> +# define slab_irq_disable_nort(cpu)	do { } while (0)
>>>> +# define slab_irq_enable_nort(cpu)	do { } while (0)
>>> And these are the PREEMPT_RT version. So basically, this patch is a nop
>>> for PREEMPT_RT. I doubt it will solve your bug ;-)
>> yes, it's nop. I'm sorry, I didn't show .config.
>> My kernel config is !PREEMPT_RT, it's same as the deadlock report.
>> I guess it's a problem only !PREEMPT_RT.
> 
> Hmm, I'll have to look deeper into this on Monday (I'm off today -
> Friday).

thanks for taking time.

> 
> Is this a bug in mainline? The !PREEMPT_RT case should be as close to
> mainline as possible, with no actual changes in object code. If this is
> not the case, then we need to fix that.

No, this issue is in -rt only.

thanks,
Hiroshi Shimamoto