Re: [RFC PATCH] mm: kasan: suppress soft lockup in slub when !CONFIG_PREEMPT

[Andrey Ryabinin]

On 12/08/2017 11:26 AM, Dmitry Vyukov wrote:
> On Fri, Dec 8, 2017 at 12:40 AM, Matthew Wilcox <willy@xxxxxxxxxxxxx> wrote:
>> On Fri, Dec 08, 2017 at 07:30:07AM +0800, Yang Shi wrote:
>>> When running stress test with KASAN enabled, the below softlockup may
>>> happen occasionally:
>>>
>>> NMI watchdog: BUG: soft lockup - CPU#7 stuck for 22s!
>>> hardirqs last  enabled at (0): [<          (null)>]      (null)
>>> hardirqs last disabled at (0): [] copy_process.part.30+0x5c6/0x1f50
>>> softirqs last  enabled at (0): [] copy_process.part.30+0x5c6/0x1f50
>>> softirqs last disabled at (0): [<          (null)>]      (null)
>>
>>> Call Trace:
>>>  [] __slab_free+0x19c/0x270
>>>  [] ___cache_free+0xa6/0xb0
>>>  [] qlist_free_all+0x47/0x80
>>>  [] quarantine_reduce+0x159/0x190
>>>  [] kasan_kmalloc+0xaf/0xc0
>>>  [] kasan_slab_alloc+0x12/0x20
>>>  [] kmem_cache_alloc+0xfa/0x360
>>>  [] ? getname_flags+0x4f/0x1f0
>>>  [] getname_flags+0x4f/0x1f0
>>>  [] getname+0x12/0x20
>>>  [] do_sys_open+0xf9/0x210
>>>  [] SyS_open+0x1e/0x20
>>>  [] entry_SYSCALL_64_fastpath+0x1f/0xc2
>>
>> This feels like papering over a problem.  KASAN only calls
>> quarantine_reduce() when it's allowed to block.  Presumably it has
>> millions of entries on the free list at this point.  I think the right
>> thing to do is for qlist_free_all() to call cond_resched() after freeing
>> every N items.
> 
> 
> Agree. Adding touch_softlockup_watchdog() to a random low-level
> function looks like a wrong thing to do.
> quarantine_reduce() already has this logic. Look at
> QUARANTINE_BATCHES. It's meant to do exactly this -- limit amount of
> work in quarantine_reduce() and in quarantine_remove_cache() to
> reasonably-sized batches. We could simply increase number of batches
> to make them smaller. But it would be good to understand what exactly
> happens in this case. Batches should on a par of ~~1MB. Why freeing
> 1MB worth of objects (smallest of which is 32b) takes 22 seconds?
> 

I think the problem here is that kernel 4.9.44-003.ali3000.alios7.x86_64.debug
doesn't have 64abdcb24351 ("kasan: eliminate long stalls during quarantine reduction").

We probably should ask that commit to be included in stable, but it would be good to hear
a confirmation from Yang that it really helps.