Re: [PATCH] mm: use this_cpu_cmpxchg_double in put_cpu_partial

[Wengang Wang <wen.gang.wang@oracle.com>]

On 2018/11/26 16:36, Wei Yang wrote:
> On Mon, Nov 26, 2018 at 08:57:54AM -0800, Wengang Wang wrote:
>>
>> On 2018/11/25 17:59, Wei Yang wrote:
>>> On Tue, Nov 20, 2018 at 10:58 AM zhong jiang <zhongjiang@huawei.com> wrote:
>>>> On 2018/11/17 9:33, Wengang Wang wrote:
>>>>> The this_cpu_cmpxchg makes the do-while loop pass as long as the
>>>>> s->cpu_slab->partial as the same value. It doesn't care what happened to
>>>>> that slab. Interrupt is not disabled, and new alloc/free can happen in the
>>>>> interrupt handlers. Theoretically, after we have a reference to the it,
>>>>> stored in _oldpage_, the first slab on the partial list on this CPU can be
>>>>> moved to kmem_cache_node and then moved to different kmem_cache_cpu and
>>>>> then somehow can be added back as head to partial list of current
>>>>> kmem_cache_cpu, though that is a very rare case. If that rare case really
>>>>> happened, the reading of oldpage->pobjects may get a 0xdead0000
>>>>> unexpectedly, stored in _pobjects_, if the reading happens just after
>>>>> another CPU removed the slab from kmem_cache_node, setting lru.prev to
>>>>> LIST_POISON2 (0xdead000000000200). The wrong _pobjects_(negative) then
>>>>> prevents slabs from being moved to kmem_cache_node and being finally freed.
>>>>>
>>>>> We see in a vmcore, there are 375210 slabs kept in the partial list of one
>>>>> kmem_cache_cpu, but only 305 in-use objects in the same list for
>>>>> kmalloc-2048 cache. We see negative values for page.pobjects, the last page
>>>>> with negative _pobjects_ has the value of 0xdead0004, the next page looks
>>>>> good (_pobjects is 1).
>>>>>
>>>>> For the fix, I wanted to call this_cpu_cmpxchg_double with
>>>>> oldpage->pobjects, but failed due to size difference between
>>>>> oldpage->pobjects and cpu_slab->partial. So I changed to call
>>>>> this_cpu_cmpxchg_double with _tid_. I don't really want no alloc/free
>>>>> happen in between, but just want to make sure the first slab did expereince
>>>>> a remove and re-add. This patch is more to call for ideas.
>>>> Have you hit the really issue or just review the code ?
>>>>
>>>> I did hit the issue and fixed in the upstream patch unpredictably by the following patch.
>>>> e5d9998f3e09 ("slub: make ->cpu_partial unsigned int")
>>>>
>>> Zhong,
>>>
>>> I took a look into your upstream patch, while I am confused how your patch
>>> fix this issue?
>>>
>>> In put_cpu_partial(), the cmpxchg compare cpu_slab->partial (a page struct)
>>> instead of the cpu_partial (an unsigned integer). I didn't get the
>>> point of this fix.
>> I think the patch can't prevent pobjects from being set as 0xdead0000 (the
>> primary 4 bytes of LIST_POISON2).
>> But if pobjects is treated as unsigned integer,
>>
>> 2266???????????????????????????????????????????????? pobjects = oldpage->pobjects;
>> 2267???????????????????????????????????????????????? pages = oldpage->pages;
>> 2268???????????????????????????????????????????????? if (drain && pobjects > s->cpu_partial) {
>> 2269???????????????????????????????????????????????????????????????? unsigned long flags;
>>
> Ehh..., you mean (0xdead0000 > 0x02) ?
Yes.

> This is really a bad thing, if it wordarounds the problem like this.
It does.
> I strongly don't agree this is a *fix*. This is too tricky.
It's tricky. I don't know for what purpose the patch went in. The commit 
message was just saying _pobjects_ should be "unsigned"...

thanks,
wengang

>> line 2268 will be true in put_cpu_partial(), thus code goes to
>> unfreeze_partials(). This way the slabs in the cpu partial list can be moved
>> to kmem_cache_nod and then freed. So it fixes (or say workarounds) the
>> problem I see here (huge number of empty slabs stay in cpu partial list).
>>
>> thanks
>> wengang
>>
>>>> Thanks,
>>>> zhong jiang